arena.h revision aa5113b1fdafd1129c22512837c6c3d66c295fc8
1/******************************************************************************/ 2#ifdef JEMALLOC_H_TYPES 3 4/* 5 * RUN_MAX_OVRHD indicates maximum desired run header overhead. Runs are sized 6 * as small as possible such that this setting is still honored, without 7 * violating other constraints. The goal is to make runs as small as possible 8 * without exceeding a per run external fragmentation threshold. 9 * 10 * We use binary fixed point math for overhead computations, where the binary 11 * point is implicitly RUN_BFP bits to the left. 12 * 13 * Note that it is possible to set RUN_MAX_OVRHD low enough that it cannot be 14 * honored for some/all object sizes, since when heap profiling is enabled 15 * there is one pointer of header overhead per object (plus a constant). This 16 * constraint is relaxed (ignored) for runs that are so small that the 17 * per-region overhead is greater than: 18 * 19 * (RUN_MAX_OVRHD / (reg_interval << (3+RUN_BFP)) 20 */ 21#define RUN_BFP 12 22/* \/ Implicit binary fixed point. */ 23#define RUN_MAX_OVRHD 0x0000003dU 24#define RUN_MAX_OVRHD_RELAX 0x00001800U 25 26/* Maximum number of regions in one run. */ 27#define LG_RUN_MAXREGS 11 28#define RUN_MAXREGS (1U << LG_RUN_MAXREGS) 29 30/* 31 * Minimum redzone size. Redzones may be larger than this if necessary to 32 * preserve region alignment. 33 */ 34#define REDZONE_MINSIZE 16 35 36/* 37 * The minimum ratio of active:dirty pages per arena is computed as: 38 * 39 * (nactive >> opt_lg_dirty_mult) >= ndirty 40 * 41 * So, supposing that opt_lg_dirty_mult is 3, there can be no less than 8 times 42 * as many active pages as dirty pages. 43 */ 44#define LG_DIRTY_MULT_DEFAULT 3 45 46typedef struct arena_chunk_map_s arena_chunk_map_t; 47typedef struct arena_chunk_s arena_chunk_t; 48typedef struct arena_run_s arena_run_t; 49typedef struct arena_bin_info_s arena_bin_info_t; 50typedef struct arena_bin_s arena_bin_t; 51typedef struct arena_s arena_t; 52 53#endif /* JEMALLOC_H_TYPES */ 54/******************************************************************************/ 55#ifdef JEMALLOC_H_STRUCTS 56 57/* Each element of the chunk map corresponds to one page within the chunk. */ 58struct arena_chunk_map_s { 59#ifndef JEMALLOC_PROF 60 /* 61 * Overlay prof_ctx in order to allow it to be referenced by dead code. 62 * Such antics aren't warranted for per arena data structures, but 63 * chunk map overhead accounts for a percentage of memory, rather than 64 * being just a fixed cost. 65 */ 66 union { 67#endif 68 union { 69 /* 70 * Linkage for run trees. There are two disjoint uses: 71 * 72 * 1) arena_t's runs_avail tree. 73 * 2) arena_run_t conceptually uses this linkage for in-use 74 * non-full runs, rather than directly embedding linkage. 75 */ 76 rb_node(arena_chunk_map_t) rb_link; 77 /* 78 * List of runs currently in purgatory. arena_chunk_purge() 79 * temporarily allocates runs that contain dirty pages while 80 * purging, so that other threads cannot use the runs while the 81 * purging thread is operating without the arena lock held. 82 */ 83 ql_elm(arena_chunk_map_t) ql_link; 84 } u; 85 86 /* Profile counters, used for large object runs. */ 87 prof_ctx_t *prof_ctx; 88#ifndef JEMALLOC_PROF 89 }; /* union { ... }; */ 90#endif 91 92 /* 93 * Run address (or size) and various flags are stored together. The bit 94 * layout looks like (assuming 32-bit system): 95 * 96 * ???????? ???????? ????nnnn nnnndula 97 * 98 * ? : Unallocated: Run address for first/last pages, unset for internal 99 * pages. 100 * Small: Run page offset. 101 * Large: Run size for first page, unset for trailing pages. 102 * n : binind for small size class, BININD_INVALID for large size class. 103 * d : dirty? 104 * u : unzeroed? 105 * l : large? 106 * a : allocated? 