tcache.h revision ae4c7b4b4092906c641d69b4bf9fcb4a7d50790d
1/******************************************************************************/ 2#ifdef JEMALLOC_H_TYPES 3 4typedef struct tcache_bin_info_s tcache_bin_info_t; 5typedef struct tcache_bin_s tcache_bin_t; 6typedef struct tcache_s tcache_t; 7 8/* 9 * tcache pointers close to NULL are used to encode state information that is 10 * used for two purposes: preventing thread caching on a per thread basis and 11 * cleaning up during thread shutdown. 12 */ 13#define TCACHE_STATE_DISABLED ((tcache_t *)(uintptr_t)1) 14#define TCACHE_STATE_REINCARNATED ((tcache_t *)(uintptr_t)2) 15#define TCACHE_STATE_PURGATORY ((tcache_t *)(uintptr_t)3) 16#define TCACHE_STATE_MAX TCACHE_STATE_PURGATORY 17 18/* 19 * Absolute maximum number of cache slots for each small bin in the thread 20 * cache. This is an additional constraint beyond that imposed as: twice the 21 * number of regions per run for this size class. 22 * 23 * This constant must be an even number. 24 */ 25#define TCACHE_NSLOTS_SMALL_MAX 200 26 27/* Number of cache slots for large size classes. */ 28#define TCACHE_NSLOTS_LARGE 20 29 30/* (1U << opt_lg_tcache_max) is used to compute tcache_maxclass. */ 31#define LG_TCACHE_MAXCLASS_DEFAULT 15 32 33/* 34 * TCACHE_GC_SWEEP is the approximate number of allocation events between 35 * full GC sweeps. Integer rounding may cause the actual number to be 36 * slightly higher, since GC is performed incrementally. 37 */ 38#define TCACHE_GC_SWEEP 8192 39 40/* Number of tcache allocation/deallocation events between incremental GCs. */ 41#define TCACHE_GC_INCR \ 42 ((TCACHE_GC_SWEEP / NBINS) + ((TCACHE_GC_SWEEP / NBINS == 0) ? 0 : 1)) 43 44#endif /* JEMALLOC_H_TYPES */ 45/******************************************************************************/ 46#ifdef JEMALLOC_H_STRUCTS 47 48typedef enum { 49 tcache_enabled_false = 0, /* Enable cast to/from bool. */ 50 tcache_enabled_true = 1, 51 tcache_enabled_default = 2 52} tcache_enabled_t; 53 54/* 55 * Read-only information associated with each element of tcache_t's tbins array 56 * is stored separately, mainly to reduce memory usage. 57 */ 58struct tcache_bin_info_s { 59 unsigned ncached_max; /* Upper limit on ncached. */ 60}; 61 62struct tcache_bin_s { 63 tcache_bin_stats_t tstats; 64 int low_water; /* Min # cached since last GC. */ 65 unsigned lg_fill_div; /* Fill (ncached_max >> lg_fill_div). */ 66 unsigned ncached; /* # of cached objects. */ 67 void **avail; /* Stack of available objects. */ 68}; 69 70struct tcache_s { 71 ql_elm(tcache_t) link; /* Used for aggregating stats. */ 72 uint64_t prof_accumbytes;/* Cleared after arena_prof_accum() */ 73 arena_t *arena; /* This thread's arena. */ 74 unsigned ev_cnt; /* Event count since incremental GC. */ 75 unsigned next_gc_bin; /* Next bin to GC. */ 76 tcache_bin_t tbins[1]; /* Dynamically sized. */ 77 /* 78 * The pointer stacks associated with tbins follow as a contiguous 79 * array. During tcache initialization, the avail pointer in each 80 * element of tbins is initialized to point to the proper offset within 81 * this array. 