1 #define JEMALLOC_TCACHE_C_ 2 #include "jemalloc/internal/jemalloc_internal.h" 3 4 /******************************************************************************/ 5 /* Data. */ 6 7 bool opt_tcache = true; 8 ssize_t opt_lg_tcache_max = LG_TCACHE_MAXCLASS_DEFAULT; 9 10 tcache_bin_info_t *tcache_bin_info; 11 static unsigned stack_nelms; /* Total stack elms per tcache. */ 12 13 unsigned nhbins; 14 size_t tcache_maxclass; 15 16 tcaches_t *tcaches; 17 18 /* Index of first element within tcaches that has never been used. */ 19 static unsigned tcaches_past; 20 21 /* Head of singly linked list tracking available tcaches elements. */ 22 static tcaches_t *tcaches_avail; 23 24 /* Protects tcaches{,_past,_avail}. */ 25 static malloc_mutex_t tcaches_mtx; 26 27 /******************************************************************************/ 28 29 size_t 30 tcache_salloc(tsdn_t *tsdn, const void *ptr) 31 { 32 33 return (arena_salloc(tsdn, ptr, false)); 34 } 35 36 void 37 tcache_event_hard(tsd_t *tsd, tcache_t *tcache) 38 { 39 szind_t binind = tcache->next_gc_bin; 40 tcache_bin_t *tbin = &tcache->tbins[binind]; 41 tcache_bin_info_t *tbin_info = &tcache_bin_info[binind]; 42 43 if (tbin->low_water > 0) { 44 /* 45 * Flush (ceiling) 3/4 of the objects below the low water mark. 46 */ 47 if (binind < NBINS) { 48 tcache_bin_flush_small(tsd, tcache, tbin, binind, 49 tbin->ncached - tbin->low_water + (tbin->low_water 50 >> 2)); 51 } else { 52 tcache_bin_flush_large(tsd, tbin, binind, tbin->ncached 53 - tbin->low_water + (tbin->low_water >> 2), tcache); 54 } 55 /* 56 * Reduce fill count by 2X. Limit lg_fill_div such that the 57 * fill count is always at least 1. 58 */ 59 if ((tbin_info->ncached_max >> (tbin->lg_fill_div+1)) >= 1) 60 tbin->lg_fill_div++; 61 } else if (tbin->low_water < 0) { 62 /* 63 * Increase fill count by 2X. Make sure lg_fill_div stays 64 * greater than 0. 65 */ 66 if (tbin->lg_fill_div > 1) 67 tbin->lg_fill_div--; 68 } 69 tbin->low_water = tbin->ncached; 70 71 tcache->next_gc_bin++; 72 if (tcache->next_gc_bin == nhbins) 73 tcache->next_gc_bin = 0; 74 } 75 76 void * 77 tcache_alloc_small_hard(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache, 78 tcache_bin_t *tbin, szind_t binind, bool *tcache_success) 79 { 80 void *ret; 81 82 arena_tcache_fill_small(tsdn, arena, tbin, binind, config_prof ? 83 tcache->prof_accumbytes : 0); 84 if (config_prof) 85 tcache->prof_accumbytes = 0; 86 ret = tcache_alloc_easy(tbin, tcache_success); 87 88 return (ret); 89 } 90 91 void 92 tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin, 93 szind_t binind, unsigned rem) 94 { 95 arena_t *arena; 96 void *ptr; 97 unsigned i, nflush, ndeferred; 98 bool merged_stats = false; 99 100 assert(binind < NBINS); 101 assert(rem <= tbin->ncached); 102 103 arena = arena_choose(tsd, NULL); 104 assert(arena != NULL); 105 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 106 /* Lock the arena bin associated with the first object. */ 107 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 108 *(tbin->avail - 1)); 109 arena_t *bin_arena = extent_node_arena_get(&chunk->node); 110 arena_bin_t *bin = &bin_arena->bins[binind]; 111 112 if (config_prof && bin_arena == arena) { 113 if (arena_prof_accum(tsd_tsdn(tsd), arena, 114 tcache->prof_accumbytes)) 115 prof_idump(tsd_tsdn(tsd)); 116 tcache->prof_accumbytes = 0; 117 } 118 119 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 120 if (config_stats && bin_arena == arena) { 121 assert(!merged_stats); 