1/* Malloc implementation for multiple threads without lock contention.
2 Copyright (C) 2001-2022 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public License as
7 published by the Free Software Foundation; either version 2.1 of the
8 License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; see the file COPYING.LIB. If
17 not, see <https://www.gnu.org/licenses/>. */
18
19#include <stdbool.h>
20
21#if HAVE_TUNABLES
22# define TUNABLE_NAMESPACE malloc
23#endif
24#include <elf/dl-tunables.h>
25
26/* Compile-time constants. */
27
28#define HEAP_MIN_SIZE (32 * 1024)
29#ifndef HEAP_MAX_SIZE
30# ifdef DEFAULT_MMAP_THRESHOLD_MAX
31# define HEAP_MAX_SIZE (2 * DEFAULT_MMAP_THRESHOLD_MAX)
32# else
33# define HEAP_MAX_SIZE (1024 * 1024) /* must be a power of two */
34# endif
35#endif
36
37/* HEAP_MIN_SIZE and HEAP_MAX_SIZE limit the size of mmap()ed heaps
38 that are dynamically created for multi-threaded programs. The
39 maximum size must be a power of two, for fast determination of
40 which heap belongs to a chunk. It should be much larger than the
41 mmap threshold, so that requests with a size just below that
42 threshold can be fulfilled without creating too many heaps. */
43
44/* When huge pages are used to create new arenas, the maximum and minumum
45 size are based on the runtime defined huge page size. */
46
47static inline size_t
48heap_min_size (void)
49{
50#if HAVE_TUNABLES
51 return mp_.hp_pagesize == 0 ? HEAP_MIN_SIZE : mp_.hp_pagesize;
52#else
53 return HEAP_MIN_SIZE;
54#endif
55}
56
57static inline size_t
58heap_max_size (void)
59{
60#if HAVE_TUNABLES
61 return mp_.hp_pagesize == 0 ? HEAP_MAX_SIZE : mp_.hp_pagesize * 4;
62#else
63 return HEAP_MAX_SIZE;
64#endif
65}
66
67/***************************************************************************/
68
69#define top(ar_ptr) ((ar_ptr)->top)
70
71/* A heap is a single contiguous memory region holding (coalesceable)
72 malloc_chunks. It is allocated with mmap() and always starts at an
73 address aligned to HEAP_MAX_SIZE. */
74
75typedef struct _heap_info
76{
77 mstate ar_ptr; /* Arena for this heap. */
78 struct _heap_info *prev; /* Previous heap. */
79 size_t size; /* Current size in bytes. */
80 size_t mprotect_size; /* Size in bytes that has been mprotected
81 PROT_READ|PROT_WRITE. */
82 size_t pagesize; /* Page size used when allocating the arena. */
83 /* Make sure the following data is properly aligned, particularly
84 that sizeof (heap_info) + 2 * SIZE_SZ is a multiple of
85 MALLOC_ALIGNMENT. */
86 char pad[-3 * SIZE_SZ & MALLOC_ALIGN_MASK];
87} heap_info;
88
89/* Get a compile-time error if the heap_info padding is not correct
90 to make alignment work as expected in sYSMALLOc. */
91extern int sanity_check_heap_info_alignment[(sizeof (heap_info)
92 + 2 * SIZE_SZ) % MALLOC_ALIGNMENT
93 ? -1 : 1];
94
95/* Thread specific data. */
96
97static __thread mstate thread_arena attribute_tls_model_ie;
98
99/* Arena free list. free_list_lock synchronizes access to the
100 free_list variable below, and the next_free and attached_threads
101 members of struct malloc_state objects. No other locks must be
102 acquired after free_list_lock has been acquired. */
103
104__libc_lock_define_initialized (static, free_list_lock);
105#if IS_IN (libc)
106static size_t narenas = 1;
107#endif
108static mstate free_list;
109
110/* list_lock prevents concurrent writes to the next member of struct
111 malloc_state objects.
112
113 Read access to the next member is supposed to synchronize with the
114 atomic_write_barrier and the write to the next member in
115 _int_new_arena. This suffers from data races; see the FIXME
116 comments in _int_new_arena and reused_arena.
117
118 list_lock also prevents concurrent forks. At the time list_lock is
119 acquired, no arena lock must have been acquired, but it is
120 permitted to acquire arena locks subsequently, while list_lock is
121 acquired. */
122__libc_lock_define_initialized (static, list_lock);
123
124/* Already initialized? */
125static bool __malloc_initialized = false;
126
127/**************************************************************************/
128
129
130/* arena_get() acquires an arena and locks the corresponding mutex.
