malloc: Use __get_nprocs on arena_get2 (BZ 30945)
[glibc.git] / nptl / allocatestack.c
blob97d0efec101dc947434c327d5520c2aeaddf9ac2
1 /* Copyright (C) 2002-2023 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
9 The GNU C Library is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 Lesser General Public License for more details.
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library; if not, see
16 <https://www.gnu.org/licenses/>. */
18 #include <assert.h>
19 #include <errno.h>
20 #include <signal.h>
21 #include <stdint.h>
22 #include <string.h>
23 #include <unistd.h>
24 #include <sys/mman.h>
25 #include <sys/param.h>
26 #include <dl-sysdep.h>
27 #include <dl-tls.h>
28 #include <tls.h>
29 #include <list.h>
30 #include <lowlevellock.h>
31 #include <futex-internal.h>
32 #include <kernel-features.h>
33 #include <nptl-stack.h>
34 #include <libc-lock.h>
35 #include <tls-internal.h>
36 #include <intprops.h>
37 #include <setvmaname.h>
39 /* Default alignment of stack. */
40 #ifndef STACK_ALIGN
41 # define STACK_ALIGN __alignof__ (long double)
42 #endif
44 /* Default value for minimal stack size after allocating thread
45 descriptor and guard. */
46 #ifndef MINIMAL_REST_STACK
47 # define MINIMAL_REST_STACK 4096
48 #endif
51 /* Newer kernels have the MAP_STACK flag to indicate a mapping is used for
52 a stack. Use it when possible. */
53 #ifndef MAP_STACK
54 # define MAP_STACK 0
55 #endif
57 /* Get a stack frame from the cache. We have to match by size since
58 some blocks might be too small or far too large. */
59 static struct pthread *
60 get_cached_stack (size_t *sizep, void **memp)
62 size_t size = *sizep;
63 struct pthread *result = NULL;
64 list_t *entry;
66 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
68 /* Search the cache for a matching entry. We search for the
69 smallest stack which has at least the required size. Note that
70 in normal situations the size of all allocated stacks is the
71 same. As the very least there are only a few different sizes.
72 Therefore this loop will exit early most of the time with an
73 exact match. */
74 list_for_each (entry, &GL (dl_stack_cache))
76 struct pthread *curr;
78 curr = list_entry (entry, struct pthread, list);
79 if (__nptl_stack_in_use (curr) && curr->stackblock_size >= size)
81 if (curr->stackblock_size == size)
83 result = curr;
84 break;
87 if (result == NULL
88 || result->stackblock_size > curr->stackblock_size)
89 result = curr;
93 if (__builtin_expect (result == NULL, 0)
94 /* Make sure the size difference is not too excessive. In that
95 case we do not use the block. */
96 || __builtin_expect (result->stackblock_size > 4 * size, 0))
98 /* Release the lock. */
99 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
101 return NULL;
104 /* Don't allow setxid until cloned. */
105 result->setxid_futex = -1;
107 /* Dequeue the entry. */
108 __nptl_stack_list_del (&result->list);
110 /* And add to the list of stacks in use. */
111 __nptl_stack_list_add (&result->list, &GL (dl_stack_used));
113 /* And decrease the cache size. */
114 GL (dl_stack_cache_actsize) -= result->stackblock_size;
116 /* Release the lock early. */
117 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
119 /* Report size and location of the stack to the caller. */
120 *sizep = result->stackblock_size;
121 *memp = result->stackblock;
123 /* Cancellation handling is back to the default. */
124 result->cancelhandling = 0;
125 result->cleanup = NULL;
126 result->setup_failed = 0;
128 /* No pending event. */
129 result->nextevent = NULL;
131 result->exiting = false;
132 __libc_lock_init (result->exit_lock);
133 memset (&result->tls_state, 0, sizeof result->tls_state);
135 /* Clear the DTV. */
136 dtv_t *dtv = GET_DTV (TLS_TPADJ (result));
137 for (size_t cnt = 0; cnt < dtv[-1].counter; ++cnt)
138 free (dtv[1 + cnt].pointer.to_free);
139 memset (dtv, '\0', (dtv[-1].counter + 1) * sizeof (dtv_t));
141 /* Re-initialize the TLS. */
142 _dl_allocate_tls_init (TLS_TPADJ (result), true);
144 return result;
147 /* Return the guard page position on allocated stack. */