122 merged_stats = true; 123 bin->stats.nflushes++; 124 bin->stats.nrequests += tbin->tstats.nrequests; 125 tbin->tstats.nrequests = 0; 126 } 127 ndeferred = 0; 128 for (i = 0; i < nflush; i++) { 129 ptr = *(tbin->avail - 1 - i); 130 assert(ptr != NULL); 131 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 132 if (extent_node_arena_get(&chunk->node) == bin_arena) { 133 size_t pageind = ((uintptr_t)ptr - 134 (uintptr_t)chunk) >> LG_PAGE; 135 arena_chunk_map_bits_t *bitselm = 136 arena_bitselm_get_mutable(chunk, pageind); 137 arena_dalloc_bin_junked_locked(tsd_tsdn(tsd), 138 bin_arena, chunk, ptr, bitselm); 139 } else { 140 /* 141 * This object was allocated via a different 142 * arena bin than the one that is currently 143 * locked. Stash the object, so that it can be 144 * handled in a future pass. 145 */ 146 *(tbin->avail - 1 - ndeferred) = ptr; 147 ndeferred++; 148 } 149 } 150 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 151 arena_decay_ticks(tsd_tsdn(tsd), bin_arena, nflush - ndeferred); 152 } 153 if (config_stats && !merged_stats) { 154 /* 155 * The flush loop didn't happen to flush to this thread's 156 * arena, so the stats didn't get merged. Manually do so now. 157 */ 158 arena_bin_t *bin = &arena->bins[binind]; 159 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 160 bin->stats.nflushes++; 161 bin->stats.nrequests += tbin->tstats.nrequests; 162 tbin->tstats.nrequests = 0; 163 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 164 } 165 166 memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem * 167 sizeof(void *)); 168 tbin->ncached = rem; 169 if ((int)tbin->ncached < tbin->low_water) 170 tbin->low_water = tbin->ncached; 171 } 172 173 void 174 tcache_bin_flush_large(tsd_t *tsd, tcache_bin_t *tbin, szind_t binind, 175 unsigned rem, tcache_t *tcache) 176 { 177 arena_t *arena; 178 void *ptr; 179 unsigned i, nflush, ndeferred; 180 bool merged_stats = false; 181 182 assert(binind < nhbins); 183 assert(rem <= tbin->ncached); 184 185 arena = arena_choose(tsd, NULL); 186 assert(arena != NULL); 187 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 188 /* Lock the arena associated with the first object. */ 189 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 190 *(tbin->avail - 1)); 191 arena_t *locked_arena = extent_node_arena_get(&chunk->node); 192 UNUSED bool idump; 193 194 if (config_prof) 195 idump = false; 196 malloc_mutex_lock(tsd_tsdn(tsd), &locked_arena->lock); 197 if ((config_prof || config_stats) && locked_arena == arena) { 198 if (config_prof) { 199 idump = arena_prof_accum_locked(arena, 200 tcache->prof_accumbytes); 201 tcache->prof_accumbytes = 0; 202 } 203 if (config_stats) { 204 merged_stats = true; 205 arena->stats.nrequests_large += 206 tbin->tstats.nrequests; 207 arena->stats.lstats[binind - NBINS].nrequests += 208 tbin->tstats.nrequests; 209 tbin->tstats.nrequests = 0; 210 } 211 } 212 ndeferred = 0; 213 for (i = 0; i < nflush; i++) { 214 ptr = *(tbin->avail - 1 - i); 215 assert(ptr != NULL); 216 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 217 if (extent_node_arena_get(&chunk->node) == 218 locked_arena) { 219 arena_dalloc_large_junked_locked(tsd_tsdn(tsd), 220 locked_arena, chunk, ptr); 221 } else { 222 /* 223 * This object was allocated via a different 224 * arena than the one that is currently locked. 