131 First, try the one last locked successfully by this thread. (This
132 is the common case and handled with a macro for speed.) Then, loop
133 once over the circularly linked list of arenas. If no arena is
134 readily available, create a new one. In this latter case, `size'
135 is just a hint as to how much memory will be required immediately
136 in the new arena. */
137
138#define arena_get(ptr, size) do { \
139 ptr = thread_arena; \
140 arena_lock (ptr, size); \
141 } while (0)
142
143#define arena_lock(ptr, size) do { \
144 if (ptr) \
145 __libc_lock_lock (ptr->mutex); \
146 else \
147 ptr = arena_get2 ((size), NULL); \
148 } while (0)
149
150/* find the heap and corresponding arena for a given ptr */
151
152static inline heap_info *
153heap_for_ptr (void *ptr)
154{
155 size_t max_size = heap_max_size ();
156 return PTR_ALIGN_DOWN (ptr, max_size);
157}
158
159static inline struct malloc_state *
160arena_for_chunk (mchunkptr ptr)
161{
162 return chunk_main_arena (ptr) ? &main_arena : heap_for_ptr (ptr)->ar_ptr;
163}
164
165
166/**************************************************************************/
167
168/* atfork support. */
169
170/* The following three functions are called around fork from a
171 multi-threaded process. We do not use the general fork handler
172 mechanism to make sure that our handlers are the last ones being
173 called, so that other fork handlers can use the malloc
174 subsystem. */
175
176void
177__malloc_fork_lock_parent (void)
178{
179 if (!__malloc_initialized)
180 return;
181
182 /* We do not acquire free_list_lock here because we completely
183 reconstruct free_list in __malloc_fork_unlock_child. */
184
185 __libc_lock_lock (list_lock);
186
187 for (mstate ar_ptr = &main_arena;; )
188 {
189 __libc_lock_lock (ar_ptr->mutex);
190 ar_ptr = ar_ptr->next;
191 if (ar_ptr == &main_arena)
192 break;
193 }
194}
195
196void
197__malloc_fork_unlock_parent (void)
198{
199 if (!__malloc_initialized)
200 return;
201
202 for (mstate ar_ptr = &main_arena;; )
203 {
204 __libc_lock_unlock (ar_ptr->mutex);
205 ar_ptr = ar_ptr->next;
206 if (ar_ptr == &main_arena)
207 break;
208 }
209 __libc_lock_unlock (list_lock);
210}
211
212void
213__malloc_fork_unlock_child (void)
214{
215 if (!__malloc_initialized)
216 return;
217
218 /* Push all arenas to the free list, except thread_arena, which is
219 attached to the current thread. */
220 __libc_lock_init (free_list_lock);
221 if (thread_arena != NULL)
222 thread_arena->attached_threads = 1;
223 free_list = NULL;
224 for (mstate ar_ptr = &main_arena;; )
225 {
226 __libc_lock_init (ar_ptr->mutex);
227 if (ar_ptr != thread_arena)
228 {
229 /* This arena is no longer attached to any thread. */
230 ar_ptr->attached_threads = 0;
231 ar_ptr->next_free = free_list;
232 free_list = ar_ptr;
233 }
234 ar_ptr = ar_ptr->next;
235 if (ar_ptr == &main_arena)
236 break;
237 }
238
239 __libc_lock_init (list_lock);
240}
241
242#if HAVE_TUNABLES
243# define TUNABLE_CALLBACK_FNDECL(__name, __type) \
244static inline int do_ ## __name (__type value); \
245static void \
246TUNABLE_CALLBACK (__name) (tunable_val_t *valp) \
247{ \
248 __type value = (__type) (valp)->numval; \
249 do_ ## __name (value); \
250}
251
252TUNABLE_CALLBACK_FNDECL (set_mmap_threshold, size_t)
253TUNABLE_CALLBACK_FNDECL (set_mmaps_max, int32_t)
254TUNABLE_CALLBACK_FNDECL (set_top_pad, size_t)
255TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t)
256TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t)
257TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t)
258TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t)
259#if USE_TCACHE
260TUNABLE_CALLBACK_FNDECL (set_tcache_max, size_t)
261TUNABLE_CALLBACK_FNDECL (set_tcache_count, size_t)
262TUNABLE_CALLBACK_FNDECL (set_tcache_unsorted_limit, size_t)
263#endif
264TUNABLE_CALLBACK_FNDECL (set_mxfast, size_t)
265TUNABLE_CALLBACK_FNDECL (set_hugetlb, size_t)
266#else
267/* Initialization routine. */
268#include <string.h>
269extern char **_environ;
270
271static char *
272next_env_entry (char ***position)
273{
274 char **current = *position;
275 char *result = NULL;
276
277 while (*current != NULL)
278 {
279 if (__builtin_expect ((*current)[0] == 'M', 0)
280 && (*current)[1] == 'A'
281 && (*current)[2] == 'L'
282 && (*current)[3] == 'L'
283 && (*current)[4] == 'O'
284 && (*current)[5] == 'C'
285 && (*current)[6] == '_')
286 {
287 result = &(*current)[7];
288
289 /* Save current position for next visit. */
290 *position = ++current;
291
292 break;
293 }
294
295 ++current;
296 }
297
298 return result;
299}
300#endif
301
302
303#ifdef SHARED
304extern struct dl_open_hook *_dl_open_hook;
305libc_hidden_proto (_dl_open_hook);
306#endif
307
308#if USE_TCACHE
309static void tcache_key_initialize (void);
310#endif
311
312static void
313ptmalloc_init (void)
314{
315 if (__malloc_initialized)
316 return;
317
318 __malloc_initialized = true;
319
320#if USE_TCACHE
321 tcache_key_initialize ();
322#endif
323
324#ifdef USE_MTAG
325 if ((TUNABLE_GET_FULL (glibc, mem, tagging, int32_t, NULL) & 1) != 0)
326 {
327 /* If the tunable says that we should be using tagged memory
328 and that morecore does not support tagged regions, then
329 disable it. */
330 if (__MTAG_SBRK_UNTAGGED)
331 __always_fail_morecore = true;
332
333 mtag_enabled = true;
334 mtag_mmap_flags = __MTAG_MMAP_FLAGS;
335 }
336#endif
337
338#if defined SHARED && IS_IN (libc)
339 /* In case this libc copy is in a non-default namespace, never use
340 brk. Likewise if dlopened from statically linked program. The
341 generic sbrk implementation also enforces this, but it is not
342 used on Hurd. */
343 if (!__libc_initial)
344 __always_fail_morecore = true;
345#endif
346
347 thread_arena = &main_arena;
348
349 malloc_init_state (av: &main_arena);
350
351#if HAVE_TUNABLES
352 TUNABLE_GET (top_pad, size_t, TUNABLE_CALLBACK (set_top_pad));
353 TUNABLE_GET (perturb, int32_t, TUNABLE_CALLBACK (set_perturb_byte));
354 TUNABLE_GET (mmap_threshold, size_t, TUNABLE_CALLBACK (set_mmap_threshold));
355 TUNABLE_GET (trim_threshold, size_t, TUNABLE_CALLBACK (set_trim_threshold));
356 TUNABLE_GET (mmap_max, int32_t, TUNABLE_CALLBACK (set_mmaps_max));
357 TUNABLE_GET (arena_max, size_t, TUNABLE_CALLBACK (set_arena_max));
358 TUNABLE_GET (arena_test, size_t, TUNABLE_CALLBACK (set_arena_test));
359# if USE_TCACHE
360 TUNABLE_GET (tcache_max, size_t, TUNABLE_CALLBACK (set_tcache_max));
361 TUNABLE_GET (tcache_count, size_t, TUNABLE_CALLBACK (set_tcache_count));
362 TUNABLE_GET (tcache_unsorted_limit, size_t,
363 TUNABLE_CALLBACK (set_tcache_unsorted_limit));
364# endif
365 TUNABLE_GET (mxfast, size_t, TUNABLE_CALLBACK (set_mxfast));
366 TUNABLE_GET (hugetlb, size_t, TUNABLE_CALLBACK (set_hugetlb));
367 if (mp_.hp_pagesize > 0)
368 /* Force mmap for main arena instead of sbrk, so hugepages are explicitly
369 used. */
370 __always_fail_morecore = true;
371#else
372 if (__glibc_likely (_environ != NULL))
373 {
374 char **runp = _environ;
375 char *envline;
376
377 while (__builtin_expect ((envline = next_env_entry (&runp)) != NULL,
378 0))
379 {
380 size_t len = strcspn (envline, "=");
381
382 if (envline[len] != '=')
