1 | /* Malloc implementation for multiple threads without lock contention. |
2 | Copyright (C) 2001-2023 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 | |
47 | static inline size_t |
48 | heap_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 | |
57 | static inline size_t |
58 | heap_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 | |
75 | typedef 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. */ |
91 | extern int sanity_check_heap_info_alignment[(sizeof (heap_info) |
92 | + 2 * SIZE_SZ) % MALLOC_ALIGNMENT |
93 | ? -1 : 1]; |
94 | |
95 | /* Thread specific data. */ |
96 | |
97 | static __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) |
106 | static size_t narenas = 1; |
107 | #endif |
108 | static 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? */ |
125 | static 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 | |
152 | static inline heap_info * |
153 | heap_for_ptr (void *ptr) |
154 | { |
155 | size_t max_size = heap_max_size (); |
156 | return PTR_ALIGN_DOWN (ptr, max_size); |
157 | } |
158 | |
159 | static inline struct malloc_state * |
160 | arena_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 | |
176 | void |
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 | |
196 | void |
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 | |
212 | void |
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) \ |
244 | static inline int do_ ## __name (__type value); \ |
245 | static void \ |
246 | TUNABLE_CALLBACK (__name) (tunable_val_t *valp) \ |
247 | { \ |
248 | __type value = (__type) (valp)->numval; \ |
249 | do_ ## __name (value); \ |
250 | } |
251 | |
252 | TUNABLE_CALLBACK_FNDECL (set_mmap_threshold, size_t) |
253 | TUNABLE_CALLBACK_FNDECL (set_mmaps_max, int32_t) |
254 | TUNABLE_CALLBACK_FNDECL (set_top_pad, size_t) |
255 | TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t) |
256 | TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t) |
257 | TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t) |
258 | TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t) |
259 | #if USE_TCACHE |
260 | TUNABLE_CALLBACK_FNDECL (set_tcache_max, size_t) |
261 | TUNABLE_CALLBACK_FNDECL (set_tcache_count, size_t) |
262 | TUNABLE_CALLBACK_FNDECL (set_tcache_unsorted_limit, size_t) |
263 | #endif |
264 | TUNABLE_CALLBACK_FNDECL (set_mxfast, size_t) |
265 | TUNABLE_CALLBACK_FNDECL (set_hugetlb, size_t) |
266 | #else |
267 | /* Initialization routine. */ |
268 | #include <string.h> |
269 | extern char **_environ; |
270 | |
271 | static char * |
272 | next_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 | #if USE_TCACHE |
304 | static void tcache_key_initialize (void); |
305 | #endif |
306 | |
307 | static void |
308 | ptmalloc_init (void) |
309 | { |
310 | if (__malloc_initialized) |
311 | return; |
312 | |
313 | __malloc_initialized = true; |
314 | |
315 | #if USE_TCACHE |
316 | tcache_key_initialize (); |
317 | #endif |
318 | |
319 | #ifdef USE_MTAG |
320 | if ((TUNABLE_GET_FULL (glibc, mem, tagging, int32_t, NULL) & 1) != 0) |
321 | { |
322 | /* If the tunable says that we should be using tagged memory |
323 | and that morecore does not support tagged regions, then |
324 | disable it. */ |
325 | if (__MTAG_SBRK_UNTAGGED) |
326 | __always_fail_morecore = true; |
327 | |
328 | mtag_enabled = true; |
329 | mtag_mmap_flags = __MTAG_MMAP_FLAGS; |
330 | } |
331 | #endif |