225 * Stash the object, so that it can be handled 226 * in a future pass. 227 */ 228 *(tbin->avail - 1 - ndeferred) = ptr; 229 ndeferred++; 230 } 231 } 232 malloc_mutex_unlock(tsd_tsdn(tsd), &locked_arena->lock); 233 if (config_prof && idump) 234 prof_idump(tsd_tsdn(tsd)); 235 arena_decay_ticks(tsd_tsdn(tsd), locked_arena, nflush - 236 ndeferred); 237 } 238 if (config_stats && !merged_stats) { 239 /* 240 * The flush loop didn't happen to flush to this thread's 241 * arena, so the stats didn't get merged. Manually do so now. 242 */ 243 malloc_mutex_lock(tsd_tsdn(tsd), &arena->lock); 244 arena->stats.nrequests_large += tbin->tstats.nrequests; 245 arena->stats.lstats[binind - NBINS].nrequests += 246 tbin->tstats.nrequests; 247 tbin->tstats.nrequests = 0; 248 malloc_mutex_unlock(tsd_tsdn(tsd), &arena->lock); 249 } 250 251 memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem * 252 sizeof(void *)); 253 tbin->ncached = rem; 254 if ((int)tbin->ncached < tbin->low_water) 255 tbin->low_water = tbin->ncached; 256 } 257 258 static void 259 tcache_arena_associate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) 260 { 261 262 if (config_stats) { 263 /* Link into list of extant tcaches. */ 264 malloc_mutex_lock(tsdn, &arena->lock); 265 ql_elm_new(tcache, link); 266 ql_tail_insert(&arena->tcache_ql, tcache, link); 267 malloc_mutex_unlock(tsdn, &arena->lock); 268 } 269 } 270 271 static void 272 tcache_arena_dissociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) 273 { 274 275 if (config_stats) { 276 /* Unlink from list of extant tcaches. */ 277 malloc_mutex_lock(tsdn, &arena->lock); 278 if (config_debug) { 279 bool in_ql = false; 280 tcache_t *iter; 281 ql_foreach(iter, &arena->tcache_ql, link) { 282 if (iter == tcache) { 283 in_ql = true; 284 break; 285 } 286 } 287 assert(in_ql); 288 } 289 ql_remove(&arena->tcache_ql, tcache, link); 290 tcache_stats_merge(tsdn, tcache, arena); 291 malloc_mutex_unlock(tsdn, &arena->lock); 292 } 293 } 294 295 void 296 tcache_arena_reassociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *oldarena, 297 arena_t *newarena) 298 { 299 300 tcache_arena_dissociate(tsdn, tcache, oldarena); 301 tcache_arena_associate(tsdn, tcache, newarena); 302 } 303 304 tcache_t * 305 tcache_get_hard(tsd_t *tsd) 306 { 307 arena_t *arena; 308 309 if (!tcache_enabled_get()) { 310 if (tsd_nominal(tsd)) 311 tcache_enabled_set(false); /* Memoize. */ 312 return (NULL); 313 } 314 arena = arena_choose(tsd, NULL); 315 if (unlikely(arena == NULL)) 316 return (NULL); 317 return (tcache_create(tsd_tsdn(tsd), arena)); 318 } 319 320 tcache_t * 321 tcache_create(tsdn_t *tsdn, arena_t *arena) 322 { 323 tcache_t *tcache; 324 size_t size, stack_offset; 325 unsigned i; 326 327 size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins); 328 /* Naturally align the pointer stacks. */ 329 size = PTR_CEILING(size); 330 stack_offset = size; 331 size += stack_nelms * sizeof(void *); 332 /* Avoid false cacheline sharing. */ 333 size = sa2u(size, CACHELINE); 334 335 tcache = ipallocztm(tsdn, size, CACHELINE, true, NULL, true, 336 arena_get(TSDN_NULL, 0, true)); 337 if (tcache == NULL) 338 return (NULL); 339 340 tcache_arena_associate(tsdn, tcache, arena); 341 342 ticker_init(&tcache->gc_ticker, TCACHE_GC_INCR); 343 344 assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0); 345 for (i = 0; i < nhbins; i++) { 346 tcache->tbins[i].lg_fill_div = 1; 347 stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *); 348 /* 349 * avail points past the available space. Allocations will 350 * access the slots toward higher addresses (for the benefit of 351 * prefetch). 