1 /******************************************************************************/
2 #ifdef JEMALLOC_H_TYPES
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.
10 * We use binary fixed point math for overhead computations, where the binary
11 * point is implicitly RUN_BFP bits to the left.
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:
19 * (RUN_MAX_OVRHD / (reg_interval << (3+RUN_BFP))
22 /* \/ Implicit binary fixed point. */
23 #define RUN_MAX_OVRHD 0x0000003dU
24 #define RUN_MAX_OVRHD_RELAX 0x00001800U
26 /* Maximum number of regions in one run. */
27 #define LG_RUN_MAXREGS 11
28 #define RUN_MAXREGS (1U << LG_RUN_MAXREGS)
31 * Minimum redzone size. Redzones may be larger than this if necessary to
32 * preserve region alignment.
34 #define REDZONE_MINSIZE 16
37 * The minimum ratio of active:dirty pages per arena is computed as:
39 * (nactive >> opt_lg_dirty_mult) >= ndirty
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.
44 #define LG_DIRTY_MULT_DEFAULT 3
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;
53 #endif /* JEMALLOC_H_TYPES */
54 /******************************************************************************/
55 #ifdef JEMALLOC_H_STRUCTS
57 /* Each element of the chunk map corresponds to one page within the chunk. */
58 struct arena_chunk_map_s {
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.
70 * Linkage for run trees. There are two disjoint uses:
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.
76 rb_node(arena_chunk_map_t) rb_link;
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.
83 ql_elm(arena_chunk_map_t) ql_link;
86 /* Profile counters, used for large object runs. */
89 }; /* union { ... }; */
93 * Run address (or size) and various flags are stored together. The bit
94 * layout looks like (assuming 32-bit system):
96 * ???????? ???????? ????nnnn nnnndula
98 * ? : Unallocated: Run address for first/last pages, unset for internal
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.
108 * Following are example bit patterns for the three types of runs.
110 * p : run page offset
112 * n : binind for size class; large objects set these to BININD_INVALID
113 * except for promoted allocations (see prof_promote)
120 * Unallocated (clean):
121 * ssssssss ssssssss ssss++++ ++++du-a
122 * xxxxxxxx xxxxxxxx xxxxxxxx xxxx-Uxx
123 * ssssssss ssssssss ssss++++ ++++dU-a
125 * Unallocated (dirty):
126 * ssssssss ssssssss ssss++++ ++++D--a
127 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx
128 * ssssssss ssssssss ssss++++ ++++D--a
131 * pppppppp pppppppp ppppnnnn nnnnd--A
132 * pppppppp pppppppp ppppnnnn nnnn---A
133 * pppppppp pppppppp ppppnnnn nnnnd--A
136 * ssssssss ssssssss ssss++++ ++++D-LA
137 * xxxxxxxx xxxxxxxx xxxxxxxx xxxxxxxx
138 * -------- -------- ----++++ ++++D-LA
140 * Large (sampled, size <= PAGE):
141 * ssssssss ssssssss ssssnnnn nnnnD-LA
143 * Large (not sampled, size == PAGE):
144 * ssssssss ssssssss ssss++++ ++++D-LA
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
159 typedef rb_tree(arena_chunk_map_t) arena_avail_tree_t;
160 typedef rb_tree(arena_chunk_map_t) arena_run_tree_t;
162 /* Arena chunk header. */
163 struct arena_chunk_s {
164 /* Arena that owns the chunk. */
167 /* Linkage for tree of arena chunks that contain dirty runs. */
168 rb_node(arena_chunk_t) dirty_link;
170 /* Number of dirty pages. */
173 /* Number of available runs. */
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.
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).
190 arena_chunk_map_t map[1]; /* Dynamically sized. */
192 typedef rb_tree(arena_chunk_t) arena_chunk_tree_t;
195 /* Bin this run is associated with. */
198 /* Index of next region that has never been allocated, or nregs. */
201 /* Number of free regions in run. */
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.
