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 if (arena_prof_accum(arena, tcache->prof_accumbytes)) 101 prof_idump(); 102 tcache->prof_accumbytes = 0; 103 } 104 105 malloc_mutex_lock(&bin->lock); 106 if (config_stats && arena == tcache->arena) { 107 assert(merged_stats == false); 108 merged_stats = true; 109 bin->stats.nflushes++; 110 bin->stats.nrequests += tbin->tstats.nrequests; 111 tbin->tstats.nrequests = 0; 112 } 113 ndeferred = 0; 114 for (i = 0; i < nflush; i++) { 115 ptr = tbin->avail[i]; 116 assert(ptr != NULL); 117 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 118 if (chunk->arena == arena) { 119 size_t pageind = ((uintptr_t)ptr - 120 (uintptr_t)chunk) >> LG_PAGE; 121 arena_chunk_map_t *mapelm = 122 arena_mapp_get(chunk, pageind); 123 if (config_fill && opt_junk) { 124 arena_alloc_junk_small(ptr, 125 &arena_bin_info[binind], true); 126 } 127 arena_dalloc_bin_locked(arena, chunk, ptr, 128 mapelm); 129 } else { 130 /* 131 * This object was allocated via a different 132 * arena bin than the one that is currently 133 * locked. Stash the object, so that it can be 134 * handled in a future pass. 135 */ 136 tbin->avail[ndeferred] = ptr; 137 ndeferred++; 138 } 139 } 140 malloc_mutex_unlock(&bin->lock); 141 } 142 if (config_stats && merged_stats == false) { 143 /* 144 * The flush loop didn't happen to flush to this thread's 145 * arena, so the stats didn't get merged. Manually do so now. 146 */ 147 arena_bin_t *bin = &tcache->arena->bins[binind]; 148 malloc_mutex_lock(&bin->lock); 149 bin->stats.nflushes++; 150 bin->stats.nrequests += tbin->tstats.nrequests; 151 tbin->tstats.nrequests = 0; 152 malloc_mutex_unlock(&bin->lock); 153 } 154 155 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 156 rem * sizeof(void *)); 157 tbin->ncached = rem; 158 if ((int)tbin->ncached < tbin->low_water) 159 tbin->low_water = tbin->ncached; 160 } 161 162 void 163 tcache_bin_flush_large(tcache_bin_t *tbin, size_t binind, unsigned rem, 164 tcache_t *tcache) 165 { 166 void *ptr; 167 unsigned i, nflush, ndeferred; 168 bool merged_stats = false; 169 170 assert(binind < nhbins); 171 assert(rem <= tbin->ncached); 172 173 for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) { 174 /* Lock the arena associated with the first object. */ 175 arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE( 176 tbin->avail[0]); 177 arena_t *arena = chunk->arena; 178 UNUSED bool idump; 179 180 if (config_prof) 181 idump = false; 182 malloc_mutex_lock(&arena->lock); 183 if ((config_prof || config_stats) && arena == tcache->arena) { 184 if (config_prof) { 185 idump = arena_prof_accum_locked(arena, 186 tcache->prof_accumbytes); 187 tcache->prof_accumbytes = 0; 188 } 189 if (config_stats) { 190 merged_stats = true; 191 arena->stats.nrequests_large += 192 tbin->tstats.nrequests; 193 arena->stats.lstats[binind - NBINS].nrequests += 194 tbin->tstats.nrequests; 195 tbin->tstats.nrequests = 0; 196 } 197 } 198 ndeferred = 0; 199 for (i = 0; i < nflush; i++) { 200 ptr = tbin->avail[i]; 201 assert(ptr != NULL); 202 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 203 if (chunk->arena == arena) 204 arena_dalloc_large_locked(arena, chunk, ptr); 205 else { 206 /* 207 * This object was allocated via a different 208 * arena than the one that is currently locked. 