352 */ 353 tcache->tbins[i].avail = (void **)((uintptr_t)tcache + 354 (uintptr_t)stack_offset); 355 } 356 357 return (tcache); 358 } 359 360 static void 361 tcache_destroy(tsd_t *tsd, tcache_t *tcache) 362 { 363 arena_t *arena; 364 unsigned i; 365 366 arena = arena_choose(tsd, NULL); 367 tcache_arena_dissociate(tsd_tsdn(tsd), tcache, arena); 368 369 for (i = 0; i < NBINS; i++) { 370 tcache_bin_t *tbin = &tcache->tbins[i]; 371 tcache_bin_flush_small(tsd, tcache, tbin, i, 0); 372 373 if (config_stats && tbin->tstats.nrequests != 0) { 374 arena_bin_t *bin = &arena->bins[i]; 375 malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock); 376 bin->stats.nrequests += tbin->tstats.nrequests; 377 malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock); 378 } 379 } 380 381 for (; i < nhbins; i++) { 382 tcache_bin_t *tbin = &tcache->tbins[i]; 383 tcache_bin_flush_large(tsd, tbin, i, 0, tcache); 384 385 if (config_stats && tbin->tstats.nrequests != 0) { 386 malloc_mutex_lock(tsd_tsdn(tsd), &arena->lock); 387 arena->stats.nrequests_large += tbin->tstats.nrequests; 388 arena->stats.lstats[i - NBINS].nrequests += 389 tbin->tstats.nrequests; 390 malloc_mutex_unlock(tsd_tsdn(tsd), &arena->lock); 391 } 392 } 393 394 if (config_prof && tcache->prof_accumbytes > 0 && 395 arena_prof_accum(tsd_tsdn(tsd), arena, tcache->prof_accumbytes)) 396 prof_idump(tsd_tsdn(tsd)); 397 398 idalloctm(tsd_tsdn(tsd), tcache, NULL, true, true); 399 } 400 401 void 402 tcache_cleanup(tsd_t *tsd) 403 { 404 tcache_t *tcache; 405 406 if (!config_tcache) 407 return; 408 409 if ((tcache = tsd_tcache_get(tsd)) != NULL) { 410 tcache_destroy(tsd, tcache); 411 tsd_tcache_set(tsd, NULL); 412 } 413 } 414 415 void 416 tcache_enabled_cleanup(tsd_t *tsd) 417 { 418 419 /* Do nothing. */ 420 } 421 422 void 423 tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) 424 { 425 unsigned i; 426 427 cassert(config_stats); 428 429 malloc_mutex_assert_owner(tsdn, &arena->lock); 430 431 /* Merge and reset tcache stats. */ 432 for (i = 0; i < NBINS; i++) { 433 arena_bin_t *bin = &arena->bins[i]; 434 tcache_bin_t *tbin = &tcache->tbins[i]; 435 malloc_mutex_lock(tsdn, &bin->lock); 436 bin->stats.nrequests += tbin->tstats.nrequests; 437 malloc_mutex_unlock(tsdn, &bin->lock); 438 tbin->tstats.nrequests = 0; 439 } 440 441 for (; i < nhbins; i++) { 442 malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS]; 443 tcache_bin_t *tbin = &tcache->tbins[i]; 444 arena->stats.nrequests_large += tbin->tstats.nrequests; 445 lstats->nrequests += tbin->tstats.nrequests; 446 tbin->tstats.nrequests = 0; 447 } 448 } 449 450 static bool 451 tcaches_create_prep(tsd_t *tsd) { 452 bool err; 453 454 malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx); 455 456 if (tcaches == NULL) { 457 tcaches = base_alloc(tsd_tsdn(tsd), sizeof(tcache_t *) * 458 (MALLOCX_TCACHE_MAX+1)); 459 if (tcaches == NULL) { 460 err = true; 461 goto label_return; 462 } 463 } 464 465 if (tcaches_avail == NULL && tcaches_past > MALLOCX_TCACHE_MAX) { 466 err = true; 467 goto label_return; 468 } 469 470 err = false; 471 label_return: 472 malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx); 473 return err; 474 } 475 476 bool 477 tcaches_create(tsd_t *tsd, unsigned *r_ind) { 478 bool err; 479 arena_t *arena; 480 tcache_t *tcache; 481 tcaches_t *elm; 482 483 if (tcaches_create_prep(tsd)) { 484 err = true; 485 goto label_return; 486 } 487 488 arena = arena_ichoose(tsd, NULL); 489 if (unlikely(arena == NULL)) { 490 err = true; 