| 1 |
/* ---------------------------------------------------------------------------- |
| 2 |
Copyright (c) 2018-2023, Microsoft Research, Daan Leijen |
| 3 |
This is free software; you can redistribute it and/or modify it under the |
| 4 |
terms of the MIT license. A copy of the license can be found in the file |
| 5 |
"LICENSE" at the root of this distribution. |
| 6 |
-----------------------------------------------------------------------------*/ |
| 7 |
#include "mimalloc.h" |
| 8 |
#include "mimalloc/internal.h" |
| 9 |
#include "mimalloc/atomic.h" |
| 10 |
#include "mimalloc/prim.h" |
| 11 |
|
| 12 |
|
| 13 |
/* ----------------------------------------------------------- |
| 14 |
Initialization. |
| 15 |
On windows initializes support for aligned allocation and |
| 16 |
large OS pages (if MIMALLOC_LARGE_OS_PAGES is true). |
| 17 |
----------------------------------------------------------- */ |
| 18 |
|
| 19 |
static mi_os_mem_config_t mi_os_mem_config = { |
| 20 |
4096, // page size |
| 21 |
0, // large page size (usually 2MiB) |
| 22 |
4096, // allocation granularity |
| 23 |
true, // has overcommit? (if true we use MAP_NORESERVE on mmap systems) |
| 24 |
false, // must free whole? (on mmap systems we can free anywhere in a mapped range, but on Windows we must free the entire span) |
| 25 |
true // has virtual reserve? (if true we can reserve virtual address space without using commit or physical memory) |
| 26 |
}; |
| 27 |
|
| 28 |
bool _mi_os_has_overcommit(void) { |
| 29 |
return mi_os_mem_config.has_overcommit; |
| 30 |
} |
| 31 |
|
| 32 |
bool _mi_os_has_virtual_reserve(void) { |
| 33 |
return mi_os_mem_config.has_virtual_reserve; |
| 34 |
} |
| 35 |
|
| 36 |
|
| 37 |
// OS (small) page size |
| 38 |
size_t _mi_os_page_size(void) { |
| 39 |
return mi_os_mem_config.page_size; |
| 40 |
} |
| 41 |
|
| 42 |
// if large OS pages are supported (2 or 4MiB), then return the size, otherwise return the small page size (4KiB) |
| 43 |
size_t _mi_os_large_page_size(void) { |
| 44 |
return (mi_os_mem_config.large_page_size != 0 ? mi_os_mem_config.large_page_size : _mi_os_page_size()); |
| 45 |
} |
| 46 |
|
| 47 |
bool _mi_os_use_large_page(size_t size, size_t alignment) { |
| 48 |
// if we have access, check the size and alignment requirements |
| 49 |
if (mi_os_mem_config.large_page_size == 0 || !mi_option_is_enabled(mi_option_allow_large_os_pages)) return false; |
| 50 |
return ((size % mi_os_mem_config.large_page_size) == 0 && (alignment % mi_os_mem_config.large_page_size) == 0); |
| 51 |
} |
| 52 |
|
| 53 |
// round to a good OS allocation size (bounded by max 12.5% waste) |
| 54 |
size_t _mi_os_good_alloc_size(size_t size) { |
| 55 |
size_t align_size; |
| 56 |
if (size < 512*MI_KiB) align_size = _mi_os_page_size(); |
| 57 |
else if (size < 2*MI_MiB) align_size = 64*MI_KiB; |
| 58 |
else if (size < 8*MI_MiB) align_size = 256*MI_KiB; |
| 59 |
else if (size < 32*MI_MiB) align_size = 1*MI_MiB; |
| 60 |
else align_size = 4*MI_MiB; |
| 61 |
if mi_unlikely(size >= (SIZE_MAX - align_size)) return size; // possible overflow? |
| 62 |
return _mi_align_up(size, align_size); |
| 63 |
} |
| 64 |
|
| 65 |
void _mi_os_init(void) { |
| 66 |
_mi_prim_mem_init(&mi_os_mem_config); |
| 67 |
} |
| 68 |
|
| 69 |
|
| 70 |
/* ----------------------------------------------------------- |
| 71 |
Util |
| 72 |
-------------------------------------------------------------- */ |
| 73 |
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats); |
| 74 |
bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats); |
| 75 |
|
| 76 |
static void* mi_align_up_ptr(void* p, size_t alignment) { |
| 77 |
return (void*)_mi_align_up((uintptr_t)p, alignment); |
| 78 |
} |
| 79 |
|
| 80 |
static void* mi_align_down_ptr(void* p, size_t alignment) { |
| 81 |
return (void*)_mi_align_down((uintptr_t)p, alignment); |
| 82 |
} |
| 83 |
|
| 84 |
|
| 85 |
/* ----------------------------------------------------------- |
| 86 |
aligned hinting |
| 87 |
-------------------------------------------------------------- */ |
| 88 |
|
| 89 |
// On 64-bit systems, we can do efficient aligned allocation by using |
