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 arena_prof_accum(arena, tcache->prof_accumbytes); 101 tcache->prof_accumbytes = 0; 102 } 103 104 malloc_mutex_lock(&bin->lock); 105 if (config_stats && arena == tcache->arena) { 106 assert(merged_stats == false); 107 merged_stats = true; 108 bin->stats.nflushes++; 109 bin->stats.nrequests += tbin->tstats.nrequests; 110 tbin->tstats.nrequests = 0; 111 } 112 ndeferred = 0; 113 for (i = 0; i < nflush; i++) { 114 ptr = tbin->avail[i]; 115 assert(ptr != NULL); 116 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 117 if (chunk->arena == arena) { 118 size_t pageind = ((uintptr_t)ptr - 119 (uintptr_t)chunk) >> LG_PAGE; 120 arena_chunk_map_t *mapelm = 121 arena_mapp_get(chunk, pageind); 122 if (config_fill && opt_junk) { 123 arena_alloc_junk_small(ptr, 124 &arena_bin_info[binind], true); 125 } 126 arena_dalloc_bin_locked(arena, chunk, ptr, 127 mapelm); 128 } else { 129 /* 130 * This object was allocated via a different 131 * arena bin than the one that is currently 132 * locked. Stash the object, so that it can be 133 * handled in a future pass. 134 */ 135 tbin->avail[ndeferred] = ptr; 136 ndeferred++; 137 } 138 } 139 malloc_mutex_unlock(&bin->lock); 140 } 141 if (config_stats && merged_stats == false) { 142 /* 143 * The flush loop didn't happen to flush to this thread's 144 * arena, so the stats didn't get merged. Manually do so now. 145 */ 146 arena_bin_t *bin = &tcache->arena->bins[binind]; 147 malloc_mutex_lock(&bin->lock); 148 bin->stats.nflushes++; 149 bin->stats.nrequests += tbin->tstats.nrequests; 150 tbin->tstats.nrequests = 0; 151 malloc_mutex_unlock(&bin->lock); 152 } 153 154 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 155 rem * sizeof(void *)); 156 tbin->ncached = rem; 157 if ((int)tbin->ncached < tbin->low_water) 158 tbin->low_water = tbin->ncached; 159 } 160 161 void 162 tcache_bin_flush_large(tcache_bin_t *tbin, size_t binind, unsigned rem, 163 tcache_t *tcache) 164 { 165 void *ptr; 166 unsigned i, nflush, ndeferred; 167 bool merged_stats = false; 168 169 assert(binind < nhbins); 170 assert(rem <= tbin->ncached); 171 172 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 173 /* Lock the arena associated with the first object. */ 174 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 175 tbin->avail[0]); 176 arena_t *arena = chunk->arena; 177 178 malloc_mutex_lock(&arena->lock); 179 if ((config_prof || config_stats) && arena == tcache->arena) { 180 if (config_prof) { 181 arena_prof_accum_locked(arena, 182 tcache->prof_accumbytes); 183 tcache->prof_accumbytes = 0; 184 } 185 if (config_stats) { 186 merged_stats = true; 187 arena->stats.nrequests_large += 188 tbin->tstats.nrequests; 189 arena->stats.lstats[binind - NBINS].nrequests += 190 tbin->tstats.nrequests; 191 tbin->tstats.nrequests = 0; 192 } 193 } 194 ndeferred = 0; 195 for (i = 0; i < nflush; i++) { 196 ptr = tbin->avail[i]; 197 assert(ptr != NULL); 198 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 199 if (chunk->arena == arena) 200 arena_dalloc_large_locked(arena, chunk, ptr); 201 else { 202 /* 203 * This object was allocated via a different 204 * arena than the one that is currently locked. 