1#define JEMALLOC_ARENA_C_ 2#include "jemalloc/internal/jemalloc_preamble.h" 3#include "jemalloc/internal/jemalloc_internal_includes.h" 4 5#include "jemalloc/internal/assert.h" 6#include "jemalloc/internal/div.h" 7#include "jemalloc/internal/extent_dss.h" 8#include "jemalloc/internal/extent_mmap.h" 9#include "jemalloc/internal/mutex.h" 10#include "jemalloc/internal/rtree.h" 11#include "jemalloc/internal/size_classes.h" 12#include "jemalloc/internal/util.h" 13 14/******************************************************************************/ 15/* Data. */ 16 17/* 18 * Define names for both unininitialized and initialized phases, so that 19 * options and mallctl processing are straightforward. 20 */ 21const char *percpu_arena_mode_names[] = { 22 "percpu", 23 "phycpu", 24 "disabled", 25 "percpu", 26 "phycpu" 27}; 28percpu_arena_mode_t opt_percpu_arena = PERCPU_ARENA_DEFAULT; 29 30ssize_t opt_dirty_decay_ms = DIRTY_DECAY_MS_DEFAULT; 31ssize_t opt_muzzy_decay_ms = MUZZY_DECAY_MS_DEFAULT; 32 33static atomic_zd_t dirty_decay_ms_default; 34static atomic_zd_t muzzy_decay_ms_default; 35 36const uint64_t h_steps[SMOOTHSTEP_NSTEPS] = { 37#define STEP(step, h, x, y) \ 38 h, 39 SMOOTHSTEP 40#undef STEP 41}; 42 43static div_info_t arena_binind_div_info[NBINS]; 44 45/******************************************************************************/ 46/* 47 * Function prototypes for static functions that are referenced prior to 48 * definition. 49 */ 50 51static void arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, 52 arena_decay_t *decay, extents_t *extents, bool all, size_t npages_limit, 53 size_t npages_decay_max, bool is_background_thread); 54static bool arena_decay_dirty(tsdn_t *tsdn, arena_t *arena, 55 bool is_background_thread, bool all); 56static void arena_dalloc_bin_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, 57 bin_t *bin); 58static void arena_bin_lower_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, 59 bin_t *bin); 60 61/******************************************************************************/ 62 63void 64arena_basic_stats_merge(UNUSED tsdn_t *tsdn, arena_t *arena, unsigned *nthreads, 65 const char **dss, ssize_t *dirty_decay_ms, ssize_t *muzzy_decay_ms, 66 size_t *nactive, size_t *ndirty, size_t *nmuzzy) { 67 *nthreads += arena_nthreads_get(arena, false); 68 *dss = dss_prec_names[arena_dss_prec_get(arena)]; 69 *dirty_decay_ms = arena_dirty_decay_ms_get(arena); 70 *muzzy_decay_ms = arena_muzzy_decay_ms_get(arena); 71 *nactive += atomic_load_zu(&arena->nactive, ATOMIC_RELAXED); 72 *ndirty += extents_npages_get(&arena->extents_dirty); 73 *nmuzzy += extents_npages_get(&arena->extents_muzzy); 74} 75 76void 77arena_stats_merge(tsdn_t *tsdn, arena_t *arena, unsigned *nthreads, 78 const char **dss, ssize_t *dirty_decay_ms, ssize_t *muzzy_decay_ms, 79 size_t *nactive, size_t *ndirty, size_t *nmuzzy, arena_stats_t *astats, 80 bin_stats_t *bstats, arena_stats_large_t *lstats) { 81 cassert(config_stats); 82 83 arena_basic_stats_merge(tsdn, arena, nthreads, dss, dirty_decay_ms, 84 muzzy_decay_ms, nactive, ndirty, nmuzzy); 85 86 size_t base_allocated, base_resident, base_mapped, metadata_thp; 87 base_stats_get(tsdn, arena->base, &base_allocated, &base_resident, 88 &base_mapped, &metadata_thp); 89 90 arena_stats_lock(tsdn, &arena->stats); 91 92 arena_stats_accum_zu(&astats->mapped, base_mapped 93 + arena_stats_read_zu(tsdn, &arena->stats, &arena->stats.mapped)); 94 arena_stats_accum_zu(&astats->retained, 95 extents_npages_get(&arena->extents_retained) << LG_PAGE); 96 97 arena_stats_accum_u64(&astats->decay_dirty.npurge, 98 arena_stats_read_u64(tsdn, &arena->stats, 99 &arena->stats.decay_dirty.npurge)); 100 arena_stats_accum_u64(&astats->decay_dirty.nmadvise, 101 arena_stats_read_u64(tsdn, &arena->stats, 102 &arena->stats.decay_dirty.nmadvise)); 103 arena_stats_accum_u64(&astats->decay_dirty.purged, 104 arena_stats_read_u64(tsdn, &arena->stats, 105 &arena->stats.decay_dirty.purged)); 106 107 arena_stats_accum_u64(&astats->decay_muzzy.npurge, 108 arena_stats_read_u64(tsdn, &arena->stats, 109 &arena->stats.decay_muzzy.npurge)); 110 arena_stats_accum_u64(&astats->decay_muzzy.nmadvise, 111 arena_stats_read_u64(tsdn, &arena->stats, 112 &arena->stats.decay_muzzy.nmadvise)); 113 arena_stats_accum_u64(&astats->decay_muzzy.purged, 114 arena_stats_read_u64(tsdn, &arena->stats, 115 &arena->stats.decay_muzzy.purged)); 116 117 arena_stats_accum_zu(&astats->base, base_allocated); 118 arena_stats_accum_zu(&astats->internal, arena_internal_get(arena)); 119 arena_stats_accum_zu(&astats->metadata_thp, metadata_thp); 120 arena_stats_accum_zu(&astats->resident, base_resident + 121 (((atomic_load_zu(&arena->nactive, ATOMIC_RELAXED) + 122 extents_npages_get(&arena->extents_dirty) + 123 extents_npages_get(&arena->extents_muzzy)) << LG_PAGE))); 124 125 for (szind_t i = 0; i < NSIZES - NBINS; i++) { 126 uint64_t nmalloc = arena_stats_read_u64(tsdn, &arena->stats, 127 &arena->stats.lstats[i].nmalloc); 128 arena_stats_accum_u64(&lstats[i].nmalloc, nmalloc); 129 arena_stats_accum_u64(&astats->nmalloc_large, nmalloc); 130 131 uint64_t ndalloc = arena_stats_read_u64(tsdn, &arena->stats, 132 &arena->stats.lstats[i].ndalloc); 133 arena_stats_accum_u64(&lstats[i].ndalloc, ndalloc); 134 arena_stats_accum_u64(&astats->ndalloc_large, ndalloc); 135 136 uint64_t nrequests = arena_stats_read_u64(tsdn, &arena->stats, 137 &arena->stats.lstats[i].nrequests); 138 arena_stats_accum_u64(&lstats[i].nrequests, 139 nmalloc + nrequests); 140 arena_stats_accum_u64(&astats->nrequests_large, 141 nmalloc + nrequests); 142 143 assert(nmalloc >= ndalloc); 144 assert(nmalloc - ndalloc <= SIZE_T_MAX); 145 size_t curlextents = (size_t)(nmalloc - ndalloc); 146 lstats[i].curlextents += curlextents; 147 arena_stats_accum_zu(&astats->allocated_large, 148 curlextents * sz_index2size(NBINS + i)); 149 } 150 151 arena_stats_unlock(tsdn, &arena->stats); 152 153 /* tcache_bytes counts currently cached bytes. */ 154 atomic_store_zu(&astats->tcache_bytes, 0, ATOMIC_RELAXED); 155 malloc_mutex_lock(tsdn, &arena->tcache_ql_mtx); 156 cache_bin_array_descriptor_t *descriptor; 157 ql_foreach(descriptor, &arena->cache_bin_array_descriptor_ql, link) { 158 szind_t i = 0; 159 for (; i < NBINS; i++) { 160 cache_bin_t *tbin = &descriptor->bins_small[i]; 161 arena_stats_accum_zu(&astats->tcache_bytes, 162 tbin->ncached * sz_index2size(i)); 163 } 164 for (; i < nhbins; i++) { 165 cache_bin_t *tbin = &descriptor->bins_large[i]; 166 arena_stats_accum_zu(&astats->tcache_bytes, 167 tbin->ncached * sz_index2size(i)); 168 } 169 } 170 malloc_mutex_prof_read(tsdn, 171 &astats->mutex_prof_data[arena_prof_mutex_tcache_list], 172 &arena->tcache_ql_mtx); 173 malloc_mutex_unlock(tsdn, &arena->tcache_ql_mtx); 174 175#define READ_ARENA_MUTEX_PROF_DATA(mtx, ind) \ 176 malloc_mutex_lock(tsdn, &arena->mtx); \ 177 malloc_mutex_prof_read(tsdn, &astats->mutex_prof_data[ind], \ 178 &arena->mtx); \ 179 malloc_mutex_unlock(tsdn, &arena->mtx); 180 181 /* Gather per arena mutex profiling data. */ 182 READ_ARENA_MUTEX_PROF_DATA(large_mtx, arena_prof_mutex_large); 183 READ_ARENA_MUTEX_PROF_DATA(extent_avail_mtx, 184 arena_prof_mutex_extent_avail) 185 READ_ARENA_MUTEX_PROF_DATA(extents_dirty.mtx, 186 arena_prof_mutex_extents_dirty) 187 READ_ARENA_MUTEX_PROF_DATA(extents_muzzy.mtx, 188 arena_prof_mutex_extents_muzzy) 189 READ_ARENA_MUTEX_PROF_DATA(extents_retained.mtx, 190 arena_prof_mutex_extents_retained) 191 READ_ARENA_MUTEX_PROF_DATA(decay_dirty.mtx, 192 arena_prof_mutex_decay_dirty) 193 READ_ARENA_MUTEX_PROF_DATA(decay_muzzy.mtx, 194 arena_prof_mutex_decay_muzzy) 195 READ_ARENA_MUTEX_PROF_DATA(base->mtx, 196 arena_prof_mutex_base) 197#undef READ_ARENA_MUTEX_PROF_DATA 198 199 nstime_copy(&astats->uptime, &arena->create_time); 200 nstime_update(&astats->uptime); 201 nstime_subtract(&astats->uptime, &arena->create_time); 202 203 for (szind_t i = 0; i < NBINS; i++) { 204 bin_stats_merge(tsdn, &bstats[i], &arena->bins[i]); 205 } 206} 207 208void 209arena_extents_dirty_dalloc(tsdn_t *tsdn, arena_t *arena, 210 extent_hooks_t **r_extent_hooks, extent_t *extent) { 211 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 212 WITNESS_RANK_CORE, 0); 213 214 extents_dalloc(tsdn, arena, r_extent_hooks, &arena->extents_dirty, 215 extent); 216 if (arena_dirty_decay_ms_get(arena) == 0) { 217 arena_decay_dirty(tsdn, arena, false, true); 218 } else { 219 arena_background_thread_inactivity_check(tsdn, arena, false); 220 } 221} 222 223static void * 224arena_slab_reg_alloc(extent_t *slab, const bin_info_t *bin_info) { 225 void *ret; 226 arena_slab_data_t *slab_data = extent_slab_data_get(slab); 227 size_t regind; 228 229 assert(extent_nfree_get(slab) > 0); 230 assert(!bitmap_full(slab_data->bitmap, &bin_info->bitmap_info)); 231 232 regind = bitmap_sfu(slab_data->bitmap, &bin_info->bitmap_info); 233 ret = (void *)((uintptr_t)extent_addr_get(slab) + 234 (uintptr_t)(bin_info->reg_size * regind)); 235 extent_nfree_dec(slab); 236 return ret; 237} 238 239#ifndef JEMALLOC_JET 240static 241#endif 242size_t 243arena_slab_regind(extent_t *slab, szind_t binind, const void *ptr) { 244 size_t diff, regind; 245 246 /* Freeing a pointer outside the slab can cause assertion failure. */ 247 assert((uintptr_t)ptr >= (uintptr_t)extent_addr_get(slab)); 248 assert((uintptr_t)ptr < (uintptr_t)extent_past_get(slab)); 249 /* Freeing an interior pointer can cause assertion failure. */ 250 assert(((uintptr_t)ptr - (uintptr_t)extent_addr_get(slab)) % 251 (uintptr_t)bin_infos[binind].reg_size == 0); 252 253 diff = (size_t)((uintptr_t)ptr - (uintptr_t)extent_addr_get(slab)); 254 255 /* Avoid doing division with a variable divisor. */ 256 regind = div_compute(&arena_binind_div_info[binind], diff); 257 258 assert(regind < bin_infos[binind].nregs); 259 260 return regind; 261} 262 263static void 264arena_slab_reg_dalloc(extent_t *slab, arena_slab_data_t *slab_data, void *ptr) { 265 szind_t binind = extent_szind_get(slab); 266 const bin_info_t *bin_info = &bin_infos[binind]; 267 size_t regind = arena_slab_regind(slab, binind, ptr); 268 269 assert(extent_nfree_get(slab) < bin_info->nregs); 270 /* Freeing an unallocated pointer can cause assertion failure. */ 271 assert(bitmap_get(slab_data->bitmap, &bin_info->bitmap_info, regind)); 272 273 bitmap_unset(slab_data->bitmap, &bin_info->bitmap_info, regind); 274 extent_nfree_inc(slab); 275} 276 277static void 278arena_nactive_add(arena_t *arena, size_t add_pages) { 279 atomic_fetch_add_zu(&arena->nactive, add_pages, ATOMIC_RELAXED); 280} 281 282static void 283arena_nactive_sub(arena_t *arena, size_t sub_pages) { 284 assert(atomic_load_zu(&arena->nactive, ATOMIC_RELAXED) >= sub_pages); 285 atomic_fetch_sub_zu(&arena->nactive, sub_pages, ATOMIC_RELAXED); 286} 287 288static void 289arena_large_malloc_stats_update(tsdn_t *tsdn, arena_t *arena, size_t usize) { 290 szind_t index, hindex; 291 292 cassert(config_stats); 293 294 if (usize < LARGE_MINCLASS) { 295 usize = LARGE_MINCLASS; 296 } 297 index = sz_size2index(usize); 298 hindex = (index >= NBINS) ? index - NBINS : 0; 299 300 arena_stats_add_u64(tsdn, &arena->stats, 301 &arena->stats.lstats[hindex].nmalloc, 1); 302} 303 304static void 305arena_large_dalloc_stats_update(tsdn_t *tsdn, arena_t *arena, size_t usize) { 306 szind_t index, hindex; 307 308 cassert(config_stats); 309 310 if (usize < LARGE_MINCLASS) { 311 usize = LARGE_MINCLASS; 312 } 313 index = sz_size2index(usize); 314 hindex = (index >= NBINS) ? index - NBINS : 0; 315 316 arena_stats_add_u64(tsdn, &arena->stats, 317 &arena->stats.lstats[hindex].ndalloc, 1); 318} 319 320static void 321arena_large_ralloc_stats_update(tsdn_t *tsdn, arena_t *arena, size_t oldusize, 322 size_t usize) { 323 arena_large_dalloc_stats_update(tsdn, arena, oldusize); 324 arena_large_malloc_stats_update(tsdn, arena, usize); 325} 326 327extent_t * 328arena_extent_alloc_large(tsdn_t *tsdn, arena_t *arena, size_t usize, 329 size_t alignment, bool *zero) { 330 extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; 331 332 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 333 WITNESS_RANK_CORE, 0); 334 335 szind_t szind = sz_size2index(usize); 336 size_t mapped_add; 337 bool commit = true; 338 extent_t *extent = extents_alloc(tsdn, arena, &extent_hooks, 339 &arena->extents_dirty, NULL, usize, sz_large_pad, alignment, false, 340 szind, zero, &commit); 341 if (extent == NULL) { 342 extent = extents_alloc(tsdn, arena, &extent_hooks, 343 &arena->extents_muzzy, NULL, usize, sz_large_pad, alignment, 344 false, szind, zero, &commit); 345 } 346 size_t size = usize + sz_large_pad; 347 if (extent == NULL) { 348 extent = extent_alloc_wrapper(tsdn, arena, &extent_hooks, NULL, 349 usize, sz_large_pad, alignment, false, szind, zero, 350 &commit); 351 if (config_stats) { 352 /* 353 * extent may be NULL on OOM, but in that case 354 * mapped_add isn't used below, so there's no need to 355 * conditionlly set it to 0 here. 356 */ 357 mapped_add = size; 358 } 359 } else if (config_stats) { 360 mapped_add = 0; 361 } 362 363 if (extent != NULL) { 364 if (config_stats) { 365 arena_stats_lock(tsdn, &arena->stats); 366 arena_large_malloc_stats_update(tsdn, arena, usize); 367 if (mapped_add != 0) { 368 arena_stats_add_zu(tsdn, &arena->stats, 369 &arena->stats.mapped, mapped_add); 370 } 371 arena_stats_unlock(tsdn, &arena->stats); 372 } 373 arena_nactive_add(arena, size >> LG_PAGE); 374 } 375 376 return extent; 377} 378 379void 380arena_extent_dalloc_large_prep(tsdn_t *tsdn, arena_t *arena, extent_t *extent) { 381 if (config_stats) { 382 arena_stats_lock(tsdn, &arena->stats); 383 arena_large_dalloc_stats_update(tsdn, arena, 384 extent_usize_get(extent)); 385 arena_stats_unlock(tsdn, &arena->stats); 386 } 387 arena_nactive_sub(arena, extent_size_get(extent) >> LG_PAGE); 388} 389 390void 391arena_extent_ralloc_large_shrink(tsdn_t *tsdn, arena_t *arena, extent_t *extent, 392 size_t oldusize) { 393 size_t usize = extent_usize_get(extent); 394 size_t udiff = oldusize - usize; 395 396 if (config_stats) { 397 arena_stats_lock(tsdn, &arena->stats); 398 arena_large_ralloc_stats_update(tsdn, arena, oldusize, usize); 399 arena_stats_unlock(tsdn, &arena->stats); 400 } 401 arena_nactive_sub(arena, udiff >> LG_PAGE); 402} 403 404void 405arena_extent_ralloc_large_expand(tsdn_t *tsdn, arena_t *arena, extent_t *extent, 406 size_t oldusize) { 407 size_t usize = extent_usize_get(extent); 408 size_t udiff = usize - oldusize; 409 410 if (config_stats) { 411 arena_stats_lock(tsdn, &arena->stats); 412 arena_large_ralloc_stats_update(tsdn, arena, oldusize, usize); 413 arena_stats_unlock(tsdn, &arena->stats); 414 } 415 arena_nactive_add(arena, udiff >> LG_PAGE); 416} 417 418static ssize_t 419arena_decay_ms_read(arena_decay_t *decay) { 420 return atomic_load_zd(&decay->time_ms, ATOMIC_RELAXED); 421} 422 423static void 424arena_decay_ms_write(arena_decay_t *decay, ssize_t decay_ms) { 425 atomic_store_zd(&decay->time_ms, decay_ms, ATOMIC_RELAXED); 426} 427 428static void 429arena_decay_deadline_init(arena_decay_t *decay) { 430 /* 431 * Generate a new deadline that is uniformly random within the next 432 * epoch after the current one. 433 */ 434 nstime_copy(&decay->deadline, &decay->epoch); 435 nstime_add(&decay->deadline, &decay->interval); 436 if (arena_decay_ms_read(decay) > 0) { 437 nstime_t jitter; 438 439 nstime_init(&jitter, prng_range_u64(&decay->jitter_state, 440 nstime_ns(&decay->interval))); 441 nstime_add(&decay->deadline, &jitter); 442 } 443} 444 445static bool 446arena_decay_deadline_reached(const arena_decay_t *decay, const nstime_t *time) { 447 return (nstime_compare(&decay->deadline, time) <= 0); 448} 449 450static size_t 451arena_decay_backlog_npages_limit(const arena_decay_t *decay) { 452 uint64_t sum; 453 size_t npages_limit_backlog; 454 unsigned i; 455 456 /* 457 * For each element of decay_backlog, multiply by the corresponding 458 * fixed-point smoothstep decay factor. Sum the products, then divide 459 * to round down to the nearest whole number of pages. 460 */ 461 sum = 0; 462 for (i = 0; i < SMOOTHSTEP_NSTEPS; i++) { 463 sum += decay->backlog[i] * h_steps[i]; 464 } 465 npages_limit_backlog = (size_t)(sum >> SMOOTHSTEP_BFP); 466 467 return npages_limit_backlog; 468} 469 470static void 471arena_decay_backlog_update_last(arena_decay_t *decay, size_t current_npages) { 472 size_t npages_delta = (current_npages > decay->nunpurged) ? 