1/* Malloc implementation for multiple threads without lock contention.
2 Copyright (C) 2001-2020 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);
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) \
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
144__malloc_fork_lock_parent (void)
145{
146 if (__malloc_initialized < 1)
147 return;
148
149 /* We do not acquire free_list_lock here because we completely
150 reconstruct free_list in __malloc_fork_unlock_child. */
151
152 __libc_lock_lock (list_lock);
153
154 for (mstate ar_ptr = &main_arena;; )
155 {
156 __libc_lock_lock (ar_ptr->mutex);
157 ar_ptr = ar_ptr->next;
158 if (ar_ptr == &main_arena)
159 break;
160 }
161}
162
163void
164__malloc_fork_unlock_parent (void)
165{
166 if (__malloc_initialized < 1)
167 return;
168
169 for (mstate ar_ptr = &main_arena;; )
170 {
171 __libc_lock_unlock (ar_ptr->mutex);
172 ar_ptr = ar_ptr->next;
173 if (ar_ptr == &main_arena)
174 break;
175 }
176 __libc_lock_unlock (list_lock);
177}
178
179void
180__malloc_fork_unlock_child (void)
181{
182 if (__malloc_initialized < 1)
183 return;
184
185 /* Push all arenas to the free list, except thread_arena, which is
186 attached to the current thread. */
187 __libc_lock_init (free_list_lock);
188 if (thread_arena != NULL)
189 thread_arena->attached_threads = 1;
190 free_list = NULL;
191 for (mstate ar_ptr = &main_arena;; )
192 {
193 __libc_lock_init (ar_ptr->mutex);
194 if (ar_ptr != thread_arena)
195 {
196 /* This arena is no longer attached to any thread. */
197 ar_ptr->attached_threads = 0;
198 ar_ptr->next_free = free_list;
199 free_list = ar_ptr;
200 }
201 ar_ptr = ar_ptr->next;
202 if (ar_ptr == &main_arena)
203 break;
204 }
205
206 __libc_lock_init (list_lock);
207}
208
209#if HAVE_TUNABLES
210void
211TUNABLE_CALLBACK (set_mallopt_check) (tunable_val_t *valp)
212{
213 int32_t value = (int32_t) valp->numval;
214 if (value != 0)
215 __malloc_check_init ();
216}
217
218# define TUNABLE_CALLBACK_FNDECL(__name, __type) \
219static inline int do_ ## __name (__type value); \
220void \
221TUNABLE_CALLBACK (__name) (tunable_val_t *valp) \
222{ \
223 __type value = (__type) (valp)->numval; \
224 do_ ## __name (value); \
225}
226
227TUNABLE_CALLBACK_FNDECL (set_mmap_threshold, size_t)
228TUNABLE_CALLBACK_FNDECL (set_mmaps_max, int32_t)
229TUNABLE_CALLBACK_FNDECL (set_top_pad, size_t)
230TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t)
231TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t)
232TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t)
233TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t)
234#if USE_TCACHE
235TUNABLE_CALLBACK_FNDECL (set_tcache_max, size_t)
236TUNABLE_CALLBACK_FNDECL (set_tcache_count, size_t)
237TUNABLE_CALLBACK_FNDECL (set_tcache_unsorted_limit, size_t)
238#endif
239TUNABLE_CALLBACK_FNDECL (set_mxfast, size_t)
240#else
241/* Initialization routine. */
242#include <string.h>
243extern char **_environ;
244
245static char *
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 malloc_init_state (&main_arena);
311
312#if HAVE_TUNABLES
313 TUNABLE_GET (check, int32_t, TUNABLE_CALLBACK (set_mallopt_check));
314 TUNABLE_GET (top_pad, size_t, TUNABLE_CALLBACK (set_top_pad));
