arena.h revision 665769357cd77b74e00a146f196fff19243b33c4
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 that purging could coalesce. 178 * Clean and dirty available runs are not coalesced, which causes 179 * virtual memory fragmentation. The ratio of 180 * (nruns_avail-nruns_adjac):nruns_adjac is used for tracking this 181 * fragmentation. 182 */ 183 size_t nruns_adjac; 184 185 /* 186 * Map of pages within chunk that keeps track of free/large/small. The 187 * first map_bias entries are omitted, since the chunk header does not 188 * need to be tracked in the map. This omission saves a header page 189 * for common chunk sizes (e.g. 4 MiB). 190 */ 191 arena_chunk_map_t map[1]; /* Dynamically sized. */ 192}; 193typedef rb_tree(arena_chunk_t) arena_chunk_tree_t; 194 195struct arena_run_s { 196 /* Bin this run is associated with. */ 197 arena_bin_t *bin; 198 199 /* Index of next region that has never been allocated, or nregs. */ 200 uint32_t nextind; 201 202 /* Number of free regions in run. */ 203 unsigned nfree; 204}; 205 206/* 207 * Read-only information associated with each element of arena_t's bins array 208 * is stored separately, partly to reduce memory usage (only one copy, rather 209 * than one per arena), but mainly to avoid false cacheline sharing. 210 * 211 * Each run has the following layout: 212 * 213 * /--------------------\ 214 * | arena_run_t header | 215 * | ... | 216 * bitmap_offset | bitmap | 217 * | ... | 218 * ctx0_offset | ctx map | 219 * | ... | 220 * |--------------------| 221 * | redzone | 222 * reg0_offset | region 0 | 223 * | redzone | 224 * |--------------------| \ 225 * | redzone | | 226 * | region 1 | > reg_interval 227 * | redzone | / 228 * |--------------------| 229 * | ... | 230 * | ... | 231 * | ... | 232 * |--------------------| 233 * | redzone | 234 * | region nregs-1 | 235 * | redzone | 236 * |--------------------| 237 * | alignment pad? | 238 * \--------------------/ 239 * 240 * reg_interval has at least the same minimum alignment as reg_size; this 241 * preserves the alignment constraint that sa2u() depends on. Alignment pad is 242 * either 0 or redzone_size; it is present only if needed to align reg0_offset. 243 */ 244struct arena_bin_info_s { 245 /* Size of regions in a run for this bin's size class. */ 246 size_t reg_size; 247 248 /* Redzone size. */ 249 size_t redzone_size; 250 251 /* Interval between regions (reg_size + (redzone_size << 1)). */ 252 size_t reg_interval; 253 254 /* Total size of a run for this bin's size class. */ 255 size_t run_size; 256 257 /* Total number of regions in a run for this bin's size class. */ 258 uint32_t nregs; 259 260 /* 261 * Offset of first bitmap_t element in a run header for this bin's size 262 * class. 263 */ 264 uint32_t bitmap_offset; 265 266 /* 267 * Metadata used to manipulate bitmaps for runs associated with this 268 * bin. 269 */ 270 bitmap_info_t bitmap_info; 271 272 /* 273 * Offset of first (prof_ctx_t *) in a run header for this bin's size 274 * class, or 0 if (config_prof == false || opt_prof == false). 275 */ 276 uint32_t ctx0_offset; 277 278 /* Offset of first region in a run for this bin's size class. */ 279 uint32_t reg0_offset; 280}; 281 282struct arena_bin_s { 283 /* 284 * All operations on runcur, runs, and stats require that lock be 285 * locked. Run allocation/deallocation are protected by the arena lock, 286 * which may be acquired while holding one or more bin locks, but not 287 * vise versa. 288 */ 289 malloc_mutex_t lock; 290 291 /* 292 * Current run being used to service allocations of this bin's size 293 * class. 294 */ 295 arena_run_t *runcur; 296 297 /* 298 * Tree of non-full runs. This tree is used when looking for an 299 * existing run when runcur is no longer usable. We choose the 300 * non-full run that is lowest in memory; this policy tends to keep 301 * objects packed well, and it can also help reduce the number of 302 * almost-empty chunks. 