82 */ 83}; 84 85#endif /* JEMALLOC_H_STRUCTS */ 86/******************************************************************************/ 87#ifdef JEMALLOC_H_EXTERNS 88 89extern bool opt_tcache; 90extern ssize_t opt_lg_tcache_max; 91 92extern tcache_bin_info_t *tcache_bin_info; 93 94/* 95 * Number of tcache bins. There are NBINS small-object bins, plus 0 or more 96 * large-object bins. 97 */ 98extern size_t nhbins; 99 100/* Maximum cached size class. */ 101extern size_t tcache_maxclass; 102 103void tcache_bin_flush_small(tcache_bin_t *tbin, size_t binind, unsigned rem, 104 tcache_t *tcache); 105void tcache_bin_flush_large(tcache_bin_t *tbin, size_t binind, unsigned rem, 106 tcache_t *tcache); 107void tcache_arena_associate(tcache_t *tcache, arena_t *arena); 108void tcache_arena_dissociate(tcache_t *tcache); 109tcache_t *tcache_create(arena_t *arena); 110void *tcache_alloc_small_hard(tcache_t *tcache, tcache_bin_t *tbin, 111 size_t binind); 112void tcache_destroy(tcache_t *tcache); 113void tcache_thread_cleanup(void *arg); 114void tcache_stats_merge(tcache_t *tcache, arena_t *arena); 115bool tcache_boot0(void); 116bool tcache_boot1(void); 117 118#endif /* JEMALLOC_H_EXTERNS */ 119/******************************************************************************/ 120#ifdef JEMALLOC_H_INLINES 121 122#ifndef JEMALLOC_ENABLE_INLINE 123malloc_tsd_protos(JEMALLOC_ATTR(unused), tcache, tcache_t *) 124malloc_tsd_protos(JEMALLOC_ATTR(unused), tcache_enabled, tcache_enabled_t) 125 126void tcache_event(tcache_t *tcache); 127void tcache_flush(void); 128bool tcache_enabled_get(void); 129tcache_t *tcache_get(bool create); 130void tcache_enabled_set(bool enabled); 131void *tcache_alloc_easy(tcache_bin_t *tbin); 132void *tcache_alloc_small(tcache_t *tcache, size_t size, bool zero); 133void *tcache_alloc_large(tcache_t *tcache, size_t size, bool zero); 134void tcache_dalloc_small(tcache_t *tcache, void *ptr); 135void tcache_dalloc_large(tcache_t *tcache, void *ptr, size_t size); 136#endif 137 138#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_TCACHE_C_)) 139/* Map of thread-specific caches. */ 140malloc_tsd_externs(tcache, tcache_t *) 141malloc_tsd_funcs(JEMALLOC_INLINE, tcache, tcache_t *, NULL, 142 tcache_thread_cleanup) 143/* Per thread flag that allows thread caches to be disabled. */ 144malloc_tsd_externs(tcache_enabled, tcache_enabled_t) 145malloc_tsd_funcs(JEMALLOC_INLINE, tcache_enabled, tcache_enabled_t, 146 tcache_enabled_default, malloc_tsd_no_cleanup) 147 148JEMALLOC_INLINE void 149tcache_flush(void) 150{ 151 tcache_t *tcache; 152 153 cassert(config_tcache); 154 155 tcache = *tcache_tsd_get(); 156 if ((uintptr_t)tcache <= (uintptr_t)TCACHE_STATE_MAX) 157 return; 158 tcache_destroy(tcache); 159 tcache = NULL; 160 tcache_tsd_set(&tcache); 161} 162 163JEMALLOC_INLINE bool 164tcache_enabled_get(void) 165{ 166 tcache_enabled_t tcache_enabled; 167 168 cassert(config_tcache); 169 170 tcache_enabled = *tcache_enabled_tsd_get(); 171 if (tcache_enabled == tcache_enabled_default) { 172 tcache_enabled = (tcache_enabled_t)opt_tcache; 173 tcache_enabled_tsd_set(&tcache_enabled); 174 } 175 176 return ((bool)tcache_enabled); 177} 178 179JEMALLOC_INLINE void 180tcache_enabled_set(bool enabled) 181{ 182 tcache_enabled_t tcache_enabled; 183 tcache_t *tcache; 184 185 cassert(config_tcache); 186 187 tcache_enabled = (tcache_enabled_t)enabled; 188 tcache_enabled_tsd_set(&tcache_enabled); 189 tcache = *tcache_tsd_get(); 190 if (enabled) { 191 if (tcache == TCACHE_STATE_DISABLED) { 192 tcache = NULL; 193 tcache_tsd_set(&tcache); 194 } 195 } else /* disabled */ { 196 if (tcache > TCACHE_STATE_MAX) { 197 tcache_destroy(tcache); 198 tcache = NULL; 199 } 200 if (tcache == NULL) { 201 tcache = TCACHE_STATE_DISABLED; 202 tcache_tsd_set(&tcache); 203 } 204 } 205} 206 207JEMALLOC_INLINE tcache_t * 208tcache_get(bool create) 209{ 210 tcache_t *tcache; 211 212 if (config_tcache == false) 213 return (NULL); 214 if (config_lazy_lock && isthreaded == false) 215 return (NULL); 216 217 tcache = *tcache_tsd_get(); 218 if ((uintptr_t)tcache <= (uintptr_t)TCACHE_STATE_MAX) { 219 if (tcache == TCACHE_STATE_DISABLED) 220 return (NULL); 221 if (tcache == NULL) { 222 if (create == false) { 223 /* 224 * Creating a tcache here would cause 225 * allocation as a side effect of free(). 