210 * Each run has the following layout:
212 * /--------------------\
213 * | arena_run_t header |
215 * bitmap_offset | bitmap |
217 * ctx0_offset | ctx map |
219 * |--------------------|
221 * reg0_offset | region 0 |
223 * |--------------------| \
225 * | region 1 | > reg_interval
227 * |--------------------|
231 * |--------------------|
235 * |--------------------|
237 * \--------------------/
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.
243 struct arena_bin_info_s {
244 /* Size of regions in a run for this bin's size class. */
250 /* Interval between regions (reg_size + (redzone_size << 1)). */
253 /* Total size of a run for this bin's size class. */
256 /* Total number of regions in a run for this bin's size class. */
260 * Offset of first bitmap_t element in a run header for this bin's size
263 uint32_t bitmap_offset;
266 * Metadata used to manipulate bitmaps for runs associated with this
269 bitmap_info_t bitmap_info;
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).
275 uint32_t ctx0_offset;
277 /* Offset of first region in a run for this bin's size class. */
278 uint32_t reg0_offset;
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
291 * Current run being used to service allocations of this bin's size
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.
303 arena_run_tree_t runs;
305 /* Bin statistics. */
306 malloc_bin_stats_t stats;
310 /* This arena's index within the arenas array. */
314 * Number of threads currently assigned to this arena. This field is
315 * protected by arenas_lock.
320 * There are three classes of arena operations from a locking
322 * 1) Thread asssignment (modifies nthreads) is protected by
324 * 2) Bin-related operations are protected by bin locks.
325 * 3) Chunk- and run-related operations are protected by this mutex.
331 * List of tcaches for extant threads associated with this arena.
332 * Stats from these are merged incrementally, and at exit.
334 ql_head(tcache_t) tcache_ql;
336 uint64_t prof_accumbytes;
340 /* Tree of dirty-page-containing chunks this arena manages. */
341 arena_chunk_tree_t chunks_dirty;
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.
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.
353 arena_chunk_t *spare;
355 /* Number of pages in active runs. */
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.
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.
375 * Size/address-ordered trees of this arena's available runs. The trees
376 * are used for first-best-fit run allocation.
378 arena_avail_tree_t runs_avail;
380 /* bins is used to store trees of free regions. */
381 arena_bin_t bins[NBINS];
384 #endif /* JEMALLOC_H_STRUCTS */
385 /******************************************************************************/
386 #ifdef JEMALLOC_H_EXTERNS
388 extern ssize_t opt_lg_dirty_mult;
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.
394 extern uint8_t const small_size2bin[];
395 #define SMALL_SIZE2BIN(s) (small_size2bin[(s-1) >> LG_TINY_MIN])
397 extern arena_bin_info_t arena_bin_info[NBINS];
399 /* Number of large size classes. */
400 #define nlclasses (chunk_npages - map_bias)
402 void arena_purge_all(arena_t *arena);
403 void arena_tcache_fill_small(arena_t *arena, tcache_bin_t *tbin,
404 size_t binind, uint64_t prof_accumbytes);
405 void arena_alloc_junk_small(void *ptr, arena_bin_info_t *bin_info,
407 void arena_dalloc_junk_small(void *ptr, arena_bin_info_t *bin_info);
408 void *arena_malloc_small(arena_t *arena, size_t size, bool zero);
