/* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ #define MDBX_ALLOY 1 #define MDBX_BUILD_SOURCERY 37ec6726f69ef69c4e52fcfb87d700f44f37246b1b635f455d2a429776a8eec6_v0_9_2_136_g9f0ff86 #ifdef MDBX_CONFIG_H #include MDBX_CONFIG_H #endif #define LIBMDBX_INTERNALS #ifdef MDBX_TOOLS #define MDBX_DEPRECATED #endif /* MDBX_TOOLS */ /* *INDENT-OFF* */ /* clang-format off */ /* In case the MDBX_DEBUG is undefined set it corresponding to NDEBUG */ #ifndef MDBX_DEBUG # ifdef NDEBUG # define MDBX_DEBUG 0 # else # define MDBX_DEBUG 1 # endif #endif /* Undefine the NDEBUG if debugging is enforced by MDBX_DEBUG */ #if MDBX_DEBUG # undef NDEBUG #endif #ifdef MDBX_ALLOY /* Amalgamated build */ # define MDBX_INTERNAL_FUNC static # define MDBX_INTERNAL_VAR static #else /* Non-amalgamated build */ # define MDBX_INTERNAL_FUNC # define MDBX_INTERNAL_VAR extern #endif /* MDBX_ALLOY */ #ifndef MDBX_DISABLE_GNU_SOURCE #define MDBX_DISABLE_GNU_SOURCE 0 #endif #if MDBX_DISABLE_GNU_SOURCE #undef _GNU_SOURCE #elif (defined(__linux__) || defined(__gnu_linux__)) && !defined(_GNU_SOURCE) #define _GNU_SOURCE #endif /*----------------------------------------------------------------------------*/ /* Should be defined before any includes */ #ifndef _FILE_OFFSET_BITS # define _FILE_OFFSET_BITS 64 #endif #ifdef __APPLE__ #define _DARWIN_C_SOURCE #endif #ifdef _MSC_VER # if _MSC_FULL_VER < 190024234 /* Actually libmdbx was not tested with compilers older than 19.00.24234 (Visual Studio 2015 Update 3). * But you could remove this #error and try to continue at your own risk. * In such case please don't rise up an issues related ONLY to old compilers. */ # error "At least \"Microsoft C/C++ Compiler\" version 19.00.24234 (Visual Studio 2015 Update 3) is required." # endif # ifndef _CRT_SECURE_NO_WARNINGS # define _CRT_SECURE_NO_WARNINGS # endif #if _MSC_VER > 1800 # pragma warning(disable : 4464) /* relative include path contains '..' */ #endif #if _MSC_VER > 1913 # pragma warning(disable : 5045) /* Compiler will insert Spectre mitigation... */ #endif #pragma warning(disable : 4710) /* 'xyz': function not inlined */ #pragma warning(disable : 4711) /* function 'xyz' selected for automatic inline expansion */ #pragma warning(disable : 4201) /* nonstandard extension used : nameless struct / union */ #pragma warning(disable : 4702) /* unreachable code */ #pragma warning(disable : 4706) /* assignment within conditional expression */ #pragma warning(disable : 4127) /* conditional expression is constant */ #pragma warning(disable : 4324) /* 'xyz': structure was padded due to alignment specifier */ #pragma warning(disable : 4310) /* cast truncates constant value */ #pragma warning(disable : 4820) /* bytes padding added after data member for alignment */ #pragma warning(disable : 4548) /* expression before comma has no effect; expected expression with side - effect */ #pragma warning(disable : 4366) /* the result of the unary '&' operator may be unaligned */ #pragma warning(disable : 4200) /* nonstandard extension used: zero-sized array in struct/union */ #pragma warning(disable : 4204) /* nonstandard extension used: non-constant aggregate initializer */ #pragma warning(disable : 4505) /* unreferenced local function has been removed */ #endif /* _MSC_VER (warnings) */ #include "mdbx.h" /* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ /* *INDENT-OFF* */ /* clang-format off */ #ifndef __GNUC_PREREQ # if defined(__GNUC__) && defined(__GNUC_MINOR__) # define __GNUC_PREREQ(maj, min) \ ((__GNUC__ << 16) + __GNUC_MINOR__ >= ((maj) << 16) + (min)) # else # define __GNUC_PREREQ(maj, min) (0) # endif #endif /* __GNUC_PREREQ */ #ifndef __CLANG_PREREQ # ifdef __clang__ # define __CLANG_PREREQ(maj,min) \ ((__clang_major__ << 16) + __clang_minor__ >= ((maj) << 16) + (min)) # else # define __CLANG_PREREQ(maj,min) (0) # endif #endif /* __CLANG_PREREQ */ #ifndef __GLIBC_PREREQ # if defined(__GLIBC__) && defined(__GLIBC_MINOR__) # define __GLIBC_PREREQ(maj, min) \ ((__GLIBC__ << 16) + __GLIBC_MINOR__ >= ((maj) << 16) + (min)) # else # define __GLIBC_PREREQ(maj, min) (0) # endif #endif /* __GLIBC_PREREQ */ #ifndef __has_warning # define __has_warning(x) (0) #endif #ifndef __has_include # define __has_include(x) (0) #endif #if __has_feature(thread_sanitizer) # define __SANITIZE_THREAD__ 1 #endif #if __has_feature(address_sanitizer) # define __SANITIZE_ADDRESS__ 1 #endif /*----------------------------------------------------------------------------*/ #ifndef __extern_C # ifdef __cplusplus # define __extern_C extern "C" # else # define __extern_C # endif #endif /* __extern_C */ #if !defined(nullptr) && !defined(__cplusplus) || (__cplusplus < 201103L && !defined(_MSC_VER)) # define nullptr NULL #endif /*----------------------------------------------------------------------------*/ #ifndef __always_inline # if defined(__GNUC__) || __has_attribute(__always_inline__) # define __always_inline __inline __attribute__((__always_inline__)) # elif defined(_MSC_VER) # define __always_inline __forceinline # else # define __always_inline # endif #endif /* __always_inline */ #ifndef __noinline # if defined(__GNUC__) || __has_attribute(__noinline__) # define __noinline __attribute__((__noinline__)) # elif defined(_MSC_VER) # define __noinline __declspec(noinline) # else # define __noinline # endif #endif /* __noinline */ #ifndef __must_check_result # if defined(__GNUC__) || __has_attribute(__warn_unused_result__) # define __must_check_result __attribute__((__warn_unused_result__)) # else # define __must_check_result # endif #endif /* __must_check_result */ #ifndef __maybe_unused # if defined(__GNUC__) || __has_attribute(__unused__) # define __maybe_unused __attribute__((__unused__)) # else # define __maybe_unused # endif #endif /* __maybe_unused */ #if !defined(__noop) && !defined(_MSC_VER) # define __noop(...) do {} while(0) #endif /* __noop */ #ifndef __fallthrough # if defined(__cplusplus) && (__has_cpp_attribute(fallthrough) && \ (!defined(__clang__) || __clang__ > 4)) || __cplusplus >= 201703L # define __fallthrough [[fallthrough]] # elif __GNUC_PREREQ(8, 0) && defined(__cplusplus) && __cplusplus >= 201103L # define __fallthrough [[fallthrough]] # elif __GNUC_PREREQ(7, 0) && \ (!defined(__LCC__) || (__LCC__ == 124 && __LCC_MINOR__ >= 12) || \ (__LCC__ == 125 && __LCC_MINOR__ >= 5) || (__LCC__ >= 126)) # define __fallthrough __attribute__((__fallthrough__)) # elif defined(__clang__) && defined(__cplusplus) && __cplusplus >= 201103L &&\ __has_feature(cxx_attributes) && __has_warning("-Wimplicit-fallthrough") # define __fallthrough [[clang::fallthrough]] # else # define __fallthrough # endif #endif /* __fallthrough */ #ifndef __unreachable # if __GNUC_PREREQ(4,5) || __has_builtin(__builtin_unreachable) # define __unreachable() __builtin_unreachable() # elif defined(_MSC_VER) # define __unreachable() __assume(0) # else # define __unreachable() __noop() # endif #endif /* __unreachable */ #ifndef __prefetch # if defined(__GNUC__) || defined(__clang__) || __has_builtin(__builtin_prefetch) # define __prefetch(ptr) __builtin_prefetch(ptr) # else # define __prefetch(ptr) __noop(ptr) # endif #endif /* __prefetch */ #ifndef __nothrow # if defined(__cplusplus) # if __cplusplus < 201703L # define __nothrow throw() # else # define __nothrow noexcept(true) # endif /* __cplusplus */ # elif defined(__GNUC__) || __has_attribute(__nothrow__) # define __nothrow __attribute__((__nothrow__)) # elif defined(_MSC_VER) && defined(__cplusplus) # define __nothrow __declspec(nothrow) # else # define __nothrow # endif #endif /* __nothrow */ #ifndef __hidden # if defined(__GNUC__) || __has_attribute(__visibility__) # define __hidden __attribute__((__visibility__("hidden"))) # else # define __hidden # endif #endif /* __hidden */ #ifndef __optimize # if defined(__OPTIMIZE__) # if (defined(__GNUC__) && !defined(__clang__)) || __has_attribute(__optimize__) # define __optimize(ops) __attribute__((__optimize__(ops))) # else # define __optimize(ops) # endif # else # define __optimize(ops) # endif #endif /* __optimize */ #ifndef __hot # if defined(__OPTIMIZE__) # if defined(__e2k__) # define __hot __attribute__((__hot__)) __optimize(3) # elif defined(__clang__) && !__has_attribute(__hot_) \ && __has_attribute(__section__) && (defined(__linux__) || defined(__gnu_linux__)) /* just put frequently used functions in separate section */ # define __hot __attribute__((__section__("text.hot"))) __optimize("O3") # elif defined(__GNUC__) || __has_attribute(__hot__) # define __hot __attribute__((__hot__)) __optimize("O3") # else # define __hot __optimize("O3") # endif # else # define __hot # endif #endif /* __hot */ #ifndef __cold # if defined(__OPTIMIZE__) # if defined(__e2k__) # define __cold __attribute__((__cold__)) __optimize(1) # elif defined(__clang__) && !__has_attribute(cold) \ && __has_attribute(__section__) && (defined(__linux__) || defined(__gnu_linux__)) /* just put infrequently used functions in separate section */ # define __cold __attribute__((__section__("text.unlikely"))) __optimize("Os") # elif defined(__GNUC__) || __has_attribute(cold) # define __cold __attribute__((__cold__)) __optimize("Os") # else # define __cold __optimize("Os") # endif # else # define __cold # endif #endif /* __cold */ #ifndef __flatten # if defined(__OPTIMIZE__) && (defined(__GNUC__) || __has_attribute(__flatten__)) # define __flatten __attribute__((__flatten__)) # else # define __flatten # endif #endif /* __flatten */ #ifndef likely # if (defined(__GNUC__) || __has_builtin(__builtin_expect)) && !defined(__COVERITY__) # define likely(cond) __builtin_expect(!!(cond), 1) # else # define likely(x) (!!(x)) # endif #endif /* likely */ #ifndef unlikely # if (defined(__GNUC__) || __has_builtin(__builtin_expect)) && !defined(__COVERITY__) # define unlikely(cond) __builtin_expect(!!(cond), 0) # else # define unlikely(x) (!!(x)) # endif #endif /* unlikely */ #ifndef __anonymous_struct_extension__ # if defined(__GNUC__) # define __anonymous_struct_extension__ __extension__ # else # define __anonymous_struct_extension__ # endif #endif /* __anonymous_struct_extension__ */ #ifndef __Wpedantic_format_voidptr static __inline __maybe_unused const void* MDBX_PURE_FUNCTION __Wpedantic_format_voidptr(const void* ptr) {return ptr;} # define __Wpedantic_format_voidptr(ARG) __Wpedantic_format_voidptr(ARG) #endif /* __Wpedantic_format_voidptr */ /*----------------------------------------------------------------------------*/ #if defined(MDBX_USE_VALGRIND) # include # ifndef VALGRIND_DISABLE_ADDR_ERROR_REPORTING_IN_RANGE /* LY: available since Valgrind 3.10 */ # define VALGRIND_DISABLE_ADDR_ERROR_REPORTING_IN_RANGE(a,s) # define VALGRIND_ENABLE_ADDR_ERROR_REPORTING_IN_RANGE(a,s) # endif #elif !defined(RUNNING_ON_VALGRIND) # define VALGRIND_CREATE_MEMPOOL(h,r,z) # define VALGRIND_DESTROY_MEMPOOL(h) # define VALGRIND_MEMPOOL_TRIM(h,a,s) # define VALGRIND_MEMPOOL_ALLOC(h,a,s) # define VALGRIND_MEMPOOL_FREE(h,a) # define VALGRIND_MEMPOOL_CHANGE(h,a,b,s) # define VALGRIND_MAKE_MEM_NOACCESS(a,s) # define VALGRIND_MAKE_MEM_DEFINED(a,s) # define VALGRIND_MAKE_MEM_UNDEFINED(a,s) # define VALGRIND_DISABLE_ADDR_ERROR_REPORTING_IN_RANGE(a,s) # define VALGRIND_ENABLE_ADDR_ERROR_REPORTING_IN_RANGE(a,s) # define VALGRIND_CHECK_MEM_IS_ADDRESSABLE(a,s) (0) # define VALGRIND_CHECK_MEM_IS_DEFINED(a,s) (0) # define RUNNING_ON_VALGRIND (0) #endif /* MDBX_USE_VALGRIND */ #ifdef __SANITIZE_ADDRESS__ # include #elif !defined(ASAN_POISON_MEMORY_REGION) # define ASAN_POISON_MEMORY_REGION(addr, size) \ ((void)(addr), (void)(size)) # define ASAN_UNPOISON_MEMORY_REGION(addr, size) \ ((void)(addr), (void)(size)) #endif /* __SANITIZE_ADDRESS__ */ /*----------------------------------------------------------------------------*/ #ifndef ARRAY_LENGTH # ifdef __cplusplus template char (&__ArraySizeHelper(T (&array)[N]))[N]; # define ARRAY_LENGTH(array) (sizeof(::__ArraySizeHelper(array))) # else # define ARRAY_LENGTH(array) (sizeof(array) / sizeof(array[0])) # endif #endif /* ARRAY_LENGTH */ #ifndef ARRAY_END # define ARRAY_END(array) (&array[ARRAY_LENGTH(array)]) #endif /* ARRAY_END */ #ifndef STRINGIFY # define STRINGIFY_HELPER(x) #x # define STRINGIFY(x) STRINGIFY_HELPER(x) #endif /* STRINGIFY */ #define CONCAT(a,b) a##b #define XCONCAT(a,b) CONCAT(a,b) #ifndef offsetof # define offsetof(type, member) __builtin_offsetof(type, member) #endif /* offsetof */ #ifndef container_of # define container_of(ptr, type, member) \ ((type *)((char *)(ptr) - offsetof(type, member))) #endif /* container_of */ #define MDBX_TETRAD(a, b, c, d) \ ((uint32_t)(a) << 24 | (uint32_t)(b) << 16 | (uint32_t)(c) << 8 | (d)) #define MDBX_STRING_TETRAD(str) MDBX_TETRAD(str[0], str[1], str[2], str[3]) #define FIXME "FIXME: " __FILE__ ", " STRINGIFY(__LINE__) #ifndef STATIC_ASSERT_MSG # if defined(static_assert) # define STATIC_ASSERT_MSG(expr, msg) static_assert(expr, msg) # elif defined(_STATIC_ASSERT) # define STATIC_ASSERT_MSG(expr, msg) _STATIC_ASSERT(expr) # elif defined(_MSC_VER) # include # define STATIC_ASSERT_MSG(expr, msg) _STATIC_ASSERT(expr) # elif (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) \ || __has_feature(c_static_assert) # define STATIC_ASSERT_MSG(expr, msg) _Static_assert(expr, msg) # else # define STATIC_ASSERT_MSG(expr, msg) switch (0) {case 0:case (expr):;} # endif #endif /* STATIC_ASSERT */ #ifndef STATIC_ASSERT # define STATIC_ASSERT(expr) STATIC_ASSERT_MSG(expr, #expr) #endif /* *INDENT-ON* */ /* clang-format on */ #if defined(__GNUC__) && !__GNUC_PREREQ(4,2) /* Actually libmdbx was not tested with compilers older than GCC 4.2. * But you could ignore this warning at your own risk. * In such case please don't rise up an issues related ONLY to old compilers. */ # warning "libmdbx required GCC >= 4.2" #endif #if defined(__clang__) && !__CLANG_PREREQ(3,8) /* Actually libmdbx was not tested with CLANG older than 3.8. * But you could ignore this warning at your own risk. * In such case please don't rise up an issues related ONLY to old compilers. */ # warning "libmdbx required CLANG >= 3.8" #endif #if defined(__GLIBC__) && !__GLIBC_PREREQ(2,12) /* Actually libmdbx was not tested with something older than glibc 2.12. * But you could ignore this warning at your own risk. * In such case please don't rise up an issues related ONLY to old systems. */ # warning "libmdbx was only tested with GLIBC >= 2.12." #endif #ifdef __SANITIZE_THREAD__ # warning "libmdbx don't compatible with ThreadSanitizer, you will get a lot of false-positive issues." #endif /* __SANITIZE_THREAD__ */ #if __has_warning("-Wnested-anon-types") # if defined(__clang__) # pragma clang diagnostic ignored "-Wnested-anon-types" # elif defined(__GNUC__) # pragma GCC diagnostic ignored "-Wnested-anon-types" # else # pragma warning disable "nested-anon-types" # endif #endif /* -Wnested-anon-types */ #if __has_warning("-Wconstant-logical-operand") # if defined(__clang__) # pragma clang diagnostic ignored "-Wconstant-logical-operand" # elif defined(__GNUC__) # pragma GCC diagnostic ignored "-Wconstant-logical-operand" # else # pragma warning disable "constant-logical-operand" # endif #endif /* -Wconstant-logical-operand */ #if defined(__LCC__) && (__LCC__ <= 121) /* bug #2798 */ # pragma diag_suppress alignment_reduction_ignored #elif defined(__ICC) # pragma warning(disable: 3453 1366) #elif __has_warning("-Walignment-reduction-ignored") # if defined(__clang__) # pragma clang diagnostic ignored "-Walignment-reduction-ignored" # elif defined(__GNUC__) # pragma GCC diagnostic ignored "-Walignment-reduction-ignored" # else # pragma warning disable "alignment-reduction-ignored" # endif #endif /* -Walignment-reduction-ignored */ /* *INDENT-ON* */ /* clang-format on */ #ifdef __cplusplus extern "C" { #endif /* https://en.wikipedia.org/wiki/Operating_system_abstraction_layer */ /* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ /*----------------------------------------------------------------------------*/ /* Microsoft compiler generates a lot of warning for self includes... */ #ifdef _MSC_VER #pragma warning(push, 1) #pragma warning(disable : 4548) /* expression before comma has no effect; \ expected expression with side - effect */ #pragma warning(disable : 4530) /* C++ exception handler used, but unwind \ * semantics are not enabled. Specify /EHsc */ #pragma warning(disable : 4577) /* 'noexcept' used with no exception handling \ * mode specified; termination on exception is \ * not guaranteed. Specify /EHsc */ #endif /* _MSC_VER (warnings) */ #if defined(_WIN32) || defined(_WIN64) #if !defined(_CRT_SECURE_NO_WARNINGS) #define _CRT_SECURE_NO_WARNINGS #endif #if !defined(_NO_CRT_STDIO_INLINE) && MDBX_BUILD_SHARED_LIBRARY && \ !defined(MDBX_TOOLS) && MDBX_AVOID_CRT #define _NO_CRT_STDIO_INLINE #endif #elif !defined(_POSIX_C_SOURCE) #define _POSIX_C_SOURCE 200809L #endif /* Windows */ /*----------------------------------------------------------------------------*/ /* C99 includes */ #include #include #include #include #include #include #include #include #include #include /* C11 stdalign.h */ #if __has_include() #include #elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L #define alignas(N) _Alignas(N) #elif defined(_MSC_VER) #define alignas(N) __declspec(align(N)) #elif __has_attribute(__aligned__) || defined(__GNUC__) #define alignas(N) __attribute__((__aligned__(N))) #else #error "FIXME: Required _alignas() or equivalent." #endif /*----------------------------------------------------------------------------*/ /* Systems includes */ #ifdef __APPLE__ #include #endif /* Apple OSX & iOS */ #if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \ defined(__BSD__) || defined(__bsdi__) || defined(__DragonFly__) || \ defined(__APPLE__) || defined(__MACH__) #include #include #include #include #if defined(__FreeBSD__) || defined(__DragonFly__) #include #elif defined(__OpenBSD__) || defined(__NetBSD__) #include #else #define SYSCTL_LEGACY_NONCONST_MIB #endif #ifndef __MACH__ #include #endif #else #include #if !(defined(__sun) || defined(__SVR4) || defined(__svr4__) || \ defined(_WIN32) || defined(_WIN64)) #include #endif /* !Solaris */ #endif /* !xBSD */ #if defined(__FreeBSD__) || __has_include() #include #endif #if defined(__APPLE__) || defined(__MACH__) || __has_include() #include #endif /* MacOS */ #if defined(__MACH__) #include #include #include #include #undef P_DIRTY #endif #if defined(__linux__) || defined(__gnu_linux__) #include #include #include #include #endif /* Linux */ #ifndef _XOPEN_SOURCE #define _XOPEN_SOURCE 0 #endif #ifndef _XOPEN_SOURCE_EXTENDED #define _XOPEN_SOURCE_EXTENDED 0 #else #include #endif /* _XOPEN_SOURCE_EXTENDED */ #if defined(__sun) || defined(__SVR4) || defined(__svr4__) #include #include /* On Solaris, it's easier to add a missing prototype rather than find a * combination of #defines that break nothing. */ __extern_C key_t ftok(const char *, int); #endif /* SunOS/Solaris */ #if defined(_WIN32) || defined(_WIN64) #ifndef WIN32_LEAN_AND_MEAN #define WIN32_LEAN_AND_MEAN #endif #include #include #include #include #include #define HAVE_SYS_STAT_H #define HAVE_SYS_TYPES_H typedef HANDLE mdbx_thread_t; typedef unsigned mdbx_thread_key_t; #define MDBX_OSAL_SECTION HANDLE #define MAP_FAILED NULL #define HIGH_DWORD(v) ((DWORD)((sizeof(v) > 4) ? ((uint64_t)(v) >> 32) : 0)) #define THREAD_CALL WINAPI #define THREAD_RESULT DWORD typedef struct { HANDLE mutex; HANDLE event[2]; } mdbx_condpair_t; typedef CRITICAL_SECTION mdbx_fastmutex_t; #if MDBX_AVOID_CRT #ifndef mdbx_malloc static inline void *mdbx_malloc(size_t bytes) { return LocalAlloc(LMEM_FIXED, bytes); } #endif /* mdbx_malloc */ #ifndef mdbx_calloc static inline void *mdbx_calloc(size_t nelem, size_t size) { return LocalAlloc(LMEM_FIXED | LMEM_ZEROINIT, nelem * size); } #endif /* mdbx_calloc */ #ifndef mdbx_realloc static inline void *mdbx_realloc(void *ptr, size_t bytes) { return ptr ? LocalReAlloc(ptr, bytes, LMEM_MOVEABLE) : LocalAlloc(LMEM_FIXED, bytes); } #endif /* mdbx_realloc */ #ifndef mdbx_free #define mdbx_free LocalFree #endif /* mdbx_free */ #else #define mdbx_malloc malloc #define mdbx_calloc calloc #define mdbx_realloc realloc #define mdbx_free free #define mdbx_strdup _strdup #endif /* MDBX_AVOID_CRT */ #ifndef snprintf #define snprintf _snprintf /* ntdll */ #endif #ifndef vsnprintf #define vsnprintf _vsnprintf /* ntdll */ #endif #else /*----------------------------------------------------------------------*/ #include #include #include #include #include #include #include #include #include #include #include typedef pthread_t mdbx_thread_t; typedef pthread_key_t mdbx_thread_key_t; #define INVALID_HANDLE_VALUE (-1) #define THREAD_CALL #define THREAD_RESULT void * typedef struct { pthread_mutex_t mutex; pthread_cond_t cond[2]; } mdbx_condpair_t; typedef pthread_mutex_t mdbx_fastmutex_t; #define mdbx_malloc malloc #define mdbx_calloc calloc #define mdbx_realloc realloc #define mdbx_free free #define mdbx_strdup strdup #endif /* Platform */ #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) /* malloc_usable_size() already provided */ #elif defined(__APPLE__) #define malloc_usable_size(ptr) malloc_size(ptr) #elif defined(_MSC_VER) && !MDBX_AVOID_CRT #define malloc_usable_size(ptr) _msize(ptr) #endif /* malloc_usable_size */ #ifdef __ANDROID_API__ #include #if __ANDROID_API__ >= 21 #include #endif #endif /* Android */ /* *INDENT-OFF* */ /* clang-format off */ #if defined(HAVE_SYS_STAT_H) || __has_include() #include #endif #if defined(HAVE_SYS_TYPES_H) || __has_include() #include #endif #if defined(HAVE_SYS_FILE_H) || __has_include() #include #endif /* *INDENT-ON* */ /* clang-format on */ #ifndef SSIZE_MAX #define SSIZE_MAX INTPTR_MAX #endif #if !defined(MADV_DODUMP) && defined(MADV_CORE) #define MADV_DODUMP MADV_CORE #endif /* MADV_CORE -> MADV_DODUMP */ #if !defined(MADV_DONTDUMP) && defined(MADV_NOCORE) #define MADV_DONTDUMP MADV_NOCORE #endif /* MADV_NOCORE -> MADV_DONTDUMP */ #if defined(i386) || defined(__386) || defined(__i386) || defined(__i386__) || \ defined(i486) || defined(__i486) || defined(__i486__) || \ defined(i586) | defined(__i586) || defined(__i586__) || defined(i686) || \ defined(__i686) || defined(__i686__) || defined(_M_IX86) || \ defined(_X86_) || defined(__THW_INTEL__) || defined(__I86__) || \ defined(__INTEL__) || defined(__x86_64) || defined(__x86_64__) || \ defined(__amd64__) || defined(__amd64) || defined(_M_X64) || \ defined(_M_AMD64) || defined(__IA32__) || defined(__INTEL__) #ifndef __ia32__ /* LY: define neutral __ia32__ for x86 and x86-64 */ #define __ia32__ 1 #endif /* __ia32__ */ #if !defined(__amd64__) && (defined(__x86_64) || defined(__x86_64__) || \ defined(__amd64) || defined(_M_X64)) /* LY: define trusty __amd64__ for all AMD64/x86-64 arch */ #define __amd64__ 1 #endif /* __amd64__ */ #endif /* all x86 */ #if (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF #error \ "Sanity checking failed: Two's complement, reasonably sized integer types" #endif #if UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul #define MDBX_WORDBITS 64 #else #define MDBX_WORDBITS 32 #endif /* MDBX_WORDBITS */ /*----------------------------------------------------------------------------*/ /* Compiler's includes for builtins/intrinsics */ #if defined(_MSC_VER) || defined(__INTEL_COMPILER) #include #elif __GNUC_PREREQ(4, 4) || defined(__clang__) #if defined(__ia32__) || defined(__e2k__) #include #endif /* __ia32__ */ #if defined(__ia32__) #include #endif /* __ia32__ */ #elif defined(__SUNPRO_C) || defined(__sun) || defined(sun) #include #elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) && \ (defined(HP_IA64) || defined(__ia64)) #include #elif defined(__IBMC__) && defined(__powerpc) #include #elif defined(_AIX) #include #include #elif (defined(__osf__) && defined(__DECC)) || defined(__alpha) #include #include #elif defined(__MWERKS__) /* CodeWarrior - troubles ? */ #pragma gcc_extensions #elif defined(__SNC__) /* Sony PS3 - troubles ? */ #elif defined(__hppa__) || defined(__hppa) #include #else #error Unsupported C compiler, please use GNU C 4.4 or newer #endif /* Compiler */ /*----------------------------------------------------------------------------*/ /* Byteorder */ #if !defined(__BYTE_ORDER__) || !defined(__ORDER_LITTLE_ENDIAN__) || \ !defined(__ORDER_BIG_ENDIAN__) /* *INDENT-OFF* */ /* clang-format off */ #if defined(__GLIBC__) || defined(__GNU_LIBRARY__) || defined(__ANDROID_API__) || \ defined(HAVE_ENDIAN_H) || __has_include() #include #elif defined(__APPLE__) || defined(__MACH__) || defined(__OpenBSD__) || \ defined(HAVE_MACHINE_ENDIAN_H) || __has_include() #include #elif defined(HAVE_SYS_ISA_DEFS_H) || __has_include() #include #elif (defined(HAVE_SYS_TYPES_H) && defined(HAVE_SYS_ENDIAN_H)) || \ (__has_include() && __has_include()) #include #include #elif defined(__bsdi__) || defined(__DragonFly__) || defined(__FreeBSD__) || \ defined(__NetBSD__) || \ defined(HAVE_SYS_PARAM_H) || __has_include() #include #endif /* OS */ /* *INDENT-ON* */ /* clang-format on */ #if defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && defined(__BIG_ENDIAN) #define __ORDER_LITTLE_ENDIAN__ __LITTLE_ENDIAN #define __ORDER_BIG_ENDIAN__ __BIG_ENDIAN #define __BYTE_ORDER__ __BYTE_ORDER #elif defined(_BYTE_ORDER) && defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN) #define __ORDER_LITTLE_ENDIAN__ _LITTLE_ENDIAN #define __ORDER_BIG_ENDIAN__ _BIG_ENDIAN #define __BYTE_ORDER__ _BYTE_ORDER #else #define __ORDER_LITTLE_ENDIAN__ 1234 #define __ORDER_BIG_ENDIAN__ 4321 #if defined(__LITTLE_ENDIAN__) || \ (defined(_LITTLE_ENDIAN) && !defined(_BIG_ENDIAN)) || \ defined(__ARMEL__) || defined(__THUMBEL__) || defined(__AARCH64EL__) || \ defined(__MIPSEL__) || defined(_MIPSEL) || defined(__MIPSEL) || \ defined(_M_ARM) || defined(_M_ARM64) || defined(__e2k__) || \ defined(__elbrus_4c__) || defined(__elbrus_8c__) || defined(__bfin__) || \ defined(__BFIN__) || defined(__ia64__) || defined(_IA64) || \ defined(__IA64__) || defined(__ia64) || defined(_M_IA64) || \ defined(__itanium__) || defined(__ia32__) || defined(__CYGWIN__) || \ defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) || \ defined(__WINDOWS__) #define __BYTE_ORDER__ __ORDER_LITTLE_ENDIAN__ #elif defined(__BIG_ENDIAN__) || \ (defined(_BIG_ENDIAN) && !defined(_LITTLE_ENDIAN)) || \ defined(__ARMEB__) || defined(__THUMBEB__) || defined(__AARCH64EB__) || \ defined(__MIPSEB__) || defined(_MIPSEB) || defined(__MIPSEB) || \ defined(__m68k__) || defined(M68000) || defined(__hppa__) || \ defined(__hppa) || defined(__HPPA__) || defined(__sparc__) || \ defined(__sparc) || defined(__370__) || defined(__THW_370__) || \ defined(__s390__) || defined(__s390x__) || defined(__SYSC_ZARCH__) #define __BYTE_ORDER__ __ORDER_BIG_ENDIAN__ #else #error __BYTE_ORDER__ should be defined. #endif /* Arch */ #endif #endif /* __BYTE_ORDER__ || __ORDER_LITTLE_ENDIAN__ || __ORDER_BIG_ENDIAN__ */ /* Get the size of a memory page for the system. * This is the basic size that the platform's memory manager uses, and is * fundamental to the use of memory-mapped files. */ MDBX_NOTHROW_CONST_FUNCTION static __maybe_unused __inline size_t mdbx_syspagesize(void) { #if defined(_WIN32) || defined(_WIN64) SYSTEM_INFO si; GetSystemInfo(&si); return si.dwPageSize; #else return sysconf(_SC_PAGE_SIZE); #endif } typedef struct mdbx_mmap_param { union { void *address; uint8_t *dxb; struct MDBX_lockinfo *lck; }; mdbx_filehandle_t fd; size_t limit; /* mapping length, but NOT a size of file nor DB */ size_t current; /* mapped region size, i.e. the size of file and DB */ #if defined(_WIN32) || defined(_WIN64) uint64_t filesize /* in-process cache of a file size. */; #endif #ifdef MDBX_OSAL_SECTION MDBX_OSAL_SECTION section; #endif } mdbx_mmap_t; typedef union bin128 { __anonymous_struct_extension__ struct { uint64_t x, y; }; __anonymous_struct_extension__ struct { uint32_t a, b, c, d; }; } bin128_t; #if defined(_WIN32) || defined(_WIN64) typedef union MDBX_srwlock { struct { long volatile readerCount; long volatile writerCount; }; RTL_SRWLOCK native; } MDBX_srwlock; #endif /* Windows */ #ifdef __cplusplus extern void mdbx_osal_jitter(bool tiny); #else /*----------------------------------------------------------------------------*/ /* Atomics */ #if defined(__cplusplus) && !defined(__STDC_NO_ATOMICS__) && (__has_include() || __has_extension(cxx_atomic)) #include #define MDBX_HAVE_C11ATOMICS #elif !defined(__cplusplus) && \ (__STDC_VERSION__ >= 201112L || __has_extension(c_atomic)) && \ !defined(__STDC_NO_ATOMICS__) && \ (__GNUC_PREREQ(4, 9) || __CLANG_PREREQ(3, 8) || \ !(defined(__GNUC__) || defined(__clang__))) #include #define MDBX_HAVE_C11ATOMICS #elif defined(__GNUC__) || defined(__clang__) #elif defined(_MSC_VER) #pragma warning(disable : 4163) /* 'xyz': not available as an intrinsic */ #pragma warning(disable : 4133) /* 'function': incompatible types - from \ 'size_t' to 'LONGLONG' */ #pragma warning(disable : 4244) /* 'return': conversion from 'LONGLONG' to \ 'std::size_t', possible loss of data */ #pragma warning(disable : 4267) /* 'function': conversion from 'size_t' to \ 'long', possible loss of data */ #pragma intrinsic(_InterlockedExchangeAdd, _InterlockedCompareExchange) #pragma intrinsic(_InterlockedExchangeAdd64, _InterlockedCompareExchange64) #elif defined(__APPLE__) #include #else #error FIXME atomic-ops #endif /*----------------------------------------------------------------------------*/ /* Memory/Compiler barriers, cache coherence */ #if __has_include() #include #elif defined(__mips) || defined(__mips__) || defined(__mips64) || \ defined(__mips64__) || defined(_M_MRX000) || defined(_MIPS_) || \ defined(__MWERKS__) || defined(__sgi) /* MIPS should have explicit cache control */ #include #endif static __maybe_unused __inline void mdbx_compiler_barrier(void) { #if defined(__clang__) || defined(__GNUC__) __asm__ __volatile__("" ::: "memory"); #elif defined(_MSC_VER) _ReadWriteBarrier(); #elif defined(__INTEL_COMPILER) /* LY: Intel Compiler may mimic GCC and MSC */ __memory_barrier(); #elif defined(__SUNPRO_C) || defined(__sun) || defined(sun) __compiler_barrier(); #elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) && \ (defined(HP_IA64) || defined(__ia64)) _Asm_sched_fence(/* LY: no-arg meaning 'all expect ALU', e.g. 0x3D3D */); #elif defined(_AIX) || defined(__ppc__) || defined(__powerpc__) || \ defined(__ppc64__) || defined(__powerpc64__) __fence(); #else #error "Could not guess the kind of compiler, please report to us." #endif } static __maybe_unused __inline void mdbx_memory_barrier(void) { #ifdef MDBX_HAVE_C11ATOMICS atomic_thread_fence(memory_order_seq_cst); #elif defined(__ATOMIC_SEQ_CST) #ifdef __clang__ __c11_atomic_thread_fence(__ATOMIC_SEQ_CST); #else __atomic_thread_fence(__ATOMIC_SEQ_CST); #endif #elif defined(__clang__) || defined(__GNUC__) __sync_synchronize(); #elif defined(_WIN32) || defined(_WIN64) MemoryBarrier(); #elif defined(__INTEL_COMPILER) /* LY: Intel Compiler may mimic GCC and MSC */ #if defined(__ia32__) _mm_mfence(); #else __mf(); #endif #elif defined(__SUNPRO_C) || defined(__sun) || defined(sun) __machine_rw_barrier(); #elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) && \ (defined(HP_IA64) || defined(__ia64)) _Asm_mf(); #elif defined(_AIX) || defined(__ppc__) || defined(__powerpc__) || \ defined(__ppc64__) || defined(__powerpc64__) __lwsync(); #else #error "Could not guess the kind of compiler, please report to us." #endif } /*----------------------------------------------------------------------------*/ /* libc compatibility stuff */ #if (!defined(__GLIBC__) && __GLIBC_PREREQ(2, 1)) && \ (defined(_GNU_SOURCE) || defined(_BSD_SOURCE)) #define mdbx_asprintf asprintf #define mdbx_vasprintf vasprintf #else MDBX_INTERNAL_FUNC MDBX_PRINTF_ARGS(2, 3) int __maybe_unused mdbx_asprintf(char **strp, const char *fmt, ...); MDBX_INTERNAL_FUNC int mdbx_vasprintf(char **strp, const char *fmt, va_list ap); #endif /*----------------------------------------------------------------------------*/ /* OS abstraction layer stuff */ /* max bytes to write in one call */ #if defined(_WIN32) || defined(_WIN64) #define MAX_WRITE UINT32_C(0x01000000) #else #define MAX_WRITE UINT32_C(0x3fff0000) #endif #if defined(__linux__) || defined(__gnu_linux__) MDBX_INTERNAL_VAR uint32_t mdbx_linux_kernel_version; MDBX_INTERNAL_VAR bool mdbx_RunningOnWSL1 /* Windows Subsystem 1 for Linux */; #endif /* Linux */ #ifndef mdbx_strdup LIBMDBX_API char *mdbx_strdup(const char *str); #endif static __maybe_unused __inline int mdbx_get_errno(void) { #if defined(_WIN32) || defined(_WIN64) DWORD rc = GetLastError(); #else int rc = errno; #endif return rc; } #ifndef mdbx_memalign_alloc MDBX_INTERNAL_FUNC int mdbx_memalign_alloc(size_t alignment, size_t bytes, void **result); #endif #ifndef mdbx_memalign_free MDBX_INTERNAL_FUNC void mdbx_memalign_free(void *ptr); #endif MDBX_INTERNAL_FUNC int mdbx_condpair_init(mdbx_condpair_t *condpair); MDBX_INTERNAL_FUNC int mdbx_condpair_lock(mdbx_condpair_t *condpair); MDBX_INTERNAL_FUNC int mdbx_condpair_unlock(mdbx_condpair_t *condpair); MDBX_INTERNAL_FUNC int mdbx_condpair_signal(mdbx_condpair_t *condpair, bool part); MDBX_INTERNAL_FUNC int mdbx_condpair_wait(mdbx_condpair_t *condpair, bool part); MDBX_INTERNAL_FUNC int mdbx_condpair_destroy(mdbx_condpair_t *condpair); MDBX_INTERNAL_FUNC int mdbx_fastmutex_init(mdbx_fastmutex_t *fastmutex); MDBX_INTERNAL_FUNC int mdbx_fastmutex_acquire(mdbx_fastmutex_t *fastmutex); MDBX_INTERNAL_FUNC int mdbx_fastmutex_release(mdbx_fastmutex_t *fastmutex); MDBX_INTERNAL_FUNC int mdbx_fastmutex_destroy(mdbx_fastmutex_t *fastmutex); MDBX_INTERNAL_FUNC int mdbx_pwritev(mdbx_filehandle_t fd, struct iovec *iov, int iovcnt, uint64_t offset, size_t expected_written); MDBX_INTERNAL_FUNC int mdbx_pread(mdbx_filehandle_t fd, void *buf, size_t count, uint64_t offset); MDBX_INTERNAL_FUNC int mdbx_pwrite(mdbx_filehandle_t fd, const void *buf, size_t count, uint64_t offset); MDBX_INTERNAL_FUNC int mdbx_write(mdbx_filehandle_t fd, const void *buf, size_t count); MDBX_INTERNAL_FUNC int mdbx_thread_create(mdbx_thread_t *thread, THREAD_RESULT(THREAD_CALL *start_routine)(void *), void *arg); MDBX_INTERNAL_FUNC int mdbx_thread_join(mdbx_thread_t thread); enum mdbx_syncmode_bits { MDBX_SYNC_NONE = 0, MDBX_SYNC_DATA = 1, MDBX_SYNC_SIZE = 2, MDBX_SYNC_IODQ = 4 }; MDBX_INTERNAL_FUNC int mdbx_fsync(mdbx_filehandle_t fd, const enum mdbx_syncmode_bits mode_bits); MDBX_INTERNAL_FUNC int mdbx_ftruncate(mdbx_filehandle_t fd, uint64_t length); MDBX_INTERNAL_FUNC int mdbx_fseek(mdbx_filehandle_t fd, uint64_t pos); MDBX_INTERNAL_FUNC int mdbx_filesize(mdbx_filehandle_t fd, uint64_t *length); enum mdbx_openfile_purpose { MDBX_OPEN_DXB_READ = 0, MDBX_OPEN_DXB_LAZY = 1, MDBX_OPEN_DXB_DSYNC = 2, MDBX_OPEN_LCK = 3, MDBX_OPEN_COPY = 4, MDBX_OPEN_DELETE = 5 }; MDBX_INTERNAL_FUNC int mdbx_openfile(const enum mdbx_openfile_purpose purpose, const MDBX_env *env, const char *pathname, mdbx_filehandle_t *fd, mdbx_mode_t unix_mode_bits); MDBX_INTERNAL_FUNC int mdbx_closefile(mdbx_filehandle_t fd); MDBX_INTERNAL_FUNC int mdbx_removefile(const char *pathname); MDBX_INTERNAL_FUNC int mdbx_removedirectory(const char *pathname); MDBX_INTERNAL_FUNC int mdbx_is_pipe(mdbx_filehandle_t fd); MDBX_INTERNAL_FUNC int mdbx_lockfile(mdbx_filehandle_t fd, bool wait); #define MMAP_OPTION_TRUNCATE 1 #define MMAP_OPTION_SEMAPHORE 2 MDBX_INTERNAL_FUNC int mdbx_mmap(const int flags, mdbx_mmap_t *map, const size_t must, const size_t limit, const unsigned options); MDBX_INTERNAL_FUNC int mdbx_munmap(mdbx_mmap_t *map); MDBX_INTERNAL_FUNC int mdbx_mresize(int flags, mdbx_mmap_t *map, size_t current, size_t wanna, const bool may_move); #if defined(_WIN32) || defined(_WIN64) typedef struct { unsigned limit, count; HANDLE handles[31]; } mdbx_handle_array_t; MDBX_INTERNAL_FUNC int mdbx_suspend_threads_before_remap(MDBX_env *env, mdbx_handle_array_t **array); MDBX_INTERNAL_FUNC int mdbx_resume_threads_after_remap(mdbx_handle_array_t *array); #endif /* Windows */ MDBX_INTERNAL_FUNC int mdbx_msync(mdbx_mmap_t *map, size_t offset, size_t length, enum mdbx_syncmode_bits mode_bits); MDBX_INTERNAL_FUNC int mdbx_check_fs_rdonly(mdbx_filehandle_t handle, const char *pathname, int err); static __maybe_unused __inline uint32_t mdbx_getpid(void) { STATIC_ASSERT(sizeof(mdbx_pid_t) <= sizeof(uint32_t)); #if defined(_WIN32) || defined(_WIN64) return GetCurrentProcessId(); #else return getpid(); #endif } static __maybe_unused __inline uintptr_t mdbx_thread_self(void) { mdbx_tid_t thunk; STATIC_ASSERT(sizeof(uintptr_t) >= sizeof(thunk)); #if defined(_WIN32) || defined(_WIN64) thunk = GetCurrentThreadId(); #else thunk = pthread_self(); #endif return (uintptr_t)thunk; } MDBX_INTERNAL_FUNC void __maybe_unused mdbx_osal_jitter(bool tiny); MDBX_INTERNAL_FUNC uint64_t mdbx_osal_monotime(void); MDBX_INTERNAL_FUNC uint64_t mdbx_osal_16dot16_to_monotime(uint32_t seconds_16dot16); MDBX_INTERNAL_FUNC uint32_t mdbx_osal_monotime_to_16dot16(uint64_t monotime); MDBX_INTERNAL_FUNC bin128_t mdbx_osal_bootid(void); /*----------------------------------------------------------------------------*/ /* lck stuff */ /// \brief Initialization of synchronization primitives linked with MDBX_env /// instance both in LCK-file and within the current process. /// \param /// global_uniqueness_flag = true - denotes that there are no other processes /// working with DB and LCK-file. Thus the function MUST initialize /// shared synchronization objects in memory-mapped LCK-file. /// global_uniqueness_flag = false - denotes that at least one process is /// already working with DB and LCK-file, including the case when DB /// has already been opened in the current process. Thus the function /// MUST NOT initialize shared synchronization objects in memory-mapped /// LCK-file that are already in use. /// \return Error code or zero on success. MDBX_INTERNAL_FUNC int mdbx_lck_init(MDBX_env *env, MDBX_env *inprocess_neighbor, int global_uniqueness_flag); /// \brief Disconnects from shared interprocess objects and destructs /// synchronization objects linked with MDBX_env instance /// within the current process. /// \param /// inprocess_neighbor = NULL - if the current process does not have other /// instances of MDBX_env linked with the DB being closed. /// Thus the function MUST check for other processes working with DB or /// LCK-file, and keep or destroy shared synchronization objects in /// memory-mapped LCK-file depending on the result. /// inprocess_neighbor = not-NULL - pointer to another instance of MDBX_env /// (anyone of there is several) working with DB or LCK-file within the /// current process. Thus the function MUST NOT try to acquire exclusive /// lock and/or try to destruct shared synchronization objects linked with /// DB or LCK-file. Moreover, the implementation MUST ensure correct work /// of other instances of MDBX_env within the current process, e.g. /// restore POSIX-fcntl locks after the closing of file descriptors. /// \return Error code (MDBX_PANIC) or zero on success. MDBX_INTERNAL_FUNC int mdbx_lck_destroy(MDBX_env *env, MDBX_env *inprocess_neighbor); /// \brief Connects to shared interprocess locking objects and tries to acquire /// the maximum lock level (shared if exclusive is not available) /// Depending on implementation or/and platform (Windows) this function may /// acquire the non-OS super-level lock (e.g. for shared synchronization /// objects initialization), which will be downgraded to OS-exclusive or /// shared via explicit calling of mdbx_lck_downgrade(). /// \return /// MDBX_RESULT_TRUE (-1) - if an exclusive lock was acquired and thus /// the current process is the first and only after the last use of DB. /// MDBX_RESULT_FALSE (0) - if a shared lock was acquired and thus /// DB has already been opened and now is used by other processes. /// Otherwise (not 0 and not -1) - error code. MDBX_INTERNAL_FUNC int mdbx_lck_seize(MDBX_env *env); /// \brief Downgrades the level of initially acquired lock to /// operational level specified by argument. The reson for such downgrade: /// - unblocking of other processes that are waiting for access, i.e. /// if (env->me_flags & MDBX_EXCLUSIVE) != 0, then other processes /// should be made aware that access is unavailable rather than /// wait for it. /// - freeing locks that interfere file operation (especially for Windows) /// (env->me_flags & MDBX_EXCLUSIVE) == 0 - downgrade to shared lock. /// (env->me_flags & MDBX_EXCLUSIVE) != 0 - downgrade to exclusive /// operational lock. /// \return Error code or zero on success MDBX_INTERNAL_FUNC int mdbx_lck_downgrade(MDBX_env *env); /// \brief Locks LCK-file or/and table of readers for (de)registering. /// \return Error code or zero on success MDBX_INTERNAL_FUNC int mdbx_rdt_lock(MDBX_env *env); /// \brief Unlocks LCK-file or/and table of readers after (de)registering. MDBX_INTERNAL_FUNC void mdbx_rdt_unlock(MDBX_env *env); /// \brief Acquires lock for DB change (on writing transaction start) /// Reading transactions will not be blocked. /// Declared as LIBMDBX_API because it is used in mdbx_chk. /// \return Error code or zero on success LIBMDBX_API int mdbx_txn_lock(MDBX_env *env, bool dont_wait); /// \brief Releases lock once DB changes is made (after writing transaction /// has finished). /// Declared as LIBMDBX_API because it is used in mdbx_chk. LIBMDBX_API void mdbx_txn_unlock(MDBX_env *env); /// \brief Sets alive-flag of reader presence (indicative lock) for PID of /// the current process. The function does no more than needed for /// the correct working of mdbx_rpid_check() in other processes. /// \return Error code or zero on success MDBX_INTERNAL_FUNC int mdbx_rpid_set(MDBX_env *env); /// \brief Resets alive-flag of reader presence (indicative lock) /// for PID of the current process. The function does no more than needed /// for the correct working of mdbx_rpid_check() in other processes. /// \return Error code or zero on success MDBX_INTERNAL_FUNC int mdbx_rpid_clear(MDBX_env *env); /// \brief Checks for reading process status with the given pid with help of /// alive-flag of presence (indicative lock) or using another way. /// \return /// MDBX_RESULT_TRUE (-1) - if the reader process with the given PID is alive /// and working with DB (indicative lock is present). /// MDBX_RESULT_FALSE (0) - if the reader process with the given PID is absent /// or not working with DB (indicative lock is not present). /// Otherwise (not 0 and not -1) - error code. MDBX_INTERNAL_FUNC int mdbx_rpid_check(MDBX_env *env, uint32_t pid); #if defined(_WIN32) || defined(_WIN64) typedef void(WINAPI *MDBX_srwlock_function)(MDBX_srwlock *); MDBX_INTERNAL_VAR MDBX_srwlock_function mdbx_srwlock_Init, mdbx_srwlock_AcquireShared, mdbx_srwlock_ReleaseShared, mdbx_srwlock_AcquireExclusive, mdbx_srwlock_ReleaseExclusive; typedef BOOL(WINAPI *MDBX_GetFileInformationByHandleEx)( _In_ HANDLE hFile, _In_ FILE_INFO_BY_HANDLE_CLASS FileInformationClass, _Out_ LPVOID lpFileInformation, _In_ DWORD dwBufferSize); MDBX_INTERNAL_VAR MDBX_GetFileInformationByHandleEx mdbx_GetFileInformationByHandleEx; typedef BOOL(WINAPI *MDBX_GetVolumeInformationByHandleW)( _In_ HANDLE hFile, _Out_opt_ LPWSTR lpVolumeNameBuffer, _In_ DWORD nVolumeNameSize, _Out_opt_ LPDWORD lpVolumeSerialNumber, _Out_opt_ LPDWORD lpMaximumComponentLength, _Out_opt_ LPDWORD lpFileSystemFlags, _Out_opt_ LPWSTR lpFileSystemNameBuffer, _In_ DWORD nFileSystemNameSize); MDBX_INTERNAL_VAR MDBX_GetVolumeInformationByHandleW mdbx_GetVolumeInformationByHandleW; typedef DWORD(WINAPI *MDBX_GetFinalPathNameByHandleW)(_In_ HANDLE hFile, _Out_ LPWSTR lpszFilePath, _In_ DWORD cchFilePath, _In_ DWORD dwFlags); MDBX_INTERNAL_VAR MDBX_GetFinalPathNameByHandleW mdbx_GetFinalPathNameByHandleW; typedef BOOL(WINAPI *MDBX_SetFileInformationByHandle)( _In_ HANDLE hFile, _In_ FILE_INFO_BY_HANDLE_CLASS FileInformationClass, _Out_ LPVOID lpFileInformation, _In_ DWORD dwBufferSize); MDBX_INTERNAL_VAR MDBX_SetFileInformationByHandle mdbx_SetFileInformationByHandle; typedef NTSTATUS(NTAPI *MDBX_NtFsControlFile)( IN HANDLE FileHandle, IN OUT HANDLE Event, IN OUT PVOID /* PIO_APC_ROUTINE */ ApcRoutine, IN OUT PVOID ApcContext, OUT PIO_STATUS_BLOCK IoStatusBlock, IN ULONG FsControlCode, IN OUT PVOID InputBuffer, IN ULONG InputBufferLength, OUT OPTIONAL PVOID OutputBuffer, IN ULONG OutputBufferLength); MDBX_INTERNAL_VAR MDBX_NtFsControlFile mdbx_NtFsControlFile; typedef uint64_t(WINAPI *MDBX_GetTickCount64)(void); MDBX_INTERNAL_VAR MDBX_GetTickCount64 mdbx_GetTickCount64; #if !defined(_WIN32_WINNT_WIN8) || _WIN32_WINNT < _WIN32_WINNT_WIN8 typedef struct _WIN32_MEMORY_RANGE_ENTRY { PVOID VirtualAddress; SIZE_T NumberOfBytes; } WIN32_MEMORY_RANGE_ENTRY, *PWIN32_MEMORY_RANGE_ENTRY; #endif /* Windows 8.x */ typedef BOOL(WINAPI *MDBX_PrefetchVirtualMemory)( HANDLE hProcess, ULONG_PTR NumberOfEntries, PWIN32_MEMORY_RANGE_ENTRY VirtualAddresses, ULONG Flags); MDBX_INTERNAL_VAR MDBX_PrefetchVirtualMemory mdbx_PrefetchVirtualMemory; #if 0 /* LY: unused for now */ #if !defined(_WIN32_WINNT_WIN81) || _WIN32_WINNT < _WIN32_WINNT_WIN81 typedef enum OFFER_PRIORITY { VmOfferPriorityVeryLow = 1, VmOfferPriorityLow, VmOfferPriorityBelowNormal, VmOfferPriorityNormal } OFFER_PRIORITY; #endif /* Windows 8.1 */ typedef DWORD(WINAPI *MDBX_DiscardVirtualMemory)(PVOID VirtualAddress, SIZE_T Size); MDBX_INTERNAL_VAR MDBX_DiscardVirtualMemory mdbx_DiscardVirtualMemory; typedef DWORD(WINAPI *MDBX_ReclaimVirtualMemory)(PVOID VirtualAddress, SIZE_T Size); MDBX_INTERNAL_VAR MDBX_ReclaimVirtualMemory mdbx_ReclaimVirtualMemory; typedef DWORD(WINAPI *MDBX_OfferVirtualMemory( PVOID VirtualAddress, SIZE_T Size, OFFER_PRIORITY Priority ); MDBX_INTERNAL_VAR MDBX_OfferVirtualMemory mdbx_OfferVirtualMemory; #endif /* unused for now */ typedef enum _SECTION_INHERIT { ViewShare = 1, ViewUnmap = 2 } SECTION_INHERIT; typedef NTSTATUS(NTAPI *MDBX_NtExtendSection)(IN HANDLE SectionHandle, IN PLARGE_INTEGER NewSectionSize); MDBX_INTERNAL_VAR MDBX_NtExtendSection mdbx_NtExtendSection; static __inline bool mdbx_RunningUnderWine(void) { return !mdbx_NtExtendSection; } typedef LSTATUS(WINAPI *MDBX_RegGetValueA)(HKEY hkey, LPCSTR lpSubKey, LPCSTR lpValue, DWORD dwFlags, LPDWORD pdwType, PVOID pvData, LPDWORD pcbData); MDBX_INTERNAL_VAR MDBX_RegGetValueA mdbx_RegGetValueA; #endif /* Windows */ #endif /* !__cplusplus */ /*----------------------------------------------------------------------------*/ #if defined(_MSC_VER) && _MSC_VER >= 1900 /* LY: MSVC 2015/2017/2019 has buggy/inconsistent PRIuPTR/PRIxPTR macros * for internal format-args checker. */ #undef PRIuPTR #undef PRIiPTR #undef PRIdPTR #undef PRIxPTR #define PRIuPTR "Iu" #define PRIiPTR "Ii" #define PRIdPTR "Id" #define PRIxPTR "Ix" #define PRIuSIZE "zu" #define PRIiSIZE "zi" #define PRIdSIZE "zd" #define PRIxSIZE "zx" #endif /* fix PRI*PTR for _MSC_VER */ #ifndef PRIuSIZE #define PRIuSIZE PRIuPTR #define PRIiSIZE PRIiPTR #define PRIdSIZE PRIdPTR #define PRIxSIZE PRIxPTR #endif /* PRI*SIZE macros for MSVC */ #ifdef _MSC_VER #pragma warning(pop) #endif #define mdbx_sourcery_anchor XCONCAT(mdbx_sourcery_, MDBX_BUILD_SOURCERY) #if defined(MDBX_TOOLS) extern LIBMDBX_API const char *const mdbx_sourcery_anchor; #endif /******************************************************************************* ******************************************************************************* ******************************************************************************* * * * #### ##### ##### # #### # # #### * # # # # # # # # ## # # * # # # # # # # # # # # #### * # # ##### # # # # # # # # * # # # # # # # # ## # # * #### # # # #### # # #### * * */ /** \defgroup build_option Build options * The libmdbx build options. @{ */ #ifdef DOXYGEN /* !!! Actually this is a fake definitions !!! * !!! for documentation generation by Doxygen !!! */ /** Controls enabling of debugging features. * * - `MDBX_DEBUG = 0` (by default) Disables any debugging features at all, * including logging and assertion controls. * Logging level and corresponding debug flags changing * by \ref mdbx_setup_debug() will not have effect. * - `MDBX_DEBUG > 0` Enables code for the debugging features (logging, * assertions checking and internal audit). * Simultaneously sets the default logging level * to the `MDBX_DEBUG` value. * Also enables \ref MDBX_DBG_AUDIT if `MDBX_DEBUG >= 2`. * * \ingroup build_option */ #define MDBX_DEBUG 0...7 /** Disables using of GNU libc extensions. */ #define MDBX_DISABLE_GNU_SOURCE 0 or 1 #endif /* DOXYGEN */ /** Using fcntl(F_FULLFSYNC) with 5-10 times slowdown */ #define MDBX_OSX_WANNA_DURABILITY 0 /** Using fsync() with chance of data lost on power failure */ #define MDBX_OSX_WANNA_SPEED 1 #ifndef MDBX_OSX_SPEED_INSTEADOF_DURABILITY /** Choices \ref MDBX_OSX_WANNA_DURABILITY or \ref MDBX_OSX_WANNA_SPEED * for OSX & iOS */ #define MDBX_OSX_SPEED_INSTEADOF_DURABILITY MDBX_OSX_WANNA_DURABILITY #endif /* MDBX_OSX_SPEED_INSTEADOF_DURABILITY */ /** Controls checking PID against reuse DB environment after the fork() */ #ifndef MDBX_ENV_CHECKPID #if defined(MADV_DONTFORK) || defined(_WIN32) || defined(_WIN64) /* PID check could be omitted: * - on Linux when madvise(MADV_DONTFORK) is available, i.e. after the fork() * mapped pages will not be available for child process. * - in Windows where fork() not available. */ #define MDBX_ENV_CHECKPID 0 #else #define MDBX_ENV_CHECKPID 1 #endif #define MDBX_ENV_CHECKPID_CONFIG "AUTO=" STRINGIFY(MDBX_ENV_CHECKPID) #else #define MDBX_ENV_CHECKPID_CONFIG STRINGIFY(MDBX_ENV_CHECKPID) #endif /* MDBX_ENV_CHECKPID */ /** Controls checking transaction owner thread against misuse transactions from * other threads. */ #ifndef MDBX_TXN_CHECKOWNER #define MDBX_TXN_CHECKOWNER 1 #define MDBX_TXN_CHECKOWNER_CONFIG "AUTO=" STRINGIFY(MDBX_TXN_CHECKOWNER) #else #define MDBX_TXN_CHECKOWNER_CONFIG STRINGIFY(MDBX_TXN_CHECKOWNER) #endif /* MDBX_TXN_CHECKOWNER */ /** Does a system have battery-backed Real-Time Clock or just a fake. */ #ifndef MDBX_TRUST_RTC #if defined(__linux__) || defined(__gnu_linux__) || defined(__NetBSD__) || \ defined(__OpenBSD__) #define MDBX_TRUST_RTC 0 /* a lot of embedded systems have a fake RTC */ #else #define MDBX_TRUST_RTC 1 #endif #define MDBX_TRUST_RTC_CONFIG "AUTO=" STRINGIFY(MDBX_TRUST_RTC) #else #define MDBX_TRUST_RTC_CONFIG STRINGIFY(MDBX_TRUST_RTC) #endif /* MDBX_TRUST_RTC */ /** Controls online database auto-compactification during write-transactions. */ #ifndef MDBX_ENABLE_REFUND #define MDBX_ENABLE_REFUND 1 #endif #if !(MDBX_ENABLE_REFUND == 0 || MDBX_ENABLE_REFUND == 1) #error MDBX_ENABLE_REFUND must be defined as 0 or 1 #endif /* MDBX_ENABLE_REFUND */ /** Controls sort order of internal page number lists. * The database format depend on this option and libmdbx builded with different * option value are incompatible. */ #ifndef MDBX_PNL_ASCENDING #define MDBX_PNL_ASCENDING 0 #endif #if !(MDBX_PNL_ASCENDING == 0 || MDBX_PNL_ASCENDING == 1) #error MDBX_PNL_ASCENDING must be defined as 0 or 1 #endif /* MDBX_PNL_ASCENDING */ //------------------------------------------------------------------------------ /** Win32 File Locking API for \ref MDBX_LOCKING */ #define MDBX_LOCKING_WIN32FILES -1 /** SystemV IPC semaphores for \ref MDBX_LOCKING */ #define MDBX_LOCKING_SYSV 5 /** POSIX-1 Shared anonymous semaphores for \ref MDBX_LOCKING */ #define MDBX_LOCKING_POSIX1988 1988 /** POSIX-2001 Shared Mutexes for \ref MDBX_LOCKING */ #define MDBX_LOCKING_POSIX2001 2001 /** POSIX-2008 Robust Mutexes for \ref MDBX_LOCKING */ #define MDBX_LOCKING_POSIX2008 2008 /** BeOS Benaphores, aka Futexes for \ref MDBX_LOCKING */ #define MDBX_LOCKING_BENAPHORE 1995 /** Advanced: Choices the locking implementation (autodetection by default). */ #if defined(_WIN32) || defined(_WIN64) #define MDBX_LOCKING MDBX_LOCKING_WIN32FILES #else #ifndef MDBX_LOCKING #if defined(_POSIX_THREAD_PROCESS_SHARED) && \ _POSIX_THREAD_PROCESS_SHARED >= 200112L && !defined(__FreeBSD__) /* Some platforms define the EOWNERDEAD error code even though they * don't support Robust Mutexes. If doubt compile with -MDBX_LOCKING=2001. */ #if defined(EOWNERDEAD) && _POSIX_THREAD_PROCESS_SHARED >= 200809L && \ ((defined(_POSIX_THREAD_ROBUST_PRIO_INHERIT) && \ _POSIX_THREAD_ROBUST_PRIO_INHERIT > 0) || \ (defined(_POSIX_THREAD_ROBUST_PRIO_PROTECT) && \ _POSIX_THREAD_ROBUST_PRIO_PROTECT > 0) || \ defined(PTHREAD_MUTEX_ROBUST) || defined(PTHREAD_MUTEX_ROBUST_NP)) && \ (!defined(__GLIBC__) || \ __GLIBC_PREREQ(2, 10) /* troubles with Robust mutexes before 2.10 */) #define MDBX_LOCKING MDBX_LOCKING_POSIX2008 #else #define MDBX_LOCKING MDBX_LOCKING_POSIX2001 #endif #elif defined(__sun) || defined(__SVR4) || defined(__svr4__) #define MDBX_LOCKING MDBX_LOCKING_POSIX1988 #else #define MDBX_LOCKING MDBX_LOCKING_SYSV #endif #define MDBX_LOCKING_CONFIG "AUTO=" STRINGIFY(MDBX_LOCKING) #else #define MDBX_LOCKING_CONFIG STRINGIFY(MDBX_LOCKING) #endif /* MDBX_LOCKING */ #endif /* !Windows */ /** Advanced: Using POSIX OFD-locks (autodetection by default). */ #ifndef MDBX_USE_OFDLOCKS #if defined(F_OFD_SETLK) && defined(F_OFD_SETLKW) && defined(F_OFD_GETLK) && \ !defined(MDBX_SAFE4QEMU) && \ !defined(__sun) /* OFD-lock are broken on Solaris */ #define MDBX_USE_OFDLOCKS 1 #else #define MDBX_USE_OFDLOCKS 0 #endif #define MDBX_USE_OFDLOCKS_CONFIG "AUTO=" STRINGIFY(MDBX_USE_OFDLOCKS) #else #define MDBX_USE_OFDLOCKS_CONFIG STRINGIFY(MDBX_USE_OFDLOCKS) #endif /* MDBX_USE_OFDLOCKS */ /** Advanced: Using sendfile() syscall (autodetection by default). */ #ifndef MDBX_USE_SENDFILE #if ((defined(__linux__) || defined(__gnu_linux__)) && \ !defined(__ANDROID_API__)) || \ (defined(__ANDROID_API__) && __ANDROID_API__ >= 21) #define MDBX_USE_SENDFILE 1 #else #define MDBX_USE_SENDFILE 0 #endif #endif /* MDBX_USE_SENDFILE */ /** Advanced: Using copy_file_range() syscall (autodetection by default). */ #ifndef MDBX_USE_COPYFILERANGE #if __GLIBC_PREREQ(2, 27) && defined(_GNU_SOURCE) #define MDBX_USE_COPYFILERANGE 1 #else #define MDBX_USE_COPYFILERANGE 0 #endif #endif /* MDBX_USE_COPYFILERANGE */ /** Advanced: Using sync_file_range() syscall (autodetection by default). */ #ifndef MDBX_USE_SYNCFILERANGE #if ((defined(__linux__) || defined(__gnu_linux__)) && \ defined(SYNC_FILE_RANGE_WRITE) && !defined(__ANDROID_API__)) || \ (defined(__ANDROID_API__) && __ANDROID_API__ >= 26) #define MDBX_USE_SYNCFILERANGE 1 #else #define MDBX_USE_SYNCFILERANGE 0 #endif #endif /* MDBX_USE_SYNCFILERANGE */ //------------------------------------------------------------------------------ #ifndef MDBX_CPU_WRITEBACK_INCOHERENT #if defined(__ia32__) || defined(__e2k__) || defined(__hppa) || \ defined(__hppa__) || defined(DOXYGEN) #define MDBX_CPU_WRITEBACK_INCOHERENT 0 #else #define MDBX_CPU_WRITEBACK_INCOHERENT 1 #endif #endif /* MDBX_CPU_WRITEBACK_INCOHERENT */ #ifndef MDBX_MMAP_INCOHERENT_FILE_WRITE #ifdef __OpenBSD__ #define MDBX_MMAP_INCOHERENT_FILE_WRITE 1 #else #define MDBX_MMAP_INCOHERENT_FILE_WRITE 0 #endif #endif /* MDBX_MMAP_INCOHERENT_FILE_WRITE */ #ifndef MDBX_MMAP_INCOHERENT_CPU_CACHE #if defined(__mips) || defined(__mips__) || defined(__mips64) || \ defined(__mips64__) || defined(_M_MRX000) || defined(_MIPS_) || \ defined(__MWERKS__) || defined(__sgi) /* MIPS has cache coherency issues. */ #define MDBX_MMAP_INCOHERENT_CPU_CACHE 1 #else /* LY: assume no relevant mmap/dcache issues. */ #define MDBX_MMAP_INCOHERENT_CPU_CACHE 0 #endif #endif /* MDBX_MMAP_INCOHERENT_CPU_CACHE */ #ifndef MDBX_64BIT_ATOMIC #if MDBX_WORDBITS >= 64 || defined(DOXYGEN) #define MDBX_64BIT_ATOMIC 1 #else #define MDBX_64BIT_ATOMIC 0 #endif #define MDBX_64BIT_ATOMIC_CONFIG "AUTO=" STRINGIFY(MDBX_64BIT_ATOMIC) #else #define MDBX_64BIT_ATOMIC_CONFIG STRINGIFY(MDBX_64BIT_ATOMIC) #endif /* MDBX_64BIT_ATOMIC */ #ifndef MDBX_64BIT_CAS #if defined(ATOMIC_LLONG_LOCK_FREE) #if ATOMIC_LLONG_LOCK_FREE > 1 #define MDBX_64BIT_CAS 1 #else #define MDBX_64BIT_CAS 0 #endif #elif defined(__GCC_ATOMIC_LLONG_LOCK_FREE) #if __GCC_ATOMIC_LLONG_LOCK_FREE > 1 #define MDBX_64BIT_CAS 1 #else #define MDBX_64BIT_CAS 0 #endif #elif defined(__CLANG_ATOMIC_LLONG_LOCK_FREE) #if __CLANG_ATOMIC_LLONG_LOCK_FREE > 1 #define MDBX_64BIT_CAS 1 #else #define MDBX_64BIT_CAS 0 #endif #elif defined(_MSC_VER) || defined(__APPLE__) || defined(DOXYGEN) #define MDBX_64BIT_CAS 1 #else #define MDBX_64BIT_CAS MDBX_64BIT_ATOMIC #endif #define MDBX_64BIT_CAS_CONFIG "AUTO=" STRINGIFY(MDBX_64BIT_CAS) #else #define MDBX_64BIT_CAS_CONFIG STRINGIFY(MDBX_64BIT_CAS) #endif /* MDBX_64BIT_CAS */ #if !defined(MDBX_UNALIGNED_OK) #if defined(_MSC_VER) #define MDBX_UNALIGNED_OK 1 /* avoid MSVC misoptimization */ #elif __CLANG_PREREQ(5, 0) || __GNUC_PREREQ(5, 0) #define MDBX_UNALIGNED_OK 0 /* expecting optimization is well done */ #elif (defined(__ia32__) || defined(__ARM_FEATURE_UNALIGNED)) && \ !defined(__ALIGNED__) #define MDBX_UNALIGNED_OK 1 #else #define MDBX_UNALIGNED_OK 0 #endif #endif /* MDBX_UNALIGNED_OK */ #ifndef MDBX_CACHELINE_SIZE #if defined(SYSTEM_CACHE_ALIGNMENT_SIZE) #define MDBX_CACHELINE_SIZE SYSTEM_CACHE_ALIGNMENT_SIZE #elif defined(__ia64__) || defined(__ia64) || defined(_M_IA64) #define MDBX_CACHELINE_SIZE 128 #else #define MDBX_CACHELINE_SIZE 64 #endif #endif /* MDBX_CACHELINE_SIZE */ /** @} end of build options */ /******************************************************************************* ******************************************************************************* ******************************************************************************/ /*----------------------------------------------------------------------------*/ /* Basic constants and types */ typedef union { volatile uint32_t weak; #ifdef MDBX_HAVE_C11ATOMICS volatile _Atomic uint32_t c11a; #endif /* MDBX_HAVE_C11ATOMICS */ } MDBX_atomic_uint32_t; typedef union { volatile uint64_t weak; #if defined(MDBX_HAVE_C11ATOMICS) && (MDBX_64BIT_CAS || MDBX_64BIT_ATOMIC) volatile _Atomic uint64_t c11a; #endif #if !defined(MDBX_HAVE_C11ATOMICS) || !MDBX_64BIT_CAS || !MDBX_64BIT_ATOMIC __anonymous_struct_extension__ struct { #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ MDBX_atomic_uint32_t low, high; #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ MDBX_atomic_uint32_t high, low; #else #error "FIXME: Unsupported byte order" #endif /* __BYTE_ORDER__ */ }; #endif } MDBX_atomic_uint64_t; /* The minimum number of keys required in a database page. * Setting this to a larger value will place a smaller bound on the * maximum size of a data item. Data items larger than this size will * be pushed into overflow pages instead of being stored directly in * the B-tree node. This value used to default to 4. With a page size * of 4096 bytes that meant that any item larger than 1024 bytes would * go into an overflow page. That also meant that on average 2-3KB of * each overflow page was wasted space. The value cannot be lower than * 2 because then there would no longer be a tree structure. With this * value, items larger than 2KB will go into overflow pages, and on * average only 1KB will be wasted. */ #define MDBX_MINKEYS 2 /* A stamp that identifies a file as an MDBX file. * There's nothing special about this value other than that it is easily * recognizable, and it will reflect any byte order mismatches. */ #define MDBX_MAGIC UINT64_C(/* 56-bit prime */ 0x59659DBDEF4C11) /* The version number for a database's datafile format. */ #define MDBX_DATA_VERSION 2 /* The version number for a database's lockfile format. */ #define MDBX_LOCK_VERSION 3 /* handle for the DB used to track free pages. */ #define FREE_DBI 0 /* handle for the default DB. */ #define MAIN_DBI 1 /* Number of DBs in metapage (free and main) - also hardcoded elsewhere */ #define CORE_DBS 2 /* Number of meta pages - also hardcoded elsewhere */ #define NUM_METAS 3 /* A page number in the database. * * MDBX uses 32 bit for page numbers. This limits database * size up to 2^44 bytes, in case of 4K pages. */ typedef uint32_t pgno_t; typedef MDBX_atomic_uint32_t atomic_pgno_t; #define PRIaPGNO PRIu32 #define MAX_PAGENO UINT32_C(0x7FFFffff) #define MIN_PAGENO NUM_METAS #define SAFE64_INVALID_THRESHOLD UINT64_C(0xffffFFFF00000000) /* A transaction ID. */ typedef uint64_t txnid_t; typedef MDBX_atomic_uint64_t atomic_txnid_t; #define PRIaTXN PRIi64 #define MIN_TXNID UINT64_C(1) #define MAX_TXNID (SAFE64_INVALID_THRESHOLD - 1) #define INITIAL_TXNID (MIN_TXNID + NUM_METAS - 1) #define INVALID_TXNID UINT64_MAX /* LY: for testing non-atomic 64-bit txnid on 32-bit arches. * #define MDBX_TXNID_STEP (UINT32_MAX / 3) */ #ifndef MDBX_TXNID_STEP #if MDBX_64BIT_CAS #define MDBX_TXNID_STEP 1u #else #define MDBX_TXNID_STEP 2u #endif #endif /* MDBX_TXNID_STEP */ /* Used for offsets within a single page. * Since memory pages are typically 4 or 8KB in size, 12-13 bits, * this is plenty. */ typedef uint16_t indx_t; #define MEGABYTE ((size_t)1 << 20) /*----------------------------------------------------------------------------*/ /* Core structures for database and shared memory (i.e. format definition) */ #pragma pack(push, 1) /* Information about a single database in the environment. */ typedef struct MDBX_db { uint16_t md_flags; /* see mdbx_dbi_open */ uint16_t md_depth; /* depth of this tree */ uint32_t md_xsize; /* key-size for MDBX_DUPFIXED (LEAF2 pages) */ pgno_t md_root; /* the root page of this tree */ pgno_t md_branch_pages; /* number of internal pages */ pgno_t md_leaf_pages; /* number of leaf pages */ pgno_t md_overflow_pages; /* number of overflow pages */ uint64_t md_seq; /* table sequence counter */ uint64_t md_entries; /* number of data items */ uint64_t md_mod_txnid; /* txnid of last committed modification */ } MDBX_db; /* database size-related parameters */ typedef struct mdbx_geo_t { uint16_t grow; /* datafile growth step in pages */ uint16_t shrink; /* datafile shrink threshold in pages */ pgno_t lower; /* minimal size of datafile in pages */ pgno_t upper; /* maximal size of datafile in pages */ pgno_t now; /* current size of datafile in pages */ pgno_t next; /* first unused page in the datafile, * but actually the file may be shorter. */ } mdbx_geo_t; /* Meta page content. * A meta page is the start point for accessing a database snapshot. * Pages 0-1 are meta pages. Transaction N writes meta page (N % 2). */ typedef struct MDBX_meta { /* Stamp identifying this as an MDBX file. * It must be set to MDBX_MAGIC with MDBX_DATA_VERSION. */ uint32_t mm_magic_and_version[2]; /* txnid that committed this page, the first of a two-phase-update pair */ uint32_t mm_txnid_a[2]; uint16_t mm_extra_flags; /* extra DB flags, zero (nothing) for now */ uint8_t mm_validator_id; /* ID of checksum and page validation method, * zero (nothing) for now */ uint8_t mm_extra_pagehdr; /* extra bytes in the page header, * zero (nothing) for now */ mdbx_geo_t mm_geo; /* database size-related parameters */ MDBX_db mm_dbs[CORE_DBS]; /* first is free space, 2nd is main db */ /* The size of pages used in this DB */ #define mm_psize mm_dbs[FREE_DBI].md_xsize /* Any persistent environment flags, see mdbx_env */ #define mm_flags mm_dbs[FREE_DBI].md_flags MDBX_canary mm_canary; #define MDBX_DATASIGN_NONE 0u #define MDBX_DATASIGN_WEAK 1u #define SIGN_IS_STEADY(sign) ((sign) > MDBX_DATASIGN_WEAK) #define META_IS_STEADY(meta) \ SIGN_IS_STEADY(unaligned_peek_u64(4, (meta)->mm_datasync_sign)) uint32_t mm_datasync_sign[2]; /* txnid that committed this page, the second of a two-phase-update pair */ uint32_t mm_txnid_b[2]; /* Number of non-meta pages which were put in GC after COW. May be 0 in case * DB was previously handled by libmdbx without corresponding feature. * This value in couple with mr_snapshot_pages_retired allows fast estimation * of "how much reader is restraining GC recycling". */ uint32_t mm_pages_retired[2]; /* The analogue /proc/sys/kernel/random/boot_id or similar to determine * whether the system was rebooted after the last use of the database files. * If there was no reboot, but there is no need to rollback to the last * steady sync point. Zeros mean that no relevant information is available * from the system. */ bin128_t mm_bootid; } MDBX_meta; /* Common header for all page types. The page type depends on mp_flags. * * P_BRANCH and P_LEAF pages have unsorted 'MDBX_node's at the end, with * sorted mp_ptrs[] entries referring to them. Exception: P_LEAF2 pages * omit mp_ptrs and pack sorted MDBX_DUPFIXED values after the page header. * * P_OVERFLOW records occupy one or more contiguous pages where only the * first has a page header. They hold the real data of F_BIGDATA nodes. * * P_SUBP sub-pages are small leaf "pages" with duplicate data. * A node with flag F_DUPDATA but not F_SUBDATA contains a sub-page. * (Duplicate data can also go in sub-databases, which use normal pages.) * * P_META pages contain MDBX_meta, the start point of an MDBX snapshot. * * Each non-metapage up to MDBX_meta.mm_last_pg is reachable exactly once * in the snapshot: Either used by a database or listed in a GC record. */ typedef struct MDBX_page { union { struct MDBX_page *mp_next; /* for in-memory list of freed pages */ uint64_t mp_txnid; /* txnid that committed this page */ }; uint16_t mp_leaf2_ksize; /* key size if this is a LEAF2 page */ #define P_BRANCH 0x01 /* branch page */ #define P_LEAF 0x02 /* leaf page */ #define P_OVERFLOW 0x04 /* overflow page */ #define P_META 0x08 /* meta page */ #define P_DIRTY 0x10 /* dirty page, also set for P_SUBP pages */ #define P_LEAF2 0x20 /* for MDBX_DUPFIXED records */ #define P_SUBP 0x40 /* for MDBX_DUPSORT sub-pages */ #define P_LOOSE 0x4000 /* page was dirtied then freed, can be reused */ #define P_KEEP 0x8000 /* leave this page alone during spill */ uint16_t mp_flags; union { __anonymous_struct_extension__ struct { indx_t mp_lower; /* lower bound of free space */ indx_t mp_upper; /* upper bound of free space */ }; uint32_t mp_pages; /* number of overflow pages */ }; pgno_t mp_pgno; /* page number */ #if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || \ (!defined(__cplusplus) && defined(_MSC_VER)) indx_t mp_ptrs[] /* dynamic size */; #endif /* C99 */ } MDBX_page; /* Size of the page header, excluding dynamic data at the end */ #define PAGEHDRSZ ((unsigned)offsetof(MDBX_page, mp_ptrs)) #pragma pack(pop) #if MDBX_LOCKING == MDBX_LOCKING_WIN32FILES #define MDBX_CLOCK_SIGN UINT32_C(0xF10C) typedef void mdbx_ipclock_t; #elif MDBX_LOCKING == MDBX_LOCKING_SYSV #define MDBX_CLOCK_SIGN UINT32_C(0xF18D) typedef mdbx_pid_t mdbx_ipclock_t; #ifndef EOWNERDEAD #define EOWNERDEAD MDBX_RESULT_TRUE #endif #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 #define MDBX_CLOCK_SIGN UINT32_C(0x8017) typedef pthread_mutex_t mdbx_ipclock_t; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988 #define MDBX_CLOCK_SIGN UINT32_C(0xFC29) typedef sem_t mdbx_ipclock_t; #else #error "FIXME" #endif /* MDBX_LOCKING */ #if MDBX_LOCKING > MDBX_LOCKING_SYSV && !defined(__cplusplus) MDBX_INTERNAL_FUNC int mdbx_ipclock_stub(mdbx_ipclock_t *ipc); MDBX_INTERNAL_FUNC int mdbx_ipclock_destroy(mdbx_ipclock_t *ipc); #endif /* MDBX_LOCKING */ /* Reader Lock Table * * Readers don't acquire any locks for their data access. Instead, they * simply record their transaction ID in the reader table. The reader * mutex is needed just to find an empty slot in the reader table. The * slot's address is saved in thread-specific data so that subsequent * read transactions started by the same thread need no further locking to * proceed. * * If MDBX_NOTLS is set, the slot address is not saved in thread-specific data. * No reader table is used if the database is on a read-only filesystem. * * Since the database uses multi-version concurrency control, readers don't * actually need any locking. This table is used to keep track of which * readers are using data from which old transactions, so that we'll know * when a particular old transaction is no longer in use. Old transactions * that have discarded any data pages can then have those pages reclaimed * for use by a later write transaction. * * The lock table is constructed such that reader slots are aligned with the * processor's cache line size. Any slot is only ever used by one thread. * This alignment guarantees that there will be no contention or cache * thrashing as threads update their own slot info, and also eliminates * any need for locking when accessing a slot. * * A writer thread will scan every slot in the table to determine the oldest * outstanding reader transaction. Any freed pages older than this will be * reclaimed by the writer. The writer doesn't use any locks when scanning * this table. This means that there's no guarantee that the writer will * see the most up-to-date reader info, but that's not required for correct * operation - all we need is to know the upper bound on the oldest reader, * we don't care at all about the newest reader. So the only consequence of * reading stale information here is that old pages might hang around a * while longer before being reclaimed. That's actually good anyway, because * the longer we delay reclaiming old pages, the more likely it is that a * string of contiguous pages can be found after coalescing old pages from * many old transactions together. */ /* The actual reader record, with cacheline padding. */ typedef struct MDBX_reader { /* Current Transaction ID when this transaction began, or (txnid_t)-1. * Multiple readers that start at the same time will probably have the * same ID here. Again, it's not important to exclude them from * anything; all we need to know is which version of the DB they * started from so we can avoid overwriting any data used in that * particular version. */ MDBX_atomic_uint64_t /* txnid_t */ mr_txnid; /* The information we store in a single slot of the reader table. * In addition to a transaction ID, we also record the process and * thread ID that owns a slot, so that we can detect stale information, * e.g. threads or processes that went away without cleaning up. * * NOTE: We currently don't check for stale records. * We simply re-init the table when we know that we're the only process * opening the lock file. */ /* The thread ID of the thread owning this txn. */ MDBX_atomic_uint64_t mr_tid; /* The process ID of the process owning this reader txn. */ MDBX_atomic_uint32_t mr_pid; /* The number of pages used in the reader's MVCC snapshot, * i.e. the value of meta->mm_geo.next and txn->mt_next_pgno */ atomic_pgno_t mr_snapshot_pages_used; /* Number of retired pages at the time this reader starts transaction. So, * at any time the difference mm_pages_retired - mr_snapshot_pages_retired * will give the number of pages which this reader restraining from reuse. */ MDBX_atomic_uint64_t mr_snapshot_pages_retired; } MDBX_reader; /* The header for the reader table (a memory-mapped lock file). */ typedef struct MDBX_lockinfo { /* Stamp identifying this as an MDBX file. * It must be set to MDBX_MAGIC with with MDBX_LOCK_VERSION. */ uint64_t mti_magic_and_version; /* Format of this lock file. Must be set to MDBX_LOCK_FORMAT. */ uint32_t mti_os_and_format; /* Flags which environment was opened. */ MDBX_atomic_uint32_t mti_envmode; /* Threshold of un-synced-with-disk pages for auto-sync feature, * zero means no-threshold, i.e. auto-sync is disabled. */ atomic_pgno_t mti_autosync_threshold; /* Low 32-bit of txnid with which meta-pages was synced, * i.e. for sync-polling in the MDBX_NOMETASYNC mode. */ MDBX_atomic_uint32_t mti_meta_sync_txnid; /* Period for timed auto-sync feature, i.e. at the every steady checkpoint * the mti_unsynced_timeout sets to the current_time + mti_autosync_period. * The time value is represented in a suitable system-dependent form, for * example clock_gettime(CLOCK_BOOTTIME) or clock_gettime(CLOCK_MONOTONIC). * Zero means timed auto-sync is disabled. */ MDBX_atomic_uint64_t mti_autosync_period; /* Marker to distinguish uniqueness of DB/CLK.*/ MDBX_atomic_uint64_t mti_bait_uniqueness; alignas(MDBX_CACHELINE_SIZE) /* cacheline ---------------------------------*/ /* Write transaction lock. */ #if MDBX_LOCKING > 0 mdbx_ipclock_t mti_wlock; #endif /* MDBX_LOCKING > 0 */ atomic_txnid_t mti_oldest_reader; /* Timestamp of the last steady sync. Value is represented in a suitable * system-dependent form, for example clock_gettime(CLOCK_BOOTTIME) or * clock_gettime(CLOCK_MONOTONIC). */ MDBX_atomic_uint64_t mti_sync_timestamp; /* Number un-synced-with-disk pages for auto-sync feature. */ atomic_pgno_t mti_unsynced_pages; /* Number of page which was discarded last time by madvise(MADV_FREE). */ atomic_pgno_t mti_discarded_tail; /* Timestamp of the last readers check. */ MDBX_atomic_uint64_t mti_reader_check_timestamp; alignas(MDBX_CACHELINE_SIZE) /* cacheline ---------------------------------*/ /* Readeaders registration lock. */ #if MDBX_LOCKING > 0 mdbx_ipclock_t mti_rlock; #endif /* MDBX_LOCKING > 0 */ /* The number of slots that have been used in the reader table. * This always records the maximum count, it is not decremented * when readers release their slots. */ MDBX_atomic_uint32_t mti_numreaders; MDBX_atomic_uint32_t mti_readers_refresh_flag; #if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || \ (!defined(__cplusplus) && defined(_MSC_VER)) alignas(MDBX_CACHELINE_SIZE) /* cacheline ---------------------------------*/ MDBX_reader mti_readers[] /* dynamic size */; #endif /* C99 */ } MDBX_lockinfo; /* Lockfile format signature: version, features and field layout */ #define MDBX_LOCK_FORMAT \ (MDBX_CLOCK_SIGN * 27733 + (unsigned)sizeof(MDBX_reader) * 13 + \ (unsigned)offsetof(MDBX_reader, mr_snapshot_pages_used) * 251 + \ (unsigned)offsetof(MDBX_lockinfo, mti_oldest_reader) * 83 + \ (unsigned)offsetof(MDBX_lockinfo, mti_numreaders) * 37 + \ (unsigned)offsetof(MDBX_lockinfo, mti_readers) * 29) #define MDBX_DATA_MAGIC \ ((MDBX_MAGIC << 8) + MDBX_PNL_ASCENDING * 64 + MDBX_DATA_VERSION) #define MDBX_DATA_MAGIC_DEVEL ((MDBX_MAGIC << 8) + 255) #define MDBX_LOCK_MAGIC ((MDBX_MAGIC << 8) + MDBX_LOCK_VERSION) #ifndef MDBX_ASSUME_MALLOC_OVERHEAD #define MDBX_ASSUME_MALLOC_OVERHEAD (sizeof(void *) * 2u) #endif /* MDBX_ASSUME_MALLOC_OVERHEAD */ /* The maximum size of a database page. * * It is 64K, but value-PAGEHDRSZ must fit in MDBX_page.mp_upper. * * MDBX will use database pages < OS pages if needed. * That causes more I/O in write transactions: The OS must * know (read) the whole page before writing a partial page. * * Note that we don't currently support Huge pages. On Linux, * regular data files cannot use Huge pages, and in general * Huge pages aren't actually pageable. We rely on the OS * demand-pager to read our data and page it out when memory * pressure from other processes is high. So until OSs have * actual paging support for Huge pages, they're not viable. */ #define MAX_PAGESIZE MDBX_MAX_PAGESIZE #define MIN_PAGESIZE MDBX_MIN_PAGESIZE #define MIN_MAPSIZE (MIN_PAGESIZE * MIN_PAGENO) #if defined(_WIN32) || defined(_WIN64) #define MAX_MAPSIZE32 UINT32_C(0x38000000) #else #define MAX_MAPSIZE32 UINT32_C(0x7f000000) #endif #define MAX_MAPSIZE64 (MAX_PAGENO * (uint64_t)MAX_PAGESIZE) #if MDBX_WORDBITS >= 64 #define MAX_MAPSIZE MAX_MAPSIZE64 #define MDBX_READERS_LIMIT \ ((MAX_PAGESIZE - sizeof(MDBX_lockinfo)) / sizeof(MDBX_reader)) #define MDBX_PGL_LIMIT MAX_PAGENO #else #define MDBX_READERS_LIMIT 1024 #define MAX_MAPSIZE MAX_MAPSIZE32 #define MDBX_PGL_LIMIT (MAX_MAPSIZE32 / MIN_PAGESIZE) #endif /* MDBX_WORDBITS */ /*----------------------------------------------------------------------------*/ /* An PNL is an Page Number List, a sorted array of IDs. * The first element of the array is a counter for how many actual page-numbers * are in the list. By default PNLs are sorted in descending order, this allow * cut off a page with lowest pgno (at the tail) just truncating the list. The * sort order of PNLs is controlled by the MDBX_PNL_ASCENDING build option. */ typedef pgno_t *MDBX_PNL; #if MDBX_PNL_ASCENDING #define MDBX_PNL_ORDERED(first, last) ((first) < (last)) #define MDBX_PNL_DISORDERED(first, last) ((first) >= (last)) #else #define MDBX_PNL_ORDERED(first, last) ((first) > (last)) #define MDBX_PNL_DISORDERED(first, last) ((first) <= (last)) #endif /* List of txnid, only for MDBX_txn.tw.lifo_reclaimed */ typedef txnid_t *MDBX_TXL; /* An Dirty-Page list item is an pgno/pointer pair. */ typedef struct MDBX_dp { pgno_t pgno; MDBX_page *ptr; } MDBX_dp; /* An DPL (dirty-page list) is a sorted array of MDBX_DPs. */ typedef struct MDBX_dpl { unsigned sorted; unsigned length; unsigned allocated; #if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || \ (!defined(__cplusplus) && defined(_MSC_VER)) MDBX_dp items[] /* dynamic size with holes at zero and after the last */; #endif } MDBX_dpl; /* PNL sizes */ #define MDBX_PNL_GRANULATE 1024 #define MDBX_PNL_INITIAL \ (MDBX_PNL_GRANULATE - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(pgno_t)) #define MDBX_TXL_GRANULATE 32 #define MDBX_TXL_INITIAL \ (MDBX_TXL_GRANULATE - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(txnid_t)) #define MDBX_TXL_MAX \ ((1u << 17) - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(txnid_t)) #define MDBX_PNL_ALLOCLEN(pl) ((pl)[-1]) #define MDBX_PNL_SIZE(pl) ((pl)[0]) #define MDBX_PNL_FIRST(pl) ((pl)[1]) #define MDBX_PNL_LAST(pl) ((pl)[MDBX_PNL_SIZE(pl)]) #define MDBX_PNL_BEGIN(pl) (&(pl)[1]) #define MDBX_PNL_END(pl) (&(pl)[MDBX_PNL_SIZE(pl) + 1]) #if MDBX_PNL_ASCENDING #define MDBX_PNL_LEAST(pl) MDBX_PNL_FIRST(pl) #define MDBX_PNL_MOST(pl) MDBX_PNL_LAST(pl) #else #define MDBX_PNL_LEAST(pl) MDBX_PNL_LAST(pl) #define MDBX_PNL_MOST(pl) MDBX_PNL_FIRST(pl) #endif #define MDBX_PNL_SIZEOF(pl) ((MDBX_PNL_SIZE(pl) + 1) * sizeof(pgno_t)) #define MDBX_PNL_IS_EMPTY(pl) (MDBX_PNL_SIZE(pl) == 0) /*----------------------------------------------------------------------------*/ /* Internal structures */ /* Auxiliary DB info. * The information here is mostly static/read-only. There is * only a single copy of this record in the environment. */ typedef struct MDBX_dbx { MDBX_val md_name; /* name of the database */ MDBX_cmp_func *md_cmp; /* function for comparing keys */ MDBX_cmp_func *md_dcmp; /* function for comparing data items */ size_t md_klen_min, md_klen_max; /* min/max key length for the database */ size_t md_vlen_min, md_vlen_max; /* min/max value/data length for the database */ } MDBX_dbx; /* A database transaction. * Every operation requires a transaction handle. */ struct MDBX_txn { #define MDBX_MT_SIGNATURE UINT32_C(0x93D53A31) size_t mt_signature; MDBX_txn *mt_parent; /* parent of a nested txn */ /* Nested txn under this txn, set together with flag MDBX_TXN_HAS_CHILD */ MDBX_txn *mt_child; mdbx_geo_t mt_geo; /* next unallocated page */ #define mt_next_pgno mt_geo.next /* corresponding to the current size of datafile */ #define mt_end_pgno mt_geo.now /* Transaction Flags */ /* mdbx_txn_begin() flags */ #define MDBX_TXN_RO_BEGIN_FLAGS (MDBX_TXN_RDONLY | MDBX_TXN_RDONLY_PREPARE) #define MDBX_TXN_RW_BEGIN_FLAGS \ (MDBX_TXN_NOMETASYNC | MDBX_TXN_NOSYNC | MDBX_TXN_TRY) /* Additional flag for mdbx_sync_locked() */ #define MDBX_SHRINK_ALLOWED UINT32_C(0x40000000) /* internal txn flags */ #define MDBX_TXN_FINISHED 0x01 /* txn is finished or never began */ #define MDBX_TXN_ERROR 0x02 /* txn is unusable after an error */ #define MDBX_TXN_DIRTY 0x04 /* must write, even if dirty list is empty */ #define MDBX_TXN_SPILLS 0x08 /* txn or a parent has spilled pages */ #define MDBX_TXN_HAS_CHILD 0x10 /* txn has an MDBX_txn.mt_child */ /* most operations on the txn are currently illegal */ #define MDBX_TXN_BLOCKED \ (MDBX_TXN_FINISHED | MDBX_TXN_ERROR | MDBX_TXN_HAS_CHILD) #define TXN_FLAGS \ (MDBX_TXN_FINISHED | MDBX_TXN_ERROR | MDBX_TXN_DIRTY | MDBX_TXN_SPILLS | \ MDBX_TXN_HAS_CHILD) #if (TXN_FLAGS & (MDBX_TXN_RW_BEGIN_FLAGS | MDBX_TXN_RO_BEGIN_FLAGS)) || \ ((MDBX_TXN_RW_BEGIN_FLAGS | MDBX_TXN_RO_BEGIN_FLAGS | TXN_FLAGS) & \ MDBX_SHRINK_ALLOWED) #error "Oops, some flags overlapped or wrong" #endif unsigned mt_flags; /* The ID of this transaction. IDs are integers incrementing from 1. * Only committed write transactions increment the ID. If a transaction * aborts, the ID may be re-used by the next writer. */ txnid_t mt_txnid; MDBX_env *mt_env; /* the DB environment */ /* Array of records for each DB known in the environment. */ MDBX_dbx *mt_dbxs; /* Array of MDBX_db records for each known DB */ MDBX_db *mt_dbs; /* Array of sequence numbers for each DB handle */ unsigned *mt_dbiseqs; /* Transaction DBI Flags */ #define DBI_DIRTY MDBX_DBI_DIRTY /* DB was written in this txn */ #define DBI_STALE MDBX_DBI_STALE /* Named-DB record is older than txnID */ #define DBI_FRESH MDBX_DBI_FRESH /* Named-DB handle opened in this txn */ #define DBI_CREAT MDBX_DBI_CREAT /* Named-DB handle created in this txn */ #define DBI_VALID 0x10 /* DB handle is valid, see also DB_VALID */ #define DBI_USRVALID 0x20 /* As DB_VALID, but not set for FREE_DBI */ #define DBI_DUPDATA 0x40 /* DB is MDBX_DUPSORT data */ #define DBI_AUDITED 0x80 /* Internal flag for accounting during audit */ /* Array of flags for each DB */ uint8_t *mt_dbistate; /* Number of DB records in use, or 0 when the txn is finished. * This number only ever increments until the txn finishes; we * don't decrement it when individual DB handles are closed. */ MDBX_dbi mt_numdbs; size_t mt_owner; /* thread ID that owns this transaction */ MDBX_canary mt_canary; void *mt_userctx; /* User-settable context */ union { struct { /* For read txns: This thread/txn's reader table slot, or NULL. */ MDBX_reader *reader; } to; struct { /* In write txns, array of cursors for each DB */ MDBX_cursor **cursors; pgno_t *reclaimed_pglist; /* Reclaimed GC pages */ txnid_t last_reclaimed; /* ID of last used record */ #if MDBX_ENABLE_REFUND pgno_t loose_refund_wl /* FIXME: describe */; #endif /* MDBX_ENABLE_REFUND */ /* dirtylist room: Dirty array size - dirty pages visible to this txn. * Includes ancestor txns' dirty pages not hidden by other txns' * dirty/spilled pages. Thus commit(nested txn) has room to merge * dirtylist into mt_parent after freeing hidden mt_parent pages. */ unsigned dirtyroom; /* For write txns: Modified pages. Sorted when not MDBX_WRITEMAP. */ MDBX_dpl *dirtylist; /* The list of reclaimed txns from GC */ MDBX_TXL lifo_reclaimed; /* The list of pages that became unused during this transaction. */ MDBX_PNL retired_pages; /* The list of loose pages that became unused and may be reused * in this transaction, linked through `mp_next`. */ MDBX_page *loose_pages; /* Number of loose pages (tw.loose_pages) */ unsigned loose_count; /* The sorted list of dirty pages we temporarily wrote to disk * because the dirty list was full. page numbers in here are * shifted left by 1, deleted slots have the LSB set. */ MDBX_PNL spill_pages; unsigned spill_least_removed; } tw; }; }; #if MDBX_WORDBITS >= 64 #define CURSOR_STACK 32 #else #define CURSOR_STACK 24 #endif struct MDBX_xcursor; /* Cursors are used for all DB operations. * A cursor holds a path of (page pointer, key index) from the DB * root to a position in the DB, plus other state. MDBX_DUPSORT * cursors include an xcursor to the current data item. Write txns * track their cursors and keep them up to date when data moves. * Exception: An xcursor's pointer to a P_SUBP page can be stale. * (A node with F_DUPDATA but no F_SUBDATA contains a subpage). */ struct MDBX_cursor { #define MDBX_MC_LIVE UINT32_C(0xFE05D5B1) #define MDBX_MC_READY4CLOSE UINT32_C(0x2817A047) #define MDBX_MC_WAIT4EOT UINT32_C(0x90E297A7) uint32_t mc_signature; /* The database handle this cursor operates on */ MDBX_dbi mc_dbi; /* Next cursor on this DB in this txn */ MDBX_cursor *mc_next; /* Backup of the original cursor if this cursor is a shadow */ MDBX_cursor *mc_backup; /* Context used for databases with MDBX_DUPSORT, otherwise NULL */ struct MDBX_xcursor *mc_xcursor; /* The transaction that owns this cursor */ MDBX_txn *mc_txn; /* The database record for this cursor */ MDBX_db *mc_db; /* The database auxiliary record for this cursor */ MDBX_dbx *mc_dbx; /* The mt_dbistate for this database */ uint8_t *mc_dbistate; unsigned mc_snum; /* number of pushed pages */ unsigned mc_top; /* index of top page, normally mc_snum-1 */ /* Cursor state flags. */ #define C_INITIALIZED 0x01 /* cursor has been initialized and is valid */ #define C_EOF 0x02 /* No more data */ #define C_SUB 0x04 /* Cursor is a sub-cursor */ #define C_DEL 0x08 /* last op was a cursor_del */ #define C_UNTRACK 0x10 /* Un-track cursor when closing */ #define C_RECLAIMING 0x20 /* GC lookup is prohibited */ #define C_GCFREEZE 0x40 /* reclaimed_pglist must not be updated */ /* Cursor checking flags. */ #define C_COPYING 0x100 /* skip key-value length check (copying simplify) */ #define C_UPDATING 0x200 /* update/rebalance pending */ #define C_RETIRING 0x400 /* refs to child pages may be invalid */ #define C_SKIPORD 0x800 /* don't check keys ordering */ unsigned mc_flags; /* see mdbx_cursor */ MDBX_page *mc_pg[CURSOR_STACK]; /* stack of pushed pages */ indx_t mc_ki[CURSOR_STACK]; /* stack of page indices */ }; /* Context for sorted-dup records. * We could have gone to a fully recursive design, with arbitrarily * deep nesting of sub-databases. But for now we only handle these * levels - main DB, optional sub-DB, sorted-duplicate DB. */ typedef struct MDBX_xcursor { /* A sub-cursor for traversing the Dup DB */ MDBX_cursor mx_cursor; /* The database record for this Dup DB */ MDBX_db mx_db; /* The auxiliary DB record for this Dup DB */ MDBX_dbx mx_dbx; /* The mt_dbistate for this Dup DB */ uint8_t mx_dbistate; } MDBX_xcursor; typedef struct MDBX_cursor_couple { MDBX_cursor outer; void *mc_userctx; /* User-settable context */ MDBX_xcursor inner; } MDBX_cursor_couple; /* The database environment. */ struct MDBX_env { #define MDBX_ME_SIGNATURE UINT32_C(0x9A899641) MDBX_atomic_uint32_t me_signature; /* Failed to update the meta page. Probably an I/O error. */ #define MDBX_FATAL_ERROR UINT32_C(0x80000000) /* Some fields are initialized. */ #define MDBX_ENV_ACTIVE UINT32_C(0x20000000) /* me_txkey is set */ #define MDBX_ENV_TXKEY UINT32_C(0x10000000) /* Legacy MDBX_MAPASYNC (prior v0.9) */ #define MDBX_DEPRECATED_MAPASYNC UINT32_C(0x100000) #define ENV_INTERNAL_FLAGS (MDBX_FATAL_ERROR | MDBX_ENV_ACTIVE | MDBX_ENV_TXKEY) uint32_t me_flags; mdbx_mmap_t me_dxb_mmap; /* The main data file */ #define me_map me_dxb_mmap.dxb #define me_lazy_fd me_dxb_mmap.fd mdbx_filehandle_t me_dsync_fd; mdbx_mmap_t me_lck_mmap; /* The lock file */ #define me_lfd me_lck_mmap.fd #define me_lck me_lck_mmap.lck unsigned me_psize; /* DB page size, initialized from me_os_psize */ uint8_t me_psize2log; /* log2 of DB page size */ int8_t me_stuck_meta; /* recovery-only: target meta page or less that zero */ unsigned me_os_psize; /* OS page size, from mdbx_syspagesize() */ unsigned me_maxreaders; /* size of the reader table */ mdbx_fastmutex_t me_dbi_lock; MDBX_dbi me_numdbs; /* number of DBs opened */ MDBX_dbi me_maxdbs; /* size of the DB table */ uint32_t me_pid; /* process ID of this env */ mdbx_thread_key_t me_txkey; /* thread-key for readers */ char *me_pathname; /* path to the DB files */ void *me_pbuf; /* scratch area for DUPSORT put() */ MDBX_txn *me_txn; /* current write transaction */ MDBX_txn *me_txn0; /* prealloc'd write transaction */ /* write-txn lock */ #if MDBX_LOCKING == MDBX_LOCKING_SYSV union { key_t key; int semid; } me_sysv_ipc; #endif /* MDBX_LOCKING == MDBX_LOCKING_SYSV */ #if MDBX_LOCKING > 0 mdbx_ipclock_t *me_wlock; #endif /* MDBX_LOCKING > 0 */ MDBX_dbx *me_dbxs; /* array of static DB info */ uint16_t *me_dbflags; /* array of flags from MDBX_db.md_flags */ unsigned *me_dbiseqs; /* array of dbi sequence numbers */ atomic_txnid_t *me_oldest; /* ID of oldest reader last time we looked */ MDBX_page *me_dp_reserve; /* list of malloc'd blocks for re-use */ /* PNL of pages that became unused in a write txn */ MDBX_PNL me_retired_pages; /* Number of freelist items that can fit in a single overflow page */ unsigned me_maxgc_ov1page; unsigned me_branch_nodemax; /* max size of a branch-node */ uint32_t me_live_reader; /* have liveness lock in reader table */ void *me_userctx; /* User-settable context */ MDBX_atomic_uint64_t *me_sync_timestamp; MDBX_atomic_uint64_t *me_autosync_period; atomic_pgno_t *me_unsynced_pages; atomic_pgno_t *me_autosync_threshold; atomic_pgno_t *me_discarded_tail; MDBX_atomic_uint32_t *me_meta_sync_txnid; MDBX_hsr_func *me_hsr_callback; /* Callback for kicking laggard readers */ unsigned me_dp_reserve_len; struct { unsigned dp_reserve_limit; unsigned rp_augment_limit; unsigned dp_limit; unsigned dp_initial; uint8_t dp_loose_limit; uint8_t spill_max_denominator; uint8_t spill_min_denominator; uint8_t spill_parent4child_denominator; } me_options; struct { #if MDBX_LOCKING > 0 mdbx_ipclock_t wlock; #endif /* MDBX_LOCKING > 0 */ atomic_txnid_t oldest; MDBX_atomic_uint64_t sync_timestamp; MDBX_atomic_uint64_t autosync_period; atomic_pgno_t autosync_pending; atomic_pgno_t autosync_threshold; atomic_pgno_t discarded_tail; MDBX_atomic_uint32_t meta_sync_txnid; } me_lckless_stub; #if MDBX_DEBUG MDBX_assert_func *me_assert_func; /* Callback for assertion failures */ #endif #ifdef MDBX_USE_VALGRIND int me_valgrind_handle; #endif #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) pgno_t me_poison_edge; #endif /* MDBX_USE_VALGRIND || __SANITIZE_ADDRESS__ */ MDBX_env *me_lcklist_next; /* struct me_dbgeo used for accepting db-geo params from user for the new * database creation, i.e. when mdbx_env_set_geometry() was called before * mdbx_env_open(). */ struct { size_t lower; /* minimal size of datafile */ size_t upper; /* maximal size of datafile */ size_t now; /* current size of datafile */ size_t grow; /* step to grow datafile */ size_t shrink; /* threshold to shrink datafile */ } me_dbgeo; #if defined(_WIN32) || defined(_WIN64) MDBX_srwlock me_remap_guard; /* Workaround for LockFileEx and WriteFile multithread bug */ CRITICAL_SECTION me_windowsbug_lock; #else mdbx_fastmutex_t me_remap_guard; #endif }; #ifndef __cplusplus /*----------------------------------------------------------------------------*/ /* Debug and Logging stuff */ #define MDBX_RUNTIME_FLAGS_INIT \ ((MDBX_DEBUG) > 0) * MDBX_DBG_ASSERT + ((MDBX_DEBUG) > 1) * MDBX_DBG_AUDIT extern uint8_t mdbx_runtime_flags; extern uint8_t mdbx_loglevel; extern MDBX_debug_func *mdbx_debug_logger; MDBX_INTERNAL_FUNC void MDBX_PRINTF_ARGS(4, 5) mdbx_debug_log(int level, const char *function, int line, const char *fmt, ...) MDBX_PRINTF_ARGS(4, 5); MDBX_INTERNAL_FUNC void mdbx_debug_log_va(int level, const char *function, int line, const char *fmt, va_list args); #define mdbx_log_enabled(msg) unlikely(msg <= mdbx_loglevel) #if MDBX_DEBUG #define mdbx_assert_enabled() unlikely(mdbx_runtime_flags &MDBX_DBG_ASSERT) #define mdbx_audit_enabled() unlikely(mdbx_runtime_flags &MDBX_DBG_AUDIT) #else /* MDBX_DEBUG */ #define mdbx_audit_enabled() (0) #if !defined(NDEBUG) || defined(MDBX_FORCE_ASSERTIONS) #define mdbx_assert_enabled() (1) #else #define mdbx_assert_enabled() (0) #endif /* NDEBUG */ #endif /* MDBX_DEBUG */ #if !MDBX_DEBUG && defined(__ANDROID_API__) #define mdbx_assert_fail(env, msg, func, line) \ __android_log_assert(msg, "mdbx", "%s:%u", func, line) #else void mdbx_assert_fail(const MDBX_env *env, const char *msg, const char *func, int line); #endif #define mdbx_debug_extra(fmt, ...) \ do { \ if (MDBX_DEBUG && mdbx_log_enabled(MDBX_LOG_EXTRA)) \ mdbx_debug_log(MDBX_LOG_EXTRA, __func__, __LINE__, fmt, __VA_ARGS__); \ } while (0) #define mdbx_debug_extra_print(fmt, ...) \ do { \ if (MDBX_DEBUG && mdbx_log_enabled(MDBX_LOG_EXTRA)) \ mdbx_debug_log(MDBX_LOG_EXTRA, NULL, 0, fmt, __VA_ARGS__); \ } while (0) #define mdbx_trace(fmt, ...) \ do { \ if (MDBX_DEBUG && mdbx_log_enabled(MDBX_LOG_TRACE)) \ mdbx_debug_log(MDBX_LOG_TRACE, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_debug(fmt, ...) \ do { \ if (MDBX_DEBUG && mdbx_log_enabled(MDBX_LOG_DEBUG)) \ mdbx_debug_log(MDBX_LOG_DEBUG, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_verbose(fmt, ...) \ do { \ if (MDBX_DEBUG && mdbx_log_enabled(MDBX_LOG_VERBOSE)) \ mdbx_debug_log(MDBX_LOG_VERBOSE, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_notice(fmt, ...) \ do { \ if (mdbx_log_enabled(MDBX_LOG_NOTICE)) \ mdbx_debug_log(MDBX_LOG_NOTICE, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_warning(fmt, ...) \ do { \ if (mdbx_log_enabled(MDBX_LOG_WARN)) \ mdbx_debug_log(MDBX_LOG_WARN, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_error(fmt, ...) \ do { \ if (mdbx_log_enabled(MDBX_LOG_ERROR)) \ mdbx_debug_log(MDBX_LOG_ERROR, __func__, __LINE__, fmt "\n", \ __VA_ARGS__); \ } while (0) #define mdbx_fatal(fmt, ...) \ mdbx_debug_log(MDBX_LOG_FATAL, __func__, __LINE__, fmt "\n", __VA_ARGS__); #define mdbx_ensure_msg(env, expr, msg) \ do { \ if (unlikely(!(expr))) \ mdbx_assert_fail(env, msg, __func__, __LINE__); \ } while (0) #define mdbx_ensure(env, expr) mdbx_ensure_msg(env, expr, #expr) /* assert(3) variant in environment context */ #define mdbx_assert(env, expr) \ do { \ if (mdbx_assert_enabled()) \ mdbx_ensure(env, expr); \ } while (0) /* assert(3) variant in cursor context */ #define mdbx_cassert(mc, expr) mdbx_assert((mc)->mc_txn->mt_env, expr) /* assert(3) variant in transaction context */ #define mdbx_tassert(txn, expr) mdbx_assert((txn)->mt_env, expr) #ifndef MDBX_TOOLS /* Avoid using internal mdbx_assert() */ #undef assert #define assert(expr) mdbx_assert(NULL, expr) #endif /*----------------------------------------------------------------------------*/ /* Cache coherence and mmap invalidation */ #if MDBX_CPU_WRITEBACK_INCOHERENT #define mdbx_flush_incoherent_cpu_writeback() mdbx_memory_barrier() #else #define mdbx_flush_incoherent_cpu_writeback() mdbx_compiler_barrier() #endif /* MDBX_CPU_WRITEBACK_INCOHERENT */ static __maybe_unused __inline void mdbx_flush_incoherent_mmap(void *addr, size_t nbytes, const intptr_t pagesize) { #if MDBX_MMAP_INCOHERENT_FILE_WRITE char *const begin = (char *)(-pagesize & (intptr_t)addr); char *const end = (char *)(-pagesize & (intptr_t)((char *)addr + nbytes + pagesize - 1)); int err = msync(begin, end - begin, MS_SYNC | MS_INVALIDATE) ? errno : 0; mdbx_assert(nullptr, err == 0); (void)err; #else (void)pagesize; #endif /* MDBX_MMAP_INCOHERENT_FILE_WRITE */ #if MDBX_MMAP_INCOHERENT_CPU_CACHE #ifdef DCACHE /* MIPS has cache coherency issues. * Note: for any nbytes >= on-chip cache size, entire is flushed. */ cacheflush(addr, nbytes, DCACHE); #else #error "Oops, cacheflush() not available" #endif /* DCACHE */ #endif /* MDBX_MMAP_INCOHERENT_CPU_CACHE */ #if !MDBX_MMAP_INCOHERENT_FILE_WRITE && !MDBX_MMAP_INCOHERENT_CPU_CACHE (void)addr; (void)nbytes; #endif } /*----------------------------------------------------------------------------*/ /* Internal prototypes */ MDBX_INTERNAL_FUNC int mdbx_cleanup_dead_readers(MDBX_env *env, int rlocked, int *dead); MDBX_INTERNAL_FUNC int mdbx_rthc_alloc(mdbx_thread_key_t *key, MDBX_reader *begin, MDBX_reader *end); MDBX_INTERNAL_FUNC void mdbx_rthc_remove(const mdbx_thread_key_t key); MDBX_INTERNAL_FUNC void mdbx_rthc_global_init(void); MDBX_INTERNAL_FUNC void mdbx_rthc_global_dtor(void); MDBX_INTERNAL_FUNC void mdbx_rthc_thread_dtor(void *ptr); static __maybe_unused __inline void mdbx_jitter4testing(bool tiny) { #if MDBX_DEBUG if (MDBX_DBG_JITTER & mdbx_runtime_flags) mdbx_osal_jitter(tiny); #else (void)tiny; #endif } #endif /* !__cplusplus */ #define MDBX_IS_ERROR(rc) \ ((rc) != MDBX_RESULT_TRUE && (rc) != MDBX_RESULT_FALSE) /* Internal error codes, not exposed outside libmdbx */ #define MDBX_NO_ROOT (MDBX_LAST_ADDED_ERRCODE + 10) /* Debugging output value of a cursor DBI: Negative in a sub-cursor. */ #define DDBI(mc) \ (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi) /* Key size which fits in a DKBUF. */ #define DKBUF_MAXKEYSIZE 511 /* FIXME */ #if MDBX_DEBUG #define DKBUF char _kbuf[DKBUF_MAXKEYSIZE * 4 + 2] #define DKEY(x) mdbx_dump_val(x, _kbuf, DKBUF_MAXKEYSIZE * 2 + 1) #define DVAL(x) \ mdbx_dump_val(x, _kbuf + DKBUF_MAXKEYSIZE * 2 + 1, DKBUF_MAXKEYSIZE * 2 + 1) #else #define DKBUF ((void)(0)) #define DKEY(x) ("-") #define DVAL(x) ("-") #endif /* An invalid page number. * Mainly used to denote an empty tree. */ #define P_INVALID (~(pgno_t)0) /* Test if the flags f are set in a flag word w. */ #define F_ISSET(w, f) (((w) & (f)) == (f)) /* Round n up to an even number. */ #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */ /* Default size of memory map. * This is certainly too small for any actual applications. Apps should * always set the size explicitly using mdbx_env_set_mapsize(). */ #define DEFAULT_MAPSIZE MEGABYTE /* Number of slots in the reader table. * This value was chosen somewhat arbitrarily. The 61 is a prime number, * and such readers plus a couple mutexes fit into single 4KB page. * Applications should set the table size using mdbx_env_set_maxreaders(). */ #define DEFAULT_READERS 61 /* Test if a page is a leaf page */ #define IS_LEAF(p) (((p)->mp_flags & P_LEAF) != 0) /* Test if a page is a LEAF2 page */ #define IS_LEAF2(p) unlikely(((p)->mp_flags & P_LEAF2) != 0) /* Test if a page is a branch page */ #define IS_BRANCH(p) (((p)->mp_flags & P_BRANCH) != 0) /* Test if a page is an overflow page */ #define IS_OVERFLOW(p) unlikely(((p)->mp_flags & P_OVERFLOW) != 0) /* Test if a page is a sub page */ #define IS_SUBP(p) (((p)->mp_flags & P_SUBP) != 0) /* Test if a page is dirty */ #define IS_DIRTY(p) (((p)->mp_flags & P_DIRTY) != 0) #define PAGETYPE(p) ((p)->mp_flags & (P_BRANCH | P_LEAF | P_LEAF2 | P_OVERFLOW)) /* Header for a single key/data pair within a page. * Used in pages of type P_BRANCH and P_LEAF without P_LEAF2. * We guarantee 2-byte alignment for 'MDBX_node's. * * Leaf node flags describe node contents. F_BIGDATA says the node's * data part is the page number of an overflow page with actual data. * F_DUPDATA and F_SUBDATA can be combined giving duplicate data in * a sub-page/sub-database, and named databases (just F_SUBDATA). */ typedef struct MDBX_node { #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ union { uint32_t mn_dsize; uint32_t mn_pgno32; }; uint8_t mn_flags; /* see mdbx_node flags */ uint8_t mn_extra; uint16_t mn_ksize; /* key size */ #else uint16_t mn_ksize; /* key size */ uint8_t mn_extra; uint8_t mn_flags; /* see mdbx_node flags */ union { uint32_t mn_pgno32; uint32_t mn_dsize; }; #endif /* __BYTE_ORDER__ */ /* mdbx_node Flags */ #define F_BIGDATA 0x01 /* data put on overflow page */ #define F_SUBDATA 0x02 /* data is a sub-database */ #define F_DUPDATA 0x04 /* data has duplicates */ /* valid flags for mdbx_node_add() */ #define NODE_ADD_FLAGS (F_DUPDATA | F_SUBDATA | MDBX_RESERVE | MDBX_APPEND) #if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || \ (!defined(__cplusplus) && defined(_MSC_VER)) uint8_t mn_data[] /* key and data are appended here */; #endif /* C99 */ } MDBX_node; #define DB_PERSISTENT_FLAGS \ (MDBX_REVERSEKEY | MDBX_DUPSORT | MDBX_INTEGERKEY | MDBX_DUPFIXED | \ MDBX_INTEGERDUP | MDBX_REVERSEDUP) /* mdbx_dbi_open() flags */ #define DB_USABLE_FLAGS (DB_PERSISTENT_FLAGS | MDBX_CREATE | MDBX_DB_ACCEDE) #define DB_VALID 0x8000 /* DB handle is valid, for me_dbflags */ #define DB_INTERNAL_FLAGS DB_VALID #if DB_INTERNAL_FLAGS & DB_USABLE_FLAGS #error "Oops, some flags overlapped or wrong" #endif #if DB_PERSISTENT_FLAGS & ~DB_USABLE_FLAGS #error "Oops, some flags overlapped or wrong" #endif /* max number of pages to commit in one writev() call */ #define MDBX_COMMIT_PAGES 64 #if defined(IOV_MAX) && IOV_MAX < MDBX_COMMIT_PAGES /* sysconf(_SC_IOV_MAX) */ #undef MDBX_COMMIT_PAGES #define MDBX_COMMIT_PAGES IOV_MAX #endif /* * / * | -1, a < b * CMP2INT(a,b) = < 0, a == b * | 1, a > b * \ */ #if 1 /* LY: fast enough on most systems */ #define CMP2INT(a, b) (((b) > (a)) ? -1 : (a) > (b)) #else #define CMP2INT(a, b) (((a) > (b)) - ((b) > (a))) #endif /* Do not spill pages to disk if txn is getting full, may fail instead */ #define MDBX_NOSPILL 0x8000 MDBX_NOTHROW_CONST_FUNCTION static __maybe_unused __inline pgno_t pgno_add(pgno_t base, pgno_t augend) { assert(base <= MAX_PAGENO); return (augend < MAX_PAGENO - base) ? base + augend : MAX_PAGENO; } MDBX_NOTHROW_CONST_FUNCTION static __maybe_unused __inline pgno_t pgno_sub(pgno_t base, pgno_t subtrahend) { assert(base >= MIN_PAGENO); return (subtrahend < base - MIN_PAGENO) ? base - subtrahend : MIN_PAGENO; } MDBX_NOTHROW_CONST_FUNCTION static __always_inline __maybe_unused bool is_powerof2(size_t x) { return (x & (x - 1)) == 0; } MDBX_NOTHROW_CONST_FUNCTION static __always_inline __maybe_unused size_t floor_powerof2(size_t value, size_t granularity) { assert(is_powerof2(granularity)); return value & ~(granularity - 1); } MDBX_NOTHROW_CONST_FUNCTION static __always_inline __maybe_unused size_t ceil_powerof2(size_t value, size_t granularity) { return floor_powerof2(value + granularity - 1, granularity); } /* Only a subset of the mdbx_env flags can be changed * at runtime. Changing other flags requires closing the * environment and re-opening it with the new flags. */ #define ENV_CHANGEABLE_FLAGS \ (MDBX_SAFE_NOSYNC | MDBX_NOMETASYNC | MDBX_DEPRECATED_MAPASYNC | \ MDBX_NOMEMINIT | MDBX_COALESCE | MDBX_PAGEPERTURB | MDBX_ACCEDE) #define ENV_CHANGELESS_FLAGS \ (MDBX_NOSUBDIR | MDBX_RDONLY | MDBX_WRITEMAP | MDBX_NOTLS | MDBX_NORDAHEAD | \ MDBX_LIFORECLAIM | MDBX_EXCLUSIVE) #define ENV_USABLE_FLAGS (ENV_CHANGEABLE_FLAGS | ENV_CHANGELESS_FLAGS) #if !defined(__cplusplus) || defined(__cpp_constexpr) static __maybe_unused void static_checks(void) { STATIC_ASSERT_MSG(INT16_MAX - CORE_DBS == MDBX_MAX_DBI, "Oops, MDBX_MAX_DBI or CORE_DBS?"); STATIC_ASSERT_MSG((unsigned)(MDBX_DB_ACCEDE | MDBX_CREATE) == ((DB_USABLE_FLAGS | DB_INTERNAL_FLAGS) & (ENV_USABLE_FLAGS | ENV_INTERNAL_FLAGS)), "Oops, some flags overlapped or wrong"); STATIC_ASSERT_MSG((ENV_INTERNAL_FLAGS & ENV_USABLE_FLAGS) == 0, "Oops, some flags overlapped or wrong"); } #endif /* Disabled for MSVC 19.0 (VisualStudio 2015) */ #ifdef __cplusplus } #endif /* * Copyright 2015-2021 Leonid Yuriev . * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * This code is derived from "LMDB engine" written by * Howard Chu (Symas Corporation), which itself derived from btree.c * written by Martin Hedenfalk. * * --- * * Portions Copyright 2011-2015 Howard Chu, Symas Corp. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . * * --- * * Portions Copyright (c) 2009, 2010 Martin Hedenfalk * * Permission to use, copy, modify, and distribute this software for any * purpose with or without fee is hereby granted, provided that the above * copyright notice and this permission notice appear in all copies. * * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. */ /*------------------------------------------------------------------------------ * Internal inline functions */ MDBX_NOTHROW_CONST_FUNCTION static unsigned log2n(size_t value) { assert(value > 0 && value < INT32_MAX && is_powerof2(value)); assert((value & -(int32_t)value) == value); #if __GNUC_PREREQ(4, 1) || __has_builtin(__builtin_ctzl) return __builtin_ctzl(value); #elif defined(_MSC_VER) unsigned long index; _BitScanForward(&index, (unsigned long)value); return index; #else static const uint8_t debruijn_ctz32[32] = { 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8, 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9}; return debruijn_ctz32[(uint32_t)(value * 0x077CB531u) >> 27]; #endif } /*------------------------------------------------------------------------------ * Unaligned access */ MDBX_NOTHROW_CONST_FUNCTION static __maybe_unused __always_inline unsigned field_alignment(unsigned alignment_baseline, size_t field_offset) { unsigned merge = alignment_baseline | (unsigned)field_offset; return merge & -(int)merge; } /* read-thunk for UB-sanitizer */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline uint8_t peek_u8(const uint8_t *const __restrict ptr) { return *ptr; } /* write-thunk for UB-sanitizer */ static __always_inline void poke_u8(uint8_t *const __restrict ptr, const uint8_t v) { *ptr = v; } MDBX_NOTHROW_PURE_FUNCTION static __always_inline uint16_t unaligned_peek_u16(const unsigned expected_alignment, const void *const ptr) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(uint16_t)) == 0) return *(const uint16_t *)ptr; else { uint16_t v; memcpy(&v, ptr, sizeof(v)); return v; } } static __always_inline void unaligned_poke_u16(const unsigned expected_alignment, void *const __restrict ptr, const uint16_t v) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(v)) == 0) *(uint16_t *)ptr = v; else memcpy(ptr, &v, sizeof(v)); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline uint32_t unaligned_peek_u32( const unsigned expected_alignment, const void *const __restrict ptr) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(uint32_t)) == 0) return *(const uint32_t *)ptr; else if ((expected_alignment % sizeof(uint16_t)) == 0) { const uint16_t lo = ((const uint16_t *)ptr)[__BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__]; const uint16_t hi = ((const uint16_t *)ptr)[__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__]; return lo | (uint32_t)hi << 16; } else { uint32_t v; memcpy(&v, ptr, sizeof(v)); return v; } } static __always_inline void unaligned_poke_u32(const unsigned expected_alignment, void *const __restrict ptr, const uint32_t v) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(v)) == 0) *(uint32_t *)ptr = v; else if ((expected_alignment % sizeof(uint16_t)) == 0) { ((uint16_t *)ptr)[__BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__] = (uint16_t)v; ((uint16_t *)ptr)[__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__] = (uint16_t)(v >> 16); } else memcpy(ptr, &v, sizeof(v)); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline uint64_t unaligned_peek_u64( const unsigned expected_alignment, const void *const __restrict ptr) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(uint64_t)) == 0) return *(const uint64_t *)ptr; else if ((expected_alignment % sizeof(uint32_t)) == 0) { const uint32_t lo = ((const uint32_t *)ptr)[__BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__]; const uint32_t hi = ((const uint32_t *)ptr)[__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__]; return lo | (uint64_t)hi << 32; } else { uint64_t v; memcpy(&v, ptr, sizeof(v)); return v; } } static __always_inline void unaligned_poke_u64(const unsigned expected_alignment, void *const __restrict ptr, const uint64_t v) { assert((uintptr_t)ptr % expected_alignment == 0); if (MDBX_UNALIGNED_OK || (expected_alignment % sizeof(v)) == 0) *(uint64_t *)ptr = v; else if ((expected_alignment % sizeof(uint32_t)) == 0) { ((uint32_t *)ptr)[__BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__] = (uint32_t)v; ((uint32_t *)ptr)[__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__] = (uint32_t)(v >> 32); } else memcpy(ptr, &v, sizeof(v)); } #define UNALIGNED_PEEK_8(ptr, struct, field) \ peek_u8((const uint8_t *)(ptr) + offsetof(struct, field)) #define UNALIGNED_POKE_8(ptr, struct, field, value) \ poke_u8((uint8_t *)(ptr) + offsetof(struct, field), value) #define UNALIGNED_PEEK_16(ptr, struct, field) \ unaligned_peek_u16(1, (const char *)(ptr) + offsetof(struct, field)) #define UNALIGNED_POKE_16(ptr, struct, field, value) \ unaligned_poke_u16(1, (char *)(ptr) + offsetof(struct, field), value) #define UNALIGNED_PEEK_32(ptr, struct, field) \ unaligned_peek_u32(1, (const char *)(ptr) + offsetof(struct, field)) #define UNALIGNED_POKE_32(ptr, struct, field, value) \ unaligned_poke_u32(1, (char *)(ptr) + offsetof(struct, field), value) #define UNALIGNED_PEEK_64(ptr, struct, field) \ unaligned_peek_u64(1, (const char *)(ptr) + offsetof(struct, field)) #define UNALIGNED_POKE_64(ptr, struct, field, value) \ unaligned_poke_u64(1, (char *)(ptr) + offsetof(struct, field), value) /* Get the page number pointed to by a branch node */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline pgno_t node_pgno(const MDBX_node *const __restrict node) { pgno_t pgno = UNALIGNED_PEEK_32(node, MDBX_node, mn_pgno32); if (sizeof(pgno) > 4) pgno |= ((uint64_t)UNALIGNED_PEEK_8(node, MDBX_node, mn_extra)) << 32; return pgno; } /* Set the page number in a branch node */ static __always_inline void node_set_pgno(MDBX_node *const __restrict node, pgno_t pgno) { assert(pgno >= MIN_PAGENO && pgno <= MAX_PAGENO); UNALIGNED_POKE_32(node, MDBX_node, mn_pgno32, (uint32_t)pgno); if (sizeof(pgno) > 4) UNALIGNED_POKE_8(node, MDBX_node, mn_extra, (uint8_t)((uint64_t)pgno >> 32)); } /* Get the size of the data in a leaf node */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t node_ds(const MDBX_node *const __restrict node) { return UNALIGNED_PEEK_32(node, MDBX_node, mn_dsize); } /* Set the size of the data for a leaf node */ static __always_inline void node_set_ds(MDBX_node *const __restrict node, size_t size) { assert(size < INT_MAX); UNALIGNED_POKE_32(node, MDBX_node, mn_dsize, (uint32_t)size); } /* The size of a key in a node */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t node_ks(const MDBX_node *const __restrict node) { return UNALIGNED_PEEK_16(node, MDBX_node, mn_ksize); } /* Set the size of the key for a leaf node */ static __always_inline void node_set_ks(MDBX_node *const __restrict node, size_t size) { assert(size < INT16_MAX); UNALIGNED_POKE_16(node, MDBX_node, mn_ksize, (uint16_t)size); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline uint8_t node_flags(const MDBX_node *const __restrict node) { return UNALIGNED_PEEK_8(node, MDBX_node, mn_flags); } static __always_inline void node_set_flags(MDBX_node *const __restrict node, uint8_t flags) { UNALIGNED_POKE_8(node, MDBX_node, mn_flags, flags); } /* Size of the node header, excluding dynamic data at the end */ #define NODESIZE offsetof(MDBX_node, mn_data) /* Address of the key for the node */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline void * node_key(const MDBX_node *const __restrict node) { return (char *)node + NODESIZE; } /* Address of the data for a node */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline void * node_data(const MDBX_node *const __restrict node) { return (char *)node_key(node) + node_ks(node); } /* Size of a node in a leaf page with a given key and data. * This is node header plus key plus data size. */ MDBX_NOTHROW_CONST_FUNCTION static __always_inline size_t node_size_len(const size_t key_len, const size_t value_len) { return NODESIZE + EVEN(key_len + value_len); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t node_size(const MDBX_val *key, const MDBX_val *value) { return node_size_len(key ? key->iov_len : 0, value ? value->iov_len : 0); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline pgno_t peek_pgno(const void *const __restrict ptr) { if (sizeof(pgno_t) == sizeof(uint32_t)) return (pgno_t)unaligned_peek_u32(1, ptr); else if (sizeof(pgno_t) == sizeof(uint64_t)) return (pgno_t)unaligned_peek_u64(1, ptr); else { pgno_t pgno; memcpy(&pgno, ptr, sizeof(pgno)); return pgno; } } static __always_inline void poke_pgno(void *const __restrict ptr, const pgno_t pgno) { if (sizeof(pgno) == sizeof(uint32_t)) unaligned_poke_u32(1, ptr, pgno); else if (sizeof(pgno) == sizeof(uint64_t)) unaligned_poke_u64(1, ptr, pgno); else memcpy(ptr, &pgno, sizeof(pgno)); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline pgno_t node_largedata_pgno(const MDBX_node *const __restrict node) { assert(node_flags(node) & F_BIGDATA); return peek_pgno(node_data(node)); } /*------------------------------------------------------------------------------ * Key length limitation factors: * * - Branch-page must contain at least two (MDBX_MINKEYS) nodes, * within each a key and a child page number. But we can't split a page if * it contains less that 4 keys. Therefore, at least 3 branch-node should * fit in the single branch-page: * pageroom = pagesize - page_hdr_len; * branch.maxnode = even_floor(pageroom / 3 - sizeof(indx_t)); * branch.maxkey = branch.maxnode - node_hdr_len; * * - Leaf-node of non-dupsort database must fit into one leaf-page, * where a value could be placed on a large/overflow page: * leaf.maxnode = even_floor(pageroom - sizeof(indx_t)); * leaf.maxkey = leaf.maxnode - node_hdr_len - sizeof(pgno_t); * * - SubDatabase-node must fit into one leaf-page: * subdb.maxname = leaf.maxnode - node_hdr_len - sizeof(MDBX_db); * * - Dupsort values itself are a keys in a dupsort-subdb and couldn't be * longer than the branch.maxkey. But dupsort node must fit into one * leaf-page, since dupsort value couldn't be placed on a large/overflow * page. * * - So, the simplest solution is to use half of branch.maxkey as * a common maxkey value. Nevertheless, the actual values of maxkey are: * nondupsort.maxkey = even_floor(pageroom / 3) * - sizeof(indx_t) - node_hdr_len; * dupsort.maxkey(value) = min(nondupsort.maxkey, * leaf.maxnode - even_ceil(length(value))); */ #define PAGEROOM(pagesize) ((pagesize)-PAGEHDRSZ) #define EVEN_FLOOR(n) ((n) & ~(size_t)1) #define BRANCH_NODEMAX(pagesize) \ (EVEN_FLOOR(PAGEROOM(pagesize) / (MDBX_MINKEYS * 2 - 1)) - sizeof(indx_t)) #define LEAF_NODEMAX(pagesize) (PAGEROOM(pagesize) - sizeof(indx_t)) #define MAX_GC1OVPAGE(pagesize) (PAGEROOM(pagesize) / sizeof(pgno_t) - 1) __cold int mdbx_env_get_maxkeysize(const MDBX_env *env) { return mdbx_env_get_maxkeysize_ex(env, MDBX_DUPSORT); } __cold int mdbx_env_get_maxkeysize_ex(const MDBX_env *env, MDBX_db_flags_t flags) { if (unlikely(!env || env->me_signature.weak != MDBX_ME_SIGNATURE)) return -1; return (int)mdbx_limits_keysize_max((intptr_t)env->me_psize, flags); } size_t mdbx_default_pagesize(void) { size_t pagesize = mdbx_syspagesize(); mdbx_ensure(nullptr, is_powerof2(pagesize)); pagesize = (pagesize >= MIN_PAGESIZE) ? pagesize : MIN_PAGESIZE; pagesize = (pagesize <= MAX_PAGESIZE) ? pagesize : MAX_PAGESIZE; return pagesize; } __cold intptr_t mdbx_limits_keysize_max(intptr_t pagesize, MDBX_db_flags_t flags) { if (pagesize < 1) pagesize = (intptr_t)mdbx_default_pagesize(); if (unlikely(pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2((size_t)pagesize))) return -1; STATIC_ASSERT(BRANCH_NODEMAX(MIN_PAGESIZE) - NODESIZE - sizeof(pgno_t) >= 8); STATIC_ASSERT(LEAF_NODEMAX(MIN_PAGESIZE) - NODESIZE - sizeof(pgno_t) >= 8); STATIC_ASSERT(LEAF_NODEMAX(MIN_PAGESIZE) - NODESIZE >= sizeof(MDBX_db)); if (flags & MDBX_INTEGERKEY) return 8 /* sizeof(uint64_t) */; STATIC_ASSERT(BRANCH_NODEMAX(MAX_PAGESIZE) - NODESIZE - sizeof(pgno_t) < LEAF_NODEMAX(MAX_PAGESIZE) - NODESIZE - sizeof(MDBX_db)); STATIC_ASSERT(BRANCH_NODEMAX(MIN_PAGESIZE) - NODESIZE - sizeof(pgno_t) < LEAF_NODEMAX(MIN_PAGESIZE) - NODESIZE - sizeof(MDBX_db)); if (flags & (MDBX_DUPSORT | MDBX_DUPFIXED | MDBX_INTEGERDUP | MDBX_REVERSEDUP)) return BRANCH_NODEMAX(pagesize) - NODESIZE - sizeof(MDBX_db); return BRANCH_NODEMAX(pagesize) - NODESIZE - sizeof(pgno_t); } __cold int mdbx_env_get_maxvalsize_ex(const MDBX_env *env, MDBX_db_flags_t flags) { if (unlikely(!env || env->me_signature.weak != MDBX_ME_SIGNATURE)) return -1; return (int)mdbx_limits_valsize_max((intptr_t)env->me_psize, flags); } __cold intptr_t mdbx_limits_valsize_max(intptr_t pagesize, MDBX_db_flags_t flags) { if (pagesize < 1) pagesize = (intptr_t)mdbx_default_pagesize(); if (unlikely(pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2((size_t)pagesize))) return -1; if (flags & MDBX_INTEGERDUP) return 8 /* sizeof(uint64_t) */; if (flags & (MDBX_DUPSORT | MDBX_DUPFIXED | MDBX_REVERSEDUP)) return BRANCH_NODEMAX(pagesize) - NODESIZE; const unsigned page_ln2 = log2n(pagesize); const size_t hard = 0x7FF00000ul; const size_t hard_pages = hard >> page_ln2; STATIC_ASSERT(MDBX_PGL_LIMIT <= MAX_PAGENO); const size_t pages_limit = MDBX_PGL_LIMIT / 4; const size_t limit = (hard_pages < pages_limit) ? hard : (pages_limit << page_ln2); return (limit < MAX_MAPSIZE / 2) ? limit : MAX_MAPSIZE / 2; } /* Calculate the size of a leaf node. * * The size depends on the environment's page size; if a data item * is too large it will be put onto an overflow page and the node * size will only include the key and not the data. Sizes are always * rounded up to an even number of bytes, to guarantee 2-byte alignment * of the MDBX_node headers. */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t leaf_size(const MDBX_env *env, const MDBX_val *key, const MDBX_val *data) { size_t node_bytes = node_size(key, data); /* NOTE: The actual limit is LEAF_NODEMAX(env->me_psize), but it reasonable to * use env->me_branch_nodemax (which is 3 times less) as the threshold * because: * - Large threshold implies that any insertion/update could result split * a single leaf page to THREE, which requires TWO insertion into parent * branch page, then could leads to split parent page and so on up to * the root. Such double-splitting is complex, ie costly (in case simple * clear implementation) either dangerous (in case high-optimized * implementation). * - This does not affect capabilities, i.e. it does not limit the maximum * key size. * - At a lower threshold, on average, the density of keys on leaf pages * increases and the height of the tree decreases. Thus, this lead the * less number of pages participating in the search, and the search * speed increases. * - On the other hand, there is no universal gold ratio here and with a * smaller threshold, we will create more overflows/large pages, * i.e. the database size will be larger as will the IOPS volume. * * So, the lower threshold is not a silver bullet, but it allow implementation * to be much simple and robust, without adding a flaws. */ if (node_bytes > env->me_branch_nodemax) { /* put on overflow page */ node_bytes = node_size(key, nullptr) + sizeof(pgno_t); } return node_bytes + sizeof(indx_t); } /* Calculate the size of a branch node. * * The size should depend on the environment's page size but since * we currently don't support spilling large keys onto overflow * pages, it's simply the size of the MDBX_node header plus the * size of the key. Sizes are always rounded up to an even number * of bytes, to guarantee 2-byte alignment of the MDBX_node headers. * * [in] env The environment handle. * [in] key The key for the node. * * Returns The number of bytes needed to store the node. */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t branch_size(const MDBX_env *env, const MDBX_val *key) { /* Size of a node in a branch page with a given key. * This is just the node header plus the key, there is no data. */ size_t node_bytes = node_size(key, nullptr); if (unlikely(node_bytes > env->me_branch_nodemax)) { /* put on overflow page */ /* not implemented */ mdbx_assert_fail(env, "INDXSIZE(key) <= env->me_nodemax", __func__, __LINE__); node_bytes = node_size(key, nullptr) + sizeof(pgno_t); } return node_bytes + sizeof(indx_t); } MDBX_NOTHROW_CONST_FUNCTION static __always_inline uint16_t flags_db2sub(uint16_t db_flags) { uint16_t sub_flags = db_flags & MDBX_DUPFIXED; /* MDBX_INTEGERDUP => MDBX_INTEGERKEY */ #define SHIFT_INTEGERDUP_TO_INTEGERKEY 2 STATIC_ASSERT((MDBX_INTEGERDUP >> SHIFT_INTEGERDUP_TO_INTEGERKEY) == MDBX_INTEGERKEY); sub_flags |= (db_flags & MDBX_INTEGERDUP) >> SHIFT_INTEGERDUP_TO_INTEGERKEY; /* MDBX_REVERSEDUP => MDBX_REVERSEKEY */ #define SHIFT_REVERSEDUP_TO_REVERSEKEY 5 STATIC_ASSERT((MDBX_REVERSEDUP >> SHIFT_REVERSEDUP_TO_REVERSEKEY) == MDBX_REVERSEKEY); sub_flags |= (db_flags & MDBX_REVERSEDUP) >> SHIFT_REVERSEDUP_TO_REVERSEKEY; return sub_flags; } /*----------------------------------------------------------------------------*/ MDBX_NOTHROW_PURE_FUNCTION static __always_inline size_t pgno2bytes(const MDBX_env *env, pgno_t pgno) { mdbx_assert(env, (1u << env->me_psize2log) == env->me_psize); return ((size_t)pgno) << env->me_psize2log; } MDBX_NOTHROW_PURE_FUNCTION static __always_inline MDBX_page * pgno2page(const MDBX_env *env, pgno_t pgno) { return (MDBX_page *)(env->me_map + pgno2bytes(env, pgno)); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline pgno_t bytes2pgno(const MDBX_env *env, size_t bytes) { mdbx_assert(env, (env->me_psize >> env->me_psize2log) == 1); return (pgno_t)(bytes >> env->me_psize2log); } MDBX_NOTHROW_PURE_FUNCTION static size_t pgno_align2os_bytes(const MDBX_env *env, pgno_t pgno) { return ceil_powerof2(pgno2bytes(env, pgno), env->me_os_psize); } MDBX_NOTHROW_PURE_FUNCTION static pgno_t pgno_align2os_pgno(const MDBX_env *env, pgno_t pgno) { return bytes2pgno(env, pgno_align2os_bytes(env, pgno)); } MDBX_NOTHROW_PURE_FUNCTION static size_t bytes_align2os_bytes(const MDBX_env *env, size_t bytes) { return ceil_powerof2(ceil_powerof2(bytes, env->me_psize), env->me_os_psize); } /* Address of first usable data byte in a page, after the header */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline void * page_data(const MDBX_page *mp) { return (char *)mp + PAGEHDRSZ; } MDBX_NOTHROW_PURE_FUNCTION static __always_inline const MDBX_page * data_page(const void *data) { return container_of(data, MDBX_page, mp_ptrs); } MDBX_NOTHROW_PURE_FUNCTION static __always_inline MDBX_meta * page_meta(MDBX_page *mp) { return (MDBX_meta *)page_data(mp); } /* Number of nodes on a page */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline unsigned page_numkeys(const MDBX_page *mp) { return mp->mp_lower >> 1; } /* The amount of space remaining in the page */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline unsigned page_room(const MDBX_page *mp) { return mp->mp_upper - mp->mp_lower; } MDBX_NOTHROW_PURE_FUNCTION static __always_inline unsigned page_space(const MDBX_env *env) { STATIC_ASSERT(PAGEHDRSZ % 2 == 0); return env->me_psize - PAGEHDRSZ; } MDBX_NOTHROW_PURE_FUNCTION static __always_inline unsigned page_used(const MDBX_env *env, const MDBX_page *mp) { return page_space(env) - page_room(mp); } /* The percentage of space used in the page, in a percents. */ MDBX_NOTHROW_PURE_FUNCTION static __maybe_unused __inline double page_fill(const MDBX_env *env, const MDBX_page *mp) { return page_used(env, mp) * 100.0 / page_space(env); } /* The number of overflow pages needed to store the given size. */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline pgno_t number_of_ovpages(const MDBX_env *env, size_t bytes) { return bytes2pgno(env, PAGEHDRSZ - 1 + bytes) + 1; } __cold static int MDBX_PRINTF_ARGS(2, 3) bad_page(const MDBX_page *mp, const char *fmt, ...) { if (mdbx_log_enabled(MDBX_LOG_ERROR)) { static const MDBX_page *prev; if (prev != mp) { prev = mp; mdbx_debug_log(MDBX_LOG_ERROR, "badpage", 0, "corrupted page #%u, mod-txnid %" PRIaTXN "\n", mp->mp_pgno, mp->mp_txnid); } va_list args; va_start(args, fmt); mdbx_debug_log_va(MDBX_LOG_ERROR, "badpage", 0, fmt, args); va_end(args); } return MDBX_CORRUPTED; } /* Address of node i in page p */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline MDBX_node * page_node(const MDBX_page *mp, unsigned i) { assert((mp->mp_flags & (P_LEAF2 | P_OVERFLOW | P_META)) == 0); assert(page_numkeys(mp) > (unsigned)(i)); assert(mp->mp_ptrs[i] % 2 == 0); return (MDBX_node *)((char *)mp + mp->mp_ptrs[i] + PAGEHDRSZ); } /* The address of a key in a LEAF2 page. * LEAF2 pages are used for MDBX_DUPFIXED sorted-duplicate sub-DBs. * There are no node headers, keys are stored contiguously. */ MDBX_NOTHROW_PURE_FUNCTION static __always_inline void * page_leaf2key(const MDBX_page *mp, unsigned i, size_t keysize) { assert((mp->mp_flags & (P_BRANCH | P_LEAF | P_LEAF2 | P_OVERFLOW | P_META)) == (P_LEAF | P_LEAF2)); assert(mp->mp_leaf2_ksize == keysize); (void)keysize; return (char *)mp + PAGEHDRSZ + (i * mp->mp_leaf2_ksize); } /* Set the node's key into keyptr. */ static __always_inline void get_key(const MDBX_node *node, MDBX_val *keyptr) { keyptr->iov_len = node_ks(node); keyptr->iov_base = node_key(node); } /* Set the node's key into keyptr, if requested. */ static __always_inline void get_key_optional(const MDBX_node *node, MDBX_val *keyptr /* __may_null */) { if (keyptr) get_key(node, keyptr); } /*------------------------------------------------------------------------------ * Workaround for mmaped-lookahead-cross-page-boundary bug * in an obsolete versions of Elbrus's libc and kernels. */ #if defined(__e2k__) && defined(MDBX_E2K_MLHCPB_WORKAROUND) && \ MDBX_E2K_MLHCPB_WORKAROUND int __hot mdbx_e2k_memcmp_bug_workaround(const void *s1, const void *s2, size_t n) { if (unlikely(n > 42 /* LY: align followed access if reasonable possible */ && (((uintptr_t)s1) & 7) != 0 && (((uintptr_t)s1) & 7) == (((uintptr_t)s2) & 7))) { if (((uintptr_t)s1) & 1) { const int diff = *(uint8_t *)s1 - *(uint8_t *)s2; if (diff) return diff; s1 = (char *)s1 + 1; s2 = (char *)s2 + 1; n -= 1; } if (((uintptr_t)s1) & 2) { const uint16_t a = *(uint16_t *)s1; const uint16_t b = *(uint16_t *)s2; if (likely(a != b)) return (__builtin_bswap16(a) > __builtin_bswap16(b)) ? 1 : -1; s1 = (char *)s1 + 2; s2 = (char *)s2 + 2; n -= 2; } if (((uintptr_t)s1) & 4) { const uint32_t a = *(uint32_t *)s1; const uint32_t b = *(uint32_t *)s2; if (likely(a != b)) return (__builtin_bswap32(a) > __builtin_bswap32(b)) ? 1 : -1; s1 = (char *)s1 + 4; s2 = (char *)s2 + 4; n -= 4; } } while (n >= 8) { const uint64_t a = *(uint64_t *)s1; const uint64_t b = *(uint64_t *)s2; if (likely(a != b)) return (__builtin_bswap64(a) > __builtin_bswap64(b)) ? 1 : -1; s1 = (char *)s1 + 8; s2 = (char *)s2 + 8; n -= 8; } if (n & 4) { const uint32_t a = *(uint32_t *)s1; const uint32_t b = *(uint32_t *)s2; if (likely(a != b)) return (__builtin_bswap32(a) > __builtin_bswap32(b)) ? 1 : -1; s1 = (char *)s1 + 4; s2 = (char *)s2 + 4; } if (n & 2) { const uint16_t a = *(uint16_t *)s1; const uint16_t b = *(uint16_t *)s2; if (likely(a != b)) return (__builtin_bswap16(a) > __builtin_bswap16(b)) ? 1 : -1; s1 = (char *)s1 + 2; s2 = (char *)s2 + 2; } return (n & 1) ? *(uint8_t *)s1 - *(uint8_t *)s2 : 0; } int __hot mdbx_e2k_strcmp_bug_workaround(const char *s1, const char *s2) { while (true) { int diff = *(uint8_t *)s1 - *(uint8_t *)s2; if (likely(diff != 0) || *s1 == '\0') return diff; s1 += 1; s2 += 1; } } int __hot mdbx_e2k_strncmp_bug_workaround(const char *s1, const char *s2, size_t n) { while (n > 0) { int diff = *(uint8_t *)s1 - *(uint8_t *)s2; if (likely(diff != 0) || *s1 == '\0') return diff; s1 += 1; s2 += 1; n -= 1; } return 0; } size_t __hot mdbx_e2k_strlen_bug_workaround(const char *s) { size_t n = 0; while (*s) { s += 1; n += 1; } return n; } size_t __hot mdbx_e2k_strnlen_bug_workaround(const char *s, size_t maxlen) { size_t n = 0; while (maxlen > n && *s) { s += 1; n += 1; } return n; } #endif /* MDBX_E2K_MLHCPB_WORKAROUND */ /*------------------------------------------------------------------------------ * safe read/write volatile 64-bit fields on 32-bit architectures. */ enum MDBX_memory_order { mo_Relaxed, mo_AcquireRelease, mo_SequentialConsistency }; #ifdef MDBX_HAVE_C11ATOMICS #if defined(__clang__) && __clang__ < 8 #define MDBX_WORKAROUND_CLANG_C11ATOMICS_BUG(type, ptr) \ ((volatile _Atomic(type) *)(ptr)) #else #define MDBX_WORKAROUND_CLANG_C11ATOMICS_BUG(type, ptr) (ptr) #endif /* __clang__ < 8.x */ static __always_inline memory_order mo_c11_store(enum MDBX_memory_order fence) { switch (fence) { default: assert(false); __unreachable(); case mo_Relaxed: return memory_order_relaxed; case mo_AcquireRelease: return memory_order_release; case mo_SequentialConsistency: return memory_order_seq_cst; } } static __always_inline memory_order mo_c11_load(enum MDBX_memory_order fence) { switch (fence) { default: assert(false); __unreachable(); case mo_Relaxed: return memory_order_relaxed; case mo_AcquireRelease: return memory_order_acquire; case mo_SequentialConsistency: return memory_order_seq_cst; } } #endif /* MDBX_HAVE_C11ATOMICS */ static __maybe_unused __always_inline void mdbx_memory_fence(bool checkpoint, bool write) { #ifdef MDBX_HAVE_C11ATOMICS atomic_thread_fence( checkpoint ? memory_order_seq_cst : (write ? memory_order_release : memory_order_acquire)); #else /* MDBX_HAVE_C11ATOMICS */ mdbx_compiler_barrier(); if (checkpoint || (write && MDBX_CPU_WRITEBACK_INCOHERENT)) mdbx_memory_barrier(); #endif /* MDBX_HAVE_C11ATOMICS */ } static __maybe_unused __always_inline uint32_t atomic_store32(MDBX_atomic_uint32_t *ptr, const uint32_t value, enum MDBX_memory_order order) { STATIC_ASSERT(sizeof(MDBX_atomic_uint32_t) == 4); #ifdef MDBX_HAVE_C11ATOMICS assert(atomic_is_lock_free(&ptr->c11a)); atomic_store_explicit(&ptr->c11a, value, mo_c11_store(order)); #else /* MDBX_HAVE_C11ATOMICS */ if (order != mo_Relaxed) mdbx_compiler_barrier(); ptr->weak = value; if (order != mo_Relaxed) mdbx_memory_fence(order == mo_SequentialConsistency, true); #endif /* MDBX_HAVE_C11ATOMICS */ return value; } static __maybe_unused __always_inline uint32_t atomic_load32(const MDBX_atomic_uint32_t *ptr, enum MDBX_memory_order order) { STATIC_ASSERT(sizeof(MDBX_atomic_uint32_t) == 4); #ifdef MDBX_HAVE_C11ATOMICS assert(atomic_is_lock_free(&ptr->c11a)); return atomic_load_explicit( MDBX_WORKAROUND_CLANG_C11ATOMICS_BUG(uint32_t, &ptr->c11a), mo_c11_load(order)); #else /* MDBX_HAVE_C11ATOMICS */ if (order != mo_Relaxed) mdbx_memory_fence(order == mo_SequentialConsistency, false); const uint32_t value = ptr->weak; if (order != mo_Relaxed) mdbx_compiler_barrier(); return value; #endif /* MDBX_HAVE_C11ATOMICS */ } static __always_inline uint64_t atomic_store64(MDBX_atomic_uint64_t *ptr, const uint64_t value, enum MDBX_memory_order order) { STATIC_ASSERT(sizeof(MDBX_atomic_uint64_t) == 8); #if MDBX_64BIT_ATOMIC #ifdef MDBX_HAVE_C11ATOMICS assert(atomic_is_lock_free(&ptr->c11a)); atomic_store_explicit(&ptr->c11a, value, mo_c11_store(order)); #else /* MDBX_HAVE_C11ATOMICS */ if (order != mo_Relaxed) mdbx_compiler_barrier(); ptr->weak = value; if (order != mo_Relaxed) mdbx_memory_fence(order == mo_SequentialConsistency, true); #endif /* MDBX_HAVE_C11ATOMICS */ #else /* !MDBX_64BIT_ATOMIC */ atomic_store32(&ptr->low, (uint32_t)value, (order == mo_Relaxed) ? mo_Relaxed : mo_AcquireRelease); mdbx_jitter4testing(true); atomic_store32(&ptr->high, (uint32_t)(value >> 32), order); mdbx_jitter4testing(true); #endif /* !MDBX_64BIT_ATOMIC */ return value; } static __always_inline uint64_t atomic_load64(const MDBX_atomic_uint64_t *ptr, enum MDBX_memory_order order) { STATIC_ASSERT(sizeof(MDBX_atomic_uint64_t) == 8); #if MDBX_64BIT_ATOMIC #ifdef MDBX_HAVE_C11ATOMICS assert(atomic_is_lock_free(&ptr->c11a)); return atomic_load_explicit( MDBX_WORKAROUND_CLANG_C11ATOMICS_BUG(uint64_t, &ptr->c11a), mo_c11_load(order)); #else /* MDBX_HAVE_C11ATOMICS */ if (order != mo_Relaxed) mdbx_memory_fence(order == mo_SequentialConsistency, false); const uint64_t value = ptr->weak; if (order != mo_Relaxed) mdbx_compiler_barrier(); return value; #endif /* MDBX_HAVE_C11ATOMICS */ #else /* !MDBX_64BIT_ATOMIC */ uint64_t value = (uint64_t)atomic_load32(&ptr->high, order) << 32; mdbx_jitter4testing(true); value |= atomic_load32(&ptr->low, (order == mo_Relaxed) ? mo_Relaxed : mo_AcquireRelease); mdbx_jitter4testing(true); for (;;) { uint64_t again = (uint64_t)atomic_load32(&ptr->high, order) << 32; mdbx_jitter4testing(true); again |= atomic_load32( &ptr->low, (order == mo_Relaxed) ? mo_Relaxed : mo_AcquireRelease); mdbx_jitter4testing(true); if (likely(value == again)) return value; value = again; } #endif /* !MDBX_64BIT_ATOMIC */ } static __always_inline void atomic_yield(void) { #if defined(_WIN32) || defined(_WIN64) YieldProcessor(); #elif defined(__ia32__) || defined(__e2k__) __builtin_ia32_pause(); #elif defined(__ia64__) #if defined(__HP_cc__) || defined(__HP_aCC__) _Asm_hint(_HINT_PAUSE); #else __asm__ __volatile__("hint @pause"); #endif #elif defined(__aarch64__) || (defined(__ARM_ARCH) && __ARM_ARCH > 6) || \ defined(__ARM_ARCH_6K__) #ifdef __CC_ARM __yield(); #else __asm__ __volatile__("yield"); #endif #elif (defined(__mips64) || defined(__mips64__)) && defined(__mips_isa_rev) && \ __mips_isa_rev >= 2 __asm__ __volatile__("pause"); #elif defined(__mips) || defined(__mips__) || defined(__mips64) || \ defined(__mips64__) || defined(_M_MRX000) || defined(_MIPS_) || \ defined(__MWERKS__) || defined(__sgi) __asm__ __volatile__(".word 0x00000140"); #elif defined(__linux__) || defined(__gnu_linux__) || defined(_UNIX03_SOURCE) sched_yield(); #elif (defined(_GNU_SOURCE) && __GLIBC_PREREQ(2, 1)) || defined(_OPEN_THREADS) pthread_yield(); #endif } #if MDBX_64BIT_CAS static __always_inline bool atomic_cas64(MDBX_atomic_uint64_t *p, uint64_t c, uint64_t v) { #ifdef MDBX_HAVE_C11ATOMICS STATIC_ASSERT(sizeof(long long) >= sizeof(uint64_t)); #ifdef ATOMIC_LLONG_LOCK_FREE STATIC_ASSERT(ATOMIC_LLONG_LOCK_FREE > 0); #if ATOMIC_LLONG_LOCK_FREE < 2 assert(atomic_is_lock_free(&p->c11a)); #endif /* ATOMIC_LLONG_LOCK_FREE < 2 */ #else /* defined(ATOMIC_LLONG_LOCK_FREE) */ assert(atomic_is_lock_free(&p->c11a)); #endif return atomic_compare_exchange_strong(&p->c11a, &c, v); #elif defined(__GNUC__) || defined(__clang__) return __sync_bool_compare_and_swap(&p->weak, c, v); #elif defined(_MSC_VER) return c == (uint64_t)_InterlockedCompareExchange64(&p->weak, v, c); #elif defined(__APPLE__) return OSAtomicCompareAndSwap64Barrier(c, v, &p->weak); #else #error FIXME: Unsupported compiler #endif } #endif /* MDBX_64BIT_CAS */ static __always_inline bool atomic_cas32(MDBX_atomic_uint32_t *p, uint32_t c, uint32_t v) { #ifdef MDBX_HAVE_C11ATOMICS STATIC_ASSERT(sizeof(int) >= sizeof(uint32_t)); #ifdef ATOMIC_INT_LOCK_FREE STATIC_ASSERT(ATOMIC_INT_LOCK_FREE > 0); #if ATOMIC_INT_LOCK_FREE < 2 assert(atomic_is_lock_free(&p->c11a)); #endif #else assert(atomic_is_lock_free(&p->c11a)); #endif return atomic_compare_exchange_strong(&p->c11a, &c, v); #elif defined(__GNUC__) || defined(__clang__) return __sync_bool_compare_and_swap(&p->weak, c, v); #elif defined(_MSC_VER) STATIC_ASSERT(sizeof(volatile long) == sizeof(volatile uint32_t)); return c == (uint32_t)_InterlockedCompareExchange(&p->weak, v, c); #elif defined(__APPLE__) return OSAtomicCompareAndSwap32Barrier(c, v, &p->weak); #else #error FIXME: Unsupported compiler #endif } static __always_inline uint32_t atomic_add32(MDBX_atomic_uint32_t *p, uint32_t v) { #ifdef MDBX_HAVE_C11ATOMICS STATIC_ASSERT(sizeof(int) >= sizeof(uint32_t)); #ifdef ATOMIC_INT_LOCK_FREE STATIC_ASSERT(ATOMIC_INT_LOCK_FREE > 0); #if ATOMIC_INT_LOCK_FREE < 2 assert(atomic_is_lock_free(&p->c11a)); #endif #else assert(atomic_is_lock_free(&p->c11a)); #endif return atomic_fetch_add(&p->c11a, v); #elif defined(__GNUC__) || defined(__clang__) return __sync_fetch_and_add(&p->weak, v); #elif defined(_MSC_VER) STATIC_ASSERT(sizeof(volatile long) == sizeof(volatile uint32_t)); return _InterlockedExchangeAdd(&p->weak, v); #elif defined(__APPLE__) return OSAtomicAdd32Barrier(v, &p->weak); #else #error FIXME: Unsupported compiler #endif } #define atomic_sub32(p, v) atomic_add32(p, 0 - (v)) static __always_inline uint64_t safe64_txnid_next(uint64_t txnid) { txnid += MDBX_TXNID_STEP; #if !MDBX_64BIT_CAS /* avoid overflow of low-part in safe64_reset() */ txnid += (UINT32_MAX == (uint32_t)txnid); #endif return txnid; } static __always_inline void safe64_reset(MDBX_atomic_uint64_t *ptr, bool single_writer) { #if !MDBX_64BIT_CAS if (!single_writer) { STATIC_ASSERT(MDBX_TXNID_STEP > 1); /* it is safe to increment low-part to avoid ABA, since MDBX_TXNID_STEP > 1 * and overflow was preserved in safe64_txnid_next() */ atomic_add32(&ptr->low, 1) /* avoid ABA in safe64_reset_compare() */; atomic_store32( &ptr->high, UINT32_MAX, mo_AcquireRelease) /* atomically make >= SAFE64_INVALID_THRESHOLD */; atomic_add32(&ptr->low, 1) /* avoid ABA in safe64_reset_compare() */; } else #elif MDBX_64BIT_ATOMIC /* atomically make value >= SAFE64_INVALID_THRESHOLD by 64-bit operation */ atomic_store64(ptr, UINT64_MAX, single_writer ? mo_AcquireRelease : mo_SequentialConsistency); #else /* atomically make value >= SAFE64_INVALID_THRESHOLD by 32-bit operation */ atomic_store32(&ptr->high, UINT32_MAX, single_writer ? mo_AcquireRelease : mo_SequentialConsistency); #endif /* MDBX_64BIT_ATOMIC */ assert(ptr->weak >= SAFE64_INVALID_THRESHOLD); mdbx_jitter4testing(true); } static __always_inline bool safe64_reset_compare(MDBX_atomic_uint64_t *ptr, txnid_t compare) { /* LY: This function is used to reset `mr_txnid` from hsr-handler in case * the asynchronously cancellation of read transaction. Therefore, * there may be a collision between the cleanup performed here and * asynchronous termination and restarting of the read transaction * in another proces/thread. In general we MUST NOT reset the `mr_txnid` * if a new transaction was started (i.e. if `mr_txnid` was changed). */ #if MDBX_64BIT_CAS bool rc = atomic_cas64(ptr, compare, UINT64_MAX); #else /* LY: There is no gold ratio here since shared mutex is too costly, * in such way we must acquire/release it for every update of mr_txnid, * i.e. twice for each read transaction). */ bool rc = false; if (likely(atomic_load32(&ptr->low, mo_AcquireRelease) == (uint32_t)compare && atomic_cas32(&ptr->high, (uint32_t)(compare >> 32), UINT32_MAX))) { if (unlikely(atomic_load32(&ptr->low, mo_AcquireRelease) != (uint32_t)compare)) atomic_cas32(&ptr->high, UINT32_MAX, (uint32_t)(compare >> 32)); else rc = true; } #endif /* MDBX_64BIT_CAS */ mdbx_jitter4testing(true); return rc; } static __always_inline void safe64_write(MDBX_atomic_uint64_t *ptr, const uint64_t v) { assert(ptr->weak >= SAFE64_INVALID_THRESHOLD); #if MDBX_64BIT_ATOMIC atomic_store64(ptr, v, mo_AcquireRelease); #else /* MDBX_64BIT_ATOMIC */ /* update low-part but still value >= SAFE64_INVALID_THRESHOLD */ atomic_store32(&ptr->low, (uint32_t)v, mo_AcquireRelease); assert(ptr->weak >= SAFE64_INVALID_THRESHOLD); mdbx_jitter4testing(true); /* update high-part from SAFE64_INVALID_THRESHOLD to actual value */ atomic_store32(&ptr->high, (uint32_t)(v >> 32), mo_AcquireRelease); #endif /* MDBX_64BIT_ATOMIC */ assert(ptr->weak == v); mdbx_jitter4testing(true); } static __always_inline uint64_t safe64_read(const MDBX_atomic_uint64_t *ptr) { mdbx_jitter4testing(true); uint64_t v = atomic_load64(ptr, mo_AcquireRelease); mdbx_jitter4testing(true); return v; } #if 0 /* unused for now */ static __maybe_unused __always_inline bool safe64_is_valid(uint64_t v) { #if MDBX_WORDBITS >= 64 return v < SAFE64_INVALID_THRESHOLD; #else return (v >> 32) != UINT32_MAX; #endif /* MDBX_WORDBITS */ } static __maybe_unused __always_inline bool safe64_is_valid_ptr(const MDBX_atomic_uint64_t *ptr) { #if MDBX_64BIT_ATOMIC return atomic_load64(ptr, mo_AcquireRelease) < SAFE64_INVALID_THRESHOLD; #else return atomic_load32(&ptr->high, mo_AcquireRelease) != UINT32_MAX; #endif /* MDBX_64BIT_ATOMIC */ } static __always_inline void safe64_update(MDBX_atomic_uint64_t *ptr, const uint64_t v) { #if MDBX_64BIT_ATOMIC safe64_reset(ptr, true); #endif /* MDBX_64BIT_ATOMIC */ safe64_write(ptr, v); } #endif /* unused for now */ /*----------------------------------------------------------------------------*/ /* rthc (tls keys and destructors) */ typedef struct rthc_entry_t { MDBX_reader *begin; MDBX_reader *end; mdbx_thread_key_t thr_tls_key; bool key_valid; } rthc_entry_t; #if MDBX_DEBUG #define RTHC_INITIAL_LIMIT 1 #else #define RTHC_INITIAL_LIMIT 16 #endif static bin128_t bootid; #if defined(_WIN32) || defined(_WIN64) static CRITICAL_SECTION rthc_critical_section; static CRITICAL_SECTION lcklist_critical_section; #else int __cxa_thread_atexit_impl(void (*dtor)(void *), void *obj, void *dso_symbol) __attribute__((__weak__)); #ifdef __APPLE__ /* FIXME: Thread-Local Storage destructors & DSO-unloading */ int __cxa_thread_atexit_impl(void (*dtor)(void *), void *obj, void *dso_symbol) { (void)dtor; (void)obj; (void)dso_symbol; return -1; } #endif /* __APPLE__ */ static pthread_mutex_t lcklist_mutex = PTHREAD_MUTEX_INITIALIZER; static pthread_mutex_t rthc_mutex = PTHREAD_MUTEX_INITIALIZER; static pthread_cond_t rthc_cond = PTHREAD_COND_INITIALIZER; static mdbx_thread_key_t rthc_key; static MDBX_atomic_uint32_t rthc_pending; static void __cold workaround_glibc_bug21031(void) { /* Workaround for https://sourceware.org/bugzilla/show_bug.cgi?id=21031 * * Due race between pthread_key_delete() and __nptl_deallocate_tsd() * The destructor(s) of thread-local-storage object(s) may be running * in another thread(s) and be blocked or not finished yet. * In such case we get a SEGFAULT after unload this library DSO. * * So just by yielding a few timeslices we give a chance * to such destructor(s) for completion and avoids segfault. */ sched_yield(); sched_yield(); sched_yield(); } #endif static unsigned rthc_count, rthc_limit; static rthc_entry_t *rthc_table; static rthc_entry_t rthc_table_static[RTHC_INITIAL_LIMIT]; static __inline void rthc_lock(void) { #if defined(_WIN32) || defined(_WIN64) EnterCriticalSection(&rthc_critical_section); #else mdbx_ensure(nullptr, pthread_mutex_lock(&rthc_mutex) == 0); #endif } static __inline void rthc_unlock(void) { #if defined(_WIN32) || defined(_WIN64) LeaveCriticalSection(&rthc_critical_section); #else mdbx_ensure(nullptr, pthread_mutex_unlock(&rthc_mutex) == 0); #endif } static __inline int thread_key_create(mdbx_thread_key_t *key) { int rc; #if defined(_WIN32) || defined(_WIN64) *key = TlsAlloc(); rc = (*key != TLS_OUT_OF_INDEXES) ? MDBX_SUCCESS : GetLastError(); #else rc = pthread_key_create(key, nullptr); #endif mdbx_trace("&key = %p, value %" PRIuPTR ", rc %d", __Wpedantic_format_voidptr(key), (uintptr_t)*key, rc); return rc; } static __inline void thread_key_delete(mdbx_thread_key_t key) { mdbx_trace("key = %" PRIuPTR, (uintptr_t)key); #if defined(_WIN32) || defined(_WIN64) mdbx_ensure(nullptr, TlsFree(key)); #else mdbx_ensure(nullptr, pthread_key_delete(key) == 0); workaround_glibc_bug21031(); #endif } static __inline void *thread_rthc_get(mdbx_thread_key_t key) { #if defined(_WIN32) || defined(_WIN64) return TlsGetValue(key); #else return pthread_getspecific(key); #endif } static void thread_rthc_set(mdbx_thread_key_t key, const void *value) { #if defined(_WIN32) || defined(_WIN64) mdbx_ensure(nullptr, TlsSetValue(key, (void *)value)); #else #define MDBX_THREAD_RTHC_ZERO 0 #define MDBX_THREAD_RTHC_REGISTERED 1 #define MDBX_THREAD_RTHC_COUNTED 2 static __thread uint32_t thread_registration_state; if (value && unlikely(thread_registration_state == MDBX_THREAD_RTHC_ZERO)) { thread_registration_state = MDBX_THREAD_RTHC_REGISTERED; mdbx_trace("thread registered 0x%" PRIxPTR, mdbx_thread_self()); if (&__cxa_thread_atexit_impl == nullptr || __cxa_thread_atexit_impl(mdbx_rthc_thread_dtor, &thread_registration_state, (void *)&mdbx_version /* dso_anchor */)) { mdbx_ensure(nullptr, pthread_setspecific( rthc_key, &thread_registration_state) == 0); thread_registration_state = MDBX_THREAD_RTHC_COUNTED; const unsigned count_before = atomic_add32(&rthc_pending, 1); mdbx_ensure(nullptr, count_before < INT_MAX); mdbx_trace("fallback to pthreads' tsd, key %" PRIuPTR ", count %u", (uintptr_t)rthc_key, count_before); (void)count_before; } } mdbx_ensure(nullptr, pthread_setspecific(key, value) == 0); #endif } __cold void mdbx_rthc_global_init(void) { rthc_limit = RTHC_INITIAL_LIMIT; rthc_table = rthc_table_static; #if defined(_WIN32) || defined(_WIN64) InitializeCriticalSection(&rthc_critical_section); InitializeCriticalSection(&lcklist_critical_section); #else mdbx_ensure(nullptr, pthread_key_create(&rthc_key, mdbx_rthc_thread_dtor) == 0); mdbx_trace("pid %d, &mdbx_rthc_key = %p, value 0x%x", mdbx_getpid(), __Wpedantic_format_voidptr(&rthc_key), (unsigned)rthc_key); #endif /* checking time conversion, this also avoids racing on 32-bit architectures * during writing calculated 64-bit ratio(s) into memory. */ uint32_t proba = UINT32_MAX; while (true) { unsigned time_conversion_checkup = mdbx_osal_monotime_to_16dot16(mdbx_osal_16dot16_to_monotime(proba)); unsigned one_more = (proba < UINT32_MAX) ? proba + 1 : proba; unsigned one_less = (proba > 0) ? proba - 1 : proba; mdbx_ensure(nullptr, time_conversion_checkup >= one_less && time_conversion_checkup <= one_more); if (proba == 0) break; proba >>= 1; } bootid = mdbx_osal_bootid(); } /* dtor called for thread, i.e. for all mdbx's environment objects */ __cold void mdbx_rthc_thread_dtor(void *ptr) { rthc_lock(); mdbx_trace(">> pid %d, thread 0x%" PRIxPTR ", rthc %p", mdbx_getpid(), mdbx_thread_self(), ptr); const uint32_t self_pid = mdbx_getpid(); for (unsigned i = 0; i < rthc_count; ++i) { if (!rthc_table[i].key_valid) continue; const mdbx_thread_key_t key = rthc_table[i].thr_tls_key; MDBX_reader *const rthc = thread_rthc_get(key); if (rthc < rthc_table[i].begin || rthc >= rthc_table[i].end) continue; #if !defined(_WIN32) && !defined(_WIN64) if (pthread_setspecific(key, nullptr) != 0) { mdbx_trace("== thread 0x%" PRIxPTR ", rthc %p: ignore race with tsd-key deletion", mdbx_thread_self(), ptr); continue /* ignore race with tsd-key deletion by mdbx_env_close() */; } #endif mdbx_trace("== thread 0x%" PRIxPTR ", rthc %p, [%i], %p ... %p (%+i), rtch-pid %i, " "current-pid %i", mdbx_thread_self(), __Wpedantic_format_voidptr(rthc), i, __Wpedantic_format_voidptr(rthc_table[i].begin), __Wpedantic_format_voidptr(rthc_table[i].end), (int)(rthc - rthc_table[i].begin), rthc->mr_pid.weak, self_pid); if (atomic_load32(&rthc->mr_pid, mo_Relaxed) == self_pid) { mdbx_trace("==== thread 0x%" PRIxPTR ", rthc %p, cleanup", mdbx_thread_self(), __Wpedantic_format_voidptr(rthc)); atomic_store32(&rthc->mr_pid, 0, mo_AcquireRelease); } } #if defined(_WIN32) || defined(_WIN64) mdbx_trace("<< thread 0x%" PRIxPTR ", rthc %p", mdbx_thread_self(), ptr); rthc_unlock(); #else const char self_registration = *(char *)ptr; *(char *)ptr = MDBX_THREAD_RTHC_ZERO; mdbx_trace("== thread 0x%" PRIxPTR ", rthc %p, pid %d, self-status %d", mdbx_thread_self(), ptr, mdbx_getpid(), self_registration); if (self_registration == MDBX_THREAD_RTHC_COUNTED) mdbx_ensure(nullptr, atomic_sub32(&rthc_pending, 1) > 0); if (atomic_load32(&rthc_pending, mo_AcquireRelease) == 0) { mdbx_trace("== thread 0x%" PRIxPTR ", rthc %p, pid %d, wake", mdbx_thread_self(), ptr, mdbx_getpid()); mdbx_ensure(nullptr, pthread_cond_broadcast(&rthc_cond) == 0); } mdbx_trace("<< thread 0x%" PRIxPTR ", rthc %p", mdbx_thread_self(), ptr); /* Allow tail call optimization, i.e. gcc should generate the jmp instruction * instead of a call for pthread_mutex_unlock() and therefore CPU could not * return to current DSO's code section, which may be unloaded immediately * after the mutex got released. */ pthread_mutex_unlock(&rthc_mutex); #endif } __cold void mdbx_rthc_global_dtor(void) { mdbx_trace(">> pid %d", mdbx_getpid()); rthc_lock(); #if !defined(_WIN32) && !defined(_WIN64) char *rthc = (char *)pthread_getspecific(rthc_key); mdbx_trace("== thread 0x%" PRIxPTR ", rthc %p, pid %d, self-status %d", mdbx_thread_self(), __Wpedantic_format_voidptr(rthc), mdbx_getpid(), rthc ? *rthc : -1); if (rthc) { const char self_registration = *(char *)rthc; *rthc = MDBX_THREAD_RTHC_ZERO; if (self_registration == MDBX_THREAD_RTHC_COUNTED) mdbx_ensure(nullptr, atomic_sub32(&rthc_pending, 1) > 0); } struct timespec abstime; mdbx_ensure(nullptr, clock_gettime(CLOCK_REALTIME, &abstime) == 0); abstime.tv_nsec += 1000000000l / 10; if (abstime.tv_nsec >= 1000000000l) { abstime.tv_nsec -= 1000000000l; abstime.tv_sec += 1; } #if MDBX_DEBUG > 0 abstime.tv_sec += 600; #endif for (unsigned left; (left = atomic_load32(&rthc_pending, mo_AcquireRelease)) > 0;) { mdbx_trace("pid %d, pending %u, wait for...", mdbx_getpid(), left); const int rc = pthread_cond_timedwait(&rthc_cond, &rthc_mutex, &abstime); if (rc && rc != EINTR) break; } thread_key_delete(rthc_key); #endif const uint32_t self_pid = mdbx_getpid(); for (unsigned i = 0; i < rthc_count; ++i) { if (!rthc_table[i].key_valid) continue; const mdbx_thread_key_t key = rthc_table[i].thr_tls_key; thread_key_delete(key); for (MDBX_reader *rthc = rthc_table[i].begin; rthc < rthc_table[i].end; ++rthc) { mdbx_trace( "== [%i] = key %" PRIuPTR ", %p ... %p, rthc %p (%+i), " "rthc-pid %i, current-pid %i", i, (uintptr_t)key, __Wpedantic_format_voidptr(rthc_table[i].begin), __Wpedantic_format_voidptr(rthc_table[i].end), __Wpedantic_format_voidptr(rthc), (int)(rthc - rthc_table[i].begin), rthc->mr_pid.weak, self_pid); if (atomic_load32(&rthc->mr_pid, mo_Relaxed) == self_pid) { atomic_store32(&rthc->mr_pid, 0, mo_AcquireRelease); mdbx_trace("== cleanup %p", __Wpedantic_format_voidptr(rthc)); } } } rthc_limit = rthc_count = 0; if (rthc_table != rthc_table_static) mdbx_free(rthc_table); rthc_table = nullptr; rthc_unlock(); #if defined(_WIN32) || defined(_WIN64) DeleteCriticalSection(&lcklist_critical_section); DeleteCriticalSection(&rthc_critical_section); #else /* LY: yielding a few timeslices to give a more chance * to racing destructor(s) for completion. */ workaround_glibc_bug21031(); #endif mdbx_trace("<< pid %d\n", mdbx_getpid()); } __cold int mdbx_rthc_alloc(mdbx_thread_key_t *key, MDBX_reader *begin, MDBX_reader *end) { int rc; if (key) { #ifndef NDEBUG *key = (mdbx_thread_key_t)0xBADBADBAD; #endif /* NDEBUG */ rc = thread_key_create(key); if (rc != MDBX_SUCCESS) return rc; } rthc_lock(); const mdbx_thread_key_t new_key = key ? *key : 0; mdbx_trace(">> key %" PRIuPTR ", rthc_count %u, rthc_limit %u", (uintptr_t)new_key, rthc_count, rthc_limit); if (rthc_count == rthc_limit) { rthc_entry_t *new_table = mdbx_realloc((rthc_table == rthc_table_static) ? nullptr : rthc_table, sizeof(rthc_entry_t) * rthc_limit * 2); if (new_table == nullptr) { rc = MDBX_ENOMEM; goto bailout; } if (rthc_table == rthc_table_static) memcpy(new_table, rthc_table_static, sizeof(rthc_table_static)); rthc_table = new_table; rthc_limit *= 2; } mdbx_trace("== [%i] = key %" PRIuPTR ", %p ... %p", rthc_count, (uintptr_t)new_key, __Wpedantic_format_voidptr(begin), __Wpedantic_format_voidptr(end)); rthc_table[rthc_count].key_valid = key ? true : false; rthc_table[rthc_count].thr_tls_key = key ? new_key : 0; rthc_table[rthc_count].begin = begin; rthc_table[rthc_count].end = end; ++rthc_count; mdbx_trace("<< key %" PRIuPTR ", rthc_count %u, rthc_limit %u", (uintptr_t)new_key, rthc_count, rthc_limit); rthc_unlock(); return MDBX_SUCCESS; bailout: if (key) thread_key_delete(*key); rthc_unlock(); return rc; } __cold void mdbx_rthc_remove(const mdbx_thread_key_t key) { thread_key_delete(key); rthc_lock(); mdbx_trace(">> key %zu, rthc_count %u, rthc_limit %u", (size_t)key, rthc_count, rthc_limit); for (unsigned i = 0; i < rthc_count; ++i) { if (rthc_table[i].key_valid && key == rthc_table[i].thr_tls_key) { const uint32_t self_pid = mdbx_getpid(); mdbx_trace("== [%i], %p ...%p, current-pid %d", i, __Wpedantic_format_voidptr(rthc_table[i].begin), __Wpedantic_format_voidptr(rthc_table[i].end), self_pid); for (MDBX_reader *rthc = rthc_table[i].begin; rthc < rthc_table[i].end; ++rthc) { if (atomic_load32(&rthc->mr_pid, mo_Relaxed) == self_pid) { atomic_store32(&rthc->mr_pid, 0, mo_AcquireRelease); mdbx_trace("== cleanup %p", __Wpedantic_format_voidptr(rthc)); } } if (--rthc_count > 0) rthc_table[i] = rthc_table[rthc_count]; else if (rthc_table != rthc_table_static) { mdbx_free(rthc_table); rthc_table = rthc_table_static; rthc_limit = RTHC_INITIAL_LIMIT; } break; } } mdbx_trace("<< key %zu, rthc_count %u, rthc_limit %u", (size_t)key, rthc_count, rthc_limit); rthc_unlock(); } //------------------------------------------------------------------------------ #define RTHC_ENVLIST_END ((MDBX_env *)((uintptr_t)50459)) static MDBX_env *inprocess_lcklist_head = RTHC_ENVLIST_END; static __inline void lcklist_lock(void) { #if defined(_WIN32) || defined(_WIN64) EnterCriticalSection(&lcklist_critical_section); #else mdbx_ensure(nullptr, pthread_mutex_lock(&lcklist_mutex) == 0); #endif } static __inline void lcklist_unlock(void) { #if defined(_WIN32) || defined(_WIN64) LeaveCriticalSection(&lcklist_critical_section); #else mdbx_ensure(nullptr, pthread_mutex_unlock(&lcklist_mutex) == 0); #endif } MDBX_NOTHROW_CONST_FUNCTION static uint64_t rrxmrrxmsx_0(uint64_t v) { /* Pelle Evensen's mixer, https://bit.ly/2HOfynt */ v ^= (v << 39 | v >> 25) ^ (v << 14 | v >> 50); v *= UINT64_C(0xA24BAED4963EE407); v ^= (v << 40 | v >> 24) ^ (v << 15 | v >> 49); v *= UINT64_C(0x9FB21C651E98DF25); return v ^ v >> 28; } static int uniq_peek(const mdbx_mmap_t *pending, mdbx_mmap_t *scan) { int rc; uint64_t bait; if (pending->address) { bait = atomic_load64(&pending->lck->mti_bait_uniqueness, mo_AcquireRelease); rc = MDBX_SUCCESS; } else { bait = 0 /* hush MSVC warning */; rc = mdbx_msync(scan, 0, sizeof(MDBX_lockinfo), true); if (rc == MDBX_SUCCESS) rc = mdbx_pread(pending->fd, &bait, sizeof(scan->lck->mti_bait_uniqueness), offsetof(MDBX_lockinfo, mti_bait_uniqueness)); } if (likely(rc == MDBX_SUCCESS) && bait == atomic_load64(&scan->lck->mti_bait_uniqueness, mo_AcquireRelease)) rc = MDBX_RESULT_TRUE; mdbx_trace("uniq-peek: %s, bait 0x%016" PRIx64 ",%s rc %d", pending->lck ? "mem" : "file", bait, (rc == MDBX_RESULT_TRUE) ? " found," : (rc ? " FAILED," : ""), rc); return rc; } static int uniq_poke(const mdbx_mmap_t *pending, mdbx_mmap_t *scan, uint64_t *abra) { if (*abra == 0) { const uintptr_t tid = mdbx_thread_self(); uintptr_t uit = 0; memcpy(&uit, &tid, (sizeof(tid) < sizeof(uit)) ? sizeof(tid) : sizeof(uit)); *abra = rrxmrrxmsx_0(mdbx_osal_monotime() + UINT64_C(5873865991930747) * uit); } const uint64_t cadabra = rrxmrrxmsx_0(*abra + UINT64_C(7680760450171793) * (unsigned)mdbx_getpid()) << 24 | *abra >> 40; atomic_store64(&scan->lck->mti_bait_uniqueness, cadabra, mo_SequentialConsistency); *abra = *abra * UINT64_C(6364136223846793005) + 1; return uniq_peek(pending, scan); } __cold static int uniq_check(const mdbx_mmap_t *pending, MDBX_env **found) { *found = nullptr; uint64_t salt = 0; for (MDBX_env *scan = inprocess_lcklist_head; scan != RTHC_ENVLIST_END; scan = scan->me_lcklist_next) { int err = atomic_load64(&scan->me_lck_mmap.lck->mti_bait_uniqueness, mo_AcquireRelease) ? uniq_peek(pending, &scan->me_lck_mmap) : uniq_poke(pending, &scan->me_lck_mmap, &salt); if (err == MDBX_ENODATA) { uint64_t length; if (likely(mdbx_filesize(pending->fd, &length) == MDBX_SUCCESS && length == 0)) { /* LY: skip checking since LCK-file is empty, i.e. just created. */ mdbx_debug("uniq-probe: %s", "unique (new/empty lck)"); return MDBX_RESULT_TRUE; } } if (err == MDBX_RESULT_TRUE) err = uniq_poke(pending, &scan->me_lck_mmap, &salt); if (err == MDBX_RESULT_TRUE) { (void)mdbx_msync(&scan->me_lck_mmap, 0, sizeof(MDBX_lockinfo), false); err = uniq_poke(pending, &scan->me_lck_mmap, &salt); } if (err == MDBX_RESULT_TRUE) { err = uniq_poke(pending, &scan->me_lck_mmap, &salt); *found = scan; mdbx_debug("uniq-probe: found %p", __Wpedantic_format_voidptr(*found)); return MDBX_RESULT_FALSE; } if (unlikely(err != MDBX_SUCCESS)) { mdbx_debug("uniq-probe: failed rc %d", err); return err; } } mdbx_debug("uniq-probe: %s", "unique"); return MDBX_RESULT_TRUE; } static int lcklist_detach_locked(MDBX_env *env) { MDBX_env *inprocess_neighbor = nullptr; int rc = MDBX_SUCCESS; if (env->me_lcklist_next != nullptr) { mdbx_ensure(env, env->me_lcklist_next != nullptr); mdbx_ensure(env, inprocess_lcklist_head != RTHC_ENVLIST_END); for (MDBX_env **ptr = &inprocess_lcklist_head; *ptr != RTHC_ENVLIST_END; ptr = &(*ptr)->me_lcklist_next) { if (*ptr == env) { *ptr = env->me_lcklist_next; env->me_lcklist_next = nullptr; break; } } mdbx_ensure(env, env->me_lcklist_next == nullptr); } rc = likely(mdbx_getpid() == env->me_pid) ? uniq_check(&env->me_lck_mmap, &inprocess_neighbor) : MDBX_PANIC; if (!inprocess_neighbor && env->me_live_reader) (void)mdbx_rpid_clear(env); if (!MDBX_IS_ERROR(rc)) rc = mdbx_lck_destroy(env, inprocess_neighbor); return rc; } /*------------------------------------------------------------------------------ * LY: State of the art quicksort-based sorting, with internal stack * and network-sort for small chunks. * Thanks to John M. Gamble for the http://pages.ripco.net/~jgamble/nw.html */ #define SORT_CMP_SWAP(TYPE, CMP, a, b) \ do { \ const TYPE swap_tmp = (a); \ const bool swap_cmp = CMP(swap_tmp, b); \ (a) = swap_cmp ? swap_tmp : b; \ (b) = swap_cmp ? b : swap_tmp; \ } while (0) // 3 comparators, 3 parallel operations // o-----^--^--o // | | // o--^--|--v--o // | | // o--v--v-----o // // [[1,2]] // [[0,2]] // [[0,1]] #define SORT_NETWORK_3(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ } while (0) // 5 comparators, 3 parallel operations // o--^--^--------o // | | // o--v--|--^--^--o // | | | // o--^--v--|--v--o // | | // o--v-----v-----o // // [[0,1],[2,3]] // [[0,2],[1,3]] // [[1,2]] #define SORT_NETWORK_4(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ } while (0) // 9 comparators, 5 parallel operations // o--^--^-----^-----------o // | | | // o--|--|--^--v-----^--^--o // | | | | | // o--|--v--|--^--^--|--v--o // | | | | | // o--|-----v--|--v--|--^--o // | | | | // o--v--------v-----v--v--o // // [[0,4],[1,3]] // [[0,2]] // [[2,4],[0,1]] // [[2,3],[1,4]] // [[1,2],[3,4]] #define SORT_NETWORK_5(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ } while (0) // 12 comparators, 6 parallel operations // o-----^--^--^-----------------o // | | | // o--^--|--v--|--^--------^-----o // | | | | | // o--v--v-----|--|--^--^--|--^--o // | | | | | | // o-----^--^--v--|--|--|--v--v--o // | | | | | // o--^--|--v-----v--|--v--------o // | | | // o--v--v-----------v-----------o // // [[1,2],[4,5]] // [[0,2],[3,5]] // [[0,1],[3,4],[2,5]] // [[0,3],[1,4]] // [[2,4],[1,3]] // [[2,3]] #define SORT_NETWORK_6(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ } while (0) // 16 comparators, 6 parallel operations // o--^--------^-----^-----------------o // | | | // o--|--^-----|--^--v--------^--^-----o // | | | | | | // o--|--|--^--v--|--^-----^--|--v-----o // | | | | | | | // o--|--|--|-----v--|--^--v--|--^--^--o // | | | | | | | | // o--v--|--|--^-----v--|--^--v--|--v--o // | | | | | | // o-----v--|--|--------v--v-----|--^--o // | | | | // o--------v--v-----------------v--v--o // // [[0,4],[1,5],[2,6]] // [[0,2],[1,3],[4,6]] // [[2,4],[3,5],[0,1]] // [[2,3],[4,5]] // [[1,4],[3,6]] // [[1,2],[3,4],[5,6]] #define SORT_NETWORK_7(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ } while (0) // 19 comparators, 6 parallel operations // o--^--------^-----^-----------------o // | | | // o--|--^-----|--^--v--------^--^-----o // | | | | | | // o--|--|--^--v--|--^-----^--|--v-----o // | | | | | | | // o--|--|--|--^--v--|--^--v--|--^--^--o // | | | | | | | | | // o--v--|--|--|--^--v--|--^--v--|--v--o // | | | | | | | // o-----v--|--|--|--^--v--v-----|--^--o // | | | | | | // o--------v--|--v--|--^--------v--v--o // | | | // o-----------v-----v--v--------------o // // [[0,4],[1,5],[2,6],[3,7]] // [[0,2],[1,3],[4,6],[5,7]] // [[2,4],[3,5],[0,1],[6,7]] // [[2,3],[4,5]] // [[1,4],[3,6]] // [[1,2],[3,4],[5,6]] #define SORT_NETWORK_8(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ } while (0) // 25 comparators, 9 parallel operations // o--^-----^--^-----^-----------------------------------o // | | | | // o--v--^--v--|-----|--^-----^-----------^--------------o // | | | | | | // o-----v-----|-----|--|-----|--^-----^--|--^-----^--^--o // | | | | | | | | | | // o--^-----^--v--^--v--|-----|--|-----|--v--|-----|--v--o // | | | | | | | | | // o--v--^--v-----|-----v--^--v--|-----|-----|--^--v-----o // | | | | | | | // o-----v--------|--------|-----v--^--v--^--|--|--^-----o // | | | | | | | // o--^-----^-----v--------|--------|-----|--v--v--v-----o // | | | | | // o--v--^--v--------------v--------|-----v--------------o // | | // o-----v--------------------------v--------------------o // // [[0,1],[3,4],[6,7]] // [[1,2],[4,5],[7,8]] // [[0,1],[3,4],[6,7],[2,5]] // [[0,3],[1,4],[5,8]] // [[3,6],[4,7],[2,5]] // [[0,3],[1,4],[5,7],[2,6]] // [[1,3],[4,6]] // [[2,4],[5,6]] // [[2,3]] #define SORT_NETWORK_9(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ } while (0) // 29 comparators, 9 parallel operations // o--------------^-----^--^--^-----------------------o // | | | | // o-----------^--|--^--|--|--v--^--------^-----------o // | | | | | | | // o--------^--|--|--|--|--v--^--v-----^--|--^--------o // | | | | | | | | | // o-----^--|--|--|--|--v--^--|-----^--|--v--v--^-----o // | | | | | | | | | | // o--^--|--|--|--|--v-----|--v--^--|--|--^-----v--^--o // | | | | | | | | | | | // o--|--|--|--|--v--^-----|--^--|--v--v--|-----^--v--o // | | | | | | | | | | // o--|--|--|--v--^--|-----v--|--v--^-----|--^--v-----o // | | | | | | | | | // o--|--|--v-----|--|--^-----v--^--|-----v--v--------o // | | | | | | | // o--|--v--------|--v--|--^-----v--v-----------------o // | | | | // o--v-----------v-----v--v--------------------------o // // [[4,9],[3,8],[2,7],[1,6],[0,5]] // [[1,4],[6,9],[0,3],[5,8]] // [[0,2],[3,6],[7,9]] // [[0,1],[2,4],[5,7],[8,9]] // [[1,2],[4,6],[7,8],[3,5]] // [[2,5],[6,8],[1,3],[4,7]] // [[2,3],[6,7]] // [[3,4],[5,6]] // [[4,5]] #define SORT_NETWORK_10(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ } while (0) // 35 comparators, 9 parallel operations // o--^-----^-----------------^--------^--------------------o // | | | | // o--v--^--|--^--^--------^--|--------|--^-----------------o // | | | | | | | | // o--^--|--v--v--|-----^--|--|--------|--|-----^--^--------o // | | | | | | | | | | // o--v--v--------|-----|--|--|--^-----|--|--^--v--|--^--^--o // | | | | | | | | | | | // o--^-----^-----|-----|--|--v--|--^--v--v--|-----v--|--v--o // | | | | | | | | | // o--v--^--|--^--v--^--|--v-----|--|--------|--------v--^--o // | | | | | | | | | // o--^--|--v--v--^--|--v--^-----|--|--------|--------^--v--o // | | | | | | | | | // o--v--v--------|--|-----|-----v--|--^-----|-----^--|--^--o // | | | | | | | | | // o--^--^--------|--|-----|--------v--|-----v--^--|--v--v--o // | | | | | | | | // o--v--|--^-----|--v-----|-----------|--------v--v--------o // | | | | | // o-----v--v-----v--------v-----------v--------------------o // // [[0,1],[2,3],[4,5],[6,7],[8,9]] // [[1,3],[5,7],[0,2],[4,6],[8,10]] // [[1,2],[5,6],[9,10],[0,4],[3,7]] // [[1,5],[6,10],[4,8]] // [[5,9],[2,6],[0,4],[3,8]] // [[1,5],[6,10],[2,3],[8,9]] // [[1,4],[7,10],[3,5],[6,8]] // [[2,4],[7,9],[5,6]] // [[3,4],[7,8]] #define SORT_NETWORK_11(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ } while (0) // 39 comparators, parallel operations // o--^-----^-----------------^--------^--------------------o // | | | | // o--v--^--|--^--^--------^--|--------|--^-----------------o // | | | | | | | | // o--^--|--v--v--|-----^--|--|--------|--|-----^--^--------o // | | | | | | | | | | // o--v--v--------|-----|--|--|--^-----|--|--^--v--|--^--^--o // | | | | | | | | | | | // o--^-----^-----|-----|--|--v--|--^--v--v--|-----v--|--v--o // | | | | | | | | | // o--v--^--|--^--v--^--|--v-----|--|--------|--------v--^--o // | | | | | | | | | // o--^--|--v--v--^--|--v--^-----|--|--------|--------^--v--o // | | | | | | | | | // o--v--v--------|--|-----|--^--v--|--^--^--|-----^--|--^--o // | | | | | | | | | | | // o--^-----^-----|--|-----|--|-----v--|--|--v--^--|--v--v--o // | | | | | | | | | | // o--v--^--|--^--|--v-----|--|--------|--|-----v--v--------o // | | | | | | | | // o--^--|--v--v--v--------v--|--------|--v-----------------o // | | | | // o--v--v--------------------v--------v--------------------o // // [[0,1],[2,3],[4,5],[6,7],[8,9],[10,11]] // [[1,3],[5,7],[9,11],[0,2],[4,6],[8,10]] // [[1,2],[5,6],[9,10],[0,4],[7,11]] // [[1,5],[6,10],[3,7],[4,8]] // [[5,9],[2,6],[0,4],[7,11],[3,8]] // [[1,5],[6,10],[2,3],[8,9]] // [[1,4],[7,10],[3,5],[6,8]] // [[2,4],[7,9],[5,6]] // [[3,4],[7,8]] #define SORT_NETWORK_12(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ } while (0) // 45 comparators, 10 parallel operations // o--------^--^-----^-----------------------------^-----------------o // | | | | // o--^-----|--v-----|-----^--------------^-----^--|-----^-----------o // | | | | | | | | // o--|-----|--^--^--v-----|--------------|--^--|--|--^--v--^--------o // | | | | | | | | | | | // o--|--^--|--|--v-----^--|--------^-----|--|--v--|--|--^--v-----^--o // | | | | | | | | | | | | | // o--|--v--|--|--^-----|--v-----^--v-----|--|--^--|--|--|--^--^--v--o // | | | | | | | | | | | | | | // o--|--^--|--|--|--^--|--------|-----^--|--|--|--v--v--v--|--v--^--o // | | | | | | | | | | | | | | // o--|--|--|--v--v--|--|--^-----|--^--v--|--v--|--^--------v--^--v--o // | | | | | | | | | | | | // o--v--|--|-----^--|--v--|--^--|--|-----v-----v--|--^--------v-----o // | | | | | | | | | | // o-----v--|--^--|--|-----|--v--|--|--^-----^-----v--v--^-----------o // | | | | | | | | | | // o--^-----|--|--|--v-----|-----v--|--v--^--|--^--------v-----------o // | | | | | | | | | // o--|-----|--|--|--^-----|--------v--^--|--v--v--------------------o // | | | | | | | | // o--v-----|--v--|--v-----|--^--------v--v--------------------------o // | | | | // o--------v-----v--------v--v--------------------------------------o // // [[1,7],[9,11],[3,4],[5,8],[0,12],[2,6]] // [[0,1],[2,3],[4,6],[8,11],[7,12],[5,9]] // [[0,2],[3,7],[10,11],[1,4],[6,12]] // [[7,8],[11,12],[4,9],[6,10]] // [[3,4],[5,6],[8,9],[10,11],[1,7]] // [[2,6],[9,11],[1,3],[4,7],[8,10],[0,5]] // [[2,5],[6,8],[9,10]] // [[1,2],[3,5],[7,8],[4,6]] // [[2,3],[4,5],[6,7],[8,9]] // [[3,4],[5,6]] #define SORT_NETWORK_13(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ } while (0) /* *INDENT-OFF* */ /* clang-format off */ // 51 comparators, 10 parallel operations // o--^--^-----^-----------^-----------------------------------------------------------o // | | | | // o--v--|--^--|--^--------|--^-----^-----------------------^--------------------------o // | | | | | | | | // o--^--v--|--|--|--^-----|--|--^--v-----------------------|--^--^--------------------o // | | | | | | | | | | | // o--v-----v--|--|--|--^--|--|--|--^--------------^--------|--|--|--^--^--^-----------o // | | | | | | | | | | | | | | | // o--^--^-----v--|--|--|--|--|--|--|--^-----------|-----^--v--|--v--|--|--v-----------o // | | | | | | | | | | | | | | | // o--v--|--^-----v--|--|--|--|--|--|--|--^--^-----|-----|-----|--^--|--v-----^--------o // | | | | | | | | | | | | | | | | | // o--^--v--|--------v--|--|--|--|--|--|--|--|--^--|-----|-----|--v--|-----^--v-----^--o // | | | | | | | | | | | | | | | | | // o--v-----v-----------v--|--|--|--|--|--|--|--|--|--^--|--^--|-----|--^--|--^--^--v--o // | | | | | | | | | | | | | | | | | | // o--^--^-----^-----------v--|--|--|--|--|--|--|--|--|--v--|--v-----v--|--v--|--v--^--o // | | | | | | | | | | | | | | | | // o--v--|--^--|--^-----------v--|--|--|--|--|--v--|--|-----|--^--------|-----v--^--v--o // | | | | | | | | | | | | | | | // o--^--v--|--|--|--------------v--|--|--|--v-----|--|-----|--v--------|--^-----v-----o // | | | | | | | | | | | | // o--v-----v--|--|-----------------v--|--|--------|--v-----|--^--------|--|--^--------o // | | | | | | | | | | // o--^--------v--|--------------------v--|--------v--------|--|--------v--v--v--------o // | | | | | // o--v-----------v-----------------------v-----------------v--v-----------------------o // // [[0,1],[2,3],[4,5],[6,7],[8,9],[10,11],[12,13]] // [[0,2],[4,6],[8,10],[1,3],[5,7],[9,11]] // [[0,4],[8,12],[1,5],[9,13],[2,6],[3,7]] // [[0,8],[1,9],[2,10],[3,11],[4,12],[5,13]] // [[5,10],[6,9],[3,12],[7,11],[1,2],[4,8]] // [[1,4],[7,13],[2,8],[5,6],[9,10]] // [[2,4],[11,13],[3,8],[7,12]] // [[6,8],[10,12],[3,5],[7,9]] // [[3,4],[5,6],[7,8],[9,10],[11,12]] // [[6,7],[8,9]] /* *INDENT-ON* */ /* clang-format on */ #define SORT_NETWORK_14(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[12], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ } while (0) /* *INDENT-OFF* */ /* clang-format off */ // 56 comparators, 10 parallel operations // o--^--^-----^-----------^--------------------------------------------------------------o // | | | | // o--v--|--^--|--^--------|--^-----^--------------------------^--------------------------o // | | | | | | | | // o--^--v--|--|--|--^-----|--|--^--v--------------------------|--^--^--------------------o // | | | | | | | | | | | // o--v-----v--|--|--|--^--|--|--|--^-----------------^--------|--|--|--^--^--^-----------o // | | | | | | | | | | | | | | | // o--^--^-----v--|--|--|--|--|--|--|--^--------------|-----^--v--|--v--|--|--v-----------o // | | | | | | | | | | | | | | | // o--v--|--^-----v--|--|--|--|--|--|--|--^-----^-----|-----|-----|--^--|--v-----^--------o // | | | | | | | | | | | | | | | | | // o--^--v--|--------v--|--|--|--|--|--|--|--^--|--^--|-----|-----|--v--|-----^--v-----^--o // | | | | | | | | | | | | | | | | | | // o--v-----v-----------v--|--|--|--|--|--|--|--|--|--|--^--|--^--|-----|--^--|--^--^--v--o // | | | | | | | | | | | | | | | | | | | // o--^--^-----^-----------v--|--|--|--|--|--|--|--|--|--|--v--|--v-----v--|--v--|--v--^--o // | | | | | | | | | | | | | | | | | // o--v--|--^--|--^-----------v--|--|--|--|--|--|--v--|--|-----|--^--------|-----v--^--v--o // | | | | | | | | | | | | | | | | // o--^--v--|--|--|--^-----------v--|--|--|--|--v-----|--|-----|--v--------|--^-----v-----o // | | | | | | | | | | | | | | // o--v-----v--|--|--|--------------v--|--|--|--------|--v-----|--^--^-----|--|--^--------o // | | | | | | | | | | | | | // o--^--^-----v--|--|-----------------v--|--|--------v--------|--|--|-----v--v--v--------o // | | | | | | | | | // o--v--|--------v--|--------------------v--|--^--------------v--|--v--------------------o // | | | | | // o-----v-----------v-----------------------v--v-----------------v-----------------------o // // [[0,1],[2,3],[4,5],[6,7],[8,9],[10,11],[12,13]] // [[0,2],[4,6],[8,10],[12,14],[1,3],[5,7],[9,11]] // [[0,4],[8,12],[1,5],[9,13],[2,6],[10,14],[3,7]] // [[0,8],[1,9],[2,10],[3,11],[4,12],[5,13],[6,14]] // [[5,10],[6,9],[3,12],[13,14],[7,11],[1,2],[4,8]] // [[1,4],[7,13],[2,8],[11,14],[5,6],[9,10]] // [[2,4],[11,13],[3,8],[7,12]] // [[6,8],[10,12],[3,5],[7,9]] // [[3,4],[5,6],[7,8],[9,10],[11,12]] // [[6,7],[8,9]] /* *INDENT-ON* */ /* clang-format on */ #define SORT_NETWORK_15(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[12], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[12], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[13], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ } while (0) /* *INDENT-OFF* */ /* clang-format off */ // 60 comparators, 10 parallel operations // o--^--^-----^-----------^-----------------------------------------------------------------o // | | | | // o--v--|--^--|--^--------|--^-----^-----------------------------^--------------------------o // | | | | | | | | // o--^--v--|--|--|--^-----|--|--^--v-----------------------------|--^--^--------------------o // | | | | | | | | | | | // o--v-----v--|--|--|--^--|--|--|--^--------------------^--------|--|--|--^--^--^-----------o // | | | | | | | | | | | | | | | // o--^--^-----v--|--|--|--|--|--|--|--^-----------------|-----^--v--|--v--|--|--v-----------o // | | | | | | | | | | | | | | | // o--v--|--^-----v--|--|--|--|--|--|--|--^--------^-----|-----|-----|--^--|--v-----^--------o // | | | | | | | | | | | | | | | | | // o--^--v--|--------v--|--|--|--|--|--|--|--^-----|--^--|-----|-----|--v--|-----^--v-----^--o // | | | | | | | | | | | | | | | | | | // o--v-----v-----------v--|--|--|--|--|--|--|--^--|--|--|--^--|--^--|-----|--^--|--^--^--v--o // | | | | | | | | | | | | | | | | | | | | // o--^--^-----^-----------v--|--|--|--|--|--|--|--|--|--|--|--v--|--v-----v--|--v--|--v--^--o // | | | | | | | | | | | | | | | | | | // o--v--|--^--|--^-----------v--|--|--|--|--|--|--|--v--|--|-----|--^--------|-----v--^--v--o // | | | | | | | | | | | | | | | | | // o--^--v--|--|--|--^-----------v--|--|--|--|--|--v-----|--|-----|--v--------|--^-----v-----o // | | | | | | | | | | | | | | | // o--v-----v--|--|--|--^-----------v--|--|--|--|--------|--v-----|--^--^-----|--|--^--------o // | | | | | | | | | | | | | | | // o--^--^-----v--|--|--|--------------v--|--|--|--------v--------|--|--|-----v--v--v--------o // | | | | | | | | | | | // o--v--|--^-----v--|--|-----------------v--|--|--^--------------v--|--v--------------------o // | | | | | | | | // o--^--v--|--------v--|--------------------v--|--v-----------------v-----------------------o // | | | | // o--v-----v-----------v-----------------------v--------------------------------------------o // // [[0,1],[2,3],[4,5],[6,7],[8,9],[10,11],[12,13],[14,15]] // [[0,2],[4,6],[8,10],[12,14],[1,3],[5,7],[9,11],[13,15]] // [[0,4],[8,12],[1,5],[9,13],[2,6],[10,14],[3,7],[11,15]] // [[0,8],[1,9],[2,10],[3,11],[4,12],[5,13],[6,14],[7,15]] // [[5,10],[6,9],[3,12],[13,14],[7,11],[1,2],[4,8]] // [[1,4],[7,13],[2,8],[11,14],[5,6],[9,10]] // [[2,4],[11,13],[3,8],[7,12]] // [[6,8],[10,12],[3,5],[7,9]] // [[3,4],[5,6],[7,8],[9,10],[11,12]] // [[6,7],[8,9]] /* *INDENT-ON* */ /* clang-format on */ #define SORT_NETWORK_16(TYPE, CMP, begin) \ do { \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[12], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[14], begin[15]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[12], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[3]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[13], begin[15]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[15]); \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[15]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[13], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[11]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[2]); \ SORT_CMP_SWAP(TYPE, CMP, begin[4], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[1], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[14]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[2], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[13]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[10], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[5]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[9]); \ SORT_CMP_SWAP(TYPE, CMP, begin[3], begin[4]); \ SORT_CMP_SWAP(TYPE, CMP, begin[5], begin[6]); \ SORT_CMP_SWAP(TYPE, CMP, begin[7], begin[8]); \ SORT_CMP_SWAP(TYPE, CMP, begin[9], begin[10]); \ SORT_CMP_SWAP(TYPE, CMP, begin[11], begin[12]); \ SORT_CMP_SWAP(TYPE, CMP, begin[6], begin[7]); \ SORT_CMP_SWAP(TYPE, CMP, begin[8], begin[9]); \ } while (0) #define SORT_INNER(TYPE, CMP, begin, end, len) \ switch (len) { \ default: \ __unreachable(); \ case 0: \ case 1: \ break; \ case 2: \ SORT_CMP_SWAP(TYPE, CMP, begin[0], begin[1]); \ break; \ case 3: \ SORT_NETWORK_3(TYPE, CMP, begin); \ break; \ case 4: \ SORT_NETWORK_4(TYPE, CMP, begin); \ break; \ case 5: \ SORT_NETWORK_5(TYPE, CMP, begin); \ break; \ case 6: \ SORT_NETWORK_6(TYPE, CMP, begin); \ break; \ case 7: \ SORT_NETWORK_7(TYPE, CMP, begin); \ break; \ case 8: \ SORT_NETWORK_8(TYPE, CMP, begin); \ break; \ case 9: \ SORT_NETWORK_9(TYPE, CMP, begin); \ break; \ case 10: \ SORT_NETWORK_10(TYPE, CMP, begin); \ break; \ case 11: \ SORT_NETWORK_11(TYPE, CMP, begin); \ break; \ case 12: \ SORT_NETWORK_12(TYPE, CMP, begin); \ break; \ case 13: \ SORT_NETWORK_13(TYPE, CMP, begin); \ break; \ case 14: \ SORT_NETWORK_14(TYPE, CMP, begin); \ break; \ case 15: \ SORT_NETWORK_15(TYPE, CMP, begin); \ break; \ case 16: \ SORT_NETWORK_16(TYPE, CMP, begin); \ break; \ } #define SORT_SWAP(TYPE, a, b) \ do { \ const TYPE swap_tmp = (a); \ (a) = (b); \ (b) = swap_tmp; \ } while (0) #define SORT_PUSH(low, high) \ do { \ top->lo = (low); \ top->hi = (high); \ ++top; \ } while (0) #define SORT_POP(low, high) \ do { \ --top; \ low = top->lo; \ high = top->hi; \ } while (0) #define SORT_IMPL(NAME, EXPECT_LOW_CARDINALITY_OR_PRESORTED, TYPE, CMP) \ \ static __inline bool NAME##_is_sorted(const TYPE *first, const TYPE *last) { \ while (++first <= last) \ if (CMP(first[0], first[-1])) \ return false; \ return true; \ } \ \ typedef struct { \ TYPE *lo, *hi; \ } NAME##_stack; \ \ static __hot void NAME(TYPE *const begin, TYPE *const end) { \ NAME##_stack stack[sizeof(unsigned) * CHAR_BIT], *top = stack; \ \ TYPE *hi = end - 1; \ TYPE *lo = begin; \ while (true) { \ const ptrdiff_t len = hi - lo; \ if (len < 16) { \ SORT_INNER(TYPE, CMP, lo, hi + 1, len + 1); \ if (unlikely(top == stack)) \ break; \ SORT_POP(lo, hi); \ continue; \ } \ \ TYPE *mid = lo + (len >> 1); \ SORT_CMP_SWAP(TYPE, CMP, *lo, *mid); \ SORT_CMP_SWAP(TYPE, CMP, *mid, *hi); \ SORT_CMP_SWAP(TYPE, CMP, *lo, *mid); \ \ TYPE *right = hi - 1; \ TYPE *left = lo + 1; \ while (1) { \ while (CMP(*left, *mid)) \ ++left; \ while (CMP(*mid, *right)) \ --right; \ if (unlikely(left > right)) { \ if (EXPECT_LOW_CARDINALITY_OR_PRESORTED) { \ if (NAME##_is_sorted(lo, right)) \ lo = right + 1; \ if (NAME##_is_sorted(left, hi)) \ hi = left; \ } \ break; \ } \ SORT_SWAP(TYPE, *left, *right); \ mid = (mid == left) ? right : (mid == right) ? left : mid; \ ++left; \ --right; \ } \ \ if (right - lo > hi - left) { \ SORT_PUSH(lo, right); \ lo = left; \ } else { \ SORT_PUSH(left, hi); \ hi = right; \ } \ } \ \ if (mdbx_audit_enabled()) { \ for (TYPE *scan = begin + 1; scan < end; ++scan) \ assert(CMP(scan[-1], scan[0])); \ } \ } /*------------------------------------------------------------------------------ * LY: Binary search */ #define SEARCH_IMPL(NAME, TYPE_LIST, TYPE_ARG, CMP) \ static __always_inline const TYPE_LIST *NAME( \ const TYPE_LIST *first, unsigned length, const TYPE_ARG item) { \ const TYPE_LIST *const begin = first, *const end = begin + length; \ \ while (length > 3) { \ const unsigned whole = length; \ length >>= 1; \ const TYPE_LIST *const middle = first + length; \ if (CMP(*middle, item)) { \ first = middle + 1; \ length = whole - length - 1; \ } \ } \ \ switch (length) { \ case 3: \ if (!CMP(*first, item)) \ break; \ ++first; \ __fallthrough /* fall through */; \ case 2: \ if (!CMP(*first, item)) \ break; \ ++first; \ __fallthrough /* fall through */; \ case 1: \ if (!CMP(*first, item)) \ break; \ ++first; \ __fallthrough /* fall through */; \ case 0: \ break; \ default: \ __unreachable(); \ } \ \ if (mdbx_audit_enabled()) { \ for (const TYPE_LIST *scan = begin; scan < first; ++scan) \ assert(CMP(*scan, item)); \ for (const TYPE_LIST *scan = first; scan < end; ++scan) \ assert(!CMP(*scan, item)); \ (void)begin, (void)end; \ } \ \ return first; \ } /*----------------------------------------------------------------------------*/ static __always_inline size_t pnl2bytes(const size_t size) { assert(size > 0 && size <= MDBX_PGL_LIMIT); size_t bytes = ceil_powerof2(MDBX_ASSUME_MALLOC_OVERHEAD + sizeof(pgno_t) * (size + 2), MDBX_PNL_GRANULATE * sizeof(pgno_t)) - MDBX_ASSUME_MALLOC_OVERHEAD; return bytes; } static __always_inline pgno_t bytes2pnl(const size_t bytes) { size_t size = bytes / sizeof(pgno_t); assert(size > 2 && size <= MDBX_PGL_LIMIT); return (pgno_t)size - 2; } static MDBX_PNL mdbx_pnl_alloc(size_t size) { size_t bytes = pnl2bytes(size); MDBX_PNL pl = mdbx_malloc(bytes); if (likely(pl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(pl); #endif /* malloc_usable_size */ pl[0] = bytes2pnl(bytes); assert(pl[0] >= size); pl[1] = 0; pl += 1; } return pl; } static void mdbx_pnl_free(MDBX_PNL pl) { if (likely(pl)) mdbx_free(pl - 1); } /* Shrink the PNL to the default size if it has grown larger */ static void mdbx_pnl_shrink(MDBX_PNL *ppl) { assert(bytes2pnl(pnl2bytes(MDBX_PNL_INITIAL)) == MDBX_PNL_INITIAL); assert(MDBX_PNL_SIZE(*ppl) <= MDBX_PGL_LIMIT && MDBX_PNL_ALLOCLEN(*ppl) >= MDBX_PNL_SIZE(*ppl)); MDBX_PNL_SIZE(*ppl) = 0; if (unlikely(MDBX_PNL_ALLOCLEN(*ppl) > MDBX_PNL_INITIAL * 2 - MDBX_CACHELINE_SIZE / sizeof(pgno_t))) { size_t bytes = pnl2bytes(MDBX_PNL_INITIAL); MDBX_PNL pl = mdbx_realloc(*ppl - 1, bytes); if (likely(pl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(pl); #endif /* malloc_usable_size */ *pl = bytes2pnl(bytes); *ppl = pl + 1; } } } /* Grow the PNL to the size growed to at least given size */ static int mdbx_pnl_reserve(MDBX_PNL *ppl, const size_t wanna) { const size_t allocated = MDBX_PNL_ALLOCLEN(*ppl); assert(MDBX_PNL_SIZE(*ppl) <= MDBX_PGL_LIMIT && MDBX_PNL_ALLOCLEN(*ppl) >= MDBX_PNL_SIZE(*ppl)); if (likely(allocated >= wanna)) return MDBX_SUCCESS; if (unlikely(wanna > /* paranoia */ MDBX_PGL_LIMIT)) { mdbx_error("PNL too long (%zu > %zu)", wanna, (size_t)MDBX_PGL_LIMIT); return MDBX_TXN_FULL; } const size_t size = (wanna + wanna - allocated < MDBX_PGL_LIMIT) ? wanna + wanna - allocated : MDBX_PGL_LIMIT; size_t bytes = pnl2bytes(size); MDBX_PNL pl = mdbx_realloc(*ppl - 1, bytes); if (likely(pl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(pl); #endif /* malloc_usable_size */ *pl = bytes2pnl(bytes); assert(*pl >= wanna); *ppl = pl + 1; return MDBX_SUCCESS; } return MDBX_ENOMEM; } /* Make room for num additional elements in an PNL */ static __always_inline int __must_check_result mdbx_pnl_need(MDBX_PNL *ppl, size_t num) { assert(MDBX_PNL_SIZE(*ppl) <= MDBX_PGL_LIMIT && MDBX_PNL_ALLOCLEN(*ppl) >= MDBX_PNL_SIZE(*ppl)); assert(num <= MDBX_PGL_LIMIT); const size_t wanna = MDBX_PNL_SIZE(*ppl) + num; return likely(MDBX_PNL_ALLOCLEN(*ppl) >= wanna) ? MDBX_SUCCESS : mdbx_pnl_reserve(ppl, wanna); } static __always_inline void mdbx_pnl_xappend(MDBX_PNL pl, pgno_t pgno) { assert(MDBX_PNL_SIZE(pl) < MDBX_PNL_ALLOCLEN(pl)); if (mdbx_audit_enabled()) { for (unsigned i = MDBX_PNL_SIZE(pl); i > 0; --i) assert(pgno != pl[i]); } MDBX_PNL_SIZE(pl) += 1; MDBX_PNL_LAST(pl) = pgno; } /* Append an pgno range onto an unsorted PNL */ __always_inline static int __must_check_result mdbx_pnl_append_range(bool spilled, MDBX_PNL *ppl, pgno_t pgno, unsigned n) { assert(n > 0); int rc = mdbx_pnl_need(ppl, n); if (unlikely(rc != MDBX_SUCCESS)) return rc; const MDBX_PNL pnl = *ppl; #if MDBX_PNL_ASCENDING unsigned w = MDBX_PNL_SIZE(pnl); do { pnl[++w] = pgno; pgno += spilled ? 2 : 1; } while (--n); MDBX_PNL_SIZE(pnl) = w; #else unsigned w = MDBX_PNL_SIZE(pnl) + n; MDBX_PNL_SIZE(pnl) = w; do { pnl[w--] = pgno; pgno += spilled ? 2 : 1; } while (--n); #endif return MDBX_SUCCESS; } /* Append an pgno range into the sorted PNL */ static __hot int __must_check_result mdbx_pnl_insert_range(MDBX_PNL *ppl, pgno_t pgno, unsigned n) { assert(n > 0); int rc = mdbx_pnl_need(ppl, n); if (unlikely(rc != MDBX_SUCCESS)) return rc; const MDBX_PNL pnl = *ppl; unsigned r = MDBX_PNL_SIZE(pnl), w = r + n; MDBX_PNL_SIZE(pnl) = w; while (r && MDBX_PNL_DISORDERED(pnl[r], pgno)) pnl[w--] = pnl[r--]; for (pgno_t fill = MDBX_PNL_ASCENDING ? pgno + n : pgno; w > r; --w) pnl[w] = MDBX_PNL_ASCENDING ? --fill : fill++; return MDBX_SUCCESS; } static bool mdbx_pnl_check(const MDBX_PNL pl, const pgno_t limit) { if (likely(MDBX_PNL_SIZE(pl))) { assert(MDBX_PNL_LEAST(pl) >= MIN_PAGENO); assert(MDBX_PNL_MOST(pl) < limit); assert(MDBX_PNL_SIZE(pl) <= MDBX_PGL_LIMIT); if (unlikely(MDBX_PNL_SIZE(pl) > MDBX_PGL_LIMIT)) return false; if (unlikely(MDBX_PNL_LEAST(pl) < MIN_PAGENO)) return false; if (unlikely(MDBX_PNL_MOST(pl) >= limit)) return false; if (mdbx_audit_enabled()) { for (const pgno_t *scan = &MDBX_PNL_LAST(pl); --scan > pl;) { assert(MDBX_PNL_ORDERED(scan[0], scan[1])); if (unlikely(!MDBX_PNL_ORDERED(scan[0], scan[1]))) return false; } } } return true; } static __always_inline bool mdbx_pnl_check4assert(const MDBX_PNL pl, const pgno_t limit) { if (unlikely(pl == nullptr)) return true; assert(MDBX_PNL_ALLOCLEN(pl) >= MDBX_PNL_SIZE(pl)); if (unlikely(MDBX_PNL_ALLOCLEN(pl) < MDBX_PNL_SIZE(pl))) return false; return mdbx_pnl_check(pl, limit); } /* Merge an PNL onto an PNL. The destination PNL must be big enough */ static void __hot mdbx_pnl_xmerge(MDBX_PNL dst, const MDBX_PNL src) { assert(mdbx_pnl_check4assert(dst, MAX_PAGENO + 1)); assert(mdbx_pnl_check(src, MAX_PAGENO + 1)); const size_t total = MDBX_PNL_SIZE(dst) + MDBX_PNL_SIZE(src); assert(MDBX_PNL_ALLOCLEN(dst) >= total); pgno_t *w = dst + total; pgno_t *d = dst + MDBX_PNL_SIZE(dst); const pgno_t *s = src + MDBX_PNL_SIZE(src); dst[0] = /* detent for scan below */ (MDBX_PNL_ASCENDING ? 0 : ~(pgno_t)0); while (s > src) { while (MDBX_PNL_ORDERED(*s, *d)) *w-- = *d--; *w-- = *s--; } MDBX_PNL_SIZE(dst) = (pgno_t)total; assert(mdbx_pnl_check4assert(dst, MAX_PAGENO + 1)); } static void mdbx_spill_remove(MDBX_txn *txn, unsigned idx, unsigned npages) { mdbx_tassert(txn, idx > 0 && idx <= MDBX_PNL_SIZE(txn->tw.spill_pages) && txn->tw.spill_least_removed > 0); txn->tw.spill_least_removed = (idx < txn->tw.spill_least_removed) ? idx : txn->tw.spill_least_removed; txn->tw.spill_pages[idx] |= 1; MDBX_PNL_SIZE(txn->tw.spill_pages) -= (idx == MDBX_PNL_SIZE(txn->tw.spill_pages)); while (unlikely(npages > 1)) { const pgno_t pgno = (txn->tw.spill_pages[idx] >> 1) + 1; if (MDBX_PNL_ASCENDING) { if (++idx > MDBX_PNL_SIZE(txn->tw.spill_pages) || (txn->tw.spill_pages[idx] >> 1) != pgno) return; } else { if (--idx < 1 || (txn->tw.spill_pages[idx] >> 1) != pgno) return; txn->tw.spill_least_removed = (idx < txn->tw.spill_least_removed) ? idx : txn->tw.spill_least_removed; } txn->tw.spill_pages[idx] |= 1; MDBX_PNL_SIZE(txn->tw.spill_pages) -= (idx == MDBX_PNL_SIZE(txn->tw.spill_pages)); --npages; } } static MDBX_PNL mdbx_spill_purge(MDBX_txn *txn) { mdbx_tassert(txn, txn->tw.spill_least_removed > 0); const MDBX_PNL sl = txn->tw.spill_pages; if (txn->tw.spill_least_removed != INT_MAX) { unsigned len = MDBX_PNL_SIZE(sl), r, w; for (w = r = txn->tw.spill_least_removed; r <= len; ++r) { sl[w] = sl[r]; w += 1 - (sl[r] & 1); } for (size_t i = 1; i < w; ++i) mdbx_tassert(txn, (sl[i] & 1) == 0); MDBX_PNL_SIZE(sl) = w - 1; txn->tw.spill_least_removed = INT_MAX; } else { for (size_t i = 1; i <= MDBX_PNL_SIZE(sl); ++i) mdbx_tassert(txn, (sl[i] & 1) == 0); } return sl; } SORT_IMPL(pgno_sort, false, pgno_t, MDBX_PNL_ORDERED) static __hot void mdbx_pnl_sort(MDBX_PNL pnl) { pgno_sort(MDBX_PNL_BEGIN(pnl), MDBX_PNL_END(pnl)); assert(mdbx_pnl_check(pnl, MAX_PAGENO + 1)); } /* Search for an pgno in an PNL. * Returns The index of the first item greater than or equal to pgno. */ SEARCH_IMPL(pgno_bsearch, pgno_t, pgno_t, MDBX_PNL_ORDERED) static __hot unsigned mdbx_pnl_search(const MDBX_PNL pnl, pgno_t pgno) { assert(mdbx_pnl_check4assert(pnl, MAX_PAGENO + 1)); const pgno_t *begin = MDBX_PNL_BEGIN(pnl); const pgno_t *it = pgno_bsearch(begin, MDBX_PNL_SIZE(pnl), pgno); const pgno_t *end = begin + MDBX_PNL_SIZE(pnl); assert(it >= begin && it <= end); if (it != begin) assert(MDBX_PNL_ORDERED(it[-1], pgno)); if (it != end) assert(!MDBX_PNL_ORDERED(it[0], pgno)); return (unsigned)(it - begin + 1); } static __inline unsigned mdbx_pnl_exist(const MDBX_PNL pnl, pgno_t pgno) { unsigned n = mdbx_pnl_search(pnl, pgno); return (n <= MDBX_PNL_SIZE(pnl) && pnl[n] == pgno) ? n : 0; } static __inline unsigned mdbx_pnl_intersect(const MDBX_PNL pnl, pgno_t pgno, unsigned npages) { const unsigned len = MDBX_PNL_SIZE(pnl); if (mdbx_log_enabled(MDBX_LOG_EXTRA)) { mdbx_debug_extra("PNL len %u [", len); for (unsigned i = 1; i <= len; ++i) mdbx_debug_extra_print(" %" PRIaPGNO, pnl[i]); mdbx_debug_extra_print("%s\n", "]"); } const pgno_t range_last = pgno + npages - 1; #if MDBX_PNL_ASCENDING const unsigned n = mdbx_pnl_search(pnl, pgno); assert(n && (n == MDBX_PNL_SIZE(pnl) + 1 || pgno <= pnl[n])); const bool rc = n <= MDBX_PNL_SIZE(pnl) && pnl[n] <= range_last; #else const unsigned n = mdbx_pnl_search(pnl, range_last); assert(n && (n == MDBX_PNL_SIZE(pnl) + 1 || range_last >= pnl[n])); const bool rc = n <= MDBX_PNL_SIZE(pnl) && pnl[n] >= pgno; #endif if (mdbx_assert_enabled()) { bool check = false; for (unsigned i = 0; i < npages; ++i) check |= mdbx_pnl_exist(pnl, pgno + i) != 0; assert(check == rc); } return rc; } /*----------------------------------------------------------------------------*/ static __always_inline size_t txl2bytes(const size_t size) { assert(size > 0 && size <= MDBX_TXL_MAX * 2); size_t bytes = ceil_powerof2(MDBX_ASSUME_MALLOC_OVERHEAD + sizeof(txnid_t) * (size + 2), MDBX_TXL_GRANULATE * sizeof(txnid_t)) - MDBX_ASSUME_MALLOC_OVERHEAD; return bytes; } static __always_inline size_t bytes2txl(const size_t bytes) { size_t size = bytes / sizeof(txnid_t); assert(size > 2 && size <= MDBX_TXL_MAX * 2); return size - 2; } static MDBX_TXL mdbx_txl_alloc(void) { size_t bytes = txl2bytes(MDBX_TXL_INITIAL); MDBX_TXL tl = mdbx_malloc(bytes); if (likely(tl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(tl); #endif /* malloc_usable_size */ tl[0] = bytes2txl(bytes); assert(tl[0] >= MDBX_TXL_INITIAL); tl[1] = 0; tl += 1; } return tl; } static void mdbx_txl_free(MDBX_TXL tl) { if (likely(tl)) mdbx_free(tl - 1); } static int mdbx_txl_reserve(MDBX_TXL *ptl, const size_t wanna) { const size_t allocated = (size_t)MDBX_PNL_ALLOCLEN(*ptl); assert(MDBX_PNL_SIZE(*ptl) <= MDBX_TXL_MAX && MDBX_PNL_ALLOCLEN(*ptl) >= MDBX_PNL_SIZE(*ptl)); if (likely(allocated >= wanna)) return MDBX_SUCCESS; if (unlikely(wanna > /* paranoia */ MDBX_TXL_MAX)) { mdbx_error("TXL too long (%zu > %zu)", wanna, (size_t)MDBX_TXL_MAX); return MDBX_TXN_FULL; } const size_t size = (wanna + wanna - allocated < MDBX_TXL_MAX) ? wanna + wanna - allocated : MDBX_TXL_MAX; size_t bytes = txl2bytes(size); MDBX_TXL tl = mdbx_realloc(*ptl - 1, bytes); if (likely(tl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(tl); #endif /* malloc_usable_size */ *tl = bytes2txl(bytes); assert(*tl >= wanna); *ptl = tl + 1; return MDBX_SUCCESS; } return MDBX_ENOMEM; } static __always_inline int __must_check_result mdbx_txl_need(MDBX_TXL *ptl, size_t num) { assert(MDBX_PNL_SIZE(*ptl) <= MDBX_TXL_MAX && MDBX_PNL_ALLOCLEN(*ptl) >= MDBX_PNL_SIZE(*ptl)); assert(num <= MDBX_PGL_LIMIT); const size_t wanna = (size_t)MDBX_PNL_SIZE(*ptl) + num; return likely(MDBX_PNL_ALLOCLEN(*ptl) >= wanna) ? MDBX_SUCCESS : mdbx_txl_reserve(ptl, wanna); } static __always_inline void mdbx_txl_xappend(MDBX_TXL tl, txnid_t id) { assert(MDBX_PNL_SIZE(tl) < MDBX_PNL_ALLOCLEN(tl)); MDBX_PNL_SIZE(tl) += 1; MDBX_PNL_LAST(tl) = id; } #define TXNID_SORT_CMP(first, last) ((first) > (last)) SORT_IMPL(txnid_sort, false, txnid_t, TXNID_SORT_CMP) static void mdbx_txl_sort(MDBX_TXL tl) { txnid_sort(MDBX_PNL_BEGIN(tl), MDBX_PNL_END(tl)); } static int __must_check_result mdbx_txl_append(MDBX_TXL *ptl, txnid_t id) { if (unlikely(MDBX_PNL_SIZE(*ptl) == MDBX_PNL_ALLOCLEN(*ptl))) { int rc = mdbx_txl_need(ptl, MDBX_TXL_GRANULATE); if (unlikely(rc != MDBX_SUCCESS)) return rc; } mdbx_txl_xappend(*ptl, id); return MDBX_SUCCESS; } /*----------------------------------------------------------------------------*/ static __always_inline size_t dpl2bytes(const ptrdiff_t size) { assert(size > 2 && (size_t)size <= MDBX_PGL_LIMIT); size_t bytes = ceil_powerof2(MDBX_ASSUME_MALLOC_OVERHEAD + sizeof(MDBX_dpl) + (size + 2) * sizeof(MDBX_dp), MDBX_PNL_GRANULATE * sizeof(void *) * 2) - MDBX_ASSUME_MALLOC_OVERHEAD; return bytes; } static __always_inline unsigned bytes2dpl(const ptrdiff_t bytes) { size_t size = (bytes - sizeof(MDBX_dpl)) / sizeof(MDBX_dp); assert(size > 4 && size <= MDBX_PGL_LIMIT + MDBX_PNL_GRANULATE); return (unsigned)size - 2; } static __always_inline void mdbx_dpl_clear(MDBX_dpl *dl) { dl->sorted = dl->length = 0; } static void mdbx_dpl_free(MDBX_txn *txn) { if (likely(txn->tw.dirtylist)) { mdbx_free(txn->tw.dirtylist); txn->tw.dirtylist = NULL; } } static MDBX_dpl *mdbx_dpl_reserve(MDBX_txn *txn, size_t size) { mdbx_tassert(txn, txn->tw.dirtylist == NULL || txn->tw.dirtylist->length <= size); size_t bytes = dpl2bytes((size < MDBX_PGL_LIMIT) ? size : MDBX_PGL_LIMIT); MDBX_dpl *const dl = mdbx_realloc(txn->tw.dirtylist, bytes); if (likely(dl)) { #if __GLIBC_PREREQ(2, 12) || defined(__FreeBSD__) || defined(malloc_usable_size) bytes = malloc_usable_size(dl); #endif /* malloc_usable_size */ dl->allocated = bytes2dpl(bytes); dl->items[0].pgno = 0; dl->items[0].ptr = nullptr; mdbx_tassert(txn, txn->tw.dirtylist == NULL || dl->length <= dl->allocated); txn->tw.dirtylist = dl; } return dl; } static int mdbx_dpl_alloc(MDBX_txn *txn) { mdbx_tassert(txn, (txn->mt_flags & MDBX_TXN_RDONLY) == 0 && !txn->tw.dirtylist); MDBX_dpl *const dl = mdbx_dpl_reserve(txn, txn->mt_env->me_options.dp_initial); if (unlikely(!dl)) return MDBX_ENOMEM; mdbx_dpl_clear(dl); return MDBX_SUCCESS; } #define DP_SORT_CMP(first, last) ((first).pgno < (last).pgno) SORT_IMPL(dp_sort, false, MDBX_dp, DP_SORT_CMP) static __always_inline MDBX_dpl *mdbx_dpl_sort(MDBX_dpl *dl) { assert(dl->length <= MDBX_PGL_LIMIT); assert(dl->sorted <= dl->length); if (dl->sorted != dl->length) { dl->sorted = dl->length; dp_sort(dl->items + 1, dl->items + dl->length + 1); } return dl; } /* Returns the index of the first dirty-page whose pgno * member is greater than or equal to id. */ #define DP_SEARCH_CMP(dp, id) ((dp).pgno < (id)) SEARCH_IMPL(dp_bsearch, MDBX_dp, pgno_t, DP_SEARCH_CMP) static unsigned __hot mdbx_dpl_search(MDBX_dpl *dl, pgno_t pgno) { if (mdbx_audit_enabled()) { for (const MDBX_dp *ptr = dl->items + dl->sorted; --ptr > dl->items;) { assert(ptr[0].pgno < ptr[1].pgno); assert(ptr[0].pgno >= NUM_METAS); } } switch (dl->length - dl->sorted) { default: /* sort a whole */ dl->sorted = dl->length; dp_sort(dl->items + 1, dl->items + dl->length + 1); break; case 0: /* whole sorted cases */ break; #define LINEAR_SEARCH_CASE(N) \ case N: \ if (dl->items[dl->length - N + 1].pgno == pgno) \ return dl->length - N + 1; \ __fallthrough /* try linear search until the threshold */ LINEAR_SEARCH_CASE(16); /* fall through */ LINEAR_SEARCH_CASE(15); /* fall through */ LINEAR_SEARCH_CASE(14); /* fall through */ LINEAR_SEARCH_CASE(13); /* fall through */ LINEAR_SEARCH_CASE(12); /* fall through */ LINEAR_SEARCH_CASE(11); /* fall through */ LINEAR_SEARCH_CASE(10); /* fall through */ LINEAR_SEARCH_CASE(9); /* fall through */ LINEAR_SEARCH_CASE(8); /* fall through */ LINEAR_SEARCH_CASE(7); /* fall through */ LINEAR_SEARCH_CASE(6); /* fall through */ LINEAR_SEARCH_CASE(5); /* fall through */ LINEAR_SEARCH_CASE(4); /* fall through */ LINEAR_SEARCH_CASE(3); /* fall through */ LINEAR_SEARCH_CASE(2); /* fall through */ case 1: if (dl->items[dl->length].pgno == pgno) return dl->length; /* continue bsearch on the sorted part */ break; } return (unsigned)(dp_bsearch(dl->items + 1, dl->sorted, pgno) - dl->items); } static __inline bool mdbx_dpl_intersect(MDBX_dpl *dl, pgno_t pgno, unsigned npages) { assert(dl->sorted == dl->length); unsigned const n = mdbx_dpl_search(dl, pgno); assert(n >= 1 && n <= dl->length + 1); assert(n > dl->length || pgno <= dl->items[n].pgno); assert(n == 1 || pgno > dl->items[n - 1].pgno); const MDBX_page *const prev = dl->items[n - 1].ptr; const bool rc = (/* intersection with founded */ n <= dl->length && pgno + npages > dl->items[n].pgno) || (/* intersection with prev */ n > 1 && unlikely(IS_OVERFLOW(prev)) && prev->mp_pgno + prev->mp_pages > pgno); if (mdbx_assert_enabled()) { bool check = false; for (unsigned i = 1; i <= dl->length; ++i) { const MDBX_page *const dp = dl->items[i].ptr; if (!(dp->mp_pgno /* begin */ >= /* end */ pgno + npages || dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1) /* end */ <= /* begin */ pgno)) check |= true; } assert(check == rc); } return rc; } static __always_inline unsigned mdbx_dpl_exist(MDBX_dpl *dl, pgno_t pgno) { unsigned i = mdbx_dpl_search(dl, pgno); assert((int)i > 0); return (i <= dl->length && dl->items[i].pgno == pgno) ? i : 0; } static __always_inline MDBX_page *mdbx_dpl_find(MDBX_dpl *dl, pgno_t pgno) { const unsigned i = mdbx_dpl_search(dl, pgno); assert((int)i > 0); return (i <= dl->length && dl->items[i].pgno == pgno) ? dl->items[i].ptr : nullptr; } static __maybe_unused const MDBX_page *debug_dpl_find(const MDBX_dpl *dl, const pgno_t pgno) { for (unsigned i = dl->length; i > dl->sorted; --i) if (dl->items[i].pgno == pgno) return dl->items[i].ptr; if (dl->sorted) { const unsigned i = (unsigned)(dp_bsearch(dl->items + 1, dl->sorted, pgno) - dl->items); if (i <= dl->sorted && dl->items[i].pgno == pgno) return dl->items[i].ptr; } return nullptr; } static void mdbx_dpl_remove(MDBX_dpl *dl, unsigned i) { assert((int)i > 0 && i <= dl->length); dl->sorted -= dl->sorted >= i; dl->length -= 1; if (dl->length >= i) memmove(dl->items + i, dl->items + i + 1, (dl->length - i + 1) * sizeof(dl->items[0])); } static __always_inline int __must_check_result mdbx_dpl_append(MDBX_txn *txn, pgno_t pgno, MDBX_page *page) { MDBX_dpl *dl = txn->tw.dirtylist; assert(dl->length <= MDBX_PGL_LIMIT + MDBX_PNL_GRANULATE); if (mdbx_audit_enabled()) { for (unsigned i = dl->length; i > 0; --i) { assert(dl->items[i].pgno != pgno); if (unlikely(dl->items[i].pgno == pgno)) return MDBX_PROBLEM; } } if (unlikely(dl->length == dl->allocated)) { if (unlikely(dl->allocated >= MDBX_PGL_LIMIT)) { mdbx_error("DPL is full (MDBX_PGL_LIMIT %u)", MDBX_PGL_LIMIT); return MDBX_TXN_FULL; } const size_t size = (dl->allocated < MDBX_PNL_INITIAL * 42) ? dl->allocated + dl->allocated : dl->allocated + dl->allocated / 2; dl = mdbx_dpl_reserve(txn, size); if (unlikely(!dl)) return MDBX_ENOMEM; mdbx_tassert(txn, dl->length < dl->allocated); } /* append page */ const unsigned n = dl->length + 1; if (n == 1 || (dl->sorted >= dl->length && dl->items[n - 1].pgno < pgno)) dl->sorted = n; dl->length = n; dl->items[n].pgno = pgno; dl->items[n].ptr = page; return MDBX_SUCCESS; } /*----------------------------------------------------------------------------*/ uint8_t mdbx_runtime_flags = MDBX_RUNTIME_FLAGS_INIT; uint8_t mdbx_loglevel = MDBX_LOG_FATAL; MDBX_debug_func *mdbx_debug_logger; static __must_check_result int mdbx_page_retire(MDBX_cursor *mc, MDBX_page *mp); static __must_check_result int mdbx_page_loose(MDBX_txn *txn, MDBX_page *mp); static int mdbx_page_alloc(MDBX_cursor *mc, const unsigned num, MDBX_page **const mp, int flags); static txnid_t mdbx_kick_longlived_readers(MDBX_env *env, const txnid_t laggard); static int mdbx_page_new(MDBX_cursor *mc, uint32_t flags, unsigned num, MDBX_page **mp); static int mdbx_page_touch(MDBX_cursor *mc); static int mdbx_cursor_touch(MDBX_cursor *mc); #define MDBX_END_NAMES \ { \ "committed", "empty-commit", "abort", "reset", "reset-tmp", "fail-begin", \ "fail-beginchild" \ } enum { /* mdbx_txn_end operation number, for logging */ MDBX_END_COMMITTED, MDBX_END_PURE_COMMIT, MDBX_END_ABORT, MDBX_END_RESET, MDBX_END_RESET_TMP, MDBX_END_FAIL_BEGIN, MDBX_END_FAIL_BEGINCHILD }; #define MDBX_END_OPMASK 0x0F /* mask for mdbx_txn_end() operation number */ #define MDBX_END_UPDATE 0x10 /* update env state (DBIs) */ #define MDBX_END_FREE 0x20 /* free txn unless it is MDBX_env.me_txn0 */ #define MDBX_END_EOTDONE 0x40 /* txn's cursors already closed */ #define MDBX_END_SLOT 0x80 /* release any reader slot if MDBX_NOTLS */ static int mdbx_txn_end(MDBX_txn *txn, const unsigned mode); static int __must_check_result mdbx_page_get(MDBX_cursor *mc, pgno_t pgno, MDBX_page **mp, int *lvl, txnid_t pp_txnid); static int __must_check_result mdbx_page_search_root(MDBX_cursor *mc, const MDBX_val *key, int flags); #define MDBX_PS_MODIFY 1 #define MDBX_PS_ROOTONLY 2 #define MDBX_PS_FIRST 4 #define MDBX_PS_LAST 8 static int __must_check_result mdbx_page_search(MDBX_cursor *mc, const MDBX_val *key, int flags); static int __must_check_result mdbx_page_merge(MDBX_cursor *csrc, MDBX_cursor *cdst); static int __must_check_result mdbx_page_flush(MDBX_txn *txn, const unsigned keep); #define MDBX_SPLIT_REPLACE MDBX_APPENDDUP /* newkey is not new */ static int __must_check_result mdbx_page_split(MDBX_cursor *mc, const MDBX_val *newkey, MDBX_val *newdata, pgno_t newpgno, unsigned nflags); static int __must_check_result mdbx_read_header(MDBX_env *env, MDBX_meta *meta, uint64_t *filesize, const int lck_exclusive); static int __must_check_result mdbx_sync_locked(MDBX_env *env, unsigned flags, MDBX_meta *const pending); static int mdbx_env_close0(MDBX_env *env); static MDBX_node *mdbx_node_search(MDBX_cursor *mc, const MDBX_val *key, int *exactp); static int __must_check_result mdbx_node_add_branch(MDBX_cursor *mc, unsigned indx, const MDBX_val *key, pgno_t pgno); static int __must_check_result mdbx_node_add_leaf(MDBX_cursor *mc, unsigned indx, const MDBX_val *key, MDBX_val *data, unsigned flags); static int __must_check_result mdbx_node_add_leaf2(MDBX_cursor *mc, unsigned indx, const MDBX_val *key); static void mdbx_node_del(MDBX_cursor *mc, size_t ksize); static void mdbx_node_shrink(MDBX_page *mp, unsigned indx); static int __must_check_result mdbx_node_move(MDBX_cursor *csrc, MDBX_cursor *cdst, int fromleft); static int __must_check_result mdbx_node_read(MDBX_cursor *mc, MDBX_node *leaf, MDBX_val *data, const txnid_t pp_txnid); static int __must_check_result mdbx_rebalance(MDBX_cursor *mc); static int __must_check_result mdbx_update_key(MDBX_cursor *mc, const MDBX_val *key); static void mdbx_cursor_pop(MDBX_cursor *mc); static int __must_check_result mdbx_cursor_push(MDBX_cursor *mc, MDBX_page *mp); static int __must_check_result mdbx_audit_ex(MDBX_txn *txn, unsigned retired_stored, bool dont_filter_gc); static int __must_check_result mdbx_page_check(MDBX_cursor *const mc, const MDBX_page *const mp, unsigned options); static int __must_check_result mdbx_cursor_check(MDBX_cursor *mc, unsigned options); static int __must_check_result mdbx_cursor_del0(MDBX_cursor *mc); static int __must_check_result mdbx_del0(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, const MDBX_val *data, unsigned flags); #define SIBLING_LEFT 0 #define SIBLING_RIGHT 2 static int __must_check_result mdbx_cursor_sibling(MDBX_cursor *mc, int dir); static int __must_check_result mdbx_cursor_next(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op); static int __must_check_result mdbx_cursor_prev(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op); static int __must_check_result mdbx_cursor_set(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op, int *exactp); static int __must_check_result mdbx_cursor_first(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data); static int __must_check_result mdbx_cursor_last(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data); static int __must_check_result mdbx_cursor_init(MDBX_cursor *mc, MDBX_txn *txn, MDBX_dbi dbi); static int __must_check_result mdbx_xcursor_init0(MDBX_cursor *mc); static int __must_check_result mdbx_xcursor_init1(MDBX_cursor *mc, MDBX_node *node, const MDBX_page *mp); static int __must_check_result mdbx_xcursor_init2(MDBX_cursor *mc, MDBX_xcursor *src_mx, bool new_dupdata); static void cursor_copy_internal(const MDBX_cursor *csrc, MDBX_cursor *cdst); static int __must_check_result mdbx_drop0(MDBX_cursor *mc, int subs); static int __must_check_result mdbx_fetch_sdb(MDBX_txn *txn, MDBX_dbi dbi); static int __must_check_result mdbx_setup_dbx(MDBX_dbx *const dbx, const MDBX_db *const db, const unsigned pagesize); static MDBX_cmp_func cmp_lexical, cmp_reverse, cmp_int_align4, cmp_int_align2, cmp_int_unaligned, cmp_lenfast; static __inline MDBX_cmp_func *get_default_keycmp(unsigned flags); static __inline MDBX_cmp_func *get_default_datacmp(unsigned flags); __cold const char *mdbx_liberr2str(int errnum) { /* Table of descriptions for MDBX errors */ static const char *const tbl[] = { "MDBX_KEYEXIST: Key/data pair already exists", "MDBX_NOTFOUND: No matching key/data pair found", "MDBX_PAGE_NOTFOUND: Requested page not found", "MDBX_CORRUPTED: Database is corrupted", "MDBX_PANIC: Environment had fatal error", "MDBX_VERSION_MISMATCH: DB version mismatch libmdbx", "MDBX_INVALID: File is not an MDBX file", "MDBX_MAP_FULL: Environment mapsize limit reached", "MDBX_DBS_FULL: Too many DBI-handles (maxdbs reached)", "MDBX_READERS_FULL: Too many readers (maxreaders reached)", NULL /* MDBX_TLS_FULL (-30789): unused in MDBX */, "MDBX_TXN_FULL: Transaction has too many dirty pages," " i.e transaction is too big", "MDBX_CURSOR_FULL: Cursor stack limit reachedn - this usually indicates" " corruption, i.e branch-pages loop", "MDBX_PAGE_FULL: Internal error - Page has no more space", "MDBX_UNABLE_EXTEND_MAPSIZE: Database engine was unable to extend" " mapping, e.g. since address space is unavailable or busy," " or Operation system not supported such operations", "MDBX_INCOMPATIBLE: Environment or database is not compatible" " with the requested operation or the specified flags", "MDBX_BAD_RSLOT: Invalid reuse of reader locktable slot," " e.g. read-transaction already run for current thread", "MDBX_BAD_TXN: Transaction is not valid for requested operation," " e.g. had errored and be must aborted, has a child, or is invalid", "MDBX_BAD_VALSIZE: Invalid size or alignment of key or data" " for target database, either invalid subDB name", "MDBX_BAD_DBI: The specified DBI-handle is invalid" " or changed by another thread/transaction", "MDBX_PROBLEM: Unexpected internal error, transaction should be aborted", "MDBX_BUSY: Another write transaction is running," " or environment is already used while opening with MDBX_EXCLUSIVE flag", }; if (errnum >= MDBX_KEYEXIST && errnum <= MDBX_BUSY) { int i = errnum - MDBX_KEYEXIST; return tbl[i]; } switch (errnum) { case MDBX_SUCCESS: return "MDBX_SUCCESS: Successful"; case MDBX_EMULTIVAL: return "MDBX_EMULTIVAL: The specified key has" " more than one associated value"; case MDBX_EBADSIGN: return "MDBX_EBADSIGN: Wrong signature of a runtime object(s)," " e.g. memory corruption or double-free"; case MDBX_WANNA_RECOVERY: return "MDBX_WANNA_RECOVERY: Database should be recovered," " but this could NOT be done automatically for now" " since it opened in read-only mode"; case MDBX_EKEYMISMATCH: return "MDBX_EKEYMISMATCH: The given key value is mismatched to the" " current cursor position"; case MDBX_TOO_LARGE: return "MDBX_TOO_LARGE: Database is too large for current system," " e.g. could NOT be mapped into RAM"; case MDBX_THREAD_MISMATCH: return "MDBX_THREAD_MISMATCH: A thread has attempted to use a not" " owned object, e.g. a transaction that started by another thread"; case MDBX_TXN_OVERLAPPING: return "MDBX_TXN_OVERLAPPING: Overlapping read and write transactions for" " the current thread"; default: return NULL; } } const char *__cold mdbx_strerror_r(int errnum, char *buf, size_t buflen) { const char *msg = mdbx_liberr2str(errnum); if (!msg && buflen > 0 && buflen < INT_MAX) { #if defined(_WIN32) || defined(_WIN64) const DWORD size = FormatMessageA( FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, NULL, errnum, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), buf, (DWORD)buflen, NULL); return size ? buf : "FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM) failed"; #elif defined(_GNU_SOURCE) && defined(__GLIBC__) /* GNU-specific */ if (errnum > 0) msg = strerror_r(errnum, buf, buflen); #elif (_POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600) /* XSI-compliant */ if (errnum > 0 && strerror_r(errnum, buf, buflen) == 0) msg = buf; #else if (errnum > 0) { msg = strerror(errnum); if (msg) { strncpy(buf, msg, buflen); msg = buf; } } #endif if (!msg) { (void)snprintf(buf, buflen, "error %d", errnum); msg = buf; } buf[buflen - 1] = '\0'; } return msg; } const char *__cold mdbx_strerror(int errnum) { #if defined(_WIN32) || defined(_WIN64) static char buf[1024]; return mdbx_strerror_r(errnum, buf, sizeof(buf)); #else const char *msg = mdbx_liberr2str(errnum); if (!msg) { if (errnum > 0) msg = strerror(errnum); if (!msg) { static char buf[32]; (void)snprintf(buf, sizeof(buf) - 1, "error %d", errnum); msg = buf; } } return msg; #endif } #if defined(_WIN32) || defined(_WIN64) /* Bit of madness for Windows */ const char *mdbx_strerror_r_ANSI2OEM(int errnum, char *buf, size_t buflen) { const char *msg = mdbx_liberr2str(errnum); if (!msg && buflen > 0 && buflen < INT_MAX) { const DWORD size = FormatMessageA( FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, NULL, errnum, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), buf, (DWORD)buflen, NULL); if (!size) msg = "FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM) failed"; else if (!CharToOemBuffA(buf, buf, size)) msg = "CharToOemBuffA() failed"; else msg = buf; } return msg; } const char *mdbx_strerror_ANSI2OEM(int errnum) { static char buf[1024]; return mdbx_strerror_r_ANSI2OEM(errnum, buf, sizeof(buf)); } #endif /* Bit of madness for Windows */ void __cold mdbx_debug_log_va(int level, const char *function, int line, const char *fmt, va_list args) { if (mdbx_debug_logger) mdbx_debug_logger(level, function, line, fmt, args); else { #if defined(_WIN32) || defined(_WIN64) if (IsDebuggerPresent()) { int prefix_len = 0; char *prefix = nullptr; if (function && line > 0) prefix_len = mdbx_asprintf(&prefix, "%s:%d ", function, line); else if (function) prefix_len = mdbx_asprintf(&prefix, "%s: ", function); else if (line > 0) prefix_len = mdbx_asprintf(&prefix, "%d: ", line); if (prefix_len > 0 && prefix) { OutputDebugStringA(prefix); mdbx_free(prefix); } char *msg = nullptr; int msg_len = mdbx_vasprintf(&msg, fmt, args); if (msg_len > 0 && msg) { OutputDebugStringA(msg); mdbx_free(msg); } } #else if (function && line > 0) fprintf(stderr, "%s:%d ", function, line); else if (function) fprintf(stderr, "%s: ", function); else if (line > 0) fprintf(stderr, "%d: ", line); vfprintf(stderr, fmt, args); fflush(stderr); #endif } } void __cold mdbx_debug_log(int level, const char *function, int line, const char *fmt, ...) { va_list args; va_start(args, fmt); mdbx_debug_log_va(level, function, line, fmt, args); va_end(args); } /* Dump a key in ascii or hexadecimal. */ const char *mdbx_dump_val(const MDBX_val *key, char *const buf, const size_t bufsize) { if (!key) return ""; if (!key->iov_len) return ""; if (!buf || bufsize < 4) return nullptr; bool is_ascii = true; const uint8_t *const data = key->iov_base; for (unsigned i = 0; i < key->iov_len; i++) if (data[i] < ' ' || data[i] > '~') { is_ascii = false; break; } if (is_ascii) { int len = snprintf(buf, bufsize, "%.