1/* Malloc implementation for multiple threads without lock contention.
2 Copyright (C) 2001-2021 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 <https://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);
82#if IS_IN (libc)
83static size_t narenas = 1;
84#endif
85static mstate free_list;
86
87/* list_lock prevents concurrent writes to the next member of struct
88 malloc_state objects.
89
90 Read access to the next member is supposed to synchronize with the
91 atomic_write_barrier and the write to the next member in
92 _int_new_arena. This suffers from data races; see the FIXME
93 comments in _int_new_arena and reused_arena.
94
95 list_lock also prevents concurrent forks. At the time list_lock is
96 acquired, no arena lock must have been acquired, but it is
97 permitted to acquire arena locks subsequently, while list_lock is
98 acquired. */
99__libc_lock_define_initialized (static, list_lock);
100
101/* Already initialized? */
102static bool __malloc_initialized = false;
103
104/**************************************************************************/
105
106
107/* arena_get() acquires an arena and locks the corresponding mutex.
108 First, try the one last locked successfully by this thread. (This
109 is the common case and handled with a macro for speed.) Then, loop
110 once over the circularly linked list of arenas. If no arena is
111 readily available, create a new one. In this latter case, `size'
112 is just a hint as to how much memory will be required immediately
113 in the new arena. */
114
115#define arena_get(ptr, size) do { \
116 ptr = thread_arena; \
117 arena_lock (ptr, size); \
118 } while (0)
119
120#define arena_lock(ptr, size) do { \
121 if (ptr) \
122 __libc_lock_lock (ptr->mutex); \
123 else \
124 ptr = arena_get2 ((size), NULL); \
125 } while (0)
126
127/* find the heap and corresponding arena for a given ptr */
128
129#define heap_for_ptr(ptr) \
130 ((heap_info *) ((unsigned long) (ptr) & ~(HEAP_MAX_SIZE - 1)))
131#define arena_for_chunk(ptr) \
132 (chunk_main_arena (ptr) ? &main_arena : heap_for_ptr (ptr)->ar_ptr)
133
134
135/**************************************************************************/
136
137/* atfork support. */
138
139/* The following three functions are called around fork from a
140 multi-threaded process. We do not use the general fork handler
141 mechanism to make sure that our handlers are the last ones being
142 called, so that other fork handlers can use the malloc
143 subsystem. */
144
145void
146__malloc_fork_lock_parent (void)
147{
148 if (!__malloc_initialized)
149 return;
150
151 /* We do not acquire free_list_lock here because we completely
152 reconstruct free_list in __malloc_fork_unlock_child. */
153
154 __libc_lock_lock (list_lock);
155
156 for (mstate ar_ptr = &main_arena;; )
157 {
158 __libc_lock_lock (ar_ptr->mutex);
159 ar_ptr = ar_ptr->next;
160 if (ar_ptr == &main_arena)
161 break;
162 }
163}
164
165void
166__malloc_fork_unlock_parent (void)
167{
168 if (!__malloc_initialized)
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
182__malloc_fork_unlock_child (void)
183{
184 if (!__malloc_initialized)
185 return;
186