473 current_npages - decay->nunpurged : 0; 474 decay->backlog[SMOOTHSTEP_NSTEPS-1] = npages_delta; 475 476 if (config_debug) { 477 if (current_npages > decay->ceil_npages) { 478 decay->ceil_npages = current_npages; 479 } 480 size_t npages_limit = arena_decay_backlog_npages_limit(decay); 481 assert(decay->ceil_npages >= npages_limit); 482 if (decay->ceil_npages > npages_limit) { 483 decay->ceil_npages = npages_limit; 484 } 485 } 486} 487 488static void 489arena_decay_backlog_update(arena_decay_t *decay, uint64_t nadvance_u64, 490 size_t current_npages) { 491 if (nadvance_u64 >= SMOOTHSTEP_NSTEPS) { 492 memset(decay->backlog, 0, (SMOOTHSTEP_NSTEPS-1) * 493 sizeof(size_t)); 494 } else { 495 size_t nadvance_z = (size_t)nadvance_u64; 496 497 assert((uint64_t)nadvance_z == nadvance_u64); 498 499 memmove(decay->backlog, &decay->backlog[nadvance_z], 500 (SMOOTHSTEP_NSTEPS - nadvance_z) * sizeof(size_t)); 501 if (nadvance_z > 1) { 502 memset(&decay->backlog[SMOOTHSTEP_NSTEPS - 503 nadvance_z], 0, (nadvance_z-1) * sizeof(size_t)); 504 } 505 } 506 507 arena_decay_backlog_update_last(decay, current_npages); 508} 509 510static void 511arena_decay_try_purge(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 512 extents_t *extents, size_t current_npages, size_t npages_limit, 513 bool is_background_thread) { 514 if (current_npages > npages_limit) { 515 arena_decay_to_limit(tsdn, arena, decay, extents, false, 516 npages_limit, current_npages - npages_limit, 517 is_background_thread); 518 } 519} 520 521static void 522arena_decay_epoch_advance_helper(arena_decay_t *decay, const nstime_t *time, 523 size_t current_npages) { 524 assert(arena_decay_deadline_reached(decay, time)); 525 526 nstime_t delta; 527 nstime_copy(&delta, time); 528 nstime_subtract(&delta, &decay->epoch); 529 530 uint64_t nadvance_u64 = nstime_divide(&delta, &decay->interval); 531 assert(nadvance_u64 > 0); 532 533 /* Add nadvance_u64 decay intervals to epoch. */ 534 nstime_copy(&delta, &decay->interval); 535 nstime_imultiply(&delta, nadvance_u64); 536 nstime_add(&decay->epoch, &delta); 537 538 /* Set a new deadline. */ 539 arena_decay_deadline_init(decay); 540 541 /* Update the backlog. */ 542 arena_decay_backlog_update(decay, nadvance_u64, current_npages); 543} 544 545static void 546arena_decay_epoch_advance(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 547 extents_t *extents, const nstime_t *time, bool is_background_thread) { 548 size_t current_npages = extents_npages_get(extents); 549 arena_decay_epoch_advance_helper(decay, time, current_npages); 550 551 size_t npages_limit = arena_decay_backlog_npages_limit(decay); 552 /* We may unlock decay->mtx when try_purge(). Finish logging first. */ 553 decay->nunpurged = (npages_limit > current_npages) ? npages_limit : 554 current_npages; 555 556 if (!background_thread_enabled() || is_background_thread) { 557 arena_decay_try_purge(tsdn, arena, decay, extents, 558 current_npages, npages_limit, is_background_thread); 559 } 560} 561 562static void 563arena_decay_reinit(arena_decay_t *decay, ssize_t decay_ms) { 564 arena_decay_ms_write(decay, decay_ms); 565 if (decay_ms > 0) { 566 nstime_init(&decay->interval, (uint64_t)decay_ms * 567 KQU(1000000)); 568 nstime_idivide(&decay->interval, SMOOTHSTEP_NSTEPS); 569 } 570 571 nstime_init(&decay->epoch, 0); 572 nstime_update(&decay->epoch); 573 decay->jitter_state = (uint64_t)(uintptr_t)decay; 574 arena_decay_deadline_init(decay); 575 decay->nunpurged = 0; 576 memset(decay->backlog, 0, SMOOTHSTEP_NSTEPS * sizeof(size_t)); 577} 578 579static bool 580arena_decay_init(arena_decay_t *decay, ssize_t decay_ms, 581 arena_stats_decay_t *stats) { 582 if (config_debug) { 583 for (size_t i = 0; i < sizeof(arena_decay_t); i++) { 584 assert(((char *)decay)[i] == 0); 585 } 586 decay->ceil_npages = 0; 587 } 588 if (malloc_mutex_init(&decay->mtx, "decay", WITNESS_RANK_DECAY, 589 malloc_mutex_rank_exclusive)) { 590 return true; 591 } 592 decay->purging = false; 593 arena_decay_reinit(decay, decay_ms); 594 /* Memory is zeroed, so there is no need to clear stats. */ 595 if (config_stats) { 596 decay->stats = stats; 597 } 598 return false; 599} 600 601static bool 602arena_decay_ms_valid(ssize_t decay_ms) { 603 if (decay_ms < -1) { 604 return false; 605 } 606 if (decay_ms == -1 || (uint64_t)decay_ms <= NSTIME_SEC_MAX * 607 KQU(1000)) { 608 return true; 609 } 610 return false; 611} 612 613static bool 614arena_maybe_decay(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 615 extents_t *extents, bool is_background_thread) { 616 malloc_mutex_assert_owner(tsdn, &decay->mtx); 617 618 /* Purge all or nothing if the option is disabled. */ 619 ssize_t decay_ms = arena_decay_ms_read(decay); 620 if (decay_ms <= 0) { 621 if (decay_ms == 0) { 622 arena_decay_to_limit(tsdn, arena, decay, extents, false, 623 0, extents_npages_get(extents), 624 is_background_thread); 625 } 626 return false; 627 } 628 629 nstime_t time; 630 nstime_init(&time, 0); 631 nstime_update(&time); 632 if (unlikely(!nstime_monotonic() && nstime_compare(&decay->epoch, &time) 633 > 0)) { 634 /* 635 * Time went backwards. Move the epoch back in time and 636 * generate a new deadline, with the expectation that time 637 * typically flows forward for long enough periods of time that 638 * epochs complete. Unfortunately, this strategy is susceptible 639 * to clock jitter triggering premature epoch advances, but 640 * clock jitter estimation and compensation isn't feasible here 641 * because calls into this code are event-driven. 642 */ 643 nstime_copy(&decay->epoch, &time); 644 arena_decay_deadline_init(decay); 645#ifndef __NetBSD__ 646 } else { 647 /* Verify that time does not go backwards. */ 648 assert(nstime_compare(&decay->epoch, &time) <= 0); 649#endif 650 } 651 652 /* 653 * If the deadline has been reached, advance to the current epoch and 654 * purge to the new limit if necessary. Note that dirty pages created 655 * during the current epoch are not subject to purge until a future 656 * epoch, so as a result purging only happens during epoch advances, or 657 * being triggered by background threads (scheduled event). 658 */ 659 bool advance_epoch = arena_decay_deadline_reached(decay, &time); 660 if (advance_epoch) { 661 arena_decay_epoch_advance(tsdn, arena, decay, extents, &time, 662 is_background_thread); 663 } else if (is_background_thread) { 664 arena_decay_try_purge(tsdn, arena, decay, extents, 665 extents_npages_get(extents), 666 arena_decay_backlog_npages_limit(decay), 667 is_background_thread); 668 } 669 670 return advance_epoch; 671} 672 673static ssize_t 674arena_decay_ms_get(arena_decay_t *decay) { 675 return arena_decay_ms_read(decay); 676} 677 678ssize_t 679arena_dirty_decay_ms_get(arena_t *arena) { 680 return arena_decay_ms_get(&arena->decay_dirty); 681} 682 683ssize_t 684arena_muzzy_decay_ms_get(arena_t *arena) { 685 return arena_decay_ms_get(&arena->decay_muzzy); 686} 687 688static bool 689arena_decay_ms_set(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 690 extents_t *extents, ssize_t decay_ms) { 691 if (!arena_decay_ms_valid(decay_ms)) { 692 return true; 693 } 694 695 malloc_mutex_lock(tsdn, &decay->mtx); 696 /* 697 * Restart decay backlog from scratch, which may cause many dirty pages 698 * to be immediately purged. It would conceptually be possible to map 699 * the old backlog onto the new backlog, but there is no justification 700 * for such complexity since decay_ms changes are intended to be 701 * infrequent, either between the {-1, 0, >0} states, or a one-time 702 * arbitrary change during initial arena configuration. 