kiper220/mimalloc
1
0
Fork 0
mirror of https://github.com/microsoft/mimalloc.git synced 2025-05-03 22:19:32 +03:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'dev' into random-context-macos

Browse source
This commit is contained in:
Daan 2021-12-15 16:57:25 -08:00 • committed by GitHub
commit 2a4ad02d27
WARNING! Although there is a key with this ID in the database it does not verify this commit! This commit is SUSPICIOUS.
GPG key ID: 4AEE18F83AFDEB23
33 changed files with 1210 additions and 659 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

97
CMakeLists.txt
View file

@ -12,7 +12,7 @@ option(MI_XMALLOC "Enable abort() call on memory allocation failure by
option(MI_SHOW_ERRORS "Show error and warning messages by default (only enabled by default in DEBUG mode)" OFF)
option(MI_USE_CXX "Use the C++ compiler to compile the library (instead of the C compiler)" OFF)
option(MI_SEE_ASM "Generate assembly files" OFF)
option(MI_INTERPOSE "Use interpose to override standard malloc on macOS" OFF)
option(MI_OSX_INTERPOSE "Use interpose to override standard malloc on macOS" ON)
option(MI_OSX_ZONE "Use malloc zone to override standard malloc on macOS" ON)
option(MI_LOCAL_DYNAMIC_TLS "Use slightly slower, dlopen-compatible TLS mechanism (Unix)" OFF)
option(MI_BUILD_SHARED "Build shared library" ON)
@ -23,7 +23,9 @@ option(MI_DEBUG_TSAN "Build with thread sanitizer (needs clang)" OFF)
option(MI_DEBUG_UBSAN "Build with undefined-behavior sanitizer (needs clang++)" OFF)
option(MI_CHECK_FULL "Use full internal invariant checking in DEBUG mode (deprecated, use MI_DEBUG_FULL instead)" OFF)
option(MI_INSTALL_TOPLEVEL "Install directly into $CMAKE_INSTALL_PREFIX instead of PREFIX/lib/mimalloc-version" OFF)
option(MI_USE_LIBATOMIC "Explicitly link with -latomic (on older systems)" OFF)
include(GNUInstallDirs)
include("cmake/mimalloc-config-version.cmake")
set(mi_sources
@ -42,10 +44,12 @@ set(mi_sources
src/options.c
src/init.c)
# -----------------------------------------------------------------------------
# Converience: set default build type depending on the build directory
# Convenience: set default build type depending on the build directory
# -----------------------------------------------------------------------------
message(STATUS "")
if (NOT CMAKE_BUILD_TYPE)
if ("${CMAKE_BINARY_DIR}" MATCHES ".*(D|d)ebug$" OR MI_DEBUG_FULL)
message(STATUS "No build type selected, default to: Debug")
@ -61,6 +65,7 @@ if("${CMAKE_BINARY_DIR}" MATCHES ".*(S|s)ecure$")
set(MI_SECURE "ON")
endif()
# -----------------------------------------------------------------------------
# Process options
# -----------------------------------------------------------------------------
@ -77,12 +82,20 @@ if(MI_OVERRIDE)
message(STATUS " Use malloc zone to override malloc (MI_OSX_ZONE=ON)")
list(APPEND mi_sources src/alloc-override-osx.c)
list(APPEND mi_defines MI_OSX_ZONE=1)
if (NOT MI_OSX_INTERPOSE)
message(STATUS " WARNING: zone overriding usually also needs interpose (use -DMI_OSX_INTERPOSE=ON)")
endif()
endif()
if(MI_INTERPOSE)
if(MI_OSX_INTERPOSE)
# use interpose on macOS
message(STATUS " Use interpose to override malloc (MI_INTERPOSE=ON)")
message(STATUS " WARNING: interpose does not seem to work reliably on the M1; use -DMI_OSX_ZONE=ON instead")
list(APPEND mi_defines MI_INTERPOSE)
message(STATUS " Use interpose to override malloc (MI_OSX_INTERPOSE=ON)")
list(APPEND mi_defines MI_OSX_INTERPOSE=1)
if (NOT MI_OSX_ZONE)
message(STATUS " WARNING: interpose usually also needs zone overriding (use -DMI_OSX_INTERPOSE=ON)")
endif()
endif()
if((NOT MI_USE_CXX) AND MI_OVERRIDE)
message(STATUS " WARNING: if overriding C++ new/delete, it is best to build mimalloc with a C++ compiler (use -DMI_USE_CXX=ON)")
endif()
endif()
endif()
@ -166,9 +179,9 @@ endif()
# Compiler flags
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU")
list(APPEND mi_cflags -Wall -Wextra -Wno-unknown-pragmas -fvisibility=hidden)
if(CMAKE_C_COMPILER_ID MATCHES "GNU")
list(APPEND mi_cflags -Wno-invalid-memory-model)
endif()
if(NOT MI_USE_CXX)
list(APPEND mi_cflags -Wstrict-prototypes)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang")
list(APPEND mi_cflags -Wpedantic -Wno-static-in-inline)
endif()
@ -184,6 +197,9 @@ if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU|Intel" AND NOT CMAKE_SYSTEM
else()
list(APPEND mi_cflags -ftls-model=initial-exec)
endif()
if(MI_OVERRIDE)
list(APPEND mi_cflags -fno-builtin-malloc)
endif()
endif()
if (MSVC AND MSVC_VERSION GREATER_EQUAL 1914)
@ -194,12 +210,17 @@ endif()
if(WIN32)
list(APPEND mi_libraries psapi shell32 user32 advapi32 bcrypt)
else()
if(NOT ${CMAKE_C_COMPILER} MATCHES "android")
list(APPEND mi_libraries pthread)
find_library(LIBRT rt)
if(LIBRT)
list(APPEND mi_libraries ${LIBRT})
endif()
find_library(MI_LIBPTHREAD pthread)
if (MI_LIBPTHREAD)
list(APPEND mi_libraries ${MI_LIBPTHREAD})
endif()
find_library(MI_LIBRT rt)
if(MI_LIBRT)
list(APPEND mi_libraries ${MI_LIBRT})
endif()
find_library(MI_LIBATOMIC atomic)
if (MI_LIBATOMIC OR MI_USE_LIBATOMIC)
list(APPEND mi_libraries atomic)
endif()
endif()
@ -207,14 +228,19 @@ endif()
# Install and output names
# -----------------------------------------------------------------------------
# dynamic/shared library and symlinks always go to /usr/local/lib equivalent
set(mi_install_libdir "${CMAKE_INSTALL_LIBDIR}")
# static libraries and object files, includes, and cmake config files
# are either installed at top level, or use versioned directories for side-by-side installation (default)
if (MI_INSTALL_TOPLEVEL)
set(mi_install_libdir "lib")
set(mi_install_incdir "include")
set(mi_install_cmakedir "cmake")
set(mi_install_objdir "${CMAKE_INSTALL_LIBDIR}")
set(mi_install_incdir "${CMAKE_INSTALL_INCLUDEDIR}")
set(mi_install_cmakedir "${CMAKE_INSTALL_LIBDIR}/cmake/mimalloc")
else()
set(mi_install_libdir "lib/mimalloc-${mi_version}")
set(mi_install_incdir "include/mimalloc-${mi_version}")
set(mi_install_cmakedir "share/mimalloc-${mi_version}/cmake")
set(mi_install_objdir "${CMAKE_INSTALL_LIBDIR}/mimalloc-${mi_version}") # for static library and object files
set(mi_install_incdir "${CMAKE_INSTALL_INCLUDEDIR}/mimalloc-${mi_version}") # for includes
set(mi_install_cmakedir "${CMAKE_INSTALL_LIBDIR}/cmake/mimalloc-${mi_version}") # for cmake package info
endif()
if(MI_SECURE)
@ -224,7 +250,7 @@ else()
endif()
string(TOLOWER "${CMAKE_BUILD_TYPE}" CMAKE_BUILD_TYPE_LC)
if(NOT(CMAKE_BUILD_TYPE_LC MATCHES "^(release|relwithdebinfo|minsizerel)$"))
if(NOT(CMAKE_BUILD_TYPE_LC MATCHES "^(release|relwithdebinfo|minsizerel|none)$"))
set(mi_basename "${mi_basename}-${CMAKE_BUILD_TYPE_LC}") #append build type (e.g. -debug) if not a release version
endif()
if(MI_BUILD_SHARED)
@ -242,13 +268,16 @@ endif()
message(STATUS "")
message(STATUS "Library base name: ${mi_basename}")
message(STATUS "Version : ${mi_version}")
message(STATUS "Build type : ${CMAKE_BUILD_TYPE_LC}")
if(MI_USE_CXX)
message(STATUS "Compiler : ${CMAKE_CXX_COMPILER}")
message(STATUS "C++ Compiler : ${CMAKE_CXX_COMPILER}")
else()
message(STATUS "Compiler : ${CMAKE_C_COMPILER}")
message(STATUS "C Compiler : ${CMAKE_C_COMPILER}")
endif()
message(STATUS "Version : ${mi_version}")
message(STATUS "Compiler flags : ${mi_cflags}")
message(STATUS "Compiler defines : ${mi_defines}")
message(STATUS "Link libraries : ${mi_libraries}")
message(STATUS "Build targets : ${mi_build_targets}")
message(STATUS "")
@ -259,7 +288,7 @@ message(STATUS "")
# shared library
if(MI_BUILD_SHARED)
add_library(mimalloc SHARED ${mi_sources})
set_target_properties(mimalloc PROPERTIES VERSION ${mi_version} OUTPUT_NAME ${mi_basename} )
set_target_properties(mimalloc PROPERTIES VERSION ${mi_version} SOVERSION ${mi_version_major} OUTPUT_NAME ${mi_basename} )
target_compile_definitions(mimalloc PRIVATE ${mi_defines} MI_SHARED_LIB MI_SHARED_LIB_EXPORT)
target_compile_options(mimalloc PRIVATE ${mi_cflags})
target_link_libraries(mimalloc PUBLIC ${mi_libraries})
@ -305,7 +334,7 @@ if (MI_BUILD_STATIC)
set_target_properties(mimalloc-static PROPERTIES OUTPUT_NAME ${mi_basename})
endif()
install(TARGETS mimalloc-static EXPORT mimalloc DESTINATION ${mi_install_libdir} LIBRARY)
install(TARGETS mimalloc-static EXPORT mimalloc DESTINATION ${mi_install_objdir} LIBRARY)
endif()
# install include files
@ -315,16 +344,6 @@ install(FILES include/mimalloc-new-delete.h DESTINATION ${mi_install_incdir})
install(FILES cmake/mimalloc-config.cmake DESTINATION ${mi_install_cmakedir})
install(FILES cmake/mimalloc-config-version.cmake DESTINATION ${mi_install_cmakedir})
if(NOT WIN32 AND MI_BUILD_SHARED AND NOT MI_INSTALL_TOPLEVEL)
# install a symlink in the /usr/local/lib to the versioned library
# note: use delayed prefix expansion as \${CMAKE_INSTALL_PREFIX}
set(mi_symlink "${CMAKE_SHARED_MODULE_PREFIX}${mi_basename}${CMAKE_SHARED_LIBRARY_SUFFIX}")
set(mi_soname "mimalloc-${mi_version}/${mi_symlink}.${mi_version}")
install(CODE "execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink ${mi_soname} ${mi_symlink} WORKING_DIRECTORY \${CMAKE_INSTALL_PREFIX}/lib)")
install(CODE "MESSAGE(\"-- Symbolic link: \${CMAKE_INSTALL_PREFIX}/lib/${mi_symlink} -> ${mi_soname}\")")
install(CODE "execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink ${mi_soname} ${mi_symlink}.${mi_version} WORKING_DIRECTORY \${CMAKE_INSTALL_PREFIX}/lib)")
install(CODE "MESSAGE(\"-- Symbolic link: \${CMAKE_INSTALL_PREFIX}/lib/${mi_symlink}.${mi_version} -> ${mi_soname}\")")
endif()
# single object file for more predictable static overriding
if (MI_BUILD_OBJECT)
@ -343,7 +362,7 @@ if (MI_BUILD_OBJECT)
# the FILES expression can also be: $<TARGET_OBJECTS:mimalloc-obj>
# but that fails cmake versions less than 3.10 so we leave it as is for now
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles/mimalloc-obj.dir/src/static.c${CMAKE_C_OUTPUT_EXTENSION}
DESTINATION ${mi_install_libdir}
DESTINATION ${mi_install_objdir}
RENAME ${mi_basename}${CMAKE_C_OUTPUT_EXTENSION} )
endif()
@ -356,7 +375,7 @@ if (MI_BUILD_TESTS)
target_compile_definitions(mimalloc-test-api PRIVATE ${mi_defines})
target_compile_options(mimalloc-test-api PRIVATE ${mi_cflags})
target_include_directories(mimalloc-test-api PRIVATE include)
target_link_libraries(mimalloc-test-api PRIVATE mimalloc-static ${mi_libraries})
target_link_libraries(mimalloc-test-api PRIVATE mimalloc ${mi_libraries})
add_executable(mimalloc-test-stress test/test-stress.c)
target_compile_definitions(mimalloc-test-stress PRIVATE ${mi_defines})

2
azure-pipelines.yml
View file

@ -115,7 +115,7 @@ jobs:
displayName: macOS
pool:
vmImage:
macOS-10.15
macOS-latest
strategy:
matrix:
Debug:

1
cmake/mimalloc-config-version.cmake
View file

@ -1,5 +1,6 @@
set(mi_version_major 1)
set(mi_version_minor 7)
set(mi_version_patch 4)
set(mi_version ${mi_version_major}.${mi_version_minor})
set(PACKAGE_VERSION ${mi_version})

19
cmake/mimalloc-config.cmake
View file

@ -1,11 +1,14 @@
include(${CMAKE_CURRENT_LIST_DIR}/mimalloc.cmake)
get_filename_component(MIMALLOC_SHARE_DIR "${CMAKE_CURRENT_LIST_DIR}" PATH) # one up from the cmake dir, e.g. /usr/local/share/mimalloc-2.0
if (MIMALLOC_SHARE_DIR MATCHES "/share/")
string(REPLACE "/share/" "/lib/" MIMALLOC_LIBRARY_DIR ${MIMALLOC_SHARE_DIR})
string(REPLACE "/share/" "/include/" MIMALLOC_INCLUDE_DIR ${MIMALLOC_SHARE_DIR})
else()
# if MI_INSTALL_TOPLEVEL==ON
set(MIMALLOC_LIBRARY_DIR "${MIMALLOC_SHARE_DIR}/lib")
set(MIMALLOC_INCLUDE_DIR "${MIMALLOC_SHARE_DIR}/include")
get_filename_component(MIMALLOC_CMAKE_DIR "${CMAKE_CURRENT_LIST_DIR}" PATH) # one up from the cmake dir, e.g. /usr/local/lib/cmake/mimalloc-2.0
get_filename_component(MIMALLOC_VERSION_DIR "${CMAKE_CURRENT_LIST_DIR}" NAME)
string(REPLACE "/lib/cmake" "/lib" MIMALLOC_LIBRARY_DIR "${MIMALLOC_CMAKE_DIR}")
if("${MIMALLOC_VERSION_DIR}" EQUAL "mimalloc")
# top level install
string(REPLACE "/lib/cmake" "/include" MIMALLOC_INCLUDE_DIR "${MIMALLOC_CMAKE_DIR}")
set(MIMALLOC_OBJECT_DIR "${MIMALLOC_LIBRARY_DIR}")
else()
# versioned
string(REPLACE "/lib/cmake/" "/include/" MIMALLOC_INCLUDE_DIR "${CMAKE_CURRENT_LIST_DIR}")
string(REPLACE "/lib/cmake/" "/lib/" MIMALLOC_OBJECT_DIR "${CMAKE_CURRENT_LIST_DIR}")
endif()
set(MIMALLOC_TARGET_DIR "${MIMALLOC_LIBRARY_DIR}") # legacy

3
doc/mimalloc-doc.h
View file

@ -782,6 +782,7 @@ typedef enum mi_option_e {
mi_option_eager_region_commit, ///< Eagerly commit large (256MiB) memory regions (enabled by default, except on Windows)
mi_option_large_os_pages, ///< Use large OS pages (2MiB in size) if possible
mi_option_reserve_huge_os_pages, ///< The number of huge OS pages (1GiB in size) to reserve at the start of the program.
mi_option_reserve_huge_os_pages_at, ///< Reserve huge OS pages at node N.
mi_option_segment_cache, ///< The number of segments per thread to keep cached.
mi_option_page_reset, ///< Reset page memory after \a mi_option_reset_delay milliseconds when it becomes free.
mi_option_segment_reset, ///< Experimental
@ -1053,6 +1054,8 @@ or via environment variables.
`MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
and allocate just a little to take up space in the huge OS page area (which cannot be reset).
- `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N`: where N is the numa node. This reserves the huge pages at a specific numa node.
(`N` is -1 by default to reserve huge pages evenly among the given number of numa nodes (or use the available ones as detected))
Use caution when using `fork` in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write
for all pages in the original process including the huge OS pages. When any memory is now written in that area, the

6
ide/vs2019/mimalloc.vcxproj
View file

@ -92,7 +92,7 @@
</PropertyGroup>
<ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<WarningLevel>Level4</WarningLevel>
<Optimization>Disabled</Optimization>
<SDLCheck>true</SDLCheck>
<ConformanceMode>true</ConformanceMode>
@ -138,7 +138,7 @@
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<WarningLevel>Level4</WarningLevel>
<Optimization>MaxSpeed</Optimization>
<FunctionLevelLinking>true</FunctionLevelLinking>
<ConformanceMode>true</ConformanceMode>
@ -166,7 +166,7 @@
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
<ClCompile>
<WarningLevel>Level3</WarningLevel>
<WarningLevel>Level4</WarningLevel>
<Optimization>MaxSpeed</Optimization>
<FunctionLevelLinking>true</FunctionLevelLinking>
<ConformanceMode>true</ConformanceMode>

4
include/mimalloc-atomic.h
View file

@ -25,7 +25,7 @@ terms of the MIT license. A copy of the license can be found in the file
#define mi_memory_order(name) std::memory_order_##name
#elif defined(_MSC_VER)
// Use MSVC C wrapper for C11 atomics
#define _Atomic(tp) tp
#define _Atomic(tp) tp
#define ATOMIC_VAR_INIT(x) x
#define mi_atomic(name) mi_atomic_##name
#define mi_memory_order(name) mi_memory_order_##name
@ -173,7 +173,7 @@ static inline uintptr_t mi_atomic_exchange_explicit(_Atomic(uintptr_t)*p, uintpt
}
static inline void mi_atomic_thread_fence(mi_memory_order mo) {
(void)(mo);
_Atomic(uintptr_t)x = 0;
_Atomic(uintptr_t) x = 0;
mi_atomic_exchange_explicit(&x, 1, mo);
}
static inline uintptr_t mi_atomic_load_explicit(_Atomic(uintptr_t) const* p, mi_memory_order mo) {