| 90 |
// the 2TiB to 30TiB area to allocate those. |
| 91 |
#if (MI_INTPTR_SIZE >= 8) |
| 92 |
static mi_decl_cache_align _Atomic(uintptr_t)aligned_base; |
| 93 |
|
| 94 |
// Return a MI_SEGMENT_SIZE aligned address that is probably available. |
| 95 |
// If this returns NULL, the OS will determine the address but on some OS's that may not be |
| 96 |
// properly aligned which can be more costly as it needs to be adjusted afterwards. |
| 97 |
// For a size > 1GiB this always returns NULL in order to guarantee good ASLR randomization; |
| 98 |
// (otherwise an initial large allocation of say 2TiB has a 50% chance to include (known) addresses |
| 99 |
// in the middle of the 2TiB - 6TiB address range (see issue #372)) |
| 100 |
|
| 101 |
#define MI_HINT_BASE ((uintptr_t)2 << 40) // 2TiB start |
| 102 |
#define MI_HINT_AREA ((uintptr_t)4 << 40) // upto 6TiB (since before win8 there is "only" 8TiB available to processes) |
| 103 |
#define MI_HINT_MAX ((uintptr_t)30 << 40) // wrap after 30TiB (area after 32TiB is used for huge OS pages) |
| 104 |
|
| 105 |
void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size) |
| 106 |
{ |
| 107 |
if (try_alignment <= 1 || try_alignment > MI_SEGMENT_SIZE) return NULL; |
| 108 |
size = _mi_align_up(size, MI_SEGMENT_SIZE); |
| 109 |
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. |
| 110 |
#if (MI_SECURE>0) |
| 111 |
size += MI_SEGMENT_SIZE; // put in `MI_SEGMENT_SIZE` virtual gaps between hinted blocks; this splits VLA's but increases guarded areas. |
| 112 |
#endif |
| 113 |
|
| 114 |
uintptr_t hint = mi_atomic_add_acq_rel(&aligned_base, size); |
| 115 |
if (hint == 0 || hint > MI_HINT_MAX) { // wrap or initialize |
| 116 |
uintptr_t init = MI_HINT_BASE; |
| 117 |
#if (MI_SECURE>0 || MI_DEBUG==0) // security: randomize start of aligned allocations unless in debug mode |
| 118 |
mi_heap_t* heap = mi_prim_get_default_heap(); |
| 119 |
// gh-123022: default heap may not be initialized in CPython in background threads |
| 120 |
if (mi_heap_is_initialized(heap)) { |
| 121 |
uintptr_t r = _mi_heap_random_next(heap); |
| 122 |
init = init + ((MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)) % MI_HINT_AREA); // (randomly 20 bits)*4MiB == 0 to 4TiB |
| 123 |
} |
| 124 |
#endif |
| 125 |
uintptr_t expected = hint + size; |
| 126 |
mi_atomic_cas_strong_acq_rel(&aligned_base, &expected, init); |
| 127 |
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 |
| 128 |
} |
| 129 |
if (hint%try_alignment != 0) return NULL; |
| 130 |
return (void*)hint; |
| 131 |
} |
| 132 |
#else |
| 133 |
void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size) { |
| 134 |
MI_UNUSED(try_alignment); MI_UNUSED(size); |
| 135 |
return NULL; |
| 136 |
} |
| 137 |
#endif |
| 138 |
|
| 139 |
|
| 140 |
/* ----------------------------------------------------------- |
| 141 |
Free memory |
| 142 |
-------------------------------------------------------------- */ |
| 143 |
|
| 144 |
static void mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats); |
| 145 |
|
| 146 |
static void mi_os_prim_free(void* addr, size_t size, bool still_committed, mi_stats_t* tld_stats) { |
| 147 |
MI_UNUSED(tld_stats); |
| 148 |
mi_assert_internal((size % _mi_os_page_size()) == 0); |
| 149 |
if (addr == NULL || size == 0) return; // || _mi_os_is_huge_reserved(addr) |
| 150 |
int err = _mi_prim_free(addr, size); |
| 151 |
if (err != 0) { |
| 152 |
_mi_warning_message("unable to free OS memory (error: %d (0x%x), size: 0x%zx bytes, address: %p)\n", err, err, size, addr); |
| 153 |
} |
| 154 |
mi_stats_t* stats = &_mi_stats_main; |
| 155 |
if (still_committed) { _mi_stat_decrease(&stats->committed, size); } |
| 156 |
_mi_stat_decrease(&stats->reserved, size); |
| 157 |
} |
| 158 |
|
| 159 |
void _mi_os_free_ex(void* addr, size_t size, bool still_committed, mi_memid_t memid, mi_stats_t* tld_stats) { |
| 160 |
if (mi_memkind_is_os(memid.memkind)) { |
| 161 |
size_t csize = _mi_os_good_alloc_size(size); |
| 162 |
void* base = addr; |
| 163 |
// different base? (due to alignment) |
| 164 |
