arena.h revision 87a02d2bb18dbcb2955541b849bc95862e864803
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; 161 162/* Arena chunk header. */ 163struct arena_chunk_s { 164 /* Arena that owns the chunk. */ 165 arena_t *arena; 166 167 /* Linkage for tree of arena chunks that contain dirty runs. */ 168 rb_node(arena_chunk_t) dirty_link; 169 170 /* Number of dirty pages. */ 171 size_t ndirty; 172 173 /* Number of available runs. */ 174 size_t nruns_avail; 175 176 /* 177 * Number of available run adjacencies. Clean and dirty available runs 178 * are not coalesced, which causes virtual memory fragmentation. The 179 * ratio of (nruns_avail-nruns_adjac):nruns_adjac is used for tracking 180 * this fragmentation. 181 * */ 182 size_t nruns_adjac; 183 184 /* 185 * Map of pages within chunk that keeps track of free/large/small. The 186 * first map_bias entries are omitted, since the chunk header does not 187 * need to be tracked in the map. This omission saves a header page 188 * for common chunk sizes (e.g. 4 MiB). 189 */ 190 arena_chunk_map_t map[1]; /* Dynamically sized. */ 191}; 192typedef rb_tree(arena_chunk_t) arena_chunk_tree_t; 193 194struct arena_run_s { 195 /* Bin this run is associated with. */ 196 arena_bin_t *bin; 197 198 /* Index of next region that has never been allocated, or nregs. */ 199 uint32_t nextind; 200 201 /* Number of free regions in run. */ 202 unsigned nfree; 203}; 204 205/* 206 * Read-only information associated with each element of arena_t's bins array 207 * is stored separately, partly to reduce memory usage (only one copy, rather 208 * than one per arena), but mainly to avoid false cacheline sharing. 209 * 210 * Each run has the following layout: 211 * 212 * /--------------------\ 213 * | arena_run_t header | 214 * | ... | 215 * bitmap_offset | bitmap | 216 * | ... | 217 * ctx0_offset | ctx map | 218 * | ... | 219 * |--------------------| 220 * | redzone | 221 * reg0_offset | region 0 | 222 * | redzone | 223 * |--------------------| \ 224 * | redzone | | 225 * | region 1 | > reg_interval 226 * | redzone | / 227 * |--------------------| 228 * | ... | 229 * | ... | 230 * | ... | 231 * |--------------------| 232 * | redzone | 233 * | region nregs-1 | 234 * | redzone | 235 * |--------------------| 236 * | alignment pad? | 237 * \--------------------/ 238 * 239 * reg_interval has at least the same minimum alignment as reg_size; this 240 * preserves the alignment constraint that sa2u() depends on. Alignment pad is 241 * either 0 or redzone_size; it is present only if needed to align reg0_offset. 242 */ 243struct arena_bin_info_s { 244 /* Size of regions in a run for this bin's size class. */ 245 size_t reg_size; 246 247 /* Redzone size. */ 248 size_t redzone_size; 249 250 /* Interval between regions (reg_size + (redzone_size << 1)). */ 251 size_t reg_interval; 252 253 /* Total size of a run for this bin's size class. */ 254 size_t run_size; 255 256 /* Total number of regions in a run for this bin's size class. */ 257 uint32_t nregs; 258 259 /* 260 * Offset of first bitmap_t element in a run header for this bin's size 261 * class. 262 */ 263 uint32_t bitmap_offset; 264 265 /* 266 * Metadata used to manipulate bitmaps for runs associated with this 267 * bin. 268 */ 269 bitmap_info_t bitmap_info; 270 271 /* 272 * Offset of first (prof_ctx_t *) in a run header for this bin's size 273 * class, or 0 if (config_prof == false || opt_prof == false). 274 */ 275 uint32_t ctx0_offset; 276 277 /* Offset of first region in a run for this bin's size class. */ 278 uint32_t reg0_offset; 279}; 280 281struct arena_bin_s { 282 /* 283 * All operations on runcur, runs, and stats require that lock be 284 * locked. Run allocation/deallocation are protected by the arena lock, 285 * which may be acquired while holding one or more bin locks, but not 286 * vise versa. 287 */ 288 malloc_mutex_t lock; 289 290 /* 291 * Current run being used to service allocations of this bin's size 292 * class. 