112
include/mimalloc-internal.h
View file

@ -22,7 +22,7 @@ terms of the MIT license. A copy of the license can be found in the file
#define mi_decl_noinline __declspec(noinline)
#define mi_decl_thread __declspec(thread)
#define mi_decl_cache_align __declspec(align(MI_CACHE_LINE))
#elif (defined(__GNUC__) && (__GNUC__>=3)) // includes clang and icc
#elif (defined(__GNUC__) && (__GNUC__ >= 3)) || defined(__clang__) // includes clang and icc
#define mi_decl_noinline __attribute__((noinline))
#define mi_decl_thread __thread
#define mi_decl_cache_align __attribute__((aligned(MI_CACHE_LINE)))
@ -32,6 +32,16 @@ terms of the MIT license. A copy of the license can be found in the file
#define mi_decl_cache_align
#endif
#if defined(__EMSCRIPTEN__) && !defined(__wasi__)
#define __wasi__
#endif
#if defined(__cplusplus)
#define mi_decl_externc extern "C"
#else
#define mi_decl_externc
#endif
// "options.c"
void _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* message);
void _mi_fprintf(mi_output_fun* out, void* arg, const char* fmt, ...);
@ -46,14 +56,15 @@ void _mi_random_init(mi_random_ctx_t* ctx);
void _mi_random_split(mi_random_ctx_t* ctx, mi_random_ctx_t* new_ctx);
uintptr_t _mi_random_next(mi_random_ctx_t* ctx);
uintptr_t _mi_heap_random_next(mi_heap_t* heap);
uintptr_t _os_random_weak(uintptr_t extra_seed);
uintptr_t _mi_os_random_weak(uintptr_t extra_seed);
static inline uintptr_t _mi_random_shuffle(uintptr_t x);
// init.c
extern mi_decl_cache_align mi_stats_t _mi_stats_main;
extern mi_decl_cache_align const mi_page_t _mi_page_empty;
bool _mi_is_main_thread(void);
bool _mi_preloading(); // true while the C runtime is not ready
size_t _mi_current_thread_count(void);
bool _mi_preloading(void); // true while the C runtime is not ready
// os.c
size_t _mi_os_page_size(void);
@ -61,6 +72,7 @@ void _mi_os_init(void); // called fro
void* _mi_os_alloc(size_t size, mi_stats_t* stats); // to allocate thread local data
void _mi_os_free(void* p, size_t size, mi_stats_t* stats); // to free thread local data
size_t _mi_os_good_alloc_size(size_t size);
bool _mi_os_has_overcommit(void);
// memory.c
void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_pinned, bool* is_zero, size_t* id, mi_os_tld_t* tld);
@ -90,7 +102,7 @@ void _mi_abandoned_await_readers(void);
// "page.c"
void* _mi_malloc_generic(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc;
void _mi_page_retire(mi_page_t* page); // free the page if there are no other pages with many free blocks
void _mi_page_retire(mi_page_t* page) mi_attr_noexcept; // free the page if there are no other pages with many free blocks
void _mi_page_unfull(mi_page_t* page);
void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force); // free the page
void _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq); // abandon the page, to be picked up by another thread...
@ -176,11 +188,11 @@ bool _mi_page_is_valid(mi_page_t* page);
/* -----------------------------------------------------------
Inlined definitions
----------------------------------------------------------- */
#define UNUSED(x) (void)(x)
#define MI_UNUSED(x) (void)(x)
#if (MI_DEBUG>0)
#define UNUSED_RELEASE(x)
#define MI_UNUSED_RELEASE(x)
#else
#define UNUSED_RELEASE(x) UNUSED(x)
#define MI_UNUSED_RELEASE(x) MI_UNUSED(x)
#endif
#define MI_INIT4(x) x(),x(),x(),x()
@ -236,18 +248,18 @@ static inline bool mi_malloc_satisfies_alignment(size_t alignment, size_t size)
}
// Overflow detecting multiply
#if __has_builtin(__builtin_umul_overflow) || __GNUC__ >= 5
#if __has_builtin(__builtin_umul_overflow) || (defined(__GNUC__) && (__GNUC__ >= 5))
#include <limits.h> // UINT_MAX, ULONG_MAX
#if defined(_CLOCK_T) // for Illumos
#undef _CLOCK_T
#endif
static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) {
#if (SIZE_MAX == UINT_MAX)
return __builtin_umul_overflow(count, size, total);
#elif (SIZE_MAX == ULONG_MAX)
return __builtin_umull_overflow(count, size, total);
#if (SIZE_MAX == ULONG_MAX)
return __builtin_umull_overflow(count, size, (unsigned long *)total);
#elif (SIZE_MAX == UINT_MAX)
return __builtin_umul_overflow(count, size, (unsigned int *)total);
#else
return __builtin_umulll_overflow(count, size, total);
return __builtin_umulll_overflow(count, size, (unsigned long long *)total);
#endif
}
#else /* __builtin_umul_overflow is unavailable */
@ -285,7 +297,7 @@ We try to circumvent this in an efficient way:
- macOSX : we use an unused TLS slot from the OS allocated slots (MI_TLS_SLOT). On OSX, the
loader itself calls `malloc` even before the modules are initialized.
- OpenBSD: we use an unused slot from the pthread block (MI_TLS_PTHREAD_SLOT_OFS).
- DragonFly: the uniqueid use is buggy but kept for reference.
- DragonFly: defaults are working but seem slow compared to freeBSD (see PR #323)
------------------------------------------------------------------------------------------- */
extern const mi_heap_t _mi_heap_empty; // read-only empty heap, initial value of the thread local default heap
@ -295,15 +307,19 @@ mi_heap_t* _mi_heap_main_get(void); // statically allocated main backing hea
#if defined(MI_MALLOC_OVERRIDE)
#if defined(__APPLE__) // macOS
#define MI_TLS_SLOT 89 // seems unused?
// #define MI_TLS_RECURSE_GUARD 1
// other possible unused ones are 9, 29, __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY4 (94), __PTK_FRAMEWORK_GC_KEY9 (112) and __PTK_FRAMEWORK_OLDGC_KEY9 (89)
// see <https://github.com/rweichler/substrate/blob/master/include/pthread_machdep.h>
#elif defined(__OpenBSD__)
// use end bytes of a name; goes wrong if anyone uses names > 23 characters (ptrhread specifies 16)
// see <https://github.com/openbsd/src/blob/master/lib/libc/include/thread_private.h#L371>
#define MI_TLS_PTHREAD_SLOT_OFS (6*sizeof(int) + 4*sizeof(void*) + 24)
#elif defined(__DragonFly__)
#warning "mimalloc is not working correctly on DragonFly yet."
//#define MI_TLS_PTHREAD_SLOT_OFS (4 + 1*sizeof(void*)) // offset `uniqueid` (also used by gdb?) <https://github.com/DragonFlyBSD/DragonFlyBSD/blob/master/lib/libthread_xu/thread/thr_private.h#L458>
// #elif defined(__DragonFly__)
// #warning "mimalloc is not working correctly on DragonFly yet."
// #define MI_TLS_PTHREAD_SLOT_OFS (4 + 1*sizeof(void*)) // offset `uniqueid` (also used by gdb?) <https://github.com/DragonFlyBSD/DragonFlyBSD/blob/master/lib/libthread_xu/thread/thr_private.h#L458>
#elif defined(__ANDROID__)
// See issue #381
#define MI_TLS_PTHREAD
#endif
#endif
@ -332,10 +348,12 @@ extern pthread_key_t _mi_heap_default_key;
// However, on the Apple M1 we do use the address of this variable as the unique thread-id (issue #356).
extern mi_decl_thread mi_heap_t* _mi_heap_default; // default heap to allocate from
static inline mi_heap_t* mi_get_default_heap(void) {
#if defined(MI_TLS_SLOT)
mi_heap_t* heap = (mi_heap_t*)mi_tls_slot(MI_TLS_SLOT);
return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
if (mi_unlikely(heap == NULL)) { heap = (mi_heap_t*)&_mi_heap_empty; } //_mi_heap_empty_get(); }
return heap;
#elif defined(MI_TLS_PTHREAD_SLOT_OFS)
mi_heap_t* heap = *mi_tls_pthread_heap_slot();
return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
@ -343,7 +361,7 @@ static inline mi_heap_t* mi_get_default_heap(void) {
mi_heap_t* heap = (mi_unlikely(_mi_heap_default_key == (pthread_key_t)(-1)) ? _mi_heap_main_get() : (mi_heap_t*)pthread_getspecific(_mi_heap_default_key));
return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
#else
#if defined(MI_TLS_RECURSE_GUARD)
#if defined(MI_TLS_RECURSE_GUARD)
if (mi_unlikely(!_mi_process_is_initialized)) return _mi_heap_main_get();
#endif
return _mi_heap_default;
@ -399,11 +417,11 @@ static inline mi_segment_t* _mi_page_segment(const mi_page_t* page) {
}
// used internally
static inline uintptr_t _mi_segment_page_idx_of(const mi_segment_t* segment, const void* p) {
static inline size_t _mi_segment_page_idx_of(const mi_segment_t* segment, const void* p) {
// if (segment->page_size > MI_SEGMENT_SIZE) return &segment->pages[0]; // huge pages
ptrdiff_t diff = (uint8_t*)p - (uint8_t*)segment;
mi_assert_internal(diff >= 0 && (size_t)diff < MI_SEGMENT_SIZE);
uintptr_t idx = (uintptr_t)diff >> segment->page_shift;
size_t idx = (size_t)diff >> segment->page_shift;
mi_assert_internal(idx < segment->capacity);
mi_assert_internal(segment->page_kind <= MI_PAGE_MEDIUM || idx == 0);
return idx;
@ -411,7 +429,7 @@ static inline uintptr_t _mi_segment_page_idx_of(const mi_segment_t* segment, con
// Get the page containing the pointer
static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const void* p) {
uintptr_t idx = _mi_segment_page_idx_of(segment, p);
size_t idx = _mi_segment_page_idx_of(segment, p);
return &((mi_segment_t*)segment)->pages[idx];
}
@ -427,7 +445,7 @@ static inline mi_page_t* _mi_ptr_page(void* p) {
return _mi_segment_page_of(_mi_ptr_segment(p), p);
}
// Get the block size of a page (special cased for huge objects)
// Get the block size of a page (special case for huge objects)
static inline size_t mi_page_block_size(const mi_page_t* page) {
const size_t bsize = page->xblock_size;
mi_assert_internal(bsize > 0);
@ -569,8 +587,8 @@ static inline bool mi_is_in_same_page(const void* p, const void* q) {
mi_segment_t* segmentp = _mi_ptr_segment(p);
mi_segment_t* segmentq = _mi_ptr_segment(q);
if (segmentp != segmentq) return false;
uintptr_t idxp = _mi_segment_page_idx_of(segmentp, p);
uintptr_t idxq = _mi_segment_page_idx_of(segmentq, q);
size_t idxp = _mi_segment_page_idx_of(segmentp, p);
size_t idxq = _mi_segment_page_idx_of(segmentq, q);
return (idxp == idxq);
}
@ -597,7 +615,7 @@ static inline mi_block_t* mi_block_nextx( const void* null, const mi_block_t* bl
#ifdef MI_ENCODE_FREELIST
return (mi_block_t*)mi_ptr_decode(null, block->next, keys);
#else
UNUSED(keys); UNUSED(null);
MI_UNUSED(keys); MI_UNUSED(null);
return (mi_block_t*)block->next;
#endif
}
@ -606,7 +624,7 @@ static inline void mi_block_set_nextx(const void* null, mi_block_t* block, const
#ifdef MI_ENCODE_FREELIST
block->next = mi_ptr_encode(null, next, keys);
#else
UNUSED(keys); UNUSED(null);
MI_UNUSED(keys); MI_UNUSED(null);
block->next = (mi_encoded_t)next;
#endif
}
@ -622,7 +640,7 @@ static inline mi_block_t* mi_block_next(const mi_page_t* page, const mi_block_t*
}
return next;
#else
UNUSED(page);
MI_UNUSED(page);
return mi_block_nextx(page,block,NULL);
#endif
}
@ -631,7 +649,7 @@ static inline void mi_block_set_next(const mi_page_t* page, mi_block_t* block, c
#ifdef MI_ENCODE_FREELIST
mi_block_set_nextx(page,block,next, page->keys);
#else
UNUSED(page);
MI_UNUSED(page);
mi_block_set_nextx(page,block,next,NULL);
#endif
}
@ -686,7 +704,7 @@ static inline size_t _mi_os_numa_node_count(void) {
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
static inline uintptr_t _mi_thread_id(void) mi_attr_noexcept {
static inline mi_threadid_t _mi_thread_id(void) mi_attr_noexcept {
// Windows: works on Intel and ARM in both 32- and 64-bit
return (uintptr_t)NtCurrentTeb();
}
@ -707,17 +725,17 @@ static inline void* mi_tls_slot(size_t slot) mi_attr_noexcept {
#elif defined(__x86_64__)
__asm__("movq %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : ); // x86_64 Linux, BSD uses FS
#elif defined(__arm__)
void** tcb; UNUSED(ofs);
void** tcb; MI_UNUSED(ofs);
__asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
res = tcb[slot];
#elif defined(__aarch64__)
void** tcb; UNUSED(ofs);
#if defined(__APPLE__) // M1, issue #343
void** tcb; MI_UNUSED(ofs);
#if defined(__APPLE__) // M1, issue #343
__asm__ volatile ("mrs %0, tpidrro_el0" : "=r" (tcb));
tcb = (void**)((uintptr_t)tcb & ~0x07UL); // clear lower 3 bits
#else
#else
__asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
#endif
#endif
res = tcb[slot];
#endif
return res;
@ -731,28 +749,28 @@ static inline void mi_tls_slot_set(size_t slot, void* value) mi_attr_noexcept {
#elif defined(__APPLE__) && defined(__x86_64__)
__asm__("movq %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x86_64 macOSX uses GS
#elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
__asm__("movl %1,%%fs:%1" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x32 ABI
__asm__("movl %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x32 ABI
#elif defined(__x86_64__)
__asm__("movq %1,%%fs:%1" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x86_64 Linux, BSD uses FS
__asm__("movq %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x86_64 Linux, BSD uses FS
#elif defined(__arm__)
void** tcb; UNUSED(ofs);
void** tcb; MI_UNUSED(ofs);
__asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
tcb[slot] = value;
#elif defined(__aarch64__)
void** tcb; UNUSED(ofs);
#if defined(__APPLE__) // M1, issue #343
void** tcb; MI_UNUSED(ofs);
#if defined(__APPLE__) // M1, issue #343
__asm__ volatile ("mrs %0, tpidrro_el0" : "=r" (tcb));
tcb = (void**)((uintptr_t)tcb & ~0x07UL); // clear lower 3 bits
#else
#else
__asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
#endif
#endif
tcb[slot] = value;
#endif
}
static inline uintptr_t _mi_thread_id(void) mi_attr_noexcept {
#if defined(__BIONIC__) && (defined(__arm__) || defined(__aarch64__))
// on Android, slot 1 is the thread ID (pointer to pthread internal struct)
static inline mi_threadid_t _mi_thread_id(void) mi_attr_noexcept {
#if defined(__arm__) || (defined(__ANDROID__) && defined(__aarch64__))
// issue #384, #495: on arm32 and arm32/arm64 Android, slot 1 is the thread ID (pointer to pthread internal struct)
return (uintptr_t)mi_tls_slot(1);
#else
// in all our other targets, slot 0 is the pointer to the thread control block
@ -761,7 +779,7 @@ static inline uintptr_t _mi_thread_id(void) mi_attr_noexcept {
}
#else
// otherwise use standard C
static inline uintptr_t _mi_thread_id(void) mi_attr_noexcept {
static inline mi_threadid_t _mi_thread_id(void) mi_attr_noexcept {
return (uintptr_t)&_mi_heap_default;
}
#endif
@ -903,7 +921,7 @@ static inline void _mi_memcpy(void* dst, const void* src, size_t n) {
// This is used for example in `mi_realloc`.
// -------------------------------------------------------------------------------
#if (__GNUC__ >= 4) || defined(__clang__)
#if (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__)
// On GCC/CLang we provide a hint that the pointers are word aligned.
#include <string.h>
static inline void _mi_memcpy_aligned(void* dst, const void* src, size_t n) {

1
include/mimalloc-override.h
View file

@ -48,6 +48,7 @@ not accidentally mix pointers from different allocators).
#define valloc(n) mi_valloc(n)
#define pvalloc(n) mi_pvalloc(n)
#define reallocarray(p,s,n) mi_reallocarray(p,s,n)
#define reallocarr(p,s,n) mi_reallocarr(p,s,n)
#define memalign(a,n) mi_memalign(a,n)
#define aligned_alloc(a,n) mi_aligned_alloc(a,n)
#define posix_memalign(p,a,n) mi_posix_memalign(p,a,n)

88
include/mimalloc-types.h
View file

@ -17,7 +17,7 @@ terms of the MIT license. A copy of the license can be found in the file
#endif
// Minimal alignment necessary. On most platforms 16 bytes are needed
// due to SSE registers for example. This must be at least `MI_INTPTR_SIZE`
// due to SSE registers for example. This must be at least `sizeof(void*)`
#ifndef MI_MAX_ALIGN_SIZE
#define MI_MAX_ALIGN_SIZE 16 // sizeof(max_align_t)
#endif
@ -67,6 +67,7 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_ENCODE_FREELIST 1
#endif
// ------------------------------------------------------
// Platform specific values
// ------------------------------------------------------
@ -83,20 +84,43 @@ terms of the MIT license. A copy of the license can be found in the file
// or otherwise one might define an intptr_t type that is larger than a pointer...
// ------------------------------------------------------
#if INTPTR_MAX == 9223372036854775807LL
#if INTPTR_MAX > INT64_MAX
# define MI_INTPTR_SHIFT (4) // assume 128-bit (as on arm CHERI for example)
#elif INTPTR_MAX == INT64_MAX
# define MI_INTPTR_SHIFT (3)
#elif INTPTR_MAX == 2147483647LL
#elif INTPTR_MAX == INT32_MAX
# define MI_INTPTR_SHIFT (2)
#else
#error platform must be 32 or 64 bits
#error platform pointers must be 32, 64, or 128 bits
#endif
#if SIZE_MAX == UINT64_MAX
# define MI_SIZE_SHIFT (3)
typedef int64_t mi_ssize_t;
#elif SIZE_MAX == UINT32_MAX
# define MI_SIZE_SHIFT (2)
typedef int32_t mi_ssize_t;
#else
#error platform objects must be 32 or 64 bits
#endif
#if (SIZE_MAX/2) > LONG_MAX
# define MI_ZU(x) x##ULL
# define MI_ZI(x) x##LL
#else
# define MI_ZU(x) x##UL
# define MI_ZI(x) x##L
#endif
#define MI_INTPTR_SIZE (1<<MI_INTPTR_SHIFT)
#define MI_INTPTR_BITS (MI_INTPTR_SIZE*8)
#define KiB ((size_t)1024)
#define MiB (KiB*KiB)
#define GiB (MiB*KiB)
#define MI_SIZE_SIZE (1<<MI_SIZE_SHIFT)
#define MI_SIZE_BITS (MI_SIZE_SIZE*8)
#define MI_KiB (MI_ZU(1024))
#define MI_MiB (MI_KiB*MI_KiB)
#define MI_GiB (MI_MiB*MI_KiB)
// ------------------------------------------------------
@ -105,18 +129,18 @@ terms of the MIT license. A copy of the license can be found in the file
// Main tuning parameters for segment and page sizes
// Sizes for 64-bit, divide by two for 32-bit
#define MI_SMALL_PAGE_SHIFT (13 + MI_INTPTR_SHIFT) // 64kb
#define MI_MEDIUM_PAGE_SHIFT ( 3 + MI_SMALL_PAGE_SHIFT) // 512kb
#define MI_LARGE_PAGE_SHIFT ( 3 + MI_MEDIUM_PAGE_SHIFT) // 4mb
#define MI_SEGMENT_SHIFT ( MI_LARGE_PAGE_SHIFT) // 4mb
#define MI_SMALL_PAGE_SHIFT (13 + MI_INTPTR_SHIFT) // 64KiB
#define MI_MEDIUM_PAGE_SHIFT ( 3 + MI_SMALL_PAGE_SHIFT) // 512KiB
#define MI_LARGE_PAGE_SHIFT ( 3 + MI_MEDIUM_PAGE_SHIFT) // 4MiB
#define MI_SEGMENT_SHIFT ( MI_LARGE_PAGE_SHIFT) // 4MiB
// Derived constants
#define MI_SEGMENT_SIZE (1UL<<MI_SEGMENT_SHIFT)
#define MI_SEGMENT_MASK ((uintptr_t)MI_SEGMENT_SIZE - 1)
#define MI_SEGMENT_SIZE (MI_ZU(1)<<MI_SEGMENT_SHIFT)
#define MI_SEGMENT_MASK (MI_SEGMENT_SIZE - 1)
#define MI_SMALL_PAGE_SIZE (1UL<<MI_SMALL_PAGE_SHIFT)
#define MI_MEDIUM_PAGE_SIZE (1UL<<MI_MEDIUM_PAGE_SHIFT)
#define MI_LARGE_PAGE_SIZE (1UL<<MI_LARGE_PAGE_SHIFT)
#define MI_SMALL_PAGE_SIZE (MI_ZU(1)<<MI_SMALL_PAGE_SHIFT)
#define MI_MEDIUM_PAGE_SIZE (MI_ZU(1)<<MI_MEDIUM_PAGE_SHIFT)
#define MI_LARGE_PAGE_SIZE (MI_ZU(1)<<MI_LARGE_PAGE_SHIFT)
#define MI_SMALL_PAGES_PER_SEGMENT (MI_SEGMENT_SIZE/MI_SMALL_PAGE_SIZE)
#define MI_MEDIUM_PAGES_PER_SEGMENT (MI_SEGMENT_SIZE/MI_MEDIUM_PAGE_SIZE)
@ -124,9 +148,9 @@ terms of the MIT license. A copy of the license can be found in the file
// The max object size are checked to not waste more than 12.5% internally over the page sizes.
// (Except for large pages since huge objects are allocated in 4MiB chunks)
#define MI_SMALL_OBJ_SIZE_MAX (MI_SMALL_PAGE_SIZE/4) // 16kb
#define MI_MEDIUM_OBJ_SIZE_MAX (MI_MEDIUM_PAGE_SIZE/4) // 128kb
#define MI_LARGE_OBJ_SIZE_MAX (MI_LARGE_PAGE_SIZE/2) // 2mb
#define MI_SMALL_OBJ_SIZE_MAX (MI_SMALL_PAGE_SIZE/4) // 16KiB
#define MI_MEDIUM_OBJ_SIZE_MAX (MI_MEDIUM_PAGE_SIZE/4) // 128KiB
#define MI_LARGE_OBJ_SIZE_MAX (MI_LARGE_PAGE_SIZE/2) // 2MiB
#define MI_LARGE_OBJ_WSIZE_MAX (MI_LARGE_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
#define MI_HUGE_OBJ_SIZE_MAX (2*MI_INTPTR_SIZE*MI_SEGMENT_SIZE) // (must match MI_REGION_MAX_ALLOC_SIZE in memory.c)
@ -140,9 +164,17 @@ terms of the MIT license. A copy of the license can be found in the file
// Used as a special value to encode block sizes in 32 bits.
#define MI_HUGE_BLOCK_SIZE ((uint32_t)MI_HUGE_OBJ_SIZE_MAX)
// ------------------------------------------------------
// Mimalloc pages contain allocated blocks
// ------------------------------------------------------
// The free lists use encoded next fields
// (Only actually encodes when MI_ENCODED_FREELIST is defined.)
typedef uintptr_t mi_encoded_t;
typedef uintptr_t mi_encoded_t;
// thread id's
typedef size_t mi_threadid_t;
// free lists contain blocks
typedef struct mi_block_s {
@ -249,13 +281,13 @@ typedef struct mi_page_s {
typedef enum mi_page_kind_e {
MI_PAGE_SMALL, // small blocks go into 64kb pages inside a segment
MI_PAGE_MEDIUM, // medium blocks go into 512kb pages inside a segment
MI_PAGE_SMALL, // small blocks go into 64KiB pages inside a segment
MI_PAGE_MEDIUM, // medium blocks go into 512KiB pages inside a segment
MI_PAGE_LARGE, // larger blocks go into a single page spanning a whole segment
MI_PAGE_HUGE // huge blocks (>512kb) are put into a single page in a segment of the exact size (but still 2mb aligned)
MI_PAGE_HUGE // huge blocks (>512KiB) are put into a single page in a segment of the exact size (but still 2MiB aligned)
} mi_page_kind_t;
// Segments are large allocated memory blocks (2mb on 64 bit) from
// Segments are large allocated memory blocks (2MiB on 64 bit) from
// the OS. Inside segments we allocated fixed size _pages_ that
// contain blocks.
typedef struct mi_segment_s {
@ -279,8 +311,8 @@ typedef struct mi_segment_s {
uintptr_t cookie; // verify addresses in secure mode: `_mi_ptr_cookie(segment) == segment->cookie`
// layout like this to optimize access in `mi_free`
size_t page_shift; // `1 << page_shift` == the page sizes == `page->block_size * page->reserved` (unless the first page, then `-segment_info_size`).
_Atomic(uintptr_t) thread_id; // unique id of the thread owning this segment
size_t page_shift; // `1 << page_shift` == the page sizes == `page->block_size * page->reserved` (unless the first page, then `-segment_info_size`).
_Atomic(mi_threadid_t) thread_id; // unique id of the thread owning this segment
mi_page_kind_t page_kind; // kind of pages: small, large, or huge
mi_page_t pages[1]; // up to `MI_SMALL_PAGES_PER_SEGMENT` pages
} mi_segment_t;
@ -319,7 +351,7 @@ typedef struct mi_random_cxt_s {
} mi_random_ctx_t;
// In debug mode there is a padding stucture at the end of the blocks to check for buffer overflows
// In debug mode there is a padding structure at the end of the blocks to check for buffer overflows
#if (MI_PADDING)
typedef struct mi_padding_s {
uint32_t canary; // encoded block value to check validity of the padding (in case of overflow)
@ -341,7 +373,7 @@ struct mi_heap_s {
mi_page_t* pages_free_direct[MI_PAGES_DIRECT]; // optimize: array where every entry points a page with possibly free blocks in the corresponding queue for that size.
mi_page_queue_t pages[MI_BIN_FULL + 1]; // queue of pages for each size class (or "bin")
_Atomic(mi_block_t*) thread_delayed_free;
uintptr_t thread_id; // thread this heap belongs too
mi_threadid_t thread_id; // thread this heap belongs too
uintptr_t cookie; // random cookie to verify pointers (see `_mi_ptr_cookie`)
uintptr_t keys[2]; // two random keys used to encode the `thread_delayed_free` list
mi_random_ctx_t random; // random number context used for secure allocation

