1/* Malloc implementation for multiple threads without lock contention.
2 Copyright (C) 2001-2017 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Contributed by Wolfram Gloger <wg@malloc.de>, 2001.
5
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public License as
8 published by the Free Software Foundation; either version 2.1 of the
9 License, or (at your option) any later version.
10
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
15
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; see the file COPYING.LIB. If
18 not, see <http://www.gnu.org/licenses/>. */
19
20#include <stdbool.h>
21
22#if HAVE_TUNABLES
23# define TUNABLE_NAMESPACE malloc
24#endif
25#include <elf/dl-tunables.h>
26
27/* Compile-time constants. */
28
29#define HEAP_MIN_SIZE (32 * 1024)
30#ifndef HEAP_MAX_SIZE
31# ifdef DEFAULT_MMAP_THRESHOLD_MAX
32# define HEAP_MAX_SIZE (2 * DEFAULT_MMAP_THRESHOLD_MAX)
33# else
34# define HEAP_MAX_SIZE (1024 * 1024) /* must be a power of two */
35# endif
36#endif
37
38/* HEAP_MIN_SIZE and HEAP_MAX_SIZE limit the size of mmap()ed heaps
39 that are dynamically created for multi-threaded programs. The
40 maximum size must be a power of two, for fast determination of
41 which heap belongs to a chunk. It should be much larger than the
42 mmap threshold, so that requests with a size just below that
43 threshold can be fulfilled without creating too many heaps. */
44
45/***************************************************************************/
46
47#define top(ar_ptr) ((ar_ptr)->top)
48
49/* A heap is a single contiguous memory region holding (coalesceable)
50 malloc_chunks. It is allocated with mmap() and always starts at an
51 address aligned to HEAP_MAX_SIZE. */
52
53typedef struct _heap_info
54{
55 mstate ar_ptr; /* Arena for this heap. */
56 struct _heap_info *prev; /* Previous heap. */
57 size_t size; /* Current size in bytes. */
58 size_t mprotect_size; /* Size in bytes that has been mprotected
59 PROT_READ|PROT_WRITE. */
60 /* Make sure the following data is properly aligned, particularly
61 that sizeof (heap_info) + 2 * SIZE_SZ is a multiple of
62 MALLOC_ALIGNMENT. */
63 char pad[-6 * SIZE_SZ & MALLOC_ALIGN_MASK];
64} heap_info;
65
66/* Get a compile-time error if the heap_info padding is not correct
67 to make alignment work as expected in sYSMALLOc. */
68extern int sanity_check_heap_info_alignment[(sizeof (heap_info)
69 + 2 * SIZE_SZ) % MALLOC_ALIGNMENT
70 ? -1 : 1];
71
72/* Thread specific data. */
73
74static __thread mstate thread_arena attribute_tls_model_ie;
75
76/* Arena free list. free_list_lock synchronizes access to the
77 free_list variable below, and the next_free and attached_threads
78 members of struct malloc_state objects. No other locks must be
79 acquired after free_list_lock has been acquired. */
80
81__libc_lock_define_initialized (static, free_list_lock);
82static size_t narenas = 1;
83static mstate free_list;
84
85/* list_lock prevents concurrent writes to the next member of struct
86 malloc_state objects.
87
88 Read access to the next member is supposed to synchronize with the
89 atomic_write_barrier and the write to the next member in
90 _int_new_arena. This suffers from data races; see the FIXME
91 comments in _int_new_arena and reused_arena.
92
93 list_lock also prevents concurrent forks. At the time list_lock is
94 acquired, no arena lock must have been acquired, but it is
95 permitted to acquire arena locks subsequently, while list_lock is
96 acquired. */
97__libc_lock_define_initialized (static, list_lock);
98
99/* Already initialized? */
100int __malloc_initialized = -1;
101
102/**************************************************************************/
103
104
105/* arena_get() acquires an arena and locks the corresponding mutex.