*s", (key->iov_len > INT_MAX) ? INT_MAX : (int)key->iov_len, data); assert(len > 0 && (unsigned)len < bufsize); (void)len; } else { char *const detent = buf + bufsize - 2; char *ptr = buf; *ptr++ = '<'; for (unsigned i = 0; i < key->iov_len; i++) { const ptrdiff_t left = detent - ptr; assert(left > 0); int len = snprintf(ptr, left, "%02x", data[i]); if (len < 0 || len >= left) break; ptr += len; } if (ptr < detent) { ptr[0] = '>'; ptr[1] = '\0'; } } return buf; } /*------------------------------------------------------------------------------ LY: debug stuff */ static const char *mdbx_leafnode_type(MDBX_node *n) { static const char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}}; return F_ISSET(node_flags(n), F_BIGDATA) ? ": overflow page" : tp[F_ISSET(node_flags(n), F_DUPDATA)] [F_ISSET(node_flags(n), F_SUBDATA)]; } /* Display all the keys in the page. */ static __maybe_unused void mdbx_page_list(MDBX_page *mp) { pgno_t pgno = mp->mp_pgno; const char *type, *state = IS_DIRTY(mp) ? ", dirty" : ""; MDBX_node *node; unsigned i, nkeys, nsize, total = 0; MDBX_val key; DKBUF; switch (mp->mp_flags & (P_BRANCH | P_LEAF | P_LEAF2 | P_META | P_OVERFLOW | P_SUBP)) { case P_BRANCH: type = "Branch page"; break; case P_LEAF: type = "Leaf page"; break; case P_LEAF | P_SUBP: type = "Leaf sub-page"; break; case P_LEAF | P_LEAF2: type = "Leaf2 page"; break; case P_LEAF | P_LEAF2 | P_SUBP: type = "Leaf2 sub-page"; break; case P_OVERFLOW: mdbx_verbose("Overflow page %" PRIaPGNO " pages %u%s\n", pgno, mp->mp_pages, state); return; case P_META: mdbx_verbose("Meta-page %" PRIaPGNO " txnid %" PRIu64 "\n", pgno, unaligned_peek_u64(4, page_meta(mp)->mm_txnid_a)); return; default: mdbx_verbose("Bad page %" PRIaPGNO " flags 0x%X\n", pgno, mp->mp_flags); return; } nkeys = page_numkeys(mp); mdbx_verbose("%s %" PRIaPGNO " numkeys %u%s\n", type, pgno, nkeys, state); for (i = 0; i < nkeys; i++) { if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */ key.iov_len = nsize = mp->mp_leaf2_ksize; key.iov_base = page_leaf2key(mp, i, nsize); total += nsize; mdbx_verbose("key %u: nsize %u, %s\n", i, nsize, DKEY(&key)); continue; } node = page_node(mp, i); key.iov_len = node_ks(node); key.iov_base = node->mn_data; nsize = (unsigned)(NODESIZE + key.iov_len); if (IS_BRANCH(mp)) { mdbx_verbose("key %u: page %" PRIaPGNO ", %s\n", i, node_pgno(node), DKEY(&key)); total += nsize; } else { if (F_ISSET(node_flags(node), F_BIGDATA)) nsize += sizeof(pgno_t); else nsize += (unsigned)node_ds(node); total += nsize; nsize += sizeof(indx_t); mdbx_verbose("key %u: nsize %u, %s%s\n", i, nsize, DKEY(&key), mdbx_leafnode_type(node)); } total = EVEN(total); } mdbx_verbose("Total: header %u + contents %u + unused %u\n", IS_LEAF2(mp) ? PAGEHDRSZ : PAGEHDRSZ + mp->mp_lower, total, page_room(mp)); } /*----------------------------------------------------------------------------*/ /* Check if there is an initialized xcursor, so XCURSOR_REFRESH() is proper */ #define XCURSOR_INITED(mc) \ ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) /* Update sub-page pointer, if any, in mc->mc_xcursor. * Needed when the node which contains the sub-page may have moved. * Called with mp = mc->mc_pg[mc->mc_top], ki = mc->mc_ki[mc->mc_top]. */ #define XCURSOR_REFRESH(mc, mp, ki) \ do { \ MDBX_page *xr_pg = (mp); \ MDBX_node *xr_node = page_node(xr_pg, ki); \ if ((node_flags(xr_node) & (F_DUPDATA | F_SUBDATA)) == F_DUPDATA) \ (mc)->mc_xcursor->mx_cursor.mc_pg[0] = node_data(xr_node); \ } while (0) static __maybe_unused bool cursor_is_tracked(const MDBX_cursor *mc) { for (MDBX_cursor *scan = mc->mc_txn->tw.cursors[mc->mc_dbi]; scan; scan = scan->mc_next) if (mc == ((mc->mc_flags & C_SUB) ? &scan->mc_xcursor->mx_cursor : scan)) return true; return false; } /* Perform act while tracking temporary cursor mn */ #define WITH_CURSOR_TRACKING(mn, act) \ do { \ mdbx_cassert(&(mn), \ mn.mc_txn->tw.cursors != NULL /* must be not rdonly txt */); \ mdbx_cassert(&(mn), !cursor_is_tracked(&(mn))); \ MDBX_cursor mc_dummy; \ MDBX_cursor **tracking_head = &(mn).mc_txn->tw.cursors[mn.mc_dbi]; \ MDBX_cursor *tracked = &(mn); \ if ((mn).mc_flags & C_SUB) { \ mc_dummy.mc_flags = C_INITIALIZED; \ mc_dummy.mc_top = 0; \ mc_dummy.mc_xcursor = (MDBX_xcursor *)&(mn); \ tracked = &mc_dummy; \ } \ tracked->mc_next = *tracking_head; \ *tracking_head = tracked; \ { act; } \ *tracking_head = tracked->mc_next; \ } while (0) int mdbx_cmp(const MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *a, const MDBX_val *b) { mdbx_assert(NULL, txn->mt_signature == MDBX_MT_SIGNATURE); return txn->mt_dbxs[dbi].md_cmp(a, b); } int mdbx_dcmp(const MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *a, const MDBX_val *b) { mdbx_assert(NULL, txn->mt_signature == MDBX_MT_SIGNATURE); return txn->mt_dbxs[dbi].md_dcmp(a, b); } /* Allocate memory for a page. * Re-use old malloc'd pages first for singletons, otherwise just malloc. * Set MDBX_TXN_ERROR on failure. */ static MDBX_page *mdbx_page_malloc(MDBX_txn *txn, unsigned num) { MDBX_env *env = txn->mt_env; MDBX_page *np = env->me_dp_reserve; size_t size = env->me_psize; if (likely(num == 1 && np)) { mdbx_assert(env, env->me_dp_reserve_len > 0); ASAN_UNPOISON_MEMORY_REGION(np, size); VALGRIND_MEMPOOL_ALLOC(env, np, size); VALGRIND_MAKE_MEM_DEFINED(&np->mp_next, sizeof(np->mp_next)); env->me_dp_reserve = np->mp_next; env->me_dp_reserve_len -= 1; } else { size = pgno2bytes(env, num); np = mdbx_malloc(size); if (unlikely(!np)) { txn->mt_flags |= MDBX_TXN_ERROR; return np; } VALGRIND_MEMPOOL_ALLOC(env, np, size); } if ((env->me_flags & MDBX_NOMEMINIT) == 0) { /* For a single page alloc, we init everything after the page header. * For multi-page, we init the final page; if the caller needed that * many pages they will be filling in at least up to the last page. */ size_t skip = PAGEHDRSZ; if (num > 1) skip += pgno2bytes(env, num - 1); memset((char *)np + skip, 0, size - skip); } #if MDBX_DEBUG np->mp_pgno = 0; #endif VALGRIND_MAKE_MEM_UNDEFINED(np, size); np->mp_flags = 0; np->mp_pages = num; return np; } /* Free a shadow dirty page */ static void mdbx_dpage_free(MDBX_env *env, MDBX_page *dp, unsigned npages) { VALGRIND_MAKE_MEM_UNDEFINED(dp, pgno2bytes(env, npages)); ASAN_UNPOISON_MEMORY_REGION(dp, pgno2bytes(env, npages)); if (MDBX_DEBUG || unlikely(env->me_flags & MDBX_PAGEPERTURB)) memset(dp, -1, pgno2bytes(env, npages)); if (npages == 1 && env->me_dp_reserve_len < env->me_options.dp_reserve_limit) { ASAN_POISON_MEMORY_REGION((char *)dp + sizeof(dp->mp_next), pgno2bytes(env, npages) - sizeof(dp->mp_next)); dp->mp_next = env->me_dp_reserve; VALGRIND_MEMPOOL_FREE(env, dp); env->me_dp_reserve = dp; env->me_dp_reserve_len += 1; } else { /* large pages just get freed directly */ VALGRIND_MEMPOOL_FREE(env, dp); mdbx_free(dp); } } /* Return all dirty pages to dpage list */ static void mdbx_dlist_free(MDBX_txn *txn) { MDBX_env *env = txn->mt_env; MDBX_dpl *const dl = txn->tw.dirtylist; const size_t len = dl->length; for (size_t i = 1; i <= len; i++) { MDBX_page *dp = dl->items[i].ptr; mdbx_dpage_free(env, dp, IS_OVERFLOW(dp) ? dp->mp_pages : 1); } mdbx_dpl_clear(dl); } static __always_inline MDBX_db *mdbx_outer_db(MDBX_cursor *mc) { mdbx_cassert(mc, (mc->mc_flags & C_SUB) != 0); MDBX_xcursor *mx = container_of(mc->mc_db, MDBX_xcursor, mx_db); MDBX_cursor_couple *couple = container_of(mx, MDBX_cursor_couple, inner); mdbx_cassert(mc, mc->mc_db == &couple->outer.mc_xcursor->mx_db); mdbx_cassert(mc, mc->mc_dbx == &couple->outer.mc_xcursor->mx_dbx); return couple->outer.mc_db; } static __cold __maybe_unused bool mdbx_dirtylist_check(MDBX_txn *txn) { if (!mdbx_audit_enabled()) return true; const MDBX_dpl *const dl = txn->tw.dirtylist; unsigned loose = 0; for (unsigned i = dl->length; i > 0; --i) { const MDBX_page *const dp = dl->items[i].ptr; if (!dp) continue; mdbx_tassert(txn, dp->mp_pgno == dl->items[i].pgno); if (unlikely(dp->mp_pgno != dl->items[i].pgno)) return false; mdbx_tassert(txn, dp->mp_flags & P_DIRTY); if (unlikely((dp->mp_flags & P_DIRTY) == 0)) return false; if (dp->mp_flags & P_LOOSE) { mdbx_tassert(txn, dp->mp_flags == (P_LOOSE | P_DIRTY)); if (unlikely(dp->mp_flags != (P_LOOSE | P_DIRTY))) return false; loose += 1; } const unsigned num = IS_OVERFLOW(dp) ? dp->mp_pages : 1; mdbx_tassert(txn, txn->mt_next_pgno >= dp->mp_pgno + num); if (unlikely(txn->mt_next_pgno < dp->mp_pgno + num)) return false; if (i < dl->sorted) { mdbx_tassert(txn, dl->items[i + 1].pgno >= dp->mp_pgno + num); if (unlikely(dl->items[i + 1].pgno < dp->mp_pgno + num)) return false; } const unsigned rpa = mdbx_pnl_search(txn->tw.reclaimed_pglist, dp->mp_pgno); mdbx_tassert(txn, rpa > MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) || txn->tw.reclaimed_pglist[rpa] != dp->mp_pgno); if (rpa <= MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) && unlikely(txn->tw.reclaimed_pglist[rpa] == dp->mp_pgno)) return false; if (num > 1) { const unsigned rpb = mdbx_pnl_search(txn->tw.reclaimed_pglist, dp->mp_pgno + num - 1); mdbx_tassert(txn, rpa == rpb); if (unlikely(rpa != rpb)) return false; } } mdbx_tassert(txn, loose == txn->tw.loose_count); if (unlikely(loose != txn->tw.loose_count)) return false; for (unsigned i = 1; i <= MDBX_PNL_SIZE(txn->tw.retired_pages); ++i) { const MDBX_page *const dp = debug_dpl_find(dl, txn->tw.retired_pages[i]); mdbx_tassert(txn, !dp); if (unlikely(dp)) return false; } return true; } #if MDBX_ENABLE_REFUND static void mdbx_refund_reclaimed(MDBX_txn *txn) { /* Scanning in descend order */ pgno_t next_pgno = txn->mt_next_pgno; const MDBX_PNL pnl = txn->tw.reclaimed_pglist; mdbx_tassert(txn, MDBX_PNL_SIZE(pnl) && MDBX_PNL_MOST(pnl) == next_pgno - 1); #if MDBX_PNL_ASCENDING unsigned i = MDBX_PNL_SIZE(pnl); mdbx_tassert(txn, pnl[i] == next_pgno - 1); while (--next_pgno, --i > 0 && pnl[i] == next_pgno - 1) ; MDBX_PNL_SIZE(pnl) = i; #else unsigned i = 1; mdbx_tassert(txn, pnl[i] == next_pgno - 1); unsigned len = MDBX_PNL_SIZE(pnl); while (--next_pgno, ++i <= len && pnl[i] == next_pgno - 1) ; MDBX_PNL_SIZE(pnl) = len -= i - 1; for (unsigned move = 0; move < len; ++move) pnl[1 + move] = pnl[i + move]; #endif mdbx_verbose("refunded %" PRIaPGNO " pages: %" PRIaPGNO " -> %" PRIaPGNO, txn->mt_next_pgno - next_pgno, txn->mt_next_pgno, next_pgno); txn->mt_next_pgno = next_pgno; mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - 1)); } static void mdbx_refund_loose(MDBX_txn *txn) { mdbx_tassert(txn, txn->tw.loose_pages != nullptr); mdbx_tassert(txn, txn->tw.loose_count > 0); MDBX_dpl *const dl = txn->tw.dirtylist; mdbx_tassert(txn, dl->length >= txn->tw.loose_count); pgno_t onstack[MDBX_CACHELINE_SIZE * 8 / sizeof(pgno_t)]; MDBX_PNL suitable = onstack; if (dl->length - dl->sorted > txn->tw.loose_count) { /* Dirty list is useless since unsorted. */ if (bytes2pnl(sizeof(onstack)) < txn->tw.loose_count) { suitable = mdbx_pnl_alloc(txn->tw.loose_count); if (unlikely(!suitable)) return /* this is not a reason for transaction fail */; } /* Collect loose-pages which may be refunded. */ mdbx_tassert(txn, txn->mt_next_pgno >= MIN_PAGENO + txn->tw.loose_count); pgno_t most = MIN_PAGENO; unsigned w = 0; for (const MDBX_page *lp = txn->tw.loose_pages; lp; lp = lp->mp_next) { mdbx_tassert(txn, lp->mp_flags == (P_LOOSE | P_DIRTY)); mdbx_tassert(txn, txn->mt_next_pgno > lp->mp_pgno); if (likely(txn->mt_next_pgno - txn->tw.loose_count <= lp->mp_pgno)) { mdbx_tassert(txn, w < ((suitable == onstack) ? bytes2pnl(sizeof(onstack)) : MDBX_PNL_ALLOCLEN(suitable))); suitable[++w] = lp->mp_pgno; most = (lp->mp_pgno > most) ? lp->mp_pgno : most; } } if (most + 1 == txn->mt_next_pgno) { /* Sort suitable list and refund pages at the tail. */ MDBX_PNL_SIZE(suitable) = w; mdbx_pnl_sort(suitable); /* Scanning in descend order */ const int step = MDBX_PNL_ASCENDING ? -1 : 1; const int begin = MDBX_PNL_ASCENDING ? MDBX_PNL_SIZE(suitable) : 1; const int end = MDBX_PNL_ASCENDING ? 0 : MDBX_PNL_SIZE(suitable) + 1; mdbx_tassert(txn, suitable[begin] >= suitable[end - step]); mdbx_tassert(txn, most == suitable[begin]); for (int i = begin + step; i != end; i += step) { if (suitable[i] != most - 1) break; most -= 1; } const unsigned refunded = txn->mt_next_pgno - most; mdbx_debug("refund-suitable %u pages %" PRIaPGNO " -> %" PRIaPGNO, refunded, most, txn->mt_next_pgno); txn->tw.loose_count -= refunded; txn->tw.dirtyroom += refunded; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); txn->mt_next_pgno = most; /* Filter-out dirty list */ unsigned r = 0; w = 0; if (dl->sorted) { do { if (dl->items[++r].pgno < most) { if (++w != r) dl->items[w] = dl->items[r]; } } while (r < dl->sorted); dl->sorted = w; } while (r < dl->length) { if (dl->items[++r].pgno < most) { if (++w != r) dl->items[w] = dl->items[r]; } } dl->length = w; mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + dl->length == txn->mt_env->me_options.dp_limit); goto unlink_loose; } } else { /* Dirtylist is mostly sorted, just refund loose pages at the end. */ mdbx_dpl_sort(dl); mdbx_tassert(txn, dl->length < 2 || dl->items[1].pgno < dl->items[dl->length].pgno); mdbx_tassert(txn, dl->sorted == dl->length); /* Scan dirtylist tail-forward and cutoff suitable pages. */ while (dl->length && dl->items[dl->length].pgno == txn->mt_next_pgno - 1 && dl->items[dl->length].ptr->mp_flags == (P_LOOSE | P_DIRTY)) { MDBX_page *dp = dl->items[dl->length].ptr; mdbx_debug("refund-sorted page %" PRIaPGNO, dp->mp_pgno); mdbx_tassert(txn, dp->mp_pgno == dl->items[dl->length].pgno); txn->mt_next_pgno -= 1; dl->length -= 1; } if (dl->sorted != dl->length) { const unsigned refunded = dl->sorted - dl->length; dl->sorted = dl->length; txn->tw.loose_count -= refunded; txn->tw.dirtyroom += refunded; mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + dl->length == txn->mt_env->me_options.dp_limit); assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); /* Filter-out loose chain & dispose refunded pages. */ unlink_loose: for (MDBX_page **link = &txn->tw.loose_pages; *link;) { MDBX_page *dp = *link; mdbx_tassert(txn, dp->mp_flags == (P_LOOSE | P_DIRTY)); if (txn->mt_next_pgno > dp->mp_pgno) { link = &dp->mp_next; } else { *link = dp->mp_next; if ((txn->mt_flags & MDBX_WRITEMAP) == 0) mdbx_dpage_free(txn->mt_env, dp, 1); } } } } mdbx_tassert(txn, mdbx_dirtylist_check(txn)); mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + dl->length == txn->mt_env->me_options.dp_limit); if (suitable != onstack) mdbx_pnl_free(suitable); txn->tw.loose_refund_wl = txn->mt_next_pgno; } static bool mdbx_refund(MDBX_txn *txn) { const pgno_t before = txn->mt_next_pgno; if (txn->tw.loose_pages && txn->tw.loose_refund_wl > txn->mt_next_pgno) mdbx_refund_loose(txn); while (true) { if (MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) == 0 || MDBX_PNL_MOST(txn->tw.reclaimed_pglist) != txn->mt_next_pgno - 1) break; mdbx_refund_reclaimed(txn); if (!txn->tw.loose_pages || txn->tw.loose_refund_wl <= txn->mt_next_pgno) break; const pgno_t memo = txn->mt_next_pgno; mdbx_refund_loose(txn); if (memo == txn->mt_next_pgno) break; } if (before == txn->mt_next_pgno) return false; if (txn->tw.spill_pages) /* Squash deleted pagenums if we refunded any */ mdbx_spill_purge(txn); return true; } #else /* MDBX_ENABLE_REFUND */ static __inline bool mdbx_refund(MDBX_txn *txn) { (void)txn; /* No online auto-compactification. */ return false; } #endif /* MDBX_ENABLE_REFUND */ static __cold void mdbx_kill_page(MDBX_env *env, MDBX_page *mp, pgno_t pgno, unsigned npages) { mdbx_debug("kill%s %u page %" PRIaPGNO, IS_DIRTY(mp) ? " dirty" : "", npages, pgno); mdbx_assert(env, pgno >= NUM_METAS && npages); if (IS_DIRTY(mp) || (env->me_flags & MDBX_WRITEMAP)) { const size_t bytes = pgno2bytes(env, npages); memset(mp, -1, bytes); mp->mp_pgno = pgno; if ((env->me_flags & MDBX_WRITEMAP) == 0) mdbx_pwrite(env->me_lazy_fd, mp, bytes, pgno2bytes(env, pgno)); } else { struct iovec iov[MDBX_COMMIT_PAGES]; iov[0].iov_len = env->me_psize; iov[0].iov_base = (char *)env->me_pbuf + env->me_psize; size_t iov_off = pgno2bytes(env, pgno); unsigned n = 1; while (--npages) { iov[n] = iov[0]; if (++n == MDBX_COMMIT_PAGES) { mdbx_pwritev(env->me_lazy_fd, iov, MDBX_COMMIT_PAGES, iov_off, pgno2bytes(env, MDBX_COMMIT_PAGES)); iov_off += pgno2bytes(env, MDBX_COMMIT_PAGES); n = 0; } } mdbx_pwritev(env->me_lazy_fd, iov, n, iov_off, pgno2bytes(env, n)); } } /* Remove page from dirty list */ static __inline void mdbx_page_wash(MDBX_txn *txn, const unsigned di, MDBX_page *const mp, const unsigned npages) { mdbx_tassert(txn, di && di <= txn->tw.dirtylist->length && txn->tw.dirtylist->items[di].ptr == mp); mdbx_dpl_remove(txn->tw.dirtylist, di); txn->tw.dirtyroom++; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); mp->mp_txnid = INVALID_TXNID; mp->mp_flags = 0xFFFF; VALGRIND_MAKE_MEM_UNDEFINED(mp, PAGEHDRSZ); if (txn->mt_flags & MDBX_WRITEMAP) { VALGRIND_MAKE_MEM_NOACCESS(page_data(mp), pgno2bytes(txn->mt_env, npages) - PAGEHDRSZ); ASAN_POISON_MEMORY_REGION(page_data(mp), pgno2bytes(txn->mt_env, npages) - PAGEHDRSZ); } else mdbx_dpage_free(txn->mt_env, mp, npages); } /* Retire, loosen or free a single page. * * Saves single pages to a list for future reuse * in this same txn. It has been pulled from the GC * and already resides on the dirty list, but has been * deleted. Use these pages first before pulling again * from the GC. * * If the page wasn't dirtied in this txn, just add it * to this txn's free list. */ static int mdbx_page_loose(MDBX_txn *txn, MDBX_page *mp) { const unsigned npages = IS_OVERFLOW(mp) ? mp->mp_pages : 1; const pgno_t pgno = mp->mp_pgno; const bool is_dirty = IS_DIRTY(mp); if (is_dirty) { mdbx_tassert(txn, !txn->tw.spill_pages || !mdbx_pnl_exist(txn->tw.spill_pages, pgno << 1)); mdbx_tassert(txn, debug_dpl_find(txn->tw.dirtylist, pgno) == mp || txn->mt_parent); } else { mdbx_tassert(txn, !debug_dpl_find(txn->tw.dirtylist, pgno)); } const unsigned di = is_dirty ? mdbx_dpl_exist(txn->tw.dirtylist, pgno) : 0; const unsigned si = (!is_dirty && txn->tw.spill_pages) ? mdbx_pnl_exist(txn->tw.spill_pages, pgno << 1) : 0; if (MDBX_ENABLE_REFUND && unlikely(pgno + npages == txn->mt_next_pgno)) { const char *kind; if (di) { mdbx_tassert(txn, txn->tw.dirtylist->items[di].ptr == mp); if (unlikely(txn->tw.dirtylist->items[di].ptr != mp)) { wrong_dirty: mdbx_error("wrong dirty page 0x%p #%" PRIaPGNO, __Wpedantic_format_voidptr(mp), pgno); txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PROBLEM; } /* Страница испачкана в этой транзакции, но до этого могла быть * аллоцирована, испачкана и пролита в одной из родительских транзакций. * Её МОЖНО вытолкнуть в нераспределенный хвост. */ kind = "dirty"; /* Remove from dirty list */ mdbx_page_wash(txn, di, mp, npages); } else if (is_dirty) { /* The page MUST BE from parent transaction. */ if (mdbx_audit_enabled()) { const MDBX_page *parent_dp = nullptr; for (MDBX_txn *parent = txn->mt_parent; parent && !parent_dp; parent = parent->mt_parent) { mdbx_tassert(txn, !parent->tw.spill_pages || !mdbx_pnl_exist(parent->tw.spill_pages, pgno << 1)); parent_dp = debug_dpl_find(parent->tw.dirtylist, pgno); } mdbx_tassert(txn, parent_dp == mp); } kind = "parent-dirty"; } else if (si) { /* Страница пролита в этой транзакции, следовательно она аллоцирована * и запачкана в этой или одной из родительских транзакций. * Её МОЖНО вытолкнуть в нераспределенный хвост. */ kind = "spilled"; mdbx_spill_remove(txn, si, npages); } else { for (MDBX_txn *parent = txn->mt_parent; parent && (parent->mt_flags & MDBX_TXN_SPILLS); parent = parent->mt_parent) if (parent->tw.spill_pages && mdbx_pnl_exist(parent->tw.spill_pages, pgno << 1)) { /* Страница аллоцирована, запачкана и пролита в одной из родительских * транзакций. Её МОЖНО вытолкнуть в нераспределенный хвост. */ kind = "parent-spilled"; goto refund; } /* Страница используется (входит в MVCC-снимки доступные читателям). * Её НЕЛЬЗЯ вытолкнуть в нераспределенных хвост. */ goto retire; } refund: mdbx_debug("refunded %u %s page %" PRIaPGNO, npages, kind, pgno); txn->mt_next_pgno = pgno; mdbx_refund(txn); return MDBX_SUCCESS; } if (is_dirty) { if (di) { /* Dirty page from this transaction */ mdbx_tassert(txn, txn->tw.dirtylist->items[di].ptr == mp); if (unlikely(txn->tw.dirtylist->items[di].ptr != mp)) goto wrong_dirty; /* If suitable we can reuse it through loose list */ if (likely( npages == 1 && txn->tw.loose_count < txn->mt_env->me_options.dp_loose_limit && (!MDBX_ENABLE_REFUND || /* skip pages near to the end in favor of compactification */ txn->mt_next_pgno > pgno + txn->mt_env->me_options.dp_loose_limit || txn->mt_next_pgno <= txn->mt_env->me_options.dp_loose_limit))) { mdbx_debug("loosen dirty page %" PRIaPGNO, pgno); mp->mp_flags = P_LOOSE | P_DIRTY; mp->mp_next = txn->tw.loose_pages; txn->tw.loose_pages = mp; txn->tw.loose_count++; #if MDBX_ENABLE_REFUND txn->tw.loose_refund_wl = (pgno + 2 > txn->tw.loose_refund_wl) ? pgno + 2 : txn->tw.loose_refund_wl; #endif /* MDBX_ENABLE_REFUND */ if (MDBX_DEBUG || unlikely(txn->mt_env->me_flags & MDBX_PAGEPERTURB)) memset(page_data(mp), -1, txn->mt_env->me_psize - PAGEHDRSZ); VALGRIND_MAKE_MEM_NOACCESS(page_data(mp), txn->mt_env->me_psize - PAGEHDRSZ); ASAN_POISON_MEMORY_REGION(page_data(mp), txn->mt_env->me_psize - PAGEHDRSZ); return MDBX_SUCCESS; } #if !MDBX_DEBUG && !defined(MDBX_USE_VALGRIND) && !defined(__SANITIZE_ADDRESS__) if (unlikely(txn->mt_env->me_flags & MDBX_PAGEPERTURB)) #endif { /* Страница могла быть изменена в одной из родительских транзакций, * в том числе, позже выгружена и затем снова загружена и изменена. * В обоих случаях её нельзя затирать на диске и помечать недоступной * в asan и/или valgrind */ for (MDBX_txn *parent = txn->mt_parent; parent && (parent->mt_flags & MDBX_TXN_SPILLS); parent = parent->mt_parent) { if (parent->tw.spill_pages && mdbx_pnl_intersect(parent->tw.spill_pages, pgno << 1, npages << 1)) goto skip_invalidate; if (mdbx_dpl_intersect(parent->tw.dirtylist, pgno, npages)) goto skip_invalidate; } #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) if (MDBX_DEBUG || unlikely(txn->mt_env->me_flags & MDBX_PAGEPERTURB)) #endif mdbx_kill_page(txn->mt_env, mp, pgno, npages); if (!(txn->mt_flags & MDBX_WRITEMAP)) { VALGRIND_MAKE_MEM_NOACCESS(page_data(pgno2page(txn->mt_env, pgno)), pgno2bytes(txn->mt_env, npages) - PAGEHDRSZ); ASAN_POISON_MEMORY_REGION(page_data(pgno2page(txn->mt_env, pgno)), pgno2bytes(txn->mt_env, npages) - PAGEHDRSZ); } } skip_invalidate: /* Remove from dirty list */ mdbx_page_wash(txn, di, mp, npages); } else { /* Dirty page MUST BE a clone from (one of) parent transaction(s). */ if (mdbx_audit_enabled()) { const MDBX_page *parent_dp = nullptr; /* Check parent(s)'s dirty lists. */ for (MDBX_txn *parent = txn->mt_parent; parent && !parent_dp; parent = parent->mt_parent) { mdbx_tassert(txn, !parent->tw.spill_pages || !mdbx_pnl_exist(parent->tw.spill_pages, pgno << 1)); parent_dp = debug_dpl_find(parent->tw.dirtylist, pgno); } mdbx_tassert(txn, parent_dp == mp); } } reclaim: mdbx_debug("reclaim %u %s page %" PRIaPGNO, npages, "dirty", pgno); int rc = mdbx_pnl_insert_range(&txn->tw.reclaimed_pglist, pgno, npages); mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); return rc; } if (si) { /* Page ws spilled in this txn */ mdbx_spill_remove(txn, si, npages); /* Страница могла быть выделена и затем пролита в этой транзакции, * тогда её необходимо поместить в reclaimed-список. * Либо она могла быть выделена в одной из родительских транзакций и затем * пролита в этой транзакции, тогда её необходимо поместить в retired-список * для последующей фильтрации при коммите. */ for (MDBX_txn *parent = txn->mt_parent; parent; parent = parent->mt_parent) { if (mdbx_dpl_exist(parent->tw.dirtylist, pgno)) goto retire; } /* Страница точно была выделена в этой транзакции * и теперь может быть использована повторно. */ goto reclaim; } else { /* Страница может входить в доступный читателям MVCC-снимок, либо же она * могла быть выделена, а затем пролита в одной из родительских транзакций. * Поэтому пока помещаем её в retired-список, который будет фильтроваться * относительно dirty- и spilled-списков родительских транзакций при коммите * дочерних транзакций, либо же будет записан в GC в неизменном виде. */ } retire: mdbx_debug("retire %u page %" PRIaPGNO, npages, pgno); int rc = mdbx_pnl_append_range(false, &txn->tw.retired_pages, pgno, npages); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); return rc; } static int mdbx_page_retire(MDBX_cursor *mc, MDBX_page *mp) { if (unlikely(mc->mc_flags & C_SUB)) { MDBX_db *outer = mdbx_outer_db(mc); mdbx_cassert(mc, !IS_BRANCH(mp) || outer->md_branch_pages > 0); outer->md_branch_pages -= IS_BRANCH(mp); mdbx_cassert(mc, !IS_LEAF(mp) || outer->md_leaf_pages > 0); outer->md_leaf_pages -= IS_LEAF(mp); mdbx_cassert(mc, !IS_OVERFLOW(mp)); } mdbx_cassert(mc, !IS_BRANCH(mp) || mc->mc_db->md_branch_pages > 0); mc->mc_db->md_branch_pages -= IS_BRANCH(mp); mdbx_cassert(mc, !IS_LEAF(mp) || mc->mc_db->md_leaf_pages > 0); mc->mc_db->md_leaf_pages -= IS_LEAF(mp); mdbx_cassert(mc, !IS_OVERFLOW(mp) || mc->mc_db->md_overflow_pages >= mp->mp_pages); mc->mc_db->md_overflow_pages -= IS_OVERFLOW(mp) ? mp->mp_pages : 0; return mdbx_page_loose(mc->mc_txn, mp); } static __must_check_result __always_inline int mdbx_retire_pgno(MDBX_cursor *mc, const pgno_t pgno) { MDBX_page *mp; int rc; if (mdbx_audit_enabled()) { const unsigned save_flags = mc->mc_flags; mc->mc_flags |= C_RETIRING; rc = mdbx_page_get(mc, pgno, &mp, NULL, mc->mc_txn->mt_txnid); if (likely(rc == MDBX_SUCCESS)) rc = mdbx_page_retire(mc, mp); mc->mc_flags = (mc->mc_flags & ~C_RETIRING) | (save_flags & C_RETIRING); } else { rc = mdbx_page_get(mc, pgno, &mp, NULL, mc->mc_txn->mt_txnid); if (likely(rc == MDBX_SUCCESS)) rc = mdbx_page_retire(mc, mp); } return rc; } /* Toggle P_KEEP in dirty, non-overflow, non-sub pages watched by txn. * * [in] mc A cursor handle for the current operation. * [in] pflags Flags of the pages to update: * - P_DIRTY to set P_KEEP, * - P_DIRTY|P_KEEP to clear it. */ static void mdbx_cursor_xkeep(MDBX_cursor *mc, unsigned pflags) { const unsigned mask = P_SUBP | P_DIRTY | P_LOOSE | P_KEEP; if (mc->mc_flags & C_INITIALIZED) { MDBX_cursor *m3 = mc; for (;;) { MDBX_page *mp = NULL; for (unsigned j = 0; j < m3->mc_snum; j++) { mp = m3->mc_pg[j]; if ((mp->mp_flags & mask) == pflags) mp->mp_flags ^= P_KEEP; } if (!(mp && IS_LEAF(mp))) break; /* Proceed to mx if it is at a sub-database */ MDBX_xcursor *mx = m3->mc_xcursor; if (!(mx && (mx->mx_cursor.mc_flags & C_INITIALIZED))) break; const unsigned nkeys = page_numkeys(mp); unsigned ki = m3->mc_ki[m3->mc_top]; mdbx_cassert(mc, nkeys > 0 && (ki < nkeys || (ki == nkeys && (mx->mx_cursor.mc_flags & C_EOF)))); ki -= ki >= nkeys; if (!(node_flags(page_node(mp, ki)) & F_SUBDATA)) break; m3 = &mx->mx_cursor; } } } /* Mark pages seen by cursors: First m0, then tracked cursors * [in] all No shortcuts. Needed except after a full mdbx_page_flush(). */ static void mdbx_txn_xkeep(MDBX_txn *txn, MDBX_cursor *m0, const unsigned pflags, const bool all) { if (m0) mdbx_cursor_xkeep(m0, pflags); for (unsigned i = FREE_DBI; i < txn->mt_numdbs; ++i) if (txn->mt_dbistate[i] & DBI_DIRTY) for (MDBX_cursor *mc = txn->tw.cursors[i]; mc; mc = mc->mc_next) if (mc != m0) mdbx_cursor_xkeep(mc, pflags); if (all) { /* Mark dirty root pages */ const unsigned mask = P_SUBP | P_DIRTY | P_LOOSE | P_KEEP; for (unsigned i = 0; i < txn->mt_numdbs; i++) { if (txn->mt_dbistate[i] & DBI_DIRTY) { pgno_t pgno = txn->mt_dbs[i].md_root; if (pgno == P_INVALID) continue; MDBX_page *dp = mdbx_dpl_find(txn->tw.dirtylist, pgno); if (dp && (dp->mp_flags & mask) == pflags) dp->mp_flags ^= P_KEEP; } } } } /* Spill pages from the dirty list back to disk. * This is intended to prevent running into MDBX_TXN_FULL situations, * but note that they may still occur in a few cases: * * 1) our estimate of the txn size could be too small. Currently this * seems unlikely, except with a large number of MDBX_MULTIPLE items. * * 2) child txns may run out of space if their parents dirtied a * lot of pages and never spilled them. TODO: we probably should do * a preemptive spill during mdbx_txn_begin() of a child txn, if * the parent's dirtyroom is below a given threshold. * * Otherwise, if not using nested txns, it is expected that apps will * not run into MDBX_TXN_FULL any more. The pages are flushed to disk * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared. * If the txn never references them again, they can be left alone. * If the txn only reads them, they can be used without any fuss. * If the txn writes them again, they can be dirtied immediately without * going thru all of the work of mdbx_page_touch(). Such references are * handled by mdbx_page_unspill(). * * Also note, we never spill DB root pages, nor pages of active cursors, * because we'll need these back again soon anyway. And in nested txns, * we can't spill a page in a child txn if it was already spilled in a * parent txn. That would alter the parent txns' data even though * the child hasn't committed yet, and we'd have no way to undo it if * the child aborted. */ static int mdbx_txn_spill(MDBX_txn *txn, MDBX_cursor *m0, unsigned need) { #ifndef MDBX_DEBUG_SPILLING if (likely(txn->tw.dirtyroom > need)) return MDBX_SUCCESS; unsigned spill = need - txn->tw.dirtyroom; #else /* spill at least one page if defined MDBX_DEBUG_SPILLING */ unsigned spill = (need > txn->tw.dirtyroom) ? need - txn->tw.dirtyroom : 1; #endif /* MDBX_DEBUG_SPILLING */ const unsigned dirty = txn->tw.dirtylist->length; const unsigned spill_min = txn->mt_env->me_options.spill_min_denominator ? dirty / txn->mt_env->me_options.spill_min_denominator : 0; const unsigned spill_max = dirty - (txn->mt_env->me_options.spill_max_denominator ? dirty / txn->mt_env->me_options.spill_max_denominator : 0); spill = (spill > spill_min) ? spill : spill_min; spill = (spill < spill_max) ? spill : spill_max; if (!spill) return MDBX_SUCCESS; mdbx_notice("spilling %u dirty-entries (have %u dirty-room, need %u)", spill, txn->tw.dirtyroom, need); mdbx_tassert(txn, txn->tw.dirtylist->length >= spill); int rc; if (!txn->tw.spill_pages) { txn->tw.spill_least_removed = INT_MAX; txn->tw.spill_pages = mdbx_pnl_alloc(spill); if (unlikely(!txn->tw.spill_pages)) { rc = MDBX_ENOMEM; goto bailout; } } else { /* purge deleted slots */ mdbx_spill_purge(txn); rc = mdbx_pnl_reserve(&txn->tw.spill_pages, spill); (void)rc /* ignore since the resulting list may be shorter and mdbx_pnl_append() will increase pnl on demand */ ; } /* Preserve pages which may soon be dirtied again */ mdbx_txn_xkeep(txn, m0, P_DIRTY, true); MDBX_dpl *const dl = mdbx_dpl_sort(txn->tw.dirtylist); /* Save the page IDs of all the pages we're flushing */ /* flush from the tail forward, this saves a lot of shifting later on. */ const unsigned dl_len_before = dl->length; unsigned spilled = 0; unsigned keep = dl_len_before; for (; keep && spill; keep--) { const pgno_t pgno = dl->items[keep].pgno; MDBX_page *dp = dl->items[keep].ptr; const unsigned npages = IS_OVERFLOW(dp) ? dp->mp_pages : 1; if (dp->mp_flags & (P_LOOSE | P_KEEP)) { mdbx_debug("skip %s %u page %" PRIaPGNO, (dp->mp_flags & P_LOOSE) ? "loose" : "keep", npages, dp->mp_pgno); skip: continue; } /* Can't spill twice, * make sure it's not already in a parent's spill list(s). */ MDBX_txn *parent = txn->mt_parent; if (parent && (parent->mt_flags & MDBX_TXN_SPILLS)) { do if (parent->tw.spill_pages && mdbx_pnl_intersect(parent->tw.spill_pages, pgno << 1, npages << 1)) { mdbx_debug("skip parent-spilled %u page %" PRIaPGNO, npages, pgno); dp->mp_flags |= P_KEEP; goto skip; } while ((parent = parent->mt_parent) != nullptr); } mdbx_debug("spill %u page %" PRIaPGNO, npages, dp->mp_pgno); rc = mdbx_pnl_append_range(true, &txn->tw.spill_pages, pgno << 1, npages); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; spill--; spilled += 1; } mdbx_pnl_sort(txn->tw.spill_pages); /* Flush the spilled part of dirty list */ rc = mdbx_page_flush(txn, keep); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; /* Reset any dirty pages we kept that page_flush didn't see */ mdbx_tassert(txn, dl_len_before - spilled == dl->length); mdbx_txn_xkeep(txn, m0, P_DIRTY | P_KEEP, keep > 0); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); bailout: txn->mt_flags |= rc ? MDBX_TXN_ERROR : MDBX_TXN_SPILLS; return rc; } static int mdbx_cursor_spill(MDBX_cursor *mc, const MDBX_val *key, const MDBX_val *data) { if (mc->mc_flags & C_SUB) return MDBX_SUCCESS; MDBX_txn *txn = mc->mc_txn; if (txn->mt_flags & MDBX_WRITEMAP) return MDBX_SUCCESS; /* Estimate how much space this operation will take: */ /* 1) Max b-tree height, reasonable enough with including dups' sub-tree */ unsigned need = CURSOR_STACK + 3; /* 2) GC/FreeDB for any payload */ if (mc->mc_dbi > FREE_DBI) { need += txn->mt_dbs[FREE_DBI].md_depth + 3; /* 3) Named DBs also dirty the main DB */ if (mc->mc_dbi > MAIN_DBI) need += txn->mt_dbs[MAIN_DBI].md_depth + 3; } /* 4) Factor the key+data which to be put in */ need += bytes2pgno(txn->mt_env, node_size(key, data)) + 1; return mdbx_txn_spill(txn, mc, need); } /*----------------------------------------------------------------------------*/ static __always_inline bool meta_bootid_match(const MDBX_meta *meta) { return memcmp(&meta->mm_bootid, &bootid, 16) == 0 && (bootid.x | bootid.y) != 0; } static bool meta_weak_acceptable(const MDBX_env *env, const MDBX_meta *meta, const int lck_exclusive) { return lck_exclusive ? /* exclusive lock */ meta_bootid_match(meta) : /* db already opened */ env->me_lck && (env->me_lck->mti_envmode.weak & MDBX_RDONLY) == 0; } #define METAPAGE(env, n) page_meta(pgno2page(env, n)) #define METAPAGE_END(env) METAPAGE(env, NUM_METAS) static __inline txnid_t meta_txnid(const MDBX_env *env, const MDBX_meta *meta, const bool allow_volatile) { mdbx_memory_fence(mo_AcquireRelease, false); txnid_t a = unaligned_peek_u64(4, &meta->mm_txnid_a); txnid_t b = unaligned_peek_u64(4, &meta->mm_txnid_b); if (allow_volatile) return (a == b) ? a : 0; mdbx_assert(env, a == b); return a; } static __inline txnid_t mdbx_meta_txnid_stable(const MDBX_env *env, const MDBX_meta *meta) { return meta_txnid(env, meta, false); } static __inline txnid_t mdbx_meta_txnid_fluid(const MDBX_env *env, const MDBX_meta *meta) { return meta_txnid(env, meta, true); } static __inline void mdbx_meta_update_begin(const MDBX_env *env, MDBX_meta *meta, txnid_t txnid) { mdbx_assert(env, meta >= METAPAGE(env, 0) && meta < METAPAGE_END(env)); mdbx_assert(env, unaligned_peek_u64(4, meta->mm_txnid_a) < txnid && unaligned_peek_u64(4, meta->mm_txnid_b) < txnid); (void)env; unaligned_poke_u64(4, meta->mm_txnid_b, 0); mdbx_memory_fence(mo_AcquireRelease, true); unaligned_poke_u64(4, meta->mm_txnid_a, txnid); } static __inline void mdbx_meta_update_end(const MDBX_env *env, MDBX_meta *meta, txnid_t txnid) { mdbx_assert(env, meta >= METAPAGE(env, 0) && meta < METAPAGE_END(env)); mdbx_assert(env, unaligned_peek_u64(4, meta->mm_txnid_a) == txnid); mdbx_assert(env, unaligned_peek_u64(4, meta->mm_txnid_b) < txnid); (void)env; mdbx_jitter4testing(true); memcpy(&meta->mm_bootid, &bootid, 16); unaligned_poke_u64(4, meta->mm_txnid_b, txnid); mdbx_memory_fence(mo_AcquireRelease, true); } static __inline void mdbx_meta_set_txnid(const MDBX_env *env, MDBX_meta *meta, txnid_t txnid) { mdbx_assert(env, !env->me_map || meta < METAPAGE(env, 0) || meta >= METAPAGE_END(env)); (void)env; /* update inconsistent since this function used ONLY for filling meta-image * for writing, but not the actual meta-page */ memcpy(&meta->mm_bootid, &bootid, 16); unaligned_poke_u64(4, meta->mm_txnid_a, txnid); unaligned_poke_u64(4, meta->mm_txnid_b, txnid); } static __inline uint64_t mdbx_meta_sign(const MDBX_meta *meta) { uint64_t sign = MDBX_DATASIGN_NONE; #if 0 /* TODO */ sign = hippeus_hash64(...); #else (void)meta; #endif /* LY: newer returns MDBX_DATASIGN_NONE or MDBX_DATASIGN_WEAK */ return (sign > MDBX_DATASIGN_WEAK) ? sign : ~sign; } enum meta_choise_mode { prefer_last, prefer_steady }; static __inline bool mdbx_meta_ot(const enum meta_choise_mode mode, const MDBX_env *env, const MDBX_meta *a, const MDBX_meta *b) { mdbx_jitter4testing(true); txnid_t txnid_a = mdbx_meta_txnid_fluid(env, a); txnid_t txnid_b = mdbx_meta_txnid_fluid(env, b); mdbx_jitter4testing(true); switch (mode) { default: assert(false); __unreachable(); /* fall through */ __fallthrough; case prefer_steady: if (META_IS_STEADY(a) != META_IS_STEADY(b)) return META_IS_STEADY(b); /* fall through */ __fallthrough; case prefer_last: mdbx_jitter4testing(true); if (txnid_a == txnid_b) return META_IS_STEADY(b); return txnid_a < txnid_b; } } static __inline bool mdbx_meta_eq(const MDBX_env *env, const MDBX_meta *a, const MDBX_meta *b) { mdbx_jitter4testing(true); const txnid_t txnid = mdbx_meta_txnid_fluid(env, a); if (!txnid || txnid != mdbx_meta_txnid_fluid(env, b)) return false; mdbx_jitter4testing(true); if (META_IS_STEADY(a) != META_IS_STEADY(b)) return false; mdbx_jitter4testing(true); return true; } static int mdbx_meta_eq_mask(const MDBX_env *env) { MDBX_meta *m0 = METAPAGE(env, 0); MDBX_meta *m1 = METAPAGE(env, 1); MDBX_meta *m2 = METAPAGE(env, 2); int rc = mdbx_meta_eq(env, m0, m1) ? 1 : 0; if (mdbx_meta_eq(env, m1, m2)) rc += 2; if (mdbx_meta_eq(env, m2, m0)) rc += 4; return rc; } static __inline MDBX_meta *mdbx_meta_recent(const enum meta_choise_mode mode, const MDBX_env *env, MDBX_meta *a, MDBX_meta *b) { const bool a_older_that_b = mdbx_meta_ot(mode, env, a, b); mdbx_assert(env, !mdbx_meta_eq(env, a, b)); return a_older_that_b ? b : a; } static __inline MDBX_meta *mdbx_meta_ancient(const enum meta_choise_mode mode, const MDBX_env *env, MDBX_meta *a, MDBX_meta *b) { const bool a_older_that_b = mdbx_meta_ot(mode, env, a, b); mdbx_assert(env, !mdbx_meta_eq(env, a, b)); return a_older_that_b ? a : b; } static __inline MDBX_meta * mdbx_meta_mostrecent(const enum meta_choise_mode mode, const MDBX_env *env) { MDBX_meta *m0 = METAPAGE(env, 0); MDBX_meta *m1 = METAPAGE(env, 1); MDBX_meta *m2 = METAPAGE(env, 2); MDBX_meta *head = mdbx_meta_recent(mode, env, m0, m1); head = mdbx_meta_recent(mode, env, head, m2); return head; } static MDBX_meta *mdbx_meta_steady(const MDBX_env *env) { return mdbx_meta_mostrecent(prefer_steady, env); } static MDBX_meta *mdbx_meta_head(const MDBX_env *env) { return mdbx_meta_mostrecent(prefer_last, env); } static txnid_t mdbx_recent_committed_txnid(const MDBX_env *env) { while (true) { const MDBX_meta *head = mdbx_meta_head(env); const txnid_t recent = mdbx_meta_txnid_fluid(env, head); mdbx_compiler_barrier(); if (likely(head == mdbx_meta_head(env) && recent == mdbx_meta_txnid_fluid(env, head))) return recent; } } static txnid_t mdbx_recent_steady_txnid(const MDBX_env *env) { while (true) { const MDBX_meta *head = mdbx_meta_steady(env); const txnid_t recent = mdbx_meta_txnid_fluid(env, head); mdbx_compiler_barrier(); if (likely(head == mdbx_meta_steady(env) && recent == mdbx_meta_txnid_fluid(env, head))) return recent; } } static const char *mdbx_durable_str(const MDBX_meta *const meta) { if (META_IS_STEADY(meta)) return (unaligned_peek_u64(4, meta->mm_datasync_sign) == mdbx_meta_sign(meta)) ? "Steady" : "Tainted"; return "Weak"; } /*----------------------------------------------------------------------------*/ /* Find oldest txnid still referenced. */ static txnid_t mdbx_find_oldest(const MDBX_txn *txn) { mdbx_tassert(txn, (txn->mt_flags & MDBX_TXN_RDONLY) == 0); MDBX_env *env = txn->mt_env; const txnid_t edge = mdbx_recent_steady_txnid(env); mdbx_tassert(txn, edge <= txn->mt_txnid); MDBX_lockinfo *const lck = env->me_lck; if (unlikely(lck == NULL /* exclusive mode */)) return atomic_store64(&env->me_lckless_stub.oldest, edge, mo_AcquireRelease); const txnid_t last_oldest = atomic_load64(&lck->mti_oldest_reader, mo_AcquireRelease); mdbx_tassert(txn, edge >= last_oldest); if (likely(last_oldest == edge)) return edge; const uint32_t nothing_changed = MDBX_STRING_TETRAD("None"); const uint32_t snap_readers_refresh_flag = atomic_load32(&lck->mti_readers_refresh_flag, mo_AcquireRelease); mdbx_jitter4testing(false); if (snap_readers_refresh_flag == nothing_changed) return last_oldest; txnid_t oldest = edge; atomic_store32(&lck->mti_readers_refresh_flag, nothing_changed, mo_AcquireRelease); const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease)) { /* mdbx_jitter4testing(true); */ const txnid_t snap = safe64_read(&lck->mti_readers[i].mr_txnid); if (oldest > snap && last_oldest <= /* ignore pending updates */ snap) { oldest = snap; if (oldest == last_oldest) return oldest; } } } if (oldest != last_oldest) { mdbx_notice("update oldest %" PRIaTXN " -> %" PRIaTXN, last_oldest, oldest); mdbx_tassert(txn, oldest >= lck->mti_oldest_reader.weak); atomic_store64(&lck->mti_oldest_reader, oldest, mo_AcquireRelease); } return oldest; } /* Find largest mvcc-snapshot still referenced. */ static __cold pgno_t mdbx_find_largest(MDBX_env *env, pgno_t largest) { MDBX_lockinfo *const lck = env->me_lck; if (likely(lck != NULL /* exclusive mode */)) { const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { retry: if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease)) { /* mdbx_jitter4testing(true); */ const pgno_t snap_pages = atomic_load32( &lck->mti_readers[i].mr_snapshot_pages_used, mo_Relaxed); const txnid_t snap_txnid = safe64_read(&lck->mti_readers[i].mr_txnid); if (unlikely( snap_pages != atomic_load32(&lck->mti_readers[i].mr_snapshot_pages_used, mo_AcquireRelease) || snap_txnid != safe64_read(&lck->mti_readers[i].mr_txnid))) goto retry; if (largest < snap_pages && atomic_load64(&lck->mti_oldest_reader, mo_AcquireRelease) <= /* ignore pending updates */ snap_txnid && snap_txnid <= env->me_txn0->mt_txnid) largest = snap_pages; } } } return largest; } /* Add a page to the txn's dirty list */ static int __must_check_result mdbx_page_dirty(MDBX_txn *txn, MDBX_page *mp) { mp->mp_txnid = INVALID_TXNID; mp->mp_flags |= P_DIRTY; if (unlikely(txn->tw.dirtyroom == 0)) { mdbx_error("Dirtyroom is depleted, DPL length %u", txn->tw.dirtylist->length); return MDBX_TXN_FULL; } const int rc = mdbx_dpl_append(txn, mp->mp_pgno, mp); if (unlikely(rc != MDBX_SUCCESS)) { txn->mt_flags |= MDBX_TXN_ERROR; return rc; } txn->tw.dirtyroom--; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); return MDBX_SUCCESS; } #if !(defined(_WIN32) || defined(_WIN64)) static __always_inline __maybe_unused int ignore_enosys(int err) { #ifdef ENOSYS if (err == ENOSYS) return MDBX_RESULT_TRUE; #endif /* ENOSYS */ #ifdef ENOIMPL if (err == ENOIMPL) return MDBX_RESULT_TRUE; #endif /* ENOIMPL */ #ifdef ENOTSUP if (err == ENOTSUP) return MDBX_RESULT_TRUE; #endif /* ENOTSUP */ #ifdef ENOSUPP if (err == ENOSUPP) return MDBX_RESULT_TRUE; #endif /* ENOSUPP */ #ifdef EOPNOTSUPP if (err == EOPNOTSUPP) return MDBX_RESULT_TRUE; #endif /* EOPNOTSUPP */ if (err == EAGAIN) return MDBX_RESULT_TRUE; return err; } #endif /* defined(_WIN32) || defined(_WIN64) */ /* Turn on/off readahead. It's harmful when the DB is larger than RAM. */ static __cold int mdbx_set_readahead(MDBX_env *env, const size_t offset, const size_t length, const bool enable) { assert(length > 0); mdbx_notice("readahead %s %u..%u", enable ? "ON" : "OFF", bytes2pgno(env, offset), bytes2pgno(env, offset + length)); #if defined(F_RDAHEAD) if (unlikely(fcntl(env->me_lazy_fd, F_RDAHEAD, enable) == -1)) return errno; #endif /* F_RDAHEAD */ if (enable) { #if defined(F_RDADVISE) struct radvisory hint; hint.ra_offset = offset; hint.ra_count = length; (void)/* Ignore ENOTTY for DB on the ram-disk and so on */ fcntl( env->me_lazy_fd, F_RDADVISE, &hint); #endif /* F_RDADVISE */ #if defined(MADV_WILLNEED) int err = madvise(env->me_map + offset, length, MADV_WILLNEED) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(POSIX_MADV_WILLNEED) int err = ignore_enosys( posix_madvise(env->me_map + offset, length, POSIX_MADV_WILLNEED)); if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(_WIN32) || defined(_WIN64) if (mdbx_PrefetchVirtualMemory) { WIN32_MEMORY_RANGE_ENTRY hint; hint.VirtualAddress = env->me_map + offset; hint.NumberOfBytes = length; (void)mdbx_PrefetchVirtualMemory(GetCurrentProcess(), 1, &hint, 0); } #elif defined(POSIX_FADV_WILLNEED) int err = ignore_enosys( posix_fadvise(env->me_lazy_fd, offset, length, POSIX_FADV_WILLNEED)); if (unlikely(MDBX_IS_ERROR(err))) return err; #endif /* MADV_WILLNEED */ } else { #if defined(MADV_RANDOM) int err = madvise(env->me_map + offset, length, MADV_RANDOM) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(POSIX_MADV_RANDOM) int err = ignore_enosys( posix_madvise(env->me_map + offset, length, POSIX_MADV_RANDOM)); if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(POSIX_FADV_RANDOM) int err = ignore_enosys( posix_fadvise(env->me_lazy_fd, offset, length, POSIX_FADV_RANDOM)); if (unlikely(MDBX_IS_ERROR(err))) return err; #endif /* MADV_RANDOM */ } return MDBX_SUCCESS; } static __cold int mdbx_mapresize(MDBX_env *env, const pgno_t used_pgno, const pgno_t size_pgno, const pgno_t limit_pgno, const bool implicit) { if ((env->me_flags & MDBX_WRITEMAP) && env->me_unsynced_pages->weak) { int err = mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, used_pgno), true); if (unlikely(err != MDBX_SUCCESS)) return err; } const size_t limit_bytes = pgno_align2os_bytes(env, limit_pgno); const size_t size_bytes = pgno_align2os_bytes(env, size_pgno); const size_t prev_size = env->me_dxb_mmap.current; const size_t prev_limit = env->me_dxb_mmap.limit; const void *const prev_addr = env->me_map; mdbx_verbose("resize datafile/mapping: " "present %" PRIuPTR " -> %" PRIuPTR ", " "limit %" PRIuPTR " -> %" PRIuPTR, prev_size, size_bytes, prev_limit, limit_bytes); mdbx_assert(env, limit_bytes >= size_bytes); mdbx_assert(env, bytes2pgno(env, size_bytes) >= size_pgno); mdbx_assert(env, bytes2pgno(env, limit_bytes) >= limit_pgno); #if defined(_WIN32) || defined(_WIN64) /* Acquire guard in exclusive mode for: * - to avoid collision between read and write txns around env->me_dbgeo; * - to avoid attachment of new reading threads (see mdbx_rdt_lock); */ mdbx_srwlock_AcquireExclusive(&env->me_remap_guard); mdbx_handle_array_t *suspended = NULL; mdbx_handle_array_t array_onstack; int rc = MDBX_SUCCESS; if (limit_bytes == env->me_dxb_mmap.limit && size_bytes == env->me_dxb_mmap.current && size_bytes == env->me_dxb_mmap.filesize) goto bailout; /* 1) Windows allows only extending a read-write section, but not a * corresponding mapped view. Therefore in other cases we must suspend * the local threads for safe remap. * 2) At least on Windows 10 1803 the entire mapped section is unavailable * for short time during NtExtendSection() or VirtualAlloc() execution. * 3) Under Wine runtime environment on Linux a section extending is not * supported. Therefore thread suspending is always required. * * THEREFORE LOCAL THREADS SUSPENDING IS ALWAYS REQUIRED! */ array_onstack.limit = ARRAY_LENGTH(array_onstack.handles); array_onstack.count = 0; suspended = &array_onstack; rc = mdbx_suspend_threads_before_remap(env, &suspended); if (rc != MDBX_SUCCESS) { mdbx_error("failed suspend-for-remap: errcode %d", rc); goto bailout; } const bool mapping_can_be_moved = !implicit; #else /* Windows */ /* Acquire guard to avoid collision between read and write txns * around env->me_dbgeo */ bool mapping_can_be_moved = false; int rc = mdbx_fastmutex_acquire(&env->me_remap_guard); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (limit_bytes == env->me_dxb_mmap.limit && size_bytes == env->me_dxb_mmap.current) goto bailout; if (limit_bytes != env->me_dxb_mmap.limit && env->me_lck && !implicit) { int err = mdbx_rdt_lock(env) /* lock readers table until remap done */; if (unlikely(MDBX_IS_ERROR(err))) { rc = err; goto bailout; } /* looking for readers from this process */ MDBX_lockinfo *const lck = env->me_lck; const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); mapping_can_be_moved = true; for (unsigned i = 0; i < snap_nreaders; ++i) { if (lck->mti_readers[i].mr_pid.weak == env->me_pid && lck->mti_readers[i].mr_tid.weak != mdbx_thread_self()) { /* the base address of the mapping can't be changed since * the other reader thread from this process exists. */ mdbx_rdt_unlock(env); mapping_can_be_moved = false; break; } } } #endif /* ! Windows */ if (size_bytes < prev_size) { mdbx_notice("resize-MADV_%s %u..%u", (env->me_flags & MDBX_WRITEMAP) ? "REMOVE" : "DONTNEED", size_pgno, bytes2pgno(env, prev_size)); rc = MDBX_RESULT_TRUE; #if defined(MADV_REMOVE) if (env->me_flags & MDBX_WRITEMAP) rc = madvise(env->me_map + size_bytes, prev_size - size_bytes, MADV_REMOVE) ? ignore_enosys(errno) : MDBX_SUCCESS; #endif /* MADV_REMOVE */ #if defined(MADV_DONTNEED) if (rc == MDBX_RESULT_TRUE) rc = madvise(env->me_map + size_bytes, prev_size - size_bytes, MADV_DONTNEED) ? ignore_enosys(errno) : MDBX_SUCCESS; #elif defined(POSIX_MADV_DONTNEED) if (rc == MDBX_RESULT_TRUE) rc = ignore_enosys(posix_madvise(env->me_map + size_bytes, prev_size - size_bytes, POSIX_MADV_DONTNEED)); #elif defined(POSIX_FADV_DONTNEED) if (rc == MDBX_RESULT_TRUE) rc = ignore_enosys(posix_fadvise(env->me_lazy_fd, size_bytes, prev_size - size_bytes, POSIX_FADV_DONTNEED)); #endif /* MADV_DONTNEED */ if (unlikely(MDBX_IS_ERROR(rc))) goto bailout; if (env->me_discarded_tail->weak > size_pgno) env->me_discarded_tail->weak = size_pgno; } rc = mdbx_mresize(env->me_flags, &env->me_dxb_mmap, size_bytes, limit_bytes, mapping_can_be_moved); if (rc == MDBX_SUCCESS && (env->me_flags & MDBX_NORDAHEAD) == 0) { const int readahead = mdbx_is_readahead_reasonable(size_bytes, -(intptr_t)prev_size); if (readahead == MDBX_RESULT_FALSE) rc = mdbx_set_readahead( env, 0, (size_bytes > prev_size) ? size_bytes : prev_size, false); else if (readahead == MDBX_RESULT_TRUE) { const size_t readahead_pivot = (limit_bytes != prev_limit || env->me_dxb_mmap.address != prev_addr #if defined(_WIN32) || defined(_WIN64) || prev_size > size_bytes #endif /* Windows */ ) ? 0 /* reassign readahead to the entire map because it was remapped */ : prev_size; if (size_bytes > readahead_pivot) { env->me_discarded_tail->weak = size_pgno; rc = mdbx_set_readahead(env, readahead_pivot, size_bytes - readahead_pivot, true); } } } bailout: if (rc == MDBX_SUCCESS) { #if defined(_WIN32) || defined(_WIN64) mdbx_assert(env, size_bytes == env->me_dxb_mmap.current); mdbx_assert(env, size_bytes <= env->me_dxb_mmap.filesize); mdbx_assert(env, limit_bytes == env->me_dxb_mmap.limit); #endif /* Windows */ #ifdef MDBX_USE_VALGRIND if (prev_limit != env->me_dxb_mmap.limit || prev_addr != env->me_map) { VALGRIND_DISCARD(env->me_valgrind_handle); env->me_valgrind_handle = 0; if (env->me_dxb_mmap.limit) env->me_valgrind_handle = VALGRIND_CREATE_BLOCK(env->me_map, env->me_dxb_mmap.limit, "mdbx"); } #endif /* MDBX_USE_VALGRIND */ } else { if (rc != MDBX_UNABLE_EXTEND_MAPSIZE) { mdbx_error("failed resize datafile/mapping: " "present %" PRIuPTR " -> %" PRIuPTR ", " "limit %" PRIuPTR " -> %" PRIuPTR ", errcode %d", prev_size, size_bytes, prev_limit, limit_bytes, rc); } else { mdbx_warning("unable resize datafile/mapping: " "present %" PRIuPTR " -> %" PRIuPTR ", " "limit %" PRIuPTR " -> %" PRIuPTR ", errcode %d", prev_size, size_bytes, prev_limit, limit_bytes, rc); } if (!env->me_dxb_mmap.address) { env->me_flags |= MDBX_FATAL_ERROR; if (env->me_txn) env->me_txn->mt_flags |= MDBX_TXN_ERROR; rc = MDBX_PANIC; } } #if defined(_WIN32) || defined(_WIN64) int err = MDBX_SUCCESS; mdbx_srwlock_ReleaseExclusive(&env->me_remap_guard); if (suspended) { err = mdbx_resume_threads_after_remap(suspended); if (suspended != &array_onstack) mdbx_free(suspended); } #else if (env->me_lck && mapping_can_be_moved) mdbx_rdt_unlock(env); int err = mdbx_fastmutex_release(&env->me_remap_guard); #endif /* Windows */ if (err != MDBX_SUCCESS) { mdbx_fatal("failed resume-after-remap: errcode %d", err); return MDBX_PANIC; } return rc; } static __cold int mdbx_mapresize_implicit(MDBX_env *env, const pgno_t used_pgno, const pgno_t size_pgno, const pgno_t limit_pgno) { const pgno_t mapped_pgno = bytes2pgno(env, env->me_dxb_mmap.limit); mdbx_assert(env, mapped_pgno >= used_pgno); return mdbx_mapresize( env, used_pgno, size_pgno, (size_pgno > mapped_pgno) ? limit_pgno : /* The actual mapsize may be less since the geo.upper may be changed by other process. So, avoids remapping until it necessary. */ mapped_pgno, true); } static int mdbx_meta_unsteady(MDBX_env *env, const txnid_t last_steady, MDBX_meta *const meta) { const uint64_t wipe = MDBX_DATASIGN_NONE; if (unlikely(META_IS_STEADY(meta)) && mdbx_meta_txnid_stable(env, meta) <= last_steady) { mdbx_warning("wipe txn #%" PRIaTXN ", meta %" PRIaPGNO, last_steady, data_page(meta)->mp_pgno); if (env->me_flags & MDBX_WRITEMAP) unaligned_poke_u64(4, meta->mm_datasync_sign, wipe); else return mdbx_pwrite(env->me_lazy_fd, &wipe, sizeof(meta->mm_datasync_sign), (uint8_t *)&meta->mm_datasync_sign - env->me_map); } return MDBX_SUCCESS; } __cold static int mdbx_wipe_steady(MDBX_env *env, const txnid_t last_steady) { int err = mdbx_meta_unsteady(env, last_steady, METAPAGE(env, 0)); if (unlikely(err != MDBX_SUCCESS)) return err; err = mdbx_meta_unsteady(env, last_steady, METAPAGE(env, 1)); if (unlikely(err != MDBX_SUCCESS)) return err; err = mdbx_meta_unsteady(env, last_steady, METAPAGE(env, 2)); if (unlikely(err != MDBX_SUCCESS)) return err; if (env->me_flags & MDBX_WRITEMAP) { mdbx_flush_incoherent_cpu_writeback(); err = mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, NUM_METAS), false); if (unlikely(err != MDBX_SUCCESS)) return err; } else { #if MDBX_USE_SYNCFILERANGE static bool syncfilerange_unavailable; if (likely(!syncfilerange_unavailable)) { if (likely(!sync_file_range( env->me_lazy_fd, 0, pgno2bytes(env, NUM_METAS), SYNC_FILE_RANGE_WRITE | SYNC_FILE_RANGE_WAIT_AFTER))) goto done_filesync; err = errno; if (ignore_enosys(err) != MDBX_RESULT_TRUE) return err; syncfilerange_unavailable = true; } #endif /* MDBX_USE_SYNCFILERANGE */ err = mdbx_fsync(env->me_lazy_fd, MDBX_SYNC_DATA); if (unlikely(err != MDBX_SUCCESS)) return err; #if MDBX_USE_SYNCFILERANGE done_filesync: #endif /* MDBX_USE_SYNCFILERANGE */ mdbx_flush_incoherent_mmap(env->me_map, pgno2bytes(env, NUM_METAS), env->me_os_psize); } if (likely(env->me_lck)) /* force oldest refresh */ atomic_store32(&env->me_lck->mti_readers_refresh_flag, true, mo_AcquireRelease); return MDBX_SUCCESS; } static __inline txnid_t pp_txnid4chk(const MDBX_page *mp, const MDBX_txn *txn) { return IS_DIRTY(mp) ? txn->mt_txnid - 1 : (/* maybe zero in legacy DB */ mp->mp_txnid ? mp->mp_txnid : MIN_TXNID); } static __inline txnid_t pp_txnid2chk(const MDBX_txn *txn) { #ifdef MDBX_DEBUG_LEGACY if (txn->mt_txnid < 2222) return 0; #endif return txn->mt_txnid; } /* Allocate page numbers and memory for writing. Maintain mt_last_reclaimed, * mt_reclaimed_pglist and mt_next_pgno. Set MDBX_TXN_ERROR on failure. * * If there are free pages available from older transactions, they * are re-used first. Otherwise allocate a new page at mt_next_pgno. * Do not modify the GC, just merge GC records into mt_reclaimed_pglist * and move mt_last_reclaimed to say which records were consumed. Only this * function can create mt_reclaimed_pglist and move * mt_last_reclaimed/mt_next_pgno. * * [in] mc cursor A cursor handle identifying the transaction and * database for which we are allocating. * [in] num the number of pages to allocate. * [out] mp Address of the allocated page(s). Requests for multiple pages * will always be satisfied by a single contiguous chunk of memory. * * Returns 0 on success, non-zero on failure.*/ #define MDBX_ALLOC_CACHE 1 #define MDBX_ALLOC_GC 2 #define MDBX_ALLOC_NEW 4 #define MDBX_ALLOC_ALL (MDBX_ALLOC_CACHE | MDBX_ALLOC_GC | MDBX_ALLOC_NEW) __hot static int mdbx_page_alloc(MDBX_cursor *mc, const unsigned num, MDBX_page **const mp, int flags) { int rc; MDBX_txn *const txn = mc->mc_txn; MDBX_env *const env = txn->mt_env; MDBX_page *np; const unsigned coalesce_threshold = env->me_maxgc_ov1page - env->me_maxgc_ov1page / 4; if (likely(flags & MDBX_ALLOC_GC)) { flags |= env->me_flags & (MDBX_COALESCE | MDBX_LIFORECLAIM); if (MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) > coalesce_threshold) flags &= ~MDBX_COALESCE; if (unlikely( /* If mc is updating the GC, then the retired-list cannot play catch-up with itself by growing while trying to save it. */ (mc->mc_flags & C_RECLAIMING) || /* avoid (recursive) search inside empty tree and while tree is updating, https://github.com/erthink/libmdbx/issues/31 */ txn->mt_dbs[FREE_DBI].md_entries == 0 || /* If our dirty list is already full, we can't touch GC */ (txn->tw.dirtyroom < txn->mt_dbs[FREE_DBI].md_depth && !(txn->mt_dbistate[FREE_DBI] & DBI_DIRTY)))) flags &= ~(MDBX_ALLOC_GC | MDBX_COALESCE); } if (likely(num == 1 && (flags & MDBX_ALLOC_CACHE) != 0)) { /* If there are any loose pages, just use them */ mdbx_assert(env, mp && num); if (likely(txn->tw.loose_pages)) { #if MDBX_ENABLE_REFUND if (txn->tw.loose_refund_wl > txn->mt_next_pgno) { mdbx_refund(txn); if (unlikely(!txn->tw.loose_pages)) goto no_loose; } #endif /* MDBX_ENABLE_REFUND */ np = txn->tw.loose_pages; txn->tw.loose_pages = np->mp_next; txn->tw.loose_count--; mdbx_debug("db %d use loose page %" PRIaPGNO, DDBI(mc), np->mp_pgno); mdbx_tassert(txn, np->mp_pgno < txn->mt_next_pgno); mdbx_ensure(env, np->mp_pgno >= NUM_METAS); VALGRIND_MAKE_MEM_UNDEFINED(page_data(np), page_space(txn->mt_env)); ASAN_UNPOISON_MEMORY_REGION(page_data(np), page_space(txn->mt_env)); np->mp_flags = P_DIRTY; np->mp_txnid = INVALID_TXNID; *mp = np; return MDBX_SUCCESS; } } #if MDBX_ENABLE_REFUND no_loose: #endif /* MDBX_ENABLE_REFUND */ mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); pgno_t pgno, *re_list = txn->tw.reclaimed_pglist; unsigned range_begin = 0, re_len = MDBX_PNL_SIZE(re_list); txnid_t oldest = 0, last = 0; const unsigned wanna_range = num - 1; while (true) { /* hsr-kick retry loop */ MDBX_cursor_couple recur; for (MDBX_cursor_op op = MDBX_FIRST;; op = (flags & MDBX_LIFORECLAIM) ? MDBX_PREV : MDBX_NEXT) { MDBX_val key, data; /* Seek a big enough contiguous page range. * Prefer pages with lower pgno. */ mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno)); if ((flags & (MDBX_COALESCE | MDBX_ALLOC_CACHE)) == MDBX_ALLOC_CACHE && re_len > wanna_range) { mdbx_tassert(txn, MDBX_PNL_LAST(re_list) < txn->mt_next_pgno && MDBX_PNL_FIRST(re_list) < txn->mt_next_pgno); range_begin = MDBX_PNL_ASCENDING ? 1 : re_len; pgno = MDBX_PNL_LEAST(re_list); if (likely(wanna_range == 0)) goto done; #if MDBX_PNL_ASCENDING mdbx_tassert(txn, pgno == re_list[1] && range_begin == 1); while (true) { unsigned range_end = range_begin + wanna_range; if (re_list[range_end] - pgno == wanna_range) goto done; if (range_end == re_len) break; pgno = re_list[++range_begin]; } #else mdbx_tassert(txn, pgno == re_list[re_len] && range_begin == re_len); while (true) { if (re_list[range_begin - wanna_range] - pgno == wanna_range) goto done; if (range_begin == wanna_range) break; pgno = re_list[--range_begin]; } #endif /* MDBX_PNL sort-order */ } if (op == MDBX_FIRST) { /* 1st iteration, setup cursor, etc */ if (unlikely(!(flags & MDBX_ALLOC_GC))) break /* reclaiming is prohibited for now */; /* Prepare to fetch more and coalesce */ oldest = (flags & MDBX_LIFORECLAIM) ? mdbx_find_oldest(txn) : atomic_load64(env->me_oldest, mo_AcquireRelease); rc = mdbx_cursor_init(&recur.outer, txn, FREE_DBI); if (unlikely(rc != MDBX_SUCCESS)) goto fail; if (flags & MDBX_LIFORECLAIM) { /* Begin from oldest reader if any */ if (oldest > MIN_TXNID) { last = oldest - 1; op = MDBX_SET_RANGE; } } else if (txn->tw.last_reclaimed) { /* Continue lookup from txn->tw.last_reclaimed to oldest reader */ last = txn->tw.last_reclaimed; op = MDBX_SET_RANGE; } key.iov_base = &last; key.iov_len = sizeof(last); } if (!(flags & MDBX_LIFORECLAIM)) { /* Do not try fetch more if the record will be too recent */ if (op != MDBX_FIRST && ++last >= oldest) { oldest = mdbx_find_oldest(txn); if (oldest <= last) break; } } rc = mdbx_cursor_get(&recur.outer, &key, NULL, op); if (rc == MDBX_NOTFOUND && (flags & MDBX_LIFORECLAIM)) { if (op == MDBX_SET_RANGE) continue; txnid_t snap = mdbx_find_oldest(txn); if (oldest < snap) { oldest = snap; last = oldest - 1; key.iov_base = &last; key.iov_len = sizeof(last); op = MDBX_SET_RANGE; rc = mdbx_cursor_get(&recur.outer, &key, NULL, op); } } if (unlikely(rc)) { if (rc == MDBX_NOTFOUND) break; goto fail; } if (unlikely(key.iov_len != sizeof(txnid_t))) { rc = MDBX_CORRUPTED; goto fail; } last = unaligned_peek_u64(4, key.iov_base); if (unlikely(last < MIN_TXNID || last > MAX_TXNID)) { rc = MDBX_CORRUPTED; goto fail; } if (oldest <= last) { oldest = mdbx_find_oldest(txn); if (oldest <= last) { if (flags & MDBX_LIFORECLAIM) continue; break; } } if (flags & MDBX_LIFORECLAIM) { /* skip IDs of records that already reclaimed */ if (txn->tw.lifo_reclaimed) { size_t i; for (i = (size_t)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed); i > 0; --i) if (txn->tw.lifo_reclaimed[i] == last) break; if (i) continue; } } /* Reading next GC record */ np = recur.outer.mc_pg[recur.outer.mc_top]; if (unlikely((rc = mdbx_node_read( &recur.outer, page_node(np, recur.outer.mc_ki[recur.outer.mc_top]), &data, pp_txnid4chk(np, txn))) != MDBX_SUCCESS)) goto fail; if ((flags & MDBX_LIFORECLAIM) && !txn->tw.lifo_reclaimed) { txn->tw.lifo_reclaimed = mdbx_txl_alloc(); if (unlikely(!txn->tw.lifo_reclaimed)) { rc = MDBX_ENOMEM; goto fail; } } /* Append PNL from GC record to tw.reclaimed_pglist */ mdbx_cassert(mc, (mc->mc_flags & C_GCFREEZE) == 0); pgno_t *gc_pnl = (pgno_t *)data.iov_base; mdbx_tassert(txn, data.iov_len >= MDBX_PNL_SIZEOF(gc_pnl)); if (unlikely(data.iov_len < MDBX_PNL_SIZEOF(gc_pnl) || !mdbx_pnl_check(gc_pnl, txn->mt_next_pgno))) { rc = MDBX_CORRUPTED; goto fail; } const unsigned gc_len = MDBX_PNL_SIZE(gc_pnl); if (unlikely(/* resulting list is tool long */ gc_len + MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) > env->me_options.rp_augment_limit) && (((/* not a slot-request from gc-update */ mp || (flags & MDBX_LIFORECLAIM) == 0 || (txn->tw.lifo_reclaimed && MDBX_PNL_SIZE(txn->tw.lifo_reclaimed))) && /* have enough unallocated space */ pgno_add( txn->mt_next_pgno, num) <= txn->mt_geo.upper) || gc_len + MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) >= MDBX_PGL_LIMIT / 16 * 15)) { /* Stop reclaiming to avoid overflow the page list. * This is a rare case while search for a continuously multi-page region * in a large database. https://github.com/erthink/libmdbx/issues/123 */ mdbx_debug("stop reclaiming to avoid PNL overflow: %u (current) + %u " "(chunk) -> %u", MDBX_PNL_SIZE(txn->tw.reclaimed_pglist), gc_len, gc_len + MDBX_PNL_SIZE(txn->tw.reclaimed_pglist)); flags &= ~(MDBX_ALLOC_GC | MDBX_COALESCE); break; } rc = mdbx_pnl_need(&txn->tw.reclaimed_pglist, gc_len); if (unlikely(rc != MDBX_SUCCESS)) goto fail; re_list = txn->tw.reclaimed_pglist; /* Remember ID of GC record */ if (flags & MDBX_LIFORECLAIM) { rc = mdbx_txl_append(&txn->tw.lifo_reclaimed, last); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } txn->tw.last_reclaimed = last; if (mdbx_log_enabled(MDBX_LOG_EXTRA)) { mdbx_debug_extra("PNL read txn %" PRIaTXN " root %" PRIaPGNO " num %u, PNL", last, txn->mt_dbs[FREE_DBI].md_root, gc_len); for (unsigned i = gc_len; i; i--) mdbx_debug_extra_print(" %" PRIaPGNO, gc_pnl[i]); mdbx_debug_extra_print("%s\n", "."); } /* Merge in descending sorted order */ const unsigned prev_re_len = MDBX_PNL_SIZE(re_list); mdbx_pnl_xmerge(re_list, gc_pnl); /* re-check to avoid duplicates */ if (unlikely(!mdbx_pnl_check(re_list, txn->mt_next_pgno))) { rc = MDBX_CORRUPTED; goto fail; } mdbx_tassert(txn, mdbx_dirtylist_check(txn)); re_len = MDBX_PNL_SIZE(re_list); mdbx_tassert(txn, re_len == 0 || re_list[re_len] < txn->mt_next_pgno); if (MDBX_ENABLE_REFUND && re_len && unlikely(MDBX_PNL_MOST(re_list) == txn->mt_next_pgno - 1)) { /* Refund suitable pages into "unallocated" space */ mdbx_refund(txn); re_list = txn->tw.reclaimed_pglist; re_len = MDBX_PNL_SIZE(re_list); } /* Done for a kick-reclaim mode, actually no page needed */ if (unlikely((flags & MDBX_ALLOC_CACHE) == 0)) return MDBX_SUCCESS; /* Don't try to coalesce too much. */ if (re_len /* current size */ > coalesce_threshold || (re_len > prev_re_len && re_len - prev_re_len /* delta from prev */ >= coalesce_threshold / 2)) flags &= ~MDBX_COALESCE; } if (F_ISSET(flags, MDBX_COALESCE | MDBX_ALLOC_CACHE)) { flags -= MDBX_COALESCE; continue; } /* There is no suitable pages in the GC and to be able to allocate * we should CHOICE one of: * - make a new steady checkpoint if reclaiming was stopped by * the last steady-sync, or wipe it in the MDBX_UTTERLY_NOSYNC mode; * - kick lagging reader(s) if reclaiming was stopped by ones of it. * - extend the database file. */ /* Will use new pages from the map if nothing is suitable in the GC. */ range_begin = 0; pgno = txn->mt_next_pgno; const pgno_t next = pgno_add(pgno, num); if (flags & MDBX_ALLOC_GC) { const MDBX_meta *const head = mdbx_meta_head(env); MDBX_meta *const steady = mdbx_meta_steady(env); /* does reclaiming stopped at the last steady point? */ if (head != steady && META_IS_STEADY(steady) && oldest == mdbx_meta_txnid_stable(env, steady)) { mdbx_debug("gc-kick-steady: head %" PRIaTXN "-%s, tail %" PRIaTXN "-%s, oldest %" PRIaTXN, mdbx_meta_txnid_stable(env, head), mdbx_durable_str(head), mdbx_meta_txnid_stable(env, steady), mdbx_durable_str(steady), oldest); rc = MDBX_RESULT_TRUE; const pgno_t autosync_threshold = atomic_load32(env->me_autosync_threshold, mo_Relaxed); const uint64_t autosync_period = atomic_load64(env->me_autosync_period, mo_Relaxed); /* wipe the last steady-point if one of: * - UTTERLY_NOSYNC mode AND auto-sync threshold is NOT specified * - UTTERLY_NOSYNC mode AND free space at steady-point is exhausted * otherwise, make a new steady-point if one of: * - auto-sync threshold is specified and reached; * - upper limit of database size is reached; * - database is full (with the current file size) * AND auto-sync threshold it NOT specified */ if (F_ISSET(env->me_flags, MDBX_UTTERLY_NOSYNC) && ((autosync_threshold | autosync_period) == 0 || next >= steady->mm_geo.now)) { /* wipe steady checkpoint in MDBX_UTTERLY_NOSYNC mode * without any auto-sync threshold(s). */ rc = mdbx_wipe_steady(env, oldest); mdbx_debug("gc-wipe-steady, rc %d", rc); mdbx_assert(env, steady != mdbx_meta_steady(env)); } else if ((flags & MDBX_ALLOC_NEW) == 0 || (autosync_threshold && atomic_load32(env->me_unsynced_pages, mo_Relaxed) >= autosync_threshold) || (autosync_period && mdbx_osal_monotime() - atomic_load64(env->me_sync_timestamp, mo_Relaxed) >= autosync_period) || next >= txn->mt_geo.upper || (next >= txn->mt_end_pgno && (autosync_threshold | autosync_period) == 0)) { /* make steady checkpoint. */ MDBX_meta meta = *head; rc = mdbx_sync_locked(env, env->me_flags & MDBX_WRITEMAP, &meta); mdbx_debug("gc-make-steady, rc %d", rc); mdbx_assert(env, steady != mdbx_meta_steady(env)); } if (rc == MDBX_SUCCESS) { if (mdbx_find_oldest(txn) > oldest) continue; /* it is reasonable check/kick lagging reader(s) here, * since we made a new steady point or wipe the last. */ if (oldest < txn->mt_txnid - MDBX_TXNID_STEP && mdbx_kick_longlived_readers(env, oldest) > oldest) continue; } else if (unlikely(rc != MDBX_RESULT_TRUE)) goto fail; } } /* don't kick lagging reader(s) if is enough unallocated space * at the end of database file. */ if ((flags & MDBX_ALLOC_NEW) && next <= txn->mt_end_pgno) goto done; if ((flags & MDBX_ALLOC_GC) && oldest < txn->mt_txnid - MDBX_TXNID_STEP && mdbx_kick_longlived_readers(env, oldest) > oldest) continue; rc = MDBX_NOTFOUND; if (flags & MDBX_ALLOC_NEW) { rc = MDBX_MAP_FULL; if (next <= txn->mt_geo.upper && txn->mt_geo.grow) { mdbx_assert(env, next > txn->mt_end_pgno); pgno_t aligned = pgno_align2os_pgno( env, pgno_add(next, txn->mt_geo.grow - next % txn->mt_geo.grow)); if (aligned > txn->mt_geo.upper) aligned = txn->mt_geo.upper; mdbx_assert(env, aligned > txn->mt_end_pgno); mdbx_verbose("try growth datafile to %" PRIaPGNO " pages (+%" PRIaPGNO ")", aligned, aligned - txn->mt_end_pgno); rc = mdbx_mapresize_implicit(env, txn->mt_next_pgno, aligned, txn->mt_geo.upper); if (rc == MDBX_SUCCESS) { env->me_txn->mt_end_pgno = aligned; goto done; } mdbx_error("unable growth datafile to %" PRIaPGNO " pages (+%" PRIaPGNO "), errcode %d", aligned, aligned - txn->mt_end_pgno, rc); } else { mdbx_debug("gc-alloc: next %u > upper %u", next, txn->mt_geo.upper); } } fail: mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); if (likely(mp)) { *mp = nullptr; txn->mt_flags |= MDBX_TXN_ERROR; } mdbx_assert(env, rc != MDBX_SUCCESS); return rc; } done: if (unlikely(mp == nullptr)) return MDBX_SUCCESS; if (unlikely(txn->tw.dirtyroom < 1)) return MDBX_TXN_FULL; mdbx_ensure(env, pgno >= NUM_METAS); if (env->me_flags & MDBX_WRITEMAP) { np = pgno2page(env, pgno); /* LY: reset no-access flag from mdbx_page_loose() */ VALGRIND_MAKE_MEM_UNDEFINED(np, pgno2bytes(env, num)); ASAN_UNPOISON_MEMORY_REGION(np, pgno2bytes(env, num)); } else { if (unlikely(!(np = mdbx_page_malloc(txn, num)))) { rc = MDBX_ENOMEM; goto fail; } } if (range_begin) { mdbx_cassert(mc, (mc->mc_flags & C_GCFREEZE) == 0); mdbx_tassert(txn, pgno < txn->mt_next_pgno); mdbx_tassert(txn, pgno == re_list[range_begin]); /* Cutoff allocated pages from tw.reclaimed_pglist */ #if MDBX_PNL_ASCENDING for (unsigned i = range_begin + num; i <= re_len;) re_list[range_begin++] = re_list[i++]; MDBX_PNL_SIZE(re_list) = re_len = range_begin - 1; #else MDBX_PNL_SIZE(re_list) = re_len -= num; for (unsigned i = range_begin - num; i < re_len;) re_list[++i] = re_list[++range_begin]; #endif mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); } else { txn->mt_next_pgno = pgno + num; mdbx_assert(env, txn->mt_next_pgno <= txn->mt_end_pgno); } if (unlikely(env->me_flags & MDBX_PAGEPERTURB)) memset(np, -1, pgno2bytes(env, num)); VALGRIND_MAKE_MEM_UNDEFINED(np, pgno2bytes(env, num)); np->mp_pgno = pgno; np->mp_leaf2_ksize = 0; np->mp_flags = 0; if ((mdbx_assert_enabled() || mdbx_audit_enabled()) && num > 1) { np->mp_pages = num; np->mp_flags = P_OVERFLOW; } rc = mdbx_page_dirty(txn, np); if (unlikely(rc != MDBX_SUCCESS)) goto fail; *mp = np; mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); return MDBX_SUCCESS; } /* Copy the used portions of a non-overflow page. */ __hot static void mdbx_page_copy(MDBX_page *dst, MDBX_page *src, size_t psize) { STATIC_ASSERT(UINT16_MAX > MAX_PAGESIZE - PAGEHDRSZ); STATIC_ASSERT(MIN_PAGESIZE > PAGEHDRSZ + NODESIZE * 4); if (!IS_LEAF2(src)) { size_t upper = src->mp_upper, lower = src->mp_lower, unused = upper - lower; /* If page isn't full, just copy the used portion. Adjust * alignment so memcpy may copy words instead of bytes. */ if (unused >= MDBX_CACHELINE_SIZE * 2) { lower = ceil_powerof2(lower + PAGEHDRSZ, sizeof(void *)); upper = floor_powerof2(upper + PAGEHDRSZ, sizeof(void *)); memcpy(dst, src, lower); dst = (void *)((char *)dst + upper); src = (void *)((char *)src + upper); psize -= upper; } } memcpy(dst, src, psize); } /* Pull a page off the txn's spill list, if present. * * If a page being referenced was spilled to disk in this txn, bring * it back and make it dirty/writable again. * * [in] txn the transaction handle. * [in] mp the page being referenced. It must not be dirty. * [out] ret the writable page, if any. * ret is unchanged if mp wasn't spilled. */ static int __must_check_result mdbx_page_unspill(MDBX_txn *txn, MDBX_page *mp, MDBX_page **ret) { mdbx_tassert(txn, !IS_DIRTY(mp)); const pgno_t spilled_pgno = mp->mp_pgno << 1; const MDBX_txn *scan = txn; do { if ((scan->mt_flags & MDBX_TXN_SPILLS) == 0) break; if (!scan->tw.spill_pages) continue; const unsigned si = mdbx_pnl_exist(scan->tw.spill_pages, spilled_pgno); if (!si) continue; const unsigned npages = IS_OVERFLOW(mp) ? mp->mp_pages : 1; MDBX_page *np = mp; if ((txn->mt_flags & MDBX_WRITEMAP) == 0) { np = mdbx_page_malloc(txn, npages); if (unlikely(!np)) return MDBX_ENOMEM; if (likely(npages == 1)) mdbx_page_copy(np, mp, txn->mt_env->me_psize); else memcpy(np, mp, pgno2bytes(txn->mt_env, npages)); } mdbx_debug("unspill page %" PRIaPGNO, mp->mp_pgno); if (scan == txn) { /* If in current txn, this page is no longer spilled. * If it happens to be the last page, truncate the spill list. * Otherwise mark it as deleted by setting the LSB. */ mdbx_spill_remove(txn, si, npages); } /* otherwise, if belonging to a parent txn, the * page remains spilled until child commits */ int rc = mdbx_page_dirty(txn, np); if (unlikely(rc != MDBX_SUCCESS)) { if ((txn->mt_flags & MDBX_WRITEMAP) == 0) mdbx_dpage_free(txn->mt_env, np, npages); return rc; } *ret = np; break; } while ((scan = scan->mt_parent) != nullptr); return MDBX_SUCCESS; } /* Touch a page: make it dirty and re-insert into tree with updated pgno. * Set MDBX_TXN_ERROR on failure. * * [in] mc cursor pointing to the page to be touched * * Returns 0 on success, non-zero on failure. */ __hot static int mdbx_page_touch(MDBX_cursor *mc) { MDBX_page *const mp = mc->mc_pg[mc->mc_top], *np; MDBX_txn *txn = mc->mc_txn; MDBX_cursor *m2, *m3; pgno_t pgno; int rc; if (mdbx_assert_enabled()) { if (mc->mc_dbi >= CORE_DBS) { if (mc->mc_flags & C_SUB) { MDBX_xcursor *mx = container_of(mc->mc_db, MDBX_xcursor, mx_db); MDBX_cursor_couple *couple = container_of(mx, MDBX_cursor_couple, inner); mdbx_cassert(mc, mc->mc_db == &couple->outer.mc_xcursor->mx_db); mdbx_cassert(mc, mc->mc_dbx == &couple->outer.mc_xcursor->mx_dbx); mdbx_cassert(mc, *couple->outer.mc_dbistate & DBI_DIRTY); } else { mdbx_cassert(mc, *mc->mc_dbistate & DBI_DIRTY); } mdbx_cassert(mc, mc->mc_txn->mt_flags & MDBX_TXN_DIRTY); } mdbx_cassert(mc, !IS_OVERFLOW(mp)); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); } if (!IS_DIRTY(mp)) { if (txn->mt_flags & MDBX_TXN_SPILLS) { np = NULL; rc = mdbx_page_unspill(txn, mp, &np); if (unlikely(rc)) goto fail; if (likely(np)) goto done; } if (unlikely((rc = mdbx_pnl_need(&txn->tw.retired_pages, 1)) || (rc = mdbx_page_alloc(mc, 1, &np, MDBX_ALLOC_ALL)))) goto fail; pgno = np->mp_pgno; mdbx_debug("touched db %d page %" PRIaPGNO " -> %" PRIaPGNO, DDBI(mc), mp->mp_pgno, pgno); mdbx_cassert(mc, mp->mp_pgno != pgno); mdbx_pnl_xappend(txn->tw.retired_pages, mp->mp_pgno); /* Update the parent page, if any, to point to the new page */ if (mc->mc_top) { MDBX_page *parent = mc->mc_pg[mc->mc_top - 1]; MDBX_node *node = page_node(parent, mc->mc_ki[mc->mc_top - 1]); node_set_pgno(node, pgno); } else { mc->mc_db->md_root = pgno; } } else if (txn->mt_parent && !IS_SUBP(mp)) { mdbx_tassert(txn, (txn->mt_env->me_flags & MDBX_WRITEMAP) == 0); pgno = mp->mp_pgno; /* If txn has a parent, make sure the page is in our dirty list. */ const MDBX_page *const dp = mdbx_dpl_find(txn->tw.dirtylist, pgno); if (dp) { if (unlikely(mp != dp)) { /* bad cursor? */ mdbx_error("wrong page 0x%p #%" PRIaPGNO " in the dirtylist, expecting %p", __Wpedantic_format_voidptr(dp), pgno, __Wpedantic_format_voidptr(mp)); mc->mc_flags &= ~(C_INITIALIZED | C_EOF); rc = MDBX_PROBLEM; goto fail; } return MDBX_SUCCESS; } mdbx_debug("clone db %d page %" PRIaPGNO, DDBI(mc), mp->mp_pgno); mdbx_cassert(mc, txn->tw.dirtylist->length <= MDBX_PGL_LIMIT + MDBX_PNL_GRANULATE); /* No - copy it */ np = mdbx_page_malloc(txn, 1); if (unlikely(!np)) { rc = MDBX_ENOMEM; goto fail; } /* insert a clone of parent's dirty page, so don't touch dirtyroom */ rc = mdbx_dpl_append(txn, pgno, np); if (unlikely(rc)) { mdbx_dpage_free(txn->mt_env, np, 1); goto fail; } np->mp_pgno = pgno; np->mp_txnid = INVALID_TXNID; np->mp_flags |= P_DIRTY; mdbx_tassert(txn, mdbx_dirtylist_check(txn)); } else { return MDBX_SUCCESS; } mdbx_page_copy(np, mp, txn->mt_env->me_psize); np->mp_pgno = pgno; np->mp_txnid = INVALID_TXNID; np->mp_flags |= P_DIRTY; done: /* Adjust cursors pointing to mp */ mc->mc_pg[mc->mc_top] = np; m2 = txn->tw.cursors[mc->mc_dbi]; if (mc->mc_flags & C_SUB) { for (; m2; m2 = m2->mc_next) { m3 = &m2->mc_xcursor->mx_cursor; if (m3->mc_snum < mc->mc_snum) continue; if (m3->mc_pg[mc->mc_top] == mp) m3->mc_pg[mc->mc_top] = np; } } else { for (; m2; m2 = m2->mc_next) { if (m2->mc_snum < mc->mc_snum) continue; if (m2 == mc) continue; if (m2->mc_pg[mc->mc_top] == mp) { m2->mc_pg[mc->mc_top] = np; if (XCURSOR_INITED(m2) && IS_LEAF(np)) XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]); } } } return MDBX_SUCCESS; fail: txn->mt_flags |= MDBX_TXN_ERROR; return rc; } __cold static int mdbx_env_sync_internal(MDBX_env *env, bool force, bool nonblock) { unsigned flags = env->me_flags & ~MDBX_NOMETASYNC; if (unlikely(flags & (MDBX_RDONLY | MDBX_FATAL_ERROR))) return MDBX_EACCESS; if (unlikely(!env->me_map)) return MDBX_EPERM; int rc = MDBX_RESULT_TRUE /* means "nothing to sync" */; bool need_unlock = false; if (nonblock && atomic_load32(env->me_unsynced_pages, mo_AcquireRelease) == 0) goto fastpath; const bool outside_txn = (env->me_txn0->mt_owner != mdbx_thread_self()); if (outside_txn) { int err = mdbx_txn_lock(env, nonblock); if (unlikely(err != MDBX_SUCCESS)) return err; need_unlock = true; } const MDBX_meta *head = mdbx_meta_head(env); pgno_t unsynced_pages = atomic_load32(env->me_unsynced_pages, mo_Relaxed); if (!META_IS_STEADY(head) || unsynced_pages) { const pgno_t autosync_threshold = atomic_load32(env->me_autosync_threshold, mo_Relaxed); const uint64_t autosync_period = atomic_load64(env->me_autosync_period, mo_Relaxed); if (force || (autosync_threshold && unsynced_pages >= autosync_threshold) || (autosync_period && mdbx_osal_monotime() - atomic_load64(env->me_sync_timestamp, mo_Relaxed) >= autosync_period)) flags &= MDBX_WRITEMAP /* clear flags for full steady sync */; if (outside_txn) { if (unsynced_pages > /* FIXME: define threshold */ 16 && (flags & MDBX_SAFE_NOSYNC) == 0) { mdbx_assert(env, ((flags ^ env->me_flags) & MDBX_WRITEMAP) == 0); const size_t usedbytes = pgno_align2os_bytes(env, head->mm_geo.next); mdbx_txn_unlock(env); /* LY: pre-sync without holding lock to reduce latency for writer(s) */ int err = (flags & MDBX_WRITEMAP) ? mdbx_msync(&env->me_dxb_mmap, 0, usedbytes, false) : mdbx_fsync(env->me_lazy_fd, MDBX_SYNC_DATA); if (unlikely(err != MDBX_SUCCESS)) return err; err = mdbx_txn_lock(env, nonblock); if (unlikely(err != MDBX_SUCCESS)) return err; /* LY: head and unsynced_pages may be changed. */ head = mdbx_meta_head(env); unsynced_pages = atomic_load32(env->me_unsynced_pages, mo_Relaxed); } env->me_txn0->mt_txnid = meta_txnid(env, head, false); mdbx_find_oldest(env->me_txn0); rc = MDBX_RESULT_FALSE /* means "some data was synced" */; } if (!META_IS_STEADY(head) || ((flags & MDBX_SAFE_NOSYNC) == 0 && unsynced_pages)) { mdbx_debug("meta-head %" PRIaPGNO ", %s, sync_pending %" PRIaPGNO, data_page(head)->mp_pgno, mdbx_durable_str(head), unsynced_pages); MDBX_meta meta = *head; int err = mdbx_sync_locked(env, flags | MDBX_SHRINK_ALLOWED, &meta); if (unlikely(err != MDBX_SUCCESS)) { if (need_unlock) mdbx_txn_unlock(env); return err; } rc = MDBX_RESULT_FALSE /* means "some data was synced" */; } } fastpath: /* LY: sync meta-pages if MDBX_NOMETASYNC enabled * and someone was not synced above. */ if (rc == MDBX_RESULT_TRUE && (env->me_flags & MDBX_NOMETASYNC) != 0) { const txnid_t head_txnid = mdbx_recent_committed_txnid(env); if (atomic_load32(env->me_meta_sync_txnid, mo_Relaxed) != (uint32_t)head_txnid) { rc = (flags & MDBX_WRITEMAP) ? mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, NUM_METAS), false) : mdbx_fsync(env->me_lazy_fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ); if (likely(rc == MDBX_SUCCESS)) atomic_store32(env->me_meta_sync_txnid, (uint32_t)head_txnid, mo_Relaxed); } } if (need_unlock) mdbx_txn_unlock(env); return rc; } static __inline int check_env(const MDBX_env *env) { if (unlikely(!env)) return MDBX_EINVAL; if (unlikely(env->me_signature.weak != MDBX_ME_SIGNATURE)) return MDBX_EBADSIGN; #if MDBX_ENV_CHECKPID if (unlikely(env->me_pid != mdbx_getpid())) { ((MDBX_env *)env)->me_flags |= MDBX_FATAL_ERROR; return MDBX_PANIC; } #endif /* MDBX_ENV_CHECKPID */ if (unlikely(env->me_flags & MDBX_FATAL_ERROR)) return MDBX_PANIC; return MDBX_SUCCESS; } __cold int mdbx_env_sync_ex(MDBX_env *env, bool force, bool nonblock) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_env_sync_internal(env, force, nonblock); } __cold int mdbx_env_sync(MDBX_env *env) { return __inline_mdbx_env_sync(env); } __cold int mdbx_env_sync_poll(MDBX_env *env) { return __inline_mdbx_env_sync_poll(env); } /* Back up parent txn's cursors, then grab the originals for tracking */ static int mdbx_cursor_shadow(MDBX_txn *parent, MDBX_txn *nested) { for (int i = parent->mt_numdbs; --i >= 0;) { nested->tw.cursors[i] = NULL; MDBX_cursor *mc = parent->tw.cursors[i]; if (mc != NULL) { size_t size = mc->mc_xcursor ? sizeof(MDBX_cursor) + sizeof(MDBX_xcursor) : sizeof(MDBX_cursor); for (MDBX_cursor *bk; mc; mc = bk->mc_next) { bk = mc; if (mc->mc_signature != MDBX_MC_LIVE) continue; bk = mdbx_malloc(size); if (unlikely(!bk)) return MDBX_ENOMEM; *bk = *mc; mc->mc_backup = bk; /* Kill pointers into src to reduce abuse: The * user may not use mc until dst ends. But we need a valid * txn pointer here for cursor fixups to keep working. */ mc->mc_txn = nested; mc->mc_db = &nested->mt_dbs[i]; mc->mc_dbistate = &nested->mt_dbistate[i]; MDBX_xcursor *mx = mc->mc_xcursor; if (mx != NULL) { *(MDBX_xcursor *)(bk + 1) = *mx; mx->mx_cursor.mc_txn = nested; } mc->mc_next = nested->tw.cursors[i]; nested->tw.cursors[i] = mc; } } } return MDBX_SUCCESS; } /* Close this write txn's cursors, give parent txn's cursors back to parent. * * [in] txn the transaction handle. * [in] merge true to keep changes to parent cursors, false to revert. * * Returns 0 on success, non-zero on failure. */ static void mdbx_cursors_eot(MDBX_txn *txn, const bool merge) { mdbx_tassert(txn, (txn->mt_flags & MDBX_TXN_RDONLY) == 0); for (int i = txn->mt_numdbs; --i >= 0;) { MDBX_cursor *next, *mc = txn->tw.cursors[i]; if (!mc) continue; txn->tw.cursors[i] = NULL; do { const unsigned stage = mc->mc_signature; MDBX_cursor *bk = mc->mc_backup; next = mc->mc_next; mdbx_ensure(txn->mt_env, stage == MDBX_MC_LIVE || (stage == MDBX_MC_WAIT4EOT && bk)); mdbx_cassert(mc, mc->mc_dbi == (unsigned)i); if (bk) { MDBX_xcursor *mx = mc->mc_xcursor; mdbx_cassert(mc, mx == bk->mc_xcursor); mdbx_tassert(txn, txn->mt_parent != NULL); mdbx_ensure(txn->mt_env, bk->mc_signature == MDBX_MC_LIVE); if (stage == MDBX_MC_WAIT4EOT /* Cursor was closed by user */) mc->mc_signature = stage /* Promote closed state to parent txn */; else if (merge) { /* Restore pointers to parent txn */ mc->mc_next = bk->mc_next; mc->mc_backup = bk->mc_backup; mc->mc_txn = bk->mc_txn; mc->mc_db = bk->mc_db; mc->mc_dbistate = bk->mc_dbistate; if (mx) { if (mx != bk->mc_xcursor) { *bk->mc_xcursor = *mx; mx = bk->mc_xcursor; } mx->mx_cursor.mc_txn = bk->mc_txn; } } else { /* Restore from backup, i.e. rollback/abort nested txn */ *mc = *bk; if (mx) *mx = *(MDBX_xcursor *)(bk + 1); } bk->mc_signature = 0; mdbx_free(bk); } else { mdbx_ensure(txn->mt_env, stage == MDBX_MC_LIVE); mc->mc_signature = MDBX_MC_READY4CLOSE /* Cursor may be reused */; mc->mc_flags = 0 /* reset C_UNTRACK */; } } while ((mc = next) != NULL); } } #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) /* Find largest mvcc-snapshot still referenced by this process. */ static pgno_t mdbx_find_largest_this(MDBX_env *env, pgno_t largest) { MDBX_lockinfo *const lck = env->me_lck; if (likely(lck != NULL /* exclusive mode */)) { const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { retry: if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease) == env->me_pid) { /* mdbx_jitter4testing(true); */ const pgno_t snap_pages = atomic_load32( &lck->mti_readers[i].mr_snapshot_pages_used, mo_Relaxed); const txnid_t snap_txnid = safe64_read(&lck->mti_readers[i].mr_txnid); if (unlikely( snap_pages != atomic_load32(&lck->mti_readers[i].mr_snapshot_pages_used, mo_AcquireRelease) || snap_txnid != safe64_read(&lck->mti_readers[i].mr_txnid))) goto retry; if (largest < snap_pages && atomic_load64(&lck->mti_oldest_reader, mo_AcquireRelease) <= /* ignore pending updates */ snap_txnid && snap_txnid <= MAX_TXNID) largest = snap_pages; } } } return largest; } static void mdbx_txn_valgrind(MDBX_env *env, MDBX_txn *txn) { #if !defined(__SANITIZE_ADDRESS__) if (!RUNNING_ON_VALGRIND) return; #endif if (txn) { /* transaction start */ if (env->me_poison_edge < txn->mt_next_pgno) env->me_poison_edge = txn->mt_next_pgno; VALGRIND_MAKE_MEM_DEFINED(env->me_map, pgno2bytes(env, txn->mt_next_pgno)); ASAN_UNPOISON_MEMORY_REGION(env->me_map, pgno2bytes(env, txn->mt_next_pgno)); /* don't touch more, it should be already poisoned */ } else { /* transaction end */ bool should_unlock = false; pgno_t last = MAX_PAGENO; if (env->me_txn0 && env->me_txn0->mt_owner == mdbx_thread_self()) { /* inside write-txn */ MDBX_meta *head = mdbx_meta_head(env); last = head->mm_geo.next; } else if (env->me_flags & MDBX_RDONLY) { /* read-only mode, no write-txn, no wlock mutex */ last = NUM_METAS; } else if (mdbx_txn_lock(env, true) == MDBX_SUCCESS) { /* no write-txn */ last = NUM_METAS; should_unlock = true; } else { /* write txn is running, therefore shouldn't poison any memory range */ return; } last = mdbx_find_largest_this(env, last); const pgno_t edge = env->me_poison_edge; if (edge > last) { mdbx_assert(env, last >= NUM_METAS); env->me_poison_edge = last; VALGRIND_MAKE_MEM_NOACCESS(env->me_map + pgno2bytes(env, last), pgno2bytes(env, edge - last)); ASAN_POISON_MEMORY_REGION(env->me_map + pgno2bytes(env, last), pgno2bytes(env, edge - last)); } if (should_unlock) mdbx_txn_unlock(env); } } #endif /* MDBX_USE_VALGRIND || __SANITIZE_ADDRESS__ */ typedef struct { int err; MDBX_reader *rslot; } bind_rslot_result; static bind_rslot_result bind_rslot(MDBX_env *env, const uintptr_t tid) { mdbx_assert(env, env->me_lck); mdbx_assert(env, env->me_lck->mti_magic_and_version == MDBX_LOCK_MAGIC); mdbx_assert(env, env->me_lck->mti_os_and_format == MDBX_LOCK_FORMAT); bind_rslot_result result = {mdbx_rdt_lock(env), nullptr}; if (unlikely(MDBX_IS_ERROR(result.err))) return result; if (unlikely(env->me_flags & MDBX_FATAL_ERROR)) { mdbx_rdt_unlock(env); result.err = MDBX_PANIC; return result; } if (unlikely(!env->me_map)) { mdbx_rdt_unlock(env); result.err = MDBX_EPERM; return result; } if (unlikely(env->me_live_reader != env->me_pid)) { result.err = mdbx_rpid_set(env); if (unlikely(result.err != MDBX_SUCCESS)) { mdbx_rdt_unlock(env); return result; } env->me_live_reader = env->me_pid; } result.err = MDBX_SUCCESS; unsigned slot, nreaders; while (1) { nreaders = atomic_load32(&env->me_lck->mti_numreaders, mo_Relaxed); for (slot = 0; slot < nreaders; slot++) if (atomic_load32(&env->me_lck->mti_readers[slot].mr_pid, mo_Relaxed) == 0) break; if (likely(slot < env->me_maxreaders)) break; result.err = mdbx_cleanup_dead_readers(env, true, NULL); if (result.err != MDBX_RESULT_TRUE) { mdbx_rdt_unlock(env); result.err = (result.err == MDBX_SUCCESS) ? MDBX_READERS_FULL : result.err; return result; } } result.rslot = &env->me_lck->mti_readers[slot]; /* Claim the reader slot, carefully since other code * uses the reader table un-mutexed: First reset the * slot, next publish it in lck->mti_numreaders. After * that, it is safe for mdbx_env_close() to touch it. * When it will be closed, we can finally claim it. */ atomic_store32(&result.rslot->mr_pid, 0, mo_Relaxed); safe64_reset(&result.rslot->mr_txnid, true); if (slot == nreaders) atomic_store32(&env->me_lck->mti_numreaders, ++nreaders, mo_Relaxed); atomic_store64(&result.rslot->mr_tid, (env->me_flags & MDBX_NOTLS) ? 0 : tid, mo_Relaxed); atomic_store32(&result.rslot->mr_pid, env->me_pid, mo_Relaxed); mdbx_rdt_unlock(env); if (likely(env->me_flags & MDBX_ENV_TXKEY)) { mdbx_assert(env, env->me_live_reader == env->me_pid); thread_rthc_set(env->me_txkey, result.rslot); } return result; } __cold int mdbx_thread_register(const MDBX_env *env) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!env->me_lck)) return (env->me_flags & MDBX_EXCLUSIVE) ? MDBX_EINVAL : MDBX_EPERM; if (unlikely((env->me_flags & MDBX_ENV_TXKEY) == 0)) { mdbx_assert(env, !env->me_lck || (env->me_flags & MDBX_NOTLS)); return MDBX_EINVAL /* MDBX_NOTLS mode */; } mdbx_assert(env, (env->me_flags & (MDBX_NOTLS | MDBX_ENV_TXKEY | MDBX_EXCLUSIVE)) == MDBX_ENV_TXKEY); MDBX_reader *r = thread_rthc_get(env->me_txkey); if (unlikely(r != NULL)) { mdbx_assert(env, r->mr_pid.weak == env->me_pid); mdbx_assert(env, r->mr_tid.weak == mdbx_thread_self()); if (unlikely(r->mr_pid.weak != env->me_pid)) return MDBX_BAD_RSLOT; return MDBX_RESULT_TRUE /* already registered */; } const uintptr_t tid = mdbx_thread_self(); if (env->me_txn0 && unlikely(env->me_txn0->mt_owner == tid)) return MDBX_TXN_OVERLAPPING; return bind_rslot((MDBX_env *)env, tid).err; } __cold int mdbx_thread_unregister(const MDBX_env *env) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!env->me_lck)) return MDBX_RESULT_TRUE; if (unlikely((env->me_flags & MDBX_ENV_TXKEY) == 0)) { mdbx_assert(env, !env->me_lck || (env->me_flags & MDBX_NOTLS)); return MDBX_RESULT_TRUE /* MDBX_NOTLS mode */; } mdbx_assert(env, (env->me_flags & (MDBX_NOTLS | MDBX_ENV_TXKEY | MDBX_EXCLUSIVE)) == MDBX_ENV_TXKEY); MDBX_reader *r = thread_rthc_get(env->me_txkey); if (unlikely(r == NULL)) return MDBX_RESULT_TRUE /* not registered */; mdbx_assert(env, r->mr_pid.weak == env->me_pid); mdbx_assert(env, r->mr_tid.weak == mdbx_thread_self()); if (unlikely(r->mr_pid.weak != env->me_pid || r->mr_tid.weak != mdbx_thread_self())) return MDBX_BAD_RSLOT; if (unlikely(r->mr_txnid.weak < SAFE64_INVALID_THRESHOLD)) return MDBX_BUSY /* transaction is still active */; atomic_store32(&r->mr_pid, 0, mo_AcquireRelease); atomic_store32(&env->me_lck->mti_readers_refresh_flag, true, mo_Relaxed); thread_rthc_set(env->me_txkey, nullptr); return MDBX_SUCCESS; } /* Common code for mdbx_txn_begin() and mdbx_txn_renew(). */ static int mdbx_txn_renew0(MDBX_txn *txn, const unsigned flags) { MDBX_env *env = txn->mt_env; int rc; #if MDBX_ENV_CHECKPID if (unlikely(env->me_pid != mdbx_getpid())) { env->me_flags |= MDBX_FATAL_ERROR; return MDBX_PANIC; } #endif /* MDBX_ENV_CHECKPID */ STATIC_ASSERT(sizeof(MDBX_reader) == 32); #if MDBX_LOCKING > 0 STATIC_ASSERT(offsetof(MDBX_lockinfo, mti_wlock) % MDBX_CACHELINE_SIZE == 0); STATIC_ASSERT(offsetof(MDBX_lockinfo, mti_rlock) % MDBX_CACHELINE_SIZE == 0); #else STATIC_ASSERT( offsetof(MDBX_lockinfo, mti_oldest_reader) % MDBX_CACHELINE_SIZE == 0); STATIC_ASSERT(offsetof(MDBX_lockinfo, mti_numreaders) % MDBX_CACHELINE_SIZE == 0); #endif /* MDBX_LOCKING */ STATIC_ASSERT(offsetof(MDBX_lockinfo, mti_readers) % MDBX_CACHELINE_SIZE == 0); const uintptr_t tid = mdbx_thread_self(); if (flags & MDBX_TXN_RDONLY) { mdbx_assert(env, (flags & ~(MDBX_TXN_RO_BEGIN_FLAGS | MDBX_WRITEMAP)) == 0); txn->mt_flags = MDBX_TXN_RDONLY | (env->me_flags & (MDBX_NOTLS | MDBX_WRITEMAP)); MDBX_reader *r = txn->to.reader; STATIC_ASSERT(sizeof(uintptr_t) <= sizeof(r->mr_tid)); if (likely(env->me_flags & MDBX_ENV_TXKEY)) { mdbx_assert(env, !(env->me_flags & MDBX_NOTLS)); r = thread_rthc_get(env->me_txkey); if (likely(r)) { if (unlikely(!r->mr_pid.weak) && (mdbx_runtime_flags & MDBX_DBG_LEGACY_MULTIOPEN)) { thread_rthc_set(env->me_txkey, nullptr); r = nullptr; } else { mdbx_assert(env, r->mr_pid.weak == env->me_pid); mdbx_assert(env, r->mr_tid.weak == mdbx_thread_self()); } } } else { mdbx_assert(env, !env->me_lck || (env->me_flags & MDBX_NOTLS)); } if (likely(r)) { if (unlikely(r->mr_pid.weak != env->me_pid || r->mr_txnid.weak < SAFE64_INVALID_THRESHOLD)) return MDBX_BAD_RSLOT; } else if (env->me_lck) { bind_rslot_result brs = bind_rslot(env, tid); if (unlikely(brs.err != MDBX_SUCCESS)) return brs.err; r = brs.rslot; } txn->to.reader = r; if (flags & (MDBX_TXN_RDONLY_PREPARE - MDBX_TXN_RDONLY)) { mdbx_assert(env, txn->mt_txnid == 0); mdbx_assert(env, txn->mt_owner == 0); mdbx_assert(env, txn->mt_numdbs == 0); if (likely(r)) { mdbx_assert(env, r->mr_snapshot_pages_used.weak == 0); mdbx_assert(env, r->mr_txnid.weak >= SAFE64_INVALID_THRESHOLD); atomic_store32(&r->mr_snapshot_pages_used, 0, mo_Relaxed); } txn->mt_flags = MDBX_TXN_RDONLY | MDBX_TXN_FINISHED; return MDBX_SUCCESS; } /* Seek & fetch the last meta */ if (likely(/* not recovery mode */ env->me_stuck_meta < 0)) { while (1) { MDBX_meta *const meta = mdbx_meta_head(env); mdbx_jitter4testing(false); const txnid_t snap = mdbx_meta_txnid_fluid(env, meta); mdbx_jitter4testing(false); if (likely(r)) { safe64_reset(&r->mr_txnid, false); atomic_store32(&r->mr_snapshot_pages_used, meta->mm_geo.next, mo_Relaxed); atomic_store64(&r->mr_snapshot_pages_retired, unaligned_peek_u64(4, meta->mm_pages_retired), mo_Relaxed); safe64_write(&r->mr_txnid, snap); mdbx_jitter4testing(false); mdbx_assert(env, r->mr_pid.weak == mdbx_getpid()); mdbx_assert( env, r->mr_tid.weak == ((env->me_flags & MDBX_NOTLS) ? 0 : mdbx_thread_self())); mdbx_assert(env, r->mr_txnid.weak == snap); atomic_store32(&env->me_lck->mti_readers_refresh_flag, true, mo_AcquireRelease); } mdbx_jitter4testing(true); /* Snap the state from current meta-head */ txn->mt_txnid = snap; txn->mt_geo = meta->mm_geo; memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDBX_db)); txn->mt_canary = meta->mm_canary; /* LY: Retry on a race, ITS#7970. */ if (likely(meta == mdbx_meta_head(env) && snap == mdbx_meta_txnid_fluid(env, meta) && snap >= atomic_load64(env->me_oldest, mo_AcquireRelease))) { mdbx_jitter4testing(false); break; } } } else { /* recovery mode */ MDBX_meta *const meta = METAPAGE(env, env->me_stuck_meta); txn->mt_txnid = mdbx_meta_txnid_stable(env, meta); txn->mt_geo = meta->mm_geo; memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDBX_db)); txn->mt_canary = meta->mm_canary; if (likely(r)) { atomic_store32(&r->mr_snapshot_pages_used, meta->mm_geo.next, mo_Relaxed); atomic_store64(&r->mr_snapshot_pages_retired, unaligned_peek_u64(4, meta->mm_pages_retired), mo_Relaxed); atomic_store64(&r->mr_txnid, txn->mt_txnid, mo_AcquireRelease); mdbx_jitter4testing(false); mdbx_assert(env, r->mr_pid.weak == mdbx_getpid()); mdbx_assert( env, r->mr_tid.weak == ((env->me_flags & MDBX_NOTLS) ? 0 : mdbx_thread_self())); mdbx_assert(env, r->mr_txnid.weak == txn->mt_txnid); atomic_store32(&env->me_lck->mti_readers_refresh_flag, true, mo_AcquireRelease); } } if (unlikely(txn->mt_txnid < MIN_TXNID || txn->mt_txnid > MAX_TXNID)) { mdbx_error("%s", "environment corrupted by died writer, must shutdown!"); rc = MDBX_CORRUPTED; goto bailout; } mdbx_assert(env, txn->mt_txnid >= env->me_oldest->weak); txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */ mdbx_ensure(env, txn->mt_txnid >= /* paranoia is appropriate here */ env->me_oldest->weak); txn->mt_numdbs = env->me_numdbs; } else { mdbx_assert(env, (flags & ~(MDBX_TXN_RW_BEGIN_FLAGS | MDBX_TXN_SPILLS | MDBX_WRITEMAP)) == 0); if (unlikely(txn->mt_owner == tid || /* not recovery mode */ env->me_stuck_meta >= 0)) return MDBX_BUSY; MDBX_lockinfo *const lck = env->me_lck; if (lck && (env->me_flags & MDBX_NOTLS) == 0 && (mdbx_runtime_flags & MDBX_DBG_LEGACY_OVERLAP) == 0) { const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_Relaxed) == env->me_pid && unlikely(atomic_load64(&lck->mti_readers[i].mr_tid, mo_Relaxed) == tid)) { const txnid_t txnid = safe64_read(&lck->mti_readers[i].mr_txnid); if (txnid >= MIN_TXNID && txnid <= MAX_TXNID) return MDBX_TXN_OVERLAPPING; } } } /* Not yet touching txn == env->me_txn0, it may be active */ mdbx_jitter4testing(false); rc = mdbx_txn_lock(env, F_ISSET(flags, MDBX_TXN_TRY)); if (unlikely(rc)) return rc; if (unlikely(env->me_flags & MDBX_FATAL_ERROR)) { mdbx_txn_unlock(env); return MDBX_PANIC; } #if defined(_WIN32) || defined(_WIN64) if (unlikely(!env->me_map)) { mdbx_txn_unlock(env); return MDBX_EPERM; } #endif /* Windows */ mdbx_jitter4testing(false); MDBX_meta *meta = mdbx_meta_head(env); mdbx_jitter4testing(false); txn->mt_canary = meta->mm_canary; const txnid_t snap = mdbx_meta_txnid_stable(env, meta); txn->mt_txnid = safe64_txnid_next(snap); if (unlikely(txn->mt_txnid > MAX_TXNID)) { rc = MDBX_TXN_FULL; mdbx_error("txnid overflow, raise %d", rc); goto bailout; } txn->mt_flags = flags; txn->mt_child = NULL; txn->tw.loose_pages = NULL; txn->tw.loose_count = 0; #if MDBX_ENABLE_REFUND txn->tw.loose_refund_wl = 0; #endif /* MDBX_ENABLE_REFUND */ txn->tw.dirtyroom = txn->mt_env->me_options.dp_limit; mdbx_dpl_clear(txn->tw.dirtylist); MDBX_PNL_SIZE(txn->tw.retired_pages) = 0; txn->tw.spill_pages = NULL; txn->tw.spill_least_removed = 0; txn->tw.last_reclaimed = 0; if (txn->tw.lifo_reclaimed) MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) = 0; env->me_txn = txn; txn->mt_numdbs = env->me_numdbs; memcpy(txn->mt_dbiseqs, env->me_dbiseqs, txn->mt_numdbs * sizeof(unsigned)); /* Copy the DB info and flags */ memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDBX_db)); /* Moved to here to avoid a data race in read TXNs */ txn->mt_geo = meta->mm_geo; } /* Setup db info */ mdbx_compiler_barrier(); for (unsigned i = CORE_DBS; i < txn->mt_numdbs; i++) { const unsigned db_flags = env->me_dbflags[i]; txn->mt_dbs[i].md_flags = db_flags & DB_PERSISTENT_FLAGS; txn->mt_dbistate[i] = (db_flags & DB_VALID) ? DBI_VALID | DBI_USRVALID | DBI_STALE : 0; } txn->mt_dbistate[MAIN_DBI] = DBI_VALID | DBI_USRVALID; txn->mt_dbistate[FREE_DBI] = DBI_VALID; if (unlikely(env->me_flags & MDBX_FATAL_ERROR)) { mdbx_warning("%s", "environment had fatal error, must shutdown!"); rc = MDBX_PANIC; } else { const size_t size = pgno2bytes(env, (txn->mt_flags & MDBX_TXN_RDONLY) ? txn->mt_next_pgno : txn->mt_end_pgno); if (unlikely(size > env->me_dxb_mmap.limit)) { if (txn->mt_geo.upper > MAX_PAGENO || bytes2pgno(env, pgno2bytes(env, txn->mt_geo.upper)) != txn->mt_geo.upper) { rc = MDBX_UNABLE_EXTEND_MAPSIZE; goto bailout; } rc = mdbx_mapresize(env, txn->mt_next_pgno, txn->mt_end_pgno, txn->mt_geo.upper, (txn->mt_flags & MDBX_TXN_RDONLY) ? true : false); if (rc != MDBX_SUCCESS) goto bailout; } if (txn->mt_flags & MDBX_TXN_RDONLY) { #if defined(_WIN32) || defined(_WIN64) if ((size > env->me_dbgeo.lower && env->me_dbgeo.shrink) || (mdbx_RunningUnderWine() && /* under Wine acquisition of remap_guard is always required, * since Wine don't support section extending, * i.e. in both cases unmap+map are required. */ size < env->me_dbgeo.upper && env->me_dbgeo.grow)) { txn->mt_flags |= MDBX_SHRINK_ALLOWED; mdbx_srwlock_AcquireShared(&env->me_remap_guard); } #endif /* Windows */ } else { env->me_dxb_mmap.current = size; #if defined(_WIN32) || defined(_WIN64) env->me_dxb_mmap.filesize = (env->me_dxb_mmap.filesize < size) ? size : env->me_dxb_mmap.filesize; #endif /* Windows */ } #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) mdbx_txn_valgrind(env, txn); #endif txn->mt_owner = tid; return MDBX_SUCCESS; } bailout: mdbx_tassert(txn, rc != MDBX_SUCCESS); mdbx_txn_end(txn, MDBX_END_SLOT | MDBX_END_FAIL_BEGIN); return rc; } static __always_inline int check_txn(const MDBX_txn *txn, int bad_bits) { if (unlikely(!txn)) return MDBX_EINVAL; if (unlikely(txn->mt_signature != MDBX_MT_SIGNATURE)) return MDBX_EBADSIGN; if (unlikely(txn->mt_flags & bad_bits)) return MDBX_BAD_TXN; #if MDBX_TXN_CHECKOWNER if ((txn->mt_flags & MDBX_NOTLS) == 0 && unlikely(txn->mt_owner != mdbx_thread_self())) return txn->mt_owner ? MDBX_THREAD_MISMATCH : MDBX_BAD_TXN; #endif /* MDBX_TXN_CHECKOWNER */ return MDBX_SUCCESS; } static __always_inline int check_txn_rw(const MDBX_txn *txn, int bad_bits) { if (unlikely(!txn)) return MDBX_EINVAL; if (unlikely(txn->mt_signature != MDBX_MT_SIGNATURE)) return MDBX_EBADSIGN; if (unlikely(txn->mt_flags & bad_bits)) return MDBX_BAD_TXN; if (unlikely(F_ISSET(txn->mt_flags, MDBX_TXN_RDONLY))) return MDBX_EACCESS; #if MDBX_TXN_CHECKOWNER if (unlikely(txn->mt_owner != mdbx_thread_self())) return txn->mt_owner ? MDBX_THREAD_MISMATCH : MDBX_BAD_TXN; #endif /* MDBX_TXN_CHECKOWNER */ return MDBX_SUCCESS; } int mdbx_txn_renew(MDBX_txn *txn) { if (unlikely(!txn)) return MDBX_EINVAL; if (unlikely(txn->mt_signature != MDBX_MT_SIGNATURE)) return MDBX_EBADSIGN; if (unlikely((txn->mt_flags & MDBX_TXN_RDONLY) == 0)) return MDBX_EINVAL; int rc; if (unlikely(txn->mt_owner != 0 || !(txn->mt_flags & MDBX_TXN_FINISHED))) { rc = mdbx_txn_reset(txn); if (unlikely(rc != MDBX_SUCCESS)) return rc; } rc = mdbx_txn_renew0(txn, MDBX_TXN_RDONLY); if (rc == MDBX_SUCCESS) { txn->mt_owner = mdbx_thread_self(); mdbx_debug("renew txn %" PRIaTXN "%c %p on env %p, root page %" PRIaPGNO "/%" PRIaPGNO, txn->mt_txnid, (txn->mt_flags & MDBX_TXN_RDONLY) ? 'r' : 'w', (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root, txn->mt_dbs[FREE_DBI].md_root); } return rc; } int mdbx_txn_begin(MDBX_env *env, MDBX_txn *parent, MDBX_txn_flags_t flags, MDBX_txn **ret) { return __inline_mdbx_txn_begin(env, parent, flags, ret); } int mdbx_txn_set_userctx(MDBX_txn *txn, void *ctx) { int rc = check_txn(txn, MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD); if (unlikely(rc != MDBX_SUCCESS)) return rc; txn->mt_userctx = ctx; return MDBX_SUCCESS; } void *mdbx_txn_get_userctx(const MDBX_txn *txn) { return check_txn(txn, MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD) ? nullptr : txn->mt_userctx; } int mdbx_txn_begin_ex(MDBX_env *env, MDBX_txn *parent, MDBX_txn_flags_t flags, MDBX_txn **ret, void *context) { MDBX_txn *txn; unsigned size, tsize; if (unlikely(!ret)) return MDBX_EINVAL; *ret = NULL; if (unlikely((flags & ~MDBX_TXN_RW_BEGIN_FLAGS) && (flags & ~MDBX_TXN_RO_BEGIN_FLAGS))) return MDBX_EINVAL; int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(env->me_flags & MDBX_RDONLY & ~flags)) /* write txn in RDONLY env */ return MDBX_EACCESS; if (unlikely(!env->me_map)) return MDBX_EPERM; flags |= env->me_flags & MDBX_WRITEMAP; if (parent) { /* Nested transactions: Max 1 child, write txns only, no writemap */ rc = check_txn_rw(parent, MDBX_TXN_RDONLY | MDBX_WRITEMAP | MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (env->me_options.spill_parent4child_denominator) { /* Spill dirty-pages of parent to provide dirtyroom for child txn */ rc = mdbx_txn_spill(parent, nullptr, parent->tw.dirtylist->length / env->me_options.spill_parent4child_denominator); if (unlikely(rc != MDBX_SUCCESS)) return rc; } mdbx_tassert(parent, mdbx_audit_ex(parent, 0, false) == 0); flags |= parent->mt_flags & (MDBX_TXN_RW_BEGIN_FLAGS | MDBX_TXN_SPILLS); /* Child txns save MDBX_pgstate and use own copy of cursors */ size = env->me_maxdbs * (sizeof(MDBX_db) + sizeof(MDBX_cursor *) + 1); size += tsize = sizeof(MDBX_txn); } else if (flags & MDBX_TXN_RDONLY) { if (env->me_txn0 && unlikely(env->me_txn0->mt_owner == mdbx_thread_self()) && (mdbx_runtime_flags & MDBX_DBG_LEGACY_OVERLAP) == 0) return MDBX_TXN_OVERLAPPING; size = env->me_maxdbs * (sizeof(MDBX_db) + 1); size += tsize = sizeof(MDBX_txn); } else { /* Reuse preallocated write txn. However, do not touch it until * mdbx_txn_renew0() succeeds, since it currently may be active. */ txn = env->me_txn0; goto renew; } if (unlikely((txn = mdbx_malloc(size)) == NULL)) { mdbx_debug("calloc: %s", "failed"); return MDBX_ENOMEM; } memset(txn, 0, tsize); txn->mt_dbxs = env->me_dbxs; /* static */ txn->mt_dbs = (MDBX_db *)((char *)txn + tsize); txn->mt_dbistate = (uint8_t *)txn + size - env->me_maxdbs; txn->mt_flags = flags; txn->mt_env = env; if (parent) { mdbx_tassert(parent, mdbx_dirtylist_check(parent)); txn->tw.cursors = (MDBX_cursor **)(txn->mt_dbs + env->me_maxdbs); txn->mt_dbiseqs = parent->mt_dbiseqs; rc = mdbx_dpl_alloc(txn); if (likely(rc == MDBX_SUCCESS)) { const unsigned len = MDBX_PNL_SIZE(parent->tw.reclaimed_pglist) + parent->tw.loose_count; txn->tw.reclaimed_pglist = mdbx_pnl_alloc((len > MDBX_PNL_INITIAL) ? len : MDBX_PNL_INITIAL); if (unlikely(!txn->tw.reclaimed_pglist)) rc = MDBX_ENOMEM; } if (unlikely(rc != MDBX_SUCCESS)) { nested_failed: mdbx_pnl_free(txn->tw.reclaimed_pglist); mdbx_dpl_free(txn); mdbx_free(txn); return rc; } txn->tw.dirtyroom = parent->tw.dirtyroom; /* Move loose pages to reclaimed list */ if (parent->tw.loose_count) { do { MDBX_page *lp = parent->tw.loose_pages; const unsigned di = mdbx_dpl_exist(parent->tw.dirtylist, lp->mp_pgno); mdbx_tassert(parent, di && parent->tw.dirtylist->items[di].ptr == lp); mdbx_tassert(parent, lp->mp_flags == (P_LOOSE | P_DIRTY)); rc = mdbx_pnl_insert_range(&parent->tw.reclaimed_pglist, lp->mp_pgno, 1); if (unlikely(rc != MDBX_SUCCESS)) goto nested_failed; parent->tw.loose_pages = lp->mp_next; /* Remove from dirty list */ mdbx_page_wash(parent, di, lp, 1); } while (parent->tw.loose_pages); parent->tw.loose_count = 0; #if MDBX_ENABLE_REFUND parent->tw.loose_refund_wl = 0; #endif /* MDBX_ENABLE_REFUND */ mdbx_tassert(parent, mdbx_dirtylist_check(parent)); } mdbx_dpl_sort(parent->tw.dirtylist); if (parent->tw.spill_pages) mdbx_spill_purge(parent); mdbx_tassert(txn, MDBX_PNL_ALLOCLEN(txn->tw.reclaimed_pglist) >= MDBX_PNL_SIZE(parent->tw.reclaimed_pglist)); memcpy(txn->tw.reclaimed_pglist, parent->tw.reclaimed_pglist, MDBX_PNL_SIZEOF(parent->tw.reclaimed_pglist)); mdbx_assert(env, mdbx_pnl_check4assert( txn->tw.reclaimed_pglist, (txn->mt_next_pgno /* LY: intentional assignment here, only for assertion */ = parent->mt_next_pgno) - MDBX_ENABLE_REFUND)); txn->tw.last_reclaimed = parent->tw.last_reclaimed; if (parent->tw.lifo_reclaimed) { txn->tw.lifo_reclaimed = parent->tw.lifo_reclaimed; parent->tw.lifo_reclaimed = (void *)(intptr_t)MDBX_PNL_SIZE(parent->tw.lifo_reclaimed); } txn->tw.retired_pages = parent->tw.retired_pages; parent->tw.retired_pages = (void *)(intptr_t)MDBX_PNL_SIZE(parent->tw.retired_pages); txn->mt_txnid = parent->mt_txnid; txn->mt_geo = parent->mt_geo; #if MDBX_ENABLE_REFUND txn->tw.loose_refund_wl = 0; #endif /* MDBX_ENABLE_REFUND */ txn->mt_canary = parent->mt_canary; parent->mt_flags |= MDBX_TXN_HAS_CHILD; parent->mt_child = txn; txn->mt_parent = parent; txn->mt_numdbs = parent->mt_numdbs; txn->mt_owner = parent->mt_owner; memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDBX_db)); /* Copy parent's mt_dbistate, but clear DB_NEW */ for (unsigned i = 0; i < txn->mt_numdbs; i++) txn->mt_dbistate[i] = parent->mt_dbistate[i] & ~(DBI_FRESH | DBI_CREAT | DBI_DIRTY); mdbx_tassert(parent, parent->mt_parent || parent->tw.dirtyroom + parent->tw.dirtylist->length == env->me_options.dp_limit); assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); env->me_txn = txn; rc = mdbx_cursor_shadow(parent, txn); if (unlikely(rc != MDBX_SUCCESS)) mdbx_txn_end(txn, MDBX_END_FAIL_BEGINCHILD); } else { /* MDBX_TXN_RDONLY */ txn->mt_dbiseqs = env->me_dbiseqs; renew: rc = mdbx_txn_renew0(txn, flags); } if (unlikely(rc != MDBX_SUCCESS)) { if (txn != env->me_txn0) mdbx_free(txn); } else { if (flags & (MDBX_TXN_RDONLY_PREPARE - MDBX_TXN_RDONLY)) mdbx_assert(env, txn->mt_flags == (MDBX_TXN_RDONLY | MDBX_TXN_FINISHED)); else if (flags & MDBX_TXN_RDONLY) mdbx_assert(env, (txn->mt_flags & ~(MDBX_NOTLS | MDBX_TXN_RDONLY | MDBX_WRITEMAP | /* Win32: SRWL flag */ MDBX_SHRINK_ALLOWED)) == 0); else { mdbx_assert(env, (txn->mt_flags & ~(MDBX_WRITEMAP | MDBX_SHRINK_ALLOWED | MDBX_NOMETASYNC | MDBX_SAFE_NOSYNC | MDBX_TXN_SPILLS)) == 0); assert(!txn->tw.spill_pages && !txn->tw.spill_least_removed); } txn->mt_signature = MDBX_MT_SIGNATURE; txn->mt_userctx = context; *ret = txn; mdbx_debug("begin txn %" PRIaTXN "%c %p on env %p, root page %" PRIaPGNO "/%" PRIaPGNO, txn->mt_txnid, (flags & MDBX_TXN_RDONLY) ? 