15
include/mimalloc.h
View file

@ -8,7 +8,7 @@ terms of the MIT license. A copy of the license can be found in the file
#ifndef MIMALLOC_H
#define MIMALLOC_H
#define MI_MALLOC_VERSION 171 // major + 2 digits minor
#define MI_MALLOC_VERSION 174 // major + 2 digits minor
// ------------------------------------------------------
// Compiler specific attributes
@ -26,7 +26,7 @@ terms of the MIT license. A copy of the license can be found in the file
#if defined(__cplusplus) && (__cplusplus >= 201703)
#define mi_decl_nodiscard [[nodiscard]]
#elif (__GNUC__ >= 4) || defined(__clang__) // includes clang, icc, and clang-cl
#elif (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__) // includes clang, icc, and clang-cl
#define mi_decl_nodiscard __attribute__((warn_unused_result))
#elif (_MSC_VER >= 1700)
#define mi_decl_nodiscard _Check_return_
@ -58,8 +58,12 @@ terms of the MIT license. A copy of the license can be found in the file
#define mi_attr_alloc_size2(s1,s2)
#define mi_attr_alloc_align(p)
#elif defined(__GNUC__) // includes clang and icc
#if defined(MI_SHARED_LIB) && defined(MI_SHARED_LIB_EXPORT)
#define mi_decl_export __attribute__((visibility("default")))
#else
#define mi_decl_export
#endif
#define mi_cdecl // leads to warnings... __attribute__((cdecl))
#define mi_decl_export __attribute__((visibility("default")))
#define mi_decl_restrict
#define mi_attr_malloc __attribute__((malloc))
#if (defined(__clang_major__) && (__clang_major__ < 4)) || (__GNUC__ < 5)
@ -267,7 +271,6 @@ mi_decl_export int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size
mi_decl_export int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept;
mi_decl_export bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept;
// deprecated
mi_decl_export int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept;
@ -306,6 +309,7 @@ typedef enum mi_option_e {
mi_option_reset_decommits,
mi_option_large_os_pages, // implies eager commit
mi_option_reserve_huge_os_pages,
mi_option_reserve_huge_os_pages_at,
mi_option_reserve_os_memory,
mi_option_segment_cache,
mi_option_page_reset,
@ -342,6 +346,7 @@ mi_decl_export void mi_option_set_default(mi_option_t option, long value);
mi_decl_export void mi_cfree(void* p) mi_attr_noexcept;
mi_decl_export void* mi__expand(void* p, size_t newsize) mi_attr_noexcept;
mi_decl_nodiscard mi_decl_export size_t mi_malloc_size(const void* p) mi_attr_noexcept;
mi_decl_nodiscard mi_decl_export size_t mi_malloc_good_size(size_t size) mi_attr_noexcept;
mi_decl_nodiscard mi_decl_export size_t mi_malloc_usable_size(const void *p) mi_attr_noexcept;
mi_decl_export int mi_posix_memalign(void** p, size_t alignment, size_t size) mi_attr_noexcept;
@ -351,6 +356,7 @@ mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_pvalloc(size_t size)
mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_aligned_alloc(size_t alignment, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(1);
mi_decl_nodiscard mi_decl_export void* mi_reallocarray(void* p, size_t count, size_t size) mi_attr_noexcept mi_attr_alloc_size2(2,3);
mi_decl_nodiscard mi_decl_export int mi_reallocarr(void* p, size_t count, size_t size) mi_attr_noexcept mi_attr_alloc_size2(2,3);
mi_decl_nodiscard mi_decl_export void* mi_aligned_recalloc(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept;
mi_decl_nodiscard mi_decl_export void* mi_aligned_offset_recalloc(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept;
@ -383,6 +389,7 @@ mi_decl_nodiscard mi_decl_export void* mi_new_reallocn(void* p, size_t newcount,
// ---------------------------------------------------------------------------------------------
#ifdef __cplusplus
#include <cstddef> // std::size_t
#include <cstdint> // PTRDIFF_MAX
#if (__cplusplus >= 201103L) || (_MSC_VER > 1900) // C++11
#include <type_traits> // std::true_type

29
readme.md
View file

@ -12,13 +12,13 @@ is a general purpose allocator with excellent [performance](#performance) charac
Initially developed by Daan Leijen for the run-time systems of the
[Koka](https://koka-lang.github.io) and [Lean](https://github.com/leanprover/lean) languages.
Latest release tag: `v2.0.2` (beta, 2021-06-17).
Latest stable tag: `v1.7.2` (2021-06-17).
Latest release tag: `v2.0.3` (beta, 2021-11-14).
Latest stable tag: `v1.7.3` (2021-11-14).
mimalloc is a drop-in replacement for `malloc` and can be used in other programs
without code changes, for example, on dynamically linked ELF-based systems (Linux, BSD, etc.) you can use it as:
```
> LD_PRELOAD=/usr/bin/libmimalloc.so myprogram
> LD_PRELOAD=/usr/lib/libmimalloc.so myprogram
```
It also has an easy way to override the default allocator in [Windows](#override_on_windows). Notable aspects of the design include:
@ -77,6 +77,10 @@ Note: the `v2.x` beta has a new algorithm for managing internal mimalloc pages t
and fragmentation compared to mimalloc `v1.x` (especially for large workloads). Should otherwise have similar performance
(see [below](#performance)); please report if you observe any significant performance regression.
* 2021-11-14, `v1.7.3`, `v2.0.3` (beta): improved WASM support, improved macOS support and performance (including
M1), improved performance for v2 for large objects, Python integration improvements, more standard
installation directories, various small fixes.
* 2021-06-17, `v1.7.2`, `v2.0.2` (beta): support M1, better installation layout on Linux, fix
thread_id on Android, prefer 2-6TiB area for aligned allocation to work better on pre-windows 8, various small fixes.
@ -142,7 +146,7 @@ mimalloc is used in various large scale low-latency services and programs, for e
## Windows
Open `ide/vs2019/mimalloc.sln` in Visual Studio 2019 and build (or `ide/vs2017/mimalloc.sln`).
Open `ide/vs2019/mimalloc.sln` in Visual Studio 2019 and build.
The `mimalloc` project builds a static library (in `out/msvc-x64`), while the
`mimalloc-override` project builds a DLL for overriding malloc
in the entire program.
@ -191,6 +195,11 @@ Notes:
2. Install CCMake: `sudo apt-get install cmake-curses-gui`
## Single source
You can also directly build the single `src/static.c` file as part of your project without
needing `cmake` at all. Make sure to also add the mimalloc `include` directory to the include path.
# Using the library
@ -302,6 +311,9 @@ or via environment variables:
`MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
and allocate just a little to take up space in the huge OS page area (which cannot be reset).
The huge pages are usually allocated evenly among NUMA nodes.
We can use `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N` where `N` is the numa node (starting at 0) to allocate all
the huge pages at a specific numa node instead.
Use caution when using `fork` in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write
for all pages in the original process including the huge OS pages. When any memory is now written in that area, the
@ -337,9 +349,9 @@ When _mimalloc_ is built using debug mode, various checks are done at runtime to
- Corrupted free-lists and some forms of use-after-free are detected.
# Overriding Malloc
# Overriding Standard Malloc
Overriding the standard `malloc` can be done either _dynamically_ or _statically_.
Overriding the standard `malloc` (and `new`) can be done either _dynamically_ or _statically_.
## Dynamic override
@ -370,13 +382,12 @@ On macOS we can also preload the mimalloc shared
library so all calls to the standard `malloc` interface are
resolved to the _mimalloc_ library.
```
> env DYLD_FORCE_FLAT_NAMESPACE=1 DYLD_INSERT_LIBRARIES=/usr/lib/libmimalloc.dylib myprogram
> env DYLD_INSERT_LIBRARIES=/usr/lib/libmimalloc.dylib myprogram
```
Note that certain security restrictions may apply when doing this from
the [shell](https://stackoverflow.com/questions/43941322/dyld-insert-libraries-ignored-when-calling-application-through-bash).
(Note: macOS support for dynamic overriding is recent, please report any issues.)
### Override on Windows
@ -386,7 +397,7 @@ the (dynamic) C runtime allocator, including those from other DLL's or libraries
The overriding on Windows requires that you link your program explicitly with
the mimalloc DLL and use the C-runtime library as a DLL (using the `/MD` or `/MDd` switch).
Also, the `mimalloc-redirect.dll` (or `mimalloc-redirect32.dll`) must be available
Also, the `mimalloc-redirect.dll` (or `mimalloc-redirect32.dll`) must be put
in the same folder as the main `mimalloc-override.dll` at runtime (as it is a dependency).
The redirection DLL ensures that all calls to the C runtime malloc API get redirected to
mimalloc (in `mimalloc-override.dll`).

341
src/alloc-override-osx.c
View file

@ -17,17 +17,20 @@ terms of the MIT license. A copy of the license can be found in the file
/* ------------------------------------------------------
Override system malloc on macOS
This is done through the malloc zone interface.
It seems we also need to interpose (see `alloc-override.c`)
or otherwise we get zone errors as there are usually
already allocations done by the time we take over the
zone. Unfortunately, that means we need to replace
the `free` with a checked free (`cfree`) impacting
performance.
It seems to be most robust in combination with interposing
though or otherwise we may get zone errors as there are could
be allocations done by the time we take over the
zone.
------------------------------------------------------ */
#include <AvailabilityMacros.h>
#include <malloc/malloc.h>
#include <string.h> // memset
#include <stdlib.h>
#ifdef __cplusplus
extern "C" {
#endif
#if defined(MAC_OS_X_VERSION_10_6) && \
MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
@ -40,45 +43,43 @@ extern malloc_zone_t* malloc_default_purgeable_zone(void) __attribute__((weak_im
------------------------------------------------------ */
static size_t zone_size(malloc_zone_t* zone, const void* p) {
UNUSED(zone);
if (!mi_is_in_heap_region(p))
return 0; // not our pointer, bail out
MI_UNUSED(zone);
//if (!mi_is_in_heap_region(p)){ return 0; } // not our pointer, bail out
return mi_usable_size(p);
}
static void* zone_malloc(malloc_zone_t* zone, size_t size) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_malloc(size);
}
static void* zone_calloc(malloc_zone_t* zone, size_t count, size_t size) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_calloc(count, size);
}
static void* zone_valloc(malloc_zone_t* zone, size_t size) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_malloc_aligned(size, _mi_os_page_size());
}
static void zone_free(malloc_zone_t* zone, void* p) {
UNUSED(zone);
MI_UNUSED(zone);
mi_free(p);
}
static void* zone_realloc(malloc_zone_t* zone, void* p, size_t newsize) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_realloc(p, newsize);
}
static void* zone_memalign(malloc_zone_t* zone, size_t alignment, size_t size) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_malloc_aligned(size,alignment);
}
static void zone_destroy(malloc_zone_t* zone) {
UNUSED(zone);
MI_UNUSED(zone);
// todo: ignore for now?
}
@ -99,16 +100,21 @@ static void zone_batch_free(malloc_zone_t* zone, void** ps, unsigned count) {
}
static size_t zone_pressure_relief(malloc_zone_t* zone, size_t size) {
UNUSED(zone); UNUSED(size);
MI_UNUSED(zone); MI_UNUSED(size);
mi_collect(false);
return 0;
}
static void zone_free_definite_size(malloc_zone_t* zone, void* p, size_t size) {
UNUSED(size);
MI_UNUSED(size);
zone_free(zone,p);
}
static boolean_t zone_claimed_address(malloc_zone_t* zone, void* p) {
MI_UNUSED(zone);
return mi_is_in_heap_region(p);
}
/* ------------------------------------------------------
Introspection members
@ -120,43 +126,43 @@ static kern_return_t intro_enumerator(task_t task, void* p,
vm_range_recorder_t recorder)
{
// todo: enumerate all memory
UNUSED(task); UNUSED(p); UNUSED(type_mask); UNUSED(zone_address);
UNUSED(reader); UNUSED(recorder);
MI_UNUSED(task); MI_UNUSED(p); MI_UNUSED(type_mask); MI_UNUSED(zone_address);
MI_UNUSED(reader); MI_UNUSED(recorder);
return KERN_SUCCESS;
}
static size_t intro_good_size(malloc_zone_t* zone, size_t size) {
UNUSED(zone);
MI_UNUSED(zone);
return mi_good_size(size);
}
static boolean_t intro_check(malloc_zone_t* zone) {
UNUSED(zone);
MI_UNUSED(zone);
return true;
}
static void intro_print(malloc_zone_t* zone, boolean_t verbose) {
UNUSED(zone); UNUSED(verbose);
MI_UNUSED(zone); MI_UNUSED(verbose);
mi_stats_print(NULL);
}
static void intro_log(malloc_zone_t* zone, void* p) {
UNUSED(zone); UNUSED(p);
MI_UNUSED(zone); MI_UNUSED(p);
// todo?
}
static void intro_force_lock(malloc_zone_t* zone) {
UNUSED(zone);
MI_UNUSED(zone);
// todo?
}
static void intro_force_unlock(malloc_zone_t* zone) {
UNUSED(zone);
MI_UNUSED(zone);
// todo?
}
static void intro_statistics(malloc_zone_t* zone, malloc_statistics_t* stats) {
UNUSED(zone);
MI_UNUSED(zone);
// todo...
stats->blocks_in_use = 0;
stats->size_in_use = 0;
@ -165,7 +171,7 @@ static void intro_statistics(malloc_zone_t* zone, malloc_statistics_t* stats) {
}
static boolean_t intro_zone_locked(malloc_zone_t* zone) {
UNUSED(zone);
MI_UNUSED(zone);
return false;
}
@ -174,7 +180,220 @@ static boolean_t intro_zone_locked(malloc_zone_t* zone) {
At process start, override the default allocator
------------------------------------------------------ */
static malloc_zone_t* mi_get_default_zone()
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
#endif
static malloc_introspection_t mi_introspect = {
.enumerator = &intro_enumerator,
.good_size = &intro_good_size,
.check = &intro_check,
.print = &intro_print,
.log = &intro_log,
.force_lock = &intro_force_lock,
.force_unlock = &intro_force_unlock,
#if defined(MAC_OS_X_VERSION_10_6) && \
MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
.statistics = &intro_statistics,
.zone_locked = &intro_zone_locked,
#endif
};
static malloc_zone_t mi_malloc_zone = {
.size = &zone_size,
.malloc = &zone_malloc,
.calloc = &zone_calloc,
.valloc = &zone_valloc,
.free = &zone_free,
.realloc = &zone_realloc,
.destroy = &zone_destroy,
.zone_name = "mimalloc",
.batch_malloc = &zone_batch_malloc,
.batch_free = &zone_batch_free,
.introspect = &mi_introspect,
#if defined(MAC_OS_X_VERSION_10_6) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
// switch to version 9+ on OSX 10.6 to support memalign.
.memalign = &zone_memalign,
.free_definite_size = &zone_free_definite_size,
.pressure_relief = &zone_pressure_relief,
#if defined(MAC_OS_X_VERSION_10_7) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_7
.claimed_address = &zone_claimed_address,
.version = 10
#else
.version = 9
#endif
#else
.version = 4
#endif
};
#ifdef __cplusplus
}
#endif
#if defined(MI_OSX_INTERPOSE) && defined(MI_SHARED_LIB_EXPORT)
// ------------------------------------------------------
// Override malloc_xxx and malloc_zone_xxx api's to use only
// our mimalloc zone. Since even the loader uses malloc
// on macOS, this ensures that all allocations go through
// mimalloc (as all calls are interposed).
// The main `malloc`, `free`, etc calls are interposed in `alloc-override.c`,
// Here, we also override macOS specific API's like
// `malloc_zone_calloc` etc. see <https://github.com/aosm/libmalloc/blob/master/man/malloc_zone_malloc.3>
// ------------------------------------------------------
static inline malloc_zone_t* mi_get_default_zone(void)
{
static bool init;
if (mi_unlikely(!init)) {
init = true;
malloc_zone_register(&mi_malloc_zone); // by calling register we avoid a zone error on free (see <http://eatmyrandom.blogspot.com/2010/03/mallocfree-interception-on-mac-os-x.html>)
}
return &mi_malloc_zone;
}
mi_decl_externc int malloc_jumpstart(uintptr_t cookie);
mi_decl_externc void _malloc_fork_prepare(void);
mi_decl_externc void _malloc_fork_parent(void);
mi_decl_externc void _malloc_fork_child(void);
static malloc_zone_t* mi_malloc_create_zone(vm_size_t size, unsigned flags) {
MI_UNUSED(size); MI_UNUSED(flags);
return mi_get_default_zone();
}
static malloc_zone_t* mi_malloc_default_zone (void) {
return mi_get_default_zone();
}
static malloc_zone_t* mi_malloc_default_purgeable_zone(void) {
return mi_get_default_zone();
}
static void mi_malloc_destroy_zone(malloc_zone_t* zone) {
MI_UNUSED(zone);
// nothing.
}
static kern_return_t mi_malloc_get_all_zones (task_t task, memory_reader_t mr, vm_address_t** addresses, unsigned* count) {
MI_UNUSED(task); MI_UNUSED(mr);
if (addresses != NULL) *addresses = NULL;
if (count != NULL) *count = 0;
return KERN_SUCCESS;
}
static const char* mi_malloc_get_zone_name(malloc_zone_t* zone) {
return (zone == NULL ? mi_malloc_zone.zone_name : zone->zone_name);
}
static void mi_malloc_set_zone_name(malloc_zone_t* zone, const char* name) {
MI_UNUSED(zone); MI_UNUSED(name);
}
static int mi_malloc_jumpstart(uintptr_t cookie) {
MI_UNUSED(cookie);
return 1; // or 0 for no error?
}
static void mi__malloc_fork_prepare(void) {
// nothing
}
static void mi__malloc_fork_parent(void) {
// nothing
}
static void mi__malloc_fork_child(void) {
// nothing
}
static void mi_malloc_printf(const char* fmt, ...) {
MI_UNUSED(fmt);
}
static bool zone_check(malloc_zone_t* zone) {
MI_UNUSED(zone);
return true;
}
static malloc_zone_t* zone_from_ptr(const void* p) {
MI_UNUSED(p);
return mi_get_default_zone();
}
static void zone_log(malloc_zone_t* zone, void* p) {
MI_UNUSED(zone); MI_UNUSED(p);
}
static void zone_print(malloc_zone_t* zone, bool b) {
MI_UNUSED(zone); MI_UNUSED(b);
}
static void zone_print_ptr_info(void* p) {
MI_UNUSED(p);
}
static void zone_register(malloc_zone_t* zone) {
MI_UNUSED(zone);
}
static void zone_unregister(malloc_zone_t* zone) {
MI_UNUSED(zone);
}
// use interposing so `DYLD_INSERT_LIBRARIES` works without `DYLD_FORCE_FLAT_NAMESPACE=1`
// See: <https://books.google.com/books?id=K8vUkpOXhN4C&pg=PA73>
struct mi_interpose_s {
const void* replacement;
const void* target;
};
#define MI_INTERPOSE_FUN(oldfun,newfun) { (const void*)&newfun, (const void*)&oldfun }
#define MI_INTERPOSE_MI(fun) MI_INTERPOSE_FUN(fun,mi_##fun)
#define MI_INTERPOSE_ZONE(fun) MI_INTERPOSE_FUN(malloc_##fun,fun)
__attribute__((used)) static const struct mi_interpose_s _mi_zone_interposes[] __attribute__((section("__DATA, __interpose"))) =
{
MI_INTERPOSE_MI(malloc_create_zone),
MI_INTERPOSE_MI(malloc_default_purgeable_zone),
MI_INTERPOSE_MI(malloc_default_zone),
MI_INTERPOSE_MI(malloc_destroy_zone),
MI_INTERPOSE_MI(malloc_get_all_zones),
MI_INTERPOSE_MI(malloc_get_zone_name),
MI_INTERPOSE_MI(malloc_jumpstart),
MI_INTERPOSE_MI(malloc_printf),
MI_INTERPOSE_MI(malloc_set_zone_name),
MI_INTERPOSE_MI(_malloc_fork_child),
MI_INTERPOSE_MI(_malloc_fork_parent),
MI_INTERPOSE_MI(_malloc_fork_prepare),
MI_INTERPOSE_ZONE(zone_batch_free),
MI_INTERPOSE_ZONE(zone_batch_malloc),
MI_INTERPOSE_ZONE(zone_calloc),
MI_INTERPOSE_ZONE(zone_check),
MI_INTERPOSE_ZONE(zone_free),
MI_INTERPOSE_ZONE(zone_from_ptr),
MI_INTERPOSE_ZONE(zone_log),
MI_INTERPOSE_ZONE(zone_malloc),
MI_INTERPOSE_ZONE(zone_memalign),
MI_INTERPOSE_ZONE(zone_print),
MI_INTERPOSE_ZONE(zone_print_ptr_info),
MI_INTERPOSE_ZONE(zone_realloc),
MI_INTERPOSE_ZONE(zone_register),
MI_INTERPOSE_ZONE(zone_unregister),
MI_INTERPOSE_ZONE(zone_valloc)
};
#else
// ------------------------------------------------------
// hook into the zone api's without interposing
// This is the official way of adding an allocator but
// it seems less robust than using interpose.
// ------------------------------------------------------
static inline malloc_zone_t* mi_get_default_zone(void)
{
// The first returned zone is the real default
malloc_zone_t** zones = NULL;
@ -189,70 +408,21 @@ static malloc_zone_t* mi_get_default_zone()
}
}
static malloc_introspection_t mi_introspect = {
.enumerator = &intro_enumerator,
.good_size = &intro_good_size,
.check = &intro_check,
.print = &intro_print,
.log = &intro_log,
.force_lock = &intro_force_lock,
.force_unlock = &intro_force_unlock,
#if defined(MAC_OS_X_VERSION_10_6) && \
MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
.zone_locked = &intro_zone_locked,
.statistics = &intro_statistics,
#endif
};
static malloc_zone_t mi_malloc_zone = {
.size = &zone_size,
.zone_name = "mimalloc",
.introspect = &mi_introspect,
.malloc = &zone_malloc,
.calloc = &zone_calloc,
.valloc = &zone_valloc,
.free = &zone_free,
.realloc = &zone_realloc,
.destroy = &zone_destroy,
.batch_malloc = &zone_batch_malloc,
.batch_free = &zone_batch_free,
#if defined(MAC_OS_X_VERSION_10_6) && \
MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
// switch to version 9 on OSX 10.6 to support memalign.
.version = 9,
.memalign = &zone_memalign,
.free_definite_size = &zone_free_definite_size,
.pressure_relief = &zone_pressure_relief,
#if defined(__clang__)
__attribute__((constructor(0)))
#else
.version = 4,
__attribute__((constructor)) // seems not supported by g++-11 on the M1
#endif
};
#if defined(MI_SHARED_LIB_EXPORT) && defined(MI_INTERPOSE)
static malloc_zone_t *mi_malloc_default_zone(void) {
return &mi_malloc_zone;
}
// TODO: should use the macros in alloc-override but they aren't available here.
__attribute__((used)) static struct {
const void *replacement;
const void *target;
} replace_malloc_default_zone[] __attribute__((section("__DATA, __interpose"))) = {
{ (const void*)mi_malloc_default_zone, (const void*)malloc_default_zone },
};
#endif
static void __attribute__((constructor(0))) _mi_macos_override_malloc() {
static void _mi_macos_override_malloc() {
malloc_zone_t* purgeable_zone = NULL;
#if defined(MAC_OS_X_VERSION_10_6) && \
#if defined(MAC_OS_X_VERSION_10_6) && \
MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6
// force the purgeable zone to exist to avoid strange bugs
if (malloc_default_purgeable_zone) {
purgeable_zone = malloc_default_purgeable_zone();
}
#endif
#endif
// Register our zone.
// thomcc: I think this is still needed to put us in the zone list.
@ -277,5 +447,6 @@ static void __attribute__((constructor(0))) _mi_macos_override_malloc() {
}
}
#endif // MI_OSX_INTERPOSE
#endif // MI_MALLOC_OVERRIDE