106 First, try the one last locked successfully by this thread. (This
107 is the common case and handled with a macro for speed.) Then, loop
108 once over the circularly linked list of arenas. If no arena is
109 readily available, create a new one. In this latter case, `size'
110 is just a hint as to how much memory will be required immediately
111 in the new arena. */
112
113#define arena_get(ptr, size) do { \
114 ptr = thread_arena; \
115 arena_lock (ptr, size); \
116 } while (0)
117
118#define arena_lock(ptr, size) do { \
119 if (ptr && !arena_is_corrupt (ptr)) \
120 __libc_lock_lock (ptr->mutex); \
121 else \
122 ptr = arena_get2 ((size), NULL); \
123 } while (0)
124
125/* find the heap and corresponding arena for a given ptr */
126
127#define heap_for_ptr(ptr) \
128 ((heap_info *) ((unsigned long) (ptr) & ~(HEAP_MAX_SIZE - 1)))
129#define arena_for_chunk(ptr) \
130 (chunk_main_arena (ptr) ? &main_arena : heap_for_ptr (ptr)->ar_ptr)
131
132
133/**************************************************************************/
134
135/* atfork support. */
136
137/* The following three functions are called around fork from a
138 multi-threaded process. We do not use the general fork handler
139 mechanism to make sure that our handlers are the last ones being
140 called, so that other fork handlers can use the malloc
141 subsystem. */
142
143void
144internal_function
145__malloc_fork_lock_parent (void)
146{
147 if (__malloc_initialized < 1)
148 return;
149
150 /* We do not acquire free_list_lock here because we completely
151 reconstruct free_list in __malloc_fork_unlock_child. */
152
153 __libc_lock_lock (list_lock);
154
155 for (mstate ar_ptr = &main_arena;; )
156 {
157 __libc_lock_lock (ar_ptr->mutex);
158 ar_ptr = ar_ptr->next;
159 if (ar_ptr == &main_arena)
160 break;
161 }
162}
163
164void
165internal_function
166__malloc_fork_unlock_parent (void)
167{
168 if (__malloc_initialized < 1)
169 return;
170
171 for (mstate ar_ptr = &main_arena;; )
172 {
173 __libc_lock_unlock (ar_ptr->mutex);
174 ar_ptr = ar_ptr->next;
175 if (ar_ptr == &main_arena)
176 break;
177 }
178 __libc_lock_unlock (list_lock);
179}
180
181void
182internal_function
183__malloc_fork_unlock_child (void)
184{
185 if (__malloc_initialized < 1)
186 return;
187
188 /* Push all arenas to the free list, except thread_arena, which is
189 attached to the current thread. */
190 __libc_lock_init (free_list_lock);
191 if (thread_arena != NULL)
192 thread_arena->attached_threads = 1;
193 free_list = NULL;
194 for (mstate ar_ptr = &main_arena;; )
195 {
196 __libc_lock_init (ar_ptr->mutex);
197 if (ar_ptr != thread_arena)
198 {
199 /* This arena is no longer attached to any thread. */
200 ar_ptr->attached_threads = 0;
201 ar_ptr->next_free = free_list;
202 free_list = ar_ptr;
203 }
204 ar_ptr = ar_ptr->next;
205 if (ar_ptr == &main_arena)
206 break;
207 }
208
209 __libc_lock_init (list_lock);
210}
211
212#if HAVE_TUNABLES
213static inline int do_set_mallopt_check (int32_t value);
214void
215DL_TUNABLE_CALLBACK (set_mallopt_check) (tunable_val_t *valp)
216{
217 int32_t value = (int32_t) valp->numval;
218 do_set_mallopt_check (value);
219 if (check_action != 0)
220 __malloc_check_init ();
221}
222
223# define DL_TUNABLE_CALLBACK_FNDECL(__name, __type) \
224static inline int do_ ## __name (__type value); \
225void \
226DL_TUNABLE_CALLBACK (__name) (tunable_val_t *valp) \
227{ \
228 __type value = (__type) (valp)->numval; \
229 do_ ## __name (value); \
230}
231
232DL_TUNABLE_CALLBACK_FNDECL (set_mmap_threshold, size_t)
233DL_TUNABLE_CALLBACK_FNDECL (set_mmaps_max, int32_t)
234DL_TUNABLE_CALLBACK_FNDECL (set_top_pad, size_t)
235DL_TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t)
236DL_TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t)
237DL_TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t)
238DL_TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t)
239#else
240/* Initialization routine. */
241#include <string.h>
242extern char **_environ;
243
244static char *
245internal_function
246next_env_entry (char ***position)
247{
248 char **current = *position;
249 char *result = NULL;
250
251 while (*current != NULL)
252 {
253 if (__builtin_expect ((*current)[0] == 'M', 0)
254 && (*current)[1] == 'A'
255 && (*current)[2] == 'L'
256 && (*current)[3] == 'L'
257 && (*current)[4] == 'O'
258 && (*current)[5] == 'C'
259 && (*current)[6] == '_')
260 {
261 result = &(*current)[7];
262
263 /* Save current position for next visit. */
264 *position = ++current;
265
266 break;
267 }
268
269 ++current;
270 }
271
272 return result;
273}
274#endif
275
276
277#ifdef SHARED
278static void *
279__failing_morecore (ptrdiff_t d)
280{
281 return (void *) MORECORE_FAILURE;
282}
283
284extern struct dl_open_hook *_dl_open_hook;
285libc_hidden_proto (_dl_open_hook);
286#endif
287
288static void
289ptmalloc_init (void)
290{
291 if (__malloc_initialized >= 0)
292 return;
293
294 __malloc_initialized = 0;
295
296#ifdef SHARED
297 /* In case this libc copy is in a non-default namespace, never use brk.