107 * 108 * Following are example bit patterns for the three types of runs. 109 * 110 * p : run page offset 111 * s : run size 112 * n : binind for size class; large objects set these to BININD_INVALID 113 * except for promoted allocations (see prof_promote) 114 * x : don't care 115 * - : 0 116 * + : 1 117 * [DULA] : bit set 118 * [dula] : bit unset 119 * 120 * Unallocated (clean): 121 * ssssssss ssssssss ssss++++ ++++du-a 122 * xxxxxxxx xxxxxxxx xxxxxxxx xxxx-Uxx 123 * ssssssss ssssssss ssss++++ ++++dU-a 124 * 125 * Unallocated (dirty): 126 * ssssssss ssssssss ssss++++ ++++D--a 127 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx 128 * ssssssss ssssssss ssss++++ ++++D--a 129 * 130 * Small: 131 * pppppppp pppppppp ppppnnnn nnnnd--A 132 * pppppppp pppppppp ppppnnnn nnnn---A 133 * pppppppp pppppppp ppppnnnn nnnnd--A 134 * 135 * Large: 136 * ssssssss ssssssss ssss++++ ++++D-LA 137 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx 138 * -------- -------- ----++++ ++++D-LA 139 * 140 * Large (sampled, size <= PAGE): 141 * ssssssss ssssssss ssssnnnn nnnnD-LA 142 * 143 * Large (not sampled, size == PAGE): 144 * ssssssss ssssssss ssss++++ ++++D-LA 145 */ 146 size_t bits; 147#define CHUNK_MAP_BININD_SHIFT 4 148#define BININD_INVALID ((size_t)0xffU) 149/* CHUNK_MAP_BININD_MASK == (BININD_INVALID << CHUNK_MAP_BININD_SHIFT) */ 150#define CHUNK_MAP_BININD_MASK ((size_t)0xff0U) 151#define CHUNK_MAP_BININD_INVALID CHUNK_MAP_BININD_MASK 152#define CHUNK_MAP_FLAGS_MASK ((size_t)0xcU) 153#define CHUNK_MAP_DIRTY ((size_t)0x8U) 154#define CHUNK_MAP_UNZEROED ((size_t)0x4U) 155#define CHUNK_MAP_LARGE ((size_t)0x2U) 156#define CHUNK_MAP_ALLOCATED ((size_t)0x1U) 157#define CHUNK_MAP_KEY CHUNK_MAP_ALLOCATED 158}; 159typedef rb_tree(arena_chunk_map_t) arena_avail_tree_t; 160typedef rb_tree(arena_chunk_map_t) arena_run_tree_t; 161typedef ql_head(arena_chunk_map_t) arena_chunk_mapelms_t; 162 163/* Arena chunk header. */ 164struct arena_chunk_s { 165 /* Arena that owns the chunk. */ 166 arena_t *arena; 167 168 /* Linkage for tree of arena chunks that contain dirty runs. */ 169 rb_node(arena_chunk_t) dirty_link; 170 171 /* Number of dirty pages. */ 172 size_t ndirty; 173 174 /* Number of available runs. */ 175 size_t nruns_avail; 176 177 /* 178 * Number of available run adjacencies that purging could coalesce. 179 * Clean and dirty available runs are not coalesced, which causes 180 * virtual memory fragmentation. The ratio of 181 * (nruns_avail-nruns_adjac):nruns_adjac is used for tracking this 182 * fragmentation. 183 */ 184 size_t nruns_adjac; 185 186 /* 187 * Map of pages within chunk that keeps track of free/large/small. The 188 * first map_bias entries are omitted, since the chunk header does not 189 * need to be tracked in the map. This omission saves a header page 190 * for common chunk sizes (e.g. 4 MiB). 191 */ 192 arena_chunk_map_t map[1]; /* Dynamically sized. */ 193}; 194typedef rb_tree(arena_chunk_t) arena_chunk_tree_t; 195 196struct arena_run_s { 197 /* Bin this run is associated with. */ 198 arena_bin_t *bin; 199 200 /* Index of next region that has never been allocated, or nregs. */ 201 uint32_t nextind; 202 203 /* Number of free regions in run. */ 204 unsigned nfree; 205}; 206 207/* 208 * Read-only information associated with each element of arena_t's bins array 209 * is stored separately, partly to reduce memory usage (only one copy, rather 210 * than one per arena), but mainly to avoid false cacheline sharing. 211 * 212 * Each run has the following layout: 213 * 214 * /--------------------\ 215 * | arena_run_t header | 216 * | ... | 217 * bitmap_offset | bitmap | 218 * | ... | 219 * ctx0_offset | ctx map | 220 * | ... | 221 * |--------------------| 222 * | redzone | 223 * reg0_offset | region 0 | 224 * | redzone | 225 * |--------------------| \ 226 * | redzone | | 227 * | region 1 | > reg_interval 228 * | redzone | / 229 * |--------------------| 230 * | ... | 231 * | ... | 232 * | ... | 233 * |--------------------| 234 * | redzone | 235 * | region nregs-1 | 236 * | redzone | 237 * |--------------------| 238 * | alignment pad? | 239 * \--------------------/ 240 * 241 * reg_interval has at least the same minimum alignment as reg_size; this 242 * preserves the alignment constraint that sa2u() depends on. Alignment pad is 243 * either 0 or redzone_size; it is present only if needed to align reg0_offset. 244 */ 245struct arena_bin_info_s { 246 /* Size of regions in a run for this bin's size class. */ 247 size_t reg_size; 248 249 /* Redzone size. */ 250 size_t redzone_size; 251 252 /* Interval between regions (reg_size + (redzone_size << 1)). */ 253 size_t reg_interval; 254 255 /* Total size of a run for this bin's size class. */ 256 size_t run_size; 257 258 /* Total number of regions in a run for this bin's size class. */ 259 uint32_t nregs; 260 261 /* 262 * Offset of first bitmap_t element in a run header for this bin's size 263 * class. 