303 */ 304 arena_run_tree_t runs; 305 306 /* Bin statistics. */ 307 malloc_bin_stats_t stats; 308}; 309 310struct arena_s { 311 /* This arena's index within the arenas array. */ 312 unsigned ind; 313 314 /* 315 * Number of threads currently assigned to this arena. This field is 316 * protected by arenas_lock. 317 */ 318 unsigned nthreads; 319 320 /* 321 * There are three classes of arena operations from a locking 322 * perspective: 323 * 1) Thread asssignment (modifies nthreads) is protected by 324 * arenas_lock. 325 * 2) Bin-related operations are protected by bin locks. 326 * 3) Chunk- and run-related operations are protected by this mutex. 327 */ 328 malloc_mutex_t lock; 329 330 arena_stats_t stats; 331 /* 332 * List of tcaches for extant threads associated with this arena. 333 * Stats from these are merged incrementally, and at exit. 334 */ 335 ql_head(tcache_t) tcache_ql; 336 337 uint64_t prof_accumbytes; 338 339 dss_prec_t dss_prec; 340 341 /* Tree of dirty-page-containing chunks this arena manages. */ 342 arena_chunk_tree_t chunks_dirty; 343 344 /* 345 * In order to avoid rapid chunk allocation/deallocation when an arena 346 * oscillates right on the cusp of needing a new chunk, cache the most 347 * recently freed chunk. The spare is left in the arena's chunk trees 348 * until it is deleted. 349 * 350 * There is one spare chunk per arena, rather than one spare total, in 351 * order to avoid interactions between multiple threads that could make 352 * a single spare inadequate. 353 */ 354 arena_chunk_t *spare; 355 356 /* Number of pages in active runs. */ 357 size_t nactive; 358 359 /* 360 * Current count of pages within unused runs that are potentially 361 * dirty, and for which madvise(... MADV_DONTNEED) has not been called. 362 * By tracking this, we can institute a limit on how much dirty unused 363 * memory is mapped for each arena. 364 */ 365 size_t ndirty; 366 367 /* 368 * Approximate number of pages being purged. It is possible for 369 * multiple threads to purge dirty pages concurrently, and they use 370 * npurgatory to indicate the total number of pages all threads are 371 * attempting to purge. 372 */ 373 size_t npurgatory; 374 375 /* 376 * Size/address-ordered trees of this arena's available runs. The trees 377 * are used for first-best-fit run allocation. 378 */ 379 arena_avail_tree_t runs_avail; 380 381 /* bins is used to store trees of free regions. */ 382 arena_bin_t bins[NBINS]; 383}; 384 385#endif /* JEMALLOC_H_STRUCTS */ 386/******************************************************************************/ 387#ifdef JEMALLOC_H_EXTERNS 388 389extern ssize_t opt_lg_dirty_mult; 390/* 391 * small_size2bin is a compact lookup table that rounds request sizes up to 392 * size classes. In order to reduce cache footprint, the table is compressed, 393 * and all accesses are via the SMALL_SIZE2BIN macro. 394 */ 395extern uint8_t const small_size2bin[]; 396#define SMALL_SIZE2BIN(s) (small_size2bin[(s-1) >> LG_TINY_MIN]) 397 398extern arena_bin_info_t arena_bin_info[NBINS]; 399 400/* Number of large size classes. */ 401#define nlclasses (chunk_npages - map_bias) 402 403void arena_purge_all(arena_t *arena); 404void arena_tcache_fill_small(arena_t *arena, tcache_bin_t *tbin, 405 size_t binind, uint64_t prof_accumbytes); 406void arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, 407 bool zero); 408void arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info); 409void *arena_malloc_small(arena_t *arena, size_t size, bool zero); 410void *arena_malloc_large(arena_t *arena, size_t size, bool zero); 411void *arena_palloc(arena_t *arena, size_t size, size_t alignment, bool zero); 412void arena_prof_promoted(const void *ptr, size_t size); 413void arena_dalloc_bin_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr, 414 arena_chunk_map_t *mapelm); 415void arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr, 416 size_t pageind, arena_chunk_map_t *mapelm); 417void arena_dalloc_small(arena_t *arena, arena_chunk_t *chunk, void *ptr, 418 size_t pageind); 419void arena_dalloc_large_locked(arena_t *arena, arena_chunk_t *chunk, 420 void *ptr); 421void arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr); 422void *arena_ralloc_no_move(void *ptr, size_t oldsize, size_t size, 423 size_t extra, bool zero); 424void *arena_ralloc(arena_t *arena, void *ptr, size_t oldsize, size_t size, 425 size_t extra, size_t alignment, bool zero, bool try_tcache_alloc, 426 bool try_tcache_dalloc); 427dss_prec_t arena_dss_prec_get(arena_t *arena); 428void arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec); 429void arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive, 430 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats, 431 malloc_large_stats_t *lstats); 432bool arena_new(arena_t *arena, unsigned ind); 433void arena_boot(void); 434void arena_prefork(arena_t *arena); 435void arena_postfork_parent(arena_t *arena); 436void arena_postfork_child(arena_t *arena); 437 438#endif /* JEMALLOC_H_EXTERNS */ 439/******************************************************************************/ 440#ifdef JEMALLOC_H_INLINES 441 442#ifndef JEMALLOC_ENABLE_INLINE 443arena_chunk_map_t *arena_mapp_get(arena_chunk_t *chunk, size_t pageind); 444size_t *arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind); 445size_t arena_mapbitsp_read(size_t *mapbitsp); 446size_t arena_mapbits_get(arena_chunk_t *chunk, size_t pageind); 447size_t arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, 448 size_t pageind); 449size_t arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind); 450size_t arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind); 451size_t arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind); 452size_t arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind); 453size_t arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind); 454size_t arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind); 455size_t arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind); 456void arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits); 457void arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, 458 size_t size, size_t flags); 459void arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 460 size_t size); 461void arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, 462 size_t size, size_t flags); 463void arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 464 size_t binind); 465void arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, 466 size_t runind, size_t binind, size_t flags); 467void arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 468 size_t unzeroed); 469bool arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes); 470bool arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes); 471bool arena_prof_accum(arena_t *arena, uint64_t accumbytes); 472size_t arena_ptr_small_binind_get(const void *ptr, size_t mapbits); 473size_t arena_bin_index(arena_t *arena, arena_bin_t *bin); 474unsigned arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, 475 const void *ptr); 476prof_ctx_t *arena_prof_ctx_get(const void *ptr); 477void arena_prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx); 478void *arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache); 479size_t arena_salloc(const void *ptr, bool demote); 480void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, 481 bool try_tcache); 482#endif 483 484#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_)) 485# ifdef JEMALLOC_ARENA_INLINE_A 486JEMALLOC_ALWAYS_INLINE arena_chunk_map_t * 487arena_mapp_get(arena_chunk_t *chunk, size_t pageind) 488{ 489 490 assert(pageind >= map_bias); 491 assert(pageind < chunk_npages); 492 493 return (&chunk->map[pageind-map_bias]); 494} 495 496JEMALLOC_ALWAYS_INLINE size_t * 497arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind) 498{ 499 500 return (&arena_mapp_get(chunk, pageind)->bits); 501} 502 503JEMALLOC_ALWAYS_INLINE size_t 504arena_mapbitsp_read(size_t *mapbitsp) 505{ 506 507 return (*mapbitsp); 508} 509 510JEMALLOC_ALWAYS_INLINE size_t 511arena_mapbits_get(arena_chunk_t *chunk, size_t pageind) 512{ 