383 /* This is a "MALLOC_" variable at the end of the string
384 without a '=' character. Ignore it since otherwise we
385 will access invalid memory below. */
386 continue;
387
388 switch (len)
389 {
390 case 8:
391 if (!__builtin_expect (__libc_enable_secure, 0))
392 {
393 if (memcmp (envline, "TOP_PAD_", 8) == 0)
394 __libc_mallopt (M_TOP_PAD, atoi (&envline[9]));
395 else if (memcmp (envline, "PERTURB_", 8) == 0)
396 __libc_mallopt (M_PERTURB, atoi (&envline[9]));
397 }
398 break;
399 case 9:
400 if (!__builtin_expect (__libc_enable_secure, 0))
401 {
402 if (memcmp (envline, "MMAP_MAX_", 9) == 0)
403 __libc_mallopt (M_MMAP_MAX, atoi (&envline[10]));
404 else if (memcmp (envline, "ARENA_MAX", 9) == 0)
405 __libc_mallopt (M_ARENA_MAX, atoi (&envline[10]));
406 }
407 break;
408 case 10:
409 if (!__builtin_expect (__libc_enable_secure, 0))
410 {
411 if (memcmp (envline, "ARENA_TEST", 10) == 0)
412 __libc_mallopt (M_ARENA_TEST, atoi (&envline[11]));
413 }
414 break;
415 case 15:
416 if (!__builtin_expect (__libc_enable_secure, 0))
417 {
418 if (memcmp (envline, "TRIM_THRESHOLD_", 15) == 0)
419 __libc_mallopt (M_TRIM_THRESHOLD, atoi (&envline[16]));
420 else if (memcmp (envline, "MMAP_THRESHOLD_", 15) == 0)
421 __libc_mallopt (M_MMAP_THRESHOLD, atoi (&envline[16]));
422 }
423 break;
424 default:
425 break;
426 }
427 }
428 }
429#endif
430}
431
432/* Managing heaps and arenas (for concurrent threads) */
433
434#if MALLOC_DEBUG > 1
435
436/* Print the complete contents of a single heap to stderr. */
437
438static void
439dump_heap (heap_info *heap)
440{
441 char *ptr;
442 mchunkptr p;
443
444 fprintf (stderr, "Heap %p, size %10lx:\n", heap, (long) heap->size);
445 ptr = (heap->ar_ptr != (mstate) (heap + 1)) ?
446 (char *) (heap + 1) : (char *) (heap + 1) + sizeof (struct malloc_state);
447 p = (mchunkptr) (((unsigned long) ptr + MALLOC_ALIGN_MASK) &
448 ~MALLOC_ALIGN_MASK);
449 for (;; )
450 {
451 fprintf (stderr, "chunk %p size %10lx", p, (long) chunksize_nomask(p));
452 if (p == top (heap->ar_ptr))
453 {
454 fprintf (stderr, " (top)\n");
455 break;
456 }
457 else if (chunksize_nomask(p) == (0 | PREV_INUSE))
458 {
459 fprintf (stderr, " (fence)\n");
460 break;
461 }
462 fprintf (stderr, "\n");
463 p = next_chunk (p);
464 }
465}
466#endif /* MALLOC_DEBUG > 1 */
467
468/* If consecutive mmap (0, HEAP_MAX_SIZE << 1, ...) calls return decreasing
469 addresses as opposed to increasing, new_heap would badly fragment the
470 address space. In that case remember the second HEAP_MAX_SIZE part
471 aligned to HEAP_MAX_SIZE from last mmap (0, HEAP_MAX_SIZE << 1, ...)
472 call (if it is already aligned) and try to reuse it next time. We need
473 no locking for it, as kernel ensures the atomicity for us - worst case
474 we'll call mmap (addr, HEAP_MAX_SIZE, ...) for some value of addr in
475 multiple threads, but only one will succeed. */
476static char *aligned_heap_area;
477
478/* Create a new heap. size is automatically rounded up to a multiple
479 of the page size. */
480
481static heap_info *
482alloc_new_heap (size_t size, size_t top_pad, size_t pagesize,
483 int mmap_flags)
484{
485 char *p1, *p2;
486 unsigned long ul;
487 heap_info *h;
488 size_t min_size = heap_min_size ();
489 size_t max_size = heap_max_size ();
490
491 if (size + top_pad < min_size)
492 size = min_size;
493 else if (size + top_pad <= max_size)
494 size += top_pad;
495 else if (size > max_size)
496 return 0;
497 else
498 size = max_size;
499 size = ALIGN_UP (size, pagesize);
500
501 /* A memory region aligned to a multiple of max_size is needed.