'r' : 'w', (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root, txn->mt_dbs[FREE_DBI].md_root); } return rc; } int mdbx_txn_info(const MDBX_txn *txn, MDBX_txn_info *info, bool scan_rlt) { int rc = check_txn(txn, MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!info)) return MDBX_EINVAL; MDBX_env *const env = txn->mt_env; #if MDBX_ENV_CHECKPID if (unlikely(env->me_pid != mdbx_getpid())) { env->me_flags |= MDBX_FATAL_ERROR; return MDBX_PANIC; } #endif /* MDBX_ENV_CHECKPID */ info->txn_id = txn->mt_txnid; info->txn_space_used = pgno2bytes(env, txn->mt_geo.next); if (txn->mt_flags & MDBX_TXN_RDONLY) { const MDBX_meta *head_meta; txnid_t head_txnid; uint64_t head_retired; do { /* fetch info from volatile head */ head_meta = mdbx_meta_head(env); head_txnid = mdbx_meta_txnid_fluid(env, head_meta); head_retired = unaligned_peek_u64(4, head_meta->mm_pages_retired); info->txn_space_limit_soft = pgno2bytes(env, head_meta->mm_geo.now); info->txn_space_limit_hard = pgno2bytes(env, head_meta->mm_geo.upper); info->txn_space_leftover = pgno2bytes(env, head_meta->mm_geo.now - head_meta->mm_geo.next); mdbx_compiler_barrier(); } while (unlikely(head_meta != mdbx_meta_head(env) || head_txnid != mdbx_meta_txnid_fluid(env, head_meta))); info->txn_reader_lag = head_txnid - info->txn_id; info->txn_space_dirty = info->txn_space_retired = 0; uint64_t reader_snapshot_pages_retired; if (txn->to.reader && head_retired > (reader_snapshot_pages_retired = atomic_load64( &txn->to.reader->mr_snapshot_pages_retired, mo_Relaxed))) { info->txn_space_dirty = info->txn_space_retired = pgno2bytes( env, (pgno_t)(head_retired - reader_snapshot_pages_retired)); size_t retired_next_reader = 0; MDBX_lockinfo *const lck = env->me_lck; if (scan_rlt && info->txn_reader_lag > 1 && lck) { /* find next more recent reader */ txnid_t next_reader = head_txnid; const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { retry: if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease)) { mdbx_jitter4testing(true); const txnid_t snap_txnid = safe64_read(&lck->mti_readers[i].mr_txnid); const uint64_t snap_retired = atomic_load64(&lck->mti_readers[i].mr_snapshot_pages_retired, mo_AcquireRelease); if (unlikely(snap_retired != atomic_load64( &lck->mti_readers[i].mr_snapshot_pages_retired, mo_Relaxed)) || snap_txnid != safe64_read(&lck->mti_readers[i].mr_txnid)) goto retry; if (snap_txnid <= txn->mt_txnid) { retired_next_reader = 0; break; } if (snap_txnid < next_reader) { next_reader = snap_txnid; retired_next_reader = pgno2bytes( env, (pgno_t)(snap_retired - atomic_load64( &txn->to.reader->mr_snapshot_pages_retired, mo_Relaxed))); } } } } info->txn_space_dirty = retired_next_reader; } } else { info->txn_space_limit_soft = pgno2bytes(env, txn->mt_geo.now); info->txn_space_limit_hard = pgno2bytes(env, txn->mt_geo.upper); info->txn_space_retired = pgno2bytes( env, txn->mt_child ? (unsigned)(uintptr_t)txn->tw.retired_pages : MDBX_PNL_SIZE(txn->tw.retired_pages)); info->txn_space_leftover = pgno2bytes(env, txn->tw.dirtyroom); info->txn_space_dirty = pgno2bytes(env, txn->mt_env->me_options.dp_limit - txn->tw.dirtyroom); info->txn_reader_lag = INT64_MAX; MDBX_lockinfo *const lck = env->me_lck; if (scan_rlt && lck) { txnid_t oldest_snapshot = txn->mt_txnid; const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); if (snap_nreaders) { oldest_snapshot = mdbx_find_oldest(txn); if (oldest_snapshot == txn->mt_txnid - 1) { /* check if there is at least one reader */ bool exists = false; for (unsigned i = 0; i < snap_nreaders; ++i) { if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_Relaxed) && txn->mt_txnid > safe64_read(&lck->mti_readers[i].mr_txnid)) { exists = true; break; } } oldest_snapshot += !exists; } } info->txn_reader_lag = txn->mt_txnid - oldest_snapshot; } } return MDBX_SUCCESS; } MDBX_env *mdbx_txn_env(const MDBX_txn *txn) { if (unlikely(!txn || txn->mt_signature != MDBX_MT_SIGNATURE || txn->mt_env->me_signature.weak != MDBX_ME_SIGNATURE)) return NULL; return txn->mt_env; } uint64_t mdbx_txn_id(const MDBX_txn *txn) { if (unlikely(!txn || txn->mt_signature != MDBX_MT_SIGNATURE)) return 0; return txn->mt_txnid; } int mdbx_txn_flags(const MDBX_txn *txn) { if (unlikely(!txn || txn->mt_signature != MDBX_MT_SIGNATURE)) return -1; return txn->mt_flags; } /* Check for misused dbi handles */ #define TXN_DBI_CHANGED(txn, dbi) \ ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi]) static void dbi_import_locked(MDBX_txn *txn) { MDBX_env *const env = txn->mt_env; const unsigned n = env->me_numdbs; for (unsigned i = CORE_DBS; i < n; ++i) { if (i >= txn->mt_numdbs) { txn->mt_dbistate[i] = 0; if (!(txn->mt_flags & MDBX_TXN_RDONLY)) txn->tw.cursors[i] = NULL; } if ((env->me_dbflags[i] & DB_VALID) && !(txn->mt_dbistate[i] & DBI_USRVALID)) { txn->mt_dbiseqs[i] = env->me_dbiseqs[i]; txn->mt_dbs[i].md_flags = env->me_dbflags[i] & DB_PERSISTENT_FLAGS; txn->mt_dbistate[i] = DBI_VALID | DBI_USRVALID | DBI_STALE; mdbx_tassert(txn, txn->mt_dbxs[i].md_cmp != NULL); } } txn->mt_numdbs = n; } /* Import DBI which opened after txn started into context */ static __cold bool dbi_import(MDBX_txn *txn, MDBX_dbi dbi) { if (dbi < CORE_DBS || dbi >= txn->mt_env->me_numdbs) return false; mdbx_ensure(txn->mt_env, mdbx_fastmutex_acquire(&txn->mt_env->me_dbi_lock) == MDBX_SUCCESS); dbi_import_locked(txn); mdbx_ensure(txn->mt_env, mdbx_fastmutex_release(&txn->mt_env->me_dbi_lock) == MDBX_SUCCESS); return txn->mt_dbistate[dbi] & DBI_USRVALID; } /* Export or close DBI handles opened in this txn. */ static void dbi_update(MDBX_txn *txn, int keep) { mdbx_tassert(txn, !txn->mt_parent && txn == txn->mt_env->me_txn0); MDBX_dbi n = txn->mt_numdbs; if (n) { bool locked = false; MDBX_env *const env = txn->mt_env; for (unsigned i = n; --i >= CORE_DBS;) { if (likely((txn->mt_dbistate[i] & DBI_CREAT) == 0)) continue; if (!locked) { mdbx_ensure(env, mdbx_fastmutex_acquire(&env->me_dbi_lock) == MDBX_SUCCESS); locked = true; } if (env->me_numdbs <= i || txn->mt_dbiseqs[i] != env->me_dbiseqs[i]) continue /* dbi explicitly closed and/or then re-opened by other txn */; if (keep) { env->me_dbflags[i] = txn->mt_dbs[i].md_flags | DB_VALID; } else { char *ptr = env->me_dbxs[i].md_name.iov_base; if (ptr) { env->me_dbxs[i].md_name.iov_len = 0; mdbx_memory_fence(false, true); mdbx_assert(env, env->me_dbflags[i] == 0); env->me_dbiseqs[i]++; env->me_dbxs[i].md_name.iov_base = NULL; mdbx_free(ptr); } } } n = env->me_numdbs; if (n > CORE_DBS && unlikely(!(env->me_dbflags[n - 1] & DB_VALID))) { if (!locked) { mdbx_ensure(env, mdbx_fastmutex_acquire(&env->me_dbi_lock) == MDBX_SUCCESS); locked = true; } n = env->me_numdbs; while (n > CORE_DBS && !(env->me_dbflags[n - 1] & DB_VALID)) --n; env->me_numdbs = n; } if (unlikely(locked)) mdbx_ensure(env, mdbx_fastmutex_release(&env->me_dbi_lock) == MDBX_SUCCESS); } } /* Filter-out pgno list from transaction's dirty-page list */ static void mdbx_dpl_sift(MDBX_txn *const txn, MDBX_PNL pl, const bool spilled) { if (MDBX_PNL_SIZE(pl) && txn->tw.dirtylist->length) { mdbx_tassert(txn, mdbx_pnl_check4assert(pl, txn->mt_next_pgno << spilled)); MDBX_dpl *dl = mdbx_dpl_sort(txn->tw.dirtylist); /* Scanning in ascend order */ const int step = MDBX_PNL_ASCENDING ? 1 : -1; const int begin = MDBX_PNL_ASCENDING ? 1 : MDBX_PNL_SIZE(pl); const int end = MDBX_PNL_ASCENDING ? MDBX_PNL_SIZE(pl) + 1 : 0; mdbx_tassert(txn, pl[begin] <= pl[end - step]); unsigned r = mdbx_dpl_search(dl, pl[begin] >> spilled); mdbx_tassert(txn, dl->sorted == dl->length); for (int i = begin; r <= dl->length;) { /* scan loop */ assert(i != end); mdbx_tassert(txn, !spilled || (pl[i] & 1) == 0); pgno_t pl_pgno = pl[i] >> spilled; pgno_t dp_pgno = dl->items[r].pgno; if (likely(dp_pgno != pl_pgno)) { const bool cmp = dp_pgno < pl_pgno; r += cmp; i += cmp ? 0 : step; if (likely(i != end)) continue; return; } /* update loop */ unsigned w = r; remove_dl: if ((txn->mt_env->me_flags & MDBX_WRITEMAP) == 0) { MDBX_page *dp = dl->items[r].ptr; mdbx_dpage_free(txn->mt_env, dp, IS_OVERFLOW(dp) ? dp->mp_pages : 1); } ++r; next_i: i += step; if (unlikely(i == end)) { while (r <= dl->length) dl->items[w++] = dl->items[r++]; } else { while (r <= dl->length) { assert(i != end); mdbx_tassert(txn, !spilled || (pl[i] & 1) == 0); pl_pgno = pl[i] >> spilled; dp_pgno = dl->items[r].pgno; if (dp_pgno < pl_pgno) dl->items[w++] = dl->items[r++]; else if (dp_pgno > pl_pgno) goto next_i; else goto remove_dl; } } dl->sorted = dl->length = w - 1; txn->tw.dirtyroom += r - w; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); return; } } } /* End a transaction, except successful commit of a nested transaction. * May be called twice for readonly txns: First reset it, then abort. * [in] txn the transaction handle to end * [in] mode why and how to end the transaction */ static int mdbx_txn_end(MDBX_txn *txn, const unsigned mode) { MDBX_env *env = txn->mt_env; static const char *const names[] = MDBX_END_NAMES; #if MDBX_ENV_CHECKPID if (unlikely(txn->mt_env->me_pid != mdbx_getpid())) { env->me_flags |= MDBX_FATAL_ERROR; return MDBX_PANIC; } #endif /* MDBX_ENV_CHECKPID */ mdbx_debug("%s txn %" PRIaTXN "%c %p on mdbenv %p, root page %" PRIaPGNO "/%" PRIaPGNO, names[mode & MDBX_END_OPMASK], txn->mt_txnid, (txn->mt_flags & MDBX_TXN_RDONLY) ? 'r' : 'w', (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root, txn->mt_dbs[FREE_DBI].md_root); mdbx_ensure(env, txn->mt_txnid >= /* paranoia is appropriate here */ env->me_oldest->weak); int rc = MDBX_SUCCESS; if (F_ISSET(txn->mt_flags, MDBX_TXN_RDONLY)) { if (txn->to.reader) { MDBX_reader *slot = txn->to.reader; mdbx_assert(env, slot->mr_pid.weak == env->me_pid); if (likely(!F_ISSET(txn->mt_flags, MDBX_TXN_FINISHED))) { mdbx_assert(env, txn->mt_txnid == slot->mr_txnid.weak && slot->mr_txnid.weak >= env->me_lck->mti_oldest_reader.weak); #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) mdbx_txn_valgrind(env, nullptr); #endif atomic_store32(&slot->mr_snapshot_pages_used, 0, mo_Relaxed); safe64_reset(&slot->mr_txnid, false); atomic_store32(&env->me_lck->mti_readers_refresh_flag, true, mo_AcquireRelease); } else { mdbx_assert(env, slot->mr_pid.weak == env->me_pid); mdbx_assert(env, slot->mr_txnid.weak >= SAFE64_INVALID_THRESHOLD); } if (mode & MDBX_END_SLOT) { if ((env->me_flags & MDBX_ENV_TXKEY) == 0) atomic_store32(&slot->mr_pid, 0, mo_AcquireRelease); txn->to.reader = NULL; } } #if defined(_WIN32) || defined(_WIN64) if (txn->mt_flags & MDBX_SHRINK_ALLOWED) mdbx_srwlock_ReleaseShared(&env->me_remap_guard); #endif txn->mt_numdbs = 0; /* prevent further DBI activity */ txn->mt_flags = MDBX_TXN_RDONLY | MDBX_TXN_FINISHED; txn->mt_owner = 0; } else if (!F_ISSET(txn->mt_flags, MDBX_TXN_FINISHED)) { #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) if (txn == env->me_txn0) mdbx_txn_valgrind(env, nullptr); #endif if (!(mode & MDBX_END_EOTDONE)) /* !(already closed cursors) */ mdbx_cursors_eot(txn, false); txn->mt_flags = MDBX_TXN_FINISHED; txn->mt_owner = 0; env->me_txn = txn->mt_parent; mdbx_pnl_free(txn->tw.spill_pages); txn->tw.spill_pages = nullptr; if (txn == env->me_txn0) { mdbx_assert(env, txn->mt_parent == NULL); /* Export or close DBI handles created in this txn */ dbi_update(txn, mode & MDBX_END_UPDATE); mdbx_pnl_shrink(&txn->tw.retired_pages); mdbx_pnl_shrink(&txn->tw.reclaimed_pglist); if (!(env->me_flags & MDBX_WRITEMAP)) mdbx_dlist_free(txn); /* The writer mutex was locked in mdbx_txn_begin. */ mdbx_txn_unlock(env); } else { mdbx_assert(env, txn->mt_parent != NULL); MDBX_txn *const parent = txn->mt_parent; mdbx_assert(env, parent->mt_signature == MDBX_MT_SIGNATURE); mdbx_assert(env, parent->mt_child == txn && (parent->mt_flags & MDBX_TXN_HAS_CHILD) != 0); mdbx_assert( env, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); if (txn->tw.lifo_reclaimed) { mdbx_assert(env, MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) >= (unsigned)(uintptr_t)parent->tw.lifo_reclaimed); MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) = (unsigned)(uintptr_t)parent->tw.lifo_reclaimed; parent->tw.lifo_reclaimed = txn->tw.lifo_reclaimed; } if (txn->tw.retired_pages) { mdbx_assert(env, MDBX_PNL_SIZE(txn->tw.retired_pages) >= (unsigned)(uintptr_t)parent->tw.retired_pages); MDBX_PNL_SIZE(txn->tw.retired_pages) = (unsigned)(uintptr_t)parent->tw.retired_pages; parent->tw.retired_pages = txn->tw.retired_pages; } parent->mt_child = nullptr; parent->mt_flags &= ~MDBX_TXN_HAS_CHILD; mdbx_tassert(parent, mdbx_dirtylist_check(parent)); mdbx_tassert(parent, mdbx_audit_ex(parent, 0, false) == 0); if (!(env->me_flags & MDBX_WRITEMAP)) mdbx_dlist_free(txn); mdbx_dpl_free(txn); mdbx_pnl_free(txn->tw.reclaimed_pglist); if (parent->mt_geo.upper != txn->mt_geo.upper || parent->mt_geo.now != txn->mt_geo.now) { /* undo resize performed by child txn */ rc = mdbx_mapresize_implicit(env, parent->mt_next_pgno, parent->mt_geo.now, parent->mt_geo.upper); if (rc == MDBX_RESULT_TRUE) { /* unable undo resize (it is regular for Windows), * therefore promote size changes from child to the parent txn */ mdbx_warning("unable undo resize performed by child txn, promote to " "the parent (%u->%u, %u->%u)", txn->mt_geo.now, parent->mt_geo.now, txn->mt_geo.upper, parent->mt_geo.upper); parent->mt_geo.now = txn->mt_geo.now; parent->mt_geo.upper = txn->mt_geo.upper; rc = MDBX_SUCCESS; } else if (unlikely(rc != MDBX_SUCCESS)) { mdbx_error("error %d while undo resize performed by child txn, fail " "the parent", rc); parent->mt_flags |= MDBX_TXN_ERROR; if (!env->me_dxb_mmap.address) env->me_flags |= MDBX_FATAL_ERROR; } } } } mdbx_assert(env, txn == env->me_txn0 || txn->mt_owner == 0); if ((mode & MDBX_END_FREE) != 0 && txn != env->me_txn0) { txn->mt_signature = 0; mdbx_free(txn); } return rc; } int mdbx_txn_reset(MDBX_txn *txn) { int rc = check_txn(txn, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; /* This call is only valid for read-only txns */ if (unlikely((txn->mt_flags & MDBX_TXN_RDONLY) == 0)) return MDBX_EINVAL; /* LY: don't close DBI-handles */ rc = mdbx_txn_end(txn, MDBX_END_RESET | MDBX_END_UPDATE); if (rc == MDBX_SUCCESS) { mdbx_tassert(txn, txn->mt_signature == MDBX_MT_SIGNATURE); mdbx_tassert(txn, txn->mt_owner == 0); } return rc; } int mdbx_txn_break(MDBX_txn *txn) { do { int rc = check_txn(txn, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; txn->mt_flags |= MDBX_TXN_ERROR; if (txn->mt_flags & MDBX_TXN_RDONLY) break; txn = txn->mt_child; } while (txn); return MDBX_SUCCESS; } int mdbx_txn_abort(MDBX_txn *txn) { int rc = check_txn(txn, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (F_ISSET(txn->mt_flags, MDBX_TXN_RDONLY)) /* LY: don't close DBI-handles */ return mdbx_txn_end(txn, MDBX_END_ABORT | MDBX_END_UPDATE | MDBX_END_SLOT | MDBX_END_FREE); if (txn->mt_child) mdbx_txn_abort(txn->mt_child); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); return mdbx_txn_end(txn, MDBX_END_ABORT | MDBX_END_SLOT | MDBX_END_FREE); } /* Count all the pages in each DB and in the freelist and make sure * it matches the actual number of pages being used. * All named DBs must be open for a correct count. */ static __cold int mdbx_audit_ex(MDBX_txn *txn, unsigned retired_stored, bool dont_filter_gc) { pgno_t pending = 0; if ((txn->mt_flags & MDBX_TXN_RDONLY) == 0) { pending = txn->tw.loose_count + MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) + (MDBX_PNL_SIZE(txn->tw.retired_pages) - retired_stored); } MDBX_cursor_couple cx; int rc = mdbx_cursor_init(&cx.outer, txn, FREE_DBI); if (unlikely(rc != MDBX_SUCCESS)) return rc; pgno_t freecount = 0; MDBX_val key, data; while ((rc = mdbx_cursor_get(&cx.outer, &key, &data, MDBX_NEXT)) == 0) { if (!dont_filter_gc) { if (unlikely(key.iov_len != sizeof(txnid_t))) return MDBX_CORRUPTED; txnid_t id = unaligned_peek_u64(4, key.iov_base); if (txn->tw.lifo_reclaimed) { for (unsigned i = 1; i <= MDBX_PNL_SIZE(txn->tw.lifo_reclaimed); ++i) if (id == txn->tw.lifo_reclaimed[i]) goto skip; } else if (id <= txn->tw.last_reclaimed) goto skip; } freecount += *(pgno_t *)data.iov_base; skip:; } mdbx_tassert(txn, rc == MDBX_NOTFOUND); for (MDBX_dbi i = FREE_DBI; i < txn->mt_numdbs; i++) txn->mt_dbistate[i] &= ~DBI_AUDITED; pgno_t count = 0; for (MDBX_dbi i = FREE_DBI; i <= MAIN_DBI; i++) { if (!(txn->mt_dbistate[i] & DBI_VALID)) continue; rc = mdbx_cursor_init(&cx.outer, txn, i); if (unlikely(rc != MDBX_SUCCESS)) return rc; txn->mt_dbistate[i] |= DBI_AUDITED; if (txn->mt_dbs[i].md_root == P_INVALID) continue; count += txn->mt_dbs[i].md_branch_pages + txn->mt_dbs[i].md_leaf_pages + txn->mt_dbs[i].md_overflow_pages; if (i != MAIN_DBI) continue; rc = mdbx_page_search(&cx.outer, NULL, MDBX_PS_FIRST); while (rc == MDBX_SUCCESS) { MDBX_page *mp = cx.outer.mc_pg[cx.outer.mc_top]; for (unsigned j = 0; j < page_numkeys(mp); j++) { MDBX_node *node = page_node(mp, j); if (node_flags(node) == F_SUBDATA) { if (unlikely(node_ds(node) != sizeof(MDBX_db))) return MDBX_CORRUPTED; MDBX_db db_copy, *db; memcpy(db = &db_copy, node_data(node), sizeof(db_copy)); if ((txn->mt_flags & MDBX_TXN_RDONLY) == 0) { for (MDBX_dbi k = txn->mt_numdbs; --k > MAIN_DBI;) { if ((txn->mt_dbistate[k] & DBI_VALID) && /* txn->mt_dbxs[k].md_name.iov_len > 0 && */ node_ks(node) == txn->mt_dbxs[k].md_name.iov_len && memcmp(node_key(node), txn->mt_dbxs[k].md_name.iov_base, node_ks(node)) == 0) { txn->mt_dbistate[k] |= DBI_AUDITED; if (!(txn->mt_dbistate[k] & MDBX_DBI_STALE)) db = txn->mt_dbs + k; break; } } } count += db->md_branch_pages + db->md_leaf_pages + db->md_overflow_pages; } } rc = mdbx_cursor_sibling(&cx.outer, SIBLING_RIGHT); } mdbx_tassert(txn, rc == MDBX_NOTFOUND); } for (MDBX_dbi i = FREE_DBI; i < txn->mt_numdbs; i++) { if ((txn->mt_dbistate[i] & (DBI_VALID | DBI_AUDITED | DBI_STALE)) != DBI_VALID) continue; for (MDBX_txn *t = txn; t; t = t->mt_parent) if (F_ISSET(t->mt_dbistate[i], DBI_DIRTY | DBI_CREAT)) { count += t->mt_dbs[i].md_branch_pages + t->mt_dbs[i].md_leaf_pages + t->mt_dbs[i].md_overflow_pages; txn->mt_dbistate[i] |= DBI_AUDITED; break; } if (!(txn->mt_dbistate[i] & DBI_AUDITED)) { mdbx_warning("audit %s@%" PRIaTXN ": unable account dbi %d / \"%*s\", state 0x%02x", txn->mt_parent ? "nested-" : "", txn->mt_txnid, i, (int)txn->mt_dbxs[i].md_name.iov_len, (const char *)txn->mt_dbxs[i].md_name.iov_base, txn->mt_dbistate[i]); } } if (pending + freecount + count + NUM_METAS == txn->mt_next_pgno) return MDBX_SUCCESS; if ((txn->mt_flags & MDBX_TXN_RDONLY) == 0) mdbx_error("audit @%" PRIaTXN ": %u(pending) = %u(loose-count) + " "%u(reclaimed-list) + %u(retired-pending) - %u(retired-stored)", txn->mt_txnid, pending, txn->tw.loose_count, MDBX_PNL_SIZE(txn->tw.reclaimed_pglist), txn->tw.retired_pages ? MDBX_PNL_SIZE(txn->tw.retired_pages) : 0, retired_stored); mdbx_error("audit @%" PRIaTXN ": %" PRIaPGNO "(pending) + %" PRIaPGNO "(free) + %" PRIaPGNO "(count) = %" PRIaPGNO "(total) <> %" PRIaPGNO "(next-pgno)", txn->mt_txnid, pending, freecount, count + NUM_METAS, pending + freecount + count + NUM_METAS, txn->mt_next_pgno); return MDBX_PROBLEM; } static __always_inline unsigned backlog_size(MDBX_txn *txn) { return MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) + txn->tw.loose_count; } /* LY: Prepare a backlog of pages to modify GC itself, * while reclaiming is prohibited. It should be enough to prevent search * in mdbx_page_alloc() during a deleting, when GC tree is unbalanced. */ static int mdbx_prep_backlog(MDBX_txn *txn, MDBX_cursor *gc_cursor, const size_t pnl_bytes) { const unsigned linear4list = number_of_ovpages(txn->mt_env, pnl_bytes); const unsigned backlog4cow = txn->mt_dbs[FREE_DBI].md_depth; const unsigned backlog4rebalance = backlog4cow + 1; if (likely(linear4list == 1 && backlog_size(txn) > (pnl_bytes ? backlog4rebalance : (backlog4cow + backlog4rebalance)))) return MDBX_SUCCESS; mdbx_trace(">> pnl_bytes %zu, backlog %u, 4list %u, 4cow %u, 4rebalance %u", pnl_bytes, backlog_size(txn), linear4list, backlog4cow, backlog4rebalance); gc_cursor->mc_flags &= ~C_RECLAIMING; int err = mdbx_cursor_touch(gc_cursor); mdbx_trace("== after-touch, backlog %u, err %d", backlog_size(txn), err); if (linear4list > 1 && err == MDBX_SUCCESS) { err = mdbx_page_alloc(gc_cursor, linear4list, nullptr, MDBX_ALLOC_GC | MDBX_ALLOC_CACHE); mdbx_trace("== after-4linear, backlog %u, err %d", backlog_size(txn), err); } while (backlog_size(txn) < backlog4cow + linear4list && err == MDBX_SUCCESS) err = mdbx_page_alloc(gc_cursor, 1, NULL, MDBX_ALLOC_GC); gc_cursor->mc_flags |= C_RECLAIMING; mdbx_trace("<< backlog %u, err %d", backlog_size(txn), err); return (err != MDBX_NOTFOUND) ? err : MDBX_SUCCESS; } static __inline void clean_reserved_gc_pnl(MDBX_env *env, MDBX_val pnl) { /* PNL is initially empty, zero out at least the length */ memset(pnl.iov_base, 0, sizeof(pgno_t)); if ((env->me_flags & (MDBX_WRITEMAP | MDBX_NOMEMINIT)) == 0) /* zero out to avoid leaking values from uninitialized malloc'ed memory * to the file in non-writemap mode if length of the saving page-list * was changed during space reservation. */ memset(pnl.iov_base, 0, pnl.iov_len); } /* Cleanup reclaimed GC records, than save the retired-list as of this * transaction to GC (aka freeDB). This recursive changes the reclaimed-list * loose-list and retired-list. Keep trying until it stabilizes. */ static int mdbx_update_gc(MDBX_txn *txn) { /* txn->tw.reclaimed_pglist[] can grow and shrink during this call. * txn->tw.last_reclaimed and txn->tw.retired_pages[] can only grow. * Page numbers cannot disappear from txn->tw.retired_pages[]. */ MDBX_env *const env = txn->mt_env; const bool lifo = (env->me_flags & MDBX_LIFORECLAIM) != 0; const char *dbg_prefix_mode = lifo ? " lifo" : " fifo"; (void)dbg_prefix_mode; mdbx_trace("\n>>> @%" PRIaTXN, txn->mt_txnid); unsigned retired_stored = 0, loop = 0; MDBX_cursor_couple couple; int rc = mdbx_cursor_init(&couple.outer, txn, FREE_DBI); if (unlikely(rc != MDBX_SUCCESS)) goto bailout_notracking; couple.outer.mc_flags |= C_RECLAIMING; couple.outer.mc_next = txn->tw.cursors[FREE_DBI]; txn->tw.cursors[FREE_DBI] = &couple.outer; retry: ++loop; retry_noaccount: mdbx_trace("%s", " >> restart"); mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); mdbx_tassert(txn, txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); if (unlikely(/* paranoia */ loop > ((MDBX_DEBUG > 0) ? 9 : 99))) { mdbx_error("too more loops %u, bailout", loop); rc = MDBX_PROBLEM; goto bailout; } rc = mdbx_prep_backlog(txn, &couple.outer, MDBX_PNL_SIZEOF(txn->tw.retired_pages)); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; unsigned settled = 0, cleaned_gc_slot = 0, reused_gc_slot = 0, filled_gc_slot = ~0u; txnid_t cleaned_gc_id = 0, gc_rid = txn->tw.last_reclaimed; while (true) { /* Come back here after each Put() in case retired-list changed */ MDBX_val key, data; mdbx_trace("%s", " >> continue"); mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); if (lifo) { if (cleaned_gc_slot < (txn->tw.lifo_reclaimed ? MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : 0)) { settled = 0; cleaned_gc_slot = 0; reused_gc_slot = 0; filled_gc_slot = ~0u; /* LY: cleanup reclaimed records. */ do { cleaned_gc_id = txn->tw.lifo_reclaimed[++cleaned_gc_slot]; mdbx_tassert(txn, cleaned_gc_slot > 0 && cleaned_gc_id < env->me_oldest->weak); key.iov_base = &cleaned_gc_id; key.iov_len = sizeof(cleaned_gc_id); rc = mdbx_cursor_get(&couple.outer, &key, NULL, MDBX_SET); if (rc == MDBX_NOTFOUND) continue; if (unlikely(rc != MDBX_SUCCESS)) goto bailout; rc = mdbx_prep_backlog(txn, &couple.outer, 0); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; mdbx_tassert(txn, cleaned_gc_id < env->me_oldest->weak); mdbx_trace("%s.cleanup-reclaimed-id [%u]%" PRIaTXN, dbg_prefix_mode, cleaned_gc_slot, cleaned_gc_id); rc = mdbx_cursor_del(&couple.outer, 0); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } while (cleaned_gc_slot < MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)); mdbx_txl_sort(txn->tw.lifo_reclaimed); } } else { /* If using records from GC which we have not yet deleted, * now delete them and any we reserved for tw.reclaimed_pglist. */ while (cleaned_gc_id <= txn->tw.last_reclaimed) { gc_rid = cleaned_gc_id; settled = 0; rc = mdbx_cursor_first(&couple.outer, &key, NULL); if (unlikely(rc != MDBX_SUCCESS)) { if (rc == MDBX_NOTFOUND) break; goto bailout; } if (unlikely(key.iov_len != sizeof(txnid_t))) { rc = MDBX_CORRUPTED; goto bailout; } cleaned_gc_id = unaligned_peek_u64(4, key.iov_base); if (unlikely(cleaned_gc_id < MIN_TXNID || cleaned_gc_id > MAX_TXNID)) { rc = MDBX_CORRUPTED; goto bailout; } if (cleaned_gc_id > txn->tw.last_reclaimed) break; if (cleaned_gc_id < txn->tw.last_reclaimed) { rc = mdbx_prep_backlog(txn, &couple.outer, 0); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } mdbx_tassert(txn, cleaned_gc_id <= txn->tw.last_reclaimed); mdbx_tassert(txn, cleaned_gc_id < env->me_oldest->weak); mdbx_trace("%s.cleanup-reclaimed-id %" PRIaTXN, dbg_prefix_mode, cleaned_gc_id); rc = mdbx_cursor_del(&couple.outer, 0); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } } mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); mdbx_tassert(txn, txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); if (mdbx_audit_enabled()) { rc = mdbx_audit_ex(txn, retired_stored, false); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } /* return suitable into unallocated space */ if (mdbx_refund(txn)) { mdbx_tassert( txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); if (mdbx_audit_enabled()) { rc = mdbx_audit_ex(txn, retired_stored, false); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } } /* handle loose pages - put ones into the reclaimed- or retired-list */ if (txn->tw.loose_pages) { /* Return loose page numbers to tw.reclaimed_pglist, * though usually none are left at this point. * The pages themselves remain in dirtylist. */ if (unlikely(!txn->tw.lifo_reclaimed && txn->tw.last_reclaimed < 1)) { if (txn->tw.loose_count > 0) { /* Put loose page numbers in tw.retired_pages, * since unable to return them to tw.reclaimed_pglist. */ if (unlikely((rc = mdbx_pnl_need(&txn->tw.retired_pages, txn->tw.loose_count)) != 0)) goto bailout; for (MDBX_page *mp = txn->tw.loose_pages; mp; mp = mp->mp_next) mdbx_pnl_xappend(txn->tw.retired_pages, mp->mp_pgno); mdbx_trace("%s: append %u loose-pages to retired-pages", dbg_prefix_mode, txn->tw.loose_count); } } else { /* Room for loose pages + temp PNL with same */ rc = mdbx_pnl_need(&txn->tw.reclaimed_pglist, 2 * txn->tw.loose_count + 2); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; MDBX_PNL loose = txn->tw.reclaimed_pglist + MDBX_PNL_ALLOCLEN(txn->tw.reclaimed_pglist) - txn->tw.loose_count - 1; unsigned count = 0; for (MDBX_page *mp = txn->tw.loose_pages; mp; mp = mp->mp_next) { mdbx_tassert(txn, mp->mp_flags == (P_LOOSE | P_DIRTY)); loose[++count] = mp->mp_pgno; } mdbx_tassert(txn, count == txn->tw.loose_count); MDBX_PNL_SIZE(loose) = count; mdbx_pnl_sort(loose); mdbx_pnl_xmerge(txn->tw.reclaimed_pglist, loose); mdbx_trace("%s: append %u loose-pages to reclaimed-pages", dbg_prefix_mode, txn->tw.loose_count); } /* filter-out list of dirty-pages from loose-pages */ MDBX_dpl *const dl = txn->tw.dirtylist; unsigned w = 0; for (unsigned r = w; ++r <= dl->length;) { MDBX_page *dp = dl->items[r].ptr; mdbx_tassert(txn, (dp->mp_flags & P_DIRTY)); mdbx_tassert(txn, dl->items[r].pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1) <= txn->mt_next_pgno); if ((dp->mp_flags & P_LOOSE) == 0) { if (++w != r) dl->items[w] = dl->items[r]; } else { mdbx_tassert(txn, dp->mp_flags == (P_LOOSE | P_DIRTY)); if ((env->me_flags & MDBX_WRITEMAP) == 0) mdbx_dpage_free(env, dp, 1); } } mdbx_trace("%s: filtered-out loose-pages from %u -> %u dirty-pages", dbg_prefix_mode, dl->length, w); mdbx_tassert(txn, txn->tw.loose_count == dl->length - w); dl->length = w; dl->sorted = 0; txn->tw.dirtyroom += txn->tw.loose_count; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); txn->tw.loose_pages = NULL; txn->tw.loose_count = 0; #if MDBX_ENABLE_REFUND txn->tw.loose_refund_wl = 0; #endif /* MDBX_ENABLE_REFUND */ } const unsigned amount = (unsigned)MDBX_PNL_SIZE(txn->tw.reclaimed_pglist); /* handle retired-list - store ones into single gc-record */ if (retired_stored < MDBX_PNL_SIZE(txn->tw.retired_pages)) { if (unlikely(!retired_stored)) { /* Make sure last page of GC is touched and on retired-list */ couple.outer.mc_flags &= ~C_RECLAIMING; rc = mdbx_page_search(&couple.outer, NULL, MDBX_PS_LAST | MDBX_PS_MODIFY); couple.outer.mc_flags |= C_RECLAIMING; if (unlikely(rc != MDBX_SUCCESS) && rc != MDBX_NOTFOUND) goto bailout; } /* Write to last page of GC */ key.iov_len = sizeof(txn->mt_txnid); key.iov_base = &txn->mt_txnid; do { data.iov_len = MDBX_PNL_SIZEOF(txn->tw.retired_pages); mdbx_prep_backlog(txn, &couple.outer, data.iov_len); rc = mdbx_cursor_put(&couple.outer, &key, &data, MDBX_RESERVE); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; /* Retry if tw.retired_pages[] grew during the Put() */ } while (data.iov_len < MDBX_PNL_SIZEOF(txn->tw.retired_pages)); retired_stored = (unsigned)MDBX_PNL_SIZE(txn->tw.retired_pages); mdbx_pnl_sort(txn->tw.retired_pages); mdbx_assert(env, data.iov_len == MDBX_PNL_SIZEOF(txn->tw.retired_pages)); memcpy(data.iov_base, txn->tw.retired_pages, data.iov_len); mdbx_trace("%s.put-retired #%u @ %" PRIaTXN, dbg_prefix_mode, retired_stored, txn->mt_txnid); if (mdbx_log_enabled(MDBX_LOG_EXTRA)) { unsigned i = retired_stored; mdbx_debug_extra("PNL write txn %" PRIaTXN " root %" PRIaPGNO " num %u, PNL", txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i); for (; i; i--) mdbx_debug_extra_print(" %" PRIaPGNO, txn->tw.retired_pages[i]); mdbx_debug_extra_print("%s\n", "."); } if (unlikely(amount != MDBX_PNL_SIZE(txn->tw.reclaimed_pglist))) { mdbx_trace("%s.reclaimed-list changed %u -> %u, retry", dbg_prefix_mode, amount, (unsigned)MDBX_PNL_SIZE(txn->tw.reclaimed_pglist)); goto retry_noaccount /* rare case, but avoids GC fragmentation and one cycle. */ ; } continue; } /* handle reclaimed and lost pages - merge and store both into gc */ mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); mdbx_tassert(txn, txn->tw.loose_count == 0); mdbx_trace("%s", " >> reserving"); if (mdbx_audit_enabled()) { rc = mdbx_audit_ex(txn, retired_stored, false); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } const unsigned left = amount - settled; mdbx_trace("%s: amount %u, settled %d, left %d, lifo-reclaimed-slots %u, " "reused-gc-slots %u", dbg_prefix_mode, amount, settled, (int)left, txn->tw.lifo_reclaimed ? (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : 0, reused_gc_slot); if (0 >= (int)left) break; const unsigned prefer_max_scatter = 257; txnid_t reservation_gc_id; if (lifo) { if (txn->tw.lifo_reclaimed == nullptr) { txn->tw.lifo_reclaimed = mdbx_txl_alloc(); if (unlikely(!txn->tw.lifo_reclaimed)) { rc = MDBX_ENOMEM; goto bailout; } } if ((unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) < prefer_max_scatter && left > ((unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot) * env->me_maxgc_ov1page) { /* LY: need just a txn-id for save page list. */ bool need_cleanup = false; txnid_t snap_oldest; retry_rid: couple.outer.mc_flags &= ~C_RECLAIMING; do { snap_oldest = mdbx_find_oldest(txn); rc = mdbx_page_alloc(&couple.outer, 0, NULL, MDBX_ALLOC_GC); if (likely(rc == MDBX_SUCCESS)) { mdbx_trace("%s: took @%" PRIaTXN " from GC", dbg_prefix_mode, MDBX_PNL_LAST(txn->tw.lifo_reclaimed)); need_cleanup = true; } } while (rc == MDBX_SUCCESS && (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) < prefer_max_scatter && left > ((unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot) * env->me_maxgc_ov1page); couple.outer.mc_flags |= C_RECLAIMING; if (likely(rc == MDBX_SUCCESS)) { mdbx_trace("%s: got enough from GC.", dbg_prefix_mode); continue; } else if (unlikely(rc != MDBX_NOTFOUND)) /* LY: some troubles... */ goto bailout; if (MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)) { if (need_cleanup) { mdbx_txl_sort(txn->tw.lifo_reclaimed); cleaned_gc_slot = 0; } gc_rid = MDBX_PNL_LAST(txn->tw.lifo_reclaimed); } else { mdbx_tassert(txn, txn->tw.last_reclaimed == 0); if (unlikely(mdbx_find_oldest(txn) != snap_oldest)) /* should retry mdbx_page_alloc(MDBX_ALLOC_GC) * if the oldest reader changes since the last attempt */ goto retry_rid; /* no reclaimable GC entries, * therefore no entries with ID < mdbx_find_oldest(txn) */ txn->tw.last_reclaimed = gc_rid = snap_oldest - 1; mdbx_trace("%s: none recycled yet, set rid to @%" PRIaTXN, dbg_prefix_mode, gc_rid); } /* LY: GC is empty, will look any free txn-id in high2low order. */ while (MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) < prefer_max_scatter && left > ((unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot) * env->me_maxgc_ov1page) { if (unlikely(gc_rid < 2)) { if (unlikely(MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) <= reused_gc_slot)) { mdbx_notice("** restart: reserve depleted (reused_gc_slot %u >= " "lifo_reclaimed %u" PRIaTXN, reused_gc_slot, (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)); goto retry; } break; } mdbx_tassert(txn, gc_rid >= MIN_TXNID && gc_rid <= MAX_TXNID); --gc_rid; key.iov_base = &gc_rid; key.iov_len = sizeof(gc_rid); rc = mdbx_cursor_get(&couple.outer, &key, &data, MDBX_SET_KEY); if (unlikely(rc == MDBX_SUCCESS)) { mdbx_debug("%s: GC's id %" PRIaTXN " is used, continue bottom-up search", dbg_prefix_mode, gc_rid); ++gc_rid; rc = mdbx_cursor_get(&couple.outer, &key, &data, MDBX_FIRST); if (rc == MDBX_NOTFOUND) { mdbx_debug("%s: GC is empty", dbg_prefix_mode); break; } if (unlikely(rc != MDBX_SUCCESS || key.iov_len != sizeof(mdbx_tid_t))) { rc = MDBX_CORRUPTED; goto bailout; } txnid_t gc_first = unaligned_peek_u64(4, key.iov_base); if (unlikely(gc_first < MIN_TXNID || gc_first > MAX_TXNID)) { rc = MDBX_CORRUPTED; goto bailout; } if (gc_first < 2) { mdbx_debug("%s: no free GC's id(s) less than %" PRIaTXN, dbg_prefix_mode, gc_rid); break; } gc_rid = gc_first - 1; } rc = mdbx_txl_append(&txn->tw.lifo_reclaimed, gc_rid); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; if (reused_gc_slot) /* rare case, but it is better to clear and re-create GC entries * with less fragmentation. */ need_cleanup = true; else cleaned_gc_slot += 1 /* mark cleanup is not needed for added slot. */; mdbx_trace("%s: append @%" PRIaTXN " to lifo-reclaimed, cleaned-gc-slot = %u", dbg_prefix_mode, gc_rid, cleaned_gc_slot); } if (need_cleanup) { cleaned_gc_slot = 0; mdbx_trace("%s: restart inner-loop to clear and re-create GC entries", dbg_prefix_mode); continue; } } const unsigned i = (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot; mdbx_tassert(txn, i > 0 && i <= MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)); reservation_gc_id = txn->tw.lifo_reclaimed[i]; mdbx_trace("%s: take @%" PRIaTXN " from lifo-reclaimed[%u]", dbg_prefix_mode, reservation_gc_id, i); } else { mdbx_tassert(txn, txn->tw.lifo_reclaimed == NULL); if (unlikely(gc_rid == 0)) { gc_rid = mdbx_find_oldest(txn) - 1; rc = mdbx_cursor_get(&couple.outer, &key, NULL, MDBX_FIRST); if (rc == MDBX_SUCCESS) { if (unlikely(key.iov_len != sizeof(txnid_t))) { rc = MDBX_CORRUPTED; goto bailout; } txnid_t gc_first = unaligned_peek_u64(4, key.iov_base); if (unlikely(gc_first < MIN_TXNID || gc_first > MAX_TXNID)) { rc = MDBX_CORRUPTED; goto bailout; } if (gc_rid >= gc_first) gc_rid = gc_first - 1; if (unlikely(gc_rid == 0)) { mdbx_error("%s", "** no GC tail-space to store"); goto retry; } } else if (rc != MDBX_NOTFOUND) goto bailout; txn->tw.last_reclaimed = gc_rid; } reservation_gc_id = gc_rid--; mdbx_trace("%s: take @%" PRIaTXN " from head-gc-id", dbg_prefix_mode, reservation_gc_id); } ++reused_gc_slot; unsigned chunk = left; if (unlikely(chunk > env->me_maxgc_ov1page)) { const unsigned avail_gc_slots = txn->tw.lifo_reclaimed ? (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot + 1 : (gc_rid < INT16_MAX) ? (unsigned)gc_rid : INT16_MAX; if (avail_gc_slots > 1) { if (chunk < env->me_maxgc_ov1page * 2) chunk /= 2; else { const unsigned threshold = env->me_maxgc_ov1page * ((avail_gc_slots < prefer_max_scatter) ? avail_gc_slots : prefer_max_scatter); if (left < threshold) chunk = env->me_maxgc_ov1page; else { const unsigned tail = left - threshold + env->me_maxgc_ov1page + 1; unsigned span = 1; unsigned avail = (unsigned)((pgno2bytes(env, span) - PAGEHDRSZ) / sizeof(pgno_t)) /*- 1 + span */; if (tail > avail) { for (unsigned i = amount - span; i > 0; --i) { if (MDBX_PNL_ASCENDING ? (txn->tw.reclaimed_pglist[i] + span) : (txn->tw.reclaimed_pglist[i] - span) == txn->tw.reclaimed_pglist[i + span]) { span += 1; avail = (unsigned)((pgno2bytes(env, span) - PAGEHDRSZ) / sizeof(pgno_t)) - 1 + span; if (avail >= tail) break; } } } chunk = (avail >= tail) ? tail - span : (avail_gc_slots > 3 && reused_gc_slot < prefer_max_scatter - 3) ? avail - span : tail; } } } } mdbx_tassert(txn, chunk > 0); mdbx_trace("%s: gc_rid %" PRIaTXN ", reused_gc_slot %u, reservation-id " "%" PRIaTXN, dbg_prefix_mode, gc_rid, reused_gc_slot, reservation_gc_id); mdbx_trace("%s: chunk %u, gc-per-ovpage %u", dbg_prefix_mode, chunk, env->me_maxgc_ov1page); mdbx_tassert(txn, reservation_gc_id < env->me_oldest->weak); if (unlikely(reservation_gc_id < 1 || reservation_gc_id >= atomic_load64(env->me_oldest, mo_Relaxed))) { mdbx_error("%s", "** internal error (reservation_gc_id)"); rc = MDBX_PROBLEM; goto bailout; } key.iov_len = sizeof(reservation_gc_id); key.iov_base = &reservation_gc_id; data.iov_len = (chunk + 1) * sizeof(pgno_t); mdbx_trace("%s.reserve: %u [%u...%u] @%" PRIaTXN, dbg_prefix_mode, chunk, settled + 1, settled + chunk + 1, reservation_gc_id); mdbx_prep_backlog(txn, &couple.outer, data.iov_len); rc = mdbx_cursor_put(&couple.outer, &key, &data, MDBX_RESERVE | MDBX_NOOVERWRITE); mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; clean_reserved_gc_pnl(env, data); settled += chunk; mdbx_trace("%s.settled %u (+%u), continue", dbg_prefix_mode, settled, chunk); if (txn->tw.lifo_reclaimed && unlikely(amount < MDBX_PNL_SIZE(txn->tw.reclaimed_pglist))) { mdbx_notice("** restart: reclaimed-list growth %u -> %u", amount, (unsigned)MDBX_PNL_SIZE(txn->tw.reclaimed_pglist)); goto retry_noaccount; } continue; } mdbx_tassert( txn, cleaned_gc_slot == (txn->tw.lifo_reclaimed ? MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : 0)); mdbx_trace("%s", " >> filling"); /* Fill in the reserved records */ filled_gc_slot = txn->tw.lifo_reclaimed ? (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) - reused_gc_slot : reused_gc_slot; rc = MDBX_SUCCESS; mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.reclaimed_pglist, txn->mt_next_pgno - MDBX_ENABLE_REFUND)); mdbx_tassert(txn, mdbx_dirtylist_check(txn)); if (MDBX_PNL_SIZE(txn->tw.reclaimed_pglist)) { MDBX_val key, data; key.iov_len = data.iov_len = 0; /* avoid MSVC warning */ key.iov_base = data.iov_base = NULL; const unsigned amount = MDBX_PNL_SIZE(txn->tw.reclaimed_pglist); unsigned left = amount; if (txn->tw.lifo_reclaimed == nullptr) { mdbx_tassert(txn, lifo == 0); rc = mdbx_cursor_first(&couple.outer, &key, &data); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } else { mdbx_tassert(txn, lifo != 0); } while (true) { txnid_t fill_gc_id; mdbx_trace("%s: left %u of %u", dbg_prefix_mode, left, (unsigned)MDBX_PNL_SIZE(txn->tw.reclaimed_pglist)); if (txn->tw.lifo_reclaimed == nullptr) { mdbx_tassert(txn, lifo == 0); fill_gc_id = unaligned_peek_u64(4, key.iov_base); if (filled_gc_slot-- == 0 || fill_gc_id > txn->tw.last_reclaimed) { mdbx_notice( "** restart: reserve depleted (filled_slot %u, fill_id %" PRIaTXN " > last_reclaimed %" PRIaTXN, filled_gc_slot, fill_gc_id, txn->tw.last_reclaimed); goto retry; } } else { mdbx_tassert(txn, lifo != 0); if (++filled_gc_slot > (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)) { mdbx_notice("** restart: reserve depleted (filled_gc_slot %u > " "lifo_reclaimed %u" PRIaTXN, filled_gc_slot, (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)); goto retry; } fill_gc_id = txn->tw.lifo_reclaimed[filled_gc_slot]; mdbx_trace("%s.seek-reservation @%" PRIaTXN " at lifo_reclaimed[%u]", dbg_prefix_mode, fill_gc_id, filled_gc_slot); key.iov_base = &fill_gc_id; key.iov_len = sizeof(fill_gc_id); rc = mdbx_cursor_get(&couple.outer, &key, &data, MDBX_SET_KEY); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } mdbx_tassert(txn, cleaned_gc_slot == (txn->tw.lifo_reclaimed ? MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : 0)); mdbx_tassert(txn, fill_gc_id > 0 && fill_gc_id < env->me_oldest->weak); key.iov_base = &fill_gc_id; key.iov_len = sizeof(fill_gc_id); mdbx_tassert(txn, data.iov_len >= sizeof(pgno_t) * 2); couple.outer.mc_flags |= C_GCFREEZE; unsigned chunk = (unsigned)(data.iov_len / sizeof(pgno_t)) - 1; if (unlikely(chunk > left)) { mdbx_trace("%s: chunk %u > left %u, @%" PRIaTXN, dbg_prefix_mode, chunk, left, fill_gc_id); if ((loop < 5 && chunk - left > loop / 2) || chunk - left > env->me_maxgc_ov1page) { data.iov_len = (left + 1) * sizeof(pgno_t); if (loop < 7) couple.outer.mc_flags &= ~C_GCFREEZE; } chunk = left; } rc = mdbx_cursor_put(&couple.outer, &key, &data, MDBX_CURRENT | MDBX_RESERVE); couple.outer.mc_flags &= ~C_GCFREEZE; if (unlikely(rc != MDBX_SUCCESS)) goto bailout; clean_reserved_gc_pnl(env, data); if (unlikely(txn->tw.loose_count || amount != MDBX_PNL_SIZE(txn->tw.reclaimed_pglist))) { mdbx_notice("** restart: reclaimed-list growth (%u -> %u, loose +%u)", amount, MDBX_PNL_SIZE(txn->tw.reclaimed_pglist), txn->tw.loose_count); goto retry; } if (unlikely(txn->tw.lifo_reclaimed ? cleaned_gc_slot < MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : cleaned_gc_id < txn->tw.last_reclaimed)) { mdbx_notice("%s", "** restart: reclaimed-slots changed"); goto retry; } if (unlikely(retired_stored != MDBX_PNL_SIZE(txn->tw.retired_pages))) { mdbx_tassert(txn, retired_stored < MDBX_PNL_SIZE(txn->tw.retired_pages)); mdbx_notice("** restart: retired-list growth (%u -> %u)", retired_stored, MDBX_PNL_SIZE(txn->tw.retired_pages)); goto retry; } pgno_t *dst = data.iov_base; *dst++ = chunk; pgno_t *src = MDBX_PNL_BEGIN(txn->tw.reclaimed_pglist) + left - chunk; memcpy(dst, src, chunk * sizeof(pgno_t)); pgno_t *from = src, *to = src + chunk; mdbx_trace("%s.fill: %u [ %u:%" PRIaPGNO "...%u:%" PRIaPGNO "] @%" PRIaTXN, dbg_prefix_mode, chunk, (unsigned)(from - txn->tw.reclaimed_pglist), from[0], (unsigned)(to - txn->tw.reclaimed_pglist), to[-1], fill_gc_id); left -= chunk; if (mdbx_audit_enabled()) { rc = mdbx_audit_ex(txn, retired_stored + amount - left, true); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } if (left == 0) { rc = MDBX_SUCCESS; break; } if (txn->tw.lifo_reclaimed == nullptr) { mdbx_tassert(txn, lifo == 0); rc = mdbx_cursor_next(&couple.outer, &key, &data, MDBX_NEXT); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } else { mdbx_tassert(txn, lifo != 0); } } } mdbx_tassert(txn, rc == MDBX_SUCCESS); if (unlikely(txn->tw.loose_count != 0 || filled_gc_slot != (txn->tw.lifo_reclaimed ? (unsigned)MDBX_PNL_SIZE(txn->tw.lifo_reclaimed) : 0))) { mdbx_notice("** restart: reserve excess (filled-slot %u, loose-count %u)", filled_gc_slot, txn->tw.loose_count); goto retry; } mdbx_tassert(txn, txn->tw.lifo_reclaimed == NULL || cleaned_gc_slot == MDBX_PNL_SIZE(txn->tw.lifo_reclaimed)); bailout: txn->tw.cursors[FREE_DBI] = couple.outer.mc_next; bailout_notracking: MDBX_PNL_SIZE(txn->tw.reclaimed_pglist) = 0; mdbx_trace("<<< %u loops, rc = %d", loop, rc); return rc; } static int mdbx_flush_iov(MDBX_txn *const txn, struct iovec *iov, unsigned iov_items, size_t iov_off, size_t iov_bytes) { MDBX_env *const env = txn->mt_env; mdbx_assert(env, iov_items > 0); int rc; if (likely(iov_items == 1)) { mdbx_assert(env, iov->iov_len == iov_bytes); rc = mdbx_pwrite(env->me_lazy_fd, iov->iov_base, iov_bytes, iov_off); } else { rc = mdbx_pwritev(env->me_lazy_fd, iov, iov_items, iov_off, iov_bytes); } if (unlikely(rc != MDBX_SUCCESS)) mdbx_error("Write error: %s", mdbx_strerror(rc)); else { VALGRIND_MAKE_MEM_DEFINED(txn->mt_env->me_map + iov_off, iov_bytes); ASAN_UNPOISON_MEMORY_REGION(txn->mt_env->me_map + iov_off, iov_bytes); } for (unsigned i = 0; i < iov_items; i++) mdbx_dpage_free(env, (MDBX_page *)iov[i].iov_base, bytes2pgno(env, iov[i].iov_len)); return rc; } /* Flush (some) dirty pages to the map, after clearing their dirty flag. * [in] txn the transaction that's being committed * [in] keep number of initial pages in dirtylist to keep dirty. * Returns 0 on success, non-zero on failure. */ __hot static int mdbx_page_flush(MDBX_txn *txn, const unsigned keep) { struct iovec iov[MDBX_COMMIT_PAGES]; MDBX_dpl *const dl = mdbx_dpl_sort(txn->tw.dirtylist); MDBX_env *const env = txn->mt_env; pgno_t flush_begin = MAX_PAGENO; pgno_t flush_end = MIN_PAGENO; int rc = MDBX_SUCCESS; unsigned iov_items = 0; size_t iov_bytes = 0; size_t iov_off = 0; unsigned r, w; for (r = w = keep; ++r <= dl->length;) { MDBX_page *dp = dl->items[r].ptr; mdbx_tassert(txn, dp->mp_pgno >= MIN_PAGENO && dp->mp_pgno < txn->mt_next_pgno); mdbx_tassert(txn, dp->mp_flags & P_DIRTY); /* Don't flush this page yet */ if (dp->mp_flags & P_KEEP) { dp->mp_flags -= P_KEEP; dl->items[++w] = dl->items[r]; continue; } if (dp->mp_flags & P_LOOSE) { dl->items[++w] = dl->items[r]; continue; } const unsigned npages = IS_OVERFLOW(dp) ? dp->mp_pages : 1; flush_begin = (flush_begin < dp->mp_pgno) ? flush_begin : dp->mp_pgno; flush_end = (flush_end > dp->mp_pgno + npages) ? flush_end : dp->mp_pgno + npages; env->me_unsynced_pages->weak += npages; dp->mp_flags -= P_DIRTY; dp->mp_txnid = pp_txnid2chk(txn); if ((env->me_flags & MDBX_WRITEMAP) == 0) { const size_t size = pgno2bytes(env, npages); if (iov_off + iov_bytes != pgno2bytes(env, dp->mp_pgno) || iov_items == ARRAY_LENGTH(iov) || iov_bytes + size > MAX_WRITE) { if (iov_items) { rc = mdbx_flush_iov(txn, iov, iov_items, iov_off, iov_bytes); #if defined(__linux__) || defined(__gnu_linux__) if (mdbx_linux_kernel_version >= 0x02060b00) /* Linux kernels older than version 2.6.11 ignore the addr and nbytes * arguments, making this function fairly expensive. Therefore, the * whole cache is always flushed. */ #endif /* Linux */ mdbx_flush_incoherent_mmap(env->me_map + iov_off, iov_bytes, env->me_os_psize); iov_items = 0; iov_bytes = 0; if (unlikely(rc != MDBX_SUCCESS)) { do dl->items[++w] = dl->items[r]; while (++r <= dl->length); break; } } iov_off = pgno2bytes(env, dp->mp_pgno); } iov[iov_items].iov_base = (void *)dp; iov[iov_items].iov_len = size; iov_items += 1; iov_bytes += size; } } mdbx_tassert(txn, dl->sorted == dl->length && r == dl->length + 1); txn->tw.dirtyroom += dl->length - w; assert(txn->tw.dirtyroom <= txn->mt_env->me_options.dp_limit); dl->sorted = dl->length = w; mdbx_tassert(txn, txn->mt_parent || txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); if (iov_items) rc = mdbx_flush_iov(txn, iov, iov_items, iov_off, iov_bytes); if (unlikely(rc != MDBX_SUCCESS)) { txn->mt_flags |= MDBX_TXN_ERROR; return rc; } #if defined(__linux__) || defined(__gnu_linux__) if ((env->me_flags & MDBX_WRITEMAP) == 0 && mdbx_linux_kernel_version < 0x02060b00) /* Linux kernels older than version 2.6.11 ignore the addr and nbytes * arguments, making this function fairly expensive. Therefore, the * whole cache is always flushed. */ mdbx_flush_incoherent_mmap(env->me_map + pgno2bytes(env, flush_begin), pgno2bytes(env, flush_end - flush_begin), env->me_os_psize); #endif /* Linux */ /* TODO: use flush_begin & flush_end for msync() & sync_file_range(). */ (void)flush_begin; (void)flush_end; return MDBX_SUCCESS; } /* Check txn and dbi arguments to a function */ static __always_inline bool mdbx_txn_dbi_exists(MDBX_txn *txn, MDBX_dbi dbi, unsigned validity) { if (likely(dbi < txn->mt_numdbs && (txn->mt_dbistate[dbi] & validity))) return true; return dbi_import(txn, dbi); } int mdbx_txn_commit(MDBX_txn *txn) { return __inline_mdbx_txn_commit(txn); } /* Merge child txn into parent */ static __inline void mdbx_txn_merge(MDBX_txn *const parent, MDBX_txn *const txn, const unsigned parent_retired_len) { MDBX_dpl *const src = mdbx_dpl_sort(txn->tw.dirtylist); /* Remove refunded pages from parent's dirty list */ MDBX_dpl *const dst = mdbx_dpl_sort(parent->tw.dirtylist); while (MDBX_ENABLE_REFUND && dst->length && dst->items[dst->length].pgno >= parent->mt_next_pgno) { if (!(txn->mt_env->me_flags & MDBX_WRITEMAP)) { MDBX_page *dp = dst->items[dst->length].ptr; mdbx_dpage_free(txn->mt_env, dp, IS_OVERFLOW(dp) ? dp->mp_pages : 1); } dst->length -= 1; } parent->tw.dirtyroom += dst->sorted - dst->length; assert(parent->tw.dirtyroom <= parent->mt_env->me_options.dp_limit); dst->sorted = dst->length; mdbx_tassert(parent, parent->mt_parent || parent->tw.dirtyroom + parent->tw.dirtylist->length == parent->mt_env->me_options.dp_limit); /* Remove reclaimed pages from parent's dirty list */ const MDBX_PNL reclaimed_list = parent->tw.reclaimed_pglist; mdbx_dpl_sift(parent, reclaimed_list, false); /* Move retired pages from parent's dirty & spilled list to reclaimed */ unsigned r, w, d, s, l; for (r = w = parent_retired_len; ++r <= MDBX_PNL_SIZE(parent->tw.retired_pages);) { const pgno_t pgno = parent->tw.retired_pages[r]; const unsigned di = mdbx_dpl_exist(parent->tw.dirtylist, pgno); const unsigned si = (!di && unlikely(parent->tw.spill_pages)) ? mdbx_pnl_exist(parent->tw.spill_pages, pgno << 1) : 0; unsigned npages; const char *kind; if (di) { MDBX_page *dp = parent->tw.dirtylist->items[di].ptr; mdbx_tassert(parent, (dp->mp_flags & ~(P_LEAF | P_LEAF2 | P_BRANCH | P_OVERFLOW | P_DIRTY)) == 0); npages = IS_OVERFLOW(dp) ? dp->mp_pages : 1; mdbx_page_wash(parent, di, dp, npages); kind = "dirty"; l = 1; if (unlikely(npages > l)) { /* OVERFLOW-страница могла быть переиспользована по частям. Тогда * в retired-списке может быть только начало последовательности, * а остаток растащен по dirty, spilled и reclaimed спискам. Поэтому * переносим в reclaimed с проверкой на обрыв последовательности. * В любом случае, все осколки будут учтены и отфильтрованы, т.е. если * страница была разбита на части, то важно удалить dirty-элемент, * а все осколки будут учтены отдельно. */ /* Список retired страниц не сортирован, но для ускорения сортировки * дополняется в соответствии с MDBX_PNL_ASCENDING */ #if MDBX_PNL_ASCENDING const unsigned len = MDBX_PNL_SIZE(parent->tw.retired_pages); while (r < len && parent->tw.retired_pages[r + 1] == pgno + l) { ++r; if (++l == npages) break; } #else while (w > parent_retired_len && parent->tw.retired_pages[w - 1] == pgno + l) { --w; if (++l == npages) break; } #endif } } else if (unlikely(si)) { l = npages = 1; mdbx_spill_remove(parent, si, 1); kind = "spilled"; } else { parent->tw.retired_pages[++w] = pgno; continue; } mdbx_debug("reclaim retired parent's %u->%u %s page %" PRIaPGNO, npages, l, kind, pgno); int err = mdbx_pnl_insert_range(&parent->tw.reclaimed_pglist, pgno, l); mdbx_ensure(txn->mt_env, err == MDBX_SUCCESS); } MDBX_PNL_SIZE(parent->tw.retired_pages) = w; /* Filter-out parent spill list */ if (parent->tw.spill_pages && MDBX_PNL_SIZE(parent->tw.spill_pages) > 0) { const MDBX_PNL sl = mdbx_spill_purge(parent); unsigned len = MDBX_PNL_SIZE(sl); if (len) { /* Remove refunded pages from parent's spill list */ if (MDBX_ENABLE_REFUND && MDBX_PNL_MOST(sl) >= (parent->mt_next_pgno << 1)) { #if MDBX_PNL_ASCENDING unsigned i = MDBX_PNL_SIZE(sl); assert(MDBX_PNL_MOST(sl) == MDBX_PNL_LAST(sl)); do { if ((sl[i] & 1) == 0) mdbx_debug("refund parent's spilled page %" PRIaPGNO, sl[i] >> 1); i -= 1; } while (i && sl[i] >= (parent->mt_next_pgno << 1)); MDBX_PNL_SIZE(sl) = i; #else assert(MDBX_PNL_MOST(sl) == MDBX_PNL_FIRST(sl)); unsigned i = 0; do { ++i; if ((sl[i] & 1) == 0) mdbx_debug("refund parent's spilled page %" PRIaPGNO, sl[i] >> 1); } while (i < len && sl[i + 1] >= (parent->mt_next_pgno << 1)); MDBX_PNL_SIZE(sl) = len -= i; memmove(sl + 1, sl + 1 + i, len * sizeof(sl[0])); #endif } mdbx_tassert(txn, mdbx_pnl_check4assert(sl, parent->mt_next_pgno << 1)); /* Remove reclaimed pages from parent's spill list */ s = MDBX_PNL_SIZE(sl), r = MDBX_PNL_SIZE(reclaimed_list); /* Scanning from end to begin */ while (s && r) { if (sl[s] & 1) { --s; continue; } const pgno_t spilled_pgno = sl[s] >> 1; const pgno_t reclaimed_pgno = reclaimed_list[r]; if (reclaimed_pgno != spilled_pgno) { const bool cmp = MDBX_PNL_ORDERED(spilled_pgno, reclaimed_pgno); s -= !cmp; r -= cmp; } else { mdbx_debug("remove reclaimed parent's spilled page %" PRIaPGNO, reclaimed_pgno); mdbx_spill_remove(parent, s, 1); --s; --r; } } /* Remove anything in our dirty list from parent's spill list */ /* Scanning spill list in descend order */ const int step = MDBX_PNL_ASCENDING ? -1 : 1; s = MDBX_PNL_ASCENDING ? MDBX_PNL_SIZE(sl) : 1; d = src->length; while (d && (MDBX_PNL_ASCENDING ? s > 0 : s <= MDBX_PNL_SIZE(sl))) { if (sl[s] & 1) { s += step; continue; } const pgno_t spilled_pgno = sl[s] >> 1; const pgno_t dirty_pgno_form = src->items[d].pgno; MDBX_page *dp = src->items[d].ptr; const unsigned npages = IS_OVERFLOW(dp) ? dp->mp_pages : 1; const pgno_t dirty_pgno_to = dirty_pgno_form + npages; if (dirty_pgno_form > spilled_pgno) { --d; continue; } if (dirty_pgno_to <= spilled_pgno) { s += step; continue; } mdbx_debug("remove dirtied parent's spilled %u page %" PRIaPGNO, npages, dirty_pgno_form); mdbx_spill_remove(parent, s, 1); s += step; } /* Squash deleted pagenums if we deleted any */ mdbx_spill_purge(parent); } } /* Remove anything in our spill list from parent's dirty list */ if (txn->tw.spill_pages) { mdbx_tassert(txn, mdbx_pnl_check4assert(txn->tw.spill_pages, parent->mt_next_pgno << 1)); mdbx_dpl_sift(parent, txn->tw.spill_pages, true); mdbx_tassert(parent, parent->mt_parent || parent->tw.dirtyroom + parent->tw.dirtylist->length == parent->mt_env->me_options.dp_limit); } /* Find length of merging our dirty list with parent's and release * filter-out pages */ for (l = 0, d = dst->length, s = src->length; d > 0 && s > 0;) { const MDBX_page *sp = src->items[s].ptr; mdbx_tassert(parent, (sp->mp_flags & ~(P_LEAF | P_LEAF2 | P_BRANCH | P_OVERFLOW | P_DIRTY | P_LOOSE)) == 0); const unsigned s_npages = IS_OVERFLOW(sp) ? sp->mp_pages : 1; const pgno_t s_pgno = src->items[s].pgno; MDBX_page *dp = dst->items[d].ptr; mdbx_tassert(parent, (dp->mp_flags & ~(P_LEAF | P_LEAF2 | P_BRANCH | P_OVERFLOW | P_DIRTY)) == 0); const unsigned d_npages = IS_OVERFLOW(dp) ? dp->mp_pages : 1; const pgno_t d_pgno = dst->items[d].pgno; if (d_pgno >= s_pgno + s_npages) { --d; ++l; } else if (d_pgno + d_npages <= s_pgno) { --s; ++l; } else { dst->items[d--].ptr = nullptr; if ((txn->mt_flags & MDBX_WRITEMAP) == 0) mdbx_dpage_free(txn->mt_env, dp, d_npages); } } assert(dst->sorted == dst->length); mdbx_tassert(parent, dst->allocated >= l + d + s); dst->sorted = l + d + s; /* the merged length */ /* Merge our dirty list into parent's, i.e. merge(dst, src) -> dst */ if (dst->sorted >= dst->length) { /* from end to begin with dst extending */ for (l = dst->sorted, s = src->length, d = dst->length; s > 0 && d > 0;) { if (unlikely(l <= d)) { /* squash to get a gap of free space for merge */ for (r = w = 1; r <= d; ++r) if (dst->items[r].ptr) { if (w != r) { dst->items[w] = dst->items[r]; dst->items[r].ptr = nullptr; } ++w; } mdbx_notice("squash to begin for extending-merge %u -> %u", d, w - 1); d = w - 1; continue; } assert(l > d); if (dst->items[d].ptr) { dst->items[l--] = (dst->items[d].pgno > src->items[s].pgno) ? dst->items[d--] : src->items[s--]; } else --d; } if (s > 0) { assert(l == s); while (d > 0) { assert(dst->items[d].ptr == nullptr); --d; } do { assert(l > 0); dst->items[l--] = src->items[s--]; } while (s > 0); } else { assert(l == d); while (l > 0) { assert(dst->items[l].ptr != nullptr); --l; } } } else { /* from begin to end with dst shrinking (a lot of new overflow pages) */ for (l = s = d = 1; s <= src->length && d <= dst->length;) { if (unlikely(l >= d)) { /* squash to get a gap of free space for merge */ for (r = w = dst->length; r >= d; --r) if (dst->items[r].ptr) { if (w != r) { dst->items[w] = dst->items[r]; dst->items[r].ptr = nullptr; } --w; } mdbx_notice("squash to end for shrinking-merge %u -> %u", d, w + 1); d = w + 1; continue; } assert(l < d); if (dst->items[d].ptr) { dst->items[l++] = (dst->items[d].pgno < src->items[s].pgno) ? dst->items[d++] : src->items[s++]; } else ++d; } if (s <= src->length) { assert(dst->sorted - l == src->length - s); while (d <= dst->length) { assert(dst->items[d].ptr == nullptr); --d; } do { assert(l <= dst->sorted); dst->items[l++] = src->items[s++]; } while (s <= src->length); } else { assert(dst->sorted - l == dst->length - d); while (l <= dst->sorted) { assert(l <= d && d <= dst->length && dst->items[d].ptr); dst->items[l++] = dst->items[d++]; } } } parent->tw.dirtyroom -= dst->sorted - dst->length; assert(parent->tw.dirtyroom <= parent->mt_env->me_options.dp_limit); dst->length = dst->sorted; mdbx_tassert(parent, parent->mt_parent || parent->tw.dirtyroom + parent->tw.dirtylist->length == parent->mt_env->me_options.dp_limit); mdbx_tassert(parent, mdbx_dirtylist_check(parent)); mdbx_dpl_free(txn); if (txn->tw.spill_pages) { if (parent->tw.spill_pages) { /* Must not fail since space was preserved above. */ mdbx_pnl_xmerge(parent->tw.spill_pages, txn->tw.spill_pages); mdbx_pnl_free(txn->tw.spill_pages); } else { parent->tw.spill_pages = txn->tw.spill_pages; parent->tw.spill_least_removed = txn->tw.spill_least_removed; } mdbx_tassert(parent, mdbx_dirtylist_check(parent)); } parent->mt_flags &= ~MDBX_TXN_HAS_CHILD; if (parent->tw.spill_pages) { assert(mdbx_pnl_check4assert(parent->tw.spill_pages, parent->mt_next_pgno << 1)); if (MDBX_PNL_SIZE(parent->tw.spill_pages)) parent->mt_flags |= MDBX_TXN_SPILLS; } } int mdbx_txn_commit_ex(MDBX_txn *txn, MDBX_commit_latency *latency) { STATIC_ASSERT(MDBX_TXN_FINISHED == MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD - MDBX_TXN_ERROR); const uint64_t ts_0 = latency ? mdbx_osal_monotime() : 0; uint64_t ts_1 = 0, ts_2 = 0, ts_3 = 0, ts_4 = 0; uint32_t audit_duration = 0; int rc = check_txn(txn, MDBX_TXN_FINISHED); if (unlikely(rc != MDBX_SUCCESS)) goto provide_latency; if (unlikely(txn->mt_flags & MDBX_TXN_ERROR)) { rc = MDBX_RESULT_TRUE; goto fail; } MDBX_env *env = txn->mt_env; #if MDBX_ENV_CHECKPID if (unlikely(env->me_pid != mdbx_getpid())) { env->me_flags |= MDBX_FATAL_ERROR; rc = MDBX_PANIC; goto provide_latency; } #endif /* MDBX_ENV_CHECKPID */ /* mdbx_txn_end() mode for a commit which writes nothing */ unsigned end_mode = MDBX_END_PURE_COMMIT | MDBX_END_UPDATE | MDBX_END_SLOT | MDBX_END_FREE; if (unlikely(F_ISSET(txn->mt_flags, MDBX_TXN_RDONLY))) goto done; if (txn->mt_child) { rc = mdbx_txn_commit_ex(txn->mt_child, NULL); mdbx_tassert(txn, txn->mt_child == NULL); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } if (unlikely(txn != env->me_txn)) { mdbx_debug("%s", "attempt to commit unknown transaction"); rc = MDBX_EINVAL; goto fail; } if (txn->mt_parent) { mdbx_tassert(txn, mdbx_audit_ex(txn, 0, false) == 0); mdbx_assert(env, txn != env->me_txn0); MDBX_txn *const parent = txn->mt_parent; mdbx_assert(env, parent->mt_signature == MDBX_MT_SIGNATURE); mdbx_assert(env, parent->mt_child == txn && (parent->mt_flags & MDBX_TXN_HAS_CHILD) != 0); mdbx_assert(env, mdbx_dirtylist_check(txn)); if (txn->tw.dirtylist->length == 0 && !(txn->mt_flags & MDBX_TXN_DIRTY) && parent->mt_numdbs == txn->mt_numdbs) { for (int i = txn->mt_numdbs; --i >= 0;) { mdbx_tassert(txn, (txn->mt_dbistate[i] & DBI_DIRTY) == 0); if ((txn->mt_dbistate[i] & DBI_STALE) && !(parent->mt_dbistate[i] & DBI_STALE)) mdbx_tassert(txn, memcmp(&parent->mt_dbs[i], &txn->mt_dbs[i], sizeof(MDBX_db)) == 0); } mdbx_tassert(txn, memcmp(&parent->mt_geo, &txn->mt_geo, sizeof(parent->mt_geo)) == 0); mdbx_tassert(txn, memcmp(&parent->mt_canary, &txn->mt_canary, sizeof(parent->mt_canary)) == 0); mdbx_tassert(txn, !txn->tw.spill_pages || MDBX_PNL_SIZE(txn->tw.spill_pages) == 0); mdbx_tassert(txn, txn->tw.loose_count == 0); /* fast completion of pure nested transaction */ end_mode = MDBX_END_PURE_COMMIT | MDBX_END_SLOT | MDBX_END_FREE; goto done; } /* Preserve space for spill list to avoid parent's state corruption * if allocation fails. */ const unsigned parent_retired_len = (unsigned)(uintptr_t)parent->tw.retired_pages; mdbx_tassert(txn, parent_retired_len <= MDBX_PNL_SIZE(txn->tw.retired_pages)); const unsigned retired_delta = MDBX_PNL_SIZE(txn->tw.retired_pages) - parent_retired_len; if (retired_delta) { rc = mdbx_pnl_need(&txn->tw.reclaimed_pglist, retired_delta); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } if (txn->tw.spill_pages) { if (parent->tw.spill_pages) { rc = mdbx_pnl_need(&parent->tw.spill_pages, MDBX_PNL_SIZE(txn->tw.spill_pages)); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } mdbx_spill_purge(txn); } if (unlikely(txn->tw.dirtylist->length + parent->tw.dirtylist->length > parent->tw.dirtylist->allocated && !mdbx_dpl_reserve(parent, txn->tw.dirtylist->length + parent->tw.dirtylist->length))) { rc = MDBX_ENOMEM; goto fail; } //------------------------------------------------------------------------- parent->tw.lifo_reclaimed = txn->tw.lifo_reclaimed; txn->tw.lifo_reclaimed = NULL; parent->tw.retired_pages = txn->tw.retired_pages; txn->tw.retired_pages = NULL; mdbx_pnl_free(parent->tw.reclaimed_pglist); parent->tw.reclaimed_pglist = txn->tw.reclaimed_pglist; txn->tw.reclaimed_pglist = NULL; parent->tw.last_reclaimed = txn->tw.last_reclaimed; parent->mt_geo = txn->mt_geo; parent->mt_canary = txn->mt_canary; parent->mt_flags |= txn->mt_flags & MDBX_TXN_DIRTY; /* Move loose pages to parent */ #if MDBX_ENABLE_REFUND parent->tw.loose_refund_wl = txn->tw.loose_refund_wl; #endif /* MDBX_ENABLE_REFUND */ parent->tw.loose_count = txn->tw.loose_count; parent->tw.loose_pages = txn->tw.loose_pages; /* Merge our cursors into parent's and close them */ mdbx_cursors_eot(txn, true); end_mode |= MDBX_END_EOTDONE; /* Update parent's DBs array */ memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDBX_db)); parent->mt_numdbs = txn->mt_numdbs; parent->mt_dbistate[FREE_DBI] = txn->mt_dbistate[FREE_DBI]; parent->mt_dbistate[MAIN_DBI] = txn->mt_dbistate[MAIN_DBI]; for (unsigned i = CORE_DBS; i < txn->mt_numdbs; i++) { /* preserve parent's status */ const uint8_t state = txn->mt_dbistate[i] | (parent->mt_dbistate[i] & (DBI_CREAT | DBI_FRESH | DBI_DIRTY)); mdbx_debug("db %u dbi-state %s 0x%02x -> 0x%02x", i, (parent->mt_dbistate[i] != state) ? "update" : "still", parent->mt_dbistate[i], state); parent->mt_dbistate[i] = state; } ts_1 = latency ? mdbx_osal_monotime() : 0; mdbx_txn_merge(parent, txn, parent_retired_len); ts_2 = latency ? mdbx_osal_monotime() : 0; env->me_txn = parent; parent->mt_child = NULL; mdbx_tassert(parent, mdbx_dirtylist_check(parent)); #if MDBX_ENABLE_REFUND mdbx_refund(parent); if (mdbx_assert_enabled()) { /* Check parent's loose pages not suitable for refund */ for (MDBX_page *lp = parent->tw.loose_pages; lp; lp = lp->mp_next) mdbx_tassert(parent, lp->mp_pgno < parent->tw.loose_refund_wl && lp->mp_pgno + 1 < parent->mt_next_pgno); /* Check parent's reclaimed pages not suitable for refund */ if (MDBX_PNL_SIZE(parent->tw.reclaimed_pglist)) mdbx_tassert(parent, MDBX_PNL_MOST(parent->tw.reclaimed_pglist) + 1 < parent->mt_next_pgno); } #endif /* MDBX_ENABLE_REFUND */ ts_4 = ts_3 = latency ? mdbx_osal_monotime() : 0; txn->mt_signature = 0; mdbx_free(txn); mdbx_tassert(parent, mdbx_audit_ex(parent, 0, false) == 0); rc = MDBX_SUCCESS; goto provide_latency; } mdbx_tassert(txn, txn->tw.dirtyroom + txn->tw.dirtylist->length == txn->mt_env->me_options.dp_limit); mdbx_cursors_eot(txn, false); end_mode |= MDBX_END_EOTDONE; if (txn->tw.dirtylist->length == 0 && (txn->mt_flags & (MDBX_TXN_DIRTY | MDBX_TXN_SPILLS)) == 0) { for (int i = txn->mt_numdbs; --i >= 0;) mdbx_tassert(txn, (txn->mt_dbistate[i] & DBI_DIRTY) == 0); rc = MDBX_SUCCESS; goto done; } mdbx_debug("committing txn %" PRIaTXN " %p on mdbenv %p, root page %" PRIaPGNO "/%" PRIaPGNO, txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root, txn->mt_dbs[FREE_DBI].md_root); /* Update DB root pointers */ if (txn->mt_numdbs > CORE_DBS) { MDBX_cursor_couple couple; MDBX_val data; data.iov_len = sizeof(MDBX_db); rc = mdbx_cursor_init(&couple.outer, txn, MAIN_DBI); if (unlikely(rc != MDBX_SUCCESS)) goto fail; for (MDBX_dbi i = CORE_DBS; i < txn->mt_numdbs; i++) { if (txn->mt_dbistate[i] & DBI_DIRTY) { if (unlikely(TXN_DBI_CHANGED(txn, i))) { rc = MDBX_BAD_DBI; goto fail; } MDBX_db *db = &txn->mt_dbs[i]; mdbx_debug("update main's entry for sub-db %u, mod_txnid %" PRIaTXN " -> %" PRIaTXN, i, pp_txnid2chk(txn), db->md_mod_txnid); db->md_mod_txnid = pp_txnid2chk(txn); data.iov_base = db; WITH_CURSOR_TRACKING(couple.outer, rc = mdbx_cursor_put(&couple.outer, &txn->mt_dbxs[i].md_name, &data, F_SUBDATA)); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } } } ts_1 = latency ? mdbx_osal_monotime() : 0; rc = mdbx_update_gc(txn); if (unlikely(rc != MDBX_SUCCESS)) goto fail; ts_2 = latency ? mdbx_osal_monotime() : 0; if (mdbx_audit_enabled()) { rc = mdbx_audit_ex(txn, MDBX_PNL_SIZE(txn->tw.retired_pages), true); const uint64_t audit_end = mdbx_osal_monotime(); audit_duration = mdbx_osal_monotime_to_16dot16(audit_end - ts_2); ts_2 = audit_end; if (unlikely(rc != MDBX_SUCCESS)) goto fail; } rc = mdbx_page_flush(txn, 0); ts_3 = latency ? mdbx_osal_monotime() : 0; if (likely(rc == MDBX_SUCCESS)) { if (txn->mt_dbs[MAIN_DBI].md_flags & DBI_DIRTY) txn->mt_dbs[MAIN_DBI].md_mod_txnid = pp_txnid2chk(txn); MDBX_meta meta, *head = mdbx_meta_head(env); memcpy(meta.mm_magic_and_version, head->mm_magic_and_version, 8); meta.mm_extra_flags = head->mm_extra_flags; meta.mm_validator_id = head->mm_validator_id; meta.mm_extra_pagehdr = head->mm_extra_pagehdr; unaligned_poke_u64(4, meta.mm_pages_retired, unaligned_peek_u64(4, head->mm_pages_retired) + MDBX_PNL_SIZE(txn->tw.retired_pages)); meta.mm_geo = txn->mt_geo; meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI]; meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; meta.mm_canary = txn->mt_canary; mdbx_meta_set_txnid(env, &meta, txn->mt_txnid); rc = mdbx_sync_locked( env, env->me_flags | txn->mt_flags | MDBX_SHRINK_ALLOWED, &meta); } ts_4 = latency ? mdbx_osal_monotime() : 0; if (unlikely(rc != MDBX_SUCCESS)) { env->me_flags |= MDBX_FATAL_ERROR; goto fail; } end_mode = MDBX_END_COMMITTED | MDBX_END_UPDATE | MDBX_END_EOTDONE; done: rc = mdbx_txn_end(txn, end_mode); provide_latency: if (latency) { latency->audit = audit_duration; latency->preparation = ts_1 ? mdbx_osal_monotime_to_16dot16(ts_1 - ts_0) : 0; latency->gc = (ts_1 && ts_2) ? mdbx_osal_monotime_to_16dot16(ts_2 - ts_1) : 0; latency->write = (ts_2 && ts_3) ? mdbx_osal_monotime_to_16dot16(ts_3 - ts_2) : 0; latency->sync = (ts_3 && ts_4) ? mdbx_osal_monotime_to_16dot16(ts_4 - ts_3) : 0; const uint64_t ts_5 = mdbx_osal_monotime(); latency->ending = ts_4 ? mdbx_osal_monotime_to_16dot16(ts_5 - ts_4) : 0; latency->whole = mdbx_osal_monotime_to_16dot16(ts_5 - ts_0); } return rc; fail: mdbx_txn_abort(txn); goto provide_latency; } static __cold int mdbx_validate_meta(MDBX_env *env, MDBX_meta *const meta, uint64_t *filesize, const MDBX_page *const page, const unsigned meta_number, MDBX_meta *dest, const unsigned guess_pagesize) { const uint64_t magic_and_version = unaligned_peek_u64(4, &meta->mm_magic_and_version); if (magic_and_version != MDBX_DATA_MAGIC && magic_and_version != MDBX_DATA_MAGIC_DEVEL) { mdbx_error("meta[%u] has invalid magic/version %" PRIx64, meta_number, magic_and_version); return ((magic_and_version >> 8) != MDBX_MAGIC) ? MDBX_INVALID : MDBX_VERSION_MISMATCH; } if (page->mp_pgno != meta_number) { mdbx_error("meta[%u] has invalid pageno %" PRIaPGNO, meta_number, page->mp_pgno); return MDBX_INVALID; } if (page->mp_flags != P_META) { mdbx_error("page #%u not a meta-page", meta_number); return MDBX_INVALID; } /* LY: check pagesize */ if (!is_powerof2(meta->mm_psize) || meta->mm_psize < MIN_PAGESIZE || meta->mm_psize > MAX_PAGESIZE) { mdbx_warning("meta[%u] has invalid pagesize (%u), skip it", meta_number, meta->mm_psize); return is_powerof2(meta->mm_psize) ? MDBX_VERSION_MISMATCH : MDBX_INVALID; } if (dest && meta_number == 0 && guess_pagesize != meta->mm_psize) { dest->mm_psize = meta->mm_psize; mdbx_verbose("meta[%u] took pagesize %u", meta_number, meta->mm_psize); } if (unaligned_peek_u64(4, &meta->mm_txnid_a) != unaligned_peek_u64(4, &meta->mm_txnid_b)) { mdbx_warning("meta[%u] not completely updated, skip it", meta_number); return MDBX_RESULT_TRUE; } /* LY: check signature as a checksum */ if (META_IS_STEADY(meta) && unaligned_peek_u64(4, &meta->mm_datasync_sign) != mdbx_meta_sign(meta)) { mdbx_warning("meta[%u] has invalid steady-checksum (0x%" PRIx64 " != 0x%" PRIx64 "), skip it", meta_number, unaligned_peek_u64(4, &meta->mm_datasync_sign), mdbx_meta_sign(meta)); return MDBX_RESULT_TRUE; } mdbx_debug("read meta%" PRIaPGNO " = root %" PRIaPGNO "/%" PRIaPGNO ", geo %" PRIaPGNO "/%" PRIaPGNO "-%" PRIaPGNO "/%" PRIaPGNO " +%u -%u, txn_id %" PRIaTXN ", %s", page->mp_pgno, meta->mm_dbs[MAIN_DBI].md_root, meta->mm_dbs[FREE_DBI].md_root, meta->mm_geo.lower, meta->mm_geo.next, meta->mm_geo.now, meta->mm_geo.upper, meta->mm_geo.grow, meta->mm_geo.shrink, unaligned_peek_u64(4, meta->mm_txnid_a), mdbx_durable_str(meta)); /* LY: check min-pages value */ if (meta->mm_geo.lower < MIN_PAGENO || meta->mm_geo.lower > MAX_PAGENO) { mdbx_warning("meta[%u] has invalid min-pages (%" PRIaPGNO "), skip it", meta_number, meta->mm_geo.lower); return MDBX_INVALID; } /* LY: check max-pages value */ if (meta->mm_geo.upper < MIN_PAGENO || meta->mm_geo.upper > MAX_PAGENO || meta->mm_geo.upper < meta->mm_geo.lower) { mdbx_warning("meta[%u] has invalid max-pages (%" PRIaPGNO "), skip it", meta_number, meta->mm_geo.upper); return MDBX_INVALID; } /* LY: check last_pgno */ if (meta->mm_geo.next < MIN_PAGENO || meta->mm_geo.next - 1 > MAX_PAGENO) { mdbx_warning("meta[%u] has invalid next-pageno (%" PRIaPGNO "), skip it", meta_number, meta->mm_geo.next); return MDBX_CORRUPTED; } /* LY: check filesize & used_bytes */ const uint64_t used_bytes = meta->mm_geo.next * (uint64_t)meta->mm_psize; if (used_bytes > *filesize) { /* Here could be a race with DB-shrinking performed by other process */ int err = mdbx_filesize(env->me_lazy_fd, filesize); if (unlikely(err != MDBX_SUCCESS)) return err; if (used_bytes > *filesize) { mdbx_warning("meta[%u] used-bytes (%" PRIu64 ") beyond filesize (%" PRIu64 "), skip it", meta_number, used_bytes, *filesize); return MDBX_CORRUPTED; } } if (meta->mm_geo.next - 1 > MAX_PAGENO || used_bytes > MAX_MAPSIZE) { mdbx_warning("meta[%u] has too large used-space (%" PRIu64 "), skip it", meta_number, used_bytes); return MDBX_TOO_LARGE; } /* LY: check mapsize limits */ const uint64_t mapsize_min = meta->mm_geo.lower * (uint64_t)meta->mm_psize; STATIC_ASSERT(MAX_MAPSIZE < PTRDIFF_MAX - MAX_PAGESIZE); STATIC_ASSERT(MIN_MAPSIZE < MAX_MAPSIZE); if (mapsize_min < MIN_MAPSIZE || mapsize_min > MAX_MAPSIZE) { if (MAX_MAPSIZE != MAX_MAPSIZE64 && mapsize_min > MAX_MAPSIZE && mapsize_min <= MAX_MAPSIZE64) { mdbx_assert(env, meta->mm_geo.next - 1 <= MAX_PAGENO && used_bytes <= MAX_MAPSIZE); mdbx_warning("meta[%u] has too large min-mapsize (%" PRIu64 "), " "but size of used space still acceptable (%" PRIu64 ")", meta_number, mapsize_min, used_bytes); meta->mm_geo.lower = (pgno_t)(MAX_MAPSIZE / meta->mm_psize); } else { mdbx_warning("meta[%u] has invalid min-mapsize (%" PRIu64 "), skip it", meta_number, mapsize_min); return MDBX_VERSION_MISMATCH; } } const uint64_t mapsize_max = meta->mm_geo.upper * (uint64_t)meta->mm_psize; STATIC_ASSERT(MIN_MAPSIZE < MAX_MAPSIZE); if (mapsize_max > MAX_MAPSIZE || MAX_PAGENO < ceil_powerof2((size_t)mapsize_max, env->me_os_psize) / (size_t)meta->mm_psize) { /* allow to open large DB from a 32-bit environment */ mdbx_assert(env, meta->mm_geo.next - 1 <= MAX_PAGENO && used_bytes <= MAX_MAPSIZE); mdbx_warning("meta[%u] has too large max-mapsize (%" PRIu64 "), " "but size of used space still acceptable (%" PRIu64 ")", meta_number, mapsize_max, used_bytes); meta->mm_geo.upper = (pgno_t)(MAX_MAPSIZE / meta->mm_psize); } /* LY: check and silently put mm_geo.now into [geo.lower...geo.upper]. * * Copy-with-compaction by previous version of libmdbx could produce DB-file * less than meta.geo.lower bound, in case actual filling is low or no data * at all. This is not a problem as there is no damage or loss of data. * Therefore it is better not to consider such situation as an error, but * silently correct it. */ if (meta->mm_geo.now < meta->mm_geo.lower) meta->mm_geo.now = meta->mm_geo.lower; if (meta->mm_geo.now > meta->mm_geo.upper && meta->mm_geo.next <= meta->mm_geo.upper) meta->mm_geo.now = meta->mm_geo.upper; if (meta->mm_geo.next > meta->mm_geo.now) { mdbx_warning("meta[%u] next-pageno (%" PRIaPGNO ") is beyond end-pgno (%" PRIaPGNO "), skip it", meta_number, meta->mm_geo.next, meta->mm_geo.now); return MDBX_CORRUPTED; } /* LY: GC root */ if (meta->mm_dbs[FREE_DBI].md_root == P_INVALID) { if (meta->mm_dbs[FREE_DBI].md_branch_pages || meta->mm_dbs[FREE_DBI].md_depth || meta->mm_dbs[FREE_DBI].md_entries || meta->mm_dbs[FREE_DBI].md_leaf_pages || meta->mm_dbs[FREE_DBI].md_overflow_pages) { mdbx_warning("meta[%u] has false-empty GC, skip it", meta_number); return MDBX_CORRUPTED; } } else if (meta->mm_dbs[FREE_DBI].md_root >= meta->mm_geo.next) { mdbx_warning("meta[%u] has invalid GC-root %" PRIaPGNO ", skip it", meta_number, meta->mm_dbs[FREE_DBI].md_root); return MDBX_CORRUPTED; } /* LY: MainDB root */ if (meta->mm_dbs[MAIN_DBI].md_root == P_INVALID) { if (meta->mm_dbs[MAIN_DBI].md_branch_pages || meta->mm_dbs[MAIN_DBI].md_depth || meta->mm_dbs[MAIN_DBI].md_entries || meta->mm_dbs[MAIN_DBI].md_leaf_pages || meta->mm_dbs[MAIN_DBI].md_overflow_pages) { mdbx_warning("meta[%u] has false-empty maindb", meta_number); return MDBX_CORRUPTED; } } else if (meta->mm_dbs[MAIN_DBI].md_root >= meta->mm_geo.next) { mdbx_warning("meta[%u] has invalid maindb-root %" PRIaPGNO ", skip it", meta_number, meta->mm_dbs[MAIN_DBI].md_root); return MDBX_CORRUPTED; } if (unaligned_peek_u64(4, &meta->mm_txnid_a) == 0) { mdbx_warning("meta[%u] has zero txnid, skip it", meta_number); return MDBX_RESULT_TRUE; } return MDBX_SUCCESS; } /* Read the environment parameters of a DB environment * before mapping it into memory. */ static __cold int mdbx_read_header(MDBX_env *env, MDBX_meta *dest, uint64_t *filesize, const int lck_exclusive) { int rc = mdbx_filesize(env->me_lazy_fd, filesize); if (unlikely(rc != MDBX_SUCCESS)) return rc; memset(dest, 0, sizeof(MDBX_meta)); unaligned_poke_u64(4, dest->mm_datasync_sign, MDBX_DATASIGN_WEAK); rc = MDBX_CORRUPTED; /* Read twice all meta pages so we can find the latest one. */ unsigned loop_limit = NUM_METAS * 2; for (unsigned loop_count = 0; loop_count < loop_limit; ++loop_count) { /* We don't know the page size on first time. * So, just guess it. */ unsigned guess_pagesize = dest->mm_psize; if (guess_pagesize == 0) guess_pagesize = (loop_count > NUM_METAS) ? env->me_psize : env->me_os_psize; const unsigned meta_number = loop_count % NUM_METAS; const unsigned offset = guess_pagesize * meta_number; char buffer[MIN_PAGESIZE]; unsigned retryleft = 42; while (1) { mdbx_trace("reading meta[%d]: offset %u, bytes %u, retry-left %u", meta_number, offset, MIN_PAGESIZE, retryleft); int err = mdbx_pread(env->me_lazy_fd, buffer, MIN_PAGESIZE, offset); if (err != MDBX_SUCCESS) { if (err == MDBX_ENODATA && offset == 0 && loop_count == 0 && *filesize == 0 && (env->me_flags & MDBX_RDONLY) == 0) mdbx_warning("read meta: empty file (%d, %s)", err, mdbx_strerror(err)); else mdbx_error("read meta[%u,%u]: %i, %s", offset, MIN_PAGESIZE, err, mdbx_strerror(err)); return err; } char again[MIN_PAGESIZE]; err = mdbx_pread(env->me_lazy_fd, again, MIN_PAGESIZE, offset); if (err != MDBX_SUCCESS) { mdbx_error("read meta[%u,%u]: %i, %s", offset, MIN_PAGESIZE, err, mdbx_strerror(err)); return err; } if (memcmp(buffer, again, MIN_PAGESIZE) == 0 || --retryleft == 0) break; mdbx_verbose("meta[%u] was updated, re-read it", meta_number); } if (!retryleft) { mdbx_error("meta[%u] is too volatile, skip it", meta_number); continue; } MDBX_page *const page = (MDBX_page *)buffer; MDBX_meta *const meta = page_meta(page); rc = mdbx_validate_meta(env, meta, filesize, page, meta_number, dest, guess_pagesize); if (rc != MDBX_SUCCESS) continue; if ((env->me_stuck_meta < 0) ? mdbx_meta_ot(meta_bootid_match(meta) ? prefer_last : prefer_steady, env, dest, meta) : (meta_number == (unsigned)env->me_stuck_meta)) { *dest = *meta; if (!lck_exclusive && !META_IS_STEADY(dest)) loop_limit += 1; /* LY: should re-read to hush race with update */ mdbx_verbose("latch meta[%u]", meta_number); } } if (dest->mm_psize == 0 || ((env->me_stuck_meta < 0) ? (!META_IS_STEADY(dest) && !meta_weak_acceptable(env, dest, lck_exclusive)) : false)) { mdbx_error("%s", "no usable meta-pages, database is corrupted"); if (rc == MDBX_SUCCESS) { /* TODO: try to restore the database by fully checking b-tree structure * for the each meta page, if the corresponding option was given */ return MDBX_CORRUPTED; } return rc; } return MDBX_SUCCESS; } static MDBX_page *__cold mdbx_meta_model(const MDBX_env *env, MDBX_page *model, unsigned num) { mdbx_ensure(env, is_powerof2(env->me_psize)); mdbx_ensure(env, env->me_psize >= MIN_PAGESIZE); mdbx_ensure(env, env->me_psize <= MAX_PAGESIZE); mdbx_ensure(env, env->me_dbgeo.lower >= MIN_MAPSIZE); mdbx_ensure(env, env->me_dbgeo.upper <= MAX_MAPSIZE); mdbx_ensure(env, env->me_dbgeo.now >= env->me_dbgeo.lower); mdbx_ensure(env, env->me_dbgeo.now <= env->me_dbgeo.upper); memset(model, 0, sizeof(*model)); model->mp_pgno = num; model->mp_flags = P_META; MDBX_meta *const model_meta = page_meta(model); unaligned_poke_u64(4, model_meta->mm_magic_and_version, MDBX_DATA_MAGIC); model_meta->mm_geo.lower = bytes2pgno(env, env->me_dbgeo.lower); model_meta->mm_geo.upper = bytes2pgno(env, env->me_dbgeo.upper); model_meta->mm_geo.grow = (uint16_t)bytes2pgno(env, env->me_dbgeo.grow); model_meta->mm_geo.shrink = (uint16_t)bytes2pgno(env, env->me_dbgeo.shrink); model_meta->mm_geo.now = bytes2pgno(env, env->me_dbgeo.now); model_meta->mm_geo.next = NUM_METAS; mdbx_ensure(env, model_meta->mm_geo.lower >= MIN_PAGENO); mdbx_ensure(env, model_meta->mm_geo.upper <= MAX_PAGENO); mdbx_ensure(env, model_meta->mm_geo.now >= model_meta->mm_geo.lower); mdbx_ensure(env, model_meta->mm_geo.now <= model_meta->mm_geo.upper); mdbx_ensure(env, model_meta->mm_geo.next >= MIN_PAGENO); mdbx_ensure(env, model_meta->mm_geo.next <= model_meta->mm_geo.now); mdbx_ensure(env, model_meta->mm_geo.grow == bytes2pgno(env, env->me_dbgeo.grow)); mdbx_ensure(env, model_meta->mm_geo.shrink == bytes2pgno(env, env->me_dbgeo.shrink)); model_meta->mm_psize = env->me_psize; model_meta->mm_flags = (uint16_t)env->me_flags; model_meta->mm_flags |= MDBX_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */ model_meta->mm_dbs[FREE_DBI].md_root = P_INVALID; model_meta->mm_dbs[MAIN_DBI].md_root = P_INVALID; mdbx_meta_set_txnid(env, model_meta, MIN_TXNID + num); unaligned_poke_u64(4, model_meta->mm_datasync_sign, mdbx_meta_sign(model_meta)); return (MDBX_page *)((uint8_t *)model + env->me_psize); } /* Fill in most of the zeroed meta-pages for an empty database environment. * Return pointer to recently (head) meta-page. */ static MDBX_meta *__cold mdbx_init_metas(const MDBX_env *env, void *buffer) { MDBX_page *page0 = (MDBX_page *)buffer; MDBX_page *page1 = mdbx_meta_model(env, page0, 0); MDBX_page *page2 = mdbx_meta_model(env, page1, 1); mdbx_meta_model(env, page2, 2); mdbx_assert(env, !mdbx_meta_eq(env, page_meta(page0), page_meta(page1))); mdbx_assert(env, !mdbx_meta_eq(env, page_meta(page1), page_meta(page2))); mdbx_assert(env, !mdbx_meta_eq(env, page_meta(page2), page_meta(page0))); return page_meta(page2); } static size_t mdbx_madvise_threshold(const MDBX_env *env, const size_t largest_bytes) { /* TODO: use options */ const unsigned factor = 9; const size_t threshold = (largest_bytes < (65536ul << factor)) ? 65536 /* minimal threshold */ : (largest_bytes > (MEGABYTE * 4 << factor)) ? MEGABYTE * 4 /* maximal threshold */ : largest_bytes >> factor; return bytes_align2os_bytes(env, threshold); } static int mdbx_sync_locked(MDBX_env *env, unsigned flags, MDBX_meta *const pending) { mdbx_assert(env, ((env->me_flags ^ flags) & MDBX_WRITEMAP) == 0); MDBX_meta *const meta0 = METAPAGE(env, 0); MDBX_meta *const meta1 = METAPAGE(env, 1); MDBX_meta *const meta2 = METAPAGE(env, 2); MDBX_meta *const head = mdbx_meta_head(env); int rc; mdbx_assert(env, mdbx_meta_eq_mask(env) == 0); mdbx_assert(env, pending < METAPAGE(env, 0) || pending > METAPAGE(env, NUM_METAS)); mdbx_assert(env, (env->me_flags & (MDBX_RDONLY | MDBX_FATAL_ERROR)) == 0); mdbx_assert(env, pending->mm_geo.next <= pending->mm_geo.now); if (flags & MDBX_SAFE_NOSYNC) { /* Check auto-sync conditions */ const pgno_t autosync_threshold = atomic_load32(env->me_autosync_threshold, mo_Relaxed); const uint64_t autosync_period = atomic_load64(env->me_autosync_period, mo_Relaxed); if ((autosync_threshold && atomic_load32(env->me_unsynced_pages, mo_Relaxed) >= autosync_threshold) || (autosync_period && mdbx_osal_monotime() - atomic_load64(env->me_sync_timestamp, mo_Relaxed) >= autosync_period)) flags &= MDBX_WRITEMAP | MDBX_SHRINK_ALLOWED; /* force steady */ } pgno_t shrink = 0; if (flags & MDBX_SHRINK_ALLOWED) { /* LY: check conditions to discard unused pages */ const pgno_t largest_pgno = mdbx_find_largest( env, (head->mm_geo.next > pending->mm_geo.next) ? head->mm_geo.next : pending->mm_geo.next); mdbx_assert(env, largest_pgno >= NUM_METAS); #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) const pgno_t edge = env->me_poison_edge; if (edge > largest_pgno) { env->me_poison_edge = largest_pgno; VALGRIND_MAKE_MEM_NOACCESS(env->me_map + pgno2bytes(env, largest_pgno), pgno2bytes(env, edge - largest_pgno)); ASAN_POISON_MEMORY_REGION(env->me_map + pgno2bytes(env, largest_pgno), pgno2bytes(env, edge - largest_pgno)); } #endif /* MDBX_USE_VALGRIND || __SANITIZE_ADDRESS__ */ #if defined(MADV_DONTNEED) const size_t largest_bytes = pgno2bytes(env, largest_pgno); /* threshold to avoid unreasonable frequent madvise() calls */ const size_t madvise_threshold = mdbx_madvise_threshold(env, largest_bytes); const size_t discard_edge_bytes = bytes_align2os_bytes( env, ((MDBX_RDONLY & (env->me_lck ? env->me_lck->mti_envmode.weak : env->me_flags)) ? largest_bytes : largest_bytes + madvise_threshold)); const pgno_t discard_edge_pgno = bytes2pgno(env, discard_edge_bytes); const pgno_t prev_discarded_pgno = atomic_load32(env->me_discarded_tail, mo_AcquireRelease); if (prev_discarded_pgno >= discard_edge_pgno + bytes2pgno(env, madvise_threshold)) { mdbx_notice("open-MADV_%s %u..%u", "DONTNEED", prev_discarded_pgno, largest_pgno); atomic_store32(env->me_discarded_tail, discard_edge_pgno, mo_AcquireRelease); const size_t prev_discarded_bytes = ceil_powerof2(pgno2bytes(env, prev_discarded_pgno), env->me_os_psize); mdbx_ensure(env, prev_discarded_bytes > discard_edge_bytes); int advise = MADV_DONTNEED; #if defined(MADV_FREE) && \ 0 /* MADV_FREE works for only anonymous vma at the moment */ if ((env->me_flags & MDBX_WRITEMAP) && mdbx_linux_kernel_version > 0x04050000) advise = MADV_FREE; #endif /* MADV_FREE */ int err = madvise(env->me_map + discard_edge_bytes, prev_discarded_bytes - discard_edge_bytes, advise) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; } #endif /* MADV_FREE || MADV_DONTNEED */ /* LY: check conditions to shrink datafile */ const pgno_t backlog_gap = 3 + pending->mm_dbs[FREE_DBI].md_depth * 3; if (pending->mm_geo.shrink && pending->mm_geo.now - pending->mm_geo.next > pending->mm_geo.shrink + backlog_gap) { if (pending->mm_geo.now > largest_pgno && pending->mm_geo.now - largest_pgno > pending->mm_geo.shrink + backlog_gap) { const pgno_t aligner = pending->mm_geo.grow ? pending->mm_geo.grow : pending->mm_geo.shrink; const pgno_t with_backlog_gap = largest_pgno + backlog_gap; const pgno_t aligned = pgno_align2os_pgno( env, with_backlog_gap + aligner - with_backlog_gap % aligner); const pgno_t bottom = (aligned > pending->mm_geo.lower) ? aligned : pending->mm_geo.lower; if (pending->mm_geo.now > bottom) { if (META_IS_STEADY(mdbx_meta_steady(env))) /* force steady, but only if steady-checkpoint is present */ flags &= MDBX_WRITEMAP | MDBX_SHRINK_ALLOWED; shrink = pending->mm_geo.now - bottom; pending->mm_geo.now = bottom; if (unlikely(mdbx_meta_txnid_stable(env, head) == unaligned_peek_u64(4, pending->mm_txnid_a))) { const txnid_t txnid = safe64_txnid_next(unaligned_peek_u64(4, pending->mm_txnid_a)); if (unlikely(txnid > MAX_TXNID)) { rc = MDBX_TXN_FULL; mdbx_error("txnid overflow, raise %d", rc); goto fail; } mdbx_meta_set_txnid(env, pending, txnid); } } } } } /* LY: step#1 - sync previously written/updated data-pages */ rc = MDBX_RESULT_FALSE /* carry steady */; if (atomic_load32(env->me_unsynced_pages, mo_AcquireRelease)) { mdbx_assert(env, ((flags ^ env->me_flags) & MDBX_WRITEMAP) == 0); enum mdbx_syncmode_bits mode_bits = MDBX_SYNC_NONE; if ((flags & MDBX_SAFE_NOSYNC) == 0) { mode_bits = MDBX_SYNC_DATA; if (pending->mm_geo.next > mdbx_meta_steady(env)->mm_geo.now) mode_bits |= MDBX_SYNC_SIZE; if (flags & MDBX_NOMETASYNC) mode_bits |= MDBX_SYNC_IODQ; } if (flags & MDBX_WRITEMAP) rc = mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, pending->mm_geo.next), mode_bits); else rc = mdbx_fsync(env->me_lazy_fd, mode_bits); if (unlikely(rc != MDBX_SUCCESS)) goto fail; rc = (flags & MDBX_SAFE_NOSYNC) ? MDBX_RESULT_TRUE /* carry non-steady */ : MDBX_RESULT_FALSE /* carry steady */; } /* Steady or Weak */ if (rc == MDBX_RESULT_FALSE /* carry steady */) { atomic_store64(env->me_sync_timestamp, mdbx_osal_monotime(), mo_Relaxed); unaligned_poke_u64(4, pending->mm_datasync_sign, mdbx_meta_sign(pending)); atomic_store32(env->me_unsynced_pages, 0, mo_AcquireRelease); } else { assert(rc == MDBX_RESULT_TRUE /* carry non-steady */); unaligned_poke_u64(4, pending->mm_datasync_sign, MDBX_DATASIGN_WEAK); } MDBX_meta *target = nullptr; if (mdbx_meta_txnid_stable(env, head) == unaligned_peek_u64(4, pending->mm_txnid_a)) { mdbx_assert(env, memcmp(&head->mm_dbs, &pending->mm_dbs, sizeof(head->mm_dbs)) == 0); mdbx_assert(env, memcmp(&head->mm_canary, &pending->mm_canary, sizeof(head->mm_canary)) == 0); mdbx_assert(env, memcmp(&head->mm_geo, &pending->mm_geo, sizeof(pending->mm_geo)) == 0); if (!META_IS_STEADY(head) && META_IS_STEADY(pending)) target = head; else { mdbx_ensure(env, mdbx_meta_eq(env, head, pending)); mdbx_debug("%s", "skip update meta"); return MDBX_SUCCESS; } } else if (head == meta0) target = mdbx_meta_ancient(prefer_steady, env, meta1, meta2); else if (head == meta1) target = mdbx_meta_ancient(prefer_steady, env, meta0, meta2); else { mdbx_assert(env, head == meta2); target = mdbx_meta_ancient(prefer_steady, env, meta0, meta1); } /* LY: step#2 - update meta-page. */ mdbx_debug("writing meta%" PRIaPGNO " = root %" PRIaPGNO "/%" PRIaPGNO ", geo %" PRIaPGNO "/%" PRIaPGNO "-%" PRIaPGNO "/%" PRIaPGNO " +%u -%u, txn_id %" PRIaTXN ", %s", data_page(target)->mp_pgno, pending->mm_dbs[MAIN_DBI].md_root, pending->mm_dbs[FREE_DBI].md_root, pending->mm_geo.lower, pending->mm_geo.next, pending->mm_geo.now, pending->mm_geo.upper, pending->mm_geo.grow, pending->mm_geo.shrink, unaligned_peek_u64(4, pending->mm_txnid_a), mdbx_durable_str(pending)); mdbx_debug("meta0: %s, %s, txn_id %" PRIaTXN ", root %" PRIaPGNO "/%" PRIaPGNO, (meta0 == head) ? "head" : (meta0 == target) ? "tail" : "stay", mdbx_durable_str(meta0), mdbx_meta_txnid_fluid(env, meta0), meta0->mm_dbs[MAIN_DBI].md_root, meta0->mm_dbs[FREE_DBI].md_root); mdbx_debug("meta1: %s, %s, txn_id %" PRIaTXN ", root %" PRIaPGNO "/%" PRIaPGNO, (meta1 == head) ? "head" : (meta1 == target) ? "tail" : "stay", mdbx_durable_str(meta1), mdbx_meta_txnid_fluid(env, meta1), meta1->mm_dbs[MAIN_DBI].md_root, meta1->mm_dbs[FREE_DBI].md_root); mdbx_debug("meta2: %s, %s, txn_id %" PRIaTXN ", root %" PRIaPGNO "/%" PRIaPGNO, (meta2 == head) ? "head" : (meta2 == target) ? "tail" : "stay", mdbx_durable_str(meta2), mdbx_meta_txnid_fluid(env, meta2), meta2->mm_dbs[MAIN_DBI].md_root, meta2->mm_dbs[FREE_DBI].md_root); mdbx_assert(env, !mdbx_meta_eq(env, pending, meta0)); mdbx_assert(env, !mdbx_meta_eq(env, pending, meta1)); mdbx_assert(env, !mdbx_meta_eq(env, pending, meta2)); mdbx_assert(env, ((env->me_flags ^ flags) & MDBX_WRITEMAP) == 0); mdbx_ensure(env, target == head || mdbx_meta_txnid_stable(env, target) < unaligned_peek_u64(4, pending->mm_txnid_a)); if (flags & MDBX_WRITEMAP) { mdbx_jitter4testing(true); if (likely(target != head)) { /* LY: 'invalidate' the meta. */ mdbx_meta_update_begin(env, target, unaligned_peek_u64(4, pending->mm_txnid_a)); unaligned_poke_u64(4, target->mm_datasync_sign, MDBX_DATASIGN_WEAK); #ifndef NDEBUG /* debug: provoke failure to catch a violators, but don't touch mm_psize * and mm_flags to allow readers catch actual pagesize. */ uint8_t *provoke_begin = (uint8_t *)&target->mm_dbs[FREE_DBI].md_root; uint8_t *provoke_end = (uint8_t *)&target->mm_datasync_sign; memset(provoke_begin, 0xCC, provoke_end - provoke_begin); mdbx_jitter4testing(false); #endif /* LY: update info */ target->mm_geo = pending->mm_geo; target->mm_dbs[FREE_DBI] = pending->mm_dbs[FREE_DBI]; target->mm_dbs[MAIN_DBI] = pending->mm_dbs[MAIN_DBI]; target->mm_canary = pending->mm_canary; memcpy(target->mm_pages_retired, pending->mm_pages_retired, 8); mdbx_jitter4testing(true); /* LY: 'commit' the meta */ mdbx_meta_update_end(env, target, unaligned_peek_u64(4, pending->mm_txnid_b)); mdbx_jitter4testing(true); } else { /* dangerous case (target == head), only mm_datasync_sign could * me updated, check assertions once again */ mdbx_ensure(env, mdbx_meta_txnid_stable(env, head) == unaligned_peek_u64(4, pending->mm_txnid_a) && !META_IS_STEADY(head) && META_IS_STEADY(pending)); mdbx_ensure(env, memcmp(&head->mm_geo, &pending->mm_geo, sizeof(head->mm_geo)) == 0); mdbx_ensure(env, memcmp(&head->mm_dbs, &pending->mm_dbs, sizeof(head->mm_dbs)) == 0); mdbx_ensure(env, memcmp(&head->mm_canary, &pending->mm_canary, sizeof(head->mm_canary)) == 0); } memcpy(target->mm_datasync_sign, pending->mm_datasync_sign, 8); mdbx_flush_incoherent_cpu_writeback(); mdbx_jitter4testing(true); /* sync meta-pages */ rc = mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, NUM_METAS), (flags & MDBX_NOMETASYNC) ? MDBX_SYNC_NONE : MDBX_SYNC_DATA | MDBX_SYNC_IODQ); if (unlikely(rc != MDBX_SUCCESS)) goto fail; } else { const MDBX_meta undo_meta = *target; const mdbx_filehandle_t fd = (env->me_dsync_fd != INVALID_HANDLE_VALUE) ? env->me_dsync_fd : env->me_lazy_fd; rc = mdbx_pwrite(fd, pending, sizeof(MDBX_meta), (uint8_t *)target - env->me_map); if (unlikely(rc != MDBX_SUCCESS)) { undo: mdbx_debug("%s", "write failed, disk error?"); /* On a failure, the pagecache still contains the new data. * Try write some old data back, to prevent it from being used. */ mdbx_pwrite(fd, &undo_meta, sizeof(MDBX_meta), (uint8_t *)target - env->me_map); goto fail; } mdbx_flush_incoherent_mmap(target, sizeof(MDBX_meta), env->me_os_psize); /* sync meta-pages */ if ((flags & MDBX_NOMETASYNC) == 0 && fd == env->me_lazy_fd) { rc = mdbx_fsync(env->me_lazy_fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ); if (rc != MDBX_SUCCESS) goto undo; } } if (flags & MDBX_NOMETASYNC) env->me_unsynced_pages->weak += 1; else env->me_meta_sync_txnid->weak = (uint32_t)unaligned_peek_u64(4, pending->mm_txnid_a); /* LY: shrink datafile if needed */ if (unlikely(shrink)) { mdbx_verbose("shrink to %" PRIaPGNO " pages (-%" PRIaPGNO ")", pending->mm_geo.now, shrink); rc = mdbx_mapresize_implicit(env, pending->mm_geo.next, pending->mm_geo.now, pending->mm_geo.upper); if (MDBX_IS_ERROR(rc)) goto fail; } if (likely(env->me_lck)) /* toggle oldest refresh */ atomic_store32(&env->me_lck->mti_readers_refresh_flag, false, mo_AcquireRelease); return MDBX_SUCCESS; fail: env->me_flags |= MDBX_FATAL_ERROR; return rc; } static void __cold mdbx_setup_pagesize(MDBX_env *env, const size_t pagesize) { STATIC_ASSERT(PTRDIFF_MAX > MAX_MAPSIZE); STATIC_ASSERT(MIN_PAGESIZE > sizeof(MDBX_page) + sizeof(MDBX_meta)); mdbx_ensure(env, is_powerof2(pagesize)); mdbx_ensure(env, pagesize >= MIN_PAGESIZE); mdbx_ensure(env, pagesize <= MAX_PAGESIZE); env->me_psize = (unsigned)pagesize; STATIC_ASSERT(MAX_GC1OVPAGE(MIN_PAGESIZE) > 4); STATIC_ASSERT(MAX_GC1OVPAGE(MAX_PAGESIZE) < MDBX_PGL_LIMIT / 4); const intptr_t maxgc_ov1page = (pagesize - PAGEHDRSZ) / sizeof(pgno_t) - 1; mdbx_ensure(env, maxgc_ov1page > 42 && maxgc_ov1page < (intptr_t)MDBX_PGL_LIMIT / 4); env->me_maxgc_ov1page = (unsigned)maxgc_ov1page; STATIC_ASSERT(LEAF_NODEMAX(MIN_PAGESIZE) > sizeof(MDBX_db) + NODESIZE + 42); STATIC_ASSERT(LEAF_NODEMAX(MAX_PAGESIZE) < UINT16_MAX); STATIC_ASSERT(LEAF_NODEMAX(MIN_PAGESIZE) > BRANCH_NODEMAX(MIN_PAGESIZE)); STATIC_ASSERT(BRANCH_NODEMAX(MAX_PAGESIZE) > NODESIZE + 42); STATIC_ASSERT(BRANCH_NODEMAX(MAX_PAGESIZE) < UINT16_MAX); const intptr_t branch_nodemax = BRANCH_NODEMAX(pagesize); mdbx_ensure(env, branch_nodemax > 42 && branch_nodemax < (int)UINT16_MAX && branch_nodemax % 2 == 0); env->me_branch_nodemax = (unsigned)branch_nodemax; env->me_psize2log = (uint8_t)log2n(pagesize); mdbx_assert(env, pgno2bytes(env, 1) == pagesize); mdbx_assert(env, bytes2pgno(env, pagesize + pagesize) == 2); const pgno_t max_pgno = bytes2pgno(env, MAX_MAPSIZE); if (env->me_options.dp_limit > max_pgno - NUM_METAS) env->me_options.dp_limit = max_pgno - NUM_METAS; if (env->me_options.dp_initial > env->me_options.dp_limit) env->me_options.dp_initial = env->me_options.dp_limit; } __cold int mdbx_env_create(MDBX_env **penv) { MDBX_env *env = mdbx_calloc(1, sizeof(MDBX_env)); if (unlikely(!env)) return MDBX_ENOMEM; env->me_maxreaders = DEFAULT_READERS; env->me_maxdbs = env->me_numdbs = CORE_DBS; env->me_lazy_fd = INVALID_HANDLE_VALUE; env->me_dsync_fd = INVALID_HANDLE_VALUE; env->me_lfd = INVALID_HANDLE_VALUE; env->me_pid = mdbx_getpid(); env->me_stuck_meta = -1; env->me_options.dp_reserve_limit = 1024; env->me_options.rp_augment_limit = 256 * 1024; env->me_options.dp_limit = 64 * 1024; if (env->me_options.dp_limit > MAX_PAGENO - NUM_METAS) env->me_options.dp_limit = MAX_PAGENO - NUM_METAS; env->me_options.dp_initial = MDBX_PNL_INITIAL; if (env->me_options.dp_initial > env->me_options.dp_limit) env->me_options.dp_initial = env->me_options.dp_limit; env->me_options.spill_max_denominator = 8; env->me_options.spill_min_denominator = 8; env->me_options.spill_parent4child_denominator = 0; env->me_options.dp_loose_limit = 64; int rc; const size_t os_psize = mdbx_syspagesize(); if (unlikely(!is_powerof2(os_psize) || os_psize < MIN_PAGESIZE)) { mdbx_error("unsuitable system pagesize %" PRIuPTR, os_psize); rc = MDBX_INCOMPATIBLE; goto bailout; } env->me_os_psize = (unsigned)os_psize; mdbx_setup_pagesize(env, (env->me_os_psize < MAX_PAGESIZE) ? env->me_os_psize : MAX_PAGESIZE); rc = mdbx_fastmutex_init(&env->me_dbi_lock); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; #if defined(_WIN32) || defined(_WIN64) mdbx_srwlock_Init(&env->me_remap_guard); InitializeCriticalSection(&env->me_windowsbug_lock); #else rc = mdbx_fastmutex_init(&env->me_remap_guard); if (unlikely(rc != MDBX_SUCCESS)) { mdbx_fastmutex_destroy(&env->me_dbi_lock); goto bailout; } #if MDBX_LOCKING > MDBX_LOCKING_SYSV rc = mdbx_ipclock_stub(&env->me_lckless_stub.wlock); #endif /* MDBX_LOCKING */ if (unlikely(rc != MDBX_SUCCESS)) { mdbx_fastmutex_destroy(&env->me_remap_guard); mdbx_fastmutex_destroy(&env->me_dbi_lock); goto bailout; } #endif /* Windows */ VALGRIND_CREATE_MEMPOOL(env, 0, 0); env->me_signature.weak = MDBX_ME_SIGNATURE; *penv = env; return MDBX_SUCCESS; bailout: mdbx_free(env); *penv = nullptr; return rc; } __cold LIBMDBX_API int mdbx_env_set_geometry(MDBX_env *env, intptr_t size_lower, intptr_t size_now, intptr_t size_upper, intptr_t growth_step, intptr_t shrink_threshold, intptr_t pagesize) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; const bool inside_txn = (env->me_txn0 && env->me_txn0->mt_owner == mdbx_thread_self()); #if MDBX_DEBUG if (growth_step < 0) growth_step = 1; if (shrink_threshold < 0) shrink_threshold = 1; #endif bool need_unlock = false; rc = MDBX_PROBLEM; if (env->me_map) { /* env already mapped */ if (unlikely(env->me_flags & MDBX_RDONLY)) return MDBX_EACCESS; if (!inside_txn) { int err = mdbx_txn_lock(env, false); if (unlikely(err != MDBX_SUCCESS)) return err; need_unlock = true; } MDBX_meta *head = mdbx_meta_head(env); if (!inside_txn) { env->me_txn0->mt_txnid = meta_txnid(env, head, false); mdbx_find_oldest(env->me_txn0); } /* get untouched params from DB */ if (pagesize <= 0 || pagesize >= INT_MAX) pagesize = env->me_psize; if (size_lower < 0) size_lower = pgno2bytes(env, head->mm_geo.lower); if (size_now < 0) size_now = pgno2bytes(env, head->mm_geo.now); if (size_upper < 0) size_upper = pgno2bytes(env, head->mm_geo.upper); if (growth_step < 0) growth_step = pgno2bytes(env, head->mm_geo.grow); if (shrink_threshold < 0) shrink_threshold = pgno2bytes(env, head->mm_geo.shrink); if (pagesize != (intptr_t)env->me_psize) { rc = MDBX_EINVAL; goto bailout; } const size_t usedbytes = pgno2bytes(env, mdbx_find_largest(env, head->mm_geo.next)); if ((size_t)size_upper < usedbytes) { rc = MDBX_MAP_FULL; goto bailout; } if ((size_t)size_now < usedbytes) size_now = usedbytes; } else { /* env NOT yet mapped */ if (unlikely(inside_txn)) return MDBX_PANIC; /* is requested some auto-value for pagesize ? */ if (pagesize >= INT_MAX /* maximal */) pagesize = MAX_PAGESIZE; else if (pagesize <= 0) { if (pagesize < 0 /* default */) { pagesize = env->me_os_psize; if ((uintptr_t)pagesize > MAX_PAGESIZE) pagesize = MAX_PAGESIZE; mdbx_assert(env, (uintptr_t)pagesize >= MIN_PAGESIZE); } else if (pagesize == 0 /* minimal */) pagesize = MIN_PAGESIZE; /* choose pagesize */ intptr_t max_size = (size_now > size_lower) ? size_now : size_lower; max_size = (size_upper > max_size) ? size_upper : max_size; if (max_size < 0 /* default */) max_size = DEFAULT_MAPSIZE; else if (max_size == 0 /* minimal */) max_size = MIN_MAPSIZE; else if (max_size >= (intptr_t)MAX_MAPSIZE /* maximal */) max_size = MAX_MAPSIZE; while (max_size > pagesize * (int64_t)MAX_PAGENO && pagesize < MAX_PAGESIZE) pagesize <<= 1; } } if (pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2(pagesize)) { rc = MDBX_EINVAL; goto bailout; } if (size_lower <= 0) { size_lower = MIN_MAPSIZE; if (MIN_MAPSIZE / pagesize < MIN_PAGENO) size_lower = MIN_PAGENO * pagesize; } if (size_lower == INTPTR_MAX) size_lower = MAX_MAPSIZE; if (size_now <= 0) { size_now = DEFAULT_MAPSIZE; if (size_now < size_lower) size_now = size_lower; if (size_upper >= size_lower && size_now > size_upper) size_now = size_upper; } if (size_now == INTPTR_MAX) size_now = MAX_MAPSIZE; if (size_upper <= 0) { if ((size_t)size_now >= MAX_MAPSIZE / 2) size_upper = MAX_MAPSIZE; else if (MAX_MAPSIZE != MAX_MAPSIZE32 && (size_t)size_now >= MAX_MAPSIZE32 / 2 && (size_t)size_now <= MAX_MAPSIZE32 / 4 * 3) size_upper = MAX_MAPSIZE32; else { size_upper = size_now + size_now; if ((size_t)size_upper < DEFAULT_MAPSIZE * 2) size_upper = DEFAULT_MAPSIZE * 2; } if ((size_t)size_upper / pagesize > MAX_PAGENO) size_upper = pagesize * MAX_PAGENO; } if (unlikely(size_lower < (intptr_t)MIN_MAPSIZE || size_lower > size_upper)) { rc = MDBX_EINVAL; goto bailout; } if ((uint64_t)size_lower / pagesize < MIN_PAGENO) { rc = MDBX_EINVAL; goto bailout; } if (unlikely((size_t)size_upper > MAX_MAPSIZE || (uint64_t)size_upper / pagesize > MAX_PAGENO)) { rc = MDBX_TOO_LARGE; goto bailout; } const size_t unit = (env->me_os_psize > (size_t)pagesize) ? env->me_os_psize : (size_t)pagesize; size_lower = ceil_powerof2(size_lower, unit); size_upper = ceil_powerof2(size_upper, unit); size_now = ceil_powerof2(size_now, unit); /* LY: подбираем значение size_upper: * - кратное размеру страницы * - без нарушения MAX_MAPSIZE и MAX_PAGENO */ while (unlikely((size_t)size_upper > MAX_MAPSIZE || (uint64_t)size_upper / pagesize > MAX_PAGENO)) { if ((size_t)size_upper < unit + MIN_MAPSIZE || (size_t)size_upper < (size_t)pagesize * (MIN_PAGENO + 1)) { /* паранойа на случай переполнения при невероятных значениях */ rc = MDBX_EINVAL; goto bailout; } size_upper -= unit; if ((size_t)size_upper < (size_t)size_lower) size_lower = size_upper; } mdbx_assert(env, (size_upper - size_lower) % env->me_os_psize == 0); if (size_now < size_lower) size_now = size_lower; if (size_now > size_upper) size_now = size_upper; if (growth_step < 0) { growth_step = ((size_t)(size_upper - size_lower)) / 42; if (growth_step > size_lower) growth_step = size_lower; if (growth_step < 65536) growth_step = 65536; if ((size_t)growth_step > MEGABYTE * 16) growth_step = MEGABYTE * 16; } if (growth_step == 0 && shrink_threshold > 0) growth_step = 1; growth_step = ceil_powerof2(growth_step, unit); if (bytes2pgno(env, growth_step) > UINT16_MAX) growth_step = pgno2bytes(env, UINT16_MAX); if (shrink_threshold < 0) shrink_threshold = growth_step + growth_step; shrink_threshold = ceil_powerof2(shrink_threshold, unit); if (bytes2pgno(env, shrink_threshold) > UINT16_MAX) shrink_threshold = pgno2bytes(env, UINT16_MAX); /* save user's geo-params for future open/create */ env->me_dbgeo.lower = size_lower; env->me_dbgeo.now = size_now; env->me_dbgeo.upper = size_upper; env->me_dbgeo.grow = growth_step; env->me_dbgeo.shrink = shrink_threshold; rc = MDBX_SUCCESS; mdbx_ensure(env, pagesize >= MIN_PAGESIZE); mdbx_ensure(env, pagesize <= MAX_PAGESIZE); mdbx_ensure(env, is_powerof2(pagesize)); mdbx_ensure(env, is_powerof2(env->me_os_psize)); mdbx_ensure(env, env->me_dbgeo.lower >= MIN_MAPSIZE); mdbx_ensure(env, env->me_dbgeo.lower / pagesize >= MIN_PAGENO); mdbx_ensure(env, env->me_dbgeo.lower % pagesize == 0); mdbx_ensure(env, env->me_dbgeo.lower % env->me_os_psize == 0); mdbx_ensure(env, env->me_dbgeo.upper <= MAX_MAPSIZE); mdbx_ensure(env, env->me_dbgeo.upper / pagesize <= MAX_PAGENO); mdbx_ensure(env, env->me_dbgeo.upper % pagesize == 0); mdbx_ensure(env, env->me_dbgeo.upper % env->me_os_psize == 0); mdbx_ensure(env, env->me_dbgeo.now >= env->me_dbgeo.lower); mdbx_ensure(env, env->me_dbgeo.now <= env->me_dbgeo.upper); mdbx_ensure(env, env->me_dbgeo.now % pagesize == 0); mdbx_ensure(env, env->me_dbgeo.now % env->me_os_psize == 0); mdbx_ensure(env, env->me_dbgeo.grow % pagesize == 0); mdbx_ensure(env, env->me_dbgeo.grow % env->me_os_psize == 0); mdbx_ensure(env, env->me_dbgeo.shrink % pagesize == 0); mdbx_ensure(env, env->me_dbgeo.shrink % env->me_os_psize == 0); if (env->me_map) { /* apply new params to opened environment */ mdbx_ensure(env, pagesize == (intptr_t)env->me_psize); MDBX_meta meta; MDBX_meta *head = nullptr; const mdbx_geo_t *current_geo; if (inside_txn) { current_geo = &env->me_txn->mt_geo; } else { head = mdbx_meta_head(env); meta = *head; current_geo = &meta.mm_geo; } mdbx_geo_t new_geo; new_geo.lower = bytes2pgno(env, env->me_dbgeo.lower); new_geo.now = bytes2pgno(env, env->me_dbgeo.now); new_geo.upper = bytes2pgno(env, env->me_dbgeo.upper); new_geo.grow = (uint16_t)bytes2pgno(env, env->me_dbgeo.grow); new_geo.shrink = (uint16_t)bytes2pgno(env, env->me_dbgeo.shrink); new_geo.next = current_geo->next; mdbx_ensure(env, pgno_align2os_bytes(env, new_geo.lower) == env->me_dbgeo.lower); mdbx_ensure(env, pgno_align2os_bytes(env, new_geo.upper) == env->me_dbgeo.upper); mdbx_ensure(env, pgno_align2os_bytes(env, new_geo.now) == env->me_dbgeo.now); mdbx_ensure(env, pgno_align2os_bytes(env, new_geo.grow) == env->me_dbgeo.grow); mdbx_ensure(env, pgno_align2os_bytes(env, new_geo.shrink) == env->me_dbgeo.shrink); mdbx_ensure(env, env->me_dbgeo.lower >= MIN_MAPSIZE); mdbx_ensure(env, new_geo.lower >= MIN_PAGENO); mdbx_ensure(env, env->me_dbgeo.upper <= MAX_MAPSIZE); mdbx_ensure(env, new_geo.upper <= MAX_PAGENO); mdbx_ensure(env, new_geo.now >= new_geo.next); mdbx_ensure(env, new_geo.upper >= new_geo.now); mdbx_ensure(env, new_geo.now >= new_geo.lower); if (memcmp(current_geo, &new_geo, sizeof(mdbx_geo_t)) != 0) { #if defined(_WIN32) || defined(_WIN64) /* Was DB shrinking disabled before and now it will be enabled? */ if (new_geo.lower < new_geo.upper && new_geo.shrink && !