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 46 typedef struct arena_chunk_map_s arena_chunk_map_t; 47 typedef struct arena_chunk_s arena_chunk_t; 48 typedef struct arena_run_s arena_run_t; 49 typedef struct arena_bin_info_s arena_bin_info_t; 50 typedef struct arena_bin_s arena_bin_t; 51 typedef 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. */ 58 struct 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 * x : don't care 114 * - : 0 115 * + : 1 116 * [DULA] : bit set 117 * [dula] : bit unset 118 * 119 * Unallocated (clean): 120 * ssssssss ssssssss ssss++++ ++++du-a 121 * xxxxxxxx xxxxxxxx xxxxxxxx xxxx-Uxx 122 * ssssssss ssssssss ssss++++ ++++dU-a 123 * 124 * Unallocated (dirty): 125 * ssssssss ssssssss ssss++++ ++++D--a 126 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx 127 * ssssssss ssssssss ssss++++ ++++D--a 128 * 129 * Small: 130 * pppppppp pppppppp ppppnnnn nnnnd--A 131 * pppppppp pppppppp ppppnnnn nnnn---A 132 * pppppppp pppppppp ppppnnnn nnnnd--A 133 * 134 * Large: 135 * ssssssss ssssssss ssss++++ ++++D-LA 136 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx 137 * -------- -------- ----++++ ++++D-LA 138 * 139 * Large (sampled, size <= PAGE): 140 * ssssssss ssssssss ssssnnnn nnnnD-LA 141 * 142 * Large (not sampled, size == PAGE): 143 * ssssssss ssssssss ssss++++ ++++D-LA 144 */ 145 size_t bits; 146 #define CHUNK_MAP_BININD_SHIFT 4 147 #define BININD_INVALID ((size_t)0xffU) 148 /* CHUNK_MAP_BININD_MASK == (BININD_INVALID << CHUNK_MAP_BININD_SHIFT) */ 149 #define CHUNK_MAP_BININD_MASK ((size_t)0xff0U) 150 #define CHUNK_MAP_BININD_INVALID CHUNK_MAP_BININD_MASK 151 #define CHUNK_MAP_FLAGS_MASK ((size_t)0xcU) 152 #define CHUNK_MAP_DIRTY ((size_t)0x8U) 153 #define CHUNK_MAP_UNZEROED ((size_t)0x4U) 154 #define CHUNK_MAP_LARGE ((size_t)0x2U) 155 #define CHUNK_MAP_ALLOCATED ((size_t)0x1U) 156 #define CHUNK_MAP_KEY CHUNK_MAP_ALLOCATED 157 }; 158 typedef rb_tree(arena_chunk_map_t) arena_avail_tree_t; 159 typedef rb_tree(arena_chunk_map_t) arena_run_tree_t; 160 typedef ql_head(arena_chunk_map_t) arena_chunk_mapelms_t; 161 162 /* Arena chunk header. */ 163 struct 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 }; 193 typedef rb_tree(arena_chunk_t) arena_chunk_tree_t; 194 195 struct 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 * |--------------------| 219 * | redzone | 220 * reg0_offset | region 0 | 221 * | redzone | 222 * |--------------------| \ 223 * | redzone | | 224 * | region 1 | > reg_interval 225 * | redzone | / 226 * |--------------------| 227 * | ... | 228 * | ... | 229 * | ... | 230 * |--------------------| 231 * | redzone | 232 * | region nregs-1 | 233 * | redzone | 234 * |--------------------| 235 * | alignment pad? | 236 * \--------------------/ 237 * 238 * reg_interval has at least the same minimum alignment as reg_size; this 239 * preserves the alignment constraint that sa2u() depends on. Alignment pad is 240 * either 0 or redzone_size; it is present only if needed to align reg0_offset. 241 */ 242 struct arena_bin_info_s { 243 /* Size of regions in a run for this bin's size class. */ 244 size_t reg_size; 245 246 /* Redzone size. */ 247 size_t redzone_size; 248 249 /* Interval between regions (reg_size + (redzone_size << 1)). */ 250 size_t reg_interval; 251 252 /* Total size of a run for this bin's size class. */ 253 size_t run_size; 254 255 /* Total number of regions in a run for this bin's size class. */ 256 uint32_t nregs; 257 258 /* 259 * Offset of first bitmap_t element in a run header for this bin's size 260 * class. 261 */ 262 uint32_t bitmap_offset; 263 264 /* 265 * Metadata used to manipulate bitmaps for runs associated with this 266 * bin. 267 */ 268 bitmap_info_t bitmap_info; 269 270 /* Offset of first region in a run for this bin's size class. */ 271 uint32_t reg0_offset; 272 }; 273 274 struct arena_bin_s { 275 /* 276 * All operations on runcur, runs, and stats require that lock be 277 * locked. Run allocation/deallocation are protected by the arena lock, 278 * which may be acquired while holding one or more bin locks, but not 279 * vise versa. 280 */ 281 malloc_mutex_t lock; 282 283 /* 284 * Current run being used to service allocations of this bin's size 285 * class. 