226 * Ordinarily that would be okay since 227 * tcache_create() failure is a soft failure 228 * that doesn't propagate. However, if TLS 229 * data are freed via free() as in glibc, 230 * subtle corruption could result from setting 231 * a TLS variable after its backing memory is 232 * freed. 233 */ 234 return (NULL); 235 } 236 if (tcache_enabled_get() == false) { 237 tcache_enabled_set(false); /* Memoize. */ 238 return (NULL); 239 } 240 return (tcache_create(choose_arena())); 241 } 242 if (tcache == TCACHE_STATE_PURGATORY) { 243 /* 244 * Make a note that an allocator function was called 245 * after tcache_thread_cleanup() was called. 246 */ 247 tcache = TCACHE_STATE_REINCARNATED; 248 tcache_tsd_set(&tcache); 249 return (NULL); 250 } 251 if (tcache == TCACHE_STATE_REINCARNATED) 252 return (NULL); 253 not_reached(); 254 } 255 256 return (tcache); 257} 258 259JEMALLOC_INLINE void 260tcache_event(tcache_t *tcache) 261{ 262 263 if (TCACHE_GC_INCR == 0) 264 return; 265 266 tcache->ev_cnt++; 267 assert(tcache->ev_cnt <= TCACHE_GC_INCR); 268 if (tcache->ev_cnt == TCACHE_GC_INCR) { 269 size_t binind = tcache->next_gc_bin; 270 tcache_bin_t *tbin = &tcache->tbins[binind]; 271 tcache_bin_info_t *tbin_info = &tcache_bin_info[binind]; 272 273 if (tbin->low_water > 0) { 274 /* 275 * Flush (ceiling) 3/4 of the objects below the low 276 * water mark. 277 */ 278 if (binind < NBINS) { 279 tcache_bin_flush_small(tbin, binind, 280 tbin->ncached - tbin->low_water + 281 (tbin->low_water >> 2), tcache); 282 } else { 283 tcache_bin_flush_large(tbin, binind, 284 tbin->ncached - tbin->low_water + 285 (tbin->low_water >> 2), tcache); 286 } 287 /* 288 * Reduce fill count by 2X. Limit lg_fill_div such that 289 * the fill count is always at least 1. 290 */ 291 if ((tbin_info->ncached_max >> (tbin->lg_fill_div+1)) 292 >= 1) 293 tbin->lg_fill_div++; 294 } else if (tbin->low_water < 0) { 295 /* 296 * Increase fill count by 2X. Make sure lg_fill_div 297 * stays greater than 0. 298 */ 299 if (tbin->lg_fill_div > 1) 300 tbin->lg_fill_div--; 301 } 302 tbin->low_water = tbin->ncached; 303 304 tcache->next_gc_bin++; 305 if (tcache->next_gc_bin == nhbins) 306 tcache->next_gc_bin = 0; 307 tcache->ev_cnt = 0; 308 } 309} 310 311JEMALLOC_INLINE void * 312tcache_alloc_easy(tcache_bin_t *tbin) 313{ 314 void *ret; 315 316 if (tbin->ncached == 0) { 317 tbin->low_water = -1; 318 return (NULL); 319 } 320 tbin->ncached--; 321 if ((int)tbin->ncached < tbin->low_water) 322 tbin->low_water = tbin->ncached; 323 ret = tbin->avail[tbin->ncached]; 324 return (ret); 325} 326 327JEMALLOC_INLINE void * 328tcache_alloc_small(tcache_t *tcache, size_t size, bool zero) 329{ 330 void *ret; 331 size_t binind; 332 tcache_bin_t *tbin; 333 334 binind = SMALL_SIZE2BIN(size); 335 assert(binind < NBINS); 336 tbin = &tcache->tbins[binind]; 337 ret = tcache_alloc_easy(tbin); 338 if (ret == NULL) { 339 ret = tcache_alloc_small_hard(tcache, tbin, binind); 340 if (ret == NULL) 341 return (NULL); 342 } 343 assert(arena_salloc(ret) == arena_bin_info[binind].reg_size); 344 345 if (zero == false) { 346 if (config_fill) { 347 if (opt_junk) 348 memset(ret, 0xa5, size); 349 else if (opt_zero) 350 memset(ret, 0, size); 351 } 352 } else 353 memset(ret, 0, size); 354 355 if (config_stats) 356 