187 /* Push all arenas to the free list, except thread_arena, which is
188 attached to the current thread. */
189 __libc_lock_init (free_list_lock);
190 if (thread_arena != NULL)
191 thread_arena->attached_threads = 1;
192 free_list = NULL;
193 for (mstate ar_ptr = &main_arena;; )
194 {
195 __libc_lock_init (ar_ptr->mutex);
196 if (ar_ptr != thread_arena)
197 {
198 /* This arena is no longer attached to any thread. */
199 ar_ptr->attached_threads = 0;
200 ar_ptr->next_free = free_list;
201 free_list = ar_ptr;
202 }
203 ar_ptr = ar_ptr->next;
204 if (ar_ptr == &main_arena)
205 break;
206 }
207
208 __libc_lock_init (list_lock);
209}
210
211#if HAVE_TUNABLES
212# define TUNABLE_CALLBACK_FNDECL(__name, __type) \
213static inline int do_ ## __name (__type value); \
214static void \
215TUNABLE_CALLBACK (__name) (tunable_val_t *valp) \
216{ \
217 __type value = (__type) (valp)->numval; \
218 do_ ## __name (value); \
219}
220
221TUNABLE_CALLBACK_FNDECL (set_mmap_threshold, size_t)
222TUNABLE_CALLBACK_FNDECL (set_mmaps_max, int32_t)
223TUNABLE_CALLBACK_FNDECL (set_top_pad, size_t)
224TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t)
225TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t)
226TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t)
227TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t)
228#if USE_TCACHE
229TUNABLE_CALLBACK_FNDECL (set_tcache_max, size_t)
230TUNABLE_CALLBACK_FNDECL (set_tcache_count, size_t)
231TUNABLE_CALLBACK_FNDECL (set_tcache_unsorted_limit, size_t)
232#endif
233TUNABLE_CALLBACK_FNDECL (set_mxfast, size_t)
234#else
235/* Initialization routine. */
236#include <string.h>
237extern char **_environ;
238
239static char *
240next_env_entry (char ***position)
241{
242 char **current = *position;
243 char *result = NULL;
244
245 while (*current != NULL)
246 {
247 if (__builtin_expect ((*current)[0] == 'M', 0)
248 && (*current)[1] == 'A'
249 && (*current)[2] == 'L'
250 && (*current)[3] == 'L'
251 && (*current)[4] == 'O'
252 && (*current)[5] == 'C'
253 && (*current)[6] == '_')
254 {
255 result = &(*current)[7];
256
257 /* Save current position for next visit. */
258 *position = ++current;
259
260 break;
261 }
262
263 ++current;
264 }
265
266 return result;
267}
268#endif
269
270
271#ifdef SHARED
272extern struct dl_open_hook *_dl_open_hook;
273libc_hidden_proto (_dl_open_hook);
274#endif
275
276#if USE_TCACHE
277static void tcache_key_initialize (void);
278#endif
279
280static void
281ptmalloc_init (void)
282{
283 if (__malloc_initialized)
284 return;
285
286 __malloc_initialized = true;
287
288#if USE_TCACHE
289 tcache_key_initialize ();
290#endif
291
292#ifdef USE_MTAG
293 if ((TUNABLE_GET_FULL (glibc, mem, tagging, int32_t, NULL) & 1) != 0)
294 {
295 /* If the tunable says that we should be using tagged memory
296 and that morecore does not support tagged regions, then
297 disable it. */
298 if (__MTAG_SBRK_UNTAGGED)
299 __always_fail_morecore = true;
300
301 mtag_enabled = true;
302 mtag_mmap_flags = __MTAG_MMAP_FLAGS;
303 }
304#endif
305
306#if defined SHARED && IS_IN (libc)
307 /* In case this libc copy is in a non-default namespace, never use