286 */ 287 arena_run_t *runcur; 288 289 /* 290 * Tree of non-full runs. This tree is used when looking for an 291 * existing run when runcur is no longer usable. We choose the 292 * non-full run that is lowest in memory; this policy tends to keep 293 * objects packed well, and it can also help reduce the number of 294 * almost-empty chunks. 295 */ 296 arena_run_tree_t runs; 297 298 /* Bin statistics. */ 299 malloc_bin_stats_t stats; 300 }; 301 302 struct arena_s { 303 /* This arena's index within the arenas array. */ 304 unsigned ind; 305 306 /* 307 * Number of threads currently assigned to this arena. This field is 308 * protected by arenas_lock. 309 */ 310 unsigned nthreads; 311 312 /* 313 * There are three classes of arena operations from a locking 314 * perspective: 315 * 1) Thread asssignment (modifies nthreads) is protected by 316 * arenas_lock. 317 * 2) Bin-related operations are protected by bin locks. 318 * 3) Chunk- and run-related operations are protected by this mutex. 319 */ 320 malloc_mutex_t lock; 321 322 arena_stats_t stats; 323 /* 324 * List of tcaches for extant threads associated with this arena. 325 * Stats from these are merged incrementally, and at exit. 326 */ 327 ql_head(tcache_t) tcache_ql; 328 329 uint64_t prof_accumbytes; 330 331 dss_prec_t dss_prec; 332 333 /* Tree of dirty-page-containing chunks this arena manages. */ 334 arena_chunk_tree_t chunks_dirty; 335 336 /* 337 * In order to avoid rapid chunk allocation/deallocation when an arena 338 * oscillates right on the cusp of needing a new chunk, cache the most 339 * recently freed chunk. The spare is left in the arena's chunk trees 340 * until it is deleted. 341 * 342 * There is one spare chunk per arena, rather than one spare total, in 343 * order to avoid interactions between multiple threads that could make 344 * a single spare inadequate. 345 */ 346 arena_chunk_t *spare; 347 348 /* Number of pages in active runs and huge regions. */ 349 size_t nactive; 350 351 /* 352 * Current count of pages within unused runs that are potentially 353 * dirty, and for which madvise(... MADV_DONTNEED) has not been called. 354 * By tracking this, we can institute a limit on how much dirty unused 355 * memory is mapped for each arena. 356 */ 357 size_t ndirty; 358 359 /* 360 * Approximate number of pages being purged. It is possible for 361 * multiple threads to purge dirty pages concurrently, and they use 362 * npurgatory to indicate the total number of pages all threads are 363 * attempting to purge. 364 */ 365 size_t npurgatory; 366 367 /* 368 * Size/address-ordered trees of this arena's available runs. The trees 369 * are used for first-best-fit run allocation. 370 */ 371 arena_avail_tree_t runs_avail; 372 373 /* 374 * user-configureable chunk allocation and deallocation functions. 375 */ 376 chunk_alloc_t *chunk_alloc; 377 chunk_dalloc_t *chunk_dalloc; 378 379 /* bins is used to store trees of free regions. */ 380 arena_bin_t bins[NBINS]; 381 }; 382 383 #endif /* JEMALLOC_H_STRUCTS */ 384 /******************************************************************************/ 385 #ifdef JEMALLOC_H_EXTERNS 386 387 extern ssize_t opt_lg_dirty_mult; 388 /* 389 * small_size2bin_tab is a compact lookup table that rounds request sizes up to 390 * size classes. In order to reduce cache footprint, the table is compressed, 391 * and all accesses are via small_size2bin(). 392 */ 393 extern uint8_t const small_size2bin_tab[]; 394 /* 395 * small_bin2size_tab duplicates information in arena_bin_info, but in a const 396 * array, for which it is easier for the compiler to optimize repeated 397 * dereferences. 