409 void *arena_malloc_large(arena_t *arena, size_t size, bool zero);
410 void *arena_palloc(arena_t *arena, size_t size, size_t alignment, bool zero);
411 void arena_prof_promoted(const void *ptr, size_t size);
412 void arena_dalloc_bin_locked(arena_t *arena, arena_chunk_t *chunk, void *ptr,
413 arena_chunk_map_t *mapelm);
414 void arena_dalloc_bin(arena_t *arena, arena_chunk_t *chunk, void *ptr,
415 size_t pageind, arena_chunk_map_t *mapelm);
416 void arena_dalloc_small(arena_t *arena, arena_chunk_t *chunk, void *ptr,
418 void arena_dalloc_large_locked(arena_t *arena, arena_chunk_t *chunk,
420 void arena_dalloc_large(arena_t *arena, arena_chunk_t *chunk, void *ptr);
421 void *arena_ralloc_no_move(void *ptr, size_t oldsize, size_t size,
422 size_t extra, bool zero);
423 void *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);
426 dss_prec_t arena_dss_prec_get(arena_t *arena);
427 void arena_dss_prec_set(arena_t *arena, dss_prec_t dss_prec);
428 void 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);
431 bool arena_new(arena_t *arena, unsigned ind);
432 void arena_boot(void);
433 void arena_prefork(arena_t *arena);
434 void arena_postfork_parent(arena_t *arena);
435 void arena_postfork_child(arena_t *arena);
437 #endif /* JEMALLOC_H_EXTERNS */
438 /******************************************************************************/
439 #ifdef JEMALLOC_H_INLINES
441 #ifndef JEMALLOC_ENABLE_INLINE
442 arena_chunk_map_t *arena_mapp_get(arena_chunk_t *chunk, size_t pageind);
443 size_t *arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind);
444 size_t arena_mapbits_get(arena_chunk_t *chunk, size_t pageind);
445 size_t arena_mapbits_unallocated_size_get(arena_chunk_t *chunk,
447 size_t arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind);
448 size_t arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind);
449 size_t arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind);
450 size_t arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind);
451 size_t arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind);
452 size_t arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind);
453 size_t arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind);
454 void arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind,
455 size_t size, size_t flags);
456 void arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind,
458 void arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind,
459 size_t size, size_t flags);
460 void arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind,
462 void arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind,
463 size_t runind, size_t binind, size_t flags);
464 void arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind,
466 bool arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes);
467 bool arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes);
468 bool arena_prof_accum(arena_t *arena, uint64_t accumbytes);
469 size_t arena_ptr_small_binind_get(const void *ptr, size_t mapbits);
470 size_t arena_bin_index(arena_t *arena, arena_bin_t *bin);
471 unsigned arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info,
473 prof_ctx_t *arena_prof_ctx_get(const void *ptr);
474 void arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx);
475 void *arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache);
476 size_t arena_salloc(const void *ptr, bool demote);
477 void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr,
481 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_ARENA_C_))
482 # ifdef JEMALLOC_ARENA_INLINE_A