188
src/alloc-override.c
View file

@ -13,15 +13,25 @@ terms of the MIT license. A copy of the license can be found in the file
#error "It is only possible to override "malloc" on Windows when building as a DLL (and linking the C runtime as a DLL)"
#endif
#if defined(MI_MALLOC_OVERRIDE) && !(defined(_WIN32)) // || (defined(__APPLE__) && !defined(MI_INTERPOSE)))
#if defined(MI_MALLOC_OVERRIDE) && !(defined(_WIN32))
#if defined(__APPLE__)
mi_decl_externc void vfree(void* p);
mi_decl_externc size_t malloc_size(const void* p);
mi_decl_externc size_t malloc_good_size(size_t size);
#endif
// helper definition for C override of C++ new
typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
// ------------------------------------------------------
// Override system malloc
// ------------------------------------------------------
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__)
// use aliasing to alias the exported function to one of our `mi_` functions
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__) && !defined(MI_VALGRIND)
// gcc, clang: use aliasing to alias the exported function to one of our `mi_` functions
#if (defined(__GNUC__) && __GNUC__ >= 9)
#pragma GCC diagnostic ignored "-Wattributes" // or we get warnings that nodiscard is ignored on a forward
#define MI_FORWARD(fun) __attribute__((alias(#fun), used, visibility("default"), copy(fun)));
#else
#define MI_FORWARD(fun) __attribute__((alias(#fun), used, visibility("default")));
@ -32,7 +42,7 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_FORWARD0(fun,x) MI_FORWARD(fun)
#define MI_FORWARD02(fun,x,y) MI_FORWARD(fun)
#else
// use forwarding by calling our `mi_` function
// otherwise use forwarding by calling our `mi_` function
#define MI_FORWARD1(fun,x) { return fun(x); }
#define MI_FORWARD2(fun,x,y) { return fun(x,y); }
#define MI_FORWARD3(fun,x,y,z) { return fun(x,y,z); }
@ -40,7 +50,11 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_FORWARD02(fun,x,y) { fun(x,y); }
#endif
#if defined(__APPLE__) && defined(MI_SHARED_LIB_EXPORT) && defined(MI_INTERPOSE)
#if defined(__APPLE__) && defined(MI_SHARED_LIB_EXPORT) && defined(MI_OSX_INTERPOSE)
// define MI_OSX_IS_INTERPOSED as we should not provide forwarding definitions for
// functions that are interposed (or the interposing does not work)
#define MI_OSX_IS_INTERPOSED
// use interposing so `DYLD_INSERT_LIBRARIES` works without `DYLD_FORCE_FLAT_NAMESPACE=1`
// See: <https://books.google.com/books?id=K8vUkpOXhN4C&pg=PA73>
struct mi_interpose_s {
@ -49,6 +63,7 @@ terms of the MIT license. A copy of the license can be found in the file
};
#define MI_INTERPOSE_FUN(oldfun,newfun) { (const void*)&newfun, (const void*)&oldfun }
#define MI_INTERPOSE_MI(fun) MI_INTERPOSE_FUN(fun,mi_##fun)
__attribute__((used)) static struct mi_interpose_s _mi_interposes[] __attribute__((section("__DATA, __interpose"))) =
{
MI_INTERPOSE_MI(malloc),
@ -60,21 +75,49 @@ terms of the MIT license. A copy of the license can be found in the file
MI_INTERPOSE_MI(posix_memalign),
MI_INTERPOSE_MI(reallocf),
MI_INTERPOSE_MI(valloc),
#ifndef MI_OSX_ZONE
// some code allocates from default zone but deallocates using plain free :-( (like NxHashResizeToCapacity <https://github.com/nneonneo/osx-10.9-opensource/blob/master/objc4-551.1/runtime/hashtable2.mm>)
MI_INTERPOSE_FUN(free,mi_cfree), // use safe free that checks if pointers are from us
#else
// We interpose malloc_default_zone in alloc-override-osx.c
MI_INTERPOSE_MI(malloc_size),
MI_INTERPOSE_MI(malloc_good_size),
MI_INTERPOSE_MI(aligned_alloc),
#ifdef MI_OSX_ZONE
// we interpose malloc_default_zone in alloc-override-osx.c so we can use mi_free safely
MI_INTERPOSE_MI(free),
#endif
// some code allocates from a zone but deallocates using plain free :-( (like NxHashResizeToCapacity <https://github.com/nneonneo/osx-10.9-opensource/blob/master/objc4-551.1/runtime/hashtable2.mm>)
MI_INTERPOSE_FUN(vfree,mi_free),
#else
// sometimes code allocates from default zone but deallocates using plain free :-( (like NxHashResizeToCapacity <https://github.com/nneonneo/osx-10.9-opensource/blob/master/objc4-551.1/runtime/hashtable2.mm>)
MI_INTERPOSE_FUN(free,mi_cfree), // use safe free that checks if pointers are from us
MI_INTERPOSE_FUN(vfree,mi_cfree),
#endif
};
#ifdef __cplusplus
extern "C" {
void _ZdlPv(void* p); // delete
void _ZdaPv(void* p); // delete[]
void _ZdlPvm(void* p, size_t n); // delete
void _ZdaPvm(void* p, size_t n); // delete[]
void* _Znwm(size_t n); // new
void* _Znam(size_t n); // new[]
void* _ZnwmRKSt9nothrow_t(size_t n, mi_nothrow_t tag); // new nothrow
void* _ZnamRKSt9nothrow_t(size_t n, mi_nothrow_t tag); // new[] nothrow
}
__attribute__((used)) static struct mi_interpose_s _mi_cxx_interposes[] __attribute__((section("__DATA, __interpose"))) =
{
MI_INTERPOSE_FUN(_ZdlPv,mi_free),
MI_INTERPOSE_FUN(_ZdaPv,mi_free),
MI_INTERPOSE_FUN(_ZdlPvm,mi_free_size),
MI_INTERPOSE_FUN(_ZdaPvm,mi_free_size),
MI_INTERPOSE_FUN(_Znwm,mi_new),
MI_INTERPOSE_FUN(_Znam,mi_new),
MI_INTERPOSE_FUN(_ZnwmRKSt9nothrow_t,mi_new_nothrow),
MI_INTERPOSE_FUN(_ZnamRKSt9nothrow_t,mi_new_nothrow),
};
#endif // __cplusplus
#elif defined(_MSC_VER)
// cannot override malloc unless using a dll.
// we just override new/delete which does work in a static library.
#else
// On all other systems forward to our API
// On all other systems forward to our API
void* malloc(size_t size) MI_FORWARD1(mi_malloc, size)
void* calloc(size_t size, size_t n) MI_FORWARD2(mi_calloc, size, n)
void* realloc(void* p, size_t newsize) MI_FORWARD2(mi_realloc, p, newsize)
@ -96,18 +139,21 @@ terms of the MIT license. A copy of the license can be found in the file
// see <https://en.cppreference.com/w/cpp/memory/new/operator_new>
// ------------------------------------------------------
#include <new>
void operator delete(void* p) noexcept MI_FORWARD0(mi_free,p)
void operator delete[](void* p) noexcept MI_FORWARD0(mi_free,p)
void* operator new(std::size_t n) noexcept(false) MI_FORWARD1(mi_new,n)
void* operator new[](std::size_t n) noexcept(false) MI_FORWARD1(mi_new,n)
#ifndef MI_OSX_IS_INTERPOSED
void operator delete(void* p) noexcept MI_FORWARD0(mi_free,p)
void operator delete[](void* p) noexcept MI_FORWARD0(mi_free,p)
void* operator new (std::size_t n, const std::nothrow_t& tag) noexcept { UNUSED(tag); return mi_new_nothrow(n); }
void* operator new[](std::size_t n, const std::nothrow_t& tag) noexcept { UNUSED(tag); return mi_new_nothrow(n); }
void* operator new(std::size_t n) noexcept(false) MI_FORWARD1(mi_new,n)
void* operator new[](std::size_t n) noexcept(false) MI_FORWARD1(mi_new,n)
#if (__cplusplus >= 201402L || _MSC_VER >= 1916)
void operator delete (void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
void operator delete[](void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
void* operator new (std::size_t n, const std::nothrow_t& tag) noexcept { MI_UNUSED(tag); return mi_new_nothrow(n); }
void* operator new[](std::size_t n, const std::nothrow_t& tag) noexcept { MI_UNUSED(tag); return mi_new_nothrow(n); }
#if (__cplusplus >= 201402L || _MSC_VER >= 1916)
void operator delete (void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
void operator delete[](void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
#endif
#endif
#if (__cplusplus > 201402L && defined(__cpp_aligned_new)) && (!defined(__GNUC__) || (__GNUC__ > 5))
@ -122,12 +168,13 @@ terms of the MIT license. A copy of the license can be found in the file
void* operator new[](std::size_t n, std::align_val_t al, const std::nothrow_t&) noexcept { return mi_new_aligned_nothrow(n, static_cast<size_t>(al)); }
#endif
#elif (defined(__GNUC__) || defined(__clang__))
#elif (defined(__GNUC__) || defined(__clang__))
// ------------------------------------------------------
// Override by defining the mangled C++ names of the operators (as
// used by GCC and CLang).
// See <https://itanium-cxx-abi.github.io/cxx-abi/abi.html#mangling>
// ------------------------------------------------------
void _ZdlPv(void* p) MI_FORWARD0(mi_free,p) // delete
void _ZdaPv(void* p) MI_FORWARD0(mi_free,p) // delete[]
void _ZdlPvm(void* p, size_t n) MI_FORWARD02(mi_free_size,p,n)
@ -136,78 +183,83 @@ terms of the MIT license. A copy of the license can be found in the file
void _ZdaPvSt11align_val_t(void* p, size_t al) { mi_free_aligned(p,al); }
void _ZdlPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
void _ZdaPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
#if (MI_INTPTR_SIZE==8)
void* _Znwm(size_t n) MI_FORWARD1(mi_new,n) // new 64-bit
void* _Znam(size_t n) MI_FORWARD1(mi_new,n) // new[] 64-bit
void* _ZnwmRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnamRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnwmSt11align_val_t(size_t n, size_t al) MI_FORWARD2(mi_new_aligned, n, al)
void* _ZnamSt11align_val_t(size_t n, size_t al) MI_FORWARD2(mi_new_aligned, n, al)
void* _ZnwmRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnamRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnwmSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnamSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnwmSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnamSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
#elif (MI_INTPTR_SIZE==4)
void* _Znwj(size_t n) MI_FORWARD1(mi_new,n) // new 64-bit
void* _Znaj(size_t n) MI_FORWARD1(mi_new,n) // new[] 64-bit
void* _ZnwjRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnajRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnwjSt11align_val_t(size_t n, size_t al) MI_FORWARD2(mi_new_aligned, n, al)
void* _ZnajSt11align_val_t(size_t n, size_t al) MI_FORWARD2(mi_new_aligned, n, al)
void* _ZnwjRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnajRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { UNUSED(tag); return mi_new_nothrow(n); }
void* _ZnwjSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnajSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnwjSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
void* _ZnajSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
#else
#error "define overloads for new/delete for this platform (just for performance, can be skipped)"
#error "define overloads for new/delete for this platform (just for performance, can be skipped)"
#endif
#endif // __cplusplus
// ------------------------------------------------------
// Further Posix & Unix functions definitions
// ------------------------------------------------------
#ifdef __cplusplus
extern "C" {
#endif
// ------------------------------------------------------
// Posix & Unix functions definitions
// ------------------------------------------------------
#ifndef MI_OSX_IS_INTERPOSED
// Forward Posix/Unix calls as well
void* reallocf(void* p, size_t newsize) MI_FORWARD2(mi_reallocf,p,newsize)
size_t malloc_size(const void* p) MI_FORWARD1(mi_usable_size,p)
#if !defined(__ANDROID__) && !defined(__FreeBSD__)
size_t malloc_usable_size(void *p) MI_FORWARD1(mi_usable_size,p)
#else
size_t malloc_usable_size(const void *p) MI_FORWARD1(mi_usable_size,p)
#endif
void cfree(void* p) MI_FORWARD0(mi_free, p)
void* reallocf(void* p, size_t newsize) MI_FORWARD2(mi_reallocf,p,newsize)
size_t malloc_size(const void* p) MI_FORWARD1(mi_usable_size,p)
#if !defined(__ANDROID__)
size_t malloc_usable_size(void *p) MI_FORWARD1(mi_usable_size,p)
#else
size_t malloc_usable_size(const void *p) MI_FORWARD1(mi_usable_size,p)
// No forwarding here due to aliasing/name mangling issues
void* valloc(size_t size) { return mi_valloc(size); }
void vfree(void* p) { mi_free(p); }
size_t malloc_good_size(size_t size) { return mi_malloc_good_size(size); }
int posix_memalign(void** p, size_t alignment, size_t size) { return mi_posix_memalign(p, alignment, size); }
// `aligned_alloc` is only available when __USE_ISOC11 is defined.
// Note: Conda has a custom glibc where `aligned_alloc` is declared `static inline` and we cannot
// override it, but both _ISOC11_SOURCE and __USE_ISOC11 are undefined in Conda GCC7 or GCC9.
// Fortunately, in the case where `aligned_alloc` is declared as `static inline` it
// uses internally `memalign`, `posix_memalign`, or `_aligned_malloc` so we can avoid overriding it ourselves.
#if __USE_ISOC11
void* aligned_alloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
#endif
#endif
// no forwarding here due to aliasing/name mangling issues
void* valloc(size_t size) { return mi_valloc(size); }
void* pvalloc(size_t size) { return mi_pvalloc(size); }
void* reallocarray(void* p, size_t count, size_t size) { return mi_reallocarray(p, count, size); }
void* memalign(size_t alignment, size_t size) { return mi_memalign(alignment, size); }
int posix_memalign(void** p, size_t alignment, size_t size) { return mi_posix_memalign(p, alignment, size); }
void* _aligned_malloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
// `aligned_alloc` is only available when __USE_ISOC11 is defined.
// Note: Conda has a custom glibc where `aligned_alloc` is declared `static inline` and we cannot
// override it, but both _ISOC11_SOURCE and __USE_ISOC11 are undefined in Conda GCC7 or GCC9.
// Fortunately, in the case where `aligned_alloc` is declared as `static inline` it
// uses internally `memalign`, `posix_memalign`, or `_aligned_malloc` so we can avoid overriding it ourselves.
#if __USE_ISOC11
void* aligned_alloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
#endif
void cfree(void* p) { mi_free(p); }
void* pvalloc(size_t size) { return mi_pvalloc(size); }
void* reallocarray(void* p, size_t count, size_t size) { return mi_reallocarray(p, count, size); }
int reallocarr(void* p, size_t count, size_t size) { return mi_reallocarr(p, count, size); }
void* memalign(size_t alignment, size_t size) { return mi_memalign(alignment, size); }
void* _aligned_malloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
#if defined(__GLIBC__) && defined(__linux__)
// forward __libc interface (needed for glibc-based Linux distributions)
void* __libc_malloc(size_t size) MI_FORWARD1(mi_malloc,size)
void* __libc_calloc(size_t count, size_t size) MI_FORWARD2(mi_calloc,count,size)
void* __libc_realloc(void* p, size_t size) MI_FORWARD2(mi_realloc,p,size)
void __libc_free(void* p) MI_FORWARD0(mi_free,p)
void __libc_cfree(void* p) MI_FORWARD0(mi_free,p)
void* __libc_malloc(size_t size) MI_FORWARD1(mi_malloc,size)
void* __libc_calloc(size_t count, size_t size) MI_FORWARD2(mi_calloc,count,size)
void* __libc_realloc(void* p, size_t size) MI_FORWARD2(mi_realloc,p,size)
void __libc_free(void* p) MI_FORWARD0(mi_free,p)
void __libc_cfree(void* p) MI_FORWARD0(mi_free,p)
void* __libc_valloc(size_t size) { return mi_valloc(size); }
void* __libc_pvalloc(size_t size) { return mi_pvalloc(size); }
void* __libc_memalign(size_t alignment, size_t size) { return mi_memalign(alignment,size); }
void* __libc_valloc(size_t size) { return mi_valloc(size); }
void* __libc_pvalloc(size_t size) { return mi_pvalloc(size); }
void* __libc_memalign(size_t alignment, size_t size) { return mi_memalign(alignment,size); }
int __posix_memalign(void** p, size_t alignment, size_t size) { return mi_posix_memalign(p,alignment,size); }
#endif

20
src/alloc-posix.c
View file

@ -33,13 +33,19 @@ terms of the MIT license. A copy of the license can be found in the file
size_t mi_malloc_size(const void* p) mi_attr_noexcept {
//if (!mi_is_in_heap_region(p)) return 0;
return mi_usable_size(p);
}
size_t mi_malloc_usable_size(const void *p) mi_attr_noexcept {
//if (!mi_is_in_heap_region(p)) return 0;
return mi_usable_size(p);
}
size_t mi_malloc_good_size(size_t size) mi_attr_noexcept {
return mi_good_size(size);
}
void mi_cfree(void* p) mi_attr_noexcept {
if (mi_is_in_heap_region(p)) {
mi_free(p);
@ -86,13 +92,23 @@ mi_decl_restrict void* mi_aligned_alloc(size_t alignment, size_t size) mi_attr_n
void* mi_reallocarray( void* p, size_t count, size_t size ) mi_attr_noexcept { // BSD
void* newp = mi_reallocn(p,count,size);
if (newp==NULL) errno = ENOMEM;
if (newp==NULL) { errno = ENOMEM; }
return newp;
}
int mi_reallocarr( void* p, size_t count, size_t size ) mi_attr_noexcept { // NetBSD
mi_assert(p != NULL);
if (p == NULL) return EINVAL; // should we set errno as well?
void** op = (void**)p;
void* newp = mi_reallocarray(*op, count, size);
if (mi_unlikely(newp == NULL)) return errno;
*op = newp;
return 0;
}
void* mi__expand(void* p, size_t newsize) mi_attr_noexcept { // Microsoft
void* res = mi_expand(p, newsize);
if (res == NULL) errno = ENOMEM;
if (res == NULL) { errno = ENOMEM; }
return res;
}

65
src/alloc.c
View file

@ -4,6 +4,10 @@ This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
#ifndef _DEFAULT_SOURCE
#define _DEFAULT_SOURCE // for realpath() on Linux
#endif
#include "mimalloc.h"
#include "mimalloc-internal.h"
#include "mimalloc-atomic.h"
@ -119,7 +123,7 @@ extern inline mi_decl_restrict void* mi_malloc(size_t size) mi_attr_noexcept {
void _mi_block_zero_init(const mi_page_t* page, void* p, size_t size) {
// note: we need to initialize the whole usable block size to zero, not just the requested size,
// or the recalloc/rezalloc functions cannot safely expand in place (see issue #63)
UNUSED(size);
MI_UNUSED(size);
mi_assert_internal(p != NULL);
mi_assert_internal(mi_usable_size(p) >= size); // size can be zero
mi_assert_internal(_mi_ptr_page(p)==page);
@ -201,8 +205,8 @@ static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block
}
#else
static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
UNUSED(page);
UNUSED(block);
MI_UNUSED(page);
MI_UNUSED(block);
return false;
}
#endif
@ -274,19 +278,19 @@ static void mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, co
}
#else
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
UNUSED(page);
UNUSED(block);
MI_UNUSED(page);
MI_UNUSED(block);
}
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
UNUSED(block);
MI_UNUSED(block);
return mi_page_usable_block_size(page);
}
static void mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
UNUSED(page);
UNUSED(block);
UNUSED(min_size);
MI_UNUSED(page);
MI_UNUSED(block);
MI_UNUSED(min_size);
}
#endif
@ -294,7 +298,7 @@ static void mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, co
#if (MI_STAT>0)
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
#if (MI_STAT < 2)
UNUSED(block);
MI_UNUSED(block);
#endif
mi_heap_t* const heap = mi_heap_get_default();
const size_t bsize = mi_page_usable_block_size(page);
@ -311,7 +315,7 @@ static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
}
#else
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
UNUSED(page); UNUSED(block);
MI_UNUSED(page); MI_UNUSED(block);
}
#endif
@ -329,7 +333,7 @@ static void mi_stat_huge_free(const mi_page_t* page) {
}
#else
static void mi_stat_huge_free(const mi_page_t* page) {
UNUSED(page);
MI_UNUSED(page);
}
#endif
@ -431,7 +435,7 @@ mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* p
}
static void mi_decl_noinline mi_free_generic(const mi_segment_t* segment, bool local, void* p) {
static void mi_decl_noinline mi_free_generic(const mi_segment_t* segment, bool local, void* p) mi_attr_noexcept {
mi_page_t* const page = _mi_segment_page_of(segment, p);
mi_block_t* const block = (mi_page_has_aligned(page) ? _mi_page_ptr_unalign(segment, page, p) : (mi_block_t*)p);
mi_stat_free(page, block);
@ -443,7 +447,7 @@ static void mi_decl_noinline mi_free_generic(const mi_segment_t* segment, bool l
// (and secure mode) if this was a valid pointer.
static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* msg)
{
UNUSED(msg);
MI_UNUSED(msg);
#if (MI_DEBUG>0)
if (mi_unlikely(((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0)) {
_mi_error_message(EINVAL, "%s: invalid (unaligned) pointer: %p\n", msg, p);
@ -465,24 +469,23 @@ static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* ms
#endif
#if (MI_DEBUG>0 || MI_SECURE>=4)
if (mi_unlikely(_mi_ptr_cookie(segment) != segment->cookie)) {
_mi_error_message(EINVAL, "%s: pointer does not point to a valid heap space: %p\n", p);
_mi_error_message(EINVAL, "%s: pointer does not point to a valid heap space: %p\n", msg, p);
}
#endif
return segment;
}
// Free a block
// Free a block
void mi_free(void* p) mi_attr_noexcept
{
const mi_segment_t* const segment = mi_checked_ptr_segment(p,"mi_free");
if (mi_unlikely(segment == NULL)) return;
const uintptr_t tid = _mi_thread_id();
mi_threadid_t tid = _mi_thread_id();
mi_page_t* const page = _mi_segment_page_of(segment, p);
mi_block_t* const block = (mi_block_t*)p;
if (mi_likely(tid == segment->thread_id && page->flags.full_aligned == 0)) { // the thread id matches and it is not a full page, nor has aligned blocks
if (mi_likely(tid == mi_atomic_load_relaxed(&segment->thread_id) && page->flags.full_aligned == 0)) { // the thread id matches and it is not a full page, nor has aligned blocks
// local, and not full or aligned
if (mi_unlikely(mi_check_is_double_free(page,block))) return;
mi_check_padding(page, block);
@ -570,19 +573,19 @@ void* _mi_externs[] = {
// ------------------------------------------------------
void mi_free_size(void* p, size_t size) mi_attr_noexcept {
UNUSED_RELEASE(size);
MI_UNUSED_RELEASE(size);
mi_assert(p == NULL || size <= _mi_usable_size(p,"mi_free_size"));
mi_free(p);
}
void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept {
UNUSED_RELEASE(alignment);
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free_size(p,size);
}
void mi_free_aligned(void* p, size_t alignment) mi_attr_noexcept {
UNUSED_RELEASE(alignment);
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free(p);
}
@ -747,7 +750,7 @@ mi_decl_restrict char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char
}
#else
#include <unistd.h> // pathconf
static size_t mi_path_max() {
static size_t mi_path_max(void) {
static size_t path_max = 0;
if (path_max <= 0) {
long m = pathconf("/",_PC_PATH_MAX);
@ -798,7 +801,10 @@ static bool mi_try_new_handler(bool nothrow) {
std::set_new_handler(h);
#endif
if (h==NULL) {
if (!nothrow) throw std::bad_alloc();
_mi_error_message(ENOMEM, "out of memory in 'new'");
if (!nothrow) {
throw std::bad_alloc();
}
return false;
}
else {
@ -807,13 +813,13 @@ static bool mi_try_new_handler(bool nothrow) {
}
}
#else
typedef void (*std_new_handler_t)();
typedef void (*std_new_handler_t)(void);
#if (defined(__GNUC__) || defined(__clang__))
std_new_handler_t __attribute((weak)) _ZSt15get_new_handlerv() {
std_new_handler_t __attribute((weak)) _ZSt15get_new_handlerv(void) {
return NULL;
}
static std_new_handler_t mi_get_new_handler() {
static std_new_handler_t mi_get_new_handler(void) {
return _ZSt15get_new_handlerv();
}
#else
@ -826,7 +832,10 @@ static std_new_handler_t mi_get_new_handler() {
static bool mi_try_new_handler(bool nothrow) {
std_new_handler_t h = mi_get_new_handler();
if (h==NULL) {
if (!nothrow) exit(ENOMEM); // cannot throw in plain C, use exit as we are out of memory anyway.
_mi_error_message(ENOMEM, "out of memory in 'new'");
if (!nothrow) {
abort(); // cannot throw in plain C, use abort
}
return false;
}
else {