(current_geo->lower < current_geo->upper && current_geo->shrink)) { if (!env->me_lck) { rc = MDBX_EPERM; goto bailout; } int err = mdbx_rdt_lock(env); if (unlikely(MDBX_IS_ERROR(err))) { rc = err; goto bailout; } /* Check if there are any reading threads that do not use the SRWL */ const size_t CurrentTid = GetCurrentThreadId(); const MDBX_reader *const begin = env->me_lck->mti_readers; const MDBX_reader *const end = begin + atomic_load32(&env->me_lck->mti_numreaders, mo_AcquireRelease); for (const MDBX_reader *reader = begin; reader < end; ++reader) { if (reader->mr_pid.weak == env->me_pid && reader->mr_tid.weak && reader->mr_tid.weak != CurrentTid) { /* At least one thread may don't use SRWL */ rc = MDBX_EPERM; break; } } mdbx_rdt_unlock(env); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } #endif if (new_geo.now != current_geo->now || new_geo.upper != current_geo->upper) { rc = mdbx_mapresize(env, current_geo->next, new_geo.now, new_geo.upper, false); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; mdbx_assert(env, (head == nullptr) == inside_txn); if (head) head = /* base address could be changed */ mdbx_meta_head(env); } if (inside_txn) { env->me_txn->mt_geo = new_geo; env->me_txn->mt_flags |= MDBX_TXN_DIRTY; } else { meta.mm_geo = new_geo; const txnid_t txnid = safe64_txnid_next(mdbx_meta_txnid_stable(env, head)); if (unlikely(txnid > MAX_TXNID)) { rc = MDBX_TXN_FULL; mdbx_error("txnid overflow, raise %d", rc); goto bailout; } mdbx_meta_set_txnid(env, &meta, txnid); rc = mdbx_sync_locked(env, env->me_flags, &meta); } } } else if (pagesize != (intptr_t)env->me_psize) { mdbx_setup_pagesize(env, pagesize); } bailout: if (need_unlock) mdbx_txn_unlock(env); return rc; } __cold int mdbx_env_set_mapsize(MDBX_env *env, size_t size) { return __inline_mdbx_env_set_mapsize(env, size); } __cold int mdbx_env_set_maxdbs(MDBX_env *env, MDBX_dbi dbs) { return __inline_mdbx_env_set_maxdbs(env, dbs); } __cold int mdbx_env_get_maxdbs(const MDBX_env *env, MDBX_dbi *dbs) { return __inline_mdbx_env_get_maxdbs(env, dbs); } __cold int mdbx_env_set_maxreaders(MDBX_env *env, unsigned readers) { return __inline_mdbx_env_set_maxreaders(env, readers); } __cold int mdbx_env_get_maxreaders(const MDBX_env *env, unsigned *readers) { return __inline_mdbx_env_get_maxreaders(env, readers); } /* Further setup required for opening an MDBX environment */ static __cold int mdbx_setup_dxb(MDBX_env *env, const int lck_rc) { uint64_t filesize_before; MDBX_meta meta; int rc = MDBX_RESULT_FALSE; int err = mdbx_read_header(env, &meta, &filesize_before, lck_rc); if (unlikely(err != MDBX_SUCCESS)) { if (lck_rc != /* lck exclusive */ MDBX_RESULT_TRUE || err != MDBX_ENODATA || (env->me_flags & MDBX_RDONLY) != 0 || /* recovery mode */ env->me_stuck_meta >= 0) return err; mdbx_debug("%s", "create new database"); rc = /* new database */ MDBX_RESULT_TRUE; if (!env->me_dbgeo.now) { /* set defaults if not configured */ err = mdbx_env_set_geometry(env, 0, -1, DEFAULT_MAPSIZE, -1, -1, -1); if (unlikely(err != MDBX_SUCCESS)) return err; } void *buffer = mdbx_calloc(NUM_METAS, env->me_psize); if (!buffer) return MDBX_ENOMEM; meta = *mdbx_init_metas(env, buffer); err = mdbx_pwrite(env->me_lazy_fd, buffer, env->me_psize * NUM_METAS, 0); mdbx_free(buffer); if (unlikely(err != MDBX_SUCCESS)) return err; err = mdbx_ftruncate(env->me_lazy_fd, filesize_before = env->me_dbgeo.now); if (unlikely(err != MDBX_SUCCESS)) return err; #ifndef NDEBUG /* just for checking */ err = mdbx_read_header(env, &meta, &filesize_before, lck_rc); if (unlikely(err != MDBX_SUCCESS)) return err; #endif } mdbx_verbose("header: root %" PRIaPGNO "/%" PRIaPGNO ", geo %" PRIaPGNO "/%" PRIaPGNO "-%" PRIaPGNO "/%" PRIaPGNO " +%u -%u, txn_id %" PRIaTXN ", %s", meta.mm_dbs[MAIN_DBI].md_root, meta.mm_dbs[FREE_DBI].md_root, meta.mm_geo.lower, meta.mm_geo.next, meta.mm_geo.now, meta.mm_geo.upper, meta.mm_geo.grow, meta.mm_geo.shrink, unaligned_peek_u64(4, meta.mm_txnid_a), mdbx_durable_str(&meta)); mdbx_setup_pagesize(env, meta.mm_psize); const size_t used_bytes = pgno2bytes(env, meta.mm_geo.next); const size_t used_aligned2os_bytes = ceil_powerof2(used_bytes, env->me_os_psize); if ((env->me_flags & MDBX_RDONLY) /* readonly */ || lck_rc != MDBX_RESULT_TRUE /* not exclusive */ || /* recovery mode */ env->me_stuck_meta >= 0) { /* use present params from db */ const size_t pagesize = meta.mm_psize; err = mdbx_env_set_geometry( env, meta.mm_geo.lower * pagesize, meta.mm_geo.now * pagesize, meta.mm_geo.upper * pagesize, meta.mm_geo.grow * pagesize, meta.mm_geo.shrink * pagesize, meta.mm_psize); if (unlikely(err != MDBX_SUCCESS)) { mdbx_error("%s: err %d", "could not apply preconfigured geometry from db", err); return (err == MDBX_EINVAL) ? MDBX_INCOMPATIBLE : err; } } else if (env->me_dbgeo.now) { /* silently growth to last used page */ if (env->me_dbgeo.now < used_aligned2os_bytes) env->me_dbgeo.now = used_aligned2os_bytes; if (env->me_dbgeo.upper < used_aligned2os_bytes) env->me_dbgeo.upper = used_aligned2os_bytes; /* apply preconfigured params, but only if substantial changes: * - upper or lower limit changes * - shrink threshold or growth step * But ignore change just a 'now/current' size. */ if (bytes_align2os_bytes(env, env->me_dbgeo.upper) != pgno2bytes(env, meta.mm_geo.upper) || bytes_align2os_bytes(env, env->me_dbgeo.lower) != pgno2bytes(env, meta.mm_geo.lower) || bytes_align2os_bytes(env, env->me_dbgeo.shrink) != pgno2bytes(env, meta.mm_geo.shrink) || bytes_align2os_bytes(env, env->me_dbgeo.grow) != pgno2bytes(env, meta.mm_geo.grow)) { if (env->me_dbgeo.shrink && env->me_dbgeo.now > used_bytes) /* pre-shrink if enabled */ env->me_dbgeo.now = used_bytes + env->me_dbgeo.shrink - used_bytes % env->me_dbgeo.shrink; err = mdbx_env_set_geometry(env, env->me_dbgeo.lower, env->me_dbgeo.now, env->me_dbgeo.upper, env->me_dbgeo.grow, env->me_dbgeo.shrink, meta.mm_psize); if (unlikely(err != MDBX_SUCCESS)) { mdbx_error("%s: err %d", "could not apply preconfigured db-geometry", err); return (err == MDBX_EINVAL) ? MDBX_INCOMPATIBLE : err; } /* update meta fields */ meta.mm_geo.now = bytes2pgno(env, env->me_dbgeo.now); meta.mm_geo.lower = bytes2pgno(env, env->me_dbgeo.lower); meta.mm_geo.upper = bytes2pgno(env, env->me_dbgeo.upper); meta.mm_geo.grow = (uint16_t)bytes2pgno(env, env->me_dbgeo.grow); meta.mm_geo.shrink = (uint16_t)bytes2pgno(env, env->me_dbgeo.shrink); mdbx_verbose("amended: root %" PRIaPGNO "/%" PRIaPGNO ", geo %" PRIaPGNO "/%" PRIaPGNO "-%" PRIaPGNO "/%" PRIaPGNO " +%u -%u, txn_id %" PRIaTXN ", %s", meta.mm_dbs[MAIN_DBI].md_root, meta.mm_dbs[FREE_DBI].md_root, meta.mm_geo.lower, meta.mm_geo.next, meta.mm_geo.now, meta.mm_geo.upper, meta.mm_geo.grow, meta.mm_geo.shrink, unaligned_peek_u64(4, meta.mm_txnid_a), mdbx_durable_str(&meta)); } else { /* fetch back 'now/current' size, since it was ignored during comparison * and may differ. */ env->me_dbgeo.now = pgno_align2os_bytes(env, meta.mm_geo.now); } mdbx_ensure(env, meta.mm_geo.now >= meta.mm_geo.next); } else { /* geo-params are not pre-configured by user, * get current values from the meta. */ env->me_dbgeo.now = pgno2bytes(env, meta.mm_geo.now); env->me_dbgeo.lower = pgno2bytes(env, meta.mm_geo.lower); env->me_dbgeo.upper = pgno2bytes(env, meta.mm_geo.upper); env->me_dbgeo.grow = pgno2bytes(env, meta.mm_geo.grow); env->me_dbgeo.shrink = pgno2bytes(env, meta.mm_geo.shrink); } mdbx_ensure(env, pgno_align2os_bytes(env, meta.mm_geo.now) == env->me_dbgeo.now); mdbx_ensure(env, env->me_dbgeo.now >= used_bytes); if (unlikely(filesize_before != env->me_dbgeo.now)) { if (lck_rc != /* lck exclusive */ MDBX_RESULT_TRUE) { mdbx_verbose("filesize mismatch (expect %" PRIuPTR "b/%" PRIaPGNO "p, have %" PRIu64 "b/%" PRIaPGNO "p), " "assume other process working", env->me_dbgeo.now, bytes2pgno(env, env->me_dbgeo.now), filesize_before, bytes2pgno(env, (size_t)filesize_before)); } else { mdbx_warning("filesize mismatch (expect %" PRIuSIZE "b/%" PRIaPGNO "p, have %" PRIu64 "b/%" PRIaPGNO "p)", env->me_dbgeo.now, bytes2pgno(env, env->me_dbgeo.now), filesize_before, bytes2pgno(env, (size_t)filesize_before)); if (filesize_before < used_bytes) { mdbx_error("last-page beyond end-of-file (last %" PRIaPGNO ", have %" PRIaPGNO ")", meta.mm_geo.next, bytes2pgno(env, (size_t)filesize_before)); return MDBX_CORRUPTED; } if (env->me_flags & MDBX_RDONLY) { if (filesize_before & (env->me_os_psize - 1)) { mdbx_error("%s", "filesize should be rounded-up to system page"); return MDBX_WANNA_RECOVERY; } mdbx_warning("%s", "ignore filesize mismatch in readonly-mode"); } else { mdbx_verbose("will resize datafile to %" PRIuSIZE " bytes, %" PRIaPGNO " pages", env->me_dbgeo.now, bytes2pgno(env, env->me_dbgeo.now)); } } } mdbx_verbose("current boot-id %" PRIx64 "-%" PRIx64 " (%savailable)", bootid.x, bootid.y, (bootid.x | bootid.y) ? "" : "not-"); /* calculate readahead hint before mmap with zero redundant pages */ const bool readahead = (env->me_flags & MDBX_NORDAHEAD) == 0 && mdbx_is_readahead_reasonable(used_bytes, 0) == MDBX_RESULT_TRUE; err = mdbx_mmap(env->me_flags, &env->me_dxb_mmap, env->me_dbgeo.now, env->me_dbgeo.upper, lck_rc ? MMAP_OPTION_TRUNCATE : 0); if (unlikely(err != MDBX_SUCCESS)) return err; #if defined(MADV_DONTDUMP) err = madvise(env->me_map, env->me_dxb_mmap.limit, MADV_DONTDUMP) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; #endif /* MADV_DONTDUMP */ #if defined(MADV_DODUMP) if (mdbx_runtime_flags & MDBX_DBG_DUMP) { const size_t meta_length_aligned2os = pgno_align2os_bytes(env, NUM_METAS); err = madvise(env->me_map, meta_length_aligned2os, MADV_DODUMP) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; } #endif /* MADV_DODUMP */ #ifdef MDBX_USE_VALGRIND env->me_valgrind_handle = VALGRIND_CREATE_BLOCK(env->me_map, env->me_dxb_mmap.limit, "mdbx"); #endif /* MDBX_USE_VALGRIND */ mdbx_assert(env, used_bytes >= pgno2bytes(env, NUM_METAS) && used_bytes <= env->me_dxb_mmap.limit); #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) VALGRIND_MAKE_MEM_NOACCESS(env->me_map + used_bytes, env->me_dxb_mmap.limit - used_bytes); ASAN_POISON_MEMORY_REGION(env->me_map + used_bytes, env->me_dxb_mmap.limit - used_bytes); env->me_poison_edge = bytes2pgno(env, env->me_dxb_mmap.limit); #endif /* MDBX_USE_VALGRIND || __SANITIZE_ADDRESS__ */ const unsigned meta_clash_mask = mdbx_meta_eq_mask(env); if (unlikely(meta_clash_mask)) { if (/* not recovery mode */ env->me_stuck_meta < 0) { mdbx_error("meta-pages are clashed: mask 0x%d", meta_clash_mask); return MDBX_CORRUPTED; } else { mdbx_warning("ignore meta-pages clashing (mask 0x%d) in recovery mode", meta_clash_mask); } } while (likely(/* not recovery mode */ env->me_stuck_meta < 0)) { MDBX_meta *const head = mdbx_meta_head(env); const txnid_t head_txnid = mdbx_meta_txnid_fluid(env, head); MDBX_meta *const steady = mdbx_meta_steady(env); const txnid_t steady_txnid = mdbx_meta_txnid_fluid(env, steady); if (head_txnid == steady_txnid) break; if (lck_rc == /* lck exclusive */ MDBX_RESULT_TRUE) { mdbx_assert(env, META_IS_STEADY(steady) && !META_IS_STEADY(head)); if (meta_bootid_match(head)) { MDBX_meta clone = *head; uint64_t filesize = env->me_dbgeo.now; err = mdbx_validate_meta( env, &clone, &filesize, data_page(head), bytes2pgno(env, (uint8_t *)data_page(head) - env->me_map), nullptr, env->me_psize); if (err == MDBX_SUCCESS) { mdbx_warning( "opening after an unclean shutdown, but boot-id(%016" PRIx64 "-%016" PRIx64 ") is MATCH: rollback NOT needed, steady-sync NEEDED%s", bootid.x, bootid.y, (env->me_flags & MDBX_RDONLY) ? ", but unable in read-only mode" : ""); if (env->me_flags & MDBX_RDONLY) return MDBX_WANNA_RECOVERY /* LY: could not recovery/sync */; meta = clone; atomic_store32(env->me_unsynced_pages, meta.mm_geo.next, mo_Relaxed); break; } mdbx_warning("opening after an unclean shutdown, " "but boot-id(%016" PRIx64 "-%016" PRIx64 ") is MATCH, " "but last meta not valid, rollback needed", bootid.x, bootid.y); } if (env->me_flags & MDBX_RDONLY) { mdbx_error("rollback needed: (from head %" PRIaTXN " to steady %" PRIaTXN "), but unable in read-only mode", head_txnid, steady_txnid); return MDBX_WANNA_RECOVERY /* LY: could not recovery/rollback */; } const MDBX_meta *const meta0 = METAPAGE(env, 0); const MDBX_meta *const meta1 = METAPAGE(env, 1); const MDBX_meta *const meta2 = METAPAGE(env, 2); txnid_t undo_txnid = 0 /* zero means undo is unneeded */; while ( (head != meta0 && mdbx_meta_txnid_fluid(env, meta0) == undo_txnid) || (head != meta1 && mdbx_meta_txnid_fluid(env, meta1) == undo_txnid) || (head != meta2 && mdbx_meta_txnid_fluid(env, meta2) == undo_txnid)) undo_txnid = safe64_txnid_next(undo_txnid); if (unlikely(undo_txnid >= steady_txnid)) { mdbx_fatal("rollback failed: no suitable txnid (0,1,2) < %" PRIaTXN, steady_txnid); return MDBX_PANIC /* LY: could not recovery/rollback */; } /* LY: rollback weak checkpoint */ mdbx_notice("rollback: from %" PRIaTXN ", to %" PRIaTXN " as %" PRIaTXN, head_txnid, steady_txnid, undo_txnid); mdbx_ensure(env, head_txnid == mdbx_meta_txnid_stable(env, head)); if (env->me_flags & MDBX_WRITEMAP) { /* It is possible to update txnid without safe64_write(), * since DB opened exclusive for now */ unaligned_poke_u64(4, head->mm_txnid_a, undo_txnid); unaligned_poke_u64(4, head->mm_datasync_sign, MDBX_DATASIGN_WEAK); unaligned_poke_u64(4, head->mm_txnid_b, undo_txnid); const size_t offset = (uint8_t *)data_page(head) - env->me_dxb_mmap.dxb; const size_t paged_offset = floor_powerof2(offset, env->me_os_psize); const size_t paged_length = ceil_powerof2( env->me_psize + offset - paged_offset, env->me_os_psize); err = mdbx_msync(&env->me_dxb_mmap, paged_offset, paged_length, false); } else { MDBX_meta rollback = *head; mdbx_meta_set_txnid(env, &rollback, undo_txnid); unaligned_poke_u64(4, rollback.mm_datasync_sign, MDBX_DATASIGN_WEAK); err = mdbx_pwrite(env->me_lazy_fd, &rollback, sizeof(MDBX_meta), (uint8_t *)head - (uint8_t *)env->me_map); } if (err) { mdbx_error("error %d rollback from %" PRIaTXN ", to %" PRIaTXN " as %" PRIaTXN, err, head_txnid, steady_txnid, undo_txnid); return err; } mdbx_flush_incoherent_mmap(env->me_map, pgno2bytes(env, NUM_METAS), env->me_os_psize); mdbx_ensure(env, undo_txnid == mdbx_meta_txnid_fluid(env, head)); mdbx_ensure(env, 0 == mdbx_meta_eq_mask(env)); continue; } if (!env->me_lck) { /* LY: without-lck (read-only) mode, so it is impossible that other * process made weak checkpoint. */ mdbx_error("%s", "without-lck, unable recovery/rollback"); return MDBX_WANNA_RECOVERY; } /* LY: assume just have a collision with other running process, * or someone make a weak checkpoint */ mdbx_verbose("%s", "assume collision or online weak checkpoint"); break; } const MDBX_meta *head = mdbx_meta_head(env); if (lck_rc == /* lck exclusive */ MDBX_RESULT_TRUE) { /* re-check size after mmap */ if ((env->me_dxb_mmap.current & (env->me_os_psize - 1)) != 0 || env->me_dxb_mmap.current < used_bytes) { mdbx_error("unacceptable/unexpected datafile size %" PRIuPTR, env->me_dxb_mmap.current); return MDBX_PROBLEM; } if (env->me_dxb_mmap.current != env->me_dbgeo.now) { meta.mm_geo.now = bytes2pgno(env, env->me_dxb_mmap.current); mdbx_notice("need update meta-geo to filesize %" PRIuPTR " bytes, %" PRIaPGNO " pages", env->me_dxb_mmap.current, meta.mm_geo.now); } if (memcmp(&meta.mm_geo, &head->mm_geo, sizeof(meta.mm_geo))) { if ((env->me_flags & MDBX_RDONLY) != 0 || /* recovery mode */ env->me_stuck_meta >= 0) { mdbx_warning("skipped update meta.geo in %s mode: from l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u, to l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u", (env->me_stuck_meta < 0) ? "read-only" : "recovery", head->mm_geo.lower, head->mm_geo.now, head->mm_geo.upper, head->mm_geo.shrink, head->mm_geo.grow, meta.mm_geo.lower, meta.mm_geo.now, meta.mm_geo.upper, meta.mm_geo.shrink, meta.mm_geo.grow); } else { const txnid_t txnid = mdbx_meta_txnid_stable(env, head); const txnid_t next_txnid = safe64_txnid_next(txnid); if (unlikely(txnid > MAX_TXNID)) { mdbx_error("txnid overflow, raise %d", MDBX_TXN_FULL); return MDBX_TXN_FULL; } mdbx_notice("updating meta.geo: " "from l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u (txn#%" PRIaTXN "), " "to l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u (txn#%" PRIaTXN ")", head->mm_geo.lower, head->mm_geo.now, head->mm_geo.upper, head->mm_geo.shrink, head->mm_geo.grow, txnid, meta.mm_geo.lower, meta.mm_geo.now, meta.mm_geo.upper, meta.mm_geo.shrink, meta.mm_geo.grow, next_txnid); mdbx_ensure(env, mdbx_meta_eq(env, &meta, head)); mdbx_meta_set_txnid(env, &meta, next_txnid); err = mdbx_sync_locked(env, env->me_flags | MDBX_SHRINK_ALLOWED, &meta); if (err) { mdbx_error("error %d, while updating meta.geo: " "from l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u (txn#%" PRIaTXN "), " "to l%" PRIaPGNO "-n%" PRIaPGNO "-u%" PRIaPGNO "/s%u-g%u (txn#%" PRIaTXN ")", err, head->mm_geo.lower, head->mm_geo.now, head->mm_geo.upper, head->mm_geo.shrink, head->mm_geo.grow, txnid, meta.mm_geo.lower, meta.mm_geo.now, meta.mm_geo.upper, meta.mm_geo.shrink, meta.mm_geo.grow, next_txnid); return err; } } } } atomic_store32(env->me_discarded_tail, bytes2pgno(env, used_aligned2os_bytes), mo_AcquireRelease); if (used_aligned2os_bytes < env->me_dxb_mmap.current) { #if defined(MADV_REMOVE) if (lck_rc && (env->me_flags & MDBX_WRITEMAP) != 0 && /* not recovery mode */ env->me_stuck_meta < 0) { mdbx_notice("open-MADV_%s %u..%u", "REMOVE (deallocate file space)", env->me_discarded_tail->weak, bytes2pgno(env, env->me_dxb_mmap.current)); err = madvise(env->me_map + used_aligned2os_bytes, env->me_dxb_mmap.current - used_aligned2os_bytes, MADV_REMOVE) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; } #endif /* MADV_REMOVE */ #if defined(MADV_DONTNEED) mdbx_notice("open-MADV_%s %u..%u", "DONTNEED", env->me_discarded_tail->weak, bytes2pgno(env, env->me_dxb_mmap.current)); err = madvise(env->me_map + used_aligned2os_bytes, env->me_dxb_mmap.current - used_aligned2os_bytes, MADV_DONTNEED) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(POSIX_MADV_DONTNEED) err = ignore_enosys(posix_madvise( env->me_map + used_aligned2os_bytes, env->me_dxb_mmap.current - used_aligned2os_bytes, POSIX_MADV_DONTNEED)); if (unlikely(MDBX_IS_ERROR(err))) return err; #elif defined(POSIX_FADV_DONTNEED) err = ignore_enosys(posix_fadvise( env->me_lazy_fd, used_aligned2os_bytes, env->me_dxb_mmap.current - used_aligned2os_bytes, POSIX_FADV_DONTNEED)); if (unlikely(MDBX_IS_ERROR(err))) return err; #endif /* MADV_DONTNEED */ } err = mdbx_set_readahead(env, 0, used_bytes, readahead); if (err != MDBX_SUCCESS && lck_rc == /* lck exclusive */ MDBX_RESULT_TRUE) return err; return rc; } /******************************************************************************/ /* Open and/or initialize the lock region for the environment. */ static __cold int mdbx_setup_lck(MDBX_env *env, char *lck_pathname, mdbx_mode_t mode) { mdbx_assert(env, env->me_lazy_fd != INVALID_HANDLE_VALUE); mdbx_assert(env, env->me_lfd == INVALID_HANDLE_VALUE); int err = mdbx_openfile(MDBX_OPEN_LCK, env, lck_pathname, &env->me_lfd, mode); if (err != MDBX_SUCCESS) { if (!(err == MDBX_ENOFILE && (env->me_flags & MDBX_EXCLUSIVE)) && !((err == MDBX_EROFS || err == MDBX_EACCESS || err == MDBX_EPERM) && (env->me_flags & MDBX_RDONLY))) return err; /* ensure the file system is read-only */ err = mdbx_check_fs_rdonly(env->me_lazy_fd, lck_pathname, err); if (err != MDBX_SUCCESS && /* ignore ERROR_NOT_SUPPORTED for exclusive mode */ !(err == MDBX_ENOSYS && (env->me_flags & MDBX_EXCLUSIVE))) return err; /* LY: without-lck mode (e.g. exclusive or on read-only filesystem) */ /* beginning of a locked section ---------------------------------------- */ lcklist_lock(); mdbx_assert(env, env->me_lcklist_next == nullptr); env->me_lfd = INVALID_HANDLE_VALUE; const int rc = mdbx_lck_seize(env); if (MDBX_IS_ERROR(rc)) { /* Calling lcklist_detach_locked() is required to restore POSIX-filelock * and this job will be done by mdbx_env_close0(). */ lcklist_unlock(); return rc; } /* insert into inprocess lck-list */ env->me_lcklist_next = inprocess_lcklist_head; inprocess_lcklist_head = env; lcklist_unlock(); /* end of a locked section ---------------------------------------------- */ env->me_oldest = &env->me_lckless_stub.oldest; env->me_sync_timestamp = &env->me_lckless_stub.sync_timestamp; env->me_autosync_period = &env->me_lckless_stub.autosync_period; env->me_unsynced_pages = &env->me_lckless_stub.autosync_pending; env->me_autosync_threshold = &env->me_lckless_stub.autosync_threshold; env->me_discarded_tail = &env->me_lckless_stub.discarded_tail; env->me_meta_sync_txnid = &env->me_lckless_stub.meta_sync_txnid; env->me_maxreaders = UINT_MAX; #if MDBX_LOCKING > 0 env->me_wlock = &env->me_lckless_stub.wlock; #endif /* MDBX_LOCKING > 0 */ mdbx_debug("lck-setup:%s%s%s", " lck-less", (env->me_flags & MDBX_RDONLY) ? " readonly" : "", (rc == MDBX_RESULT_TRUE) ? " exclusive" : " cooperative"); return rc; } /* beginning of a locked section ------------------------------------------ */ lcklist_lock(); mdbx_assert(env, env->me_lcklist_next == nullptr); /* Try to get exclusive lock. If we succeed, then * nobody is using the lock region and we should initialize it. */ err = mdbx_lck_seize(env); if (MDBX_IS_ERROR(err)) { bailout: /* Calling lcklist_detach_locked() is required to restore POSIX-filelock * and this job will be done by mdbx_env_close0(). */ lcklist_unlock(); return err; } MDBX_env *inprocess_neighbor = nullptr; if (err == MDBX_RESULT_TRUE) { err = uniq_check(&env->me_lck_mmap, &inprocess_neighbor); if (MDBX_IS_ERROR(err)) goto bailout; if (inprocess_neighbor && ((mdbx_runtime_flags & MDBX_DBG_LEGACY_MULTIOPEN) == 0 || (inprocess_neighbor->me_flags & MDBX_EXCLUSIVE) != 0)) { err = MDBX_BUSY; goto bailout; } } const int lck_seize_rc = err; mdbx_debug("lck-setup:%s%s%s", " with-lck", (env->me_flags & MDBX_RDONLY) ? " readonly" : "", (lck_seize_rc == MDBX_RESULT_TRUE) ? " exclusive" : " cooperative"); uint64_t size = 0; err = mdbx_filesize(env->me_lfd, &size); if (unlikely(err != MDBX_SUCCESS)) goto bailout; if (lck_seize_rc == MDBX_RESULT_TRUE) { size = ceil_powerof2(env->me_maxreaders * sizeof(MDBX_reader) + sizeof(MDBX_lockinfo), env->me_os_psize); mdbx_jitter4testing(false); } else { if (env->me_flags & MDBX_EXCLUSIVE) { err = MDBX_BUSY; goto bailout; } if (size > INT_MAX || (size & (env->me_os_psize - 1)) != 0 || size < env->me_os_psize) { mdbx_error("lck-file has invalid size %" PRIu64 " bytes", size); err = MDBX_PROBLEM; goto bailout; } } const size_t maxreaders = ((size_t)size - sizeof(MDBX_lockinfo)) / sizeof(MDBX_reader); if (maxreaders < 4) { mdbx_error("lck-size too small (up to %" PRIuPTR " readers)", maxreaders); err = MDBX_PROBLEM; goto bailout; } env->me_maxreaders = (maxreaders <= MDBX_READERS_LIMIT) ? (unsigned)maxreaders : (unsigned)MDBX_READERS_LIMIT; err = mdbx_mmap((env->me_flags & MDBX_EXCLUSIVE) | MDBX_WRITEMAP, &env->me_lck_mmap, (size_t)size, (size_t)size, lck_seize_rc ? MMAP_OPTION_TRUNCATE | MMAP_OPTION_SEMAPHORE : MMAP_OPTION_SEMAPHORE); if (unlikely(err != MDBX_SUCCESS)) goto bailout; #ifdef MADV_DODUMP err = madvise(env->me_lck, size, MADV_DODUMP) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) goto bailout; #endif /* MADV_DODUMP */ #ifdef MADV_WILLNEED err = madvise(env->me_lck, size, MADV_WILLNEED) ? ignore_enosys(errno) : MDBX_SUCCESS; if (unlikely(MDBX_IS_ERROR(err))) goto bailout; #endif /* MADV_WILLNEED */ struct MDBX_lockinfo *const lck = env->me_lck; if (lck_seize_rc == MDBX_RESULT_TRUE) { /* LY: exclusive mode, check and reset lck content */ memset(lck, 0, (size_t)size); mdbx_jitter4testing(false); lck->mti_magic_and_version = MDBX_LOCK_MAGIC; lck->mti_os_and_format = MDBX_LOCK_FORMAT; err = mdbx_msync(&env->me_lck_mmap, 0, (size_t)size, false); if (unlikely(err != MDBX_SUCCESS)) { mdbx_error("initial-%s for lck-file failed", "msync"); goto bailout; } err = mdbx_fsync(env->me_lck_mmap.fd, MDBX_SYNC_SIZE); if (unlikely(err != MDBX_SUCCESS)) { mdbx_error("initial-%s for lck-file failed", "fsync"); goto bailout; } } else { if (lck->mti_magic_and_version != MDBX_LOCK_MAGIC) { mdbx_error("%s", "lock region has invalid magic/version"); err = ((lck->mti_magic_and_version >> 8) != MDBX_MAGIC) ? MDBX_INVALID : MDBX_VERSION_MISMATCH; goto bailout; } if (lck->mti_os_and_format != MDBX_LOCK_FORMAT) { mdbx_error("lock region has os/format 0x%" PRIx32 ", expected 0x%" PRIx32, lck->mti_os_and_format, MDBX_LOCK_FORMAT); err = MDBX_VERSION_MISMATCH; goto bailout; } } err = mdbx_lck_init(env, inprocess_neighbor, lck_seize_rc); if (MDBX_IS_ERROR(err)) goto bailout; mdbx_ensure(env, env->me_lcklist_next == nullptr); /* insert into inprocess lck-list */ env->me_lcklist_next = inprocess_lcklist_head; inprocess_lcklist_head = env; lcklist_unlock(); /* end of a locked section ------------------------------------------------ */ mdbx_assert(env, !MDBX_IS_ERROR(lck_seize_rc)); env->me_oldest = &lck->mti_oldest_reader; env->me_sync_timestamp = &lck->mti_sync_timestamp; env->me_autosync_period = &lck->mti_autosync_period; env->me_unsynced_pages = &lck->mti_unsynced_pages; env->me_autosync_threshold = &lck->mti_autosync_threshold; env->me_discarded_tail = &lck->mti_discarded_tail; env->me_meta_sync_txnid = &lck->mti_meta_sync_txnid; #if MDBX_LOCKING > 0 env->me_wlock = &lck->mti_wlock; #endif /* MDBX_LOCKING > 0 */ return lck_seize_rc; } __cold int mdbx_is_readahead_reasonable(size_t volume, intptr_t redundancy) { if (volume <= 1024 * 1024 * 4ul) return MDBX_RESULT_TRUE; const intptr_t pagesize = mdbx_syspagesize(); if (unlikely(pagesize < MIN_PAGESIZE || !is_powerof2(pagesize))) return MDBX_INCOMPATIBLE; #if defined(_WIN32) || defined(_WIN64) MEMORYSTATUSEX info; memset(&info, 0, sizeof(info)); info.dwLength = sizeof(info); if (!GlobalMemoryStatusEx(&info)) return GetLastError(); #endif const int log2page = log2n(pagesize); #if defined(_WIN32) || defined(_WIN64) const intptr_t total_ram_pages = (intptr_t)(info.ullTotalPhys >> log2page); #elif defined(_SC_PHYS_PAGES) const intptr_t total_ram_pages = sysconf(_SC_PHYS_PAGES); if (total_ram_pages == -1) return errno; #elif defined(_SC_AIX_REALMEM) const intptr_t total_ram_Kb = sysconf(_SC_AIX_REALMEM); if (total_ram_Kb == -1) return errno; const intptr_t total_ram_pages = (total_ram_Kb << 10) >> log2page; #elif defined(HW_USERMEM) || defined(HW_PHYSMEM64) || defined(HW_MEMSIZE) || \ defined(HW_PHYSMEM) size_t ram, len = sizeof(ram); static const int mib[] = { CTL_HW, #if defined(HW_USERMEM) HW_USERMEM #elif defined(HW_PHYSMEM64) HW_PHYSMEM64 #elif defined(HW_MEMSIZE) HW_MEMSIZE #else HW_PHYSMEM #endif }; if (sysctl( #ifdef SYSCTL_LEGACY_NONCONST_MIB (int *) #endif mib, ARRAY_LENGTH(mib), &ram, &len, NULL, 0) != 0) return errno; if (len != sizeof(ram)) return MDBX_ENOSYS; const intptr_t total_ram_pages = (intptr_t)(ram >> log2page); #else #error "FIXME: Get User-accessible or physical RAM" #endif if (total_ram_pages < 1) return MDBX_ENOSYS; const intptr_t volume_pages = (volume + pagesize - 1) >> log2page; const intptr_t redundancy_pages = (redundancy < 0) ? -(intptr_t)((-redundancy + pagesize - 1) >> log2page) : (intptr_t)(redundancy + pagesize - 1) >> log2page; if (volume_pages >= total_ram_pages || volume_pages + redundancy_pages >= total_ram_pages) return MDBX_RESULT_FALSE; #if defined(_WIN32) || defined(_WIN64) const intptr_t avail_ram_pages = (intptr_t)(info.ullAvailPhys >> log2page); #elif defined(_SC_AVPHYS_PAGES) const intptr_t avail_ram_pages = sysconf(_SC_AVPHYS_PAGES); if (avail_ram_pages == -1) return errno; #elif defined(__MACH__) mach_msg_type_number_t count = HOST_VM_INFO_COUNT; vm_statistics_data_t vmstat; mach_port_t mport = mach_host_self(); kern_return_t kerr = host_statistics(mach_host_self(), HOST_VM_INFO, (host_info_t)&vmstat, &count); mach_port_deallocate(mach_task_self(), mport); if (unlikely(kerr != KERN_SUCCESS)) return MDBX_ENOSYS; const intptr_t avail_ram_pages = vmstat.free_count; #elif defined(VM_TOTAL) || defined(VM_METER) struct vmtotal info; size_t len = sizeof(info); static const int mib[] = { CTL_VM, #if defined(VM_TOTAL) VM_TOTAL #elif defined(VM_METER) VM_METER #endif }; if (sysctl( #ifdef SYSCTL_LEGACY_NONCONST_MIB (int *) #endif mib, ARRAY_LENGTH(mib), &info, &len, NULL, 0) != 0) return errno; if (len != sizeof(info)) return MDBX_ENOSYS; const intptr_t avail_ram_pages = info.t_free; #else #error "FIXME: Get Available RAM" #endif if (avail_ram_pages < 1) return MDBX_ENOSYS; return (volume_pages + redundancy_pages >= avail_ram_pages) ? MDBX_RESULT_FALSE : MDBX_RESULT_TRUE; } /* Merge sync flags */ static uint32_t merge_sync_flags(const uint32_t a, const uint32_t b) { uint32_t r = a | b; /* avoid false MDBX_UTTERLY_NOSYNC */ if (F_ISSET(r, MDBX_UTTERLY_NOSYNC) && !F_ISSET(a, MDBX_UTTERLY_NOSYNC) && !F_ISSET(b, MDBX_UTTERLY_NOSYNC)) r = (r - MDBX_UTTERLY_NOSYNC) | MDBX_SAFE_NOSYNC; /* convert MDBX_DEPRECATED_MAPASYNC to MDBX_SAFE_NOSYNC */ if ((r & (MDBX_WRITEMAP | MDBX_DEPRECATED_MAPASYNC)) == (MDBX_WRITEMAP | MDBX_DEPRECATED_MAPASYNC) && !F_ISSET(r, MDBX_UTTERLY_NOSYNC)) r = (r - MDBX_DEPRECATED_MAPASYNC) | MDBX_SAFE_NOSYNC; /* force MDBX_NOMETASYNC if MDBX_SAFE_NOSYNC enabled */ if (r & MDBX_SAFE_NOSYNC) r |= MDBX_NOMETASYNC; assert(!(F_ISSET(r, MDBX_UTTERLY_NOSYNC) && !F_ISSET(a, MDBX_UTTERLY_NOSYNC) && !F_ISSET(b, MDBX_UTTERLY_NOSYNC))); return r; } __cold int mdbx_env_turn_for_recovery(MDBX_env *env, unsigned target_meta) { if (unlikely(target_meta >= NUM_METAS)) return MDBX_EINVAL; int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if ((env->me_flags & (MDBX_EXCLUSIVE | MDBX_RDONLY)) != MDBX_EXCLUSIVE) return MDBX_EPERM; MDBX_page *page = (env->me_flags & MDBX_WRITEMAP) ? pgno2page(env, target_meta) : memcpy(env->me_pbuf, pgno2page(env, target_meta), env->me_psize); page->mp_pgno = target_meta; page->mp_flags = P_META; MDBX_meta *meta = page_meta(page); unaligned_poke_u64(4, meta->mm_magic_and_version, MDBX_DATA_MAGIC); meta->mm_psize = env->me_psize; txnid_t txnid = mdbx_meta_txnid_stable(env, meta); const txnid_t txnid0 = mdbx_meta_txnid_stable(env, METAPAGE(env, 0)); if (target_meta != 0 && txnid <= txnid0) txnid = safe64_txnid_next(txnid0); const txnid_t txnid1 = mdbx_meta_txnid_stable(env, METAPAGE(env, 1)); if (target_meta != 1 && txnid <= txnid1) txnid = safe64_txnid_next(txnid1); const txnid_t txnid2 = mdbx_meta_txnid_stable(env, METAPAGE(env, 2)); if (target_meta != 2 && txnid <= txnid2) txnid = safe64_txnid_next(txnid2); if (!META_IS_STEADY(meta) || mdbx_recent_committed_txnid(env) != txnid) { if (unlikely(txnid > MAX_TXNID)) { mdbx_error("txnid overflow, raise %d", MDBX_TXN_FULL); return MDBX_TXN_FULL; } mdbx_meta_set_txnid(env, meta, txnid); unaligned_poke_u64(4, meta->mm_datasync_sign, mdbx_meta_sign(meta)); } if (env->me_flags & MDBX_WRITEMAP) { mdbx_flush_incoherent_cpu_writeback(); rc = mdbx_msync(&env->me_dxb_mmap, 0, pgno_align2os_bytes(env, target_meta), MDBX_SYNC_DATA | MDBX_SYNC_IODQ); } else { const mdbx_filehandle_t fd = (env->me_dsync_fd != INVALID_HANDLE_VALUE) ? env->me_dsync_fd : env->me_lazy_fd; rc = mdbx_pwrite(fd, page, env->me_psize, pgno2bytes(env, target_meta)); if (rc == MDBX_SUCCESS && fd == env->me_lazy_fd) rc = mdbx_fsync(env->me_lazy_fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ); } return rc; } __cold int mdbx_env_open_for_recovery(MDBX_env *env, const char *pathname, unsigned target_meta, bool writeable) { if (unlikely(target_meta >= NUM_METAS)) return MDBX_EINVAL; int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; env->me_stuck_meta = (int8_t)target_meta; return mdbx_env_open( env, pathname, writeable ? MDBX_EXCLUSIVE : MDBX_EXCLUSIVE | MDBX_RDONLY, 0); } typedef struct { void *buffer_for_free; char *lck, *dxb; size_t ent_len; } MDBX_handle_env_pathname; __cold static int mdbx_handle_env_pathname(MDBX_handle_env_pathname *ctx, const char *pathname, MDBX_env_flags_t *flags, const mdbx_mode_t mode) { int rc; memset(ctx, 0, sizeof(*ctx)); if (unlikely(!pathname)) return MDBX_EINVAL; #if defined(_WIN32) || defined(_WIN64) const size_t wlen = mbstowcs(nullptr, pathname, INT_MAX); if (wlen < 1 || wlen > /* MAX_PATH */ INT16_MAX) return ERROR_INVALID_NAME; wchar_t *const pathnameW = _alloca((wlen + 1) * sizeof(wchar_t)); if (wlen != mbstowcs(pathnameW, pathname, wlen + 1)) return ERROR_INVALID_NAME; const DWORD dwAttrib = GetFileAttributesW(pathnameW); if (dwAttrib == INVALID_FILE_ATTRIBUTES) { rc = GetLastError(); if (rc != MDBX_ENOFILE) return rc; if (mode == 0 || (*flags & MDBX_RDONLY) != 0) /* can't open existing */ return rc; /* auto-create directory if requested */ if ((*flags & MDBX_NOSUBDIR) == 0 && !CreateDirectoryW(pathnameW, nullptr)) { rc = GetLastError(); if (rc != ERROR_ALREADY_EXISTS) return rc; } } else { /* ignore passed MDBX_NOSUBDIR flag and set it automatically */ *flags |= MDBX_NOSUBDIR; if (dwAttrib & FILE_ATTRIBUTE_DIRECTORY) *flags -= MDBX_NOSUBDIR; } #else struct stat st; if (stat(pathname, &st)) { rc = errno; if (rc != MDBX_ENOFILE) return rc; if (mode == 0 || (*flags & MDBX_RDONLY) != 0) /* can't open existing */ return rc; /* auto-create directory if requested */ const mdbx_mode_t dir_mode = (/* inherit read/write permissions for group and others */ mode & (S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH)) | /* always add read/write/search for owner */ S_IRWXU | ((mode & S_IRGRP) ? /* +search if readable by group */ S_IXGRP : 0) | ((mode & S_IROTH) ? /* +search if readable by others */ S_IXOTH : 0); if ((*flags & MDBX_NOSUBDIR) == 0 && mkdir(pathname, dir_mode)) { rc = errno; if (rc != EEXIST) return rc; } } else { /* ignore passed MDBX_NOSUBDIR flag and set it automatically */ *flags |= MDBX_NOSUBDIR; if (S_ISDIR(st.st_mode)) *flags -= MDBX_NOSUBDIR; } #endif static const char dxb_name[] = MDBX_DATANAME; static const size_t dxb_name_len = sizeof(dxb_name) - 1; static const char lck_name[] = MDBX_LOCKNAME; static const char lock_suffix[] = MDBX_LOCK_SUFFIX; ctx->ent_len = strlen(pathname); if ((*flags & MDBX_NOSUBDIR) && ctx->ent_len >= dxb_name_len && !memcmp(dxb_name, pathname + ctx->ent_len - dxb_name_len, dxb_name_len)) { *flags -= MDBX_NOSUBDIR; ctx->ent_len -= dxb_name_len; } const size_t bytes_needed = ctx->ent_len * 2 + ((*flags & MDBX_NOSUBDIR) ? sizeof(lock_suffix) + 1 : sizeof(lck_name) + sizeof(dxb_name)); ctx->buffer_for_free = mdbx_malloc(bytes_needed); if (!ctx->buffer_for_free) return MDBX_ENOMEM; ctx->lck = ctx->buffer_for_free; if (*flags & MDBX_NOSUBDIR) { ctx->dxb = ctx->lck + ctx->ent_len + sizeof(lock_suffix); sprintf(ctx->lck, "%s%s", pathname, lock_suffix); strcpy(ctx->dxb, pathname); } else { ctx->dxb = ctx->lck + ctx->ent_len + sizeof(lck_name); sprintf(ctx->lck, "%.*s%s", (int)ctx->ent_len, pathname, lck_name); sprintf(ctx->dxb, "%.*s%s", (int)ctx->ent_len, pathname, dxb_name); } return MDBX_SUCCESS; } __cold int mdbx_env_delete(const char *pathname, MDBX_env_delete_mode_t mode) { switch (mode) { default: return MDBX_EINVAL; case MDBX_ENV_JUST_DELETE: case MDBX_ENV_ENSURE_UNUSED: case MDBX_ENV_WAIT_FOR_UNUSED: break; } MDBX_env dummy_env; memset(&dummy_env, 0, sizeof(dummy_env)); dummy_env.me_flags = (mode == MDBX_ENV_ENSURE_UNUSED) ? MDBX_EXCLUSIVE : MDBX_ENV_DEFAULTS; dummy_env.me_os_psize = (unsigned)mdbx_syspagesize(); dummy_env.me_psize = (unsigned)mdbx_default_pagesize(); dummy_env.me_pathname = (char *)pathname; MDBX_handle_env_pathname env_pathname; STATIC_ASSERT(sizeof(dummy_env.me_flags) == sizeof(MDBX_env_flags_t)); int rc = MDBX_RESULT_TRUE, err = mdbx_handle_env_pathname( &env_pathname, pathname, (MDBX_env_flags_t *)&dummy_env.me_flags, 0); if (likely(err == MDBX_SUCCESS)) { mdbx_filehandle_t clk_handle = INVALID_HANDLE_VALUE, dxb_handle = INVALID_HANDLE_VALUE; if (mode > MDBX_ENV_JUST_DELETE) { err = mdbx_openfile(MDBX_OPEN_DELETE, &dummy_env, env_pathname.dxb, &dxb_handle, 0); err = (err == MDBX_ENOFILE) ? MDBX_SUCCESS : err; if (err == MDBX_SUCCESS) { err = mdbx_openfile(MDBX_OPEN_DELETE, &dummy_env, env_pathname.lck, &clk_handle, 0); err = (err == MDBX_ENOFILE) ? MDBX_SUCCESS : err; } if (err == MDBX_SUCCESS && clk_handle != INVALID_HANDLE_VALUE) err = mdbx_lockfile(clk_handle, mode == MDBX_ENV_WAIT_FOR_UNUSED); if (err == MDBX_SUCCESS && dxb_handle != INVALID_HANDLE_VALUE) err = mdbx_lockfile(dxb_handle, mode == MDBX_ENV_WAIT_FOR_UNUSED); } if (err == MDBX_SUCCESS) { err = mdbx_removefile(env_pathname.dxb); if (err == MDBX_SUCCESS) rc = MDBX_SUCCESS; else if (err == MDBX_ENOFILE) err = MDBX_SUCCESS; } if (err == MDBX_SUCCESS) { err = mdbx_removefile(env_pathname.lck); if (err == MDBX_SUCCESS) rc = MDBX_SUCCESS; else if (err == MDBX_ENOFILE) err = MDBX_SUCCESS; } if (err == MDBX_SUCCESS && !(dummy_env.me_flags & MDBX_NOSUBDIR)) { err = mdbx_removedirectory(pathname); if (err == MDBX_SUCCESS) rc = MDBX_SUCCESS; else if (err == MDBX_ENOFILE) err = MDBX_SUCCESS; } if (dxb_handle != INVALID_HANDLE_VALUE) mdbx_closefile(dxb_handle); if (clk_handle != INVALID_HANDLE_VALUE) mdbx_closefile(clk_handle); } else if (err == MDBX_ENOFILE) err = MDBX_SUCCESS; mdbx_free(env_pathname.buffer_for_free); return (err == MDBX_SUCCESS) ? rc : err; } __cold int mdbx_env_open(MDBX_env *env, const char *pathname, MDBX_env_flags_t flags, mdbx_mode_t mode) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (flags & ~ENV_USABLE_FLAGS) return MDBX_EINVAL; if (env->me_lazy_fd != INVALID_HANDLE_VALUE || (env->me_flags & MDBX_ENV_ACTIVE) != 0) return MDBX_EPERM; /* pickup previously mdbx_env_set_flags(), * but avoid MDBX_UTTERLY_NOSYNC by disjunction */ const uint32_t saved_me_flags = env->me_flags; flags = merge_sync_flags(flags, env->me_flags); MDBX_handle_env_pathname env_pathname; rc = mdbx_handle_env_pathname(&env_pathname, pathname, &flags, mode); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; if (flags & MDBX_RDONLY) { /* LY: silently ignore irrelevant flags when * we're only getting read access */ flags &= ~(MDBX_WRITEMAP | MDBX_DEPRECATED_MAPASYNC | MDBX_SAFE_NOSYNC | MDBX_NOMETASYNC | MDBX_COALESCE | MDBX_LIFORECLAIM | MDBX_NOMEMINIT | MDBX_ACCEDE); } else { #if MDBX_MMAP_INCOHERENT_FILE_WRITE /* Temporary `workaround` for OpenBSD kernel's flaw. * See https://github.com/erthink/libmdbx/issues/67 */ if ((flags & MDBX_WRITEMAP) == 0) { if (flags & MDBX_ACCEDE) flags |= MDBX_WRITEMAP; else { mdbx_debug_log(MDBX_LOG_ERROR, __func__, __LINE__, "System (i.e. OpenBSD) requires MDBX_WRITEMAP because " "of an internal flaw(s) in a file/buffer/page cache.\n"); rc = 42 /* ENOPROTOOPT */; goto bailout; } } #endif /* MDBX_MMAP_INCOHERENT_FILE_WRITE */ } env->me_flags = (flags & ~MDBX_FATAL_ERROR) | MDBX_ENV_ACTIVE; env->me_pathname = mdbx_calloc(env_pathname.ent_len + 1, 1); env->me_dbxs = mdbx_calloc(env->me_maxdbs, sizeof(MDBX_dbx)); env->me_dbflags = mdbx_calloc(env->me_maxdbs, sizeof(env->me_dbflags[0])); env->me_dbiseqs = mdbx_calloc(env->me_maxdbs, sizeof(env->me_dbiseqs[0])); if (!(env->me_dbxs && env->me_pathname && env->me_dbflags && env->me_dbiseqs)) { rc = MDBX_ENOMEM; goto bailout; } memcpy(env->me_pathname, env_pathname.dxb, env_pathname.ent_len); env->me_dbxs[FREE_DBI].md_cmp = cmp_int_align4; /* aligned MDBX_INTEGERKEY */ env->me_dbxs[FREE_DBI].md_dcmp = cmp_lenfast; rc = mdbx_openfile(F_ISSET(flags, MDBX_RDONLY) ? MDBX_OPEN_DXB_READ : MDBX_OPEN_DXB_LAZY, env, env_pathname.dxb, &env->me_lazy_fd, mode); if (rc != MDBX_SUCCESS) goto bailout; mdbx_assert(env, env->me_dsync_fd == INVALID_HANDLE_VALUE); if ((flags & (MDBX_RDONLY | MDBX_SAFE_NOSYNC | MDBX_NOMETASYNC)) == 0) { rc = mdbx_openfile(MDBX_OPEN_DXB_DSYNC, env, env_pathname.dxb, &env->me_dsync_fd, 0); mdbx_ensure(env, (rc != MDBX_SUCCESS) == (env->me_dsync_fd == INVALID_HANDLE_VALUE)); } #if MDBX_LOCKING == MDBX_LOCKING_SYSV env->me_sysv_ipc.key = ftok(env_pathname.dxb, 42); if (env->me_sysv_ipc.key == -1) { rc = errno; goto bailout; } #endif /* MDBX_LOCKING */ #if !(defined(_WIN32) || defined(_WIN64)) if (mode == 0) { /* pickup mode for lck-file */ struct stat st; if (fstat(env->me_lazy_fd, &st)) { rc = errno; goto bailout; } mode = st.st_mode; } mode = (/* inherit read permissions for group and others */ mode & (S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH)) | /* always add read/write/search for owner */ S_IRUSR | S_IWUSR | ((mode & S_IRGRP) ? /* +write if readable by group */ S_IWGRP : 0) | ((mode & S_IROTH) ? /* +write if readable by others */ S_IWOTH : 0); #endif /* !Windows */ const int lck_rc = mdbx_setup_lck(env, env_pathname.lck, mode); if (MDBX_IS_ERROR(lck_rc)) { rc = lck_rc; goto bailout; } /* Set the position in files outside of the data to avoid corruption * due to erroneous use of file descriptors in the application code. */ mdbx_fseek(env->me_lfd, UINT64_C(1) << 63); mdbx_fseek(env->me_lazy_fd, UINT64_C(1) << 63); if (env->me_dsync_fd != INVALID_HANDLE_VALUE) mdbx_fseek(env->me_dsync_fd, UINT64_C(1) << 63); const MDBX_env_flags_t rigorous_flags = MDBX_SAFE_NOSYNC | MDBX_DEPRECATED_MAPASYNC; const MDBX_env_flags_t mode_flags = rigorous_flags | MDBX_NOMETASYNC | MDBX_LIFORECLAIM | MDBX_COALESCE | MDBX_NORDAHEAD; if (env->me_lck && lck_rc != MDBX_RESULT_TRUE && (env->me_flags & MDBX_RDONLY) == 0) { while (atomic_load32(&env->me_lck->mti_envmode, mo_AcquireRelease) == MDBX_RDONLY) { if (atomic_cas32(&env->me_lck->mti_envmode, MDBX_RDONLY, env->me_flags & mode_flags)) break; atomic_yield(); } if (env->me_flags & MDBX_ACCEDE) { /* pickup current mode-flags, including MDBX_LIFORECLAIM | * MDBX_COALESCE | MDBX_NORDAHEAD */ const unsigned diff = (env->me_lck->mti_envmode.weak ^ env->me_flags) & mode_flags; mdbx_notice("accede mode-flags: 0x%X, 0x%X -> 0x%X", diff, env->me_flags, env->me_flags ^ diff); env->me_flags ^= diff; } if ((env->me_lck->mti_envmode.weak ^ env->me_flags) & rigorous_flags) { mdbx_error("%s", "current mode/flags incompatible with requested"); rc = MDBX_INCOMPATIBLE; goto bailout; } } const int dxb_rc = mdbx_setup_dxb(env, lck_rc); if (MDBX_IS_ERROR(dxb_rc)) { rc = dxb_rc; goto bailout; } if (unlikely(/* recovery mode */ env->me_stuck_meta >= 0) && (lck_rc != /* exclusive */ MDBX_RESULT_TRUE || (flags & MDBX_EXCLUSIVE) == 0)) { mdbx_error("%s", "recovery requires exclusive mode"); rc = MDBX_BUSY; goto bailout; } mdbx_debug("opened dbenv %p", (void *)env); if (env->me_lck) { if (lck_rc == MDBX_RESULT_TRUE) { env->me_lck->mti_envmode.weak = env->me_flags & (mode_flags | MDBX_RDONLY); rc = mdbx_lck_downgrade(env); mdbx_debug("lck-downgrade-%s: rc %i", (env->me_flags & MDBX_EXCLUSIVE) ? "partial" : "full", rc); if (rc != MDBX_SUCCESS) goto bailout; } else { rc = mdbx_cleanup_dead_readers(env, false, NULL); if (MDBX_IS_ERROR(rc)) goto bailout; } if ((env->me_flags & MDBX_NOTLS) == 0) { rc = mdbx_rthc_alloc(&env->me_txkey, &env->me_lck->mti_readers[0], &env->me_lck->mti_readers[env->me_maxreaders]); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; env->me_flags |= MDBX_ENV_TXKEY; } } if ((flags & MDBX_RDONLY) == 0) { const size_t tsize = sizeof(MDBX_txn), size = tsize + env->me_maxdbs * (sizeof(MDBX_db) + sizeof(MDBX_cursor *) + sizeof(unsigned) + 1); rc = mdbx_memalign_alloc( env->me_os_psize, env->me_psize * (1 /* page buffer */ + 1 /* page killer buffer */), &env->me_pbuf); if (rc == MDBX_SUCCESS) { memset(env->me_pbuf, -1, env->me_psize * 2); MDBX_txn *txn = mdbx_calloc(1, size); if (txn) { txn->mt_dbs = (MDBX_db *)((char *)txn + tsize); txn->tw.cursors = (MDBX_cursor **)(txn->mt_dbs + env->me_maxdbs); txn->mt_dbiseqs = (unsigned *)(txn->tw.cursors + env->me_maxdbs); txn->mt_dbistate = (uint8_t *)(txn->mt_dbiseqs + env->me_maxdbs); txn->mt_env = env; txn->mt_dbxs = env->me_dbxs; txn->mt_flags = MDBX_TXN_FINISHED; env->me_txn0 = txn; rc = mdbx_dpl_alloc(txn); if (likely(rc == MDBX_SUCCESS)) { txn->tw.retired_pages = mdbx_pnl_alloc(MDBX_PNL_INITIAL); txn->tw.reclaimed_pglist = mdbx_pnl_alloc(MDBX_PNL_INITIAL); if (unlikely(!txn->tw.retired_pages || !txn->tw.reclaimed_pglist)) rc = MDBX_ENOMEM; } } else rc = MDBX_ENOMEM; } } #if MDBX_DEBUG if (rc == MDBX_SUCCESS) { MDBX_meta *meta = mdbx_meta_head(env); MDBX_db *db = &meta->mm_dbs[MAIN_DBI]; mdbx_debug("opened database version %u, pagesize %u", (uint8_t)unaligned_peek_u64(4, meta->mm_magic_and_version), env->me_psize); mdbx_debug("using meta page %" PRIaPGNO ", txn %" PRIaTXN, data_page(meta)->mp_pgno, mdbx_meta_txnid_fluid(env, meta)); mdbx_debug("depth: %u", db->md_depth); mdbx_debug("entries: %" PRIu64, db->md_entries); mdbx_debug("branch pages: %" PRIaPGNO, db->md_branch_pages); mdbx_debug("leaf pages: %" PRIaPGNO, db->md_leaf_pages); mdbx_debug("overflow pages: %" PRIaPGNO, db->md_overflow_pages); mdbx_debug("root: %" PRIaPGNO, db->md_root); mdbx_debug("schema_altered: %" PRIaTXN, db->md_mod_txnid); } #endif bailout: if (rc != MDBX_SUCCESS) { rc = mdbx_env_close0(env) ? MDBX_PANIC : rc; env->me_flags = saved_me_flags | ((rc != MDBX_PANIC) ? 0 : MDBX_FATAL_ERROR); } else { #if defined(MDBX_USE_VALGRIND) || defined(__SANITIZE_ADDRESS__) mdbx_txn_valgrind(env, nullptr); #endif } mdbx_free(env_pathname.buffer_for_free); return rc; } /* Destroy resources from mdbx_env_open(), clear our readers & DBIs */ static __cold int mdbx_env_close0(MDBX_env *env) { env->me_stuck_meta = -1; if (!(env->me_flags & MDBX_ENV_ACTIVE)) { mdbx_ensure(env, env->me_lcklist_next == nullptr); return MDBX_SUCCESS; } env->me_flags &= ~MDBX_ENV_ACTIVE; env->me_oldest = nullptr; env->me_sync_timestamp = nullptr; env->me_autosync_period = nullptr; env->me_unsynced_pages = nullptr; env->me_autosync_threshold = nullptr; env->me_discarded_tail = nullptr; env->me_meta_sync_txnid = nullptr; if (env->me_flags & MDBX_ENV_TXKEY) mdbx_rthc_remove(env->me_txkey); lcklist_lock(); const int rc = lcklist_detach_locked(env); lcklist_unlock(); if (env->me_map) { mdbx_munmap(&env->me_dxb_mmap); #ifdef MDBX_USE_VALGRIND VALGRIND_DISCARD(env->me_valgrind_handle); env->me_valgrind_handle = -1; #endif } if (env->me_dsync_fd != INVALID_HANDLE_VALUE) { (void)mdbx_closefile(env->me_dsync_fd); env->me_dsync_fd = INVALID_HANDLE_VALUE; } if (env->me_lazy_fd != INVALID_HANDLE_VALUE) { (void)mdbx_closefile(env->me_lazy_fd); env->me_lazy_fd = INVALID_HANDLE_VALUE; } if (env->me_lck) mdbx_munmap(&env->me_lck_mmap); if (env->me_lfd != INVALID_HANDLE_VALUE) { (void)mdbx_closefile(env->me_lfd); env->me_lfd = INVALID_HANDLE_VALUE; } if (env->me_dbxs) { for (unsigned i = env->me_numdbs; --i >= CORE_DBS;) mdbx_free(env->me_dbxs[i].md_name.iov_base); mdbx_free(env->me_dbxs); } mdbx_memalign_free(env->me_pbuf); mdbx_free(env->me_dbiseqs); mdbx_free(env->me_dbflags); mdbx_free(env->me_pathname); if (env->me_txn0) { mdbx_dpl_free(env->me_txn0); mdbx_txl_free(env->me_txn0->tw.lifo_reclaimed); mdbx_pnl_free(env->me_txn0->tw.retired_pages); mdbx_pnl_free(env->me_txn0->tw.spill_pages); mdbx_pnl_free(env->me_txn0->tw.reclaimed_pglist); mdbx_free(env->me_txn0); } env->me_flags = 0; return rc; } __cold int mdbx_env_close_ex(MDBX_env *env, bool dont_sync) { MDBX_page *dp; int rc = MDBX_SUCCESS; if (unlikely(!env)) return MDBX_EINVAL; if (unlikely(env->me_signature.weak != MDBX_ME_SIGNATURE)) return MDBX_EBADSIGN; #if MDBX_ENV_CHECKPID || !(defined(_WIN32) || defined(_WIN64)) /* Check the PID even if MDBX_ENV_CHECKPID=0 on non-Windows * platforms (i.e. where fork() is available). * This is required to legitimize a call after fork() * from a child process, that should be allowed to free resources. */ if (unlikely(env->me_pid != mdbx_getpid())) env->me_flags |= MDBX_FATAL_ERROR; #endif /* MDBX_ENV_CHECKPID */ if ((env->me_flags & (MDBX_RDONLY | MDBX_FATAL_ERROR)) == 0 && env->me_txn0) { if (env->me_txn0->mt_owner && env->me_txn0->mt_owner != mdbx_thread_self()) return MDBX_BUSY; } else dont_sync = true; if (!atomic_cas32(&env->me_signature, MDBX_ME_SIGNATURE, 0)) return MDBX_EBADSIGN; if (!dont_sync) { #if defined(_WIN32) || defined(_WIN64) /* On windows, without blocking is impossible to determine whether another * process is running a writing transaction or not. * Because in the "owner died" condition kernel don't release * file lock immediately. */ rc = mdbx_env_sync_internal(env, true, false); rc = (rc == MDBX_RESULT_TRUE) ? MDBX_SUCCESS : rc; #else struct stat st; if (unlikely(fstat(env->me_lazy_fd, &st))) rc = errno; else if (st.st_nlink > 0 /* don't sync deleted files */) { rc = mdbx_env_sync_internal(env, true, true); rc = (rc == MDBX_BUSY || rc == EAGAIN || rc == EACCES || rc == EBUSY || rc == EWOULDBLOCK || rc == MDBX_RESULT_TRUE) ? MDBX_SUCCESS : rc; } #endif } mdbx_assert(env, env->me_signature.weak == 0); rc = mdbx_env_close0(env) ? MDBX_PANIC : rc; mdbx_ensure(env, mdbx_fastmutex_destroy(&env->me_dbi_lock) == MDBX_SUCCESS); #if defined(_WIN32) || defined(_WIN64) /* me_remap_guard don't have destructor (Slim Reader/Writer Lock) */ DeleteCriticalSection(&env->me_windowsbug_lock); #else mdbx_ensure(env, mdbx_fastmutex_destroy(&env->me_remap_guard) == MDBX_SUCCESS); #endif /* Windows */ #if MDBX_LOCKING > MDBX_LOCKING_SYSV mdbx_ensure(env, mdbx_ipclock_destroy(&env->me_lckless_stub.wlock) == 0); #endif /* MDBX_LOCKING */ while ((dp = env->me_dp_reserve) != NULL) { ASAN_UNPOISON_MEMORY_REGION(dp, env->me_psize); VALGRIND_MAKE_MEM_DEFINED(&dp->mp_next, sizeof(dp->mp_next)); env->me_dp_reserve = dp->mp_next; mdbx_free(dp); } VALGRIND_DESTROY_MEMPOOL(env); mdbx_ensure(env, env->me_lcklist_next == nullptr); env->me_pid = 0; mdbx_free(env); return rc; } __cold int mdbx_env_close(MDBX_env *env) { return __inline_mdbx_env_close(env); } /* Compare two items pointing at aligned unsigned int's. */ static int __hot cmp_int_align4(const MDBX_val *a, const MDBX_val *b) { mdbx_assert(NULL, a->iov_len == b->iov_len); switch (a->iov_len) { case 4: return CMP2INT(unaligned_peek_u32(4, a->iov_base), unaligned_peek_u32(4, b->iov_base)); case 8: return CMP2INT(unaligned_peek_u64(4, a->iov_base), unaligned_peek_u64(4, b->iov_base)); default: mdbx_assert_fail(NULL, "invalid size for INTEGERKEY/INTEGERDUP", __func__, __LINE__); return 0; } } /* Compare two items pointing at 2-byte aligned unsigned int's. */ static int __hot cmp_int_align2(const MDBX_val *a, const MDBX_val *b) { mdbx_assert(NULL, a->iov_len == b->iov_len); switch (a->iov_len) { case 4: return CMP2INT(unaligned_peek_u32(2, a->iov_base), unaligned_peek_u32(2, b->iov_base)); case 8: return CMP2INT(unaligned_peek_u64(2, a->iov_base), unaligned_peek_u64(2, b->iov_base)); default: mdbx_assert_fail(NULL, "invalid size for INTEGERKEY/INTEGERDUP", __func__, __LINE__); return 0; } } /* Compare two items pointing at unsigned values with unknown alignment. * * This is also set as MDBX_INTEGERDUP|MDBX_DUPFIXED's MDBX_dbx.md_dcmp. */ static int __hot cmp_int_unaligned(const MDBX_val *a, const MDBX_val *b) { mdbx_assert(NULL, a->iov_len == b->iov_len); switch (a->iov_len) { case 4: return CMP2INT(unaligned_peek_u32(1, a->iov_base), unaligned_peek_u32(1, b->iov_base)); case 8: return CMP2INT(unaligned_peek_u64(1, a->iov_base), unaligned_peek_u64(1, b->iov_base)); default: mdbx_assert_fail(NULL, "invalid size for INTEGERKEY/INTEGERDUP", __func__, __LINE__); return 0; } } /* Compare two items lexically */ static int __hot cmp_lexical(const MDBX_val *a, const MDBX_val *b) { if (a->iov_len == b->iov_len) return memcmp(a->iov_base, b->iov_base, a->iov_len); const int diff_len = (a->iov_len < b->iov_len) ? -1 : 1; const size_t shortest = (a->iov_len < b->iov_len) ? a->iov_len : b->iov_len; int diff_data = memcmp(a->iov_base, b->iov_base, shortest); return likely(diff_data) ? diff_data : diff_len; } /* Compare two items in reverse byte order */ static int __hot cmp_reverse(const MDBX_val *a, const MDBX_val *b) { const uint8_t *pa = (const uint8_t *)a->iov_base + a->iov_len; const uint8_t *pb = (const uint8_t *)b->iov_base + b->iov_len; const size_t shortest = (a->iov_len < b->iov_len) ? a->iov_len : b->iov_len; const uint8_t *const end = pa - shortest; while (pa != end) { int diff = *--pa - *--pb; if (likely(diff)) return diff; } return CMP2INT(a->iov_len, b->iov_len); } /* Fast non-lexically comparator */ static int __hot cmp_lenfast(const MDBX_val *a, const MDBX_val *b) { int diff = CMP2INT(a->iov_len, b->iov_len); return likely(diff) ? diff : memcmp(a->iov_base, b->iov_base, a->iov_len); } static bool unsure_equal(MDBX_cmp_func cmp, const MDBX_val *a, const MDBX_val *b) { return cmp == cmp_lenfast || cmp == cmp_lexical || cmp == cmp_reverse || cmp == cmp_int_unaligned || cmp_lenfast(a, b) == 0; } /* Search for key within a page, using binary search. * Returns the smallest entry larger or equal to the key. * If exactp is non-null, stores whether the found entry was an exact match * in *exactp (1 or 0). * Updates the cursor index with the index of the found entry. * If no entry larger or equal to the key is found, returns NULL. */ static MDBX_node *__hot mdbx_node_search(MDBX_cursor *mc, const MDBX_val *key, int *exactp) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; const int nkeys = page_numkeys(mp); DKBUF; mdbx_debug("searching %u keys in %s %spage %" PRIaPGNO, nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "", mp->mp_pgno); int low = IS_LEAF(mp) ? 0 : 1; int high = nkeys - 1; *exactp = 0; if (unlikely(high < low)) { mc->mc_ki[mc->mc_top] = 0; return NULL; } int rc = 0, i = 0; MDBX_cmp_func *cmp = mc->mc_dbx->md_cmp; MDBX_val nodekey; if (unlikely(IS_LEAF2(mp))) { mdbx_cassert(mc, mp->mp_leaf2_ksize == mc->mc_db->md_xsize); nodekey.iov_len = mp->mp_leaf2_ksize; do { i = (low + high) >> 1; nodekey.iov_base = page_leaf2key(mp, i, nodekey.iov_len); mdbx_cassert(mc, (char *)mp + mc->mc_txn->mt_env->me_psize >= (char *)nodekey.iov_base + nodekey.iov_len); rc = cmp(key, &nodekey); mdbx_debug("found leaf index %u [%s], rc = %i", i, DKEY(&nodekey), rc); if (unlikely(rc == 0)) { *exactp = 1; break; } low = (rc < 0) ? low : i + 1; high = (rc < 0) ? i - 1 : high; } while (likely(low <= high)); /* Found entry is less than the key. */ /* Skip to get the smallest entry larger than key. */ i += rc > 0; /* store the key index */ mc->mc_ki[mc->mc_top] = (indx_t)i; return (i < nkeys) ? /* fake for LEAF2 */ (MDBX_node *)(intptr_t)-1 : /* There is no entry larger or equal to the key. */ NULL; } if (cmp == cmp_int_align2 && IS_BRANCH(mp)) /* Branch pages have no data, so if using integer keys, * alignment is guaranteed. Use faster mdbx_cmp_int_align4(). */ cmp = cmp_int_align4; MDBX_node *node; do { i = (low + high) >> 1; node = page_node(mp, i); nodekey.iov_len = node_ks(node); nodekey.iov_base = node_key(node); mdbx_cassert(mc, (char *)mp + mc->mc_txn->mt_env->me_psize >= (char *)nodekey.iov_base + nodekey.iov_len); rc = cmp(key, &nodekey); if (IS_LEAF(mp)) mdbx_debug("found leaf index %u [%s], rc = %i", i, DKEY(&nodekey), rc); else mdbx_debug("found branch index %u [%s -> %" PRIaPGNO "], rc = %i", i, DKEY(&nodekey), node_pgno(node), rc); if (unlikely(rc == 0)) { *exactp = 1; break; } low = (rc < 0) ? low : i + 1; high = (rc < 0) ? i - 1 : high; } while (likely(low <= high)); /* Found entry is less than the key. */ /* Skip to get the smallest entry larger than key. */ i += rc > 0; /* store the key index */ mc->mc_ki[mc->mc_top] = (indx_t)i; return (i < nkeys) ? page_node(mp, i) : /* There is no entry larger or equal to the key. */ NULL; } #if 0 /* unused for now */ static void mdbx_cursor_adjust(MDBX_cursor *mc, func) { MDBX_cursor *m2; for (m2 = mc->mc_txn->tw.cursors[mc->mc_dbi]; m2; m2 = m2->mc_next) { if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) { func(mc, m2); } } } #endif /* Pop a page off the top of the cursor's stack. */ static __inline void mdbx_cursor_pop(MDBX_cursor *mc) { if (mc->mc_snum) { mdbx_debug("popped page %" PRIaPGNO " off db %d cursor %p", mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *)mc); if (--mc->mc_snum) { mc->mc_top--; } else { mc->mc_flags &= ~C_INITIALIZED; } } } /* Push a page onto the top of the cursor's stack. * Set MDBX_TXN_ERROR on failure. */ static __inline int mdbx_cursor_push(MDBX_cursor *mc, MDBX_page *mp) { mdbx_debug("pushing page %" PRIaPGNO " on db %d cursor %p", mp->mp_pgno, DDBI(mc), (void *)mc); if (unlikely(mc->mc_snum >= CURSOR_STACK)) { mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_CURSOR_FULL; } mdbx_cassert(mc, mc->mc_snum < UINT16_MAX); mc->mc_top = mc->mc_snum++; mc->mc_pg[mc->mc_top] = mp; mc->mc_ki[mc->mc_top] = 0; return MDBX_SUCCESS; } /* Find the address of the page corresponding to a given page number. * Set MDBX_TXN_ERROR on failure. * * [in] mc the cursor accessing the page. * [in] pgno the page number for the page to retrieve. * [out] ret address of a pointer where the page's address will be * stored. * [out] lvl dirtylist inheritance level of found page. 1=current txn, * 0=mapped page. * * Returns 0 on success, non-zero on failure. */ __hot static int mdbx_page_get(MDBX_cursor *mc, pgno_t pgno, MDBX_page **ret, int *lvl, txnid_t pp_txnid) { MDBX_txn *txn = mc->mc_txn; if (unlikely(pgno >= txn->mt_next_pgno)) { mdbx_error("page #%" PRIaPGNO " beyond next-pgno", pgno); notfound: *ret = nullptr; txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PAGE_NOTFOUND; } MDBX_env *const env = txn->mt_env; MDBX_page *p = nullptr; mdbx_assert(env, ((txn->mt_flags ^ env->me_flags) & MDBX_WRITEMAP) == 0); mdbx_assert(env, pp_txnid >= MIN_TXNID && pp_txnid <= txn->mt_txnid); const uint16_t illegal_bits = (txn->mt_flags & MDBX_TXN_RDONLY) ? ~(P_BRANCH | P_LEAF | P_LEAF2 | P_OVERFLOW) : ~(P_BRANCH | P_LEAF | P_LEAF2 | P_OVERFLOW | P_DIRTY); int level; if (unlikely((txn->mt_flags & (MDBX_TXN_RDONLY | MDBX_WRITEMAP)) == 0)) { level = 1; do { /* Spilled pages were dirtied in this txn and flushed * because the dirty list got full. Bring this page * back in from the map (but don't unspill it here, * leave that unless page_touch happens again). */ if (unlikely(txn->mt_flags & MDBX_TXN_SPILLS) && txn->tw.spill_pages && mdbx_pnl_exist(txn->tw.spill_pages, pgno << 1)) { pp_txnid = txn->mt_txnid; goto spilled; } p = mdbx_dpl_find(txn->tw.dirtylist, pgno); if (p) goto dirty; level++; } while ((txn = txn->mt_parent) != NULL); } level = 0; spilled: p = pgno2page(env, pgno); dirty: *(lvl ? lvl : &level) = level; *ret = p; if (unlikely(p->mp_pgno != pgno)) { bad_page( p, "mismatch actual pgno (%" PRIaPGNO ") != expected (%" PRIaPGNO ")\n", p->mp_pgno, pgno); goto notfound; } if (unlikely(p->mp_flags & illegal_bits)) { bad_page(p, "invalid page's flags (%u)\n", p->mp_flags); goto corrupted; } if (unlikely(p->mp_txnid > ((p->mp_flags & P_DIRTY) ? UINT64_MAX : pp_txnid))) { bad_page(p, "page mod-txnid (%" PRIaTXN ") > parent (%" PRIaTXN ")\n", p->mp_txnid, pp_txnid); goto corrupted; } if (unlikely((p->mp_upper < p->mp_lower || ((p->mp_lower | p->mp_upper) & 1) != 0 || PAGEHDRSZ + p->mp_upper > env->me_psize) && !IS_OVERFLOW(p))) { bad_page(p, "invalid page lower(%u)/upper(%u) with limit (%u)\n", p->mp_lower, p->mp_upper, page_space(env)); goto corrupted; } if (mdbx_audit_enabled()) return mdbx_page_check(mc, p, C_UPDATING); return MDBX_SUCCESS; corrupted: mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_CORRUPTED; } /* Finish mdbx_page_search() / mdbx_page_search_lowest(). * The cursor is at the root page, set up the rest of it. */ __hot static int mdbx_page_search_root(MDBX_cursor *mc, const MDBX_val *key, int flags) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; int rc; DKBUF; while (IS_BRANCH(mp)) { MDBX_node *node; int i; mdbx_debug("branch page %" PRIaPGNO " has %u keys", mp->mp_pgno, page_numkeys(mp)); /* Don't assert on branch pages in the GC. We can get here * while in the process of rebalancing a GC branch page; we must * let that proceed. ITS#8336 */ mdbx_cassert(mc, !mc->mc_dbi || page_numkeys(mp) > 1); mdbx_debug("found index 0 to page %" PRIaPGNO, node_pgno(page_node(mp, 0))); if (flags & (MDBX_PS_FIRST | MDBX_PS_LAST)) { i = 0; if (flags & MDBX_PS_LAST) { i = page_numkeys(mp) - 1; /* if already init'd, see if we're already in right place */ if (mc->mc_flags & C_INITIALIZED) { if (mc->mc_ki[mc->mc_top] == i) { mc->mc_top = mc->mc_snum++; mp = mc->mc_pg[mc->mc_top]; goto ready; } } } } else { int exact; node = mdbx_node_search(mc, key, &exact); if (node == NULL) i = page_numkeys(mp) - 1; else { i = mc->mc_ki[mc->mc_top]; if (!exact) { mdbx_cassert(mc, i > 0); i--; } } mdbx_debug("following index %u for key [%s]", i, DKEY(key)); } mdbx_cassert(mc, i < (int)page_numkeys(mp)); node = page_node(mp, i); if (unlikely((rc = mdbx_page_get(mc, node_pgno(node), &mp, NULL, pp_txnid4chk(mp, mc->mc_txn))) != 0)) return rc; mc->mc_ki[mc->mc_top] = (indx_t)i; if (unlikely(rc = mdbx_cursor_push(mc, mp))) return rc; ready: if (flags & MDBX_PS_MODIFY) { if (unlikely((rc = mdbx_page_touch(mc)) != 0)) return rc; mp = mc->mc_pg[mc->mc_top]; } } if (unlikely(!IS_LEAF(mp))) { mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return bad_page(mp, "index points to a page with 0x%02x flags\n", mp->mp_flags); } mdbx_debug("found leaf page %" PRIaPGNO " for key [%s]", mp->mp_pgno, DKEY(key)); mc->mc_flags |= C_INITIALIZED; mc->mc_flags &= ~C_EOF; return MDBX_SUCCESS; } static int mdbx_setup_dbx(MDBX_dbx *const dbx, const MDBX_db *const db, const unsigned pagesize) { if (unlikely(!dbx->md_cmp)) { dbx->md_cmp = get_default_keycmp(db->md_flags); dbx->md_dcmp = get_default_datacmp(db->md_flags); } dbx->md_klen_min = (db->md_flags & MDBX_INTEGERKEY) ? 4 /* sizeof(uint32_t) */ : 0; dbx->md_klen_max = mdbx_limits_keysize_max(pagesize, db->md_flags); assert(dbx->md_klen_max != (unsigned)-1); dbx->md_vlen_min = (db->md_flags & MDBX_INTEGERDUP) ? 4 /* sizeof(uint32_t) */ : ((db->md_flags & MDBX_DUPFIXED) ? 1 : 0); dbx->md_vlen_max = mdbx_limits_valsize_max(pagesize, db->md_flags); assert(dbx->md_vlen_max != (unsigned)-1); if ((db->md_flags & (MDBX_DUPFIXED | MDBX_INTEGERDUP)) != 0 && db->md_xsize) { if (unlikely(db->md_xsize < dbx->md_vlen_min || db->md_xsize > dbx->md_vlen_max)) { mdbx_error("db.md_xsize (%u) <> min/max value-length (%zu/%zu)", db->md_xsize, dbx->md_vlen_min, dbx->md_vlen_max); return MDBX_CORRUPTED; } dbx->md_vlen_min = dbx->md_vlen_max = db->md_xsize; } return MDBX_SUCCESS; } static int mdbx_fetch_sdb(MDBX_txn *txn, MDBX_dbi dbi) { MDBX_cursor_couple couple; if (unlikely(TXN_DBI_CHANGED(txn, dbi))) return MDBX_BAD_DBI; int rc = mdbx_cursor_init(&couple.outer, txn, MAIN_DBI); if (unlikely(rc != MDBX_SUCCESS)) return rc; MDBX_dbx *const dbx = &txn->mt_dbxs[dbi]; rc = mdbx_page_search(&couple.outer, &dbx->md_name, 0); if (unlikely(rc != MDBX_SUCCESS)) return (rc == MDBX_NOTFOUND) ? MDBX_BAD_DBI : rc; MDBX_val data; int exact = 0; MDBX_node *node = mdbx_node_search(&couple.outer, &dbx->md_name, &exact); if (unlikely(!exact)) return MDBX_BAD_DBI; if (unlikely((node_flags(node) & (F_DUPDATA | F_SUBDATA)) != F_SUBDATA)) return MDBX_INCOMPATIBLE; /* not a named DB */ const txnid_t pp_txnid = pp_txnid4chk(couple.outer.mc_pg[couple.outer.mc_top], txn); rc = mdbx_node_read(&couple.outer, node, &data, pp_txnid); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(data.iov_len != sizeof(MDBX_db))) return MDBX_INCOMPATIBLE; /* not a named DB */ uint16_t md_flags = UNALIGNED_PEEK_16(data.iov_base, MDBX_db, md_flags); /* The txn may not know this DBI, or another process may * have dropped and recreated the DB with other flags. */ MDBX_db *const db = &txn->mt_dbs[dbi]; if (unlikely((db->md_flags & DB_PERSISTENT_FLAGS) != md_flags)) return MDBX_INCOMPATIBLE; memcpy(db, data.iov_base, sizeof(MDBX_db)); mdbx_tassert(txn, txn->mt_txnid >= pp_txnid); if (unlikely(db->md_mod_txnid > pp_txnid)) { mdbx_error("db.md_mod_txnid (%" PRIaTXN ") > page-txnid (%" PRIaTXN ")", db->md_mod_txnid, pp_txnid); return MDBX_CORRUPTED; } rc = mdbx_setup_dbx(dbx, db, txn->mt_env->me_psize); if (unlikely(rc != MDBX_SUCCESS)) return rc; txn->mt_dbistate[dbi] &= ~DBI_STALE; return MDBX_SUCCESS; } /* Search for the lowest key under the current branch page. * This just bypasses a numkeys check in the current page * before calling mdbx_page_search_root(), because the callers * are all in situations where the current page is known to * be underfilled. */ __hot static int mdbx_page_search_lowest(MDBX_cursor *mc) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; mdbx_cassert(mc, IS_BRANCH(mp)); MDBX_node *node = page_node(mp, 0); int rc; if (unlikely((rc = mdbx_page_get(mc, node_pgno(node), &mp, NULL, pp_txnid4chk(mp, mc->mc_txn))) != 0)) return rc; mc->mc_ki[mc->mc_top] = 0; if (unlikely(rc = mdbx_cursor_push(mc, mp))) return rc; return mdbx_page_search_root(mc, NULL, MDBX_PS_FIRST); } /* Search for the page a given key should be in. * Push it and its parent pages on the cursor stack. * * [in,out] mc the cursor for this operation. * [in] key the key to search for, or NULL for first/last page. * [in] flags If MDBX_PS_MODIFY is set, visited pages in the DB * are touched (updated with new page numbers). * If MDBX_PS_FIRST or MDBX_PS_LAST is set, find first or last * leaf. * This is used by mdbx_cursor_first() and mdbx_cursor_last(). * If MDBX_PS_ROOTONLY set, just fetch root node, no further * lookups. * * Returns 0 on success, non-zero on failure. */ __hot static int mdbx_page_search(MDBX_cursor *mc, const MDBX_val *key, int flags) { int rc; pgno_t root; /* Make sure the txn is still viable, then find the root from * the txn's db table and set it as the root of the cursor's stack. */ if (unlikely(mc->mc_txn->mt_flags & MDBX_TXN_BLOCKED)) { mdbx_debug("%s", "transaction has failed, must abort"); return MDBX_BAD_TXN; } /* Make sure we're using an up-to-date root */ if (unlikely(*mc->mc_dbistate & DBI_STALE)) { rc = mdbx_fetch_sdb(mc->mc_txn, mc->mc_dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; } root = mc->mc_db->md_root; if (unlikely(root == P_INVALID)) { /* Tree is empty. */ mdbx_debug("%s", "tree is empty"); return MDBX_NOTFOUND; } mdbx_cassert(mc, root >= NUM_METAS); if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) { if (unlikely((rc = mdbx_page_get( mc, root, &mc->mc_pg[0], NULL, (/* maybe zero in legacy DB */ mc->mc_db->md_mod_txnid && !(*mc->mc_dbistate & DBI_DIRTY)) ? mc->mc_db->md_mod_txnid : mc->mc_txn->mt_txnid)) != 0)) return rc; } mc->mc_snum = 1; mc->mc_top = 0; mdbx_debug("db %d root page %" PRIaPGNO " has flags 0x%X", DDBI(mc), root, mc->mc_pg[0]->mp_flags); if (flags & MDBX_PS_MODIFY) { if (unlikely(rc = mdbx_page_touch(mc))) return rc; } if (flags & MDBX_PS_ROOTONLY) return MDBX_SUCCESS; return mdbx_page_search_root(mc, key, flags); } /* Return the data associated with a given node. * * [in] mc The cursor for this operation. * [in] leaf The node being read. * [out] data Updated to point to the node's data. * * Returns 0 on success, non-zero on failure. */ static __always_inline int mdbx_node_read(MDBX_cursor *mc, MDBX_node *node, MDBX_val *data, const txnid_t pp_txnid) { data->iov_len = node_ds(node); data->iov_base = node_data(node); if (unlikely(F_ISSET(node_flags(node), F_BIGDATA))) { /* Read overflow data. */ MDBX_page *omp; /* overflow page */ int rc = mdbx_page_get(mc, node_largedata_pgno(node), &omp, NULL, pp_txnid); if (unlikely((rc != MDBX_SUCCESS))) { mdbx_debug("read overflow page %" PRIaPGNO " failed", node_largedata_pgno(node)); return rc; } data->iov_base = page_data(omp); } return MDBX_SUCCESS; } int mdbx_get(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, MDBX_val *data) { DKBUF; mdbx_debug("===> get db %u key [%s]", dbi, DKEY(key)); int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key || !data)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; int exact = 0; return mdbx_cursor_set(&cx.outer, (MDBX_val *)key, data, MDBX_SET, &exact); } int mdbx_get_equal_or_great(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *key, MDBX_val *data) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key || !data)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(txn->mt_flags & MDBX_TXN_BLOCKED)) return MDBX_BAD_TXN; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_cursor_get(&cx.outer, key, data, MDBX_SET_LOWERBOUND); } int mdbx_get_ex(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *key, MDBX_val *data, size_t *values_count) { DKBUF; mdbx_debug("===> get db %u key [%s]", dbi, DKEY(key)); int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key || !data)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; int exact = 0; rc = mdbx_cursor_set(&cx.outer, key, data, MDBX_SET_KEY, &exact); if (unlikely(rc != MDBX_SUCCESS)) { if (rc == MDBX_NOTFOUND && values_count) *values_count = 0; return rc; } if (values_count) { *values_count = 1; if (cx.outer.mc_xcursor != NULL) { MDBX_node *node = page_node(cx.outer.mc_pg[cx.outer.mc_top], cx.outer.mc_ki[cx.outer.mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { mdbx_tassert(txn, cx.outer.mc_xcursor == &cx.inner && (cx.inner.mx_cursor.mc_flags & C_INITIALIZED)); *values_count = (sizeof(*values_count) >= sizeof(cx.inner.mx_db.md_entries) || cx.inner.mx_db.md_entries <= PTRDIFF_MAX) ? (size_t)cx.inner.mx_db.md_entries : PTRDIFF_MAX; } } } return MDBX_SUCCESS; } /* Find a sibling for a page. * Replaces the page at the top of the cursor's stack with the specified * sibling, if one exists. * * [in] mc The cursor for this operation. * [in] dir SIBLING_LEFT or SIBLING_RIGHT. * * Returns 0 on success, non-zero on failure. */ static int mdbx_cursor_sibling(MDBX_cursor *mc, int dir) { int rc; MDBX_node *node; MDBX_page *mp; assert(dir == SIBLING_LEFT || dir == SIBLING_RIGHT); if (unlikely(mc->mc_snum < 2)) return MDBX_NOTFOUND; /* root has no siblings */ mdbx_cursor_pop(mc); mdbx_debug("parent page is page %" PRIaPGNO ", index %u", mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]); if ((dir == SIBLING_RIGHT) ? (mc->mc_ki[mc->mc_top] + 1u >= page_numkeys(mc->mc_pg[mc->mc_top])) : (mc->mc_ki[mc->mc_top] == 0)) { mdbx_debug("no more keys aside, moving to next %s sibling", dir ? "right" : "left"); if (unlikely((rc = mdbx_cursor_sibling(mc, dir)) != MDBX_SUCCESS)) { /* undo cursor_pop before returning */ mc->mc_top++; mc->mc_snum++; return rc; } } else { assert((dir - 1) == -1 || (dir - 1) == 1); mc->mc_ki[mc->mc_top] += (indx_t)(dir - 1); mdbx_debug("just moving to %s index key %u", (dir == SIBLING_RIGHT) ? "right" : "left", mc->mc_ki[mc->mc_top]); } mdbx_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top])); node = page_node(mp = mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (unlikely((rc = mdbx_page_get(mc, node_pgno(node), &mp, NULL, pp_txnid4chk(mp, mc->mc_txn))) != 0)) { /* mc will be inconsistent if caller does mc_snum++ as above */ mc->mc_flags &= ~(C_INITIALIZED | C_EOF); return rc; } rc = mdbx_cursor_push(mc, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; mc->mc_ki[mc->mc_top] = (indx_t)((dir == SIBLING_LEFT) ? page_numkeys(mp) - 1 : 0); return MDBX_SUCCESS; } /* Move the cursor to the next data item. */ static int mdbx_cursor_next(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op) { MDBX_page *mp; MDBX_node *node; int rc; if ((mc->mc_flags & C_DEL) && op == MDBX_NEXT_DUP) return MDBX_NOTFOUND; if (!(mc->mc_flags & C_INITIALIZED)) return mdbx_cursor_first(mc, key, data); mp = mc->mc_pg[mc->mc_top]; if (mc->mc_flags & C_EOF) { if (mc->mc_ki[mc->mc_top] + 1u >= page_numkeys(mp)) return MDBX_NOTFOUND; mc->mc_flags ^= C_EOF; } if (mc->mc_db->md_flags & MDBX_DUPSORT) { node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { if (op == MDBX_NEXT || op == MDBX_NEXT_DUP) { rc = mdbx_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDBX_NEXT); if (op != MDBX_NEXT || rc != MDBX_NOTFOUND) { if (likely(rc == MDBX_SUCCESS)) get_key_optional(node, key); return rc; } } } else { mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); if (op == MDBX_NEXT_DUP) return MDBX_NOTFOUND; } } mdbx_debug("cursor_next: top page is %" PRIaPGNO " in cursor %p", mp->mp_pgno, (void *)mc); if (mc->mc_flags & C_DEL) { mc->mc_flags ^= C_DEL; goto skip; } int ki = mc->mc_ki[mc->mc_top]; mc->mc_ki[mc->mc_top] = (indx_t)++ki; const int numkeys = page_numkeys(mp); if (unlikely(ki >= numkeys)) { mdbx_debug("%s", "=====> move to next sibling page"); mc->mc_ki[mc->mc_top] = (indx_t)(numkeys - 1); if (unlikely((rc = mdbx_cursor_sibling(mc, SIBLING_RIGHT)) != MDBX_SUCCESS)) { mc->mc_flags |= C_EOF; return rc; } mp = mc->mc_pg[mc->mc_top]; mdbx_debug("next page is %" PRIaPGNO ", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]); } skip: mdbx_debug("==> cursor points to page %" PRIaPGNO " with %u keys, key index %u", mp->mp_pgno, page_numkeys(mp), mc->mc_ki[mc->mc_top]); if (unlikely(!IS_LEAF(mp))) return MDBX_CORRUPTED; if (IS_LEAF2(mp)) { if (likely(key)) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mp, mc->mc_ki[mc->mc_top], key->iov_len); } return MDBX_SUCCESS; } node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { rc = mdbx_xcursor_init1(mc, node, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); if (unlikely(rc != MDBX_SUCCESS)) return rc; } else if (likely(data)) { if (unlikely((rc = mdbx_node_read(mc, node, data, pp_txnid4chk(mp, mc->mc_txn))) != MDBX_SUCCESS)) return rc; } get_key_optional(node, key); return MDBX_SUCCESS; } /* Move the cursor to the previous data item. */ static int mdbx_cursor_prev(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op) { MDBX_page *mp; MDBX_node *node; int rc; if ((mc->mc_flags & C_DEL) && op == MDBX_PREV_DUP) return MDBX_NOTFOUND; if (!(mc->mc_flags & C_INITIALIZED)) { rc = mdbx_cursor_last(mc, key, data); if (unlikely(rc)) return rc; mc->mc_ki[mc->mc_top]++; } mp = mc->mc_pg[mc->mc_top]; if ((mc->mc_db->md_flags & MDBX_DUPSORT) && mc->mc_ki[mc->mc_top] < page_numkeys(mp)) { node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { if (op == MDBX_PREV || op == MDBX_PREV_DUP) { rc = mdbx_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDBX_PREV); if (op != MDBX_PREV || rc != MDBX_NOTFOUND) { if (likely(rc == MDBX_SUCCESS)) { get_key_optional(node, key); mc->mc_flags &= ~C_EOF; } return rc; } } } else { mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); if (op == MDBX_PREV_DUP) return MDBX_NOTFOUND; } } mdbx_debug("cursor_prev: top page is %" PRIaPGNO " in cursor %p", mp->mp_pgno, (void *)mc); mc->mc_flags &= ~(C_EOF | C_DEL); int ki = mc->mc_ki[mc->mc_top]; mc->mc_ki[mc->mc_top] = (indx_t)--ki; if (unlikely(ki < 0)) { mc->mc_ki[mc->mc_top] = 0; mdbx_debug("%s", "=====> move to prev sibling page"); if ((rc = mdbx_cursor_sibling(mc, SIBLING_LEFT)) != MDBX_SUCCESS) return rc; mp = mc->mc_pg[mc->mc_top]; mdbx_debug("prev page is %" PRIaPGNO ", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]); } mdbx_debug("==> cursor points to page %" PRIaPGNO " with %u keys, key index %u", mp->mp_pgno, page_numkeys(mp), mc->mc_ki[mc->mc_top]); if (unlikely(!IS_LEAF(mp))) return MDBX_CORRUPTED; if (IS_LEAF2(mp)) { if (likely(key)) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mp, mc->mc_ki[mc->mc_top], key->iov_len); } return MDBX_SUCCESS; } node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { rc = mdbx_xcursor_init1(mc, node, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); if (unlikely(rc != MDBX_SUCCESS)) return rc; } else if (likely(data)) { if (unlikely((rc = mdbx_node_read(mc, node, data, pp_txnid4chk(mp, mc->mc_txn))) != MDBX_SUCCESS)) return rc; } get_key_optional(node, key); return MDBX_SUCCESS; } /* Set the cursor on a specific data item. */ static int mdbx_cursor_set(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op, int *exactp) { int rc; MDBX_page *mp; MDBX_node *node = NULL; int stub_exactp; DKBUF; exactp = exactp ? exactp : &stub_exactp; if (unlikely(key->iov_len < mc->mc_dbx->md_klen_min || key->iov_len > mc->mc_dbx->md_klen_max)) { mdbx_cassert(mc, !"Invalid key-size"); return MDBX_BAD_VALSIZE; } MDBX_val aligned_key = *key; uint64_t aligned_keybytes; if (mc->mc_db->md_flags & MDBX_INTEGERKEY) { switch (aligned_key.iov_len) { default: mdbx_cassert(mc, !"key-size is invalid for MDBX_INTEGERKEY"); return MDBX_BAD_VALSIZE; case 4: if (unlikely(3 & (uintptr_t)aligned_key.iov_base)) /* copy instead of return error to avoid break compatibility */ aligned_key.iov_base = memcpy(&aligned_keybytes, aligned_key.iov_base, 4); break; case 8: if (unlikely(7 & (uintptr_t)aligned_key.iov_base)) /* copy instead of return error to avoid break compatibility */ aligned_key.iov_base = memcpy(&aligned_keybytes, aligned_key.iov_base, 8); break; } } if (mc->mc_xcursor) mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); /* See if we're already on the right page */ if (mc->mc_flags & C_INITIALIZED) { MDBX_val nodekey; mdbx_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); mp = mc->mc_pg[mc->mc_top]; if (unlikely(!page_numkeys(mp))) { mc->mc_ki[mc->mc_top] = 0; mc->mc_flags |= C_EOF; return MDBX_NOTFOUND; } if (IS_LEAF2(mp)) { nodekey.iov_len = mc->mc_db->md_xsize; nodekey.iov_base = page_leaf2key(mp, 0, nodekey.iov_len); } else { node = page_node(mp, 0); get_key(node, &nodekey); } rc = mc->mc_dbx->md_cmp(&aligned_key, &nodekey); if (unlikely(rc == 0)) { /* Probably happens rarely, but first node on the page * was the one we wanted. */ mc->mc_ki[mc->mc_top] = 0; *exactp = 1; mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); goto set1; } if (rc > 0) { const unsigned nkeys = page_numkeys(mp); unsigned i; if (nkeys > 1) { if (IS_LEAF2(mp)) { nodekey.iov_base = page_leaf2key(mp, nkeys - 1, nodekey.iov_len); } else { node = page_node(mp, nkeys - 1); get_key(node, &nodekey); } rc = mc->mc_dbx->md_cmp(&aligned_key, &nodekey); if (rc == 0) { /* last node was the one we wanted */ mdbx_cassert(mc, nkeys >= 1 && nkeys <= UINT16_MAX + 1); mc->mc_ki[mc->mc_top] = (indx_t)(nkeys - 1); *exactp = 1; mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); goto set1; } if (rc < 0) { if (mc->mc_ki[mc->mc_top] < page_numkeys(mp)) { /* This is definitely the right page, skip search_page */ if (IS_LEAF2(mp)) { nodekey.iov_base = page_leaf2key(mp, mc->mc_ki[mc->mc_top], nodekey.iov_len); } else { node = page_node(mp, mc->mc_ki[mc->mc_top]); get_key(node, &nodekey); } rc = mc->mc_dbx->md_cmp(&aligned_key, &nodekey); if (rc == 0) { /* current node was the one we wanted */ *exactp = 1; mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); goto set1; } } rc = 0; mc->mc_flags &= ~C_EOF; goto set2; } } /* If any parents have right-sibs, search. * Otherwise, there's nothing further. */ for (i = 0; i < mc->mc_top; i++) if (mc->mc_ki[i] < page_numkeys(mc->mc_pg[i]) - 1) break; if (i == mc->mc_top) { /* There are no other pages */ mdbx_cassert(mc, nkeys <= UINT16_MAX); mc->mc_ki[mc->mc_top] = (uint16_t)nkeys; mc->mc_flags |= C_EOF; return MDBX_NOTFOUND; } } if (!mc->mc_top) { /* There are no other pages */ mc->mc_ki[mc->mc_top] = 0; if (op == MDBX_SET_RANGE) { rc = 0; goto set1; } else { mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); return MDBX_NOTFOUND; } } } else { mc->mc_pg[0] = 0; } rc = mdbx_page_search(mc, &aligned_key, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; mp = mc->mc_pg[mc->mc_top]; mdbx_cassert(mc, IS_LEAF(mp)); set2: node = mdbx_node_search(mc, &aligned_key, exactp); if (!*exactp && op != MDBX_SET_RANGE) { /* MDBX_SET specified and not an exact match. */ if (unlikely(mc->mc_ki[mc->mc_top] >= page_numkeys(mc->mc_pg[mc->mc_top]))) mc->mc_flags |= C_EOF; return MDBX_NOTFOUND; } if (node == NULL) { mdbx_debug("%s", "===> inexact leaf not found, goto sibling"); if (unlikely((rc = mdbx_cursor_sibling(mc, SIBLING_RIGHT)) != MDBX_SUCCESS)) { mc->mc_flags |= C_EOF; return rc; /* no entries matched */ } mp = mc->mc_pg[mc->mc_top]; mdbx_cassert(mc, IS_LEAF(mp)); if (!IS_LEAF2(mp)) node = page_node(mp, 0); } mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); set1: mc->mc_flags |= C_INITIALIZED; mc->mc_flags &= ~C_EOF; if (IS_LEAF2(mp)) { if (op == MDBX_SET_RANGE || op == MDBX_SET_KEY) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mp, mc->mc_ki[mc->mc_top], key->iov_len); } return MDBX_SUCCESS; } if (F_ISSET(node_flags(node), F_DUPDATA)) { rc = mdbx_xcursor_init1(mc, node, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (op == MDBX_SET || op == MDBX_SET_KEY || op == MDBX_SET_RANGE) { rc = mdbx_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); } else { rc = mdbx_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDBX_SET_RANGE, NULL); } if (unlikely(rc != MDBX_SUCCESS)) return rc; } else if (likely(data)) { if (op == MDBX_GET_BOTH || op == MDBX_GET_BOTH_RANGE) { if (unlikely(data->iov_len < mc->mc_dbx->md_vlen_min || data->iov_len > mc->mc_dbx->md_vlen_max)) { mdbx_cassert(mc, !"Invalid data-size"); return MDBX_BAD_VALSIZE; } MDBX_val aligned_data = *data; uint64_t aligned_databytes; if (mc->mc_db->md_flags & MDBX_INTEGERDUP) { switch (aligned_data.iov_len) { default: mdbx_cassert(mc, !"data-size is invalid for MDBX_INTEGERDUP"); return MDBX_BAD_VALSIZE; case 4: if (unlikely(3 & (uintptr_t)aligned_data.iov_base)) /* copy instead of return error to avoid break compatibility */ aligned_data.iov_base = memcpy(&aligned_databytes, aligned_data.iov_base, 4); break; case 8: if (unlikely(7 & (uintptr_t)aligned_data.iov_base)) /* copy instead of return error to avoid break compatibility */ aligned_data.iov_base = memcpy(&aligned_databytes, aligned_data.iov_base, 8); break; } } MDBX_val olddata; if (unlikely((rc = mdbx_node_read( mc, node, &olddata, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn))) != MDBX_SUCCESS)) return rc; rc = mc->mc_dbx->md_dcmp(&aligned_data, &olddata); if (rc) { mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); if (op != MDBX_GET_BOTH_RANGE || rc > 0) return MDBX_NOTFOUND; *exactp = 0; rc = 0; } *data = olddata; } else if (unlikely((rc = mdbx_node_read(mc, node, data, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn))) != MDBX_SUCCESS)) return rc; } /* The key already matches in all other cases */ if (op == MDBX_SET_RANGE || op == MDBX_SET_KEY) get_key_optional(node, key); mdbx_debug("==> cursor placed on key [%s], data [%s]", DKEY(key), DVAL(data)); return rc; } /* Move the cursor to the first item in the database. */ static int mdbx_cursor_first(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data) { int rc; if (mc->mc_xcursor) mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { rc = mdbx_page_search(mc, NULL, MDBX_PS_FIRST); if (unlikely(rc != MDBX_SUCCESS)) return rc; } if (unlikely(!IS_LEAF(mc->mc_pg[mc->mc_top]))) return MDBX_CORRUPTED; mc->mc_flags |= C_INITIALIZED; mc->mc_flags &= ~C_EOF; mc->mc_ki[mc->mc_top] = 0; if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { if (likely(key)) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mc->mc_pg[mc->mc_top], 0, key->iov_len); } return MDBX_SUCCESS; } MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], 0); if (F_ISSET(node_flags(node), F_DUPDATA)) { rc = mdbx_xcursor_init1(mc, node, mc->mc_pg[mc->mc_top]); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); if (unlikely(rc)) return rc; } else if (likely(data)) { if (unlikely((rc = mdbx_node_read( mc, node, data, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn))) != MDBX_SUCCESS)) return rc; } get_key_optional(node, key); return MDBX_SUCCESS; } /* Move the cursor to the last item in the database. */ static int mdbx_cursor_last(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data) { int rc; if (mc->mc_xcursor) mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { rc = mdbx_page_search(mc, NULL, MDBX_PS_LAST); if (unlikely(rc != MDBX_SUCCESS)) return rc; } if (unlikely(!IS_LEAF(mc->mc_pg[mc->mc_top]))) return MDBX_CORRUPTED; mc->mc_ki[mc->mc_top] = (indx_t)page_numkeys(mc->mc_pg[mc->mc_top]) - 1; mc->mc_flags |= C_INITIALIZED | C_EOF; if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { if (likely(key)) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->iov_len); } return MDBX_SUCCESS; } MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { rc = mdbx_xcursor_init1(mc, node, mc->mc_pg[mc->mc_top]); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); if (unlikely(rc)) return rc; } else if (likely(data)) { if (unlikely((rc = mdbx_node_read( mc, node, data, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn))) != MDBX_SUCCESS)) return rc; } get_key_optional(node, key); return MDBX_SUCCESS; } int mdbx_cursor_get(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, MDBX_cursor_op op) { if (unlikely(mc == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn(mc->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; int exact = 0; int (*mfunc)(MDBX_cursor * mc, MDBX_val * key, MDBX_val * data); switch (op) { case MDBX_GET_CURRENT: { if (unlikely(!(mc->mc_flags & C_INITIALIZED))) return MDBX_ENODATA; MDBX_page *mp = mc->mc_pg[mc->mc_top]; const unsigned nkeys = page_numkeys(mp); if (mc->mc_ki[mc->mc_top] >= nkeys) { mdbx_cassert(mc, nkeys <= UINT16_MAX); mc->mc_ki[mc->mc_top] = (uint16_t)nkeys; mc->mc_flags |= C_EOF; return MDBX_NOTFOUND; } mdbx_cassert(mc, nkeys > 0); rc = MDBX_SUCCESS; if (IS_LEAF2(mp)) { key->iov_len = mc->mc_db->md_xsize; key->iov_base = page_leaf2key(mp, mc->mc_ki[mc->mc_top], key->iov_len); } else { MDBX_node *node = page_node(mp, mc->mc_ki[mc->mc_top]); get_key_optional(node, key); if (data) { if (F_ISSET(node_flags(node), F_DUPDATA)) { if (unlikely(!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))) { rc = mdbx_xcursor_init1(mc, node, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); if (unlikely(rc)) return rc; } else { rc = mdbx_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDBX_GET_CURRENT); if (unlikely(rc)) return rc; } } else { rc = mdbx_node_read(mc, node, data, pp_txnid4chk(mp, mc->mc_txn)); if (unlikely(rc)) return rc; } } } break; } case MDBX_GET_BOTH: case MDBX_GET_BOTH_RANGE: if (unlikely(data == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_xcursor == NULL)) return MDBX_INCOMPATIBLE; /* fall through */ __fallthrough; case MDBX_SET: case MDBX_SET_KEY: case MDBX_SET_RANGE: if (unlikely(key == NULL)) return MDBX_EINVAL; rc = mdbx_cursor_set(mc, key, data, op, &exact); if (mc->mc_flags & C_INITIALIZED) { mdbx_cassert(mc, mc->mc_snum > 0 && mc->mc_top < mc->mc_snum); mdbx_cassert(mc, mc->mc_ki[mc->mc_top] < page_numkeys(mc->mc_pg[mc->mc_top]) || (mc->mc_flags & C_EOF)); } break; case MDBX_GET_MULTIPLE: if (unlikely(data == NULL || !(mc->mc_flags & C_INITIALIZED))) return MDBX_EINVAL; if (unlikely(!(mc->mc_db->md_flags & MDBX_DUPFIXED))) return MDBX_INCOMPATIBLE; rc = MDBX_SUCCESS; if ((mc->mc_xcursor->mx_cursor.mc_flags & (C_INITIALIZED | C_EOF)) != C_INITIALIZED) break; goto fetchm; case MDBX_NEXT_MULTIPLE: if (unlikely(data == NULL)) return MDBX_EINVAL; if (unlikely(!(mc->mc_db->md_flags & MDBX_DUPFIXED))) return MDBX_INCOMPATIBLE; rc = mdbx_cursor_next(mc, key, data, MDBX_NEXT_DUP); if (rc == MDBX_SUCCESS) { if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { MDBX_cursor *mx; fetchm: mx = &mc->mc_xcursor->mx_cursor; data->iov_len = page_numkeys(mx->mc_pg[mx->mc_top]) * mx->mc_db->md_xsize; data->iov_base = page_data(mx->mc_pg[mx->mc_top]); mx->mc_ki[mx->mc_top] = (indx_t)page_numkeys(mx->mc_pg[mx->mc_top]) - 1; } else { rc = MDBX_NOTFOUND; } } break; case MDBX_PREV_MULTIPLE: if (data == NULL) return MDBX_EINVAL; if (!(mc->mc_db->md_flags & MDBX_DUPFIXED)) return MDBX_INCOMPATIBLE; rc = MDBX_SUCCESS; if (!(mc->mc_flags & C_INITIALIZED)) rc = mdbx_cursor_last(mc, key, data); if (rc == MDBX_SUCCESS) { MDBX_cursor *mx = &mc->mc_xcursor->mx_cursor; if (mx->mc_flags & C_INITIALIZED) { rc = mdbx_cursor_sibling(mx, SIBLING_LEFT); if (rc == MDBX_SUCCESS) goto fetchm; } else { rc = MDBX_NOTFOUND; } } break; case MDBX_NEXT: case MDBX_NEXT_DUP: case MDBX_NEXT_NODUP: rc = mdbx_cursor_next(mc, key, data, op); break; case MDBX_PREV: case MDBX_PREV_DUP: case MDBX_PREV_NODUP: rc = mdbx_cursor_prev(mc, key, data, op); break; case MDBX_FIRST: rc = mdbx_cursor_first(mc, key, data); break; case MDBX_FIRST_DUP: mfunc = mdbx_cursor_first; mmove: if (unlikely(data == NULL || !(mc->mc_flags & C_INITIALIZED))) return MDBX_EINVAL; if (unlikely(mc->mc_xcursor == NULL)) return MDBX_INCOMPATIBLE; if (mc->mc_ki[mc->mc_top] >= page_numkeys(mc->mc_pg[mc->mc_top])) { mc->mc_ki[mc->mc_top] = (indx_t)page_numkeys(mc->mc_pg[mc->mc_top]); mc->mc_flags |= C_EOF; return MDBX_NOTFOUND; } { MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (!F_ISSET(node_flags(node), F_DUPDATA)) { get_key_optional(node, key); rc = mdbx_node_read(mc, node, data, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn)); break; } } if (unlikely(!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))) return MDBX_EINVAL; rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL); break; case MDBX_LAST: rc = mdbx_cursor_last(mc, key, data); break; case MDBX_LAST_DUP: mfunc = mdbx_cursor_last; goto mmove; case MDBX_SET_LOWERBOUND: { if (unlikely(key == NULL || data == NULL)) return MDBX_EINVAL; MDBX_val save_data = *data; rc = mdbx_cursor_set(mc, key, data, MDBX_SET_RANGE, &exact); if (rc == MDBX_SUCCESS && exact && mc->mc_xcursor) { mc->mc_flags &= ~C_DEL; if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { *data = save_data; exact = 0; rc = mdbx_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDBX_SET_RANGE, &exact); if (rc == MDBX_NOTFOUND) { mdbx_cassert(mc, !exact); rc = mdbx_cursor_next(mc, key, data, MDBX_NEXT_NODUP); } } else { int cmp = mc->mc_dbx->md_dcmp(&save_data, data); exact = (cmp == 0); if (cmp > 0) rc = mdbx_cursor_next(mc, key, data, MDBX_NEXT_NODUP); } } if (rc == MDBX_SUCCESS && !exact) rc = MDBX_RESULT_TRUE; break; } default: mdbx_debug("unhandled/unimplemented cursor operation %u", op); return MDBX_EINVAL; } mc->mc_flags &= ~C_DEL; return rc; } /* Touch all the pages in the cursor stack. Set mc_top. * Makes sure all the pages are writable, before attempting a write operation. * [in] mc The cursor to operate on. */ static int mdbx_cursor_touch(MDBX_cursor *mc) { int rc = MDBX_SUCCESS; if (mc->mc_dbi >= CORE_DBS && (*mc->mc_dbistate & (DBI_DIRTY | DBI_DUPDATA)) == 0) { mdbx_cassert(mc, (mc->mc_flags & C_RECLAIMING) == 0); /* Touch DB record of named DB */ MDBX_cursor_couple cx; if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi)) return MDBX_BAD_DBI; rc = mdbx_cursor_init(&cx.outer, mc->mc_txn, MAIN_DBI); if (unlikely(rc != MDBX_SUCCESS)) return rc; *mc->mc_dbistate |= DBI_DIRTY; mc->mc_txn->mt_flags |= MDBX_TXN_DIRTY; rc = mdbx_page_search(&cx.outer, &mc->mc_dbx->md_name, MDBX_PS_MODIFY); if (unlikely(rc)) return rc; } mc->mc_top = 0; if (mc->mc_snum) { do { rc = mdbx_page_touch(mc); } while (!rc && ++(mc->mc_top) < mc->mc_snum); mc->mc_top = mc->mc_snum - 1; } return rc; } int mdbx_cursor_put(MDBX_cursor *mc, const MDBX_val *key, MDBX_val *data, unsigned flags) { MDBX_env *env; MDBX_page *sub_root = NULL; MDBX_val xdata, *rdata, dkey, olddata; MDBX_db nested_dupdb; int rc2; DKBUF; if (unlikely(mc == NULL || key == NULL || data == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn_rw(mc->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))) return MDBX_BAD_DBI; mdbx_cassert(mc, cursor_is_tracked(mc)); env = mc->mc_txn->mt_env; /* Check this first so counter will always be zero on any early failures. */ size_t mcount = 0, dcount = 0; if (unlikely(flags & MDBX_MULTIPLE)) { if (unlikely(flags & MDBX_RESERVE)) return MDBX_EINVAL; if (unlikely(!F_ISSET(mc->mc_db->md_flags, MDBX_DUPFIXED))) return MDBX_INCOMPATIBLE; dcount = data[1].iov_len; if (unlikely(dcount < 2 || data->iov_len == 0)) return MDBX_BAD_VALSIZE; if (unlikely(mc->mc_db->md_xsize != data->iov_len) && mc->mc_db->md_xsize) return MDBX_BAD_VALSIZE; if (unlikely(dcount > MAX_MAPSIZE / 2 / (BRANCH_NODEMAX(MAX_PAGESIZE) - NODESIZE))) { /* checking for multiplication overflow */ if (unlikely(dcount > MAX_MAPSIZE / 2 / data->iov_len)) return MDBX_TOO_LARGE; } data[1].iov_len = 0 /* reset done item counter */; } if (flags & MDBX_RESERVE) { if (unlikely(mc->mc_db->md_flags & (MDBX_DUPSORT | MDBX_REVERSEDUP | MDBX_INTEGERDUP | MDBX_DUPFIXED))) return MDBX_INCOMPATIBLE; data->iov_base = nullptr; } const unsigned nospill = flags & MDBX_NOSPILL; flags -= nospill; if (unlikely(mc->mc_txn->mt_flags & (MDBX_TXN_RDONLY | MDBX_TXN_BLOCKED))) return (mc->mc_txn->mt_flags & MDBX_TXN_RDONLY) ? MDBX_EACCESS : MDBX_BAD_TXN; uint64_t aligned_keybytes, aligned_databytes; MDBX_val aligned_key, aligned_data; if (likely((mc->mc_flags & C_SUB) == 0)) { if (unlikely(key->iov_len < mc->mc_dbx->md_klen_min || key->iov_len > mc->mc_dbx->md_klen_max)) { mdbx_cassert(mc, !"Invalid key-size"); return MDBX_BAD_VALSIZE; } if (unlikely(data->iov_len < mc->mc_dbx->md_vlen_min || data->iov_len > mc->mc_dbx->md_vlen_max)) { mdbx_cassert(mc, !"Invalid data-size"); return MDBX_BAD_VALSIZE; } if (mc->mc_db->md_flags & MDBX_INTEGERKEY) { switch (key->iov_len) { default: mdbx_cassert(mc, !"key-size is invalid for MDBX_INTEGERKEY"); return MDBX_BAD_VALSIZE; case 4: if (unlikely(3 & (uintptr_t)key->iov_base)) { /* copy instead of return error to avoid break compatibility */ aligned_key.iov_base = memcpy(&aligned_keybytes, key->iov_base, aligned_key.iov_len = 4); key = &aligned_key; } break; case 8: if (unlikely(7 & (uintptr_t)key->iov_base)) { /* copy instead of return error to avoid break compatibility */ aligned_key.iov_base = memcpy(&aligned_keybytes, key->iov_base, aligned_key.iov_len = 8); key = &aligned_key; } break; } } if (mc->mc_db->md_flags & MDBX_INTEGERDUP) { switch (data->iov_len) { default: mdbx_cassert(mc, !"data-size is invalid for MDBX_INTEGERKEY"); return MDBX_BAD_VALSIZE; case 4: if (unlikely(3 & (uintptr_t)data->iov_base)) { if (unlikely(flags & MDBX_MULTIPLE)) return MDBX_BAD_VALSIZE; /* copy instead of return error to avoid break compatibility */ aligned_data.iov_base = memcpy(&aligned_databytes, data->iov_base, aligned_data.iov_len = 4); data = &aligned_data; } break; case 8: if (unlikely(7 & (uintptr_t)data->iov_base)) { if (unlikely(flags & MDBX_MULTIPLE)) return MDBX_BAD_VALSIZE; /* copy instead of return error to avoid break compatibility */ aligned_data.iov_base = memcpy(&aligned_databytes, data->iov_base, aligned_data.iov_len = 8); data = &aligned_data; } break; } } } mdbx_debug("==> put db %d key [%s], size %" PRIuPTR ", data [%s] size %" PRIuPTR, DDBI(mc), DKEY(key), key->iov_len, DVAL((flags & MDBX_RESERVE) ? nullptr : data), data->iov_len); int dupdata_flag = 0; if ((flags & MDBX_CURRENT) != 0 && (mc->mc_flags & C_SUB) == 0) { if (unlikely(flags & (MDBX_APPEND | MDBX_NOOVERWRITE))) return MDBX_EINVAL; /* Опция MDBX_CURRENT означает, что запрошено обновление текущей записи, * на которой сейчас стоит курсор. Проверяем что переданный ключ совпадает * со значением в текущей позиции курсора. * Здесь проще вызвать mdbx_cursor_get(), так как для обслуживания таблиц * с MDBX_DUPSORT также требуется текущий размер данных. */ MDBX_val current_key, current_data; rc = mdbx_cursor_get(mc, ¤t_key, ¤t_data, MDBX_GET_CURRENT); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (mc->mc_dbx->md_cmp(key, ¤t_key) != 0) return MDBX_EKEYMISMATCH; if (unlikely((flags & MDBX_MULTIPLE))) goto drop_current; if (F_ISSET(mc->mc_db->md_flags, MDBX_DUPSORT)) { MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { mdbx_cassert(mc, mc->mc_xcursor != NULL && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)); /* Если за ключом более одного значения, либо если размер данных * отличается, то вместо обновления требуется удаление и * последующая вставка. */ if (mc->mc_xcursor->mx_db.md_entries > 1 || current_data.iov_len != data->iov_len) { drop_current: rc = mdbx_cursor_del(mc, flags & MDBX_ALLDUPS); if (unlikely(rc != MDBX_SUCCESS)) return rc; flags -= MDBX_CURRENT; goto skip_check_samedata; } } else if (unlikely(node_size(key, data) > /* See note inside leaf_size() */ env->me_branch_nodemax)) { rc = mdbx_cursor_del(mc, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; flags -= MDBX_CURRENT; goto skip_check_samedata; } } if (!(flags & MDBX_RESERVE) && unlikely(cmp_lenfast(¤t_data, data) == 0)) return MDBX_SUCCESS /* the same data, nothing to update */; skip_check_samedata:; } if (mc->mc_db->md_root == P_INVALID) { /* new database, cursor has nothing to point to */ mc->mc_snum = 0; mc->mc_top = 0; mc->mc_flags &= ~C_INITIALIZED; rc = MDBX_NO_ROOT; } else if ((flags & MDBX_CURRENT) == 0) { int exact = false; if ((flags & MDBX_APPEND) && mc->mc_db->md_entries > 0) { rc = mdbx_cursor_last(mc, &dkey, &olddata); if (likely(rc == MDBX_SUCCESS)) { rc = mc->mc_dbx->md_cmp(key, &dkey); if (likely(rc > 0)) { mc->mc_ki[mc->mc_top]++; /* step forward for appending */ rc = MDBX_NOTFOUND; } else { if (unlikely(rc != MDBX_SUCCESS || !(flags & MDBX_APPENDDUP))) /* new-key < last-key * or new-key == last-key without MDBX_APPENDDUP */ return MDBX_EKEYMISMATCH; exact = true; } } } else { rc = mdbx_cursor_set(mc, (MDBX_val *)key, &olddata, MDBX_SET, &exact); } if (likely(rc == MDBX_SUCCESS)) { if (exact) { if (unlikely(flags & MDBX_NOOVERWRITE)) { mdbx_debug("duplicate key [%s]", DKEY(key)); *data = olddata; return MDBX_KEYEXIST; } if (unlikely(mc->mc_flags & C_SUB)) { /* nested subtree of DUPSORT-database with the same key, * nothing to update */ mdbx_assert(env, data->iov_len == 0 && olddata.iov_len == 0); return MDBX_SUCCESS; } if (unlikely(flags & MDBX_ALLDUPS) && mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) { rc = mdbx_cursor_del(mc, MDBX_ALLDUPS); if (unlikely(rc != MDBX_SUCCESS)) return rc; flags -= MDBX_ALLDUPS; rc = MDBX_NOTFOUND; exact = false; } else /* checking for early exit without dirtying pages */ if (!(flags & (MDBX_RESERVE | MDBX_MULTIPLE)) && unlikely(mc->mc_dbx->md_dcmp(data, &olddata) == 0)) { if (!mc->mc_xcursor) /* the same data, nothing to update */ return MDBX_SUCCESS; if (flags & MDBX_NODUPDATA) return MDBX_KEYEXIST; if (flags & MDBX_APPENDDUP) return MDBX_EKEYMISMATCH; if (likely(unsure_equal(mc->mc_dbx->md_dcmp, data, &olddata))) /* data is match exactly byte-to-byte, nothing to update */ return MDBX_SUCCESS; else { /* The data has differences, but the user-provided comparator * considers them equal. So continue update since called without. * Continue to update since was called without MDBX_NODUPDATA. */ } } } } else if (unlikely(rc != MDBX_NOTFOUND)) return rc; } mc->mc_flags &= ~C_DEL; /* Cursor is positioned, check for room in the dirty list */ if (!nospill) { if (unlikely(flags & MDBX_MULTIPLE)) { rdata = &xdata; xdata.iov_len = data->iov_len * dcount; } else { rdata = data; } if (unlikely(rc2 = mdbx_cursor_spill(mc, key, rdata))) return rc2; } if (unlikely(rc == MDBX_NO_ROOT)) { MDBX_page *np; /* new database, write a root leaf page */ mdbx_debug("%s", "allocating new root leaf page"); if (unlikely(rc2 = mdbx_page_new(mc, P_LEAF, 1, &np))) { return rc2; } rc2 = mdbx_cursor_push(mc, np); if (unlikely(rc2 != MDBX_SUCCESS)) return rc2; mc->mc_db->md_root = np->mp_pgno; mc->mc_db->md_depth++; if (mc->mc_db->md_flags & MDBX_INTEGERKEY) { assert(key->iov_len >= mc->mc_dbx->md_klen_min && key->iov_len <= mc->mc_dbx->md_klen_max); mc->mc_dbx->md_klen_min = mc->mc_dbx->md_klen_max = key->iov_len; } if (mc->mc_db->md_flags & (MDBX_INTEGERDUP | MDBX_DUPFIXED)) { assert(data->iov_len >= mc->mc_dbx->md_vlen_min && data->iov_len <= mc->mc_dbx->md_vlen_max); assert(mc->mc_xcursor != NULL); mc->mc_db->md_xsize = mc->mc_xcursor->mx_db.md_xsize = (unsigned)(mc->mc_dbx->md_vlen_min = mc->mc_dbx->md_vlen_max = mc->mc_xcursor->mx_dbx.md_klen_min = mc->mc_xcursor->mx_dbx.md_klen_max = data->iov_len); } if ((mc->mc_db->md_flags & (MDBX_DUPSORT | MDBX_DUPFIXED)) == MDBX_DUPFIXED) np->mp_flags |= P_LEAF2; mc->mc_flags |= C_INITIALIZED; } else { /* make sure all cursor pages are writable */ rc2 = mdbx_cursor_touch(mc); if (unlikely(rc2)) return rc2; } bool insert_key, insert_data, do_sub = false; insert_key = insert_data = (rc != MDBX_SUCCESS); uint16_t fp_flags = P_LEAF | P_DIRTY; MDBX_page *fp = env->me_pbuf; fp->mp_txnid = pp_txnid2chk(mc->mc_txn); if (insert_key) { /* The key does not exist */ mdbx_debug("inserting key at index %i", mc->mc_ki[mc->mc_top]); if ((mc->mc_db->md_flags & MDBX_DUPSORT) && node_size(key, data) > /* See note inside leaf_size() */ env->me_branch_nodemax) { /* Too big for a node, insert in sub-DB. Set up an empty * "old sub-page" for prep_subDB to expand to a full page. */ fp->mp_leaf2_ksize = (mc->mc_db->md_flags & MDBX_DUPFIXED) ? (uint16_t)data->iov_len : 0; fp->mp_lower = fp->mp_upper = 0; olddata.iov_len = PAGEHDRSZ; goto prep_subDB; } } else { /* there's only a key anyway, so this is a no-op */ if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { char *ptr; unsigned ksize = mc->mc_db->md_xsize; if (unlikely(key->iov_len != ksize)) return MDBX_BAD_VALSIZE; ptr = page_leaf2key(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize); memcpy(ptr, key->iov_base, ksize); fix_parent: /* if overwriting slot 0 of leaf, need to * update branch key if there is a parent page */ if (mc->mc_top && !mc->mc_ki[mc->mc_top]) { unsigned dtop = 1; mc->mc_top--; /* slot 0 is always an empty key, find real slot */ while (mc->mc_top && !mc->mc_ki[mc->mc_top]) { mc->mc_top--; dtop++; } if (mc->mc_ki[mc->mc_top]) rc2 = mdbx_update_key(mc, key); else rc2 = MDBX_SUCCESS; mdbx_cassert(mc, mc->mc_top + dtop < UINT16_MAX); mc->mc_top += (uint16_t)dtop; if (rc2) return rc2; } if (mdbx_audit_enabled()) { int err = mdbx_cursor_check(mc, 0); if (unlikely(err != MDBX_SUCCESS)) return err; } return MDBX_SUCCESS; } more:; if (mdbx_audit_enabled()) { int err = mdbx_cursor_check(mc, 0); if (unlikely(err != MDBX_SUCCESS)) return err; } MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); /* overflow page overwrites need special handling */ if (unlikely(F_ISSET(node_flags(node), F_BIGDATA))) { int level, ovpages, dpages = (node_size(key, data) > /* See note inside leaf_size() */ env->me_branch_nodemax) ? number_of_ovpages(env, data->iov_len) : 0; const pgno_t pg = node_largedata_pgno(node); MDBX_page *omp; if (unlikely((rc2 = mdbx_page_get( mc, pg, &omp, &level, pp_txnid4chk(mc->mc_pg[mc->mc_top], mc->mc_txn))) != 0)) return rc2; ovpages = omp->mp_pages; /* Is the ov page large enough? */ if (unlikely(mc->mc_flags & C_GCFREEZE) ? ovpages >= dpages : ovpages == /* LY: add configurable threshold to keep reserve space */ dpages) { if (!IS_DIRTY(omp) && (level || (env->me_flags & MDBX_WRITEMAP))) { rc2 = mdbx_page_unspill(mc->mc_txn, omp, &omp); if (unlikely(rc2)) return rc2; level = 0; /* dirty in this txn or clean */ } /* Is it dirty? */ if (IS_DIRTY(omp)) { mdbx_cassert(mc, omp->mp_txnid > SAFE64_INVALID_THRESHOLD); /* yes, overwrite it. Note in this case we don't * bother to try shrinking the page if the new data * is smaller than the overflow threshold. */ if (unlikely(level > 1)) { /* It is writable only in a parent txn */ MDBX_page *np = mdbx_page_malloc(mc->mc_txn, ovpages); if (unlikely(!np)) return MDBX_ENOMEM; /* Note - this page is already counted in parent's dirtyroom */ rc2 = mdbx_dpl_append(mc->mc_txn, pg, np); if (unlikely(rc2 != MDBX_SUCCESS)) { rc = rc2; mdbx_dpage_free(env, np, ovpages); goto fail; } /* Currently we make the page look as with put() in the * parent txn, in case the user peeks at MDBX_RESERVEd * or unused parts. Some users treat ovpages specially. */ const size_t whole = pgno2bytes(env, ovpages); /* Skip the part where MDBX will put *data. * Copy end of page, adjusting alignment so * compiler may copy words instead of bytes. */ const size_t off = (PAGEHDRSZ + data->iov_len) & -(intptr_t)sizeof(size_t); memcpy((size_t *)((char *)np + off), (size_t *)((char *)omp + off), whole - off); memcpy(np, omp, PAGEHDRSZ); /* Copy header of page */ omp = np; mdbx_cassert(mc, mdbx_dirtylist_check(mc->mc_txn)); } node_set_ds(node, data->iov_len); if (F_ISSET(flags, MDBX_RESERVE)) data->iov_base = page_data(omp); else memcpy(page_data(omp), data->iov_base, data->iov_len); if (mdbx_audit_enabled()) { int err = mdbx_cursor_check(mc, 0); if (unlikely(err != MDBX_SUCCESS)) return err; } return MDBX_SUCCESS; } } if ((rc2 = mdbx_page_retire(mc, omp)) != MDBX_SUCCESS) return rc2; } else { olddata.iov_len = node_ds(node); olddata.iov_base = node_data(node); mdbx_cassert(mc, (char *)olddata.iov_base + olddata.iov_len <= (char *)(mc->mc_pg[mc->mc_top]) + env->me_psize); /* DB has dups? */ if (F_ISSET(mc->mc_db->md_flags, MDBX_DUPSORT)) { /* Prepare (sub-)page/sub-DB to accept the new item, if needed. * fp: old sub-page or a header faking it. * mp: new (sub-)page. offset: growth in page size. * xdata: node data with new page or DB. */ unsigned i; size_t offset = 0; MDBX_page *mp = fp = xdata.iov_base = env->me_pbuf; mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno; /* Was a single item before, must convert now */ if (!F_ISSET(node_flags(node), F_DUPDATA)) { /* does data match? */ const int cmp = mc->mc_dbx->md_dcmp(data, &olddata); if ((flags & MDBX_APPENDDUP) && unlikely(cmp <= 0)) return MDBX_EKEYMISMATCH; if (cmp == 0) { if (flags & MDBX_NODUPDATA) return MDBX_KEYEXIST; if (likely(unsure_equal(mc->mc_dbx->md_dcmp, data, &olddata))) { /* data is match exactly byte-to-byte, nothing to update */ if (unlikely(flags & MDBX_MULTIPLE)) { rc = MDBX_SUCCESS; goto continue_multiple; } return MDBX_SUCCESS; } else { /* The data has differences, but the user-provided comparator * considers them equal. So continue update since called without. * Continue to update since was called without MDBX_NODUPDATA. */ } mdbx_cassert( mc, node_size(key, data) <= /* See note inside leaf_size() */ env->me_branch_nodemax); goto current; } /* Just overwrite the current item */ if (flags & MDBX_CURRENT) { mdbx_cassert( mc, node_size(key, data) <= /* See note inside leaf_size() */ env->me_branch_nodemax); goto current; } /* Back up original data item */ memcpy(dkey.iov_base = fp + 1, olddata.iov_base, dkey.iov_len = olddata.iov_len); dupdata_flag = 1; /* Make sub-page header for the dup items, with dummy body */ fp->mp_flags = P_LEAF | P_DIRTY | P_SUBP; fp->mp_lower = 0; xdata.iov_len = PAGEHDRSZ + dkey.iov_len + data->iov_len; if (mc->mc_db->md_flags & MDBX_DUPFIXED) { fp->mp_flags |= P_LEAF2; fp->mp_leaf2_ksize = (uint16_t)data->iov_len; xdata.iov_len += 2 * data->iov_len; /* leave space for 2 more */ mdbx_cassert(mc, xdata.iov_len <= env->me_psize); } else { xdata.iov_len += 2 * (sizeof(indx_t) + NODESIZE) + (dkey.iov_len & 1) + (data->iov_len & 1); mdbx_cassert(mc, xdata.iov_len <= env->me_psize); } fp->mp_upper = (uint16_t)(xdata.iov_len - PAGEHDRSZ); olddata.iov_len = xdata.iov_len; /* pretend olddata is fp */ } else if (node_flags(node) & F_SUBDATA) { /* Data is on sub-DB, just store it */ flags |= F_DUPDATA | F_SUBDATA; goto put_sub; } else { /* Data is on sub-page */ fp = olddata.iov_base; switch (flags) { default: if (!(mc->mc_db->md_flags & MDBX_DUPFIXED)) { offset = node_size(data, nullptr) + sizeof(indx_t); break; } offset = fp->mp_leaf2_ksize; if (page_room(fp) < offset) { offset *= 4; /* space for 4 more */ break; } /* FALLTHRU: Big enough MDBX_DUPFIXED sub-page */ __fallthrough; case MDBX_CURRENT | MDBX_NODUPDATA: case MDBX_CURRENT: fp->mp_flags |= P_DIRTY; fp->mp_txnid = INVALID_TXNID; fp->mp_pgno = mp->mp_pgno; mc->mc_xcursor->mx_cursor.mc_pg[0] = fp; flags |= F_DUPDATA; goto put_sub; } xdata.iov_len = olddata.iov_len + offset; } fp_flags = fp->mp_flags; if (NODESIZE + node_ks(node) + xdata.iov_len > /* See note inside leaf_size() */ env->me_branch_nodemax) { /* Too big for a sub-page, convert to sub-DB */ fp_flags &= ~P_SUBP; prep_subDB: nested_dupdb.md_xsize = 0; nested_dupdb.md_flags = flags_db2sub(mc->mc_db->md_flags); if (mc->mc_db->md_flags & MDBX_DUPFIXED) { fp_flags |= P_LEAF2; nested_dupdb.md_xsize = fp->mp_leaf2_ksize; } nested_dupdb.md_depth = 1; nested_dupdb.md_branch_pages = 0; nested_dupdb.md_leaf_pages = 1; nested_dupdb.md_overflow_pages = 0; nested_dupdb.md_entries = page_numkeys(fp); xdata.iov_len = sizeof(nested_dupdb); xdata.iov_base = &nested_dupdb; if ((rc2 = mdbx_page_alloc(mc, 1, &mp, MDBX_ALLOC_ALL))) return rc2; mc->mc_db->md_leaf_pages += 1; mdbx_cassert(mc, env->me_psize > olddata.iov_len); offset = env->me_psize - (unsigned)olddata.iov_len; flags |= F_DUPDATA | F_SUBDATA; nested_dupdb.md_root = mp->mp_pgno; nested_dupdb.md_seq = 0; nested_dupdb.md_mod_txnid = pp_txnid2chk(mc->mc_txn); sub_root = mp; } if (mp != fp) { mp->mp_flags = fp_flags | P_DIRTY; mp->mp_txnid = INVALID_TXNID; mp->mp_leaf2_ksize = fp->mp_leaf2_ksize; mp->mp_lower = fp->mp_lower; mdbx_cassert(mc, fp->mp_upper + offset <= UINT16_MAX); mp->mp_upper = (indx_t)(fp->mp_upper + offset); if (unlikely(fp_flags & P_LEAF2)) { memcpy(page_data(mp), page_data(fp), page_numkeys(fp) * fp->mp_leaf2_ksize); } else { memcpy((char *)mp + mp->mp_upper + PAGEHDRSZ, (char *)fp + fp->mp_upper + PAGEHDRSZ, olddata.iov_len - fp->mp_upper - PAGEHDRSZ); memcpy((char *)(&mp->mp_ptrs), (char *)(&fp->mp_ptrs), page_numkeys(fp) * sizeof(mp->mp_ptrs[0])); for (i = 0; i < page_numkeys(fp); i++) { mdbx_cassert(mc, mp->mp_ptrs[i] + offset <= UINT16_MAX); mp->mp_ptrs[i] += (indx_t)offset; } } } rdata = &xdata; flags |= F_DUPDATA; do_sub = true; if (!insert_key) mdbx_node_del(mc, 0); goto new_sub; } /* MDBX passes F_SUBDATA in 'flags' to write a DB record */ if (unlikely((node_flags(node) ^ flags) & F_SUBDATA)) return MDBX_INCOMPATIBLE; current: if (data->iov_len == olddata.iov_len) { mdbx_cassert(mc, EVEN(key->iov_len) == EVEN(node_ks(node))); /* same size, just replace it. Note that we could * also reuse this node if the new data is smaller, * but instead we opt to shrink the node in that case. */ if (F_ISSET(flags, MDBX_RESERVE)) data->iov_base = olddata.iov_base; else if (!