332 | |
333 | #if defined SHARED && IS_IN (libc) |
334 | /* In case this libc copy is in a non-default namespace, never use |
335 | brk. Likewise if dlopened from statically linked program. The |
336 | generic sbrk implementation also enforces this, but it is not |
337 | used on Hurd. */ |
338 | if (!__libc_initial) |
339 | __always_fail_morecore = true; |
340 | #endif |
341 | |
342 | thread_arena = &main_arena; |
343 | |
344 | malloc_init_state (&main_arena); |
345 | |
346 | #if HAVE_TUNABLES |
347 | TUNABLE_GET (top_pad, size_t, TUNABLE_CALLBACK (set_top_pad)); |
348 | TUNABLE_GET (perturb, int32_t, TUNABLE_CALLBACK (set_perturb_byte)); |
349 | TUNABLE_GET (mmap_threshold, size_t, TUNABLE_CALLBACK (set_mmap_threshold)); |
350 | TUNABLE_GET (trim_threshold, size_t, TUNABLE_CALLBACK (set_trim_threshold)); |
351 | TUNABLE_GET (mmap_max, int32_t, TUNABLE_CALLBACK (set_mmaps_max)); |
352 | TUNABLE_GET (arena_max, size_t, TUNABLE_CALLBACK (set_arena_max)); |
353 | TUNABLE_GET (arena_test, size_t, TUNABLE_CALLBACK (set_arena_test)); |
354 | # if USE_TCACHE |
355 | TUNABLE_GET (tcache_max, size_t, TUNABLE_CALLBACK (set_tcache_max)); |
356 | TUNABLE_GET (tcache_count, size_t, TUNABLE_CALLBACK (set_tcache_count)); |
357 | TUNABLE_GET (tcache_unsorted_limit, size_t, |
358 | TUNABLE_CALLBACK (set_tcache_unsorted_limit)); |
359 | # endif |
360 | TUNABLE_GET (mxfast, size_t, TUNABLE_CALLBACK (set_mxfast)); |
361 | TUNABLE_GET (hugetlb, size_t, TUNABLE_CALLBACK (set_hugetlb)); |
362 | if (mp_.hp_pagesize > 0) |
363 | /* Force mmap for main arena instead of sbrk, so hugepages are explicitly |
364 | used. */ |
365 | __always_fail_morecore = true; |
366 | #else |
367 | if (__glibc_likely (_environ != NULL)) |
368 | { |
369 | char **runp = _environ; |
370 | char *envline; |
371 | |
372 | while (__builtin_expect ((envline = next_env_entry (&runp)) != NULL, |
373 | 0)) |
374 | { |
375 | size_t len = strcspn (envline, "=" ); |
376 | |
377 | if (envline[len] != '=') |
378 | /* This is a "MALLOC_" variable at the end of the string |
379 | without a '=' character. Ignore it since otherwise we |
380 | will access invalid memory below. */ |
381 | continue; |
382 | |
383 | switch (len) |
384 | { |
385 | case 8: |
386 | if (!__builtin_expect (__libc_enable_secure, 0)) |
387 | { |
388 | if (memcmp (envline, "TOP_PAD_" , 8) == 0) |
389 | __libc_mallopt (M_TOP_PAD, strtol (&envline[9], NULL, 10)); |
390 | else if (memcmp (envline, "PERTURB_" , 8) == 0) |
391 | __libc_mallopt (M_PERTURB, strtol (&envline[9], NULL, 10)); |
392 | } |
393 | break; |
394 | case 9: |
395 | if (!__builtin_expect (__libc_enable_secure, 0)) |
396 | { |
397 | if (memcmp (envline, "MMAP_MAX_" , 9) == 0) |
398 | __libc_mallopt (M_MMAP_MAX, strtol (&envline[10], |
399 | NULL, 10)); |
400 | else if (memcmp (envline, "ARENA_MAX" , 9) == 0) |
401 | __libc_mallopt (M_ARENA_MAX, strtol (&envline[10], |
402 | NULL, 10)); |
403 | } |
404 | break; |
405 | case 10: |
406 | if (!__builtin_expect (__libc_enable_secure, 0)) |
407 | { |
408 | if (memcmp (envline, "ARENA_TEST" , 10) == 0) |
409 | __libc_mallopt (M_ARENA_TEST, strtol (&envline[11], |
410 | NULL, 10)); |
411 | } |
412 | break; |
413 | case 15: |
414 | if (!__builtin_expect (__libc_enable_secure, 0)) |
415 | { |
416 | if (memcmp (envline, "TRIM_THRESHOLD_" , 15) == 0) |
417 | __libc_mallopt (M_TRIM_THRESHOLD, strtol (&envline[16], |