45
src/arena.c
View file

@ -20,7 +20,7 @@ which is sometimes needed for embedded devices or shared memory for example.
The arena allocation needs to be thread safe and we use an atomic
bitmap to allocate. The current implementation of the bitmap can
only do this within a field (`uintptr_t`) so we can allocate at most
only do this within a field (`size_t`) so we can allocate at most
blocks of 2GiB (64*32MiB) and no object can cross the boundary. This
can lead to fragmentation but fortunately most objects will be regions
of 256MiB in practice.
@ -62,18 +62,18 @@ typedef struct mi_arena_s {
size_t field_count; // number of bitmap fields (where `field_count * MI_BITMAP_FIELD_BITS >= block_count`)
int numa_node; // associated NUMA node
bool is_zero_init; // is the arena zero initialized?
bool is_committed; // is the memory fully committed? (if so, block_committed == NULL)
bool allow_decommit; // is decommit allowed? if true, is_large should be false and blocks_committed != NULL
bool is_large; // large- or huge OS pages (always committed)
_Atomic(uintptr_t) search_idx; // optimization to start the search for free blocks
_Atomic(size_t) search_idx; // optimization to start the search for free blocks
mi_bitmap_field_t* blocks_dirty; // are the blocks potentially non-zero?
mi_bitmap_field_t* blocks_committed; // if `!is_committed`, are the blocks committed?
mi_bitmap_field_t* blocks_committed; // are the blocks committed? (can be NULL for memory that cannot be decommitted)
mi_bitmap_field_t blocks_inuse[1]; // in-place bitmap of in-use blocks (of size `field_count`)
} mi_arena_t;
// The available arenas
static mi_decl_cache_align _Atomic(mi_arena_t*) mi_arenas[MI_MAX_ARENAS];
static mi_decl_cache_align _Atomic(uintptr_t) mi_arena_count; // = 0
static mi_decl_cache_align _Atomic(size_t) mi_arena_count; // = 0
/* -----------------------------------------------------------
@ -129,8 +129,8 @@ static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t n
*memid = mi_arena_id_create(arena_index, bitmap_index);
*is_zero = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
*large = arena->is_large;
*is_pinned = (arena->is_large || arena->is_committed);
if (arena->is_committed) {
*is_pinned = (arena->is_large || !arena->allow_decommit);
if (arena->blocks_committed == NULL) {
// always committed
*commit = true;
}
@ -245,12 +245,13 @@ void _mi_arena_free(void* p, size_t size, size_t memid, bool all_committed, mi_s
return;
}
// potentially decommit
if (arena->is_committed) {
mi_assert_internal(all_committed);
if (!arena->allow_decommit || arena->blocks_committed == NULL) {
mi_assert_internal(all_committed); // note: may be not true as we may "pretend" to be not committed (in segment.c)
}
else {
mi_assert_internal(arena->blocks_committed != NULL);
_mi_os_decommit(p, blocks * MI_ARENA_BLOCK_SIZE, stats); // ok if this fails
// todo: use reset instead of decommit on windows?
_mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
}
// and make it available to others again
@ -271,7 +272,7 @@ static bool mi_arena_add(mi_arena_t* arena) {
mi_assert_internal((uintptr_t)mi_atomic_load_ptr_relaxed(uint8_t,&arena->start) % MI_SEGMENT_ALIGN == 0);
mi_assert_internal(arena->block_count > 0);
uintptr_t i = mi_atomic_increment_acq_rel(&mi_arena_count);
size_t i = mi_atomic_increment_acq_rel(&mi_arena_count);
if (i >= MI_MAX_ARENAS) {
mi_atomic_decrement_acq_rel(&mi_arena_count);
return false;
@ -282,12 +283,14 @@ static bool mi_arena_add(mi_arena_t* arena) {
bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept
{
if (size < MI_ARENA_BLOCK_SIZE) return false;
if (is_large) {
mi_assert_internal(is_committed);
is_committed = true;
}
const size_t bcount = mi_block_count_of_size(size);
const size_t bcount = size / MI_ARENA_BLOCK_SIZE;
const size_t fields = _mi_divide_up(bcount, MI_BITMAP_FIELD_BITS);
const size_t bitmaps = (is_committed ? 2 : 3);
const size_t asize = sizeof(mi_arena_t) + (bitmaps*fields*sizeof(mi_bitmap_field_t));
@ -300,12 +303,16 @@ bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_la
arena->numa_node = numa_node; // TODO: or get the current numa node if -1? (now it allows anyone to allocate on -1)
arena->is_large = is_large;
arena->is_zero_init = is_zero;
arena->is_committed = is_committed;
arena->allow_decommit = !is_large && !is_committed; // only allow decommit for initially uncommitted memory
arena->search_idx = 0;
arena->blocks_dirty = &arena->blocks_inuse[fields]; // just after inuse bitmap
arena->blocks_committed = (is_committed ? NULL : &arena->blocks_inuse[2*fields]); // just after dirty bitmap
arena->blocks_committed = (!arena->allow_decommit ? NULL : &arena->blocks_inuse[2*fields]); // just after dirty bitmap
// the bitmaps are already zero initialized due to os_alloc
// just claim leftover blocks if needed
// initialize committed bitmap?
if (arena->blocks_committed != NULL && is_committed) {
memset((void*)arena->blocks_committed, 0xFF, fields*sizeof(mi_bitmap_field_t)); // cast to void* to avoid atomic warning
}
// and claim leftover blocks if needed (so we never allocate there)
ptrdiff_t post = (fields * MI_BITMAP_FIELD_BITS) - bcount;
mi_assert_internal(post >= 0);
if (post > 0) {
@ -321,7 +328,7 @@ bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_la
// Reserve a range of regular OS memory
int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept
{
size = _mi_os_good_alloc_size(size);
size = _mi_align_up(size, MI_ARENA_BLOCK_SIZE); // at least one block
bool large = allow_large;
void* start = _mi_os_alloc_aligned(size, MI_SEGMENT_ALIGN, commit, &large, &_mi_stats_main);
if (start==NULL) return ENOMEM;
@ -330,7 +337,7 @@ int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noe
_mi_verbose_message("failed to reserve %zu k memory\n", _mi_divide_up(size,1024));
return ENOMEM;
}
_mi_verbose_message("reserved %zu kb memory%s\n", _mi_divide_up(size,1024), large ? " (in large os pages)" : "");
_mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size,1024), large ? " (in large os pages)" : "");
return 0;
}
@ -347,10 +354,10 @@ int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msec
size_t pages_reserved = 0;
void* p = _mi_os_alloc_huge_os_pages(pages, numa_node, timeout_msecs, &pages_reserved, &hsize);
if (p==NULL || pages_reserved==0) {
_mi_warning_message("failed to reserve %zu gb huge pages\n", pages);
_mi_warning_message("failed to reserve %zu GiB huge pages\n", pages);
return ENOMEM;
}
_mi_verbose_message("numa node %i: reserved %zu gb huge pages (of the %zu gb requested)\n", numa_node, pages_reserved, pages);
_mi_verbose_message("numa node %i: reserved %zu GiB huge pages (of the %zu GiB requested)\n", numa_node, pages_reserved, pages);
if (!mi_manage_os_memory(p, hsize, true, true, true, numa_node)) {
_mi_os_free_huge_pages(p, hsize, &_mi_stats_main);
@ -389,7 +396,7 @@ int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t
}
int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept {
UNUSED(max_secs);
MI_UNUSED(max_secs);
_mi_warning_message("mi_reserve_huge_os_pages is deprecated: use mi_reserve_huge_os_pages_interleave/at instead\n");
if (pages_reserved != NULL) *pages_reserved = 0;
int err = mi_reserve_huge_os_pages_interleave(pages, 0, (size_t)(max_secs * 1000.0));

90
src/bitmap.c
View file

@ -7,7 +7,7 @@ terms of the MIT license. A copy of the license can be found in the file
/* ----------------------------------------------------------------------------
Concurrent bitmap that can set/reset sequences of bits atomically,
represeted as an array of fields where each field is a machine word (`uintptr_t`)
represeted as an array of fields where each field is a machine word (`size_t`)
There are two api's; the standard one cannot have sequences that cross
between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
@ -26,12 +26,12 @@ between the fields. (This is used in arena allocation)
----------------------------------------------------------- */
// The bit mask for a given number of blocks at a specified bit index.
static inline uintptr_t mi_bitmap_mask_(size_t count, size_t bitidx) {
static inline size_t mi_bitmap_mask_(size_t count, size_t bitidx) {
mi_assert_internal(count + bitidx <= MI_BITMAP_FIELD_BITS);
mi_assert_internal(count > 0);
if (count >= MI_BITMAP_FIELD_BITS) return MI_BITMAP_FIELD_FULL;
if (count == 0) return 0;
return ((((uintptr_t)1 << count) - 1) << bitidx);
return ((((size_t)1 << count) - 1) << bitidx);
}
@ -46,27 +46,27 @@ bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_
{
mi_assert_internal(bitmap_idx != NULL);
mi_assert_internal(count <= MI_BITMAP_FIELD_BITS);
_Atomic(uintptr_t)* field = &bitmap[idx];
uintptr_t map = mi_atomic_load_relaxed(field);
mi_bitmap_field_t* field = &bitmap[idx];
size_t map = mi_atomic_load_relaxed(field);
if (map==MI_BITMAP_FIELD_FULL) return false; // short cut
// search for 0-bit sequence of length count
const uintptr_t mask = mi_bitmap_mask_(count, 0);
const size_t bitidx_max = MI_BITMAP_FIELD_BITS - count;
const size_t mask = mi_bitmap_mask_(count, 0);
const size_t bitidx_max = MI_BITMAP_FIELD_BITS - count;
#ifdef MI_HAVE_FAST_BITSCAN
size_t bitidx = mi_ctz(~map); // quickly find the first zero bit if possible
#else
size_t bitidx = 0; // otherwise start at 0
#endif
uintptr_t m = (mask << bitidx); // invariant: m == mask shifted by bitidx
size_t m = (mask << bitidx); // invariant: m == mask shifted by bitidx
// scan linearly for a free range of zero bits
while (bitidx <= bitidx_max) {
const uintptr_t mapm = map & m;
const size_t mapm = map & m;
if (mapm == 0) { // are the mask bits free at bitidx?
mi_assert_internal((m >> bitidx) == mask); // no overflow?
const uintptr_t newmap = map | m;
const size_t newmap = map | m;
mi_assert_internal((newmap^map) >> bitidx == mask);
if (!mi_atomic_cas_weak_acq_rel(field, &map, newmap)) { // TODO: use strong cas here?
// no success, another thread claimed concurrently.. keep going (with updated `map`)
@ -121,10 +121,10 @@ bool _mi_bitmap_try_find_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, c
bool mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
const size_t idx = mi_bitmap_index_field(bitmap_idx);
const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); UNUSED(bitmap_fields);
const size_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
// mi_assert_internal((bitmap[idx] & mask) == mask);
uintptr_t prev = mi_atomic_and_acq_rel(&bitmap[idx], ~mask);
size_t prev = mi_atomic_and_acq_rel(&bitmap[idx], ~mask);
return ((prev & mask) == mask);
}
@ -134,10 +134,10 @@ bool mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, m
bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero) {
const size_t idx = mi_bitmap_index_field(bitmap_idx);
const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); UNUSED(bitmap_fields);
const size_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
//mi_assert_internal(any_zero != NULL || (bitmap[idx] & mask) == 0);
uintptr_t prev = mi_atomic_or_acq_rel(&bitmap[idx], mask);
size_t prev = mi_atomic_or_acq_rel(&bitmap[idx], mask);
if (any_zero != NULL) *any_zero = ((prev & mask) != mask);
return ((prev & mask) == 0);
}
@ -146,9 +146,9 @@ bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi
static bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_ones) {
const size_t idx = mi_bitmap_index_field(bitmap_idx);
const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); UNUSED(bitmap_fields);
uintptr_t field = mi_atomic_load_relaxed(&bitmap[idx]);
const size_t mask = mi_bitmap_mask_(count, bitidx);
mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
size_t field = mi_atomic_load_relaxed(&bitmap[idx]);
if (any_ones != NULL) *any_ones = ((field & mask) != 0);
return ((field & mask) == mask);
}
@ -176,8 +176,8 @@ static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bit
mi_assert_internal(bitmap_idx != NULL);
// check initial trailing zeros
_Atomic(uintptr_t)* field = &bitmap[idx];
uintptr_t map = mi_atomic_load_relaxed(field);
mi_bitmap_field_t* field = &bitmap[idx];
size_t map = mi_atomic_load_relaxed(field);
const size_t initial = mi_clz(map); // count of initial zeros starting at idx
mi_assert_internal(initial <= MI_BITMAP_FIELD_BITS);
if (initial == 0) return false;
@ -186,11 +186,11 @@ static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bit
// scan ahead
size_t found = initial;
uintptr_t mask = 0; // mask bits for the final field
size_t mask = 0; // mask bits for the final field
while(found < count) {
field++;
map = mi_atomic_load_relaxed(field);
const uintptr_t mask_bits = (found + MI_BITMAP_FIELD_BITS <= count ? MI_BITMAP_FIELD_BITS : (count - found));
const size_t mask_bits = (found + MI_BITMAP_FIELD_BITS <= count ? MI_BITMAP_FIELD_BITS : (count - found));
mask = mi_bitmap_mask_(mask_bits, 0);
if ((map & mask) != 0) return false;
found += mask_bits;
@ -199,13 +199,13 @@ static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bit
// found range of zeros up to the final field; mask contains mask in the final field
// now claim it atomically
_Atomic(uintptr_t)* const final_field = field;
const uintptr_t final_mask = mask;
_Atomic(uintptr_t)* const initial_field = &bitmap[idx];
const uintptr_t initial_mask = mi_bitmap_mask_(initial, MI_BITMAP_FIELD_BITS - initial);
mi_bitmap_field_t* const final_field = field;
const size_t final_mask = mask;
mi_bitmap_field_t* const initial_field = &bitmap[idx];
const size_t initial_mask = mi_bitmap_mask_(initial, MI_BITMAP_FIELD_BITS - initial);
// initial field
uintptr_t newmap;
size_t newmap;
field = initial_field;
map = mi_atomic_load_relaxed(field);
do {
@ -280,8 +280,8 @@ bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitm
}
// Helper for masks across fields; returns the mid count, post_mask may be 0
static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, uintptr_t* pre_mask, uintptr_t* mid_mask, uintptr_t* post_mask) {
UNUSED_RELEASE(bitmap_fields);
static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, size_t* pre_mask, size_t* mid_mask, size_t* post_mask) {
MI_UNUSED_RELEASE(bitmap_fields);
const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
if (mi_likely(bitidx + count <= MI_BITMAP_FIELD_BITS)) {
*pre_mask = mi_bitmap_mask_(count, bitidx);
@ -308,13 +308,13 @@ static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_
// Returns `true` if all `count` bits were 1 previously.
bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
size_t idx = mi_bitmap_index_field(bitmap_idx);
uintptr_t pre_mask;
uintptr_t mid_mask;
uintptr_t post_mask;
size_t pre_mask;
size_t mid_mask;
size_t post_mask;
size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
bool all_one = true;
_Atomic(uintptr_t)*field = &bitmap[idx];
uintptr_t prev = mi_atomic_and_acq_rel(field++, ~pre_mask);
mi_bitmap_field_t* field = &bitmap[idx];
size_t prev = mi_atomic_and_acq_rel(field++, ~pre_mask);
if ((prev & pre_mask) != pre_mask) all_one = false;
while(mid_count-- > 0) {
prev = mi_atomic_and_acq_rel(field++, ~mid_mask);
@ -331,14 +331,14 @@ bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t
// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
size_t idx = mi_bitmap_index_field(bitmap_idx);
uintptr_t pre_mask;
uintptr_t mid_mask;
uintptr_t post_mask;
size_t pre_mask;
size_t mid_mask;
size_t post_mask;
size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
bool all_zero = true;
bool any_zero = false;
_Atomic(uintptr_t)*field = &bitmap[idx];
uintptr_t prev = mi_atomic_or_acq_rel(field++, pre_mask);
_Atomic(size_t)*field = &bitmap[idx];
size_t prev = mi_atomic_or_acq_rel(field++, pre_mask);
if ((prev & pre_mask) != 0) all_zero = false;
if ((prev & pre_mask) != pre_mask) any_zero = true;
while (mid_count-- > 0) {
@ -360,14 +360,14 @@ bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t co
// `any_ones` is `true` if there was at least one bit set to one.
static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
size_t idx = mi_bitmap_index_field(bitmap_idx);
uintptr_t pre_mask;
uintptr_t mid_mask;
uintptr_t post_mask;
size_t pre_mask;
size_t mid_mask;
size_t post_mask;
size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
bool all_ones = true;
bool any_ones = false;
_Atomic(uintptr_t)* field = &bitmap[idx];
uintptr_t prev = mi_atomic_load_relaxed(field++);
mi_bitmap_field_t* field = &bitmap[idx];
size_t prev = mi_atomic_load_relaxed(field++);
if ((prev & pre_mask) != pre_mask) all_ones = false;
if ((prev & pre_mask) != 0) any_ones = true;
while (mid_count-- > 0) {

10
src/bitmap.h
View file

@ -7,7 +7,7 @@ terms of the MIT license. A copy of the license can be found in the file
/* ----------------------------------------------------------------------------
Concurrent bitmap that can set/reset sequences of bits atomically,
represeted as an array of fields where each field is a machine word (`uintptr_t`)
represeted as an array of fields where each field is a machine word (`size_t`)
There are two api's; the standard one cannot have sequences that cross
between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
@ -24,11 +24,11 @@ between the fields. (This is used in arena allocation)
Bitmap definition
----------------------------------------------------------- */
#define MI_BITMAP_FIELD_BITS (8*MI_INTPTR_SIZE)
#define MI_BITMAP_FIELD_FULL (~((uintptr_t)0)) // all bits set
#define MI_BITMAP_FIELD_BITS (8*MI_SIZE_SIZE)
#define MI_BITMAP_FIELD_FULL (~((size_t)0)) // all bits set
// An atomic bitmap of `uintptr_t` fields
typedef _Atomic(uintptr_t) mi_bitmap_field_t;
// An atomic bitmap of `size_t` fields
typedef _Atomic(size_t) mi_bitmap_field_t;
typedef mi_bitmap_field_t* mi_bitmap_t;
// A bitmap index is the index of the bit in a bitmap.

36
src/heap.c
View file

@ -50,9 +50,9 @@ static bool mi_heap_visit_pages(mi_heap_t* heap, heap_page_visitor_fun* fn, void
#if MI_DEBUG>=2
static bool mi_heap_page_is_valid(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
UNUSED(arg1);
UNUSED(arg2);
UNUSED(pq);
MI_UNUSED(arg1);
MI_UNUSED(arg2);
MI_UNUSED(pq);
mi_assert_internal(mi_page_heap(page) == heap);
mi_segment_t* segment = _mi_page_segment(page);
mi_assert_internal(segment->thread_id == heap->thread_id);
@ -86,8 +86,8 @@ typedef enum mi_collect_e {
static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg_collect, void* arg2 ) {
UNUSED(arg2);
UNUSED(heap);
MI_UNUSED(arg2);
MI_UNUSED(heap);
mi_assert_internal(mi_heap_page_is_valid(heap, pq, page, NULL, NULL));
mi_collect_t collect = *((mi_collect_t*)arg_collect);
_mi_page_free_collect(page, collect >= MI_FORCE);
@ -104,10 +104,10 @@ static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t
}
static bool mi_heap_page_never_delayed_free(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
UNUSED(arg1);
UNUSED(arg2);
UNUSED(heap);
UNUSED(pq);
MI_UNUSED(arg1);
MI_UNUSED(arg2);
MI_UNUSED(heap);
MI_UNUSED(pq);
_mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
return true; // don't break
}
@ -262,10 +262,10 @@ static void mi_heap_free(mi_heap_t* heap) {
----------------------------------------------------------- */
static bool _mi_heap_page_destroy(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
UNUSED(arg1);
UNUSED(arg2);
UNUSED(heap);
UNUSED(pq);
MI_UNUSED(arg1);
MI_UNUSED(arg2);
MI_UNUSED(heap);
MI_UNUSED(pq);
// ensure no more thread_delayed_free will be added
_mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
@ -333,7 +333,7 @@ void mi_heap_destroy(mi_heap_t* heap) {
Safe Heap delete
----------------------------------------------------------- */
// Tranfer the pages from one heap to the other
// Transfer the pages from one heap to the other
static void mi_heap_absorb(mi_heap_t* heap, mi_heap_t* from) {
mi_assert_internal(heap!=NULL);
if (from==NULL || from->page_count == 0) return;
@ -422,8 +422,8 @@ bool mi_heap_contains_block(mi_heap_t* heap, const void* p) {
static bool mi_heap_page_check_owned(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* p, void* vfound) {
UNUSED(heap);
UNUSED(pq);
MI_UNUSED(heap);
MI_UNUSED(pq);
bool* found = (bool*)vfound;
mi_segment_t* segment = _mi_page_segment(page);
void* start = _mi_page_start(segment, page, NULL);
@ -521,8 +521,8 @@ typedef bool (mi_heap_area_visit_fun)(const mi_heap_t* heap, const mi_heap_area_
static bool mi_heap_visit_areas_page(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* vfun, void* arg) {
UNUSED(heap);
UNUSED(pq);
MI_UNUSED(heap);
MI_UNUSED(pq);
mi_heap_area_visit_fun* fun = (mi_heap_area_visit_fun*)vfun;
mi_heap_area_ex_t xarea;
const size_t bsize = mi_page_block_size(page);

33
src/init.c
View file

@ -102,6 +102,7 @@ mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
false
};
// the thread-local default heap for allocation
mi_decl_thread mi_heap_t* _mi_heap_default = (mi_heap_t*)&_mi_heap_empty;
@ -141,7 +142,7 @@ mi_stats_t _mi_stats_main = { MI_STATS_NULL };
static void mi_heap_main_init(void) {
if (_mi_heap_main.cookie == 0) {
_mi_heap_main.thread_id = _mi_thread_id();
_mi_heap_main.cookie = _os_random_weak((uintptr_t)&mi_heap_main_init);
_mi_heap_main.cookie = _mi_os_random_weak((uintptr_t)&mi_heap_main_init);
_mi_random_init(&_mi_heap_main.random);
_mi_heap_main.keys[0] = _mi_heap_random_next(&_mi_heap_main);
_mi_heap_main.keys[1] = _mi_heap_random_next(&_mi_heap_main);
@ -331,6 +332,12 @@ bool _mi_is_main_thread(void) {
return (_mi_heap_main.thread_id==0 || _mi_heap_main.thread_id == _mi_thread_id());
}
static _Atomic(size_t) thread_count = ATOMIC_VAR_INIT(1);
size_t _mi_current_thread_count(void) {
return mi_atomic_load_relaxed(&thread_count);
}
// This is called from the `mi_malloc_generic`
void mi_thread_init(void) mi_attr_noexcept
{
@ -343,6 +350,7 @@ void mi_thread_init(void) mi_attr_noexcept
if (_mi_heap_init()) return; // returns true if already initialized
_mi_stat_increase(&_mi_stats_main.threads, 1);
mi_atomic_increment_relaxed(&thread_count);
//_mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id());
}
@ -351,6 +359,7 @@ void mi_thread_done(void) mi_attr_noexcept {
}
static void _mi_thread_done(mi_heap_t* heap) {
mi_atomic_decrement_relaxed(&thread_count);
_mi_stat_decrease(&_mi_stats_main.threads, 1);
// check thread-id as on Windows shutdown with FLS the main (exit) thread may call this on thread-local heaps...
@ -441,7 +450,7 @@ static void mi_process_load(void) {
mi_heap_main_init();
#if defined(MI_TLS_RECURSE_GUARD)
volatile mi_heap_t* dummy = _mi_heap_default; // access TLS to allocate it before setting tls_initialized to true;
UNUSED(dummy);
MI_UNUSED(dummy);
#endif
os_preloading = false;
atexit(&mi_process_done);
@ -478,10 +487,11 @@ static void mi_detect_cpu_features(void) {
void mi_process_init(void) mi_attr_noexcept {
// ensure we are called once
if (_mi_process_is_initialized) return;
_mi_verbose_message("process init: 0x%zx\n", _mi_thread_id());
_mi_process_is_initialized = true;
mi_process_setup_auto_thread_done();
_mi_verbose_message("process init: 0x%zx\n", _mi_thread_id());
mi_detect_cpu_features();
_mi_os_init();
mi_heap_main_init();
@ -494,11 +504,18 @@ void mi_process_init(void) mi_attr_noexcept {
if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
size_t pages = mi_option_get(mi_option_reserve_huge_os_pages);
mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
long reserve_at = mi_option_get(mi_option_reserve_huge_os_pages_at);
if (reserve_at != -1) {
mi_reserve_huge_os_pages_at(pages, reserve_at, pages*500);
} else {
mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
}
}
if (mi_option_is_enabled(mi_option_reserve_os_memory)) {
long ksize = mi_option_get(mi_option_reserve_os_memory);
if (ksize > 0) mi_reserve_os_memory((size_t)ksize*KiB, true, true);
if (ksize > 0) {
mi_reserve_os_memory((size_t)ksize*MI_KiB, true, true);
}
}
}
@ -536,8 +553,8 @@ static void mi_process_done(void) {
#if defined(_WIN32) && defined(MI_SHARED_LIB)
// Windows DLL: easy to hook into process_init and thread_done
__declspec(dllexport) BOOL WINAPI DllMain(HINSTANCE inst, DWORD reason, LPVOID reserved) {
UNUSED(reserved);
UNUSED(inst);
MI_UNUSED(reserved);
MI_UNUSED(inst);
if (reason==DLL_PROCESS_ATTACH) {
mi_process_load();
}
@ -569,7 +586,7 @@ static void mi_process_done(void) {
return 0;
}
typedef int(*_crt_cb)(void);
#ifdef _M_X64
#if defined(_M_X64) || defined(_M_ARM64)
__pragma(comment(linker, "/include:" "_mi_msvc_initu"))
#pragma section(".CRT$XIU", long, read)
#else