293 */ 294 arena_run_t *runcur; 295 296 /* 297 * Tree of non-full runs. This tree is used when looking for an 298 * existing run when runcur is no longer usable. We choose the 299 * non-full run that is lowest in memory; this policy tends to keep 300 * objects packed well, and it can also help reduce the number of 301 * almost-empty chunks. 302 */ 303 arena_run_tree_t runs; 304 305 /* Bin statistics. */ 306 malloc_bin_stats_t stats; 307}; 308 309struct arena_s { 310 /* This arena's index within the arenas array. */ 311 unsigned ind; 312 313 /* 314 * Number of threads currently assigned to this arena. This field is 315 * protected by arenas_lock. 316 */ 317 unsigned nthreads; 318 319 /* 320 * There are three classes of arena operations from a locking 321 * perspective: 322 * 1) Thread asssignment (modifies nthreads) is protected by 323 * arenas_lock. 324 * 2) Bin-related operations are protected by bin locks. 325 * 3) Chunk- and run-related operations are protected by this mutex. 326 */ 327 malloc_mutex_t lock; 328 329 arena_stats_t stats; 330 /* 331 * List of tcaches for extant threads associated with this arena. 332 * Stats from these are merged incrementally, and at exit. 333 */ 334 ql_head(tcache_t) tcache_ql; 335 336 uint64_t prof_accumbytes; 337 338 dss_prec_t dss_prec; 339 340 /* Tree of dirty-page-containing chunks this arena manages. */ 341 arena_chunk_tree_t chunks_dirty; 342 343 /* 344 * In order to avoid rapid chunk allocation/deallocation when an arena 345 * oscillates right on the cusp of needing a new chunk, cache the most 346 * recently freed chunk. The spare is left in the arena's chunk trees 347 * until it is deleted. 348 * 349 * There is one spare chunk per arena, rather than one spare total, in 350 * order to avoid interactions between multiple threads that could make 351 * a single spare inadequate. 352 */ 353 arena_chunk_t *spare; 354 355 /* Number of pages in active runs. */ 356 size_t nactive; 357 358 /* 359 * Current count of pages within unused runs that are potentially 360 * dirty, and for which madvise(... MADV_DONTNEED) has not been called. 361 * By tracking this, we can institute a limit on how much dirty unused 362 * memory is mapped for each arena. 363 */ 364 size_t ndirty; 365 366 /* 367 * Approximate number of pages being purged. It is possible for 368 * multiple threads to purge dirty pages concurrently, and they use 369 * npurgatory to indicate the total number of pages all threads are 370 * attempting to purge. 371 */ 372 size_t npurgatory; 373 374 /* 375 * Size/address-ordered trees of this arena's available runs. The trees 376 * are used for first-best-fit run allocation. 377 */ 378 arena_avail_tree_t runs_avail; 379 380 /* bins is used to store trees of free regions. */ 381 arena_bin_t bins[NBINS]; 382}; 383 384#endif /* JEMALLOC_H_STRUCTS */ 385/******************************************************************************/ 386#ifdef JEMALLOC_H_EXTERNS 387 388extern ssize_t opt_lg_dirty_mult; 389/* 390 * small_size2bin is a compact lookup table that rounds request sizes up to 391 * size classes. In order to reduce cache footprint, the table is compressed, 392 * and all accesses are via the SMALL_SIZE2BIN macro. 393 */ 394extern uint8_t const small_size2bin[]; 395#define SMALL_SIZE2BIN(s) (small_size2bin[(s-1) >> LG_TINY_MIN]) 396 397extern arena_bin_info_t arena_bin_info[NBINS]; 398 399/* Number of large size classes. */ 400#define nlclasses (chunk_npages - map_bias) 401 402void arena_purge_all(arena_t *arena); 403void arena_tcache_fill_small(arena_t *arena, tcache_bin_t *tbin, 404 size_t binind, uint64_t prof_accumbytes); 405void arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, 406 bool zero); 407void arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info); 408void *arena_malloc_small(arena_t *arena, size_t size, bool zero); 409void *arena_malloc_large(arena_t *arena, size_t size, bool zero); 410void *arena_palloc(arena_t *arena, size_t size, size_t alignment, bool zero); 411void arena_prof_promoted(const void *ptr, size_t size); 412void arena_dalloc_bin_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr, 413 arena_chunk_map_t *mapelm); 414void arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr, 415 size_t pageind, arena_chunk_map_t *mapelm); 416void arena_dalloc_small(arena_t *arena, arena_chunk_t *chunk, void *ptr, 417 size_t pageind); 418void arena_dalloc_large_locked(arena_t *arena, arena_chunk_t *chunk, 419 void *ptr); 420void arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr); 421void *arena_ralloc_no_move(void *ptr, size_t oldsize, size_t size, 422 size_t extra, bool zero); 423void *arena_ralloc(arena_t *arena, void *ptr, size_t oldsize, size_t size, 424 size_t extra, size_t alignment, bool zero, bool try_tcache_alloc, 425 bool try_tcache_dalloc); 426dss_prec_t arena_dss_prec_get(arena_t *arena); 427void arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec); 428void arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive, 429 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats, 430 malloc_large_stats_t *lstats); 431bool arena_new(arena_t *arena, unsigned ind); 432void arena_boot(void); 433void arena_prefork(arena_t *arena); 434void arena_postfork_parent(arena_t *arena); 435void arena_postfork_child(arena_t *arena); 436 437#endif /* JEMALLOC_H_EXTERNS */ 438/******************************************************************************/ 439#ifdef JEMALLOC_H_INLINES 440 441#ifndef JEMALLOC_ENABLE_INLINE 442arena_chunk_map_t *arena_mapp_get(arena_chunk_t *chunk, size_t pageind); 443size_t *arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind); 444size_t arena_mapbitsp_read(size_t *mapbitsp); 445size_t arena_mapbits_get(arena_chunk_t *chunk, size_t pageind); 446size_t arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, 447 size_t pageind); 448size_t arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind); 449size_t arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind); 450size_t arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind); 451size_t arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind); 452size_t arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind); 453size_t arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind); 454size_t arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind); 455void arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits); 456void arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, 457 size_t size, size_t flags); 458void arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 459 size_t size); 460void arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, 461 size_t size, size_t flags); 462void arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 463 size_t binind); 464void arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, 465 size_t runind, size_t binind, size_t flags); 466void arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 467 size_t unzeroed); 468bool arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes); 469bool arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes); 470bool arena_prof_accum(arena_t *arena, uint64_t accumbytes); 471size_t arena_ptr_small_binind_get(const void *ptr, size_t mapbits); 472size_t arena_bin_index(arena_t *arena, arena_bin_t *bin); 473unsigned arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, 474 const void *ptr); 475prof_ctx_t *arena_prof_ctx_get(const void *ptr); 476void arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx); 477void *arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache); 478size_t arena_salloc(const void *ptr, bool demote); 479void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, 480 bool try_tcache); 481#endif 482 483#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_)) 484# ifdef JEMALLOC_ARENA_INLINE_A 485JEMALLOC_ALWAYS_INLINE arena_chunk_map_t * 486arena_mapp_get(arena_chunk_t *chunk, size_t pageind) 487{ 488 489 assert(pageind >= map_bias); 490 assert(pageind < chunk_npages); 491 492 return (&chunk->map[pageind-map_bias]); 493} 494 495JEMALLOC_ALWAYS_INLINE size_t * 496arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind) 497{ 498 499 return (&arena_mapp_get(chunk, pageind)->bits); 500} 501 502JEMALLOC_ALWAYS_INLINE