diff --git a/ide/vs2017/mimalloc-override.vcxproj b/ide/vs2017/mimalloc-override.vcxproj index d9bce9c0..ee44993c 100644 --- a/ide/vs2017/mimalloc-override.vcxproj +++ b/ide/vs2017/mimalloc-override.vcxproj @@ -146,7 +146,6 @@ MaxSpeed true true - true true ../../include MI_SHARED_LIB;MI_SHARED_LIB_EXPORT;MI_MALLOC_OVERRIDE;%(PreprocessorDefinitions);NDEBUG @@ -155,6 +154,7 @@ false MultiThreadedDLL Default + false true @@ -173,7 +173,6 @@ MaxSpeed true true - true true ../../include MI_SHARED_LIB;MI_SHARED_LIB_EXPORT;MI_MALLOC_OVERRIDE;%(PreprocessorDefinitions);NDEBUG @@ -182,6 +181,7 @@ false MultiThreadedDLL Default + false true diff --git a/ide/vs2017/mimalloc.vcxproj b/ide/vs2017/mimalloc.vcxproj index 5b151da7..4e68b628 100644 --- a/ide/vs2017/mimalloc.vcxproj +++ b/ide/vs2017/mimalloc.vcxproj @@ -141,8 +141,6 @@ Level3 MaxSpeed true - true - true true ../../include %(PreprocessorDefinitions);NDEBUG @@ -150,11 +148,9 @@ $(IntDir) false false - AnySuitable - Neither - false - false + Default CompileAsCpp + true true @@ -172,8 +168,6 @@ Level3 MaxSpeed true - true - true true ../../include %(PreprocessorDefinitions);NDEBUG @@ -181,11 +175,9 @@ $(IntDir) false false - AnySuitable - Neither - false - false + Default CompileAsCpp + true true diff --git a/include/mimalloc-atomic.h b/include/mimalloc-atomic.h index d504634c..b20f47b6 100644 --- a/include/mimalloc-atomic.h +++ b/include/mimalloc-atomic.h @@ -68,6 +68,9 @@ static inline void* mi_atomic_exchange_ptr(volatile void** p, void* exchange) { return (void*)mi_atomic_exchange((volatile uintptr_t*)p, (uintptr_t)exchange); } +static inline intptr_t mi_atomic_iread(volatile intptr_t* p) { + return (intptr_t)mi_atomic_read( (volatile uintptr_t*)p ); +} #ifdef _MSC_VER #define WIN32_LEAN_AND_MEAN diff --git a/include/mimalloc-internal.h b/include/mimalloc-internal.h index 3aee4ae1..f9bd7f10 100644 --- a/include/mimalloc-internal.h +++ b/include/mimalloc-internal.h @@ -102,6 +102,8 @@ uintptr_t _mi_heap_random(mi_heap_t* heap); // "stats.c" void _mi_stats_done(mi_stats_t* stats); +double _mi_clock_end(double start); +double _mi_clock_start(void); // "alloc.c" void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size) mi_attr_noexcept; // called from `_mi_malloc_generic` @@ -161,15 +163,15 @@ bool _mi_page_is_valid(mi_page_t* page); // Overflow detecting multiply #define MI_MUL_NO_OVERFLOW ((size_t)1 << (4*sizeof(size_t))) // sqrt(SIZE_MAX) -static inline bool mi_mul_overflow(size_t size, size_t count, size_t* total) { +static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) { #if __has_builtin(__builtin_umul_overflow) || __GNUC__ >= 5 #if (MI_INTPTR_SIZE == 4) - return __builtin_umul_overflow(size, count, total); + return __builtin_umul_overflow(count, size, total); #else - return __builtin_umull_overflow(size, count, total); + return __builtin_umull_overflow(count, size, total); #endif #else /* __builtin_umul_overflow is unavailable */ - *total = size * count; + *total = count * size; return ((size >= MI_MUL_NO_OVERFLOW || count >= MI_MUL_NO_OVERFLOW) && size > 0 && (SIZE_MAX / size) < count); #endif diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h index 7a240b7e..2044d873 100644 --- a/include/mimalloc-types.h +++ b/include/mimalloc-types.h @@ -84,18 +84,19 @@ terms of the MIT license. A copy of the license can be found in the file // Derived constants #define MI_SEGMENT_SIZE ((size_t)1<= 655360) + +#if (MI_MEDIUM_OBJ_WSIZE_MAX >= 655360) #error "define more bins" #endif @@ -170,7 +172,7 @@ typedef struct mi_page_s { #endif mi_page_flags_t flags; size_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`) - + mi_block_t* local_free; // list of deferred free blocks by this thread (migrates to `free`) volatile uintptr_t thread_freed; // at least this number of blocks are in `thread_free` volatile mi_thread_free_t thread_free; // list of deferred free blocks freed by other threads @@ -213,7 +215,7 @@ typedef struct mi_segment_s { struct mi_segment_s* prev; struct mi_segment_s* abandoned_next; // abandoned segment stack: `used == abandoned` size_t abandoned; // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`) - size_t used; // count of pages in use + size_t used; // count of pages in use size_t segment_size;// for huge pages this may be different from `MI_SEGMENT_SIZE` size_t segment_info_size; // space we are using from the first page for segment meta-data and possible guard pages. uintptr_t cookie; // verify addresses in debug mode: `mi_ptr_cookie(segment) == segment->cookie` @@ -221,7 +223,7 @@ typedef struct mi_segment_s { bool all_committed; // layout like this to optimize access in `mi_free` - mi_segment_kind_t kind; + mi_segment_kind_t kind; uintptr_t thread_id; size_t slice_count; // slices in this segment (at most MI_SLICES_PER_SEGMENT) mi_slice_t slices[MI_SLICES_PER_SEGMENT]; diff --git a/include/mimalloc.h b/include/mimalloc.h index c6b7b5f8..fb41b037 100644 --- a/include/mimalloc.h +++ b/include/mimalloc.h @@ -195,7 +195,7 @@ typedef bool (mi_cdecl mi_block_visit_fun)(const mi_heap_t* heap, const mi_heap_ mi_decl_export bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_all_blocks, mi_block_visit_fun* visitor, void* arg); mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept; - +mi_decl_export int mi_reserve_huge_os_pages(size_t pages, double max_secs) mi_attr_noexcept; // ------------------------------------------------------ // Convenience @@ -220,25 +220,27 @@ mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept; typedef enum mi_option_e { // stable options - mi_option_show_stats, mi_option_show_errors, + mi_option_show_stats, mi_option_verbose, // the following options are experimental mi_option_secure, mi_option_eager_commit, mi_option_eager_region_commit, - mi_option_large_os_pages, // implies eager commit + mi_option_large_os_pages, // implies eager commit + mi_option_reserve_huge_os_pages, mi_option_page_reset, mi_option_cache_reset, mi_option_reset_decommits, - mi_option_reset_discards, _mi_option_last } mi_option_t; mi_decl_export bool mi_option_is_enabled(mi_option_t option); -mi_decl_export void mi_option_enable(mi_option_t option, bool enable); -mi_decl_export void mi_option_enable_default(mi_option_t option, bool enable); +mi_decl_export void mi_option_enable(mi_option_t option); +mi_decl_export void mi_option_disable(mi_option_t option); +mi_decl_export void mi_option_set_enabled(mi_option_t option, bool enable); +mi_decl_export void mi_option_set_enabled_default(mi_option_t option, bool enable); mi_decl_export long mi_option_get(mi_option_t option); mi_decl_export void mi_option_set(mi_option_t option, long value); diff --git a/src/alloc-override-win.c b/src/alloc-override-win.c index d1d51b9a..0bd05deb 100644 --- a/src/alloc-override-win.c +++ b/src/alloc-override-win.c @@ -666,7 +666,7 @@ static void mi_patches_at_quick_exit(void) { mi_patches_enable_term(); // enter termination phase and patch realloc/free with a no-op } -__declspec(dllexport) BOOL WINAPI DllEntry(HINSTANCE inst, DWORD reason, LPVOID reserved) { +BOOL WINAPI DllEntry(HINSTANCE inst, DWORD reason, LPVOID reserved) { if (reason == DLL_PROCESS_ATTACH) { __security_init_cookie(); } diff --git a/src/alloc.c b/src/alloc.c index b5a48bde..8fd8c4b3 100644 --- a/src/alloc.c +++ b/src/alloc.c @@ -38,7 +38,7 @@ extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t siz block->next = 0; #endif #if (MI_STAT>1) - if(size <= MI_LARGE_SIZE_MAX) { + if(size <= MI_LARGE_OBJ_SIZE_MAX) { size_t bin = _mi_bin(size); mi_heap_stat_increase(heap,normal[bin], 1); } @@ -230,7 +230,7 @@ void mi_free(void* p) mi_attr_noexcept #if (MI_STAT>1) mi_heap_t* heap = mi_heap_get_default(); mi_heap_stat_decrease( heap, malloc, mi_usable_size(p)); - if (page->block_size <= MI_LARGE_SIZE_MAX) { + if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { mi_heap_stat_decrease( heap, normal[_mi_bin(page->block_size)], 1); } // huge page stat is accounted for in `_mi_page_retire` diff --git a/src/heap.c b/src/heap.c index 69084731..d10677eb 100644 --- a/src/heap.c +++ b/src/heap.c @@ -130,19 +130,19 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect) // if abandoning, mark all pages to no longer add to delayed_free if (collect == ABANDON) { //for (mi_page_t* page = heap->pages[MI_BIN_FULL].first; page != NULL; page = page->next) { - // _mi_page_use_delayed_free(page, false); // set thread_free.delayed to MI_NO_DELAYED_FREE - //} + // _mi_page_use_delayed_free(page, false); // set thread_free.delayed to MI_NO_DELAYED_FREE + //} mi_heap_visit_pages(heap, &mi_heap_page_never_delayed_free, NULL, NULL); } - // free thread delayed blocks. + // free thread delayed blocks. // (if abandoning, after this there are no more local references into the pages.) _mi_heap_delayed_free(heap); // collect all pages owned by this thread mi_heap_visit_pages(heap, &mi_heap_page_collect, &collect, NULL); mi_assert_internal( collect != ABANDON || heap->thread_delayed_free == NULL ); - + // collect segment caches if (collect >= FORCE) { _mi_segment_thread_collect(&heap->tld->segments); @@ -172,7 +172,7 @@ void mi_collect(bool force) mi_attr_noexcept { ----------------------------------------------------------- */ mi_heap_t* mi_heap_get_default(void) { - mi_thread_init(); + mi_thread_init(); return mi_get_default_heap(); } @@ -221,7 +221,7 @@ static void mi_heap_reset_pages(mi_heap_t* heap) { static void mi_heap_free(mi_heap_t* heap) { mi_assert_internal(mi_heap_is_initialized(heap)); if (mi_heap_is_backing(heap)) return; // dont free the backing heap - + // reset default if (mi_heap_is_default(heap)) { _mi_heap_default = heap->tld->heap_backing; @@ -242,11 +242,11 @@ static bool _mi_heap_page_destroy(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_ UNUSED(pq); // ensure no more thread_delayed_free will be added - _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE); + _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE); // stats - if (page->block_size > MI_MEDIUM_SIZE_MAX) { - if (page->block_size <= MI_LARGE_SIZE_MAX) { + if (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX) { + if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { _mi_stat_decrease(&heap->tld->stats.large,page->block_size); } else { @@ -255,7 +255,7 @@ static bool _mi_heap_page_destroy(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_ } #if (MI_STAT>1) size_t inuse = page->used - page->thread_freed; - if (page->block_size <= MI_LARGE_SIZE_MAX) { + if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { mi_heap_stat_decrease(heap,normal[_mi_bin(page->block_size)], inuse); } mi_heap_stat_decrease(heap,malloc, page->block_size * inuse); // todo: off for aligned blocks... @@ -303,20 +303,20 @@ 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; - // unfull all full pages - mi_page_t* page = heap->pages[MI_BIN_FULL].first; + // unfull all full pages in the `from` heap + mi_page_t* page = from->pages[MI_BIN_FULL].first; while (page != NULL) { mi_page_t* next = page->next; _mi_page_unfull(page); page = next; } - mi_assert_internal(heap->pages[MI_BIN_FULL].first == NULL); + mi_assert_internal(from->pages[MI_BIN_FULL].first == NULL); // free outstanding thread delayed free blocks _mi_heap_delayed_free(from); // transfer all pages by appending the queues; this will set - // a new heap field which is ok as all pages are unfull'd and thus + // a new heap field which is ok as all pages are unfull'd and thus // other threads won't access this field anymore (see `mi_free_block_mt`) for (size_t i = 0; i < MI_BIN_FULL; i++) { mi_page_queue_t* pq = &heap->pages[i]; @@ -327,7 +327,7 @@ static void mi_heap_absorb(mi_heap_t* heap, mi_heap_t* from) { } mi_assert_internal(from->thread_delayed_free == NULL); mi_assert_internal(from->page_count == 0); - + // and reset the `from` heap mi_heap_reset_pages(from); } @@ -525,4 +525,3 @@ bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_blocks, mi_block_vis mi_visit_blocks_args_t args = { visit_blocks, visitor, arg }; return mi_heap_visit_areas(heap, &mi_heap_area_visitor, &args); } - diff --git a/src/init.c b/src/init.c index d48c1a30..d941f55d 100644 --- a/src/init.c +++ b/src/init.c @@ -43,8 +43,8 @@ const mi_page_t _mi_page_empty = { QNULL( 10240), QNULL( 12288), QNULL( 14336), QNULL( 16384), QNULL( 20480), QNULL( 24576), QNULL( 28672), QNULL( 32768), /* 56 */ \ QNULL( 40960), QNULL( 49152), QNULL( 57344), QNULL( 65536), QNULL( 81920), QNULL( 98304), QNULL(114688), QNULL(131072), /* 64 */ \ QNULL(163840), QNULL(196608), QNULL(229376), QNULL(262144), QNULL(327680), QNULL(393216), QNULL(458752), QNULL(524288), /* 72 */ \ - QNULL(MI_MEDIUM_WSIZE_MAX + 1 /* 655360, Huge queue */), \ - QNULL(MI_MEDIUM_WSIZE_MAX + 2) /* Full queue */ } + QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 1 /* 655360, Huge queue */), \ + QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 2) /* Full queue */ } #define MI_STAT_COUNT_NULL() {0,0,0,0} @@ -116,14 +116,14 @@ mi_heap_t _mi_heap_main = { MI_SMALL_PAGES_EMPTY, MI_PAGE_QUEUES_EMPTY, NULL, - 0, - 0, + 0, // thread id #if MI_INTPTR_SIZE==8 // the cookie of the main heap can be fixed (unlike page cookies that need to be secure!) 0xCDCDCDCDCDCDCDCDUL, #else 0xCDCDCDCDUL, #endif - 0, + 0, // random + 0, // page count false // can reclaim }; @@ -432,6 +432,12 @@ static void mi_process_load(void) { const char* msg = NULL; mi_allocator_init(&msg); if (msg != NULL) _mi_verbose_message(msg); + + if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) { + size_t pages = mi_option_get(mi_option_reserve_huge_os_pages); + double max_secs = (double)pages / 5.0; // 0.2s per page + mi_reserve_huge_os_pages(pages, max_secs); + } } // Initialize the process; called by thread_init or the process loader diff --git a/src/memory.c b/src/memory.c index bf99ef42..26f87092 100644 --- a/src/memory.c +++ b/src/memory.c @@ -128,6 +128,7 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit size_t mask = mi_region_block_mask(blocks,bitidx); mi_assert_internal(mask != 0); mi_assert_internal((mask & mi_atomic_read(®ion->map)) == mask); + mi_assert_internal(®ions[idx] == region); // ensure the region is reserved void* start = mi_atomic_read_ptr(®ion->start); @@ -149,9 +150,23 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit mi_atomic_increment(®ions_count); } else { - // failed, another thread allocated just before us, free our allocated memory - // TODO: should we keep the allocated memory and assign it to some other region? - _mi_os_free(start, MI_REGION_SIZE, tld->stats); + // failed, another thread allocated just before us! + // we assign it to a later slot instead (up to 4 tries). + // note: we don't need to increment the region count, this will happen on another allocation + for(size_t i = 1; i <= 4 && idx + i < MI_REGION_MAX; i++) { + void* s = mi_atomic_read_ptr(®ions[idx+i].start); + if (s == NULL) { // quick test + if (mi_atomic_compare_exchange_ptr(®ions[idx+i].start, start, s)) { + start = NULL; + break; + } + } + } + if (start != NULL) { + // free it if we didn't succeed