513 514 return (arena_mapbitsp_read(arena_mapbitsp_get(chunk, pageind))); 515} 516 517JEMALLOC_ALWAYS_INLINE size_t 518arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, size_t pageind) 519{ 520 size_t mapbits; 521 522 mapbits = arena_mapbits_get(chunk, pageind); 523 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 524 return (mapbits & ~PAGE_MASK); 525} 526 527JEMALLOC_ALWAYS_INLINE size_t 528arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind) 529{ 530 size_t mapbits; 531 532 mapbits = arena_mapbits_get(chunk, pageind); 533 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 534 (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)); 535 return (mapbits & ~PAGE_MASK); 536} 537 538JEMALLOC_ALWAYS_INLINE size_t 539arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind) 540{ 541 size_t mapbits; 542 543 mapbits = arena_mapbits_get(chunk, pageind); 544 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 545 CHUNK_MAP_ALLOCATED); 546 return (mapbits >> LG_PAGE); 547} 548 549JEMALLOC_ALWAYS_INLINE size_t 550arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind) 551{ 552 size_t mapbits; 553 size_t binind; 554 555 mapbits = arena_mapbits_get(chunk, pageind); 556 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 557 assert(binind < NBINS || binind == BININD_INVALID); 558 return (binind); 559} 560 561JEMALLOC_ALWAYS_INLINE size_t 562arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind) 563{ 564 size_t mapbits; 565 566 mapbits = arena_mapbits_get(chunk, pageind); 567 return (mapbits & CHUNK_MAP_DIRTY); 568} 569 570JEMALLOC_ALWAYS_INLINE size_t 571arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind) 572{ 573 size_t mapbits; 574 575 mapbits = arena_mapbits_get(chunk, pageind); 576 return (mapbits & CHUNK_MAP_UNZEROED); 577} 578 579JEMALLOC_ALWAYS_INLINE size_t 580arena_mapbits_large_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_LARGE); 586} 587 588JEMALLOC_ALWAYS_INLINE size_t 589arena_mapbits_allocated_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_ALLOCATED); 595} 596 597JEMALLOC_ALWAYS_INLINE void 598arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits) 599{ 600 601 *mapbitsp = mapbits; 602} 603 604JEMALLOC_ALWAYS_INLINE void 605arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size, 606 size_t flags) 607{ 608 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 609 610 assert((size & PAGE_MASK) == 0); 611 assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0); 612 assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags); 613 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags); 614} 615 616JEMALLOC_ALWAYS_INLINE void 617arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 618 size_t size) 619{ 620 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 621 size_t mapbits = arena_mapbitsp_read(mapbitsp); 622 623 assert((size & PAGE_MASK) == 0); 624 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 625 arena_mapbitsp_write(mapbitsp, size | (mapbits & PAGE_MASK)); 626} 627 628JEMALLOC_ALWAYS_INLINE void 629arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size, 630 size_t flags) 631{ 632 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 633 size_t mapbits = arena_mapbitsp_read(mapbitsp); 634 size_t unzeroed; 635 636 assert((size & PAGE_MASK) == 0); 637 assert((flags & CHUNK_MAP_DIRTY) == flags); 638 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 639 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags 640 | unzeroed | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED); 641} 642 643JEMALLOC_ALWAYS_INLINE void 644arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 645 size_t binind) 646{ 647 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 648 size_t mapbits = arena_mapbitsp_read(mapbitsp); 649 650 assert(binind <= BININD_INVALID); 651 assert(arena_mapbits_large_size_get(chunk, pageind) == PAGE); 652 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_BININD_MASK) | 653 (binind << CHUNK_MAP_BININD_SHIFT)); 654} 655 