308 brk. Likewise if dlopened from statically linked program. The
309 generic sbrk implementation also enforces this, but it is not
310 used on Hurd. */
311 if (!__libc_initial)
312 __always_fail_morecore = true;
313#endif
314
315 thread_arena = &main_arena;
316
317 malloc_init_state (&main_arena);
318
319#if HAVE_TUNABLES
320 TUNABLE_GET (top_pad, size_t, TUNABLE_CALLBACK (set_top_pad));
321 TUNABLE_GET (perturb, int32_t, TUNABLE_CALLBACK (set_perturb_byte));
322 TUNABLE_GET (mmap_threshold, size_t, TUNABLE_CALLBACK (set_mmap_threshold));
323 TUNABLE_GET (trim_threshold, size_t, TUNABLE_CALLBACK (set_trim_threshold));
324 TUNABLE_GET (mmap_max, int32_t, TUNABLE_CALLBACK (set_mmaps_max));
325 TUNABLE_GET (arena_max, size_t, TUNABLE_CALLBACK (set_arena_max));
326 TUNABLE_GET (arena_test, size_t, TUNABLE_CALLBACK (set_arena_test));
327# if USE_TCACHE
328 TUNABLE_GET (tcache_max, size_t, TUNABLE_CALLBACK (set_tcache_max));
329 TUNABLE_GET (tcache_count, size_t, TUNABLE_CALLBACK (set_tcache_count));
330 TUNABLE_GET (tcache_unsorted_limit, size_t,
331 TUNABLE_CALLBACK (set_tcache_unsorted_limit));
332# endif
333 TUNABLE_GET (mxfast, size_t, TUNABLE_CALLBACK (set_mxfast));
334#else
335 if (__glibc_likely (_environ != NULL))
336 {
337 char **runp = _environ;
338 char *envline;
339
340 while (__builtin_expect ((envline = next_env_entry (&runp)) != NULL,
341 0))
342 {
343 size_t len = strcspn (envline, "=");
344
345 if (envline[len] != '=')
346 /* This is a "MALLOC_" variable at the end of the string
347 without a '=' character. Ignore it since otherwise we
348 will access invalid memory below. */
349 continue;
350
351 switch (len)
352 {
353 case 8:
354 if (!__builtin_expect (__libc_enable_secure, 0))
355 {
356 if (memcmp (envline, "TOP_PAD_", 8) == 0)
357 __libc_mallopt (M_TOP_PAD, atoi (&envline[9]));
358 else if (memcmp (envline, "PERTURB_", 8) == 0)
359 __libc_mallopt (M_PERTURB, atoi (&envline[9]));
360 }
361 break;
362 case 9:
363 if (!__builtin_expect (__libc_enable_secure, 0))
364 {
365 if (memcmp (envline, "MMAP_MAX_", 9) == 0)
366 __libc_mallopt (M_MMAP_MAX, atoi (&envline[10]));
367 else if (memcmp (envline, "ARENA_MAX", 9) == 0)
368 __libc_mallopt (M_ARENA_MAX, atoi (&envline[10]));
369 }
370 break;
371 case 10:
372 if (!__builtin_expect (__libc_enable_secure, 0))
373 {
374 if (memcmp (envline, "ARENA_TEST", 10) == 0)
375 __libc_mallopt (M_ARENA_TEST, atoi (&envline[11]));
376 }
377 break;
378 case 15:
379 if (!__builtin_expect (__libc_enable_secure, 0))
380 {
381 if (memcmp (envline, "TRIM_THRESHOLD_", 15) == 0)
382 __libc_mallopt (M_TRIM_THRESHOLD, atoi (&envline[16]));
383 else if (memcmp (envline, "MMAP_THRESHOLD_", 15) == 0)
384 __libc_mallopt (M_MMAP_THRESHOLD, atoi (&envline[16]));
385 }
386 break;
387 default:
388 break;
389 }
390 }
391 }
392#endif
393}
394
395/* Managing heaps and arenas (for concurrent threads) */
396
397#if MALLOC_DEBUG > 1
398
399/* Print the complete contents of a single heap to stderr. */
400
401static void
402dump_heap (heap_info *heap)
403{
404 char *ptr;
405 mchunkptr p;
406
407 fprintf (stderr, "Heap %p, size %10lx:\n", heap, (long) heap->size);
408 ptr = (heap->ar_ptr != (mstate) (heap + 1)) ?
