1 /* Copyright (C) 2002, 2003, 2004 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Ulrich Drepper <drepper@redhat.com>, 2002.
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
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
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.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, write to the Free
17 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
26 #include <sys/param.h>
27 #include <dl-sysdep.h>
32 #ifndef NEED_SEPARATE_REGISTER_STACK
34 /* Most architectures have exactly one stack pointer. Some have more. */
35 # define STACK_VARIABLES void *stackaddr
37 /* How to pass the values to the 'create_thread' function. */
38 # define STACK_VARIABLES_ARGS stackaddr
40 /* How to declare function which gets there parameters. */
41 # define STACK_VARIABLES_PARMS void *stackaddr
43 /* How to declare allocate_stack. */
44 # define ALLOCATE_STACK_PARMS void **stack
46 /* This is how the function is called. We do it this way to allow
47 other variants of the function to have more parameters. */
48 # define ALLOCATE_STACK(attr, pd) allocate_stack (attr, pd, &stackaddr)
52 /* We need two stacks. The kernel will place them but we have to tell
53 the kernel about the size of the reserved address space. */
54 # define STACK_VARIABLES void *stackaddr; size_t stacksize
56 /* How to pass the values to the 'create_thread' function. */
57 # define STACK_VARIABLES_ARGS stackaddr, stacksize
59 /* How to declare function which gets there parameters. */
60 # define STACK_VARIABLES_PARMS void *stackaddr, size_t stacksize
62 /* How to declare allocate_stack. */
63 # define ALLOCATE_STACK_PARMS void **stack, size_t *stacksize
65 /* This is how the function is called. We do it this way to allow
66 other variants of the function to have more parameters. */
67 # define ALLOCATE_STACK(attr, pd) \
68 allocate_stack (attr, pd, &stackaddr, &stacksize)
73 /* Default alignment of stack. */
75 # define STACK_ALIGN __alignof__ (long double)
78 /* Default value for minimal stack size after allocating thread
79 descriptor and guard. */
80 #ifndef MINIMAL_REST_STACK
81 # define MINIMAL_REST_STACK 4096
85 /* Let the architecture add some flags to the mmap() call used to
87 #ifndef ARCH_MAP_FLAGS
88 # define ARCH_MAP_FLAGS 0
91 /* This yields the pointer that TLS support code calls the thread pointer. */
93 # define TLS_TPADJ(pd) (pd)
95 # define TLS_TPADJ(pd) ((struct pthread *)((char *) (pd) + TLS_PRE_TCB_SIZE))
98 /* Cache handling for not-yet free stacks. */
100 /* Maximum size in kB of cache. */
101 static size_t stack_cache_maxsize = 40 * 1024 * 1024; /* 40MiBi by default. */
102 static size_t stack_cache_actsize;
104 /* Mutex protecting this variable. */
105 static lll_lock_t stack_cache_lock = LLL_LOCK_INITIALIZER;
107 /* List of queued stack frames. */
108 static LIST_HEAD (stack_cache);
110 /* List of the stacks in use. */
111 static LIST_HEAD (stack_used);
113 /* List of the threads with user provided stacks in use. No need to
114 initialize this, since it's done in __pthread_initialize_minimal. */
115 list_t __stack_user __attribute__ ((nocommon));
116 hidden_data_def (__stack_user)
118 #if COLORING_INCREMENT != 0
119 /* Number of threads created. */
120 static unsigned int nptl_ncreated;
124 /* Check whether the stack is still used or not. */
125 #define FREE_P(descr) ((descr)->tid <= 0)
128 /* We create a double linked list of all cache entries. Double linked
129 because this allows removing entries from the end. */
132 /* Get a stack frame from the cache. We have to match by size since
133 some blocks might be too small or far too large. */
134 static struct pthread *
135 get_cached_stack (size_t *sizep, void **memp)
137 size_t size = *sizep;
138 struct pthread *result = NULL;
141 lll_lock (stack_cache_lock);
143 /* Search the cache for a matching entry. We search for the
144 smallest stack which has at least the required size. Note that
145 in normal situations the size of all allocated stacks is the
146 same. As the very least there are only a few different sizes.