315 TUNABLE_GET (perturb, int32_t, TUNABLE_CALLBACK (set_perturb_byte));
316 TUNABLE_GET (mmap_threshold, size_t, TUNABLE_CALLBACK (set_mmap_threshold));
317 TUNABLE_GET (trim_threshold, size_t, TUNABLE_CALLBACK (set_trim_threshold));
318 TUNABLE_GET (mmap_max, int32_t, TUNABLE_CALLBACK (set_mmaps_max));
319 TUNABLE_GET (arena_max, size_t, TUNABLE_CALLBACK (set_arena_max));
320 TUNABLE_GET (arena_test, size_t, TUNABLE_CALLBACK (set_arena_test));
321# if USE_TCACHE
322 TUNABLE_GET (tcache_max, size_t, TUNABLE_CALLBACK (set_tcache_max));
323 TUNABLE_GET (tcache_count, size_t, TUNABLE_CALLBACK (set_tcache_count));
324 TUNABLE_GET (tcache_unsorted_limit, size_t,
325 TUNABLE_CALLBACK (set_tcache_unsorted_limit));
326# endif
327 TUNABLE_GET (mxfast, size_t, TUNABLE_CALLBACK (set_mxfast));
328#else
329 const char *s = NULL;
330 if (__glibc_likely (_environ != NULL))
331 {
332 char **runp = _environ;
333 char *envline;
334
335 while (__builtin_expect ((envline = next_env_entry (&runp)) != NULL,
336 0))
337 {
338 size_t len = strcspn (envline, "=");
339
340 if (envline[len] != '=')
341 /* This is a "MALLOC_" variable at the end of the string
342 without a '=' character. Ignore it since otherwise we
343 will access invalid memory below. */
344 continue;
345
346 switch (len)
347 {
348 case 6:
349 if (memcmp (envline, "CHECK_", 6) == 0)
350 s = &envline[7];
351 break;
352 case 8:
353 if (!__builtin_expect (__libc_enable_secure, 0))
354 {
355 if (memcmp (envline, "TOP_PAD_", 8) == 0)
356 __libc_mallopt (M_TOP_PAD, atoi (&envline[9]));
357 else if (memcmp (envline, "PERTURB_", 8) == 0)
358 __libc_mallopt (M_PERTURB, atoi (&envline[9]));
359 }
360 break;
361 case 9:
362 if (!__builtin_expect (__libc_enable_secure, 0))
363 {
364 if (memcmp (envline, "MMAP_MAX_", 9) == 0)
365 __libc_mallopt (M_MMAP_MAX, atoi (&envline[10]));
366 else if (memcmp (envline, "ARENA_MAX", 9) == 0)
367 __libc_mallopt (M_ARENA_MAX, atoi (&envline[10]));
368 }
369 break;
370 case 10:
371 if (!__builtin_expect (__libc_enable_secure, 0))
372 {
373 if (memcmp (envline, "ARENA_TEST", 10) == 0)
374 __libc_mallopt (M_ARENA_TEST, atoi (&envline[11]));
375 }
376 break;
377 case 15:
378 if (!__builtin_expect (__libc_enable_secure, 0))
379 {
380 if (memcmp (envline, "TRIM_THRESHOLD_", 15) == 0)
381 __libc_mallopt (M_TRIM_THRESHOLD, atoi (&envline[16]));
382 else if (memcmp (envline, "MMAP_THRESHOLD_", 15) == 0)
383 __libc_mallopt (M_MMAP_THRESHOLD, atoi (&envline[16]));
384 }
385 break;
386 default:
387 break;
388 }
389 }
390 }
391 if (s && s[0] != '\0' && s[0] != '0')
392 __malloc_check_init ();
393#endif
394
395#if HAVE_MALLOC_INIT_HOOK
396 void (*hook) (void) = atomic_forced_read (__malloc_initialize_hook);
397 if (hook != NULL)
398 (*hook)();
399#endif
400 __malloc_initialized = 1;
401}
402
403/* Managing heaps and arenas (for concurrent threads) */
404
405#if MALLOC_DEBUG > 1
406
407/* Print the complete contents of a single heap to stderr. */
408
409static void
410dump_heap (heap_info *heap)
411{
412 char *ptr;
413 mchunkptr p;
414
415 fprintf (stderr, "Heap %p, size %10lx:\n", heap, (long) heap->size);
416 ptr = (heap->ar_ptr != (mstate) (heap + 1)) ?