483 JEMALLOC_ALWAYS_INLINE arena_chunk_map_t *
484 arena_mapp_get(arena_chunk_t *chunk, size_t pageind)
487 assert(pageind >= map_bias);
488 assert(pageind < chunk_npages);
490 return (&chunk->map[pageind-map_bias]);
493 JEMALLOC_ALWAYS_INLINE size_t *
494 arena_mapbitsp_get(arena_chunk_t *chunk, size_t pageind)
497 return (&arena_mapp_get(chunk, pageind)->bits);
500 JEMALLOC_ALWAYS_INLINE size_t
501 arena_mapbits_get(arena_chunk_t *chunk, size_t pageind)
504 return (*arena_mapbitsp_get(chunk, pageind));
507 JEMALLOC_ALWAYS_INLINE size_t
508 arena_mapbits_unallocated_size_get(arena_chunk_t *chunk, size_t pageind)
512 mapbits = arena_mapbits_get(chunk, pageind);
513 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0);
514 return (mapbits & ~PAGE_MASK);
517 JEMALLOC_ALWAYS_INLINE size_t
518 arena_mapbits_large_size_get(arena_chunk_t *chunk, size_t pageind)
522 mapbits = arena_mapbits_get(chunk, pageind);
523 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) ==
524 (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED));
525 return (mapbits & ~PAGE_MASK);
528 JEMALLOC_ALWAYS_INLINE size_t
529 arena_mapbits_small_runind_get(arena_chunk_t *chunk, size_t pageind)
533 mapbits = arena_mapbits_get(chunk, pageind);
534 assert((mapbits & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) ==
535 CHUNK_MAP_ALLOCATED);
536 return (mapbits >> LG_PAGE);
539 JEMALLOC_ALWAYS_INLINE size_t
540 arena_mapbits_binind_get(arena_chunk_t *chunk, size_t pageind)
545 mapbits = arena_mapbits_get(chunk, pageind);
546 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT;
547 assert(binind < NBINS || binind == BININD_INVALID);
551 JEMALLOC_ALWAYS_INLINE size_t
552 arena_mapbits_dirty_get(arena_chunk_t *chunk, size_t pageind)
556 mapbits = arena_mapbits_get(chunk, pageind);
557 return (mapbits & CHUNK_MAP_DIRTY);
560 JEMALLOC_ALWAYS_INLINE size_t
561 arena_mapbits_unzeroed_get(arena_chunk_t *chunk, size_t pageind)
565 mapbits = arena_mapbits_get(chunk, pageind);
566 return (mapbits & CHUNK_MAP_UNZEROED);
569 JEMALLOC_ALWAYS_INLINE size_t
570 arena_mapbits_large_get(arena_chunk_t *chunk, size_t pageind)
574 mapbits = arena_mapbits_get(chunk, pageind);
575 return (mapbits & CHUNK_MAP_LARGE);
578 JEMALLOC_ALWAYS_INLINE size_t
579 arena_mapbits_allocated_get(arena_chunk_t *chunk, size_t pageind)
583 mapbits = arena_mapbits_get(chunk, pageind);
584 return (mapbits & CHUNK_MAP_ALLOCATED);
587 JEMALLOC_ALWAYS_INLINE void
588 arena_mapbits_unallocated_set(arena_chunk_t *chunk, size_t pageind, size_t size,
593 mapbitsp = arena_mapbitsp_get(chunk, pageind);
594 assert((size & PAGE_MASK) == 0);
595 assert((flags & ~CHUNK_MAP_FLAGS_MASK) == 0);
596 assert((flags & (CHUNK_MAP_DIRTY|CHUNK_MAP_UNZEROED)) == flags);
597 *mapbitsp = size | CHUNK_MAP_BININD_INVALID | flags;
600 JEMALLOC_ALWAYS_INLINE void
601 arena_mapbits_unallocated_size_set(arena_chunk_t *chunk, size_t pageind,
606 mapbitsp = arena_mapbitsp_get(chunk, pageind);
607 assert((size & PAGE_MASK) == 0);
608 assert((*mapbitsp & (CHUNK_MAP_LARGE|CHUNK_MAP_ALLOCATED)) == 0);
609 *mapbitsp = size | (*mapbitsp & PAGE_MASK);
612 JEMALLOC_ALWAYS_INLINE void
613 arena_mapbits_large_set(arena_chunk_t *chunk, size_t pageind, size_t size,
619 mapbitsp = arena_mapbitsp_get(chunk, pageind);
620 assert((size & PAGE_MASK) == 0);
621 assert((flags & CHUNK_MAP_DIRTY) == flags);
622 unzeroed = *mapbitsp & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */
623 *mapbitsp = size | CHUNK_MAP_BININD_INVALID | flags | unzeroed |