502 No swap space needs to be reserved for the following large
503 mapping (on Linux, this is the case for all non-writable mappings
504 anyway). */
505 p2 = MAP_FAILED;
506 if (aligned_heap_area)
507 {
508 p2 = (char *) MMAP (aligned_heap_area, max_size, PROT_NONE, mmap_flags);
509 aligned_heap_area = NULL;
510 if (p2 != MAP_FAILED && ((unsigned long) p2 & (max_size - 1)))
511 {
512 __munmap (addr: p2, len: max_size);
513 p2 = MAP_FAILED;
514 }
515 }
516 if (p2 == MAP_FAILED)
517 {
518 p1 = (char *) MMAP (0, max_size << 1, PROT_NONE, mmap_flags);
519 if (p1 != MAP_FAILED)
520 {
521 p2 = (char *) (((unsigned long) p1 + (max_size - 1))
522 & ~(max_size - 1));
523 ul = p2 - p1;
524 if (ul)
525 __munmap (addr: p1, len: ul);
526 else
527 aligned_heap_area = p2 + max_size;
528 __munmap (addr: p2 + max_size, len: max_size - ul);
529 }
530 else
531 {
532 /* Try to take the chance that an allocation of only max_size
533 is already aligned. */
534 p2 = (char *) MMAP (0, max_size, PROT_NONE, mmap_flags);
535 if (p2 == MAP_FAILED)
536 return 0;
537
538 if ((unsigned long) p2 & (max_size - 1))
539 {
540 __munmap (addr: p2, len: max_size);
541 return 0;
542 }
543 }
544 }
545 if (__mprotect (addr: p2, len: size, mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0)
546 {
547 __munmap (addr: p2, len: max_size);
548 return 0;
549 }
550
551 madvise_thp (p: p2, size);
552
553 h = (heap_info *) p2;
554 h->size = size;
555 h->mprotect_size = size;
556 h->pagesize = pagesize;
557 LIBC_PROBE (memory_heap_new, 2, h, h->size);
558 return h;
559}
560
561static heap_info *
562new_heap (size_t size, size_t top_pad)
563{
564#if HAVE_TUNABLES
565 if (__glibc_unlikely (mp_.hp_pagesize != 0))
566 {
567 /* MAP_NORESERVE is not used for huge pages because some kernel may
568 not reserve the mmap region and a subsequent access may trigger
569 a SIGBUS if there is no free pages in the pool. */
570 heap_info *h = alloc_new_heap (size, top_pad, pagesize: mp_.hp_pagesize,
571 mmap_flags: mp_.hp_flags);
572 if (h != NULL)
573 return h;
574 }
575#endif
576 return alloc_new_heap (size, top_pad, GLRO (dl_pagesize), MAP_NORESERVE);
577}
578
579/* Grow a heap. size is automatically rounded up to a
580 multiple of the page size. */
581
582static int
583grow_heap (heap_info *h, long diff)
584{
585 size_t pagesize = h->pagesize;
586 size_t max_size = heap_max_size ();
587 long new_size;
588
589 diff = ALIGN_UP (diff, pagesize);
590 new_size = (long) h->size + diff;
591 if ((unsigned long) new_size > (unsigned long) max_size)
592 return -1;
593
594 if ((unsigned long) new_size > h->mprotect_size)
595 {
596 if (__mprotect (addr: (char *) h + h->mprotect_size,
597 len: (unsigned long) new_size - h->mprotect_size,
598 mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0)
599 return -2;
600
601 h->mprotect_size = new_size;
602 }
603
604 h->size = new_size;
605 LIBC_PROBE (memory_heap_more, 2, h, h->size);
606 return 0;
607}
608
609/* Shrink a heap. */
610
611static int
612shrink_heap (heap_info *h, long diff)
613{
614 long new_size;
615
616 new_size = (long) h->size - diff;
617 if (new_size < (long) sizeof (*h))
618 return -1;
619
620 /* Try to re-map the extra heap space freshly to save memory, and make it
621 inaccessible. See malloc-sysdep.h to know when this is true. */
622 if (__glibc_unlikely (check_may_shrink_heap ()))
623 {
624 if ((char *) MMAP ((char *) h + new_size, diff, PROT_NONE,
625 MAP_FIXED) == (char *) MAP_FAILED)
626 return -2;
627
628 h->mprotect_size = new_size;
629 }
630 else
631 __madvise (addr: (char *) h + new_size, len: diff, MADV_DONTNEED);