703 */ 704 arena_decay_reinit(decay, decay_ms); 705 arena_maybe_decay(tsdn, arena, decay, extents, false); 706 malloc_mutex_unlock(tsdn, &decay->mtx); 707 708 return false; 709} 710 711bool 712arena_dirty_decay_ms_set(tsdn_t *tsdn, arena_t *arena, 713 ssize_t decay_ms) { 714 return arena_decay_ms_set(tsdn, arena, &arena->decay_dirty, 715 &arena->extents_dirty, decay_ms); 716} 717 718bool 719arena_muzzy_decay_ms_set(tsdn_t *tsdn, arena_t *arena, 720 ssize_t decay_ms) { 721 return arena_decay_ms_set(tsdn, arena, &arena->decay_muzzy, 722 &arena->extents_muzzy, decay_ms); 723} 724 725static size_t 726arena_stash_decayed(tsdn_t *tsdn, arena_t *arena, 727 extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_limit, 728 size_t npages_decay_max, extent_list_t *decay_extents) { 729 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 730 WITNESS_RANK_CORE, 0); 731 732 /* Stash extents according to npages_limit. */ 733 size_t nstashed = 0; 734 extent_t *extent; 735 while (nstashed < npages_decay_max && 736 (extent = extents_evict(tsdn, arena, r_extent_hooks, extents, 737 npages_limit)) != NULL) { 738 extent_list_append(decay_extents, extent); 739 nstashed += extent_size_get(extent) >> LG_PAGE; 740 } 741 return nstashed; 742} 743 744static size_t 745arena_decay_stashed(tsdn_t *tsdn, arena_t *arena, 746 extent_hooks_t **r_extent_hooks, arena_decay_t *decay, extents_t *extents, 747 bool all, extent_list_t *decay_extents, bool is_background_thread) { 748 UNUSED size_t nmadvise, nunmapped; 749 size_t npurged; 750 751 if (config_stats) { 752 nmadvise = 0; 753 nunmapped = 0; 754 } 755 npurged = 0; 756 757 ssize_t muzzy_decay_ms = arena_muzzy_decay_ms_get(arena); 758 for (extent_t *extent = extent_list_first(decay_extents); extent != 759 NULL; extent = extent_list_first(decay_extents)) { 760 if (config_stats) { 761 nmadvise++; 762 } 763 size_t npages = extent_size_get(extent) >> LG_PAGE; 764 npurged += npages; 765 extent_list_remove(decay_extents, extent); 766 switch (extents_state_get(extents)) { 767 case extent_state_active: 768 not_reached(); 769 case extent_state_dirty: 770 if (!all && muzzy_decay_ms != 0 && 771 !extent_purge_lazy_wrapper(tsdn, arena, 772 r_extent_hooks, extent, 0, 773 extent_size_get(extent))) { 774 extents_dalloc(tsdn, arena, r_extent_hooks, 775 &arena->extents_muzzy, extent); 776 arena_background_thread_inactivity_check(tsdn, 777 arena, is_background_thread); 778 break; 779 } 780 /* Fall through. */ 781 case extent_state_muzzy: 782 extent_dalloc_wrapper(tsdn, arena, r_extent_hooks, 783 extent); 784 if (config_stats) { 785 nunmapped += npages; 786 } 787 break; 788 case extent_state_retained: 789 default: 790 not_reached(); 791 } 792 } 793 794 if (config_stats) { 795 arena_stats_lock(tsdn, &arena->stats); 796 arena_stats_add_u64(tsdn, &arena->stats, &decay->stats->npurge, 797 1); 798 arena_stats_add_u64(tsdn, &arena->stats, 799 &decay->stats->nmadvise, nmadvise); 800 arena_stats_add_u64(tsdn, &arena->stats, &decay->stats->purged, 801 npurged); 802 arena_stats_sub_zu(tsdn, &arena->stats, &arena->stats.mapped, 803 nunmapped << LG_PAGE); 804 arena_stats_unlock(tsdn, &arena->stats); 805 } 806 807 return npurged; 808} 809 810/* 811 * npages_limit: Decay at most npages_decay_max pages without violating the 812 * invariant: (extents_npages_get(extents) >= npages_limit). We need an upper 813 * bound on number of pages in order to prevent unbounded growth (namely in 814 * stashed), otherwise unbounded new pages could be added to extents during the 815 * current decay run, so that the purging thread never finishes. 816 */ 817static void 818arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 819 extents_t *extents, bool all, size_t npages_limit, size_t npages_decay_max, 820 bool is_background_thread) { 821 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 822 WITNESS_RANK_CORE, 1); 823 malloc_mutex_assert_owner(tsdn, &decay->mtx); 824 825 if (decay->purging) { 826 return; 827 } 828 decay->purging = true; 829 malloc_mutex_unlock(tsdn, &decay->mtx); 830 831 extent_hooks_t *extent_hooks = extent_hooks_get(arena); 832 833 extent_list_t decay_extents; 834 extent_list_init(&decay_extents); 835 836 size_t npurge = arena_stash_decayed(tsdn, arena, &extent_hooks, extents, 837 npages_limit, npages_decay_max, &decay_extents); 838 if (npurge != 0) { 839 UNUSED size_t npurged = arena_decay_stashed(tsdn, arena, 840 &extent_hooks, decay, extents, all, &decay_extents, 841 is_background_thread); 842 assert(npurged == npurge); 843 } 844 845 malloc_mutex_lock(tsdn, &decay->mtx); 846 decay->purging = false; 847} 848 849static bool 850arena_decay_impl(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay, 851 extents_t *extents, bool is_background_thread, bool all) { 852 if (all) { 853 malloc_mutex_lock(tsdn, &decay->mtx); 854 arena_decay_to_limit(tsdn, arena, decay, extents, all, 0, 855 extents_npages_get(extents), is_background_thread); 856 malloc_mutex_unlock(tsdn, &decay->mtx); 857 858 return false; 859 } 860 861 if (malloc_mutex_trylock(tsdn, &decay->mtx)) { 862 /* No need to wait if another thread is in progress. */ 863 return true; 864 } 865 866 bool epoch_advanced = arena_maybe_decay(tsdn, arena, decay, extents, 867 is_background_thread); 868 UNUSED size_t npages_new; 869 if (epoch_advanced) { 870 /* Backlog is updated on epoch advance. */ 871 npages_new = decay->backlog[SMOOTHSTEP_NSTEPS-1]; 872 } else 873 npages_new = 0; // XXX: gcc without -O 874 malloc_mutex_unlock(tsdn, &decay->mtx); 875 876 if (have_background_thread && background_thread_enabled() && 877 epoch_advanced && !is_background_thread) { 878 background_thread_interval_check(tsdn, arena, decay, 879 npages_new); 880 } 881 882 return false; 883} 884 885static bool 886arena_decay_dirty(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, 887 bool all) { 888 return arena_decay_impl(tsdn, arena, &arena->decay_dirty, 889 &arena->extents_dirty, is_background_thread, all); 890} 891 892static bool 893arena_decay_muzzy(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, 894 bool all) { 895 return arena_decay_impl(tsdn, arena, &arena->decay_muzzy, 896 &arena->extents_muzzy, is_background_thread, all); 897} 898 899void 900arena_decay(tsdn_t *tsdn, arena_t *arena, bool is_background_thread, bool all) { 901 if (arena_decay_dirty(tsdn, arena, is_background_thread, all)) { 902 return; 903 } 904 arena_decay_muzzy(tsdn, arena, is_background_thread, all); 905} 906 907static void 908arena_slab_dalloc(tsdn_t *tsdn, arena_t *arena, extent_t *slab) { 909 arena_nactive_sub(arena, extent_size_get(slab) >> LG_PAGE); 910 911 extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; 912 arena_extents_dirty_dalloc(tsdn, arena, &extent_hooks, slab); 913} 914 915static void 916arena_bin_slabs_nonfull_insert(bin_t *bin, extent_t *slab) { 917 assert(extent_nfree_get(slab) > 0); 918 extent_heap_insert(&bin->slabs_nonfull, slab); 919} 920 921static void 922arena_bin_slabs_nonfull_remove(bin_t *bin, extent_t *slab) { 923 extent_heap_remove(&bin->slabs_nonfull, slab); 924} 925 926static extent_t * 927arena_bin_slabs_nonfull_tryget(bin_t *bin) { 928 extent_t *slab = extent_heap_remove_first(&bin->slabs_nonfull); 929 if (slab == NULL) { 930 return NULL; 931 } 932 if (config_stats) { 933 bin->stats.reslabs++; 934 } 935 return slab; 936} 937 938static void 939arena_bin_slabs_full_insert(arena_t *arena, bin_t *bin, extent_t *slab) { 940 assert(extent_nfree_get(slab) == 0); 941 /* 942 * Tracking extents is required by arena_reset, which is not allowed 943 * for auto arenas. Bypass this step to avoid touching the extent 944 * linkage (often results in cache misses) for auto arenas. 945 */ 946 if (arena_is_auto(arena)) { 947 return; 948 } 949 extent_list_append(&bin->slabs_full, slab); 950} 951 952static void 953arena_bin_slabs_full_remove(arena_t *arena, bin_t *bin, extent_t *slab) { 954 if (arena_is_auto(arena)) { 955 return; 956 } 957 extent_list_remove(&bin->slabs_full, slab); 958} 959 960void 961arena_reset(tsd_t *tsd, arena_t *arena) { 962 /* 963 * Locking in this function is unintuitive. The caller guarantees that 964 * no concurrent operations are happening in this arena, but there are 965 * still reasons that some locking is necessary: 966 * 967 * - Some of the functions in the transitive closure of calls assume 968 * appropriate locks are held, and in some cases these locks are 969 * temporarily dropped to avoid lock order reversal or deadlock due to 970 * reentry. 971 * - mallctl("epoch", ...) may concurrently refresh stats. While 972 * strictly speaking this is a "concurrent operation", disallowing 973 * stats refreshes would impose an inconvenient burden. 