if (memid.mem.os.base != NULL) { |
| 165 |
mi_assert(memid.mem.os.base <= addr); |
| 166 |
mi_assert((uint8_t*)memid.mem.os.base + memid.mem.os.alignment >= (uint8_t*)addr); |
| 167 |
base = memid.mem.os.base; |
| 168 |
csize += ((uint8_t*)addr - (uint8_t*)memid.mem.os.base); |
| 169 |
} |
| 170 |
// free it |
| 171 |
if (memid.memkind == MI_MEM_OS_HUGE) { |
| 172 |
mi_assert(memid.is_pinned); |
| 173 |
mi_os_free_huge_os_pages(base, csize, tld_stats); |
| 174 |
} |
| 175 |
else { |
| 176 |
mi_os_prim_free(base, csize, still_committed, tld_stats); |
| 177 |
} |
| 178 |
} |
| 179 |
else { |
| 180 |
// nothing to do |
| 181 |
mi_assert(memid.memkind < MI_MEM_OS); |
| 182 |
} |
| 183 |
} |
| 184 |
|
| 185 |
void _mi_os_free(void* p, size_t size, mi_memid_t memid, mi_stats_t* tld_stats) { |
| 186 |
_mi_os_free_ex(p, size, true, memid, tld_stats); |
| 187 |
} |
| 188 |
|
| 189 |
|
| 190 |
/* ----------------------------------------------------------- |
| 191 |
Primitive allocation from the OS. |
| 192 |
-------------------------------------------------------------- */ |
| 193 |
|
| 194 |
// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned. |
| 195 |
static void* mi_os_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* stats) { |
| 196 |
mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0); |
| 197 |
mi_assert_internal(is_zero != NULL); |
| 198 |
mi_assert_internal(is_large != NULL); |
| 199 |
if (size == 0) return NULL; |
| 200 |
if (!commit) { allow_large = false; } |
| 201 |
if (try_alignment == 0) { try_alignment = 1; } // avoid 0 to ensure there will be no divide by zero when aligning |
| 202 |
|
| 203 |
*is_zero = false; |
| 204 |
void* p = NULL; |
| 205 |
int err = _mi_prim_alloc(size, try_alignment, commit, allow_large, is_large, is_zero, &p); |
| 206 |
if (err != 0) { |
| 207 |
_mi_warning_message("unable to allocate OS memory (error: %d (0x%x), size: 0x%zx bytes, align: 0x%zx, commit: %d, allow large: %d)\n", err, err, size, try_alignment, commit, allow_large); |
| 208 |
} |
| 209 |
mi_stat_counter_increase(stats->mmap_calls, 1); |
| 210 |
if (p != NULL) { |
| 211 |
_mi_stat_increase(&stats->reserved, size); |
| 212 |
if (commit) { |
| 213 |
_mi_stat_increase(&stats->committed, size); |
| 214 |
// seems needed for asan (or `mimalloc-test-api` fails) |
| 215 |
#ifdef MI_TRACK_ASAN |
| 216 |
if (*is_zero) { mi_track_mem_defined(p,size); } |
| 217 |
else { mi_track_mem_undefined(p,size); } |
| 218 |
#endif |
| 219 |
} |
| 220 |
} |
| 221 |
return p; |
| 222 |
} |
| 223 |
|
| 224 |
|
| 225 |
// Primitive aligned allocation from the OS. |
| 226 |
// This function guarantees the allocated memory is aligned. |
| 227 |
static void* mi_os_prim_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** base, mi_stats_t* stats) { |
| 228 |
mi_assert_internal(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0)); |
| 229 |
mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0); |
| 230 |
mi_assert_internal(is_large != NULL); |
| 231 |
mi_assert_internal(is_zero != NULL); |
| 232 |
mi_assert_internal(base != NULL); |
| 233 |
if (!commit) allow_large = false; |
| 234 |
if (!(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0))) return NULL; |
| 235 |
size = _mi_align_up(size, _mi_os_page_size()); |
| 236 |
|
| 237 |
// try first with a hint (this will be aligned directly on Win 10+ or BSD) |
| 238 |
void* p = mi_os_prim_alloc(size, alignment, commit, allow_large, is_large, is_zero, stats); |
| 239 |
if (p == NULL) return NULL; |
| 240 |
|
| 241 |
// aligned already? |
| 242 |
if (((uintptr_t)p % alignment) == 0) { |
| 243 |
*base = p; |
| 244 |
} |
| 245 |
else { |
| 246 |
// if not aligned, free it, overallocate, and unmap around it |
| 247 |
// NOTE(sgross): this warning causes issues in Python tests |
| 248 |
// _mi_warning_message("unable to allocate aligned OS memory directly, fall back to over-allocation (size: 0x%zx bytes, address: %p, alignment: 0x%zx, commit: %d)\n", size, p, alignment, commit); |
| 249 |
mi_os_prim_free(p, size, commit, stats); |
| 250 |
if (size >= (SIZE_MAX - alignment)) return NULL; // overflow |