264 */ 265 uint32_t bitmap_offset; 266 267 /* 268 * Metadata used to manipulate bitmaps for runs associated with this 269 * bin. 270 */ 271 bitmap_info_t bitmap_info; 272 273 /* 274 * Offset of first (prof_ctx_t *) in a run header for this bin's size 275 * class, or 0 if (config_prof == false || opt_prof == false). 276 */ 277 uint32_t ctx0_offset; 278 279 /* Offset of first region in a run for this bin's size class. */ 280 uint32_t reg0_offset; 281}; 282 283struct arena_bin_s { 284 /* 285 * All operations on runcur, runs, and stats require that lock be 286 * locked. Run allocation/deallocation are protected by the arena lock, 287 * which may be acquired while holding one or more bin locks, but not 288 * vise versa. 289 */ 290 malloc_mutex_t lock; 291 292 /* 293 * Current run being used to service allocations of this bin's size 294 * class. 295 */ 296 arena_run_t *runcur; 297 298 /* 299 * Tree of non-full runs. This tree is used when looking for an 300 * existing run when runcur is no longer usable. We choose the 301 * non-full run that is lowest in memory; this policy tends to keep 302 * objects packed well, and it can also help reduce the number of 303 * almost-empty chunks. 304 */ 305 arena_run_tree_t runs; 306 307 /* Bin statistics. */ 308 malloc_bin_stats_t stats; 309}; 310 311struct arena_s { 312 /* This arena's index within the arenas array. */ 313 unsigned ind; 314 315 /* 316 * Number of threads currently assigned to this arena. This field is 317 * protected by arenas_lock. 318 */ 319 unsigned nthreads; 320 321 /* 322 * There are three classes of arena operations from a locking 323 * perspective: 324 * 1) Thread asssignment (modifies nthreads) is protected by 325 * arenas_lock. 326 * 2) Bin-related operations are protected by bin locks. 327 * 3) Chunk- and run-related operations are protected by this mutex. 328 */ 329 malloc_mutex_t lock; 330 331 arena_stats_t stats; 332 /* 333 * List of tcaches for extant threads associated with this arena. 334 * Stats from these are merged incrementally, and at exit. 335 */ 336 ql_head(tcache_t) tcache_ql; 337 338 uint64_t prof_accumbytes; 339 340 dss_prec_t dss_prec; 341 342 /* Tree of dirty-page-containing chunks this arena manages. */ 343 arena_chunk_tree_t chunks_dirty; 344 345 /* 346 * In order to avoid rapid chunk allocation/deallocation when an arena 347 * oscillates right on the cusp of needing a new chunk, cache the most 348 * recently freed chunk. The spare is left in the arena's chunk trees 349 * until it is deleted. 350 * 351 * There is one spare chunk per arena, rather than one spare total, in 352 * order to avoid interactions between multiple threads that could make 353 * a single spare inadequate. 354 */ 355 arena_chunk_t *spare; 356 357 /* Number of pages in active runs. */ 358 size_t nactive; 359 360 /* 361 * Current count of pages within unused runs that are potentially 362 * dirty, and for which madvise(... MADV_DONTNEED) has not been called. 363 * By tracking this, we can institute a limit on how much dirty unused 364 * memory is mapped for each arena. 365 */ 366 size_t ndirty; 367 368 /* 369 * Approximate number of pages being purged. It is possible for 370 * multiple threads to purge dirty pages concurrently, and they use 371 * npurgatory to indicate the total number of pages all threads are 372 * attempting to purge. 373 */ 374 size_t npurgatory; 375 376 /* 377 * Size/address-ordered trees of this arena's available runs. The trees 378 * are used for first-best-fit run allocation. 379 */ 380 arena_avail_tree_t runs_avail; 381 382 /* bins is used to store trees of free regions. */ 383 arena_bin_t bins[NBINS]; 384}; 385 386#endif /* JEMALLOC_H_STRUCTS */ 387/******************************************************************************/ 388#ifdef JEMALLOC_H_EXTERNS 389 390extern ssize_t opt_lg_dirty_mult; 391/* 392 * small_size2bin is a compact lookup table that rounds request sizes up to 393 * size classes. In order to reduce cache footprint, the table is compressed, 394 * and all accesses are via the SMALL_SIZE2BIN macro. 