632 /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/
633
634 h->size = new_size;
635 LIBC_PROBE (memory_heap_less, 2, h, h->size);
636 return 0;
637}
638
639/* Delete a heap. */
640
641static int
642heap_trim (heap_info *heap, size_t pad)
643{
644 mstate ar_ptr = heap->ar_ptr;
645 mchunkptr top_chunk = top (ar_ptr), p;
646 heap_info *prev_heap;
647 long new_size, top_size, top_area, extra, prev_size, misalign;
648 size_t max_size = heap_max_size ();
649
650 /* Can this heap go away completely? */
651 while (top_chunk == chunk_at_offset (heap, sizeof (*heap)))
652 {
653 prev_heap = heap->prev;
654 prev_size = prev_heap->size - (MINSIZE - 2 * SIZE_SZ);
655 p = chunk_at_offset (prev_heap, prev_size);
656 /* fencepost must be properly aligned. */
657 misalign = ((long) p) & MALLOC_ALIGN_MASK;
658 p = chunk_at_offset (prev_heap, prev_size - misalign);
659 assert (chunksize_nomask (p) == (0 | PREV_INUSE)); /* must be fencepost */
660 p = prev_chunk (p);
661 new_size = chunksize (p) + (MINSIZE - 2 * SIZE_SZ) + misalign;
662 assert (new_size > 0 && new_size < (long) (2 * MINSIZE));
663 if (!prev_inuse (p))
664 new_size += prev_size (p);
665 assert (new_size > 0 && new_size < max_size);
666 if (new_size + (max_size - prev_heap->size) < pad + MINSIZE
667 + heap->pagesize)
668 break;
669 ar_ptr->system_mem -= heap->size;
670 LIBC_PROBE (memory_heap_free, 2, heap, heap->size);
671 if ((char *) heap + max_size == aligned_heap_area)
672 aligned_heap_area = NULL;
673 __munmap (addr: heap, len: max_size);
674 heap = prev_heap;
675 if (!prev_inuse (p)) /* consolidate backward */
676 {
677 p = prev_chunk (p);
678 unlink_chunk (av: ar_ptr, p);
679 }
680 assert (((unsigned long) ((char *) p + new_size) & (heap->pagesize - 1))
681 == 0);
682 assert (((char *) p + new_size) == ((char *) heap + heap->size));
683 top (ar_ptr) = top_chunk = p;
684 set_head (top_chunk, new_size | PREV_INUSE);
685 /*check_chunk(ar_ptr, top_chunk);*/
686 }
687
688 /* Uses similar logic for per-thread arenas as the main arena with systrim
689 and _int_free by preserving the top pad and rounding down to the nearest
690 page. */
691 top_size = chunksize (top_chunk);
692 if ((unsigned long)(top_size) <
693 (unsigned long)(mp_.trim_threshold))
694 return 0;
695
696 top_area = top_size - MINSIZE - 1;
697 if (top_area < 0 || (size_t) top_area <= pad)
698 return 0;
699
700 /* Release in pagesize units and round down to the nearest page. */
701 extra = ALIGN_DOWN(top_area - pad, heap->pagesize);
702 if (extra == 0)
703 return 0;
704
705 /* Try to shrink. */
706 if (shrink_heap (h: heap, diff: extra) != 0)
707 return 0;
708
709 ar_ptr->system_mem -= extra;
710
711 /* Success. Adjust top accordingly. */
712 set_head (top_chunk, (top_size - extra) | PREV_INUSE);
713 /*check_chunk(ar_ptr, top_chunk);*/
714 return 1;
715}
716
717/* Create a new arena with initial size "size". */
718
719#if IS_IN (libc)
720/* If REPLACED_ARENA is not NULL, detach it from this thread. Must be
721 called while free_list_lock is held. */
722static void
723detach_arena (mstate replaced_arena)
724{
725 if (replaced_arena != NULL)
726 {
727 assert (replaced_arena->attached_threads > 0);
728 /* The current implementation only detaches from main_arena in
729 case of allocation failure. This means that it is likely not
730 beneficial to put the arena on free_list even if the
731 reference count reaches zero. */
732 --replaced_arena->attached_threads;
733 }
734}
735
736static mstate
737_int_new_arena (size_t size)
738{
739 mstate a;