298 Likewise if dlopened from statically linked program. */
299 Dl_info di;
300 struct link_map *l;
301
302 if (_dl_open_hook != NULL
303 || (_dl_addr (ptmalloc_init, &di, &l, NULL) != 0
304 && l->l_ns != LM_ID_BASE))
305 __morecore = __failing_morecore;
306#endif
307
308 thread_arena = &main_arena;
309
310#if HAVE_TUNABLES
311 /* Ensure initialization/consolidation and do it under a lock so that a
312 thread attempting to use the arena in parallel waits on us till we
313 finish. */
314 __libc_lock_lock (main_arena.mutex);
315 malloc_consolidate (&main_arena);
316
317 TUNABLE_SET_VAL_WITH_CALLBACK (check, NULL, set_mallopt_check);
318 TUNABLE_SET_VAL_WITH_CALLBACK (top_pad, NULL, set_top_pad);
319 TUNABLE_SET_VAL_WITH_CALLBACK (perturb, NULL, set_perturb_byte);
320 TUNABLE_SET_VAL_WITH_CALLBACK (mmap_threshold, NULL, set_mmap_threshold);
321 TUNABLE_SET_VAL_WITH_CALLBACK (trim_threshold, NULL, set_trim_threshold);
322 TUNABLE_SET_VAL_WITH_CALLBACK (mmap_max, NULL, set_mmaps_max);
323 TUNABLE_SET_VAL_WITH_CALLBACK (arena_max, NULL, set_arena_max);
324 TUNABLE_SET_VAL_WITH_CALLBACK (arena_test, NULL, set_arena_test);
325 __libc_lock_unlock (main_arena.mutex);
326#else
327 const char *s = NULL;
328 if (__glibc_likely (_environ != NULL))
329 {
330 char **runp = _environ;
331 char *envline;
332
333 while (__builtin_expect ((envline = next_env_entry (&runp)) != NULL,
334 0))
335 {
336 size_t len = strcspn (envline, "=");
337
338 if (envline[len] != '=')
339 /* This is a "MALLOC_" variable at the end of the string
340 without a '=' character. Ignore it since otherwise we
341 will access invalid memory below. */
342 continue;
343
344 switch (len)
345 {
346 case 6:
347 if (memcmp (envline, "CHECK_", 6) == 0)
348 s = &envline[7];
349 break;
350 case 8:
351 if (!__builtin_expect (__libc_enable_secure, 0))
352 {
353 if (memcmp (envline, "TOP_PAD_", 8) == 0)
354 __libc_mallopt (M_TOP_PAD, atoi (&envline[9]));
355 else if (memcmp (envline, "PERTURB_", 8) == 0)
356 __libc_mallopt (M_PERTURB, atoi (&envline[9]));
357 }
358 break;
359 case 9:
360 if (!__builtin_expect (__libc_enable_secure, 0))
361 {
362 if (memcmp (envline, "MMAP_MAX_", 9) == 0)
363 __libc_mallopt (M_MMAP_MAX, atoi (&envline[10]));
364 else if (memcmp (envline, "ARENA_MAX", 9) == 0)
365 __libc_mallopt (M_ARENA_MAX, atoi (&envline[10]));
366 }
367 break;
368 case 10:
369 if (!__builtin_expect (__libc_enable_secure, 0))
370 {
371 if (memcmp (envline, "ARENA_TEST", 10) == 0)
372 __libc_mallopt (M_ARENA_TEST, atoi (&envline[11]));
373 }
374 break;
375 case 15:
376 if (!__builtin_expect (__libc_enable_secure, 0))
377 {
378 if (memcmp (envline, "TRIM_THRESHOLD_", 15) == 0)
379 __libc_mallopt (M_TRIM_THRESHOLD, atoi (&envline[16]));