395 */ 396extern uint8_t const small_size2bin[]; 397#define SMALL_SIZE2BIN(s) (small_size2bin[(s-1) >> LG_TINY_MIN]) 398 399extern arena_bin_info_t arena_bin_info[NBINS]; 400 401/* Number of large size classes. */ 402#define nlclasses (chunk_npages - map_bias) 403 404void arena_purge_all(arena_t *arena); 405void arena_tcache_fill_small(arena_t *arena, tcache_bin_t *tbin, 406 size_t binind, uint64_t prof_accumbytes); 407void arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, 408 bool zero); 409#ifdef JEMALLOC_JET 410typedef void (arena_redzone_corruption_t)(void *, size_t, bool, size_t, 411 uint8_t); 412extern arena_redzone_corruption_t *arena_redzone_corruption; 413typedef void (arena_dalloc_junk_small_t)(void *, arena_bin_info_t *); 414extern arena_dalloc_junk_small_t *arena_dalloc_junk_small; 415#else 416void arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info); 417#endif 418void arena_quarantine_junk_small(void *ptr, size_t usize); 419void *arena_malloc_small(arena_t *arena, size_t size, bool zero); 420void *arena_malloc_large(arena_t *arena, size_t size, bool zero); 421void *arena_palloc(arena_t *arena, size_t size, size_t alignment, bool zero); 422void arena_prof_promoted(const void *ptr, size_t size); 423void arena_dalloc_bin_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr, 424 arena_chunk_map_t *mapelm); 425void arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr, 426 size_t pageind, arena_chunk_map_t *mapelm); 427void arena_dalloc_small(arena_t *arena, arena_chunk_t *chunk, void *ptr, 428 size_t pageind); 429#ifdef JEMALLOC_JET 430typedef void (arena_dalloc_junk_large_t)(void *, size_t); 431extern arena_dalloc_junk_large_t *arena_dalloc_junk_large; 432#endif 433void arena_dalloc_large_locked(arena_t *arena, arena_chunk_t *chunk, 434 void *ptr); 435void arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr); 436#ifdef JEMALLOC_JET 437typedef void (arena_ralloc_junk_large_t)(void *, size_t, size_t); 438extern arena_ralloc_junk_large_t *arena_ralloc_junk_large; 439#endif 440bool arena_ralloc_no_move(void *ptr, size_t oldsize, size_t size, 441 size_t extra, bool zero); 442void *arena_ralloc(arena_t *arena, void *ptr, size_t oldsize, size_t size, 443 size_t extra, size_t alignment, bool zero, bool try_tcache_alloc, 444 bool try_tcache_dalloc); 445dss_prec_t arena_dss_prec_get(arena_t *arena); 446void arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec); 447void arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive, 448 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats, 449 malloc_large_stats_t *lstats); 450bool arena_new(arena_t *arena, unsigned ind); 451void arena_boot(void); 452void arena_prefork(arena_t *arena); 453void arena_postfork_parent(arena_t *arena); 454void arena_postfork_child(arena_t *arena); 455 456#endif /* JEMALLOC_H_EXTERNS */ 457/******************************************************************************/ 458#ifdef JEMALLOC_H_INLINES 459 460#ifndef JEMALLOC_ENABLE_INLINE 461arena_chunk_map_t *arena_mapp_get(arena_chunk_t *chunk, size_t pageind); 462size_t *arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind); 463size_t arena_mapbitsp_read(size_t *mapbitsp); 464size_t arena_mapbits_get(arena_chunk_t *chunk, size_t pageind); 465size_t arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, 466 size_t pageind); 467size_t arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind); 468size_t arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind); 469size_t arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind); 470size_t arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind); 471size_t arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind); 472size_t arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind); 473size_t arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind); 474void arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits); 475void arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, 476 size_t size, size_t flags); 477void arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 478 size_t size); 479void arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, 480 size_t size, size_t flags); 481void arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 482 size_t binind); 483void arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, 484 size_t runind, size_t binind, size_t flags); 485void arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 486 size_t unzeroed); 487bool arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes); 488bool arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes); 489bool arena_prof_accum(arena_t *arena, uint64_t accumbytes); 490size_t arena_ptr_small_binind_get(const void *ptr, size_t mapbits); 491size_t arena_bin_index(arena_t *arena, arena_bin_t *bin); 492unsigned arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, 493 const void *ptr); 494prof_ctx_t *arena_prof_ctx_get(const void *ptr); 495void arena_prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx); 496void *arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache); 497size_t arena_salloc(const void *ptr, bool demote); 498void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, 499 bool try_tcache); 500#endif 501 502#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_)) 503# ifdef JEMALLOC_ARENA_INLINE_A 504JEMALLOC_ALWAYS_INLINE arena_chunk_map_t * 505arena_mapp_get(arena_chunk_t *chunk, size_t pageind) 506{ 507 508 assert(pageind >= map_bias); 509 assert(pageind < chunk_npages); 510 511 return (&chunk->map[pageind-map_bias]); 512} 513 514JEMALLOC_ALWAYS_INLINE size_t * 515arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind) 516{ 517 518 return (&arena_mapp_get(chunk, pageind)->bits); 519} 520 521JEMALLOC_ALWAYS_INLINE size_t 522arena_mapbitsp_read(size_t *mapbitsp) 523{ 524 525 return (*mapbitsp); 526} 527 528JEMALLOC_ALWAYS_INLINE size_t 529arena_mapbits_get(arena_chunk_t *chunk, size_t pageind) 530{ 531 532 return (arena_mapbitsp_read(arena_mapbitsp_get(chunk, pageind))); 533} 534 535JEMALLOC_ALWAYS_INLINE size_t 536arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, size_t pageind) 537{ 538 size_t mapbits; 539 540 mapbits = arena_mapbits_get(chunk, pageind); 541 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 542 return (mapbits & ~PAGE_MASK); 543} 544 545JEMALLOC_ALWAYS_INLINE size_t 546arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind) 547{ 548 size_t mapbits; 549 550 mapbits = arena_mapbits_get(chunk, pageind); 551 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 552 (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)); 553 return (mapbits & ~PAGE_MASK); 554} 555 556JEMALLOC_ALWAYS_INLINE size_t 557arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind) 558{ 559 size_t mapbits; 560 561 mapbits = arena_mapbits_get(chunk, pageind); 562 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 563 CHUNK_MAP_ALLOCATED); 564 return (mapbits >> LG_PAGE); 565} 566 567JEMALLOC_ALWAYS_INLINE size_t 568arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind) 569{ 570 size_t mapbits; 571 size_t binind; 572 573 mapbits = arena_mapbits_get(chunk, pageind); 574 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 575 assert(binind < NBINS || binind == BININD_INVALID); 576 return (binind); 577} 578 579JEMALLOC_ALWAYS_INLINE size_t 580arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind) 581{ 582 size_t mapbits; 583 584 mapbits = arena_mapbits_get(chunk, pageind); 585 return (mapbits & CHUNK_MAP_DIRTY); 586} 587 588JEMALLOC_ALWAYS_INLINE size_t 589arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind) 590{ 591 size_t mapbits; 592 593 mapbits = arena_mapbits_get(chunk, pageind); 594 return (mapbits & CHUNK_MAP_UNZEROED); 595} 596 597JEMALLOC_ALWAYS_INLINE size_t 598arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind) 599{ 600 size_t mapbits; 601 602 mapbits = arena_mapbits_get(chunk, pageind); 603 return (mapbits & CHUNK_MAP_LARGE); 604} 605 606JEMALLOC_ALWAYS_INLINE size_t 607arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind) 608{ 609 size_t mapbits; 610 611 mapbits = arena_mapbits_get(chunk, pageind); 612 return (mapbits & CHUNK_MAP_ALLOCATED); 613} 614 615JEMALLOC_ALWAYS_INLINE void 616arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits) 617{ 618 619 *mapbitsp = mapbits; 620} 621 622JEMALLOC_ALWAYS_INLINE void 623arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size, 624 size_t flags) 625{ 626 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 627 628 assert((size & PAGE_MASK) == 0); 629 assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0); 630 assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags); 631 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags); 632} 633 634JEMALLOC_ALWAYS_INLINE void 635arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 636 size_t size) 637{ 638 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 639 size_t mapbits = arena_mapbitsp_read(mapbitsp); 640 641 assert((size & PAGE_MASK) == 0); 642 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 643 