398 */ 399 extern uint32_t const small_bin2size_tab[NBINS]; 400 401 extern arena_bin_info_t arena_bin_info[NBINS]; 402 403 /* Number of large size classes. */ 404 #define nlclasses (chunk_npages - map_bias) 405 406 void *arena_chunk_alloc_huge(arena_t *arena, size_t size, size_t alignment, 407 bool *zero); 408 void arena_chunk_dalloc_huge(arena_t *arena, void *chunk, size_t size); 409 void arena_purge_all(arena_t *arena); 410 void arena_tcache_fill_small(arena_t *arena, tcache_bin_t *tbin, 411 size_t binind, uint64_t prof_accumbytes); 412 void arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info, 413 bool zero); 414 #ifdef JEMALLOC_JET 415 typedef void (arena_redzone_corruption_t)(void *, size_t, bool, size_t, 416 uint8_t); 417 extern arena_redzone_corruption_t *arena_redzone_corruption; 418 typedef void (arena_dalloc_junk_small_t)(void *, arena_bin_info_t *); 419 extern arena_dalloc_junk_small_t *arena_dalloc_junk_small; 420 #else 421 void arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info); 422 #endif 423 void arena_quarantine_junk_small(void *ptr, size_t usize); 424 void *arena_malloc_small(arena_t *arena, size_t size, bool zero); 425 void *arena_malloc_large(arena_t *arena, size_t size, bool zero); 426 void *arena_palloc(arena_t *arena, size_t size, size_t alignment, bool zero); 427 void arena_prof_promoted(const void *ptr, size_t size); 428 void arena_dalloc_bin_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr, 429 arena_chunk_map_t *mapelm); 430 void arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr, 431 size_t pageind, arena_chunk_map_t *mapelm); 432 void arena_dalloc_small(arena_t *arena, arena_chunk_t *chunk, void *ptr, 433 size_t pageind); 434 #ifdef JEMALLOC_JET 435 typedef void (arena_dalloc_junk_large_t)(void *, size_t); 436 extern arena_dalloc_junk_large_t *arena_dalloc_junk_large; 437 #endif 438 void arena_dalloc_large_locked(arena_t *arena, arena_chunk_t *chunk, 439 void *ptr); 440 void arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr); 441 #ifdef JEMALLOC_JET 442 typedef void (arena_ralloc_junk_large_t)(void *, size_t, size_t); 443 extern arena_ralloc_junk_large_t *arena_ralloc_junk_large; 444 #endif 445 bool arena_ralloc_no_move(void *ptr, size_t oldsize, size_t size, 446 size_t extra, bool zero); 447 void *arena_ralloc(arena_t *arena, void *ptr, size_t oldsize, size_t size, 448 size_t extra, size_t alignment, bool zero, bool try_tcache_alloc, 449 bool try_tcache_dalloc); 450 dss_prec_t arena_dss_prec_get(arena_t *arena); 451 bool arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec); 452 void arena_stats_merge(arena_t *arena, const char **dss, size_t *nactive, 453 size_t *ndirty, arena_stats_t *astats, malloc_bin_stats_t *bstats, 454 malloc_large_stats_t *lstats); 455 bool arena_new(arena_t *arena, unsigned ind); 456 void arena_boot(void); 457 void arena_prefork(arena_t *arena); 458 void arena_postfork_parent(arena_t *arena); 459 void arena_postfork_child(arena_t *arena); 460 461 #endif /* JEMALLOC_H_EXTERNS */ 462 /******************************************************************************/ 463 #ifdef JEMALLOC_H_INLINES 464 465 #ifndef JEMALLOC_ENABLE_INLINE 466 size_t small_size2bin_compute(size_t size); 467 size_t small_size2bin_lookup(size_t size); 468 size_t small_size2bin(size_t size); 469 size_t small_bin2size_compute(size_t binind); 470 size_t small_bin2size_lookup(size_t binind); 471 size_t small_bin2size(size_t binind); 472 size_t small_s2u_compute(size_t size); 473 size_t small_s2u_lookup(size_t size); 474 size_t small_s2u(size_t size); 475 arena_chunk_map_t *arena_mapp_get(arena_chunk_t *chunk, size_t pageind); 476 size_t *arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind); 477 size_t arena_mapbitsp_read(size_t *mapbitsp); 478 size_t arena_mapbits_get(arena_chunk_t *chunk, size_t pageind); 479 size_t arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, 480 size_t pageind); 481 size_t arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind); 482 size_t arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind); 483 size_t arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind); 484 size_t arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind); 485 size_t arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind); 486 size_t arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind); 487 size_t arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind); 488 void arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits); 489 void arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, 490 size_t size, size_t flags); 491 void arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 