380 else if (memcmp (envline, "MMAP_THRESHOLD_", 15) == 0)
381 __libc_mallopt (M_MMAP_THRESHOLD, atoi (&envline[16]));
382 }
383 break;
384 default:
385 break;
386 }
387 }
388 }
389 if (s && s[0])
390 {
391 __libc_mallopt (M_CHECK_ACTION, (int) (s[0] - '0'));
392 if (check_action != 0)
393 __malloc_check_init ();
394 }
395#endif
396
397#if HAVE_MALLOC_INIT_HOOK
398 void (*hook) (void) = atomic_forced_read (__malloc_initialize_hook);
399 if (hook != NULL)
400 (*hook)();
401#endif
402 __malloc_initialized = 1;
403}
404
405/* Managing heaps and arenas (for concurrent threads) */
406
407#if MALLOC_DEBUG > 1
408
409/* Print the complete contents of a single heap to stderr. */
410
411static void
412dump_heap (heap_info *heap)
413{
414 char *ptr;
415 mchunkptr p;
416
417 fprintf (stderr, "Heap %p, size %10lx:\n", heap, (long) heap->size);
418 ptr = (heap->ar_ptr != (mstate) (heap + 1)) ?
419 (char *) (heap + 1) : (char *) (heap + 1) + sizeof (struct malloc_state);
420 p = (mchunkptr) (((unsigned long) ptr + MALLOC_ALIGN_MASK) &
421 ~MALLOC_ALIGN_MASK);
422 for (;; )
423 {
424 fprintf (stderr, "chunk %p size %10lx", p, (long) p->size);
425 if (p == top (heap->ar_ptr))
426 {
427 fprintf (stderr, " (top)\n");
428 break;
429 }
430 else if (p->size == (0 | PREV_INUSE))
431 {
432 fprintf (stderr, " (fence)\n");
433 break;
434 }
435 fprintf (stderr, "\n");
436 p = next_chunk (p);
437 }
438}
439#endif /* MALLOC_DEBUG > 1 */
440
441/* If consecutive mmap (0, HEAP_MAX_SIZE << 1, ...) calls return decreasing
442 addresses as opposed to increasing, new_heap would badly fragment the
443 address space. In that case remember the second HEAP_MAX_SIZE part
444 aligned to HEAP_MAX_SIZE from last mmap (0, HEAP_MAX_SIZE << 1, ...)
445 call (if it is already aligned) and try to reuse it next time. We need
446 no locking for it, as kernel ensures the atomicity for us - worst case
447 we'll call mmap (addr, HEAP_MAX_SIZE, ...) for some value of addr in
448 multiple threads, but only one will succeed. */
449static char *aligned_heap_area;
450
451/* Create a new heap. size is automatically rounded up to a multiple
452 of the page size. */
453
454static heap_info *
455internal_function
456new_heap (size_t size, size_t top_pad)
457{
458 size_t pagesize = GLRO (dl_pagesize);
459 char *p1, *p2;
460 unsigned long ul;
461 heap_info *h;
462
463 if (size + top_pad < HEAP_MIN_SIZE)
464 size = HEAP_MIN_SIZE;
465 else if (size + top_pad <= HEAP_MAX_SIZE)
466 size += top_pad;
467 else if (size > HEAP_MAX_SIZE)
468 return 0;
469 else
470 size = HEAP_MAX_SIZE;
471 size = ALIGN_UP (size, pagesize);
472
473 /* A memory region aligned to a multiple of HEAP_MAX_SIZE is needed.