491 goto label_return; 492 } 493 tcache = tcache_create(tsd_tsdn(tsd), arena); 494 if (tcache == NULL) { 495 err = true; 496 goto label_return; 497 } 498 499 malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx); 500 if (tcaches_avail != NULL) { 501 elm = tcaches_avail; 502 tcaches_avail = tcaches_avail->next; 503 elm->tcache = tcache; 504 *r_ind = (unsigned)(elm - tcaches); 505 } else { 506 elm = &tcaches[tcaches_past]; 507 elm->tcache = tcache; 508 *r_ind = tcaches_past; 509 tcaches_past++; 510 } 511 malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx); 512 513 err = false; 514 label_return: 515 malloc_mutex_assert_not_owner(tsd_tsdn(tsd), &tcaches_mtx); 516 return err; 517 } 518 519 static void 520 tcaches_elm_flush(tsd_t *tsd, tcaches_t *elm) { 521 malloc_mutex_assert_owner(tsd_tsdn(tsd), &tcaches_mtx); 522 523 if (elm->tcache == NULL) { 524 return; 525 } 526 tcache_destroy(tsd, elm->tcache); 527 elm->tcache = NULL; 528 } 529 530 void 531 tcaches_flush(tsd_t *tsd, unsigned ind) { 532 malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx); 533 tcaches_elm_flush(tsd, &tcaches[ind]); 534 malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx); 535 } 536 537 void 538 tcaches_destroy(tsd_t *tsd, unsigned ind) { 539 tcaches_t *elm; 540 541 malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx); 542 elm = &tcaches[ind]; 543 tcaches_elm_flush(tsd, elm); 544 elm->next = tcaches_avail; 545 tcaches_avail = elm; 546 malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx); 547 } 548 549 bool 550 tcache_boot(tsdn_t *tsdn) { 551 unsigned i; 552 553 cassert(config_tcache); 554 555 /* 556 * If necessary, clamp opt_lg_tcache_max, now that large_maxclass is 557 * known. 558 */ 559 if (opt_lg_tcache_max < 0 || (ZU(1) << opt_lg_tcache_max) < SMALL_MAXCLASS) 560 tcache_maxclass = SMALL_MAXCLASS; 561 else if ((ZU(1) << opt_lg_tcache_max) > large_maxclass) 562 tcache_maxclass = large_maxclass; 563 else 564 tcache_maxclass = (ZU(1) << opt_lg_tcache_max); 565 566 if (malloc_mutex_init(&tcaches_mtx, "tcaches", WITNESS_RANK_TCACHES)) { 567 return true; 568 } 569 570 nhbins = size2index(tcache_maxclass) + 1; 571 572 /* Initialize tcache_bin_info. */ 573 tcache_bin_info = (tcache_bin_info_t *)base_alloc(tsdn, nhbins * 574 sizeof(tcache_bin_info_t)); 575 if (tcache_bin_info == NULL) 576 return (true); 577 stack_nelms = 0; 578 for (i = 0; i < NBINS; i++) { 579 if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MIN) { 580 tcache_bin_info[i].ncached_max = 581 TCACHE_NSLOTS_SMALL_MIN; 582 } else if ((arena_bin_info[i].nregs << 1) <= 583 TCACHE_NSLOTS_SMALL_MAX) { 584 tcache_bin_info[i].ncached_max = 585 (arena_bin_info[i].nregs << 1); 586 } else { 587 tcache_bin_info[i].ncached_max = 588 TCACHE_NSLOTS_SMALL_MAX; 589 } 590 stack_nelms += tcache_bin_info[i].ncached_max; 591 } 592 for (; i < nhbins; i++) { 593 tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_LARGE; 594 stack_nelms += tcache_bin_info[i].ncached_max; 595 } 596 597 return (false); 598 } 599 600 void 601 tcache_prefork(tsdn_t *tsdn) { 602 if (!config_prof && opt_tcache) { 603 malloc_mutex_prefork(tsdn, &tcaches_mtx); 604 } 605 } 606 607 void 608 tcache_postfork_parent(tsdn_t *tsdn) { 609 if (!config_prof && opt_tcache) { 610 malloc_mutex_postfork_parent(tsdn, &tcaches_mtx); 611 } 612 } 613 614 void 615 tcache_postfork_child(tsdn_t *tsdn) { 616 if (!config_prof && opt_tcache) { 617 malloc_mutex_postfork_child(tsdn, &tcaches_mtx); 618 } 619 } 620
