diff --git a/src/os.c b/src/os.c index f6a3cc86..69eb49df 100644 --- a/src/os.c +++ b/src/os.c @@ -308,7 +308,57 @@ static int mi_madvise(void* addr, size_t length, int advice) { /* ----------------------------------------------------------- - free memory + aligned hinting +-------------------------------------------------------------- */ + +// On 64-bit systems, we can do efficient aligned allocation by using +// the 2TiB to 30TiB area to allocate those. +#if (MI_INTPTR_SIZE >= 8) +static mi_decl_cache_align _Atomic(uintptr_t)aligned_base; + +// Return a MI_SEGMENT_SIZE aligned address that is probably available. +// If this returns NULL, the OS will determine the address but on some OS's that may not be +// properly aligned which can be more costly as it needs to be adjusted afterwards. +// For a size > 1GiB this always returns NULL in order to guarantee good ASLR randomization; +// (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 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 <= 1 || try_alignment > MI_SEGMENT_SIZE) return NULL; + 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 > 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)) % 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 > MI_HINT_MAX but that is ok, it is a hint after all + } + if (hint%try_alignment != 0) return NULL; + return (void*)hint; +} +#else +static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size) { + MI_UNUSED(try_alignment); MI_UNUSED(size); + return NULL; +} +#endif + +/* ----------------------------------------------------------- + Free memory -------------------------------------------------------------- */ static bool mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats_t* stats) @@ -347,9 +397,6 @@ static bool mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats return !err; } -#if !(defined(__wasi__) || defined(MI_USE_SBRK) || defined(MAP_ALIGNED) || MI_INTPTR_SIZE < 8) -static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size); -#endif /* ----------------------------------------------------------- Raw allocation on Windows (VirtualAlloc) @@ -652,53 +699,6 @@ 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 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. -// If this returns NULL, the OS will determine the address but on some OS's that may not be -// properly aligned which can be more costly as it needs to be adjusted afterwards. -// For a size > 1GiB this always returns NULL in order to guarantee good ASLR randomization; -// (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 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 <= 1 || try_alignment > MI_SEGMENT_SIZE) return NULL; - 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 > 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)) % 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 > 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) || defined(MAP_ALIGNED) || (MI_INTPTR_SIZE < 8) -// no need for mi_os_get_aligned_hint -#else -static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size) { - MI_UNUSED(try_alignment); MI_UNUSED(size); - return NULL; -} -#endif /* ----------------------------------------------------------- Primitive allocation from the OS. @@ -799,6 +799,7 @@ static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, return p; } + /* ----------------------------------------------------------- OS API: alloc, free, alloc_aligned ----------------------------------------------------------- */ @@ -826,6 +827,7 @@ 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) { + MI_UNUSED(&mi_os_get_aligned_hint); // suppress unused warnings MI_UNUSED(tld_stats); if (size == 0) return NULL; size = _mi_os_good_alloc_size(size);