624 CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
627 JEMALLOC_ALWAYS_INLINE void
628 arena_mapbits_large_binind_set(arena_chunk_t *chunk, size_t pageind,
633 assert(binind <= BININD_INVALID);
634 mapbitsp = arena_mapbitsp_get(chunk, pageind);
635 assert(arena_mapbits_large_size_get(chunk, pageind) == PAGE);
636 *mapbitsp = (*mapbitsp & ~CHUNK_MAP_BININD_MASK) | (binind <<
637 CHUNK_MAP_BININD_SHIFT);
640 JEMALLOC_ALWAYS_INLINE void
641 arena_mapbits_small_set(arena_chunk_t *chunk, size_t pageind, size_t runind,
642 size_t binind, size_t flags)
647 assert(binind < BININD_INVALID);
648 mapbitsp = arena_mapbitsp_get(chunk, pageind);
649 assert(pageind - runind >= map_bias);
650 assert((flags & CHUNK_MAP_DIRTY) == flags);
651 unzeroed = *mapbitsp & CHUNK_MAP_UNZEROED; /* Preserve unzeroed. */
652 *mapbitsp = (runind << LG_PAGE) | (binind << CHUNK_MAP_BININD_SHIFT) |
653 flags | unzeroed | CHUNK_MAP_ALLOCATED;
656 JEMALLOC_ALWAYS_INLINE void
657 arena_mapbits_unzeroed_set(arena_chunk_t *chunk, size_t pageind,
662 mapbitsp = arena_mapbitsp_get(chunk, pageind);
663 *mapbitsp = (*mapbitsp & ~CHUNK_MAP_UNZEROED) | unzeroed;
667 arena_prof_accum_impl(arena_t *arena, uint64_t accumbytes)
670 cassert(config_prof);
671 assert(prof_interval != 0);
673 arena->prof_accumbytes += accumbytes;
674 if (arena->prof_accumbytes >= prof_interval) {
675 arena->prof_accumbytes -= prof_interval;
682 arena_prof_accum_locked(arena_t *arena, uint64_t accumbytes)
685 cassert(config_prof);
687 if (prof_interval == 0)
689 return (arena_prof_accum_impl(arena, accumbytes));
693 arena_prof_accum(arena_t *arena, uint64_t accumbytes)
696 cassert(config_prof);
698 if (prof_interval == 0)
704 malloc_mutex_lock(&arena->lock);
705 ret = arena_prof_accum_impl(arena, accumbytes);
706 malloc_mutex_unlock(&arena->lock);
711 JEMALLOC_ALWAYS_INLINE size_t
712 arena_ptr_small_binind_get(const void *ptr, size_t mapbits)
716 binind = (mapbits & CHUNK_MAP_BININD_MASK) >> CHUNK_MAP_BININD_SHIFT;
719 arena_chunk_t *chunk;
722 size_t actual_mapbits;
725 size_t actual_binind;
726 arena_bin_info_t *bin_info;
728 assert(binind != BININD_INVALID);
729 assert(binind < NBINS);
730 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
731 arena = chunk->arena;
732 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
733 actual_mapbits = arena_mapbits_get(chunk, pageind);
734 assert(mapbits == actual_mapbits);
735 assert(arena_mapbits_large_get(chunk, pageind) == 0);
736 assert(arena_mapbits_allocated_get(chunk, pageind) != 0);
737 run = (arena_run_t *)((uintptr_t)chunk + (uintptr_t)((pageind -
738 (actual_mapbits >> LG_PAGE)) << LG_PAGE));
740 actual_binind = bin - arena->bins;
741 assert(binind == actual_binind);
742 bin_info = &arena_bin_info[actual_binind];
743 assert(((uintptr_t)ptr - ((uintptr_t)run +
744 (uintptr_t)bin_info->reg0_offset)) % bin_info->reg_interval
750 # endif /* JEMALLOC_ARENA_INLINE_A */
752 # ifdef JEMALLOC_ARENA_INLINE_B
753 JEMALLOC_INLINE size_t
754 arena_bin_index(arena_t *arena, arena_bin_t *bin)
756 size_t binind = bin - arena->bins;
757 assert(binind < NBINS);
761 JEMALLOC_INLINE unsigned
762 arena_run_regind(arena_run_t *run, arena_bin_info_t *bin_info, const void *ptr)
764 unsigned shift, diff, regind;
768 * Freeing a pointer lower than region zero can cause assertion
771 assert((uintptr_t)ptr >= (uintptr_t)run +
772 (uintptr_t)bin_info->reg0_offset);
775 * Avoid doing division with a variable divisor if possible. Using
776 * actual division here can reduce allocator throughput by over 20%!