417 (char *) (heap + 1) : (char *) (heap + 1) + sizeof (struct malloc_state);
418 p = (mchunkptr) (((unsigned long) ptr + MALLOC_ALIGN_MASK) &
419 ~MALLOC_ALIGN_MASK);
420 for (;; )
421 {
422 fprintf (stderr, "chunk %p size %10lx", p, (long) p->size);
423 if (p == top (heap->ar_ptr))
424 {
425 fprintf (stderr, " (top)\n");
426 break;
427 }
428 else if (p->size == (0 | PREV_INUSE))
429 {
430 fprintf (stderr, " (fence)\n");
431 break;
432 }
433 fprintf (stderr, "\n");
434 p = next_chunk (p);
435 }
436}
437#endif /* MALLOC_DEBUG > 1 */
438
439/* If consecutive mmap (0, HEAP_MAX_SIZE << 1, ...) calls return decreasing
440 addresses as opposed to increasing, new_heap would badly fragment the
441 address space. In that case remember the second HEAP_MAX_SIZE part
442 aligned to HEAP_MAX_SIZE from last mmap (0, HEAP_MAX_SIZE << 1, ...)
443 call (if it is already aligned) and try to reuse it next time. We need
444 no locking for it, as kernel ensures the atomicity for us - worst case
445 we'll call mmap (addr, HEAP_MAX_SIZE, ...) for some value of addr in
446 multiple threads, but only one will succeed. */
447static char *aligned_heap_area;
448
449/* Create a new heap. size is automatically rounded up to a multiple
450 of the page size. */
451
452static heap_info *
453new_heap (size_t size, size_t top_pad)
454{
455 size_t pagesize = GLRO (dl_pagesize);
456 char *p1, *p2;
457 unsigned long ul;
458 heap_info *h;
459
460 if (size + top_pad < HEAP_MIN_SIZE)
461 size = HEAP_MIN_SIZE;
462 else if (size + top_pad <= HEAP_MAX_SIZE)
463 size += top_pad;
464 else if (size > HEAP_MAX_SIZE)
465 return 0;
466 else
467 size = HEAP_MAX_SIZE;
468 size = ALIGN_UP (size, pagesize);
469
470 /* A memory region aligned to a multiple of HEAP_MAX_SIZE is needed.
471 No swap space needs to be reserved for the following large
472 mapping (on Linux, this is the case for all non-writable mappings
473 anyway). */
474 p2 = MAP_FAILED;
475 if (aligned_heap_area)
476 {
477 p2 = (char *) MMAP (aligned_heap_area, HEAP_MAX_SIZE, PROT_NONE,
478 MAP_NORESERVE);
479 aligned_heap_area = NULL;
480 if (p2 != MAP_FAILED && ((unsigned long) p2 & (HEAP_MAX_SIZE - 1)))
481 {
482 __munmap (p2, HEAP_MAX_SIZE);
483 p2 = MAP_FAILED;
484 }
485 }
486 if (p2 == MAP_FAILED)
487 {
488 p1 = (char *) MMAP (0, HEAP_MAX_SIZE << 1, PROT_NONE, MAP_NORESERVE);
489 if (p1 != MAP_FAILED)
490 {
491 p2 = (char *) (((unsigned long) p1 + (HEAP_MAX_SIZE - 1))
492 & ~(HEAP_MAX_SIZE - 1));
493 ul = p2 - p1;
494 if (ul)
495 __munmap (p1, ul);
496 else
497 aligned_heap_area = p2 + HEAP_MAX_SIZE;
498 __munmap (p2 + HEAP_MAX_SIZE, HEAP_MAX_SIZE - ul);
499 }
500 else
501 {
502 /* Try to take the chance that an allocation of only HEAP_MAX_SIZE
503 is already aligned. */
504 p2 = (char *) MMAP (0, HEAP_MAX_SIZE, PROT_NONE, MAP_NORESERVE);
505 if (p2 == MAP_FAILED)
506 return 0;
507
508 if ((unsigned long) p2 & (HEAP_MAX_SIZE - 1))
509 {
510 __munmap (p2, HEAP_MAX_SIZE);
511 return 0;
512 }
513 }
514 }
515 if (__mprotect (p2, size, PROT_READ | PROT_WRITE) != 0)
516 {
517 __munmap (p2, HEAP_MAX_SIZE);
518 return 0;
519 }
520 h = (heap_info *) p2;
521 h->size = size;
522 h->mprotect_size = size;
523 LIBC_PROBE (memory_heap_new, 2, h, h->size);
524 return h;
525}
526
527/* Grow a heap. size is automatically rounded up to a
528 multiple of the page size. */
529