474 No swap space needs to be reserved for the following large
475 mapping (on Linux, this is the case for all non-writable mappings
476 anyway). */
477 p2 = MAP_FAILED;
478 if (aligned_heap_area)
479 {
480 p2 = (char *) MMAP (aligned_heap_area, HEAP_MAX_SIZE, PROT_NONE,
481 MAP_NORESERVE);
482 aligned_heap_area = NULL;
483 if (p2 != MAP_FAILED && ((unsigned long) p2 & (HEAP_MAX_SIZE - 1)))
484 {
485 __munmap (p2, HEAP_MAX_SIZE);
486 p2 = MAP_FAILED;
487 }
488 }
489 if (p2 == MAP_FAILED)
490 {
491 p1 = (char *) MMAP (0, HEAP_MAX_SIZE << 1, PROT_NONE, MAP_NORESERVE);
492 if (p1 != MAP_FAILED)
493 {
494 p2 = (char *) (((unsigned long) p1 + (HEAP_MAX_SIZE - 1))
495 & ~(HEAP_MAX_SIZE - 1));
496 ul = p2 - p1;
497 if (ul)
498 __munmap (p1, ul);
499 else
500 aligned_heap_area = p2 + HEAP_MAX_SIZE;
501 __munmap (p2 + HEAP_MAX_SIZE, HEAP_MAX_SIZE - ul);
502 }
503 else
504 {
505 /* Try to take the chance that an allocation of only HEAP_MAX_SIZE
506 is already aligned. */
507 p2 = (char *) MMAP (0, HEAP_MAX_SIZE, PROT_NONE, MAP_NORESERVE);
508 if (p2 == MAP_FAILED)
509 return 0;
510
511 if ((unsigned long) p2 & (HEAP_MAX_SIZE - 1))
512 {
513 __munmap (p2, HEAP_MAX_SIZE);
514 return 0;
515 }
516 }
517 }
518 if (__mprotect (p2, size, PROT_READ | PROT_WRITE) != 0)
519 {
520 __munmap (p2, HEAP_MAX_SIZE);
521 return 0;
522 }
523 h = (heap_info *) p2;
524 h->size = size;
525 h->mprotect_size = size;
526 LIBC_PROBE (memory_heap_new, 2, h, h->size);
527 return h;
528}
529
530/* Grow a heap. size is automatically rounded up to a
531 multiple of the page size. */
532
533static int
534grow_heap (heap_info *h, long diff)
535{
536 size_t pagesize = GLRO (dl_pagesize);
537 long new_size;
538
539 diff = ALIGN_UP (diff, pagesize);
540 new_size = (long) h->size + diff;
541 if ((unsigned long) new_size > (unsigned long) HEAP_MAX_SIZE)
542 return -1;
543
544 if ((unsigned long) new_size > h->mprotect_size)
545 {
546 if (__mprotect ((char *) h + h->mprotect_size,
547 (unsigned long) new_size - h->mprotect_size,
548 PROT_READ | PROT_WRITE) != 0)
549 return -2;
550
551 h->mprotect_size = new_size;
552 }
553
554 h->size = new_size;
555 LIBC_PROBE (memory_heap_more, 2, h, h->size);
556 return 0;
557}
558
559/* Shrink a heap. */
560
561static int
562shrink_heap (heap_info *h, long diff)
563{
564 long new_size;
565
566 new_size = (long) h->size - diff;
567 if (new_size < (long) sizeof (*h))
568 return -1;
569
570 /* Try to re-map the extra heap space freshly to save memory, and make it
571 inaccessible. See malloc-sysdep.h to know when this is true. */
572 if (__glibc_unlikely (check_may_shrink_heap ()))
573 {
574 if ((char *) MMAP ((char *) h + new_size, diff, PROT_NONE,
575 MAP_FIXED) == (char *) MAP_FAILED)
576 return -2;
577
578 h->mprotect_size = new_size;
579 }
580 else
581 __madvise ((char *) h + new_size, diff, MADV_DONTNEED);
582 /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/
583
584 h->size = new_size;
585 LIBC_PROBE (memory_heap_less, 2, h, h->size);
586 return 0;
587}
588
589/* Delete a heap. */
590
591#define delete_heap(heap) \
592 do { \
593 if ((char *) (heap) + HEAP_MAX_SIZE == aligned_heap_area) \
594 aligned_heap_area = NULL; \
595 __munmap ((char *) (heap), HEAP_MAX_SIZE); \
596 } while (0)
597
598static int
599internal_function
600heap_trim (heap_info *heap, size_t pad)
601{
602 mstate ar_ptr = heap->ar_ptr;
603 unsigned long pagesz = GLRO (dl_pagesize);
604 mchunkptr top_chunk = top (ar_ptr), p, bck, fwd;
605 heap_info *prev_heap;
606 long new_size, top_size, top_area, extra, prev_size, misalign;
607
608 /* Can this heap go away completely? */
609 while (top_chunk == chunk_at_offset (heap, sizeof (*heap)))
610 {
611 prev_heap = heap->prev;
612 prev_size = prev_heap->size - (MINSIZE - 2 * SIZE_SZ);
613 p = chunk_at_offset (prev_heap, prev_size);
614 /* fencepost must be properly aligned. */
615 misalign = ((long) p) & MALLOC_ALIGN_MASK;
616 p = chunk_at_offset (prev_heap, prev_size - misalign);
617 assert (chunksize_nomask (p) == (0 | PREV_INUSE)); /* must be fencepost */
618 p = prev_chunk (p);
619 new_size = chunksize (p) + (MINSIZE - 2 * SIZE_SZ) + misalign;