492 size_t size); 493 void arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, 494 size_t size, size_t flags); 495 void arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 496 size_t binind); 497 void arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, 498 size_t runind, size_t binind, size_t flags); 499 void arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 500 size_t unzeroed); 501 bool arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes); 502 bool arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes); 503 bool arena_prof_accum(arena_t *arena, uint64_t accumbytes); 504 size_t arena_ptr_small_binind_get(const void *ptr, size_t mapbits); 505 size_t arena_bin_index(arena_t *arena, arena_bin_t *bin); 506 unsigned arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, 507 const void *ptr); 508 prof_ctx_t *arena_prof_ctx_get(const void *ptr); 509 void arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx); 510 void *arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache); 511 size_t arena_salloc(const void *ptr, bool demote); 512 void arena_dalloc(arena_chunk_t *chunk, void *ptr, bool try_tcache); 513 #endif 514 515 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_)) 516 # ifdef JEMALLOC_ARENA_INLINE_A 517 JEMALLOC_INLINE size_t 518 small_size2bin_compute(size_t size) 519 { 520 #if (NTBINS != 0) 521 if (size <= (ZU(1) << LG_TINY_MAXCLASS)) { 522 size_t lg_tmin = LG_TINY_MAXCLASS - NTBINS + 1; 523 size_t lg_ceil = lg_floor(pow2_ceil(size)); 524 return (lg_ceil < lg_tmin ? 0 : lg_ceil - lg_tmin); 525 } else 526 #endif 527 { 528 size_t x = lg_floor((size<<1)-1); 529 size_t shift = (x < LG_SIZE_CLASS_GROUP + LG_QUANTUM) ? 0 : 530 x - (LG_SIZE_CLASS_GROUP + LG_QUANTUM); 531 size_t grp = shift << LG_SIZE_CLASS_GROUP; 532 533 size_t lg_delta = (x < LG_SIZE_CLASS_GROUP + LG_QUANTUM + 1) 534 ? LG_QUANTUM : x - LG_SIZE_CLASS_GROUP - 1; 535 536 size_t delta_inverse_mask = ZI(-1) << lg_delta; 537 size_t mod = ((((size-1) & delta_inverse_mask) >> lg_delta)) & 538 ((ZU(1) << LG_SIZE_CLASS_GROUP) - 1); 539 540 size_t bin = NTBINS + grp + mod; 541 return (bin); 542 } 543 } 544 545 JEMALLOC_ALWAYS_INLINE size_t 546 small_size2bin_lookup(size_t size) 547 { 548 549 assert(size <= LOOKUP_MAXCLASS); 550 { 551 size_t ret = ((size_t)(small_size2bin_tab[(size-1) >> 552 LG_TINY_MIN])); 553 assert(ret == small_size2bin_compute(size)); 554 return (ret); 555 } 556 } 557 558 JEMALLOC_ALWAYS_INLINE size_t 559 small_size2bin(size_t size) 560 { 561 562 assert(size > 0); 563 if (size <= LOOKUP_MAXCLASS) 564 return (small_size2bin_lookup(size)); 565 else 566 return (small_size2bin_compute(size)); 567 } 568 569 JEMALLOC_INLINE size_t 570 small_bin2size_compute(size_t binind) 571 { 572 #if (NTBINS > 0) 573 if (binind < NTBINS) 574 return (ZU(1) << (LG_TINY_MAXCLASS - NTBINS + 1 + binind)); 575 else 576 #endif 577 { 578 size_t reduced_binind = binind - NTBINS; 579 size_t grp = reduced_binind >> LG_SIZE_CLASS_GROUP; 580 size_t mod = reduced_binind & ((ZU(1) << LG_SIZE_CLASS_GROUP) - 581 1); 582 583 size_t grp_size_mask = ~((!!grp)-1); 584 size_t grp_size = ((ZU(1) << (LG_QUANTUM + 585 (LG_SIZE_CLASS_GROUP-1))) << grp) & grp_size_mask; 586 587 size_t shift = (grp == 0) ? 1 : grp; 588 size_t lg_delta = shift + (LG_QUANTUM-1); 589 size_t mod_size = (mod+1) << lg_delta; 590 591 size_t usize = grp_size + mod_size; 592 return (usize); 593 } 594 } 595 596 JEMALLOC_ALWAYS_INLINE size_t 597 small_bin2size_lookup(size_t binind) 598 { 599 600 assert(binind < NBINS); 601 { 602 size_t ret = ((size_t)(small_bin2size_tab[binind])); 603 assert(ret == small_bin2size_compute(binind)); 604 return (ret); 605 } 606 } 607 608 JEMALLOC_ALWAYS_INLINE size_t 609 small_bin2size(size_t binind) 610 { 611 612 return (small_bin2size_lookup(binind)); 613 } 614 615 JEMALLOC_ALWAYS_INLINE size_t 616 small_s2u_compute(size_t size) 617 { 618 #if (NTBINS > 0) 619 if (size <= (ZU(1) << LG_TINY_MAXCLASS)) { 620 size_t lg_tmin = LG_TINY_MAXCLASS - NTBINS + 1; 621 size_t lg_ceil = lg_floor(pow2_ceil(size)); 622 return (lg_ceil < lg_tmin ? (ZU(1) << lg_tmin) : 623 (ZU(1) << lg_ceil)); 624 } else 