530static int
531grow_heap (heap_info *h, long diff)
532{
533 size_t pagesize = GLRO (dl_pagesize);
534 long new_size;
535
536 diff = ALIGN_UP (diff, pagesize);
537 new_size = (long) h->size + diff;
538 if ((unsigned long) new_size > (unsigned long) HEAP_MAX_SIZE)
539 return -1;
540
541 if ((unsigned long) new_size > h->mprotect_size)
542 {
543 if (__mprotect ((char *) h + h->mprotect_size,
544 (unsigned long) new_size - h->mprotect_size,
545 PROT_READ | PROT_WRITE) != 0)
546 return -2;
547
548 h->mprotect_size = new_size;
549 }
550
551 h->size = new_size;
552 LIBC_PROBE (memory_heap_more, 2, h, h->size);
553 return 0;
554}
555
556/* Shrink a heap. */
557
558static int
559shrink_heap (heap_info *h, long diff)
560{
561 long new_size;
562
563 new_size = (long) h->size - diff;
564 if (new_size < (long) sizeof (*h))
565 return -1;
566
567 /* Try to re-map the extra heap space freshly to save memory, and make it
568 inaccessible. See malloc-sysdep.h to know when this is true. */
569 if (__glibc_unlikely (check_may_shrink_heap ()))
570 {
571 if ((char *) MMAP ((char *) h + new_size, diff, PROT_NONE,
572 MAP_FIXED) == (char *) MAP_FAILED)
573 return -2;
574
575 h->mprotect_size = new_size;
576 }
577 else
578 __madvise ((char *) h + new_size, diff, MADV_DONTNEED);
579 /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/
580
581 h->size = new_size;
582 LIBC_PROBE (memory_heap_less, 2, h, h->size);
583 return 0;
584}
585
586/* Delete a heap. */
587
588#define delete_heap(heap) \
589 do { \
590 if ((char *) (heap) + HEAP_MAX_SIZE == aligned_heap_area) \
591 aligned_heap_area = NULL; \
592 __munmap ((char *) (heap), HEAP_MAX_SIZE); \
593 } while (0)
594
595static int
596heap_trim (heap_info *heap, size_t pad)
597{
598 mstate ar_ptr = heap->ar_ptr;
599 unsigned long pagesz = GLRO (dl_pagesize);
600 mchunkptr top_chunk = top (ar_ptr), p;
601 heap_info *prev_heap;
602 long new_size, top_size, top_area, extra, prev_size, misalign;
603
604 /* Can this heap go away completely? */
605 while (top_chunk == chunk_at_offset (heap, sizeof (*heap)))
606 {
607 prev_heap = heap->prev;
608 prev_size = prev_heap->size - (MINSIZE - 2 * SIZE_SZ);
609 p = chunk_at_offset (prev_heap, prev_size);
610 /* fencepost must be properly aligned. */
611 misalign = ((long) p) & MALLOC_ALIGN_MASK;
612 p = chunk_at_offset (prev_heap, prev_size - misalign);
613 assert (chunksize_nomask (p) == (0 | PREV_INUSE)); /* must be fencepost */
614 p = prev_chunk (p);
615 new_size = chunksize (p) + (MINSIZE - 2 * SIZE_SZ) + misalign;
616 assert (new_size > 0 && new_size < (long) (2 * MINSIZE));
617 if (!prev_inuse (p))
618 new_size += prev_size (p);
619 assert (new_size > 0 && new_size < HEAP_MAX_SIZE);
620 if (new_size + (HEAP_MAX_SIZE - prev_heap->size) < pad + MINSIZE + pagesz)
621 break;
622 ar_ptr->system_mem -= heap->size;
623 LIBC_PROBE (memory_heap_free, 2, heap, heap->size);
624 delete_heap (heap);
625 heap = prev_heap;
626 if (!prev_inuse (p)) /* consolidate backward */
627 {
628 p = prev_chunk (p);
629 unlink_chunk (ar_ptr, p);
630 }
631 assert (((unsigned long) ((char *) p + new_size) & (pagesz - 1)) == 0);
632 assert (((char *) p + new_size) == ((char *) heap + heap->size));
633 top (ar_ptr) = top_chunk = p;
634 set_head (top_chunk, new_size | PREV_INUSE);
635 /*check_chunk(ar_ptr, top_chunk);*/
636 }
637