arena_mapbitsp_write(mapbitsp, size | (mapbits & PAGE_MASK)); 644} 645 646JEMALLOC_ALWAYS_INLINE void 647arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size, 648 size_t flags) 649{ 650 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 651 size_t mapbits = arena_mapbitsp_read(mapbitsp); 652 size_t unzeroed; 653 654 assert((size & PAGE_MASK) == 0); 655 assert((flags & CHUNK_MAP_DIRTY) == flags); 656 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 657 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags 658 | unzeroed | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED); 659} 660 661JEMALLOC_ALWAYS_INLINE void 662arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 663 size_t binind) 664{ 665 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 666 size_t mapbits = arena_mapbitsp_read(mapbitsp); 667 668 assert(binind <= BININD_INVALID); 669 assert(arena_mapbits_large_size_get(chunk, pageind) == PAGE); 670 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_BININD_MASK) | 671 (binind << CHUNK_MAP_BININD_SHIFT)); 672} 673 674JEMALLOC_ALWAYS_INLINE void 675arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, size_t runind, 676 size_t binind, size_t flags) 677{ 678 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 679 size_t mapbits = arena_mapbitsp_read(mapbitsp); 680 size_t unzeroed; 681 682 assert(binind < BININD_INVALID); 683 assert(pageind - runind >= map_bias); 684 assert((flags & CHUNK_MAP_DIRTY) == flags); 685 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 686 arena_mapbitsp_write(mapbitsp, (runind << LG_PAGE) | (binind << 687 CHUNK_MAP_BININD_SHIFT) | flags | unzeroed | CHUNK_MAP_ALLOCATED); 688} 689 690JEMALLOC_ALWAYS_INLINE void 691arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 692 size_t unzeroed) 693{ 694 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 695 size_t mapbits = arena_mapbitsp_read(mapbitsp); 696 697 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_UNZEROED) | 698 unzeroed); 699} 700 701JEMALLOC_INLINE bool 702arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes) 703{ 704 705 cassert(config_prof); 706 assert(prof_interval != 0); 707 708 arena->prof_accumbytes += accumbytes; 709 if (arena->prof_accumbytes >= prof_interval) { 710 arena->prof_accumbytes -= prof_interval; 711 return (true); 712 } 713 return (false); 714} 715 716JEMALLOC_INLINE bool 717arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes) 718{ 719 720 cassert(config_prof); 721 722 if (prof_interval == 0) 723 return (false); 724 return (arena_prof_accum_impl(arena, accumbytes)); 725} 726 727JEMALLOC_INLINE bool 728arena_prof_accum(arena_t *arena, uint64_t accumbytes) 729{ 730 731 cassert(config_prof); 732 733 if (prof_interval == 0) 734 return (false); 735 736 { 737 bool ret; 738 739 malloc_mutex_lock(&arena->lock); 740 ret = arena_prof_accum_impl(arena, accumbytes); 741 malloc_mutex_unlock(&arena->lock); 742 return (ret); 743 } 744} 745 746JEMALLOC_ALWAYS_INLINE size_t 747arena_ptr_small_binind_get(const void *ptr, size_t mapbits) 748{ 749 size_t binind; 750 751 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 752 753 if (config_debug) { 754 arena_chunk_t *chunk; 755 arena_t *arena; 756 size_t pageind; 757 size_t actual_mapbits; 758 arena_run_t *run; 759 arena_bin_t *bin; 760 size_t actual_binind; 761 arena_bin_info_t *bin_info; 762 763 assert(binind != BININD_INVALID); 764 assert(binind < NBINS); 765 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 766 arena = chunk->arena; 767 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 768 actual_mapbits = arena_mapbits_get(chunk, pageind); 769 assert(mapbits == actual_mapbits); 770 assert(arena_mapbits_large_get(chunk, pageind) == 0); 771 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 772 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind - 773 (actual_mapbits >> LG_PAGE)) << LG_PAGE)); 774 bin = run->bin; 775 actual_binind = bin - arena->bins; 776 assert(binind == actual_binind); 777 bin_info = &arena_bin_info[actual_binind]; 778 assert(((uintptr_t)ptr - ((uintptr_t)run + 779 (uintptr_t)bin_info->reg0_offset)) % bin_info->reg_interval 780 == 0); 781 } 782 783 return (binind); 784} 785# endif /* JEMALLOC_ARENA_INLINE_A */ 786 787# ifdef JEMALLOC_ARENA_INLINE_B 788JEMALLOC_INLINE size_t 789arena_bin_index(arena_t *arena, arena_bin_t *bin) 790{ 791 size_t binind = bin - arena->bins; 792 assert(binind < NBINS); 793 return (binind); 794} 795 796JEMALLOC_INLINE unsigned 797arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, const void *ptr) 798{ 799 unsigned shift, diff, regind; 800 size_t interval; 801 802 /* 803 * Freeing a pointer lower than region zero can cause assertion 804 * failure. 