(mc->mc_flags & C_SUB)) memcpy(olddata.iov_base, data->iov_base, data->iov_len); else { mdbx_cassert(mc, page_numkeys(mc->mc_pg[mc->mc_top]) == 1); mdbx_cassert(mc, PAGETYPE(mc->mc_pg[mc->mc_top]) == P_LEAF); mdbx_cassert(mc, node_ds(node) == 0); mdbx_cassert(mc, node_flags(node) == 0); mdbx_cassert(mc, key->iov_len < UINT16_MAX); node_set_ks(node, key->iov_len); memcpy(node_key(node), key->iov_base, key->iov_len); mdbx_cassert(mc, (char *)node_key(node) + node_ds(node) < (char *)(mc->mc_pg[mc->mc_top]) + env->me_psize); goto fix_parent; } if (mdbx_audit_enabled()) { int err = mdbx_cursor_check(mc, 0); if (unlikely(err != MDBX_SUCCESS)) return err; } return MDBX_SUCCESS; } } mdbx_node_del(mc, 0); } rdata = data; new_sub:; unsigned nflags = flags & NODE_ADD_FLAGS; size_t nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->iov_len : leaf_size(env, key, rdata); if (page_room(mc->mc_pg[mc->mc_top]) < nsize) { if ((flags & (F_DUPDATA | F_SUBDATA)) == F_DUPDATA) nflags &= ~MDBX_APPEND; /* sub-page may need room to grow */ if (!insert_key) nflags |= MDBX_SPLIT_REPLACE; rc = mdbx_page_split(mc, key, rdata, P_INVALID, nflags); if (rc == MDBX_SUCCESS && mdbx_audit_enabled()) rc = mdbx_cursor_check(mc, 0); } else { /* There is room already in this leaf page. */ if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { mdbx_cassert(mc, (nflags & (F_BIGDATA | F_SUBDATA | F_DUPDATA)) == 0 && rdata->iov_len == 0); rc = mdbx_node_add_leaf2(mc, mc->mc_ki[mc->mc_top], key); } else rc = mdbx_node_add_leaf(mc, mc->mc_ki[mc->mc_top], key, rdata, nflags); if (likely(rc == 0)) { /* Adjust other cursors pointing to mp */ const MDBX_dbi dbi = mc->mc_dbi; const unsigned i = mc->mc_top; MDBX_page *const mp = mc->mc_pg[i]; for (MDBX_cursor *m2 = mc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { MDBX_cursor *m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue; if (m3->mc_ki[i] >= mc->mc_ki[i]) m3->mc_ki[i] += insert_key; if (XCURSOR_INITED(m3)) XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]); } } } if (likely(rc == MDBX_SUCCESS)) { /* Now store the actual data in the child DB. Note that we're * storing the user data in the keys field, so there are strict * size limits on dupdata. The actual data fields of the child * DB are all zero size. */ if (do_sub) { int xflags; size_t ecount; put_sub: xdata.iov_len = 0; xdata.iov_base = nullptr; MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); #define SHIFT_MDBX_NODUPDATA_TO_MDBX_NOOVERWRITE 1 STATIC_ASSERT( (MDBX_NODUPDATA >> SHIFT_MDBX_NODUPDATA_TO_MDBX_NOOVERWRITE) == MDBX_NOOVERWRITE); xflags = MDBX_CURRENT | MDBX_NOSPILL | ((flags & MDBX_NODUPDATA) >> SHIFT_MDBX_NODUPDATA_TO_MDBX_NOOVERWRITE); if ((flags & MDBX_CURRENT) == 0) { xflags -= MDBX_CURRENT; rc2 = mdbx_xcursor_init1(mc, node, mc->mc_pg[mc->mc_top]); if (unlikely(rc2 != MDBX_SUCCESS)) return rc2; } if (sub_root) mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root; /* converted, write the original data first */ if (dupdata_flag) { rc = mdbx_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags); if (unlikely(rc)) goto bad_sub; /* we've done our job */ dkey.iov_len = 0; } if (!(node_flags(node) & F_SUBDATA) || sub_root) { /* Adjust other cursors pointing to mp */ MDBX_cursor *m2; MDBX_xcursor *mx = mc->mc_xcursor; unsigned i = mc->mc_top; MDBX_page *mp = mc->mc_pg[i]; const int nkeys = page_numkeys(mp); for (m2 = mc->mc_txn->tw.cursors[mc->mc_dbi]; m2; m2 = m2->mc_next) { if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; if (!(m2->mc_flags & C_INITIALIZED)) continue; if (m2->mc_pg[i] == mp) { if (m2->mc_ki[i] == mc->mc_ki[i]) { rc2 = mdbx_xcursor_init2(m2, mx, dupdata_flag); if (unlikely(rc2 != MDBX_SUCCESS)) return rc2; } else if (!insert_key && m2->mc_ki[i] < nkeys) { XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]); } } } } mdbx_cassert(mc, mc->mc_xcursor->mx_db.md_entries < PTRDIFF_MAX); ecount = (size_t)mc->mc_xcursor->mx_db.md_entries; #define SHIFT_MDBX_APPENDDUP_TO_MDBX_APPEND 1 STATIC_ASSERT((MDBX_APPENDDUP >> SHIFT_MDBX_APPENDDUP_TO_MDBX_APPEND) == MDBX_APPEND); xflags |= (flags & MDBX_APPENDDUP) >> SHIFT_MDBX_APPENDDUP_TO_MDBX_APPEND; rc = mdbx_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags); if (flags & F_SUBDATA) { void *db = node_data(node); mc->mc_xcursor->mx_db.md_mod_txnid = pp_txnid2chk(mc->mc_txn); memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDBX_db)); } insert_data = (ecount != (size_t)mc->mc_xcursor->mx_db.md_entries); } /* Increment count unless we just replaced an existing item. */ if (insert_data) mc->mc_db->md_entries++; if (insert_key) { /* Invalidate txn if we created an empty sub-DB */ if (unlikely(rc)) goto bad_sub; /* If we succeeded and the key didn't exist before, * make sure the cursor is marked valid. */ mc->mc_flags |= C_INITIALIZED; } if (flags & MDBX_MULTIPLE) { if (!rc) { continue_multiple: mcount++; /* let caller know how many succeeded, if any */ data[1].iov_len = mcount; if (mcount < dcount) { data[0].iov_base = (char *)data[0].iov_base + data[0].iov_len; insert_key = insert_data = false; goto more; } } } if (rc == MDBX_SUCCESS && mdbx_audit_enabled()) rc = mdbx_cursor_check(mc, 0); return rc; bad_sub: if (unlikely(rc == MDBX_KEYEXIST)) mdbx_error("unexpected %s", "MDBX_KEYEXIST"); /* should not happen, we deleted that item */ rc = MDBX_PROBLEM; } fail: mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return rc; } int mdbx_cursor_del(MDBX_cursor *mc, MDBX_put_flags_t flags) { if (unlikely(!mc)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn_rw(mc->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))) return MDBX_BAD_DBI; if (unlikely(!(mc->mc_flags & C_INITIALIZED))) return MDBX_ENODATA; if (unlikely(mc->mc_ki[mc->mc_top] >= page_numkeys(mc->mc_pg[mc->mc_top]))) return MDBX_NOTFOUND; if (unlikely(!(flags & MDBX_NOSPILL) && (rc = mdbx_cursor_spill(mc, NULL, NULL)))) return rc; rc = mdbx_cursor_touch(mc); if (unlikely(rc != MDBX_SUCCESS)) return rc; MDBX_page *mp = mc->mc_pg[mc->mc_top]; if (unlikely(!IS_LEAF(mp))) return MDBX_CORRUPTED; if (IS_LEAF2(mp)) goto del_key; MDBX_node *node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { if (flags & (MDBX_ALLDUPS | /* for compatibility */ MDBX_NODUPDATA)) { /* mdbx_cursor_del0() will subtract the final entry */ mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1; mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; } else { if (!F_ISSET(node_flags(node), F_SUBDATA)) mc->mc_xcursor->mx_cursor.mc_pg[0] = node_data(node); rc = mdbx_cursor_del(&mc->mc_xcursor->mx_cursor, MDBX_NOSPILL); if (unlikely(rc)) return rc; /* If sub-DB still has entries, we're done */ if (mc->mc_xcursor->mx_db.md_entries) { if (node_flags(node) & F_SUBDATA) { /* update subDB info */ void *db = node_data(node); mc->mc_xcursor->mx_db.md_mod_txnid = pp_txnid2chk(mc->mc_txn); memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDBX_db)); } else { MDBX_cursor *m2; /* shrink fake page */ mdbx_node_shrink(mp, mc->mc_ki[mc->mc_top]); node = page_node(mp, mc->mc_ki[mc->mc_top]); mc->mc_xcursor->mx_cursor.mc_pg[0] = node_data(node); /* fix other sub-DB cursors pointed at fake pages on this page */ for (m2 = mc->mc_txn->tw.cursors[mc->mc_dbi]; m2; m2 = m2->mc_next) { if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; if (!(m2->mc_flags & C_INITIALIZED)) continue; if (m2->mc_pg[mc->mc_top] == mp) { MDBX_node *inner = node; if (m2->mc_ki[mc->mc_top] >= page_numkeys(mp)) continue; if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) { inner = page_node(mp, m2->mc_ki[mc->mc_top]); if (node_flags(inner) & F_SUBDATA) continue; } m2->mc_xcursor->mx_cursor.mc_pg[0] = node_data(inner); } } } mc->mc_db->md_entries--; mdbx_cassert(mc, mc->mc_db->md_entries > 0 && mc->mc_db->md_depth > 0 && mc->mc_db->md_root != P_INVALID); return rc; } else { mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; } /* otherwise fall thru and delete the sub-DB */ } if (node_flags(node) & F_SUBDATA) { /* add all the child DB's pages to the free list */ rc = mdbx_drop0(&mc->mc_xcursor->mx_cursor, 0); if (unlikely(rc)) goto fail; } } /* MDBX passes F_SUBDATA in 'flags' to delete a DB record */ else if (unlikely((node_flags(node) ^ flags) & F_SUBDATA)) return MDBX_INCOMPATIBLE; /* add overflow pages to free list */ if (F_ISSET(node_flags(node), F_BIGDATA)) { MDBX_page *omp; if (unlikely((rc = mdbx_page_get(mc, node_largedata_pgno(node), &omp, NULL, pp_txnid4chk(mp, mc->mc_txn))) || (rc = mdbx_page_retire(mc, omp)))) goto fail; } del_key: return mdbx_cursor_del0(mc); fail: mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return rc; } /* Allocate and initialize new pages for a database. * Set MDBX_TXN_ERROR on failure. * * [in] mc a cursor on the database being added to. * [in] flags flags defining what type of page is being allocated. * [in] num the number of pages to allocate. This is usually 1, * unless allocating overflow pages for a large record. * [out] mp Address of a page, or NULL on failure. * * Returns 0 on success, non-zero on failure. */ static int mdbx_page_new(MDBX_cursor *mc, unsigned flags, unsigned num, MDBX_page **mp) { MDBX_page *np; int rc; if (unlikely((rc = mdbx_page_alloc(mc, num, &np, MDBX_ALLOC_ALL)))) return rc; *mp = np; mdbx_debug("db %u allocated new page %" PRIaPGNO ", num %u", mc->mc_dbi, np->mp_pgno, num); np->mp_flags = (uint16_t)(flags | P_DIRTY); np->mp_txnid = INVALID_TXNID; np->mp_lower = 0; np->mp_upper = (indx_t)(mc->mc_txn->mt_env->me_psize - PAGEHDRSZ); *mc->mc_dbistate |= DBI_DIRTY; mc->mc_txn->mt_flags |= MDBX_TXN_DIRTY; mc->mc_db->md_branch_pages += IS_BRANCH(np); mc->mc_db->md_leaf_pages += IS_LEAF(np); if (unlikely(IS_OVERFLOW(np))) { mc->mc_db->md_overflow_pages += num; np->mp_pages = num; mdbx_cassert(mc, !(mc->mc_flags & C_SUB)); } else if (unlikely(mc->mc_flags & C_SUB)) { MDBX_db *outer = mdbx_outer_db(mc); outer->md_branch_pages += IS_BRANCH(np); outer->md_leaf_pages += IS_LEAF(np); } return MDBX_SUCCESS; } static int __must_check_result mdbx_node_add_leaf2(MDBX_cursor *mc, unsigned indx, const MDBX_val *key) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; DKBUF; mdbx_debug("add to leaf2-%spage %" PRIaPGNO " index %i, " " key size %" PRIuPTR " [%s]", IS_SUBP(mp) ? "sub-" : "", mp->mp_pgno, indx, key ? key->iov_len : 0, DKEY(key)); mdbx_cassert(mc, key); mdbx_cassert(mc, PAGETYPE(mp) == (P_LEAF | P_LEAF2)); const unsigned ksize = mc->mc_db->md_xsize; mdbx_cassert(mc, ksize == key->iov_len); const unsigned nkeys = page_numkeys(mp); /* Just using these for counting */ const intptr_t lower = mp->mp_lower + sizeof(indx_t); const intptr_t upper = mp->mp_upper - (ksize - sizeof(indx_t)); if (unlikely(lower > upper)) { mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PAGE_FULL; } mp->mp_lower = (indx_t)lower; mp->mp_upper = (indx_t)upper; char *const ptr = page_leaf2key(mp, indx, ksize); mdbx_cassert(mc, nkeys >= indx); const unsigned diff = nkeys - indx; if (likely(diff > 0)) /* Move higher keys up one slot. */ memmove(ptr + ksize, ptr, diff * ksize); /* insert new key */ memcpy(ptr, key->iov_base, ksize); return MDBX_SUCCESS; } static int __must_check_result mdbx_node_add_branch(MDBX_cursor *mc, unsigned indx, const MDBX_val *key, pgno_t pgno) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; DKBUF; mdbx_debug("add to branch-%spage %" PRIaPGNO " index %i, node-pgno %" PRIaPGNO " key size %" PRIuPTR " [%s]", IS_SUBP(mp) ? "sub-" : "", mp->mp_pgno, indx, pgno, key ? key->iov_len : 0, DKEY(key)); mdbx_cassert(mc, PAGETYPE(mp) == P_BRANCH); STATIC_ASSERT(NODESIZE % 2 == 0); /* Move higher pointers up one slot. */ const unsigned nkeys = page_numkeys(mp); mdbx_cassert(mc, nkeys >= indx); for (unsigned i = nkeys; i > indx; --i) mp->mp_ptrs[i] = mp->mp_ptrs[i - 1]; /* Adjust free space offsets. */ const size_t branch_bytes = branch_size(mc->mc_txn->mt_env, key); const intptr_t lower = mp->mp_lower + sizeof(indx_t); const intptr_t upper = mp->mp_upper - (branch_bytes - sizeof(indx_t)); if (unlikely(lower > upper)) { mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PAGE_FULL; } mp->mp_lower = (indx_t)lower; mp->mp_ptrs[indx] = mp->mp_upper = (indx_t)upper; /* Write the node data. */ MDBX_node *node = page_node(mp, indx); node_set_pgno(node, pgno); node_set_flags(node, 0); UNALIGNED_POKE_8(node, MDBX_node, mn_extra, 0); node_set_ks(node, 0); if (likely(key != NULL)) { node_set_ks(node, key->iov_len); memcpy(node_key(node), key->iov_base, key->iov_len); } return MDBX_SUCCESS; } static int __must_check_result mdbx_node_add_leaf(MDBX_cursor *mc, unsigned indx, const MDBX_val *key, MDBX_val *data, unsigned flags) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; DKBUF; mdbx_debug("add to leaf-%spage %" PRIaPGNO " index %i, data size %" PRIuPTR " key size %" PRIuPTR " [%s]", IS_SUBP(mp) ? "sub-" : "", mp->mp_pgno, indx, data ? data->iov_len : 0, key ? key->iov_len : 0, DKEY(key)); mdbx_cassert(mc, key != NULL && data != NULL); mdbx_cassert(mc, PAGETYPE(mp) == P_LEAF); MDBX_page *largepage = NULL; size_t leaf_bytes; if (unlikely(flags & F_BIGDATA)) { /* Data already on overflow page. */ STATIC_ASSERT(sizeof(pgno_t) % 2 == 0); leaf_bytes = node_size(key, nullptr) + sizeof(pgno_t) + sizeof(indx_t); } else if (unlikely(node_size(key, data) > /* See note inside leaf_size() */ mc->mc_txn->mt_env->me_branch_nodemax)) { /* Put data on overflow page. */ mdbx_ensure(mc->mc_txn->mt_env, !F_ISSET(mc->mc_db->md_flags, MDBX_DUPSORT)); if (unlikely(flags & (F_DUPDATA | F_SUBDATA))) return MDBX_PROBLEM; const pgno_t ovpages = number_of_ovpages(mc->mc_txn->mt_env, data->iov_len); int rc = mdbx_page_new(mc, P_OVERFLOW, ovpages, &largepage); if (unlikely(rc != MDBX_SUCCESS)) return rc; mdbx_debug("allocated %u overflow page(s) %" PRIaPGNO "for %" PRIuPTR " data bytes", largepage->mp_pages, largepage->mp_pgno, data->iov_len); flags |= F_BIGDATA; leaf_bytes = node_size(key, nullptr) + sizeof(pgno_t) + sizeof(indx_t); } else { leaf_bytes = node_size(key, data) + sizeof(indx_t); } mdbx_cassert(mc, leaf_bytes == leaf_size(mc->mc_txn->mt_env, key, data)); /* Move higher pointers up one slot. */ const unsigned nkeys = page_numkeys(mp); mdbx_cassert(mc, nkeys >= indx); for (unsigned i = nkeys; i > indx; --i) mp->mp_ptrs[i] = mp->mp_ptrs[i - 1]; /* Adjust free space offsets. */ const intptr_t lower = mp->mp_lower + sizeof(indx_t); const intptr_t upper = mp->mp_upper - (leaf_bytes - sizeof(indx_t)); if (unlikely(lower > upper)) { mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PAGE_FULL; } mp->mp_lower = (indx_t)lower; mp->mp_ptrs[indx] = mp->mp_upper = (indx_t)upper; /* Write the node data. */ MDBX_node *node = page_node(mp, indx); node_set_ks(node, key->iov_len); node_set_flags(node, (uint8_t)flags); UNALIGNED_POKE_8(node, MDBX_node, mn_extra, 0); node_set_ds(node, data->iov_len); memcpy(node_key(node), key->iov_base, key->iov_len); void *nodedata = node_data(node); if (likely(largepage == NULL)) { if (unlikely(flags & F_BIGDATA)) memcpy(nodedata, data->iov_base, sizeof(pgno_t)); else if (unlikely(flags & MDBX_RESERVE)) data->iov_base = nodedata; else if (likely(nodedata != data->iov_base && data->iov_len /* to avoid UBSAN traps*/ != 0)) memcpy(nodedata, data->iov_base, data->iov_len); } else { poke_pgno(nodedata, largepage->mp_pgno); nodedata = page_data(largepage); if (unlikely(flags & MDBX_RESERVE)) data->iov_base = nodedata; else if (likely(nodedata != data->iov_base && data->iov_len /* to avoid UBSAN traps*/ != 0)) memcpy(nodedata, data->iov_base, data->iov_len); } return MDBX_SUCCESS; } /* Delete the specified node from a page. * [in] mc Cursor pointing to the node to delete. * [in] ksize The size of a node. Only used if the page is * part of a MDBX_DUPFIXED database. */ static void mdbx_node_del(MDBX_cursor *mc, size_t ksize) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; int indx = mc->mc_ki[mc->mc_top]; int i, j, nkeys, ptr; MDBX_node *node; char *base; mdbx_debug("delete node %u on %s page %" PRIaPGNO, indx, IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno); nkeys = page_numkeys(mp); mdbx_cassert(mc, indx < nkeys); if (IS_LEAF2(mp)) { mdbx_cassert(mc, ksize >= sizeof(indx_t)); unsigned diff = nkeys - 1 - indx; base = page_leaf2key(mp, indx, ksize); if (diff) memmove(base, base + ksize, diff * ksize); mdbx_cassert(mc, mp->mp_lower >= sizeof(indx_t)); mp->mp_lower -= sizeof(indx_t); mdbx_cassert(mc, (size_t)UINT16_MAX - mp->mp_upper >= ksize - sizeof(indx_t)); mp->mp_upper += (indx_t)(ksize - sizeof(indx_t)); return; } node = page_node(mp, indx); size_t sz = NODESIZE + node_ks(node); if (IS_LEAF(mp)) { if (F_ISSET(node_flags(node), F_BIGDATA)) sz += sizeof(pgno_t); else sz += node_ds(node); } sz = EVEN(sz); ptr = mp->mp_ptrs[indx]; for (i = j = 0; i < nkeys; i++) { if (i != indx) { mp->mp_ptrs[j] = mp->mp_ptrs[i]; if (mp->mp_ptrs[i] < ptr) { mdbx_cassert(mc, (size_t)UINT16_MAX - mp->mp_ptrs[j] >= sz); mp->mp_ptrs[j] += (indx_t)sz; } j++; } } base = (char *)mp + mp->mp_upper + PAGEHDRSZ; memmove(base + sz, base, ptr - mp->mp_upper); mdbx_cassert(mc, mp->mp_lower >= sizeof(indx_t)); mp->mp_lower -= sizeof(indx_t); mdbx_cassert(mc, (size_t)UINT16_MAX - mp->mp_upper >= sz); mp->mp_upper += (indx_t)sz; #if MDBX_DEBUG > 0 if (mdbx_audit_enabled()) { int page_check_err = mdbx_page_check(mc, mp, C_UPDATING); mdbx_cassert(mc, page_check_err == MDBX_SUCCESS); } #endif } /* Compact the main page after deleting a node on a subpage. * [in] mp The main page to operate on. * [in] indx The index of the subpage on the main page. */ static void mdbx_node_shrink(MDBX_page *mp, unsigned indx) { MDBX_node *node; MDBX_page *sp, *xp; char *base; size_t nsize, delta, len, ptr; int i; node = page_node(mp, indx); sp = (MDBX_page *)node_data(node); delta = page_room(sp); assert(delta > 0); /* Prepare to shift upward, set len = length(subpage part to shift) */ if (IS_LEAF2(sp)) { delta &= /* do not make the node uneven-sized */ ~(size_t)1; if (unlikely(delta) == 0) return; nsize = node_ds(node) - delta; assert(nsize % 1 == 0); len = nsize; } else { xp = (MDBX_page *)((char *)sp + delta); /* destination subpage */ for (i = page_numkeys(sp); --i >= 0;) { assert(sp->mp_ptrs[i] >= delta); xp->mp_ptrs[i] = (indx_t)(sp->mp_ptrs[i] - delta); } nsize = node_ds(node) - delta; len = PAGEHDRSZ; } sp->mp_upper = sp->mp_lower; sp->mp_pgno = mp->mp_pgno; node_set_ds(node, nsize); /* Shift upward */ base = (char *)mp + mp->mp_upper + PAGEHDRSZ; memmove(base + delta, base, (char *)sp + len - base); ptr = mp->mp_ptrs[indx]; for (i = page_numkeys(mp); --i >= 0;) { if (mp->mp_ptrs[i] <= ptr) { assert((size_t)UINT16_MAX - mp->mp_ptrs[i] >= delta); mp->mp_ptrs[i] += (indx_t)delta; } } assert((size_t)UINT16_MAX - mp->mp_upper >= delta); mp->mp_upper += (indx_t)delta; } /* Initial setup of a sorted-dups cursor. * * Sorted duplicates are implemented as a sub-database for the given key. * The duplicate data items are actually keys of the sub-database. * Operations on the duplicate data items are performed using a sub-cursor * initialized when the sub-database is first accessed. This function does * the preliminary setup of the sub-cursor, filling in the fields that * depend only on the parent DB. * * [in] mc The main cursor whose sorted-dups cursor is to be initialized. */ static int mdbx_xcursor_init0(MDBX_cursor *mc) { MDBX_xcursor *mx = mc->mc_xcursor; if (unlikely(mx == nullptr)) { mdbx_error("unexpected dupsort-page for non-dupsort db/cursor (dbi %u)", mc->mc_dbi); return MDBX_CORRUPTED; } mx->mx_cursor.mc_xcursor = NULL; mx->mx_cursor.mc_txn = mc->mc_txn; mx->mx_cursor.mc_db = &mx->mx_db; mx->mx_cursor.mc_dbx = &mx->mx_dbx; mx->mx_cursor.mc_dbi = mc->mc_dbi; mx->mx_cursor.mc_dbistate = &mx->mx_dbistate; mx->mx_cursor.mc_snum = 0; mx->mx_cursor.mc_top = 0; mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_COPYING | C_SKIPORD)); mx->mx_dbx.md_name.iov_len = 0; mx->mx_dbx.md_name.iov_base = NULL; mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp; mx->mx_dbx.md_dcmp = NULL; mx->mx_dbx.md_klen_min = INT_MAX; mx->mx_dbx.md_vlen_min = mx->mx_dbx.md_klen_max = mx->mx_dbx.md_vlen_max = 0; return MDBX_SUCCESS; } /* Final setup of a sorted-dups cursor. * Sets up the fields that depend on the data from the main cursor. * [in] mc The main cursor whose sorted-dups cursor is to be initialized. * [in] node The data containing the MDBX_db record for the sorted-dup database. */ static int mdbx_xcursor_init1(MDBX_cursor *mc, MDBX_node *node, const MDBX_page *mp) { MDBX_xcursor *mx = mc->mc_xcursor; if (unlikely(mx == nullptr)) { mdbx_error("unexpected dupsort-page for non-dupsort db/cursor (dbi %u)", mc->mc_dbi); return MDBX_CORRUPTED; } const uint8_t flags = node_flags(node); switch (flags) { default: mdbx_error("invalid node flags %u", flags); return MDBX_CORRUPTED; case F_DUPDATA | F_SUBDATA: if (unlikely(node_ds(node) != sizeof(MDBX_db))) { mdbx_error("invalid nested-db record size %zu", node_ds(node)); return MDBX_CORRUPTED; } memcpy(&mx->mx_db, node_data(node), sizeof(MDBX_db)); const txnid_t pp_txnid = IS_DIRTY(mp) ? mc->mc_txn->mt_txnid : mp->mp_txnid; if (unlikely(mx->mx_db.md_mod_txnid > pp_txnid)) { mdbx_error("nested-db.md_mod_txnid (%" PRIaTXN ") > page-txnid (%" PRIaTXN ")", mx->mx_db.md_mod_txnid, pp_txnid); return MDBX_CORRUPTED; } mx->mx_cursor.mc_pg[0] = 0; mx->mx_cursor.mc_snum = 0; mx->mx_cursor.mc_top = 0; mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_COPYING | C_SKIPORD)); break; case F_DUPDATA: if (unlikely(node_ds(node) <= PAGEHDRSZ)) { mdbx_error("invalid nested-page size %zu", node_ds(node)); return MDBX_CORRUPTED; } MDBX_page *fp = node_data(node); mx->mx_db.md_depth = 1; mx->mx_db.md_branch_pages = 0; mx->mx_db.md_leaf_pages = 1; mx->mx_db.md_overflow_pages = 0; mx->mx_db.md_entries = page_numkeys(fp); mx->mx_db.md_root = fp->mp_pgno; mx->mx_db.md_mod_txnid = mp->mp_txnid; mx->mx_cursor.mc_snum = 1; mx->mx_cursor.mc_top = 0; mx->mx_cursor.mc_flags = C_INITIALIZED | C_SUB | (mc->mc_flags & (C_COPYING | C_SKIPORD)); mx->mx_cursor.mc_pg[0] = fp; mx->mx_cursor.mc_ki[0] = 0; mx->mx_db.md_flags = flags_db2sub(mc->mc_db->md_flags); mx->mx_db.md_xsize = (mc->mc_db->md_flags & MDBX_DUPFIXED) ? fp->mp_leaf2_ksize : 0; break; } if (unlikely(mx->mx_db.md_xsize != mc->mc_db->md_xsize)) { if (unlikely(mc->mc_db->md_xsize != 0)) { mdbx_error("cursor mismatched nested-db md_xsize %u", mc->mc_db->md_xsize); return MDBX_CORRUPTED; } if (unlikely((mc->mc_db->md_flags & MDBX_DUPFIXED) == 0)) { mdbx_error("mismatched nested-db md_flags %u", mc->mc_db->md_flags); return MDBX_CORRUPTED; } if (unlikely(mx->mx_db.md_xsize < mc->mc_dbx->md_vlen_min || mx->mx_db.md_xsize > mc->mc_dbx->md_vlen_max)) { mdbx_error("mismatched nested-db.md_xsize (%u) <> min/max value-length " "(%zu/%zu)", mx->mx_db.md_xsize, mc->mc_dbx->md_vlen_min, mc->mc_dbx->md_vlen_max); return MDBX_CORRUPTED; } mc->mc_db->md_xsize = mx->mx_db.md_xsize; mc->mc_dbx->md_vlen_min = mc->mc_dbx->md_vlen_max = mx->mx_db.md_xsize; } mx->mx_dbx.md_klen_min = mc->mc_dbx->md_vlen_min; mx->mx_dbx.md_klen_max = mc->mc_dbx->md_vlen_max; mdbx_debug("Sub-db -%u root page %" PRIaPGNO, mx->mx_cursor.mc_dbi, mx->mx_db.md_root); mx->mx_dbistate = DBI_VALID | DBI_USRVALID | DBI_DUPDATA; return MDBX_SUCCESS; } /* Fixup a sorted-dups cursor due to underlying update. * Sets up some fields that depend on the data from the main cursor. * Almost the same as init1, but skips initialization steps if the * xcursor had already been used. * [in] mc The main cursor whose sorted-dups cursor is to be fixed up. * [in] src_mx The xcursor of an up-to-date cursor. * [in] new_dupdata True if converting from a non-F_DUPDATA item. */ static int mdbx_xcursor_init2(MDBX_cursor *mc, MDBX_xcursor *src_mx, bool new_dupdata) { MDBX_xcursor *mx = mc->mc_xcursor; if (unlikely(mx == nullptr)) { mdbx_error("unexpected dupsort-page for non-dupsort db/cursor (dbi %u)", mc->mc_dbi); return MDBX_CORRUPTED; } if (new_dupdata) { mx->mx_cursor.mc_snum = 1; mx->mx_cursor.mc_top = 0; mx->mx_cursor.mc_flags |= C_INITIALIZED; mx->mx_cursor.mc_ki[0] = 0; mx->mx_dbistate = DBI_VALID | DBI_USRVALID | DBI_DUPDATA; } mx->mx_dbx.md_klen_min = src_mx->mx_dbx.md_klen_min; mx->mx_dbx.md_klen_max = src_mx->mx_dbx.md_klen_max; mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp; mx->mx_db = src_mx->mx_db; mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0]; if (mx->mx_cursor.mc_flags & C_INITIALIZED) { mdbx_debug("Sub-db -%u root page %" PRIaPGNO, mx->mx_cursor.mc_dbi, mx->mx_db.md_root); } return MDBX_SUCCESS; } static __inline int mdbx_couple_init(MDBX_cursor_couple *couple, const MDBX_dbi dbi, MDBX_txn *const txn, MDBX_db *const db, MDBX_dbx *const dbx, uint8_t *const dbstate) { couple->outer.mc_signature = MDBX_MC_LIVE; couple->outer.mc_next = NULL; couple->outer.mc_backup = NULL; couple->outer.mc_dbi = dbi; couple->outer.mc_txn = txn; couple->outer.mc_db = db; couple->outer.mc_dbx = dbx; couple->outer.mc_dbistate = dbstate; couple->outer.mc_snum = 0; couple->outer.mc_top = 0; couple->outer.mc_pg[0] = 0; couple->outer.mc_flags = 0; couple->outer.mc_ki[0] = 0; couple->outer.mc_xcursor = NULL; int rc = MDBX_SUCCESS; if (unlikely(*couple->outer.mc_dbistate & DBI_STALE)) { rc = mdbx_page_search(&couple->outer, NULL, MDBX_PS_ROOTONLY); rc = (rc != MDBX_NOTFOUND) ? rc : MDBX_SUCCESS; } else if (unlikely(couple->outer.mc_dbx->md_klen_max == 0)) { rc = mdbx_setup_dbx(couple->outer.mc_dbx, couple->outer.mc_db, txn->mt_env->me_psize); } if (couple->outer.mc_db->md_flags & MDBX_DUPSORT) { couple->inner.mx_cursor.mc_signature = MDBX_MC_LIVE; couple->outer.mc_xcursor = &couple->inner; rc = mdbx_xcursor_init0(&couple->outer); if (unlikely(rc != MDBX_SUCCESS)) return rc; couple->inner.mx_dbx.md_klen_min = couple->outer.mc_dbx->md_vlen_min; couple->inner.mx_dbx.md_klen_max = couple->outer.mc_dbx->md_vlen_max; } return rc; } /* Initialize a cursor for a given transaction and database. */ static int mdbx_cursor_init(MDBX_cursor *mc, MDBX_txn *txn, MDBX_dbi dbi) { STATIC_ASSERT(offsetof(MDBX_cursor_couple, outer) == 0); if (unlikely(TXN_DBI_CHANGED(txn, dbi))) return MDBX_BAD_DBI; return mdbx_couple_init(container_of(mc, MDBX_cursor_couple, outer), dbi, txn, &txn->mt_dbs[dbi], &txn->mt_dbxs[dbi], &txn->mt_dbistate[dbi]); } MDBX_cursor *mdbx_cursor_create(void *context) { MDBX_cursor_couple *couple = mdbx_calloc(1, sizeof(MDBX_cursor_couple)); if (unlikely(!couple)) return nullptr; couple->outer.mc_signature = MDBX_MC_READY4CLOSE; couple->outer.mc_dbi = UINT_MAX; couple->mc_userctx = context; return &couple->outer; } int mdbx_cursor_set_userctx(MDBX_cursor *mc, void *ctx) { if (unlikely(!mc)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_READY4CLOSE && mc->mc_signature != MDBX_MC_LIVE)) return MDBX_EBADSIGN; MDBX_cursor_couple *couple = container_of(mc, MDBX_cursor_couple, outer); couple->mc_userctx = ctx; return MDBX_SUCCESS; } void *mdbx_cursor_get_userctx(const MDBX_cursor *mc) { if (unlikely(!mc)) return nullptr; if (unlikely(mc->mc_signature != MDBX_MC_READY4CLOSE && mc->mc_signature != MDBX_MC_LIVE)) return nullptr; MDBX_cursor_couple *couple = container_of(mc, MDBX_cursor_couple, outer); return couple->mc_userctx; } int mdbx_cursor_bind(MDBX_txn *txn, MDBX_cursor *mc, MDBX_dbi dbi) { if (unlikely(!mc)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_READY4CLOSE && mc->mc_signature != MDBX_MC_LIVE)) return MDBX_EBADSIGN; int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_VALID))) return MDBX_BAD_DBI; if (unlikely(dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDBX_TXN_RDONLY))) return MDBX_EACCESS; if (unlikely(mc->mc_backup)) /* Cursor from parent transaction */ { mdbx_cassert(mc, mc->mc_signature == MDBX_MC_LIVE); if (unlikely(mc->mc_dbi != dbi || /* paranoia */ mc->mc_signature != MDBX_MC_LIVE || mc->mc_txn != txn)) return MDBX_EINVAL; assert(mc->mc_db == &txn->mt_dbs[dbi]); assert(mc->mc_dbx == &txn->mt_dbxs[dbi]); assert(mc->mc_dbi == dbi); assert(mc->mc_dbistate == &txn->mt_dbistate[dbi]); return likely(mc->mc_dbi == dbi && /* paranoia */ mc->mc_signature == MDBX_MC_LIVE && mc->mc_txn == txn) ? MDBX_SUCCESS : MDBX_EINVAL /* Disallow change DBI in nested transactions */; } if (mc->mc_signature == MDBX_MC_LIVE) { if (unlikely(!mc->mc_txn || mc->mc_txn->mt_signature != MDBX_MT_SIGNATURE)) return MDBX_PROBLEM; if (mc->mc_flags & C_UNTRACK) { mdbx_cassert(mc, !(mc->mc_txn->mt_flags & MDBX_TXN_RDONLY)); MDBX_cursor **prev = &mc->mc_txn->tw.cursors[mc->mc_dbi]; while (*prev && *prev != mc) prev = &(*prev)->mc_next; mdbx_cassert(mc, *prev == mc); *prev = mc->mc_next; } mc->mc_signature = MDBX_MC_READY4CLOSE; mc->mc_flags = 0; mc->mc_dbi = UINT_MAX; mc->mc_next = NULL; mc->mc_db = NULL; mc->mc_dbx = NULL; mc->mc_dbistate = NULL; } mdbx_cassert(mc, !(mc->mc_flags & C_UNTRACK)); rc = mdbx_cursor_init(mc, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (!(txn->mt_flags & MDBX_TXN_RDONLY)) { mc->mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = mc; mc->mc_flags |= C_UNTRACK; } return MDBX_SUCCESS; } int mdbx_cursor_open(MDBX_txn *txn, MDBX_dbi dbi, MDBX_cursor **ret) { if (unlikely(!ret)) return MDBX_EINVAL; *ret = NULL; MDBX_cursor *const mc = mdbx_cursor_create(nullptr); if (unlikely(!mc)) return MDBX_ENOMEM; int rc = mdbx_cursor_bind(txn, mc, dbi); if (unlikely(rc != MDBX_SUCCESS)) { mdbx_cursor_close(mc); return rc; } *ret = mc; return MDBX_SUCCESS; } int mdbx_cursor_renew(MDBX_txn *txn, MDBX_cursor *mc) { return likely(mc) ? mdbx_cursor_bind(txn, mc, mc->mc_dbi) : MDBX_EINVAL; } int mdbx_cursor_copy(const MDBX_cursor *src, MDBX_cursor *dest) { if (unlikely(!src)) return MDBX_EINVAL; if (unlikely(src->mc_signature != MDBX_MC_LIVE)) return (src->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = mdbx_cursor_bind(src->mc_txn, dest, src->mc_dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; assert(dest->mc_db == src->mc_db); assert(dest->mc_dbi == src->mc_dbi); assert(dest->mc_dbx == src->mc_dbx); assert(dest->mc_dbistate == src->mc_dbistate); again: assert(dest->mc_txn == src->mc_txn); dest->mc_flags ^= (dest->mc_flags ^ src->mc_flags) & ~C_UNTRACK; dest->mc_top = src->mc_top; dest->mc_snum = src->mc_snum; for (unsigned i = 0; i < src->mc_snum; ++i) { dest->mc_ki[i] = src->mc_ki[i]; dest->mc_pg[i] = src->mc_pg[i]; } if (src->mc_xcursor) { dest->mc_xcursor->mx_db = src->mc_xcursor->mx_db; dest->mc_xcursor->mx_dbx = src->mc_xcursor->mx_dbx; src = &src->mc_xcursor->mx_cursor; dest = &dest->mc_xcursor->mx_cursor; goto again; } return MDBX_SUCCESS; } void mdbx_cursor_close(MDBX_cursor *mc) { if (likely(mc)) { mdbx_ensure(NULL, mc->mc_signature == MDBX_MC_LIVE || mc->mc_signature == MDBX_MC_READY4CLOSE); MDBX_txn *const txn = mc->mc_txn; if (!mc->mc_backup) { mc->mc_txn = NULL; /* Remove from txn, if tracked. * A read-only txn (!C_UNTRACK) may have been freed already, * so do not peek inside it. Only write txns track cursors. */ if (mc->mc_flags & C_UNTRACK) { mdbx_ensure(txn->mt_env, check_txn_rw(txn, 0) == MDBX_SUCCESS); MDBX_cursor **prev = &txn->tw.cursors[mc->mc_dbi]; while (*prev && *prev != mc) prev = &(*prev)->mc_next; mdbx_tassert(txn, *prev == mc); *prev = mc->mc_next; } mc->mc_signature = 0; mc->mc_next = mc; mdbx_free(mc); } else { /* Cursor closed before nested txn ends */ mdbx_tassert(txn, mc->mc_signature == MDBX_MC_LIVE); mdbx_ensure(txn->mt_env, check_txn_rw(txn, 0) == MDBX_SUCCESS); mc->mc_signature = MDBX_MC_WAIT4EOT; } } } MDBX_txn *mdbx_cursor_txn(const MDBX_cursor *mc) { if (unlikely(!mc || mc->mc_signature != MDBX_MC_LIVE)) return NULL; MDBX_txn *txn = mc->mc_txn; if (unlikely(!txn || txn->mt_signature != MDBX_MT_SIGNATURE)) return NULL; if (unlikely(txn->mt_flags & MDBX_TXN_FINISHED)) return NULL; return txn; } MDBX_dbi mdbx_cursor_dbi(const MDBX_cursor *mc) { if (unlikely(!mc || mc->mc_signature != MDBX_MC_LIVE)) return UINT_MAX; return mc->mc_dbi; } /* Return the count of duplicate data items for the current key */ int mdbx_cursor_count(const MDBX_cursor *mc, size_t *countp) { if (unlikely(mc == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn(mc->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(countp == NULL || !(mc->mc_flags & C_INITIALIZED))) return MDBX_EINVAL; if (!mc->mc_snum) { *countp = 0; return MDBX_NOTFOUND; } MDBX_page *mp = mc->mc_pg[mc->mc_top]; if ((mc->mc_flags & C_EOF) && mc->mc_ki[mc->mc_top] >= page_numkeys(mp)) { *countp = 0; return MDBX_NOTFOUND; } *countp = 1; if (mc->mc_xcursor != NULL) { MDBX_node *node = page_node(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { mdbx_cassert(mc, mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)); *countp = unlikely(mc->mc_xcursor->mx_db.md_entries > PTRDIFF_MAX) ? PTRDIFF_MAX : (size_t)mc->mc_xcursor->mx_db.md_entries; } } return MDBX_SUCCESS; } /* Replace the key for a branch node with a new key. * Set MDBX_TXN_ERROR on failure. * [in] mc Cursor pointing to the node to operate on. * [in] key The new key to use. * Returns 0 on success, non-zero on failure. */ static int mdbx_update_key(MDBX_cursor *mc, const MDBX_val *key) { MDBX_page *mp; MDBX_node *node; char *base; size_t len; int delta, ksize, oksize; int ptr, i, nkeys, indx; DKBUF; mdbx_cassert(mc, cursor_is_tracked(mc)); indx = mc->mc_ki[mc->mc_top]; mp = mc->mc_pg[mc->mc_top]; node = page_node(mp, indx); ptr = mp->mp_ptrs[indx]; if (MDBX_DEBUG) { MDBX_val k2; char kbuf2[DKBUF_MAXKEYSIZE * 2 + 1]; k2.iov_base = node_key(node); k2.iov_len = node_ks(node); mdbx_debug("update key %u (offset %u) [%s] to [%s] on page %" PRIaPGNO, indx, ptr, mdbx_dump_val(&k2, kbuf2, sizeof(kbuf2)), DKEY(key), mp->mp_pgno); } /* Sizes must be 2-byte aligned. */ ksize = EVEN(key->iov_len); oksize = EVEN(node_ks(node)); delta = ksize - oksize; /* Shift node contents if EVEN(key length) changed. */ if (delta) { if (delta > (int)page_room(mp)) { /* not enough space left, do a delete and split */ mdbx_debug("Not enough room, delta = %d, splitting...", delta); pgno_t pgno = node_pgno(node); mdbx_node_del(mc, 0); int rc = mdbx_page_split(mc, key, NULL, pgno, MDBX_SPLIT_REPLACE); if (rc == MDBX_SUCCESS && mdbx_audit_enabled()) rc = mdbx_cursor_check(mc, C_UPDATING); return rc; } nkeys = page_numkeys(mp); for (i = 0; i < nkeys; i++) { if (mp->mp_ptrs[i] <= ptr) { mdbx_cassert(mc, mp->mp_ptrs[i] >= delta); mp->mp_ptrs[i] -= (indx_t)delta; } } base = (char *)mp + mp->mp_upper + PAGEHDRSZ; len = ptr - mp->mp_upper + NODESIZE; memmove(base - delta, base, len); mdbx_cassert(mc, mp->mp_upper >= delta); mp->mp_upper -= (indx_t)delta; node = page_node(mp, indx); } /* But even if no shift was needed, update ksize */ node_set_ks(node, key->iov_len); if (likely(key->iov_len /* to avoid UBSAN traps*/ != 0)) memcpy(node_key(node), key->iov_base, key->iov_len); return MDBX_SUCCESS; } /* Move a node from csrc to cdst. */ static int mdbx_node_move(MDBX_cursor *csrc, MDBX_cursor *cdst, int fromleft) { int rc; DKBUF; MDBX_page *psrc = csrc->mc_pg[csrc->mc_top]; MDBX_page *pdst = cdst->mc_pg[cdst->mc_top]; mdbx_cassert(csrc, PAGETYPE(psrc) == PAGETYPE(pdst)); mdbx_cassert(csrc, csrc->mc_dbi == cdst->mc_dbi); mdbx_cassert(csrc, csrc->mc_top == cdst->mc_top); if (unlikely(PAGETYPE(psrc) != PAGETYPE(pdst))) { bailout: csrc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return MDBX_PROBLEM; } MDBX_val key4move; switch (PAGETYPE(psrc)) { case P_BRANCH: { const MDBX_node *srcnode = page_node(psrc, csrc->mc_ki[csrc->mc_top]); mdbx_cassert(csrc, node_flags(srcnode) == 0); const pgno_t srcpg = node_pgno(srcnode); key4move.iov_len = node_ks(srcnode); key4move.iov_base = node_key(srcnode); if (csrc->mc_ki[csrc->mc_top] == 0) { const unsigned snum = csrc->mc_snum; mdbx_cassert(csrc, snum > 0); /* must find the lowest key below src */ rc = mdbx_page_search_lowest(csrc); MDBX_page *lowest_page = csrc->mc_pg[csrc->mc_top]; if (unlikely(rc)) return rc; mdbx_cassert(csrc, IS_LEAF(lowest_page)); if (unlikely(!IS_LEAF(lowest_page))) goto bailout; if (IS_LEAF2(lowest_page)) { key4move.iov_len = csrc->mc_db->md_xsize; key4move.iov_base = page_leaf2key(lowest_page, 0, key4move.iov_len); } else { const MDBX_node *lowest_node = page_node(lowest_page, 0); key4move.iov_len = node_ks(lowest_node); key4move.iov_base = node_key(lowest_node); } /* restore cursor after mdbx_page_search_lowest() */ csrc->mc_snum = snum; csrc->mc_top = snum - 1; csrc->mc_ki[csrc->mc_top] = 0; /* paranoia */ mdbx_cassert(csrc, psrc == csrc->mc_pg[csrc->mc_top]); mdbx_cassert(csrc, IS_BRANCH(psrc)); if (unlikely(!IS_BRANCH(psrc))) goto bailout; } if (cdst->mc_ki[cdst->mc_top] == 0) { const unsigned snum = cdst->mc_snum; mdbx_cassert(csrc, snum > 0); MDBX_cursor mn; cursor_copy_internal(cdst, &mn); mn.mc_xcursor = NULL; /* must find the lowest key below dst */ rc = mdbx_page_search_lowest(&mn); if (unlikely(rc)) return rc; MDBX_page *const lowest_page = mn.mc_pg[mn.mc_top]; mdbx_cassert(cdst, IS_LEAF(lowest_page)); if (unlikely(!IS_LEAF(lowest_page))) goto bailout; MDBX_val key; if (IS_LEAF2(lowest_page)) { key.iov_len = mn.mc_db->md_xsize; key.iov_base = page_leaf2key(lowest_page, 0, key.iov_len); } else { MDBX_node *lowest_node = page_node(lowest_page, 0); key.iov_len = node_ks(lowest_node); key.iov_base = node_key(lowest_node); } /* restore cursor after mdbx_page_search_lowest() */ mn.mc_snum = snum; mn.mc_top = snum - 1; mn.mc_ki[mn.mc_top] = 0; const intptr_t delta = EVEN(key.iov_len) - EVEN(node_ks(page_node(mn.mc_pg[mn.mc_top], 0))); const intptr_t needed = branch_size(cdst->mc_txn->mt_env, &key4move) + delta; const intptr_t have = page_room(pdst); if (unlikely(needed > have)) return MDBX_RESULT_TRUE; if (unlikely((rc = mdbx_page_touch(csrc)) || (rc = mdbx_page_touch(cdst)))) return rc; psrc = csrc->mc_pg[csrc->mc_top]; pdst = cdst->mc_pg[cdst->mc_top]; WITH_CURSOR_TRACKING(mn, rc = mdbx_update_key(&mn, &key)); if (unlikely(rc)) return rc; } else { const size_t needed = branch_size(cdst->mc_txn->mt_env, &key4move); const size_t have = page_room(pdst); if (unlikely(needed > have)) return MDBX_RESULT_TRUE; if (unlikely((rc = mdbx_page_touch(csrc)) || (rc = mdbx_page_touch(cdst)))) return rc; psrc = csrc->mc_pg[csrc->mc_top]; pdst = cdst->mc_pg[cdst->mc_top]; } mdbx_debug("moving %s-node %u [%s] on page %" PRIaPGNO " to node %u on page %" PRIaPGNO, "branch", csrc->mc_ki[csrc->mc_top], DKEY(&key4move), psrc->mp_pgno, cdst->mc_ki[cdst->mc_top], pdst->mp_pgno); /* Add the node to the destination page. */ rc = mdbx_node_add_branch(cdst, cdst->mc_ki[cdst->mc_top], &key4move, srcpg); } break; case P_LEAF: { /* Mark src and dst as dirty. */ if (unlikely((rc = mdbx_page_touch(csrc)) || (rc = mdbx_page_touch(cdst)))) return rc; psrc = csrc->mc_pg[csrc->mc_top]; pdst = cdst->mc_pg[cdst->mc_top]; const MDBX_node *srcnode = page_node(psrc, csrc->mc_ki[csrc->mc_top]); MDBX_val data; data.iov_len = node_ds(srcnode); data.iov_base = node_data(srcnode); key4move.iov_len = node_ks(srcnode); key4move.iov_base = node_key(srcnode); mdbx_debug("moving %s-node %u [%s] on page %" PRIaPGNO " to node %u on page %" PRIaPGNO, "leaf", csrc->mc_ki[csrc->mc_top], DKEY(&key4move), psrc->mp_pgno, cdst->mc_ki[cdst->mc_top], pdst->mp_pgno); /* Add the node to the destination page. */ rc = mdbx_node_add_leaf(cdst, cdst->mc_ki[cdst->mc_top], &key4move, &data, node_flags(srcnode)); } break; case P_LEAF | P_LEAF2: { /* Mark src and dst as dirty. */ if (unlikely((rc = mdbx_page_touch(csrc)) || (rc = mdbx_page_touch(cdst)))) return rc; psrc = csrc->mc_pg[csrc->mc_top]; pdst = cdst->mc_pg[cdst->mc_top]; key4move.iov_len = csrc->mc_db->md_xsize; key4move.iov_base = page_leaf2key(psrc, csrc->mc_ki[csrc->mc_top], key4move.iov_len); mdbx_debug("moving %s-node %u [%s] on page %" PRIaPGNO " to node %u on page %" PRIaPGNO, "leaf2", csrc->mc_ki[csrc->mc_top], DKEY(&key4move), psrc->mp_pgno, cdst->mc_ki[cdst->mc_top], pdst->mp_pgno); /* Add the node to the destination page. */ rc = mdbx_node_add_leaf2(cdst, cdst->mc_ki[cdst->mc_top], &key4move); } break; default: goto bailout; } if (unlikely(rc != MDBX_SUCCESS)) return rc; /* Delete the node from the source page. */ mdbx_node_del(csrc, key4move.iov_len); mdbx_cassert(csrc, psrc == csrc->mc_pg[csrc->mc_top]); mdbx_cassert(cdst, pdst == cdst->mc_pg[cdst->mc_top]); mdbx_cassert(csrc, PAGETYPE(psrc) == PAGETYPE(pdst)); { /* Adjust other cursors pointing to mp */ MDBX_cursor *m2, *m3; const MDBX_dbi dbi = csrc->mc_dbi; mdbx_cassert(csrc, csrc->mc_top == cdst->mc_top); if (fromleft) { /* If we're adding on the left, bump others up */ for (m2 = csrc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { m3 = (csrc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) continue; if (m3 != cdst && m3->mc_pg[csrc->mc_top] == pdst && m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) { m3->mc_ki[csrc->mc_top]++; } if (m3 != csrc && m3->mc_pg[csrc->mc_top] == psrc && m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) { m3->mc_pg[csrc->mc_top] = pdst; m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; mdbx_cassert(csrc, csrc->mc_top > 0); m3->mc_ki[csrc->mc_top - 1]++; } if (XCURSOR_INITED(m3) && IS_LEAF(psrc)) XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]); } } else { /* Adding on the right, bump others down */ for (m2 = csrc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { m3 = (csrc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == csrc) continue; if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) continue; if (m3->mc_pg[csrc->mc_top] == psrc) { if (!m3->mc_ki[csrc->mc_top]) { m3->mc_pg[csrc->mc_top] = pdst; m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; mdbx_cassert(csrc, csrc->mc_top > 0); m3->mc_ki[csrc->mc_top - 1]--; } else { m3->mc_ki[csrc->mc_top]--; } if (XCURSOR_INITED(m3) && IS_LEAF(psrc)) XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]); } } } } /* Update the parent separators. */ if (csrc->mc_ki[csrc->mc_top] == 0) { mdbx_cassert(csrc, csrc->mc_top > 0); if (csrc->mc_ki[csrc->mc_top - 1] != 0) { MDBX_val key; if (IS_LEAF2(psrc)) { key.iov_len = psrc->mp_leaf2_ksize; key.iov_base = page_leaf2key(psrc, 0, key.iov_len); } else { MDBX_node *srcnode = page_node(psrc, 0); key.iov_len = node_ks(srcnode); key.iov_base = node_key(srcnode); } mdbx_debug("update separator for source page %" PRIaPGNO " to [%s]", psrc->mp_pgno, DKEY(&key)); MDBX_cursor mn; cursor_copy_internal(csrc, &mn); mn.mc_xcursor = NULL; mdbx_cassert(csrc, mn.mc_snum > 0); mn.mc_snum--; mn.mc_top--; /* We want mdbx_rebalance to find mn when doing fixups */ WITH_CURSOR_TRACKING(mn, rc = mdbx_update_key(&mn, &key)); if (unlikely(rc != MDBX_SUCCESS)) return rc; } if (IS_BRANCH(psrc)) { const MDBX_val nullkey = {0, 0}; const indx_t ix = csrc->mc_ki[csrc->mc_top]; csrc->mc_ki[csrc->mc_top] = 0; rc = mdbx_update_key(csrc, &nullkey); csrc->mc_ki[csrc->mc_top] = ix; mdbx_cassert(csrc, rc == MDBX_SUCCESS); } } if (cdst->mc_ki[cdst->mc_top] == 0) { mdbx_cassert(cdst, cdst->mc_top > 0); if (cdst->mc_ki[cdst->mc_top - 1] != 0) { MDBX_val key; if (IS_LEAF2(pdst)) { key.iov_len = pdst->mp_leaf2_ksize; key.iov_base = page_leaf2key(pdst, 0, key.iov_len); } else { MDBX_node *srcnode = page_node(pdst, 0); key.iov_len = node_ks(srcnode); key.iov_base = node_key(srcnode); } mdbx_debug("update separator for destination page %" PRIaPGNO " to [%s]", pdst->mp_pgno, DKEY(&key)); MDBX_cursor mn; cursor_copy_internal(cdst, &mn); mn.mc_xcursor = NULL; mdbx_cassert(cdst, mn.mc_snum > 0); mn.mc_snum--; mn.mc_top--; /* We want mdbx_rebalance to find mn when doing fixups */ WITH_CURSOR_TRACKING(mn, rc = mdbx_update_key(&mn, &key)); if (unlikely(rc != MDBX_SUCCESS)) return rc; } if (IS_BRANCH(pdst)) { const MDBX_val nullkey = {0, 0}; const indx_t ix = cdst->mc_ki[cdst->mc_top]; cdst->mc_ki[cdst->mc_top] = 0; rc = mdbx_update_key(cdst, &nullkey); cdst->mc_ki[cdst->mc_top] = ix; mdbx_cassert(cdst, rc == MDBX_SUCCESS); } } return MDBX_SUCCESS; } /* Merge one page into another. * * The nodes from the page pointed to by csrc will be copied to the page * pointed to by cdst and then the csrc page will be freed. * * [in] csrc Cursor pointing to the source page. * [in] cdst Cursor pointing to the destination page. * * Returns 0 on success, non-zero on failure. */ static int mdbx_page_merge(MDBX_cursor *csrc, MDBX_cursor *cdst) { MDBX_val key; int rc; mdbx_cassert(csrc, csrc != cdst); mdbx_cassert(csrc, cursor_is_tracked(csrc)); mdbx_cassert(cdst, cursor_is_tracked(cdst)); const MDBX_page *const psrc = csrc->mc_pg[csrc->mc_top]; MDBX_page *pdst = cdst->mc_pg[cdst->mc_top]; mdbx_debug("merging page %" PRIaPGNO " into %" PRIaPGNO, psrc->mp_pgno, pdst->mp_pgno); mdbx_cassert(csrc, PAGETYPE(psrc) == PAGETYPE(pdst)); mdbx_cassert(csrc, csrc->mc_dbi == cdst->mc_dbi && csrc->mc_db == cdst->mc_db); mdbx_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */ mdbx_cassert(cdst, cdst->mc_snum > 1); mdbx_cassert(cdst, cdst->mc_snum < cdst->mc_db->md_depth || IS_LEAF(cdst->mc_pg[cdst->mc_db->md_depth - 1])); mdbx_cassert(csrc, csrc->mc_snum < csrc->mc_db->md_depth || IS_LEAF(csrc->mc_pg[csrc->mc_db->md_depth - 1])); mdbx_cassert(cdst, page_room(pdst) >= page_used(cdst->mc_txn->mt_env, psrc)); const int pagetype = PAGETYPE(psrc); /* Move all nodes from src to dst */ const unsigned dst_nkeys = page_numkeys(pdst); const unsigned src_nkeys = page_numkeys(psrc); mdbx_cassert(cdst, dst_nkeys + src_nkeys >= (unsigned)(IS_LEAF(psrc) ? 1 : MDBX_MINKEYS)); if (likely(src_nkeys)) { unsigned j = dst_nkeys; if (unlikely(pagetype & P_LEAF2)) { /* Mark dst as dirty. */ if (unlikely(rc = mdbx_page_touch(cdst))) return rc; key.iov_len = csrc->mc_db->md_xsize; key.iov_base = page_data(psrc); unsigned i = 0; do { rc = mdbx_node_add_leaf2(cdst, j++, &key); if (unlikely(rc != MDBX_SUCCESS)) return rc; key.iov_base = (char *)key.iov_base + key.iov_len; } while (++i != src_nkeys); } else { MDBX_node *srcnode = page_node(psrc, 0); key.iov_len = node_ks(srcnode); key.iov_base = node_key(srcnode); if (pagetype & P_BRANCH) { MDBX_cursor mn; cursor_copy_internal(csrc, &mn); mn.mc_xcursor = NULL; /* must find the lowest key below src */ rc = mdbx_page_search_lowest(&mn); if (unlikely(rc)) return rc; const MDBX_page *mp = mn.mc_pg[mn.mc_top]; if (likely(!IS_LEAF2(mp))) { mdbx_cassert(&mn, IS_LEAF(mp)); const MDBX_node *lowest = page_node(mp, 0); key.iov_len = node_ks(lowest); key.iov_base = node_key(lowest); } else { mdbx_cassert(&mn, mn.mc_top > csrc->mc_top); key.iov_len = mp->mp_leaf2_ksize; key.iov_base = page_leaf2key(mp, mn.mc_ki[mn.mc_top], key.iov_len); } mdbx_cassert(&mn, key.iov_len >= csrc->mc_dbx->md_klen_min); mdbx_cassert(&mn, key.iov_len <= csrc->mc_dbx->md_klen_max); const size_t dst_room = page_room(pdst); const size_t src_used = page_used(cdst->mc_txn->mt_env, psrc); const size_t space_needed = src_used - node_ks(srcnode) + key.iov_len; if (unlikely(space_needed > dst_room)) return MDBX_RESULT_TRUE; } /* Mark dst as dirty. */ if (unlikely(rc = mdbx_page_touch(cdst))) return rc; unsigned i = 0; while (true) { if (pagetype & P_LEAF) { MDBX_val data; data.iov_len = node_ds(srcnode); data.iov_base = node_data(srcnode); rc = mdbx_node_add_leaf(cdst, j++, &key, &data, node_flags(srcnode)); } else { mdbx_cassert(csrc, node_flags(srcnode) == 0); rc = mdbx_node_add_branch(cdst, j++, &key, node_pgno(srcnode)); } if (unlikely(rc != MDBX_SUCCESS)) return rc; if (++i == src_nkeys) break; srcnode = page_node(psrc, i); key.iov_len = node_ks(srcnode); key.iov_base = node_key(srcnode); } } pdst = cdst->mc_pg[cdst->mc_top]; mdbx_debug("dst page %" PRIaPGNO " now has %u keys (%.1f%% filled)", pdst->mp_pgno, page_numkeys(pdst), page_fill(cdst->mc_txn->mt_env, pdst)); mdbx_cassert(csrc, psrc == csrc->mc_pg[csrc->mc_top]); mdbx_cassert(cdst, pdst == cdst->mc_pg[cdst->mc_top]); } /* Unlink the src page from parent and add to free list. */ csrc->mc_top--; mdbx_node_del(csrc, 0); if (csrc->mc_ki[csrc->mc_top] == 0) { const MDBX_val nullkey = {0, 0}; rc = mdbx_update_key(csrc, &nullkey); if (unlikely(rc)) { csrc->mc_top++; return rc; } } csrc->mc_top++; mdbx_cassert(csrc, psrc == csrc->mc_pg[csrc->mc_top]); mdbx_cassert(cdst, pdst == cdst->mc_pg[cdst->mc_top]); { /* Adjust other cursors pointing to mp */ MDBX_cursor *m2, *m3; const MDBX_dbi dbi = csrc->mc_dbi; const unsigned top = csrc->mc_top; for (m2 = csrc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { m3 = (csrc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == csrc || top >= m3->mc_snum) continue; if (m3->mc_pg[top] == psrc) { m3->mc_pg[top] = pdst; mdbx_cassert(m3, dst_nkeys + m3->mc_ki[top] <= UINT16_MAX); m3->mc_ki[top] += (indx_t)dst_nkeys; m3->mc_ki[top - 1] = cdst->mc_ki[top - 1]; } else if (m3->mc_pg[top - 1] == csrc->mc_pg[top - 1] && m3->mc_ki[top - 1] > csrc->mc_ki[top - 1]) { m3->mc_ki[top - 1]--; } if (XCURSOR_INITED(m3) && IS_LEAF(psrc)) XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]); } } /* If not operating on GC, allow this page to be reused * in this txn. Otherwise just add to free list. */ rc = mdbx_page_retire(csrc, (MDBX_page *)psrc); if (unlikely(rc)) return rc; mdbx_cassert(cdst, cdst->mc_db->md_entries > 0); mdbx_cassert(cdst, cdst->mc_snum <= cdst->mc_db->md_depth); mdbx_cassert(cdst, cdst->mc_top > 0); mdbx_cassert(cdst, cdst->mc_snum == cdst->mc_top + 1); MDBX_page *const top_page = cdst->mc_pg[cdst->mc_top]; const indx_t top_indx = cdst->mc_ki[cdst->mc_top]; const unsigned save_snum = cdst->mc_snum; const uint16_t save_depth = cdst->mc_db->md_depth; mdbx_cursor_pop(cdst); rc = mdbx_rebalance(cdst); if (unlikely(rc)) return rc; mdbx_cassert(cdst, cdst->mc_db->md_entries > 0); mdbx_cassert(cdst, cdst->mc_snum <= cdst->mc_db->md_depth); mdbx_cassert(cdst, cdst->mc_snum == cdst->mc_top + 1); if (IS_LEAF(cdst->mc_pg[cdst->mc_top])) { /* LY: don't touch cursor if top-page is a LEAF */ mdbx_cassert(cdst, IS_LEAF(cdst->mc_pg[cdst->mc_top]) || PAGETYPE(cdst->mc_pg[cdst->mc_top]) == pagetype); return MDBX_SUCCESS; } mdbx_cassert(cdst, page_numkeys(top_page) == dst_nkeys + src_nkeys); if (pagetype != PAGETYPE(top_page)) { /* LY: LEAF-page becomes BRANCH, unable restore cursor's stack */ goto bailout; } if (top_page == cdst->mc_pg[cdst->mc_top]) { /* LY: don't touch cursor if prev top-page already on the top */ mdbx_cassert(cdst, cdst->mc_ki[cdst->mc_top] == top_indx); mdbx_cassert(cdst, IS_LEAF(cdst->mc_pg[cdst->mc_top]) || PAGETYPE(cdst->mc_pg[cdst->mc_top]) == pagetype); return MDBX_SUCCESS; } const int new_snum = save_snum - save_depth + cdst->mc_db->md_depth; if (unlikely(new_snum < 1 || new_snum > cdst->mc_db->md_depth)) { /* LY: out of range, unable restore cursor's stack */ goto bailout; } if (top_page == cdst->mc_pg[new_snum - 1]) { mdbx_cassert(cdst, cdst->mc_ki[new_snum - 1] == top_indx); /* LY: restore cursor stack */ cdst->mc_snum = (uint16_t)new_snum; cdst->mc_top = (uint16_t)new_snum - 1; mdbx_cassert(cdst, cdst->mc_snum < cdst->mc_db->md_depth || IS_LEAF(cdst->mc_pg[cdst->mc_db->md_depth - 1])); mdbx_cassert(cdst, IS_LEAF(cdst->mc_pg[cdst->mc_top]) || PAGETYPE(cdst->mc_pg[cdst->mc_top]) == pagetype); return MDBX_SUCCESS; } MDBX_page *const stub_page = (MDBX_page *)(~(uintptr_t)top_page); const indx_t stub_indx = top_indx; if (save_depth > cdst->mc_db->md_depth && ((cdst->mc_pg[save_snum - 1] == top_page && cdst->mc_ki[save_snum - 1] == top_indx) || (cdst->mc_pg[save_snum - 1] == stub_page && cdst->mc_ki[save_snum - 1] == stub_indx))) { /* LY: restore cursor stack */ cdst->mc_pg[new_snum - 1] = top_page; cdst->mc_ki[new_snum - 1] = top_indx; cdst->mc_pg[new_snum] = (MDBX_page *)(~(uintptr_t)cdst->mc_pg[new_snum]); cdst->mc_ki[new_snum] = ~cdst->mc_ki[new_snum]; cdst->mc_snum = (uint16_t)new_snum; cdst->mc_top = (uint16_t)new_snum - 1; mdbx_cassert(cdst, cdst->mc_snum < cdst->mc_db->md_depth || IS_LEAF(cdst->mc_pg[cdst->mc_db->md_depth - 1])); mdbx_cassert(cdst, IS_LEAF(cdst->mc_pg[cdst->mc_top]) || PAGETYPE(cdst->mc_pg[cdst->mc_top]) == pagetype); return MDBX_SUCCESS; } bailout: /* LY: unable restore cursor's stack */ cdst->mc_flags &= ~C_INITIALIZED; return MDBX_CURSOR_FULL; } /* Copy the contents of a cursor. * [in] csrc The cursor to copy from. * [out] cdst The cursor to copy to. */ static void cursor_copy_internal(const MDBX_cursor *csrc, MDBX_cursor *cdst) { mdbx_cassert(csrc, csrc->mc_txn->mt_txnid >= csrc->mc_txn->mt_env->me_oldest->weak); cdst->mc_txn = csrc->mc_txn; cdst->mc_dbi = csrc->mc_dbi; cdst->mc_db = csrc->mc_db; cdst->mc_dbx = csrc->mc_dbx; cdst->mc_snum = csrc->mc_snum; cdst->mc_top = csrc->mc_top; cdst->mc_flags = csrc->mc_flags; cdst->mc_dbistate = csrc->mc_dbistate; for (unsigned i = 0; i < csrc->mc_snum; i++) { cdst->mc_pg[i] = csrc->mc_pg[i]; cdst->mc_ki[i] = csrc->mc_ki[i]; } } /* Rebalance the tree after a delete operation. * [in] mc Cursor pointing to the page where rebalancing should begin. * Returns 0 on success, non-zero on failure. */ static int mdbx_rebalance(MDBX_cursor *mc) { mdbx_cassert(mc, cursor_is_tracked(mc)); mdbx_cassert(mc, mc->mc_snum > 0); mdbx_cassert(mc, mc->mc_snum < mc->mc_db->md_depth || IS_LEAF(mc->mc_pg[mc->mc_db->md_depth - 1])); const int pagetype = PAGETYPE(mc->mc_pg[mc->mc_top]); STATIC_ASSERT(P_BRANCH == 1); const unsigned minkeys = (pagetype & P_BRANCH) + 1; /* The threshold of minimum page fill factor, in form of a negative binary * exponent, i.e. X = 2 means 1/(2**X) == 1/(2**2) == 1/4 == 25%. * Pages emptier than this are candidates for merging. */ const unsigned threshold_fill_exp2 = 2; /* The threshold of minimum page fill factor, as a number of free bytes on a * page. Pages emptier than this are candidates for merging. */ const unsigned room_threshold = page_space(mc->mc_txn->mt_env) - (page_space(mc->mc_txn->mt_env) >> threshold_fill_exp2); const MDBX_page *const tp = mc->mc_pg[mc->mc_top]; mdbx_debug("rebalancing %s page %" PRIaPGNO " (has %u keys, %.1f%% full)", (pagetype & P_LEAF) ? "leaf" : "branch", tp->mp_pgno, page_numkeys(tp), page_fill(mc->mc_txn->mt_env, tp)); if (unlikely(page_numkeys(tp) < minkeys)) { mdbx_debug("page %" PRIaPGNO " must be merged due keys < %u threshold", tp->mp_pgno, minkeys); } else if (unlikely(page_room(tp) > room_threshold)) { mdbx_debug("page %" PRIaPGNO " should be merged due room %u > %u threshold", tp->mp_pgno, page_room(tp), room_threshold); } else { mdbx_debug("no need to rebalance page %" PRIaPGNO ", room %u < %u threshold", tp->mp_pgno, page_room(tp), room_threshold); mdbx_cassert(mc, mc->mc_db->md_entries > 0); return MDBX_SUCCESS; } int rc; if (mc->mc_snum < 2) { MDBX_page *const mp = mc->mc_pg[0]; const unsigned nkeys = page_numkeys(mp); mdbx_cassert(mc, (mc->mc_db->md_entries == 0) == (nkeys == 0)); if (IS_SUBP(mp)) { mdbx_debug("%s", "Can't rebalance a subpage, ignoring"); mdbx_cassert(mc, pagetype & P_LEAF); return MDBX_SUCCESS; } if (nkeys == 0) { mdbx_cassert(mc, IS_LEAF(mp)); mdbx_debug("%s", "tree is completely empty"); mc->mc_db->md_root = P_INVALID; mc->mc_db->md_depth = 0; mdbx_cassert(mc, mc->mc_db->md_branch_pages == 0 && mc->mc_db->md_overflow_pages == 0 && mc->mc_db->md_leaf_pages == 1); /* Adjust cursors pointing to mp */ for (MDBX_cursor *m2 = mc->mc_txn->tw.cursors[mc->mc_dbi]; m2; m2 = m2->mc_next) { MDBX_cursor *m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == mc || !(m3->mc_flags & C_INITIALIZED)) continue; if (m3->mc_pg[0] == mp) { m3->mc_snum = 0; m3->mc_top = 0; m3->mc_flags &= ~C_INITIALIZED; } } mc->mc_snum = 0; mc->mc_top = 0; mc->mc_flags &= ~C_INITIALIZED; rc = mdbx_page_retire(mc, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; } else if (IS_BRANCH(mp) && nkeys == 1) { mdbx_debug("%s", "collapsing root page!"); mc->mc_db->md_root = node_pgno(page_node(mp, 0)); rc = mdbx_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL, pp_txnid4chk(mp, mc->mc_txn)); if (unlikely(rc != MDBX_SUCCESS)) return rc; mc->mc_db->md_depth--; mc->mc_ki[0] = mc->mc_ki[1]; for (int i = 1; i < mc->mc_db->md_depth; i++) { mc->mc_pg[i] = mc->mc_pg[i + 1]; mc->mc_ki[i] = mc->mc_ki[i + 1]; } /* Adjust other cursors pointing to mp */ for (MDBX_cursor *m2 = mc->mc_txn->tw.cursors[mc->mc_dbi]; m2; m2 = m2->mc_next) { MDBX_cursor *m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == mc || !(m3->mc_flags & C_INITIALIZED)) continue; if (m3->mc_pg[0] == mp) { for (int i = 0; i < mc->mc_db->md_depth; i++) { m3->mc_pg[i] = m3->mc_pg[i + 1]; m3->mc_ki[i] = m3->mc_ki[i + 1]; } m3->mc_snum--; m3->mc_top--; } } mdbx_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]) || PAGETYPE(mc->mc_pg[mc->mc_top]) == pagetype); mdbx_cassert(mc, mc->mc_snum < mc->mc_db->md_depth || IS_LEAF(mc->mc_pg[mc->mc_db->md_depth - 1])); rc = mdbx_page_retire(mc, mp); if (unlikely(rc != MDBX_SUCCESS)) return rc; } else { mdbx_debug("root page %" PRIaPGNO " doesn't need rebalancing (flags 0x%x)", mp->mp_pgno, mp->mp_flags); } return MDBX_SUCCESS; } /* The parent (branch page) must have at least 2 pointers, * otherwise the tree is invalid. */ const unsigned pre_top = mc->mc_top - 1; mdbx_cassert(mc, IS_BRANCH(mc->mc_pg[pre_top])); mdbx_cassert(mc, !IS_SUBP(mc->mc_pg[0])); mdbx_cassert(mc, page_numkeys(mc->mc_pg[pre_top]) > 1); /* Leaf page fill factor is below the threshold. * Try to move keys from left or right neighbor, or * merge with a neighbor page. */ /* Find neighbors. */ MDBX_cursor mn; cursor_copy_internal(mc, &mn); mn.mc_xcursor = NULL; MDBX_page *left = nullptr, *right = nullptr; if (mn.mc_ki[pre_top] > 0) { rc = mdbx_page_get( &mn, node_pgno(page_node(mn.mc_pg[pre_top], mn.mc_ki[pre_top] - 1)), &left, NULL, pp_txnid4chk(mn.mc_pg[pre_top], mc->mc_txn)); if (unlikely(rc != MDBX_SUCCESS)) return rc; mdbx_cassert(mc, PAGETYPE(left) == PAGETYPE(mc->mc_pg[mc->mc_top])); } if (mn.mc_ki[pre_top] + 1u < page_numkeys(mn.mc_pg[pre_top])) { rc = mdbx_page_get( &mn, node_pgno(page_node(mn.mc_pg[pre_top], mn.mc_ki[pre_top] + 1)), &right, NULL, pp_txnid4chk(mn.mc_pg[pre_top], mc->mc_txn)); if (unlikely(rc != MDBX_SUCCESS)) return rc; mdbx_cassert(mc, PAGETYPE(right) == PAGETYPE(mc->mc_pg[mc->mc_top])); } const indx_t ki_top = mc->mc_ki[mc->mc_top]; const indx_t ki_pre_top = mn.mc_ki[pre_top]; const indx_t nkeys = (indx_t)page_numkeys(mn.mc_pg[mn.mc_top]); if (left && page_room(left) > room_threshold && (!right || page_room(right) < page_room(left))) { /* try merge with left */ mdbx_cassert(mc, page_numkeys(left) >= minkeys); mn.mc_pg[mn.mc_top] = left; mn.mc_ki[mn.mc_top - 1] = ki_pre_top - 1; mn.mc_ki[mn.mc_top] = (indx_t)(page_numkeys(left) - 1); mc->mc_ki[mc->mc_top] = 0; const indx_t new_ki = (indx_t)(ki_top + page_numkeys(left)); mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1; /* We want mdbx_rebalance to find mn when doing fixups */ WITH_CURSOR_TRACKING(mn, rc = mdbx_page_merge(mc, &mn)); if (likely(rc != MDBX_RESULT_TRUE)) { cursor_copy_internal(&mn, mc); mc->mc_ki[mc->mc_top] = new_ki; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } if (right && page_room(right) > room_threshold) { /* try merge with right */ mdbx_cassert(mc, page_numkeys(right) >= minkeys); mn.mc_pg[mn.mc_top] = right; mn.mc_ki[mn.mc_top - 1] = ki_pre_top + 1; mn.mc_ki[mn.mc_top] = 0; mc->mc_ki[mc->mc_top] = nkeys; WITH_CURSOR_TRACKING(mn, rc = mdbx_page_merge(&mn, mc)); if (likely(rc != MDBX_RESULT_TRUE)) { mc->mc_ki[mc->mc_top] = ki_top; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } if (left && page_numkeys(left) > minkeys && (!right || page_numkeys(right) <= minkeys || page_room(right) > page_room(left))) { /* try move from left */ mn.mc_pg[mn.mc_top] = left; mn.mc_ki[mn.mc_top - 1] = ki_pre_top - 1; mn.mc_ki[mn.mc_top] = (indx_t)(page_numkeys(left) - 1); mc->mc_ki[mc->mc_top] = 0; WITH_CURSOR_TRACKING(mn, rc = mdbx_node_move(&mn, mc, true)); if (likely(rc != MDBX_RESULT_TRUE)) { mc->mc_ki[mc->mc_top] = ki_top + 1; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } if (right && page_numkeys(right) > minkeys) { /* try move from right */ mn.mc_pg[mn.mc_top] = right; mn.mc_ki[mn.mc_top - 1] = ki_pre_top + 1; mn.mc_ki[mn.mc_top] = 0; mc->mc_ki[mc->mc_top] = nkeys; WITH_CURSOR_TRACKING(mn, rc = mdbx_node_move(&mn, mc, false)); if (likely(rc != MDBX_RESULT_TRUE)) { mc->mc_ki[mc->mc_top] = ki_top; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } if (nkeys >= minkeys) { #if MDBX_DEBUG > 0 if (mdbx_audit_enabled()) return mdbx_cursor_check(mc, C_UPDATING); #endif return MDBX_SUCCESS; } if (left && (!right || page_room(left) > page_room(right))) { /* try merge with left */ mdbx_cassert(mc, page_numkeys(left) >= minkeys); mn.mc_pg[mn.mc_top] = left; mn.mc_ki[mn.mc_top - 1] = ki_pre_top - 1; mn.mc_ki[mn.mc_top] = (indx_t)(page_numkeys(left) - 1); mc->mc_ki[mc->mc_top] = 0; const indx_t new_ki = (indx_t)(ki_top + page_numkeys(left)); mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1; /* We want mdbx_rebalance to find mn when doing fixups */ WITH_CURSOR_TRACKING(mn, rc = mdbx_page_merge(mc, &mn)); if (likely(rc != MDBX_RESULT_TRUE)) { cursor_copy_internal(&mn, mc); mc->mc_ki[mc->mc_top] = new_ki; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } if (likely(right)) { /* try merge with right */ mdbx_cassert(mc, page_numkeys(right) >= minkeys); mn.mc_pg[mn.mc_top] = right; mn.mc_ki[mn.mc_top - 1] = ki_pre_top + 1; mn.mc_ki[mn.mc_top] = 0; mc->mc_ki[mc->mc_top] = nkeys; WITH_CURSOR_TRACKING(mn, rc = mdbx_page_merge(&mn, mc)); if (likely(rc != MDBX_RESULT_TRUE)) { mc->mc_ki[mc->mc_top] = ki_top; mdbx_cassert(mc, rc || page_numkeys(mc->mc_pg[mc->mc_top]) >= minkeys); return rc; } } return MDBX_PROBLEM; } static __cold int mdbx_page_check(MDBX_cursor *const mc, const MDBX_page *const mp, unsigned options) { options |= mc->mc_flags & (C_COPYING | C_UPDATING | C_RETIRING | C_SKIPORD); MDBX_env *const env = mc->mc_txn->mt_env; const unsigned nkeys = page_numkeys(mp); char *const end_of_page = (char *)mp + env->me_psize; if (unlikely(mp->mp_pgno < MIN_PAGENO || mp->mp_pgno > MAX_PAGENO)) return bad_page(mp, "invalid pgno (%u)\n", mp->mp_pgno); if (IS_OVERFLOW(mp)) { if (unlikely(mp->mp_pages < 1 && mp->mp_pages >= MAX_PAGENO / 2)) return bad_page(mp, "invalid overflow n-pages (%u)\n", mp->mp_pages); if (unlikely(mp->mp_pgno + mp->mp_pages > mc->mc_txn->mt_next_pgno)) return bad_page(mp, "overflow page beyond (%u) next-pgno\n", mp->mp_pgno + mp->mp_pages); return MDBX_SUCCESS; } int rc = MDBX_SUCCESS; if ((options & C_UPDATING) == 0 || !IS_DIRTY(mp)) { if (unlikely(nkeys < 2 && IS_BRANCH(mp))) rc = bad_page(mp, "branch-page nkey (%u) < 2\n", nkeys); } MDBX_val here, prev = {0, 0}; for (unsigned i = 0; i < nkeys; ++i) { if (IS_LEAF2(mp)) { const size_t ksize = mp->mp_leaf2_ksize; char *const key = page_leaf2key(mp, i, ksize); if (unlikely(end_of_page < key + ksize)) { rc = bad_page(mp, "leaf2-key beyond (%zu) page-end\n", key + ksize - end_of_page); continue; } if ((options & C_COPYING) == 0) { if (unlikely(ksize != mc->mc_dbx->md_klen_min)) { if (unlikely(ksize < mc->mc_dbx->md_klen_min || ksize > mc->mc_dbx->md_klen_max)) rc = bad_page( mp, "leaf2-key size (%zu) <> min/max key-length (%zu/%zu)\n", ksize, mc->mc_dbx->md_klen_min, mc->mc_dbx->md_klen_max); else mc->mc_dbx->md_klen_min = mc->mc_dbx->md_klen_max = ksize; } if ((options & C_SKIPORD) == 0) { here.iov_len = ksize; here.iov_base = key; if (prev.iov_base && unlikely(mc->mc_dbx->md_cmp(&here, &prev) <= 0)) rc = bad_page(mp, "leaf2-key #%u wrong order\n", i); prev = here; } } } else { const MDBX_node *const node = page_node(mp, i); const char *node_end = (char *)node + NODESIZE; if (unlikely(node_end > end_of_page)) { rc = bad_page(mp, "node (%zu) beyond page-end\n", node_end - end_of_page); continue; } if (IS_LEAF(mp) || i > 0) { size_t ksize = node_ks(node); char *key = node_key(node); if (unlikely(end_of_page < key + ksize)) { rc = bad_page(mp, "node-key (%zu) beyond page-end\n", key + ksize - end_of_page); continue; } if ((options & C_COPYING) == 0) { if (unlikely(ksize < mc->mc_dbx->md_klen_min || ksize > mc->mc_dbx->md_klen_max)) rc = bad_page( mp, "node-key size (%zu) <> min/max key-length (%zu/%zu)\n", ksize, mc->mc_dbx->md_klen_min, mc->mc_dbx->md_klen_max); if ((options & C_SKIPORD) == 0) { here.iov_base = key; here.iov_len = ksize; if (prev.iov_base && unlikely(mc->mc_dbx->md_cmp(&here, &prev) <= 0)) rc = bad_page(mp, "node-key #%u wrong order\n", i); prev = here; } } } if (IS_BRANCH(mp)) { if ((options & C_RETIRING) == 0) { const pgno_t ref = node_pgno(node); if (unlikely(ref < MIN_PAGENO || ref >= mc->mc_txn->mt_next_pgno)) rc = bad_page(mp, "branch-node wrong pgno (%u)\n", ref); } continue; } switch (node_flags(node)) { default: rc = bad_page(mp, "invalid node flags (%u)\n", node_flags(node)); break; case F_BIGDATA /* data on large-page */: case 0 /* usual */: case F_SUBDATA /* sub-db */: case F_SUBDATA | F_DUPDATA /* dupsorted sub-tree */: case F_DUPDATA /* short sub-page */: break; } if (node_flags(node) & F_BIGDATA) { const size_t dsize = node_ds(node); if ((options & C_COPYING) == 0) { if (unlikely(dsize <= mc->mc_dbx->md_vlen_min || dsize > mc->mc_dbx->md_vlen_max)) rc = bad_page( mp, "big-node data size (%zu) <> min/max value-length (%zu/%zu)\n", dsize, mc->mc_dbx->md_vlen_min, mc->mc_dbx->md_vlen_max); } if ((options & C_RETIRING) == 0) { MDBX_page *lp; int err = mdbx_page_get(mc, node_largedata_pgno(node), &lp, NULL, pp_txnid4chk(mp, mc->mc_txn)); if (unlikely(err != MDBX_SUCCESS)) return err; if (unlikely(!IS_OVERFLOW(lp))) { rc = bad_page(mp, "big-node refs to non-overflow page (%u)\n", lp->mp_pgno); continue; } if (unlikely(number_of_ovpages(env, dsize) > lp->mp_pages)) rc = bad_page(mp, "big-node size (%zu) mismatch n-pages size (%u)\n", dsize, lp->mp_pages); } continue; } const size_t dsize = node_ds(node); const char *const data = node_data(node); if (unlikely(end_of_page < data + dsize)) { rc = bad_page(mp, "node-data(%u of %u, %zu bytes) beyond (%zu) page-end\n", i, nkeys, dsize, data + dsize - end_of_page); continue; } switch (node_flags(node)) { default: /* wrong, but already handled */ continue; case 0 /* usual */: if ((options & C_COPYING) == 0) { if (unlikely(dsize < mc->mc_dbx->md_vlen_min || dsize > mc->mc_dbx->md_vlen_max)) { rc = bad_page( mp, "node-data size (%zu) <> min/max value-length (%zu/%zu)\n", dsize, mc->mc_dbx->md_vlen_min, mc->mc_dbx->md_vlen_max); continue; } } break; case F_SUBDATA /* sub-db */: if (unlikely(dsize != sizeof(MDBX_db))) { rc = bad_page(mp, "invalid sub-db record size (%zu)\n", dsize); continue; } break; case F_SUBDATA | F_DUPDATA /* dupsorted sub-tree */: if (unlikely(dsize != sizeof(MDBX_db))) { rc = bad_page(mp, "invalid nested-db record size (%zu)\n", dsize); continue; } break; case F_DUPDATA /* short sub-page */: if (unlikely(dsize <= PAGEHDRSZ)) { rc = bad_page(mp, "invalid nested-page record size (%zu)\n", dsize); continue; } else { const MDBX_page *const sp = (MDBX_page *)data; const char *const end_of_subpage = data + dsize; const int nsubkeys = page_numkeys(sp); switch (sp->mp_flags & ~P_DIRTY /* ignore for sub-pages */) { case P_LEAF | P_SUBP: case P_LEAF | P_LEAF2 | P_SUBP: break; default: rc = bad_page(mp, "invalid nested-page flags (%u)\n", sp->mp_flags); continue; } MDBX_val sub_here, sub_prev = {0, 0}; for (int j = 0; j < nsubkeys; j++) { if (IS_LEAF2(sp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */ size_t sub_ksize = sp->mp_leaf2_ksize; char *sub_key = page_leaf2key(sp, j, sub_ksize); if (unlikely(end_of_subpage < sub_key + sub_ksize)) { rc = bad_page(mp, "nested-leaf2-key beyond (%zu) nested-page\n", sub_key + sub_ksize - end_of_subpage); continue; } if ((options & C_COPYING) == 0) { if (unlikely(sub_ksize != mc->mc_dbx->md_vlen_min)) { if (unlikely(sub_ksize < mc->mc_dbx->md_vlen_min || sub_ksize > mc->mc_dbx->md_vlen_max)) { rc = bad_page(mp, "nested-leaf2-key size (%zu) <> min/max " "value-length (%zu/%zu)\n", sub_ksize, mc->mc_dbx->md_vlen_min, mc->mc_dbx->md_vlen_max); continue; } mc->mc_dbx->md_vlen_min = mc->mc_dbx->md_vlen_max = sub_ksize; } if ((options & C_SKIPORD) == 0) { sub_here.iov_len = sub_ksize; sub_here.iov_base = sub_key; if (sub_prev.iov_base && unlikely(mc->mc_dbx->md_dcmp(&sub_prev, &sub_here) >= 0)) rc = bad_page(mp, "nested-leaf2-key #%u wrong order\n", j); sub_prev = sub_here; } } } else { const MDBX_node *const sub_node = page_node(sp, j); const char *sub_node_end = (char *)sub_node + NODESIZE; if (unlikely(sub_node_end > end_of_subpage)) { rc = bad_page(mp, "nested-node beyond (%zu) nested-page\n", end_of_subpage - sub_node_end); continue; } if (unlikely(node_flags(sub_node) != 0)) rc = bad_page(mp, "nested-node invalid flags (%u)\n", node_flags(sub_node)); size_t sub_ksize = node_ks(sub_node); char *sub_key = node_key(sub_node); size_t sub_dsize = node_ds(sub_node); /* char *sub_data = node_data(sub_node); */ if ((options & C_COPYING) == 0) { if (unlikely(sub_ksize < mc->mc_dbx->md_vlen_min || sub_ksize > mc->mc_dbx->md_vlen_max)) rc = bad_page(mp, "nested-node-key size (%zu) <> min/max " "value-length (%zu/%zu)\n", sub_ksize, mc->mc_dbx->md_vlen_min, mc->mc_dbx->md_vlen_max); if ((options & C_SKIPORD) == 0) { sub_here.iov_len = sub_ksize; sub_here.iov_base = sub_key; if (sub_prev.iov_base && unlikely(mc->mc_dbx->md_dcmp(&sub_prev, &sub_here) >= 0)) rc = bad_page(mp, "nested-node-key #%u wrong order\n", j); sub_prev = sub_here; } } if (unlikely(sub_dsize != 0)) rc = bad_page(mp, "nested-node non-empty data size (%zu)\n", sub_dsize); if (unlikely(end_of_subpage < sub_key + sub_ksize)) rc = bad_page(mp, "nested-node-key beyond (%zu) nested-page\n", sub_key + sub_ksize - end_of_subpage); } } } break; } } } return rc; } static __cold int mdbx_cursor_check(MDBX_cursor *mc, unsigned options) { mdbx_tassert(mc->mc_txn, mc->mc_txn->mt_parent || mc->mc_txn->tw.dirtyroom + mc->mc_txn->tw.dirtylist->length == mc->mc_txn->mt_env->me_options.dp_limit); mdbx_cassert(mc, mc->mc_top == mc->mc_snum - 1); if (unlikely(mc->mc_top != mc->mc_snum - 1)) return MDBX_CURSOR_FULL; mdbx_cassert(mc, (options & C_UPDATING) ? mc->mc_snum <= mc->mc_db->md_depth : mc->mc_snum == mc->mc_db->md_depth); if (unlikely((options & C_UPDATING) ? mc->mc_snum > mc->mc_db->md_depth : mc->mc_snum != mc->mc_db->md_depth)) return MDBX_CURSOR_FULL; for (int n = 0; n < (int)mc->mc_snum; ++n) { MDBX_page *mp = mc->mc_pg[n]; const unsigned nkeys = page_numkeys(mp); const bool expect_branch = (n < mc->mc_db->md_depth - 1) ? true : false; const bool expect_nested_leaf = (n + 1 == mc->mc_db->md_depth - 1) ? true : false; const bool branch = IS_BRANCH(mp) ? true : false; mdbx_cassert(mc, branch == expect_branch); if (unlikely(branch != expect_branch)) return MDBX_CURSOR_FULL; if ((options & C_UPDATING) == 0) { mdbx_cassert(mc, nkeys > mc->mc_ki[n] || (!branch && nkeys == mc->mc_ki[n] && (mc->mc_flags & C_EOF) != 0)); if (unlikely(nkeys <= mc->mc_ki[n] && !(!branch && nkeys == mc->mc_ki[n] && (mc->mc_flags & C_EOF) != 0))) return MDBX_CURSOR_FULL; } else { mdbx_cassert(mc, nkeys + 1 >= mc->mc_ki[n]); if (unlikely(nkeys + 1 < mc->mc_ki[n])) return MDBX_CURSOR_FULL; } int err = mdbx_page_check(mc, mp, options); if (unlikely(err != MDBX_SUCCESS)) return err; for (unsigned i = 0; i < nkeys; ++i) { if (branch) { MDBX_node *node = page_node(mp, i); mdbx_cassert(mc, node_flags(node) == 0); if (unlikely(node_flags(node) != 0)) return MDBX_CURSOR_FULL; pgno_t pgno = node_pgno(node); MDBX_page *np; int rc = mdbx_page_get(mc, pgno, &np, NULL, pp_txnid4chk(mp, mc->mc_txn)); mdbx_cassert(mc, rc == MDBX_SUCCESS); if (unlikely(rc != MDBX_SUCCESS)) return rc; const bool nested_leaf = IS_LEAF(np) ? true : false; mdbx_cassert(mc, nested_leaf == expect_nested_leaf); if (unlikely(nested_leaf != expect_nested_leaf)) return MDBX_CURSOR_FULL; err = mdbx_page_check(mc, np, options); if (unlikely(err != MDBX_SUCCESS)) return err; } } } return MDBX_SUCCESS; } /* Complete a delete operation started by mdbx_cursor_del(). */ static int mdbx_cursor_del0(MDBX_cursor *mc) { int rc; MDBX_page *mp; indx_t ki; unsigned nkeys; MDBX_dbi dbi = mc->mc_dbi; mdbx_cassert(mc, cursor_is_tracked(mc)); mdbx_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); ki = mc->mc_ki[mc->mc_top]; mp = mc->mc_pg[mc->mc_top]; mdbx_node_del(mc, mc->mc_db->md_xsize); mc->mc_db->md_entries--; /* Adjust other cursors pointing to mp */ for (MDBX_cursor *m2 = mc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { MDBX_cursor *m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == mc || !(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) continue; if (m3->mc_snum < mc->mc_snum) continue; if (m3->mc_pg[mc->mc_top] == mp) { if (m3->mc_ki[mc->mc_top] == ki) { m3->mc_flags |= C_DEL; if (mc->mc_db->md_flags & MDBX_DUPSORT) { /* Sub-cursor referred into dataset which is gone */ m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED | C_EOF); } continue; } else if (m3->mc_ki[mc->mc_top] > ki) { m3->mc_ki[mc->mc_top]--; } if (XCURSOR_INITED(m3)) XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]); } } rc = mdbx_rebalance(mc); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; if (unlikely(!mc->mc_snum)) { /* DB is totally empty now, just bail out. * Other cursors adjustments were already done * by mdbx_rebalance and aren't needed here. */ mdbx_cassert(mc, mc->mc_db->md_entries == 0 && mc->mc_db->md_depth == 0 && mc->mc_db->md_root == P_INVALID); mc->mc_flags |= C_EOF; return MDBX_SUCCESS; } ki = mc->mc_ki[mc->mc_top]; mp = mc->mc_pg[mc->mc_top]; mdbx_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); nkeys = page_numkeys(mp); mdbx_cassert(mc, (mc->mc_db->md_entries > 0 && nkeys > 0) || ((mc->mc_flags & C_SUB) && mc->mc_db->md_entries == 0 && nkeys == 0)); /* Adjust this and other cursors pointing to mp */ for (MDBX_cursor *m2 = mc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { MDBX_cursor *m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) continue; if (m3->mc_snum < mc->mc_snum) continue; if (m3->mc_pg[mc->mc_top] == mp) { /* if m3 points past last node in page, find next sibling */ if (m3->mc_ki[mc->mc_top] >= nkeys) { rc = mdbx_cursor_sibling(m3, SIBLING_RIGHT); if (rc == MDBX_NOTFOUND) { m3->mc_flags |= C_EOF; rc = MDBX_SUCCESS; continue; } if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } if (m3->mc_ki[mc->mc_top] >= ki || /* moved to right sibling */ m3->mc_pg[mc->mc_top] != mp) { if (m3->mc_xcursor && !(m3->mc_flags & C_EOF)) { MDBX_node *node = page_node(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]); /* If this node has dupdata, it may need to be reinited * because its data has moved. * If the xcursor was not inited it must be reinited. * Else if node points to a subDB, nothing is needed. */ if (node_flags(node) & F_DUPDATA) { if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { if (!(node_flags(node) & F_SUBDATA)) m3->mc_xcursor->mx_cursor.mc_pg[0] = node_data(node); } else { rc = mdbx_xcursor_init1(m3, node, m3->mc_pg[m3->mc_top]); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; rc = mdbx_cursor_first(&m3->mc_xcursor->mx_cursor, NULL, NULL); if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } } m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL; } m3->mc_flags |= C_DEL; } } } mdbx_cassert(mc, rc == MDBX_SUCCESS); if (mdbx_audit_enabled()) rc = mdbx_cursor_check(mc, 0); return rc; bailout: mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return rc; } int mdbx_del(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, const MDBX_val *data) { int rc = check_txn_rw(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(txn->mt_flags & (MDBX_TXN_RDONLY | MDBX_TXN_BLOCKED))) return (txn->mt_flags & MDBX_TXN_RDONLY) ? MDBX_EACCESS : MDBX_BAD_TXN; return mdbx_del0(txn, dbi, key, data, 0); } static int mdbx_del0(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, const MDBX_val *data, unsigned flags) { MDBX_cursor_couple cx; MDBX_cursor_op op; MDBX_val rdata; int rc, exact = 0; DKBUF; mdbx_debug("====> delete db %u key [%s], data [%s]", dbi, DKEY(key), DVAL(data)); rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (data) { op = MDBX_GET_BOTH; rdata = *data; data = &rdata; } else { op = MDBX_SET; flags |= MDBX_ALLDUPS; } rc = mdbx_cursor_set(&cx.outer, (MDBX_val *)key, (MDBX_val *)data, op, &exact); if (likely(rc == MDBX_SUCCESS)) { /* let mdbx_page_split know about this cursor if needed: * delete will trigger a rebalance; if it needs to move * a node from one page to another, it will have to * update the parent's separator key(s). If the new sepkey * is larger than the current one, the parent page may * run out of space, triggering a split. We need this * cursor to be consistent until the end of the rebalance. */ cx.outer.mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = &cx.outer; rc = mdbx_cursor_del(&cx.outer, flags); txn->tw.cursors[dbi] = cx.outer.mc_next; } return rc; } /* Split a page and insert a new node. * Set MDBX_TXN_ERROR on failure. * [in,out] mc Cursor pointing to the page and desired insertion index. * The cursor will be updated to point to the actual page and index where * the node got inserted after the split. * [in] newkey The key for the newly inserted node. * [in] newdata The data for the newly inserted node. * [in] newpgno The page number, if the new node is a branch node. * [in] nflags The NODE_ADD_FLAGS for the new node. * Returns 0 on success, non-zero on failure. */ static int mdbx_page_split(MDBX_cursor *mc, const MDBX_val *newkey, MDBX_val *newdata, pgno_t newpgno, unsigned nflags) { unsigned flags; int rc = MDBX_SUCCESS, foliage = 0, did_split = 0; pgno_t pgno = 0; unsigned i, ptop; MDBX_env *env = mc->mc_txn->mt_env; MDBX_val sepkey, rkey, xdata; MDBX_page *copy = NULL; MDBX_page *rp, *pp; MDBX_cursor mn; DKBUF; MDBX_page *mp = mc->mc_pg[mc->mc_top]; unsigned newindx = mc->mc_ki[mc->mc_top]; unsigned nkeys = page_numkeys(mp); if (mdbx_audit_enabled()) { rc = mdbx_cursor_check(mc, C_UPDATING); if (unlikely(rc != MDBX_SUCCESS)) return rc; } mdbx_cassert(mc, nkeys >= (unsigned)(IS_BRANCH(mp) ? MDBX_MINKEYS * 2 - 1 : 1)); mdbx_debug("-----> splitting %s page %" PRIaPGNO " and adding [%s] at index %i/%i", IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno, DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys); /* Create a right sibling. */ if ((rc = mdbx_page_new(mc, mp->mp_flags, 1, &rp))) return rc; rp->mp_leaf2_ksize = mp->mp_leaf2_ksize; mdbx_debug("new right sibling: page %" PRIaPGNO, rp->mp_pgno); /* Usually when splitting the root page, the cursor * height is 1. But when called from mdbx_update_key, * the cursor height may be greater because it walks * up the stack while finding the branch slot to update. */ if (mc->mc_top < 1) { if ((rc = mdbx_page_new(mc, P_BRANCH, 1, &pp))) goto done; /* shift current top to make room for new parent */ mdbx_cassert(mc, mc->mc_snum < 2 && mc->mc_db->md_depth > 0); mc->mc_pg[2] = mc->mc_pg[1]; mc->mc_ki[2] = mc->mc_ki[1]; mc->mc_pg[1] = mc->mc_pg[0]; mc->mc_ki[1] = mc->mc_ki[0]; mc->mc_pg[0] = pp; mc->mc_ki[0] = 0; mc->mc_db->md_root = pp->mp_pgno; mdbx_debug("root split! new root = %" PRIaPGNO, pp->mp_pgno); foliage = mc->mc_db->md_depth++; /* Add left (implicit) pointer. */ if (unlikely((rc = mdbx_node_add_branch(mc, 0, NULL, mp->mp_pgno)) != MDBX_SUCCESS)) { /* undo the pre-push */ mc->mc_pg[0] = mc->mc_pg[1]; mc->mc_ki[0] = mc->mc_ki[1]; mc->mc_db->md_root = mp->mp_pgno; mc->mc_db->md_depth--; goto done; } mc->mc_snum++; mc->mc_top++; ptop = 0; } else { ptop = mc->mc_top - 1; mdbx_debug("parent branch page is %" PRIaPGNO, mc->mc_pg[ptop]->mp_pgno); } cursor_copy_internal(mc, &mn); mn.mc_xcursor = NULL; mn.mc_pg[mn.mc_top] = rp; mn.mc_ki[mn.mc_top] = 0; mn.mc_ki[ptop] = mc->mc_ki[ptop] + 1; unsigned split_indx; if (nflags & MDBX_APPEND) { mn.mc_ki[mn.mc_top] = 0; sepkey = *newkey; split_indx = newindx; nkeys = 0; } else { split_indx = (nkeys + 1) / 2; if (IS_LEAF2(rp)) { char *split, *ins; unsigned lsize, rsize, ksize; /* Move half of the keys to the right sibling */ const int x = mc->mc_ki[mc->mc_top] - split_indx; ksize = mc->mc_db->md_xsize; split = page_leaf2key(mp, split_indx, ksize); rsize = (nkeys - split_indx) * ksize; lsize = (nkeys - split_indx) * sizeof(indx_t); mdbx_cassert(mc, mp->mp_lower >= lsize); mp->mp_lower -= (indx_t)lsize; mdbx_cassert(mc, rp->mp_lower + lsize <= UINT16_MAX); rp->mp_lower += (indx_t)lsize; mdbx_cassert(mc, mp->mp_upper + rsize - lsize <= UINT16_MAX); mp->mp_upper += (indx_t)(rsize - lsize); mdbx_cassert(mc, rp->mp_upper >= rsize - lsize); rp->mp_upper -= (indx_t)(rsize - lsize); sepkey.iov_len = ksize; if (newindx == split_indx) { sepkey.iov_base = newkey->iov_base; } else { sepkey.iov_base = split; } if (x < 0) { mdbx_cassert(mc, ksize >= sizeof(indx_t)); ins = page_leaf2key(mp, mc->mc_ki[mc->mc_top], ksize); memcpy(rp->mp_ptrs, split, rsize); sepkey.iov_base = rp->mp_ptrs; memmove(ins + ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize); memcpy(ins, newkey->iov_base, ksize); mdbx_cassert(mc, UINT16_MAX - mp->mp_lower >= (int)sizeof(indx_t)); mp->mp_lower += sizeof(indx_t); mdbx_cassert(mc, mp->mp_upper >= ksize - sizeof(indx_t)); mp->mp_upper -= (indx_t)(ksize - sizeof(indx_t)); } else { memcpy(rp->mp_ptrs, split, x * ksize); ins = page_leaf2key(rp, x, ksize); memcpy(ins, newkey->iov_base, ksize); memcpy(ins + ksize, split + x * ksize, rsize - x * ksize); mdbx_cassert(mc, UINT16_MAX - rp->mp_lower >= (int)sizeof(indx_t)); rp->mp_lower += sizeof(indx_t); mdbx_cassert(mc, rp->mp_upper >= ksize - sizeof(indx_t)); rp->mp_upper -= (indx_t)(ksize - sizeof(indx_t)); mdbx_cassert(mc, x <= (int)UINT16_MAX); mc->mc_ki[mc->mc_top] = (indx_t)x; } } else { /* Maximum free space in an empty page */ const unsigned pmax = page_space(env); const size_t nsize = IS_LEAF(mp) ? leaf_size(env, newkey, newdata) : branch_size(env, newkey); /* grab a page to hold a temporary copy */ copy = mdbx_page_malloc(mc->mc_txn, 1); if (unlikely(copy == NULL)) { rc = MDBX_ENOMEM; goto done; } copy->mp_pgno = mp->mp_pgno; copy->mp_flags = mp->mp_flags; copy->mp_txnid = INVALID_TXNID; copy->mp_lower = 0; copy->mp_upper = (indx_t)page_space(env); /* prepare to insert */ for (unsigned j = i = 0; i < nkeys; i++) { if (i == newindx) copy->mp_ptrs[j++] = 0; copy->mp_ptrs[j++] = mp->mp_ptrs[i]; } /* When items are relatively large the split point needs * to be checked, because being off-by-one will make the * difference between success or failure in mdbx_node_add. * * It's also relevant if a page happens to be laid out * such that one half of its nodes are all "small" and * the other half of its nodes are "large." If the new * item is also "large" and falls on the half with * "large" nodes, it also may not fit. * * As a final tweak, if the new item goes on the last * spot on the page (and thus, onto the new page), bias * the split so the new page is emptier than the old page. * This yields better packing during sequential inserts. */ if (nkeys < 32 || nsize > pmax / 16 || newindx >= nkeys) { /* Find split point */ int dir; size_t psize = 0; unsigned k; if (newindx <= split_indx || newindx >= nkeys) { i = 0; dir = 1; k = (newindx >= nkeys) ? nkeys : split_indx + 1 + IS_LEAF(mp); split_indx = k - 1; } else { i = nkeys; dir = -1; k = split_indx - 1; split_indx += 1; } do { if (i == newindx) { psize += nsize; } else { MDBX_node *node = (MDBX_node *)((char *)mp + copy->mp_ptrs[i] + PAGEHDRSZ); psize += NODESIZE + node_ks(node) + sizeof(indx_t); if (IS_LEAF(mp)) psize += F_ISSET(node_flags(node), F_BIGDATA) ? sizeof(pgno_t) : node_ds(node); psize = EVEN(psize); } if (psize > pmax) { split_indx = i + (dir < 0); break; } i += dir; } while (i != k); } if (split_indx == newindx) { sepkey.iov_len = newkey->iov_len; sepkey.iov_base = newkey->iov_base; } else { MDBX_node *node = (MDBX_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEHDRSZ); sepkey.iov_len = node_ks(node); sepkey.iov_base = node_key(node); } } } mdbx_debug("separator is %d [%s]", split_indx, DKEY(&sepkey)); if (mdbx_audit_enabled()) { rc = mdbx_cursor_check(mc, C_UPDATING); if (unlikely(rc != MDBX_SUCCESS)) goto done; rc = mdbx_cursor_check(&mn, C_UPDATING); if (unlikely(rc != MDBX_SUCCESS)) goto done; } /* Copy separator key to the parent. */ if (page_room(mn.mc_pg[ptop]) < branch_size(env, &sepkey)) { const int snum = mc->mc_snum; const int depth = mc->mc_db->md_depth; mn.mc_snum--; mn.mc_top--; did_split = 1; /* We want other splits to find mn when doing fixups */ WITH_CURSOR_TRACKING( mn, rc = mdbx_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0)); if (unlikely(rc != MDBX_SUCCESS)) goto done; mdbx_cassert(mc, (int)mc->mc_snum - snum == mc->mc_db->md_depth - depth); if (mdbx_audit_enabled()) { rc = mdbx_cursor_check(mc, C_UPDATING); if (unlikely(rc != MDBX_SUCCESS)) goto done; } /* root split? */ ptop += mc->mc_snum - snum; /* Right page might now have changed parent. * Check if left page also changed parent. */ if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && mc->mc_ki[ptop] >= page_numkeys(mc->mc_pg[ptop])) { for (i = 0; i < ptop; i++) { mc->mc_pg[i] = mn.mc_pg[i]; mc->mc_ki[i] = mn.mc_ki[i]; } mc->mc_pg[ptop] = mn.mc_pg[ptop]; if (mn.mc_ki[ptop]) { mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1; } else { /* find right page's left sibling */ mc->mc_ki[ptop] = mn.mc_ki[ptop]; rc = mdbx_cursor_sibling(mc, SIBLING_LEFT); } } } else { mn.mc_top--; rc = mdbx_node_add_branch(&mn, mn.mc_ki[ptop], &sepkey, rp->mp_pgno); mn.mc_top++; } if (unlikely(rc != MDBX_SUCCESS)) { if (rc == MDBX_NOTFOUND) /* improper mdbx_cursor_sibling() result */ { mdbx_error("unexpected %s", "MDBX_NOTFOUND"); rc = MDBX_PROBLEM; } goto done; } if (nflags & MDBX_APPEND) { mc->mc_pg[mc->mc_top] = rp; mc->mc_ki[mc->mc_top] = 0; switch (PAGETYPE(rp)) { case P_BRANCH: { mdbx_cassert(mc, (nflags & (F_BIGDATA | F_SUBDATA | F_DUPDATA)) == 0); mdbx_cassert(mc, newpgno != 0 && newpgno != P_INVALID); rc = mdbx_node_add_branch(mc, 0, newkey, newpgno); } break; case P_LEAF: { mdbx_cassert(mc, newpgno == 0 || newpgno == P_INVALID); rc = mdbx_node_add_leaf(mc, 0, newkey, newdata, nflags); } break; case P_LEAF | P_LEAF2: { mdbx_cassert(mc, (nflags & (F_BIGDATA | F_SUBDATA | F_DUPDATA)) == 0); mdbx_cassert(mc, newpgno == 0 || newpgno == P_INVALID); rc = mdbx_node_add_leaf2(mc, 0, newkey); } break; default: rc = bad_page(rp, "wrong page-type %u\n", PAGETYPE(rp)); } if (rc) goto done; for (i = 0; i < mc->mc_top; i++) mc->mc_ki[i] = mn.mc_ki[i]; } else if (!IS_LEAF2(mp)) { /* Move nodes */ mc->mc_pg[mc->mc_top] = rp; i = split_indx; indx_t n = 0; do { MDBX_val *rdata = NULL; if (i == newindx) { rkey.iov_base = newkey->iov_base; rkey.iov_len = newkey->iov_len; if (IS_LEAF(mp)) { rdata = newdata; } else pgno = newpgno; flags = nflags; /* Update index for the new key. */ mc->mc_ki[mc->mc_top] = n; } else { MDBX_node *node = (MDBX_node *)((char *)mp + copy->mp_ptrs[i] + PAGEHDRSZ); rkey.iov_base = node_key(node); rkey.iov_len = node_ks(node); if (IS_LEAF(mp)) { xdata.iov_base = node_data(node); xdata.iov_len = node_ds(node); rdata = &xdata; } else pgno = node_pgno(node); flags = node_flags(node); } switch (PAGETYPE(rp)) { case P_BRANCH: { mdbx_cassert(mc, 0 == (uint16_t)flags); if (n == 0) { /* First branch index doesn't need key data. */ rkey.iov_len = 0; } rc = mdbx_node_add_branch(mc, n, &rkey, pgno); } break; case P_LEAF: { mdbx_cassert(mc, pgno == 0); mdbx_cassert(mc, rdata != NULL); rc = mdbx_node_add_leaf(mc, n, &rkey, rdata, flags); } break; /* case P_LEAF | P_LEAF2: { mdbx_cassert(mc, (nflags & (F_BIGDATA | F_SUBDATA | F_DUPDATA)) == 0); mdbx_cassert(mc, gno == 0); rc = mdbx_node_add_leaf2(mc, n, &rkey); } break; */ default: rc = bad_page(rp, "wrong page-type %u\n", PAGETYPE(rp)); } if (rc) goto done; if (i == nkeys) { i = 0; n = 0; mc->mc_pg[mc->mc_top] = copy; } else { i++; n++; } } while (i != split_indx); nkeys = page_numkeys(copy); for (i = 0; i < nkeys; i++) mp->mp_ptrs[i] = copy->mp_ptrs[i]; mp->mp_lower = copy->mp_lower; mp->mp_upper = copy->mp_upper; memcpy(page_node(mp, nkeys - 1), page_node(copy, nkeys - 1), env->me_psize - copy->mp_upper - PAGEHDRSZ); /* reset back to original page */ if (newindx < split_indx) { mc->mc_pg[mc->mc_top] = mp; } else { mc->mc_pg[mc->mc_top] = rp; mc->mc_ki[ptop]++; /* Make sure mc_ki is still valid. */ if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && mc->mc_ki[ptop] >= page_numkeys(mc->mc_pg[ptop])) { for (i = 0; i <= ptop; i++) { mc->mc_pg[i] = mn.mc_pg[i]; mc->mc_ki[i] = mn.mc_ki[i]; } } } if (nflags & MDBX_RESERVE) { MDBX_node *node = page_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (!(node_flags(node) & F_BIGDATA)) newdata->iov_base = node_data(node); } } else { if (newindx >= split_indx) { mc->mc_pg[mc->mc_top] = rp; mc->mc_ki[ptop]++; /* Make sure mc_ki is still valid. */ if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && mc->mc_ki[ptop] >= page_numkeys(mc->mc_pg[ptop])) { for (i = 0; i <= ptop; i++) { mc->mc_pg[i] = mn.mc_pg[i]; mc->mc_ki[i] = mn.mc_ki[i]; } } } } { /* Adjust other cursors pointing to mp */ MDBX_cursor *m2, *m3; MDBX_dbi dbi = mc->mc_dbi; nkeys = page_numkeys(mp); for (m2 = mc->mc_txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) { m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; if (m3 == mc) continue; if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) continue; if (foliage) { int k; /* sub cursors may be on different DB */ if (m3->mc_pg[0] != mp) continue; /* root split */ for (k = foliage; k >= 0; k--) { m3->mc_ki[k + 1] = m3->mc_ki[k]; m3->mc_pg[k + 1] = m3->mc_pg[k]; } m3->mc_ki[0] = (m3->mc_ki[0] >= nkeys) ? 1 : 0; m3->mc_pg[0] = mc->mc_pg[0]; m3->mc_snum++; m3->mc_top++; } if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) { if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDBX_SPLIT_REPLACE)) m3->mc_ki[mc->mc_top]++; if (m3->mc_ki[mc->mc_top] >= nkeys) { m3->mc_pg[mc->mc_top] = rp; mdbx_cassert(mc, m3->mc_ki[mc->mc_top] >= nkeys); m3->mc_ki[mc->mc_top] -= (indx_t)nkeys; for (i = 0; i < mc->mc_top; i++) { m3->mc_ki[i] = mn.mc_ki[i]; m3->mc_pg[i] = mn.mc_pg[i]; } } } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] && m3->mc_ki[ptop] >= mc->mc_ki[ptop]) { m3->mc_ki[ptop]++; } if (XCURSOR_INITED(m3) && IS_LEAF(mp)) XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]); } } mdbx_debug("mp left: %d, rp left: %d", page_room(mp), page_room(rp)); done: if (copy) /* tmp page */ mdbx_dpage_free(env, copy, 1); if (unlikely(rc)) mc->mc_txn->mt_flags |= MDBX_TXN_ERROR; return rc; } int mdbx_put(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, MDBX_val *data, unsigned flags) { int rc = check_txn_rw(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key || !data)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(flags & ~(MDBX_NOOVERWRITE | MDBX_NODUPDATA | MDBX_ALLDUPS | MDBX_ALLDUPS | MDBX_RESERVE | MDBX_APPEND | MDBX_APPENDDUP | MDBX_CURRENT | MDBX_MULTIPLE))) return MDBX_EINVAL; if (unlikely(txn->mt_flags & (MDBX_TXN_RDONLY | MDBX_TXN_BLOCKED))) return (txn->mt_flags & MDBX_TXN_RDONLY) ? MDBX_EACCESS : MDBX_BAD_TXN; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; cx.outer.mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = &cx.outer; /* LY: support for update (explicit overwrite) */ if (flags & MDBX_CURRENT) { rc = mdbx_cursor_get(&cx.outer, (MDBX_val *)key, NULL, MDBX_SET); if (likely(rc == MDBX_SUCCESS) && (txn->mt_dbs[dbi].md_flags & MDBX_DUPSORT) && (flags & MDBX_ALLDUPS) == 0) { /* LY: allows update (explicit overwrite) only for unique keys */ MDBX_node *node = page_node(cx.outer.mc_pg[cx.outer.mc_top], cx.outer.mc_ki[cx.outer.mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { mdbx_tassert(txn, XCURSOR_INITED(&cx.outer) && cx.outer.mc_xcursor->mx_db.md_entries > 1); rc = MDBX_EMULTIVAL; } } } if (likely(rc == MDBX_SUCCESS)) rc = mdbx_cursor_put(&cx.outer, key, data, flags); txn->tw.cursors[dbi] = cx.outer.mc_next; return rc; } /**** COPYING *****************************************************************/ #ifndef MDBX_WBUF #define MDBX_WBUF ((size_t)1024 * 1024) #endif /* State needed for a double-buffering compacting copy. */ typedef struct mdbx_copy { MDBX_env *mc_env; MDBX_txn *mc_txn; mdbx_condpair_t mc_condpair; uint8_t *mc_wbuf[2]; uint8_t *mc_over[2]; size_t mc_wlen[2]; size_t mc_olen[2]; mdbx_filehandle_t mc_fd; /* Error code. Never cleared if set. Both threads can set nonzero * to fail the copy. Not mutex-protected, MDBX expects atomic int. */ volatile int mc_error; pgno_t mc_next_pgno; volatile unsigned mc_head; volatile unsigned mc_tail; } mdbx_copy; /* Dedicated writer thread for compacting copy. */ static THREAD_RESULT __cold THREAD_CALL mdbx_env_copythr(void *arg) { mdbx_copy *my = arg; #if defined(EPIPE) && !(defined(_WIN32) || defined(_WIN64)) sigset_t sigset; sigemptyset(&sigset); sigaddset(&sigset, SIGPIPE); my->mc_error = pthread_sigmask(SIG_BLOCK, &sigset, NULL); #endif /* EPIPE */ mdbx_condpair_lock(&my->mc_condpair); while (!my->mc_error) { while (my->mc_tail == my->mc_head && !my->mc_error) { int err = mdbx_condpair_wait(&my->mc_condpair, true); if (err != MDBX_SUCCESS) { my->mc_error = err; goto bailout; } } const unsigned toggle = my->mc_tail & 1; size_t wsize = my->mc_wlen[toggle]; if (wsize == 0) { my->mc_tail += 1; break /* EOF */; } my->mc_wlen[toggle] = 0; uint8_t *ptr = my->mc_wbuf[toggle]; again: if (!my->mc_error) { int err = mdbx_write(my->mc_fd, ptr, wsize); if (err != MDBX_SUCCESS) { #if defined(EPIPE) && !