| 251 |
const size_t over_size = size + alignment; |
| 252 |
|
| 253 |
if (mi_os_mem_config.must_free_whole) { // win32 virtualAlloc cannot free parts of an allocate block |
| 254 |
// over-allocate uncommitted (virtual) memory |
| 255 |
p = mi_os_prim_alloc(over_size, 1 /*alignment*/, false /* commit? */, false /* allow_large */, is_large, is_zero, stats); |
| 256 |
if (p == NULL) return NULL; |
| 257 |
|
| 258 |
// set p to the aligned part in the full region |
| 259 |
// note: this is dangerous on Windows as VirtualFree needs the actual base pointer |
| 260 |
// this is handled though by having the `base` field in the memid's |
| 261 |
*base = p; // remember the base |
| 262 |
p = mi_align_up_ptr(p, alignment); |
| 263 |
|
| 264 |
// explicitly commit only the aligned part |
| 265 |
if (commit) { |
| 266 |
_mi_os_commit(p, size, NULL, stats); |
| 267 |
} |
| 268 |
} |
| 269 |
else { // mmap can free inside an allocation |
| 270 |
// overallocate... |
| 271 |
p = mi_os_prim_alloc(over_size, 1, commit, false, is_large, is_zero, stats); |
| 272 |
if (p == NULL) return NULL; |
| 273 |
|
| 274 |
// and selectively unmap parts around the over-allocated area. (noop on sbrk) |
| 275 |
void* aligned_p = mi_align_up_ptr(p, alignment); |
| 276 |
size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p; |
| 277 |
size_t mid_size = _mi_align_up(size, _mi_os_page_size()); |
| 278 |
size_t post_size = over_size - pre_size - mid_size; |
| 279 |
mi_assert_internal(pre_size < over_size&& post_size < over_size&& mid_size >= size); |
| 280 |
if (pre_size > 0) { mi_os_prim_free(p, pre_size, commit, stats); } |
| 281 |
if (post_size > 0) { mi_os_prim_free((uint8_t*)aligned_p + mid_size, post_size, commit, stats); } |
| 282 |
// we can return the aligned pointer on `mmap` (and sbrk) systems |
| 283 |
p = aligned_p; |
| 284 |
*base = aligned_p; // since we freed the pre part, `*base == p`. |
| 285 |
} |
| 286 |
} |
| 287 |
|
| 288 |
mi_assert_internal(p == NULL || (p != NULL && *base != NULL && ((uintptr_t)p % alignment) == 0)); |
| 289 |
return p; |
| 290 |
} |
| 291 |
|
| 292 |
|
| 293 |
/* ----------------------------------------------------------- |
| 294 |
OS API: alloc and alloc_aligned |
| 295 |
----------------------------------------------------------- */ |
| 296 |
|
| 297 |
void* _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* tld_stats) { |
| 298 |
MI_UNUSED(tld_stats); |
| 299 |
*memid = _mi_memid_none(); |
| 300 |
mi_stats_t* stats = &_mi_stats_main; |
| 301 |
if (size == 0) return NULL; |
| 302 |
size = _mi_os_good_alloc_size(size); |
| 303 |
bool os_is_large = false; |
| 304 |
bool os_is_zero = false; |
| 305 |
void* p = mi_os_prim_alloc(size, 0, true, false, &os_is_large, &os_is_zero, stats); |
| 306 |
if (p != NULL) { |
| 307 |
*memid = _mi_memid_create_os(true, os_is_zero, os_is_large); |
| 308 |
} |
| 309 |
return p; |
| 310 |
} |
| 311 |
|
| 312 |
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats) |
| 313 |
{ |
| 314 |
MI_UNUSED(&_mi_os_get_aligned_hint); // suppress unused warnings |
| 315 |
MI_UNUSED(tld_stats); |
| 316 |
*memid = _mi_memid_none(); |
| 317 |
if (size == 0) return NULL; |
| 318 |
size = _mi_os_good_alloc_size(size); |
| 319 |
alignment = _mi_align_up(alignment, _mi_os_page_size()); |
| 320 |
|
| 321 |
bool os_is_large = false; |
| 322 |
bool os_is_zero = false; |
| 323 |
void* os_base = NULL; |
| 324 |
void* p = mi_os_prim_alloc_aligned(size, alignment, commit, allow_large, &os_is_large, &os_is_zero, &os_base, &_mi_stats_main /*tld->stats*/ ); |
| 325 |
if (p != NULL) { |
| 326 |
*memid = _mi_memid_create_os(commit, os_is_zero, os_is_large); |
| 327 |
memid->mem.os.base = os_base; |
| 328 |
memid->mem.os.alignment = alignment; |
| 329 |
} |
| 330 |
return p; |
| 331 |
} |
| 332 |
|
| 333 |
/* ----------------------------------------------------------- |
| 334 |
OS aligned allocation with an offset. This is used |
| 335 |
for large alignments > MI_ALIGNMENT_MAX. We use a large mimalloc |
| 336 |
page where the object can be aligned at an offset from the start of the segment. |
| 337 |