409 (char *) (heap + 1) : (char *) (heap + 1) + sizeof (struct malloc_state);
410 p = (mchunkptr) (((unsigned long) ptr + MALLOC_ALIGN_MASK) &
411 ~MALLOC_ALIGN_MASK);
412 for (;; )
413 {
414 fprintf (stderr, "chunk %p size %10lx", p, (long) chunksize_nomask(p));
415 if (p == top (heap->ar_ptr))
416 {
417 fprintf (stderr, " (top)\n");
418 break;
419 }
420 else if (chunksize_nomask(p) == (0 | PREV_INUSE))
421 {
422 fprintf (stderr, " (fence)\n");
423 break;
424 }
425 fprintf (stderr, "\n");
426 p = next_chunk (p);
427 }
428}
429#endif /* MALLOC_DEBUG > 1 */
430
431/* If consecutive mmap (0, HEAP_MAX_SIZE << 1, ...) calls return decreasing
432 addresses as opposed to increasing, new_heap would badly fragment the
433 address space. In that case remember the second HEAP_MAX_SIZE part
434 aligned to HEAP_MAX_SIZE from last mmap (0, HEAP_MAX_SIZE << 1, ...)
435 call (if it is already aligned) and try to reuse it next time. We need
436 no locking for it, as kernel ensures the atomicity for us - worst case
437 we'll call mmap (addr, HEAP_MAX_SIZE, ...) for some value of addr in
438 multiple threads, but only one will succeed. */
439static char *aligned_heap_area;
440
441/* Create a new heap. size is automatically rounded up to a multiple
442 of the page size. */
443
444static heap_info *
445new_heap (size_t size, size_t top_pad)
446{
447 size_t pagesize = GLRO (dl_pagesize);
448 char *p1, *p2;
449 unsigned long ul;
450 heap_info *h;
451
452 if (size + top_pad < HEAP_MIN_SIZE)
453 size = HEAP_MIN_SIZE;
454 else if (size + top_pad <= HEAP_MAX_SIZE)
455 size += top_pad;
456 else if (size > HEAP_MAX_SIZE)
457 return 0;
458 else
459 size = HEAP_MAX_SIZE;
460 size = ALIGN_UP (size, pagesize);
461
462 /* A memory region aligned to a multiple of HEAP_MAX_SIZE is needed.
463 No swap space needs to be reserved for the following large
464 mapping (on Linux, this is the case for all non-writable mappings
465 anyway). */
466 p2 = MAP_FAILED;
467 if (aligned_heap_area)
468 {
469 p2 = (char *) MMAP (aligned_heap_area, HEAP_MAX_SIZE, PROT_NONE,
470 MAP_NORESERVE);
471 aligned_heap_area = NULL;
472 if (p2 != MAP_FAILED && ((unsigned long) p2 & (HEAP_MAX_SIZE - 1)))
473 {
474 __munmap (p2, HEAP_MAX_SIZE);
475 p2 = MAP_FAILED;
476 }
477 }
478 if (p2 == MAP_FAILED)
479 {
480 p1 = (char *) MMAP (0, HEAP_MAX_SIZE << 1, PROT_NONE, MAP_NORESERVE);
481 if (p1 != MAP_FAILED)
482 {
483 p2 = (char *) (((unsigned long) p1 + (HEAP_MAX_SIZE - 1))
484 & ~(HEAP_MAX_SIZE - 1));
485 ul = p2 - p1;
486 if (ul)
487 __munmap (p1, ul);
488 else
489 aligned_heap_area = p2 + HEAP_MAX_SIZE;
490 __munmap (p2 + HEAP_MAX_SIZE, HEAP_MAX_SIZE - ul);
491 }
492 else
493 {
494 /* Try to take the chance that an allocation of only HEAP_MAX_SIZE
495 is already aligned. */
496 p2 = (char *) MMAP (0, HEAP_MAX_SIZE, PROT_NONE, MAP_NORESERVE);
497 if (p2 == MAP_FAILED)
498 return 0;
499
500 if ((unsigned long) p2 & (HEAP_MAX_SIZE - 1))
501 {
502 __munmap (p2, HEAP_MAX_SIZE);
503 return 0;
504 }
505 }
506 }
507 if (__mprotect (p2, size, mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0)
508 {
509 __munmap (p2, HEAP_MAX_SIZE);
510 return 0;
511 }
512 h = (heap_info *) p2;
513 h->size = size;
514 h->mprotect_size = size;
515 LIBC_PROBE (memory_heap_new, 2, h, h->size);
516 return h;
517}
518
519/* Grow a heap. size is automatically rounded up to a
520 multiple of the page size. */