778 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run -
779 bin_info->reg0_offset);
781 /* Rescale (factor powers of 2 out of the numerator and denominator). */
782 interval = bin_info->reg_interval;
783 shift = ffs(interval) - 1;
788 /* The divisor was a power of 2. */
792 * To divide by a number D that is not a power of two we
793 * multiply by (2^21 / D) and then right shift by 21 positions.
799 * (X * interval_invs[D - 3]) >> SIZE_INV_SHIFT
801 * We can omit the first three elements, because we never
802 * divide by 0, and 1 and 2 are both powers of two, which are
805 #define SIZE_INV_SHIFT ((sizeof(unsigned) << 3) - LG_RUN_MAXREGS)
806 #define SIZE_INV(s) (((1U << SIZE_INV_SHIFT) / (s)) + 1)
807 static const unsigned interval_invs[] = {
809 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7),
810 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11),
811 SIZE_INV(12), SIZE_INV(13), SIZE_INV(14), SIZE_INV(15),
812 SIZE_INV(16), SIZE_INV(17), SIZE_INV(18), SIZE_INV(19),
813 SIZE_INV(20), SIZE_INV(21), SIZE_INV(22), SIZE_INV(23),
814 SIZE_INV(24), SIZE_INV(25), SIZE_INV(26), SIZE_INV(27),
815 SIZE_INV(28), SIZE_INV(29), SIZE_INV(30), SIZE_INV(31)
818 if (interval <= ((sizeof(interval_invs) / sizeof(unsigned)) +
820 regind = (diff * interval_invs[interval - 3]) >>
823 regind = diff / interval;
825 #undef SIZE_INV_SHIFT
827 assert(diff == regind * interval);
828 assert(regind < bin_info->nregs);
833 JEMALLOC_INLINE prof_ctx_t *
834 arena_prof_ctx_get(const void *ptr)
837 arena_chunk_t *chunk;
838 size_t pageind, mapbits;
840 cassert(config_prof);
842 assert(CHUNK_ADDR2BASE(ptr) != ptr);
844 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
845 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
846 mapbits = arena_mapbits_get(chunk, pageind);
847 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0);
848 if ((mapbits & CHUNK_MAP_LARGE) == 0) {
850 ret = (prof_ctx_t *)(uintptr_t)1U;
852 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk +
853 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) <<
855 size_t binind = arena_ptr_small_binind_get(ptr,
857 arena_bin_info_t *bin_info = &arena_bin_info[binind];
860 regind = arena_run_regind(run, bin_info, ptr);
861 ret = *(prof_ctx_t **)((uintptr_t)run +
862 bin_info->ctx0_offset + (regind *
863 sizeof(prof_ctx_t *)));
866 ret = arena_mapp_get(chunk, pageind)->prof_ctx;
872 arena_prof_ctx_set(const void *ptr, prof_ctx_t *ctx)
874 arena_chunk_t *chunk;
875 size_t pageind, mapbits;
877 cassert(config_prof);
879 assert(CHUNK_ADDR2BASE(ptr) != ptr);
881 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
882 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
883 mapbits = arena_mapbits_get(chunk, pageind);
884 assert((mapbits & CHUNK_MAP_ALLOCATED) != 0);
885 if ((mapbits & CHUNK_MAP_LARGE) == 0) {
886 if (prof_promote == false) {
887 arena_run_t *run = (arena_run_t *)((uintptr_t)chunk +
888 (uintptr_t)((pageind - (mapbits >> LG_PAGE)) <<
891 arena_bin_info_t *bin_info;
894 binind = arena_ptr_small_binind_get(ptr, mapbits);
895 bin_info = &arena_bin_info[binind];
896 regind = arena_run_regind(run, bin_info, ptr);
898 *((prof_ctx_t **)((uintptr_t)run + bin_info->ctx0_offset
899 + (regind * sizeof(prof_ctx_t *)))) = ctx;
901 assert((uintptr_t)ctx == (uintptr_t)1U);
903 arena_mapp_get(chunk, pageind)->prof_ctx = ctx;
906 JEMALLOC_ALWAYS_INLINE void *
907 arena_malloc(arena_t *arena, size_t size, bool zero, bool try_tcache)
912 assert(size <= arena_maxclass);
914 if (size <= SMALL_MAXCLASS) {
915 if (try_tcache && (tcache = tcache_get(true)) != NULL)
916 return (tcache_alloc_small(tcache, size, zero));
918 return (arena_malloc_small(choose_arena(arena), size,
923 * Initialize tcache after checking size in order to avoid
924 * infinite recursion during tcache initialization.