209 * Stash the object, so that it can be handled 210 * in a future pass. 211 */ 212 tbin->avail[ndeferred] = ptr; 213 ndeferred++; 214 } 215 } 216 malloc_mutex_unlock(&arena->lock); 217 if (config_prof && idump) 218 prof_idump(); 219 } 220 if (config_stats && merged_stats == false) { 221 /* 222 * The flush loop didn't happen to flush to this thread's 223 * arena, so the stats didn't get merged. Manually do so now. 224 */ 225 arena_t *arena = tcache->arena; 226 malloc_mutex_lock(&arena->lock); 227 arena->stats.nrequests_large += tbin->tstats.nrequests; 228 arena->stats.lstats[binind - NBINS].nrequests += 229 tbin->tstats.nrequests; 230 tbin->tstats.nrequests = 0; 231 malloc_mutex_unlock(&arena->lock); 232 } 233 234 memmove(tbin->avail, &tbin->avail[tbin->ncached - rem], 235 rem * sizeof(void *)); 236 tbin->ncached = rem; 237 if ((int)tbin->ncached < tbin->low_water) 238 tbin->low_water = tbin->ncached; 239 } 240 241 void 242 tcache_arena_associate(tcache_t *tcache, arena_t *arena) 243 { 244 245 if (config_stats) { 246 /* Link into list of extant tcaches. */ 247 malloc_mutex_lock(&arena->lock); 248 ql_elm_new(tcache, link); 249 ql_tail_insert(&arena->tcache_ql, tcache, link); 250 malloc_mutex_unlock(&arena->lock); 251 } 252 tcache->arena = arena; 253 } 254 255 void 256 tcache_arena_dissociate(tcache_t *tcache) 257 { 258 259 if (config_stats) { 260 /* Unlink from list of extant tcaches. */ 261 malloc_mutex_lock(&tcache->arena->lock); 262 ql_remove(&tcache->arena->tcache_ql, tcache, link); 263 malloc_mutex_unlock(&tcache->arena->lock); 264 tcache_stats_merge(tcache, tcache->arena); 265 } 266 } 267 268 tcache_t * 269 tcache_create(arena_t *arena) 270 { 271 tcache_t *tcache; 272 size_t size, stack_offset; 273 unsigned i; 274 275 size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins); 276 /* Naturally align the pointer stacks. */ 277 size = PTR_CEILING(size); 278 stack_offset = size; 279 size += stack_nelms * sizeof(void *); 280 /* 281 * Round up to the nearest multiple of the cacheline size, in order to 282 * avoid the possibility of false cacheline sharing. 283 * 284 * That this works relies on the same logic as in ipalloc(), but we 285 * cannot directly call ipalloc() here due to tcache bootstrapping 286 * issues. 287 */ 288 size = (size + CACHELINE_MASK) & (-CACHELINE); 289 290 if (size <= SMALL_MAXCLASS) 291 tcache = (tcache_t *)arena_malloc_small(arena, size, true); 292 else if (size <= tcache_maxclass) 293 tcache = (tcache_t *)arena_malloc_large(arena, size, true); 294 else 295 tcache = (tcache_t *)icallocx(size, false, arena); 296 297 if (tcache == NULL) 298 return (NULL); 299 300 tcache_arena_associate(tcache, arena); 301 302 assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0); 303 for (i = 0; i < nhbins; i++) { 304 tcache->tbins[i].lg_fill_div = 1; 305 tcache->tbins[i].avail = (void **)((uintptr_t)tcache + 306 (uintptr_t)stack_offset); 307 stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *); 308 } 309 310 tcache_tsd_set(&tcache); 311 312 return (tcache); 313 } 314 315 void 316 tcache_destroy(tcache_t *tcache) 317 { 318 unsigned i; 319 size_t tcache_size; 320 321 tcache_arena_dissociate(tcache); 322 323 for (i = 0; i < NBINS; i++) { 324 tcache_bin_t *tbin = &tcache->tbins[i]; 325 tcache_bin_flush_small(tbin, i, 0, tcache); 326 327 if (config_stats && tbin->tstats.nrequests != 0) { 328 arena_t *arena = tcache->arena; 329 arena_bin_t *bin = &arena->bins[i]; 330 malloc_mutex_lock(&bin->lock); 331 bin->stats.nrequests += tbin->tstats.nrequests; 332 malloc_mutex_unlock(&bin->lock); 333 } 334 } 335 336 for (; i < nhbins; i++) { 337 tcache_bin_t *tbin = &tcache->tbins[i]; 338 tcache_bin_flush_large(tbin, i, 0, tcache); 339 340 if (config_stats && tbin->tstats.nrequests != 0) { 341 arena_t *arena = tcache->arena; 342 malloc_mutex_lock(&arena->lock); 343 arena->stats.nrequests_large += tbin->tstats.nrequests; 344 arena->stats.lstats[i - NBINS].nrequests += 345 tbin->tstats.nrequests; 346 malloc_mutex_unlock(&arena->lock); 347 } 348 } 349 350 if (config_prof && tcache->prof_accumbytes > 0 && 351 