148 static inline char *
149 __attribute ((always_inline))
150 guard_position (void *mem, size_t size, size_t guardsize, struct pthread *pd,
151 size_t pagesize_m1)
153 #ifdef NEED_SEPARATE_REGISTER_STACK
154 return mem + (((size - guardsize) / 2) & ~pagesize_m1);
155 #elif _STACK_GROWS_DOWN
156 return mem;
157 #elif _STACK_GROWS_UP
158 return (char *) (((uintptr_t) pd - guardsize) & ~pagesize_m1);
159 #endif
162 /* Based on stack allocated with PROT_NONE, setup the required portions with
163 'prot' flags based on the guard page position. */
164 static inline int
165 setup_stack_prot (char *mem, size_t size, char *guard, size_t guardsize,
166 const int prot)
168 char *guardend = guard + guardsize;
169 #if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
170 /* As defined at guard_position, for architectures with downward stack
171 the guard page is always at start of the allocated area. */
172 if (__mprotect (guardend, size - guardsize, prot) != 0)
173 return errno;
174 #else
175 size_t mprots1 = (uintptr_t) guard - (uintptr_t) mem;
176 if (__mprotect (mem, mprots1, prot) != 0)
177 return errno;
178 size_t mprots2 = ((uintptr_t) mem + size) - (uintptr_t) guardend;
179 if (__mprotect (guardend, mprots2, prot) != 0)
180 return errno;
181 #endif
182 return 0;
185 /* Mark the memory of the stack as usable to the kernel. It frees everything
186 except for the space used for the TCB itself. */
187 static __always_inline void
188 advise_stack_range (void *mem, size_t size, uintptr_t pd, size_t guardsize)
190 uintptr_t sp = (uintptr_t) CURRENT_STACK_FRAME;
191 size_t pagesize_m1 = __getpagesize () - 1;
192 #if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
193 size_t freesize = (sp - (uintptr_t) mem) & ~pagesize_m1;
194 assert (freesize < size);
195 if (freesize > PTHREAD_STACK_MIN)
196 __madvise (mem, freesize - PTHREAD_STACK_MIN, MADV_DONTNEED);
197 #else
198 /* Page aligned start of memory to free (higher than or equal
199 to current sp plus the minimum stack size). */
200 uintptr_t freeblock = (sp + PTHREAD_STACK_MIN + pagesize_m1) & ~pagesize_m1;
201 uintptr_t free_end = (pd - guardsize) & ~pagesize_m1;
202 if (free_end > freeblock)
204 size_t freesize = free_end - freeblock;
205 assert (freesize < size);
206 __madvise ((void*) freeblock, freesize, MADV_DONTNEED);
208 #endif
211 /* Returns a usable stack for a new thread either by allocating a
212 new stack or reusing a cached stack of sufficient size.
213 ATTR must be non-NULL and point to a valid pthread_attr.
214 PDP must be non-NULL. */
215 static int
216 allocate_stack (const struct pthread_attr *attr, struct pthread **pdp,
217 void **stack, size_t *stacksize)
219 struct pthread *pd;
220 size_t size;
221 size_t pagesize_m1 = __getpagesize () - 1;
222 size_t tls_static_size_for_stack = __nptl_tls_static_size_for_stack ();
223 size_t tls_static_align_m1 = GLRO (dl_tls_static_align) - 1;
225 assert (powerof2 (pagesize_m1 + 1));
226 assert (TCB_ALIGNMENT >= STACK_ALIGN);
228 /* Get the stack size from the attribute if it is set. Otherwise we
229 use the default we determined at start time. */
230 if (attr->stacksize != 0)
231 size = attr->stacksize;
232 else
234 lll_lock (__default_pthread_attr_lock, LLL_PRIVATE);
235 size = __default_pthread_attr.internal.stacksize;
236 lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
239 /* Get memory for the stack. */
240 if (__glibc_unlikely (attr->flags & ATTR_FLAG_STACKADDR))
242 uintptr_t adj;
243 char *stackaddr = (char *) attr->stackaddr;
245 /* Assume the same layout as the _STACK_GROWS_DOWN case, with struct
246 pthread at the top of the stack block. Later we adjust the guard
247 location and stack address to match the _STACK_GROWS_UP case. */
248 if (_STACK_GROWS_UP)
249 stackaddr += attr->stacksize;
251 /* If the user also specified the size of the stack make sure it
252 is large enough. */
253 if (attr->stacksize != 0
254 && attr->stacksize < (tls_static_size_for_stack
255 + MINIMAL_REST_STACK))
256 return EINVAL;
258 /* Adjust stack size for alignment of the TLS block. */