205 * Stash the object, so that it can be handled 206 * in a future pass. 207 */ 208 tbin->avail[ndeferred] = ptr; 209 ndeferred++; 210 } 211 } 212 malloc_mutex_unlock(&arena->lock); 213 } 214 if (config_stats && merged_stats == false) { 215 /* 216 * The flush loop didn't happen to flush to this thread's 217 * arena, so the stats didn't get merged. Manually do so now. 218 */ 219 arena_t *arena = tcache->arena; 220 malloc_mutex_lock(&arena->lock); 221 arena->stats.nrequests_large += tbin->tstats.nrequests; 222 arena->stats.lstats[binind - NBINS].nrequests += 223 tbin->tstats.nrequests; 224 tbin->tstats.nrequests = 0; 225 malloc_mutex_unlock(&arena->lock); 226 } 227 228 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 229 rem * sizeof(void *)); 230 tbin->ncached = rem; 231 if ((int)tbin->ncached < tbin->low_water) 232 tbin->low_water = tbin->ncached; 233 } 234 235 void 236 tcache_arena_associate(tcache_t *tcache, arena_t *arena) 237 { 238 239 if (config_stats) { 240 /* Link into list of extant tcaches. */ 241 malloc_mutex_lock(&arena->lock); 242 ql_elm_new(tcache, link); 243 ql_tail_insert(&arena->tcache_ql, tcache, link); 244 malloc_mutex_unlock(&arena->lock); 245 } 246 tcache->arena = arena; 247 } 248 249 void 250 tcache_arena_dissociate(tcache_t *tcache) 251 { 252 253 if (config_stats) { 254 /* Unlink from list of extant tcaches. */ 255 malloc_mutex_lock(&tcache->arena->lock); 256 ql_remove(&tcache->arena->tcache_ql, tcache, link); 257 malloc_mutex_unlock(&tcache->arena->lock); 258 tcache_stats_merge(tcache, tcache->arena); 259 } 260 } 261 262 tcache_t * 263 tcache_create(arena_t *arena) 264 { 265 tcache_t *tcache; 266 size_t size, stack_offset; 267 unsigned i; 268 269 size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins); 270 /* Naturally align the pointer stacks. */ 271 size = PTR_CEILING(size); 272 stack_offset = size; 273 size += stack_nelms * sizeof(void *); 274 /* 275 * Round up to the nearest multiple of the cacheline size, in order to 276 * avoid the possibility of false cacheline sharing. 277 * 278 * That this works relies on the same logic as in ipalloc(), but we 279 * cannot directly call ipalloc() here due to tcache bootstrapping 280 * issues. 281 */ 282 size = (size + CACHELINE_MASK) & (-CACHELINE); 283 284 if (size <= SMALL_MAXCLASS) 285 tcache = (tcache_t *)arena_malloc_small(arena, size, true); 286 else if (size <= tcache_maxclass) 287 tcache = (tcache_t *)arena_malloc_large(arena, size, true); 288 else 289 tcache = (tcache_t *)icallocx(size, false, arena); 290 291 if (tcache == NULL) 292 return (NULL); 293 294 tcache_arena_associate(tcache, arena); 295 296 assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0); 297 for (i = 0; i < nhbins; i++) { 298 tcache->tbins[i].lg_fill_div = 1; 299 tcache->tbins[i].avail = (void **)((uintptr_t)tcache + 300 (uintptr_t)stack_offset); 301 stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *); 302 } 303 304 tcache_tsd_set(&tcache); 305 306 return (tcache); 307 } 308 309 void 310 tcache_destroy(tcache_t *tcache) 311 { 312 unsigned i; 313 size_t tcache_size; 314 315 tcache_arena_dissociate(tcache); 316 317 for (i = 0; i < NBINS; i++) { 318 tcache_bin_t *tbin = &tcache->tbins[i]; 319 tcache_bin_flush_small(tbin, i, 0, tcache); 320 321 if (config_stats && tbin->tstats.nrequests != 0) { 322 arena_t *arena = tcache->arena; 323 arena_bin_t *bin = &arena->bins[i]; 324 malloc_mutex_lock(&bin->lock); 325 bin->stats.nrequests += tbin->tstats.nrequests; 326 malloc_mutex_unlock(&bin->lock); 327 } 328 } 329 330 for (; i < nhbins; i++) { 331 tcache_bin_t *tbin = &tcache->tbins[i]; 332 tcache_bin_flush_large(tbin, i, 0, tcache); 333 334 if (config_stats && tbin->tstats.nrequests != 0) { 335 arena_t *arena = tcache->arena; 336 malloc_mutex_lock(&arena->lock); 337 arena->stats.nrequests_large += tbin->tstats.nrequests; 338 arena->stats.lstats[i - NBINS].nrequests += 339 tbin->tstats.nrequests; 340 malloc_mutex_unlock(&arena->lock); 341 } 342 } 343 344 if (config_prof && tcache->prof_accumbytes > 0) 345 arena_prof_accum(tcache->arena, tcache->prof_accumbytes); 346 347 tcache_size = arena_salloc(tcache, false); 348 if (tcache_size <= SMALL_MAXCLASS) { 349 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 350 arena_t *arena = chunk->arena; 351 size_t pageind = ((uintptr_t)tcache - (uintptr_t)chunk) >> 352 LG_PAGE; 353 arena_chunk_map_t *mapelm = arena_mapp_get(chunk, pageind); 354 355 arena_dalloc_bin(arena, chunk, tcache, pageind, mapelm); 356 } else if (tcache_size <= tcache_maxclass) { 357 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 358 arena_t *arena = chunk->arena; 359 360 arena_dalloc_large(arena, chunk, tcache); 361 } else 362 idallocx(tcache, false); 363 } 364 365 void 366 tcache_thread_cleanup(void *arg) 367 { 368 tcache_t *tcache = *(tcache_t **)arg; 369 370 if (tcache == TCACHE_STATE_DISABLED) { 371 /* Do nothing. */ 372 } else if (tcache == TCACHE_STATE_REINCARNATED) { 373 /* 374 * Another destructor called an allocator function after this 375 * destructor was called. Reset tcache to 376 * TCACHE_STATE_PURGATORY in order to receive another callback. 377 */ 378 tcache = TCACHE_STATE_PURGATORY; 379 tcache_tsd_set(&tcache); 380 } else if (tcache == TCACHE_STATE_PURGATORY) { 381 /* 382 * The previous time this destructor was called, we set the key 383 * to TCACHE_STATE_PURGATORY so that other destructors wouldn't 384 * cause re-creation of the tcache. This time, do nothing, so 385 * that the destructor will not be called again. 386 */ 387 } else if (tcache != NULL) { 388 assert(tcache != TCACHE_STATE_PURGATORY); 389 tcache_destroy(tcache); 390 tcache = TCACHE_STATE_PURGATORY; 391 tcache_tsd_set(&tcache); 392 } 393 } 394 395 void 396 tcache_stats_merge(tcache_t *tcache, arena_t *arena) 397 { 398 unsigned i; 399 400 /* Merge and reset tcache stats. */ 401 for (i = 0; i < NBINS; i++) { 402 arena_bin_t *bin = &arena->bins[i]; 403 tcache_bin_t *tbin = &tcache->tbins[i]; 404 malloc_mutex_lock(&bin->lock); 405 bin->stats.nrequests += tbin->tstats.nrequests; 406 malloc_mutex_unlock(&bin->lock); 407 tbin->tstats.nrequests = 0; 408 } 409 410 for (; i < nhbins; i++) { 411 malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS]; 412 tcache_bin_t *tbin = &tcache->tbins[i]; 413 arena->stats.nrequests_large += tbin->tstats.nrequests; 414 lstats->nrequests += tbin->tstats.nrequests; 415 tbin->tstats.nrequests = 0; 416 } 417 } 418 419 bool 420 tcache_boot0(void) 421 { 422 unsigned i; 423 424 /* 425 * If necessary, clamp opt_lg_tcache_max, now that arena_maxclass is 426 * known. 427 */ 428 if (opt_lg_tcache_max < 0 || (1U << opt_lg_tcache_max) < SMALL_MAXCLASS) 429 tcache_maxclass = SMALL_MAXCLASS; 430 else if ((1U << opt_lg_tcache_max) > arena_maxclass) 431 tcache_maxclass = arena_maxclass; 432 else 433 tcache_maxclass = (1U << opt_lg_tcache_max); 434 435 nhbins = NBINS + (tcache_maxclass >> LG_PAGE); 436 437 /* Initialize tcache_bin_info. */ 438 tcache_bin_info = (tcache_bin_info_t *)base_alloc(nhbins * 439 sizeof(tcache_bin_info_t)); 440 if (tcache_bin_info == NULL) 441 return (true); 442 stack_nelms = 0; 443 for (i = 0; i < NBINS; i++) { 444 if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MAX) { 445 tcache_bin_info[i].ncached_max = 446 (arena_bin_info[i].nregs << 1); 447 } else { 448 tcache_bin_info[i].ncached_max = 449 TCACHE_NSLOTS_SMALL_MAX; 450 } 451 stack_nelms += tcache_bin_info[i].ncached_max; 452 } 453 for (; i < nhbins; i++) { 454 tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_LARGE; 455 stack_nelms += tcache_bin_info[i].ncached_max; 456 } 457 458 return (false); 459 } 460 461 bool 462 tcache_boot1(void) 463 { 464 465 if (tcache_tsd_boot() || tcache_enabled_tsd_boot()) 466 return (true); 467 468 return (false); 469 } 470