740 heap_info *h;
741 char *ptr;
742 unsigned long misalign;
743
744 h = new_heap (size + (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT),
745 mp_.top_pad);
746 if (!h)
747 {
748 /* Maybe size is too large to fit in a single heap. So, just try
749 to create a minimally-sized arena and let _int_malloc() attempt
750 to deal with the large request via mmap_chunk(). */
751 h = new_heap (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT, mp_.top_pad);
752 if (!h)
753 return 0;
754 }
755 a = h->ar_ptr = (mstate) (h + 1);
756 malloc_init_state (a);
757 a->attached_threads = 1;
758 /*a->next = NULL;*/
759 a->system_mem = a->max_system_mem = h->size;
760
761 /* Set up the top chunk, with proper alignment. */
762 ptr = (char *) (a + 1);
763 misalign = (unsigned long) chunk2mem (ptr) & MALLOC_ALIGN_MASK;
764 if (misalign > 0)
765 ptr += MALLOC_ALIGNMENT - misalign;
766 top (a) = (mchunkptr) ptr;
767 set_head (top (a), (((char *) h + h->size) - ptr) | PREV_INUSE);
768
769 LIBC_PROBE (memory_arena_new, 2, a, size);
770 mstate replaced_arena = thread_arena;
771 thread_arena = a;
772 __libc_lock_init (a->mutex);
773
774 __libc_lock_lock (list_lock);
775
776 /* Add the new arena to the global list. */
777 a->next = main_arena.next;
778 /* FIXME: The barrier is an attempt to synchronize with read access
779 in reused_arena, which does not acquire list_lock while
780 traversing the list. */
781 atomic_write_barrier ();
782 main_arena.next = a;
783
784 __libc_lock_unlock (list_lock);
785
786 __libc_lock_lock (free_list_lock);
787 detach_arena (replaced_arena);
788 __libc_lock_unlock (free_list_lock);
789
790 /* Lock this arena. NB: Another thread may have been attached to
791 this arena because the arena is now accessible from the
792 main_arena.next list and could have been picked by reused_arena.
793 This can only happen for the last arena created (before the arena
794 limit is reached). At this point, some arena has to be attached
795 to two threads. We could acquire the arena lock before list_lock
796 to make it less likely that reused_arena picks this new arena,
797 but this could result in a deadlock with
798 __malloc_fork_lock_parent. */
799
800 __libc_lock_lock (a->mutex);
801
802 return a;
803}
804
805
806/* Remove an arena from free_list. */
807static mstate
808get_free_list (void)
809{
810 mstate replaced_arena = thread_arena;
811 mstate result = free_list;
812 if (result != NULL)
813 {
814 __libc_lock_lock (free_list_lock);
815 result = free_list;
816 if (result != NULL)
817 {
818 free_list = result->next_free;
819
820 /* The arena will be attached to this thread. */
821 assert (result->attached_threads == 0);
822 result->attached_threads = 1;
823
824 detach_arena (replaced_arena);
825 }
826 __libc_lock_unlock (free_list_lock);
827
828 if (result != NULL)
829 {
830 LIBC_PROBE (memory_arena_reuse_free_list, 1, result);
831 __libc_lock_lock (result->mutex);
832 thread_arena = result;
833 }
834 }
835
836 return result;
837}
838
839/* Remove the arena from the free list (if it is present).
840 free_list_lock must have been acquired by the caller. */
841static void
842remove_from_free_list (mstate arena)
843{
844 mstate *previous = &free_list;
845 for (mstate p = free_list; p != NULL; p = p->next_free)
846 {
847 assert (p->attached_threads == 0);
848 if (p == arena)
849 {
850 /* Remove the requested arena from the list. */
851 *previous = p->next_free;
852 break;
853 }
854 else
855 previous = &p->next_free;
856 }
857}
858
859/* Lock and return an arena that can be reused for memory allocation.