259 #if TLS_TCB_AT_TP
260 adj = ((uintptr_t) stackaddr - TLS_TCB_SIZE)
261 & tls_static_align_m1;
262 assert (size > adj + TLS_TCB_SIZE);
263 #elif TLS_DTV_AT_TP
264 adj = ((uintptr_t) stackaddr - tls_static_size_for_stack)
265 & tls_static_align_m1;
266 assert (size > adj);
267 #endif
269 /* The user provided some memory. Let's hope it matches the
270 size... We do not allocate guard pages if the user provided
271 the stack. It is the user's responsibility to do this if it
272 is wanted. */
273 #if TLS_TCB_AT_TP
274 pd = (struct pthread *) ((uintptr_t) stackaddr
275 - TLS_TCB_SIZE - adj);
276 #elif TLS_DTV_AT_TP
277 pd = (struct pthread *) (((uintptr_t) stackaddr
278 - tls_static_size_for_stack - adj)
279 - TLS_PRE_TCB_SIZE);
280 #endif
282 /* The user provided stack memory needs to be cleared. */
283 memset (pd, '\0', sizeof (struct pthread));
285 /* The first TSD block is included in the TCB. */
286 pd->specific[0] = pd->specific_1stblock;
288 /* Remember the stack-related values. */
289 pd->stackblock = (char *) stackaddr - size;
290 pd->stackblock_size = size;
292 /* This is a user-provided stack. It will not be queued in the
293 stack cache nor will the memory (except the TLS memory) be freed. */
294 pd->user_stack = true;
296 /* This is at least the second thread. */
297 pd->header.multiple_threads = 1;
299 #ifdef NEED_DL_SYSINFO
300 SETUP_THREAD_SYSINFO (pd);
301 #endif
303 /* Don't allow setxid until cloned. */
304 pd->setxid_futex = -1;
306 /* Allocate the DTV for this thread. */
307 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
309 /* Something went wrong. */
310 assert (errno == ENOMEM);
311 return errno;
315 /* Prepare to modify global data. */
316 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
318 /* And add to the list of stacks in use. */
319 list_add (&pd->list, &GL (dl_stack_user));
321 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
323 else
325 /* Allocate some anonymous memory. If possible use the cache. */
326 size_t guardsize;
327 size_t reported_guardsize;
328 size_t reqsize;
329 void *mem;
330 const int prot = (PROT_READ | PROT_WRITE
331 | ((GL(dl_stack_flags) & PF_X) ? PROT_EXEC : 0));
333 /* Adjust the stack size for alignment. */
334 size &= ~tls_static_align_m1;
335 assert (size != 0);
337 /* Make sure the size of the stack is enough for the guard and
338 eventually the thread descriptor. On some targets there is
339 a minimum guard size requirement, ARCH_MIN_GUARD_SIZE, so
340 internally enforce it (unless the guard was disabled), but
341 report the original guard size for backward compatibility:
342 before POSIX 2008 the guardsize was specified to be one page
343 by default which is observable via pthread_attr_getguardsize
344 and pthread_getattr_np. */
345 guardsize = (attr->guardsize + pagesize_m1) & ~pagesize_m1;
346 reported_guardsize = guardsize;
347 if (guardsize > 0 && guardsize < ARCH_MIN_GUARD_SIZE)
348 guardsize = ARCH_MIN_GUARD_SIZE;
349 if (guardsize < attr->guardsize || size + guardsize < guardsize)
350 /* Arithmetic overflow. */
351 return EINVAL;
352 size += guardsize;
353 if (__builtin_expect (size < ((guardsize + tls_static_size_for_stack
354 + MINIMAL_REST_STACK + pagesize_m1)
355 & ~pagesize_m1),
357 /* The stack is too small (or the guard too large). */
358 return EINVAL;
360 /* Try to get a stack from the cache. */
361 reqsize = size;
362 pd = get_cached_stack (&size, &mem);
363 if (pd == NULL)
365 /* If a guard page is required, avoid committing memory by first
366 allocate with PROT_NONE and then reserve with required permission
367 excluding the guard page. */
368 mem = __mmap (NULL, size, (guardsize == 0) ? prot : PROT_NONE,
369 MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
371 if (__glibc_unlikely (mem == MAP_FAILED))
372 return errno;
374 /* Do madvise in case the tunable glibc.pthread.stack_hugetlb is
375 set to 0, disabling hugetlb. */
376 if (__glibc_unlikely (__nptl_stack_hugetlb == 0)
377 && __madvise (mem, size, MADV_NOHUGEPAGE) != 0)
378 return errno;
380 /* SIZE is guaranteed to be greater than zero.