926 if (try_tcache && size <= tcache_maxclass && (tcache =
927 tcache_get(true)) != NULL)
928 return (tcache_alloc_large(tcache, size, zero));
930 return (arena_malloc_large(choose_arena(arena), size,
936 /* Return the size of the allocation pointed to by ptr. */
937 JEMALLOC_ALWAYS_INLINE size_t
938 arena_salloc(const void *ptr, bool demote)
941 arena_chunk_t *chunk;
942 size_t pageind, binind;
945 assert(CHUNK_ADDR2BASE(ptr) != ptr);
947 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
948 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
949 assert(arena_mapbits_allocated_get(chunk, pageind) != 0);
950 binind = arena_mapbits_binind_get(chunk, pageind);
951 if (binind == BININD_INVALID || (config_prof && demote == false &&
952 prof_promote && arena_mapbits_large_get(chunk, pageind) != 0)) {
954 * Large allocation. In the common case (demote == true), and
955 * as this is an inline function, most callers will only end up
956 * looking at binind to determine that ptr is a small
959 assert(((uintptr_t)ptr & PAGE_MASK) == 0);
960 ret = arena_mapbits_large_size_get(chunk, pageind);
962 assert(pageind + (ret>>LG_PAGE) <= chunk_npages);
963 assert(ret == PAGE || arena_mapbits_large_size_get(chunk,
964 pageind+(ret>>LG_PAGE)-1) == 0);
965 assert(binind == arena_mapbits_binind_get(chunk,
966 pageind+(ret>>LG_PAGE)-1));
967 assert(arena_mapbits_dirty_get(chunk, pageind) ==
968 arena_mapbits_dirty_get(chunk, pageind+(ret>>LG_PAGE)-1));
971 * Small allocation (possibly promoted to a large object due to
974 assert(arena_mapbits_large_get(chunk, pageind) != 0 ||
975 arena_ptr_small_binind_get(ptr, arena_mapbits_get(chunk,
976 pageind)) == binind);
977 ret = arena_bin_info[binind].reg_size;
983 JEMALLOC_ALWAYS_INLINE void
984 arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr, bool try_tcache)
986 size_t pageind, mapbits;
989 assert(arena != NULL);
990 assert(chunk->arena == arena);
992 assert(CHUNK_ADDR2BASE(ptr) != ptr);
994 pageind = ((uintptr_t)ptr - (uintptr_t)chunk) >> LG_PAGE;
995 mapbits = arena_mapbits_get(chunk, pageind);
996 assert(arena_mapbits_allocated_get(chunk, pageind) != 0);
997 if ((mapbits & CHUNK_MAP_LARGE) == 0) {
998 /* Small allocation. */
999 if (try_tcache && (tcache = tcache_get(false)) != NULL) {
1002 binind = arena_ptr_small_binind_get(ptr, mapbits);
1003 tcache_dalloc_small(tcache, ptr, binind);
1005 arena_dalloc_small(arena, chunk, ptr, pageind);
1007 size_t size = arena_mapbits_large_size_get(chunk, pageind);
1009 assert(((uintptr_t)ptr & PAGE_MASK) == 0);
1011 if (try_tcache && size <= tcache_maxclass && (tcache =
1012 tcache_get(false)) != NULL) {
1013 tcache_dalloc_large(tcache, ptr, size);
1015 arena_dalloc_large(arena, chunk, ptr);
1018 # endif /* JEMALLOC_ARENA_INLINE_B */
1021 #endif /* JEMALLOC_H_INLINES */
1022 /******************************************************************************/