625 #endif 626 { 627 size_t x = lg_floor((size<<1)-1); 628 size_t lg_delta = (x < LG_SIZE_CLASS_GROUP + LG_QUANTUM + 1) 629 ? LG_QUANTUM : x - LG_SIZE_CLASS_GROUP - 1; 630 size_t delta = ZU(1) << lg_delta; 631 size_t delta_mask = delta - 1; 632 size_t usize = (size + delta_mask) & ~delta_mask; 633 return (usize); 634 } 635 } 636 637 JEMALLOC_ALWAYS_INLINE size_t 638 small_s2u_lookup(size_t size) 639 { 640 size_t ret = (small_bin2size(small_size2bin(size))); 641 642 assert(ret == small_s2u_compute(size)); 643 return (ret); 644 } 645 646 JEMALLOC_ALWAYS_INLINE size_t 647 small_s2u(size_t size) 648 { 649 650 assert(size > 0); 651 if (size <= LOOKUP_MAXCLASS) 652 return (small_s2u_lookup(size)); 653 else 654 return (small_s2u_compute(size)); 655 } 656 # endif /* JEMALLOC_ARENA_INLINE_A */ 657 658 # ifdef JEMALLOC_ARENA_INLINE_B 659 JEMALLOC_ALWAYS_INLINE arena_chunk_map_t * 660 arena_mapp_get(arena_chunk_t *chunk, size_t pageind) 661 { 662 663 assert(pageind >= map_bias); 664 assert(pageind < chunk_npages); 665 666 return (&chunk->map[pageind-map_bias]); 667 } 668 669 JEMALLOC_ALWAYS_INLINE size_t * 670 arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind) 671 { 672 673 return (&arena_mapp_get(chunk, pageind)->bits); 674 } 675 676 JEMALLOC_ALWAYS_INLINE size_t 677 arena_mapbitsp_read(size_t *mapbitsp) 678 { 679 680 return (*mapbitsp); 681 } 682 683 JEMALLOC_ALWAYS_INLINE size_t 684 arena_mapbits_get(arena_chunk_t *chunk, size_t pageind) 685 { 686 687 return (arena_mapbitsp_read(arena_mapbitsp_get(chunk, pageind))); 688 } 689 690 JEMALLOC_ALWAYS_INLINE size_t 691 arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, size_t pageind) 692 { 693 size_t mapbits; 694 695 mapbits = arena_mapbits_get(chunk, pageind); 696 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 697 return (mapbits & ~PAGE_MASK); 698 } 699 700 JEMALLOC_ALWAYS_INLINE size_t 701 arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind) 702 { 703 size_t mapbits; 704 705 mapbits = arena_mapbits_get(chunk, pageind); 706 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 707 (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)); 708 return (mapbits & ~PAGE_MASK); 709 } 710 711 JEMALLOC_ALWAYS_INLINE size_t 712 arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind) 713 { 714 size_t mapbits; 715 716 mapbits = arena_mapbits_get(chunk, pageind); 717 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 718 CHUNK_MAP_ALLOCATED); 719 return (mapbits >> LG_PAGE); 720 } 721 722 JEMALLOC_ALWAYS_INLINE size_t 723 arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind) 724 { 725 size_t mapbits; 726 size_t binind; 727 728 mapbits = arena_mapbits_get(chunk, pageind); 729 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 730 assert(binind < NBINS || binind == BININD_INVALID); 731 return (binind); 732 } 733 734 JEMALLOC_ALWAYS_INLINE size_t 735 arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind) 736 { 737 size_t mapbits; 738 739 mapbits = arena_mapbits_get(chunk, pageind); 740 return (mapbits & CHUNK_MAP_DIRTY); 741 } 742 743 JEMALLOC_ALWAYS_INLINE size_t 744 arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind) 745 { 746 size_t mapbits; 747 748 mapbits = arena_mapbits_get(chunk, pageind); 749 return (mapbits & CHUNK_MAP_UNZEROED); 750 } 751 752 JEMALLOC_ALWAYS_INLINE size_t 753 arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind) 754 { 755 size_t mapbits; 756 757 mapbits = arena_mapbits_get(chunk, pageind); 758 return (mapbits & CHUNK_MAP_LARGE); 759 } 760 761 JEMALLOC_ALWAYS_INLINE size_t 762 arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind) 763 { 764 size_t mapbits; 765 766 mapbits = arena_mapbits_get(chunk, pageind); 767 return (mapbits & CHUNK_MAP_ALLOCATED); 768 } 769 770 JEMALLOC_ALWAYS_INLINE void 771 arena_mapbitsp_write(size_t *mapbitsp, size_t mapbits) 772 { 773 774 *mapbitsp = mapbits; 775 } 776 777 JEMALLOC_ALWAYS_INLINE void 778 arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size, 779 size_t flags) 780 { 781 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 782 783 assert((size & PAGE_MASK) == 0); 784 assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0); 