47
src/options.c
View file

@ -19,10 +19,10 @@ terms of the MIT license. A copy of the license can be found in the file
#endif
static uintptr_t mi_max_error_count = 16; // stop outputting errors after this
static uintptr_t mi_max_warning_count = 16; // stop outputting warnings after this
static size_t mi_max_error_count = 16; // stop outputting errors after this
static size_t mi_max_warning_count = 16; // stop outputting warnings after this
static void mi_add_stderr_output();
static void mi_add_stderr_output(void);
int mi_version(void) mi_attr_noexcept {
return MI_MALLOC_VERSION;
@ -76,6 +76,7 @@ static mi_option_desc_t options[_mi_option_last] =
#endif
{ 0, UNINIT, MI_OPTION(large_os_pages) }, // use large OS pages, use only with eager commit to prevent fragmentation of VMA's
{ 0, UNINIT, MI_OPTION(reserve_huge_os_pages) }, // per 1GiB huge pages
{ -1, UNINIT, MI_OPTION(reserve_huge_os_pages_at) }, // reserve huge pages at node N
{ 0, UNINIT, MI_OPTION(reserve_os_memory) },
{ 0, UNINIT, MI_OPTION(segment_cache) }, // cache N segments per thread
{ 1, UNINIT, MI_OPTION(page_reset) }, // reset page memory on free
@ -103,7 +104,7 @@ void _mi_options_init(void) {
mi_add_stderr_output(); // now it safe to use stderr for output
for(int i = 0; i < _mi_option_last; i++ ) {
mi_option_t option = (mi_option_t)i;
long l = mi_option_get(option); UNUSED(l); // initialize
long l = mi_option_get(option); MI_UNUSED(l); // initialize
if (option != mi_option_verbose) {
mi_option_desc_t* desc = &options[option];
_mi_verbose_message("option '%s': %ld\n", desc->name, desc->value);
@ -113,7 +114,7 @@ void _mi_options_init(void) {
mi_max_warning_count = mi_option_get(mi_option_max_warnings);
}
long mi_option_get(mi_option_t option) {
mi_decl_nodiscard long mi_option_get(mi_option_t option) {
mi_assert(option >= 0 && option < _mi_option_last);
mi_option_desc_t* desc = &options[option];
mi_assert(desc->option == option); // index should match the option
@ -139,7 +140,7 @@ void mi_option_set_default(mi_option_t option, long value) {
}
}
bool mi_option_is_enabled(mi_option_t option) {
mi_decl_nodiscard bool mi_option_is_enabled(mi_option_t option) {
return (mi_option_get(option) != 0);
}
@ -161,7 +162,7 @@ void mi_option_disable(mi_option_t option) {
static void mi_out_stderr(const char* msg, void* arg) {
UNUSED(arg);
MI_UNUSED(arg);
#ifdef _WIN32
// on windows with redirection, the C runtime cannot handle locale dependent output
// after the main thread closes so we use direct console output.
@ -176,19 +177,19 @@ static void mi_out_stderr(const char* msg, void* arg) {
// an output function is registered it is called immediately with
// the output up to that point.
#ifndef MI_MAX_DELAY_OUTPUT
#define MI_MAX_DELAY_OUTPUT ((uintptr_t)(32*1024))
#define MI_MAX_DELAY_OUTPUT ((size_t)(32*1024))
#endif
static char out_buf[MI_MAX_DELAY_OUTPUT+1];
static _Atomic(uintptr_t) out_len;
static _Atomic(size_t) out_len;
static void mi_out_buf(const char* msg, void* arg) {
UNUSED(arg);
MI_UNUSED(arg);
if (msg==NULL) return;
if (mi_atomic_load_relaxed(&out_len)>=MI_MAX_DELAY_OUTPUT) return;
size_t n = strlen(msg);
if (n==0) return;
// claim space
uintptr_t start = mi_atomic_add_acq_rel(&out_len, n);
size_t start = mi_atomic_add_acq_rel(&out_len, n);
if (start >= MI_MAX_DELAY_OUTPUT) return;
// check bound
if (start+n >= MI_MAX_DELAY_OUTPUT) {
@ -251,8 +252,8 @@ static void mi_add_stderr_output() {
// --------------------------------------------------------
// Messages, all end up calling `_mi_fputs`.
// --------------------------------------------------------
static _Atomic(uintptr_t) error_count; // = 0; // when >= max_error_count stop emitting errors
static _Atomic(uintptr_t) warning_count; // = 0; // when >= max_warning_count stop emitting warnings
static _Atomic(size_t) error_count; // = 0; // when >= max_error_count stop emitting errors
static _Atomic(size_t) warning_count; // = 0; // when >= max_warning_count stop emitting warnings
// When overriding malloc, we may recurse into mi_vfprintf if an allocation
// inside the C runtime causes another message.
@ -353,7 +354,7 @@ static mi_error_fun* volatile mi_error_handler; // = NULL
static _Atomic(void*) mi_error_arg; // = NULL
static void mi_error_default(int err) {
UNUSED(err);
MI_UNUSED(err);
#if (MI_DEBUG>0)
if (err==EFAULT) {
#ifdef _MSC_VER
@ -409,6 +410,14 @@ static void mi_strlcat(char* dest, const char* src, size_t dest_size) {
dest[dest_size - 1] = 0;
}
#ifdef MI_NO_GETENV
static bool mi_getenv(const char* name, char* result, size_t result_size) {
MI_UNUSED(name);
MI_UNUSED(result);
MI_UNUSED(result_size);
return false;
}
#else
static inline int mi_strnicmp(const char* s, const char* t, size_t n) {
if (n==0) return 0;
for (; *s != 0 && *t != 0 && n > 0; s++, t++, n--) {
@ -416,7 +425,6 @@ static inline int mi_strnicmp(const char* s, const char* t, size_t n) {
}
return (n==0 ? 0 : *s - *t);
}
#if defined _WIN32
// On Windows use GetEnvironmentVariable instead of getenv to work
// reliably even when this is invoked before the C runtime is initialized.
@ -484,7 +492,8 @@ static bool mi_getenv(const char* name, char* result, size_t result_size) {
return false;
}
}
#endif
#endif // !MI_USE_ENVIRON
#endif // !MI_NO_GETENV
static void mi_option_init(mi_option_desc_t* desc) {
// Read option value from the environment
@ -513,9 +522,9 @@ static void mi_option_init(mi_option_desc_t* desc) {
if (desc->option == mi_option_reserve_os_memory) {
// this option is interpreted in KiB to prevent overflow of `long`
if (*end == 'K') { end++; }
else if (*end == 'M') { value *= KiB; end++; }
else if (*end == 'G') { value *= MiB; end++; }
else { value = (value + KiB - 1) / KiB; }
else if (*end == 'M') { value *= MI_KiB; end++; }
else if (*end == 'G') { value *= MI_MiB; end++; }
else { value = (value + MI_KiB - 1) / MI_KiB; }
if (*end == 'B') { end++; }
}
if (*end == 0) {