size_t 503arena_mapbitsp_read(size_t *mapbitsp) 504{ 505 506 return (*mapbitsp); 507} 508 509JEMALLOC_ALWAYS_INLINE size_t 510arena_mapbits_get(arena_chunk_t *chunk, size_t pageind) 511{ 512 513 return (arena_mapbitsp_read(arena_mapbitsp_get(chunk, pageind))); 514} 515 516JEMALLOC_ALWAYS_INLINE size_t 517arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, size_t pageind) 518{ 519 size_t mapbits; 520 521 mapbits = arena_mapbits_get(chunk, pageind); 522 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 523 return (mapbits & ~PAGE_MASK); 524} 525 526JEMALLOC_ALWAYS_INLINE size_t 527arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind) 528{ 529 size_t mapbits; 530 531 mapbits = arena_mapbits_get(chunk, pageind); 532 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 533 (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)); 534 return (mapbits & ~PAGE_MASK); 535} 536 537JEMALLOC_ALWAYS_INLINE size_t 538arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind) 539{ 540 size_t mapbits; 541 542 mapbits = arena_mapbits_get(chunk, pageind); 543 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 544 CHUNK_MAP_ALLOCATED); 545 return (mapbits >> LG_PAGE); 546} 547 548JEMALLOC_ALWAYS_INLINE size_t 549arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind) 550{ 551 size_t mapbits; 552 size_t binind; 553 554 mapbits = arena_mapbits_get(chunk, pageind); 555 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 556 assert(binind < NBINS || binind == BININD_INVALID); 557 return (binind); 558} 559 560JEMALLOC_ALWAYS_INLINE size_t 561arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind) 562{ 563 size_t mapbits; 564 565 mapbits = arena_mapbits_get(chunk, pageind); 566 return (mapbits & CHUNK_MAP_DIRTY); 567} 568 569JEMALLOC_ALWAYS_INLINE size_t 570arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind) 571{ 572 size_t mapbits; 573 574 mapbits = arena_mapbits_get(chunk, pageind); 575 return (mapbits & CHUNK_MAP_UNZEROED); 576} 577 578JEMALLOC_ALWAYS_INLINE size_t 579arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind) 580{ 581 size_t mapbits; 582 583 mapbits = arena_mapbits_get(chunk, pageind); 584 return (mapbits & CHUNK_MAP_LARGE); 585} 586 587JEMALLOC_ALWAYS_INLINE size_t 588arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind) 589{ 590 size_t mapbits; 591 592 mapbits = arena_mapbits_get(chunk, pageind); 593 return (mapbits & CHUNK_MAP_ALLOCATED); 594} 595 596JEMALLOC_ALWAYS_INLINE void 597arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits) 598{ 599 600 *mapbitsp = mapbits; 601} 602 603JEMALLOC_ALWAYS_INLINE void 604arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size, 605 size_t flags) 606{ 607 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 608 609 assert((size & PAGE_MASK) == 0); 610 assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0); 611 assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags); 612 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags); 613} 614 615JEMALLOC_ALWAYS_INLINE void 616arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 617 size_t size) 618{ 619 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 620 size_t mapbits = arena_mapbitsp_read(mapbitsp); 621 622 assert((size & PAGE_MASK) == 0); 623 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 624 arena_mapbitsp_write(mapbitsp, size | (mapbits & PAGE_MASK)); 625} 626 627JEMALLOC_ALWAYS_INLINE void 628arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size, 629 size_t flags) 630{ 631 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 632 size_t mapbits = arena_mapbitsp_read(mapbitsp); 633 size_t unzeroed; 634 635 assert((size & PAGE_MASK) == 0); 636 assert((flags & CHUNK_MAP_DIRTY) == flags); 637 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 638 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags 639 | unzeroed | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED); 640} 641 642JEMALLOC_ALWAYS_INLINE void 643arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 644 size_t binind) 645{ 646 