974 */ 975 976 /* Large allocations. */ 977 malloc_mutex_lock(tsd_tsdn(tsd), &arena->large_mtx); 978 979 for (extent_t *extent = extent_list_first(&arena->large); extent != 980 NULL; extent = extent_list_first(&arena->large)) { 981 void *ptr = extent_base_get(extent); 982 size_t usize; 983 984 malloc_mutex_unlock(tsd_tsdn(tsd), &arena->large_mtx); 985 alloc_ctx_t alloc_ctx; 986 rtree_ctx_t *rtree_ctx = tsd_rtree_ctx(tsd); 987 rtree_szind_slab_read(tsd_tsdn(tsd), &extents_rtree, rtree_ctx, 988 (uintptr_t)ptr, true, &alloc_ctx.szind, &alloc_ctx.slab); 989 assert(alloc_ctx.szind != NSIZES); 990 991 if (config_stats || (config_prof && opt_prof)) { 992 usize = sz_index2size(alloc_ctx.szind); 993 assert(usize == isalloc(tsd_tsdn(tsd), ptr)); 994 } 995 /* Remove large allocation from prof sample set. */ 996 if (config_prof && opt_prof) { 997 prof_free(tsd, ptr, usize, &alloc_ctx); 998 } 999 large_dalloc(tsd_tsdn(tsd), extent); 1000 malloc_mutex_lock(tsd_tsdn(tsd), &arena->large_mtx); 1001 } 1002 malloc_mutex_unlock(tsd_tsdn(tsd), &arena->large_mtx); 1003 1004 /* Bins. */ 1005 for (unsigned i = 0; i < NBINS; i++) { 1006 extent_t *slab; 1007 bin_t *bin = &arena->bins[i]; 1008 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 1009 if (bin->slabcur != NULL) { 1010 slab = bin->slabcur; 1011 bin->slabcur = NULL; 1012 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 1013 arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); 1014 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 1015 } 1016 while ((slab = extent_heap_remove_first(&bin->slabs_nonfull)) != 1017 NULL) { 1018 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 1019 arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); 1020 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 1021 } 1022 for (slab = extent_list_first(&bin->slabs_full); slab != NULL; 1023 slab = extent_list_first(&bin->slabs_full)) { 1024 arena_bin_slabs_full_remove(arena, bin, slab); 1025 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 1026 arena_slab_dalloc(tsd_tsdn(tsd), arena, slab); 1027 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 1028 } 1029 if (config_stats) { 1030 bin->stats.curregs = 0; 1031 bin->stats.curslabs = 0; 1032 } 1033 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 1034 } 1035 1036 atomic_store_zu(&arena->nactive, 0, ATOMIC_RELAXED); 1037} 1038 1039static void 1040arena_destroy_retained(tsdn_t *tsdn, arena_t *arena) { 1041 /* 1042 * Iterate over the retained extents and destroy them. This gives the 1043 * extent allocator underlying the extent hooks an opportunity to unmap 1044 * all retained memory without having to keep its own metadata 1045 * structures. In practice, virtual memory for dss-allocated extents is 1046 * leaked here, so best practice is to avoid dss for arenas to be 1047 * destroyed, or provide custom extent hooks that track retained 1048 * dss-based extents for later reuse. 1049 */ 1050 extent_hooks_t *extent_hooks = extent_hooks_get(arena); 1051 extent_t *extent; 1052 while ((extent = extents_evict(tsdn, arena, &extent_hooks, 1053 &arena->extents_retained, 0)) != NULL) { 1054 extent_destroy_wrapper(tsdn, arena, &extent_hooks, extent); 1055 } 1056} 1057 1058void 1059arena_destroy(tsd_t *tsd, arena_t *arena) { 1060 assert(base_ind_get(arena->base) >= narenas_auto); 1061 assert(arena_nthreads_get(arena, false) == 0); 1062 assert(arena_nthreads_get(arena, true) == 0); 1063 1064 /* 1065 * No allocations have occurred since arena_reset() was called. 1066 * Furthermore, the caller (arena_i_destroy_ctl()) purged all cached 1067 * extents, so only retained extents may remain. 1068 */ 1069 assert(extents_npages_get(&arena->extents_dirty) == 0); 1070 assert(extents_npages_get(&arena->extents_muzzy) == 0); 1071 1072 /* Deallocate retained memory. */ 1073 arena_destroy_retained(tsd_tsdn(tsd), arena); 1074 1075 /* 1076 * Remove the arena pointer from the arenas array. We rely on the fact 1077 * that there is no way for the application to get a dirty read from the 1078 * arenas array unless there is an inherent race in the application 1079 * involving access of an arena being concurrently destroyed. The 1080 * application must synchronize knowledge of the arena's validity, so as 1081 * long as we use an atomic write to update the arenas array, the 1082 * application will get a clean read any time after it synchronizes 1083 * knowledge that the arena is no longer valid. 1084 */ 1085 arena_set(base_ind_get(arena->base), NULL); 1086 1087 /* 1088 * Destroy the base allocator, which manages all metadata ever mapped by 1089 * this arena. 1090 */ 1091 base_delete(tsd_tsdn(tsd), arena->base); 1092} 1093 1094static extent_t * 1095arena_slab_alloc_hard(tsdn_t *tsdn, arena_t *arena, 1096 extent_hooks_t **r_extent_hooks, const bin_info_t *bin_info, 1097 szind_t szind) { 1098 extent_t *slab; 1099 bool zero, commit; 1100 1101 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 1102 WITNESS_RANK_CORE, 0); 1103 1104 zero = false; 1105 commit = true; 1106 slab = extent_alloc_wrapper(tsdn, arena, r_extent_hooks, NULL, 1107 bin_info->slab_size, 0, PAGE, true, szind, &zero, &commit); 1108 1109 if (config_stats && slab != NULL) { 1110 arena_stats_mapped_add(tsdn, &arena->stats, 1111 bin_info->slab_size); 1112 } 1113 1114 return slab; 1115} 1116 1117static extent_t * 1118arena_slab_alloc(tsdn_t *tsdn, arena_t *arena, szind_t binind, 1119 const bin_info_t *bin_info) { 1120 witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn), 1121 WITNESS_RANK_CORE, 0); 1122 1123 extent_hooks_t *extent_hooks = EXTENT_HOOKS_INITIALIZER; 1124 szind_t szind = sz_size2index(bin_info->reg_size); 1125 bool zero = false; 1126 bool commit = true; 1127 extent_t *slab = extents_alloc(tsdn, arena, &extent_hooks, 1128 &arena->extents_dirty, NULL, bin_info->slab_size, 0, PAGE, true, 1129 binind, &zero, &commit); 1130 if (slab == NULL) { 1131 slab = extents_alloc(tsdn, arena, &extent_hooks, 1132 &arena->extents_muzzy, NULL, bin_info->slab_size, 0, PAGE, 1133 true, binind, &zero, &commit); 1134 } 1135 if (slab == NULL) { 1136 slab = arena_slab_alloc_hard(tsdn, arena, &extent_hooks, 1137 bin_info, szind); 1138 if (slab == NULL) { 1139 return NULL; 1140 } 1141 } 1142 assert(extent_slab_get(slab)); 1143 1144 /* Initialize slab internals. */ 1145 arena_slab_data_t *slab_data = extent_slab_data_get(slab); 1146 extent_nfree_set(slab, bin_info->nregs); 1147 bitmap_init(slab_data->bitmap, &bin_info->bitmap_info, false); 1148 1149 arena_nactive_add(arena, extent_size_get(slab) >> LG_PAGE); 1150 1151 return slab; 1152} 1153 1154static extent_t * 1155arena_bin_nonfull_slab_get(tsdn_t *tsdn, arena_t *arena, bin_t *bin, 1156 szind_t binind) { 1157 extent_t *slab; 1158 const bin_info_t *bin_info; 1159 1160 /* Look for a usable slab. */ 1161 slab = arena_bin_slabs_nonfull_tryget(bin); 1162 if (slab != NULL) { 1163 return slab; 1164 } 1165 /* No existing slabs have any space available. */ 1166 1167 bin_info = &bin_infos[binind]; 1168 1169 /* Allocate a new slab. */ 1170 malloc_mutex_unlock(tsdn, &bin->lock); 1171 /******************************/ 1172 slab = arena_slab_alloc(tsdn, arena, binind, bin_info); 1173 /********************************/ 1174 malloc_mutex_lock(tsdn, &bin->lock); 1175 if (slab != NULL) { 1176 if (config_stats) { 1177 bin->stats.nslabs++; 1178 bin->stats.curslabs++; 1179 } 1180 return slab; 1181 } 1182 1183 /* 1184 * arena_slab_alloc() failed, but another thread may have made 1185 * sufficient memory available while this one dropped bin->lock above, 1186 * so search one more time. 1187 */ 1188 slab = arena_bin_slabs_nonfull_tryget(bin); 1189 if (slab != NULL) { 1190 return slab; 1191 } 1192 1193 return NULL; 1194} 1195 1196/* Re-fill bin->slabcur, then call arena_slab_reg_alloc(). */ 1197static void * 1198arena_bin_malloc_hard(tsdn_t *tsdn, arena_t *arena, bin_t *bin, 1199 szind_t binind) { 1200 const bin_info_t *bin_info; 1201 extent_t *slab; 1202 1203 bin_info = &bin_infos[binind]; 1204 if (!arena_is_auto(arena) && bin->slabcur != NULL) { 1205 arena_bin_slabs_full_insert(arena, bin, bin->slabcur); 1206 bin->slabcur = NULL; 1207 } 1208 slab = arena_bin_nonfull_slab_get(tsdn, arena, bin, binind); 1209 if (bin->slabcur != NULL) { 1210 /* 1211 * Another thread updated slabcur while this one ran without the 1212 * bin lock in arena_bin_nonfull_slab_get(). 