620 assert (new_size > 0 && new_size < (long) (2 * MINSIZE));
621 if (!prev_inuse (p))
622 new_size += prev_size (p);
623 assert (new_size > 0 && new_size < HEAP_MAX_SIZE);
624 if (new_size + (HEAP_MAX_SIZE - prev_heap->size) < pad + MINSIZE + pagesz)
625 break;
626 ar_ptr->system_mem -= heap->size;
627 LIBC_PROBE (memory_heap_free, 2, heap, heap->size);
628 delete_heap (heap);
629 heap = prev_heap;
630 if (!prev_inuse (p)) /* consolidate backward */
631 {
632 p = prev_chunk (p);
633 unlink (ar_ptr, p, bck, fwd);
634 }
635 assert (((unsigned long) ((char *) p + new_size) & (pagesz - 1)) == 0);
636 assert (((char *) p + new_size) == ((char *) heap + heap->size));
637 top (ar_ptr) = top_chunk = p;
638 set_head (top_chunk, new_size | PREV_INUSE);
639 /*check_chunk(ar_ptr, top_chunk);*/
640 }
641
642 /* Uses similar logic for per-thread arenas as the main arena with systrim
643 and _int_free by preserving the top pad and rounding down to the nearest
644 page. */
645 top_size = chunksize (top_chunk);
646 if ((unsigned long)(top_size) <
647 (unsigned long)(mp_.trim_threshold))
648 return 0;
649
650 top_area = top_size - MINSIZE - 1;
651 if (top_area < 0 || (size_t) top_area <= pad)
652 return 0;
653
654 /* Release in pagesize units and round down to the nearest page. */
655 extra = ALIGN_DOWN(top_area - pad, pagesz);
656 if (extra == 0)
657 return 0;
658
659 /* Try to shrink. */
660 if (shrink_heap (heap, extra) != 0)
661 return 0;
662
663 ar_ptr->system_mem -= extra;
664
665 /* Success. Adjust top accordingly. */
666 set_head (top_chunk, (top_size - extra) | PREV_INUSE);
667 /*check_chunk(ar_ptr, top_chunk);*/
668 return 1;
669}
670
671/* Create a new arena with initial size "size". */
672
673/* If REPLACED_ARENA is not NULL, detach it from this thread. Must be
674 called while free_list_lock is held. */
675static void
676detach_arena (mstate replaced_arena)
677{
678 if (replaced_arena != NULL)
679 {
680 assert (replaced_arena->attached_threads > 0);
681 /* The current implementation only detaches from main_arena in
682 case of allocation failure. This means that it is likely not
683 beneficial to put the arena on free_list even if the
684 reference count reaches zero. */
685 --replaced_arena->attached_threads;
686 }
687}
688
689static mstate
690_int_new_arena (size_t size)
691{
692 mstate a;
693 heap_info *h;
694 char *ptr;
695 unsigned long misalign;
696
697 h = new_heap (size + (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT),
698 mp_.top_pad);
699 if (!h)
700 {
701 /* Maybe size is too large to fit in a single heap. So, just try
702 to create a minimally-sized arena and let _int_malloc() attempt
703 to deal with the large request via mmap_chunk(). */
704 h = new_heap (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT, mp_.top_pad);
705 if (!h)
706 return 0;
707 }
708 a = h->ar_ptr = (mstate) (h + 1);
709 malloc_init_state (a);
710 a->attached_threads = 1;
711 /*a->next = NULL;*/
712 a->system_mem = a->max_system_mem = h->size;
713
714 /* Set up the top chunk, with proper alignment. */
715 ptr = (char *) (a + 1);
716 misalign = (unsigned long) chunk2mem (ptr) & MALLOC_ALIGN_MASK;
717 if (misalign > 0)
718 ptr += MALLOC_ALIGNMENT - misalign;
719 top (a) = (mchunkptr) ptr;
720 set_head (top (a), (((char *) h + h->size) - ptr) | PREV_INUSE);
721
722 LIBC_PROBE (memory_arena_new, 2, a, size);
723 mstate replaced_arena = thread_arena;
724 thread_arena = a;
725 __libc_lock_init (a->mutex);
726
727 __libc_lock_lock (list_lock);
728
729 /* Add the new arena to the global list. */
730 a->next = main_arena.next;
731 /* FIXME: The barrier is an attempt to synchronize with read access
732 in reused_arena, which does not acquire list_lock while
733 traversing the list. */
734 atomic_write_barrier ();
735 main_arena.next = a;
736
737 __libc_lock_unlock (list_lock);
738
739 __libc_lock_lock (free_list_lock);
740 detach_arena (replaced_arena);
741 __libc_lock_unlock (free_list_lock);
742
743 /* Lock this arena. NB: Another thread may have been attached to
744 this arena because the arena is now accessible from the
745 main_arena.next list and could have been picked by reused_arena.