521
522static int
523grow_heap (heap_info *h, long diff)
524{
525 size_t pagesize = GLRO (dl_pagesize);
526 long new_size;
527
528 diff = ALIGN_UP (diff, pagesize);
529 new_size = (long) h->size + diff;
530 if ((unsigned long) new_size > (unsigned long) HEAP_MAX_SIZE)
531 return -1;
532
533 if ((unsigned long) new_size > h->mprotect_size)
534 {
535 if (__mprotect ((char *) h + h->mprotect_size,
536 (unsigned long) new_size - h->mprotect_size,
537 mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0)
538 return -2;
539
540 h->mprotect_size = new_size;
541 }
542
543 h->size = new_size;
544 LIBC_PROBE (memory_heap_more, 2, h, h->size);
545 return 0;
546}
547
548/* Shrink a heap. */
549
550static int
551shrink_heap (heap_info *h, long diff)
552{
553 long new_size;
554
555 new_size = (long) h->size - diff;
556 if (new_size < (long) sizeof (*h))
557 return -1;
558
559 /* Try to re-map the extra heap space freshly to save memory, and make it
560 inaccessible. See malloc-sysdep.h to know when this is true. */
561 if (__glibc_unlikely (check_may_shrink_heap ()))
562 {
563 if ((char *) MMAP ((char *) h + new_size, diff, PROT_NONE,
564 MAP_FIXED) == (char *) MAP_FAILED)
565 return -2;
566
567 h->mprotect_size = new_size;
568 }
569 else
570 __madvise ((char *) h + new_size, diff, MADV_DONTNEED);
571 /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/
572
573 h->size = new_size;
574 LIBC_PROBE (memory_heap_less, 2, h, h->size);
575 return 0;
576}
577
578/* Delete a heap. */
579
580#define delete_heap(heap) \
581 do { \
582 if ((char *) (heap) + HEAP_MAX_SIZE == aligned_heap_area) \
583 aligned_heap_area = NULL; \
584 __munmap ((char *) (heap), HEAP_MAX_SIZE); \
585 } while (0)
586
587static int
588heap_trim (heap_info *heap, size_t pad)
589{
590 mstate ar_ptr = heap->ar_ptr;
591 unsigned long pagesz = GLRO (dl_pagesize);
592 mchunkptr top_chunk = top (ar_ptr), p;
593 heap_info *prev_heap;
594 long new_size, top_size, top_area, extra, prev_size, misalign;
595
596 /* Can this heap go away completely? */
597 while (top_chunk == chunk_at_offset (heap, sizeof (*heap)))
598 {
599 prev_heap = heap->prev;
600 prev_size = prev_heap->size - (MINSIZE - 2 * SIZE_SZ);
601 p = chunk_at_offset (prev_heap, prev_size);
602 /* fencepost must be properly aligned. */
603 misalign = ((long) p) & MALLOC_ALIGN_MASK;
604 p = chunk_at_offset (prev_heap, prev_size - misalign);
605 assert (chunksize_nomask (p) == (0 | PREV_INUSE)); /* must be fencepost */
606 p = prev_chunk (p);
607 new_size = chunksize (p) + (MINSIZE - 2 * SIZE_SZ) + misalign;
608 assert (new_size > 0 && new_size < (long) (2 * MINSIZE));
609 if (!prev_inuse (p))
610 new_size += prev_size (p);
611 assert (new_size > 0 && new_size < HEAP_MAX_SIZE);
612 if (new_size + (HEAP_MAX_SIZE - prev_heap->size) < pad + MINSIZE + pagesz)
613 break;
614 ar_ptr->system_mem -= heap->size;
615 LIBC_PROBE (memory_heap_free, 2, heap, heap->size);
616 delete_heap (heap);
617 heap = prev_heap;
618 if (!prev_inuse (p)) /* consolidate backward */
619 {
620 p = prev_chunk (p);
621 unlink_chunk (ar_ptr, p);
622 }
623 assert (((unsigned long) ((char *) p + new_size) & (pagesz - 1)) == 0);
624 assert (((char *) p + new_size) == ((char *) heap + heap->size));
625 top (ar_ptr) = top_chunk = p;
626 set_head (top_chunk, new_size | PREV_INUSE);
627 /*check_chunk(ar_ptr, top_chunk);*/
628 }
629