418 | NULL, 10)); |
419 | else if (memcmp (envline, "MMAP_THRESHOLD_" , 15) == 0) |
420 | __libc_mallopt (M_MMAP_THRESHOLD, strtol (&envline[16], |
421 | NULL, 10)); |
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 | |
438 | static void |
439 | dump_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) (((uintptr_t) 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. */ |
476 | static char *aligned_heap_area; |
477 | |
478 | /* Create a new heap. size is automatically rounded up to a multiple |
479 | of the page size. */ |
480 | |
481 | static heap_info * |
482 | alloc_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 (p2, 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 *) (((uintptr_t) p1 + (max_size - 1)) |
522 | & ~(max_size - 1)); |
523 | ul = p2 - p1; |
524 | if (ul) |
525 | __munmap (p1, ul); |
526 | else |
527 | aligned_heap_area = p2 + max_size; |
528 | __munmap (p2 + max_size, 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 (p2, max_size); |
541 | return 0; |
542 | } |
543 | } |
544 | } |
545 | if (__mprotect (p2, size, mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0) |
546 | { |
547 | __munmap (p2, max_size); |
548 | return 0; |
549 | } |
550 | |
551 | madvise_thp (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 | |
561 | static heap_info * |
562 | new_heap (size_t size, size_t top_pad) |
563 | { |
564 | #if HAVE_TUNABLES |
565 | if (__glibc_unlikely (mp_.hp_pagesize != 0)) |
566 | { |
567 | heap_info *h = alloc_new_heap (size, top_pad, mp_.hp_pagesize, |
568 | mp_.hp_flags); |
569 | if (h != NULL) |
570 | return h; |
571 | } |
572 | #endif |
573 | return alloc_new_heap (size, top_pad, GLRO (dl_pagesize), 0); |
574 | } |
575 | |
576 | /* Grow a heap. size is automatically rounded up to a |
577 | multiple of the page size. */ |
578 | |
579 | static int |
580 | grow_heap (heap_info *h, long diff) |
581 | { |
582 | size_t pagesize = h->pagesize; |
583 | size_t max_size = heap_max_size (); |
584 | long new_size; |
585 | |
586 | diff = ALIGN_UP (diff, pagesize); |
587 | new_size = (long) h->size + diff; |
588 | if ((unsigned long) new_size > (unsigned long) max_size) |
589 | return -1; |
590 | |
591 | if ((unsigned long) new_size > h->mprotect_size) |
592 | { |
593 | if (__mprotect ((char *) h + h->mprotect_size, |
594 | (unsigned long) new_size - h->mprotect_size, |
595 | mtag_mmap_flags | PROT_READ | PROT_WRITE) != 0) |
596 | return -2; |
597 | |
598 | h->mprotect_size = new_size; |
599 | } |
600 | |
601 | h->size = new_size; |
602 | LIBC_PROBE (memory_heap_more, 2, h, h->size); |
603 | return 0; |
604 | } |
605 | |
606 | /* Shrink a heap. */ |
607 | |
608 | static int |
609 | shrink_heap (heap_info *h, long diff) |
610 | { |
611 | long new_size; |
612 | |
613 | new_size = (long) h->size - diff; |
614 | if (new_size < (long) sizeof (*h)) |
615 | return -1; |
616 | |
617 | /* Try to re-map the extra heap space freshly to save memory, and make it |
618 | inaccessible. See malloc-sysdep.h to know when this is true. */ |
619 | if (__glibc_unlikely (check_may_shrink_heap ())) |
620 | { |
621 | if ((char *) MMAP ((char *) h + new_size, diff, PROT_NONE, |
622 | MAP_FIXED) == (char *) MAP_FAILED) |
623 | return -2; |
624 | |
625 | h->mprotect_size = new_size; |
626 | } |
627 | else |
628 | __madvise ((char *) h + new_size, diff, MADV_DONTNEED); |
629 | /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/ |