(defined(_WIN32) || defined(_WIN64)) if (err == EPIPE) { /* Collect the pending SIGPIPE, * otherwise at least OS X gives it to the process on thread-exit. */ int unused; sigwait(&sigset, &unused); } #endif /* EPIPE */ my->mc_error = err; goto bailout; } } /* If there's an overflow page tail, write it too */ wsize = my->mc_olen[toggle]; if (wsize) { my->mc_olen[toggle] = 0; ptr = my->mc_over[toggle]; goto again; } my->mc_tail += 1; mdbx_condpair_signal(&my->mc_condpair, false); } bailout: mdbx_condpair_unlock(&my->mc_condpair); return (THREAD_RESULT)0; } /* Give buffer and/or MDBX_EOF to writer thread, await unused buffer. */ static __cold int mdbx_env_cthr_toggle(mdbx_copy *my) { mdbx_condpair_lock(&my->mc_condpair); mdbx_assert(my->mc_env, my->mc_head - my->mc_tail < 2 || my->mc_error); my->mc_head += 1; mdbx_condpair_signal(&my->mc_condpair, true); while (!my->mc_error && my->mc_head - my->mc_tail == 2 /* both buffers in use */) { int err = mdbx_condpair_wait(&my->mc_condpair, false); if (err != MDBX_SUCCESS) my->mc_error = err; } mdbx_condpair_unlock(&my->mc_condpair); return my->mc_error; } /* Depth-first tree traversal for compacting copy. * [in] my control structure. * [in,out] pg database root. * [in] flags includes F_DUPDATA if it is a sorted-duplicate sub-DB. */ static __cold int mdbx_env_cwalk(mdbx_copy *my, pgno_t *pg, int flags) { MDBX_cursor_couple couple; MDBX_page *mo, *mp, *leaf; char *buf, *ptr; int rc; unsigned i; /* Empty DB, nothing to do */ if (*pg == P_INVALID) return MDBX_SUCCESS; memset(&couple, 0, sizeof(couple)); couple.outer.mc_snum = 1; couple.outer.mc_txn = my->mc_txn; couple.outer.mc_flags = couple.inner.mx_cursor.mc_flags = C_COPYING | C_SKIPORD; rc = mdbx_page_get(&couple.outer, *pg, &couple.outer.mc_pg[0], NULL, my->mc_txn->mt_txnid); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_page_search_root(&couple.outer, NULL, MDBX_PS_FIRST); if (unlikely(rc != MDBX_SUCCESS)) return rc; /* Make cursor pages writable */ buf = ptr = mdbx_malloc(pgno2bytes(my->mc_env, couple.outer.mc_snum)); if (buf == NULL) return MDBX_ENOMEM; for (i = 0; i < couple.outer.mc_top; i++) { mdbx_page_copy((MDBX_page *)ptr, couple.outer.mc_pg[i], my->mc_env->me_psize); couple.outer.mc_pg[i] = (MDBX_page *)ptr; ptr += my->mc_env->me_psize; } /* This is writable space for a leaf page. Usually not needed. */ leaf = (MDBX_page *)ptr; while (couple.outer.mc_snum > 0) { mp = couple.outer.mc_pg[couple.outer.mc_top]; unsigned n = page_numkeys(mp); if (IS_LEAF(mp)) { if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) { for (i = 0; i < n; i++) { MDBX_node *node = page_node(mp, i); if (node_flags(node) & F_BIGDATA) { MDBX_page *omp; /* Need writable leaf */ if (mp != leaf) { couple.outer.mc_pg[couple.outer.mc_top] = leaf; mdbx_page_copy(leaf, mp, my->mc_env->me_psize); mp = leaf; node = page_node(mp, i); } const pgno_t pgno = node_largedata_pgno(node); poke_pgno(node_data(node), my->mc_next_pgno); rc = mdbx_page_get(&couple.outer, pgno, &omp, NULL, pp_txnid4chk(mp, my->mc_txn)); if (unlikely(rc != MDBX_SUCCESS)) goto done; unsigned toggle = my->mc_head & 1; if (my->mc_wlen[toggle] + my->mc_env->me_psize > MDBX_WBUF) { rc = mdbx_env_cthr_toggle(my); if (unlikely(rc != MDBX_SUCCESS)) goto done; toggle = my->mc_head & 1; } mo = (MDBX_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); memcpy(mo, omp, my->mc_env->me_psize); mo->mp_pgno = my->mc_next_pgno; my->mc_next_pgno += omp->mp_pages; my->mc_wlen[toggle] += my->mc_env->me_psize; if (omp->mp_pages > 1) { my->mc_olen[toggle] = pgno2bytes(my->mc_env, omp->mp_pages - 1); my->mc_over[toggle] = (uint8_t *)omp + my->mc_env->me_psize; rc = mdbx_env_cthr_toggle(my); if (unlikely(rc != MDBX_SUCCESS)) goto done; toggle = my->mc_head & 1; } } else if (node_flags(node) & F_SUBDATA) { if (node_ds(node) != sizeof(MDBX_db)) { rc = MDBX_CORRUPTED; goto done; } /* Need writable leaf */ if (mp != leaf) { couple.outer.mc_pg[couple.outer.mc_top] = leaf; mdbx_page_copy(leaf, mp, my->mc_env->me_psize); mp = leaf; node = page_node(mp, i); } MDBX_db db; memcpy(&db, node_data(node), sizeof(MDBX_db)); rc = mdbx_env_cwalk(my, &db.md_root, node_flags(node) & F_DUPDATA); if (rc) goto done; memcpy(node_data(node), &db, sizeof(MDBX_db)); } } } } else { couple.outer.mc_ki[couple.outer.mc_top]++; if (couple.outer.mc_ki[couple.outer.mc_top] < n) { again: rc = mdbx_page_get( &couple.outer, node_pgno(page_node(mp, couple.outer.mc_ki[couple.outer.mc_top])), &mp, NULL, pp_txnid4chk(mp, my->mc_txn)); if (unlikely(rc != MDBX_SUCCESS)) goto done; couple.outer.mc_top++; couple.outer.mc_snum++; couple.outer.mc_ki[couple.outer.mc_top] = 0; if (IS_BRANCH(mp)) { /* Whenever we advance to a sibling branch page, * we must proceed all the way down to its first leaf. */ mdbx_page_copy(couple.outer.mc_pg[couple.outer.mc_top], mp, my->mc_env->me_psize); goto again; } else couple.outer.mc_pg[couple.outer.mc_top] = mp; continue; } } unsigned toggle = my->mc_head & 1; if (my->mc_wlen[toggle] + my->mc_wlen[toggle] > MDBX_WBUF) { rc = mdbx_env_cthr_toggle(my); if (unlikely(rc != MDBX_SUCCESS)) goto done; toggle = my->mc_head & 1; } mo = (MDBX_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); mdbx_page_copy(mo, mp, my->mc_env->me_psize); mo->mp_pgno = my->mc_next_pgno++; my->mc_wlen[toggle] += my->mc_env->me_psize; if (couple.outer.mc_top) { /* Update parent if there is one */ node_set_pgno(page_node(couple.outer.mc_pg[couple.outer.mc_top - 1], couple.outer.mc_ki[couple.outer.mc_top - 1]), mo->mp_pgno); mdbx_cursor_pop(&couple.outer); } else { /* Otherwise we're done */ *pg = mo->mp_pgno; break; } } done: mdbx_free(buf); return rc; } static __cold void compact_fixup_meta(MDBX_env *env, MDBX_meta *meta) { /* Calculate filesize taking in account shrink/growing thresholds */ if (meta->mm_geo.next != meta->mm_geo.now) { meta->mm_geo.now = meta->mm_geo.next; const pgno_t aligner = meta->mm_geo.grow ? meta->mm_geo.grow : meta->mm_geo.shrink; if (aligner) { const pgno_t aligned = pgno_align2os_pgno( env, meta->mm_geo.next + aligner - meta->mm_geo.next % aligner); meta->mm_geo.now = aligned; } } if (meta->mm_geo.now < meta->mm_geo.lower) meta->mm_geo.now = meta->mm_geo.lower; if (meta->mm_geo.now > meta->mm_geo.upper) meta->mm_geo.now = meta->mm_geo.upper; /* Update signature */ assert(meta->mm_geo.now >= meta->mm_geo.next); unaligned_poke_u64(4, meta->mm_datasync_sign, mdbx_meta_sign(meta)); } /* Make resizeable */ static __cold void make_sizeable(MDBX_meta *meta) { meta->mm_geo.lower = MIN_PAGENO; if (meta->mm_geo.grow == 0) { const size_t step = 1 + (meta->mm_geo.upper - meta->mm_geo.lower) / 42; meta->mm_geo.grow = (step < UINT16_MAX) ? (uint16_t)step : UINT16_MAX; } if (meta->mm_geo.shrink == 0) { const size_t step = meta->mm_geo.grow + meta->mm_geo.grow; meta->mm_geo.shrink = (step < UINT16_MAX) ? (uint16_t)step : UINT16_MAX; } } /* Copy environment with compaction. */ static __cold int mdbx_env_compact(MDBX_env *env, MDBX_txn *read_txn, mdbx_filehandle_t fd, uint8_t *buffer, const bool dest_is_pipe, const int flags) { const size_t meta_bytes = pgno2bytes(env, NUM_METAS); uint8_t *const data_buffer = buffer + ceil_powerof2(meta_bytes, env->me_os_psize); MDBX_meta *const meta = mdbx_init_metas(env, buffer); mdbx_meta_set_txnid(env, meta, read_txn->mt_txnid); if (flags & MDBX_CP_FORCE_DYNAMIC_SIZE) make_sizeable(meta); /* copy canary sequences if present */ if (read_txn->mt_canary.v) { meta->mm_canary = read_txn->mt_canary; meta->mm_canary.v = mdbx_meta_txnid_stable(env, meta); } /* Set metapage 1 with current main DB */ pgno_t new_root, root = read_txn->mt_dbs[MAIN_DBI].md_root; if ((new_root = root) == P_INVALID) { /* When the DB is empty, handle it specially to * fix any breakage like page leaks from ITS#8174. */ meta->mm_dbs[MAIN_DBI].md_flags = read_txn->mt_dbs[MAIN_DBI].md_flags; compact_fixup_meta(env, meta); if (dest_is_pipe) { int rc = mdbx_write(fd, buffer, meta_bytes); if (rc != MDBX_SUCCESS) return rc; } } else { /* Count free pages + GC pages. Subtract from last_pg * to find the new last_pg, which also becomes the new root. */ pgno_t freecount = 0; MDBX_cursor_couple couple; MDBX_val key, data; int rc = mdbx_cursor_init(&couple.outer, read_txn, FREE_DBI); if (unlikely(rc != MDBX_SUCCESS)) return rc; while ((rc = mdbx_cursor_get(&couple.outer, &key, &data, MDBX_NEXT)) == 0) freecount += *(pgno_t *)data.iov_base; if (unlikely(rc != MDBX_NOTFOUND)) return rc; freecount += read_txn->mt_dbs[FREE_DBI].md_branch_pages + read_txn->mt_dbs[FREE_DBI].md_leaf_pages + read_txn->mt_dbs[FREE_DBI].md_overflow_pages; new_root = read_txn->mt_next_pgno - 1 - freecount; meta->mm_geo.next = new_root + 1; meta->mm_dbs[MAIN_DBI] = read_txn->mt_dbs[MAIN_DBI]; meta->mm_dbs[MAIN_DBI].md_root = new_root; mdbx_copy ctx; memset(&ctx, 0, sizeof(ctx)); rc = mdbx_condpair_init(&ctx.mc_condpair); if (unlikely(rc != MDBX_SUCCESS)) return rc; memset(data_buffer, 0, MDBX_WBUF * 2); ctx.mc_wbuf[0] = data_buffer; ctx.mc_wbuf[1] = data_buffer + MDBX_WBUF; ctx.mc_next_pgno = NUM_METAS; ctx.mc_env = env; ctx.mc_fd = fd; ctx.mc_txn = read_txn; mdbx_thread_t thread; int thread_err = mdbx_thread_create(&thread, mdbx_env_copythr, &ctx); if (likely(thread_err == MDBX_SUCCESS)) { if (dest_is_pipe) { compact_fixup_meta(env, meta); rc = mdbx_write(fd, buffer, meta_bytes); } if (rc == MDBX_SUCCESS) rc = mdbx_env_cwalk(&ctx, &root, 0); mdbx_env_cthr_toggle(&ctx); mdbx_env_cthr_toggle(&ctx); thread_err = mdbx_thread_join(thread); mdbx_assert(env, (ctx.mc_tail == ctx.mc_head && ctx.mc_wlen[ctx.mc_head & 1] == 0) || ctx.mc_error); mdbx_condpair_destroy(&ctx.mc_condpair); } if (unlikely(thread_err != MDBX_SUCCESS)) return thread_err; if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(ctx.mc_error != MDBX_SUCCESS)) return ctx.mc_error; if (dest_is_pipe) { if (root != new_root) { mdbx_error("post-compactification root %" PRIaPGNO " NE expected %" PRIaPGNO " (source DB corrupted or has a page leak(s))", root, new_root); return MDBX_CORRUPTED; /* page leak or corrupt DB */ } } else { if (root > new_root) { mdbx_error("post-compactification root %" PRIaPGNO " GT expected %" PRIaPGNO " (source DB corrupted)", root, new_root); return MDBX_CORRUPTED; /* page leak or corrupt DB */ } if (root < new_root) { mdbx_warning("post-compactification root %" PRIaPGNO " LT expected %" PRIaPGNO " (page leak(s) in source DB)", root, new_root); /* fixup meta */ meta->mm_dbs[MAIN_DBI].md_root = root; meta->mm_geo.next = root + 1; } compact_fixup_meta(env, meta); } } /* Extend file if required */ if (meta->mm_geo.now != meta->mm_geo.next) { const size_t whole_size = pgno2bytes(env, meta->mm_geo.now); if (!dest_is_pipe) return mdbx_ftruncate(fd, whole_size); const size_t used_size = pgno2bytes(env, meta->mm_geo.next); memset(data_buffer, 0, MDBX_WBUF); for (size_t offset = used_size; offset < whole_size;) { const size_t chunk = (MDBX_WBUF < whole_size - offset) ? MDBX_WBUF : whole_size - offset; /* copy to avoid EFAULT in case swapped-out */ int rc = mdbx_write(fd, data_buffer, chunk); if (unlikely(rc != MDBX_SUCCESS)) return rc; offset += chunk; } } return MDBX_SUCCESS; } /* Copy environment as-is. */ static __cold int mdbx_env_copy_asis(MDBX_env *env, MDBX_txn *read_txn, mdbx_filehandle_t fd, uint8_t *buffer, const bool dest_is_pipe, const int flags) { /* We must start the actual read txn after blocking writers */ int rc = mdbx_txn_end(read_txn, MDBX_END_RESET_TMP); if (unlikely(rc != MDBX_SUCCESS)) return rc; /* Temporarily block writers until we snapshot the meta pages */ rc = mdbx_txn_lock(env, false); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_txn_renew0(read_txn, MDBX_TXN_RDONLY); if (unlikely(rc != MDBX_SUCCESS)) { mdbx_txn_unlock(env); return rc; } mdbx_jitter4testing(false); const size_t meta_bytes = pgno2bytes(env, NUM_METAS); /* Make a snapshot of meta-pages, * but writing ones after the data was flushed */ memcpy(buffer, env->me_map, meta_bytes); MDBX_meta *const headcopy = /* LY: get pointer to the snapshot copy */ (MDBX_meta *)(buffer + ((uint8_t *)mdbx_meta_head(env) - env->me_map)); mdbx_txn_unlock(env); if (flags & MDBX_CP_FORCE_DYNAMIC_SIZE) make_sizeable(headcopy); /* Update signature to steady */ unaligned_poke_u64(4, headcopy->mm_datasync_sign, mdbx_meta_sign(headcopy)); /* Copy the data */ const size_t whole_size = pgno_align2os_bytes(env, read_txn->mt_end_pgno); const size_t used_size = pgno2bytes(env, read_txn->mt_next_pgno); mdbx_jitter4testing(false); if (dest_is_pipe) rc = mdbx_write(fd, buffer, meta_bytes); uint8_t *const data_buffer = buffer + ceil_powerof2(meta_bytes, env->me_os_psize); for (size_t offset = meta_bytes; rc == MDBX_SUCCESS && offset < used_size;) { #if MDBX_USE_SENDFILE static bool sendfile_unavailable; if (dest_is_pipe && likely(!sendfile_unavailable)) { off_t in_offset = offset; const ssize_t written = sendfile(fd, env->me_lazy_fd, &in_offset, used_size - offset); if (likely(written > 0)) { offset = in_offset; continue; } rc = MDBX_ENODATA; if (written == 0 || ignore_enosys(rc = errno) != MDBX_RESULT_TRUE) break; sendfile_unavailable = true; } #endif /* MDBX_USE_SENDFILE */ #if MDBX_USE_COPYFILERANGE static bool copyfilerange_unavailable; if (!dest_is_pipe && likely(!copyfilerange_unavailable)) { off_t in_offset = offset, out_offset = offset; ssize_t bytes_copied = copy_file_range( env->me_lazy_fd, &in_offset, fd, &out_offset, used_size - offset, 0); if (likely(bytes_copied > 0)) { offset = in_offset; continue; } rc = MDBX_ENODATA; if (bytes_copied == 0 || ignore_enosys(rc = errno) != MDBX_RESULT_TRUE) break; copyfilerange_unavailable = true; } #endif /* MDBX_USE_COPYFILERANGE */ /* fallback to portable */ const size_t chunk = (MDBX_WBUF < used_size - offset) ? MDBX_WBUF : used_size - offset; /* copy to avoid EFAULT in case swapped-out */ memcpy(data_buffer, env->me_map + offset, chunk); rc = mdbx_write(fd, data_buffer, chunk); offset += chunk; } /* Extend file if required */ if (likely(rc == MDBX_SUCCESS) && whole_size != used_size) { if (!dest_is_pipe) rc = mdbx_ftruncate(fd, whole_size); else { memset(data_buffer, 0, MDBX_WBUF); for (size_t offset = used_size; rc == MDBX_SUCCESS && offset < whole_size;) { const size_t chunk = (MDBX_WBUF < whole_size - offset) ? MDBX_WBUF : whole_size - offset; /* copy to avoid EFAULT in case swapped-out */ rc = mdbx_write(fd, data_buffer, chunk); offset += chunk; } } } return rc; } __cold int mdbx_env_copy2fd(MDBX_env *env, mdbx_filehandle_t fd, unsigned flags) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; const int dest_is_pipe = mdbx_is_pipe(fd); if (MDBX_IS_ERROR(dest_is_pipe)) return dest_is_pipe; if (!dest_is_pipe) { rc = mdbx_fseek(fd, 0); if (unlikely(rc != MDBX_SUCCESS)) return rc; } const size_t buffer_size = pgno_align2os_bytes(env, NUM_METAS) + ceil_powerof2(((flags & MDBX_CP_COMPACT) ? MDBX_WBUF * 2 : MDBX_WBUF), env->me_os_psize); uint8_t *buffer = NULL; rc = mdbx_memalign_alloc(env->me_os_psize, buffer_size, (void **)&buffer); if (unlikely(rc != MDBX_SUCCESS)) return rc; MDBX_txn *read_txn = NULL; /* Do the lock/unlock of the reader mutex before starting the * write txn. Otherwise other read txns could block writers. */ rc = mdbx_txn_begin(env, NULL, MDBX_TXN_RDONLY, &read_txn); if (unlikely(rc != MDBX_SUCCESS)) { mdbx_memalign_free(buffer); return rc; } if (!dest_is_pipe) { /* Firstly write a stub to meta-pages. * Now we sure to incomplete copy will not be used. */ memset(buffer, -1, pgno2bytes(env, NUM_METAS)); rc = mdbx_write(fd, buffer, pgno2bytes(env, NUM_METAS)); } if (likely(rc == MDBX_SUCCESS)) { memset(buffer, 0, pgno2bytes(env, NUM_METAS)); rc = ((flags & MDBX_CP_COMPACT) ? mdbx_env_compact : mdbx_env_copy_asis)( env, read_txn, fd, buffer, dest_is_pipe, flags); } mdbx_txn_abort(read_txn); if (!dest_is_pipe) { if (likely(rc == MDBX_SUCCESS)) rc = mdbx_fsync(fd, MDBX_SYNC_DATA | MDBX_SYNC_SIZE); /* Write actual meta */ if (likely(rc == MDBX_SUCCESS)) rc = mdbx_pwrite(fd, buffer, pgno2bytes(env, NUM_METAS), 0); if (likely(rc == MDBX_SUCCESS)) rc = mdbx_fsync(fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ); } mdbx_memalign_free(buffer); return rc; } __cold int mdbx_env_copy(MDBX_env *env, const char *dest_path, MDBX_copy_flags_t flags) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!dest_path)) return MDBX_EINVAL; /* The destination path must exist, but the destination file must not. * We don't want the OS to cache the writes, since the source data is * already in the OS cache. */ mdbx_filehandle_t newfd; rc = mdbx_openfile(MDBX_OPEN_COPY, env, dest_path, &newfd, #if defined(_WIN32) || defined(_WIN64) (mdbx_mode_t)-1 #else S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP #endif ); if (rc == MDBX_SUCCESS) { #if defined(_WIN32) || defined(_WIN64) OVERLAPPED ov; memset(&ov, 0, sizeof(ov)); if (!LockFileEx(newfd, LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, 0, INT32_MAX, &ov)) rc = GetLastError(); #else struct flock lock_op; memset(&lock_op, 0, sizeof(lock_op)); lock_op.l_type = F_WRLCK; lock_op.l_whence = SEEK_SET; lock_op.l_start = 0; lock_op.l_len = (sizeof(lock_op.l_len) > 4 ? INT64_MAX : INT32_MAX) & ~(size_t)0xffff; if (fcntl(newfd, F_SETLK, &lock_op) #if (defined(__linux__) || defined(__gnu_linux__)) && defined(LOCK_EX) && \ (!defined(__ANDROID_API__) || __ANDROID_API__ >= 24) || flock(newfd, LOCK_EX | LOCK_NB) #endif /* Linux */ ) rc = errno; #endif /* Windows / POSIX */ } if (rc == MDBX_SUCCESS) rc = mdbx_env_copy2fd(env, newfd, flags); if (newfd != INVALID_HANDLE_VALUE) { int err = mdbx_closefile(newfd); if (rc == MDBX_SUCCESS && err != rc) rc = err; if (rc != MDBX_SUCCESS) (void)mdbx_removefile(dest_path); } return rc; } /******************************************************************************/ __cold int mdbx_env_set_flags(MDBX_env *env, MDBX_env_flags_t flags, bool onoff) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(flags & ~ENV_CHANGEABLE_FLAGS)) return MDBX_EPERM; if (unlikely(env->me_flags & MDBX_RDONLY)) return MDBX_EACCESS; if (unlikely(env->me_txn0->mt_owner == mdbx_thread_self())) return MDBX_BUSY; rc = mdbx_txn_lock(env, false); if (unlikely(rc)) return rc; if (onoff) env->me_flags = merge_sync_flags(env->me_flags, flags); else env->me_flags &= ~flags; mdbx_txn_unlock(env); return MDBX_SUCCESS; } __cold int mdbx_env_get_flags(const MDBX_env *env, unsigned *arg) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!arg)) return MDBX_EINVAL; *arg = env->me_flags & ENV_USABLE_FLAGS; return MDBX_SUCCESS; } __cold int mdbx_env_set_userctx(MDBX_env *env, void *ctx) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; env->me_userctx = ctx; return MDBX_SUCCESS; } void *__cold mdbx_env_get_userctx(const MDBX_env *env) { return env ? env->me_userctx : NULL; } __cold int mdbx_env_set_assert(MDBX_env *env, MDBX_assert_func *func) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; #if MDBX_DEBUG env->me_assert_func = func; return MDBX_SUCCESS; #else (void)func; return MDBX_ENOSYS; #endif } __cold int mdbx_env_get_path(const MDBX_env *env, const char **arg) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!arg)) return MDBX_EINVAL; *arg = env->me_pathname; return MDBX_SUCCESS; } __cold int mdbx_env_get_fd(const MDBX_env *env, mdbx_filehandle_t *arg) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!arg)) return MDBX_EINVAL; *arg = env->me_lazy_fd; return MDBX_SUCCESS; } /* Common code for mdbx_dbi_stat() and mdbx_env_stat(). * [in] env the environment to operate in. * [in] db the MDBX_db record containing the stats to return. * [out] arg the address of an MDBX_stat structure to receive the stats. * Returns 0, this function always succeeds. */ static void mdbx_stat0(const MDBX_env *env, const MDBX_db *db, MDBX_stat *dest, size_t bytes) { dest->ms_psize = env->me_psize; dest->ms_depth = db->md_depth; dest->ms_branch_pages = db->md_branch_pages; dest->ms_leaf_pages = db->md_leaf_pages; dest->ms_overflow_pages = db->md_overflow_pages; dest->ms_entries = db->md_entries; if (likely(bytes >= offsetof(MDBX_stat, ms_mod_txnid) + sizeof(dest->ms_mod_txnid))) dest->ms_mod_txnid = db->md_mod_txnid; } __cold int mdbx_env_stat(const MDBX_env *env, MDBX_stat *stat, size_t bytes) { return __inline_mdbx_env_stat(env, stat, bytes); } __cold int mdbx_env_stat_ex(const MDBX_env *env, const MDBX_txn *txn, MDBX_stat *dest, size_t bytes) { if (unlikely((env == NULL && txn == NULL) || dest == NULL)) return MDBX_EINVAL; if (txn) { int err = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(err != MDBX_SUCCESS)) return err; } if (env) { int err = check_env(env); if (unlikely(err != MDBX_SUCCESS)) return err; if (txn && unlikely(txn->mt_env != env)) return MDBX_EINVAL; } const size_t size_before_modtxnid = offsetof(MDBX_stat, ms_mod_txnid); if (unlikely(bytes != sizeof(MDBX_stat)) && bytes != size_before_modtxnid) return MDBX_EINVAL; if (txn) { mdbx_stat0(txn->mt_env, &txn->mt_dbs[MAIN_DBI], dest, bytes); return MDBX_SUCCESS; } while (1) { const MDBX_meta *const recent_meta = mdbx_meta_head(env); const txnid_t txnid = mdbx_meta_txnid_fluid(env, recent_meta); mdbx_stat0(env, &recent_meta->mm_dbs[MAIN_DBI], dest, bytes); mdbx_compiler_barrier(); if (likely(txnid == mdbx_meta_txnid_fluid(env, recent_meta) && recent_meta == mdbx_meta_head(env))) return MDBX_SUCCESS; } } __cold int mdbx_dbi_dupsort_depthmask(MDBX_txn *txn, MDBX_dbi dbi, uint32_t *mask) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mask)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_VALID))) return MDBX_BAD_DBI; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; if ((cx.outer.mc_db->md_flags & MDBX_DUPSORT) == 0) return MDBX_RESULT_TRUE; MDBX_val key, data; rc = mdbx_cursor_first(&cx.outer, &key, &data); *mask = 0; while (rc == MDBX_SUCCESS) { const MDBX_node *node = page_node(cx.outer.mc_pg[cx.outer.mc_top], cx.outer.mc_ki[cx.outer.mc_top]); const MDBX_db *db = node_data(node); const unsigned flags = node_flags(node); switch (flags) { case F_BIGDATA: case 0: /* single-value entry, deep = 0 */ *mask |= 1 << 0; break; case F_DUPDATA: /* single sub-page, deep = 1 */ *mask |= 1 << 1; break; case F_DUPDATA | F_SUBDATA: /* sub-tree */ *mask |= 1 << unaligned_peek_u16(1, &db->md_depth); break; default: mdbx_error("wrong node-flags %u", flags); return MDBX_CORRUPTED; } rc = mdbx_cursor_next(&cx.outer, &key, &data, MDBX_NEXT_NODUP); } return (rc == MDBX_NOTFOUND) ? MDBX_SUCCESS : rc; } __cold int mdbx_env_info(const MDBX_env *env, MDBX_envinfo *info, size_t bytes) { return __inline_mdbx_env_info(env, info, bytes); } __cold int mdbx_env_info_ex(const MDBX_env *env, const MDBX_txn *txn, MDBX_envinfo *arg, size_t bytes) { if (unlikely((env == NULL && txn == NULL) || arg == NULL)) return MDBX_EINVAL; if (txn) { int err = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(err != MDBX_SUCCESS)) return err; } if (env) { int err = check_env(env); if (unlikely(err != MDBX_SUCCESS)) return err; if (txn && unlikely(txn->mt_env != env)) return MDBX_EINVAL; } else { env = txn->mt_env; } const size_t size_before_bootid = offsetof(MDBX_envinfo, mi_bootid); if (unlikely(bytes != sizeof(MDBX_envinfo)) && bytes != size_before_bootid) return MDBX_EINVAL; const MDBX_meta *const meta0 = METAPAGE(env, 0); const MDBX_meta *const meta1 = METAPAGE(env, 1); const MDBX_meta *const meta2 = METAPAGE(env, 2); pgno_t unsynced_pages; while (1) { if (unlikely(env->me_flags & MDBX_FATAL_ERROR)) return MDBX_PANIC; const MDBX_meta *const recent_meta = mdbx_meta_head(env); arg->mi_recent_txnid = mdbx_meta_txnid_fluid(env, recent_meta); arg->mi_meta0_txnid = mdbx_meta_txnid_fluid(env, meta0); arg->mi_meta0_sign = unaligned_peek_u64(4, meta0->mm_datasync_sign); arg->mi_meta1_txnid = mdbx_meta_txnid_fluid(env, meta1); arg->mi_meta1_sign = unaligned_peek_u64(4, meta1->mm_datasync_sign); arg->mi_meta2_txnid = mdbx_meta_txnid_fluid(env, meta2); arg->mi_meta2_sign = unaligned_peek_u64(4, meta2->mm_datasync_sign); if (likely(bytes > size_before_bootid)) { memcpy(&arg->mi_bootid.meta0, &meta0->mm_bootid, 16); memcpy(&arg->mi_bootid.meta1, &meta1->mm_bootid, 16); memcpy(&arg->mi_bootid.meta2, &meta2->mm_bootid, 16); } const MDBX_meta *txn_meta = recent_meta; arg->mi_last_pgno = txn_meta->mm_geo.next - 1; arg->mi_geo.current = pgno2bytes(env, txn_meta->mm_geo.now); if (txn) { arg->mi_last_pgno = txn->mt_next_pgno - 1; arg->mi_geo.current = pgno2bytes(env, txn->mt_end_pgno); const txnid_t wanna_meta_txnid = (txn->mt_flags & MDBX_TXN_RDONLY) ? txn->mt_txnid : txn->mt_txnid - MDBX_TXNID_STEP; txn_meta = (arg->mi_meta0_txnid == wanna_meta_txnid) ? meta0 : txn_meta; txn_meta = (arg->mi_meta1_txnid == wanna_meta_txnid) ? meta1 : txn_meta; txn_meta = (arg->mi_meta2_txnid == wanna_meta_txnid) ? meta2 : txn_meta; } arg->mi_geo.lower = pgno2bytes(env, txn_meta->mm_geo.lower); arg->mi_geo.upper = pgno2bytes(env, txn_meta->mm_geo.upper); arg->mi_geo.shrink = pgno2bytes(env, txn_meta->mm_geo.shrink); arg->mi_geo.grow = pgno2bytes(env, txn_meta->mm_geo.grow); unsynced_pages = atomic_load32(env->me_unsynced_pages, mo_Relaxed) + (atomic_load32(env->me_meta_sync_txnid, mo_Relaxed) != (uint32_t)arg->mi_last_pgno); arg->mi_mapsize = env->me_dxb_mmap.limit; mdbx_compiler_barrier(); if (likely(arg->mi_meta0_txnid == mdbx_meta_txnid_fluid(env, meta0) && arg->mi_meta0_sign == unaligned_peek_u64(4, meta0->mm_datasync_sign) && arg->mi_meta1_txnid == mdbx_meta_txnid_fluid(env, meta1) && arg->mi_meta1_sign == unaligned_peek_u64(4, meta1->mm_datasync_sign) && arg->mi_meta2_txnid == mdbx_meta_txnid_fluid(env, meta2) && arg->mi_meta2_sign == unaligned_peek_u64(4, meta2->mm_datasync_sign) && recent_meta == mdbx_meta_head(env) && arg->mi_recent_txnid == mdbx_meta_txnid_fluid(env, recent_meta))) break; } arg->mi_maxreaders = env->me_maxreaders; arg->mi_numreaders = env->me_lck ? atomic_load32(&env->me_lck->mti_numreaders, mo_Relaxed) : INT32_MAX; arg->mi_dxb_pagesize = env->me_psize; arg->mi_sys_pagesize = env->me_os_psize; const MDBX_lockinfo *const lck = env->me_lck; if (likely(bytes > size_before_bootid)) { arg->mi_unsync_volume = pgno2bytes(env, unsynced_pages); const uint64_t monotime_now = mdbx_osal_monotime(); arg->mi_since_sync_seconds16dot16 = mdbx_osal_monotime_to_16dot16( monotime_now - atomic_load64(env->me_sync_timestamp, mo_Relaxed)); arg->mi_since_reader_check_seconds16dot16 = lck ? mdbx_osal_monotime_to_16dot16( monotime_now - atomic_load64(&lck->mti_reader_check_timestamp, mo_Relaxed)) : 0; arg->mi_autosync_threshold = pgno2bytes(env, atomic_load32(env->me_autosync_threshold, mo_Relaxed)); arg->mi_autosync_period_seconds16dot16 = mdbx_osal_monotime_to_16dot16( atomic_load64(env->me_autosync_period, mo_Relaxed)); arg->mi_bootid.current.x = bootid.x; arg->mi_bootid.current.y = bootid.y; arg->mi_mode = lck ? lck->mti_envmode.weak : env->me_flags; } arg->mi_self_latter_reader_txnid = arg->mi_latter_reader_txnid = 0; if (lck) { arg->mi_self_latter_reader_txnid = arg->mi_latter_reader_txnid = arg->mi_recent_txnid; for (unsigned i = 0; i < arg->mi_numreaders; ++i) { const uint32_t pid = atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease); if (pid) { const txnid_t txnid = safe64_read(&lck->mti_readers[i].mr_txnid); if (arg->mi_latter_reader_txnid > txnid) arg->mi_latter_reader_txnid = txnid; if (pid == env->me_pid && arg->mi_self_latter_reader_txnid > txnid) arg->mi_self_latter_reader_txnid = txnid; } } } return MDBX_SUCCESS; } static __inline MDBX_cmp_func *get_default_keycmp(unsigned flags) { return (flags & MDBX_REVERSEKEY) ? cmp_reverse : (flags & MDBX_INTEGERKEY) ? cmp_int_align2 : cmp_lexical; } static __inline MDBX_cmp_func *get_default_datacmp(unsigned flags) { return !(flags & MDBX_DUPSORT) ? cmp_lenfast : ((flags & MDBX_INTEGERDUP) ? cmp_int_unaligned : ((flags & MDBX_REVERSEDUP) ? cmp_reverse : cmp_lexical)); } static int mdbx_dbi_bind(MDBX_txn *txn, const MDBX_dbi dbi, unsigned user_flags, MDBX_cmp_func *keycmp, MDBX_cmp_func *datacmp) { /* LY: so, accepting only three cases for the table's flags: * 1) user_flags and both comparators are zero * = assume that a by-default mode/flags is requested for reading; * 2) user_flags exactly the same * = assume that the target mode/flags are requested properly; * 3) user_flags differs, but table is empty and MDBX_CREATE is provided * = assume that a properly create request with custom flags; */ if ((user_flags ^ txn->mt_dbs[dbi].md_flags) & DB_PERSISTENT_FLAGS) { /* flags are differs, check other conditions */ if ((!user_flags && (!keycmp || keycmp == txn->mt_dbxs[dbi].md_cmp) && (!datacmp || datacmp == txn->mt_dbxs[dbi].md_dcmp)) || user_flags == MDBX_ACCEDE) { /* no comparators were provided and flags are zero, * seems that is case #1 above */ user_flags = txn->mt_dbs[dbi].md_flags; } else if ((user_flags & MDBX_CREATE) && txn->mt_dbs[dbi].md_entries == 0) { if (txn->mt_flags & MDBX_TXN_RDONLY) return /* FIXME: return extended info */ MDBX_EACCESS; /* make sure flags changes get committed */ txn->mt_dbs[dbi].md_flags = user_flags & DB_PERSISTENT_FLAGS; txn->mt_flags |= MDBX_TXN_DIRTY; } else { return /* FIXME: return extended info */ MDBX_INCOMPATIBLE; } } if (!keycmp) keycmp = txn->mt_dbxs[dbi].md_cmp ? txn->mt_dbxs[dbi].md_cmp : get_default_keycmp(user_flags); if (txn->mt_dbxs[dbi].md_cmp != keycmp) { if (txn->mt_dbxs[dbi].md_cmp) return MDBX_EINVAL; txn->mt_dbxs[dbi].md_cmp = keycmp; } if (!datacmp) datacmp = txn->mt_dbxs[dbi].md_dcmp ? txn->mt_dbxs[dbi].md_dcmp : get_default_datacmp(user_flags); if (txn->mt_dbxs[dbi].md_dcmp != datacmp) { if (txn->mt_dbxs[dbi].md_dcmp) return MDBX_EINVAL; txn->mt_dbxs[dbi].md_dcmp = datacmp; } return MDBX_SUCCESS; } static int dbi_open(MDBX_txn *txn, const char *table_name, unsigned user_flags, MDBX_dbi *dbi, MDBX_cmp_func *keycmp, MDBX_cmp_func *datacmp) { int rc = MDBX_EINVAL; if (unlikely(!dbi)) return rc; if (unlikely((user_flags & ~DB_USABLE_FLAGS) != 0)) { early_bailout: *dbi = 0; return rc; } rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) goto early_bailout; switch (user_flags & (MDBX_INTEGERDUP | MDBX_DUPFIXED | MDBX_DUPSORT | MDBX_REVERSEDUP | MDBX_ACCEDE)) { case MDBX_ACCEDE: if ((user_flags & MDBX_CREATE) == 0) break; __fallthrough /* fall through */; default: rc = MDBX_EINVAL; goto early_bailout; case MDBX_DUPSORT: case MDBX_DUPSORT | MDBX_REVERSEDUP: case MDBX_DUPSORT | MDBX_DUPFIXED: case MDBX_DUPSORT | MDBX_DUPFIXED | MDBX_REVERSEDUP: case MDBX_DUPSORT | MDBX_DUPFIXED | MDBX_INTEGERDUP: case MDBX_DUPSORT | MDBX_DUPFIXED | MDBX_INTEGERDUP | MDBX_REVERSEDUP: case 0: break; } /* main table? */ if (!table_name) { rc = mdbx_dbi_bind(txn, MAIN_DBI, user_flags, keycmp, datacmp); if (unlikely(rc != MDBX_SUCCESS)) goto early_bailout; *dbi = MAIN_DBI; return rc; } if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) { txn->mt_dbxs[MAIN_DBI].md_cmp = get_default_keycmp(txn->mt_dbs[MAIN_DBI].md_flags); txn->mt_dbxs[MAIN_DBI].md_dcmp = get_default_datacmp(txn->mt_dbs[MAIN_DBI].md_flags); } /* Is the DB already open? */ size_t len = strlen(table_name); MDBX_dbi scan, slot; for (slot = scan = txn->mt_numdbs; --scan >= CORE_DBS;) { if (!txn->mt_dbxs[scan].md_name.iov_len) { /* Remember this free slot */ slot = scan; continue; } if (len == txn->mt_dbxs[scan].md_name.iov_len && !strncmp(table_name, txn->mt_dbxs[scan].md_name.iov_base, len)) { rc = mdbx_dbi_bind(txn, scan, user_flags, keycmp, datacmp); if (unlikely(rc != MDBX_SUCCESS)) goto early_bailout; *dbi = scan; return rc; } } /* Fail, if no free slot and max hit */ MDBX_env *env = txn->mt_env; if (unlikely(slot >= env->me_maxdbs)) { rc = MDBX_DBS_FULL; goto early_bailout; } /* Cannot mix named table with some main-table flags */ if (unlikely(txn->mt_dbs[MAIN_DBI].md_flags & (MDBX_DUPSORT | MDBX_INTEGERKEY))) { rc = (user_flags & MDBX_CREATE) ? MDBX_INCOMPATIBLE : MDBX_NOTFOUND; goto early_bailout; } /* Find the DB info */ int exact = 0; MDBX_val key, data; key.iov_len = len; key.iov_base = (void *)table_name; MDBX_cursor_couple couple; rc = mdbx_cursor_init(&couple.outer, txn, MAIN_DBI); if (unlikely(rc != MDBX_SUCCESS)) goto early_bailout; rc = mdbx_cursor_set(&couple.outer, &key, &data, MDBX_SET, &exact); if (unlikely(rc != MDBX_SUCCESS)) { if (rc != MDBX_NOTFOUND || !(user_flags & MDBX_CREATE)) goto early_bailout; } else { /* make sure this is actually a table */ MDBX_node *node = page_node(couple.outer.mc_pg[couple.outer.mc_top], couple.outer.mc_ki[couple.outer.mc_top]); if (unlikely((node_flags(node) & (F_DUPDATA | F_SUBDATA)) != F_SUBDATA)) { rc = MDBX_INCOMPATIBLE; goto early_bailout; } if (unlikely(data.iov_len != sizeof(MDBX_db))) { rc = MDBX_CORRUPTED; goto early_bailout; } } if (rc != MDBX_SUCCESS && unlikely(txn->mt_flags & MDBX_TXN_RDONLY)) { rc = MDBX_EACCESS; goto early_bailout; } /* Done here so we cannot fail after creating a new DB */ char *namedup = mdbx_strdup(table_name); if (unlikely(!namedup)) { rc = MDBX_ENOMEM; goto early_bailout; } int err = mdbx_fastmutex_acquire(&env->me_dbi_lock); if (unlikely(err != MDBX_SUCCESS)) { rc = err; mdbx_free(namedup); goto early_bailout; } /* Import handles from env */ dbi_import_locked(txn); /* Rescan after mutex acquisition & import handles */ for (slot = scan = txn->mt_numdbs; --scan >= CORE_DBS;) { if (!txn->mt_dbxs[scan].md_name.iov_len) { /* Remember this free slot */ slot = scan; continue; } if (len == txn->mt_dbxs[scan].md_name.iov_len && !strncmp(table_name, txn->mt_dbxs[scan].md_name.iov_base, len)) { rc = mdbx_dbi_bind(txn, scan, user_flags, keycmp, datacmp); if (unlikely(rc != MDBX_SUCCESS)) goto later_bailout; *dbi = scan; goto later_exit; } } if (unlikely(slot >= env->me_maxdbs)) { rc = MDBX_DBS_FULL; goto later_bailout; } unsigned dbiflags = DBI_FRESH | DBI_VALID | DBI_USRVALID; MDBX_db db_dummy; if (unlikely(rc)) { /* MDBX_NOTFOUND and MDBX_CREATE: Create new DB */ mdbx_tassert(txn, rc == MDBX_NOTFOUND); memset(&db_dummy, 0, sizeof(db_dummy)); db_dummy.md_root = P_INVALID; db_dummy.md_flags = user_flags & DB_PERSISTENT_FLAGS; data.iov_len = sizeof(db_dummy); data.iov_base = &db_dummy; WITH_CURSOR_TRACKING(couple.outer, rc = mdbx_cursor_put(&couple.outer, &key, &data, F_SUBDATA | MDBX_NOOVERWRITE)); if (unlikely(rc != MDBX_SUCCESS)) goto later_bailout; dbiflags |= DBI_DIRTY | DBI_CREAT; txn->mt_flags |= MDBX_TXN_DIRTY; } /* Got info, register DBI in this txn */ memset(txn->mt_dbxs + slot, 0, sizeof(MDBX_dbx)); txn->mt_dbs[slot] = *(MDBX_db *)data.iov_base; env->me_dbflags[slot] = 0; rc = mdbx_dbi_bind(txn, slot, user_flags, keycmp, datacmp); if (unlikely(rc != MDBX_SUCCESS)) { mdbx_tassert(txn, (dbiflags & DBI_CREAT) == 0); later_bailout: *dbi = 0; later_exit: mdbx_free(namedup); } else { txn->mt_dbistate[slot] = (uint8_t)dbiflags; txn->mt_dbxs[slot].md_name.iov_base = namedup; txn->mt_dbxs[slot].md_name.iov_len = len; txn->mt_dbiseqs[slot] = ++env->me_dbiseqs[slot]; if (!(dbiflags & DBI_CREAT)) env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags | DB_VALID; if (txn->mt_numdbs == slot) { mdbx_compiler_barrier(); txn->mt_numdbs = env->me_numdbs = slot + 1; if (!(txn->mt_flags & MDBX_TXN_RDONLY)) txn->tw.cursors[slot] = NULL; } mdbx_assert(env, env->me_numdbs > slot); *dbi = slot; } mdbx_ensure(env, mdbx_fastmutex_release(&env->me_dbi_lock) == MDBX_SUCCESS); return rc; } int mdbx_dbi_open(MDBX_txn *txn, const char *table_name, MDBX_db_flags_t table_flags, MDBX_dbi *dbi) { return dbi_open(txn, table_name, table_flags, dbi, nullptr, nullptr); } int mdbx_dbi_open_ex(MDBX_txn *txn, const char *table_name, MDBX_db_flags_t table_flags, MDBX_dbi *dbi, MDBX_cmp_func *keycmp, MDBX_cmp_func *datacmp) { return dbi_open(txn, table_name, table_flags, dbi, keycmp, datacmp); } __cold int mdbx_dbi_stat(MDBX_txn *txn, MDBX_dbi dbi, MDBX_stat *dest, size_t bytes) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!dest)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_VALID))) return MDBX_BAD_DBI; const size_t size_before_modtxnid = offsetof(MDBX_stat, ms_mod_txnid); if (unlikely(bytes != sizeof(MDBX_stat)) && bytes != size_before_modtxnid) return MDBX_EINVAL; if (unlikely(txn->mt_flags & MDBX_TXN_BLOCKED)) return MDBX_BAD_TXN; if (unlikely(txn->mt_dbistate[dbi] & DBI_STALE)) { rc = mdbx_fetch_sdb(txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; } mdbx_stat0(txn->mt_env, &txn->mt_dbs[dbi], dest, bytes); return MDBX_SUCCESS; } static int mdbx_dbi_close_locked(MDBX_env *env, MDBX_dbi dbi) { mdbx_assert(env, dbi >= CORE_DBS); if (unlikely(dbi >= env->me_numdbs)) return MDBX_BAD_DBI; char *ptr = env->me_dbxs[dbi].md_name.iov_base; /* If there was no name, this was already closed */ if (unlikely(!ptr)) return MDBX_BAD_DBI; env->me_dbflags[dbi] = 0; env->me_dbiseqs[dbi]++; env->me_dbxs[dbi].md_name.iov_len = 0; mdbx_memory_fence(false, true); env->me_dbxs[dbi].md_name.iov_base = NULL; mdbx_free(ptr); if (env->me_numdbs == dbi + 1) env->me_numdbs = dbi; return MDBX_SUCCESS; } int mdbx_dbi_close(MDBX_env *env, MDBX_dbi dbi) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(dbi < CORE_DBS || dbi >= env->me_maxdbs)) return MDBX_BAD_DBI; rc = mdbx_fastmutex_acquire(&env->me_dbi_lock); if (likely(rc == MDBX_SUCCESS)) { rc = (dbi < env->me_maxdbs && (env->me_dbflags[dbi] & DB_VALID)) ? mdbx_dbi_close_locked(env, dbi) : MDBX_BAD_DBI; mdbx_ensure(env, mdbx_fastmutex_release(&env->me_dbi_lock) == MDBX_SUCCESS); } return rc; } int mdbx_dbi_flags_ex(MDBX_txn *txn, MDBX_dbi dbi, unsigned *flags, unsigned *state) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!flags || !state)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_VALID))) return MDBX_BAD_DBI; *flags = txn->mt_dbs[dbi].md_flags & DB_PERSISTENT_FLAGS; *state = txn->mt_dbistate[dbi] & (DBI_FRESH | DBI_CREAT | DBI_DIRTY | DBI_STALE); return MDBX_SUCCESS; } int mdbx_dbi_flags(MDBX_txn *txn, MDBX_dbi dbi, unsigned *flags) { return __inline_mdbx_dbi_flags(txn, dbi, flags); } /* Add all the DB's pages to the free list. * [in] mc Cursor on the DB to free. * [in] subs non-Zero to check for sub-DBs in this DB. * Returns 0 on success, non-zero on failure. */ static int mdbx_drop0(MDBX_cursor *mc, int subs) { int rc = mdbx_page_search(mc, NULL, MDBX_PS_FIRST); if (likely(rc == MDBX_SUCCESS)) { MDBX_txn *txn = mc->mc_txn; MDBX_cursor mx; unsigned i; /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves. * This also avoids any P_LEAF2 pages, which have no nodes. * Also if the DB doesn't have sub-DBs and has no overflow * pages, omit scanning leaves. */ if ((mc->mc_flags & C_SUB) || (subs | mc->mc_db->md_overflow_pages) == 0) mdbx_cursor_pop(mc); rc = mdbx_pnl_need(&txn->tw.retired_pages, mc->mc_db->md_branch_pages + mc->mc_db->md_leaf_pages + mc->mc_db->md_overflow_pages); if (unlikely(rc)) goto done; cursor_copy_internal(mc, &mx); while (mc->mc_snum > 0) { MDBX_page *mp = mc->mc_pg[mc->mc_top]; unsigned n = page_numkeys(mp); if (IS_LEAF(mp)) { for (i = 0; i < n; i++) { MDBX_node *node = page_node(mp, i); if (node_flags(node) & F_BIGDATA) { MDBX_page *omp; rc = mdbx_page_get(mc, node_largedata_pgno(node), &omp, NULL, pp_txnid4chk(mp, mc->mc_txn)); if (unlikely(rc)) goto done; mdbx_cassert(mc, IS_OVERFLOW(omp)); rc = mdbx_page_retire(mc, omp); if (unlikely(rc)) goto done; if (!mc->mc_db->md_overflow_pages && !subs) break; } else if (subs && (node_flags(node) & F_SUBDATA)) { rc = mdbx_xcursor_init1(mc, node, mp); if (unlikely(rc != MDBX_SUCCESS)) goto done; rc = mdbx_drop0(&mc->mc_xcursor->mx_cursor, 0); if (unlikely(rc)) goto done; } } if (!subs && !mc->mc_db->md_overflow_pages) goto pop; } else { for (i = 0; i < n; i++) { /* free it */ rc = mdbx_retire_pgno(mc, node_pgno(page_node(mp, i))); if (unlikely(rc)) goto done; } } if (!mc->mc_top) break; mdbx_cassert(mc, i <= UINT16_MAX); mc->mc_ki[mc->mc_top] = (indx_t)i; rc = mdbx_cursor_sibling(mc, SIBLING_RIGHT); if (rc) { if (unlikely(rc != MDBX_NOTFOUND)) goto done; /* no more siblings, go back to beginning * of previous level. */ pop: mdbx_cursor_pop(mc); mc->mc_ki[0] = 0; for (i = 1; i < mc->mc_snum; i++) { mc->mc_ki[i] = 0; mc->mc_pg[i] = mx.mc_pg[i]; } } } /* free it */ rc = mdbx_retire_pgno(mc, mc->mc_db->md_root); done: if (unlikely(rc)) txn->mt_flags |= MDBX_TXN_ERROR; } else if (rc == MDBX_NOTFOUND) { rc = MDBX_SUCCESS; } mc->mc_flags &= ~C_INITIALIZED; return rc; } int mdbx_drop(MDBX_txn *txn, MDBX_dbi dbi, bool del) { int rc = check_txn_rw(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(TXN_DBI_CHANGED(txn, dbi))) return MDBX_BAD_DBI; MDBX_cursor *mc; rc = mdbx_cursor_open(txn, dbi, &mc); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) { rc = MDBX_BAD_DBI; goto bailout; } if (unlikely(TXN_DBI_CHANGED(txn, dbi))) { rc = MDBX_BAD_DBI; goto bailout; } rc = mdbx_drop0(mc, mc->mc_db->md_flags & MDBX_DUPSORT); /* Invalidate the dropped DB's cursors */ for (MDBX_cursor *m2 = txn->tw.cursors[dbi]; m2; m2 = m2->mc_next) m2->mc_flags &= ~(C_INITIALIZED | C_EOF); if (unlikely(rc)) goto bailout; /* Can't delete the main DB */ if (del && dbi >= CORE_DBS) { rc = mdbx_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA); if (likely(rc == MDBX_SUCCESS)) { txn->mt_dbistate[dbi] = DBI_STALE; MDBX_env *env = txn->mt_env; rc = mdbx_fastmutex_acquire(&env->me_dbi_lock); if (unlikely(rc != MDBX_SUCCESS)) { txn->mt_flags |= MDBX_TXN_ERROR; goto bailout; } mdbx_dbi_close_locked(env, dbi); mdbx_ensure(env, mdbx_fastmutex_release(&env->me_dbi_lock) == MDBX_SUCCESS); } else { txn->mt_flags |= MDBX_TXN_ERROR; } } else { /* reset the DB record, mark it dirty */ txn->mt_dbistate[dbi] |= DBI_DIRTY; txn->mt_dbs[dbi].md_depth = 0; txn->mt_dbs[dbi].md_branch_pages = 0; txn->mt_dbs[dbi].md_leaf_pages = 0; txn->mt_dbs[dbi].md_overflow_pages = 0; txn->mt_dbs[dbi].md_entries = 0; txn->mt_dbs[dbi].md_root = P_INVALID; txn->mt_dbs[dbi].md_seq = 0; txn->mt_flags |= MDBX_TXN_DIRTY; } bailout: mdbx_cursor_close(mc); return rc; } int mdbx_set_compare(MDBX_txn *txn, MDBX_dbi dbi, MDBX_cmp_func *cmp) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; txn->mt_dbxs[dbi].md_cmp = cmp; return MDBX_SUCCESS; } int mdbx_set_dupsort(MDBX_txn *txn, MDBX_dbi dbi, MDBX_cmp_func *cmp) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; txn->mt_dbxs[dbi].md_dcmp = cmp; return MDBX_SUCCESS; } __cold int mdbx_reader_list(const MDBX_env *env, MDBX_reader_list_func *func, void *ctx) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!func)) return MDBX_EINVAL; rc = MDBX_RESULT_TRUE; int serial = 0; if (likely(env->me_lck)) { const unsigned snap_nreaders = atomic_load32(&env->me_lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; i++) { const MDBX_reader *r = env->me_lck->mti_readers + i; retry_reader:; const uint32_t pid = atomic_load32(&r->mr_pid, mo_AcquireRelease); if (!pid) continue; txnid_t txnid = safe64_read(&r->mr_txnid); const uint64_t tid = atomic_load64(&r->mr_tid, mo_Relaxed); const pgno_t pages_used = atomic_load32(&r->mr_snapshot_pages_used, mo_Relaxed); const uint64_t reader_pages_retired = atomic_load64(&r->mr_snapshot_pages_retired, mo_Relaxed); if (unlikely( txnid != safe64_read(&r->mr_txnid) || pid != atomic_load32(&r->mr_pid, mo_AcquireRelease) || tid != atomic_load64(&r->mr_tid, mo_Relaxed) || pages_used != atomic_load32(&r->mr_snapshot_pages_used, mo_Relaxed) || reader_pages_retired != atomic_load64(&r->mr_snapshot_pages_retired, mo_Relaxed))) goto retry_reader; mdbx_assert(env, txnid > 0); if (txnid >= SAFE64_INVALID_THRESHOLD) txnid = 0; size_t bytes_used = 0; size_t bytes_retained = 0; uint64_t lag = 0; if (txnid) { retry_header:; const MDBX_meta *const recent_meta = mdbx_meta_head(env); const uint64_t head_pages_retired = unaligned_peek_u64(4, recent_meta->mm_pages_retired); const txnid_t head_txnid = mdbx_meta_txnid_fluid(env, recent_meta); mdbx_compiler_barrier(); if (unlikely( recent_meta != mdbx_meta_head(env) || head_pages_retired != unaligned_peek_u64(4, recent_meta->mm_pages_retired)) || head_txnid != mdbx_meta_txnid_fluid(env, recent_meta)) goto retry_header; lag = (head_txnid - txnid) / MDBX_TXNID_STEP; bytes_used = pgno2bytes(env, pages_used); bytes_retained = (head_pages_retired > reader_pages_retired) ? pgno2bytes(env, (pgno_t)(head_pages_retired - reader_pages_retired)) : 0; } rc = func(ctx, ++serial, i, pid, (mdbx_tid_t)tid, txnid, lag, bytes_used, bytes_retained); if (unlikely(rc != MDBX_SUCCESS)) break; } } return rc; } /* Insert pid into list if not already present. * return -1 if already present. */ static bool __cold mdbx_pid_insert(uint32_t *ids, uint32_t pid) { /* binary search of pid in list */ unsigned base = 0; unsigned cursor = 1; int val = 0; unsigned n = ids[0]; while (n > 0) { unsigned pivot = n >> 1; cursor = base + pivot + 1; val = pid - ids[cursor]; if (val < 0) { n = pivot; } else if (val > 0) { base = cursor; n -= pivot + 1; } else { /* found, so it's a duplicate */ return false; } } if (val > 0) ++cursor; ids[0]++; for (n = ids[0]; n > cursor; n--) ids[n] = ids[n - 1]; ids[n] = pid; return true; } __cold int mdbx_reader_check(MDBX_env *env, int *dead) { if (dead) *dead = 0; return mdbx_cleanup_dead_readers(env, false, dead); } /* Return: * MDBX_RESULT_TRUE - done and mutex recovered * MDBX_SUCCESS - done * Otherwise errcode. */ MDBX_INTERNAL_FUNC __cold int mdbx_cleanup_dead_readers(MDBX_env *env, int rdt_locked, int *dead) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; mdbx_assert(env, rdt_locked >= 0); MDBX_lockinfo *const lck = env->me_lck; if (unlikely(lck == NULL)) { /* exclusive mode */ if (dead) *dead = 0; return MDBX_SUCCESS; } const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); uint32_t pidsbuf_onstask[142]; uint32_t *const pids = (snap_nreaders < ARRAY_LENGTH(pidsbuf_onstask)) ? pidsbuf_onstask : mdbx_malloc((snap_nreaders + 1) * sizeof(uint32_t)); if (unlikely(!pids)) return MDBX_ENOMEM; pids[0] = 0; int count = 0; for (unsigned i = 0; i < snap_nreaders; i++) { const uint32_t pid = atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease); if (pid == 0) continue /* skip empty */; if (pid == env->me_pid) continue /* skip self */; if (!mdbx_pid_insert(pids, pid)) continue /* such pid already processed */; int err = mdbx_rpid_check(env, pid); if (err == MDBX_RESULT_TRUE) continue /* reader is live */; if (err != MDBX_SUCCESS) { rc = err; break /* mdbx_rpid_check() failed */; } /* stale reader found */ if (!rdt_locked) { err = mdbx_rdt_lock(env); if (MDBX_IS_ERROR(err)) { rc = err; break; } rdt_locked = -1; if (err == MDBX_RESULT_TRUE) { /* mutex recovered, the mdbx_ipclock_failed() checked all readers */ rc = MDBX_RESULT_TRUE; break; } /* a other process may have clean and reused slot, recheck */ if (lck->mti_readers[i].mr_pid.weak != pid) continue; err = mdbx_rpid_check(env, pid); if (MDBX_IS_ERROR(err)) { rc = err; break; } if (err != MDBX_SUCCESS) continue /* the race with other process, slot reused */; } /* clean it */ for (unsigned j = i; j < snap_nreaders; j++) { if (lck->mti_readers[j].mr_pid.weak == pid) { mdbx_debug("clear stale reader pid %" PRIuPTR " txn %" PRIaTXN, (size_t)pid, lck->mti_readers[j].mr_txnid.weak); atomic_store32(&lck->mti_readers[j].mr_pid, 0, mo_AcquireRelease); atomic_store32(&lck->mti_readers_refresh_flag, true, mo_AcquireRelease); count++; } } } if (likely(!MDBX_IS_ERROR(rc))) atomic_store64(&lck->mti_reader_check_timestamp, mdbx_osal_monotime(), mo_AcquireRelease); if (rdt_locked < 0) mdbx_rdt_unlock(env); if (pids != pidsbuf_onstask) mdbx_free(pids); if (dead) *dead = count; return rc; } __cold int mdbx_setup_debug(int loglevel, int flags, MDBX_debug_func *logger) { const int rc = mdbx_runtime_flags | (mdbx_loglevel << 16); if (loglevel != MDBX_LOG_DONTCHANGE) mdbx_loglevel = (uint8_t)loglevel; if (flags != MDBX_DBG_DONTCHANGE) { flags &= #if MDBX_DEBUG MDBX_DBG_ASSERT | MDBX_DBG_AUDIT | MDBX_DBG_JITTER | #endif MDBX_DBG_DUMP | MDBX_DBG_LEGACY_MULTIOPEN | MDBX_DBG_LEGACY_OVERLAP; mdbx_runtime_flags = (uint8_t)flags; } if (logger != MDBX_LOGGER_DONTCHANGE) mdbx_debug_logger = logger; return rc; } static txnid_t __cold mdbx_kick_longlived_readers(MDBX_env *env, const txnid_t laggard) { mdbx_debug("DB size maxed out by reading #%" PRIaTXN, laggard); int retry; for (retry = 0; retry < INT_MAX; ++retry) { txnid_t oldest = mdbx_recent_steady_txnid(env); mdbx_assert(env, oldest < env->me_txn0->mt_txnid); mdbx_assert(env, oldest >= laggard); mdbx_assert(env, oldest >= env->me_oldest->weak); if (oldest == laggard || unlikely(!env->me_lck /* without-LCK mode */)) return oldest; if (MDBX_IS_ERROR(mdbx_cleanup_dead_readers(env, false, NULL))) break; MDBX_reader *asleep = nullptr; MDBX_lockinfo *const lck = env->me_lck; uint64_t oldest_retired = UINT64_MAX; const unsigned snap_nreaders = atomic_load32(&lck->mti_numreaders, mo_AcquireRelease); for (unsigned i = 0; i < snap_nreaders; ++i) { retry: if (atomic_load32(&lck->mti_readers[i].mr_pid, mo_AcquireRelease)) { /* mdbx_jitter4testing(true); */ const uint64_t snap_retired = atomic_load64( &lck->mti_readers[i].mr_snapshot_pages_retired, mo_Relaxed); const txnid_t snap_txnid = safe64_read(&lck->mti_readers[i].mr_txnid); if (unlikely(snap_retired != atomic_load64( &lck->mti_readers[i].mr_snapshot_pages_retired, mo_AcquireRelease) || snap_txnid != safe64_read(&lck->mti_readers[i].mr_txnid))) goto retry; if (oldest > snap_txnid && laggard <= /* ignore pending updates */ snap_txnid) { oldest = snap_txnid; oldest_retired = snap_retired; asleep = &lck->mti_readers[i]; } } } if (laggard < oldest || !asleep) { if (retry && env->me_hsr_callback) { /* LY: notify end of hsr-loop */ const txnid_t gap = oldest - laggard; env->me_hsr_callback(env, env->me_txn, 0, 0, laggard, (gap < UINT_MAX) ? (unsigned)gap : UINT_MAX, 0, -retry); } mdbx_notice("hsr-kick: update oldest %" PRIaTXN " -> %" PRIaTXN, env->me_oldest->weak, oldest); mdbx_assert(env, env->me_oldest->weak <= oldest); return atomic_store64(env->me_oldest, oldest, mo_AcquireRelease); } if (!env->me_hsr_callback) break; uint32_t pid = atomic_load32(&asleep->mr_pid, mo_AcquireRelease); uint64_t tid = asleep->mr_tid.weak; if (safe64_read(&asleep->mr_txnid) != laggard || pid <= 0) continue; const MDBX_meta *head_meta = mdbx_meta_head(env); const txnid_t gap = (mdbx_meta_txnid_stable(env, head_meta) - laggard) / MDBX_TXNID_STEP; const uint64_t head_retired = unaligned_peek_u64(4, head_meta->mm_pages_retired); const size_t space = (oldest_retired > head_retired) ? pgno2bytes(env, (pgno_t)(oldest_retired - head_retired)) : 0; int rc = env->me_hsr_callback( env, env->me_txn, pid, (mdbx_tid_t)tid, laggard, (gap < UINT_MAX) ? (unsigned)gap : UINT_MAX, space, retry); if (rc < 0) break; if (rc > 0) { if (rc == 1) { safe64_reset_compare(&asleep->mr_txnid, laggard); } else { safe64_reset(&asleep->mr_txnid, true); atomic_store64(&asleep->mr_tid, 0, mo_Relaxed); atomic_store32(&asleep->mr_pid, 0, mo_Relaxed); } atomic_store32(&lck->mti_readers_refresh_flag, true, mo_AcquireRelease); } } if (retry && env->me_hsr_callback) { /* LY: notify end of hsr-loop */ env->me_hsr_callback(env, env->me_txn, 0, 0, laggard, 0, 0, -retry); } return mdbx_find_oldest(env->me_txn); } __cold int mdbx_env_set_syncbytes(MDBX_env *env, size_t threshold) { return __inline_mdbx_env_set_syncbytes(env, threshold); } __cold int mdbx_env_set_syncperiod(MDBX_env *env, unsigned seconds_16dot16) { return __inline_mdbx_env_set_syncperiod(env, seconds_16dot16); } __cold int mdbx_env_set_hsr(MDBX_env *env, MDBX_hsr_func *hsr) { int rc = check_env(env); if (unlikely(rc != MDBX_SUCCESS)) return rc; env->me_hsr_callback = hsr; return MDBX_SUCCESS; } MDBX_hsr_func *__cold mdbx_env_get_hsr(const MDBX_env *env) { return likely(env && env->me_signature.weak == MDBX_ME_SIGNATURE) ? env->me_hsr_callback : NULL; } #ifdef __SANITIZE_THREAD__ /* LY: avoid tsan-trap by me_txn, mm_last_pg and mt_next_pgno */ __attribute__((__no_sanitize_thread__, __noinline__)) #endif int mdbx_txn_straggler(const MDBX_txn *txn, int *percent) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return (rc > 0) ? -rc : rc; MDBX_env *env = txn->mt_env; if (unlikely((txn->mt_flags & MDBX_TXN_RDONLY) == 0)) { if (percent) *percent = (int)((txn->mt_next_pgno * UINT64_C(100) + txn->mt_end_pgno / 2) / txn->mt_end_pgno); return 0; } txnid_t recent; MDBX_meta *meta; do { meta = mdbx_meta_head(env); recent = mdbx_meta_txnid_fluid(env, meta); if (percent) { const pgno_t maxpg = meta->mm_geo.now; *percent = (int)((meta->mm_geo.next * UINT64_C(100) + maxpg / 2) / maxpg); } } while (unlikely(recent != mdbx_meta_txnid_fluid(env, meta))); txnid_t lag = (recent - txn->mt_txnid) / MDBX_TXNID_STEP; return (lag > INT_MAX) ? INT_MAX : (int)lag; } typedef struct mdbx_walk_ctx { void *mw_user; MDBX_pgvisitor_func *mw_visitor; MDBX_txn *mw_txn; MDBX_cursor *mw_cursor; bool mw_dont_check_keys_ordering; } mdbx_walk_ctx_t; static __cold int mdbx_walk_sdb(mdbx_walk_ctx_t *ctx, MDBX_db *const db, const char *name, int deep); static MDBX_page_type_t walk_page_type(const MDBX_page *mp) { if (mp) switch (mp->mp_flags) { case P_BRANCH: return MDBX_page_branch; case P_LEAF: return MDBX_page_leaf; case P_LEAF | P_LEAF2: return MDBX_page_dupfixed_leaf; case P_OVERFLOW: return MDBX_page_large; case P_META: return MDBX_page_meta; } return MDBX_page_broken; } /* Depth-first tree traversal. */ static __cold int mdbx_walk_tree(mdbx_walk_ctx_t *ctx, const pgno_t pgno, const char *name, int deep, txnid_t parent_txnid) { assert(pgno != P_INVALID); MDBX_page *mp = nullptr; int rc, err = mdbx_page_get(ctx->mw_cursor, pgno, &mp, NULL, parent_txnid); if (err == MDBX_SUCCESS) err = mdbx_page_check(ctx->mw_cursor, mp, 0); MDBX_page_type_t type = walk_page_type(mp); const int nentries = (mp && !IS_OVERFLOW(mp)) ? page_numkeys(mp) : 1; unsigned npages = (mp && IS_OVERFLOW(mp)) ? mp->mp_pages : 1; size_t pagesize = pgno2bytes(ctx->mw_txn->mt_env, npages); size_t header_size = (mp && !IS_LEAF2(mp) && !IS_OVERFLOW(mp)) ? PAGEHDRSZ + mp->mp_lower : PAGEHDRSZ; size_t payload_size = 0; size_t unused_size = (mp && !IS_OVERFLOW(mp) ? page_room(mp) : pagesize - header_size) - payload_size; size_t align_bytes = 0; if (err == MDBX_SUCCESS) { /* LY: Don't use mask here, e.g bitwise * (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP). * Pages should not me marked dirty/loose or otherwise. */ switch (mp->mp_flags) { default: err = MDBX_CORRUPTED; break; case P_BRANCH: if (unlikely(nentries < 2)) err = MDBX_CORRUPTED; case P_LEAF: case P_LEAF | P_LEAF2: break; } } for (int i = 0; err == MDBX_SUCCESS && i < nentries; align_bytes += ((payload_size + align_bytes) & 1), i++) { if (type == MDBX_page_dupfixed_leaf) { /* LEAF2 pages have no mp_ptrs[] or node headers */ payload_size += mp->mp_leaf2_ksize; continue; } MDBX_node *node = page_node(mp, i); payload_size += NODESIZE + node_ks(node); if (type == MDBX_page_branch) { assert(i > 0 || node_ks(node) == 0); continue; } assert(type == MDBX_page_leaf); switch (node_flags(node)) { case 0 /* usual node */: payload_size += node_ds(node); break; case F_BIGDATA /* long data on the large/overflow page */: { payload_size += sizeof(pgno_t); const pgno_t large_pgno = node_largedata_pgno(node); const size_t over_payload = node_ds(node); const size_t over_header = PAGEHDRSZ; npages = 1; MDBX_page *op; err = mdbx_page_get(ctx->mw_cursor, large_pgno, &op, NULL, pp_txnid4chk(mp, ctx->mw_txn)); if (err == MDBX_SUCCESS) err = mdbx_page_check(ctx->mw_cursor, op, 0); if (err == MDBX_SUCCESS) { /* LY: Don't use mask here, e.g bitwise * (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP). * Pages should not me marked dirty/loose or otherwise. */ if (P_OVERFLOW != op->mp_flags) err = bad_page(mp, "wrong page type %d for large data", op->mp_flags); else npages = op->mp_pages; } pagesize = pgno2bytes(ctx->mw_txn->mt_env, npages); const size_t over_unused = pagesize - over_payload - over_header; rc = ctx->mw_visitor(large_pgno, npages, ctx->mw_user, deep, name, pagesize, MDBX_page_large, err, 1, over_payload, over_header, over_unused); if (unlikely(rc != MDBX_SUCCESS)) return (rc == MDBX_RESULT_TRUE) ? MDBX_SUCCESS : rc; } break; case F_SUBDATA /* sub-db */: { const size_t namelen = node_ks(node); payload_size += node_ds(node); if (unlikely(namelen == 0 || node_ds(node) != sizeof(MDBX_db))) err = MDBX_CORRUPTED; } break; case F_SUBDATA | F_DUPDATA /* dupsorted sub-tree */: payload_size += sizeof(MDBX_db); if (unlikely(node_ds(node) != sizeof(MDBX_db))) err = MDBX_CORRUPTED; break; case F_DUPDATA /* short sub-page */: { if (unlikely(node_ds(node) <= PAGEHDRSZ)) { err = MDBX_CORRUPTED; break; } MDBX_page *sp = node_data(node); const int nsubkeys = page_numkeys(sp); size_t subheader_size = IS_LEAF2(sp) ? PAGEHDRSZ : PAGEHDRSZ + sp->mp_lower; size_t subunused_size = page_room(sp); size_t subpayload_size = 0; size_t subalign_bytes = 0; MDBX_page_type_t subtype; switch (sp->mp_flags & ~P_DIRTY /* ignore for sub-pages */) { case P_LEAF | P_SUBP: subtype = MDBX_subpage_leaf; break; case P_LEAF | P_LEAF2 | P_SUBP: subtype = MDBX_subpage_dupfixed_leaf; break; default: subtype = MDBX_subpage_broken; err = MDBX_CORRUPTED; } for (int j = 0; err == MDBX_SUCCESS && j < nsubkeys; subalign_bytes += ((subpayload_size + subalign_bytes) & 1), j++) { if (subtype == MDBX_subpage_dupfixed_leaf) { /* LEAF2 pages have no mp_ptrs[] or node headers */ subpayload_size += sp->mp_leaf2_ksize; } else { assert(subtype == MDBX_subpage_leaf); MDBX_node *subnode = page_node(sp, j); subpayload_size += NODESIZE + node_ks(subnode) + node_ds(subnode); if (unlikely(node_flags(subnode) != 0)) err = MDBX_CORRUPTED; } } rc = ctx->mw_visitor(pgno, 0, ctx->mw_user, deep + 1, name, node_ds(node), subtype, err, nsubkeys, subpayload_size, subheader_size, subunused_size + subalign_bytes); if (unlikely(rc != MDBX_SUCCESS)) return (rc == MDBX_RESULT_TRUE) ? MDBX_SUCCESS : rc; header_size += subheader_size; unused_size += subunused_size; payload_size += subpayload_size; align_bytes += subalign_bytes; } break; default: err = MDBX_CORRUPTED; } } rc = ctx->mw_visitor(pgno, 1, ctx->mw_user, deep, name, ctx->mw_txn->mt_env->me_psize, type, err, nentries, payload_size, header_size, unused_size + align_bytes); if (unlikely(rc != MDBX_SUCCESS)) return (rc == MDBX_RESULT_TRUE) ? MDBX_SUCCESS : rc; for (int i = 0; err == MDBX_SUCCESS && i < nentries; i++) { if (type == MDBX_page_dupfixed_leaf) continue; MDBX_node *node = page_node(mp, i); if (type == MDBX_page_branch) { err = mdbx_walk_tree(ctx, node_pgno(node), name, deep + 1, pp_txnid4chk(mp, ctx->mw_txn)); if (unlikely(err != MDBX_SUCCESS)) { if (err == MDBX_RESULT_TRUE) break; return err; } continue; } assert(type == MDBX_page_leaf); MDBX_db db; switch (node_flags(node)) { default: continue; case F_SUBDATA /* sub-db */: { const size_t namelen = node_ks(node); if (unlikely(namelen == 0 || node_ds(node) != sizeof(MDBX_db))) { err = MDBX_CORRUPTED; break; } char namebuf_onstask[64]; char *const sub_name = (namelen < sizeof(namebuf_onstask)) ? namebuf_onstask : mdbx_malloc(namelen + 1); if (sub_name) { memcpy(sub_name, node_key(node), namelen); sub_name[namelen] = 0; memcpy(&db, node_data(node), sizeof(db)); err = mdbx_walk_sdb(ctx, &db, sub_name, deep + 1); if (sub_name != namebuf_onstask) mdbx_free(sub_name); } else { err = MDBX_ENOMEM; } } break; case F_SUBDATA | F_DUPDATA /* dupsorted sub-tree */: if (unlikely(node_ds(node) != sizeof(MDBX_db) || ctx->mw_cursor->mc_xcursor == NULL)) err = MDBX_CORRUPTED; else { memcpy(&db, node_data(node), sizeof(db)); assert(ctx->mw_cursor->mc_xcursor == &container_of(ctx->mw_cursor, MDBX_cursor_couple, outer)->inner); ctx->mw_cursor = &ctx->mw_cursor->mc_xcursor->mx_cursor; err = mdbx_walk_tree(ctx, db.md_root, name, deep + 1, pp_txnid4chk(mp, ctx->mw_txn)); MDBX_xcursor *inner_xcursor = container_of(ctx->mw_cursor, MDBX_xcursor, mx_cursor); MDBX_cursor_couple *couple = container_of(inner_xcursor, MDBX_cursor_couple, inner); ctx->mw_cursor = &couple->outer; } break; } } return MDBX_SUCCESS; } static __cold int mdbx_walk_sdb(mdbx_walk_ctx_t *ctx, MDBX_db *const db, const char *name, int deep) { if (unlikely(db->md_root == P_INVALID)) return MDBX_SUCCESS; /* empty db */ MDBX_cursor_couple couple; MDBX_dbx dbx = {.md_klen_min = INT_MAX}; uint8_t dbistate = DBI_VALID | DBI_AUDITED; int rc = mdbx_couple_init(&couple, ~0u, ctx->mw_txn, db, &dbx, &dbistate); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (ctx->mw_dont_check_keys_ordering) { couple.outer.mc_flags |= C_SKIPORD; couple.inner.mx_cursor.mc_flags |= C_SKIPORD; } couple.outer.mc_next = ctx->mw_cursor; ctx->mw_cursor = &couple.outer; rc = mdbx_walk_tree(ctx, db->md_root, name, deep, ctx->mw_txn->mt_txnid); ctx->mw_cursor = couple.outer.mc_next; return rc; } __cold int mdbx_env_pgwalk(MDBX_txn *txn, MDBX_pgvisitor_func *visitor, void *user, bool dont_check_keys_ordering) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; mdbx_walk_ctx_t ctx; memset(&ctx, 0, sizeof(ctx)); ctx.mw_txn = txn; ctx.mw_user = user; ctx.mw_visitor = visitor; ctx.mw_dont_check_keys_ordering = dont_check_keys_ordering; rc = visitor(0, NUM_METAS, user, 0, MDBX_PGWALK_META, pgno2bytes(txn->mt_env, NUM_METAS), MDBX_page_meta, MDBX_SUCCESS, NUM_METAS, sizeof(MDBX_meta) * NUM_METAS, PAGEHDRSZ * NUM_METAS, (txn->mt_env->me_psize - sizeof(MDBX_meta) - PAGEHDRSZ) * NUM_METAS); if (!MDBX_IS_ERROR(rc)) rc = mdbx_walk_sdb(&ctx, &txn->mt_dbs[FREE_DBI], MDBX_PGWALK_GC, 0); if (!MDBX_IS_ERROR(rc)) rc = mdbx_walk_sdb(&ctx, &txn->mt_dbs[MAIN_DBI], MDBX_PGWALK_MAIN, 0); return rc; } int mdbx_canary_put(MDBX_txn *txn, const MDBX_canary *canary) { int rc = check_txn_rw(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (likely(canary)) { if (txn->mt_canary.x == canary->x && txn->mt_canary.y == canary->y && txn->mt_canary.z == canary->z) return MDBX_SUCCESS; txn->mt_canary.x = canary->x; txn->mt_canary.y = canary->y; txn->mt_canary.z = canary->z; } txn->mt_canary.v = txn->mt_txnid; txn->mt_flags |= MDBX_TXN_DIRTY; return MDBX_SUCCESS; } int mdbx_canary_get(const MDBX_txn *txn, MDBX_canary *canary) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(canary == NULL)) return MDBX_EINVAL; *canary = txn->mt_canary; return MDBX_SUCCESS; } int mdbx_cursor_on_first(const MDBX_cursor *mc) { if (unlikely(mc == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; if (!(mc->mc_flags & C_INITIALIZED)) return mc->mc_db->md_entries ? MDBX_RESULT_FALSE : MDBX_RESULT_TRUE; for (unsigned i = 0; i < mc->mc_snum; ++i) { if (mc->mc_ki[i]) return MDBX_RESULT_FALSE; } return MDBX_RESULT_TRUE; } int mdbx_cursor_on_last(const MDBX_cursor *mc) { if (unlikely(mc == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; if (!(mc->mc_flags & C_INITIALIZED)) return mc->mc_db->md_entries ? MDBX_RESULT_FALSE : MDBX_RESULT_TRUE; for (unsigned i = 0; i < mc->mc_snum; ++i) { unsigned nkeys = page_numkeys(mc->mc_pg[i]); if (mc->mc_ki[i] < nkeys - 1) return MDBX_RESULT_FALSE; } return MDBX_RESULT_TRUE; } int mdbx_cursor_eof(const MDBX_cursor *mc) { if (unlikely(mc == NULL)) return MDBX_EINVAL; if (unlikely(mc->mc_signature != MDBX_MC_LIVE)) return (mc->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; if ((mc->mc_flags & C_INITIALIZED) == 0) return MDBX_RESULT_TRUE; if (mc->mc_snum == 0) return MDBX_RESULT_TRUE; if ((mc->mc_flags & C_EOF) && mc->mc_ki[mc->mc_top] >= page_numkeys(mc->mc_pg[mc->mc_top])) return MDBX_RESULT_TRUE; return MDBX_RESULT_FALSE; } //------------------------------------------------------------------------------ struct diff_result { ptrdiff_t diff; unsigned level; int root_nkeys; }; /* calculates: r = x - y */ __hot static int cursor_diff(const MDBX_cursor *const __restrict x, const MDBX_cursor *const __restrict y, struct diff_result *const __restrict r) { r->diff = 0; r->level = 0; r->root_nkeys = 0; if (unlikely(x->mc_signature != MDBX_MC_LIVE)) return (x->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; if (unlikely(y->mc_signature != MDBX_MC_LIVE)) return (y->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn(x->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(x->mc_txn != y->mc_txn)) return MDBX_BAD_TXN; if (unlikely(y->mc_dbi != x->mc_dbi)) return MDBX_EINVAL; if (unlikely(!(y->mc_flags & x->mc_flags & C_INITIALIZED))) return MDBX_ENODATA; while (likely(r->level < y->mc_snum && r->level < x->mc_snum)) { if (unlikely(y->mc_pg[r->level] != x->mc_pg[r->level])) return MDBX_PROBLEM; int nkeys = page_numkeys(y->mc_pg[r->level]); assert(nkeys > 0); if (r->level == 0) r->root_nkeys = nkeys; const int limit_ki = nkeys - 1; const int x_ki = x->mc_ki[r->level]; const int y_ki = y->mc_ki[r->level]; r->diff = ((x_ki < limit_ki) ? x_ki : limit_ki) - ((y_ki < limit_ki) ? y_ki : limit_ki); if (r->diff == 0) { r->level += 1; continue; } while (unlikely(r->diff == 1) && likely(r->level + 1 < y->mc_snum && r->level + 1 < x->mc_snum)) { r->level += 1; /* DB'PAGEs: 0------------------>MAX * * CURSORs: y < x * STACK[i ]: | * STACK[+1]: ...y++N|0++x... */ nkeys = page_numkeys(y->mc_pg[r->level]); r->diff = (nkeys - y->mc_ki[r->level]) + x->mc_ki[r->level]; assert(r->diff > 0); } while (unlikely(r->diff == -1) && likely(r->level + 1 < y->mc_snum && r->level + 1 < x->mc_snum)) { r->level += 1; /* DB'PAGEs: 0------------------>MAX * * CURSORs: x < y * STACK[i ]: | * STACK[+1]: ...x--N|0--y... */ nkeys = page_numkeys(x->mc_pg[r->level]); r->diff = -(nkeys - x->mc_ki[r->level]) - y->mc_ki[r->level]; assert(r->diff < 0); } return MDBX_SUCCESS; } r->diff = CMP2INT(x->mc_flags & C_EOF, y->mc_flags & C_EOF); return MDBX_SUCCESS; } __hot static ptrdiff_t estimate(const MDBX_db *db, struct diff_result *const __restrict dr) { /* root: branch-page => scale = leaf-factor * branch-factor^(N-1) * level-1: branch-page(s) => scale = leaf-factor * branch-factor^2 * level-2: branch-page(s) => scale = leaf-factor * branch-factor * level-N: branch-page(s) => scale = leaf-factor * leaf-level: leaf-page(s) => scale = 1 */ ptrdiff_t btree_power = (ptrdiff_t)db->md_depth - 2 - (ptrdiff_t)dr->level; if (btree_power < 0) return dr->diff; ptrdiff_t estimated = (ptrdiff_t)db->md_entries * dr->diff / (ptrdiff_t)db->md_leaf_pages; if (btree_power == 0) return estimated; if (db->md_depth < 4) { assert(dr->level == 0 && btree_power == 1); return (ptrdiff_t)db->md_entries * dr->diff / (ptrdiff_t)dr->root_nkeys; } /* average_branchpage_fillfactor = total(branch_entries) / branch_pages total(branch_entries) = leaf_pages + branch_pages - 1 (root page) */ const size_t log2_fixedpoint = sizeof(size_t) - 1; const size_t half = UINT64_C(1) << (log2_fixedpoint - 1); const size_t factor = ((db->md_leaf_pages + db->md_branch_pages - 1) << log2_fixedpoint) / db->md_branch_pages; while (1) { switch ((size_t)btree_power) { default: { const size_t square = (factor * factor + half) >> log2_fixedpoint; const size_t quad = (square * square + half) >> log2_fixedpoint; do { estimated = estimated * quad + half; estimated >>= log2_fixedpoint; btree_power -= 4; } while (btree_power >= 4); continue; } case 3: estimated = estimated * factor + half; estimated >>= log2_fixedpoint; __fallthrough /* fall through */; case 2: estimated = estimated * factor + half; estimated >>= log2_fixedpoint; __fallthrough /* fall through */; case 1: estimated = estimated * factor + half; estimated >>= log2_fixedpoint; __fallthrough /* fall through */; case 0: if (unlikely(estimated > (ptrdiff_t)db->md_entries)) return (ptrdiff_t)db->md_entries; if (unlikely(estimated < -(ptrdiff_t)db->md_entries)) return -(ptrdiff_t)db->md_entries; return estimated; } } } int mdbx_estimate_distance(const MDBX_cursor *first, const MDBX_cursor *last, ptrdiff_t *distance_items) { if (unlikely(first == NULL || last == NULL || distance_items == NULL)) return MDBX_EINVAL; *distance_items = 0; struct diff_result dr; int rc = cursor_diff(last, first, &dr); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(dr.diff == 0) && F_ISSET(first->mc_db->md_flags & last->mc_db->md_flags, MDBX_DUPSORT | C_INITIALIZED)) { first = &first->mc_xcursor->mx_cursor; last = &last->mc_xcursor->mx_cursor; rc = cursor_diff(first, last, &dr); if (unlikely(rc != MDBX_SUCCESS)) return rc; } if (likely(dr.diff != 0)) *distance_items = estimate(first->mc_db, &dr); return MDBX_SUCCESS; } int mdbx_estimate_move(const MDBX_cursor *cursor, MDBX_val *key, MDBX_val *data, MDBX_cursor_op move_op, ptrdiff_t *distance_items) { if (unlikely(cursor == NULL || distance_items == NULL || move_op == MDBX_GET_CURRENT || move_op == MDBX_GET_MULTIPLE)) return MDBX_EINVAL; if (unlikely(cursor->mc_signature != MDBX_MC_LIVE)) return (cursor->mc_signature == MDBX_MC_READY4CLOSE) ? MDBX_EINVAL : MDBX_EBADSIGN; int rc = check_txn(cursor->mc_txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (!(cursor->mc_flags & C_INITIALIZED)) return MDBX_ENODATA; MDBX_cursor_couple next; cursor_copy_internal(cursor, &next.outer); next.outer.mc_xcursor = NULL; if (cursor->mc_db->md_flags & MDBX_DUPSORT) { next.outer.mc_xcursor = &next.inner; rc = mdbx_xcursor_init0(&next.outer); if (unlikely(rc != MDBX_SUCCESS)) return rc; MDBX_xcursor *mx = &container_of(cursor, MDBX_cursor_couple, outer)->inner; cursor_copy_internal(&mx->mx_cursor, &next.inner.mx_cursor); } MDBX_val stub = {0, 0}; if (data == NULL) { const unsigned mask = 1 << MDBX_GET_BOTH | 1 << MDBX_GET_BOTH_RANGE | 1 << MDBX_SET_KEY; if (unlikely(mask & (1 << move_op))) return MDBX_EINVAL; data = &stub; } if (key == NULL) { const unsigned mask = 1 << MDBX_GET_BOTH | 1 << MDBX_GET_BOTH_RANGE | 1 << MDBX_SET_KEY | 1 << MDBX_SET | 1 << MDBX_SET_RANGE; if (unlikely(mask & (1 << move_op))) return MDBX_EINVAL; key = &stub; } rc = mdbx_cursor_get(&next.outer, key, data, move_op); if (unlikely(rc != MDBX_SUCCESS && (rc != MDBX_NOTFOUND || !(next.outer.mc_flags & C_INITIALIZED)))) return rc; return mdbx_estimate_distance(cursor, &next.outer, distance_items); } int mdbx_estimate_range(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *begin_key, MDBX_val *begin_data, MDBX_val *end_key, MDBX_val *end_data, ptrdiff_t *size_items) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!size_items)) return MDBX_EINVAL; if (unlikely(begin_data && (begin_key == NULL || begin_key == MDBX_EPSILON))) return MDBX_EINVAL; if (unlikely(end_data && (end_key == NULL || end_key == MDBX_EPSILON))) return MDBX_EINVAL; if (unlikely(begin_key == MDBX_EPSILON && end_key == MDBX_EPSILON)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; MDBX_cursor_couple begin; /* LY: first, initialize cursor to refresh a DB in case it have DB_STALE */ rc = mdbx_cursor_init(&begin.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(begin.outer.mc_db->md_entries == 0)) { *size_items = 0; return MDBX_SUCCESS; } if (!begin_key) { if (unlikely(!end_key)) { /* LY: FIRST..LAST case */ *size_items = (ptrdiff_t)begin.outer.mc_db->md_entries; return MDBX_SUCCESS; } MDBX_val stub = {0, 0}; rc = mdbx_cursor_first(&begin.outer, &stub, &stub); if (unlikely(end_key == MDBX_EPSILON)) { /* LY: FIRST..+epsilon case */ return (rc == MDBX_SUCCESS) ? mdbx_cursor_count(&begin.outer, (size_t *)size_items) : rc; } } else { if (unlikely(begin_key == MDBX_EPSILON)) { if (end_key == NULL) { /* LY: -epsilon..LAST case */ MDBX_val stub = {0, 0}; rc = mdbx_cursor_last(&begin.outer, &stub, &stub); return (rc == MDBX_SUCCESS) ? mdbx_cursor_count(&begin.outer, (size_t *)size_items) : rc; } /* LY: -epsilon..value case */ assert(end_key != MDBX_EPSILON); begin_key = end_key; } else if (unlikely(end_key == MDBX_EPSILON)) { /* LY: value..+epsilon case */ assert(begin_key != MDBX_EPSILON); end_key = begin_key; } if (end_key && !begin_data && !end_data && (begin_key == end_key || begin.outer.mc_dbx->md_cmp(begin_key, end_key) == 0)) { /* LY: single key case */ int exact = 0; rc = mdbx_cursor_set(&begin.outer, begin_key, NULL, MDBX_SET, &exact); if (unlikely(rc != MDBX_SUCCESS)) { *size_items = 0; return (rc == MDBX_NOTFOUND) ? MDBX_SUCCESS : rc; } *size_items = 1; if (begin.outer.mc_xcursor != NULL) { MDBX_node *node = page_node(begin.outer.mc_pg[begin.outer.mc_top], begin.outer.mc_ki[begin.outer.mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { /* LY: return the number of duplicates for given key */ mdbx_tassert(txn, begin.outer.mc_xcursor == &begin.inner && (begin.inner.mx_cursor.mc_flags & C_INITIALIZED)); *size_items = (sizeof(*size_items) >= sizeof(begin.inner.mx_db.md_entries) || begin.inner.mx_db.md_entries <= PTRDIFF_MAX) ? (size_t)begin.inner.mx_db.md_entries : PTRDIFF_MAX; } } return MDBX_SUCCESS; } else { rc = mdbx_cursor_set(&begin.outer, begin_key, begin_data, begin_data ? MDBX_GET_BOTH_RANGE : MDBX_SET_RANGE, NULL); } } if (unlikely(rc != MDBX_SUCCESS)) { if (rc != MDBX_NOTFOUND || !(begin.outer.mc_flags & C_INITIALIZED)) return rc; } MDBX_cursor_couple end; rc = mdbx_cursor_init(&end.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (!end_key) { MDBX_val stub = {0, 0}; rc = mdbx_cursor_last(&end.outer, &stub, &stub); } else { rc = mdbx_cursor_set(&end.outer, end_key, end_data, end_data ? MDBX_GET_BOTH_RANGE : MDBX_SET_RANGE, NULL); } if (unlikely(rc != MDBX_SUCCESS)) { if (rc != MDBX_NOTFOUND || !(end.outer.mc_flags & C_INITIALIZED)) return rc; } rc = mdbx_estimate_distance(&begin.outer, &end.outer, size_items); if (unlikely(rc != MDBX_SUCCESS)) return rc; assert(*size_items >= -(ptrdiff_t)begin.outer.mc_db->md_entries && *size_items <= (ptrdiff_t)begin.outer.mc_db->md_entries); #if 0 /* LY: Was decided to returns as-is (i.e. negative) the estimation \ * results for an inverted ranges. */ /* Commit 8ddfd1f34ad7cf7a3c4aa75d2e248ca7e639ed63 Change-Id: If59eccf7311123ab6384c4b93f9b1fed5a0a10d1 */ if (*size_items < 0) { /* LY: inverted range case */ *size_items += (ptrdiff_t)begin.outer.mc_db->md_entries; } else if (*size_items == 0 && begin_key && end_key) { int cmp = begin.outer.mc_dbx->md_cmp(&origin_begin_key, &origin_end_key); if (cmp == 0 && (begin.inner.mx_cursor.mc_flags & C_INITIALIZED) && begin_data && end_data) cmp = begin.outer.mc_dbx->md_dcmp(&origin_begin_data, &origin_end_data); if (cmp > 0) { /* LY: inverted range case with empty scope */ *size_items = (ptrdiff_t)begin.outer.mc_db->md_entries; } } assert(*size_items >= 0 && *size_items <= (ptrdiff_t)begin.outer.mc_db->md_entries); #endif return MDBX_SUCCESS; } //------------------------------------------------------------------------------ /* Позволяет обновить или удалить существующую запись с получением * в old_data предыдущего значения данных. При этом если new_data равен * нулю, то выполняется удаление, иначе обновление/вставка. * * Текущее значение может находиться в уже измененной (грязной) странице. * В этом случае страница будет перезаписана при обновлении, а само старое * значение утрачено. Поэтому исходно в old_data должен быть передан * дополнительный буфер для копирования старого значения. * Если переданный буфер слишком мал, то функция вернет -1, установив * old_data->iov_len в соответствующее значение. * * Для не-уникальных ключей также возможен второй сценарий использования, * когда посредством old_data из записей с одинаковым ключом для * удаления/обновления выбирается конкретная. Для выбора этого сценария * во flags следует одновременно указать MDBX_CURRENT и MDBX_NOOVERWRITE. * Именно эта комбинация выбрана, так как она лишена смысла, и этим позволяет * идентифицировать запрос такого сценария. * * Функция может быть замещена соответствующими операциями с курсорами * после двух доработок (TODO): * - внешняя аллокация курсоров, в том числе на стеке (без malloc). * - получения статуса страницы по адресу (знать о P_DIRTY). */ int mdbx_replace_ex(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, MDBX_val *new_data, MDBX_val *old_data, MDBX_put_flags_t flags, MDBX_preserve_func preserver, void *preserver_context) { int rc = check_txn_rw(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!key || !old_data || old_data == new_data)) return MDBX_EINVAL; if (unlikely(old_data->iov_base == NULL && old_data->iov_len)) return MDBX_EINVAL; if (unlikely(new_data == NULL && (flags & (MDBX_CURRENT | MDBX_RESERVE)) != MDBX_CURRENT)) return MDBX_EINVAL; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(flags & ~(MDBX_NOOVERWRITE | MDBX_NODUPDATA | MDBX_ALLDUPS | MDBX_RESERVE | MDBX_APPEND | MDBX_APPENDDUP | MDBX_CURRENT))) return MDBX_EINVAL; MDBX_cursor_couple cx; rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; cx.outer.mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = &cx.outer; MDBX_val present_key = *key; if (F_ISSET(flags, MDBX_CURRENT | MDBX_NOOVERWRITE)) { /* в old_data значение для выбора конкретного дубликата */ if (unlikely(!(txn->mt_dbs[dbi].md_flags & MDBX_DUPSORT))) { rc = MDBX_EINVAL; goto bailout; } /* убираем лишний бит, он был признаком запрошенного режима */ flags -= MDBX_NOOVERWRITE; rc = mdbx_cursor_get(&cx.outer, &present_key, old_data, MDBX_GET_BOTH); if (rc != MDBX_SUCCESS) goto bailout; } else { /* в old_data буфер для сохранения предыдущего значения */ if (unlikely(new_data && old_data->iov_base == new_data->iov_base)) return MDBX_EINVAL; MDBX_val present_data; rc = mdbx_cursor_get(&cx.outer, &present_key, &present_data, MDBX_SET_KEY); if (unlikely(rc != MDBX_SUCCESS)) { old_data->iov_base = NULL; old_data->iov_len = 0; if (rc != MDBX_NOTFOUND || (flags & MDBX_CURRENT)) goto bailout; } else if (flags & MDBX_NOOVERWRITE) { rc = MDBX_KEYEXIST; *old_data = present_data; goto bailout; } else { MDBX_page *page = cx.outer.mc_pg[cx.outer.mc_top]; if (txn->mt_dbs[dbi].md_flags & MDBX_DUPSORT) { if (flags & MDBX_CURRENT) { /* disallow update/delete for multi-values */ MDBX_node *node = page_node(page, cx.outer.mc_ki[cx.outer.mc_top]); if (F_ISSET(node_flags(node), F_DUPDATA)) { mdbx_tassert(txn, XCURSOR_INITED(&cx.outer) && cx.outer.mc_xcursor->mx_db.md_entries > 1); if (cx.outer.mc_xcursor->mx_db.md_entries > 1) { rc = MDBX_EMULTIVAL; goto bailout; } } /* В оригинальной LMDB флажок MDBX_CURRENT здесь приведет * к замене данных без учета MDBX_DUPSORT сортировки, * но здесь это в любом случае допустимо, так как мы * проверили что для ключа есть только одно значение. */ } } if (IS_DIRTY(page)) { if (new_data && cmp_lenfast(&present_data, new_data) == 0) { /* если данные совпадают, то ничего делать не надо */ *old_data = *new_data; goto bailout; } rc = preserver ? preserver(preserver_context, old_data, present_data.iov_base, present_data.iov_len) : MDBX_SUCCESS; if (unlikely(rc != MDBX_SUCCESS)) goto bailout; } else { *old_data = present_data; } flags |= MDBX_CURRENT; } } if (likely(new_data)) rc = mdbx_cursor_put(&cx.outer, key, new_data, flags); else rc = mdbx_cursor_del(&cx.outer, flags & MDBX_ALLDUPS); bailout: txn->tw.cursors[dbi] = cx.outer.mc_next; return rc; } static int default_value_preserver(void *context, MDBX_val *target, const void *src, size_t bytes) { (void)context; if (unlikely(target->iov_len < bytes)) { target->iov_base = nullptr; target->iov_len = bytes; return MDBX_RESULT_TRUE; } memcpy(target->iov_base, src, target->iov_len = bytes); return MDBX_SUCCESS; } int mdbx_replace(MDBX_txn *txn, MDBX_dbi dbi, const MDBX_val *key, MDBX_val *new_data, MDBX_val *old_data, MDBX_put_flags_t flags) { return mdbx_replace_ex(txn, dbi, key, new_data, old_data, flags, default_value_preserver, nullptr); } /* Функция сообщает находится ли указанный адрес в "грязной" странице у * заданной пишущей транзакции. В конечном счете это позволяет избавиться от * лишнего копирования данных из НЕ-грязных страниц. * * "Грязные" страницы - это те, которые уже были изменены в ходе пишущей * транзакции. Соответственно, какие-либо дальнейшие изменения могут привести * к перезаписи таких страниц. Поэтому все функции, выполняющие изменения, в * качестве аргументов НЕ должны получать указатели на данные в таких * страницах. В свою очередь "НЕ грязные" страницы перед модификацией будут * скопированы. * * Другими словами, данные из "грязных" страниц должны быть либо скопированы * перед передачей в качестве аргументов для дальнейших модификаций, либо * отвергнуты на стадии проверки корректности аргументов. * * Таким образом, функция позволяет как избавиться от лишнего копирования, * так и выполнить более полную проверку аргументов. * * ВАЖНО: Передаваемый указатель должен указывать на начало данных. Только * так гарантируется что актуальный заголовок страницы будет физически * расположен в той-же странице памяти, в том числе для многостраничных * P_OVERFLOW страниц с длинными данными. */ int mdbx_is_dirty(const MDBX_txn *txn, const void *ptr) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; const MDBX_env *env = txn->mt_env; const ptrdiff_t offset = (uint8_t *)ptr - env->me_map; if (offset >= 0) { const pgno_t pgno = bytes2pgno(env, offset); if (likely(pgno < txn->mt_next_pgno)) { if (txn->mt_flags & MDBX_TXN_RDONLY) return MDBX_RESULT_FALSE; const MDBX_page *page = pgno2page(env, pgno); if (unlikely(page->mp_pgno != pgno)) { /* The ptr pointed into middle of a large page, * not to the beginning of a data. */ return MDBX_EINVAL; } if (unlikely(page->mp_flags & (P_DIRTY | P_LOOSE | P_KEEP | P_META))) return MDBX_RESULT_TRUE; if (likely(txn->tw.spill_pages == nullptr)) return MDBX_RESULT_FALSE; return mdbx_pnl_exist(txn->tw.spill_pages, pgno << 1) ? MDBX_RESULT_TRUE : MDBX_RESULT_FALSE; } if ((size_t)offset < env->me_dxb_mmap.limit) { /* Указатель адресует что-то в пределах mmap, но за границей * распределенных страниц. Такое может случится если mdbx_is_dirty() * вызывается после операции, в ходе которой грязная страница попала * в loose и затем была возвращена в нераспределенное пространство. */ return MDBX_RESULT_TRUE; } } /* Страница вне используемого mmap-диапазона, т.е. либо в функцию был * передан некорректный адрес, либо адрес в теневой странице, которая была * выделена посредством malloc(). * * Для режима WRITE_MAP режима страница однозначно "не грязная", * а для режимов без WRITE_MAP следует просматривать списки dirty * и spilled страниц у каких-либо транзакций (в том числе дочерних). * * Поэтому для WRITE_MAP возвращаем false, а для остальных режимов * всегда true. Такая логика имеет ряд преимуществ: * - не тратим время на просмотр списков; * - результат всегда безопасен (может быть ложно-положительным, * но не ложно-отрицательным); * - результат не зависит от вложенности транзакций и от относительного * положения переданной транзакции в этой рекурсии. */ return (env->me_flags & MDBX_WRITEMAP) ? MDBX_RESULT_FALSE : MDBX_RESULT_TRUE; } int mdbx_dbi_sequence(MDBX_txn *txn, MDBX_dbi dbi, uint64_t *result, uint64_t increment) { int rc = check_txn(txn, MDBX_TXN_BLOCKED); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (unlikely(!mdbx_txn_dbi_exists(txn, dbi, DBI_USRVALID))) return MDBX_BAD_DBI; if (unlikely(TXN_DBI_CHANGED(txn, dbi))) return MDBX_BAD_DBI; if (unlikely(txn->mt_dbistate[dbi] & DBI_STALE)) { rc = mdbx_fetch_sdb(txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; } MDBX_db *dbs = &txn->mt_dbs[dbi]; if (likely(result)) *result = dbs->md_seq; if (likely(increment > 0)) { if (unlikely(txn->mt_flags & MDBX_TXN_RDONLY)) return MDBX_EACCESS; uint64_t new = dbs->md_seq + increment; if (unlikely(new < increment)) return MDBX_RESULT_TRUE; mdbx_tassert(txn, new > dbs->md_seq); dbs->md_seq = new; txn->mt_flags |= MDBX_TXN_DIRTY; txn->mt_dbistate[dbi] |= DBI_DIRTY; } return MDBX_SUCCESS; } /*----------------------------------------------------------------------------*/ __cold MDBX_NOTHROW_CONST_FUNCTION intptr_t mdbx_limits_pgsize_min(void) { return __inline_mdbx_limits_pgsize_min(); } __cold MDBX_NOTHROW_CONST_FUNCTION intptr_t mdbx_limits_pgsize_max(void) { return __inline_mdbx_limits_pgsize_max(); } __cold intptr_t mdbx_limits_dbsize_min(intptr_t pagesize) { if (pagesize < 1) pagesize = (intptr_t)mdbx_default_pagesize(); else if (unlikely(pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2((size_t)pagesize))) return -1; return MIN_PAGENO * pagesize; } __cold intptr_t mdbx_limits_dbsize_max(intptr_t pagesize) { if (pagesize < 1) pagesize = (intptr_t)mdbx_default_pagesize(); else if (unlikely(pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2((size_t)pagesize))) return -1; STATIC_ASSERT(MAX_MAPSIZE < INTPTR_MAX); const uint64_t limit = MAX_PAGENO * (uint64_t)pagesize; return (limit < (intptr_t)MAX_MAPSIZE) ? (intptr_t)limit : (intptr_t)MAX_MAPSIZE; } __cold intptr_t mdbx_limits_txnsize_max(intptr_t pagesize) { if (pagesize < 1) pagesize = (intptr_t)mdbx_default_pagesize(); else if (unlikely(pagesize < (intptr_t)MIN_PAGESIZE || pagesize > (intptr_t)MAX_PAGESIZE || !is_powerof2((size_t)pagesize))) return -1; STATIC_ASSERT(MAX_MAPSIZE < INTPTR_MAX); const uint64_t limit = pagesize * (uint64_t)(MDBX_PGL_LIMIT - 1); return (limit < (intptr_t)MAX_MAPSIZE) ? (intptr_t)limit : (intptr_t)MAX_MAPSIZE; } /*** Key-making functions to avoid custom comparators *************************/ static __always_inline double key2double(const int64_t key) { union { uint64_t u; double f; } casting; casting.u = (key < 0) ? key + UINT64_C(0x8000000000000000) : UINT64_C(0xffffFFFFffffFFFF) - key; return casting.f; } static __always_inline uint64_t double2key(const double *const ptr) { STATIC_ASSERT(sizeof(double) == sizeof(int64_t)); const int64_t i = *(const int64_t *)ptr; const uint64_t u = (i < 0) ? UINT64_C(0xffffFFFFffffFFFF) - i : i + UINT64_C(0x8000000000000000); if (mdbx_assert_enabled()) { const double f = key2double(u); assert(memcmp(&f, ptr, 8) == 0); } return u; } static __always_inline float key2float(const int32_t key) { union { uint32_t u; float f; } casting; casting.u = (key < 0) ? key + UINT32_C(0x80000000) : UINT32_C(0xffffFFFF) - key; return casting.f; } static __always_inline uint32_t float2key(const float *const ptr) { STATIC_ASSERT(sizeof(float) == sizeof(int32_t)); const int32_t i = *(const int32_t *)ptr; const uint32_t u = (i < 0) ? UINT32_C(0xffffFFFF) - i : i + UINT32_C(0x80000000); if (mdbx_assert_enabled()) { const float f = key2float(u); assert(memcmp(&f, ptr, 4) == 0); } return u; } uint64_t mdbx_key_from_double(const double ieee754_64bit) { return double2key(&ieee754_64bit); } uint64_t mdbx_key_from_ptrdouble(const double *const ieee754_64bit) { return double2key(ieee754_64bit); } uint32_t mdbx_key_from_float(const float ieee754_32bit) { return float2key(&ieee754_32bit); } uint32_t mdbx_key_from_ptrfloat(const float *const ieee754_32bit) { return float2key(ieee754_32bit); } MDBX_NOTHROW_CONST_FUNCTION uint64_t mdbx_key_from_int64(const int64_t i64) { return __inline_mdbx_key_from_int64(i64); } MDBX_NOTHROW_CONST_FUNCTION uint32_t mdbx_key_from_int32(const int32_t i32) { return __inline_mdbx_key_from_int32(i32); } #define IEEE754_DOUBLE_MANTISSA_SIZE 52 #define IEEE754_DOUBLE_EXPONENTA_BIAS 0x3FF #define IEEE754_DOUBLE_EXPONENTA_MAX 0x7FF #define IEEE754_DOUBLE_IMPLICIT_LEAD UINT64_C(0x0010000000000000) #define IEEE754_DOUBLE_MANTISSA_MASK UINT64_C(0x000FFFFFFFFFFFFF) #define IEEE754_DOUBLE_MANTISSA_AMAX UINT64_C(0x001FFFFFFFFFFFFF) static __inline int clz64(uint64_t value) { #if __GNUC_PREREQ(4, 1) || __has_builtin(__builtin_clzl) if (sizeof(value) == sizeof(int)) return __builtin_clz(value); if (sizeof(value) == sizeof(long)) return __builtin_clzl(value); #if (defined(__SIZEOF_LONG_LONG__) && __SIZEOF_LONG_LONG__ == 8) || \ __has_builtin(__builtin_clzll) return __builtin_clzll(value); #endif /* have(long long) && long long == uint64_t */ #endif /* GNU C */ #if defined(_MSC_VER) unsigned long index; #if defined(_M_AMD64) || defined(_M_ARM64) || defined(_M_X64) _BitScanReverse64(&index, value); return 63 - index; #else if (value > UINT32_MAX) { _BitScanReverse(&index, (uint32_t)(value >> 32)); return 31 - index; } _BitScanReverse(&index, (uint32_t)value); return 63 - index; #endif #endif /* MSVC */ value |= value >> 1; value |= value >> 2; value |= value >> 4; value |= value >> 8; value |= value >> 16; value |= value >> 32; static const uint8_t debruijn_clz64[64] = { 63, 16, 62, 7, 15, 36, 61, 3, 6, 14, 22, 26, 35, 47, 60, 2, 9, 5, 28, 11, 13, 21, 42, 19, 25, 31, 34, 40, 46, 52, 59, 1, 17, 8, 37, 4, 23, 27, 48, 10, 29, 12, 43, 20, 32, 41, 53, 18, 38, 24, 49, 30, 44, 33, 54, 39, 50, 45, 55, 51, 56, 57, 58, 0}; return debruijn_clz64[value * UINT64_C(0x03F79D71B4CB0A89) >> 58]; } static __inline uint64_t round_mantissa(const uint64_t u64, int shift) { assert(shift < 0 && u64 > 0); shift = -shift; const unsigned half = 1 << (shift - 1); const unsigned lsb = 1 & (unsigned)(u64 >> shift); const unsigned tie2even = 1 ^ lsb; return (u64 + half - tie2even) >> shift; } uint64_t mdbx_key_from_jsonInteger(const int64_t json_integer) { const uint64_t bias = UINT64_C(0x8000000000000000); if (json_integer > 0) { const uint64_t u64 = json_integer; int shift = clz64(u64) - (64 - IEEE754_DOUBLE_MANTISSA_SIZE - 1); uint64_t mantissa = u64 << shift; if (unlikely(shift < 0)) { mantissa = round_mantissa(u64, shift); if (mantissa > IEEE754_DOUBLE_MANTISSA_AMAX) mantissa = round_mantissa(u64, --shift); } assert(mantissa >= IEEE754_DOUBLE_IMPLICIT_LEAD && mantissa <= IEEE754_DOUBLE_MANTISSA_AMAX); const uint64_t exponent = IEEE754_DOUBLE_EXPONENTA_BIAS + IEEE754_DOUBLE_MANTISSA_SIZE - shift; assert(exponent > 0 && exponent <= IEEE754_DOUBLE_EXPONENTA_MAX); const uint64_t key = bias + (exponent << IEEE754_DOUBLE_MANTISSA_SIZE) + (mantissa - IEEE754_DOUBLE_IMPLICIT_LEAD); #if !defined(_MSC_VER) || \ defined( \ _DEBUG) /* Workaround for MSVC error LNK2019: unresolved external \ symbol __except1 referenced in function __ftol3_except */ assert(key == mdbx_key_from_double((double)json_integer)); #endif /* Workaround for MSVC */ return key; } if (json_integer < 0) { const uint64_t u64 = -json_integer; int shift = clz64(u64) - (64 - IEEE754_DOUBLE_MANTISSA_SIZE - 1); uint64_t mantissa = u64 << shift; if (unlikely(shift < 0)) { mantissa = round_mantissa(u64, shift); if (mantissa > IEEE754_DOUBLE_MANTISSA_AMAX) mantissa = round_mantissa(u64, --shift); } assert(mantissa >= IEEE754_DOUBLE_IMPLICIT_LEAD && mantissa <= IEEE754_DOUBLE_MANTISSA_AMAX); const uint64_t exponent = IEEE754_DOUBLE_EXPONENTA_BIAS + IEEE754_DOUBLE_MANTISSA_SIZE - shift; assert(exponent > 0 && exponent <= IEEE754_DOUBLE_EXPONENTA_MAX); const uint64_t key = bias - 1 - (exponent << IEEE754_DOUBLE_MANTISSA_SIZE) - (mantissa - IEEE754_DOUBLE_IMPLICIT_LEAD); #if !defined(_MSC_VER) || \ defined( \ _DEBUG) /* Workaround for MSVC error LNK2019: unresolved external \ symbol __except1 referenced in function __ftol3_except */ assert(key == mdbx_key_from_double((double)json_integer)); #endif /* Workaround for MSVC */ return key; } return bias; } int64_t mdbx_jsonInteger_from_key(const MDBX_val v) { assert(v.iov_len == 8); const uint64_t key = unaligned_peek_u64(2, v.iov_base); const uint64_t bias = UINT64_C(0x8000000000000000); const uint64_t covalent = (key > bias) ? key - bias : bias - key - 1; const int shift = IEEE754_DOUBLE_EXPONENTA_BIAS + 63 - (IEEE754_DOUBLE_EXPONENTA_MAX & (int)(covalent >> IEEE754_DOUBLE_MANTISSA_SIZE)); if (unlikely(shift < 1)) return (key < bias) ? INT64_MIN : INT64_MAX; if (unlikely(shift > 63)) return 0; const uint64_t unscaled = ((covalent & IEEE754_DOUBLE_MANTISSA_MASK) << (63 - IEEE754_DOUBLE_MANTISSA_SIZE)) + bias; const int64_t absolute = unscaled >> shift; const int64_t value = (key < bias) ? -absolute : absolute; assert(key == mdbx_key_from_jsonInteger(value) || (mdbx_key_from_jsonInteger(value - 1) < key && key < mdbx_key_from_jsonInteger(value + 1))); return value; } double mdbx_double_from_key(const MDBX_val v) { assert(v.iov_len == 8); return key2double(unaligned_peek_u64(2, v.iov_base)); } float mdbx_float_from_key(const MDBX_val v) { assert(v.iov_len == 4); return key2float(unaligned_peek_u32(2, v.iov_base)); } int32_t mdbx_int32_from_key(const MDBX_val v) { assert(v.iov_len == 4); return (int32_t)(unaligned_peek_u32(2, v.iov_base) - UINT32_C(0x80000000)); } int64_t mdbx_int64_from_key(const MDBX_val v) { assert(v.iov_len == 8); return (int64_t)(unaligned_peek_u64(2, v.iov_base) - UINT64_C(0x8000000000000000)); } __cold MDBX_cmp_func *mdbx_get_keycmp(unsigned flags) { return get_default_keycmp(flags); } __cold MDBX_cmp_func *mdbx_get_datacmp(unsigned flags) { return get_default_datacmp(flags); } __cold int mdbx_env_set_option(MDBX_env *env, const MDBX_option_t option, const uint64_t value) { int err = check_env(env); if (unlikely(err != MDBX_SUCCESS)) return err; const bool lock_needed = (env->me_map && env->me_txn0 && env->me_txn0->mt_owner != mdbx_thread_self()); bool should_unlock = false; switch (option) { case MDBX_opt_sync_bytes: if (unlikely(env->me_flags & MDBX_RDONLY)) return MDBX_EACCESS; if (unlikely(!env->me_autosync_threshold)) return MDBX_EPERM; if (sizeof(value) > sizeof(size_t) && unlikely(value != (size_t)value)) return MDBX_TOO_LARGE; if (atomic_store32(env->me_autosync_threshold, bytes2pgno(env, (size_t)value + env->me_psize - 1), mo_Relaxed) != 0) { err = mdbx_env_sync_poll(env); if (unlikely(MDBX_IS_ERROR(err))) return err; } break; case MDBX_opt_sync_period: if (unlikely(env->me_flags & MDBX_RDONLY)) return MDBX_EACCESS; if (unlikely(!env->me_autosync_period)) return MDBX_EPERM; if (unlikely(value > UINT32_MAX)) return MDBX_TOO_LARGE; if (atomic_store64(env->me_autosync_period, mdbx_osal_16dot16_to_monotime((uint32_t)value), mo_Relaxed) != 0) { err = mdbx_env_sync_poll(env); if (unlikely(MDBX_IS_ERROR(err))) return err; } break; case MDBX_opt_max_db: if (unlikely(value > MDBX_MAX_DBI)) return MDBX_EINVAL; if (unlikely(env->me_map)) return MDBX_EPERM; env->me_maxdbs = (unsigned)value + CORE_DBS; break; case MDBX_opt_max_readers: if (unlikely(value < 1 || value > MDBX_READERS_LIMIT)) return MDBX_EINVAL; if (unlikely(env->me_map)) return MDBX_EPERM; env->me_maxreaders = (unsigned)value; break; case MDBX_opt_dp_reserve_limit: if (unlikely(value > INT_MAX)) return MDBX_EINVAL; if (env->me_options.dp_reserve_limit != (unsigned)value) { if (lock_needed) { err = mdbx_txn_lock(env, false); if (unlikely(err != MDBX_SUCCESS)) return err; should_unlock = true; } env->me_options.dp_reserve_limit = (unsigned)value; while (env->me_dp_reserve_len > env->me_options.dp_reserve_limit) { mdbx_assert(env, env->me_dp_reserve != NULL); MDBX_page *dp = env->me_dp_reserve; ASAN_UNPOISON_MEMORY_REGION(dp, env->me_psize); VALGRIND_MAKE_MEM_DEFINED(&dp->mp_next, sizeof(dp->mp_next)); env->me_dp_reserve = dp->mp_next; VALGRIND_MEMPOOL_FREE(env, dp); mdbx_free(dp); env->me_dp_reserve_len -= 1; } } break; case MDBX_opt_rp_augment_limit: if (unlikely(value > MDBX_PGL_LIMIT)) return MDBX_EINVAL; env->me_options.rp_augment_limit = (unsigned)value; break; case MDBX_opt_txn_dp_limit: case MDBX_opt_txn_dp_initial: if (unlikely(value > MDBX_PGL_LIMIT || value < CURSOR_STACK * 4 || value > bytes2pgno(env, env->me_dbgeo.upper) - NUM_METAS)) return MDBX_EINVAL; if (unlikely(env->me_txn0 == NULL)) return MDBX_EACCESS; if (lock_needed) { err = mdbx_txn_lock(env, false); if (unlikely(err != MDBX_SUCCESS)) return err; should_unlock = true; } if (env->me_txn) err = MDBX_EPERM /* unable change during transaction */; else { mdbx_dpl_clear(env->me_txn0->tw.dirtylist); const unsigned value32 = (unsigned)value; if (option == MDBX_opt_txn_dp_initial && env->me_options.dp_initial != value32) { if (env->me_options.dp_limit < value32) env->me_options.dp_limit = value32; if (env->me_txn0->tw.dirtylist->allocated < value32 && !mdbx_dpl_reserve(env->me_txn0, value32)) err = MDBX_ENOMEM; else env->me_options.dp_initial = value32; } if (option == MDBX_opt_txn_dp_limit && env->me_options.dp_limit != value32) { if (env->me_txn0->tw.dirtylist->allocated > value32 && !mdbx_dpl_reserve(env->me_txn0, value32)) err = MDBX_ENOMEM; else { if (env->me_options.dp_initial > value32) env->me_options.dp_initial = value32; env->me_options.dp_limit = value32; } } } break; case MDBX_opt_spill_max_denominator: if (unlikely(value > 255)) return MDBX_EINVAL; env->me_options.spill_max_denominator = (uint8_t)value; break; case MDBX_opt_spill_min_denominator: if (unlikely(value > 255)) return MDBX_EINVAL; env->me_options.spill_min_denominator = (uint8_t)value; break; case MDBX_opt_spill_parent4child_denominator: if (unlikely(value > 255)) return MDBX_EINVAL; env->me_options.spill_parent4child_denominator = (uint8_t)value; break; case MDBX_opt_loose_limit: if (unlikely(value > 255)) return MDBX_EINVAL; env->me_options.dp_loose_limit = (uint8_t)value; break; default: return MDBX_EINVAL; } if (should_unlock) mdbx_txn_unlock(env); return err; } __cold int mdbx_env_get_option(const MDBX_env *env, const MDBX_option_t option, uint64_t *value) { int err = check_env(env); if (unlikely(err != MDBX_SUCCESS)) return err; if (unlikely(!value)) return MDBX_EINVAL; switch (option) { case MDBX_opt_sync_bytes: if (unlikely(!env->me_autosync_threshold)) return MDBX_EPERM; *value = pgno2bytes(env, atomic_load32(env->me_autosync_threshold, mo_Relaxed)); break; case MDBX_opt_sync_period: if (unlikely(!env->me_autosync_period)) return MDBX_EPERM; *value = mdbx_osal_monotime_to_16dot16( atomic_load64(env->me_autosync_period, mo_Relaxed)); break; case MDBX_opt_max_db: *value = env->me_maxdbs - CORE_DBS; break; case MDBX_opt_max_readers: *value = env->me_maxreaders; break; case MDBX_opt_dp_reserve_limit: *value = env->me_options.dp_reserve_limit; break; case MDBX_opt_rp_augment_limit: *value = env->me_options.rp_augment_limit; break; case MDBX_opt_txn_dp_limit: *value = env->me_options.dp_limit; break; case MDBX_opt_txn_dp_initial: *value = env->me_options.dp_initial; break; case MDBX_opt_spill_max_denominator: *value = env->me_options.spill_max_denominator; break; case MDBX_opt_spill_min_denominator: *value = env->me_options.spill_min_denominator; break; case MDBX_opt_spill_parent4child_denominator: *value = env->me_options.spill_parent4child_denominator; break; case MDBX_opt_loose_limit: *value = env->me_options.dp_loose_limit; break; default: return MDBX_EINVAL; } return MDBX_SUCCESS; } /*** Attribute support functions for Nexenta **********************************/ #ifdef MDBX_NEXENTA_ATTRS static __inline int mdbx_attr_peek(MDBX_val *data, mdbx_attr_t *attrptr) { if (unlikely(data->iov_len < sizeof(mdbx_attr_t))) return MDBX_INCOMPATIBLE; if (likely(attrptr != NULL)) *attrptr = *(mdbx_attr_t *)data->iov_base; data->iov_len -= sizeof(mdbx_attr_t); data->iov_base = likely(data->iov_len > 0) ? ((mdbx_attr_t *)data->iov_base) + 1 : NULL; return MDBX_SUCCESS; } static __inline int mdbx_attr_poke(MDBX_val *reserved, MDBX_val *data, mdbx_attr_t attr, MDBX_put_flags_t flags) { mdbx_attr_t *space = reserved->iov_base; if (flags & MDBX_RESERVE) { if (likely(data != NULL)) { data->iov_base = data->iov_len ? space + 1 : NULL; } } else { *space = attr; if (likely(data != NULL)) { memcpy(space + 1, data->iov_base, data->iov_len); } } return MDBX_SUCCESS; } int mdbx_cursor_get_attr(MDBX_cursor *mc, MDBX_val *key, MDBX_val *data, mdbx_attr_t *attrptr, MDBX_cursor_op op) { int rc = mdbx_cursor_get(mc, key, data, op); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_attr_peek(data, attrptr); } int mdbx_get_attr(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *key, MDBX_val *data, uint64_t *attrptr) { int rc = mdbx_get(txn, dbi, key, data); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_attr_peek(data, attrptr); } int mdbx_put_attr(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *key, MDBX_val *data, mdbx_attr_t attr, MDBX_put_flags_t flags) { MDBX_val reserve; reserve.iov_base = NULL; reserve.iov_len = (data ? data->iov_len : 0) + sizeof(mdbx_attr_t); int rc = mdbx_put(txn, dbi, key, &reserve, flags | MDBX_RESERVE); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_attr_poke(&reserve, data, attr, flags); } int mdbx_cursor_put_attr(MDBX_cursor *cursor, MDBX_val *key, MDBX_val *data, mdbx_attr_t attr, MDBX_put_flags_t flags) { MDBX_val reserve; reserve.iov_base = NULL; reserve.iov_len = (data ? data->iov_len : 0) + sizeof(mdbx_attr_t); int rc = mdbx_cursor_put(cursor, key, &reserve, flags | MDBX_RESERVE); if (unlikely(rc != MDBX_SUCCESS)) return rc; return mdbx_attr_poke(&reserve, data, attr, flags); } int mdbx_set_attr(MDBX_txn *txn, MDBX_dbi dbi, MDBX_val *key, MDBX_val *data, mdbx_attr_t attr) { if (unlikely(!key || !txn)) return MDBX_EINVAL; if (unlikely(txn->mt_signature != MDBX_MT_SIGNATURE)) return MDBX_EBADSIGN; if (unlikely(!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))) return MDBX_BAD_DBI; if (unlikely(txn->mt_flags & (MDBX_TXN_RDONLY | MDBX_TXN_BLOCKED))) return (txn->mt_flags & MDBX_TXN_RDONLY) ? MDBX_EACCESS : MDBX_BAD_TXN; MDBX_cursor_couple cx; MDBX_val old_data; int rc = mdbx_cursor_init(&cx.outer, txn, dbi); if (unlikely(rc != MDBX_SUCCESS)) return rc; rc = mdbx_cursor_set(&cx.outer, key, &old_data, MDBX_SET, NULL); if (unlikely(rc != MDBX_SUCCESS)) { if (rc == MDBX_NOTFOUND && data) { cx.outer.mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = &cx.outer; rc = mdbx_cursor_put_attr(&cx.outer, key, data, attr, 0); txn->tw.cursors[dbi] = cx.outer.mc_next; } return rc; } mdbx_attr_t old_attr = 0; rc = mdbx_attr_peek(&old_data, &old_attr); if (unlikely(rc != MDBX_SUCCESS)) return rc; if (old_attr == attr && (!data || (data->iov_len == old_data.iov_len && memcmp(data->iov_base, old_data.iov_base, old_data.iov_len) == 0))) return MDBX_SUCCESS; cx.outer.mc_next = txn->tw.cursors[dbi]; txn->tw.cursors[dbi] = &cx.outer; rc = mdbx_cursor_put_attr(&cx.outer, key, data ? data : &old_data, attr, MDBX_CURRENT); txn->tw.cursors[dbi] = cx.outer.mc_next; return rc; } #endif /* MDBX_NEXENTA_ATTRS */ /******************************************************************************/ /* *INDENT-OFF* */ /* clang-format off */ __dll_export #ifdef __attribute_used__ __attribute_used__ #elif defined(__GNUC__) || __has_attribute(__used__) __attribute__((__used__)) #endif #ifdef __attribute_externally_visible__ __attribute_externally_visible__ #elif (defined(__GNUC__) && !defined(__clang__)) || \ __has_attribute(__externally_visible__) __attribute__((__externally_visible__)) #endif const struct MDBX_build_info mdbx_build = { #ifdef MDBX_BUILD_TIMESTAMP MDBX_BUILD_TIMESTAMP #else __DATE__ " " __TIME__ #endif /* MDBX_BUILD_TIMESTAMP */ , #ifdef MDBX_BUILD_TARGET MDBX_BUILD_TARGET #else #if defined(__ANDROID_API__) "Android" STRINGIFY(__ANDROID_API__) #elif defined(__linux__) || defined(__gnu_linux__) "Linux" #elif defined(EMSCRIPTEN) || defined(__EMSCRIPTEN__) "webassembly" #elif defined(__CYGWIN__) "CYGWIN" #elif defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) \ || defined(__WINDOWS__) "Windows" #elif defined(__APPLE__) #if (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE) \ || (defined(TARGET_IPHONE_SIMULATOR) && TARGET_IPHONE_SIMULATOR) "iOS" #else "MacOS" #endif #elif defined(__FreeBSD__) "FreeBSD" #elif defined(__DragonFly__) "DragonFlyBSD" #elif defined(__NetBSD__) "NetBSD" #elif defined(__OpenBSD__) "OpenBSD" #elif defined(__bsdi__) "UnixBSDI" #elif defined(__MACH__) "MACH" #elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) "HPUX" #elif defined(_AIX) "AIX" #elif defined(__sun) && defined(__SVR4) "Solaris" #elif defined(__BSD__) || defined(BSD) "UnixBSD" #elif defined(__unix__) || defined(UNIX) || defined(__unix) \ || defined(__UNIX) || defined(__UNIX__) "UNIX" #elif defined(_POSIX_VERSION) "POSIX" STRINGIFY(_POSIX_VERSION) #else "UnknownOS" #endif /* Target OS */ "-" #if defined(__amd64__) "AMD64" #elif defined(__ia32__) "IA32" #elif defined(__e2k__) || defined(__elbrus__) "Elbrus" #elif defined(__alpha__) || defined(__alpha) || defined(_M_ALPHA) "Alpha" #elif defined(__aarch64__) || defined(_M_ARM64) "ARM64" #elif defined(__arm__) || defined(__thumb__) || defined(__TARGET_ARCH_ARM) \ || defined(__TARGET_ARCH_THUMB) || defined(_ARM) || defined(_M_ARM) \ || defined(_M_ARMT) || defined(__arm) "ARM" #elif defined(__mips64) || defined(__mips64__) || (defined(__mips) && (__mips >= 64)) "MIPS64" #elif if defined(__mips__) || defined(__mips) || defined(_R4000) || defined(__MIPS__) "MIPS" #elif defined(__hppa64__) || defined(__HPPA64__) || defined(__hppa64) "PARISC64" #elif defined(__hppa__) || defined(__HPPA__) || defined(__hppa) "PARISC" #elif defined(__ia64__) || defined(__ia64) || defined(_IA64) \ || defined(__IA64__) || defined(_M_IA64) || defined(__itanium__) "Itanium" #elif defined(__powerpc64__) || defined(__ppc64__) || defined(__ppc64) \ || defined(__powerpc64) || defined(_ARCH_PPC64) "PowerPC64" #elif defined(__powerpc__) || defined(__ppc__) || defined(__powerpc) \ || defined(__ppc) || defined(_ARCH_PPC) || defined(__PPC__) || defined(__POWERPC__) "PowerPC" #elif defined(__sparc64__) || defined(__sparc64) "SPARC64" #elif defined(__sparc__) || defined(__sparc) "SPARC" #elif defined(__s390__) || defined(__s390) || defined(__zarch__) || defined(__zarch) "S390" #else "UnknownARCH" #endif #endif /* MDBX_BUILD_TARGET */ #ifdef MDBX_BUILD_TYPE # if defined(_MSC_VER) # pragma message("Configuration-depended MDBX_BUILD_TYPE: " MDBX_BUILD_TYPE) # endif "-" MDBX_BUILD_TYPE #endif /* MDBX_BUILD_TYPE */ , "MDBX_DEBUG=" STRINGIFY(MDBX_DEBUG) " MDBX_WORDBITS=" STRINGIFY(MDBX_WORDBITS) " BYTE_ORDER=" #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ "LITTLE_ENDIAN" #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ "BIG_ENDIAN" #else #error "FIXME: Unsupported byte order" #endif /* __BYTE_ORDER__ */ " MDBX_ENV_CHECKPID=" MDBX_ENV_CHECKPID_CONFIG " MDBX_TXN_CHECKOWNER=" MDBX_TXN_CHECKOWNER_CONFIG " MDBX_64BIT_ATOMIC=" MDBX_64BIT_ATOMIC_CONFIG " MDBX_64BIT_CAS=" MDBX_64BIT_CAS_CONFIG " MDBX_TRUST_RTC=" MDBX_TRUST_RTC_CONFIG " MDBX_ENABLE_REFUND=" STRINGIFY(MDBX_ENABLE_REFUND) #ifdef __SANITIZE_ADDRESS__ " SANITIZE_ADDRESS=YES" #endif /* __SANITIZE_ADDRESS__ */ #ifdef MDBX_USE_VALGRIND " MDBX_USE_VALGRIND=YES" #endif /* MDBX_USE_VALGRIND */ #ifdef MDBX_FORCE_ASSERTIONS " MDBX_FORCE_ASSERTIONS=YES" #endif /* MDBX_FORCE_ASSERTIONS */ #ifdef _GNU_SOURCE " _GNU_SOURCE=YES" #else " _GNU_SOURCE=NO" #endif /* _GNU_SOURCE */ #ifdef __APPLE__ " MDBX_OSX_SPEED_INSTEADOF_DURABILITY=" STRINGIFY(MDBX_OSX_SPEED_INSTEADOF_DURABILITY) #endif /* MacOS */ #if defined(_WIN32) || defined(_WIN64) " MDBX_AVOID_CRT=" STRINGIFY(MDBX_AVOID_CRT) " MDBX_CONFIG_MANUAL_TLS_CALLBACK=" STRINGIFY(MDBX_CONFIG_MANUAL_TLS_CALLBACK) " MDBX_BUILD_SHARED_LIBRARY=" STRINGIFY(MDBX_BUILD_SHARED_LIBRARY) " WINVER=" STRINGIFY(WINVER) #else /* Windows */ " MDBX_LOCKING=" MDBX_LOCKING_CONFIG " MDBX_USE_OFDLOCKS=" MDBX_USE_OFDLOCKS_CONFIG #endif /* !Windows */ " MDBX_CACHELINE_SIZE=" STRINGIFY(MDBX_CACHELINE_SIZE) " MDBX_CPU_WRITEBACK_INCOHERENT=" STRINGIFY(MDBX_CPU_WRITEBACK_INCOHERENT) " MDBX_MMAP_INCOHERENT_CPU_CACHE=" STRINGIFY(MDBX_MMAP_INCOHERENT_CPU_CACHE) " MDBX_MMAP_INCOHERENT_FILE_WRITE=" STRINGIFY(MDBX_MMAP_INCOHERENT_FILE_WRITE) " MDBX_UNALIGNED_OK=" STRINGIFY(MDBX_UNALIGNED_OK) " MDBX_PNL_ASCENDING=" STRINGIFY(MDBX_PNL_ASCENDING) , #ifdef MDBX_BUILD_COMPILER MDBX_BUILD_COMPILER #else #ifdef __INTEL_COMPILER "Intel C/C++ " STRINGIFY(__INTEL_COMPILER) #elsif defined(__apple_build_version__) "Apple clang " STRINGIFY(__apple_build_version__) #elif defined(__ibmxl__) "IBM clang C " STRINGIFY(__ibmxl_version__) "." STRINGIFY(__ibmxl_release__) "." STRINGIFY(__ibmxl_modification__) "." STRINGIFY(__ibmxl_ptf_fix_level__) #elif defined(__clang__) "clang " STRINGIFY(__clang_version__) #elif defined(__MINGW64__) "MINGW-64 " STRINGIFY(__MINGW64_MAJOR_VERSION) "." STRINGIFY(__MINGW64_MINOR_VERSION) #elif defined(__MINGW32__) "MINGW-32 " STRINGIFY(__MINGW32_MAJOR_VERSION) "." STRINGIFY(__MINGW32_MINOR_VERSION) #elif defined(__IBMC__) "IBM C " STRINGIFY(__IBMC__) #elif defined(__GNUC__) "GNU C/C++ " #ifdef __VERSION__ __VERSION__ #else STRINGIFY(__GNUC__) "." STRINGIFY(__GNUC_MINOR__) "." STRINGIFY(__GNUC_PATCHLEVEL__) #endif #elif defined(_MSC_VER) "MSVC " STRINGIFY(_MSC_FULL_VER) "-" STRINGIFY(_MSC_BUILD) #else "Unknown compiler" #endif #endif /* MDBX_BUILD_COMPILER */ , #ifdef MDBX_BUILD_FLAGS_CONFIG MDBX_BUILD_FLAGS_CONFIG #endif /* MDBX_BUILD_FLAGS_CONFIG */ #ifdef MDBX_BUILD_FLAGS MDBX_BUILD_FLAGS #endif /* MDBX_BUILD_FLAGS */ #if !(defined(MDBX_BUILD_FLAGS_CONFIG) || defined(MDBX_BUILD_FLAGS)) "undefined (please use correct build script)" #ifdef _MSC_VER #pragma message("warning: Build flags undefined. Please use correct build script") #else #warning "Build flags undefined. Please use correct build script" #endif // _MSC_VER #endif }; #ifdef __SANITIZE_ADDRESS__ LIBMDBX_API __attribute__((__weak__)) const char *__asan_default_options() { return "symbolize=1:allow_addr2line=1:" #ifdef _DEBUG "debug=1:" #endif /* _DEBUG */ "report_globals=1:" "replace_str=1:replace_intrin=1:" "malloc_context_size=9:" "detect_leaks=1:" "check_printf=1:" "detect_deadlocks=1:" #ifndef LTO_ENABLED "check_initialization_order=1:" #endif "detect_stack_use_after_return=1:" "intercept_tls_get_addr=1:" "decorate_proc_maps=1:" "abort_on_error=1"; } #endif /* __SANITIZE_ADDRESS__ */ /* *INDENT-ON* */ /* clang-format on */ /* https://en.wikipedia.org/wiki/Operating_system_abstraction_layer */ /* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ #if defined(_WIN32) || defined(_WIN64) #include static int waitstatus2errcode(DWORD result) { switch (result) { case WAIT_OBJECT_0: return MDBX_SUCCESS; case WAIT_FAILED: return GetLastError(); case WAIT_ABANDONED: return ERROR_ABANDONED_WAIT_0; case WAIT_IO_COMPLETION: return ERROR_USER_APC; case WAIT_TIMEOUT: return ERROR_TIMEOUT; default: return ERROR_UNHANDLED_ERROR; } } /* Map a result from an NTAPI call to WIN32 error code. */ static int ntstatus2errcode(NTSTATUS status) { DWORD dummy; OVERLAPPED ov; memset(&ov, 0, sizeof(ov)); ov.Internal = status; return GetOverlappedResult(NULL, &ov, &dummy, FALSE) ? MDBX_SUCCESS : GetLastError(); } /* We use native NT APIs to setup the memory map, so that we can * let the DB file grow incrementally instead of always preallocating * the full size. These APIs are defined in and * but those headers are meant for driver-level development and * conflict with the regular user-level headers, so we explicitly * declare them here. Using these APIs also means we must link to * ntdll.dll, which is not linked by default in user code. */ extern NTSTATUS NTAPI NtCreateSection( OUT PHANDLE SectionHandle, IN ACCESS_MASK DesiredAccess, IN OPTIONAL POBJECT_ATTRIBUTES ObjectAttributes, IN OPTIONAL PLARGE_INTEGER MaximumSize, IN ULONG SectionPageProtection, IN ULONG AllocationAttributes, IN OPTIONAL HANDLE FileHandle); typedef struct _SECTION_BASIC_INFORMATION { ULONG Unknown; ULONG SectionAttributes; LARGE_INTEGER SectionSize; } SECTION_BASIC_INFORMATION, *PSECTION_BASIC_INFORMATION; extern NTSTATUS NTAPI NtMapViewOfSection( IN HANDLE SectionHandle, IN HANDLE ProcessHandle, IN OUT PVOID *BaseAddress, IN ULONG_PTR ZeroBits, IN SIZE_T CommitSize, IN OUT OPTIONAL PLARGE_INTEGER SectionOffset, IN OUT PSIZE_T ViewSize, IN SECTION_INHERIT InheritDisposition, IN ULONG AllocationType, IN ULONG Win32Protect); extern NTSTATUS NTAPI NtUnmapViewOfSection(IN HANDLE ProcessHandle, IN OPTIONAL PVOID BaseAddress); extern NTSTATUS NTAPI NtClose(HANDLE Handle); extern NTSTATUS NTAPI NtAllocateVirtualMemory( IN HANDLE ProcessHandle, IN OUT PVOID *BaseAddress, IN ULONG_PTR ZeroBits, IN OUT PSIZE_T RegionSize, IN ULONG AllocationType, IN ULONG Protect); extern NTSTATUS NTAPI NtFreeVirtualMemory(IN HANDLE ProcessHandle, IN PVOID *BaseAddress, IN OUT PSIZE_T RegionSize, IN ULONG FreeType); #ifndef WOF_CURRENT_VERSION typedef struct _WOF_EXTERNAL_INFO { DWORD Version; DWORD Provider; } WOF_EXTERNAL_INFO, *PWOF_EXTERNAL_INFO; #endif /* WOF_CURRENT_VERSION */ #ifndef WIM_PROVIDER_CURRENT_VERSION #define WIM_PROVIDER_HASH_SIZE 20 typedef struct _WIM_PROVIDER_EXTERNAL_INFO { DWORD Version; DWORD Flags; LARGE_INTEGER DataSourceId; BYTE ResourceHash[WIM_PROVIDER_HASH_SIZE]; } WIM_PROVIDER_EXTERNAL_INFO, *PWIM_PROVIDER_EXTERNAL_INFO; #endif /* WIM_PROVIDER_CURRENT_VERSION */ #ifndef FILE_PROVIDER_CURRENT_VERSION typedef struct _FILE_PROVIDER_EXTERNAL_INFO_V1 { ULONG Version; ULONG Algorithm; ULONG Flags; } FILE_PROVIDER_EXTERNAL_INFO_V1, *PFILE_PROVIDER_EXTERNAL_INFO_V1; #endif /* FILE_PROVIDER_CURRENT_VERSION */ #ifndef STATUS_OBJECT_NOT_EXTERNALLY_BACKED #define STATUS_OBJECT_NOT_EXTERNALLY_BACKED ((NTSTATUS)0xC000046DL) #endif #ifndef STATUS_INVALID_DEVICE_REQUEST #define STATUS_INVALID_DEVICE_REQUEST ((NTSTATUS)0xC0000010L) #endif #ifndef STATUS_NOT_SUPPORTED #define STATUS_NOT_SUPPORTED ((NTSTATUS)0xC00000BBL) #endif #ifndef FILE_DEVICE_FILE_SYSTEM #define FILE_DEVICE_FILE_SYSTEM 0x00000009 #endif #ifndef FSCTL_GET_EXTERNAL_BACKING #define FSCTL_GET_EXTERNAL_BACKING \ CTL_CODE(FILE_DEVICE_FILE_SYSTEM, 196, METHOD_BUFFERED, FILE_ANY_ACCESS) #endif #endif /* _WIN32 || _WIN64 */ /*----------------------------------------------------------------------------*/ #if defined(__UCLIBC__) __extern_C void __assert(const char *, const char *, unsigned int, const char *) #ifdef __THROW __THROW #else __nothrow #endif /* __THROW */ MDBX_NORETURN; #define __assert_fail(assertion, file, line, function) \ __assert(assertion, file, line, function) #elif _POSIX_C_SOURCE > 200212 && \ /* workaround for avoid musl libc wrong prototype */ ( \ defined(__GLIBC__) || defined(__GNU_LIBRARY__)) /* Prototype should match libc runtime. ISO POSIX (2003) & LSB 1.x-3.x */ __extern_C void __assert_fail(const char *assertion, const char *file, unsigned line, const char *function) #ifdef __THROW __THROW #else __nothrow #endif /* __THROW */ MDBX_NORETURN; #elif defined(__APPLE__) || defined(__MACH__) __extern_C void __assert_rtn(const char *function, const char *file, int line, const char *assertion) /* __nothrow */ #ifdef __dead2 __dead2 #else MDBX_NORETURN #endif /* __dead2 */ #ifdef __disable_tail_calls __disable_tail_calls #endif /* __disable_tail_calls */ ; #define __assert_fail(assertion, file, line, function) \ __assert_rtn(function, file, line, assertion) #elif defined(__sun) || defined(__SVR4) || defined(__svr4__) __extern_C void __assert_c99(const char *assection, const char *file, int line, const char *function) MDBX_NORETURN; #define __assert_fail(assertion, file, line, function) \ __assert_c99(assertion, file, line, function) #elif defined(__OpenBSD__) __extern_C __dead void __assert2(const char *file, int line, const char *function, const char *assertion) /* __nothrow */; #define __assert_fail(assertion, file, line, function) \ __assert2(file, line, function, assertion) #elif defined(__NetBSD__) __extern_C __dead void __assert13(const char *file, int line, const char *function, const char *assertion) /* __nothrow */; #define __assert_fail(assertion, file, line, function) \ __assert13(file, line, function, assertion) #elif defined(__FreeBSD__) || defined(__BSD__) || defined(__bsdi__) || \ defined(__DragonFly__) __extern_C void __assert(const char *function, const char *file, int line, const char *assertion) /* __nothrow */ #ifdef __dead2 __dead2 #else MDBX_NORETURN #endif /* __dead2 */ #ifdef __disable_tail_calls __disable_tail_calls #endif /* __disable_tail_calls */ ; #define __assert_fail(assertion, file, line, function) \ __assert(function, file, line, assertion) #endif /* __assert_fail */ #if !defined(__ANDROID_API__) || MDBX_DEBUG void __cold mdbx_assert_fail(const MDBX_env *env, const char *msg, const char *func, int line) { #if MDBX_DEBUG if (env && env->me_assert_func) { env->me_assert_func(env, msg, func, line); return; } #else (void)env; #endif /* MDBX_DEBUG */ if (mdbx_debug_logger) mdbx_debug_log(MDBX_LOG_FATAL, func, line, "assert: %s\n", msg); else { #if defined(_WIN32) || defined(_WIN64) char *message = nullptr; const int num = mdbx_asprintf(&message, "\r\nMDBX-ASSERTION: %s, %s:%u", msg, func ? func : "unknown", line); if (num < 1 || !message) message = ""; OutputDebugStringA(message); if (IsDebuggerPresent()) DebugBreak(); #elif defined(__ANDROID_API__) __android_log_assert(msg, "mdbx", "%s:%u", func, line); #else __assert_fail(msg, "mdbx", line, func); #endif } #if defined(_WIN32) || defined(_WIN64) FatalExit(ERROR_UNHANDLED_ERROR); #else abort(); #endif } #endif /* __ANDROID_API__ || MDBX_DEBUG */ __cold void mdbx_panic(const char *fmt, ...) { va_list ap; va_start(ap, fmt); char *message = nullptr; const int num = mdbx_vasprintf(&message, fmt, ap); va_end(ap); const char *const const_message = (num < 1 || !message) ? "" : message; #if defined(_WIN32) || defined(_WIN64) OutputDebugStringA("\r\nMDBX-PANIC: "); OutputDebugStringA(const_message); if (IsDebuggerPresent()) DebugBreak(); FatalExit(ERROR_UNHANDLED_ERROR); #else #if defined(__ANDROID_API__) __android_log_assert("panic", "mdbx", "%s", const_message); #else __assert_fail(const_message, "mdbx", 0, "panic"); #endif /* __ANDROID_API__ */ abort(); #endif } /*----------------------------------------------------------------------------*/ #ifndef mdbx_vasprintf MDBX_INTERNAL_FUNC int mdbx_vasprintf(char **strp, const char *fmt, va_list ap) { va_list ones; va_copy(ones, ap); int needed = vsnprintf(nullptr, 0, fmt, ap); if (unlikely(needed < 0 || needed >= INT_MAX)) { *strp = nullptr; va_end(ones); return needed; } *strp = mdbx_malloc(needed + 1); if (unlikely(*strp == nullptr)) { va_end(ones); #if defined(_WIN32) || defined(_WIN64) SetLastError(MDBX_ENOMEM); #else errno = MDBX_ENOMEM; #endif return -1; } int actual = vsnprintf(*strp, needed + 1, fmt, ones); va_end(ones); assert(actual == needed); if (unlikely(actual < 0)) { mdbx_free(*strp); *strp = nullptr; } return actual; } #endif /* mdbx_vasprintf */ #ifndef mdbx_asprintf MDBX_INTERNAL_FUNC int mdbx_asprintf(char **strp, const char *fmt, ...) { va_list ap; va_start(ap, fmt); int rc = mdbx_vasprintf(strp, fmt, ap); va_end(ap); return rc; } #endif /* mdbx_asprintf */ #ifndef mdbx_memalign_alloc MDBX_INTERNAL_FUNC int mdbx_memalign_alloc(size_t alignment, size_t bytes, void **result) { assert(is_powerof2(alignment) && alignment >= sizeof(void *)); #if defined(_WIN32) || defined(_WIN64) (void)alignment; *result = VirtualAlloc(NULL, bytes, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE); return *result ? MDBX_SUCCESS : MDBX_ENOMEM /* ERROR_OUTOFMEMORY */; #elif defined(_ISOC11_SOURCE) *result = aligned_alloc(alignment, ceil_powerof2(bytes, alignment)); return *result ? MDBX_SUCCESS : errno; #elif _POSIX_VERSION >= 200112L && \ (!defined(__ANDROID_API__) || __ANDROID_API__ >= 17) *result = nullptr; return posix_memalign(result, alignment, bytes); #elif __GLIBC_PREREQ(2, 16) || __STDC_VERSION__ >= 201112L *result = memalign(alignment, bytes); return *result ? MDBX_SUCCESS : errno; #else #error FIXME #endif } #endif /* mdbx_memalign_alloc */ #ifndef mdbx_memalign_free MDBX_INTERNAL_FUNC void mdbx_memalign_free(void *ptr) { #if defined(_WIN32) || defined(_WIN64) VirtualFree(ptr, 0, MEM_RELEASE); #else mdbx_free(ptr); #endif } #endif /* mdbx_memalign_free */ #ifndef mdbx_strdup char *mdbx_strdup(const char *str) { if (!str) return NULL; size_t bytes = strlen(str) + 1; char *dup = mdbx_malloc(bytes); if (dup) memcpy(dup, str, bytes); return dup; } #endif /* mdbx_strdup */ /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int mdbx_condpair_init(mdbx_condpair_t *condpair) { int rc; memset(condpair, 0, sizeof(mdbx_condpair_t)); #if defined(_WIN32) || defined(_WIN64) if ((condpair->mutex = CreateMutexW(NULL, FALSE, NULL)) == NULL) { rc = GetLastError(); goto bailout_mutex; } if ((condpair->event[0] = CreateEventW(NULL, FALSE, FALSE, NULL)) == NULL) { rc = GetLastError(); goto bailout_event; } if ((condpair->event[1] = CreateEventW(NULL, FALSE, FALSE, NULL)) != NULL) return MDBX_SUCCESS; rc = GetLastError(); (void)CloseHandle(condpair->event[0]); bailout_event: (void)CloseHandle(condpair->mutex); #else rc = pthread_mutex_init(&condpair->mutex, NULL); if (unlikely(rc != 0)) goto bailout_mutex; rc = pthread_cond_init(&condpair->cond[0], NULL); if (unlikely(rc != 0)) goto bailout_cond; rc = pthread_cond_init(&condpair->cond[1], NULL); if (likely(rc == 0)) return MDBX_SUCCESS; (void)pthread_cond_destroy(&condpair->cond[0]); bailout_cond: (void)pthread_mutex_destroy(&condpair->mutex); #endif bailout_mutex: memset(condpair, 0, sizeof(mdbx_condpair_t)); return rc; } MDBX_INTERNAL_FUNC int mdbx_condpair_destroy(mdbx_condpair_t *condpair) { #if defined(_WIN32) || defined(_WIN64) int rc = CloseHandle(condpair->mutex) ? MDBX_SUCCESS : GetLastError(); rc = CloseHandle(condpair->event[0]) ? rc : GetLastError(); rc = CloseHandle(condpair->event[1]) ? rc : GetLastError(); #else int err, rc = pthread_mutex_destroy(&condpair->mutex); rc = (err = pthread_cond_destroy(&condpair->cond[0])) ? err : rc; rc = (err = pthread_cond_destroy(&condpair->cond[1])) ? err : rc; #endif memset(condpair, 0, sizeof(mdbx_condpair_t)); return rc; } MDBX_INTERNAL_FUNC int mdbx_condpair_lock(mdbx_condpair_t *condpair) { #if defined(_WIN32) || defined(_WIN64) DWORD code = WaitForSingleObject(condpair->mutex, INFINITE); return waitstatus2errcode(code); #else return pthread_mutex_lock(&condpair->mutex); #endif } MDBX_INTERNAL_FUNC int mdbx_condpair_unlock(mdbx_condpair_t *condpair) { #if defined(_WIN32) || defined(_WIN64) return ReleaseMutex(condpair->mutex) ? MDBX_SUCCESS : GetLastError(); #else return pthread_mutex_unlock(&condpair->mutex); #endif } MDBX_INTERNAL_FUNC int mdbx_condpair_signal(mdbx_condpair_t *condpair, bool part) { #if defined(_WIN32) || defined(_WIN64) return SetEvent(condpair->event[part]) ? MDBX_SUCCESS : GetLastError(); #else return pthread_cond_signal(&condpair->cond[part]); #endif } MDBX_INTERNAL_FUNC int mdbx_condpair_wait(mdbx_condpair_t *condpair, bool part) { #if defined(_WIN32) || defined(_WIN64) DWORD code = SignalObjectAndWait(condpair->mutex, condpair->event[part], INFINITE, FALSE); if (code == WAIT_OBJECT_0) { code = WaitForSingleObject(condpair->mutex, INFINITE); if (code == WAIT_OBJECT_0) return MDBX_SUCCESS; } return waitstatus2errcode(code); #else return pthread_cond_wait(&condpair->cond[part], &condpair->mutex); #endif } /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int mdbx_fastmutex_init(mdbx_fastmutex_t *fastmutex) { #if defined(_WIN32) || defined(_WIN64) InitializeCriticalSection(fastmutex); return MDBX_SUCCESS; #else return pthread_mutex_init(fastmutex, NULL); #endif } MDBX_INTERNAL_FUNC int mdbx_fastmutex_destroy(mdbx_fastmutex_t *fastmutex) { #if defined(_WIN32) || defined(_WIN64) DeleteCriticalSection(fastmutex); return MDBX_SUCCESS; #else return pthread_mutex_destroy(fastmutex); #endif } MDBX_INTERNAL_FUNC int mdbx_fastmutex_acquire(mdbx_fastmutex_t *fastmutex) { #if defined(_WIN32) || defined(_WIN64) __try { EnterCriticalSection(fastmutex); } __except ( (GetExceptionCode() == 0xC0000194 /* STATUS_POSSIBLE_DEADLOCK / EXCEPTION_POSSIBLE_DEADLOCK */) ? EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH) { return ERROR_POSSIBLE_DEADLOCK; } return MDBX_SUCCESS; #else return pthread_mutex_lock(fastmutex); #endif } MDBX_INTERNAL_FUNC int mdbx_fastmutex_release(mdbx_fastmutex_t *fastmutex) { #if defined(_WIN32) || defined(_WIN64) LeaveCriticalSection(fastmutex); return MDBX_SUCCESS; #else return pthread_mutex_unlock(fastmutex); #endif } /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int mdbx_removefile(const char *pathname) { #if defined(_WIN32) || defined(_WIN64) const size_t wlen = mbstowcs(nullptr, pathname, INT_MAX); if (wlen < 1 || wlen > /* MAX_PATH */ INT16_MAX) return ERROR_INVALID_NAME; wchar_t *const pathnameW = _alloca((wlen + 1) * sizeof(wchar_t)); if (wlen != mbstowcs(pathnameW, pathname, wlen + 1)) return ERROR_INVALID_NAME; return DeleteFileW(pathnameW) ? MDBX_SUCCESS : GetLastError(); #else return unlink(pathname) ? errno : MDBX_SUCCESS; #endif } #if !