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 647 size_t mapbits = arena_mapbitsp_read(mapbitsp); 648 649 assert(binind <= BININD_INVALID); 650 assert(arena_mapbits_large_size_get(chunk, pageind) == PAGE); 651 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_BININD_MASK) | 652 (binind << CHUNK_MAP_BININD_SHIFT)); 653} 654 655JEMALLOC_ALWAYS_INLINE void 656arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, size_t runind, 657 size_t binind, size_t flags) 658{ 659 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 660 size_t mapbits = arena_mapbitsp_read(mapbitsp); 661 size_t unzeroed; 662 663 assert(binind < BININD_INVALID); 664 assert(pageind - runind >= map_bias); 665 assert((flags & CHUNK_MAP_DIRTY) == flags); 666 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 667 arena_mapbitsp_write(mapbitsp, (runind << LG_PAGE) | (binind << 668 CHUNK_MAP_BININD_SHIFT) | flags | unzeroed | CHUNK_MAP_ALLOCATED); 669} 670 671JEMALLOC_ALWAYS_INLINE void 672arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 673 size_t unzeroed) 674{ 675 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 676 size_t mapbits = arena_mapbitsp_read(mapbitsp); 677 678 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_UNZEROED) | 679 unzeroed); 680} 681 682JEMALLOC_INLINE bool 683arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes) 684{ 685 686 cassert(config_prof); 687 assert(prof_interval != 0); 688 689 arena->prof_accumbytes += accumbytes; 690 if (arena->prof_accumbytes >= prof_interval) { 691 arena->prof_accumbytes -= prof_interval; 692 return (true); 693 } 694 return (false); 695} 696 697JEMALLOC_INLINE bool 698arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes) 699{ 700 701 cassert(config_prof); 702 703 if (prof_interval == 0) 704 return (false); 705 return (arena_prof_accum_impl(arena, accumbytes)); 706} 707 708JEMALLOC_INLINE bool 709arena_prof_accum(arena_t *arena, uint64_t accumbytes) 710{ 711 712 cassert(config_prof); 713 714 if (prof_interval == 0) 715 return (false); 716 717 { 718 bool ret; 719 720 malloc_mutex_lock(&arena->lock); 721 ret = arena_prof_accum_impl(arena, accumbytes); 722 malloc_mutex_unlock(&arena->lock); 723 return (ret); 724 } 725} 726 727JEMALLOC_ALWAYS_INLINE size_t 728arena_ptr_small_binind_get(const void *ptr, size_t mapbits) 729{ 730 size_t binind; 731 732 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 733 734 if (config_debug) { 735 arena_chunk_t *chunk; 736 arena_t *arena; 737 size_t pageind; 738 size_t actual_mapbits; 739 arena_run_t *run; 740 arena_bin_t *bin; 741 size_t actual_binind; 742 arena_bin_info_t *bin_info; 743 744 assert(binind != BININD_INVALID); 745 assert(binind < NBINS); 746 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 747 arena = chunk->arena; 748 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 749 actual_mapbits = arena_mapbits_get(chunk, pageind); 750 assert(mapbits == actual_mapbits); 751 assert(arena_mapbits_large_get(chunk, pageind) == 0); 752 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 753 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind - 754 (actual_mapbits >> LG_PAGE)) << LG_PAGE)); 755 bin = run->bin; 756 actual_binind = bin - arena->bins; 757 assert(binind == actual_binind); 758 bin_info = &arena_bin_info[actual_binind]; 759 assert(((uintptr_t)ptr - ((uintptr_t)run + 760 (uintptr_t)bin_info->reg0_offset)) % bin_info->reg_interval 761 == 0); 762 } 763 764 return (binind); 765} 766# endif /* JEMALLOC_ARENA_INLINE_A */ 767 768# ifdef JEMALLOC_ARENA_INLINE_B 769JEMALLOC_INLINE size_t 770arena_bin_index(arena_t *arena, arena_bin_t *bin) 771{ 772 size_t binind = bin - arena->bins; 773 assert(binind < NBINS); 774 return (binind); 775} 776 777JEMALLOC_INLINE unsigned 778arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, const void *ptr) 779{ 780 unsigned shift, diff, regind; 781 size_t interval; 782 783 /* 784 * Freeing a pointer lower than region zero can cause assertion 785 * failure. 