630 | |
631 | h->size = new_size; |
632 | LIBC_PROBE (memory_heap_less, 2, h, h->size); |
633 | return 0; |
634 | } |
635 | |
636 | /* Delete a heap. */ |
637 | |
638 | static int |
639 | heap_trim (heap_info *heap, size_t pad) |
640 | { |
641 | mstate ar_ptr = heap->ar_ptr; |
642 | mchunkptr top_chunk = top (ar_ptr), p; |
643 | heap_info *prev_heap; |
644 | long new_size, top_size, top_area, , prev_size, misalign; |
645 | size_t max_size = heap_max_size (); |
646 | |
647 | /* Can this heap go away completely? */ |
648 | while (top_chunk == chunk_at_offset (heap, sizeof (*heap))) |
649 | { |
650 | prev_heap = heap->prev; |
651 | prev_size = prev_heap->size - (MINSIZE - 2 * SIZE_SZ); |
652 | p = chunk_at_offset (prev_heap, prev_size); |
653 | /* fencepost must be properly aligned. */ |
654 | misalign = ((long) p) & MALLOC_ALIGN_MASK; |
655 | p = chunk_at_offset (prev_heap, prev_size - misalign); |
656 | assert (chunksize_nomask (p) == (0 | PREV_INUSE)); /* must be fencepost */ |
657 | p = prev_chunk (p); |
658 | new_size = chunksize (p) + (MINSIZE - 2 * SIZE_SZ) + misalign; |
659 | assert (new_size > 0 && new_size < (long) (2 * MINSIZE)); |
660 | if (!prev_inuse (p)) |
661 | new_size += prev_size (p); |
662 | assert (new_size > 0 && new_size < max_size); |
663 | if (new_size + (max_size - prev_heap->size) < pad + MINSIZE |
664 | + heap->pagesize) |
665 | break; |
666 | ar_ptr->system_mem -= heap->size; |
667 | LIBC_PROBE (memory_heap_free, 2, heap, heap->size); |
668 | if ((char *) heap + max_size == aligned_heap_area) |
669 | aligned_heap_area = NULL; |
670 | __munmap (heap, max_size); |
671 | heap = prev_heap; |
672 | if (!prev_inuse (p)) /* consolidate backward */ |
673 | { |
674 | p = prev_chunk (p); |
675 | unlink_chunk (ar_ptr, p); |
676 | } |
677 | assert (((unsigned long) ((char *) p + new_size) & (heap->pagesize - 1)) |
678 | == 0); |
679 | assert (((char *) p + new_size) == ((char *) heap + heap->size)); |
680 | top (ar_ptr) = top_chunk = p; |
681 | set_head (top_chunk, new_size | PREV_INUSE); |
682 | /*check_chunk(ar_ptr, top_chunk);*/ |
683 | } |
684 | |
685 | /* Uses similar logic for per-thread arenas as the main arena with systrim |
686 | and _int_free by preserving the top pad and rounding down to the nearest |
687 | page. */ |
688 | top_size = chunksize (top_chunk); |
689 | if ((unsigned long)(top_size) < |
690 | (unsigned long)(mp_.trim_threshold)) |
691 | return 0; |
692 | |
693 | top_area = top_size - MINSIZE - 1; |
694 | if (top_area < 0 || (size_t) top_area <= pad) |
695 | return 0; |
696 | |
697 | /* Release in pagesize units and round down to the nearest page. */ |
698 | extra = ALIGN_DOWN(top_area - pad, heap->pagesize); |
699 | if (extra == 0) |
700 | return 0; |
701 | |
702 | /* Try to shrink. */ |
703 | if (shrink_heap (heap, extra) != 0) |
704 | return 0; |
705 | |
706 | ar_ptr->system_mem -= extra; |
707 | |
708 | /* Success. Adjust top accordingly. */ |
709 | set_head (top_chunk, (top_size - extra) | PREV_INUSE); |
710 | /*check_chunk(ar_ptr, top_chunk);*/ |
711 | return 1; |
712 | } |
713 | |
714 | /* Create a new arena with initial size "size". */ |
715 | |
716 | #if IS_IN (libc) |
717 | /* If REPLACED_ARENA is not NULL, detach it from this thread. Must be |
718 | called while free_list_lock is held. */ |
719 | static void |
720 | detach_arena (mstate replaced_arena) |