As we may need to overallocate, we need to free such pointers using `mi_free_aligned` |
| 338 |
to use the actual start of the memory region. |
| 339 |
----------------------------------------------------------- */ |
| 340 |
|
| 341 |
void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offset, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats) { |
| 342 |
mi_assert(offset <= MI_SEGMENT_SIZE); |
| 343 |
mi_assert(offset <= size); |
| 344 |
mi_assert((alignment % _mi_os_page_size()) == 0); |
| 345 |
*memid = _mi_memid_none(); |
| 346 |
if (offset > MI_SEGMENT_SIZE) return NULL; |
| 347 |
if (offset == 0) { |
| 348 |
// regular aligned allocation |
| 349 |
return _mi_os_alloc_aligned(size, alignment, commit, allow_large, memid, tld_stats); |
| 350 |
} |
| 351 |
else { |
| 352 |
// overallocate to align at an offset |
| 353 |
const size_t extra = _mi_align_up(offset, alignment) - offset; |
| 354 |
const size_t oversize = size + extra; |
| 355 |
void* const start = _mi_os_alloc_aligned(oversize, alignment, commit, allow_large, memid, tld_stats); |
| 356 |
if (start == NULL) return NULL; |
| 357 |
|
| 358 |
void* const p = (uint8_t*)start + extra; |
| 359 |
mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment)); |
| 360 |
// decommit the overallocation at the start |
| 361 |
if (commit && extra > _mi_os_page_size()) { |
| 362 |
_mi_os_decommit(start, extra, tld_stats); |
| 363 |
} |
| 364 |
return p; |
| 365 |
} |
| 366 |
} |
| 367 |
|
| 368 |
/* ----------------------------------------------------------- |
| 369 |
OS memory API: reset, commit, decommit, protect, unprotect. |
| 370 |
----------------------------------------------------------- */ |
| 371 |
|
| 372 |
// OS page align within a given area, either conservative (pages inside the area only), |
| 373 |
// or not (straddling pages outside the area is possible) |
| 374 |
static void* mi_os_page_align_areax(bool conservative, void* addr, size_t size, size_t* newsize) { |
| 375 |
mi_assert(addr != NULL && size > 0); |
| 376 |
if (newsize != NULL) *newsize = 0; |
| 377 |
if (size == 0 || addr == NULL) return NULL; |
| 378 |
|
| 379 |
// page align conservatively within the range |
| 380 |
void* start = (conservative ? mi_align_up_ptr(addr, _mi_os_page_size()) |
| 381 |
: mi_align_down_ptr(addr, _mi_os_page_size())); |
| 382 |
void* end = (conservative ? mi_align_down_ptr((uint8_t*)addr + size, _mi_os_page_size()) |
| 383 |
: mi_align_up_ptr((uint8_t*)addr + size, _mi_os_page_size())); |
| 384 |
ptrdiff_t diff = (uint8_t*)end - (uint8_t*)start; |
| 385 |
if (diff <= 0) return NULL; |
| 386 |
|
| 387 |
mi_assert_internal((conservative && (size_t)diff <= size) || (!conservative && (size_t)diff >= size)); |
| 388 |
if (newsize != NULL) *newsize = (size_t)diff; |
| 389 |
return start; |
| 390 |
} |
| 391 |
|
| 392 |
static void* mi_os_page_align_area_conservative(void* addr, size_t size, size_t* newsize) { |
| 393 |
return mi_os_page_align_areax(true, addr, size, newsize); |
| 394 |
} |
| 395 |
|
| 396 |
bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) { |
| 397 |
MI_UNUSED(tld_stats); |
| 398 |
mi_stats_t* stats = &_mi_stats_main; |
| 399 |
if (is_zero != NULL) { *is_zero = false; } |
| 400 |
_mi_stat_increase(&stats->committed, size); // use size for precise commit vs. decommit |
| 401 |
_mi_stat_counter_increase(&stats->commit_calls, 1); |
| 402 |
|
| 403 |
// page align range |
| 404 |
size_t csize; |
| 405 |
void* start = mi_os_page_align_areax(false /* conservative? */, addr, size, &csize); |
| 406 |
if (csize == 0) return true; |
| 407 |
|
| 408 |
// commit |
| 409 |
bool os_is_zero = false; |
| 410 |
int err = _mi_prim_commit(start, csize, &os_is_zero); |
| 411 |
if (err != 0) { |
| 412 |
_mi_warning_message("cannot commit OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize); |
| 413 |
return false; |
| 414 |
} |
| 415 |
if (os_is_zero && is_zero != NULL) { |
| 416 |
*is_zero = true; |
| 417 |
mi_assert_expensive(mi_mem_is_zero(start, csize)); |
| 418 |
} |
| 419 |
// note: the following seems required for asan (otherwise `mimalloc-test-stress` fails) |