(defined(_WIN32) || defined(_WIN64)) static bool is_valid_fd(int fd) { return !(isatty(fd) < 0 && errno == EBADF); } #endif /*! Windows */ MDBX_INTERNAL_FUNC int mdbx_removedirectory(const char *pathname) { #if defined(_WIN32) || defined(_WIN64) const size_t wlen = mbstowcs(nullptr, pathname, INT_MAX); if (wlen < 1 || wlen > /* MAX_PATH */ INT16_MAX) return ERROR_INVALID_NAME; wchar_t *const pathnameW = _alloca((wlen + 1) * sizeof(wchar_t)); if (wlen != mbstowcs(pathnameW, pathname, wlen + 1)) return ERROR_INVALID_NAME; return RemoveDirectoryW(pathnameW) ? MDBX_SUCCESS : GetLastError(); #else return rmdir(pathname) ? errno : MDBX_SUCCESS; #endif } MDBX_INTERNAL_FUNC int mdbx_openfile(const enum mdbx_openfile_purpose purpose, const MDBX_env *env, const char *pathname, mdbx_filehandle_t *fd, mdbx_mode_t unix_mode_bits) { *fd = INVALID_HANDLE_VALUE; #if defined(_WIN32) || defined(_WIN64) const size_t wlen = mbstowcs(nullptr, pathname, INT_MAX); if (wlen < 1 || wlen > /* MAX_PATH */ INT16_MAX) return ERROR_INVALID_NAME; wchar_t *const pathnameW = _alloca((wlen + 1) * sizeof(wchar_t)); if (wlen != mbstowcs(pathnameW, pathname, wlen + 1)) return ERROR_INVALID_NAME; DWORD CreationDisposition = unix_mode_bits ? OPEN_ALWAYS : OPEN_EXISTING; DWORD FlagsAndAttributes = FILE_FLAG_POSIX_SEMANTICS | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED; DWORD DesiredAccess = FILE_READ_ATTRIBUTES; DWORD ShareMode = (env->me_flags & MDBX_EXCLUSIVE) ? 0 : (FILE_SHARE_READ | FILE_SHARE_WRITE); switch (purpose) { default: return ERROR_INVALID_PARAMETER; case MDBX_OPEN_LCK: CreationDisposition = OPEN_ALWAYS; DesiredAccess |= GENERIC_READ | GENERIC_WRITE; FlagsAndAttributes |= FILE_ATTRIBUTE_HIDDEN | FILE_ATTRIBUTE_TEMPORARY; break; case MDBX_OPEN_DXB_READ: CreationDisposition = OPEN_EXISTING; DesiredAccess |= GENERIC_READ; ShareMode |= FILE_SHARE_READ; break; case MDBX_OPEN_DXB_LAZY: DesiredAccess |= GENERIC_READ | GENERIC_WRITE; break; case MDBX_OPEN_DXB_DSYNC: CreationDisposition = OPEN_EXISTING; DesiredAccess |= GENERIC_WRITE; FlagsAndAttributes |= FILE_FLAG_WRITE_THROUGH; break; case MDBX_OPEN_COPY: CreationDisposition = CREATE_NEW; ShareMode = 0; DesiredAccess |= GENERIC_WRITE; FlagsAndAttributes |= (env->me_psize < env->me_os_psize) ? 0 : FILE_FLAG_NO_BUFFERING; break; case MDBX_OPEN_DELETE: CreationDisposition = OPEN_EXISTING; ShareMode |= FILE_SHARE_DELETE; DesiredAccess = FILE_READ_ATTRIBUTES | FILE_WRITE_ATTRIBUTES | DELETE | SYNCHRONIZE; break; } *fd = CreateFileW(pathnameW, DesiredAccess, ShareMode, NULL, CreationDisposition, FlagsAndAttributes, NULL); if (*fd == INVALID_HANDLE_VALUE) return GetLastError(); BY_HANDLE_FILE_INFORMATION info; if (!GetFileInformationByHandle(*fd, &info)) { int err = GetLastError(); CloseHandle(*fd); *fd = INVALID_HANDLE_VALUE; return err; } const DWORD AttributesDiff = (info.dwFileAttributes ^ FlagsAndAttributes) & (FILE_ATTRIBUTE_HIDDEN | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED | FILE_ATTRIBUTE_TEMPORARY | FILE_ATTRIBUTE_COMPRESSED); if (AttributesDiff) (void)SetFileAttributesW(pathnameW, info.dwFileAttributes ^ AttributesDiff); #else int flags = unix_mode_bits ? O_CREAT : 0; switch (purpose) { default: return EINVAL; case MDBX_OPEN_LCK: flags |= O_RDWR; break; case MDBX_OPEN_DXB_READ: flags = O_RDONLY; break; case MDBX_OPEN_DXB_LAZY: flags |= O_RDWR; break; case MDBX_OPEN_COPY: flags = O_CREAT | O_WRONLY | O_EXCL; break; case MDBX_OPEN_DXB_DSYNC: flags |= O_WRONLY; #if defined(O_DSYNC) flags |= O_DSYNC; #elif defined(O_SYNC) flags |= O_SYNC; #elif defined(O_FSYNC) flags |= O_FSYNC; #endif break; case MDBX_OPEN_DELETE: flags = O_RDWR; break; } const bool direct_nocache_for_copy = env->me_psize >= env->me_os_psize && purpose == MDBX_OPEN_COPY; if (direct_nocache_for_copy) { #if defined(O_DIRECT) flags |= O_DIRECT; #endif /* O_DIRECT */ #if defined(O_NOCACHE) flags |= O_NOCACHE; #endif /* O_NOCACHE */ } #ifdef O_CLOEXEC flags |= O_CLOEXEC; #endif /* O_CLOEXEC */ /* Safeguard for https://github.com/erthink/libmdbx/issues/144 */ #if STDIN_FILENO == 0 && STDOUT_FILENO == 1 && STDERR_FILENO == 2 int stub_fd0 = -1, stub_fd1 = -1, stub_fd2 = -1; static const char dev_null[] = "/dev/null"; if (!is_valid_fd(STDIN_FILENO)) { mdbx_warning("STD%s_FILENO/%d is invalid, open %s for temporary stub", "IN", STDIN_FILENO, dev_null); stub_fd0 = open(dev_null, O_RDONLY | O_NOCTTY); } if (!is_valid_fd(STDOUT_FILENO)) { mdbx_warning("STD%s_FILENO/%d is invalid, open %s for temporary stub", "OUT", STDOUT_FILENO, dev_null); stub_fd1 = open(dev_null, O_WRONLY | O_NOCTTY); } if (!is_valid_fd(STDERR_FILENO)) { mdbx_warning("STD%s_FILENO/%d is invalid, open %s for temporary stub", "ERR", STDERR_FILENO, dev_null); stub_fd2 = open(dev_null, O_WRONLY | O_NOCTTY); } #else #error "Unexpected or unsupported UNIX or POSIX system" #endif /* STDIN_FILENO == 0 && STDERR_FILENO == 2 */ *fd = open(pathname, flags, unix_mode_bits); #if defined(O_DIRECT) if (*fd < 0 && (flags & O_DIRECT) && (errno == EINVAL || errno == EAFNOSUPPORT)) { flags &= ~(O_DIRECT | O_EXCL); *fd = open(pathname, flags, unix_mode_bits); } #endif /* O_DIRECT */ /* Safeguard for https://github.com/erthink/libmdbx/issues/144 */ #if STDIN_FILENO == 0 && STDOUT_FILENO == 1 && STDERR_FILENO == 2 if (*fd == STDIN_FILENO) { mdbx_warning("Got STD%s_FILENO/%d, avoid using it by dup(fd)", "IN", STDIN_FILENO); assert(stub_fd0 == -1); *fd = dup(stub_fd0 = *fd); } if (*fd == STDOUT_FILENO) { mdbx_warning("Got STD%s_FILENO/%d, avoid using it by dup(fd)", "OUT", STDOUT_FILENO); assert(stub_fd1 == -1); *fd = dup(stub_fd1 = *fd); } if (*fd == STDERR_FILENO) { mdbx_warning("Got STD%s_FILENO/%d, avoid using it by dup(fd)", "ERR", STDERR_FILENO); assert(stub_fd2 == -1); *fd = dup(stub_fd2 = *fd); } if (stub_fd0 != -1) close(stub_fd0); if (stub_fd1 != -1) close(stub_fd1); if (stub_fd2 != -1) close(stub_fd2); if (*fd >= STDIN_FILENO && *fd <= STDERR_FILENO) { mdbx_error( "Rejecting the use of a FD in the range " "STDIN_FILENO/%d..STDERR_FILENO/%d to prevent database corruption", STDIN_FILENO, STDERR_FILENO); close(*fd); return EBADF; } #else #error "Unexpected or unsupported UNIX or POSIX system" #endif /* STDIN_FILENO == 0 && STDERR_FILENO == 2 */ if (*fd < 0) return errno; #if defined(FD_CLOEXEC) && !defined(O_CLOEXEC) const int fd_flags = fcntl(*fd, F_GETFD); if (fd_flags != -1) (void)fcntl(*fd, F_SETFD, fd_flags | FD_CLOEXEC); #endif /* FD_CLOEXEC && !O_CLOEXEC */ if (direct_nocache_for_copy) { #if defined(F_NOCACHE) && !defined(O_NOCACHE) (void)fcntl(*fd, F_NOCACHE, 1); #endif /* F_NOCACHE */ } #endif return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_closefile(mdbx_filehandle_t fd) { #if defined(_WIN32) || defined(_WIN64) return CloseHandle(fd) ? MDBX_SUCCESS : GetLastError(); #else assert(fd > STDERR_FILENO); return (close(fd) == 0) ? MDBX_SUCCESS : errno; #endif } MDBX_INTERNAL_FUNC int mdbx_pread(mdbx_filehandle_t fd, void *buf, size_t bytes, uint64_t offset) { if (bytes > MAX_WRITE) return MDBX_EINVAL; #if defined(_WIN32) || defined(_WIN64) OVERLAPPED ov; ov.hEvent = 0; ov.Offset = (DWORD)offset; ov.OffsetHigh = HIGH_DWORD(offset); DWORD read = 0; if (unlikely(!ReadFile(fd, buf, (DWORD)bytes, &read, &ov))) { int rc = GetLastError(); return (rc == MDBX_SUCCESS) ? /* paranoia */ ERROR_READ_FAULT : rc; } #else STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); intptr_t read = pread(fd, buf, bytes, offset); if (read < 0) { int rc = errno; return (rc == MDBX_SUCCESS) ? /* paranoia */ MDBX_EIO : rc; } #endif return (bytes == (size_t)read) ? MDBX_SUCCESS : MDBX_ENODATA; } MDBX_INTERNAL_FUNC int mdbx_pwrite(mdbx_filehandle_t fd, const void *buf, size_t bytes, uint64_t offset) { while (true) { #if defined(_WIN32) || defined(_WIN64) OVERLAPPED ov; ov.hEvent = 0; ov.Offset = (DWORD)offset; ov.OffsetHigh = HIGH_DWORD(offset); DWORD written; if (unlikely(!WriteFile( fd, buf, likely(bytes <= MAX_WRITE) ? (DWORD)bytes : MAX_WRITE, &written, &ov))) return GetLastError(); if (likely(bytes == written)) return MDBX_SUCCESS; #else STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); const intptr_t written = pwrite(fd, buf, likely(bytes <= MAX_WRITE) ? bytes : MAX_WRITE, offset); if (likely(bytes == (size_t)written)) return MDBX_SUCCESS; if (written < 0) { const int rc = errno; if (rc != EINTR) return rc; continue; } #endif bytes -= written; offset += written; buf = (char *)buf + written; } } MDBX_INTERNAL_FUNC int mdbx_write(mdbx_filehandle_t fd, const void *buf, size_t bytes) { while (true) { #if defined(_WIN32) || defined(_WIN64) DWORD written; if (unlikely(!WriteFile( fd, buf, likely(bytes <= MAX_WRITE) ? (DWORD)bytes : MAX_WRITE, &written, nullptr))) return GetLastError(); if (likely(bytes == written)) return MDBX_SUCCESS; #else STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); const intptr_t written = write(fd, buf, likely(bytes <= MAX_WRITE) ? bytes : MAX_WRITE); if (likely(bytes == (size_t)written)) return MDBX_SUCCESS; if (written < 0) { const int rc = errno; if (rc != EINTR) return rc; continue; } #endif bytes -= written; buf = (char *)buf + written; } } int mdbx_pwritev(mdbx_filehandle_t fd, struct iovec *iov, int iovcnt, uint64_t offset, size_t expected_written) { #if defined(_WIN32) || defined(_WIN64) || defined(__APPLE__) || \ (defined(__ANDROID_API__) && __ANDROID_API__ < 24) size_t written = 0; for (int i = 0; i < iovcnt; ++i) { int rc = mdbx_pwrite(fd, iov[i].iov_base, iov[i].iov_len, offset); if (unlikely(rc != MDBX_SUCCESS)) return rc; written += iov[i].iov_len; offset += iov[i].iov_len; } return (expected_written == written) ? MDBX_SUCCESS : MDBX_EIO /* ERROR_WRITE_FAULT */; #else int rc; intptr_t written; do { STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); written = pwritev(fd, iov, iovcnt, offset); if (likely(expected_written == (size_t)written)) return MDBX_SUCCESS; rc = errno; } while (rc == EINTR); return (written < 0) ? rc : MDBX_EIO /* Use which error code? */; #endif } MDBX_INTERNAL_FUNC int mdbx_fsync(mdbx_filehandle_t fd, enum mdbx_syncmode_bits mode_bits) { #if defined(_WIN32) || defined(_WIN64) if ((mode_bits & (MDBX_SYNC_DATA | MDBX_SYNC_IODQ)) && !FlushFileBuffers(fd)) return GetLastError(); return MDBX_SUCCESS; #else #if defined(__APPLE__) && \ MDBX_OSX_SPEED_INSTEADOF_DURABILITY == MDBX_OSX_WANNA_DURABILITY if (mode_bits & MDBX_SYNC_IODQ) return likely(fcntl(fd, F_FULLFSYNC) != -1) ? MDBX_SUCCESS : errno; #endif /* MacOS */ /* LY: This approach is always safe and without appreciable performance * degradation, even on a kernel with fdatasync's bug. * * For more info about of a corresponding fdatasync() bug * see http://www.spinics.net/lists/linux-ext4/msg33714.html */ while (1) { switch (mode_bits & (MDBX_SYNC_DATA | MDBX_SYNC_SIZE)) { case MDBX_SYNC_NONE: return MDBX_SUCCESS /* nothing to do */; #if defined(_POSIX_SYNCHRONIZED_IO) && _POSIX_SYNCHRONIZED_IO > 0 case MDBX_SYNC_DATA: if (fdatasync(fd) == 0) return MDBX_SUCCESS; break /* error */; #if defined(__linux__) || defined(__gnu_linux__) case MDBX_SYNC_SIZE: if (mdbx_linux_kernel_version >= 0x03060000) return MDBX_SUCCESS; __fallthrough /* fall through */; #endif /* Linux */ #endif /* _POSIX_SYNCHRONIZED_IO > 0 */ default: if (fsync(fd) == 0) return MDBX_SUCCESS; } int rc = errno; if (rc != EINTR) return rc; } #endif } int mdbx_filesize(mdbx_filehandle_t fd, uint64_t *length) { #if defined(_WIN32) || defined(_WIN64) BY_HANDLE_FILE_INFORMATION info; if (!GetFileInformationByHandle(fd, &info)) return GetLastError(); *length = info.nFileSizeLow | (uint64_t)info.nFileSizeHigh << 32; #else struct stat st; STATIC_ASSERT_MSG(sizeof(off_t) <= sizeof(uint64_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); if (fstat(fd, &st)) return errno; *length = st.st_size; #endif return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_is_pipe(mdbx_filehandle_t fd) { #if defined(_WIN32) || defined(_WIN64) switch (GetFileType(fd)) { case FILE_TYPE_DISK: return MDBX_RESULT_FALSE; case FILE_TYPE_CHAR: case FILE_TYPE_PIPE: return MDBX_RESULT_TRUE; default: return GetLastError(); } #else struct stat info; if (fstat(fd, &info)) return errno; switch (info.st_mode & S_IFMT) { case S_IFBLK: case S_IFREG: return MDBX_RESULT_FALSE; case S_IFCHR: case S_IFIFO: case S_IFSOCK: return MDBX_RESULT_TRUE; case S_IFDIR: case S_IFLNK: default: return MDBX_INCOMPATIBLE; } #endif } MDBX_INTERNAL_FUNC int mdbx_ftruncate(mdbx_filehandle_t fd, uint64_t length) { #if defined(_WIN32) || defined(_WIN64) if (mdbx_SetFileInformationByHandle) { FILE_END_OF_FILE_INFO EndOfFileInfo; EndOfFileInfo.EndOfFile.QuadPart = length; return mdbx_SetFileInformationByHandle(fd, FileEndOfFileInfo, &EndOfFileInfo, sizeof(FILE_END_OF_FILE_INFO)) ? MDBX_SUCCESS : GetLastError(); } else { LARGE_INTEGER li; li.QuadPart = length; return (SetFilePointerEx(fd, li, NULL, FILE_BEGIN) && SetEndOfFile(fd)) ? MDBX_SUCCESS : GetLastError(); } #else STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); return ftruncate(fd, length) == 0 ? MDBX_SUCCESS : errno; #endif } MDBX_INTERNAL_FUNC int mdbx_fseek(mdbx_filehandle_t fd, uint64_t pos) { #if defined(_WIN32) || defined(_WIN64) LARGE_INTEGER li; li.QuadPart = pos; return SetFilePointerEx(fd, li, NULL, FILE_BEGIN) ? MDBX_SUCCESS : GetLastError(); #else STATIC_ASSERT_MSG(sizeof(off_t) >= sizeof(size_t), "libmdbx requires 64-bit file I/O on 64-bit systems"); return (lseek(fd, pos, SEEK_SET) < 0) ? errno : MDBX_SUCCESS; #endif } /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int mdbx_thread_create(mdbx_thread_t *thread, THREAD_RESULT(THREAD_CALL *start_routine)(void *), void *arg) { #if defined(_WIN32) || defined(_WIN64) *thread = CreateThread(NULL, 0, start_routine, arg, 0, NULL); return *thread ? MDBX_SUCCESS : GetLastError(); #else return pthread_create(thread, NULL, start_routine, arg); #endif } MDBX_INTERNAL_FUNC int mdbx_thread_join(mdbx_thread_t thread) { #if defined(_WIN32) || defined(_WIN64) DWORD code = WaitForSingleObject(thread, INFINITE); return waitstatus2errcode(code); #else void *unused_retval = &unused_retval; return pthread_join(thread, &unused_retval); #endif } /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int mdbx_msync(mdbx_mmap_t *map, size_t offset, size_t length, enum mdbx_syncmode_bits mode_bits) { uint8_t *ptr = (uint8_t *)map->address + offset; #if defined(_WIN32) || defined(_WIN64) if (!FlushViewOfFile(ptr, length)) return GetLastError(); #else #if defined(__linux__) || defined(__gnu_linux__) if (mode_bits == MDBX_SYNC_NONE && mdbx_linux_kernel_version > 0x02061300) /* Since Linux 2.6.19, MS_ASYNC is in fact a no-op. The kernel properly * tracks dirty pages and flushes them to storage as necessary. */ return MDBX_SUCCESS; #endif /* Linux */ if (msync(ptr, length, (mode_bits & MDBX_SYNC_DATA) ? MS_SYNC : MS_ASYNC)) return errno; mode_bits &= ~MDBX_SYNC_DATA; #endif return mdbx_fsync(map->fd, mode_bits); } MDBX_INTERNAL_FUNC int mdbx_check_fs_rdonly(mdbx_filehandle_t handle, const char *pathname, int err) { #if defined(_WIN32) || defined(_WIN64) (void)pathname; (void)err; if (!mdbx_GetVolumeInformationByHandleW) return MDBX_ENOSYS; DWORD unused, flags; if (!mdbx_GetVolumeInformationByHandleW(handle, nullptr, 0, nullptr, &unused, &flags, nullptr, 0)) return GetLastError(); if ((flags & FILE_READ_ONLY_VOLUME) == 0) return MDBX_EACCESS; #else struct statvfs info; if (err != MDBX_ENOFILE) { if (statvfs(pathname, &info)) return errno; if ((info.f_flag & ST_RDONLY) == 0) return err; } if (fstatvfs(handle, &info)) return errno; if ((info.f_flag & ST_RDONLY) == 0) return (err == MDBX_ENOFILE) ? MDBX_EACCESS : err; #endif /* !Windows */ return MDBX_SUCCESS; } static int mdbx_check_fs_local(mdbx_filehandle_t handle, int flags) { #if defined(_WIN32) || defined(_WIN64) if (mdbx_RunningUnderWine() && !(flags & MDBX_EXCLUSIVE)) return ERROR_NOT_CAPABLE /* workaround for Wine */; if (GetFileType(handle) != FILE_TYPE_DISK) return ERROR_FILE_OFFLINE; if (mdbx_GetFileInformationByHandleEx) { FILE_REMOTE_PROTOCOL_INFO RemoteProtocolInfo; if (mdbx_GetFileInformationByHandleEx(handle, FileRemoteProtocolInfo, &RemoteProtocolInfo, sizeof(RemoteProtocolInfo))) { if ((RemoteProtocolInfo.Flags & REMOTE_PROTOCOL_INFO_FLAG_OFFLINE) && !(flags & MDBX_RDONLY)) return ERROR_FILE_OFFLINE; if (!(RemoteProtocolInfo.Flags & REMOTE_PROTOCOL_INFO_FLAG_LOOPBACK) && !(flags & MDBX_EXCLUSIVE)) return ERROR_REMOTE_STORAGE_MEDIA_ERROR; } } if (mdbx_NtFsControlFile) { NTSTATUS rc; struct { WOF_EXTERNAL_INFO wof_info; union { WIM_PROVIDER_EXTERNAL_INFO wim_info; FILE_PROVIDER_EXTERNAL_INFO_V1 file_info; }; size_t reserved_for_microsoft_madness[42]; } GetExternalBacking_OutputBuffer; IO_STATUS_BLOCK StatusBlock; rc = mdbx_NtFsControlFile(handle, NULL, NULL, NULL, &StatusBlock, FSCTL_GET_EXTERNAL_BACKING, NULL, 0, &GetExternalBacking_OutputBuffer, sizeof(GetExternalBacking_OutputBuffer)); if (NT_SUCCESS(rc)) { if (!(flags & MDBX_EXCLUSIVE)) return ERROR_REMOTE_STORAGE_MEDIA_ERROR; } else if (rc != STATUS_OBJECT_NOT_EXTERNALLY_BACKED && rc != STATUS_INVALID_DEVICE_REQUEST && rc != STATUS_NOT_SUPPORTED) return ntstatus2errcode(rc); } if (mdbx_GetVolumeInformationByHandleW && mdbx_GetFinalPathNameByHandleW) { WCHAR *PathBuffer = mdbx_malloc(sizeof(WCHAR) * INT16_MAX); if (!PathBuffer) return MDBX_ENOMEM; int rc = MDBX_SUCCESS; DWORD VolumeSerialNumber, FileSystemFlags; if (!mdbx_GetVolumeInformationByHandleW(handle, PathBuffer, INT16_MAX, &VolumeSerialNumber, NULL, &FileSystemFlags, NULL, 0)) { rc = GetLastError(); goto bailout; } if ((flags & MDBX_RDONLY) == 0) { if (FileSystemFlags & (FILE_SEQUENTIAL_WRITE_ONCE | FILE_READ_ONLY_VOLUME | FILE_VOLUME_IS_COMPRESSED)) { rc = ERROR_REMOTE_STORAGE_MEDIA_ERROR; goto bailout; } } if (!mdbx_GetFinalPathNameByHandleW(handle, PathBuffer, INT16_MAX, FILE_NAME_NORMALIZED | VOLUME_NAME_NT)) { rc = GetLastError(); goto bailout; } if (_wcsnicmp(PathBuffer, L"\\Device\\Mup\\", 12) == 0) { if (!(flags & MDBX_EXCLUSIVE)) { rc = ERROR_REMOTE_STORAGE_MEDIA_ERROR; goto bailout; } } else if (mdbx_GetFinalPathNameByHandleW(handle, PathBuffer, INT16_MAX, FILE_NAME_NORMALIZED | VOLUME_NAME_DOS)) { UINT DriveType = GetDriveTypeW(PathBuffer); if (DriveType == DRIVE_NO_ROOT_DIR && _wcsnicmp(PathBuffer, L"\\\\?\\", 4) == 0 && _wcsnicmp(PathBuffer + 5, L":\\", 2) == 0) { PathBuffer[7] = 0; DriveType = GetDriveTypeW(PathBuffer + 4); } switch (DriveType) { case DRIVE_CDROM: if (flags & MDBX_RDONLY) break; // fall through case DRIVE_UNKNOWN: case DRIVE_NO_ROOT_DIR: case DRIVE_REMOTE: default: if (!(flags & MDBX_EXCLUSIVE)) rc = ERROR_REMOTE_STORAGE_MEDIA_ERROR; // fall through case DRIVE_REMOVABLE: case DRIVE_FIXED: case DRIVE_RAMDISK: break; } } bailout: mdbx_free(PathBuffer); return rc; } #else struct statvfs statvfs_info; if (fstatvfs(handle, &statvfs_info)) return errno; #if defined(ST_LOCAL) || defined(ST_EXPORTED) const unsigned long st_flags = statvfs_info.f_flag; #endif /* ST_LOCAL || ST_EXPORTED */ #if defined(__NetBSD__) const unsigned type = 0; const char *const name = statvfs_info.f_fstypename; const size_t name_len = VFS_NAMELEN; #elif defined(_AIX) || defined(__OS400__) const char *const name = statvfs_info.f_basetype; const size_t name_len = sizeof(statvfs_info.f_basetype); struct stat st; if (fstat(handle, &st)) return errno; const unsigned type = st.st_vfstype; if ((st.st_flag & FS_REMOTE) != 0 && !(flags & MDBX_EXCLUSIVE)) return MDBX_EREMOTE; #elif defined(FSTYPSZ) || defined(_FSTYPSZ) const unsigned type = 0; const char *const name = statvfs_info.f_basetype; const size_t name_len = sizeof(statvfs_info.f_basetype); #elif defined(__sun) || defined(__SVR4) || defined(__svr4__) || \ defined(ST_FSTYPSZ) || defined(_ST_FSTYPSZ) const unsigned type = 0; struct stat st; if (fstat(handle, &st)) return errno; const char *const name = st.st_fstype; const size_t name_len = strlen(name); #else struct statfs statfs_info; if (fstatfs(handle, &statfs_info)) return errno; #if defined(__OpenBSD__) const unsigned type = 0; #else const unsigned type = statfs_info.f_type; #endif #if defined(MNT_LOCAL) || defined(MNT_EXPORTED) const unsigned long mnt_flags = statfs_info.f_flags; #endif /* MNT_LOCAL || MNT_EXPORTED */ #if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \ defined(__BSD__) || defined(__bsdi__) || defined(__DragonFly__) || \ defined(__APPLE__) || defined(__MACH__) || defined(MFSNAMELEN) || \ defined(MFSTYPENAMELEN) || defined(VFS_NAMELEN) const char *const name = statfs_info.f_fstypename; const size_t name_len = sizeof(statfs_info.f_fstypename); #elif defined(__ANDROID_API__) && __ANDROID_API__ < 21 const char *const name = ""; const unsigned name_len = 0; #else const char *name = ""; unsigned name_len = 0; struct stat st; if (fstat(handle, &st)) return errno; char pathbuf[PATH_MAX]; FILE *mounted = nullptr; #if defined(__linux__) || defined(__gnu_linux__) mounted = setmntent("/proc/mounts", "r"); #endif /* Linux */ if (!mounted) mounted = setmntent("/etc/mtab", "r"); if (mounted) { const struct mntent *ent; #if defined(_BSD_SOURCE) || defined(_SVID_SOURCE) || defined(__BIONIC__) || \ (defined(_DEFAULT_SOURCE) && __GLIBC_PREREQ(2, 19)) struct mntent entbuf; const bool should_copy = false; while (nullptr != (ent = getmntent_r(mounted, &entbuf, pathbuf, sizeof(pathbuf)))) #else const bool should_copy = true; while (nullptr != (ent = getmntent(mounted))) #endif { struct stat mnt; if (!stat(ent->mnt_dir, &mnt) && mnt.st_dev == st.st_dev) { if (should_copy) { name = strncpy(pathbuf, ent->mnt_fsname, name_len = sizeof(pathbuf) - 1); pathbuf[name_len] = 0; } else { name = ent->mnt_fsname; name_len = strlen(name); } break; } } endmntent(mounted); } #endif /* !xBSD && !Android/Bionic */ #endif if (name_len) { if (((name_len > 2 && strncasecmp("nfs", name, 3) == 0) || strncasecmp("cifs", name, name_len) == 0 || strncasecmp("ncpfs", name, name_len) == 0 || strncasecmp("smbfs", name, name_len) == 0 || ((name_len > 3 && strncasecmp("fuse", name, 4) == 0) && strncasecmp("fuseblk", name, name_len) != 0)) && !(flags & MDBX_EXCLUSIVE)) return MDBX_EREMOTE; if (strcasecmp("ftp", name) == 0 || strcasecmp("http", name) == 0 || strcasecmp("sshfs", name) == 0) return MDBX_EREMOTE; } #ifdef ST_LOCAL if ((st_flags & ST_LOCAL) == 0 && !(flags & MDBX_EXCLUSIVE)) return MDBX_EREMOTE; #elif defined(MNT_LOCAL) if ((mnt_flags & MNT_LOCAL) == 0 && !(flags & MDBX_EXCLUSIVE)) return MDBX_EREMOTE; #endif /* ST/MNT_LOCAL */ #ifdef ST_EXPORTED if ((st_flags & ST_EXPORTED) != 0 && !(flags & MDBX_RDONLY)) return MDBX_EREMOTE; #elif defined(MNT_EXPORTED) if ((mnt_flags & MNT_EXPORTED) != 0 && !(flags & MDBX_RDONLY)) return MDBX_EREMOTE; #endif /* ST/MNT_EXPORTED */ switch (type) { case 0xFF534D42 /* CIFS_MAGIC_NUMBER */: case 0x6969 /* NFS_SUPER_MAGIC */: case 0x564c /* NCP_SUPER_MAGIC */: case 0x517B /* SMB_SUPER_MAGIC */: #if defined(__digital__) || defined(__osf__) || defined(__osf) case 0x0E /* Tru64 NFS */: #endif #ifdef ST_FST_NFS case ST_FST_NFS: #endif if ((flags & MDBX_EXCLUSIVE) == 0) return MDBX_EREMOTE; case 0: default: break; } #endif /* Unix */ return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_mmap(const int flags, mdbx_mmap_t *map, const size_t size, const size_t limit, const unsigned options) { assert(size <= limit); map->limit = 0; map->current = 0; map->address = nullptr; #if defined(_WIN32) || defined(_WIN64) map->section = NULL; map->filesize = 0; #endif /* Windows */ int err = mdbx_check_fs_local(map->fd, flags); if (unlikely(err != MDBX_SUCCESS)) return err; if ((flags & MDBX_RDONLY) == 0 && (options & MMAP_OPTION_TRUNCATE) != 0) { err = mdbx_ftruncate(map->fd, size); if (err != MDBX_SUCCESS) return err; #if defined(_WIN32) || defined(_WIN64) map->filesize = size; #else map->current = size; #endif } else { uint64_t filesize = 0; err = mdbx_filesize(map->fd, &filesize); if (err != MDBX_SUCCESS) return err; #if defined(_WIN32) || defined(_WIN64) map->filesize = filesize; #else map->current = (filesize > limit) ? limit : (size_t)filesize; #endif } #if defined(_WIN32) || defined(_WIN64) LARGE_INTEGER SectionSize; SectionSize.QuadPart = size; err = NtCreateSection( &map->section, /* DesiredAccess */ (flags & MDBX_WRITEMAP) ? SECTION_QUERY | SECTION_MAP_READ | SECTION_EXTEND_SIZE | SECTION_MAP_WRITE : SECTION_QUERY | SECTION_MAP_READ | SECTION_EXTEND_SIZE, /* ObjectAttributes */ NULL, /* MaximumSize (InitialSize) */ &SectionSize, /* SectionPageProtection */ (flags & MDBX_RDONLY) ? PAGE_READONLY : PAGE_READWRITE, /* AllocationAttributes */ SEC_RESERVE, map->fd); if (!NT_SUCCESS(err)) return ntstatus2errcode(err); SIZE_T ViewSize = (flags & MDBX_RDONLY) ? 0 : mdbx_RunningUnderWine() ? size : limit; err = NtMapViewOfSection( map->section, GetCurrentProcess(), &map->address, /* ZeroBits */ 0, /* CommitSize */ 0, /* SectionOffset */ NULL, &ViewSize, /* InheritDisposition */ ViewUnmap, /* AllocationType */ (flags & MDBX_RDONLY) ? 0 : MEM_RESERVE, /* Win32Protect */ (flags & MDBX_WRITEMAP) ? PAGE_READWRITE : PAGE_READONLY); if (!NT_SUCCESS(err)) { NtClose(map->section); map->section = 0; map->address = nullptr; return ntstatus2errcode(err); } assert(map->address != MAP_FAILED); map->current = (size_t)SectionSize.QuadPart; map->limit = ViewSize; #else #ifndef MAP_TRYFIXED #define MAP_TRYFIXED 0 #endif #ifndef MAP_HASSEMAPHORE #define MAP_HASSEMAPHORE 0 #endif #ifndef MAP_CONCEAL #define MAP_CONCEAL 0 #endif #ifndef MAP_NOSYNC #define MAP_NOSYNC 0 #endif map->address = mmap( NULL, limit, (flags & MDBX_WRITEMAP) ? PROT_READ | PROT_WRITE : PROT_READ, MAP_SHARED | MAP_FILE | (F_ISSET(flags, MDBX_UTTERLY_NOSYNC) ? MAP_NOSYNC : 0) | ((options & MMAP_OPTION_SEMAPHORE) ? MAP_HASSEMAPHORE | MAP_NOSYNC : MAP_CONCEAL), map->fd, 0); if (unlikely(map->address == MAP_FAILED)) { map->limit = 0; map->current = 0; map->address = nullptr; return errno; } map->limit = limit; #ifdef MADV_DONTFORK if (unlikely(madvise(map->address, map->limit, MADV_DONTFORK) != 0)) return errno; #endif #ifdef MADV_NOHUGEPAGE (void)madvise(map->address, map->limit, MADV_NOHUGEPAGE); #endif #endif VALGRIND_MAKE_MEM_DEFINED(map->address, map->current); ASAN_UNPOISON_MEMORY_REGION(map->address, map->current); return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_munmap(mdbx_mmap_t *map) { VALGRIND_MAKE_MEM_NOACCESS(map->address, map->current); /* Unpoisoning is required for ASAN to avoid false-positive diagnostic * when this memory will re-used by malloc or another mmapping. * See https://github.com/erthink/libmdbx/pull/93#issuecomment-613687203 */ ASAN_UNPOISON_MEMORY_REGION(map->address, map->limit); #if defined(_WIN32) || defined(_WIN64) if (map->section) NtClose(map->section); NTSTATUS rc = NtUnmapViewOfSection(GetCurrentProcess(), map->address); if (!NT_SUCCESS(rc)) ntstatus2errcode(rc); #else if (unlikely(munmap(map->address, map->limit))) return errno; #endif map->limit = 0; map->current = 0; map->address = nullptr; return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_mresize(int flags, mdbx_mmap_t *map, size_t size, size_t limit, const bool may_move) { assert(size <= limit); #if defined(_WIN32) || defined(_WIN64) assert(size != map->current || limit != map->limit || size < map->filesize); NTSTATUS status; LARGE_INTEGER SectionSize; int err, rc = MDBX_SUCCESS; if (!(flags & MDBX_RDONLY) && limit == map->limit && size > map->current && /* workaround for Wine */ mdbx_NtExtendSection) { /* growth rw-section */ SectionSize.QuadPart = size; status = mdbx_NtExtendSection(map->section, &SectionSize); if (!NT_SUCCESS(status)) return ntstatus2errcode(status); map->current = size; if (map->filesize < size) map->filesize = size; return MDBX_SUCCESS; } if (limit > map->limit) { /* check ability of address space for growth before unmap */ PVOID BaseAddress = (PBYTE)map->address + map->limit; SIZE_T RegionSize = limit - map->limit; status = NtAllocateVirtualMemory(GetCurrentProcess(), &BaseAddress, 0, &RegionSize, MEM_RESERVE, PAGE_NOACCESS); if (status == /* STATUS_CONFLICTING_ADDRESSES */ 0xC0000018) return MDBX_UNABLE_EXTEND_MAPSIZE; if (!NT_SUCCESS(status)) return ntstatus2errcode(status); status = NtFreeVirtualMemory(GetCurrentProcess(), &BaseAddress, &RegionSize, MEM_RELEASE); if (!NT_SUCCESS(status)) return ntstatus2errcode(status); } /* Windows unable: * - shrink a mapped file; * - change size of mapped view; * - extend read-only mapping; * Therefore we should unmap/map entire section. */ status = NtUnmapViewOfSection(GetCurrentProcess(), map->address); if (!NT_SUCCESS(status)) return ntstatus2errcode(status); status = NtClose(map->section); map->section = NULL; PVOID ReservedAddress = NULL; SIZE_T ReservedSize = limit; if (!NT_SUCCESS(status)) { bailout_ntstatus: err = ntstatus2errcode(status); bailout: map->address = NULL; map->current = map->limit = 0; if (ReservedAddress) { ReservedSize = 0; status = NtFreeVirtualMemory(GetCurrentProcess(), &ReservedAddress, &ReservedSize, MEM_RELEASE); assert(NT_SUCCESS(status)); (void)status; } return err; } retry_file_and_section: /* resizing of the file may take a while, * therefore we reserve address space to avoid occupy it by other threads */ ReservedAddress = map->address; status = NtAllocateVirtualMemory(GetCurrentProcess(), &ReservedAddress, 0, &ReservedSize, MEM_RESERVE, PAGE_NOACCESS); if (!NT_SUCCESS(status)) { ReservedAddress = NULL; if (status != /* STATUS_CONFLICTING_ADDRESSES */ 0xC0000018) goto bailout_ntstatus /* no way to recovery */; if (may_move) /* the base address could be changed */ map->address = NULL; } err = mdbx_filesize(map->fd, &map->filesize); if (err != MDBX_SUCCESS) goto bailout; if ((flags & MDBX_RDONLY) == 0 && map->filesize != size) { err = mdbx_ftruncate(map->fd, size); if (err == MDBX_SUCCESS) map->filesize = size; /* ignore error, because Windows unable shrink file * that already mapped (by another process) */ } SectionSize.QuadPart = size; status = NtCreateSection( &map->section, /* DesiredAccess */ (flags & MDBX_WRITEMAP) ? SECTION_QUERY | SECTION_MAP_READ | SECTION_EXTEND_SIZE | SECTION_MAP_WRITE : SECTION_QUERY | SECTION_MAP_READ | SECTION_EXTEND_SIZE, /* ObjectAttributes */ NULL, /* MaximumSize (InitialSize) */ &SectionSize, /* SectionPageProtection */ (flags & MDBX_RDONLY) ? PAGE_READONLY : PAGE_READWRITE, /* AllocationAttributes */ SEC_RESERVE, map->fd); if (!NT_SUCCESS(status)) goto bailout_ntstatus; if (ReservedAddress) { /* release reserved address space */ ReservedSize = 0; status = NtFreeVirtualMemory(GetCurrentProcess(), &ReservedAddress, &ReservedSize, MEM_RELEASE); ReservedAddress = NULL; if (!NT_SUCCESS(status)) goto bailout_ntstatus; } retry_mapview:; SIZE_T ViewSize = (flags & MDBX_RDONLY) ? size : limit; status = NtMapViewOfSection( map->section, GetCurrentProcess(), &map->address, /* ZeroBits */ 0, /* CommitSize */ 0, /* SectionOffset */ NULL, &ViewSize, /* InheritDisposition */ ViewUnmap, /* AllocationType */ (flags & MDBX_RDONLY) ? 0 : MEM_RESERVE, /* Win32Protect */ (flags & MDBX_WRITEMAP) ? PAGE_READWRITE : PAGE_READONLY); if (!NT_SUCCESS(status)) { if (status == /* STATUS_CONFLICTING_ADDRESSES */ 0xC0000018 && map->address && may_move) { /* try remap at another base address */ map->address = NULL; goto retry_mapview; } NtClose(map->section); map->section = NULL; if (map->address && (size != map->current || limit != map->limit)) { /* try remap with previously size and limit, * but will return MDBX_UNABLE_EXTEND_MAPSIZE on success */ rc = MDBX_UNABLE_EXTEND_MAPSIZE; size = map->current; ReservedSize = limit = map->limit; goto retry_file_and_section; } /* no way to recovery */ goto bailout_ntstatus; } assert(map->address != MAP_FAILED); map->current = (size_t)SectionSize.QuadPart; map->limit = ViewSize; #else uint64_t filesize = 0; int rc = mdbx_filesize(map->fd, &filesize); if (rc != MDBX_SUCCESS) return rc; if (flags & MDBX_RDONLY) { map->current = (filesize > limit) ? limit : (size_t)filesize; if (map->current != size) rc = MDBX_UNABLE_EXTEND_MAPSIZE; } else if (filesize != size) { rc = mdbx_ftruncate(map->fd, size); if (rc != MDBX_SUCCESS) return rc; map->current = size; } if (limit == map->limit) return MDBX_SUCCESS; if (limit < map->limit) { /* unmap an excess at end of mapping. */ if (unlikely(munmap(map->dxb + limit, map->limit - limit))) return errno; map->limit = limit; return MDBX_SUCCESS; } assert(limit > map->limit); uint8_t *ptr = MAP_FAILED; #if defined(MREMAP_MAYMOVE) ptr = mremap(map->address, map->limit, limit, may_move ? MREMAP_MAYMOVE : 0); if (ptr == MAP_FAILED) { const int err = errno; switch (err) { default: return err; case EAGAIN: case ENOMEM: return MDBX_UNABLE_EXTEND_MAPSIZE; case EFAULT /* MADV_DODUMP / MADV_DONTDUMP are mixed for mmap-range */: break; } } #endif /* MREMAP_MAYMOVE */ const unsigned mmap_flags = MAP_CONCEAL | MAP_SHARED | MAP_FILE | (F_ISSET(flags, MDBX_UTTERLY_NOSYNC) ? MAP_NOSYNC : 0); const unsigned mmap_prot = (flags & MDBX_WRITEMAP) ? PROT_READ | PROT_WRITE : PROT_READ; if (ptr == MAP_FAILED) { /* Try to mmap additional space beyond the end of mapping. */ ptr = mmap(map->dxb + map->limit, limit - map->limit, mmap_prot, mmap_flags #if defined(MAP_FIXED_NOREPLACE) | MAP_FIXED_NOREPLACE #endif /* MAP_FIXED_NOREPLACE */ , map->fd, map->limit); if (ptr == map->dxb + map->limit) ptr = map->dxb; else if (ptr != MAP_FAILED) { /* the desired address is busy, unmap unsuitable one */ if (unlikely(munmap(ptr, limit - map->limit))) return errno; ptr = MAP_FAILED; } else { const int err = errno; switch (err) { default: return err; case EAGAIN: case ENOMEM: return MDBX_UNABLE_EXTEND_MAPSIZE; case EEXIST: /* address busy */ case EINVAL: /* kernel don't support MAP_FIXED_NOREPLACE */ break; } } } if (ptr == MAP_FAILED) { /* unmap and map again whole region */ if (!may_move) { /* TODO: Perhaps here it is worth to implement suspend/resume threads * and perform unmap/map as like for Windows. */ return MDBX_UNABLE_EXTEND_MAPSIZE; } if (unlikely(munmap(map->address, map->limit))) return errno; ptr = mmap(map->address, limit, mmap_prot, mmap_flags, map->fd, 0); if (unlikely(ptr == MAP_FAILED)) { ptr = mmap(map->address, map->limit, mmap_prot, mmap_flags, map->fd, 0); if (unlikely(ptr == MAP_FAILED)) { VALGRIND_MAKE_MEM_NOACCESS(map->address, map->current); /* Unpoisoning is required for ASAN to avoid false-positive diagnostic * when this memory will re-used by malloc or another mmapping. * See https://github.com/erthink/libmdbx/pull/93#issuecomment-613687203 */ ASAN_UNPOISON_MEMORY_REGION(map->address, map->limit); map->limit = 0; map->current = 0; map->address = nullptr; return errno; } rc = MDBX_UNABLE_EXTEND_MAPSIZE; limit = map->limit; } } assert(ptr && ptr != MAP_FAILED); if (map->address != ptr) { VALGRIND_MAKE_MEM_NOACCESS(map->address, map->current); /* Unpoisoning is required for ASAN to avoid false-positive diagnostic * when this memory will re-used by malloc or another mmapping. * See https://github.com/erthink/libmdbx/pull/93#issuecomment-613687203 */ ASAN_UNPOISON_MEMORY_REGION(map->address, map->limit); VALGRIND_MAKE_MEM_DEFINED(ptr, map->current); ASAN_UNPOISON_MEMORY_REGION(ptr, map->current); map->address = ptr; } map->limit = limit; #ifdef MADV_DONTFORK if (unlikely(madvise(map->address, map->limit, MADV_DONTFORK) != 0)) return errno; #endif /* MADV_DONTFORK */ #ifdef MADV_NOHUGEPAGE (void)madvise(map->address, map->limit, MADV_NOHUGEPAGE); #endif /* MADV_NOHUGEPAGE */ #endif /* POSIX / Windows */ return rc; } /*----------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC __cold void mdbx_osal_jitter(bool tiny) { for (;;) { #if defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \ defined(__x86_64__) const unsigned salt = 277u * (unsigned)__rdtsc(); #else const unsigned salt = rand(); #endif const unsigned coin = salt % (tiny ? 29u : 43u); if (coin < 43 / 3) break; #if defined(_WIN32) || defined(_WIN64) SwitchToThread(); if (coin > 43 * 2 / 3) Sleep(1); #else sched_yield(); if (coin > 43 * 2 / 3) usleep(coin); #endif } } #if defined(_WIN32) || defined(_WIN64) #elif defined(__APPLE__) || defined(__MACH__) #include #elif defined(__linux__) || defined(__gnu_linux__) static __cold clockid_t choice_monoclock(void) { struct timespec probe; #if defined(CLOCK_BOOTTIME) if (clock_gettime(CLOCK_BOOTTIME, &probe) == 0) return CLOCK_BOOTTIME; #elif defined(CLOCK_MONOTONIC_RAW) if (clock_gettime(CLOCK_MONOTONIC_RAW, &probe) == 0) return CLOCK_MONOTONIC_RAW; #elif defined(CLOCK_MONOTONIC_COARSE) if (clock_gettime(CLOCK_MONOTONIC_COARSE, &probe) == 0) return CLOCK_MONOTONIC_COARSE; #endif return CLOCK_MONOTONIC; } #endif /*----------------------------------------------------------------------------*/ #if defined(_WIN32) || defined(_WIN64) static LARGE_INTEGER performance_frequency; #elif defined(__APPLE__) || defined(__MACH__) static uint64_t ratio_16dot16_to_monotine; #endif MDBX_INTERNAL_FUNC uint64_t mdbx_osal_16dot16_to_monotime(uint32_t seconds_16dot16) { #if defined(_WIN32) || defined(_WIN64) if (unlikely(performance_frequency.QuadPart == 0)) QueryPerformanceFrequency(&performance_frequency); const uint64_t ratio = performance_frequency.QuadPart; #elif defined(__APPLE__) || defined(__MACH__) if (unlikely(ratio_16dot16_to_monotine == 0)) { mach_timebase_info_data_t ti; mach_timebase_info(&ti); ratio_16dot16_to_monotine = UINT64_C(1000000000) * ti.denom / ti.numer; } const uint64_t ratio = ratio_16dot16_to_monotine; #else const uint64_t ratio = UINT64_C(1000000000); #endif return (ratio * seconds_16dot16 + 32768) >> 16; } MDBX_INTERNAL_FUNC uint32_t mdbx_osal_monotime_to_16dot16(uint64_t monotime) { static uint64_t limit; if (unlikely(monotime > limit)) { if (limit != 0) return UINT32_MAX; limit = mdbx_osal_16dot16_to_monotime(UINT32_MAX - 1); if (monotime > limit) return UINT32_MAX; } #if defined(_WIN32) || defined(_WIN64) return (uint32_t)((monotime << 16) / performance_frequency.QuadPart); #elif defined(__APPLE__) || defined(__MACH__) return (uint32_t)((monotime << 16) / ratio_16dot16_to_monotine); #else return (uint32_t)(monotime * 128 / 1953125); #endif } MDBX_INTERNAL_FUNC uint64_t mdbx_osal_monotime(void) { #if defined(_WIN32) || defined(_WIN64) LARGE_INTEGER counter; counter.QuadPart = 0; QueryPerformanceCounter(&counter); return counter.QuadPart; #elif defined(__APPLE__) || defined(__MACH__) return mach_absolute_time(); #else #if defined(__linux__) || defined(__gnu_linux__) static clockid_t posix_clockid = -1; if (unlikely(posix_clockid < 0)) posix_clockid = choice_monoclock(); #elif defined(CLOCK_MONOTONIC) #define posix_clockid CLOCK_MONOTONIC #else #define posix_clockid CLOCK_REALTIME #endif struct timespec ts; if (unlikely(clock_gettime(posix_clockid, &ts) != 0)) { ts.tv_nsec = 0; ts.tv_sec = 0; } return ts.tv_sec * UINT64_C(1000000000) + ts.tv_nsec; #endif } /*----------------------------------------------------------------------------*/ static void bootid_shake(bin128_t *p) { /* Bob Jenkins's PRNG: https://burtleburtle.net/bob/rand/smallprng.html */ const uint32_t e = p->a - (p->b << 23 | p->b >> 9); p->a = p->b ^ (p->c << 16 | p->c >> 16); p->b = p->c + (p->d << 11 | p->d >> 21); p->c = p->d + e; p->d = e + p->a; } static void bootid_collect(bin128_t *p, const void *s, size_t n) { p->y += UINT64_C(64526882297375213); bootid_shake(p); for (size_t i = 0; i < n; ++i) { bootid_shake(p); p->y ^= UINT64_C(48797879452804441) * ((const uint8_t *)s)[i]; bootid_shake(p); p->y += 14621231; } bootid_shake(p); /* minor non-linear tomfoolery */ const unsigned z = p->x % 61; p->y = p->y << z | p->y >> (64 - z); bootid_shake(p); bootid_shake(p); const unsigned q = p->x % 59; p->y = p->y << q | p->y >> (64 - q); bootid_shake(p); bootid_shake(p); bootid_shake(p); } #if defined(_WIN32) || defined(_WIN64) static uint64_t windows_systemtime_ms() { FILETIME ft; GetSystemTimeAsFileTime(&ft); return ((uint64_t)ft.dwHighDateTime << 32 | ft.dwLowDateTime) / 10000ul; } static uint64_t windows_bootime(void) { unsigned confirmed = 0; uint64_t boottime = 0; uint64_t up0 = mdbx_GetTickCount64(); uint64_t st0 = windows_systemtime_ms(); for (uint64_t fuse = st0; up0 && st0 < fuse + 1000 * 1000u / 42;) { YieldProcessor(); const uint64_t up1 = mdbx_GetTickCount64(); const uint64_t st1 = windows_systemtime_ms(); if (st1 > fuse && st1 == st0 && up1 == up0) { uint64_t diff = st1 - up1; if (boottime == diff) { if (++confirmed > 4) return boottime; } else { confirmed = 0; boottime = diff; } fuse = st1; Sleep(1); } st0 = st1; up0 = up1; } return 0; } static LSTATUS mdbx_RegGetValue(HKEY hKey, LPCSTR lpSubKey, LPCSTR lpValue, PVOID pvData, LPDWORD pcbData) { LSTATUS rc; if (!mdbx_RegGetValueA) { /* an old Windows 2000/XP */ HKEY hSubKey; rc = RegOpenKeyA(hKey, lpSubKey, &hSubKey); if (rc == ERROR_SUCCESS) { rc = RegQueryValueExA(hSubKey, lpValue, NULL, NULL, pvData, pcbData); RegCloseKey(hSubKey); } return rc; } rc = mdbx_RegGetValueA(hKey, lpSubKey, lpValue, RRF_RT_ANY, NULL, pvData, pcbData); if (rc != ERROR_FILE_NOT_FOUND) return rc; rc = mdbx_RegGetValueA(hKey, lpSubKey, lpValue, RRF_RT_ANY | 0x00010000 /* RRF_SUBKEY_WOW6464KEY */, NULL, pvData, pcbData); if (rc != ERROR_FILE_NOT_FOUND) return rc; return mdbx_RegGetValueA(hKey, lpSubKey, lpValue, RRF_RT_ANY | 0x00020000 /* RRF_SUBKEY_WOW6432KEY */, NULL, pvData, pcbData); } #endif static __cold __maybe_unused bool bootid_parse_uuid(bin128_t *s, const void *p, const size_t n) { if (n > 31) { unsigned bits = 0; for (unsigned i = 0; i < n; ++i) /* try parse an UUID in text form */ { uint8_t c = ((const uint8_t *)p)[i]; if (c >= '0' && c <= '9') c -= '0'; else if (c >= 'a' && c <= 'f') c -= 'a' - 10; else if (c >= 'A' && c <= 'F') c -= 'A' - 10; else continue; assert(c <= 15); c ^= s->y >> 60; s->y = s->y << 4 | s->x >> 60; s->x = s->x << 4 | c; bits += 4; } if (bits > 42 * 3) /* UUID parsed successfully */ return true; } if (n > 15) /* is enough handle it as a binary? */ { if (n == sizeof(bin128_t)) { bin128_t aligned; memcpy(&aligned, p, sizeof(bin128_t)); s->x += aligned.x; s->y += aligned.y; } else bootid_collect(s, p, n); return true; } if (n) bootid_collect(s, p, n); return false; } __cold MDBX_INTERNAL_FUNC bin128_t mdbx_osal_bootid(void) { bin128_t bin = {{0, 0}}; bool got_machineid = false, got_boottime = false, got_bootseq = false; #if defined(__linux__) || defined(__gnu_linux__) { const int fd = open("/proc/sys/kernel/random/boot_id", O_RDONLY | O_NOFOLLOW); if (fd != -1) { struct statfs fs; char buf[42]; const ssize_t len = (fstatfs(fd, &fs) == 0 && fs.f_type == /* procfs */ 0x9FA0) ? read(fd, buf, sizeof(buf)) : -1; const int err = close(fd); assert(err == 0); (void)err; if (len > 0 && bootid_parse_uuid(&bin, buf, len)) return bin; } } #endif /* Linux */ #if defined(__APPLE__) || defined(__MACH__) { char buf[42]; size_t len = sizeof(buf); if (!sysctlbyname("kern.bootsessionuuid", buf, &len, nullptr, 0) && bootid_parse_uuid(&bin, buf, len)) return bin; #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && \ __MAC_OS_X_VERSION_MIN_REQUIRED > 1050 uuid_t uuid; struct timespec wait = {0, 1000000000u / 42}; if (!gethostuuid(uuid, &wait) && bootid_parse_uuid(&bin, uuid, sizeof(uuid))) got_machineid = true; #endif /* > 10.5 */ struct timeval boottime; len = sizeof(boottime); if (!sysctlbyname("kern.boottime", &boottime, &len, nullptr, 0) && len == sizeof(boottime) && boottime.tv_sec) got_boottime = true; } #endif /* Apple/Darwin */ #if defined(_WIN32) || defined(_WIN64) { union buf { DWORD BootId; DWORD BaseTime; SYSTEM_TIMEOFDAY_INFORMATION SysTimeOfDayInfo; struct { LARGE_INTEGER BootTime; LARGE_INTEGER CurrentTime; LARGE_INTEGER TimeZoneBias; ULONG TimeZoneId; ULONG Reserved; ULONGLONG BootTimeBias; ULONGLONG SleepTimeBias; } SysTimeOfDayInfoHacked; wchar_t MachineGuid[42]; char DigitalProductId[248]; } buf; static const char HKLM_MicrosoftCryptography[] = "SOFTWARE\\Microsoft\\Cryptography"; DWORD len = sizeof(buf); /* Windows is madness and must die */ if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_MicrosoftCryptography, "MachineGuid", &buf.MachineGuid, &len) == ERROR_SUCCESS && len > 42 && len < sizeof(buf)) got_machineid = bootid_parse_uuid(&bin, &buf.MachineGuid, len); if (!got_machineid) { /* again, Windows is madness */ static const char HKLM_WindowsNT[] = "SOFTWARE\\Microsoft\\Windows NT\\CurrentVersion"; static const char HKLM_WindowsNT_DPK[] = "SOFTWARE\\Microsoft\\Windows " "NT\\CurrentVersion\\DefaultProductKey"; static const char HKLM_WindowsNT_DPK2[] = "SOFTWARE\\Microsoft\\Windows " "NT\\CurrentVersion\\DefaultProductKey2"; len = sizeof(buf); if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_WindowsNT, "DigitalProductId", &buf.DigitalProductId, &len) == ERROR_SUCCESS && len > 42 && len < sizeof(buf)) { bootid_collect(&bin, &buf.DigitalProductId, len); got_machineid = true; } len = sizeof(buf); if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_WindowsNT_DPK, "DigitalProductId", &buf.DigitalProductId, &len) == ERROR_SUCCESS && len > 42 && len < sizeof(buf)) { bootid_collect(&bin, &buf.DigitalProductId, len); got_machineid = true; } len = sizeof(buf); if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_WindowsNT_DPK2, "DigitalProductId", &buf.DigitalProductId, &len) == ERROR_SUCCESS && len > 42 && len < sizeof(buf)) { bootid_collect(&bin, &buf.DigitalProductId, len); got_machineid = true; } } static const char HKLM_PrefetcherParams[] = "SYSTEM\\CurrentControlSet\\Control\\Session Manager\\Memory " "Management\\PrefetchParameters"; len = sizeof(buf); if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_PrefetcherParams, "BootId", &buf.BootId, &len) == ERROR_SUCCESS && len > 1 && len < sizeof(buf)) { bootid_collect(&bin, &buf.BootId, len); got_bootseq = true; } len = sizeof(buf); if (mdbx_RegGetValue(HKEY_LOCAL_MACHINE, HKLM_PrefetcherParams, "BaseTime", &buf.BaseTime, &len) == ERROR_SUCCESS && len >= sizeof(buf.BaseTime) && buf.BaseTime) { bootid_collect(&bin, &buf.BaseTime, len); got_boottime = true; } /* BootTime from SYSTEM_TIMEOFDAY_INFORMATION */ NTSTATUS status = NtQuerySystemInformation( 0x03 /* SystemTmeOfDayInformation */, &buf.SysTimeOfDayInfo, sizeof(buf.SysTimeOfDayInfo), &len); if (NT_SUCCESS(status) && len >= offsetof(union buf, SysTimeOfDayInfoHacked.BootTime) + sizeof(buf.SysTimeOfDayInfoHacked.BootTime) && buf.SysTimeOfDayInfoHacked.BootTime.QuadPart) { bootid_collect(&bin, &buf.SysTimeOfDayInfoHacked.BootTime, sizeof(buf.SysTimeOfDayInfoHacked.BootTime)); got_boottime = true; } if (!got_boottime) { uint64_t boottime = windows_bootime(); if (boottime) { bootid_collect(&bin, &boottime, sizeof(boottime)); got_boottime = true; } } } #endif /* Windows */ #if defined(CTL_HW) && defined(HW_UUID) if (!got_machineid) { static const int mib[] = {CTL_HW, HW_UUID}; char buf[42]; size_t len = sizeof(buf); if (sysctl( #ifdef SYSCTL_LEGACY_NONCONST_MIB (int *) #endif mib, ARRAY_LENGTH(mib), &buf, &len, NULL, 0) == 0) got_machineid = bootid_parse_uuid(&bin, buf, len); } #endif /* CTL_HW && HW_UUID */ #if defined(CTL_KERN) && defined(KERN_HOSTUUID) if (!got_machineid) { static const int mib[] = {CTL_KERN, KERN_HOSTUUID}; char buf[42]; size_t len = sizeof(buf); if (sysctl( #ifdef SYSCTL_LEGACY_NONCONST_MIB (int *) #endif mib, ARRAY_LENGTH(mib), &buf, &len, NULL, 0) == 0) got_machineid = bootid_parse_uuid(&bin, buf, len); } #endif /* CTL_KERN && KERN_HOSTUUID */ #if defined(__NetBSD__) if (!got_machineid) { char buf[42]; size_t len = sizeof(buf); if (sysctlbyname("machdep.dmi.system-uuid", buf, &len, NULL, 0) == 0) got_machineid = bootid_parse_uuid(&bin, buf, len); } #endif /* __NetBSD__ */ #if _XOPEN_SOURCE_EXTENDED if (!got_machineid) { const int hostid = gethostid(); if (hostid > 0) { bootid_collect(&bin, &hostid, sizeof(hostid)); got_machineid = true; } } #endif /* _XOPEN_SOURCE_EXTENDED */ if (!got_machineid) { lack: bin.x = bin.y = 0; return bin; } /*--------------------------------------------------------------------------*/ #if defined(CTL_KERN) && defined(KERN_BOOTTIME) if (!got_boottime) { static const int mib[] = {CTL_KERN, KERN_BOOTTIME}; struct timeval boottime; size_t len = sizeof(boottime); if (sysctl( #ifdef SYSCTL_LEGACY_NONCONST_MIB (int *) #endif mib, ARRAY_LENGTH(mib), &boottime, &len, NULL, 0) == 0 && len == sizeof(boottime) && boottime.tv_sec) { bootid_collect(&bin, &boottime, len); got_boottime = true; } } #endif /* CTL_KERN && KERN_BOOTTIME */ #if defined(__sun) || defined(__SVR4) || defined(__svr4__) if (!got_boottime) { kstat_ctl_t *kc = kstat_open(); if (kc) { kstat_t *kp = kstat_lookup(kc, "unix", 0, "system_misc"); if (kp && kstat_read(kc, kp, 0) != -1) { kstat_named_t *kn = (kstat_named_t *)kstat_data_lookup(kp, "boot_time"); if (kn) { switch (kn->data_type) { case KSTAT_DATA_INT32: case KSTAT_DATA_UINT32: bootid_collect(&bin, &kn->value, sizeof(int32_t)); got_boottime = true; case KSTAT_DATA_INT64: case KSTAT_DATA_UINT64: bootid_collect(&bin, &kn->value, sizeof(int64_t)); got_boottime = true; } } } kstat_close(kc); } } #endif /* SunOS / Solaris */ #if _XOPEN_SOURCE_EXTENDED && defined(BOOT_TIME) if (!got_boottime) { setutxent(); const struct utmpx id = {.ut_type = BOOT_TIME}; const struct utmpx *entry = getutxid(&id); if (entry) { bootid_collect(&bin, entry, sizeof(*entry)); got_boottime = true; while (unlikely((entry = getutxid(&id)) != nullptr)) { /* have multiple reboot records, assuming we can distinguish next * bootsession even if RTC is wrong or absent */ bootid_collect(&bin, entry, sizeof(*entry)); got_bootseq = true; } } endutxent(); } #endif /* _XOPEN_SOURCE_EXTENDED && BOOT_TIME */ if (!got_bootseq) { if (!got_boottime || !MDBX_TRUST_RTC) goto lack; #if defined(_WIN32) || defined(_WIN64) FILETIME now; GetSystemTimeAsFileTime(&now); if (0x1CCCCCC > now.dwHighDateTime) #else struct timespec mono, real; if (clock_gettime(CLOCK_MONOTONIC, &mono) || clock_gettime(CLOCK_REALTIME, &real) || /* wrong time, RTC is mad or absent */ 1555555555l > real.tv_sec || /* seems no adjustment by RTC/NTP, i.e. a fake time */ real.tv_sec < mono.tv_sec || 1234567890l > real.tv_sec - mono.tv_sec || (real.tv_sec - mono.tv_sec) % 900u == 0) #endif goto lack; } return bin; } /* This is CMake-template for libmdbx's version.c ******************************************************************************/ #if MDBX_VERSION_MAJOR != 0 || \ MDBX_VERSION_MINOR != 9 #error "API version mismatch! Had `git fetch --tags` done?" #endif static const char sourcery[] = STRINGIFY(MDBX_BUILD_SOURCERY); __dll_export #ifdef __attribute_used__ __attribute_used__ #elif defined(__GNUC__) || __has_attribute(__used__) __attribute__((__used__)) #endif #ifdef __attribute_externally_visible__ __attribute_externally_visible__ #elif (defined(__GNUC__) && !defined(__clang__)) || \ __has_attribute(__externally_visible__) __attribute__((__externally_visible__)) #endif const struct MDBX_version_info mdbx_version = { 0, 9, 2, 136, {"2021-01-30T02:28:12+03:00", "1ab2fa795a0dee8678b168ffcd961ca5a0fff7c5", "9f0ff865e860435b243d3d1e936e3067097a294f", "v0.9.2-136-g9f0ff86"}, sourcery}; __dll_export #ifdef __attribute_used__ __attribute_used__ #elif defined(__GNUC__) || __has_attribute(__used__) __attribute__((__used__)) #endif #ifdef __attribute_externally_visible__ __attribute_externally_visible__ #elif (defined(__GNUC__) && !defined(__clang__)) || \ __has_attribute(__externally_visible__) __attribute__((__externally_visible__)) #endif const char *const mdbx_sourcery_anchor = sourcery; /* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ #if defined(_WIN32) || defined(_WIN64) /* Windows LCK-implementation */ /* PREAMBLE FOR WINDOWS: * * We are not concerned for performance here. * If you are running Windows a performance could NOT be the goal. * Otherwise please use Linux. */ static void mdbx_winnt_import(void); #if MDBX_BUILD_SHARED_LIBRARY #if MDBX_AVOID_CRT && defined(NDEBUG) /* DEBUG/CHECKED builds still require MSVC's CRT for runtime checks. * * Define dll's entry point only for Release build when NDEBUG is defined and * MDBX_AVOID_CRT=ON. if the entry point isn't defined then MSVC's will * automatically use DllMainCRTStartup() from CRT library, which also * automatically call DllMain() from our mdbx.dll */ #pragma comment(linker, "/ENTRY:DllMain") #endif /* MDBX_AVOID_CRT */ BOOL APIENTRY DllMain(HANDLE module, DWORD reason, LPVOID reserved) #else #if !MDBX_CONFIG_MANUAL_TLS_CALLBACK static #endif /* !MDBX_CONFIG_MANUAL_TLS_CALLBACK */ void NTAPI mdbx_dll_handler(PVOID module, DWORD reason, PVOID reserved) #endif /* MDBX_BUILD_SHARED_LIBRARY */ { (void)reserved; switch (reason) { case DLL_PROCESS_ATTACH: mdbx_winnt_import(); mdbx_rthc_global_init(); break; case DLL_PROCESS_DETACH: mdbx_rthc_global_dtor(); break; case DLL_THREAD_ATTACH: break; case DLL_THREAD_DETACH: mdbx_rthc_thread_dtor(module); break; } #if MDBX_BUILD_SHARED_LIBRARY return TRUE; #endif } #if !MDBX_BUILD_SHARED_LIBRARY && !MDBX_CONFIG_MANUAL_TLS_CALLBACK /* *INDENT-OFF* */ /* clang-format off */ #if defined(_MSC_VER) # pragma const_seg(push) # pragma data_seg(push) # ifdef _WIN64 /* kick a linker to create the TLS directory if not already done */ # pragma comment(linker, "/INCLUDE:_tls_used") /* Force some symbol references. */ # pragma comment(linker, "/INCLUDE:mdbx_tls_anchor") /* specific const-segment for WIN64 */ # pragma const_seg(".CRT$XLB") const # else /* kick a linker to create the TLS directory if not already done */ # pragma comment(linker, "/INCLUDE:__tls_used") /* Force some symbol references. */ # pragma comment(linker, "/INCLUDE:_mdbx_tls_anchor") /* specific data-segment for WIN32 */ # pragma data_seg(".CRT$XLB") # endif __declspec(allocate(".CRT$XLB")) PIMAGE_TLS_CALLBACK mdbx_tls_anchor = mdbx_dll_handler; # pragma data_seg(pop) # pragma const_seg(pop) #elif defined(__GNUC__) # ifdef _WIN64 const # endif PIMAGE_TLS_CALLBACK mdbx_tls_anchor __attribute__((__section__(".CRT$XLB"), used)) = mdbx_dll_handler; #else # error FIXME #endif /* *INDENT-ON* */ /* clang-format on */ #endif /* !MDBX_BUILD_SHARED_LIBRARY && !MDBX_CONFIG_MANUAL_TLS_CALLBACK */ /*----------------------------------------------------------------------------*/ #define LCK_SHARED 0 #define LCK_EXCLUSIVE LOCKFILE_EXCLUSIVE_LOCK #define LCK_WAITFOR 0 #define LCK_DONTWAIT LOCKFILE_FAIL_IMMEDIATELY static __inline BOOL flock(mdbx_filehandle_t fd, DWORD flags, uint64_t offset, size_t bytes) { OVERLAPPED ov; ov.hEvent = 0; ov.Offset = (DWORD)offset; ov.OffsetHigh = HIGH_DWORD(offset); return LockFileEx(fd, flags, 0, (DWORD)bytes, HIGH_DWORD(bytes), &ov); } static __inline BOOL funlock(mdbx_filehandle_t fd, uint64_t offset, size_t bytes) { return UnlockFile(fd, (DWORD)offset, HIGH_DWORD(offset), (DWORD)bytes, HIGH_DWORD(bytes)); } /*----------------------------------------------------------------------------*/ /* global `write` lock for write-txt processing, * exclusive locking both meta-pages) */ #define LCK_MAXLEN (1u + (size_t)(MAXSSIZE_T)) #define LCK_META_OFFSET 0 #define LCK_META_LEN (MAX_PAGESIZE * NUM_METAS) #define LCK_BODY_OFFSET LCK_META_LEN #define LCK_BODY_LEN (LCK_MAXLEN - LCK_BODY_OFFSET) #define LCK_BODY LCK_BODY_OFFSET, LCK_BODY_LEN #define LCK_WHOLE 0, LCK_MAXLEN int mdbx_txn_lock(MDBX_env *env, bool dontwait) { if (dontwait) { if (!TryEnterCriticalSection(&env->me_windowsbug_lock)) return MDBX_BUSY; } else { __try { EnterCriticalSection(&env->me_windowsbug_lock); } __except ((GetExceptionCode() == 0xC0000194 /* STATUS_POSSIBLE_DEADLOCK / EXCEPTION_POSSIBLE_DEADLOCK */) ? EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH) { return ERROR_POSSIBLE_DEADLOCK; } } if ((env->me_flags & MDBX_EXCLUSIVE) || flock(env->me_lazy_fd, dontwait ? (LCK_EXCLUSIVE | LCK_DONTWAIT) : (LCK_EXCLUSIVE | LCK_WAITFOR), LCK_BODY)) return MDBX_SUCCESS; int rc = GetLastError(); LeaveCriticalSection(&env->me_windowsbug_lock); return (!dontwait || rc != ERROR_LOCK_VIOLATION) ? rc : MDBX_BUSY; } void mdbx_txn_unlock(MDBX_env *env) { int rc = (env->me_flags & MDBX_EXCLUSIVE) ? TRUE : funlock(env->me_lazy_fd, LCK_BODY); LeaveCriticalSection(&env->me_windowsbug_lock); if (!rc) mdbx_panic("%s failed: err %u", __func__, GetLastError()); } /*----------------------------------------------------------------------------*/ /* global `read` lock for readers registration, * exclusive locking `mti_numreaders` (second) cacheline */ #define LCK_LO_OFFSET 0 #define LCK_LO_LEN offsetof(MDBX_lockinfo, mti_numreaders) #define LCK_UP_OFFSET LCK_LO_LEN #define LCK_UP_LEN (sizeof(MDBX_lockinfo) - LCK_UP_OFFSET) #define LCK_LOWER LCK_LO_OFFSET, LCK_LO_LEN #define LCK_UPPER LCK_UP_OFFSET, LCK_UP_LEN MDBX_INTERNAL_FUNC int mdbx_rdt_lock(MDBX_env *env) { mdbx_srwlock_AcquireShared(&env->me_remap_guard); if (env->me_lfd == INVALID_HANDLE_VALUE) return MDBX_SUCCESS; /* readonly database in readonly filesystem */ /* transition from S-? (used) to S-E (locked), * e.g. exclusive lock upper-part */ if ((env->me_flags & MDBX_EXCLUSIVE) || flock(env->me_lfd, LCK_EXCLUSIVE | LCK_WAITFOR, LCK_UPPER)) return MDBX_SUCCESS; int rc = GetLastError(); mdbx_srwlock_ReleaseShared(&env->me_remap_guard); return rc; } MDBX_INTERNAL_FUNC void mdbx_rdt_unlock(MDBX_env *env) { if (env->me_lfd != INVALID_HANDLE_VALUE) { /* transition from S-E (locked) to S-? (used), e.g. unlock upper-part */ if ((env->me_flags & MDBX_EXCLUSIVE) == 0 && !funlock(env->me_lfd, LCK_UPPER)) mdbx_panic("%s failed: err %u", __func__, GetLastError()); } mdbx_srwlock_ReleaseShared(&env->me_remap_guard); } MDBX_INTERNAL_FUNC int mdbx_lockfile(mdbx_filehandle_t fd, bool wait) { return flock(fd, wait ? LCK_EXCLUSIVE | LCK_WAITFOR : LCK_EXCLUSIVE | LCK_DONTWAIT, 0, LCK_MAXLEN) ? MDBX_SUCCESS : GetLastError(); } static int suspend_and_append(mdbx_handle_array_t **array, const DWORD ThreadId) { const unsigned limit = (*array)->limit; if ((*array)->count == limit) { void *ptr = mdbx_realloc( (limit > ARRAY_LENGTH((*array)->handles)) ? *array : /* don't free initial array on the stack */ NULL, sizeof(mdbx_handle_array_t) + sizeof(HANDLE) * (limit * 2 - ARRAY_LENGTH((*array)->handles))); if (!ptr) return MDBX_ENOMEM; if (limit == ARRAY_LENGTH((*array)->handles)) memcpy(ptr, *array, sizeof(mdbx_handle_array_t)); *array = (mdbx_handle_array_t *)ptr; (*array)->limit = limit * 2; } HANDLE hThread = OpenThread(THREAD_SUSPEND_RESUME | THREAD_QUERY_INFORMATION, FALSE, ThreadId); if (hThread == NULL) return GetLastError(); if (SuspendThread(hThread) == -1) { int err = GetLastError(); DWORD ExitCode; if (err == /* workaround for Win10 UCRT bug */ ERROR_ACCESS_DENIED || !GetExitCodeThread(hThread, &ExitCode) || ExitCode != STILL_ACTIVE) err = MDBX_SUCCESS; CloseHandle(hThread); return err; } (*array)->handles[(*array)->count++] = hThread; return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_suspend_threads_before_remap(MDBX_env *env, mdbx_handle_array_t **array) { const uintptr_t CurrentTid = GetCurrentThreadId(); int rc; if (env->me_lck) { /* Scan LCK for threads of the current process */ const MDBX_reader *const begin = env->me_lck->mti_readers; const MDBX_reader *const end = begin + atomic_load32(&env->me_lck->mti_numreaders, mo_AcquireRelease); const uintptr_t WriteTxnOwner = env->me_txn0 ? env->me_txn0->mt_owner : 0; for (const MDBX_reader *reader = begin; reader < end; ++reader) { if (reader->mr_pid.weak != env->me_pid || !reader->mr_tid.weak) { skip_lck: continue; } if (reader->mr_tid.weak == CurrentTid || reader->mr_tid.weak == WriteTxnOwner) goto skip_lck; if (env->me_flags & MDBX_NOTLS) { /* Skip duplicates in no-tls mode */ for (const MDBX_reader *scan = reader; --scan >= begin;) if (scan->mr_tid.weak == reader->mr_tid.weak) goto skip_lck; } rc = suspend_and_append(array, (mdbx_tid_t)reader->mr_tid.weak); if (rc != MDBX_SUCCESS) { bailout_lck: (void)mdbx_resume_threads_after_remap(*array); return rc; } } if (WriteTxnOwner && WriteTxnOwner != CurrentTid) { rc = suspend_and_append(array, (mdbx_tid_t)WriteTxnOwner); if (rc != MDBX_SUCCESS) goto bailout_lck; } } else { /* Without LCK (i.e. read-only mode). * Walk through a snapshot of all running threads */ mdbx_assert(env, env->me_txn0 == NULL || (env->me_flags & MDBX_EXCLUSIVE) != 0); const HANDLE hSnapshot = CreateToolhelp32Snapshot(TH32CS_SNAPTHREAD, 0); if (hSnapshot == INVALID_HANDLE_VALUE) return GetLastError(); THREADENTRY32 entry; entry.dwSize = sizeof(THREADENTRY32); if (!Thread32First(hSnapshot, &entry)) { rc = GetLastError(); bailout_toolhelp: CloseHandle(hSnapshot); (void)mdbx_resume_threads_after_remap(*array); return rc; } do { if (entry.th32OwnerProcessID != env->me_pid || entry.th32ThreadID == CurrentTid) continue; rc = suspend_and_append(array, entry.th32ThreadID); if (rc != MDBX_SUCCESS) goto bailout_toolhelp; } while (Thread32Next(hSnapshot, &entry)); rc = GetLastError(); if (rc != ERROR_NO_MORE_FILES) goto bailout_toolhelp; CloseHandle(hSnapshot); } return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_resume_threads_after_remap(mdbx_handle_array_t *array) { int rc = MDBX_SUCCESS; for (unsigned i = 0; i < array->count; ++i) { const HANDLE hThread = array->handles[i]; if (ResumeThread(hThread) == -1) { const int err = GetLastError(); DWORD ExitCode; if (err != /* workaround for Win10 UCRT bug */ ERROR_ACCESS_DENIED && GetExitCodeThread(hThread, &ExitCode) && ExitCode == STILL_ACTIVE) rc = err; } CloseHandle(hThread); } return rc; } /*----------------------------------------------------------------------------*/ /* global `initial` lock for lockfile initialization, * exclusive/shared locking first cacheline */ /* Briefly description of locking schema/algorithm: * - Windows does not support upgrading or downgrading for file locking. * - Therefore upgrading/downgrading is emulated by shared and exclusive * locking of upper and lower halves. * - In other words, we have FSM with possible 9 states, * i.e. free/shared/exclusive x free/shared/exclusive == 9. * Only 6 states of FSM are used, which 2 of ones are transitive. * * States: * ?-? = free, i.e. unlocked * S-? = used, i.e. shared lock * E-? = exclusive-read, i.e. operational exclusive * ?-S * ?-E = middle (transitive state) * S-S * S-E = locked (transitive state) * E-S * E-E = exclusive-write, i.e. exclusive due (re)initialization * * The mdbx_lck_seize() moves the locking-FSM from the initial free/unlocked * state to the "exclusive write" (and returns MDBX_RESULT_TRUE) if possible, * or to the "used" (and returns MDBX_RESULT_FALSE). * * The mdbx_lck_downgrade() moves the locking-FSM from "exclusive write" * state to the "used" (i.e. shared) state. * * The mdbx_lck_upgrade() moves the locking-FSM from "used" (i.e. shared) * state to the "exclusive write" state. */ static void lck_unlock(MDBX_env *env) { int err; if (env->me_lfd != INVALID_HANDLE_VALUE) { /* double `unlock` for robustly remove overlapped shared/exclusive locks */ while (funlock(env->me_lfd, LCK_LOWER)) ; err = GetLastError(); assert(err == ERROR_NOT_LOCKED || (mdbx_RunningUnderWine() && err == ERROR_LOCK_VIOLATION)); (void)err; SetLastError(ERROR_SUCCESS); while (funlock(env->me_lfd, LCK_UPPER)) ; err = GetLastError(); assert(err == ERROR_NOT_LOCKED || (mdbx_RunningUnderWine() && err == ERROR_LOCK_VIOLATION)); (void)err; SetLastError(ERROR_SUCCESS); } if (env->me_lazy_fd != INVALID_HANDLE_VALUE) { /* explicitly unlock to avoid latency for other processes (windows kernel * releases such locks via deferred queues) */ while (funlock(env->me_lazy_fd, LCK_BODY)) ; err = GetLastError(); assert(err == ERROR_NOT_LOCKED || (mdbx_RunningUnderWine() && err == ERROR_LOCK_VIOLATION)); (void)err; SetLastError(ERROR_SUCCESS); while (funlock(env->me_lazy_fd, LCK_WHOLE)) ; err = GetLastError(); assert(err == ERROR_NOT_LOCKED || (mdbx_RunningUnderWine() && err == ERROR_LOCK_VIOLATION)); (void)err; SetLastError(ERROR_SUCCESS); } } /* Seize state as 'exclusive-write' (E-E and returns MDBX_RESULT_TRUE) * or as 'used' (S-? and returns MDBX_RESULT_FALSE). * Otherwise returns an error. */ static int internal_seize_lck(HANDLE lfd) { int rc; assert(lfd != INVALID_HANDLE_VALUE); /* 1) now on ?-? (free), get ?-E (middle) */ mdbx_jitter4testing(false); if (!flock(lfd, LCK_EXCLUSIVE | LCK_WAITFOR, LCK_UPPER)) { rc = GetLastError() /* 2) something went wrong, give up */; mdbx_error("%s, err %u", "?-?(free) >> ?-E(middle)", rc); return rc; } /* 3) now on ?-E (middle), try E-E (exclusive-write) */ mdbx_jitter4testing(false); if (flock(lfd, LCK_EXCLUSIVE | LCK_DONTWAIT, LCK_LOWER)) return MDBX_RESULT_TRUE /* 4) got E-E (exclusive-write), done */; /* 5) still on ?-E (middle) */ rc = GetLastError(); mdbx_jitter4testing(false); if (rc != ERROR_SHARING_VIOLATION && rc != ERROR_LOCK_VIOLATION) { /* 6) something went wrong, give up */ if (!funlock(lfd, LCK_UPPER)) mdbx_panic("%s(%s) failed: err %u", __func__, "?-E(middle) >> ?-?(free)", GetLastError()); return rc; } /* 7) still on ?-E (middle), try S-E (locked) */ mdbx_jitter4testing(false); rc = flock(lfd, LCK_SHARED | LCK_DONTWAIT, LCK_LOWER) ? MDBX_RESULT_FALSE : GetLastError(); mdbx_jitter4testing(false); if (rc != MDBX_RESULT_FALSE) mdbx_error("%s, err %u", "?-E(middle) >> S-E(locked)", rc); /* 8) now on S-E (locked) or still on ?-E (middle), * transition to S-? (used) or ?-? (free) */ if (!funlock(lfd, LCK_UPPER)) mdbx_panic("%s(%s) failed: err %u", __func__, "X-E(locked/middle) >> X-?(used/free)", GetLastError()); /* 9) now on S-? (used, DONE) or ?-? (free, FAILURE) */ return rc; } MDBX_INTERNAL_FUNC int mdbx_lck_seize(MDBX_env *env) { int rc; assert(env->me_lazy_fd != INVALID_HANDLE_VALUE); if (env->me_flags & MDBX_EXCLUSIVE) return MDBX_RESULT_TRUE /* nope since files were must be opened non-shareable */ ; if (env->me_lfd == INVALID_HANDLE_VALUE) { /* LY: without-lck mode (e.g. on read-only filesystem) */ mdbx_jitter4testing(false); if (!flock(env->me_lazy_fd, LCK_SHARED | LCK_DONTWAIT, LCK_WHOLE)) { rc = GetLastError(); mdbx_error("%s, err %u", "without-lck", rc); return rc; } return MDBX_RESULT_FALSE; } rc = internal_seize_lck(env->me_lfd); mdbx_jitter4testing(false); if (rc == MDBX_RESULT_TRUE && (env->me_flags & MDBX_RDONLY) == 0) { /* Check that another process don't operates in without-lck mode. * Doing such check by exclusive locking the body-part of db. Should be * noted: * - we need an exclusive lock for do so; * - we can't lock meta-pages, otherwise other process could get an error * while opening db in valid (non-conflict) mode. */ if (!flock(env->me_lazy_fd, LCK_EXCLUSIVE | LCK_DONTWAIT, LCK_BODY)) { rc = GetLastError(); mdbx_error("%s, err %u", "lock-against-without-lck", rc); mdbx_jitter4testing(false); lck_unlock(env); } else { mdbx_jitter4testing(false); if (!funlock(env->me_lazy_fd, LCK_BODY)) mdbx_panic("%s(%s) failed: err %u", __func__, "unlock-against-without-lck", GetLastError()); } } return rc; } MDBX_INTERNAL_FUNC int mdbx_lck_downgrade(MDBX_env *env) { /* Transite from exclusive-write state (E-E) to used (S-?) */ assert(env->me_lazy_fd != INVALID_HANDLE_VALUE); assert(env->me_lfd != INVALID_HANDLE_VALUE); if (env->me_flags & MDBX_EXCLUSIVE) return MDBX_SUCCESS /* nope since files were must be opened non-shareable */ ; /* 1) now at E-E (exclusive-write), transition to ?_E (middle) */ if (!funlock(env->me_lfd, LCK_LOWER)) mdbx_panic("%s(%s) failed: err %u", __func__, "E-E(exclusive-write) >> ?-E(middle)", GetLastError()); /* 2) now at ?-E (middle), transition to S-E (locked) */ if (!flock(env->me_lfd, LCK_SHARED | LCK_DONTWAIT, LCK_LOWER)) { int rc = GetLastError() /* 3) something went wrong, give up */; mdbx_error("%s, err %u", "?-E(middle) >> S-E(locked)", rc); return rc; } /* 4) got S-E (locked), continue transition to S-? (used) */ if (!funlock(env->me_lfd, LCK_UPPER)) mdbx_panic("%s(%s) failed: err %u", __func__, "S-E(locked) >> S-?(used)", GetLastError()); return MDBX_SUCCESS /* 5) now at S-? (used), done */; } MDBX_INTERNAL_FUNC int mdbx_lck_upgrade(MDBX_env *env) { /* Transite from used state (S-?) to exclusive-write (E-E) */ assert(env->me_lfd != INVALID_HANDLE_VALUE); if (env->me_flags & MDBX_EXCLUSIVE) return MDBX_SUCCESS /* nope since files were must be opened non-shareable */ ; int rc; /* 1) now on S-? (used), try S-E (locked) */ mdbx_jitter4testing(false); if (!flock(env->me_lfd, LCK_EXCLUSIVE | LCK_DONTWAIT, LCK_UPPER)) { rc = GetLastError() /* 2) something went wrong, give up */; mdbx_verbose("%s, err %u", "S-?(used) >> S-E(locked)", rc); return rc; } /* 3) now on S-E (locked), transition to ?-E (middle) */ if (!funlock(env->me_lfd, LCK_LOWER)) mdbx_panic("%s(%s) failed: err %u", __func__, "S-E(locked) >> ?-E(middle)", GetLastError()); /* 4) now on ?-E (middle), try E-E (exclusive-write) */ mdbx_jitter4testing(false); if (!flock(env->me_lfd, LCK_EXCLUSIVE | LCK_DONTWAIT, LCK_LOWER)) { rc = GetLastError() /* 5) something went wrong, give up */; mdbx_verbose("%s, err %u", "?-E(middle) >> E-E(exclusive-write)", rc); return rc; } return MDBX_SUCCESS /* 6) now at E-E (exclusive-write), done */; } MDBX_INTERNAL_FUNC int mdbx_lck_init(MDBX_env *env, MDBX_env *inprocess_neighbor, int global_uniqueness_flag) { (void)env; (void)inprocess_neighbor; (void)global_uniqueness_flag; return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_lck_destroy(MDBX_env *env, MDBX_env *inprocess_neighbor) { /* LY: should unmap before releasing the locks to avoid race condition and * STATUS_USER_MAPPED_FILE/ERROR_USER_MAPPED_FILE */ if (env->me_map) mdbx_munmap(&env->me_dxb_mmap); if (env->me_lck) { const bool synced = env->me_lck_mmap.lck->mti_unsynced_pages.weak == 0; mdbx_munmap(&env->me_lck_mmap); if (synced && !inprocess_neighbor && env->me_lfd != INVALID_HANDLE_VALUE && mdbx_lck_upgrade(env) == MDBX_SUCCESS) /* this will fail if LCK is used/mmapped by other process(es) */ mdbx_ftruncate(env->me_lfd, 0); } lck_unlock(env); return MDBX_SUCCESS; } /*----------------------------------------------------------------------------*/ /* reader checking (by pid) */ MDBX_INTERNAL_FUNC int mdbx_rpid_set(MDBX_env *env) { (void)env; return MDBX_SUCCESS; } MDBX_INTERNAL_FUNC int mdbx_rpid_clear(MDBX_env *env) { (void)env; return MDBX_SUCCESS; } /* Checks reader by pid. * * Returns: * MDBX_RESULT_TRUE, if pid is live (unable to acquire lock) * MDBX_RESULT_FALSE, if pid is dead (lock acquired) * or otherwise the errcode. */ MDBX_INTERNAL_FUNC int mdbx_rpid_check(MDBX_env *env, uint32_t pid) { (void)env; HANDLE hProcess = OpenProcess(SYNCHRONIZE, FALSE, pid); int rc; if (likely(hProcess)) { rc = WaitForSingleObject(hProcess, 0); if (unlikely(rc == WAIT_FAILED)) rc = GetLastError(); CloseHandle(hProcess); } else { rc = GetLastError(); } switch (rc) { case ERROR_INVALID_PARAMETER: /* pid seems invalid */ return MDBX_RESULT_FALSE; case WAIT_OBJECT_0: /* process just exited */ return MDBX_RESULT_FALSE; case ERROR_ACCESS_DENIED: /* The ERROR_ACCESS_DENIED would be returned for CSRSS-processes, etc. * assume pid exists */ return MDBX_RESULT_TRUE; case WAIT_TIMEOUT: /* pid running */ return MDBX_RESULT_TRUE; default: /* failure */ return rc; } } //---------------------------------------------------------------------------- // Stub for slim read-write lock // Copyright (C) 1995-2002 Brad Wilson static void WINAPI stub_srwlock_Init(MDBX_srwlock *srwl) { srwl->readerCount = srwl->writerCount = 0; } static void WINAPI stub_srwlock_AcquireShared(MDBX_srwlock *srwl) { while (true) { assert(srwl->writerCount >= 0 && srwl->readerCount >= 0); // If there's a writer already, spin without unnecessarily // interlocking the CPUs if (srwl->writerCount != 0) { YieldProcessor(); continue; } // Add to the readers list _InterlockedIncrement(&srwl->readerCount); // Check for writers again (we may have been preempted). If // there are no writers writing or waiting, then we're done. if (srwl->writerCount == 0) break; // Remove from the readers list, spin, try again _InterlockedDecrement(&srwl->readerCount); YieldProcessor(); } } static void WINAPI stub_srwlock_ReleaseShared(MDBX_srwlock *srwl) { assert(srwl->readerCount > 0); _InterlockedDecrement(&srwl->readerCount); } static void WINAPI stub_srwlock_AcquireExclusive(MDBX_srwlock *srwl) { while (true) { assert(srwl->writerCount >= 0 && srwl->readerCount >= 0); // If there's a writer already, spin without unnecessarily // interlocking the CPUs if (srwl->writerCount != 0) { YieldProcessor(); continue; } // See if we can become the writer (expensive, because it inter- // locks the CPUs, so writing should be an infrequent process) if (_InterlockedExchange(&srwl->writerCount, 1) == 0) break; } // Now we're the writer, but there may be outstanding readers. // Spin until there aren't any more; new readers will wait now // that we're the writer. while (srwl->readerCount != 0) { assert(srwl->writerCount >= 0 && srwl->readerCount >= 0); YieldProcessor(); } } static void WINAPI stub_srwlock_ReleaseExclusive(MDBX_srwlock *srwl) { assert(srwl->writerCount == 1 && srwl->readerCount >= 0); srwl->writerCount = 0; } MDBX_srwlock_function mdbx_srwlock_Init, mdbx_srwlock_AcquireShared, mdbx_srwlock_ReleaseShared, mdbx_srwlock_AcquireExclusive, mdbx_srwlock_ReleaseExclusive; /*----------------------------------------------------------------------------*/ #if 0 /* LY: unused for now */ static DWORD WINAPI stub_DiscardVirtualMemory(PVOID VirtualAddress, SIZE_T Size) { return VirtualAlloc(VirtualAddress, Size, MEM_RESET, PAGE_NOACCESS) ? ERROR_SUCCESS : GetLastError(); } #endif /* unused for now */ static uint64_t WINAPI stub_GetTickCount64(void) { LARGE_INTEGER Counter, Frequency; return (QueryPerformanceFrequency(&Frequency) && QueryPerformanceCounter(&Counter)) ? Counter.QuadPart * 1000ul / Frequency.QuadPart : 0; } /*----------------------------------------------------------------------------*/ #ifndef MDBX_ALLOY MDBX_NtExtendSection mdbx_NtExtendSection; MDBX_GetFileInformationByHandleEx mdbx_GetFileInformationByHandleEx; MDBX_GetVolumeInformationByHandleW mdbx_GetVolumeInformationByHandleW; MDBX_GetFinalPathNameByHandleW mdbx_GetFinalPathNameByHandleW; MDBX_SetFileInformationByHandle mdbx_SetFileInformationByHandle; MDBX_NtFsControlFile mdbx_NtFsControlFile; MDBX_PrefetchVirtualMemory mdbx_PrefetchVirtualMemory; MDBX_GetTickCount64 mdbx_GetTickCount64; MDBX_RegGetValueA mdbx_RegGetValueA; #if 0 /* LY: unused for now */ MDBX_DiscardVirtualMemory mdbx_DiscardVirtualMemory; MDBX_OfferVirtualMemory mdbx_OfferVirtualMemory; MDBX_ReclaimVirtualMemory mdbx_ReclaimVirtualMemory; #endif /* unused for now */ #endif /* MDBX_ALLOY */ static void mdbx_winnt_import(void) { const HINSTANCE hNtdll = GetModuleHandleA("ntdll.dll"); #define GET_PROC_ADDR(dll, ENTRY) \ mdbx_##ENTRY = (MDBX_##ENTRY)GetProcAddress(dll, #ENTRY) if (GetProcAddress(hNtdll, "wine_get_version")) { assert(mdbx_RunningUnderWine()); } else { GET_PROC_ADDR(hNtdll, NtFsControlFile); GET_PROC_ADDR(hNtdll, NtExtendSection); assert(!mdbx_RunningUnderWine()); } const HINSTANCE hKernel32dll = GetModuleHandleA("kernel32.dll"); GET_PROC_ADDR(hKernel32dll, GetFileInformationByHandleEx); GET_PROC_ADDR(hKernel32dll, GetTickCount64); if (!mdbx_GetTickCount64) mdbx_GetTickCount64 = stub_GetTickCount64; if (!mdbx_RunningUnderWine()) { GET_PROC_ADDR(hKernel32dll, SetFileInformationByHandle); GET_PROC_ADDR(hKernel32dll, GetVolumeInformationByHandleW); GET_PROC_ADDR(hKernel32dll, GetFinalPathNameByHandleW); GET_PROC_ADDR(hKernel32dll, PrefetchVirtualMemory); } const HINSTANCE hAdvapi32dll = GetModuleHandleA("advapi32.dll"); GET_PROC_ADDR(hAdvapi32dll, RegGetValueA); #if 0 /* LY: unused for now */ if (!mdbx_RunningUnderWine()) { GET_PROC_ADDR(hKernel32dll, DiscardVirtualMemory); GET_PROC_ADDR(hKernel32dll, OfferVirtualMemory); GET_PROC_ADDR(hKernel32dll, ReclaimVirtualMemory); } if (!mdbx_DiscardVirtualMemory) mdbx_DiscardVirtualMemory = stub_DiscardVirtualMemory; if (!mdbx_OfferVirtualMemory) mdbx_OfferVirtualMemory = stub_OfferVirtualMemory; if (!mdbx_ReclaimVirtualMemory) mdbx_ReclaimVirtualMemory = stub_ReclaimVirtualMemory; #endif /* unused for now */ #undef GET_PROC_ADDR const MDBX_srwlock_function init = (MDBX_srwlock_function)GetProcAddress(hKernel32dll, "InitializeSRWLock"); if (init != NULL) { mdbx_srwlock_Init = init; mdbx_srwlock_AcquireShared = (MDBX_srwlock_function)GetProcAddress( hKernel32dll, "AcquireSRWLockShared"); mdbx_srwlock_ReleaseShared = (MDBX_srwlock_function)GetProcAddress( hKernel32dll, "ReleaseSRWLockShared"); mdbx_srwlock_AcquireExclusive = (MDBX_srwlock_function)GetProcAddress( hKernel32dll, "AcquireSRWLockExclusive"); mdbx_srwlock_ReleaseExclusive = (MDBX_srwlock_function)GetProcAddress( hKernel32dll, "ReleaseSRWLockExclusive"); } else { mdbx_srwlock_Init = stub_srwlock_Init; mdbx_srwlock_AcquireShared = stub_srwlock_AcquireShared; mdbx_srwlock_ReleaseShared = stub_srwlock_ReleaseShared; mdbx_srwlock_AcquireExclusive = stub_srwlock_AcquireExclusive; mdbx_srwlock_ReleaseExclusive = stub_srwlock_ReleaseExclusive; } } #endif /* Windows LCK-implementation */ /* * Copyright 2015-2021 Leonid Yuriev * and other libmdbx authors: please see AUTHORS file. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted only as authorized by the OpenLDAP * Public License. * * A copy of this license is available in the file LICENSE in the * top-level directory of the distribution or, alternatively, at * . */ #if !(defined(_WIN32) || defined(_WIN64)) /* !Windows LCK-implementation */ #include /*----------------------------------------------------------------------------*/ /* global constructor/destructor */ #if defined(__linux__) || defined(__gnu_linux__) #include #ifndef MDBX_ALLOY uint32_t mdbx_linux_kernel_version; bool mdbx_RunningOnWSL1; #endif /* MDBX_ALLOY */ static __cold uint8_t probe_for_WSL(const char *tag) { const char *const WSL = strstr(tag, "WSL"); if (WSL && WSL[3] >= '2' && WSL[3] <= '9') return WSL[3] - '0'; const char *const wsl = strstr(tag, "wsl"); if (wsl && wsl[3] >= '2' && wsl[3] <= '9') return wsl[3] - '0'; return (WSL || wsl || strcasestr(tag, "Microsoft")) ? 1 : 0; } #endif /* Linux */ static __cold __attribute__((__constructor__)) void mdbx_global_constructor(void) { #if defined(__linux__) || defined(__gnu_linux__) struct utsname buffer; if (uname(&buffer) == 0) { /* "Official" way of detecting WSL1 but not WSL2 * https://github.com/Microsoft/WSL/issues/423#issuecomment-221627364 * * WARNING: False negative detection of WSL1 will result in DATA LOSS! * So, the REQUIREMENTS for this code: * 1. MUST detect WSL1 without false-negatives. * 2. DESIRABLE detect WSL2 but without the risk of violating the first. */ mdbx_RunningOnWSL1 = probe_for_WSL(buffer.version) == 1 || probe_for_WSL(buffer.sysname) == 1 || probe_for_WSL(buffer.release) == 1; int i = 0; char *p = buffer.release; while (*p && i < 4) { if (*p >= '0' && *p <= '9') { long number = strtol(p, &p, 10); if (number > 0) { if (number > 255) number = 255; mdbx_linux_kernel_version += number << (24 - i * 8); } ++i; } else { ++p; } } } #endif /* Linux */ mdbx_rthc_global_init(); } static __cold __attribute__((__destructor__)) void mdbx_global_destructor(void) { mdbx_rthc_global_dtor(); } /*----------------------------------------------------------------------------*/ /* lck */ /* Описание реализации блокировок для POSIX & Linux: * * lck-файл отображается в память, в нём организуется таблица читателей и * размещаются совместно используемые posix-мьютексы (futex). Посредством * этих мьютексов (см struct MDBX_lockinfo) реализуются: * - Блокировка таблицы читателей для регистрации, * т.е. функции mdbx_rdt_lock() и mdbx_rdt_unlock(). * - Блокировка БД для пишущих транзакций, * т.е. функции mdbx_txn_lock() и mdbx_txn_unlock(). * * Остальной функционал реализуется отдельно посредством файловых блокировок: * - Первоначальный захват БД в режиме exclusive/shared и последующий перевод * в операционный режим, функции mdbx_lck_seize() и mdbx_lck_downgrade(). * - Проверка присутствие процессов-читателей, * т.е. функции mdbx_rpid_set(), mdbx_rpid_clear() и mdbx_rpid_check(). * * Для блокировки файлов используется fcntl(F_SETLK), так как: * - lockf() оперирует только эксклюзивной блокировкой и требует * открытия файла в RW-режиме. * - flock() не гарантирует атомарности при смене блокировок * и оперирует только всем файлом целиком. * - Для контроля процессов-читателей используются однобайтовые * range-блокировки lck-файла посредством fcntl(F_SETLK). При этом * в качестве позиции используется pid процесса-читателя. * - Для первоначального захвата и shared/exclusive выполняется блокировка * основного файла БД и при успехе lck-файла. * * ---------------------------------------------------------------------------- * УДЕРЖИВАЕМЫЕ БЛОКИРОВКИ В ЗАВИСИМОСТИ ОТ РЕЖИМА И СОСТОЯНИЯ * * Эксклюзивный режим без lck-файла: * = заблокирован весь dxb-файл посредством F_RDLCK или F_WRLCK, * в зависимости от MDBX_RDONLY. * * Не-операционный режим на время пере-инициализации и разрушении lck-файла: * = F_WRLCK блокировка первого байта lck-файла, другие процессы ждут её * снятия при получении F_RDLCK через F_SETLKW. * - блокировки dxb-файла могут меняться до снятие эксклюзивной блокировки * lck-файла: * + для НЕ-эксклюзивного режима блокировка pid-байта в dxb-файле * посредством F_RDLCK или F_WRLCK, в зависимости от MDBX_RDONLY. * + для ЭКСКЛЮЗИВНОГО режима блокировка pid-байта всего dxb-файла * посредством F_RDLCK или F_WRLCK, в зависимости от MDBX_RDONLY. * * ОПЕРАЦИОННЫЙ режим с lck-файлом: * = F_RDLCK блокировка первого байта lck-файла, другие процессы не могут * получить F_WRLCK и таким образом видят что БД используется. * + F_WRLCK блокировка pid-байта в clk-файле после первой транзакции чтения. * + для НЕ-эксклюзивного режима блокировка pid-байта в dxb-файле * посредством F_RDLCK или F_WRLCK, в зависимости от MDBX_RDONLY. * + для ЭКСКЛЮЗИВНОГО режима блокировка pid-байта всего dxb-файла * посредством F_RDLCK или F_WRLCK, в зависимости от MDBX_RDONLY. */ #if MDBX_USE_OFDLOCKS static int op_setlk, op_setlkw, op_getlk; static void __cold choice_fcntl() { assert(!op_setlk && !op_setlkw && !op_getlk); if ((mdbx_runtime_flags & MDBX_DBG_LEGACY_MULTIOPEN) == 0 #if defined(__linux__) || defined(__gnu_linux__) && mdbx_linux_kernel_version > 0x030f0000 /* OFD locks are available since 3.15, but engages here only for 3.16 and later kernels (i.e. LTS) because of reliability reasons */ #endif /* linux */ ) { op_setlk = F_OFD_SETLK; op_setlkw = F_OFD_SETLKW; op_getlk = F_OFD_GETLK; return; } op_setlk = F_SETLK; op_setlkw = F_SETLKW; op_getlk = F_GETLK; } #else #define op_setlk F_SETLK #define op_setlkw F_SETLKW #define op_getlk F_GETLK #endif /* MDBX_USE_OFDLOCKS */ #ifndef OFF_T_MAX #define OFF_T_MAX \ ((sizeof(off_t) > 4 ? INT64_MAX : INT32_MAX) & ~(size_t)0xffff) #endif static int lck_op(mdbx_filehandle_t fd, int cmd, int lck, off_t offset, off_t len) { mdbx_jitter4testing(true); for (;;) { struct flock lock_op; memset(&lock_op, 0, sizeof(lock_op)); lock_op.l_type = lck; lock_op.l_whence = SEEK_SET; lock_op.l_start = offset; lock_op.l_len = len; int rc = fcntl(fd, cmd, &lock_op); mdbx_jitter4testing(true); if (rc != -1) { if (cmd == op_getlk) { /* Checks reader by pid. Returns: * MDBX_RESULT_TRUE - if pid is live (reader holds a lock). * MDBX_RESULT_FALSE - if pid is dead (a lock could be placed). */ return (lock_op.l_type == F_UNLCK) ? MDBX_RESULT_FALSE : MDBX_RESULT_TRUE; } return MDBX_SUCCESS; } rc = errno; if (rc != EINTR || cmd == op_setlkw) { mdbx_assert(nullptr, MDBX_IS_ERROR(rc)); return rc; } } } MDBX_INTERNAL_FUNC int mdbx_lockfile(mdbx_filehandle_t fd, bool wait) { #if MDBX_USE_OFDLOCKS if (unlikely(op_setlk == 0)) choice_fcntl(); #endif /* MDBX_USE_OFDLOCKS */ return lck_op(fd, wait ? op_setlkw : op_setlk, F_WRLCK, 0, OFF_T_MAX); } MDBX_INTERNAL_FUNC int mdbx_rpid_set(MDBX_env *env) { assert(env->me_lfd != INVALID_HANDLE_VALUE); assert(env->me_pid > 0); if (unlikely(mdbx_getpid() != env->me_pid)) return MDBX_PANIC; return lck_op(env->me_lfd, op_setlk, F_WRLCK, env->me_pid, 1); } MDBX_INTERNAL_FUNC int mdbx_rpid_clear(MDBX_env *env) { assert(env->me_lfd != INVALID_HANDLE_VALUE); assert(env->me_pid > 0); return lck_op(env->me_lfd, op_setlk, F_UNLCK, env->me_pid, 1); } MDBX_INTERNAL_FUNC int mdbx_rpid_check(MDBX_env *env, uint32_t pid) { assert(env->me_lfd != INVALID_HANDLE_VALUE); assert(pid > 0); return lck_op(env->me_lfd, op_getlk, F_WRLCK, pid, 1); } /*---------------------------------------------------------------------------*/ #if MDBX_LOCKING > MDBX_LOCKING_SYSV MDBX_INTERNAL_FUNC int mdbx_ipclock_stub(mdbx_ipclock_t *ipc) { #if MDBX_LOCKING == MDBX_LOCKING_POSIX1988 return sem_init(ipc, false, 1) ? errno : 0; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 return pthread_mutex_init(ipc, nullptr); #else #error "FIXME" #endif } MDBX_INTERNAL_FUNC int mdbx_ipclock_destroy(mdbx_ipclock_t *ipc) { #if MDBX_LOCKING == MDBX_LOCKING_POSIX1988 return sem_destroy(ipc) ? errno : 0; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 return pthread_mutex_destroy(ipc); #else #error "FIXME" #endif } #endif /* MDBX_LOCKING > MDBX_LOCKING_SYSV */ static int check_fstat(MDBX_env *env) { struct stat st; int rc = MDBX_SUCCESS; if (fstat(env->me_lazy_fd, &st)) { rc = errno; mdbx_error("fstat(%s), err %d", "DXB", rc); return rc; } if (!S_ISREG(st.st_mode) || st.st_nlink < 1) { #ifdef EBADFD rc = EBADFD; #else rc = EPERM; #endif mdbx_error("%s %s, err %d", "DXB", (st.st_nlink < 1) ? "file was removed" : "not a regular file", rc); return rc; } if (st.st_size < (off_t)(MDBX_MIN_PAGESIZE * NUM_METAS)) { mdbx_verbose("dxb-file is too short (%u), exclusive-lock needed", (unsigned)st.st_size); rc = MDBX_RESULT_TRUE; } //---------------------------------------------------------------------------- if (fstat(env->me_lfd, &st)) { rc = errno; mdbx_error("fstat(%s), err %d", "LCK", rc); return rc; } if (!S_ISREG(st.st_mode) || st.st_nlink < 1) { #ifdef EBADFD rc = EBADFD; #else rc = EPERM; #endif mdbx_error("%s %s, err %d", "LCK", (st.st_nlink < 1) ? "file was removed" : "not a regular file", rc); return rc; } /* Checking file size for detect the situation when we got the shared lock * immediately after mdbx_lck_destroy(). */ if (st.st_size < (off_t)(sizeof(MDBX_lockinfo) + sizeof(MDBX_reader))) { mdbx_verbose("lck-file is too short (%u), exclusive-lock needed", (unsigned)st.st_size); rc = MDBX_RESULT_TRUE; } return rc; } MDBX_INTERNAL_FUNC int __cold mdbx_lck_seize(MDBX_env *env) { assert(env->me_lazy_fd != INVALID_HANDLE_VALUE); if (unlikely(mdbx_getpid() != env->me_pid)) return MDBX_PANIC; #if MDBX_USE_OFDLOCKS if (unlikely(op_setlk == 0)) choice_fcntl(); #endif /* MDBX_USE_OFDLOCKS */ int rc = MDBX_SUCCESS; #if defined(__linux__) || defined(__gnu_linux__) if (unlikely(mdbx_RunningOnWSL1)) { rc = ENOLCK /* No record locks available */; mdbx_error("%s, err %u", "WSL1 (Windows Subsystem for Linux) is mad and trouble-full, " "injecting failure to avoid data loss", rc); return rc; } #endif /* Linux */ if (env->me_lfd == INVALID_HANDLE_VALUE) { /* LY: without-lck mode (e.g. exclusive or on read-only filesystem) */ rc = lck_op(env->me_lazy_fd, op_setlk, (env->me_flags & MDBX_RDONLY) ? F_RDLCK : F_WRLCK, 0, OFF_T_MAX); if (rc != MDBX_SUCCESS) { mdbx_error("%s, err %u", "without-lck", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } return MDBX_RESULT_TRUE /* Done: return with exclusive locking. */; } #if defined(_POSIX_PRIORITY_SCHEDULING) && _POSIX_PRIORITY_SCHEDULING > 0 sched_yield(); #endif retry: if (rc == MDBX_RESULT_TRUE) { rc = lck_op(env->me_lfd, op_setlk, F_UNLCK, 0, 1); if (rc != MDBX_SUCCESS) { mdbx_error("%s, err %u", "unlock-before-retry", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } } /* Firstly try to get exclusive locking. */ rc = lck_op(env->me_lfd, op_setlk, F_WRLCK, 0, 1); if (rc == MDBX_SUCCESS) { rc = check_fstat(env); if (MDBX_IS_ERROR(rc)) return rc; continue_dxb_exclusive: rc = lck_op(env->me_lazy_fd, op_setlk, (env->me_flags & MDBX_RDONLY) ? F_RDLCK : F_WRLCK, 0, OFF_T_MAX); if (rc == MDBX_SUCCESS) return MDBX_RESULT_TRUE /* Done: return with exclusive locking. */; int err = check_fstat(env); if (MDBX_IS_ERROR(err)) return err; /* the cause may be a collision with POSIX's file-lock recovery. */ if (!(rc == EAGAIN || rc == EACCES || rc == EBUSY || rc == EWOULDBLOCK || rc == EDEADLK)) { mdbx_error("%s, err %u", "dxb-exclusive", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } /* Fallback to lck-shared */ } else if (!(rc == EAGAIN || rc == EACCES || rc == EBUSY || rc == EWOULDBLOCK || rc == EDEADLK)) { mdbx_error("%s, err %u", "try-exclusive", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } /* Here could be one of two: * - mdbx_lck_destroy() from the another process was hold the lock * during a destruction. * - either mdbx_lck_seize() from the another process was got the exclusive * lock and doing initialization. * For distinguish these cases will use size of the lck-file later. */ /* Wait for lck-shared now. */ /* Here may be await during transient processes, for instance until another * competing process doesn't call lck_downgrade(). */ rc = lck_op(env->me_lfd, op_setlkw, F_RDLCK, 0, 1); if (rc != MDBX_SUCCESS) { mdbx_error("%s, err %u", "try-shared", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } rc = check_fstat(env); if (rc == MDBX_RESULT_TRUE) goto retry; if (rc != MDBX_SUCCESS) { mdbx_error("%s, err %u", "lck_fstat", rc); return rc; } /* got shared, retry exclusive */ rc = lck_op(env->me_lfd, op_setlk, F_WRLCK, 0, 1); if (rc == MDBX_SUCCESS) goto continue_dxb_exclusive; if (!(rc == EAGAIN || rc == EACCES || rc == EBUSY || rc == EWOULDBLOCK || rc == EDEADLK)) { mdbx_error("%s, err %u", "try-exclusive", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } /* Lock against another process operating in without-lck or exclusive mode. */ rc = lck_op(env->me_lazy_fd, op_setlk, (env->me_flags & MDBX_RDONLY) ? F_RDLCK : F_WRLCK, env->me_pid, 1); if (rc != MDBX_SUCCESS) { mdbx_error("%s, err %u", "lock-against-without-lck", rc); mdbx_assert(env, MDBX_IS_ERROR(rc)); return rc; } /* Done: return with shared locking. */ return MDBX_RESULT_FALSE; } MDBX_INTERNAL_FUNC int mdbx_lck_downgrade(MDBX_env *env) { assert(env->me_lfd != INVALID_HANDLE_VALUE); if (unlikely(mdbx_getpid() != env->me_pid)) return MDBX_PANIC; int rc = MDBX_SUCCESS; if ((env->me_flags & MDBX_EXCLUSIVE) == 0) { rc = lck_op(env->me_lazy_fd, op_setlk, F_UNLCK, 0, env->me_pid); if (rc == MDBX_SUCCESS) rc = lck_op(env->me_lazy_fd, op_setlk, F_UNLCK, env->me_pid + 1, OFF_T_MAX - env->me_pid - 1); } if (rc == MDBX_SUCCESS) rc = lck_op(env->me_lfd, op_setlk, F_RDLCK, 0, 1); if (unlikely(rc != 0)) { mdbx_error("%s, err %u", "lck", rc); assert(MDBX_IS_ERROR(rc)); } return rc; } MDBX_INTERNAL_FUNC int __cold mdbx_lck_destroy(MDBX_env *env, MDBX_env *inprocess_neighbor) { if (unlikely(mdbx_getpid() != env->me_pid)) return MDBX_PANIC; int rc = MDBX_SUCCESS; struct stat lck_info; if (env->me_lfd != INVALID_HANDLE_VALUE && !inprocess_neighbor && env->me_lck && /* try get exclusive access */ lck_op(env->me_lfd, op_setlk, F_WRLCK, 0, OFF_T_MAX) == 0 && /* if LCK was not removed */ fstat(env->me_lfd, &lck_info) == 0 && lck_info.st_nlink > 0 && lck_op(env->me_lazy_fd, op_setlk, (env->me_flags & MDBX_RDONLY) ? F_RDLCK : F_WRLCK, 0, OFF_T_MAX) == 0) { mdbx_verbose("%p got exclusive, drown locks", (void *)env); #if MDBX_LOCKING == MDBX_LOCKING_SYSV if (env->me_sysv_ipc.semid != -1) rc = semctl(env->me_sysv_ipc.semid, 2, IPC_RMID) ? errno : 0; #else rc = mdbx_ipclock_destroy(&env->me_lck->mti_rlock); if (rc == 0) rc = mdbx_ipclock_destroy(&env->me_lck->mti_wlock); #endif /* MDBX_LOCKING */ mdbx_assert(env, rc == 0); if (rc == 0) { const bool synced = env->me_lck_mmap.lck->mti_unsynced_pages.weak == 0; mdbx_munmap(&env->me_lck_mmap); if (synced) rc = ftruncate(env->me_lfd, 0) ? errno : 0; } mdbx_jitter4testing(false); } /* 1) POSIX's fcntl() locks (i.e. when op_setlk == F_SETLK) should be restored * after file was closed. * * 2) File locks would be released (by kernel) while the file-descriptors will * be closed. But to avoid false-positive EACCESS and EDEADLK from the kernel, * locks should be released here explicitly with properly order. */ /* close dxb and restore lock */ if (env->me_dsync_fd != INVALID_HANDLE_VALUE) { if (unlikely(close(env->me_dsync_fd) != 0) && rc == MDBX_SUCCESS) rc = errno; env->me_dsync_fd = INVALID_HANDLE_VALUE; } if (env->me_lazy_fd != INVALID_HANDLE_VALUE) { if (unlikely(close(env->me_lazy_fd) != 0) && rc == MDBX_SUCCESS) rc = errno; env->me_lazy_fd = INVALID_HANDLE_VALUE; if (op_setlk == F_SETLK && inprocess_neighbor && rc == MDBX_SUCCESS) { /* restore file-lock */ rc = lck_op( inprocess_neighbor->me_lazy_fd, F_SETLKW, (inprocess_neighbor->me_flags & MDBX_RDONLY) ? F_RDLCK : F_WRLCK, (inprocess_neighbor->me_flags & MDBX_EXCLUSIVE) ? 0 : inprocess_neighbor->me_pid, (inprocess_neighbor->me_flags & MDBX_EXCLUSIVE) ? OFF_T_MAX : 1); } } /* close clk and restore locks */ if (env->me_lfd != INVALID_HANDLE_VALUE) { if (unlikely(close(env->me_lfd) != 0) && rc == MDBX_SUCCESS) rc = errno; env->me_lfd = INVALID_HANDLE_VALUE; if (op_setlk == F_SETLK && inprocess_neighbor && rc == MDBX_SUCCESS) { /* restore file-locks */ rc = lck_op(inprocess_neighbor->me_lfd, F_SETLKW, F_RDLCK, 0, 1); if (rc == MDBX_SUCCESS && inprocess_neighbor->me_live_reader) rc = mdbx_rpid_set(inprocess_neighbor); } } if (inprocess_neighbor && rc != MDBX_SUCCESS) inprocess_neighbor->me_flags |= MDBX_FATAL_ERROR; return rc; } /*---------------------------------------------------------------------------*/ MDBX_INTERNAL_FUNC int __cold mdbx_lck_init(MDBX_env *env, MDBX_env *inprocess_neighbor, int global_uniqueness_flag) { #if MDBX_LOCKING == MDBX_LOCKING_SYSV int semid = -1; /* don't initialize semaphores twice */ (void)inprocess_neighbor; if (global_uniqueness_flag == MDBX_RESULT_TRUE) { struct stat st; if (fstat(env->me_lazy_fd, &st)) return errno; sysv_retry_create: semid = semget(env->me_sysv_ipc.key, 2, IPC_CREAT | IPC_EXCL | (st.st_mode & (S_IRWXU | S_IRWXG | S_IRWXO))); if (unlikely(semid == -1)) { int err = errno; if (err != EEXIST) return err; /* remove and re-create semaphore set */ semid = semget(env->me_sysv_ipc.key, 2, 0); if (semid == -1) { err = errno; if (err != ENOENT) return err; goto sysv_retry_create; } if (semctl(semid, 2, IPC_RMID)) { err = errno; if (err != EIDRM) return err; } goto sysv_retry_create; } unsigned short val_array[2] = {1, 1}; if (semctl(semid, 2, SETALL, val_array)) return errno; } else { semid = semget(env->me_sysv_ipc.key, 2, 0); if (semid == -1) return errno; /* check read & write access */ struct semid_ds data[2]; if (semctl(semid, 2, IPC_STAT, data) || semctl(semid, 2, IPC_SET, data)) return errno; } env->me_sysv_ipc.semid = semid; return MDBX_SUCCESS; #elif MDBX_LOCKING == MDBX_LOCKING_FUTEX (void)inprocess_neighbor; if (global_uniqueness_flag != MDBX_RESULT_TRUE) return MDBX_SUCCESS; #error "FIXME: Not implemented" #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988 /* don't initialize semaphores twice */ (void)inprocess_neighbor; if (global_uniqueness_flag == MDBX_RESULT_TRUE) { if (sem_init(&env->me_lck->mti_rlock, true, 1)) return errno; if (sem_init(&env->me_lck->mti_wlock, true, 1)) return errno; } return MDBX_SUCCESS; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 if (inprocess_neighbor) return MDBX_SUCCESS /* don't need any initialization for mutexes if LCK already opened/used inside current process */ ; /* FIXME: Unfortunately, there is no other reliable way but to long testing * on each platform. On the other hand, behavior like FreeBSD is incorrect * and we can expect it to be rare. Moreover, even on FreeBSD without * additional in-process initialization, the probability of an problem * occurring is vanishingly small, and the symptom is a return of EINVAL * while locking a mutex. In other words, in the worst case, the problem * results in an EINVAL error at the start of the transaction, but NOT data * loss, nor database corruption, nor other fatal troubles. Thus, the code * below I am inclined to think the workaround for erroneous platforms (like * FreeBSD), rather than a defect of libmdbx. */ #if defined(__FreeBSD__) /* seems that shared mutexes on FreeBSD required in-process initialization */ (void)global_uniqueness_flag; #else /* shared mutexes on many other platforms (including Darwin and Linux's * futexes) doesn't need any addition in-process initialization */ if (global_uniqueness_flag != MDBX_RESULT_TRUE) return MDBX_SUCCESS; #endif pthread_mutexattr_t ma; int rc = pthread_mutexattr_init(&ma); if (rc) return rc; rc = pthread_mutexattr_setpshared(&ma, PTHREAD_PROCESS_SHARED); if (rc) goto bailout; #if MDBX_LOCKING == MDBX_LOCKING_POSIX2008 #if defined(PTHREAD_MUTEX_ROBUST) || defined(pthread_mutexattr_setrobust) rc = pthread_mutexattr_setrobust(&ma, PTHREAD_MUTEX_ROBUST); #elif defined(PTHREAD_MUTEX_ROBUST_NP) || \ defined(pthread_mutexattr_setrobust_np) rc = pthread_mutexattr_setrobust_np(&ma, PTHREAD_MUTEX_ROBUST_NP); #elif _POSIX_THREAD_PROCESS_SHARED < 200809L rc = pthread_mutexattr_setrobust_np(&ma, PTHREAD_MUTEX_ROBUST_NP); #else rc = pthread_mutexattr_setrobust(&ma, PTHREAD_MUTEX_ROBUST); #endif if (rc) goto bailout; #endif /* MDBX_LOCKING == MDBX_LOCKING_POSIX2008 */ #if defined(_POSIX_THREAD_PRIO_INHERIT) && _POSIX_THREAD_PRIO_INHERIT >= 0 && \ !defined(MDBX_SAFE4QEMU) rc = pthread_mutexattr_setprotocol(&ma, PTHREAD_PRIO_INHERIT); if (rc == ENOTSUP) rc = pthread_mutexattr_setprotocol(&ma, PTHREAD_PRIO_NONE); if (rc && rc != ENOTSUP) goto bailout; #endif /* PTHREAD_PRIO_INHERIT */ rc = pthread_mutexattr_settype(&ma, PTHREAD_MUTEX_ERRORCHECK); if (rc && rc != ENOTSUP) goto bailout; rc = pthread_mutex_init(&env->me_lck->mti_rlock, &ma); if (rc) goto bailout; rc = pthread_mutex_init(&env->me_lck->mti_wlock, &ma); bailout: pthread_mutexattr_destroy(&ma); return rc; #else #error "FIXME" #endif /* MDBX_LOCKING > 0 */ } static int __cold mdbx_ipclock_failed(MDBX_env *env, mdbx_ipclock_t *ipc, const int err) { int rc = err; #if MDBX_LOCKING == MDBX_LOCKING_POSIX2008 || MDBX_LOCKING == MDBX_LOCKING_SYSV if (err == EOWNERDEAD) { /* We own the mutex. Clean up after dead previous owner. */ const bool rlocked = (env->me_lck && ipc == &env->me_lck->mti_rlock); rc = MDBX_SUCCESS; if (!rlocked) { if (unlikely(env->me_txn)) { /* env is hosed if the dead thread was ours */ env->me_flags |= MDBX_FATAL_ERROR; env->me_txn = NULL; rc = MDBX_PANIC; } } mdbx_warning("%clock owner died, %s", (rlocked ? 'r' : 'w'), (rc ? "this process' env is hosed" : "recovering")); int check_rc = mdbx_cleanup_dead_readers(env, rlocked, NULL); check_rc = (check_rc == MDBX_SUCCESS) ? MDBX_RESULT_TRUE : check_rc; #if MDBX_LOCKING == MDBX_LOCKING_SYSV rc = (rc == MDBX_SUCCESS) ? check_rc : rc; #else #if defined(PTHREAD_MUTEX_ROBUST) || defined(pthread_mutex_consistent) int mreco_rc = pthread_mutex_consistent(ipc); #elif defined(PTHREAD_MUTEX_ROBUST_NP) || defined(pthread_mutex_consistent_np) int mreco_rc = pthread_mutex_consistent_np(ipc); #elif _POSIX_THREAD_PROCESS_SHARED < 200809L int mreco_rc = pthread_mutex_consistent_np(ipc); #else int mreco_rc = pthread_mutex_consistent(ipc); #endif check_rc = (mreco_rc == 0) ? check_rc : mreco_rc; if (unlikely(mreco_rc)) mdbx_error("lock recovery failed, %s", mdbx_strerror(mreco_rc)); rc = (rc == MDBX_SUCCESS) ? check_rc : rc; if (MDBX_IS_ERROR(rc)) pthread_mutex_unlock(ipc); #endif /* MDBX_LOCKING == MDBX_LOCKING_POSIX2008 */ return rc; } #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 (void)ipc; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988 (void)ipc; #elif MDBX_LOCKING == MDBX_LOCKING_FUTEX #ifdef _MSC_VER #pragma message("warning: TODO") #else #warning "TODO" #endif (void)ipc; #else #error "FIXME" #endif /* MDBX_LOCKING */ mdbx_error("mutex (un)lock failed, %s", mdbx_strerror(err)); if (rc != EDEADLK) env->me_flags |= MDBX_FATAL_ERROR; return rc; } static int mdbx_ipclock_lock(MDBX_env *env, mdbx_ipclock_t *ipc, const bool dont_wait) { #if MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 int rc = dont_wait ? pthread_mutex_trylock(ipc) : pthread_mutex_lock(ipc); rc = (rc == EBUSY && dont_wait) ? MDBX_BUSY : rc; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988 int rc = MDBX_SUCCESS; if (dont_wait) { if (sem_trywait(ipc)) { rc = errno; if (rc == EAGAIN) rc = MDBX_BUSY; } } else if (sem_wait(ipc)) rc = errno; #elif MDBX_LOCKING == MDBX_LOCKING_SYSV struct sembuf op = {.sem_num = (ipc != env->me_wlock), .sem_op = -1, .sem_flg = dont_wait ? IPC_NOWAIT | SEM_UNDO : SEM_UNDO}; int rc; if (semop(env->me_sysv_ipc.semid, &op, 1)) { rc = errno; if (dont_wait && rc == EAGAIN) rc = MDBX_BUSY; } else { rc = *ipc ? EOWNERDEAD : MDBX_SUCCESS; *ipc = env->me_pid; } #else #error "FIXME" #endif /* MDBX_LOCKING */ if (unlikely(rc != MDBX_SUCCESS && rc != MDBX_BUSY)) rc = mdbx_ipclock_failed(env, ipc, rc); return rc; } static int mdbx_ipclock_unlock(MDBX_env *env, mdbx_ipclock_t *ipc) { #if MDBX_LOCKING == MDBX_LOCKING_POSIX2001 || \ MDBX_LOCKING == MDBX_LOCKING_POSIX2008 int rc = pthread_mutex_unlock(ipc); (void)env; #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988 int rc = sem_post(ipc) ? errno : MDBX_SUCCESS; (void)env; #elif MDBX_LOCKING == MDBX_LOCKING_SYSV if (unlikely(*ipc != (pid_t)env->me_pid)) return EPERM; *ipc = 0; struct sembuf op = { .sem_num = (ipc != env->me_wlock), .sem_op = 1, .sem_flg = SEM_UNDO}; int rc = semop(env->me_sysv_ipc.semid, &op, 1) ? errno : MDBX_SUCCESS; #else #error "FIXME" #endif /* MDBX_LOCKING */ return rc; } MDBX_INTERNAL_FUNC int mdbx_rdt_lock(MDBX_env *env) { mdbx_trace("%s", ">>"); mdbx_jitter4testing(true); int rc = mdbx_ipclock_lock(env, &env->me_lck->mti_rlock, false); mdbx_trace("<< rc %d", rc); return rc; } MDBX_INTERNAL_FUNC void mdbx_rdt_unlock(MDBX_env *env) { mdbx_trace("%s", ">>"); int rc = mdbx_ipclock_unlock(env, &env->me_lck->mti_rlock); mdbx_trace("<< rc %d", rc); if (unlikely(rc != MDBX_SUCCESS)) mdbx_panic("%s() failed: err %d\n", __func__, rc); mdbx_jitter4testing(true); } int mdbx_txn_lock(MDBX_env *env, bool dont_wait) { mdbx_trace("%swait %s", dont_wait ? "dont-" : "", ">>"); mdbx_jitter4testing(true); int rc = mdbx_ipclock_lock(env, env->me_wlock, dont_wait); mdbx_trace("<< rc %d", rc); return MDBX_IS_ERROR(rc) ? rc : MDBX_SUCCESS; } void mdbx_txn_unlock(MDBX_env *env) { mdbx_trace("%s", ">>"); int rc = mdbx_ipclock_unlock(env, env->me_wlock); mdbx_trace("<< rc %d", rc); if (unlikely(rc != MDBX_SUCCESS)) mdbx_panic("%s() failed: err %d\n", __func__, rc); mdbx_jitter4testing(true); } #else #ifdef _MSC_VER #pragma warning(disable : 4206) /* nonstandard extension used: translation \ unit is empty */ #endif /* _MSC_VER (warnings) */ #endif /* !Windows LCK-implementation */