721 | { |
722 | if (replaced_arena != NULL) |
723 | { |
724 | assert (replaced_arena->attached_threads > 0); |
725 | /* The current implementation only detaches from main_arena in |
726 | case of allocation failure. This means that it is likely not |
727 | beneficial to put the arena on free_list even if the |
728 | reference count reaches zero. */ |
729 | --replaced_arena->attached_threads; |
730 | } |
731 | } |
732 | |
733 | static mstate |
734 | _int_new_arena (size_t size) |
735 | { |
736 | mstate a; |
737 | heap_info *h; |
738 | char *ptr; |
739 | unsigned long misalign; |
740 | |
741 | h = new_heap (size + (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT), |
742 | mp_.top_pad); |
743 | if (!h) |
744 | { |
745 | /* Maybe size is too large to fit in a single heap. So, just try |
746 | to create a minimally-sized arena and let _int_malloc() attempt |
747 | to deal with the large request via mmap_chunk(). */ |
748 | h = new_heap (sizeof (*h) + sizeof (*a) + MALLOC_ALIGNMENT, mp_.top_pad); |
749 | if (!h) |
750 | return 0; |
751 | } |
752 | a = h->ar_ptr = (mstate) (h + 1); |
753 | malloc_init_state (a); |
754 | a->attached_threads = 1; |
755 | /*a->next = NULL;*/ |
756 | a->system_mem = a->max_system_mem = h->size; |
757 | |
758 | /* Set up the top chunk, with proper alignment. */ |
759 | ptr = (char *) (a + 1); |
760 | misalign = (uintptr_t) chunk2mem (ptr) & MALLOC_ALIGN_MASK; |
761 | if (misalign > 0) |
762 | ptr += MALLOC_ALIGNMENT - misalign; |
763 | top (a) = (mchunkptr) ptr; |
764 | set_head (top (a), (((char *) h + h->size) - ptr) | PREV_INUSE); |
765 | |
766 | LIBC_PROBE (memory_arena_new, 2, a, size); |
767 | mstate replaced_arena = thread_arena; |
768 | thread_arena = a; |
769 | __libc_lock_init (a->mutex); |
770 | |
771 | __libc_lock_lock (list_lock); |
772 | |
773 | /* Add the new arena to the global list. */ |
774 | a->next = main_arena.next; |
775 | /* FIXME: The barrier is an attempt to synchronize with read access |
776 | in reused_arena, which does not acquire list_lock while |
777 | traversing the list. */ |
778 | atomic_write_barrier (); |
779 | main_arena.next = a; |
780 | |
781 | __libc_lock_unlock (list_lock); |
782 | |
783 | __libc_lock_lock (free_list_lock); |
784 | detach_arena (replaced_arena); |
785 | __libc_lock_unlock (free_list_lock); |
786 | |
787 | /* Lock this arena. NB: Another thread may have been attached to |
788 | this arena because the arena is now accessible from the |
789 | main_arena.next list and could have been picked by reused_arena. |
790 | This can only happen for the last arena created (before the arena |
791 | limit is reached). At this point, some arena has to be attached |
792 | to two threads. We could acquire the arena lock before list_lock |
793 | to make it less likely that reused_arena picks this new arena, |
794 | but this could result in a deadlock with |
795 | __malloc_fork_lock_parent. */ |
796 | |
797 | __libc_lock_lock (a->mutex); |
798 | |
799 | return a; |
800 | } |
801 | |
802 | |
803 | /* Remove an arena from free_list. */ |
804 | static mstate |
805 | get_free_list (void) |
806 | { |
807 | mstate replaced_arena = thread_arena; |
808 | mstate result = free_list; |
809 | if (result != NULL) |
810 | { |
811 | __libc_lock_lock (free_list_lock); |
812 | result = free_list; |
813 | if (result != NULL) |
814 | { |
815 | free_list = result->next_free; |