785 assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags); 786 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags); 787 } 788 789 JEMALLOC_ALWAYS_INLINE void 790 arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind, 791 size_t size) 792 { 793 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 794 size_t mapbits = arena_mapbitsp_read(mapbitsp); 795 796 assert((size & PAGE_MASK) == 0); 797 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0); 798 arena_mapbitsp_write(mapbitsp, size | (mapbits & PAGE_MASK)); 799 } 800 801 JEMALLOC_ALWAYS_INLINE void 802 arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size, 803 size_t flags) 804 { 805 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 806 size_t mapbits = arena_mapbitsp_read(mapbitsp); 807 size_t unzeroed; 808 809 assert((size & PAGE_MASK) == 0); 810 assert((flags & CHUNK_MAP_DIRTY) == flags); 811 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 812 arena_mapbitsp_write(mapbitsp, size | CHUNK_MAP_BININD_INVALID | flags 813 | unzeroed | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED); 814 } 815 816 JEMALLOC_ALWAYS_INLINE void 817 arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind, 818 size_t binind) 819 { 820 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 821 size_t mapbits = arena_mapbitsp_read(mapbitsp); 822 823 assert(binind <= BININD_INVALID); 824 assert(arena_mapbits_large_size_get(chunk, pageind) == PAGE); 825 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_BININD_MASK) | 826 (binind << CHUNK_MAP_BININD_SHIFT)); 827 } 828 829 JEMALLOC_ALWAYS_INLINE void 830 arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, size_t runind, 831 size_t binind, size_t flags) 832 { 833 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 834 size_t mapbits = arena_mapbitsp_read(mapbitsp); 835 size_t unzeroed; 836 837 assert(binind < BININD_INVALID); 838 assert(pageind - runind >= map_bias); 839 assert((flags & CHUNK_MAP_DIRTY) == flags); 840 unzeroed = mapbits & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */ 841 arena_mapbitsp_write(mapbitsp, (runind << LG_PAGE) | (binind << 842 CHUNK_MAP_BININD_SHIFT) | flags | unzeroed | CHUNK_MAP_ALLOCATED); 843 } 844 845 JEMALLOC_ALWAYS_INLINE void 846 arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind, 847 size_t unzeroed) 848 { 849 size_t *mapbitsp = arena_mapbitsp_get(chunk, pageind); 850 size_t mapbits = arena_mapbitsp_read(mapbitsp); 851 852 arena_mapbitsp_write(mapbitsp, (mapbits & ~CHUNK_MAP_UNZEROED) | 853 unzeroed); 854 } 855 856 JEMALLOC_INLINE bool 857 arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes) 858 { 859 860 cassert(config_prof); 861 assert(prof_interval != 0); 862 863 arena->prof_accumbytes += accumbytes; 864 if (arena->prof_accumbytes >= prof_interval) { 865 arena->prof_accumbytes -= prof_interval; 866 return (true); 867 } 868 return (false); 869 } 870 871 JEMALLOC_INLINE bool 872 arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes) 873 { 874 875 cassert(config_prof); 876 877 if (prof_interval == 0) 878 return (false); 879 return (arena_prof_accum_impl(arena, accumbytes)); 880 } 881 882 JEMALLOC_INLINE bool 883 arena_prof_accum(arena_t *arena, uint64_t accumbytes) 884 { 885 886 cassert(config_prof); 887 888 if (prof_interval == 0) 889 return (false); 890 891 { 892 bool ret; 893 894 malloc_mutex_lock(&arena->lock); 895 ret = arena_prof_accum_impl(arena, accumbytes); 896 malloc_mutex_unlock(&arena->lock); 897 return (ret); 898 } 899 } 900 901 JEMALLOC_ALWAYS_INLINE size_t 902 arena_ptr_small_binind_get(const void *ptr, size_t mapbits) 903 { 904 size_t binind; 905 906 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT; 907 908 if (config_debug) { 909 arena_chunk_t *chunk; 910 arena_t *arena; 911 size_t pageind; 912 size_t actual_mapbits; 913 arena_run_t *run; 914 arena_bin_t *bin; 915 size_t actual_binind; 916 arena_bin_info_t *bin_info; 917 918 assert(binind != BININD_INVALID); 919 assert(binind < NBINS); 920 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 921 arena = chunk->arena; 922 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 923 actual_mapbits = arena_mapbits_get(chunk, pageind); 924 assert(mapbits == actual_mapbits); 925 assert(arena_mapbits_large_get(chunk, pageind) == 0); 926 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 927 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind - 928 (actual_mapbits >> LG_PAGE)) << LG_PAGE)); 929 bin = run->bin; 930 actual_binind = bin - arena->bins; 931 assert(binind == actual_binind); 932 bin_info = &arena_bin_info[actual_binind]; 933 assert(((uintptr_t)ptr - ((uintptr_t)run + 934 (uintptr_t)bin_info->reg0_offset)) % bin_info->reg_interval 935 == 0); 936 } 937 938 return (binind); 939 } 940 # endif /* JEMALLOC_ARENA_INLINE_B */ 941 942 # ifdef JEMALLOC_ARENA_INLINE_C 943 JEMALLOC_INLINE size_t 944 arena_bin_index(arena_t *arena, arena_bin_t *bin) 945 { 946 size_t binind = bin - arena->bins; 947 assert(binind < NBINS); 948 return (binind); 949 } 950 951 JEMALLOC_INLINE unsigned 952 arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, const void *ptr) 953 { 954 unsigned shift, diff, regind; 955 size_t interval; 956 957 /* 958 * Freeing a pointer lower than region zero can cause assertion 959 * failure. 960 */ 961 assert((uintptr_t)ptr >= (uintptr_t)run + 962 (uintptr_t)bin_info->reg0_offset); 963 964 /* 965 * Avoid doing division with a variable divisor if possible. Using 966 * actual division here can reduce allocator throughput by over 20%! 967 */ 968 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - 969 bin_info->reg0_offset); 970 971 /* Rescale (factor powers of 2 out of the numerator and denominator). */ 972 interval = bin_info->reg_interval; 973 shift = jemalloc_ffs(interval) - 1; 974 diff >>= shift; 975 interval >>= shift; 976 977 if (interval == 1) { 978 /* The divisor was a power of 2. */ 979 regind = diff; 980 } else { 981 /* 982 * To divide by a number D that is not a power of two we 983 * multiply by (2^21 / D) and then right shift by 21 positions. 984 * 985 * X / D 986 * 987 * becomes 988 * 989 * (X * interval_invs[D - 3]) >> SIZE_INV_SHIFT 990 * 991 * We can omit the first three elements, because we never 992 * divide by 0, and 1 and 2 are both powers of two, which are 993 * handled above. 994 */ 995 #define SIZE_INV_SHIFT ((sizeof(unsigned) << 3) - LG_RUN_MAXREGS) 996 #define SIZE_INV(s) (((1U << SIZE_INV_SHIFT) / (s)) + 1) 997 static const unsigned interval_invs[] = { 998 SIZE_INV(3), 999 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 1000 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 1001 SIZE_INV(12), SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 1002 SIZE_INV(16), SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 1003 SIZE_INV(20), SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 1004 SIZE_INV(24), SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 1005 SIZE_INV(28), SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 1006 }; 1007 1008 if (interval <= ((sizeof(interval_invs) / sizeof(unsigned)) + 1009 2)) { 1010 regind = (diff * interval_invs[interval - 3]) >> 1011 SIZE_INV_SHIFT; 1012 } else 1013 regind = diff / interval; 1014 #undef SIZE_INV 1015 #undef SIZE_INV_SHIFT 1016 } 1017 assert(diff == regind * interval); 1018 assert(regind < bin_info->nregs); 1019 1020 return (regind); 1021 } 1022 1023 JEMALLOC_INLINE prof_ctx_t * 1024 arena_prof_ctx_get(const void *ptr) 1025 { 1026 prof_ctx_t *ret; 1027 arena_chunk_t *chunk; 1028 size_t pageind, mapbits; 1029 1030 cassert(config_prof); 1031 assert(ptr != NULL); 1032 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1033 1034 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 1035 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1036 mapbits = arena_mapbits_get(chunk, pageind); 1037 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0); 1038 if ((mapbits & CHUNK_MAP_LARGE) == 0) 1039 ret = (prof_ctx_t *)(uintptr_t)1U; 1040 else 1041 ret = arena_mapp_get(chunk, pageind)->prof_ctx; 1042 1043 return (ret); 1044 } 1045 1046 JEMALLOC_INLINE void 1047 arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx) 1048 { 1049 arena_chunk_t *chunk; 1050 size_t pageind; 1051 1052 cassert(config_prof); 1053 assert(ptr != NULL); 1054 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1055 