381 So we can never get a null pointer back from mmap. */
382 assert (mem != NULL);
384 /* Place the thread descriptor at the end of the stack. */
385 #if TLS_TCB_AT_TP
386 pd = (struct pthread *) ((((uintptr_t) mem + size)
387 - TLS_TCB_SIZE)
388 & ~tls_static_align_m1);
389 #elif TLS_DTV_AT_TP
390 pd = (struct pthread *) ((((uintptr_t) mem + size
391 - tls_static_size_for_stack)
392 & ~tls_static_align_m1)
393 - TLS_PRE_TCB_SIZE);
394 #endif
396 /* Now mprotect the required region excluding the guard area. */
397 if (__glibc_likely (guardsize > 0))
399 char *guard = guard_position (mem, size, guardsize, pd,
400 pagesize_m1);
401 if (setup_stack_prot (mem, size, guard, guardsize, prot) != 0)
403 __munmap (mem, size);
404 return errno;
408 /* Remember the stack-related values. */
409 pd->stackblock = mem;
410 pd->stackblock_size = size;
411 /* Update guardsize for newly allocated guardsize to avoid
412 an mprotect in guard resize below. */
413 pd->guardsize = guardsize;
415 /* We allocated the first block thread-specific data array.
416 This address will not change for the lifetime of this
417 descriptor. */
418 pd->specific[0] = pd->specific_1stblock;
420 /* This is at least the second thread. */
421 pd->header.multiple_threads = 1;
423 #ifdef NEED_DL_SYSINFO
424 SETUP_THREAD_SYSINFO (pd);
425 #endif
427 /* Don't allow setxid until cloned. */
428 pd->setxid_futex = -1;
430 /* Allocate the DTV for this thread. */
431 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
433 /* Something went wrong. */
434 assert (errno == ENOMEM);
436 /* Free the stack memory we just allocated. */
437 (void) __munmap (mem, size);
439 return errno;
443 /* Prepare to modify global data. */
444 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
446 /* And add to the list of stacks in use. */
447 __nptl_stack_list_add (&pd->list, &GL (dl_stack_used));
449 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
452 /* There might have been a race. Another thread might have
453 caused the stacks to get exec permission while this new
454 stack was prepared. Detect if this was possible and
455 change the permission if necessary. */
456 if (__builtin_expect ((GL(dl_stack_flags) & PF_X) != 0
457 && (prot & PROT_EXEC) == 0, 0))
459 int err = __nptl_change_stack_perm (pd);
460 if (err != 0)
462 /* Free the stack memory we just allocated. */
463 (void) __munmap (mem, size);
465 return err;
470 /* Note that all of the stack and the thread descriptor is
471 zeroed. This means we do not have to initialize fields
472 with initial value zero. This is specifically true for
473 the 'tid' field which is always set back to zero once the
474 stack is not used anymore and for the 'guardsize' field
475 which will be read next. */
478 /* Create or resize the guard area if necessary. */
479 if (__glibc_unlikely (guardsize > pd->guardsize))
481 char *guard = guard_position (mem, size, guardsize, pd,
482 pagesize_m1);
483 if (__mprotect (guard, guardsize, PROT_NONE) != 0)
485 mprot_error:
486 lll_lock (GL (dl_stack_cache_lock), LLL_PRIVATE);
488 /* Remove the thread from the list. */
489 __nptl_stack_list_del (&pd->list);
491 lll_unlock (GL (dl_stack_cache_lock), LLL_PRIVATE);
493 /* Get rid of the TLS block we allocated. */
494 _dl_deallocate_tls (TLS_TPADJ (pd), false);
496 /* Free the stack memory regardless of whether the size
497 of the cache is over the limit or not. If this piece
498 of memory caused problems we better do not use it
499 anymore. Uh, and we ignore possible errors. There
500 is nothing we could do. */
501 (void) __munmap (mem, size);
503 return errno;
506 pd->guardsize = guardsize;