147 Therefore this loop will exit early most of the time with an
149 list_for_each (entry, &stack_cache)
151 struct pthread *curr;
153 curr = list_entry (entry, struct pthread, list);
154 if (FREE_P (curr) && curr->stackblock_size >= size)
156 if (curr->stackblock_size == size)
163 || result->stackblock_size > curr->stackblock_size)
168 if (__builtin_expect (result == NULL, 0)
169 /* Make sure the size difference is not too excessive. In that
170 case we do not use the block. */
171 || __builtin_expect (result->stackblock_size > 4 * size, 0))
173 /* Release the lock. */
174 lll_unlock (stack_cache_lock);
179 /* Dequeue the entry. */
180 list_del (&result->list);
182 /* And add to the list of stacks in use. */
183 list_add (&result->list, &stack_used);
185 /* And decrease the cache size. */
186 stack_cache_actsize -= result->stackblock_size;
188 /* Release the lock early. */
189 lll_unlock (stack_cache_lock);
191 /* Report size and location of the stack to the caller. */
192 *sizep = result->stackblock_size;
193 *memp = result->stackblock;
195 /* Cancellation handling is back to the default. */
196 result->cancelhandling = 0;
197 result->cleanup = NULL;
199 /* No pending event. */
200 result->nextevent = NULL;
203 dtv_t *dtv = GET_DTV (TLS_TPADJ (result));
204 memset (dtv, '\0', (dtv[-1].counter + 1) * sizeof (dtv_t));
206 /* Re-initialize the TLS. */
207 _dl_allocate_tls_init (TLS_TPADJ (result));
213 /* Add a stack frame which is not used anymore to the stack. Must be
214 called with the cache lock held. */
216 __attribute ((always_inline))
217 queue_stack (struct pthread *stack)
219 /* We unconditionally add the stack to the list. The memory may
220 still be in use but it will not be reused until the kernel marks
221 the stack as not used anymore. */
222 list_add (&stack->list, &stack_cache);
224 stack_cache_actsize += stack->stackblock_size;
225 if (__builtin_expect (stack_cache_actsize > stack_cache_maxsize, 0))
227 /* We reduce the size of the cache. Remove the last entries
228 until the size is below the limit. */
232 /* Search from the end of the list. */
233 list_for_each_prev_safe (entry, prev, &stack_cache)
235 struct pthread *curr;
237 curr = list_entry (entry, struct pthread, list);
240 /* Unlink the block. */
243 /* Account for the freed memory. */
244 stack_cache_actsize -= curr->stackblock_size;
246 /* Free the memory associated with the ELF TLS. */
247 _dl_deallocate_tls (TLS_TPADJ (curr), false);
249 /* Remove this block. This should never fail. If it
250 does something is really wrong. */
251 if (munmap (curr->stackblock, curr->stackblock_size) != 0)
254 /* Maybe we have freed enough. */
255 if (stack_cache_actsize <= stack_cache_maxsize)
265 change_stack_perm (struct pthread *pd
266 #ifdef NEED_SEPARATE_REGISTER_STACK
271 #ifdef NEED_SEPARATE_REGISTER_STACK
272 void *stack = (pd->stackblock
273 + (((((pd->stackblock_size - pd->guardsize) / 2)
274 & pagemask) + pd->guardsize) & pagemask));
275 size_t len = pd->stackblock + pd->stackblock_size - stack;
277 void *stack = pd->stackblock + pd->guardsize;
278 size_t len = pd->stackblock_size - pd->guardsize;
280 if (mprotect (stack, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
288 allocate_stack (const struct pthread_attr *attr, struct pthread **pdp,
289 ALLOCATE_STACK_PARMS)
293 size_t pagesize_m1 = __getpagesize () - 1;
296 assert (attr != NULL);
297 assert (powerof2 (pagesize_m1 + 1));
298 assert (TCB_ALIGNMENT >= STACK_ALIGN);
300 /* Get the stack size from the attribute if it is set. Otherwise we
301 use the default we determined at start time. */
302 size = attr->stacksize ?: __default_stacksize;
304 /* Get memory for the stack. */
305 if (__builtin_expect (attr->flags & ATTR_FLAG_STACKADDR, 0))
309 /* If the user also specified the size of the stack make sure it
311 if (attr->stacksize != 0
312 && attr->stacksize < (__static_tls_size + MINIMAL_REST_STACK))
315 /* Adjust stack size for alignment of the TLS block. */
317 adj = ((uintptr_t) attr->stackaddr - TLS_TCB_SIZE)
318 & __static_tls_align_m1;
319 assert (size > adj + TLS_TCB_SIZE);
321 adj = ((uintptr_t) attr->stackaddr - __static_tls_size)
322 & __static_tls_align_m1;
326 /* The user provided some memory. Let's hope it matches the
327 size... We do not allocate guard pages if the user provided
328 the stack. It is the user's responsibility to do this if it
331 pd = (struct pthread *) ((uintptr_t) attr->stackaddr
332 - TLS_TCB_SIZE - adj);
334 pd = (struct pthread *) (((uintptr_t) attr->stackaddr
335 - __static_tls_size - adj)