382
src/os.c
View file

@ -26,16 +26,20 @@ terms of the MIT license. A copy of the license can be found in the file
#pragma warning(disable:4996) // strerror
#endif
#if defined(__wasi__)
#define MI_USE_SBRK
#endif
#if defined(_WIN32)
#include <windows.h>
#elif defined(__wasi__)
// stdlib.h is all we need, and has already been included in mimalloc.h
#include <unistd.h> // sbrk
#else
#include <sys/mman.h> // mmap
#include <unistd.h> // sysconf
#if defined(__linux__)
#include <features.h>
#include <fcntl.h>
#if defined(__GLIBC__)
#include <linux/mman.h> // linux mmap flags
#else
@ -48,9 +52,13 @@ terms of the MIT license. A copy of the license can be found in the file
#include <mach/vm_statistics.h>
#endif
#endif
#if defined(__HAIKU__)
#define madvise posix_madvise
#define MADV_DONTNEED POSIX_MADV_DONTNEED
#if defined(__FreeBSD__) || defined(__DragonFly__)
#include <sys/param.h>
#if __FreeBSD_version >= 1200000
#include <sys/cpuset.h>
#include <sys/domainset.h>
#endif
#include <sys/sysctl.h>
#endif
#endif
@ -89,30 +97,40 @@ static size_t os_alloc_granularity = 4096;
// if non-zero, use large page allocation
static size_t large_os_page_size = 0;
// is memory overcommit allowed?
// set dynamically in _mi_os_init (and if true we use MAP_NORESERVE)
static bool os_overcommit = true;
bool _mi_os_has_overcommit(void) {
return os_overcommit;
}
// OS (small) page size
size_t _mi_os_page_size() {
return os_page_size;
}
// if large OS pages are supported (2 or 4MiB), then return the size, otherwise return the small page size (4KiB)
size_t _mi_os_large_page_size() {
size_t _mi_os_large_page_size(void) {
return (large_os_page_size != 0 ? large_os_page_size : _mi_os_page_size());
}
#if !defined(MI_USE_SBRK) && !defined(__wasi__)
static bool use_large_os_page(size_t size, size_t alignment) {
// if we have access, check the size and alignment requirements
if (large_os_page_size == 0 || !mi_option_is_enabled(mi_option_large_os_pages)) return false;
return ((size % large_os_page_size) == 0 && (alignment % large_os_page_size) == 0);
}
#endif
// round to a good OS allocation size (bounded by max 12.5% waste)
size_t _mi_os_good_alloc_size(size_t size) {
size_t align_size;
if (size < 512*KiB) align_size = _mi_os_page_size();
else if (size < 2*MiB) align_size = 64*KiB;
else if (size < 8*MiB) align_size = 256*KiB;
else if (size < 32*MiB) align_size = 1*MiB;
else align_size = 4*MiB;
if (size < 512*MI_KiB) align_size = _mi_os_page_size();
else if (size < 2*MI_MiB) align_size = 64*MI_KiB;
else if (size < 8*MI_MiB) align_size = 256*MI_KiB;
else if (size < 32*MI_MiB) align_size = 1*MI_MiB;
else align_size = 4*MI_MiB;
if (mi_unlikely(size >= (SIZE_MAX - align_size))) return size; // possible overflow?
return _mi_align_up(size, align_size);
}
@ -175,7 +193,9 @@ static bool mi_win_enable_large_os_pages()
return (ok!=0);
}
void _mi_os_init(void) {
void _mi_os_init(void)
{
os_overcommit = false;
// get the page size
SYSTEM_INFO si;
GetSystemInfo(&si);
@ -210,10 +230,36 @@ void _mi_os_init(void) {
}
#elif defined(__wasi__)
void _mi_os_init() {
os_page_size = 0x10000; // WebAssembly has a fixed page size: 64KB
os_overcommit = false;
os_page_size = 0x10000; // WebAssembly has a fixed page size: 64KiB
os_alloc_granularity = 16;
}
#else // generic unix
static void os_detect_overcommit(void) {
#if defined(__linux__)
int fd = open("/proc/sys/vm/overcommit_memory", O_RDONLY);
if (fd < 0) return;
char buf[128];
ssize_t nread = read(fd, &buf, sizeof(buf));
close(fd);
// <https://www.kernel.org/doc/Documentation/vm/overcommit-accounting>
// 0: heuristic overcommit, 1: always overcommit, 2: never overcommit (ignore NORESERVE)
if (nread >= 1) {
os_overcommit = (buf[0] == '0' || buf[0] == '1');
}
#elif defined(__FreeBSD__)
int val = 0;
size_t olen = sizeof(val);
if (sysctlbyname("vm.overcommit", &val, &olen, NULL, 0) == 0) {
os_overcommit = (val != 0);
}
#else
// default: overcommit is true
#endif
}
void _mi_os_init() {
// get the page size
long result = sysconf(_SC_PAGESIZE);
@ -221,7 +267,8 @@ void _mi_os_init() {
os_page_size = (size_t)result;
os_alloc_granularity = os_page_size;
}
large_os_page_size = 2*MiB; // TODO: can we query the OS for this?
large_os_page_size = 2*MI_MiB; // TODO: can we query the OS for this?
os_detect_overcommit();
}
#endif
@ -236,8 +283,8 @@ static bool mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats
bool err = false;
#if defined(_WIN32)
err = (VirtualFree(addr, 0, MEM_RELEASE) == 0);
#elif defined(__wasi__)
err = 0; // WebAssembly's heap cannot be shrunk
#elif defined(MI_USE_SBRK)
err = 0; // sbrk heap cannot be shrunk
#else
err = (munmap(addr, size) == -1);
#endif
@ -252,22 +299,29 @@ static bool mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats
}
}
#if !defined(MI_USE_SBRK) && !defined(__wasi__)
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size);
#endif
#ifdef _WIN32
static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment, DWORD flags) {
#if (MI_INTPTR_SIZE >= 8)
// on 64-bit systems, try to use the virtual address area after 4TiB for 4MiB aligned allocations
void* hint;
if (addr == NULL && (hint = mi_os_get_aligned_hint(try_alignment,size)) != NULL) {
void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE);
if (p != NULL) return p;
DWORD err = GetLastError();
if (err != ERROR_INVALID_ADDRESS && // If linked with multiple instances, we may have tried to allocate at an already allocated area (#210)
err != ERROR_INVALID_PARAMETER) { // Windows7 instability (#230)
return NULL;
// on 64-bit systems, try to use the virtual address area after 2TiB for 4MiB aligned allocations
if (addr == NULL) {
void* hint = mi_os_get_aligned_hint(try_alignment,size);
if (hint != NULL) {
void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE);
if (p != NULL) return p;
// for robustness always fall through in case of an error
/*
DWORD err = GetLastError();
if (err != ERROR_INVALID_ADDRESS && // If linked with multiple instances, we may have tried to allocate at an already allocated area (#210)
err != ERROR_INVALID_PARAMETER) { // Windows7 instability (#230)
return NULL;
}
*/
_mi_warning_message("unable to allocate hinted aligned OS memory (%zu bytes, error code: %x, address: %p, alignment: %d, flags: %x)\n", size, GetLastError(), hint, try_alignment, flags);
}
// fall through
}
#endif
#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
@ -278,7 +332,10 @@ static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment
MEM_EXTENDED_PARAMETER param = { {0, 0}, {0} };
param.Type = MemExtendedParameterAddressRequirements;
param.Pointer = &reqs;
return (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, &param, 1);
void* p = (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, &param, 1);
if (p != NULL) return p;
_mi_warning_message("unable to allocate aligned OS memory (%zu bytes, error code: %x, address: %p, alignment: %d, flags: %x)\n", size, GetLastError(), addr, try_alignment, flags);
// fall through on error
}
#endif
// last resort
@ -287,11 +344,11 @@ static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment
static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only, bool allow_large, bool* is_large) {
mi_assert_internal(!(large_only && !allow_large));
static _Atomic(uintptr_t) large_page_try_ok; // = 0;
static _Atomic(size_t) large_page_try_ok; // = 0;
void* p = NULL;
if ((large_only || use_large_os_page(size, try_alignment))
&& allow_large && (flags&MEM_COMMIT)!=0 && (flags&MEM_RESERVE)!=0) {
uintptr_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
if (!large_only && try_ok > 0) {
// if a large page allocation fails, it seems the calls to VirtualAlloc get very expensive.
// therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times.
@ -318,7 +375,32 @@ static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment,
return p;
}
#elif defined(__wasi__)
#elif defined(MI_USE_SBRK)
#define MI_SBRK_FAIL ((void*)(-1))
static void* mi_sbrk_heap_grow(size_t size, size_t try_alignment) {
void* pbase0 = sbrk(0);
if (pbase0 == MI_SBRK_FAIL) {
_mi_warning_message("unable to allocate sbrk() OS memory (%zu bytes)\n", size);
errno = ENOMEM;
return NULL;
}
uintptr_t base = (uintptr_t)pbase0;
uintptr_t aligned_base = _mi_align_up(base, (uintptr_t) try_alignment);
size_t alloc_size = _mi_align_up( aligned_base - base + size, _mi_os_page_size());
mi_assert(alloc_size >= size && (alloc_size % _mi_os_page_size()) == 0);
if (alloc_size < size) return NULL;
void* pbase1 = sbrk(alloc_size);
if (pbase1 == MI_SBRK_FAIL) {
_mi_warning_message("unable to allocate sbrk() OS memory (%zu bytes, %zu requested)\n", size, alloc_size);
errno = ENOMEM;
return NULL;
}
mi_assert(pbase0 == pbase1);
return (void*)aligned_base;
}
#elif defined(__wasi__)
// currently unused as we use sbrk() on wasm
static void* mi_wasm_heap_grow(size_t size, size_t try_alignment) {
uintptr_t base = __builtin_wasm_memory_size(0) * _mi_os_page_size();
uintptr_t aligned_base = _mi_align_up(base, (uintptr_t) try_alignment);
@ -326,31 +408,50 @@ static void* mi_wasm_heap_grow(size_t size, size_t try_alignment) {
mi_assert(alloc_size >= size && (alloc_size % _mi_os_page_size()) == 0);
if (alloc_size < size) return NULL;
if (__builtin_wasm_memory_grow(0, alloc_size / _mi_os_page_size()) == SIZE_MAX) {
_mi_warning_message("unable to allocate wasm_memory_grow() OS memory (%zu bytes, %zu requested)\n", size, alloc_size);
errno = ENOMEM;
return NULL;
}
return (void*)aligned_base;
}
#else
#define MI_OS_USE_MMAP
static void* mi_unix_mmapx(void* addr, size_t size, size_t try_alignment, int protect_flags, int flags, int fd) {
void* p = NULL;
#if (MI_INTPTR_SIZE >= 8) && !defined(MAP_ALIGNED)
// on 64-bit systems, use the virtual address area after 4TiB for 4MiB aligned allocations
void* hint;
if (addr == NULL && (hint = mi_os_get_aligned_hint(try_alignment, size)) != NULL) {
p = mmap(hint,size,protect_flags,flags,fd,0);
if (p==MAP_FAILED) p = NULL; // fall back to regular mmap
MI_UNUSED(try_alignment);
#if defined(MAP_ALIGNED) // BSD
if (addr == NULL && try_alignment > 0 && (try_alignment % _mi_os_page_size()) == 0) {
size_t n = mi_bsr(try_alignment);
if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) { // alignment is a power of 2 and 4096 <= alignment <= 1GiB
flags |= MAP_ALIGNED(n);
void* p = mmap(addr, size, protect_flags, flags | MAP_ALIGNED(n), fd, 0);
if (p!=MAP_FAILED) return p;
// fall back to regular mmap
}
}
#elif defined(MAP_ALIGN) // Solaris
if (addr == NULL && try_alignment > 0 && (try_alignment % _mi_os_page_size()) == 0) {
void* p = mmap(try_alignment, size, protect_flags, flags | MAP_ALIGN, fd, 0);
if (p!=MAP_FAILED) return p;
// fall back to regular mmap
}
#else
UNUSED(try_alignment);
UNUSED(mi_os_get_aligned_hint);
#endif
if (p==NULL) {
p = mmap(addr,size,protect_flags,flags,fd,0);
if (p==MAP_FAILED) p = NULL;
#if (MI_INTPTR_SIZE >= 8) && !defined(MAP_ALIGNED)
// on 64-bit systems, use the virtual address area after 2TiB for 4MiB aligned allocations
if (addr == NULL) {
void* hint = mi_os_get_aligned_hint(try_alignment, size);
if (hint != NULL) {
void* p = mmap(hint, size, protect_flags, flags, fd, 0);
if (p!=MAP_FAILED) return p;
// fall back to regular mmap
}
}
return p;
#endif
// regular mmap
void* p = mmap(addr, size, protect_flags, flags, fd, 0);
if (p!=MAP_FAILED) return p;
// failed to allocate
return NULL;
}
static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int protect_flags, bool large_only, bool allow_large, bool* is_large) {
@ -361,28 +462,24 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
#if !defined(MAP_NORESERVE)
#define MAP_NORESERVE 0
#endif
int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE;
int flags = MAP_PRIVATE | MAP_ANONYMOUS;
int fd = -1;
#if defined(MAP_ALIGNED) // BSD
if (try_alignment > 0) {
size_t n = mi_bsr(try_alignment);
if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) { // alignment is a power of 2 and 4096 <= alignment <= 1GiB
flags |= MAP_ALIGNED(n);
}
}
#endif
if (_mi_os_has_overcommit()) {
flags |= MAP_NORESERVE;
}
#if defined(PROT_MAX)
protect_flags |= PROT_MAX(PROT_READ | PROT_WRITE); // BSD
#endif
#if defined(VM_MAKE_TAG)
// macOS: tracking anonymous page with a specific ID. (All up to 98 are taken officially but LLVM sanitizers had taken 99)
int os_tag = (int)mi_option_get(mi_option_os_tag);
if (os_tag < 100 || os_tag > 255) os_tag = 100;
if (os_tag < 100 || os_tag > 255) { os_tag = 100; }
fd = VM_MAKE_TAG(os_tag);
#endif
// huge page allocation
if ((large_only || use_large_os_page(size, try_alignment)) && allow_large) {
static _Atomic(uintptr_t) large_page_try_ok; // = 0;
uintptr_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
static _Atomic(size_t) large_page_try_ok; // = 0;
size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
if (!large_only && try_ok > 0) {
// If the OS is not configured for large OS pages, or the user does not have
// enough permission, the `mmap` will always fail (but it might also fail for other reasons).
@ -401,7 +498,7 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
#endif
#ifdef MAP_HUGE_1GB
static bool mi_huge_pages_available = true;
if ((size % GiB) == 0 && mi_huge_pages_available) {
if ((size % MI_GiB) == 0 && mi_huge_pages_available) {
lflags |= MAP_HUGE_1GB;
}
else
@ -428,37 +525,39 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
#endif
if (large_only) return p;
if (p == NULL) {
mi_atomic_store_release(&large_page_try_ok, (uintptr_t)10); // on error, don't try again for the next N allocations
mi_atomic_store_release(&large_page_try_ok, (size_t)8); // on error, don't try again for the next N allocations
}
}
}
}
// regular allocation
if (p == NULL) {
*is_large = false;
p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, flags, fd);
#if defined(MADV_HUGEPAGE)
// Many Linux systems don't allow MAP_HUGETLB but they support instead
// transparent huge pages (THP). It is not required to call `madvise` with MADV_HUGE
// though since properly aligned allocations will already use large pages if available
// in that case -- in particular for our large regions (in `memory.c`).
// However, some systems only allow THP if called with explicit `madvise`, so
// when large OS pages are enabled for mimalloc, we call `madvise` anyways.
if (allow_large && use_large_os_page(size, try_alignment)) {
if (madvise(p, size, MADV_HUGEPAGE) == 0) {
*is_large = true; // possibly
};
}
#endif
#if defined(__sun)
if (allow_large && use_large_os_page(size, try_alignment)) {
struct memcntl_mha cmd = {0};
cmd.mha_pagesize = large_os_page_size;
cmd.mha_cmd = MHA_MAPSIZE_VA;
if (memcntl(p, size, MC_HAT_ADVISE, (caddr_t)&cmd, 0, 0) == 0) {
*is_large = true;
if (p != NULL) {
#if defined(MADV_HUGEPAGE)
// Many Linux systems don't allow MAP_HUGETLB but they support instead
// transparent huge pages (THP). Generally, it is not required to call `madvise` with MADV_HUGE
// though since properly aligned allocations will already use large pages if available
// in that case -- in particular for our large regions (in `memory.c`).
// However, some systems only allow THP if called with explicit `madvise`, so
// when large OS pages are enabled for mimalloc, we call `madvise` anyways.
if (allow_large && use_large_os_page(size, try_alignment)) {
if (madvise(p, size, MADV_HUGEPAGE) == 0) {
*is_large = true; // possibly
};
}
#elif defined(__sun)
if (allow_large && use_large_os_page(size, try_alignment)) {
struct memcntl_mha cmd = {0};
cmd.mha_pagesize = large_os_page_size;
cmd.mha_cmd = MHA_MAPSIZE_VA;
if (memcntl(p, size, MC_HAT_ADVISE, (caddr_t)&cmd, 0, 0) == 0) {
*is_large = true;
}
}
#endif
}
#endif
}
if (p == NULL) {
_mi_warning_message("unable to allocate OS memory (%zu bytes, error code: %i, address: %p, large only: %d, allow large: %d)\n", size, errno, addr, large_only, allow_large);
@ -468,8 +567,8 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
#endif
// On 64-bit systems, we can do efficient aligned allocation by using
// the 4TiB to 30TiB area to allocate them.
#if (MI_INTPTR_SIZE >= 8) && (defined(_WIN32) || (defined(MI_OS_USE_MMAP) && !defined(MAP_ALIGNED)))
// the 2TiB to 30TiB area to allocate them.
#if (MI_INTPTR_SIZE >= 8) && (defined(_WIN32) || defined(MI_OS_USE_MMAP))
static mi_decl_cache_align _Atomic(uintptr_t) aligned_base;
// Return a 4MiB aligned address that is probably available.
@ -479,36 +578,38 @@ static mi_decl_cache_align _Atomic(uintptr_t) aligned_base;
// (otherwise an initial large allocation of say 2TiB has a 50% chance to include (known) addresses
// in the middle of the 2TiB - 6TiB address range (see issue #372))
#define KK_HINT_BASE ((uintptr_t)2 << 40) // 2TiB start
#define KK_HINT_AREA ((uintptr_t)4 << 40) // upto 6TiB (since before win8 there is "only" 8TiB available to processes)
#define KK_HINT_MAX ((uintptr_t)30 << 40) // wrap after 30TiB (area after 32TiB is used for huge OS pages)
#define MI_HINT_BASE ((uintptr_t)2 << 40) // 2TiB start
#define MI_HINT_AREA ((uintptr_t)4 << 40) // upto 6TiB (since before win8 there is "only" 8TiB available to processes)
#define MI_HINT_MAX ((uintptr_t)30 << 40) // wrap after 30TiB (area after 32TiB is used for huge OS pages)
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size)
{
if (try_alignment == 0 || try_alignment > MI_SEGMENT_SIZE) return NULL;
if ((size%MI_SEGMENT_SIZE) != 0) return NULL;
if (size > 1*GiB) return NULL; // guarantee the chance of fixed valid address is at most 1/(KK_HINT_AREA / 1<<30) = 1/4096.
size = _mi_align_up(size, MI_SEGMENT_SIZE);
if (size > 1*MI_GiB) return NULL; // guarantee the chance of fixed valid address is at most 1/(MI_HINT_AREA / 1<<30) = 1/4096.
#if (MI_SECURE>0)
size += MI_SEGMENT_SIZE; // put in `MI_SEGMENT_SIZE` virtual gaps between hinted blocks; this splits VLA's but increases guarded areas.
#endif
uintptr_t hint = mi_atomic_add_acq_rel(&aligned_base, size);
if (hint == 0 || hint > KK_HINT_MAX) { // wrap or initialize
uintptr_t init = KK_HINT_BASE;
if (hint == 0 || hint > MI_HINT_MAX) { // wrap or initialize
uintptr_t init = MI_HINT_BASE;
#if (MI_SECURE>0 || MI_DEBUG==0) // security: randomize start of aligned allocations unless in debug mode
uintptr_t r = _mi_heap_random_next(mi_get_default_heap());
init = init + ((MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)) % KK_HINT_AREA); // (randomly 20 bits)*4MiB == 0 to 4TiB
init = init + ((MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)) % MI_HINT_AREA); // (randomly 20 bits)*4MiB == 0 to 4TiB
#endif
uintptr_t expected = hint + size;
mi_atomic_cas_strong_acq_rel(&aligned_base, &expected, init);
hint = mi_atomic_add_acq_rel(&aligned_base, size); // this may still give 0 or > KK_HINT_MAX but that is ok, it is a hint after all
hint = mi_atomic_add_acq_rel(&aligned_base, size); // this may still give 0 or > MI_HINT_MAX but that is ok, it is a hint after all
}
if (hint%try_alignment != 0) return NULL;
return (void*)hint;
}
#elif defined(__wasi__) || defined(MI_USE_SBRK)
// no need for mi_os_get_aligned_hint
#else
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size) {
UNUSED(try_alignment); UNUSED(size);
MI_UNUSED(try_alignment); MI_UNUSED(size);
return NULL;
}
#endif
@ -536,7 +637,12 @@ static void* mi_os_mem_alloc(size_t size, size_t try_alignment, bool commit, boo
int flags = MEM_RESERVE;
if (commit) flags |= MEM_COMMIT;
p = mi_win_virtual_alloc(NULL, size, try_alignment, flags, false, allow_large, is_large);
#elif defined(MI_USE_SBRK)
MI_UNUSED(allow_large);
*is_large = false;
p = mi_sbrk_heap_grow(size, try_alignment);
#elif defined(__wasi__)
MI_UNUSED(allow_large);
*is_large = false;
p = mi_wasm_heap_grow(size, try_alignment);
#else
@ -593,6 +699,10 @@ static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit,
mi_os_mem_free(p, over_size, commit, stats);
void* aligned_p = mi_align_up_ptr(p, alignment);
p = mi_win_virtual_alloc(aligned_p, size, alignment, flags, false, allow_large, is_large);
if (p != NULL) {
_mi_stat_increase(&stats->reserved, size);
if (commit) { _mi_stat_increase(&stats->committed, size); }
}
if (p == aligned_p) break; // success!
if (p != NULL) { // should not happen?
mi_os_mem_free(p, size, commit, stats);
@ -626,7 +736,7 @@ static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit,
----------------------------------------------------------- */
void* _mi_os_alloc(size_t size, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
if (size == 0) return NULL;
size = _mi_os_good_alloc_size(size);
@ -635,7 +745,7 @@ void* _mi_os_alloc(size_t size, mi_stats_t* tld_stats) {
}
void _mi_os_free_ex(void* p, size_t size, bool was_committed, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
if (size == 0 || p == NULL) return;
size = _mi_os_good_alloc_size(size);
@ -648,7 +758,7 @@ void _mi_os_free(void* p, size_t size, mi_stats_t* stats) {
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, mi_stats_t* tld_stats)
{
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
if (size == 0) return NULL;
size = _mi_os_good_alloc_size(size);
alignment = _mi_align_up(alignment, _mi_os_page_size());
@ -699,7 +809,7 @@ static void mi_mprotect_hint(int err) {
" > sudo sysctl -w vm.max_map_count=262144\n");
}
#else
UNUSED(err);
MI_UNUSED(err);
#endif
}
@ -744,16 +854,16 @@ static bool mi_os_commitx(void* addr, size_t size, bool commit, bool conservativ
// for commit, just change the protection
err = mprotect(start, csize, (PROT_READ | PROT_WRITE));
if (err != 0) { err = errno; }
#if defined(MADV_FREE_REUSE)
while ((err = madvise(start, csize, MADV_FREE_REUSE)) != 0 && errno == EAGAIN) { errno = 0; }
#endif
//#if defined(MADV_FREE_REUSE)
// while ((err = madvise(start, csize, MADV_FREE_REUSE)) != 0 && errno == EAGAIN) { errno = 0; }
//#endif
}
#else
err = mprotect(start, csize, (commit ? (PROT_READ | PROT_WRITE) : PROT_NONE));
if (err != 0) { err = errno; }
#endif
if (err != 0) {
_mi_warning_message("%s error: start: %p, csize: 0x%x, err: %i\n", commit ? "commit" : "decommit", start, csize, err);
_mi_warning_message("%s error: start: %p, csize: 0x%zx, err: %i\n", commit ? "commit" : "decommit", start, csize, err);
mi_mprotect_hint(err);
}
mi_assert_internal(err == 0);
@ -761,13 +871,13 @@ static bool mi_os_commitx(void* addr, size_t size, bool commit, bool conservativ
}
bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
return mi_os_commitx(addr, size, true, false /* liberal */, is_zero, stats);
}
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
bool is_zero;
return mi_os_commitx(addr, size, false, true /* conservative */, &is_zero, stats);
@ -808,18 +918,13 @@ static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats)
if (p != start) return false;
#else
#if defined(MADV_FREE)
#if defined(MADV_FREE_REUSABLE)
#define KK_MADV_FREE_INITIAL MADV_FREE_REUSABLE
#else
#define KK_MADV_FREE_INITIAL MADV_FREE
#endif
static _Atomic(uintptr_t) advice = ATOMIC_VAR_INIT(KK_MADV_FREE_INITIAL);
static _Atomic(size_t) advice = ATOMIC_VAR_INIT(MADV_FREE);
int oadvice = (int)mi_atomic_load_relaxed(&advice);
int err;
while ((err = madvise(start, csize, oadvice)) != 0 && errno == EAGAIN) { errno = 0; };
if (err != 0 && errno == EINVAL && oadvice == KK_MADV_FREE_INITIAL) {
// if MADV_FREE/MADV_FREE_REUSABLE is not supported, fall back to MADV_DONTNEED from now on
mi_atomic_store_release(&advice, (uintptr_t)MADV_DONTNEED);
if (err != 0 && errno == EINVAL && oadvice == MADV_FREE) {
// if MADV_FREE is not supported, fall back to MADV_DONTNEED from now on
mi_atomic_store_release(&advice, (size_t)MADV_DONTNEED);
err = madvise(start, csize, MADV_DONTNEED);
}
#elif defined(__wasi__)
@ -828,7 +933,7 @@ static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats)
int err = madvise(start, csize, MADV_DONTNEED);
#endif
if (err != 0) {
_mi_warning_message("madvise reset error: start: %p, csize: 0x%x, errno: %i\n", start, csize, errno);
_mi_warning_message("madvise reset error: start: %p, csize: 0x%zx, errno: %i\n", start, csize, errno);
}
//mi_assert(err == 0);
if (err != 0) return false;
@ -841,7 +946,7 @@ static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats)
// pages and reduce swapping while keeping the memory committed.
// We page align to a conservative area inside the range to reset.
bool _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
if (mi_option_is_enabled(mi_option_reset_decommits)) {
return _mi_os_decommit(addr, size, stats);
@ -852,7 +957,7 @@ bool _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats) {
}
bool _mi_os_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) {
UNUSED(tld_stats);
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
if (mi_option_is_enabled(mi_option_reset_decommits)) {
return mi_os_commit_unreset(addr, size, is_zero, stats); // re-commit it (conservatively!)
@ -887,7 +992,7 @@ static bool mi_os_protectx(void* addr, size_t size, bool protect) {
if (err != 0) { err = errno; }
#endif
if (err != 0) {
_mi_warning_message("mprotect error: start: %p, csize: 0x%x, err: %i\n", start, csize, err);
_mi_warning_message("mprotect error: start: %p, csize: 0x%zx, err: %i\n", start, csize, err);
mi_mprotect_hint(err);
}
return (err == 0);
@ -928,12 +1033,12 @@ bool _mi_os_shrink(void* p, size_t oldsize, size_t newsize, mi_stats_t* stats) {
Support for allocating huge OS pages (1Gib) that are reserved up-front
and possibly associated with a specific NUMA node. (use `numa_node>=0`)
-----------------------------------------------------------------------------*/
#define MI_HUGE_OS_PAGE_SIZE (GiB)
#define MI_HUGE_OS_PAGE_SIZE (MI_GiB)
#if defined(_WIN32) && (MI_INTPTR_SIZE >= 8)
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node)
{
mi_assert_internal(size%GiB == 0);
mi_assert_internal(size%MI_GiB == 0);
mi_assert_internal(addr != NULL);
const DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE;
@ -964,7 +1069,7 @@ static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node)
else {
// fall back to regular large pages
mi_huge_pages_available = false; // don't try further huge pages
_mi_warning_message("unable to allocate using huge (1gb) pages, trying large (2mb) pages instead (status 0x%lx)\n", err);
_mi_warning_message("unable to allocate using huge (1GiB) pages, trying large (2MiB) pages instead (status 0x%lx)\n", err);
}
}
// on modern Windows try use VirtualAlloc2 for numa aware large OS page allocation
@ -974,7 +1079,7 @@ static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node)
return (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, params, 1);
}
#else
UNUSED(numa_node);
MI_UNUSED(numa_node);
#endif
// otherwise use regular virtual alloc on older windows
return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
@ -991,30 +1096,30 @@ static long mi_os_mbind(void* start, unsigned long len, unsigned long mode, cons
}
#else
static long mi_os_mbind(void* start, unsigned long len, unsigned long mode, const unsigned long* nmask, unsigned long maxnode, unsigned flags) {
UNUSED(start); UNUSED(len); UNUSED(mode); UNUSED(nmask); UNUSED(maxnode); UNUSED(flags);
MI_UNUSED(start); MI_UNUSED(len); MI_UNUSED(mode); MI_UNUSED(nmask); MI_UNUSED(maxnode); MI_UNUSED(flags);
return 0;
}
#endif
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node) {
mi_assert_internal(size%GiB == 0);
mi_assert_internal(size%MI_GiB == 0);
bool is_large = true;
void* p = mi_unix_mmap(addr, size, MI_SEGMENT_SIZE, PROT_READ | PROT_WRITE, true, true, &is_large);
if (p == NULL) return NULL;
if (numa_node >= 0 && numa_node < 8*MI_INTPTR_SIZE) { // at most 64 nodes
uintptr_t numa_mask = (1UL << numa_node);
unsigned long numa_mask = (1UL << numa_node);
// TODO: does `mbind` work correctly for huge OS pages? should we
// use `set_mempolicy` before calling mmap instead?
// see: <https://lkml.org/lkml/2017/2/9/875>
long err = mi_os_mbind(p, size, MPOL_PREFERRED, &numa_mask, 8*MI_INTPTR_SIZE, 0);
if (err != 0) {
_mi_warning_message("failed to bind huge (1gb) pages to numa node %d: %s\n", numa_node, strerror(errno));
_mi_warning_message("failed to bind huge (1GiB) pages to numa node %d: %s\n", numa_node, strerror(errno));
}
}
return p;
}
#else
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node) {
UNUSED(addr); UNUSED(size); UNUSED(numa_node);
MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(numa_node);
return NULL;
}
#endif
@ -1050,7 +1155,7 @@ static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
}
#else
static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
UNUSED(pages);
MI_UNUSED(pages);
if (total_size != NULL) *total_size = 0;
return NULL;
}
@ -1193,6 +1298,35 @@ static size_t mi_os_numa_node_countx(void) {
}
return (node+1);
}
#elif defined(__FreeBSD__) && __FreeBSD_version >= 1200000
static size_t mi_os_numa_nodex(void) {
domainset_t dom;
size_t node;
int policy;
if (cpuset_getdomain(CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, sizeof(dom), &dom, &policy) == -1) return 0ul;
for (node = 0; node < MAXMEMDOM; node++) {
if (DOMAINSET_ISSET(node, &dom)) return node;
}
return 0ul;
}
static size_t mi_os_numa_node_countx(void) {
size_t ndomains = 0;
size_t len = sizeof(ndomains);
if (sysctlbyname("vm.ndomains", &ndomains, &len, NULL, 0) == -1) return 0ul;
return ndomains;
}
#elif defined(__DragonFly__)
static size_t mi_os_numa_nodex(void) {
// TODO: DragonFly does not seem to provide any userland means to get this information.
return 0ul;
}
static size_t mi_os_numa_node_countx(void) {
size_t ncpus = 0, nvirtcoresperphys = 0;
size_t len = sizeof(size_t);
if (sysctlbyname("hw.ncpu", &ncpus, &len, NULL, 0) == -1) return 0ul;
if (sysctlbyname("hw.cpu_topology_ht_ids", &nvirtcoresperphys, &len, NULL, 0) == -1) return 0ul;
return nvirtcoresperphys * ncpus;
}
#else
static size_t mi_os_numa_nodex(void) {
return 0;
@ -1222,7 +1356,7 @@ size_t _mi_os_numa_node_count_get(void) {
}
int _mi_os_numa_node_get(mi_os_tld_t* tld) {
UNUSED(tld);
MI_UNUSED(tld);
size_t numa_count = _mi_os_numa_node_count();
if (numa_count<=1) return 0; // optimize on single numa node systems: always node 0
// never more than the node count and >= 0

17
src/page.c
View file

@ -7,7 +7,7 @@ terms of the MIT license. A copy of the license can be found in the file
/* -----------------------------------------------------------
The core of the allocator. Every segment contains
pages of a {certain block size. The main function
pages of a certain block size. The main function
exported is `mi_malloc_generic`.
----------------------------------------------------------- */
@ -30,7 +30,7 @@ terms of the MIT license. A copy of the license can be found in the file
// Index a block in a page
static inline mi_block_t* mi_page_block_at(const mi_page_t* page, void* page_start, size_t block_size, size_t i) {
UNUSED(page);
MI_UNUSED(page);
mi_assert_internal(page != NULL);
mi_assert_internal(i <= page->reserved);
return (mi_block_t*)((uint8_t*)page_start + (i * block_size));
@ -84,9 +84,10 @@ static bool mi_page_is_valid_init(mi_page_t* page) {
mi_assert_internal(mi_page_list_is_valid(page,page->local_free));
#if MI_DEBUG>3 // generally too expensive to check this
if (page->flags.is_zero) {
for(mi_block_t* block = page->free; block != NULL; mi_block_next(page,block)) {
mi_assert_expensive(mi_mem_is_zero(block + 1, page->block_size - sizeof(mi_block_t)));
if (page->is_zero) {
const size_t ubsize = mi_page_usable_block_size(page);
for(mi_block_t* block = page->free; block != NULL; block = mi_block_next(page,block)) {
mi_assert_expensive(mi_mem_is_zero(block + 1, ubsize - sizeof(mi_block_t)));
}
}
#endif
@ -385,7 +386,7 @@ void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
// Note: called from `mi_free` and benchmarks often
// trigger this due to freeing everything and then
// allocating again so careful when changing this.
void _mi_page_retire(mi_page_t* page) {
void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
mi_assert_internal(page != NULL);
mi_assert_expensive(_mi_page_is_valid(page));
mi_assert_internal(mi_page_all_free(page));
@ -458,7 +459,7 @@ void _mi_heap_collect_retired(mi_heap_t* heap, bool force) {
#define MI_MIN_SLICES (2)
static void mi_page_free_list_extend_secure(mi_heap_t* const heap, mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats) {
UNUSED(stats);
MI_UNUSED(stats);
#if (MI_SECURE<=2)
mi_assert_internal(page->free == NULL);
mi_assert_internal(page->local_free == NULL);
@ -516,7 +517,7 @@ static void mi_page_free_list_extend_secure(mi_heap_t* const heap, mi_page_t* co
static mi_decl_noinline void mi_page_free_list_extend( mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats)
{
UNUSED(stats);
MI_UNUSED(stats);
#if (MI_SECURE <= 2)
mi_assert_internal(page->free == NULL);
mi_assert_internal(page->local_free == NULL);

21
src/random.c
View file

@ -4,6 +4,10 @@ This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
#ifndef _DEFAULT_SOURCE
#define _DEFAULT_SOURCE // for syscall() on Linux
#endif
#include "mimalloc.h"
#include "mimalloc-internal.h"
@ -165,7 +169,8 @@ If we cannot get good randomness, we fall back to weak randomness based on a tim
#if defined(_WIN32)
#if !defined(MI_USE_RTLGENRANDOM)
// We prefer BCryptGenRandom over RtlGenRandom
// We prefer to use BCryptGenRandom instead of RtlGenRandom but it can lead to a deadlock
// under the VS debugger when using dynamic overriding.
#pragma comment (lib,"bcrypt.lib")
#include <bcrypt.h>
static bool os_random_buf(void* buf, size_t buf_len) {
@ -218,14 +223,16 @@ static bool os_random_buf(void* buf, size_t buf_len) {
}
#elif defined(ANDROID) || defined(__DragonFly__) || \
defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
defined(__sun) || defined(__wasi__)
defined(__sun) // todo: what to use with __wasi__?
#include <stdlib.h>
static bool os_random_buf(void* buf, size_t buf_len) {
arc4random_buf(buf, buf_len);
return true;
}
#elif defined(__linux__)
#elif defined(__linux__) || defined(__HAIKU__)
#if defined(__linux__)
#include <sys/syscall.h>
#endif
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
@ -281,8 +288,8 @@ static bool os_random_buf(void* buf, size_t buf_len) {
#include <time.h>
#endif
uintptr_t _os_random_weak(uintptr_t extra_seed) {
uintptr_t x = (uintptr_t)&_os_random_weak ^ extra_seed; // ASLR makes the address random
uintptr_t _mi_os_random_weak(uintptr_t extra_seed) {
uintptr_t x = (uintptr_t)&_mi_os_random_weak ^ extra_seed; // ASLR makes the address random
#if defined(_WIN32)
LARGE_INTEGER pcount;
@ -310,8 +317,10 @@ void _mi_random_init(mi_random_ctx_t* ctx) {
if (!os_random_buf(key, sizeof(key))) {
// if we fail to get random data from the OS, we fall back to a
// weak random source based on the current time
#if !defined(__wasi__)
_mi_warning_message("unable to use secure randomness\n");
uintptr_t x = _os_random_weak(0);
#endif
uintptr_t x = _mi_os_random_weak(0);
for (size_t i = 0; i < 8; i++) { // key is eight 32-bit words.
x = _mi_random_shuffle(x);
((uint32_t*)key)[i] = (uint32_t)x;