arena_prof_accum(tcache->arena, tcache->prof_accumbytes)) 352 prof_idump(); 353 354 tcache_size = arena_salloc(tcache, false); 355 if (tcache_size <= SMALL_MAXCLASS) { 356 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 357 arena_t *arena = chunk->arena; 358 size_t pageind = ((uintptr_t)tcache - (uintptr_t)chunk) >> 359 LG_PAGE; 360 arena_chunk_map_t *mapelm = arena_mapp_get(chunk, pageind); 361 362 arena_dalloc_bin(arena, chunk, tcache, pageind, mapelm); 363 } else if (tcache_size <= tcache_maxclass) { 364 arena_chunk_t *chunk = CHUNK_ADDR2BASE(tcache); 365 arena_t *arena = chunk->arena; 366 367 arena_dalloc_large(arena, chunk, tcache); 368 } else 369 idallocx(tcache, false); 370 } 371 372 void 373 tcache_thread_cleanup(void *arg) 374 { 375 tcache_t *tcache = *(tcache_t **)arg; 376 377 if (tcache == TCACHE_STATE_DISABLED) { 378 /* Do nothing. */ 379 } else if (tcache == TCACHE_STATE_REINCARNATED) { 380 /* 381 * Another destructor called an allocator function after this 382 * destructor was called. Reset tcache to 383 * TCACHE_STATE_PURGATORY in order to receive another callback. 384 */ 385 tcache = TCACHE_STATE_PURGATORY; 386 tcache_tsd_set(&tcache); 387 } else if (tcache == TCACHE_STATE_PURGATORY) { 388 /* 389 * The previous time this destructor was called, we set the key 390 * to TCACHE_STATE_PURGATORY so that other destructors wouldn't 391 * cause re-creation of the tcache. This time, do nothing, so 392 * that the destructor will not be called again. 393 */ 394 } else if (tcache != NULL) { 395 assert(tcache != TCACHE_STATE_PURGATORY); 396 tcache_destroy(tcache); 397 tcache = TCACHE_STATE_PURGATORY; 398 tcache_tsd_set(&tcache); 399 } 400 } 401 402 void 403 tcache_stats_merge(tcache_t *tcache, arena_t *arena) 404 { 405 unsigned i; 406 407 /* Merge and reset tcache stats. */ 408 for (i = 0; i < NBINS; i++) { 409 arena_bin_t *bin = &arena->bins[i]; 410 tcache_bin_t *tbin = &tcache->tbins[i]; 411 malloc_mutex_lock(&bin->lock); 412 bin->stats.nrequests += tbin->tstats.nrequests; 413 malloc_mutex_unlock(&bin->lock); 414 tbin->tstats.nrequests = 0; 415 } 416 417 for (; i < nhbins; i++) { 418 malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS]; 419 tcache_bin_t *tbin = &tcache->tbins[i]; 420 arena->stats.nrequests_large += tbin->tstats.nrequests; 421 lstats->nrequests += tbin->tstats.nrequests; 422 tbin->tstats.nrequests = 0; 423 } 424 } 425 426 bool 427 tcache_boot0(void) 428 { 429 unsigned i; 430 431 /* 432 * If necessary, clamp opt_lg_tcache_max, now that arena_maxclass is 433 * known. 434 */ 435 if (opt_lg_tcache_max < 0 || (1U << opt_lg_tcache_max) < SMALL_MAXCLASS) 436 tcache_maxclass = SMALL_MAXCLASS; 437 else if ((1U << opt_lg_tcache_max) > arena_maxclass) 438 tcache_maxclass = arena_maxclass; 439 else 440 tcache_maxclass = (1U << opt_lg_tcache_max); 441 442 nhbins = NBINS + (tcache_maxclass >> LG_PAGE); 443 444 /* Initialize tcache_bin_info. */ 445 tcache_bin_info = (tcache_bin_info_t *)base_alloc(nhbins * 446 sizeof(tcache_bin_info_t)); 447 if (tcache_bin_info == NULL) 448 return (true); 449 stack_nelms = 0; 450 for (i = 0; i < NBINS; i++) { 451 if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MAX) { 452 tcache_bin_info[i].ncached_max = 453 (arena_bin_info[i].nregs << 1); 454 } else { 455 tcache_bin_info[i].ncached_max = 456 TCACHE_NSLOTS_SMALL_MAX; 457 } 458 stack_nelms += tcache_bin_info[i].ncached_max; 459 } 460 for (; i < nhbins; i++) { 461 tcache_bin_info[i].ncached_max = TCACHE_NSLOTS_LARGE; 462 stack_nelms += tcache_bin_info[i].ncached_max; 463 } 464 465 return (false); 466 } 467 468 bool 469 tcache_boot1(void) 470 { 471 472 if (tcache_tsd_boot() || tcache_enabled_tsd_boot()) 473 return (true); 474 475 return (false); 476 } 477