860 Avoid AVOID_ARENA as we have already failed to allocate memory in
861 it and it is currently locked. */
862static mstate
863reused_arena (mstate avoid_arena)
864{
865 mstate result;
866 /* FIXME: Access to next_to_use suffers from data races. */
867 static mstate next_to_use;
868 if (next_to_use == NULL)
869 next_to_use = &main_arena;
870
871 /* Iterate over all arenas (including those linked from
872 free_list). */
873 result = next_to_use;
874 do
875 {
876 if (!__libc_lock_trylock (result->mutex))
877 goto out;
878
879 /* FIXME: This is a data race, see _int_new_arena. */
880 result = result->next;
881 }
882 while (result != next_to_use);
883
884 /* Avoid AVOID_ARENA as we have already failed to allocate memory
885 in that arena and it is currently locked. */
886 if (result == avoid_arena)
887 result = result->next;
888
889 /* No arena available without contention. Wait for the next in line. */
890 LIBC_PROBE (memory_arena_reuse_wait, 3, &result->mutex, result, avoid_arena);
891 __libc_lock_lock (result->mutex);
892
893out:
894 /* Attach the arena to the current thread. */
895 {
896 /* Update the arena thread attachment counters. */
897 mstate replaced_arena = thread_arena;
898 __libc_lock_lock (free_list_lock);
899 detach_arena (replaced_arena);
900
901 /* We may have picked up an arena on the free list. We need to
902 preserve the invariant that no arena on the free list has a
903 positive attached_threads counter (otherwise,
904 arena_thread_freeres cannot use the counter to determine if the
905 arena needs to be put on the free list). We unconditionally
906 remove the selected arena from the free list. The caller of
907 reused_arena checked the free list and observed it to be empty,
908 so the list is very short. */
909 remove_from_free_list (result);
910
911 ++result->attached_threads;
912
913 __libc_lock_unlock (free_list_lock);
914 }
915
916 LIBC_PROBE (memory_arena_reuse, 2, result, avoid_arena);
917 thread_arena = result;
918 next_to_use = result->next;
919
920 return result;
921}
922
923static mstate
924arena_get2 (size_t size, mstate avoid_arena)
925{
926 mstate a;
927
928 static size_t narenas_limit;
929
930 a = get_free_list ();
931 if (a == NULL)
932 {
933 /* Nothing immediately available, so generate a new arena. */
934 if (narenas_limit == 0)
935 {
936 if (mp_.arena_max != 0)
937 narenas_limit = mp_.arena_max;
938 else if (narenas > mp_.arena_test)
939 {
940 int n = __get_nprocs_sched ();
941
942 if (n >= 1)
943 narenas_limit = NARENAS_FROM_NCORES (n);
944 else
945 /* We have no information about the system. Assume two
946 cores. */
947 narenas_limit = NARENAS_FROM_NCORES (2);
948 }
949 }
950 repeat:;
951 size_t n = narenas;
952 /* NB: the following depends on the fact that (size_t)0 - 1 is a
953 very large number and that the underflow is OK. If arena_max
954 is set the value of arena_test is irrelevant. If arena_test
955 is set but narenas is not yet larger or equal to arena_test
956 narenas_limit is 0. There is no possibility for narenas to
957 be too big for the test to always fail since there is not
958 enough address space to create that many arenas. */
959 if (__glibc_unlikely (n <= narenas_limit - 1))
960 {
961 if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n))
962 goto repeat;
963 a = _int_new_arena (size);
964 if (__glibc_unlikely (a == NULL))
965 catomic_decrement (&narenas);
966 }
967 else
968 a = reused_arena (avoid_arena);
969 }
970 return a;
971}
972
973/* If we don't have the main arena, then maybe the failure is due to running
974 out of mmapped areas, so we can try allocating on the main arena.
975 Otherwise, it is likely that sbrk() has failed and there is still a chance
976 to mmap(), so try one of the other arenas. */
977static mstate
978arena_get_retry (mstate ar_ptr, size_t bytes)
979{
980 LIBC_PROBE (memory_arena_retry, 2, bytes, ar_ptr);
981 if (ar_ptr != &main_arena)
982 {
983 __libc_lock_unlock (ar_ptr->mutex);
984 ar_ptr = &main_arena;
985 __libc_lock_lock (ar_ptr->mutex);
986 }
987 else
988 {
989 __libc_lock_unlock (ar_ptr->mutex);
990 ar_ptr = arena_get2 (bytes, ar_ptr);
991 }
992
993 return ar_ptr;
994}
995#endif
996
997void
998__malloc_arena_thread_freeres (void)
999{
1000 /* Shut down the thread cache first. This could deallocate data for
1001 the thread arena, so do this before we put the arena on the free
1002 list. */
1003 tcache_thread_shutdown ();
1004
1005 mstate a = thread_arena;
1006 thread_arena = NULL;
1007
1008 if (a != NULL)
1009 {
1010 __libc_lock_lock (free_list_lock);
1011 /* If this was the last attached thread for this arena, put the
1012 arena on the free list. */
1013 assert (a->attached_threads > 0);
1014 if (--a->attached_threads == 0)
1015 {
1016 a->next_free = free_list;
1017 free_list = a;
1018 }
1019 __libc_lock_unlock (free_list_lock);
1020 }
1021}
1022
1023/*
1024 * Local variables:
1025 * c-basic-offset: 2
1026 * End:
1027 */
1028

source code of glibc/malloc/arena.c