630 /* Uses similar logic for per-thread arenas as the main arena with systrim
631 and _int_free by preserving the top pad and rounding down to the nearest
632 page. */
633 top_size = chunksize (top_chunk);
634 if ((unsigned long)(top_size) <
635 (unsigned long)(mp_.trim_threshold))
636 return 0;
637
638 top_area = top_size - MINSIZE - 1;
639 if (top_area < 0 || (size_t) top_area <= pad)
640 return 0;
641
642 /* Release in pagesize units and round down to the nearest page. */
643 extra = ALIGN_DOWN(top_area - pad, pagesz);
644 if (extra == 0)
645 return 0;
646
647 /* Try to shrink. */
648 if (shrink_heap (heap, extra) != 0)
649 return 0;
650
651 ar_ptr->system_mem -= extra;
652
653 /* Success. Adjust top accordingly. */
654 set_head (top_chunk, (top_size - extra) | PREV_INUSE);
655 /*check_chunk(ar_ptr, top_chunk);*/
656 return 1;
657}
658
659/* Create a new arena with initial size "size". */
660
661#if IS_IN (libc)
662/* If REPLACED_ARENA is not NULL, detach it from this thread. Must be
663 called while free_list_lock is held. */
664static void
665detach_arena (mstate replaced_arena)
666{
667 if (replaced_arena != NULL)
668 {
669 assert (replaced_arena->attached_threads > 0);
670 /* The current implementation only detaches from main_arena in
671 case of allocation failure. This means that it is likely not
672 beneficial to put the arena on free_list even if the
673 reference count reaches zero. */
674 --replaced_arena->attached_threads;
675 }
676}
677
678static mstate
679_int_new_arena (size_t size)
680{
681 mstate a;
682 heap_info *h;
683 char *ptr;
684 unsigned long misalign;
685
686 h = new_heap (size + (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT),
687 mp_.top_pad);
688 if (!h)
689 {
690 /* Maybe size is too large to fit in a single heap. So, just try
691 to create a minimally-sized arena and let _int_malloc() attempt
692 to deal with the large request via mmap_chunk(). */
693 h = new_heap (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT, mp_.top_pad);
694 if (!h)
695 return 0;
696 }
697 a = h->ar_ptr = (mstate) (h + 1);
698 malloc_init_state (a);
699 a->attached_threads = 1;
700 /*a->next = NULL;*/
701 a->system_mem = a->max_system_mem = h->size;
702
703 /* Set up the top chunk, with proper alignment. */
704 ptr = (char *) (a + 1);
705 misalign = (unsigned long) chunk2mem (ptr) & MALLOC_ALIGN_MASK;
706 if (misalign > 0)
707 ptr += MALLOC_ALIGNMENT - misalign;
708 top (a) = (mchunkptr) ptr;
709 set_head (top (a), (((char *) h + h->size) - ptr) | PREV_INUSE);
710
711 LIBC_PROBE (memory_arena_new, 2, a, size);
712 mstate replaced_arena = thread_arena;
713 thread_arena = a;
714 __libc_lock_init (a->mutex);
715
716 __libc_lock_lock (list_lock);
717
718 /* Add the new arena to the global list. */
719 a->next = main_arena.next;
720 /* FIXME: The barrier is an attempt to synchronize with read access
721 in reused_arena, which does not acquire list_lock while
722 traversing the list. */
723 atomic_write_barrier ();
724 main_arena.next = a;
725
726 __libc_lock_unlock (list_lock);
727
728 __libc_lock_lock (free_list_lock);
729 detach_arena (replaced_arena);
730 __libc_lock_unlock (free_list_lock);
731
732 /* Lock this arena. NB: Another thread may have been attached to
733 this arena because the arena is now accessible from the
734 main_arena.next list and could have been picked by reused_arena.