638 /* Uses similar logic for per-thread arenas as the main arena with systrim
639 and _int_free by preserving the top pad and rounding down to the nearest
640 page. */
641 top_size = chunksize (top_chunk);
642 if ((unsigned long)(top_size) <
643 (unsigned long)(mp_.trim_threshold))
644 return 0;
645
646 top_area = top_size - MINSIZE - 1;
647 if (top_area < 0 || (size_t) top_area <= pad)
648 return 0;
649
650 /* Release in pagesize units and round down to the nearest page. */
651 extra = ALIGN_DOWN(top_area - pad, pagesz);
652 if (extra == 0)
653 return 0;
654
655 /* Try to shrink. */
656 if (shrink_heap (heap, extra) != 0)
657 return 0;
658
659 ar_ptr->system_mem -= extra;
660
661 /* Success. Adjust top accordingly. */
662 set_head (top_chunk, (top_size - extra) | PREV_INUSE);
663 /*check_chunk(ar_ptr, top_chunk);*/
664 return 1;
665}
666
667/* Create a new arena with initial size "size". */
668
669/* If REPLACED_ARENA is not NULL, detach it from this thread. Must be
670 called while free_list_lock is held. */
671static void
672detach_arena (mstate replaced_arena)
673{
674 if (replaced_arena != NULL)
675 {
676 assert (replaced_arena->attached_threads > 0);
677 /* The current implementation only detaches from main_arena in
678 case of allocation failure. This means that it is likely not
679 beneficial to put the arena on free_list even if the
680 reference count reaches zero. */
681 --replaced_arena->attached_threads;
682 }
683}
684
685static mstate
686_int_new_arena (size_t size)
687{
688 mstate a;
689 heap_info *h;
690 char *ptr;
691 unsigned long misalign;
692
693 h = new_heap (size + (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT),
694 mp_.top_pad);
695 if (!h)
696 {
697 /* Maybe size is too large to fit in a single heap. So, just try
698 to create a minimally-sized arena and let _int_malloc() attempt
699 to deal with the large request via mmap_chunk(). */
700 h = new_heap (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT, mp_.top_pad);
701 if (!h)
702 return 0;
703 }
704 a = h->ar_ptr = (mstate) (h + 1);
705 malloc_init_state (a);
706 a->attached_threads = 1;
707 /*a->next = NULL;*/
708 a->system_mem = a->max_system_mem = h->size;
709
710 /* Set up the top chunk, with proper alignment. */
711 ptr = (char *) (a + 1);
712 misalign = (unsigned long) chunk2mem (ptr) & MALLOC_ALIGN_MASK;
713 if (misalign > 0)
714 ptr += MALLOC_ALIGNMENT - misalign;
715 top (a) = (mchunkptr) ptr;
716 set_head (top (a), (((char *) h + h->size) - ptr) | PREV_INUSE);
717
718 LIBC_PROBE (memory_arena_new, 2, a, size);
719 mstate replaced_arena = thread_arena;
720 thread_arena = a;
721 __libc_lock_init (a->mutex);
722
723 __libc_lock_lock (list_lock);
724
725 /* Add the new arena to the global list. */
726 a->next = main_arena.next;
727 /* FIXME: The barrier is an attempt to synchronize with read access
728 in reused_arena, which does not acquire list_lock while
729 traversing the list. */
730 atomic_write_barrier ();
731 main_arena.next = a;
732
733 __libc_lock_unlock (list_lock);
734
735 __libc_lock_lock (free_list_lock);
736 detach_arena (replaced_arena);
737 __libc_lock_unlock (free_list_lock);
738
739 /* Lock this arena. NB: Another thread may have been attached to
740 this arena because the arena is now accessible from the
741 main_arena.next list and could have been picked by reused_arena.