786 */ 787 assert((uintptr_t)ptr >= (uintptr_t)run + 788 (uintptr_t)bin_info->reg0_offset); 789 790 /* 791 * Avoid doing division with a variable divisor if possible. Using 792 * actual division here can reduce allocator throughput by over 20%! 793 */ 794 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - 795 bin_info->reg0_offset); 796 797 /* Rescale (factor powers of 2 out of the numerator and denominator). */ 798 interval = bin_info->reg_interval; 799 shift = ffs(interval) - 1; 800 diff >>= shift; 801 interval >>= shift; 802 803 if (interval == 1) { 804 /* The divisor was a power of 2. */ 805 regind = diff; 806 } else { 807 /* 808 * To divide by a number D that is not a power of two we 809 * multiply by (2^21 / D) and then right shift by 21 positions. 810 * 811 * X / D 812 * 813 * becomes 814 * 815 * (X * interval_invs[D - 3]) >> SIZE_INV_SHIFT 816 * 817 * We can omit the first three elements, because we never 818 * divide by 0, and 1 and 2 are both powers of two, which are 819 * handled above. 820 */ 821#define SIZE_INV_SHIFT ((sizeof(unsigned) << 3) - LG_RUN_MAXREGS) 822#define SIZE_INV(s) (((1U << SIZE_INV_SHIFT) / (s)) + 1) 823 static const unsigned interval_invs[] = { 824 SIZE_INV(3), 825 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 826 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 827 SIZE_INV(12), SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 828 SIZE_INV(16), SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 829 SIZE_INV(20), SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 830 SIZE_INV(24), SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 831 SIZE_INV(28), SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 832 }; 833 834 if (interval <= ((sizeof(interval_invs) / sizeof(unsigned)) + 835 2)) { 836 regind = (diff * interval_invs[interval - 3]) >> 837 SIZE_INV_SHIFT; 838 } else 839 regind = diff / interval; 840#undef SIZE_INV 841#undef SIZE_INV_SHIFT 842 } 843 assert(diff == regind * interval); 844 assert(regind < bin_info->nregs); 845 846 return (regind); 847} 848 849JEMALLOC_INLINE prof_ctx_t * 850arena_prof_ctx_get(const void *ptr) 851{ 852 prof_ctx_t *ret; 853 arena_chunk_t *chunk; 854 size_t pageind, mapbits; 855 856 cassert(config_prof); 857 assert(ptr != NULL); 858 assert(CHUNK_ADDR2BASE(ptr) != ptr); 859 860 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 861 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 862 mapbits = arena_mapbits_get(chunk, pageind); 863 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 864 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 865 if (prof_promote) 866 ret = (prof_ctx_t *)(uintptr_t)1U; 867 else { 868 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 869 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 870 LG_PAGE)); 871 size_t binind = arena_ptr_small_binind_get(ptr, 872 mapbits); 873 arena_bin_info_t *bin_info = &arena_bin_info[binind]; 874 unsigned regind; 875 876 regind = arena_run_regind(run, bin_info, ptr); 877 ret = *(prof_ctx_t **)((uintptr_t)run + 878 bin_info->ctx0_offset + (regind * 879 sizeof(prof_ctx_t *))); 880 } 881 } else 882 ret = arena_mapp_get(chunk, pageind)->prof_ctx; 883 884 return (ret); 885} 886 887JEMALLOC_INLINE void 888arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx) 889{ 890 arena_chunk_t *chunk; 891 size_t pageind, mapbits; 892 893 cassert(config_prof); 894 assert(ptr != NULL); 895 assert(CHUNK_ADDR2BASE(ptr) != ptr); 896 897 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 898 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 899 mapbits = arena_mapbits_get(chunk, pageind); 900 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 901 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 902 if (prof_promote == false) { 903 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 904 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 905 LG_PAGE)); 906 size_t binind; 907 arena_bin_info_t *bin_info; 908 unsigned regind; 909 910 binind = arena_ptr_small_binind_get(ptr, mapbits); 911 bin_info = &arena_bin_info[binind]; 912 regind = arena_run_regind(run, bin_info, ptr); 913 914 *((prof_ctx_t **)((uintptr_t)run + bin_info->ctx0_offset 915 + (regind * sizeof(prof_ctx_t *)))) = ctx; 916 } else 917 assert((uintptr_t)ctx == (uintptr_t)1U); 918 } else 919 arena_mapp_get(chunk, pageind)->prof_ctx = ctx; 920} 921 922JEMALLOC_ALWAYS_INLINE void * 923arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache) 924{ 925 tcache_t *tcache; 926 927 assert(size != 0); 928 assert(size <= arena_maxclass); 929 930 if (size <= SMALL_MAXCLASS) { 931 if (try_tcache && (tcache = tcache_get(true)) != NULL) 932 return (tcache_alloc_small(tcache, size, zero)); 933 else { 934 return (arena_malloc_small(choose_arena(arena), size, 935 zero)); 936 } 937 } else { 938 /* 939 * Initialize tcache after checking size in order to avoid 940 * infinite recursion during tcache initialization. 941 */ 942 if (try_tcache && size <= tcache_maxclass && (tcache = 943 tcache_get(true)) != NULL) 944 return (tcache_alloc_large(tcache, size, zero)); 945 else { 946 return (arena_malloc_large(choose_arena(arena), size, 947 zero)); 948 } 949 } 950} 951 952/* Return the size of the allocation pointed to by ptr. */ 953JEMALLOC_ALWAYS_INLINE size_t 954arena_salloc(const void *ptr, bool demote) 955{ 956 size_t ret; 957 arena_chunk_t *chunk; 958 size_t pageind, binind; 959 960 assert(ptr != NULL); 961 assert(CHUNK_ADDR2BASE(ptr) != ptr); 962 963 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 964 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 965 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 966 binind = arena_mapbits_binind_get(chunk, pageind); 967 if (binind == BININD_INVALID || (config_prof && demote == false && 968 prof_promote && arena_mapbits_large_get(chunk, pageind) != 0)) { 969 /* 970 * Large allocation. In the common case (demote == true), and 971 * as this is an inline function, most callers will only end up 972 * looking at binind to determine that ptr is a small 973 * allocation. 974 */ 975 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 976 ret = arena_mapbits_large_size_get(chunk, pageind); 977 assert(ret != 0); 978 assert(pageind + (ret>>LG_PAGE) <= chunk_npages); 979 assert(ret == PAGE || arena_mapbits_large_size_get(chunk, 980 pageind+(ret>>LG_PAGE)-1) == 0); 981 assert(binind == arena_mapbits_binind_get(chunk, 982 pageind+(ret>>LG_PAGE)-1)); 983 assert(arena_mapbits_dirty_get(chunk, pageind) == 984 arena_mapbits_dirty_get(chunk, pageind+(ret>>LG_PAGE)-1)); 985 } else { 986 /* 987 * Small allocation (possibly promoted to a large object due to 988 * prof_promote). 989 */ 990 assert(arena_mapbits_large_get(chunk, pageind) != 0 || 991 arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk, 992 pageind)) == binind); 993 ret = arena_bin_info[binind].reg_size; 994 } 995 996 return (ret); 997} 998 999JEMALLOC_ALWAYS_INLINE void 1000arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, bool try_tcache) 1001{ 1002 size_t pageind, mapbits; 1003 tcache_t *tcache; 1004 1005 assert(arena != NULL); 1006 assert(chunk->arena == arena); 1007 assert(ptr != NULL); 1008 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1009 1010 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1011 mapbits = arena_mapbits_get(chunk, pageind); 1012 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1013 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 1014 /* Small allocation. */ 1015 if (try_tcache && (tcache = tcache_get(false)) != NULL) { 1016 size_t binind; 1017 1018 binind = arena_ptr_small_binind_get(ptr, mapbits); 1019 tcache_dalloc_small(tcache, ptr, binind); 1020 } else 1021 arena_dalloc_small(arena, chunk, ptr, pageind); 1022 } else { 1023 size_t size = arena_mapbits_large_size_get(chunk, pageind); 1024 1025 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1026 1027 if (try_tcache && size <= tcache_maxclass && (tcache = 1028 tcache_get(false)) != NULL) { 1029 tcache_dalloc_large(tcache, ptr, size); 1030 } else 1031 arena_dalloc_large(arena, chunk, ptr); 1032 } 1033} 1034# endif /* JEMALLOC_ARENA_INLINE_B */ 1035#endif 1036 1037#endif /* JEMALLOC_H_INLINES */ 1038/******************************************************************************/ 1039