735 This can only happen for the last arena created (before the arena
736 limit is reached). At this point, some arena has to be attached
737 to two threads. We could acquire the arena lock before list_lock
738 to make it less likely that reused_arena picks this new arena,
739 but this could result in a deadlock with
740 __malloc_fork_lock_parent. */
741
742 __libc_lock_lock (a->mutex);
743
744 return a;
745}
746
747
748/* Remove an arena from free_list. */
749static mstate
750get_free_list (void)
751{
752 mstate replaced_arena = thread_arena;
753 mstate result = free_list;
754 if (result != NULL)
755 {
756 __libc_lock_lock (free_list_lock);
757 result = free_list;
758 if (result != NULL)
759 {
760 free_list = result->next_free;
761
762 /* The arena will be attached to this thread. */
763 assert (result->attached_threads == 0);
764 result->attached_threads = 1;
765
766 detach_arena (replaced_arena);
767 }
768 __libc_lock_unlock (free_list_lock);
769
770 if (result != NULL)
771 {
772 LIBC_PROBE (memory_arena_reuse_free_list, 1, result);
773 __libc_lock_lock (result->mutex);
774 thread_arena = result;
775 }
776 }
777
778 return result;
779}
780
781/* Remove the arena from the free list (if it is present).
782 free_list_lock must have been acquired by the caller. */
783static void
784remove_from_free_list (mstate arena)
785{
786 mstate *previous = &free_list;
787 for (mstate p = free_list; p != NULL; p = p->next_free)
788 {
789 assert (p->attached_threads == 0);
790 if (p == arena)
791 {
792 /* Remove the requested arena from the list. */
793 *previous = p->next_free;
794 break;
795 }
796 else
797 previous = &p->next_free;
798 }
799}
800
801/* Lock and return an arena that can be reused for memory allocation.
802 Avoid AVOID_ARENA as we have already failed to allocate memory in
803 it and it is currently locked. */
804static mstate
805reused_arena (mstate avoid_arena)
806{
807 mstate result;
808 /* FIXME: Access to next_to_use suffers from data races. */
809 static mstate next_to_use;
810 if (next_to_use == NULL)
811 next_to_use = &main_arena;
812
813 /* Iterate over all arenas (including those linked from
814 free_list). */
815 result = next_to_use;
816 do
817 {
818 if (!__libc_lock_trylock (result->mutex))
819 goto out;
820
821 /* FIXME: This is a data race, see _int_new_arena. */
822 result = result->next;
823 }
824 while (result != next_to_use);
825
826 /* Avoid AVOID_ARENA as we have already failed to allocate memory
827 in that arena and it is currently locked. */
828 if (result == avoid_arena)
829 result = result->next;
830
831 /* No arena available without contention. Wait for the next in line. */
832 LIBC_PROBE (memory_arena_reuse_wait, 3, &result->mutex, result, avoid_arena);
833 __libc_lock_lock (result->mutex);
834
835out:
836 /* Attach the arena to the current thread. */
837 {
838 /* Update the arena thread attachment counters. */
839 mstate replaced_arena = thread_arena;
840 __libc_lock_lock (free_list_lock);
841 detach_arena (replaced_arena);
842
843 /* We may have picked up an arena on the free list. We need to
844 preserve the invariant that no arena on the free list has a
845 positive attached_threads counter (otherwise,
846 arena_thread_freeres cannot use the counter to determine if the
847 arena needs to be put on the free list). We unconditionally