1213 */ 1214 if (extent_nfree_get(bin->slabcur) > 0) { 1215 void *ret = arena_slab_reg_alloc(bin->slabcur, 1216 bin_info); 1217 if (slab != NULL) { 1218 /* 1219 * arena_slab_alloc() may have allocated slab, 1220 * or it may have been pulled from 1221 * slabs_nonfull. Therefore it is unsafe to 1222 * make any assumptions about how slab has 1223 * previously been used, and 1224 * arena_bin_lower_slab() must be called, as if 1225 * a region were just deallocated from the slab. 1226 */ 1227 if (extent_nfree_get(slab) == bin_info->nregs) { 1228 arena_dalloc_bin_slab(tsdn, arena, slab, 1229 bin); 1230 } else { 1231 arena_bin_lower_slab(tsdn, arena, slab, 1232 bin); 1233 } 1234 } 1235 return ret; 1236 } 1237 1238 arena_bin_slabs_full_insert(arena, bin, bin->slabcur); 1239 bin->slabcur = NULL; 1240 } 1241 1242 if (slab == NULL) { 1243 return NULL; 1244 } 1245 bin->slabcur = slab; 1246 1247 assert(extent_nfree_get(bin->slabcur) > 0); 1248 1249 return arena_slab_reg_alloc(slab, bin_info); 1250} 1251 1252void 1253arena_tcache_fill_small(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, 1254 cache_bin_t *tbin, szind_t binind, uint64_t prof_accumbytes) { 1255 unsigned i, nfill; 1256 bin_t *bin; 1257 1258 assert(tbin->ncached == 0); 1259 1260 if (config_prof && arena_prof_accum(tsdn, arena, prof_accumbytes)) { 1261 prof_idump(tsdn); 1262 } 1263 bin = &arena->bins[binind]; 1264 malloc_mutex_lock(tsdn, &bin->lock); 1265 for (i = 0, nfill = (tcache_bin_info[binind].ncached_max >> 1266 tcache->lg_fill_div[binind]); i < nfill; i++) { 1267 extent_t *slab; 1268 void *ptr; 1269 if ((slab = bin->slabcur) != NULL && extent_nfree_get(slab) > 1270 0) { 1271 ptr = arena_slab_reg_alloc(slab, &bin_infos[binind]); 1272 } else { 1273 ptr = arena_bin_malloc_hard(tsdn, arena, bin, binind); 1274 } 1275 if (ptr == NULL) { 1276 /* 1277 * OOM. tbin->avail isn't yet filled down to its first 1278 * element, so the successful allocations (if any) must 1279 * be moved just before tbin->avail before bailing out. 1280 */ 1281 if (i > 0) { 1282 memmove(tbin->avail - i, tbin->avail - nfill, 1283 i * sizeof(void *)); 1284 } 1285 break; 1286 } 1287 if (config_fill && unlikely(opt_junk_alloc)) { 1288 arena_alloc_junk_small(ptr, &bin_infos[binind], true); 1289 } 1290 /* Insert such that low regions get used first. */ 1291 *(tbin->avail - nfill + i) = ptr; 1292 } 1293 if (config_stats) { 1294 bin->stats.nmalloc += i; 1295 bin->stats.nrequests += tbin->tstats.nrequests; 1296 bin->stats.curregs += i; 1297 bin->stats.nfills++; 1298 tbin->tstats.nrequests = 0; 1299 } 1300 malloc_mutex_unlock(tsdn, &bin->lock); 1301 tbin->ncached = i; 1302 arena_decay_tick(tsdn, arena); 1303} 1304 1305void 1306arena_alloc_junk_small(void *ptr, const bin_info_t *bin_info, bool zero) { 1307 if (!zero) { 1308 memset(ptr, JEMALLOC_ALLOC_JUNK, bin_info->reg_size); 1309 } 1310} 1311 1312static void 1313arena_dalloc_junk_small_impl(void *ptr, const bin_info_t *bin_info) { 1314 memset(ptr, JEMALLOC_FREE_JUNK, bin_info->reg_size); 1315} 1316arena_dalloc_junk_small_t *JET_MUTABLE arena_dalloc_junk_small = 1317 arena_dalloc_junk_small_impl; 1318 1319static void * 1320arena_malloc_small(tsdn_t *tsdn, arena_t *arena, szind_t binind, bool zero) { 1321 void *ret; 1322 bin_t *bin; 1323 size_t usize; 1324 extent_t *slab; 1325 1326 assert(binind < NBINS); 1327 bin = &arena->bins[binind]; 1328 usize = sz_index2size(binind); 1329 1330 malloc_mutex_lock(tsdn, &bin->lock); 1331 if ((slab = bin->slabcur) != NULL && extent_nfree_get(slab) > 0) { 1332 ret = arena_slab_reg_alloc(slab, &bin_infos[binind]); 1333 } else { 1334 ret = arena_bin_malloc_hard(tsdn, arena, bin, binind); 1335 } 1336 1337 if (ret == NULL) { 1338 malloc_mutex_unlock(tsdn, &bin->lock); 1339 return NULL; 1340 } 1341 1342 if (config_stats) { 1343 bin->stats.nmalloc++; 1344 bin->stats.nrequests++; 1345 bin->stats.curregs++; 1346 } 1347 malloc_mutex_unlock(tsdn, &bin->lock); 1348 if (config_prof && arena_prof_accum(tsdn, arena, usize)) { 1349 prof_idump(tsdn); 1350 } 1351 1352 if (!zero) { 1353 if (config_fill) { 1354 if (unlikely(opt_junk_alloc)) { 1355 arena_alloc_junk_small(ret, 1356 &bin_infos[binind], false); 1357 } else if (unlikely(opt_zero)) { 1358 memset(ret, 0, usize); 1359 } 1360 } 1361 } else { 1362 if (config_fill && unlikely(opt_junk_alloc)) { 1363 arena_alloc_junk_small(ret, &bin_infos[binind], 1364 true); 1365 } 1366 memset(ret, 0, usize); 1367 } 1368 1369 arena_decay_tick(tsdn, arena); 1370 return ret; 1371} 1372 1373void * 1374arena_malloc_hard(tsdn_t *tsdn, arena_t *arena, size_t size, szind_t ind, 1375 bool zero) { 1376 assert(!tsdn_null(tsdn) || arena != NULL); 1377 1378 if (likely(!tsdn_null(tsdn))) { 1379 arena = arena_choose(tsdn_tsd(tsdn), arena); 1380 } 1381 if (unlikely(arena == NULL)) { 1382 return NULL; 1383 } 1384 1385 if (likely(size <= SMALL_MAXCLASS)) { 1386 return arena_malloc_small(tsdn, arena, ind, zero); 1387 } 1388 return large_malloc(tsdn, arena, sz_index2size(ind), zero); 1389} 1390 1391void * 1392arena_palloc(tsdn_t *tsdn, arena_t *arena, size_t usize, size_t alignment, 1393 bool zero, tcache_t *tcache) { 1394 void *ret; 1395 1396 if (usize <= SMALL_MAXCLASS && (alignment < PAGE || (alignment == PAGE 1397 && (usize & PAGE_MASK) == 0))) { 1398 /* Small; alignment doesn't require special slab placement. */ 1399 ret = arena_malloc(tsdn, arena, usize, sz_size2index(usize), 1400 zero, tcache, true); 1401 } else { 1402 if (likely(alignment <= CACHELINE)) { 1403 ret = large_malloc(tsdn, arena, usize, zero); 1404 } else { 1405 ret = large_palloc(tsdn, arena, usize, alignment, zero); 1406 } 1407 } 1408 return ret; 1409} 1410 1411void 1412arena_prof_promote(tsdn_t *tsdn, const void *ptr, size_t usize) { 1413 cassert(config_prof); 1414 assert(ptr != NULL); 1415 assert(isalloc(tsdn, ptr) == LARGE_MINCLASS); 1416 assert(usize <= SMALL_MAXCLASS); 1417 1418 rtree_ctx_t rtree_ctx_fallback; 1419 rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); 1420 1421 extent_t *extent = rtree_extent_read(tsdn, &extents_rtree, rtree_ctx, 1422 (uintptr_t)ptr, true); 1423 arena_t *arena = extent_arena_get(extent); 1424 1425 szind_t szind = sz_size2index(usize); 1426 extent_szind_set(extent, szind); 1427 rtree_szind_slab_update(tsdn, &extents_rtree, rtree_ctx, (uintptr_t)ptr, 1428 szind, false); 1429 1430 prof_accum_cancel(tsdn, &arena->prof_accum, usize); 1431 1432 assert(isalloc(tsdn, ptr) == usize); 1433} 1434 1435static size_t 1436arena_prof_demote(tsdn_t *tsdn, extent_t *extent, const void *ptr) { 1437 cassert(config_prof); 1438 assert(ptr != NULL); 1439 1440 extent_szind_set(extent, NBINS); 1441 rtree_ctx_t rtree_ctx_fallback; 1442 rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback); 1443 rtree_szind_slab_update(tsdn, &extents_rtree, rtree_ctx, (uintptr_t)ptr, 1444 NBINS, false); 1445 1446 assert(isalloc(tsdn, ptr) == LARGE_MINCLASS); 1447 1448 return LARGE_MINCLASS; 1449} 1450 1451void 1452arena_dalloc_promoted(tsdn_t *tsdn, void *ptr, tcache_t *tcache, 1453 bool slow_path) { 1454 cassert(config_prof); 1455 assert(opt_prof); 1456 1457 extent_t *extent = iealloc(tsdn, ptr); 1458 size_t usize = arena_prof_demote(tsdn, extent, ptr); 1459 if (usize <= tcache_maxclass) { 1460 tcache_dalloc_large(tsdn_tsd(tsdn), tcache, ptr, 1461 sz_size2index(usize), slow_path); 1462 } else { 1463 large_dalloc(tsdn, extent); 1464 } 1465} 1466 1467static void 1468arena_dissociate_bin_slab(arena_t *arena, extent_t *slab, bin_t *bin) { 1469 /* Dissociate slab from bin. */ 1470 if (slab == bin->slabcur) { 1471 bin->slabcur = NULL; 1472 } else { 1473 szind_t binind = extent_szind_get(slab); 1474 const bin_info_t *bin_info = &bin_infos[binind]; 1475 1476 /* 1477 * The following block's conditional is necessary because if the 1478 * slab only contains one region, then it never gets inserted 1479 * into the non-full slabs heap. 1480 */ 1481 if (bin_info->nregs == 1) { 1482 arena_bin_slabs_full_remove(arena, bin, slab); 1483 } else { 1484 arena_bin_slabs_nonfull_remove(bin, slab); 1485 } 1486 } 1487} 1488 1489static void 1490arena_dalloc_bin_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab, 1491 bin_t *bin) { 1492 assert(slab != bin->slabcur); 1493 1494 malloc_mutex_unlock(tsdn, &bin->lock); 1495 /******************************/ 1496 arena_slab_dalloc(tsdn, arena, slab); 1497 /****************************/ 1498 malloc_mutex_lock(tsdn, &bin->lock); 1499 if (config_stats) { 1500 bin->stats.curslabs--; 1501 } 1502} 1503 1504static void 1505arena_bin_lower_slab(UNUSED tsdn_t *tsdn, arena_t *arena, extent_t *slab, 1506 bin_t *bin) { 1507 assert(extent_nfree_get(slab) > 0); 1508 1509 /* 1510 * Make sure that if bin->slabcur is non-NULL, it refers to the 1511 * oldest/lowest non-full slab. It is okay to NULL slabcur out rather 1512 * than proactively keeping it pointing at the oldest/lowest non-full 