805 */ 806 assert((uintptr_t)ptr >= (uintptr_t)run + 807 (uintptr_t)bin_info->reg0_offset); 808 809 /* 810 * Avoid doing division with a variable divisor if possible. Using 811 * actual division here can reduce allocator throughput by over 20%! 812 */ 813 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - 814 bin_info->reg0_offset); 815 816 /* Rescale (factor powers of 2 out of the numerator and denominator). */ 817 interval = bin_info->reg_interval; 818 shift = ffs(interval) - 1; 819 diff >>= shift; 820 interval >>= shift; 821 822 if (interval == 1) { 823 /* The divisor was a power of 2. */ 824 regind = diff; 825 } else { 826 /* 827 * To divide by a number D that is not a power of two we 828 * multiply by (2^21 / D) and then right shift by 21 positions. 829 * 830 * X / D 831 * 832 * becomes 833 * 834 * (X * interval_invs[D - 3]) >> SIZE_INV_SHIFT 835 * 836 * We can omit the first three elements, because we never 837 * divide by 0, and 1 and 2 are both powers of two, which are 838 * handled above. 839 */ 840#define SIZE_INV_SHIFT ((sizeof(unsigned) << 3) - LG_RUN_MAXREGS) 841#define SIZE_INV(s) (((1U << SIZE_INV_SHIFT) / (s)) + 1) 842 static const unsigned interval_invs[] = { 843 SIZE_INV(3), 844 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 845 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 846 SIZE_INV(12), SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 847 SIZE_INV(16), SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 848 SIZE_INV(20), SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 849 SIZE_INV(24), SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 850 SIZE_INV(28), SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 851 }; 852 853 if (interval <= ((sizeof(interval_invs) / sizeof(unsigned)) + 854 2)) { 855 regind = (diff * interval_invs[interval - 3]) >> 856 SIZE_INV_SHIFT; 857 } else 858 regind = diff / interval; 859#undef SIZE_INV 860#undef SIZE_INV_SHIFT 861 } 862 assert(diff == regind * interval); 863 assert(regind < bin_info->nregs); 864 865 return (regind); 866} 867 868JEMALLOC_INLINE prof_ctx_t * 869arena_prof_ctx_get(const void *ptr) 870{ 871 prof_ctx_t *ret; 872 arena_chunk_t *chunk; 873 size_t pageind, mapbits; 874 875 cassert(config_prof); 876 assert(ptr != NULL); 877 assert(CHUNK_ADDR2BASE(ptr) != ptr); 878 879 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 880 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 881 mapbits = arena_mapbits_get(chunk, pageind); 882 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 883 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 884 if (prof_promote) 885 ret = (prof_ctx_t *)(uintptr_t)1U; 886 else { 887 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 888 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 889 LG_PAGE)); 890 size_t binind = arena_ptr_small_binind_get(ptr, 891 mapbits); 892 arena_bin_info_t *bin_info = &arena_bin_info[binind]; 893 unsigned regind; 894 895 regind = arena_run_regind(run, bin_info, ptr); 896 ret = *(prof_ctx_t **)((uintptr_t)run + 897 bin_info->ctx0_offset + (regind * 898 sizeof(prof_ctx_t *))); 899 } 900 } else 901 ret = arena_mapp_get(chunk, pageind)->prof_ctx; 902 903 return (ret); 904} 905 906JEMALLOC_INLINE void 907arena_prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx) 908{ 909 arena_chunk_t *chunk; 910 size_t pageind; 911 912 cassert(config_prof); 913 assert(ptr != NULL); 914 assert(CHUNK_ADDR2BASE(ptr) != ptr); 915 916 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 917 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 918 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 919 920 if (usize > SMALL_MAXCLASS || (prof_promote && 921 ((uintptr_t)ctx != (uintptr_t)1U || arena_mapbits_large_get(chunk, 922 pageind) != 0))) { 923 assert(arena_mapbits_large_get(chunk, pageind) != 0); 924 arena_mapp_get(chunk, pageind)->prof_ctx = ctx; 925 } else { 926 assert(arena_mapbits_large_get(chunk, pageind) == 0); 927 if (prof_promote == false) { 928 size_t mapbits = arena_mapbits_get(chunk, pageind); 929 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 