508 else if (__builtin_expect (pd->guardsize - guardsize > size - reqsize,
511 /* The old guard area is too large. */
513 #ifdef NEED_SEPARATE_REGISTER_STACK
514 char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1);
515 char *oldguard = mem + (((size - pd->guardsize) / 2) & ~pagesize_m1);
517 if (oldguard < guard
518 && __mprotect (oldguard, guard - oldguard, prot) != 0)
519 goto mprot_error;
521 if (__mprotect (guard + guardsize,
522 oldguard + pd->guardsize - guard - guardsize,
523 prot) != 0)
524 goto mprot_error;
525 #elif _STACK_GROWS_DOWN
526 if (__mprotect ((char *) mem + guardsize, pd->guardsize - guardsize,
527 prot) != 0)
528 goto mprot_error;
529 #elif _STACK_GROWS_UP
530 char *new_guard = (char *)(((uintptr_t) pd - guardsize)
531 & ~pagesize_m1);
532 char *old_guard = (char *)(((uintptr_t) pd - pd->guardsize)
533 & ~pagesize_m1);
534 /* The guard size difference might be > 0, but once rounded
535 to the nearest page the size difference might be zero. */
536 if (new_guard > old_guard
537 && __mprotect (old_guard, new_guard - old_guard, prot) != 0)
538 goto mprot_error;
539 #endif
541 pd->guardsize = guardsize;
543 /* The pthread_getattr_np() calls need to get passed the size
544 requested in the attribute, regardless of how large the
545 actually used guardsize is. */
546 pd->reported_guardsize = reported_guardsize;
549 /* Initialize the lock. We have to do this unconditionally since the
550 stillborn thread could be canceled while the lock is taken. */
551 pd->lock = LLL_LOCK_INITIALIZER;
553 /* The robust mutex lists also need to be initialized
554 unconditionally because the cleanup for the previous stack owner
555 might have happened in the kernel. */
556 pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock)
557 - offsetof (pthread_mutex_t,
558 __data.__list.__next));
559 pd->robust_head.list_op_pending = NULL;
560 #if __PTHREAD_MUTEX_HAVE_PREV
561 pd->robust_prev = &pd->robust_head;
562 #endif
563 pd->robust_head.list = &pd->robust_head;
565 /* We place the thread descriptor at the end of the stack. */
566 *pdp = pd;
568 void *stacktop;
570 #if TLS_TCB_AT_TP
571 /* The stack begins before the TCB and the static TLS block. */
572 stacktop = ((char *) (pd + 1) - tls_static_size_for_stack);
573 #elif TLS_DTV_AT_TP
574 stacktop = (char *) (pd - 1);
575 #endif
577 *stacksize = stacktop - pd->stackblock;
578 *stack = pd->stackblock;
580 return 0;
583 /* Maximum supported name from initial kernel support, not exported
584 by user API. */
585 #define ANON_VMA_NAME_MAX_LEN 80
587 #define SET_STACK_NAME(__prefix, __stack, __stacksize, __tid) \
588 ({ \
589 char __stack_name[sizeof (__prefix) + \
590 INT_BUFSIZE_BOUND (unsigned int)]; \
591 _Static_assert (sizeof __stack_name <= ANON_VMA_NAME_MAX_LEN, \
592 "VMA name size larger than maximum supported"); \
593 __snprintf (__stack_name, sizeof (__stack_name), __prefix "%u", \
594 (unsigned int) __tid); \
595 __set_vma_name (__stack, __stacksize, __stack_name); \
598 /* Add or remove an associated name to the PD VMA stack. */
599 static void
600 name_stack_maps (struct pthread *pd, bool set)
602 #if _STACK_GROWS_DOWN && !defined(NEED_SEPARATE_REGISTER_STACK)
603 void *stack = pd->stackblock + pd->guardsize;
604 #else
605 void *stack = pd->stackblock;
606 #endif
607 size_t stacksize = pd->stackblock_size - pd->guardsize;
609 if (!set)
610 __set_vma_name (stack, stacksize, NULL);
611 else
613 unsigned int tid = pd->tid;
614 if (pd->user_stack)
615 SET_STACK_NAME (" glibc: pthread user stack: ", stack, stacksize, tid);
616 else
617 SET_STACK_NAME (" glibc: pthread stack: ", stack, stacksize, tid);