816 | |
817 | /* The arena will be attached to this thread. */ |
818 | assert (result->attached_threads == 0); |
819 | result->attached_threads = 1; |
820 | |
821 | detach_arena (replaced_arena); |
822 | } |
823 | __libc_lock_unlock (free_list_lock); |
824 | |
825 | if (result != NULL) |
826 | { |
827 | LIBC_PROBE (memory_arena_reuse_free_list, 1, result); |
828 | __libc_lock_lock (result->mutex); |
829 | thread_arena = result; |
830 | } |
831 | } |
832 | |
833 | return result; |
834 | } |
835 | |
836 | /* Remove the arena from the free list (if it is present). |
837 | free_list_lock must have been acquired by the caller. */ |
838 | static void |
839 | remove_from_free_list (mstate arena) |
840 | { |
841 | mstate *previous = &free_list; |
842 | for (mstate p = free_list; p != NULL; p = p->next_free) |
843 | { |
844 | assert (p->attached_threads == 0); |
845 | if (p == arena) |
846 | { |
847 | /* Remove the requested arena from the list. */ |
848 | *previous = p->next_free; |
849 | break; |
850 | } |
851 | else |
852 | previous = &p->next_free; |
853 | } |
854 | } |
855 | |
856 | /* Lock and return an arena that can be reused for memory allocation. |
857 | Avoid AVOID_ARENA as we have already failed to allocate memory in |
858 | it and it is currently locked. */ |
859 | static mstate |
860 | reused_arena (mstate avoid_arena) |
861 | { |
862 | mstate result; |
863 | /* FIXME: Access to next_to_use suffers from data races. */ |
864 | static mstate next_to_use; |
865 | if (next_to_use == NULL) |
866 | next_to_use = &main_arena; |
867 | |
868 | /* Iterate over all arenas (including those linked from |
869 | free_list). */ |
870 | result = next_to_use; |
871 | do |
872 | { |
873 | if (!__libc_lock_trylock (result->mutex)) |
874 | goto out; |
875 | |
876 | /* FIXME: This is a data race, see _int_new_arena. */ |
877 | result = result->next; |
878 | } |
879 | while (result != next_to_use); |
880 | |
881 | /* Avoid AVOID_ARENA as we have already failed to allocate memory |
882 | in that arena and it is currently locked. */ |
883 | if (result == avoid_arena) |
884 | result = result->next; |
885 | |
886 | /* No arena available without contention. Wait for the next in line. */ |
887 | LIBC_PROBE (memory_arena_reuse_wait, 3, &result->mutex, result, avoid_arena); |
888 | __libc_lock_lock (result->mutex); |
889 | |
890 | out: |
891 | /* Attach the arena to the current thread. */ |
892 | { |
893 | /* Update the arena thread attachment counters. */ |
894 | mstate replaced_arena = thread_arena; |
895 | __libc_lock_lock (free_list_lock); |
896 | detach_arena (replaced_arena); |
897 | |
898 | /* We may have picked up an arena on the free list. We need to |
899 | preserve the invariant that no arena on the free list has a |
900 | positive attached_threads counter (otherwise, |
901 | arena_thread_freeres cannot use the counter to determine if the |
902 | arena needs to be put on the free list). We unconditionally |
903 | remove the selected arena from the free list. The caller of |
904 | reused_arena checked the free list and observed it to be empty, |
905 | so the list is very short. */ |
906 | remove_from_free_list (result); |
907 | |
908 | ++result->attached_threads; |
909 | |
910 | __libc_lock_unlock (free_list_lock); |
911 | } |
912 | |
913 | LIBC_PROBE (memory_arena_reuse, 2, result, avoid_arena); |
914 | thread_arena = result; |