1513 * slab. 1514 */ 1515 if (bin->slabcur != NULL && extent_snad_comp(bin->slabcur, slab) > 0) { 1516 /* Switch slabcur. */ 1517 if (extent_nfree_get(bin->slabcur) > 0) { 1518 arena_bin_slabs_nonfull_insert(bin, bin->slabcur); 1519 } else { 1520 arena_bin_slabs_full_insert(arena, bin, bin->slabcur); 1521 } 1522 bin->slabcur = slab; 1523 if (config_stats) { 1524 bin->stats.reslabs++; 1525 } 1526 } else { 1527 arena_bin_slabs_nonfull_insert(bin, slab); 1528 } 1529} 1530 1531static void 1532arena_dalloc_bin_locked_impl(tsdn_t *tsdn, arena_t *arena, extent_t *slab, 1533 void *ptr, bool junked) { 1534 arena_slab_data_t *slab_data = extent_slab_data_get(slab); 1535 szind_t binind = extent_szind_get(slab); 1536 bin_t *bin = &arena->bins[binind]; 1537 const bin_info_t *bin_info = &bin_infos[binind]; 1538 1539 if (!junked && config_fill && unlikely(opt_junk_free)) { 1540 arena_dalloc_junk_small(ptr, bin_info); 1541 } 1542 1543 arena_slab_reg_dalloc(slab, slab_data, ptr); 1544 unsigned nfree = extent_nfree_get(slab); 1545 if (nfree == bin_info->nregs) { 1546 arena_dissociate_bin_slab(arena, slab, bin); 1547 arena_dalloc_bin_slab(tsdn, arena, slab, bin); 1548 } else if (nfree == 1 && slab != bin->slabcur) { 1549 arena_bin_slabs_full_remove(arena, bin, slab); 1550 arena_bin_lower_slab(tsdn, arena, slab, bin); 1551 } 1552 1553 if (config_stats) { 1554 bin->stats.ndalloc++; 1555 bin->stats.curregs--; 1556 } 1557} 1558 1559void 1560arena_dalloc_bin_junked_locked(tsdn_t *tsdn, arena_t *arena, extent_t *extent, 1561 void *ptr) { 1562 arena_dalloc_bin_locked_impl(tsdn, arena, extent, ptr, true); 1563} 1564 1565static void 1566arena_dalloc_bin(tsdn_t *tsdn, arena_t *arena, extent_t *extent, void *ptr) { 1567 szind_t binind = extent_szind_get(extent); 1568 bin_t *bin = &arena->bins[binind]; 1569 1570 malloc_mutex_lock(tsdn, &bin->lock); 1571 arena_dalloc_bin_locked_impl(tsdn, arena, extent, ptr, false); 1572 malloc_mutex_unlock(tsdn, &bin->lock); 1573} 1574 1575void 1576arena_dalloc_small(tsdn_t *tsdn, void *ptr) { 1577 extent_t *extent = iealloc(tsdn, ptr); 1578 arena_t *arena = extent_arena_get(extent); 1579 1580 arena_dalloc_bin(tsdn, arena, extent, ptr); 1581 arena_decay_tick(tsdn, arena); 1582} 1583 1584bool 1585arena_ralloc_no_move(tsdn_t *tsdn, void *ptr, size_t oldsize, size_t size, 1586 size_t extra, bool zero) { 1587 /* Calls with non-zero extra had to clamp extra. */ 1588 assert(extra == 0 || size + extra <= LARGE_MAXCLASS); 1589 1590 if (unlikely(size > LARGE_MAXCLASS)) { 1591 return true; 1592 } 1593 1594 extent_t *extent = iealloc(tsdn, ptr); 1595 size_t usize_min = sz_s2u(size); 1596 size_t usize_max = sz_s2u(size + extra); 1597 if (likely(oldsize <= SMALL_MAXCLASS && usize_min <= SMALL_MAXCLASS)) { 1598 /* 1599 * Avoid moving the allocation if the size class can be left the 1600 * same. 1601 */ 1602 assert(bin_infos[sz_size2index(oldsize)].reg_size == 1603 oldsize); 1604 if ((usize_max > SMALL_MAXCLASS || sz_size2index(usize_max) != 1605 sz_size2index(oldsize)) && (size > oldsize || usize_max < 1606 oldsize)) { 1607 return true; 1608 } 1609 1610 arena_decay_tick(tsdn, extent_arena_get(extent)); 1611 return false; 1612 } else if (oldsize >= LARGE_MINCLASS && usize_max >= LARGE_MINCLASS) { 1613 return large_ralloc_no_move(tsdn, extent, usize_min, usize_max, 1614 zero); 1615 } 1616 1617 return true; 1618} 1619 1620static void * 1621arena_ralloc_move_helper(tsdn_t *tsdn, arena_t *arena, size_t usize, 1622 size_t alignment, bool zero, tcache_t *tcache) { 1623 if (alignment == 0) { 1624 return arena_malloc(tsdn, arena, usize, sz_size2index(usize), 1625 zero, tcache, true); 1626 } 1627 usize = sz_sa2u(usize, alignment); 1628 if (unlikely(usize == 0 || usize > LARGE_MAXCLASS)) { 1629 return NULL; 1630 } 1631 return ipalloct(tsdn, usize, alignment, zero, tcache, arena); 1632} 1633 1634void * 1635arena_ralloc(tsdn_t *tsdn, arena_t *arena, void *ptr, size_t oldsize, 1636 size_t size, size_t alignment, bool zero, tcache_t *tcache) { 1637 size_t usize = sz_s2u(size); 1638 if (unlikely(usize == 0 || size > LARGE_MAXCLASS)) { 1639 return NULL; 1640 } 1641 1642 if (likely(usize <= SMALL_MAXCLASS)) { 1643 /* Try to avoid moving the allocation. */ 1644 if (!arena_ralloc_no_move(tsdn, ptr, oldsize, usize, 0, zero)) { 1645 return ptr; 1646 } 1647 } 1648 1649 if (oldsize >= LARGE_MINCLASS && usize >= LARGE_MINCLASS) { 1650 return large_ralloc(tsdn, arena, iealloc(tsdn, ptr), usize, 1651 alignment, zero, tcache); 1652 } 1653 1654 /* 1655 * size and oldsize are different enough that we need to move the 1656 * object. In that case, fall back to allocating new space and copying. 1657 */ 1658 void *ret = arena_ralloc_move_helper(tsdn, arena, usize, alignment, 1659 zero, tcache); 1660 if (ret == NULL) { 1661 return NULL; 1662 } 1663 1664 /* 1665 * Junk/zero-filling were already done by 1666 * ipalloc()/arena_malloc(). 1667 */ 1668 1669 size_t copysize = (usize < oldsize) ? usize : oldsize; 1670 memcpy(ret, ptr, copysize); 1671 isdalloct(tsdn, ptr, oldsize, tcache, NULL, true); 1672 return ret; 1673} 1674 1675dss_prec_t 1676arena_dss_prec_get(arena_t *arena) { 1677 return (dss_prec_t)atomic_load_u(&arena->dss_prec, ATOMIC_ACQUIRE); 1678} 1679 1680bool 1681arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec) { 1682 if (!have_dss) { 1683 return (dss_prec != dss_prec_disabled); 1684 } 1685 atomic_store_u(&arena->dss_prec, (unsigned)dss_prec, ATOMIC_RELEASE); 1686 return false; 1687} 1688 1689ssize_t 1690arena_dirty_decay_ms_default_get(void) { 1691 return atomic_load_zd(&dirty_decay_ms_default, ATOMIC_RELAXED); 1692} 1693 1694bool 1695arena_dirty_decay_ms_default_set(ssize_t decay_ms) { 1696 if (!arena_decay_ms_valid(decay_ms)) { 1697 return true; 1698 } 1699 atomic_store_zd(&dirty_decay_ms_default, decay_ms, ATOMIC_RELAXED); 1700 return false; 1701} 1702 1703ssize_t 1704arena_muzzy_decay_ms_default_get(void) { 1705 return atomic_load_zd(&muzzy_decay_ms_default, ATOMIC_RELAXED); 1706} 1707 1708bool 1709arena_muzzy_decay_ms_default_set(ssize_t decay_ms) { 1710 if (!arena_decay_ms_valid(decay_ms)) { 1711 return true; 1712 } 1713 atomic_store_zd(&muzzy_decay_ms_default, decay_ms, ATOMIC_RELAXED); 1714 return false; 1715} 1716 1717bool 1718arena_retain_grow_limit_get_set(tsd_t *tsd, arena_t *arena, size_t *old_limit, 1719 size_t *new_limit) { 1720 assert(opt_retain); 1721 1722 pszind_t new_ind JEMALLOC_CC_SILENCE_INIT(0); 1723 if (new_limit != NULL) { 1724 size_t limit = *new_limit; 1725 /* Grow no more than the new limit. */ 1726 if ((new_ind = sz_psz2ind(limit + 1) - 1) > 1727 EXTENT_GROW_MAX_PIND) { 1728 return true; 1729 } 1730 } 1731 1732 malloc_mutex_lock(tsd_tsdn(tsd), &arena->extent_grow_mtx); 1733 if (old_limit != NULL) { 1734 *old_limit = sz_pind2sz(arena->retain_grow_limit); 1735 } 1736 if (new_limit != NULL) { 1737 arena->retain_grow_limit = new_ind; 1738 } 1739 malloc_mutex_unlock(tsd_tsdn(tsd), &arena->extent_grow_mtx); 1740 1741 return false; 1742} 1743 1744unsigned 1745arena_nthreads_get(arena_t *arena, bool internal) { 1746 return atomic_load_u(&arena->nthreads[internal], ATOMIC_RELAXED); 1747} 1748 1749void 1750arena_nthreads_inc(arena_t *arena, bool internal) { 1751 atomic_fetch_add_u(&arena->nthreads[internal], 1, ATOMIC_RELAXED); 1752} 1753 1754void 1755arena_nthreads_dec(arena_t *arena, bool internal) { 1756 atomic_fetch_sub_u(&arena->nthreads[internal], 1, ATOMIC_RELAXED); 1757} 1758 1759size_t 1760arena_extent_sn_next(arena_t *arena) { 1761 return atomic_fetch_add_zu(&arena->extent_sn_next, 1, ATOMIC_RELAXED); 1762} 1763 1764arena_t * 1765arena_new(tsdn_t *tsdn, unsigned ind, extent_hooks_t *extent_hooks) { 1766 arena_t *arena; 1767 base_t *base; 1768 unsigned i; 1769 1770 if (ind == 0) { 1771 base = b0get(); 1772 } else { 1773 base = base_new(tsdn, ind, extent_hooks); 1774 if (base == NULL) { 1775 return NULL; 1776 } 1777 } 1778 1779 arena = (arena_t *)base_alloc(tsdn, base, sizeof(arena_t), CACHELINE); 1780 if (arena == NULL) { 1781 goto label_error; 1782 } 1783 1784 atomic_store_u(&arena->nthreads[0], 0, ATOMIC_RELAXED); 1785 atomic_store_u(&arena->nthreads[1], 0, ATOMIC_RELAXED); 1786 arena->last_thd = NULL; 1787 1788 if (config_stats) { 1789 if (arena_stats_init(tsdn, &arena->stats)) { 1790 goto label_error; 1791 } 1792 1793 ql_new(&arena->tcache_ql); 