742 This can only happen for the last arena created (before the arena
743 limit is reached). At this point, some arena has to be attached
744 to two threads. We could acquire the arena lock before list_lock
745 to make it less likely that reused_arena picks this new arena,
746 but this could result in a deadlock with
747 __malloc_fork_lock_parent. */
748
749 __libc_lock_lock (a->mutex);
750
751 return a;
752}
753
754
755/* Remove an arena from free_list. */
756static mstate
757get_free_list (void)
758{
759 mstate replaced_arena = thread_arena;
760 mstate result = free_list;
761 if (result != NULL)
762 {
763 __libc_lock_lock (free_list_lock);
764 result = free_list;
765 if (result != NULL)
766 {
767 free_list = result->next_free;
768
769 /* The arena will be attached to this thread. */
770 assert (result->attached_threads == 0);
771 result->attached_threads = 1;
772
773 detach_arena (replaced_arena);
774 }
775 __libc_lock_unlock (free_list_lock);
776
777 if (result != NULL)
778 {
779 LIBC_PROBE (memory_arena_reuse_free_list, 1, result);
780 __libc_lock_lock (result->mutex);
781 thread_arena = result;
782 }
783 }
784
785 return result;
786}
787
788/* Remove the arena from the free list (if it is present).
789 free_list_lock must have been acquired by the caller. */
790static void
791remove_from_free_list (mstate arena)
792{
793 mstate *previous = &free_list;
794 for (mstate p = free_list; p != NULL; p = p->next_free)
795 {
796 assert (p->attached_threads == 0);
797 if (p == arena)
798 {
799 /* Remove the requested arena from the list. */
800 *previous = p->next_free;
801 break;
802 }
803 else
804 previous = &p->next_free;
805 }
806}
807
808/* Lock and return an arena that can be reused for memory allocation.
809 Avoid AVOID_ARENA as we have already failed to allocate memory in
810 it and it is currently locked. */
811static mstate
812reused_arena (mstate avoid_arena)
813{
814 mstate result;
815 /* FIXME: Access to next_to_use suffers from data races. */
816 static mstate next_to_use;
817 if (next_to_use == NULL)
818 next_to_use = &main_arena;
819
820 /* Iterate over all arenas (including those linked from
821 free_list). */
822 result = next_to_use;
823 do
824 {
825 if (!__libc_lock_trylock (result->mutex))
826 goto out;
827
828 /* FIXME: This is a data race, see _int_new_arena. */
829 result = result->next;
830 }
831 while (result != next_to_use);
832
833 /* Avoid AVOID_ARENA as we have already failed to allocate memory
834 in that arena and it is currently locked. */
835 if (result == avoid_arena)
836 result = result->next;
837
838 /* No arena available without contention. Wait for the next in line. */
839 LIBC_PROBE (memory_arena_reuse_wait, 3, &result->mutex, result, avoid_arena);
840 __libc_lock_lock (result->mutex);
841
842out:
843 /* Attach the arena to the current thread. */
844 {
845 /* Update the arena thread attachment counters. */
846 mstate replaced_arena = thread_arena;
847 __libc_lock_lock (free_list_lock);
848 detach_arena (replaced_arena);
849
850 /* We may have picked up an arena on the free list. We need to
851 preserve the invariant that no arena on the free list has a
852 positive attached_threads counter (otherwise,
853 arena_thread_freeres cannot use the counter to determine if the
854 arena needs to be put on the free list). We unconditionally