| 420 |
#ifdef MI_TRACK_ASAN |
| 421 |
if (os_is_zero) { mi_track_mem_defined(start,csize); } |
| 422 |
else { mi_track_mem_undefined(start,csize); } |
| 423 |
#endif |
| 424 |
return true; |
| 425 |
} |
| 426 |
|
| 427 |
static bool mi_os_decommit_ex(void* addr, size_t size, bool* needs_recommit, mi_stats_t* tld_stats) { |
| 428 |
MI_UNUSED(tld_stats); |
| 429 |
mi_stats_t* stats = &_mi_stats_main; |
| 430 |
mi_assert_internal(needs_recommit!=NULL); |
| 431 |
_mi_stat_decrease(&stats->committed, size); |
| 432 |
|
| 433 |
// page align |
| 434 |
size_t csize; |
| 435 |
void* start = mi_os_page_align_area_conservative(addr, size, &csize); |
| 436 |
if (csize == 0) return true; |
| 437 |
|
| 438 |
// decommit |
| 439 |
*needs_recommit = true; |
| 440 |
int err = _mi_prim_decommit(start,csize,needs_recommit); |
| 441 |
if (err != 0) { |
| 442 |
_mi_warning_message("cannot decommit OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize); |
| 443 |
} |
| 444 |
mi_assert_internal(err == 0); |
| 445 |
return (err == 0); |
| 446 |
} |
| 447 |
|
| 448 |
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* tld_stats) { |
| 449 |
bool needs_recommit; |
| 450 |
return mi_os_decommit_ex(addr, size, &needs_recommit, tld_stats); |
| 451 |
} |
| 452 |
|
| 453 |
|
| 454 |
// Signal to the OS that the address range is no longer in use |
| 455 |
// but may be used later again. This will release physical memory |
| 456 |
// pages and reduce swapping while keeping the memory committed. |
| 457 |
// We page align to a conservative area inside the range to reset. |
| 458 |
bool _mi_os_reset(void* addr, size_t size, mi_stats_t* stats) { |
| 459 |
// page align conservatively within the range |
| 460 |
size_t csize; |
| 461 |
void* start = mi_os_page_align_area_conservative(addr, size, &csize); |
| 462 |
if (csize == 0) return true; // || _mi_os_is_huge_reserved(addr) |
| 463 |
_mi_stat_increase(&stats->reset, csize); |
| 464 |
_mi_stat_counter_increase(&stats->reset_calls, 1); |
| 465 |
|
| 466 |
#if (MI_DEBUG>1) && !MI_SECURE && !MI_TRACK_ENABLED // && !MI_TSAN |
| 467 |
memset(start, 0, csize); // pretend it is eagerly reset |
| 468 |
#endif |
| 469 |
|
| 470 |
int err = _mi_prim_reset(start, csize); |
| 471 |
if (err != 0) { |
| 472 |
_mi_warning_message("cannot reset OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize); |
| 473 |
} |
| 474 |
return (err == 0); |
| 475 |
} |
| 476 |
|
| 477 |
|
| 478 |
// either resets or decommits memory, returns true if the memory needs |
| 479 |
// to be recommitted if it is to be re-used later on. |
| 480 |
bool _mi_os_purge_ex(void* p, size_t size, bool allow_reset, mi_stats_t* stats) |
| 481 |
{ |
| 482 |
if (mi_option_get(mi_option_purge_delay) < 0) return false; // is purging allowed? |
| 483 |
_mi_stat_counter_increase(&stats->purge_calls, 1); |
| 484 |
_mi_stat_increase(&stats->purged, size); |
| 485 |
|
| 486 |
if (mi_option_is_enabled(mi_option_purge_decommits) && // should decommit? |
| 487 |
!_mi_preloading()) // don't decommit during preloading (unsafe) |
| 488 |
{ |
| 489 |
bool needs_recommit = true; |
| 490 |
mi_os_decommit_ex(p, size, &needs_recommit, stats); |
| 491 |
return needs_recommit; |
| 492 |
} |
| 493 |
else { |
| 494 |
if (allow_reset) { // this can sometimes be not allowed if the range is not fully committed |
| 495 |
_mi_os_reset(p, size, stats); |
| 496 |
} |
| 497 |
return false; // needs no recommit |
| 498 |
} |
| 499 |
} |
| 500 |
|
| 501 |
// either resets or decommits memory, returns true if the memory needs |
| 502 |
// to be recommitted if it is to be re-used later on. |
| 503 |
bool _mi_os_purge(void* p, size_t size, mi_stats_t * stats) { |
| 504 |
return _mi_os_purge_ex(p, size, true, stats); |
| 505 |
} |
| 506 |
|
| 507 |
// Protect a region in memory to be not accessible. |
| 508 |
static bool mi_os_protectx(void* addr, size_t size, bool protect) { |
| 509 |
// page align conservatively within the range |
| 510 |
size_t csize = 0; |
| 511 |
void* start = mi_os_page_align_area_conservative(addr, size, &csize); |
| 512 |
if (csize == 0) return false; |
| 513 |