656JEMALLOC_ALWAYS_INLINE void 657arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, size_t runind, 658 size_t binind, size_t flags) 659{ 660 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 661 size_t mapbits = arena_mapbitsp_read(mapbitsp); 662 size_t unzeroed; 663 664 assert(binind < BININD_INVALID); 665 assert(pageind - runind >= map_bias); 666 assert((flags & CHUNK_MAP_DIRTY) == flags); 667 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 668 arena_mapbitsp_write(mapbitsp, (runind << LG_PAGE) | (binind << 669 CHUNK_MAP_BININD_SHIFT) | flags | unzeroed | CHUNK_MAP_ALLOCATED); 670} 671 672JEMALLOC_ALWAYS_INLINE void 673arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 674 size_t unzeroed) 675{ 676 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 677 size_t mapbits = arena_mapbitsp_read(mapbitsp); 678 679 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_UNZEROED) | 680 unzeroed); 681} 682 683JEMALLOC_INLINE bool 684arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes) 685{ 686 687 cassert(config_prof); 688 assert(prof_interval != 0); 689 690 arena->prof_accumbytes += accumbytes; 691 if (arena->prof_accumbytes >= prof_interval) { 692 arena->prof_accumbytes -= prof_interval; 693 return (true); 694 } 695 return (false); 696} 697 698JEMALLOC_INLINE bool 699arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes) 700{ 701 702 cassert(config_prof); 703 704 if (prof_interval == 0) 705 return (false); 706 return (arena_prof_accum_impl(arena, accumbytes)); 707} 708 709JEMALLOC_INLINE bool 710arena_prof_accum(arena_t *arena, uint64_t accumbytes) 711{ 712 713 cassert(config_prof); 714 715 if (prof_interval == 0) 716 return (false); 717 718 { 719 bool ret; 720 721 malloc_mutex_lock(&arena->lock); 722 ret = arena_prof_accum_impl(arena, accumbytes); 723 malloc_mutex_unlock(&arena->lock); 724 return (ret); 725 } 726} 727 728JEMALLOC_ALWAYS_INLINE size_t 729arena_ptr_small_binind_get(const void *ptr, size_t mapbits) 730{ 731 size_t binind; 732 733 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 734 735 if (config_debug) { 736 arena_chunk_t *chunk; 737 arena_t *arena; 738 size_t pageind; 739 size_t actual_mapbits; 740 arena_run_t *run; 741 arena_bin_t *bin; 742 size_t actual_binind; 743 arena_bin_info_t *bin_info; 744 745 assert(binind != BININD_INVALID); 746 assert(binind < NBINS); 747 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 748 arena = chunk->arena; 749 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 750 actual_mapbits = arena_mapbits_get(chunk, pageind); 751 assert(mapbits == actual_mapbits); 752 assert(arena_mapbits_large_get(chunk, pageind) == 0); 753 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 754 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind - 755 (actual_mapbits >> LG_PAGE)) << LG_PAGE)); 756 bin = run->bin; 757 actual_binind = bin - arena->bins; 758 assert(binind == actual_binind); 759 bin_info = &arena_bin_info[actual_binind]; 760 assert(((uintptr_t)ptr - ((uintptr_t)run + 761 (uintptr_t)bin_info->reg0_offset)) % bin_info->reg_interval 762 == 0); 763 } 764 765 return (binind); 766} 767# endif /* JEMALLOC_ARENA_INLINE_A */ 768 769# ifdef JEMALLOC_ARENA_INLINE_B 770JEMALLOC_INLINE size_t 771arena_bin_index(arena_t *arena, arena_bin_t *bin) 772{ 773 size_t binind = bin - arena->bins; 774 assert(binind < NBINS); 775 return (binind); 776} 777 778JEMALLOC_INLINE unsigned 779arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, const void *ptr) 780{ 781 unsigned shift, diff, regind; 782 size_t interval; 783 784 /* 785 * Freeing a pointer lower than region zero can cause assertion 786 * failure. 787 */ 788 assert((uintptr_t)ptr >= (uintptr_t)run + 789 (uintptr_t)bin_info->reg0_offset); 790 791 /* 792 * Avoid doing division with a variable divisor if possible. Using 793 * actual division here can reduce allocator throughput by over 20%! 