746 This can only happen for the last arena created (before the arena
747 limit is reached). At this point, some arena has to be attached
748 to two threads. We could acquire the arena lock before list_lock
749 to make it less likely that reused_arena picks this new arena,
750 but this could result in a deadlock with
751 __malloc_fork_lock_parent. */
752
753 __libc_lock_lock (a->mutex);
754
755 return a;
756}
757
758
759/* Remove an arena from free_list. */
760static mstate
761get_free_list (void)
762{
763 mstate replaced_arena = thread_arena;
764 mstate result = free_list;
765 if (result != NULL)
766 {
767 __libc_lock_lock (free_list_lock);
768 result = free_list;
769 if (result != NULL)
770 {
771 free_list = result->next_free;
772
773 /* The arena will be attached to this thread. */
774 assert (result->attached_threads == 0);
775 result->attached_threads = 1;
776
777 detach_arena (replaced_arena);
778 }
779 __libc_lock_unlock (free_list_lock);
780
781 if (result != NULL)
782 {
783 LIBC_PROBE (memory_arena_reuse_free_list, 1, result);
784 __libc_lock_lock (result->mutex);
785 thread_arena = result;
786 }
787 }
788
789 return result;
790}
791
792/* Remove the arena from the free list (if it is present).
793 free_list_lock must have been acquired by the caller. */
794static void
795remove_from_free_list (mstate arena)
796{
797 mstate *previous = &free_list;
798 for (mstate p = free_list; p != NULL; p = p->next_free)
799 {
800 assert (p->attached_threads == 0);
801 if (p == arena)
802 {
803 /* Remove the requested arena from the list. */
804 *previous = p->next_free;
805 break;
806 }
807 else
808 previous = &p->next_free;
809 }
810}
811
812/* Lock and return an arena that can be reused for memory allocation.
813 Avoid AVOID_ARENA as we have already failed to allocate memory in
814 it and it is currently locked. */
815static mstate
816reused_arena (mstate avoid_arena)
817{
818 mstate result;
819 /* FIXME: Access to next_to_use suffers from data races. */
820 static mstate next_to_use;
821 if (next_to_use == NULL)
822 next_to_use = &main_arena;
823
824 /* Iterate over all arenas (including those linked from
825 free_list). */
826 result = next_to_use;
827 do
828 {
829 if (!arena_is_corrupt (result) && !__libc_lock_trylock (result->mutex))
830 goto out;
831
832 /* FIXME: This is a data race, see _int_new_arena. */
833 result = result->next;
834 }
835 while (result != next_to_use);
836
837 /* Avoid AVOID_ARENA as we have already failed to allocate memory
838 in that arena and it is currently locked. */
839 if (result == avoid_arena)
840 result = result->next;
841
842 /* Make sure that the arena we get is not corrupted. */
843 mstate begin = result;
844 while (arena_is_corrupt (result) || result == avoid_arena)
845 {
846 result = result->next;
847 if (result == begin)
848 /* We looped around the arena list. We could not find any
849 arena that was either not corrupted or not the one we
850 wanted to avoid. */
851 return NULL;
852 }
853
854 /* No arena available without contention. Wait for the next in line. */
855 LIBC_PROBE (memory_arena_reuse_wait, 3, &result->mutex, result, avoid_arena);
856 __libc_lock_lock (result->mutex);
857
858out:
859 /* Attach the arena to the current thread. */
860 {
861 /* Update the arena thread attachment counters. */
862 mstate replaced_arena = thread_arena;
863 __libc_lock_lock (free_list_lock);
864 detach_arena (replaced_arena);
865
866 /* We may have picked up an arena on the free list. We need to
867 preserve the invariant that no arena on the free list has a