/* |
| 514 |
if (_mi_os_is_huge_reserved(addr)) { |
| 515 |
_mi_warning_message("cannot mprotect memory allocated in huge OS pages\n"); |
| 516 |
} |
| 517 |
*/ |
| 518 |
int err = _mi_prim_protect(start,csize,protect); |
| 519 |
if (err != 0) { |
| 520 |
_mi_warning_message("cannot %s OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", (protect ? "protect" : "unprotect"), err, err, start, csize); |
| 521 |
} |
| 522 |
return (err == 0); |
| 523 |
} |
| 524 |
|
| 525 |
bool _mi_os_protect(void* addr, size_t size) { |
| 526 |
return mi_os_protectx(addr, size, true); |
| 527 |
} |
| 528 |
|
| 529 |
bool _mi_os_unprotect(void* addr, size_t size) { |
| 530 |
return mi_os_protectx(addr, size, false); |
| 531 |
} |
| 532 |
|
| 533 |
|
| 534 |
|
| 535 |
/* ---------------------------------------------------------------------------- |
| 536 |
Support for allocating huge OS pages (1Gib) that are reserved up-front |
| 537 |
and possibly associated with a specific NUMA node. (use `numa_node>=0`) |
| 538 |
-----------------------------------------------------------------------------*/ |
| 539 |
#define MI_HUGE_OS_PAGE_SIZE (MI_GiB) |
| 540 |
|
| 541 |
|
| 542 |
#if (MI_INTPTR_SIZE >= 8) |
| 543 |
// To ensure proper alignment, use our own area for huge OS pages |
| 544 |
static mi_decl_cache_align _Atomic(uintptr_t) mi_huge_start; // = 0 |
| 545 |
|
| 546 |
// Claim an aligned address range for huge pages |
| 547 |
static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) { |
| 548 |
if (total_size != NULL) *total_size = 0; |
| 549 |
const size_t size = pages * MI_HUGE_OS_PAGE_SIZE; |
| 550 |
|
| 551 |
uintptr_t start = 0; |
| 552 |
uintptr_t end = 0; |
| 553 |
uintptr_t huge_start = mi_atomic_load_relaxed(&mi_huge_start); |
| 554 |
do { |
| 555 |
start = huge_start; |
| 556 |
if (start == 0) { |
| 557 |
// Initialize the start address after the 32TiB area |
| 558 |
start = ((uintptr_t)32 << 40); // 32TiB virtual start address |
| 559 |
#if (MI_SECURE>0 || MI_DEBUG==0) // security: randomize start of huge pages unless in debug mode |
| 560 |
mi_heap_t* heap = mi_prim_get_default_heap(); |
| 561 |
// gh-123022: default heap may not be initialized in CPython in background threads |
| 562 |
if (mi_heap_is_initialized(heap)) { |
| 563 |
uintptr_t r = _mi_heap_random_next(heap); |
| 564 |
start = start + ((uintptr_t)MI_HUGE_OS_PAGE_SIZE * ((r>>17) & 0x0FFF)); // (randomly 12bits)*1GiB == between 0 to 4TiB |
| 565 |
} |
| 566 |
#endif |
| 567 |
} |
| 568 |
end = start + size; |
| 569 |
mi_assert_internal(end % MI_SEGMENT_SIZE == 0); |
| 570 |
} while (!mi_atomic_cas_strong_acq_rel(&mi_huge_start, &huge_start, end)); |
| 571 |
|
| 572 |
if (total_size != NULL) *total_size = size; |
| 573 |
return (uint8_t*)start; |
| 574 |
} |
| 575 |
#else |
| 576 |
static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) { |
| 577 |
MI_UNUSED(pages); |
| 578 |
if (total_size != NULL) *total_size = 0; |
| 579 |
return NULL; |
| 580 |
} |
| 581 |
#endif |
| 582 |
|
| 583 |
// Allocate MI_SEGMENT_SIZE aligned huge pages |
| 584 |
void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_msecs, size_t* pages_reserved, size_t* psize, mi_memid_t* memid) { |
| 585 |
*memid = _mi_memid_none(); |
| 586 |
if (psize != NULL) *psize = 0; |
| 587 |
if (pages_reserved != NULL) *pages_reserved = 0; |
| 588 |
size_t size = 0; |
| 589 |
uint8_t* start = mi_os_claim_huge_pages(pages, &size); |
| 590 |
if (start == NULL) return NULL; // or 32-bit systems |
| 591 |
|
| 592 |
// Allocate one page at the time but try to place them contiguously |
| 593 |
// We allocate one page at the time to be able to abort if it takes too long |
| 594 |
// or to at least allocate as many as available on the system. |
| 595 |
mi_msecs_t start_t = _mi_clock_start(); |
| 596 |
size_t page = 0; |
| 597 |
bool all_zero = true; |
| 598 |
while (page < pages) { |
| 599 |
// allocate a page |
| 600 |
bool is_zero = false; |
| 601 |
void* addr = start + (page * MI_HUGE_OS_PAGE_SIZE); |
| 602 |
void* p = NULL; |
| 603 |
int err = _mi_prim_alloc_huge_os_pages(addr, MI_HUGE_OS_PAGE_SIZE, numa_node, &is_zero, &p); |