794 */ 795 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - 796 bin_info->reg0_offset); 797 798 /* Rescale (factor powers of 2 out of the numerator and denominator). */ 799 interval = bin_info->reg_interval; 800 shift = ffs(interval) - 1; 801 diff >>= shift; 802 interval >>= shift; 803 804 if (interval == 1) { 805 /* The divisor was a power of 2. */ 806 regind = diff; 807 } else { 808 /* 809 * To divide by a number D that is not a power of two we 810 * multiply by (2^21 / D) and then right shift by 21 positions. 811 * 812 * X / D 813 * 814 * becomes 815 * 816 * (X * interval_invs[D - 3]) >> SIZE_INV_SHIFT 817 * 818 * We can omit the first three elements, because we never 819 * divide by 0, and 1 and 2 are both powers of two, which are 820 * handled above. 821 */ 822#define SIZE_INV_SHIFT ((sizeof(unsigned) << 3) - LG_RUN_MAXREGS) 823#define SIZE_INV(s) (((1U << SIZE_INV_SHIFT) / (s)) + 1) 824 static const unsigned interval_invs[] = { 825 SIZE_INV(3), 826 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 827 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 828 SIZE_INV(12), SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 829 SIZE_INV(16), SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 830 SIZE_INV(20), SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 831 SIZE_INV(24), SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 832 SIZE_INV(28), SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 833 }; 834 835 if (interval <= ((sizeof(interval_invs) / sizeof(unsigned)) + 836 2)) { 837 regind = (diff * interval_invs[interval - 3]) >> 838 SIZE_INV_SHIFT; 839 } else 840 regind = diff / interval; 841#undef SIZE_INV 842#undef SIZE_INV_SHIFT 843 } 844 assert(diff == regind * interval); 845 assert(regind < bin_info->nregs); 846 847 return (regind); 848} 849 850JEMALLOC_INLINE prof_ctx_t * 851arena_prof_ctx_get(const void *ptr) 852{ 853 prof_ctx_t *ret; 854 arena_chunk_t *chunk; 855 size_t pageind, mapbits; 856 857 cassert(config_prof); 858 assert(ptr != NULL); 859 assert(CHUNK_ADDR2BASE(ptr) != ptr); 860 861 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 862 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 863 mapbits = arena_mapbits_get(chunk, pageind); 864 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 865 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 866 if (prof_promote) 867 ret = (prof_ctx_t *)(uintptr_t)1U; 868 else { 869 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 870 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 871 LG_PAGE)); 872 size_t binind = arena_ptr_small_binind_get(ptr, 873 mapbits); 874 arena_bin_info_t *bin_info = &arena_bin_info[binind]; 875 unsigned regind; 876 877 regind = arena_run_regind(run, bin_info, ptr); 878 ret = *(prof_ctx_t **)((uintptr_t)run + 879 bin_info->ctx0_offset + (regind * 880 sizeof(prof_ctx_t *))); 881 } 882 } else 883 ret = arena_mapp_get(chunk, pageind)->prof_ctx; 884 885 return (ret); 886} 887 888JEMALLOC_INLINE void 889arena_prof_ctx_set(const void *ptr, size_t usize, prof_ctx_t *ctx) 890{ 891 arena_chunk_t *chunk; 892 size_t pageind, mapbits; 893 894 cassert(config_prof); 895 assert(ptr != NULL); 896 assert(CHUNK_ADDR2BASE(ptr) != ptr); 897 898 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 899 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 900 mapbits = arena_mapbits_get(chunk, pageind); 901 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 902 903 if (usize > SMALL_MAXCLASS || (prof_promote && 904 ((uintptr_t)ctx != (uintptr_t)1U || ((mapbits & CHUNK_MAP_LARGE) != 905 0)))) { 906 assert((mapbits & CHUNK_MAP_LARGE) != 0); 907 arena_mapp_get(chunk, pageind)->prof_ctx = ctx; 908 } else { 909 assert((mapbits & CHUNK_MAP_LARGE) == 0); 910 if (prof_promote == false) { 911 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk + 912 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) << 913 LG_PAGE)); 914 size_t binind; 915 arena_bin_info_t *bin_info; 916 unsigned regind; 917 918 binind = arena_ptr_small_binind_get(ptr, mapbits); 919 bin_info = &arena_bin_info[binind]; 920 regind = arena_run_regind(run, bin_info, ptr); 921 922 *((prof_ctx_t **)((uintptr_t)run + 923 bin_info->ctx0_offset + (regind * sizeof(prof_ctx_t 924 *)))) = ctx; 925 } 926 } 927} 928 929JEMALLOC_ALWAYS_INLINE void * 930arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache) 931{ 932 tcache_t *tcache; 933 934 assert(size != 0); 935 assert(size <= arena_maxclass); 936 937 if (size <= SMALL_MAXCLASS) { 938 if (try_tcache && (tcache = tcache_get(true)) != NULL) 939 return (tcache_alloc_small(tcache, size, zero)); 940 else { 941 return (arena_malloc_small(choose_arena(arena), size, 942 zero)); 943 } 944 } else { 945 /* 946 * Initialize tcache after checking size in order to avoid 947 * infinite recursion during tcache initialization. 948 */ 949 if (try_tcache && size <= tcache_maxclass && (tcache = 950 tcache_get(true)) != NULL) 951 return (tcache_alloc_large(tcache, size, zero)); 952 else { 953 return (arena_malloc_large(choose_arena(arena), size, 954 zero)); 955 } 956 } 957} 958 959/* Return the size of the allocation pointed to by ptr. */ 960JEMALLOC_ALWAYS_INLINE size_t 961arena_salloc(const void *ptr, bool demote) 962{ 963 size_t ret; 964 arena_chunk_t *chunk; 965 size_t pageind, binind; 966 967 assert(ptr != NULL); 968 assert(CHUNK_ADDR2BASE(ptr) != ptr); 969 970 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 971 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 972 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 973 binind = arena_mapbits_binind_get(chunk, pageind); 974 if (binind == BININD_INVALID || (config_prof && demote == false && 975 prof_promote && arena_mapbits_large_get(chunk, pageind) != 0)) { 976 /* 977 * Large allocation. In the common case (demote == true), and 978 * as this is an inline function, most callers will only end up 979 * looking at binind to determine that ptr is a small 980 * allocation. 981 */ 982 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 983 ret = arena_mapbits_large_size_get(chunk, pageind); 984 assert(ret != 0); 985 assert(pageind + (ret>>LG_PAGE) <= chunk_npages); 986 assert(ret == PAGE || arena_mapbits_large_size_get(chunk, 987 pageind+(ret>>LG_PAGE)-1) == 0); 988 assert(binind == arena_mapbits_binind_get(chunk, 989 pageind+(ret>>LG_PAGE)-1)); 990 assert(arena_mapbits_dirty_get(chunk, pageind) == 991 arena_mapbits_dirty_get(chunk, pageind+(ret>>LG_PAGE)-1)); 992 } else { 993 /* 994 * Small allocation (possibly promoted to a large object due to 995 * prof_promote). 996 */ 997 assert(arena_mapbits_large_get(chunk, pageind) != 0 || 998 arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk, 999 pageind)) == binind); 1000 ret = arena_bin_info[binind].reg_size; 1001 } 1002 1003 return (ret); 1004} 1005 1006JEMALLOC_ALWAYS_INLINE void 1007arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, bool try_tcache) 1008{ 1009 size_t pageind, mapbits; 1010 tcache_t *tcache; 1011 1012 assert(arena != NULL); 1013 assert(chunk->arena == arena); 1014 assert(ptr != NULL); 1015 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1016 1017 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1018 mapbits = arena_mapbits_get(chunk, pageind); 1019 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1020 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 1021 /* Small allocation. */ 1022 if (try_tcache && (tcache = tcache_get(false)) != NULL) { 1023 size_t binind; 1024 1025 binind = arena_ptr_small_binind_get(ptr, mapbits); 1026 tcache_dalloc_small(tcache, ptr, binind); 1027 } else 1028 arena_dalloc_small(arena, chunk, ptr, pageind); 1029 } else { 1030 size_t size = arena_mapbits_large_size_get(chunk, pageind); 1031 1032 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1033 1034 if (try_tcache && size <= tcache_maxclass && (tcache = 1035 tcache_get(false)) != NULL) { 1036 tcache_dalloc_large(tcache, ptr, size); 1037 } else 1038 arena_dalloc_large(arena, chunk, ptr); 1039 } 1040} 1041# endif /* JEMALLOC_ARENA_INLINE_B */ 1042#endif 1043 1044#endif /* JEMALLOC_H_INLINES */ 1045/******************************************************************************/ 1046