855 remove the selected arena from the free list. The caller of
856 reused_arena checked the free list and observed it to be empty,
857 so the list is very short. */
858 remove_from_free_list (result);
859
860 ++result->attached_threads;
861
862 __libc_lock_unlock (free_list_lock);
863 }
864
865 LIBC_PROBE (memory_arena_reuse, 2, result, avoid_arena);
866 thread_arena = result;
867 next_to_use = result->next;
868
869 return result;
870}
871
872static mstate
873arena_get2 (size_t size, mstate avoid_arena)
874{
875 mstate a;
876
877 static size_t narenas_limit;
878
879 a = get_free_list ();
880 if (a == NULL)
881 {
882 /* Nothing immediately available, so generate a new arena. */
883 if (narenas_limit == 0)
884 {
885 if (mp_.arena_max != 0)
886 narenas_limit = mp_.arena_max;
887 else if (narenas > mp_.arena_test)
888 {
889 int n = __get_nprocs ();
890
891 if (n >= 1)
892 narenas_limit = NARENAS_FROM_NCORES (n);
893 else
894 /* We have no information about the system. Assume two
895 cores. */
896 narenas_limit = NARENAS_FROM_NCORES (2);
897 }
898 }
899 repeat:;
900 size_t n = narenas;
901 /* NB: the following depends on the fact that (size_t)0 - 1 is a
902 very large number and that the underflow is OK. If arena_max
903 is set the value of arena_test is irrelevant. If arena_test
904 is set but narenas is not yet larger or equal to arena_test
905 narenas_limit is 0. There is no possibility for narenas to
906 be too big for the test to always fail since there is not
907 enough address space to create that many arenas. */
908 if (__glibc_unlikely (n <= narenas_limit - 1))
909 {
910 if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n))
911 goto repeat;
912 a = _int_new_arena (size);
913 if (__glibc_unlikely (a == NULL))
914 catomic_decrement (&narenas);
915 }
916 else
917 a = reused_arena (avoid_arena);
918 }
919 return a;
920}
921
922/* If we don't have the main arena, then maybe the failure is due to running
923 out of mmapped areas, so we can try allocating on the main arena.
924 Otherwise, it is likely that sbrk() has failed and there is still a chance
925 to mmap(), so try one of the other arenas. */
926static mstate
927arena_get_retry (mstate ar_ptr, size_t bytes)
928{
929 LIBC_PROBE (memory_arena_retry, 2, bytes, ar_ptr);
930 if (ar_ptr != &main_arena)
931 {
932 __libc_lock_unlock (ar_ptr->mutex);
933 ar_ptr = &main_arena;
934 __libc_lock_lock (ar_ptr->mutex);
935 }
936 else
937 {
938 __libc_lock_unlock (ar_ptr->mutex);
939 ar_ptr = arena_get2 (bytes, ar_ptr);
940 }
941
942 return ar_ptr;
943}
944
945void
946__malloc_arena_thread_freeres (void)
947{
948 /* Shut down the thread cache first. This could deallocate data for
949 the thread arena, so do this before we put the arena on the free
950 list. */
951 tcache_thread_shutdown ();
952
953 mstate a = thread_arena;
954 thread_arena = NULL;
955
956 if (a != NULL)
957 {
958 __libc_lock_lock (free_list_lock);
959 /* If this was the last attached thread for this arena, put the
960 arena on the free list. */
961 assert (a->attached_threads > 0);
962 if (--a->attached_threads == 0)
963 {
964 a->next_free = free_list;
965 free_list = a;
966 }
967 __libc_lock_unlock (free_list_lock);
968 }
969}
970
971/*
972 * Local variables:
973 * c-basic-offset: 2
974 * End:
975 */
976