| 604 |
if (!is_zero) { all_zero = false; } |
| 605 |
if (err != 0) { |
| 606 |
_mi_warning_message("unable to allocate huge OS page (error: %d (0x%x), address: %p, size: %zx bytes)\n", err, err, addr, MI_HUGE_OS_PAGE_SIZE); |
| 607 |
break; |
| 608 |
} |
| 609 |
|
| 610 |
// Did we succeed at a contiguous address? |
| 611 |
if (p != addr) { |
| 612 |
// no success, issue a warning and break |
| 613 |
if (p != NULL) { |
| 614 |
_mi_warning_message("could not allocate contiguous huge OS page %zu at %p\n", page, addr); |
| 615 |
mi_os_prim_free(p, MI_HUGE_OS_PAGE_SIZE, true, &_mi_stats_main); |
| 616 |
} |
| 617 |
break; |
| 618 |
} |
| 619 |
|
| 620 |
// success, record it |
| 621 |
page++; // increase before timeout check (see issue #711) |
| 622 |
_mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE); |
| 623 |
_mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE); |
| 624 |
|
| 625 |
// check for timeout |
| 626 |
if (max_msecs > 0) { |
| 627 |
mi_msecs_t elapsed = _mi_clock_end(start_t); |
| 628 |
if (page >= 1) { |
| 629 |
mi_msecs_t estimate = ((elapsed / (page+1)) * pages); |
| 630 |
if (estimate > 2*max_msecs) { // seems like we are going to timeout, break |
| 631 |
elapsed = max_msecs + 1; |
| 632 |
} |
| 633 |
} |
| 634 |
if (elapsed > max_msecs) { |
| 635 |
_mi_warning_message("huge OS page allocation timed out (after allocating %zu page(s))\n", page); |
| 636 |
break; |
| 637 |
} |
| 638 |
} |
| 639 |
} |
| 640 |
mi_assert_internal(page*MI_HUGE_OS_PAGE_SIZE <= size); |
| 641 |
if (pages_reserved != NULL) { *pages_reserved = page; } |
| 642 |
if (psize != NULL) { *psize = page * MI_HUGE_OS_PAGE_SIZE; } |
| 643 |
if (page != 0) { |
| 644 |
mi_assert(start != NULL); |
| 645 |
*memid = _mi_memid_create_os(true /* is committed */, all_zero, true /* is_large */); |
| 646 |
memid->memkind = MI_MEM_OS_HUGE; |
| 647 |
mi_assert(memid->is_pinned); |
| 648 |
#ifdef MI_TRACK_ASAN |
| 649 |
if (all_zero) { mi_track_mem_defined(start,size); } |
| 650 |
#endif |
| 651 |
} |
| 652 |
return (page == 0 ? NULL : start); |
| 653 |
} |
| 654 |
|
| 655 |
// free every huge page in a range individually (as we allocated per page) |
| 656 |
// note: needed with VirtualAlloc but could potentially be done in one go on mmap'd systems. |
| 657 |
static void mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats) { |
| 658 |
if (p==NULL || size==0) return; |
| 659 |
uint8_t* base = (uint8_t*)p; |
| 660 |
while (size >= MI_HUGE_OS_PAGE_SIZE) { |
| 661 |
mi_os_prim_free(base, MI_HUGE_OS_PAGE_SIZE, true, stats); |
| 662 |
size -= MI_HUGE_OS_PAGE_SIZE; |
| 663 |
base += MI_HUGE_OS_PAGE_SIZE; |
| 664 |
} |
| 665 |
} |
| 666 |
|
| 667 |
/* ---------------------------------------------------------------------------- |
| 668 |
Support NUMA aware allocation |
| 669 |
-----------------------------------------------------------------------------*/ |
| 670 |
|
| 671 |
_Atomic(size_t) _mi_numa_node_count; // = 0 // cache the node count |
| 672 |
|
| 673 |
size_t _mi_os_numa_node_count_get(void) { |
| 674 |
size_t count = mi_atomic_load_acquire(&_mi_numa_node_count); |
| 675 |
if (count <= 0) { |
| 676 |
long ncount = mi_option_get(mi_option_use_numa_nodes); // given explicitly? |
| 677 |
if (ncount > 0) { |
| 678 |
count = (size_t)ncount; |
| 679 |
} |
| 680 |
else { |
| 681 |
count = _mi_prim_numa_node_count(); // or detect dynamically |
| 682 |
if (count == 0) count = 1; |
| 683 |
} |
| 684 |
mi_atomic_store_release(&_mi_numa_node_count, count); // save it |
| 685 |
_mi_verbose_message("using %zd numa regions\n", count); |
| 686 |
} |
| 687 |
return count; |
| 688 |
} |
| 689 |
|
| 690 |
int _mi_os_numa_node_get(mi_os_tld_t* tld) { |
| 691 |
MI_UNUSED(tld); |
| 692 |
size_t numa_count = _mi_os_numa_node_count(); |
| 693 |
if (numa_count<=1) return 0; // optimize on single numa node systems: always node 0 |
| 694 |
// never more than the node count and >= 0 |
| 695 |
size_t numa_node = _mi_prim_numa_node(); |
| 696 |
if (numa_node >= numa_count) { numa_node = numa_node % numa_count; } |
| 697 |
return (int)numa_node; |
| 698 |
} |