1 #define JEMALLOC_TCACHE_C_ 2 #include "jemalloc/internal/jemalloc_internal.h" 3 4 /******************************************************************************/ 5 /* Data. */ 6 7 malloc_tsd_data(, tcache, tcache_t *, NULL) 8 malloc_tsd_data(, tcache_enabled, tcache_enabled_t, tcache_enabled_default) 9 10 bool opt_tcache = true; 11 ssize_t opt_lg_tcache_max = LG_TCACHE_MAXCLASS_DEFAULT; 12 13 tcache_bin_info_t *tcache_bin_info; 14 static unsigned stack_nelms; /* Total stack elms per tcache. */ 15 16 size_t nhbins; 17 size_t tcache_maxclass; 18 19 /******************************************************************************/ 20 21 size_t tcache_salloc(const void *ptr) 22 { 23 24 return (arena_salloc(ptr, false)); 25 } 26 27 void 28 tcache_event_hard(tcache_t *tcache) 29 { 30 size_t binind = tcache->next_gc_bin; 31 tcache_bin_t *tbin = &tcache->tbins[binind]; 32 tcache_bin_info_t *tbin_info = &tcache_bin_info[binind]; 33 34 if (tbin->low_water > 0) { 35 /* 36 * Flush (ceiling) 3/4 of the objects below the low water mark. 37 */ 38 if (binind < NBINS) { 39 tcache_bin_flush_small(tbin, binind, tbin->ncached - 40 tbin->low_water + (tbin->low_water >> 2), tcache); 41 } else { 42 tcache_bin_flush_large(tbin, binind, tbin->ncached - 43 tbin->low_water + (tbin->low_water >> 2), tcache); 44 } 45 /* 46 * Reduce fill count by 2X. Limit lg_fill_div such that the 47 * fill count is always at least 1. 48 */ 49 if ((tbin_info->ncached_max >> (tbin->lg_fill_div+1)) >= 1) 50 tbin->lg_fill_div++; 51 } else if (tbin->low_water < 0) { 52 /* 53 * Increase fill count by 2X. Make sure lg_fill_div stays 54 * greater than 0. 55 */ 56 if (tbin->lg_fill_div > 1) 57 tbin->lg_fill_div--; 58 } 59 tbin->low_water = tbin->ncached; 60 61 tcache->next_gc_bin++; 62 if (tcache->next_gc_bin == nhbins) 63 tcache->next_gc_bin = 0; 64 tcache->ev_cnt = 0; 65 } 66 67 void * 68 tcache_alloc_small_hard(tcache_t *tcache, tcache_bin_t *tbin, size_t binind) 69 { 70 void *ret; 71 72 arena_tcache_fill_small(tcache->arena, tbin, binind, 73 config_prof ? tcache->prof_accumbytes : 0); 74 if (config_prof) 75 tcache->prof_accumbytes = 0; 76 ret = tcache_alloc_easy(tbin); 77 78 return (ret); 79 } 80 81 void 82 tcache_bin_flush_small(tcache_bin_t *tbin, size_t binind, unsigned rem, 83 tcache_t *tcache) 84 { 85 void *ptr; 86 unsigned i, nflush, ndeferred; 87 bool merged_stats = false; 88 89 assert(binind < NBINS); 90 assert(rem <= tbin->ncached); 91 92 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 93 /* Lock the arena bin associated with the first object. */ 94 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 95 tbin->avail[0]); 96 arena_t *arena = chunk->arena; 97 arena_bin_t *bin = &arena->bins[binind]; 98 99 if (config_prof && arena == tcache->arena) { 100 malloc_mutex_lock(&arena->lock); 101 arena_prof_accum(arena, tcache->prof_accumbytes); 102 malloc_mutex_unlock(&arena->lock); 103 tcache->prof_accumbytes = 0; 104 } 105 106 malloc_mutex_lock(&bin->lock); 107 if (config_stats && arena == tcache->arena) { 108 assert(merged_stats == false); 109 merged_stats = true; 110 bin->stats.nflushes++; 111 bin->stats.nrequests += tbin->tstats.nrequests; 112 tbin->tstats.nrequests = 0; 113 } 114 ndeferred = 0; 115 for (i = 0; i < nflush; i++) { 116 ptr = tbin->avail[i]; 117 assert(ptr != NULL); 118 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 119 if (chunk->arena == arena) { 120 size_t pageind = ((uintptr_t)ptr - 121 (uintptr_t)chunk) >> LG_PAGE; 122 arena_chunk_map_t *mapelm = 123 arena_mapp_get(chunk, pageind); 124 if (config_fill && opt_junk) { 125 arena_alloc_junk_small(ptr, 126 &arena_bin_info[binind], true); 127 } 128 arena_dalloc_bin_locked(arena, chunk, ptr, 129 mapelm); 130 } else { 131 /* 132 * This object was allocated via a different 133 * arena bin than the one that is currently 134 * locked. Stash the object, so that it can be 135 * handled in a future pass. 136 */ 137 tbin->avail[ndeferred] = ptr; 138 ndeferred++; 139 } 140 } 141 malloc_mutex_unlock(&bin->lock); 142 } 143 if (config_stats && merged_stats == false) { 144 /* 145 * The flush loop didn't happen to flush to this thread's 146 * arena, so the stats didn't get merged. Manually do so now. 147 */ 148 arena_bin_t *bin = &tcache->arena->bins[binind]; 149 malloc_mutex_lock(&bin->lock); 150 bin->stats.nflushes++; 151 bin->stats.nrequests += tbin->tstats.nrequests; 152 tbin->tstats.nrequests = 0; 153 malloc_mutex_unlock(&bin->lock); 154 } 155 156 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 157 rem * sizeof(void *)); 158 tbin->ncached = rem; 159 if ((int)tbin->ncached < tbin->low_water) 160 tbin->low_water = tbin->ncached; 161 } 162 163 void 164 tcache_bin_flush_large(tcache_bin_t *tbin, size_t binind, unsigned rem, 165 tcache_t *tcache) 166 { 167 void *ptr; 168 unsigned i, nflush, ndeferred; 169 bool merged_stats = false; 170 171 assert(binind < nhbins); 172 assert(rem <= tbin->ncached); 173 174 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 175 /* Lock the arena associated with the first object. */ 176 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 177 tbin->avail[0]); 178 arena_t *arena = chunk->arena; 179 180 malloc_mutex_lock(&arena->lock); 181 if ((config_prof || config_stats) && arena == tcache->arena) { 182 if (config_prof) { 183 arena_prof_accum(arena, 184 tcache->prof_accumbytes); 185 tcache->prof_accumbytes = 0; 186 } 187 if (config_stats) { 188 merged_stats = true; 189 arena->stats.nrequests_large += 190 tbin->tstats.nrequests; 191 arena->stats.lstats[binind - NBINS].nrequests += 192 tbin->tstats.nrequests; 193 tbin->tstats.nrequests = 0; 194 } 195 } 196 ndeferred = 0; 197 for (i = 0; i < nflush; i++) { 198 ptr = tbin->avail[i]; 199 assert(ptr != NULL); 200 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 201 if (chunk->arena == arena) 202 arena_dalloc_large_locked(arena, chunk, ptr); 203 else { 204 /* 205 * This object was allocated via a different 206 * arena than the one that is currently locked. 207 * Stash the object, so that it can be handled 208 * in a future pass. 209 */ 210 tbin->avail[ndeferred] = ptr; 211 ndeferred++; 212 } 213 } 214 malloc_mutex_unlock(&arena->lock); 215 } 216 if (config_stats && merged_stats == false) { 217 /* 218 * The flush loop didn't happen to flush to this thread's 219 * arena, so the stats didn't get merged. Manually do so now. 220 */ 221 arena_t *arena = tcache->arena; 222 malloc_mutex_lock(&arena->lock); 223 arena->stats.nrequests_large += tbin->tstats.nrequests; 224 arena->stats.lstats[binind - NBINS].nrequests += 225 tbin->tstats.nrequests; 226 tbin->tstats.nrequests = 0; 227 malloc_mutex_unlock(&arena->lock); 228 } 229 230 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 231 rem * sizeof(void *)); 232 tbin->ncached = rem; 233 if ((int)tbin->ncached < tbin->low_water) 234 tbin->low_water = tbin->ncached; 235 } 236 237 void 238 tcache_arena_associate(tcache_t *tcache, arena_t *arena) 239 { 240 241 if (config_stats) { 242 /* Link into list of extant tcaches. */ 243 malloc_mutex_lock(&arena->lock); 244 ql_elm_new(tcache, link); 245 ql_tail_insert(&arena->tcache_ql, tcache, link); 246 malloc_mutex_unlock(&arena->lock); 247 } 248 tcache->arena = arena; 249 } 250 251 void 252 tcache_arena_dissociate(tcache_t *tcache) 253 { 254 255 if (config_stats) { 256 /* Unlink from list of extant tcaches. */ 257 malloc_mutex_lock(&tcache->arena->lock); 258 ql_remove(&tcache->arena->tcache_ql, tcache, link); 259 malloc_mutex_unlock(&tcache->arena->lock); 260 tcache_stats_merge(tcache, tcache->arena); 261 } 262 } 263 264 tcache_t * 265 tcache_create(arena_t *arena) 266 { 267 tcache_t *tcache; 268 size_t size, stack_offset; 269 unsigned i; 270 271 size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins); 272 /* Naturally align the pointer stacks. */ 273 size = PTR_CEILING(size); 274 stack_offset = size; 275 size += stack_nelms * sizeof(void *); 276 /* 277 * Round up to the nearest multiple of the cacheline size, in order to 278 * avoid the possibility of false cacheline sharing. 279 * 280 * That this works relies on the same logic as in ipalloc(), but we 281 * cannot directly call ipalloc() here due to tcache bootstrapping 282 * issues. 283 */ 284 size = (size + CACHELINE_MASK) & (-CACHELINE); 285 286 if (size <= SMALL_MAXCLASS) 287 tcache = (tcache_t *)arena_malloc_small(arena, size, true); 288 else if (size <= tcache_maxclass) 289 tcache = (tcache_t *)arena_malloc_large(arena, size, true); 290 else 291 tcache = (tcache_t *)icalloc(size); 292 293 if (tcache == NULL) 294 return (NULL); 295 296 tcache_arena_associate(tcache, arena); 297 298 assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0); 299 for (i = 0; i < nhbins; i++) { 300 tcache->tbins[i].lg_fill_div = 1; 301 tcache->tbins[i].avail = (void **)((uintptr_t)tcache + 302 (uintptr_t)stack_offset); 303 stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *); 304 } 305 306 tcache_tsd_set(&tcache); 307 308 return (tcache); 309 } 310 311 void 312 tcache_destroy(tcache_t *tcache) 313 { 314 unsigned i; 315 size_t tcache_size; 316 317 tcache_arena_dissociate(tcache); 318 319 for (i = 0; i < NBINS; i++) { 320 tcache_bin_t *tbin = &tcache->tbins[i]; 321 tcache_bin_flush_small(tbin, i, 0, tcache); 322 323 if (config_stats && tbin->tstats.nrequests != 0) { 324 arena_t *arena = tcache->arena; 325 arena_bin_t *bin = &arena->bins[i]; 326 malloc_mutex_lock(&bin->lock); 327 bin->stats.nrequests += tbin->tstats.nrequests; 328 malloc_mutex_unlock(&bin->lock); 329 } 330 } 331 332 for (; i < nhbins; i++) { 333 tcache_bin_t *tbin = &tcache->tbins[i]; 334 tcache_bin_flush_large(tbin, i, 0, tcache); 335 336 if (config_stats && tbin->tstats.nrequests != 0) { 337 arena_t *arena = tcache->arena; 338 malloc_mutex_lock(&arena->lock); 339 arena->stats.nrequests_large += tbin->tstats.nrequests; 340 arena->stats.lstats[i - NBINS].nrequests += 341 tbin->tstats.nrequests; 342 malloc_mutex_unlock(&arena->lock); 343 } 344 } 345 346 if (config_prof && tcache->prof_accumbytes > 0) { 347 malloc_mutex_lock(&tcache->arena->lock); 348 arena_prof_accum(tcache->arena, tcache->prof_accumbytes); 349 malloc_mutex_unlock(&tcache->arena->lock); 350 } 351 352 tcache_size = arena_salloc(tcache, false); 353 if (tcache_size <= SMALL_MAXCLASS) { 354 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 355 arena_t *arena = chunk->arena; 356 size_t pageind = ((uintptr_t)tcache - (uintptr_t)chunk) >> 357 LG_PAGE; 358 arena_chunk_map_t *mapelm = arena_mapp_get(chunk, pageind); 359 360 arena_dalloc_bin(arena, chunk, tcache, pageind, mapelm); 361 } else if (tcache_size <= tcache_maxclass) { 362 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 363 arena_t *arena = chunk->arena; 364 365 arena_dalloc_large(arena, chunk, tcache); 366 } else 367 idalloc(tcache); 368 } 369 370 void 371 tcache_thread_cleanup(void *arg) 372 { 373 tcache_t *tcache = *(tcache_t **)arg; 374 375 if (tcache == TCACHE_STATE_DISABLED) { 376 /* Do nothing. */ 377 } else if (tcache == TCACHE_STATE_REINCARNATED) { 378 /* 379 * Another destructor called an allocator function after this 380 * destructor was called. Reset tcache to 381 * TCACHE_STATE_PURGATORY in order to receive another callback. 382 */ 383 tcache = TCACHE_STATE_PURGATORY; 384 tcache_tsd_set(&tcache); 385 } else if (tcache == TCACHE_STATE_PURGATORY) { 386 /* 387 * The previous time this destructor was called, we set the key 388 * to TCACHE_STATE_PURGATORY so that other destructors wouldn't 389 * cause re-creation of the tcache. This time, do nothing, so 390 * that the destructor will not be called again. 391 */ 392 } else if (tcache != NULL) { 393 assert(tcache != TCACHE_STATE_PURGATORY); 394 tcache_destroy(tcache); 395 tcache = TCACHE_STATE_PURGATORY; 396 tcache_tsd_set(&tcache); 397 } 398 } 399 400 void 401 tcache_stats_merge(tcache_t *tcache, arena_t *arena) 402 { 403 unsigned i; 404 405 /* Merge and reset tcache stats. */ 406 for (i = 0; i < NBINS; i++) { 407 arena_bin_t *bin = &arena->bins[i]; 408 tcache_bin_t *tbin = &tcache->tbins[i]; 409 malloc_mutex_lock(&bin->lock); 410 bin->stats.nrequests += tbin->tstats.nrequests; 411 malloc_mutex_unlock(&bin->lock); 412 tbin->tstats.nrequests = 0; 413 } 414 415 for (; i < nhbins; i++) { 416 malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS]; 417 tcache_bin_t *tbin = &tcache->tbins[i]; 418 arena->stats.nrequests_large += tbin->tstats.nrequests; 419 lstats->nrequests += tbin->tstats.nrequests; 420 tbin->tstats.nrequests = 0; 421 } 422 } 423 424 bool 425 tcache_boot0(void) 426 { 427 unsigned i; 428 429 /* 430 * If necessary, clamp opt_lg_tcache_max, now that arena_maxclass is 431 * known. 432 */ 433 if (opt_lg_tcache_max < 0 || (1U << opt_lg_tcache_max) < SMALL_MAXCLASS) 434 tcache_maxclass = SMALL_MAXCLASS; 435 else if ((1U << opt_lg_tcache_max) > arena_maxclass) 436 tcache_maxclass = arena_maxclass; 437 else 438 tcache_maxclass = (1U << opt_lg_tcache_max); 439 440 nhbins = NBINS + (tcache_maxclass >> LG_PAGE); 441 442 /* Initialize tcache_bin_info. */ 443 tcache_bin_info = (tcache_bin_info_t *)base_alloc(nhbins * 444 sizeof(tcache_bin_info_t)); 445 if (tcache_bin_info == NULL) 446 return (true); 447 stack_nelms = 0; 448 for (i = 0; i < NBINS; i++) { 449 if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MAX) { 450 tcache_bin_info[i].ncached_max = 451 (arena_bin_info[i].nregs << 1); 452 } else { 453 tcache_bin_info[i].ncached_max = 454 TCACHE_NSLOTS_SMALL_MAX; 455 } 456 stack_nelms += tcache_bin_info[i].ncached_max; 457 } 458 for (; i < nhbins; i++) { 459 tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_LARGE; 460 stack_nelms += tcache_bin_info[i].ncached_max; 461 } 462 463 return (false); 464 } 465 466 bool 467 tcache_boot1(void) 468 { 469 470 if (tcache_tsd_boot() || tcache_enabled_tsd_boot()) 471 return (true); 472 473 return (false); 474 } 475