868 positive attached_threads counter (otherwise,
869 arena_thread_freeres cannot use the counter to determine if the
870 arena needs to be put on the free list). We unconditionally
871 remove the selected arena from the free list. The caller of
872 reused_arena checked the free list and observed it to be empty,
873 so the list is very short. */
874 remove_from_free_list (result);
875
876 ++result->attached_threads;
877
878 __libc_lock_unlock (free_list_lock);
879 }
880
881 LIBC_PROBE (memory_arena_reuse, 2, result, avoid_arena);
882 thread_arena = result;
883 next_to_use = result->next;
884
885 return result;
886}
887
888static mstate
889internal_function
890arena_get2 (size_t size, mstate avoid_arena)
891{
892 mstate a;
893
894 static size_t narenas_limit;
895
896 a = get_free_list ();
897 if (a == NULL)
898 {
899 /* Nothing immediately available, so generate a new arena. */
900 if (narenas_limit == 0)
901 {
902 if (mp_.arena_max != 0)
903 narenas_limit = mp_.arena_max;
904 else if (narenas > mp_.arena_test)
905 {
906 int n = __get_nprocs ();
907
908 if (n >= 1)
909 narenas_limit = NARENAS_FROM_NCORES (n);
910 else
911 /* We have no information about the system. Assume two
912 cores. */
913 narenas_limit = NARENAS_FROM_NCORES (2);
914 }
915 }
916 repeat:;
917 size_t n = narenas;
918 /* NB: the following depends on the fact that (size_t)0 - 1 is a
919 very large number and that the underflow is OK. If arena_max
920 is set the value of arena_test is irrelevant. If arena_test
921 is set but narenas is not yet larger or equal to arena_test
922 narenas_limit is 0. There is no possibility for narenas to
923 be too big for the test to always fail since there is not
924 enough address space to create that many arenas. */
925 if (__glibc_unlikely (n <= narenas_limit - 1))
926 {
927 if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n))
928 goto repeat;
929 a = _int_new_arena (size);
930 if (__glibc_unlikely (a == NULL))
931 catomic_decrement (&narenas);
932 }
933 else
934 a = reused_arena (avoid_arena);
935 }
936 return a;
937}
938
939/* If we don't have the main arena, then maybe the failure is due to running
940 out of mmapped areas, so we can try allocating on the main arena.
941 Otherwise, it is likely that sbrk() has failed and there is still a chance
942 to mmap(), so try one of the other arenas. */
943static mstate
944arena_get_retry (mstate ar_ptr, size_t bytes)
945{
946 LIBC_PROBE (memory_arena_retry, 2, bytes, ar_ptr);
947 if (ar_ptr != &main_arena)
948 {
949 __libc_lock_unlock (ar_ptr->mutex);
950 /* Don't touch the main arena if it is corrupt. */
951 if (arena_is_corrupt (&main_arena))
952 return NULL;
953
954 ar_ptr = &main_arena;
955 __libc_lock_lock (ar_ptr->mutex);
956 }
957 else
958 {
959 __libc_lock_unlock (ar_ptr->mutex);
960 ar_ptr = arena_get2 (bytes, ar_ptr);
961 }
962
963 return ar_ptr;
964}
965
966static void __attribute__ ((section ("__libc_thread_freeres_fn")))
967arena_thread_freeres (void)
968{
969 mstate a = thread_arena;
970 thread_arena = NULL;
971
972 if (a != NULL)
973 {
974 __libc_lock_lock (free_list_lock);
975 /* If this was the last attached thread for this arena, put the
976 arena on the free list. */
977 assert (a->attached_threads > 0);
978 if (--a->attached_threads == 0)
979 {
980 a->next_free = free_list;
981 free_list = a;
982 }
983 __libc_lock_unlock (free_list_lock);
984 }
985}
986text_set_element (__libc_thread_subfreeres, arena_thread_freeres);
987
988/*
989 * Local variables:
990 * c-basic-offset: 2
991 * End:
992 */
993