930 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 931 LG_PAGE)); 932 size_t binind; 933 arena_bin_info_t *bin_info; 934 unsigned regind; 935 936 binind = arena_ptr_small_binind_get(ptr, mapbits); 937 bin_info = &arena_bin_info[binind]; 938 regind = arena_run_regind(run, bin_info, ptr); 939 940 *((prof_ctx_t **)((uintptr_t)run + 941 bin_info->ctx0_offset + (regind * sizeof(prof_ctx_t 942 *)))) = ctx; 943 } 944 } 945} 946 947JEMALLOC_ALWAYS_INLINE void * 948arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache) 949{ 950 tcache_t *tcache; 951 952 assert(size != 0); 953 assert(size <= arena_maxclass); 954 955 if (size <= SMALL_MAXCLASS) { 956 if (try_tcache && (tcache = tcache_get(true)) != NULL) 957 return (tcache_alloc_small(tcache, size, zero)); 958 else { 959 return (arena_malloc_small(choose_arena(arena), size, 960 zero)); 961 } 962 } else { 963 /* 964 * Initialize tcache after checking size in order to avoid 965 * infinite recursion during tcache initialization. 966 */ 967 if (try_tcache && size <= tcache_maxclass && (tcache = 968 tcache_get(true)) != NULL) 969 return (tcache_alloc_large(tcache, size, zero)); 970 else { 971 return (arena_malloc_large(choose_arena(arena), size, 972 zero)); 973 } 974 } 975} 976 977/* Return the size of the allocation pointed to by ptr. */ 978JEMALLOC_ALWAYS_INLINE size_t 979arena_salloc(const void *ptr, bool demote) 980{ 981 size_t ret; 982 arena_chunk_t *chunk; 983 size_t pageind, binind; 984 985 assert(ptr != NULL); 986 assert(CHUNK_ADDR2BASE(ptr) != ptr); 987 988 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 989 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 990 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 991 binind = arena_mapbits_binind_get(chunk, pageind); 992 if (binind == BININD_INVALID || (config_prof && demote == false && 993 prof_promote && arena_mapbits_large_get(chunk, pageind) != 0)) { 994 /* 995 * Large allocation. In the common case (demote == true), and 996 * as this is an inline function, most callers will only end up 997 * looking at binind to determine that ptr is a small 998 * allocation. 999 */ 1000 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1001 ret = arena_mapbits_large_size_get(chunk, pageind); 1002 assert(ret != 0); 1003 assert(pageind + (ret>>LG_PAGE) <= chunk_npages); 1004 assert(ret == PAGE || arena_mapbits_large_size_get(chunk, 1005 pageind+(ret>>LG_PAGE)-1) == 0); 1006 assert(binind == arena_mapbits_binind_get(chunk, 1007 pageind+(ret>>LG_PAGE)-1)); 1008 assert(arena_mapbits_dirty_get(chunk, pageind) == 1009 arena_mapbits_dirty_get(chunk, pageind+(ret>>LG_PAGE)-1)); 1010 } else { 1011 /* 1012 * Small allocation (possibly promoted to a large object due to 1013 * prof_promote). 1014 */ 1015 assert(arena_mapbits_large_get(chunk, pageind) != 0 || 1016 arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk, 1017 pageind)) == binind); 1018 ret = arena_bin_info[binind].reg_size; 1019 } 1020 1021 return (ret); 1022} 1023 1024JEMALLOC_ALWAYS_INLINE void 1025arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, bool try_tcache) 1026{ 1027 size_t pageind, mapbits; 1028 tcache_t *tcache; 1029 1030 assert(arena != NULL); 1031 assert(chunk->arena == arena); 1032 assert(ptr != NULL); 1033 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1034 1035 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1036 mapbits = arena_mapbits_get(chunk, pageind); 1037 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1038 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 1039 /* Small allocation. */ 1040 if (try_tcache && (tcache = tcache_get(false)) != NULL) { 1041 size_t binind; 1042 1043 binind = arena_ptr_small_binind_get(ptr, mapbits); 1044 tcache_dalloc_small(tcache, ptr, binind); 1045 } else 1046 arena_dalloc_small(arena, chunk, ptr, pageind); 1047 } else { 1048 size_t size = arena_mapbits_large_size_get(chunk, pageind); 1049 1050 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1051 1052 if (try_tcache && size <= tcache_maxclass && (tcache = 1053 tcache_get(false)) != NULL) { 1054 tcache_dalloc_large(tcache, ptr, size); 1055 } else 1056 arena_dalloc_large(arena, chunk, ptr); 1057 } 1058} 1059# endif /* JEMALLOC_ARENA_INLINE_B */ 1060#endif 1061 1062#endif /* JEMALLOC_H_INLINES */ 1063/******************************************************************************/ 1064