1794 ql_new(&arena->cache_bin_array_descriptor_ql); 1795 if (malloc_mutex_init(&arena->tcache_ql_mtx, "tcache_ql", 1796 WITNESS_RANK_TCACHE_QL, malloc_mutex_rank_exclusive)) { 1797 goto label_error; 1798 } 1799 } 1800 1801 if (config_prof) { 1802 if (prof_accum_init(tsdn, &arena->prof_accum)) { 1803 goto label_error; 1804 } 1805 } 1806 1807 if (config_cache_oblivious) { 1808 /* 1809 * A nondeterministic seed based on the address of arena reduces 1810 * the likelihood of lockstep non-uniform cache index 1811 * utilization among identical concurrent processes, but at the 1812 * cost of test repeatability. For debug builds, instead use a 1813 * deterministic seed. 1814 */ 1815 atomic_store_zu(&arena->offset_state, config_debug ? ind : 1816 (size_t)(uintptr_t)arena, ATOMIC_RELAXED); 1817 } 1818 1819 atomic_store_zu(&arena->extent_sn_next, 0, ATOMIC_RELAXED); 1820 1821 atomic_store_u(&arena->dss_prec, (unsigned)extent_dss_prec_get(), 1822 ATOMIC_RELAXED); 1823 1824 atomic_store_zu(&arena->nactive, 0, ATOMIC_RELAXED); 1825 1826 extent_list_init(&arena->large); 1827 if (malloc_mutex_init(&arena->large_mtx, "arena_large", 1828 WITNESS_RANK_ARENA_LARGE, malloc_mutex_rank_exclusive)) { 1829 goto label_error; 1830 } 1831 1832 /* 1833 * Delay coalescing for dirty extents despite the disruptive effect on 1834 * memory layout for best-fit extent allocation, since cached extents 1835 * are likely to be reused soon after deallocation, and the cost of 1836 * merging/splitting extents is non-trivial. 1837 */ 1838 if (extents_init(tsdn, &arena->extents_dirty, extent_state_dirty, 1839 true)) { 1840 goto label_error; 1841 } 1842 /* 1843 * Coalesce muzzy extents immediately, because operations on them are in 1844 * the critical path much less often than for dirty extents. 1845 */ 1846 if (extents_init(tsdn, &arena->extents_muzzy, extent_state_muzzy, 1847 false)) { 1848 goto label_error; 1849 } 1850 /* 1851 * Coalesce retained extents immediately, in part because they will 1852 * never be evicted (and therefore there's no opportunity for delayed 1853 * coalescing), but also because operations on retained extents are not 1854 * in the critical path. 1855 */ 1856 if (extents_init(tsdn, &arena->extents_retained, extent_state_retained, 1857 false)) { 1858 goto label_error; 1859 } 1860 1861 if (arena_decay_init(&arena->decay_dirty, 1862 arena_dirty_decay_ms_default_get(), &arena->stats.decay_dirty)) { 1863 goto label_error; 1864 } 1865 if (arena_decay_init(&arena->decay_muzzy, 1866 arena_muzzy_decay_ms_default_get(), &arena->stats.decay_muzzy)) { 1867 goto label_error; 1868 } 1869 1870 arena->extent_grow_next = sz_psz2ind(HUGEPAGE); 1871 arena->retain_grow_limit = EXTENT_GROW_MAX_PIND; 1872 if (malloc_mutex_init(&arena->extent_grow_mtx, "extent_grow", 1873 WITNESS_RANK_EXTENT_GROW, malloc_mutex_rank_exclusive)) { 1874 goto label_error; 1875 } 1876 1877 extent_avail_new(&arena->extent_avail); 1878 if (malloc_mutex_init(&arena->extent_avail_mtx, "extent_avail", 1879 WITNESS_RANK_EXTENT_AVAIL, malloc_mutex_rank_exclusive)) { 1880 goto label_error; 1881 } 1882 1883 /* Initialize bins. */ 1884 for (i = 0; i < NBINS; i++) { 1885 bool err = bin_init(&arena->bins[i]); 1886 if (err) { 1887 goto label_error; 1888 } 1889 } 1890 1891 arena->base = base; 1892 /* Set arena before creating background threads. */ 1893 arena_set(ind, arena); 1894 1895 nstime_init(&arena->create_time, 0); 1896 nstime_update(&arena->create_time); 1897 1898 /* We don't support reentrancy for arena 0 bootstrapping. */ 1899 if (ind != 0) { 1900 /* 1901 * If we're here, then arena 0 already exists, so bootstrapping 1902 * is done enough that we should have tsd. 1903 */ 1904 assert(!tsdn_null(tsdn)); 1905 pre_reentrancy(tsdn_tsd(tsdn), arena); 1906 if (hooks_arena_new_hook) { 1907 hooks_arena_new_hook(); 1908 } 1909 post_reentrancy(tsdn_tsd(tsdn)); 1910 } 1911 1912 return arena; 1913label_error: 1914 if (ind != 0) { 1915 base_delete(tsdn, base); 1916 } 1917 return NULL; 1918} 1919 1920void 1921arena_boot(void) { 1922 arena_dirty_decay_ms_default_set(opt_dirty_decay_ms); 1923 arena_muzzy_decay_ms_default_set(opt_muzzy_decay_ms); 1924#define REGIND_bin_yes(index, reg_size) \ 1925 div_init(&arena_binind_div_info[(index)], (reg_size)); 1926#define REGIND_bin_no(index, reg_size) 1927#define SC(index, lg_grp, lg_delta, ndelta, psz, bin, pgs, \ 1928 lg_delta_lookup) \ 1929 REGIND_bin_##bin(index, (1U<<lg_grp) + (ndelta << lg_delta)) 1930 SIZE_CLASSES 1931#undef REGIND_bin_yes 1932#undef REGIND_bin_no 1933#undef SC 1934} 1935 1936void 1937arena_prefork0(tsdn_t *tsdn, arena_t *arena) { 1938 malloc_mutex_prefork(tsdn, &arena->decay_dirty.mtx); 1939 malloc_mutex_prefork(tsdn, &arena->decay_muzzy.mtx); 1940} 1941 1942void 1943arena_prefork1(tsdn_t *tsdn, arena_t *arena) { 1944 if (config_stats) { 1945 malloc_mutex_prefork(tsdn, &arena->tcache_ql_mtx); 1946 } 1947} 1948 1949void 1950arena_prefork2(tsdn_t *tsdn, arena_t *arena) { 1951 malloc_mutex_prefork(tsdn, &arena->extent_grow_mtx); 1952} 1953 1954void 1955arena_prefork3(tsdn_t *tsdn, arena_t *arena) { 1956 extents_prefork(tsdn, &arena->extents_dirty); 1957 extents_prefork(tsdn, &arena->extents_muzzy); 1958 extents_prefork(tsdn, &arena->extents_retained); 1959} 1960 1961void 1962arena_prefork4(tsdn_t *tsdn, arena_t *arena) { 1963 malloc_mutex_prefork(tsdn, &arena->extent_avail_mtx); 1964} 1965 1966void 1967arena_prefork5(tsdn_t *tsdn, arena_t *arena) { 1968 base_prefork(tsdn, arena->base); 1969} 1970 1971void 1972arena_prefork6(tsdn_t *tsdn, arena_t *arena) { 1973 malloc_mutex_prefork(tsdn, &arena->large_mtx); 1974} 1975 1976void 1977arena_prefork7(tsdn_t *tsdn, arena_t *arena) { 1978 for (unsigned i = 0; i < NBINS; i++) { 1979 bin_prefork(tsdn, &arena->bins[i]); 1980 } 1981} 1982 1983void 1984arena_postfork_parent(tsdn_t *tsdn, arena_t *arena) { 1985 unsigned i; 1986 1987 for (i = 0; i < NBINS; i++) { 1988 bin_postfork_parent(tsdn, &arena->bins[i]); 1989 } 1990 malloc_mutex_postfork_parent(tsdn, &arena->large_mtx); 1991 base_postfork_parent(tsdn, arena->base); 1992 malloc_mutex_postfork_parent(tsdn, &arena->extent_avail_mtx); 1993 extents_postfork_parent(tsdn, &arena->extents_dirty); 1994 extents_postfork_parent(tsdn, &arena->extents_muzzy); 1995 extents_postfork_parent(tsdn, &arena->extents_retained); 1996 malloc_mutex_postfork_parent(tsdn, &arena->extent_grow_mtx); 1997 malloc_mutex_postfork_parent(tsdn, &arena->decay_dirty.mtx); 1998 malloc_mutex_postfork_parent(tsdn, &arena->decay_muzzy.mtx); 1999 if (config_stats) { 2000 malloc_mutex_postfork_parent(tsdn, &arena->tcache_ql_mtx); 2001 } 2002} 2003 2004void 2005arena_postfork_child(tsdn_t *tsdn, arena_t *arena) { 2006 unsigned i; 2007 2008 atomic_store_u(&arena->nthreads[0], 0, ATOMIC_RELAXED); 2009 atomic_store_u(&arena->nthreads[1], 0, ATOMIC_RELAXED); 2010 if (tsd_arena_get(tsdn_tsd(tsdn)) == arena) { 2011 arena_nthreads_inc(arena, false); 2012 } 2013 if (tsd_iarena_get(tsdn_tsd(tsdn)) == arena) { 2014 arena_nthreads_inc(arena, true); 2015 } 2016 if (config_stats) { 2017 ql_new(&arena->tcache_ql); 2018 ql_new(&arena->cache_bin_array_descriptor_ql); 2019 tcache_t *tcache = tcache_get(tsdn_tsd(tsdn)); 2020 if (tcache != NULL && tcache->arena == arena) { 2021 ql_elm_new(tcache, link); 2022 ql_tail_insert(&arena->tcache_ql, tcache, link); 2023 cache_bin_array_descriptor_init( 2024 &tcache->cache_bin_array_descriptor, 2025 tcache->bins_small, tcache->bins_large); 2026 ql_tail_insert(&arena->cache_bin_array_descriptor_ql, 2027 &tcache->cache_bin_array_descriptor, link); 2028 } 2029 } 2030 2031 for (i = 0; i < NBINS; i++) { 2032 bin_postfork_child(tsdn, &arena->bins[i]); 2033 } 2034 malloc_mutex_postfork_child(tsdn, &arena->large_mtx); 2035 base_postfork_child(tsdn, arena->base); 2036 malloc_mutex_postfork_child(tsdn, &arena->extent_avail_mtx); 2037 extents_postfork_child(tsdn, &arena->extents_dirty); 2038 extents_postfork_child(tsdn, &arena->extents_muzzy); 2039 extents_postfork_child(tsdn, &arena->extents_retained); 2040 malloc_mutex_postfork_child(tsdn, &arena->extent_grow_mtx); 2041 malloc_mutex_postfork_child(tsdn, &arena->decay_dirty.mtx); 2042 malloc_mutex_postfork_child(tsdn, &arena->decay_muzzy.mtx); 2043 if (config_stats) { 2044 malloc_mutex_postfork_child(tsdn, &arena->tcache_ql_mtx); 2045 } 2046} 2047