915 | next_to_use = result->next; |
916 | |
917 | return result; |
918 | } |
919 | |
920 | static mstate |
921 | arena_get2 (size_t size, mstate avoid_arena) |
922 | { |
923 | mstate a; |
924 | |
925 | static size_t narenas_limit; |
926 | |
927 | a = get_free_list (); |
928 | if (a == NULL) |
929 | { |
930 | /* Nothing immediately available, so generate a new arena. */ |
931 | if (narenas_limit == 0) |
932 | { |
933 | if (mp_.arena_max != 0) |
934 | narenas_limit = mp_.arena_max; |
935 | else if (narenas > mp_.arena_test) |
936 | { |
937 | int n = __get_nprocs_sched (); |
938 | |
939 | if (n >= 1) |
940 | narenas_limit = NARENAS_FROM_NCORES (n); |
941 | else |
942 | /* We have no information about the system. Assume two |
943 | cores. */ |
944 | narenas_limit = NARENAS_FROM_NCORES (2); |
945 | } |
946 | } |
947 | repeat:; |
948 | size_t n = narenas; |
949 | /* NB: the following depends on the fact that (size_t)0 - 1 is a |
950 | very large number and that the underflow is OK. If arena_max |
951 | is set the value of arena_test is irrelevant. If arena_test |
952 | is set but narenas is not yet larger or equal to arena_test |
953 | narenas_limit is 0. There is no possibility for narenas to |
954 | be too big for the test to always fail since there is not |
955 | enough address space to create that many arenas. */ |
956 | if (__glibc_unlikely (n <= narenas_limit - 1)) |
957 | { |
958 | if (catomic_compare_and_exchange_bool_acq (&narenas, n + 1, n)) |
959 | goto repeat; |
960 | a = _int_new_arena (size); |
961 | if (__glibc_unlikely (a == NULL)) |
962 | catomic_decrement (&narenas); |
963 | } |
964 | else |
965 | a = reused_arena (avoid_arena); |
966 | } |
967 | return a; |
968 | } |
969 | |
970 | /* If we don't have the main arena, then maybe the failure is due to running |
971 | out of mmapped areas, so we can try allocating on the main arena. |
972 | Otherwise, it is likely that sbrk() has failed and there is still a chance |
973 | to mmap(), so try one of the other arenas. */ |
974 | static mstate |
975 | arena_get_retry (mstate ar_ptr, size_t bytes) |
976 | { |
977 | LIBC_PROBE (memory_arena_retry, 2, bytes, ar_ptr); |
978 | if (ar_ptr != &main_arena) |
979 | { |
980 | __libc_lock_unlock (ar_ptr->mutex); |
981 | ar_ptr = &main_arena; |
982 | __libc_lock_lock (ar_ptr->mutex); |
983 | } |
984 | else |
985 | { |
986 | __libc_lock_unlock (ar_ptr->mutex); |
987 | ar_ptr = arena_get2 (bytes, ar_ptr); |
988 | } |
989 | |
990 | return ar_ptr; |
991 | } |
992 | #endif |
993 | |
994 | void |
995 | __malloc_arena_thread_freeres (void) |
996 | { |
997 | /* Shut down the thread cache first. This could deallocate data for |
998 | the thread arena, so do this before we put the arena on the free |
999 | list. */ |
1000 | tcache_thread_shutdown (); |
1001 | |
1002 | mstate a = thread_arena; |
1003 | thread_arena = NULL; |
1004 | |
1005 | if (a != NULL) |
1006 | { |
1007 | __libc_lock_lock (free_list_lock); |
1008 | /* If this was the last attached thread for this arena, put the |
1009 | arena on the free list. */ |
1010 | assert (a->attached_threads > 0); |
1011 | if (--a->attached_threads == 0) |
1012 | { |
1013 | a->next_free = free_list; |
1014 | free_list = a; |
1015 | } |
1016 | __libc_lock_unlock (free_list_lock); |
1017 | } |
1018 | } |
1019 | |
1020 | /* |
1021 | * Local variables: |
1022 | * c-basic-offset: 2 |
1023 | * End: |
1024 | */ |
1025 | |