1056 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 1057 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1058 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1059 1060 if (arena_mapbits_large_get(chunk, pageind) != 0) 1061 arena_mapp_get(chunk, pageind)->prof_ctx = ctx; 1062 } 1063 1064 JEMALLOC_ALWAYS_INLINE void * 1065 arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache) 1066 { 1067 tcache_t *tcache; 1068 1069 assert(size != 0); 1070 assert(size <= arena_maxclass); 1071 1072 if (size <= SMALL_MAXCLASS) { 1073 if (try_tcache && (tcache = tcache_get(true)) != NULL) 1074 return (tcache_alloc_small(tcache, size, zero)); 1075 else { 1076 return (arena_malloc_small(choose_arena(arena), size, 1077 zero)); 1078 } 1079 } else { 1080 /* 1081 * Initialize tcache after checking size in order to avoid 1082 * infinite recursion during tcache initialization. 1083 */ 1084 if (try_tcache && size <= tcache_maxclass && (tcache = 1085 tcache_get(true)) != NULL) 1086 return (tcache_alloc_large(tcache, size, zero)); 1087 else { 1088 return (arena_malloc_large(choose_arena(arena), size, 1089 zero)); 1090 } 1091 } 1092 } 1093 1094 /* Return the size of the allocation pointed to by ptr. */ 1095 JEMALLOC_ALWAYS_INLINE size_t 1096 arena_salloc(const void *ptr, bool demote) 1097 { 1098 size_t ret; 1099 arena_chunk_t *chunk; 1100 size_t pageind, binind; 1101 1102 assert(ptr != NULL); 1103 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1104 1105 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 1106 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1107 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1108 binind = arena_mapbits_binind_get(chunk, pageind); 1109 if (binind == BININD_INVALID || (config_prof && demote == false && 1110 arena_mapbits_large_get(chunk, pageind) != 0)) { 1111 /* 1112 * Large allocation. In the common case (demote == true), and 1113 * as this is an inline function, most callers will only end up 1114 * looking at binind to determine that ptr is a small 1115 * allocation. 1116 */ 1117 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1118 ret = arena_mapbits_large_size_get(chunk, pageind); 1119 assert(ret != 0); 1120 assert(pageind + (ret>>LG_PAGE) <= chunk_npages); 1121 assert(ret == PAGE || arena_mapbits_large_size_get(chunk, 1122 pageind+(ret>>LG_PAGE)-1) == 0); 1123 assert(binind == arena_mapbits_binind_get(chunk, 1124 pageind+(ret>>LG_PAGE)-1)); 1125 assert(arena_mapbits_dirty_get(chunk, pageind) == 1126 arena_mapbits_dirty_get(chunk, pageind+(ret>>LG_PAGE)-1)); 1127 } else { 1128 /* Small allocation (possibly promoted to a large object). */ 1129 assert(arena_mapbits_large_get(chunk, pageind) != 0 || 1130 arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk, 1131 pageind)) == binind); 1132 ret = small_bin2size(binind); 1133 } 1134 1135 return (ret); 1136 } 1137 1138 JEMALLOC_ALWAYS_INLINE void 1139 arena_dalloc(arena_chunk_t *chunk, void *ptr, bool try_tcache) 1140 { 1141 size_t pageind, mapbits; 1142 tcache_t *tcache; 1143 1144 assert(ptr != NULL); 1145 assert(CHUNK_ADDR2BASE(ptr) != ptr); 1146 1147 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE; 1148 mapbits = arena_mapbits_get(chunk, pageind); 1149 assert(arena_mapbits_allocated_get(chunk, pageind) != 0); 1150 if ((mapbits & CHUNK_MAP_LARGE) == 0) { 1151 /* Small allocation. */ 1152 if (try_tcache && (tcache = tcache_get(false)) != NULL) { 1153 size_t binind; 1154 1155 binind = arena_ptr_small_binind_get(ptr, mapbits); 1156 tcache_dalloc_small(tcache, ptr, binind); 1157 } else 1158 arena_dalloc_small(chunk->arena, chunk, ptr, pageind); 1159 } else { 1160 size_t size = arena_mapbits_large_size_get(chunk, pageind); 1161 1162 assert(((uintptr_t)ptr & PAGE_MASK) == 0); 1163 1164 if (try_tcache && size <= tcache_maxclass && (tcache = 1165 tcache_get(false)) != NULL) { 1166 tcache_dalloc_large(tcache, ptr, size); 1167 } else 1168 arena_dalloc_large(chunk->arena, chunk, ptr); 1169 } 1170 } 1171 # endif /* JEMALLOC_ARENA_INLINE_C */ 1172 #endif 1173 1174 #endif /* JEMALLOC_H_INLINES */ 1175 /******************************************************************************/ 1176