36
src/region.c
View file

@ -40,7 +40,7 @@ Possible issues:
#include "bitmap.h"
// Internal raw OS interface
size_t _mi_os_large_page_size();
size_t _mi_os_large_page_size(void);
bool _mi_os_protect(void* addr, size_t size);
bool _mi_os_unprotect(void* addr, size_t size);
bool _mi_os_commit(void* p, size_t size, bool* is_zero, mi_stats_t* stats);
@ -57,9 +57,9 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, boo
// Constants
#if (MI_INTPTR_SIZE==8)
#define MI_HEAP_REGION_MAX_SIZE (256 * GiB) // 64KiB for the region map
#define MI_HEAP_REGION_MAX_SIZE (256 * MI_GiB) // 64KiB for the region map
#elif (MI_INTPTR_SIZE==4)
#define MI_HEAP_REGION_MAX_SIZE (3 * GiB) // ~ KiB for the region map
#define MI_HEAP_REGION_MAX_SIZE (3 * MI_GiB) // ~ KiB for the region map
#else
#error "define the maximum heap space allowed for regions on this platform"
#endif
@ -74,7 +74,7 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, boo
// Region info
typedef union mi_region_info_u {
uintptr_t value;
size_t value;
struct {
bool valid; // initialized?
bool is_large:1; // allocated in fixed large/huge OS pages
@ -87,21 +87,21 @@ typedef union mi_region_info_u {
// A region owns a chunk of REGION_SIZE (256MiB) (virtual) memory with
// a bit map with one bit per MI_SEGMENT_SIZE (4MiB) block.
typedef struct mem_region_s {
_Atomic(uintptr_t) info; // mi_region_info_t.value
_Atomic(size_t) info; // mi_region_info_t.value
_Atomic(void*) start; // start of the memory area
mi_bitmap_field_t in_use; // bit per in-use block
mi_bitmap_field_t dirty; // track if non-zero per block
mi_bitmap_field_t commit; // track if committed per block
mi_bitmap_field_t reset; // track if reset per block
_Atomic(uintptr_t) arena_memid; // if allocated from a (huge page) arena
uintptr_t padding; // round to 8 fields
_Atomic(size_t) arena_memid; // if allocated from a (huge page) arena
size_t padding; // round to 8 fields
} mem_region_t;
// The region map
static mem_region_t regions[MI_REGION_MAX];
// Allocated regions
static _Atomic(uintptr_t) regions_count; // = 0;
static _Atomic(size_t) regions_count; // = 0;
/* ----------------------------------------------------------------------------
@ -186,21 +186,21 @@ static bool mi_region_try_alloc_os(size_t blocks, bool commit, bool allow_large,
mi_assert_internal(!region_large || region_commit);
// claim a fresh slot
const uintptr_t idx = mi_atomic_increment_acq_rel(&regions_count);
const size_t idx = mi_atomic_increment_acq_rel(&regions_count);
if (idx >= MI_REGION_MAX) {
mi_atomic_decrement_acq_rel(&regions_count);
_mi_arena_free(start, MI_REGION_SIZE, arena_memid, region_commit, tld->stats);
_mi_warning_message("maximum regions used: %zu GiB (perhaps recompile with a larger setting for MI_HEAP_REGION_MAX_SIZE)", _mi_divide_up(MI_HEAP_REGION_MAX_SIZE, GiB));
_mi_warning_message("maximum regions used: %zu GiB (perhaps recompile with a larger setting for MI_HEAP_REGION_MAX_SIZE)", _mi_divide_up(MI_HEAP_REGION_MAX_SIZE, MI_GiB));
return false;
}
// allocated, initialize and claim the initial blocks
mem_region_t* r = &regions[idx];
r->arena_memid = arena_memid;
mi_atomic_store_release(&r->in_use, (uintptr_t)0);
mi_atomic_store_release(&r->in_use, (size_t)0);
mi_atomic_store_release(&r->dirty, (is_zero ? 0 : MI_BITMAP_FIELD_FULL));
mi_atomic_store_release(&r->commit, (region_commit ? MI_BITMAP_FIELD_FULL : 0));
mi_atomic_store_release(&r->reset, (uintptr_t)0);
mi_atomic_store_release(&r->reset, (size_t)0);
*bit_idx = 0;
_mi_bitmap_claim(&r->in_use, 1, blocks, *bit_idx, NULL);
mi_atomic_store_ptr_release(void,&r->start, start);
@ -441,7 +441,7 @@ void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_re
// and unclaim
bool all_unclaimed = mi_bitmap_unclaim(&region->in_use, 1, blocks, bit_idx);
mi_assert_internal(all_unclaimed); UNUSED(all_unclaimed);
mi_assert_internal(all_unclaimed); MI_UNUSED(all_unclaimed);
}
}
@ -451,21 +451,21 @@ void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_re
-----------------------------------------------------------------------------*/
void _mi_mem_collect(mi_os_tld_t* tld) {
// free every region that has no segments in use.
uintptr_t rcount = mi_atomic_load_relaxed(&regions_count);
size_t rcount = mi_atomic_load_relaxed(&regions_count);
for (size_t i = 0; i < rcount; i++) {
mem_region_t* region = &regions[i];
if (mi_atomic_load_relaxed(&region->info) != 0) {
// if no segments used, try to claim the whole region
uintptr_t m = mi_atomic_load_relaxed(&region->in_use);
size_t m = mi_atomic_load_relaxed(&region->in_use);
while (m == 0 && !mi_atomic_cas_weak_release(&region->in_use, &m, MI_BITMAP_FIELD_FULL)) { /* nothing */ };
if (m == 0) {
// on success, free the whole region
uint8_t* start = (uint8_t*)mi_atomic_load_ptr_acquire(uint8_t,&regions[i].start);
size_t arena_memid = mi_atomic_load_relaxed(&regions[i].arena_memid);
uintptr_t commit = mi_atomic_load_relaxed(&regions[i].commit);
memset(&regions[i], 0, sizeof(mem_region_t));
size_t commit = mi_atomic_load_relaxed(&regions[i].commit);
memset((void*)&regions[i], 0, sizeof(mem_region_t)); // cast to void* to avoid atomic warning
// and release the whole region
mi_atomic_store_release(&region->info, (uintptr_t)0);
mi_atomic_store_release(&region->info, (size_t)0);
if (start != NULL) { // && !_mi_os_is_huge_reserved(start)) {
_mi_abandoned_await_readers(); // ensure no pending reads
_mi_arena_free(start, MI_REGION_SIZE, arena_memid, (~commit == 0), tld->stats);

29
src/segment.c
View file

@ -17,14 +17,14 @@ static uint8_t* mi_segment_raw_page_start(const mi_segment_t* segment, const mi_
/* --------------------------------------------------------------------------------
Segment allocation
We allocate pages inside bigger "segments" (4mb on 64-bit). This is to avoid
We allocate pages inside bigger "segments" (4MiB on 64-bit). This is to avoid
splitting VMA's on Linux and reduce fragmentation on other OS's.
Each thread owns its own segments.
Currently we have:
- small pages (64kb), 64 in one segment
- medium pages (512kb), 8 in one segment
- large pages (4mb), 1 in one segment
- small pages (64KiB), 64 in one segment
- medium pages (512KiB), 8 in one segment
- large pages (4MiB), 1 in one segment
- huge blocks > MI_LARGE_OBJ_SIZE_MAX become large segment with 1 page
In any case the memory for a segment is virtual and usually committed on demand.
@ -579,7 +579,10 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
mi_assert_internal(segment_size >= required);
// Initialize parameters
const bool eager_delayed = (page_kind <= MI_PAGE_MEDIUM && tld->count < (size_t)mi_option_get(mi_option_eager_commit_delay));
const bool eager_delayed = (page_kind <= MI_PAGE_MEDIUM && // don't delay for large objects
!_mi_os_has_overcommit() && // never delay on overcommit systems
_mi_current_thread_count() > 2 && // do not delay for the first N threads
tld->count < (size_t)mi_option_get(mi_option_eager_commit_delay));
const bool eager = !eager_delayed && mi_option_is_enabled(mi_option_eager_commit);
bool commit = eager; // || (page_kind >= MI_PAGE_LARGE);
bool pages_still_good = false;
@ -695,7 +698,7 @@ static mi_segment_t* mi_segment_alloc(size_t required, mi_page_kind_t page_kind,
}
static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t* tld) {
UNUSED(force);
MI_UNUSED(force);
mi_assert(segment != NULL);
// note: don't reset pages even on abandon as the whole segment is freed? (and ready for reuse)
bool force_reset = (force && mi_option_is_enabled(mi_option_abandoned_page_reset));
@ -896,13 +899,13 @@ static mi_decl_cache_align _Atomic(mi_segment_t*) abandoned_visited; // =
static mi_decl_cache_align _Atomic(mi_tagged_segment_t) abandoned; // = NULL
// Maintain these for debug purposes (these counts may be a bit off)
static mi_decl_cache_align _Atomic(uintptr_t) abandoned_count;
static mi_decl_cache_align _Atomic(uintptr_t) abandoned_visited_count;
static mi_decl_cache_align _Atomic(size_t) abandoned_count;
static mi_decl_cache_align _Atomic(size_t) abandoned_visited_count;
// We also maintain a count of current readers of the abandoned list
// in order to prevent resetting/decommitting segment memory if it might
// still be read.
static mi_decl_cache_align _Atomic(uintptr_t) abandoned_readers; // = 0
static mi_decl_cache_align _Atomic(size_t) abandoned_readers; // = 0
// Push on the visited list
static void mi_abandoned_visited_push(mi_segment_t* segment) {
@ -931,7 +934,7 @@ static bool mi_abandoned_visited_revisit(void)
mi_tagged_segment_t afirst;
mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
if (mi_tagged_segment_ptr(ts)==NULL) {
uintptr_t count = mi_atomic_load_relaxed(&abandoned_visited_count);
size_t count = mi_atomic_load_relaxed(&abandoned_visited_count);
afirst = mi_tagged_segment(first, ts);
if (mi_atomic_cas_strong_acq_rel(&abandoned, &ts, afirst)) {
mi_atomic_add_relaxed(&abandoned_count, count);
@ -950,7 +953,7 @@ static bool mi_abandoned_visited_revisit(void)
// and atomically prepend to the abandoned list
// (no need to increase the readers as we don't access the abandoned segments)
mi_tagged_segment_t anext = mi_atomic_load_relaxed(&abandoned);
uintptr_t count;
size_t count;
do {
count = mi_atomic_load_relaxed(&abandoned_visited_count);
mi_atomic_store_ptr_release(mi_segment_t, &last->abandoned_next, mi_tagged_segment_ptr(anext));
@ -978,7 +981,7 @@ static void mi_abandoned_push(mi_segment_t* segment) {
// Wait until there are no more pending reads on segments that used to be in the abandoned list
void _mi_abandoned_await_readers(void) {
uintptr_t n;
size_t n;
do {
n = mi_atomic_load_acquire(&abandoned_readers);
if (n != 0) mi_atomic_yield();
@ -1326,7 +1329,7 @@ void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block
// claim it and free
mi_heap_t* heap = mi_heap_get_default(); // issue #221; don't use the internal get_default_heap as we need to ensure the thread is initialized.
// paranoia: if this it the last reference, the cas should always succeed
uintptr_t expected_tid = 0;
size_t expected_tid = 0;
if (mi_atomic_cas_strong_acq_rel(&segment->thread_id, &expected_tid, heap->thread_id)) {
mi_block_set_next(page, block, page->free);
page->free = block;

33
src/stats.c
View file

@ -133,25 +133,29 @@ static void mi_stats_add(mi_stats_t* stats, const mi_stats_t* src) {
// unit == 0: count as decimal
// unit < 0 : count in binary
static void mi_printf_amount(int64_t n, int64_t unit, mi_output_fun* out, void* arg, const char* fmt) {
char buf[32];
char buf[32]; buf[0] = 0;
int len = 32;
const char* suffix = (unit <= 0 ? " " : "b");
const char* suffix = (unit <= 0 ? " " : "B");
const int64_t base = (unit == 0 ? 1000 : 1024);
if (unit>0) n *= unit;
const int64_t pos = (n < 0 ? -n : n);
if (pos < base) {
snprintf(buf, len, "%d %s ", (int)n, suffix);
if (n!=1 || suffix[0] != 'B') { // skip printing 1 B for the unit column
snprintf(buf, len, "%d %-3s", (int)n, (n==0 ? "" : suffix));
}
}
else {
int64_t divider = base;
const char* magnitude = "k";
if (pos >= divider*base) { divider *= base; magnitude = "m"; }
if (pos >= divider*base) { divider *= base; magnitude = "g"; }
int64_t divider = base;
const char* magnitude = "K";
if (pos >= divider*base) { divider *= base; magnitude = "M"; }
if (pos >= divider*base) { divider *= base; magnitude = "G"; }
const int64_t tens = (n / (divider/10));
const long whole = (long)(tens/10);
const long frac1 = (long)(tens%10);
snprintf(buf, len, "%ld.%ld %s%s", whole, (frac1 < 0 ? -frac1 : frac1), magnitude, suffix);
char unitdesc[8];
snprintf(unitdesc, 8, "%s%s%s", magnitude, (base==1024 ? "i" : ""), suffix);
snprintf(buf, len, "%ld.%ld %-3s", whole, (frac1 < 0 ? -frac1 : frac1), unitdesc);
}
_mi_fprintf(out, arg, (fmt==NULL ? "%11s" : fmt), buf);
}
@ -221,7 +225,7 @@ static void mi_stat_counter_print_avg(const mi_stat_counter_t* stat, const char*
static void mi_print_header(mi_output_fun* out, void* arg ) {
_mi_fprintf(out, arg, "%10s: %10s %10s %10s %10s %10s %10s\n", "heap stats", "peak ", "total ", "freed ", "current ", "unit ", "count ");
_mi_fprintf(out, arg, "%10s: %10s %10s %10s %10s %10s %10s\n", "heap stats", "peak ", "total ", "freed ", "current ", "unit ", "count ");
}
#if MI_STAT>1
@ -323,7 +327,7 @@ static void _mi_stats_print(mi_stats_t* stats, mi_output_fun* out0, void* arg0)
mi_stat_counter_print(&stats->commit_calls, "commits", out, arg);
mi_stat_print(&stats->threads, "threads", -1, out, arg);
mi_stat_counter_print_avg(&stats->searches, "searches", out, arg);
_mi_fprintf(out, arg, "%10s: %7i\n", "numa nodes", _mi_os_numa_node_count());
_mi_fprintf(out, arg, "%10s: %7zu\n", "numa nodes", _mi_os_numa_node_count());
mi_msecs_t elapsed;
mi_msecs_t user_time;
@ -412,10 +416,14 @@ mi_msecs_t _mi_clock_now(void) {
}
#else
#include <time.h>
#ifdef CLOCK_REALTIME
#if defined(CLOCK_REALTIME) || defined(CLOCK_MONOTONIC)
mi_msecs_t _mi_clock_now(void) {
struct timespec t;
#ifdef CLOCK_MONOTONIC
clock_gettime(CLOCK_MONOTONIC, &t);
#else
clock_gettime(CLOCK_REALTIME, &t);
#endif
return ((mi_msecs_t)t.tv_sec * 1000) + ((mi_msecs_t)t.tv_nsec / 1000000);
}
#else
@ -479,7 +487,7 @@ static void mi_stat_process_info(mi_msecs_t* elapsed, mi_msecs_t* utime, mi_msec
*page_faults = (size_t)info.PageFaultCount;
}
#elif defined(__unix__) || defined(__unix) || defined(unix) || defined(__APPLE__) || defined(__HAIKU__)
#elif !defined(__wasi__) && (defined(__unix__) || defined(__unix) || defined(unix) || defined(__APPLE__) || defined(__HAIKU__))
#include <stdio.h>
#include <unistd.h>
#include <sys/resource.h>
@ -520,6 +528,7 @@ static void mi_stat_process_info(mi_msecs_t* elapsed, mi_msecs_t* utime, mi_msec
while (get_next_area_info(tid.team, &c, &mem) == B_OK) {
*peak_rss += mem.ram_size;
}
*page_faults = 0;
#elif defined(__APPLE__)
*peak_rss = rusage.ru_maxrss; // BSD reports in bytes
struct mach_task_basic_info info;

7
test/CMakeLists.txt
View file

@ -1,6 +1,9 @@
cmake_minimum_required(VERSION 3.0)
project(mimalloc-test C CXX)
set(CMAKE_C_STANDARD 11)
set(CMAKE_CXX_STANDARD 17)
# Set default build type
if (NOT CMAKE_BUILD_TYPE)
if ("${CMAKE_BINARY_DIR}" MATCHES ".*(D|d)ebug$")
@ -14,7 +17,7 @@ endif()
# Import mimalloc (if installed)
find_package(mimalloc 1.7 REQUIRED NO_SYSTEM_ENVIRONMENT_PATH)
message(STATUS "Found mimalloc installed at: ${MIMALLOC_LIBRARY_DIR}")
message(STATUS "Found mimalloc installed at: ${MIMALLOC_LIBRARY_DIR} (${MIMALLOC_VERSION_DIR})")
# overriding with a dynamic library
add_executable(dynamic-override main-override.c)
@ -26,7 +29,7 @@ target_link_libraries(dynamic-override-cxx PUBLIC mimalloc)
# overriding with a static object file works reliable as the symbols in the
# object file have priority over those in library files
add_executable(static-override-obj main-override.c ${MIMALLOC_LIBRARY_DIR}/mimalloc.o)
add_executable(static-override-obj main-override.c ${MIMALLOC_OBJECT_DIR}/mimalloc.o)
target_include_directories(static-override-obj PUBLIC ${MIMALLOC_INCLUDE_DIR})
target_link_libraries(static-override-obj PUBLIC pthread)

8
test/main-override.c
View file

@ -3,7 +3,7 @@
#include <assert.h>
#include <string.h>
#include <mimalloc.h>
#include <mimalloc-override.h>
int main() {
mi_version(); // ensure mimalloc library is linked
@ -25,6 +25,12 @@ int main() {
//free(p1);
//p2 = malloc(32);
//mi_free(p2);
p1 = malloc(24);
p2 = reallocarray(p1, 16, 16);
free(p2);
p1 = malloc(24);
assert(reallocarr(&p1, 16, 16) == 0);
free(p1);
mi_stats_print(NULL);
return 0;
}

67
test/main-override.cpp
View file

@ -26,14 +26,15 @@ static void msleep(unsigned long msecs) { Sleep(msecs); }
static void msleep(unsigned long msecs) { usleep(msecs * 1000UL); }
#endif
void heap_thread_free_large(); // issue #221
void heap_no_delete(); // issue #202
void heap_late_free(); // issue #204
void padding_shrink(); // issue #209
void various_tests();
void test_mt_shutdown();
void fail_aslr(); // issue #372
void tsan_numa_test(); // issue #414
static void heap_thread_free_large(); // issue #221
static void heap_no_delete(); // issue #202
static void heap_late_free(); // issue #204
static void padding_shrink(); // issue #209
static void various_tests();
static void test_mt_shutdown();
static void fail_aslr(); // issue #372
static void tsan_numa_test(); // issue #414
static void strdup_test(); // issue #445
int main() {
mi_stats_reset(); // ignore earlier allocations
@ -43,6 +44,7 @@ int main() {
padding_shrink();
various_tests();
tsan_numa_test();
strdup_test();
//test_mt_shutdown();
//fail_aslr();
@ -66,7 +68,7 @@ public:
};
void various_tests() {
static void various_tests() {
atexit(free_p);
void* p1 = malloc(78);
void* p2 = mi_malloc_aligned(16, 24);
@ -74,18 +76,13 @@ void various_tests() {
p1 = malloc(8);
char* s = mi_strdup("hello\n");
//char* s = _strdup("hello\n");
//char* buf = NULL;
//size_t len;
//_dupenv_s(&buf,&len,"MIMALLOC_VERBOSE");
//mi_free(buf);
mi_free(p2);
p2 = malloc(16);
p1 = realloc(p1, 32);
free(p1);
free(p2);
mi_free(s);
Test* t = new Test(42);
delete t;
t = new (std::nothrow) Test(42);
@ -109,7 +106,7 @@ public:
static Static s = Static();
bool test_stl_allocator1() {
static bool test_stl_allocator1() {
std::vector<int, mi_stl_allocator<int> > vec;
vec.push_back(1);
vec.pop_back();
@ -118,38 +115,48 @@ bool test_stl_allocator1() {
struct some_struct { int i; int j; double z; };
bool test_stl_allocator2() {
static bool test_stl_allocator2() {
std::vector<some_struct, mi_stl_allocator<some_struct> > vec;
vec.push_back(some_struct());
vec.pop_back();
return vec.size() == 0;
}
// issue 445
static void strdup_test() {
#ifdef _MSC_VER
char* s = _strdup("hello\n");
char* buf = NULL;
size_t len;
_dupenv_s(&buf, &len, "MIMALLOC_VERBOSE");
mi_free(buf);
mi_free(s);
#endif
}
// Issue #202
void heap_no_delete_worker() {
static void heap_no_delete_worker() {
mi_heap_t* heap = mi_heap_new();
void* q = mi_heap_malloc(heap, 1024);
// mi_heap_delete(heap); // uncomment to prevent assertion
}
void heap_no_delete() {
static void heap_no_delete() {
auto t1 = std::thread(heap_no_delete_worker);
t1.join();
}
// Issue #204
volatile void* global_p;
static volatile void* global_p;
void t1main() {
static void t1main() {
mi_heap_t* heap = mi_heap_new();
global_p = mi_heap_malloc(heap, 1024);
mi_heap_delete(heap);
}
void heap_late_free() {
static void heap_late_free() {
auto t1 = std::thread(t1main);
msleep(2000);
@ -166,7 +173,7 @@ static void alloc0(/* void* arg */)
shared_p = mi_malloc(8);
}
void padding_shrink(void)
static void padding_shrink(void)
{
auto t1 = std::thread(alloc0);
t1.join();
@ -175,11 +182,11 @@ void padding_shrink(void)
// Issue #221
void heap_thread_free_large_worker() {
static void heap_thread_free_large_worker() {
mi_free(shared_p);
}
void heap_thread_free_large() {
static void heap_thread_free_large() {
for (int i = 0; i < 100; i++) {
shared_p = mi_malloc_aligned(2*1024*1024 + 1, 8);
auto t1 = std::thread(heap_thread_free_large_worker);
@ -189,7 +196,7 @@ void heap_thread_free_large() {
void test_mt_shutdown()
static void test_mt_shutdown()
{
const int threads = 5;
std::vector< std::future< std::vector< char* > > > ts;
@ -214,7 +221,7 @@ void test_mt_shutdown()
}
// issue #372
void fail_aslr() {
static void fail_aslr() {
size_t sz = (4ULL << 40); // 4TiB
void* p = malloc(sz);
printf("pointer p: %p: area up to %p\n", p, (uint8_t*)p + sz);
@ -222,12 +229,12 @@ void fail_aslr() {
}
// issues #414
void dummy_worker() {
static void dummy_worker() {
void* p = mi_malloc(0);
mi_free(p);
}
void tsan_numa_test() {
static void tsan_numa_test() {
auto t1 = std::thread(dummy_worker);
dummy_worker();
t1.join();

15
test/test-api.c
View file

@ -4,6 +4,9 @@ This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic ignored "-Walloc-size-larger-than="
#endif
/*
Testing allocators is difficult as bugs may only surface after particular
@ -64,15 +67,15 @@ static int failed = 0;
// ---------------------------------------------------------------------------
// Test functions
// ---------------------------------------------------------------------------
bool test_heap1();
bool test_heap2();
bool test_stl_allocator1();
bool test_stl_allocator2();
bool test_heap1(void);
bool test_heap2(void);
bool test_stl_allocator1(void);
bool test_stl_allocator2(void);
// ---------------------------------------------------------------------------
// Main testing
// ---------------------------------------------------------------------------
int main() {
int main(void) {
mi_option_disable(mi_option_verbose);
// ---------------------------------------------------
@ -83,7 +86,7 @@ int main() {
void* p = mi_malloc(0); mi_free(p);
});
CHECK_BODY("malloc-nomem1",{
result = (mi_malloc(SIZE_MAX/2) == NULL);
result = (mi_malloc((size_t)PTRDIFF_MAX + (size_t)1) == NULL);
});
CHECK_BODY("malloc-null",{
mi_free(NULL);

2
test/test-stress.c
View file

@ -25,7 +25,7 @@ terms of the MIT license.
//
// argument defaults
static int THREADS = 32; // more repeatable if THREADS <= #processors
static int SCALE = 10; // scaling factor
static int SCALE = 25; // scaling factor
static int ITER = 50; // N full iterations destructing and re-creating all threads
// static int THREADS = 8; // more repeatable if THREADS <= #processors