tbin->tstats.nrequests++; 357 if (config_prof) 358 tcache->prof_accumbytes += arena_bin_info[binind].reg_size; 359 tcache_event(tcache); 360 return (ret); 361} 362 363JEMALLOC_INLINE void * 364tcache_alloc_large(tcache_t *tcache, size_t size, bool zero) 365{ 366 void *ret; 367 size_t binind; 368 tcache_bin_t *tbin; 369 370 size = PAGE_CEILING(size); 371 assert(size <= tcache_maxclass); 372 binind = NBINS + (size >> LG_PAGE) - 1; 373 assert(binind < nhbins); 374 tbin = &tcache->tbins[binind]; 375 ret = tcache_alloc_easy(tbin); 376 if (ret == NULL) { 377 /* 378 * Only allocate one large object at a time, because it's quite 379 * expensive to create one and not use it. 380 */ 381 ret = arena_malloc_large(tcache->arena, size, zero); 382 if (ret == NULL) 383 return (NULL); 384 } else { 385 if (config_prof) { 386 arena_chunk_t *chunk = 387 (arena_chunk_t *)CHUNK_ADDR2BASE(ret); 388 size_t pageind = (((uintptr_t)ret - (uintptr_t)chunk) >> 389 LG_PAGE); 390 chunk->map[pageind-map_bias].bits &= 391 ~CHUNK_MAP_CLASS_MASK; 392 } 393 if (zero == false) { 394 if (config_fill) { 395 if (opt_junk) 396 memset(ret, 0xa5, size); 397 else if (opt_zero) 398 memset(ret, 0, size); 399 } 400 } else 401 memset(ret, 0, size); 402 403 if (config_stats) 404 tbin->tstats.nrequests++; 405 if (config_prof) 406 tcache->prof_accumbytes += size; 407 } 408 409 tcache_event(tcache); 410 return (ret); 411} 412 413JEMALLOC_INLINE void 414tcache_dalloc_small(tcache_t *tcache, void *ptr) 415{ 416 arena_t *arena; 417 arena_chunk_t *chunk; 418 arena_run_t *run; 419 arena_bin_t *bin; 420 tcache_bin_t *tbin; 421 tcache_bin_info_t *tbin_info; 422 size_t pageind, binind; 423 arena_chunk_map_t *mapelm; 424 425 assert(arena_salloc(ptr) <= SMALL_MAXCLASS); 426 427 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 428 arena = chunk->arena; 429 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 430 mapelm = &chunk->map[pageind-map_bias]; 431 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind - 432 (mapelm->bits >> LG_PAGE)) << LG_PAGE)); 433 bin = run->bin; 434 binind = ((uintptr_t)bin - (uintptr_t)&arena->bins) / 435 sizeof(arena_bin_t); 436 assert(binind < NBINS); 437 438 if (config_fill && opt_junk) 439 memset(ptr, 0x5a, arena_bin_info[binind].reg_size); 440 441 tbin = &tcache->tbins[binind]; 442 tbin_info = &tcache_bin_info[binind]; 443 if (tbin->ncached == tbin_info->ncached_max) { 444 tcache_bin_flush_small(tbin, binind, (tbin_info->ncached_max >> 445 1), tcache); 446 } 447 assert(tbin->ncached < tbin_info->ncached_max); 448 tbin->avail[tbin->ncached] = ptr; 449 tbin->ncached++; 450 451 tcache_event(tcache); 452} 453 454JEMALLOC_INLINE void 455tcache_dalloc_large(tcache_t *tcache, void *ptr, size_t size) 456{ 457 size_t binind; 458 tcache_bin_t *tbin; 459 tcache_bin_info_t *tbin_info; 460 461 assert((size & PAGE_MASK) == 0); 462 assert(arena_salloc(ptr) > SMALL_MAXCLASS); 463 assert(arena_salloc(ptr) <= tcache_maxclass); 464 465 binind = NBINS + (size >> LG_PAGE) - 1; 466 467 if (config_fill && opt_junk) 468 memset(ptr, 0x5a, size); 469 470 tbin = &tcache->tbins[binind]; 471 tbin_info = &tcache_bin_info[binind]; 472 if (tbin->ncached == tbin_info->ncached_max) { 473 tcache_bin_flush_large(tbin, binind, (tbin_info->ncached_max >> 474 1), tcache); 475 } 476 assert(tbin->ncached < tbin_info->ncached_max); 477 tbin->avail[tbin->ncached] = ptr; 478 tbin->ncached++; 479 480 tcache_event(tcache); 481} 482#endif 483 484#endif /* JEMALLOC_H_INLINES */ 485/******************************************************************************/ 486