to save it to some other region + _mi_os_free(start, MI_REGION_SIZE, tld->stats); + } + // and continue with the memory at our index start = mi_atomic_read_ptr(®ion->start); } } diff --git a/src/options.c b/src/options.c index cd7e5da1..ff65c3f5 100644 --- a/src/options.c +++ b/src/options.c @@ -34,34 +34,38 @@ typedef enum mi_init_e { typedef struct mi_option_desc_s { long value; // the value mi_init_t init; // is it initialized yet? (from the environment) + mi_option_t option; // for debugging: the option index should match the option const char* name; // option name without `mimalloc_` prefix } mi_option_desc_t; +#define MI_OPTION(opt) mi_option_##opt, #opt +#define MI_OPTION_DESC(opt) {0, UNINIT, MI_OPTION(opt) } + static mi_option_desc_t options[_mi_option_last] = { // stable options - { 0, UNINIT, "show_stats" }, - { MI_DEBUG, UNINIT, "show_errors" }, - { 0, UNINIT, "verbose" }, + { MI_DEBUG, UNINIT, MI_OPTION(show_errors) }, + { 0, UNINIT, MI_OPTION(show_stats) }, + { 0, UNINIT, MI_OPTION(verbose) }, #if MI_SECURE - { MI_SECURE, INITIALIZED, "secure" }, // in a secure build the environment setting is ignored + { MI_SECURE, INITIALIZED, MI_OPTION(secure) }, // in a secure build the environment setting is ignored #else - { 0, UNINIT, "secure" }, + { 0, UNINIT, MI_OPTION(secure) }, #endif // the following options are experimental and not all combinations make sense. - { 1, UNINIT, "eager_commit" }, // note: if eager_region_commit is on, this should be on too. + { 1, UNINIT, MI_OPTION(eager_commit) }, // note: if eager_region_commit is on, this should be on too. #ifdef _WIN32 // and BSD? - { 0, UNINIT, "eager_region_commit" }, // don't commit too eagerly on windows (just for looks...) + { 1, UNINIT, MI_OPTION(eager_region_commit) }, // don't commit too eagerly on windows (just for looks...) #else - { 1, UNINIT, "eager_region_commit" }, + { 1, UNINIT, MI_OPTION(eager_region_commit) }, #endif - { 0, UNINIT, "large_os_pages" }, // use large OS pages, use only with eager commit to prevent fragmentation of VMA's - { 0, UNINIT, "page_reset" }, - { 0, UNINIT, "cache_reset" }, - { 0, UNINIT, "reset_decommits" }, // note: cannot enable this if secure is on - { 0, UNINIT, "reset_discards" } // note: cannot enable this if secure is on + { 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) }, + { 0, UNINIT, MI_OPTION(page_reset) }, + { 0, UNINIT, MI_OPTION(cache_reset) }, + { 0, UNINIT, MI_OPTION(reset_decommits) } // note: cannot enable this if secure is on }; static void mi_option_init(mi_option_desc_t* desc); @@ -69,6 +73,7 @@ static void mi_option_init(mi_option_desc_t* desc); 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 if (mi_unlikely(desc->init == UNINIT)) { mi_option_init(desc); if (option != mi_option_verbose) { @@ -81,6 +86,7 @@ long mi_option_get(mi_option_t option) { void mi_option_set(mi_option_t option, long value) { 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 desc->value = value; desc->init = INITIALIZED; } @@ -97,14 +103,23 @@ bool mi_option_is_enabled(mi_option_t option) { return (mi_option_get(option) != 0); } -void mi_option_enable(mi_option_t option, bool enable) { +void mi_option_set_enabled(mi_option_t option, bool enable) { mi_option_set(option, (enable ? 1 : 0)); } -void mi_option_enable_default(mi_option_t option, bool enable) { +void mi_option_set_enabled_default(mi_option_t option, bool enable) { mi_option_set_default(option, (enable ? 1 : 0)); } +void mi_option_enable(mi_option_t option) { + mi_option_set_enabled(option,true); +} + +void mi_option_disable(mi_option_t option) { + mi_option_set_enabled(option,false); +} + + // -------------------------------------------------------- // Messages // -------------------------------------------------------- diff --git a/src/os.c b/src/os.c index bee5ac64..b15d58d0 100644 --- a/src/os.c +++ b/src/os.c @@ -22,6 +22,9 @@ terms of the MIT license. A copy of the license can be found in the file #else #include // mmap #include // sysconf +#if defined(__linux__) +#include // linux mmap flags +#endif #if defined(__APPLE__) #include #endif @@ -34,6 +37,9 @@ terms of the MIT license. A copy of the license can be found in the file ----------------------------------------------------------- */ bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats); +static bool mi_os_is_huge_reserved(void* p); +static void* mi_os_alloc_from_huge_reserved(size_t size, size_t try_alignment, bool commit); + static void* mi_align_up_ptr(void* p, size_t alignment) { return (void*)_mi_align_up((uintptr_t)p, alignment); } @@ -67,7 +73,7 @@ size_t _mi_os_large_page_size() { 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) return false; + 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); } @@ -81,11 +87,13 @@ static size_t mi_os_good_alloc_size(size_t size, size_t alignment) { #if defined(_WIN32) // We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016. // So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility) -// Same for DiscardVirtualMemory. (hide MEM_EXTENDED_PARAMETER to compile with older SDK's) -typedef PVOID(__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, /* MEM_EXTENDED_PARAMETER* */ void*, ULONG); -typedef DWORD(__stdcall *PDiscardVirtualMemory)(PVOID,SIZE_T); +// NtAllocateVirtualAllocEx is used for huge OS page allocation (1GiB) +// We hide MEM_EXTENDED_PARAMETER to compile with older SDK's. +#include +typedef PVOID (__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, /* MEM_EXTENDED_PARAMETER* */ void*, ULONG); +typedef NTSTATUS (__stdcall *PNtAllocateVirtualMemoryEx)(HANDLE, PVOID*, SIZE_T*, ULONG, ULONG, /* MEM_EXTENDED_PARAMETER* */ PVOID, ULONG); static PVirtualAlloc2 pVirtualAlloc2 = NULL; -static PDiscardVirtualMemory pDiscardVirtualMemory = NULL; +static PNtAllocateVirtualMemoryEx pNtAllocateVirtualMemoryEx = NULL; void _mi_os_init(void) { // get the page size @@ -100,12 +108,16 @@ void _mi_os_init(void) { // use VirtualAlloc2FromApp if possible as it is available to Windows store apps pVirtualAlloc2 = (PVirtualAlloc2)GetProcAddress(hDll, "VirtualAlloc2FromApp"); if (pVirtualAlloc2==NULL) pVirtualAlloc2 = (PVirtualAlloc2)GetProcAddress(hDll, "VirtualAlloc2"); - pDiscardVirtualMemory = (PDiscardVirtualMemory)GetProcAddress(hDll, "DiscardVirtualMemory"); + FreeLibrary(hDll); + } + hDll = LoadLibrary(TEXT("ntdll.dll")); + if (hDll != NULL) { + pNtAllocateVirtualMemoryEx = (PNtAllocateVirtualMemoryEx)GetProcAddress(hDll, "NtAllocateVirtualMemoryEx"); FreeLibrary(hDll); } // Try to see if large OS pages are supported unsigned long err = 0; - bool ok = mi_option_is_enabled(mi_option_large_os_pages); + bool ok = mi_option_is_enabled(mi_option_large_os_pages) || mi_option_is_enabled(mi_option_reserve_huge_os_pages); if (ok) { // To use large pages on Windows, we first need access permission // Set "Lock pages in memory" permission in the group policy editor @@ -161,7 +173,7 @@ void _mi_os_init() { static bool mi_os_mem_free(void* addr, size_t size, mi_stats_t* stats) { - if (addr == NULL || size == 0) return true; + if (addr == NULL || size == 0 || mi_os_is_huge_reserved(addr)) return true; bool err = false; #if defined(_WIN32) err = (VirtualFree(addr, 0, MEM_RELEASE) == 0); @@ -185,25 +197,49 @@ static bool mi_os_mem_free(void* addr, size_t size, mi_stats_t* stats) #ifdef _WIN32 static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment, DWORD flags) { #if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS) + // on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages + if ((size % (uintptr_t)1 << 30) == 0 /* 1GiB multiple */ + && (flags & MEM_LARGE_PAGES) != 0 && (flags & MEM_COMMIT) != 0 + && (addr != NULL || try_alignment == 0 || try_alignment % _mi_os_page_size() == 0) + && pNtAllocateVirtualMemoryEx != NULL) + { + #ifndef MEM_EXTENDED_PARAMETER_NONPAGED_HUGE + #define MEM_EXTENDED_PARAMETER_NONPAGED_HUGE (0x10) + #endif + MEM_EXTENDED_PARAMETER param = { 0, 0 }; + param.Type = 5; // == MemExtendedParameterAttributeFlags; + param.ULong64 = MEM_EXTENDED_PARAMETER_NONPAGED_HUGE; + SIZE_T psize = size; + void* base = addr; + NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags | MEM_RESERVE, PAGE_READWRITE, ¶m, 1); + if (err == 0) { + return base; + } + else { + // else fall back to regular large OS pages + _mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) page instead (error %lx)\n", err); + } + } + + // on modern Windows try use VirtualAlloc2 for aligned allocation if (try_alignment > 0 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) { - // on modern Windows try use VirtualAlloc2 for aligned allocation MEM_ADDRESS_REQUIREMENTS reqs = { 0 }; reqs.Alignment = try_alignment; MEM_EXTENDED_PARAMETER param = { 0 }; param.Type = MemExtendedParameterAddressRequirements; param.Pointer = &reqs; - return (*pVirtualAlloc2)(addr, NULL, size, flags, PAGE_READWRITE, ¶m, 1); + return (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, ¶m, 1); } #endif return VirtualAlloc(addr, size, flags, PAGE_READWRITE); } -static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags) { +static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only) { static volatile uintptr_t large_page_try_ok = 0; void* p = NULL; - if (use_large_os_page(size, try_alignment)) { + if (large_only || use_large_os_page(size, try_alignment)) { uintptr_t try_ok = mi_atomic_read(&large_page_try_ok); - if (try_ok > 0) { + 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. mi_atomic_compare_exchange(&large_page_try_ok, try_ok - 1, try_ok); @@ -211,6 +247,7 @@ static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, else { // large OS pages must always reserve and commit. p = mi_win_virtual_allocx(addr, size, try_alignment, MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE | flags); + if (large_only) return p; // fall back to non-large page allocation on error (`p == NULL`). if (p == NULL) { mi_atomic_write(&large_page_try_ok,10); // on error, don't try again for the next N allocations @@ -237,12 +274,13 @@ static void* mi_wasm_heap_grow(size_t size, size_t try_alignment) { return (void*)aligned_base; } #else -static void* mi_unix_mmapx(size_t size, size_t try_alignment, int protect_flags, int flags, int fd) { +#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 - static volatile intptr_t aligned_base = ((intptr_t)1 << 42); // starting at 4TiB - if (try_alignment <= MI_SEGMENT_SIZE && (size%MI_SEGMENT_SIZE)==0) { + static volatile intptr_t aligned_base = ((intptr_t)4 << 40); // starting at 4TiB + if (addr==NULL && try_alignment <= MI_SEGMENT_SIZE && (size%MI_SEGMENT_SIZE)==0) { intptr_t hint = mi_atomic_add(&aligned_base,size) - size; if (hint%try_alignment == 0) { p = mmap((void*)hint,size,protect_flags,flags,fd,0); @@ -251,12 +289,13 @@ static void* mi_unix_mmapx(size_t size, size_t try_alignment, int protect_flags, } #endif if (p==NULL) { - p = mmap(NULL,size,protect_flags,flags,fd,0); + p = mmap(addr,size,protect_flags,flags,fd,0); + if (p==MAP_FAILED) p = NULL; } return p; } -static void* mi_unix_mmap(size_t size, size_t try_alignment, int protect_flags) { +static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int protect_flags, bool large_only) { void* p = NULL; #if !defined(MAP_ANONYMOUS) #define MAP_ANONYMOUS MAP_ANON @@ -278,10 +317,10 @@ static void* mi_unix_mmap(size_t size, size_t try_alignment, int protect_flags) // macOS: tracking anonymous page with a specific ID. (All up to 98 are taken officially but LLVM sanitizers had taken 99) fd = VM_MAKE_TAG(100); #endif - if (use_large_os_page(size, try_alignment)) { + if (large_only || use_large_os_page(size, try_alignment)) { static volatile uintptr_t large_page_try_ok = 0; uintptr_t try_ok = mi_atomic_read(&large_page_try_ok); - if (try_ok > 0) { + 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). // Therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times @@ -297,27 +336,32 @@ static void* mi_unix_mmap(size_t size, size_t try_alignment, int protect_flags) #ifdef MAP_HUGETLB lflags |= MAP_HUGETLB; #endif - #ifdef MAP_HUGE_2MB - lflags |= MAP_HUGE_2MB; + #ifdef MAP_HUGE_1GB + if ((size % (uintptr_t)1 << 20) == 0) { + lflags |= MAP_HUGE_1GB; + } + else #endif + { + #ifdef MAP_HUGE_2MB + lflags |= MAP_HUGE_2MB; + #endif + } #ifdef VM_FLAGS_SUPERPAGE_SIZE_2MB lfd |= VM_FLAGS_SUPERPAGE_SIZE_2MB; #endif - if (lflags != flags) { + if (large_only || lflags != flags) { // try large OS page allocation - p = mi_unix_mmapx(size, try_alignment, protect_flags, lflags, lfd); - if (p == MAP_FAILED) { + p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, lflags, lfd); + if (large_only) return p; + if (p == NULL) { mi_atomic_write(&large_page_try_ok, 10); // on error, don't try again for the next N allocations - p = NULL; // and fall back to regular mmap } } } } if (p == NULL) { - p = mi_unix_mmapx(size, try_alignment, protect_flags, flags, fd); - if (p == MAP_FAILED) { - p = NULL; - } + 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 (TPH). It is not required to call `madvise` with MADV_HUGE @@ -325,7 +369,7 @@ static void* mi_unix_mmap(size_t size, size_t try_alignment, int protect_flags) // in that case -- in particular for our large regions (in `memory.c`). // However, some systems only allow TPH if called with explicit `madvise`, so // when large OS pages are enabled for mimalloc, we call `madvice` anyways. - else if (use_large_os_page(size, try_alignment)) { + if (use_large_os_page(size, try_alignment)) { madvise(p, size, MADV_HUGEPAGE); } #endif @@ -340,17 +384,19 @@ static void* mi_os_mem_alloc(size_t size, size_t try_alignment, bool commit, mi_ mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0); if (size == 0) return NULL; - void* p = NULL; -#if defined(_WIN32) - int flags = MEM_RESERVE; - if (commit) flags |= MEM_COMMIT; - p = mi_win_virtual_alloc(NULL, size, try_alignment, flags); -#elif defined(__wasi__) - p = mi_wasm_heap_grow(size, try_alignment); -#else - int protect_flags = (commit ? (PROT_WRITE | PROT_READ) : PROT_NONE); - p = mi_unix_mmap(size, try_alignment, protect_flags); -#endif + void* p = mi_os_alloc_from_huge_reserved(size, try_alignment, commit); + if (p != NULL) return p; + + #if defined(_WIN32) + int flags = MEM_RESERVE; + if (commit) flags |= MEM_COMMIT; + p = mi_win_virtual_alloc(NULL, size, try_alignment, flags, false); + #elif defined(__wasi__) + p = mi_wasm_heap_grow(size, try_alignment); + #else + int protect_flags = (commit ? (PROT_WRITE | PROT_READ) : PROT_NONE); + p = mi_unix_mmap(NULL, size, try_alignment, protect_flags, false); + #endif _mi_stat_increase(&stats->mmap_calls, 1); if (p != NULL) { _mi_stat_increase(&stats->reserved, size); @@ -399,7 +445,7 @@ static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, // otherwise free and allocate at an aligned address in there mi_os_mem_free(p, over_size, stats); void* aligned_p = mi_align_up_ptr(p, alignment); - p = mi_win_virtual_alloc(aligned_p, size, alignment, flags); + p = mi_win_virtual_alloc(aligned_p, size, alignment, flags, false); if (p == aligned_p) break; // success! if (p != NULL) { // should not happen? mi_os_mem_free(p, size, stats); @@ -555,17 +601,9 @@ static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats) #if defined(_WIN32) // Testing shows that for us (on `malloc-large`) MEM_RESET is 2x faster than DiscardVirtualMemory - // (but this is for an access pattern that immediately reuses the memory) - if (mi_option_is_enabled(mi_option_reset_discards) && pDiscardVirtualMemory != NULL) { - DWORD ok = (*pDiscardVirtualMemory)(start, csize); - mi_assert_internal(ok == ERROR_SUCCESS); - if (ok != ERROR_SUCCESS) return false; - } - else { - void* p = VirtualAlloc(start, csize, MEM_RESET, PAGE_READWRITE); - mi_assert_internal(p == start); - if (p != start) return false; - } + void* p = VirtualAlloc(start, csize, MEM_RESET, PAGE_READWRITE); + mi_assert_internal(p == start); + if (p != start) return false; #else #if defined(MADV_FREE) static int advice = MADV_FREE; @@ -664,3 +702,127 @@ bool _mi_os_shrink(void* p, size_t oldsize, size_t newsize, mi_stats_t* stats) { return mi_os_mem_free(start, size, stats); #endif } + + +/* ---------------------------------------------------------------------------- + +-----------------------------------------------------------------------------*/ +#define MI_HUGE_OS_PAGE_SIZE ((size_t)1 << 30) // 1GiB + +typedef struct mi_huge_info_s { + uint8_t* start; + ptrdiff_t reserved; + volatile ptrdiff_t used; +} mi_huge_info_t; + +static mi_huge_info_t os_huge_reserved = { NULL, 0, 0 }; + +static bool mi_os_is_huge_reserved(void* p) { + return (os_huge_reserved.start != NULL && + (uint8_t*)p >= os_huge_reserved.start && + (uint8_t*)p < os_huge_reserved.start + os_huge_reserved.reserved); +} + +static void* mi_os_alloc_from_huge_reserved(size_t size, size_t try_alignment, bool commit) +{ + // only allow large aligned allocations + if (size < MI_SEGMENT_SIZE || (size % MI_SEGMENT_SIZE) != 0) return NULL; + if (try_alignment > MI_SEGMENT_SIZE) return NULL; + if (!commit) return NULL; + if (os_huge_reserved.start==NULL) return NULL; + if (mi_atomic_iread(&os_huge_reserved.used) >= os_huge_reserved.reserved) return NULL; // already full + + // always aligned + mi_assert_internal( os_huge_reserved.used % MI_SEGMENT_SIZE == 0 ); + mi_assert_internal( (uintptr_t)os_huge_reserved.start % MI_SEGMENT_SIZE == 0 ); + + // try to reserve space + ptrdiff_t next = mi_atomic_add( &os_huge_reserved.used, (ptrdiff_t)size ); + if (next > os_huge_reserved.reserved) { + // "free" our over-allocation + mi_atomic_add( &os_huge_reserved.used, -((ptrdiff_t)size) ); + return NULL; + } + + // success! + uint8_t* p = os_huge_reserved.start + next - (ptrdiff_t)size; + mi_assert_internal( (uintptr_t)p % MI_SEGMENT_SIZE == 0 ); + return p; +} + +/* +static void mi_os_free_huge_reserved() { + uint8_t* addr = os_huge_reserved.start; + size_t total = os_huge_reserved.reserved; + os_huge_reserved.reserved = 0; + os_huge_reserved.start = NULL; + for( size_t current = 0; current < total; current += MI_HUGE_OS_PAGE_SIZE) { + _mi_os_free(addr + current, MI_HUGE_OS_PAGE_SIZE, &_mi_stats_main); + } +} +*/ + +#if !(MI_INTPTR_SIZE >= 8 && (defined(_WIN32) || defined(MI_OS_USE_MMAP))) +int mi_reserve_huge_os_pages(size_t pages, size_t max_secs) { + return -2; // cannot allocate +} +#else +int mi_reserve_huge_os_pages( size_t pages, double max_secs ) mi_attr_noexcept +{ + if (max_secs==0) return -1; // timeout + if (pages==0) return 0; // ok + if (os_huge_reserved.start != NULL) return -2; // already reserved + + // Allocate one page at the time but try to place them contiguously + // We allocate one page at the time to be able to abort if it takes too long + double start_t = _mi_clock_start(); + uint8_t* start = (uint8_t*)((uintptr_t)8 << 40); // 8TiB virtual start address + uint8_t* addr = start; // current top of the allocations + for (size_t page = 0; page < pages; page++, addr += MI_HUGE_OS_PAGE_SIZE ) { + // allocate lorgu pages + void* p = NULL; + #ifdef _WIN32 + p = mi_win_virtual_alloc(addr, MI_HUGE_OS_PAGE_SIZE, 0, MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE, true); + #elif defined(MI_OS_USE_MMAP) + p = mi_unix_mmap(addr, MI_HUGE_OS_PAGE_SIZE, 0, PROT_READ | PROT_WRITE, true); + #else + // always fail + #endif + + // Did we succeed at a contiguous address? + if (p != addr) { + if (p != NULL) { + _mi_warning_message("could not allocate contiguous huge page %zu at 0x%p\n", page, addr); + _mi_os_free(p, MI_HUGE_OS_PAGE_SIZE, &_mi_stats_main ); + } + else { + #ifdef _WIN32 + int err = GetLastError(); + #else + int err = errno; + #endif + _mi_warning_message("could not allocate huge page %zu at 0x%p, error: %i\n", page, addr, err); + } + return -2; + } + // success, record it + if (page==0) { + os_huge_reserved.start = addr; + } + os_huge_reserved.reserved += MI_HUGE_OS_PAGE_SIZE; + _mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE); + _mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE); + + // check for timeout + double elapsed = _mi_clock_end(start_t); + if (elapsed > max_secs) return (-1); // timeout + if (page >= 1) { + double estimate = ((elapsed / (double)(page+1)) * (double)pages); + if (estimate > 1.5*max_secs) return (-1); // seems like we are going to timeout + } + } + _mi_verbose_message("reserved %zu huge pages\n", pages); + return 0; +} +#endif + diff --git a/src/page-queue.c b/src/page-queue.c index f396e233..c53edf82 100644 --- a/src/page-queue.c +++ b/src/page-queue.c @@ -34,15 +34,15 @@ terms of the MIT license. A copy of the license can be found in the file static inline bool mi_page_queue_is_huge(const mi_page_queue_t* pq) { - return (pq->block_size == (MI_MEDIUM_SIZE_MAX+sizeof(uintptr_t))); + return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+sizeof(uintptr_t))); } static inline bool mi_page_queue_is_full(const mi_page_queue_t* pq) { - return (pq->block_size == (MI_MEDIUM_SIZE_MAX+(2*sizeof(uintptr_t)))); + return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+(2*sizeof(uintptr_t)))); } static inline bool mi_page_queue_is_special(const mi_page_queue_t* pq) { - return (pq->block_size > MI_MEDIUM_SIZE_MAX); + return (pq->block_size > MI_MEDIUM_OBJ_SIZE_MAX); } /* ----------------------------------------------------------- @@ -116,11 +116,11 @@ extern inline uint8_t _mi_bin(size_t size) { bin = (uint8_t)wsize; } #endif - else if (wsize > MI_MEDIUM_WSIZE_MAX) { + else if (wsize > MI_MEDIUM_OBJ_WSIZE_MAX) { bin = MI_BIN_HUGE; } else { - #if defined(MI_ALIGN4W) + #if defined(MI_ALIGN4W) if (wsize <= 16) { wsize = (wsize+3)&~3; } // round to 4x word sizes #endif wsize--; @@ -147,7 +147,7 @@ size_t _mi_bin_size(uint8_t bin) { // Good size for allocation size_t mi_good_size(size_t size) mi_attr_noexcept { - if (size <= MI_MEDIUM_SIZE_MAX) { + if (size <= MI_MEDIUM_OBJ_SIZE_MAX) { return _mi_bin_size(_mi_bin(size)); } else { @@ -245,7 +245,7 @@ static bool mi_page_queue_is_empty(mi_page_queue_t* queue) { static void mi_page_queue_remove(mi_page_queue_t* queue, mi_page_t* page) { mi_assert_internal(page != NULL); mi_assert_expensive(mi_page_queue_contains(queue, page)); - mi_assert_internal(page->block_size == queue->block_size || (page->block_size > MI_MEDIUM_SIZE_MAX && mi_page_queue_is_huge(queue)) || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue))); + mi_assert_internal(page->block_size == queue->block_size || (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(queue)) || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue))); if (page->prev != NULL) page->prev->next = page->next; if (page->next != NULL) page->next->prev = page->prev; if (page == queue->last) queue->last = page->prev; @@ -268,7 +268,7 @@ static void mi_page_queue_push(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_ mi_assert_internal(page->heap == NULL); mi_assert_internal(!mi_page_queue_contains(queue, page)); mi_assert_internal(page->block_size == queue->block_size || - (page->block_size > MI_MEDIUM_SIZE_MAX && mi_page_queue_is_huge(queue)) || + (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(queue)) || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue))); mi_page_set_in_full(page, mi_page_queue_is_full(queue)); @@ -297,8 +297,8 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro mi_assert_internal((page->block_size == to->block_size && page->block_size == from->block_size) || (page->block_size == to->block_size && mi_page_queue_is_full(from)) || (page->block_size == from->block_size && mi_page_queue_is_full(to)) || - (page->block_size > MI_MEDIUM_SIZE_MAX && mi_page_queue_is_huge(to)) || - (page->block_size > MI_MEDIUM_SIZE_MAX && mi_page_queue_is_full(to))); + (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(to)) || + (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_full(to))); if (page->prev != NULL) page->prev->next = page->next; if (page->next != NULL) page->next->prev = page->prev; diff --git a/src/page.c b/src/page.c index 554c82be..7ed90232 100644 --- a/src/page.c +++ b/src/page.c @@ -71,7 +71,7 @@ static bool mi_page_is_valid_init(mi_page_t* page) { mi_assert_internal(page->block_size > 0); mi_assert_internal(page->used <= page->capacity); mi_assert_internal(page->capacity <= page->reserved); - + mi_segment_t* segment = _mi_page_segment(page); uint8_t* start = _mi_page_start(segment,page,NULL); mi_assert_internal(start == _mi_segment_page_start(segment,page,NULL)); @@ -102,7 +102,7 @@ bool _mi_page_is_valid(mi_page_t* page) { mi_assert_internal(!_mi_process_is_initialized || segment->thread_id==0 || segment->thread_id == page->heap->thread_id); mi_page_queue_t* pq = mi_page_queue_of(page); mi_assert_internal(mi_page_queue_contains(pq, page)); - mi_assert_internal(pq->block_size==page->block_size || page->block_size > MI_MEDIUM_SIZE_MAX || mi_page_is_in_full(page)); + mi_assert_internal(pq->block_size==page->block_size || page->block_size > MI_MEDIUM_OBJ_SIZE_MAX || mi_page_is_in_full(page)); mi_assert_internal(mi_heap_contains_queue(page->heap,pq)); } return true; @@ -170,7 +170,7 @@ static void _mi_page_thread_free_collect(mi_page_t* page) void _mi_page_free_collect(mi_page_t* page, bool force) { mi_assert_internal(page!=NULL); - + // collect the thread free list if (force || mi_tf_block(page->thread_free) != NULL) { // quick test to avoid an atomic operation _mi_page_thread_free_collect(page); @@ -193,7 +193,7 @@ void _mi_page_free_collect(mi_page_t* page, bool force) { mi_block_set_next(page, tail, page->free); page->free = page->local_free; page->local_free = NULL; - } + } } mi_assert_internal(!force || page->local_free == NULL); @@ -356,8 +356,8 @@ void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) { mi_page_set_has_aligned(page, false); // account for huge pages here - if (page->block_size > MI_MEDIUM_SIZE_MAX) { - if (page->block_size <= MI_LARGE_SIZE_MAX) { + if (page->block_size > MI_MEDIUM_OBJ_SIZE_MAX) { + if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { _mi_stat_decrease(&page->heap->tld->stats.large, page->block_size); } else { @@ -394,7 +394,7 @@ void _mi_page_retire(mi_page_t* page) { // is the only page left with free blocks. It is not clear // how to check this efficiently though... for now we just check // if its neighbours are almost fully used. - if (mi_likely(page->block_size <= MI_MEDIUM_SIZE_MAX)) { + if (mi_likely(page->block_size <= MI_SMALL_SIZE_MAX)) { if (mi_page_mostly_used(page->prev) && mi_page_mostly_used(page->next)) { _mi_stat_counter_increase(&_mi_stats_main.page_no_retire,1); return; // dont't retire after all @@ -713,7 +713,7 @@ static mi_page_t* mi_large_page_alloc(mi_heap_t* heap, size_t size) { if (page != NULL) { mi_assert_internal(mi_page_immediate_available(page)); mi_assert_internal(page->block_size == block_size); - if (page->block_size <= MI_LARGE_SIZE_MAX) { + if (page->block_size <= MI_LARGE_OBJ_SIZE_MAX) { _mi_stat_increase(&heap->tld->stats.large, block_size); _mi_stat_counter_increase(&heap->tld->stats.large_count, 1); } @@ -740,14 +740,14 @@ void* _mi_malloc_generic(mi_heap_t* heap, size_t size) mi_attr_noexcept // call potential deferred free routines _mi_deferred_free(heap, false); - + // free delayed frees from other threads _mi_heap_delayed_free(heap); - + // huge allocation? mi_page_t* page; - if (mi_unlikely(size > MI_MEDIUM_SIZE_MAX)) { - if (mi_unlikely(size >= (SIZE_MAX - MI_MAX_ALIGN_SIZE))) { + if (mi_unlikely(size > MI_MEDIUM_OBJ_SIZE_MAX)) { + if (mi_unlikely(size > PTRDIFF_MAX)) { page = NULL; } else { diff --git a/src/segment.c b/src/segment.c index fd16e2e9..4381efd9 100644 --- a/src/segment.c +++ b/src/segment.c @@ -21,7 +21,6 @@ static void mi_segment_map_freed_at(const mi_segment_t* segment); /* ----------------------------------------------------------- Segment allocation - In any case the memory for a segment is virtual and only committed on demand (i.e. we are careful to not touch the memory @@ -113,7 +112,7 @@ static void mi_page_queue_enqueue(mi_page_queue_t* pq, mi_page_t* page) { pq->first = page; if (page->next != NULL) page->next->prev = page; else pq->last = page; - page->block_size = 0; // free + page->block_size = 0; // free } static mi_page_queue_t* mi_page_queue_for(size_t slice_count, mi_segments_tld_t* tld) { @@ -159,7 +158,7 @@ static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) { mi_page_queue_t* pq; while(slice < &segment->slices[segment->slice_count]) { mi_assert_internal(slice->slice_count > 0); - mi_assert_internal(slice->slice_offset == 0); + mi_assert_internal(slice->slice_offset == 0); size_t index = mi_slice_index(slice); size_t maxindex = (index + slice->slice_count >= segment->slice_count ? segment->slice_count : index + slice->slice_count) - 1; if (slice->block_size > 0) { // a page in use, all slices need their back offset set @@ -179,7 +178,7 @@ static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) { pq = mi_page_queue_for(slice->slice_count,tld); mi_assert_internal(mi_page_queue_contains(pq,mi_slice_to_page(slice))); } - } + } slice = &segment->slices[maxindex+1]; } mi_assert_internal(slice == &segment->slices[segment->slice_count]); @@ -193,7 +192,7 @@ static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) { ----------------------------------------------------------- */ // Start of the page available memory; can be used on uninitialized pages -uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size) +uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size) { mi_slice_t* slice = mi_page_to_slice((mi_page_t*)page); ptrdiff_t idx = slice - segment->slices; @@ -233,7 +232,7 @@ static size_t mi_segment_size(size_t required, size_t* pre_size, size_t* info_si size_t page_size = _mi_os_page_size(); size_t isize = _mi_align_up(sizeof(mi_segment_t), page_size); size_t guardsize = 0; - + if (mi_option_is_enabled(mi_option_secure)) { // in secure mode, we set up a protected page in between the segment info // and the page data (and one at the end of the segment) @@ -296,7 +295,7 @@ static mi_segment_t* mi_segment_cache_pop(size_t segment_size, mi_segments_tld_t } static bool mi_segment_cache_full(mi_segments_tld_t* tld) { - if (tld->cache_count < MI_SEGMENT_CACHE_MAX + if (tld->cache_count < MI_SEGMENT_CACHE_MAX && tld->cache_count < (1 + (tld->peak_count / MI_SEGMENT_CACHE_FRACTION)) ) { // always allow 1 element cache return false; @@ -339,7 +338,7 @@ void _mi_segment_thread_collect(mi_segments_tld_t* tld) { /* ----------------------------------------------------------- - Slices + Slices ----------------------------------------------------------- */ @@ -399,11 +398,11 @@ static void mi_segment_page_split(mi_page_t* page, size_t slice_count, mi_segmen mi_assert_internal(segment->kind != MI_SEGMENT_HUGE); size_t next_index = mi_slice_index(mi_page_to_slice(page)) + slice_count; size_t next_count = page->slice_count - slice_count; - mi_segment_page_init( segment, next_index, next_count, tld ); + mi_segment_page_init( segment, next_index, next_count, tld ); page->slice_count = (uint32_t)slice_count; } -static mi_page_t* mi_segment_page_find(size_t slice_count, mi_segments_tld_t* tld) { +static mi_page_t* mi_segment_page_find(size_t slice_count, mi_segments_tld_t* tld) { mi_assert_internal(slice_count*MI_SEGMENT_SLICE_SIZE <= MI_LARGE_SIZE_MAX); // search from best fit up mi_page_queue_t* pq = mi_page_queue_for(slice_count,tld); @@ -422,8 +421,8 @@ static mi_page_t* mi_segment_page_find(size_t slice_count, mi_segments_tld_t* tl } pq++; } - // could not find a page.. - return NULL; + // could not find a page.. + return NULL; } static void mi_segment_page_delete(mi_slice_t* slice, mi_segments_tld_t* tld) { @@ -452,7 +451,7 @@ static mi_segment_t* mi_segment_alloc(size_t required, mi_segments_tld_t* tld, m //mi_assert_internal(pre_size % MI_SEGMENT_SLICE_SIZE == 0); // Try to get it from our thread local cache first - bool commit = mi_option_is_enabled(mi_option_eager_commit) || mi_option_is_enabled(mi_option_eager_region_commit) + bool commit = mi_option_is_enabled(mi_option_eager_commit) || mi_option_is_enabled(mi_option_eager_region_commit) || required > 0; // huge page mi_segment_t* segment = mi_segment_cache_pop(segment_size, tld); if (segment==NULL) { @@ -482,7 +481,7 @@ static mi_segment_t* mi_segment_alloc(size_t required, mi_segments_tld_t* tld, m _mi_os_protect((uint8_t*)segment + info_size, (pre_size - info_size)); size_t os_page_size = _mi_os_page_size(); // and protect the last page too - _mi_os_protect((uint8_t*)segment + segment_size - os_page_size, os_page_size); + _mi_os_protect((uint8_t*)segment + segment_size - os_page_size, os_page_size); slice_count--; // don't use the last slice :-( } @@ -519,7 +518,7 @@ static mi_segment_t* mi_segment_alloc(size_t required, mi_segments_tld_t* tld, m static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t* tld) { - mi_assert_internal(segment != NULL); + mi_assert_internal(segment != NULL); mi_assert_internal(segment->next == NULL); mi_assert_internal(segment->prev == NULL); mi_assert_internal(segment->used == 0); @@ -541,7 +540,7 @@ static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t // stats _mi_stat_decrease(&tld->stats->page_committed, segment->segment_info_size); - + if (!force && mi_segment_cache_push(segment, tld)) { // it is put in our cache } @@ -555,7 +554,7 @@ static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t Page allocation ----------------------------------------------------------- */ -static mi_page_t* mi_segment_page_alloc(mi_page_kind_t page_kind, size_t required, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) +static mi_page_t* mi_segment_page_alloc(mi_page_kind_t page_kind, size_t required, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) { mi_assert_internal(required <= MI_LARGE_SIZE_MAX && page_kind <= MI_PAGE_LARGE); @@ -565,7 +564,7 @@ static mi_page_t* mi_segment_page_alloc(mi_page_kind_t page_kind, size_t require mi_page_t* page = mi_segment_page_find(slices_needed,tld); //(required <= MI_SMALL_SIZE_MAX ? 0 : slices_needed), tld); if (page==NULL) { // no free page, allocate a new segment and try again - if (mi_segment_alloc(0, tld, os_tld) == NULL) return NULL; // OOM + if (mi_segment_alloc(0, tld, os_tld) == NULL) return NULL; // OOM return mi_segment_page_alloc(page_kind, required, tld, os_tld); } mi_assert_internal(page != NULL && page->slice_count*MI_SEGMENT_SLICE_SIZE == page_size); @@ -581,11 +580,11 @@ static mi_page_t* mi_segment_page_alloc(mi_page_kind_t page_kind, size_t require if (i > 0) slice->slice_count = 0; if (!segment->all_committed && !slice->is_committed) { slice->is_committed = true; - commit = true; + commit = true; } if (slice->is_reset) { slice->is_reset = false; - unreset = true; + unreset = true; } } uint8_t* page_start = mi_slice_start(mi_page_to_slice(page)); @@ -593,7 +592,7 @@ static mi_page_t* mi_segment_page_alloc(mi_page_kind_t page_kind, size_t require if(unreset){ _mi_os_unreset(page_start, page_size, tld->stats); } // initialize the page and return - mi_assert_internal(segment->thread_id == _mi_thread_id()); + mi_assert_internal(segment->thread_id == _mi_thread_id()); segment->used++; mi_page_init_flags(page, segment->thread_id); return page; @@ -604,7 +603,7 @@ static mi_slice_t* mi_segment_page_free_coalesce(mi_page_t* page, mi_segments_tl mi_segment_t* segment = _mi_page_segment(page); mi_assert_internal(segment->used > 0); segment->used--; - + // free and coalesce the page mi_slice_t* slice = mi_page_to_slice(page); size_t slice_count = slice->slice_count; @@ -617,7 +616,7 @@ static mi_slice_t* mi_segment_page_free_coalesce(mi_page_t* page, mi_segments_tl mi_segment_page_delete(next, tld); } if (slice > segment->slices) { - mi_slice_t* prev = mi_slice_first(slice - 1); + mi_slice_t* prev = mi_slice_first(slice - 1); mi_assert_internal(prev >= segment->slices); if (prev->block_size==0) { // free previous slice -- remove it from free and merge @@ -627,7 +626,7 @@ static mi_slice_t* mi_segment_page_free_coalesce(mi_page_t* page, mi_segments_tl slice = prev; } } - + // and add the new free page mi_segment_page_init(segment, mi_slice_index(slice), slice_count, tld); mi_assert_expensive(mi_segment_is_valid(segment,tld)); @@ -649,7 +648,7 @@ static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld size_t inuse = page->capacity * page->block_size; _mi_stat_decrease(&tld->stats->page_committed, inuse); _mi_stat_decrease(&tld->stats->pages, 1); - + // reset the page memory to reduce memory pressure? if (!page->is_reset && mi_option_is_enabled(mi_option_page_reset)) { size_t psize; @@ -696,7 +695,7 @@ void _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld) else if (segment->used == segment->abandoned) { // only abandoned pages; remove from free list and abandon mi_segment_abandon(segment,tld); - } + } } @@ -718,7 +717,7 @@ static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) { mi_assert_expensive(mi_segment_is_valid(segment,tld)); // remove the free pages from our lists - mi_slice_t* slice = &segment->slices[0]; + mi_slice_t* slice = &segment->slices[0]; while (slice <= mi_segment_last_slice(segment)) { mi_assert_internal(slice->slice_count > 0); mi_assert_internal(slice->slice_offset == 0); @@ -807,14 +806,14 @@ bool _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segmen segment->abandoned--; if (mi_page_all_free(page)) { // if everything free by now, free the page - slice = mi_segment_page_clear(page, tld); // set slice again due to coalesceing + slice = mi_segment_page_clear(page, tld); // set slice again due to coalesceing } else { // otherwise reclaim it mi_page_init_flags(page, segment->thread_id); _mi_page_reclaim(heap, page); } - } + } mi_assert_internal(slice->slice_count>0 && slice->slice_offset==0); slice = slice + slice->slice_count; } @@ -824,7 +823,7 @@ bool _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segmen mi_segment_free(segment,false,tld); } else { - reclaimed++; + reclaimed++; } } return (reclaimed>0); @@ -847,15 +846,15 @@ static mi_page_t* mi_segment_huge_page_alloc(size_t size, mi_segments_tld_t* tld page = page + initial_count; page->slice_count = (uint32_t)((segment->segment_size - segment->segment_info_size)/MI_SEGMENT_SLICE_SIZE); page->slice_offset = 0; - page->block_size = size; + page->block_size = size; mi_assert_internal(page->slice_count * MI_SEGMENT_SLICE_SIZE >= size); mi_assert_internal(page->slice_count >= segment->slice_count - initial_count); - // set back pointers + // set back pointers for (size_t i = 1; i slice_count; i++) { mi_slice_t* slice = (mi_slice_t*)(page + i); slice->slice_offset = (uint32_t)(sizeof(mi_page_t)*i); slice->block_size = 1; - slice->slice_count = 0; + slice->slice_count = 0; } mi_page_init_flags(page,segment->thread_id); return page; @@ -873,13 +872,13 @@ static bool mi_is_good_fit(size_t bsize, size_t size) { mi_page_t* _mi_segment_page_alloc(size_t block_size, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) { mi_page_t* page; - if (block_size <= MI_SMALL_SIZE_MAX) {// || mi_is_good_fit(block_size,MI_SMALL_PAGE_SIZE)) { + if (block_size <= MI_SMALL_OBJ_SIZE_MAX) {// || mi_is_good_fit(block_size,MI_SMALL_PAGE_SIZE)) { page = mi_segment_page_alloc(MI_PAGE_SMALL,block_size,tld,os_tld); } - else if (block_size <= MI_MEDIUM_SIZE_MAX) {// || mi_is_good_fit(block_size, MI_MEDIUM_PAGE_SIZE)) { + else if (block_size <= MI_MEDIUM_OBJ_SIZE_MAX) {// || mi_is_good_fit(block_size, MI_MEDIUM_PAGE_SIZE)) { page = mi_segment_page_alloc(MI_PAGE_MEDIUM,MI_MEDIUM_PAGE_SIZE,tld, os_tld); } - else if (block_size <= MI_LARGE_SIZE_MAX) { + else if (block_size <= MI_LARGE_OBJ_SIZE_MAX) { page = mi_segment_page_alloc(MI_PAGE_LARGE,block_size,tld, os_tld); } else { @@ -899,7 +898,7 @@ mi_page_t* _mi_segment_page_alloc(size_t block_size, mi_segments_tld_t* tld, mi_ ----------------------------------------------------------- */ #if (MI_INTPTR_SIZE==8) -#define MI_MAX_ADDRESS ((size_t)1 << 44) // 16TB +#define MI_MAX_ADDRESS ((size_t)1 << 44) // 16TB #else #define MI_MAX_ADDRESS ((size_t)1 << 31) // 2Gb #endif @@ -911,7 +910,7 @@ mi_page_t* _mi_segment_page_alloc(size_t block_size, mi_segments_tld_t* tld, mi_ static volatile uintptr_t mi_segment_map[MI_SEGMENT_MAP_WSIZE]; // 1KiB per TB with 128MiB segments static size_t mi_segment_map_index_of(const mi_segment_t* segment, size_t* bitidx) { - mi_assert_internal(_mi_ptr_segment(segment) == segment); // is it aligned on 128MiB? + mi_assert_internal(_mi_ptr_segment(segment) == segment); // is it aligned on 128MiB? uintptr_t segindex = ((uintptr_t)segment % MI_MAX_ADDRESS) / MI_SEGMENT_SIZE; *bitidx = segindex % (8*MI_INTPTR_SIZE); return (segindex / (8*MI_INTPTR_SIZE)); @@ -953,7 +952,7 @@ static mi_segment_t* _mi_segment_of(const void* p) { return segment; // yes, allocated by us } if (index==0) return NULL; - // search downwards for the first segment in case it is an interior pointer + // search downwards for the first segment in case it is an interior pointer // could be slow but searches in 256MiB steps trough valid huge objects // note: we could maintain a lowest index to speed up the path for invalid pointers? size_t lobitidx; @@ -1006,5 +1005,3 @@ static void* mi_segment_range_of(const void* p, size_t* size) { } } */ - - diff --git a/src/stats.c b/src/stats.c index aa0c393b..8267799f 100644 --- a/src/stats.c +++ b/src/stats.c @@ -28,11 +28,14 @@ void _mi_stats_done(mi_stats_t* stats) { Statistics operations ----------------------------------------------------------- */ +static bool mi_is_in_main(void* stat) { + return ((uint8_t*)stat >= (uint8_t*)&_mi_stats_main + && (uint8_t*)stat < ((uint8_t*)&_mi_stats_main + sizeof(mi_stats_t))); +} + static void mi_stat_update(mi_stat_count_t* stat, int64_t amount) { if (amount == 0) return; - bool in_main = ((uint8_t*)stat >= (uint8_t*)&_mi_stats_main - && (uint8_t*)stat < ((uint8_t*)&_mi_stats_main + sizeof(mi_stats_t))); - if (in_main) + if (mi_is_in_main(stat)) { // add atomically (for abandoned pages) int64_t current = mi_atomic_add(&stat->current,amount); @@ -58,11 +61,16 @@ static void mi_stat_update(mi_stat_count_t* stat, int64_t amount) { } void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount) { - mi_atomic_add( &stat->count, 1 ); - mi_atomic_add( &stat->total, (int64_t)amount ); + if (mi_is_in_main(stat)) { + mi_atomic_add( &stat->count, 1 ); + mi_atomic_add( &stat->total, (int64_t)amount ); + } + else { + stat->count++; + stat->total += amount; + } } - void _mi_stat_increase(mi_stat_count_t* stat, size_t amount) { mi_stat_update(stat, (int64_t)amount); } @@ -276,8 +284,8 @@ static void _mi_stats_print(mi_stats_t* stats, double secs, FILE* out) mi_attr_n _mi_fprintf(out,"\n"); } -static double mi_clock_end(double start); -static double mi_clock_start(void); +double _mi_clock_end(double start); +double _mi_clock_start(void); static double mi_time_start = 0.0; static mi_stats_t* mi_stats_get_default(void) { @@ -289,7 +297,7 @@ void mi_stats_reset(void) mi_attr_noexcept { mi_stats_t* stats = mi_stats_get_default(); if (stats != &_mi_stats_main) { memset(stats, 0, sizeof(mi_stats_t)); } memset(&_mi_stats_main, 0, sizeof(mi_stats_t)); - mi_time_start = mi_clock_start(); + mi_time_start = _mi_clock_start(); } static void mi_stats_print_ex(mi_stats_t* stats, double secs, FILE* out) { @@ -301,11 +309,11 @@ static void mi_stats_print_ex(mi_stats_t* stats, double secs, FILE* out) { } void mi_stats_print(FILE* out) mi_attr_noexcept { - mi_stats_print_ex(mi_stats_get_default(),mi_clock_end(mi_time_start),out); + mi_stats_print_ex(mi_stats_get_default(),_mi_clock_end(mi_time_start),out); } void mi_thread_stats_print(FILE* out) mi_attr_noexcept { - _mi_stats_print(mi_stats_get_default(), mi_clock_end(mi_time_start), out); + _mi_stats_print(mi_stats_get_default(), _mi_clock_end(mi_time_start), out); } @@ -350,7 +358,7 @@ static double mi_clock_now(void) { static double mi_clock_diff = 0.0; -static double mi_clock_start(void) { +double _mi_clock_start(void) { if (mi_clock_diff == 0.0) { double t0 = mi_clock_now(); mi_clock_diff = mi_clock_now() - t0; @@ -358,7 +366,7 @@ static double mi_clock_start(void) { return mi_clock_now(); } -static double mi_clock_end(double start) { +double _mi_clock_end(double start) { double end = mi_clock_now(); return (end - start - mi_clock_diff); } diff --git a/test/test-api.c b/test/test-api.c index f4e32746..a5f61074 100644 --- a/test/test-api.c +++ b/test/test-api.c @@ -66,7 +66,7 @@ bool test_heap2(); // Main testing // --------------------------------------------------------------------------- int main() { - mi_option_enable(mi_option_verbose,false); + mi_option_disable(mi_option_verbose); // --------------------------------------------------- // Malloc diff --git a/test/test-stress.c b/test/test-stress.c index 511679ac..ad487538 100644 --- a/test/test-stress.c +++ b/test/test-stress.c @@ -154,6 +154,9 @@ int main(int argc, char** argv) { if (n > 0) N = n; } printf("start with %i threads with a %i%% load-per-thread\n", THREADS, N); + //int res = mi_reserve_huge_os_pages(4,1); + //printf("(reserve huge: %i\n)", res); + //bench_start_program(); memset((void*)transfer, 0, TRANSFERS*sizeof(void*)); run_os_threads(THREADS);