848 remove the selected arena from the free list. The caller of
849 reused_arena checked the free list and observed it to be empty,
850 so the list is very short. */
851 remove_from_free_list (result);
852
853 ++result->attached_threads;
854
855 __libc_lock_unlock (free_list_lock);
856 }
857
858 LIBC_PROBE (memory_arena_reuse, 2, result, avoid_arena);
859 thread_arena = result;
860 next_to_use = result->next;
861
862 return result;
863}
864
865static mstate
866arena_get2 (size_t size, mstate avoid_arena)
867{
868 mstate a;
869
870 static size_t narenas_limit;
871
872 a = get_free_list ();
873 if (a == NULL)
874 {
875 /* Nothing immediately available, so generate a new arena. */
876 if (narenas_limit == 0)
877 {
878 if (mp_.arena_max != 0)
879 narenas_limit = mp_.arena_max;
880 else if (narenas > mp_.arena_test)
881 {
882 int n = __get_nprocs ();
883
884 if (n >= 1)
885 narenas_limit = NARENAS_FROM_NCORES (n);
886 else
887 /* We have no information about the system. Assume two
888 cores. */
889 narenas_limit = NARENAS_FROM_NCORES (2);
890 }
891 }
892 repeat:;
893 size_t n = narenas;
894 /* NB: the following depends on the fact that (size_t)0 - 1 is a
895 very large number and that the underflow is OK. If arena_max
896 is set the value of arena_test is irrelevant. If arena_test
897 is set but narenas is not yet larger or equal to arena_test
898 narenas_limit is 0. There is no possibility for narenas to
899 be too big for the test to always fail since there is not
900 enough address space to create that many arenas. */
901 if (__glibc_unlikely (n <= narenas_limit - 1))
902 {
903 if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n))
904 goto repeat;
905 a = _int_new_arena (size);
906 if (__glibc_unlikely (a == NULL))
907 catomic_decrement (&narenas);
908 }
909 else
910 a = reused_arena (avoid_arena);
911 }
912 return a;
913}
914
915/* If we don't have the main arena, then maybe the failure is due to running
916 out of mmapped areas, so we can try allocating on the main arena.
917 Otherwise, it is likely that sbrk() has failed and there is still a chance
918 to mmap(), so try one of the other arenas. */
919static mstate
920arena_get_retry (mstate ar_ptr, size_t bytes)
921{
922 LIBC_PROBE (memory_arena_retry, 2, bytes, ar_ptr);
923 if (ar_ptr != &main_arena)
924 {
925 __libc_lock_unlock (ar_ptr->mutex);
926 ar_ptr = &main_arena;
927 __libc_lock_lock (ar_ptr->mutex);
928 }
929 else
930 {
931 __libc_lock_unlock (ar_ptr->mutex);
932 ar_ptr = arena_get2 (bytes, ar_ptr);
933 }
934
935 return ar_ptr;
936}
937#endif
938
939void
940__malloc_arena_thread_freeres (void)
941{
942 /* Shut down the thread cache first. This could deallocate data for
943 the thread arena, so do this before we put the arena on the free
944 list. */
945 tcache_thread_shutdown ();
946
947 mstate a = thread_arena;
948 thread_arena = NULL;
949
950 if (a != NULL)
951 {
952 __libc_lock_lock (free_list_lock);
953 /* If this was the last attached thread for this arena, put the
954 arena on the free list. */
955 assert (a->attached_threads > 0);
956 if (--a->attached_threads == 0)
957 {
958 a->next_free = free_list;
959 free_list = a;
960 }
961 __libc_lock_unlock (free_list_lock);
962 }
963}
964
965/*
966 * Local variables:
967 * c-basic-offset: 2
968 * End:
969 */
970