339 /* The user provided stack memory needs to be cleared. */
340 memset (pd, '\0', sizeof (struct pthread));
342 /* The first TSD block is included in the TCB. */
343 pd->specific[0] = pd->specific_1stblock;
345 #if defined __ASSUME_CLONE_STOPPED && LLL_LOCK_INITIALIZER != 0
346 /* Initialize the lock. */
347 pd->lock = LLL_LOCK_INITIALIZER;
350 /* Remember the stack-related values. */
351 pd->stackblock = (char *) attr->stackaddr - size;
352 pd->stackblock_size = size;
354 /* This is a user-provided stack. It will not be queued in the
355 stack cache nor will the memory (except the TLS memory) be freed. */
356 pd->user_stack = true;
358 /* This is at least the second thread. */
359 pd->header.multiple_threads = 1;
360 #ifndef TLS_MULTIPLE_THREADS_IN_TCB
361 __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1;
364 #ifdef NEED_DL_SYSINFO
365 /* Copy the sysinfo value from the parent. */
366 THREAD_SYSINFO(pd) = THREAD_SELF_SYSINFO;
369 /* The process ID is also the same as that of the caller. */
370 pd->pid = THREAD_GETMEM (THREAD_SELF, pid);
372 /* Allocate the DTV for this thread. */
373 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
374 /* Something went wrong. */
378 /* Prepare to modify global data. */
379 lll_lock (stack_cache_lock);
381 /* And add to the list of stacks in use. */
382 list_add (&pd->list, &__stack_user);
384 lll_unlock (stack_cache_lock);
388 /* Allocate some anonymous memory. If possible use the cache. */
392 const int prot = (PROT_READ | PROT_WRITE
393 | ((GL(dl_stack_flags) & PF_X) ? PROT_EXEC : 0));
395 #if COLORING_INCREMENT != 0
396 /* Add one more page for stack coloring. Don't do it for stacks
397 with 16 times pagesize or larger. This might just cause
398 unnecessary misalignment. */
399 if (size <= 16 * pagesize_m1)
400 size += pagesize_m1 + 1;
403 /* Adjust the stack size for alignment. */
404 size &= ~__static_tls_align_m1;
407 /* Make sure the size of the stack is enough for the guard and
408 eventually the thread descriptor. */
409 guardsize = (attr->guardsize + pagesize_m1) & ~pagesize_m1;
410 if (__builtin_expect (size < (guardsize + __static_tls_size
411 + MINIMAL_REST_STACK + pagesize_m1 + 1),
413 /* The stack is too small (or the guard too large). */
416 /* Try to get a stack from the cache. */
418 pd = get_cached_stack (&size, &mem);
421 /* To avoid aliasing effects on a larger scale then pages we
422 adjust the allocated stack size if necessary. This way
423 allocations directly following each other will not have
424 aliasing problems. */
425 #if MULTI_PAGE_ALIASING != 0
426 if ((size % MULTI_PAGE_ALIASING) == 0)
427 size += pagesize_m1 + 1;
430 mem = mmap (NULL, size, prot,
431 MAP_PRIVATE | MAP_ANONYMOUS | ARCH_MAP_FLAGS, -1, 0);
433 if (__builtin_expect (mem == MAP_FAILED, 0))
435 #ifdef ARCH_RETRY_MMAP
436 mem = ARCH_RETRY_MMAP (size);
437 if (__builtin_expect (mem == MAP_FAILED, 0))
442 /* SIZE is guaranteed to be greater than zero.
443 So we can never get a null pointer back from mmap. */
444 assert (mem != NULL);
446 #if COLORING_INCREMENT != 0
447 /* Atomically increment NCREATED. */
448 unsigned int ncreated = (atomic_exchange_and_add (&nptl_ncreated, 1)
451 /* We chose the offset for coloring by incrementing it for
452 every new thread by a fixed amount. The offset used
453 module the page size. Even if coloring would be better
454 relative to higher alignment values it makes no sense to
455 do it since the mmap() interface does not allow us to
456 specify any alignment for the returned memory block. */
457 size_t coloring = (ncreated * COLORING_INCREMENT) & pagesize_m1;
459 /* Make sure the coloring offsets does not disturb the alignment
460 of the TCB and static TLS block. */
461 if (__builtin_expect ((coloring & __static_tls_align_m1) != 0, 0))
462 coloring = (((coloring + __static_tls_align_m1)
463 & ~(__static_tls_align_m1))
466 /* Unless specified we do not make any adjustments. */
470 /* Place the thread descriptor at the end of the stack. */
472 pd = (struct pthread *) ((char *) mem + size - coloring) - 1;
474 pd = (struct pthread *) ((((uintptr_t) mem + size - coloring
476 & ~__static_tls_align_m1)
480 /* Remember the stack-related values. */
481 pd->stackblock = mem;
482 pd->stackblock_size = size;
484 /* We allocated the first block thread-specific data array.
485 This address will not change for the lifetime of this
487 pd->specific[0] = pd->specific_1stblock;
489 #if defined __ASSUME_CLONE_STOPPED && LLL_LOCK_INITIALIZER != 0
490 /* Initialize the lock. */
491 pd->lock = LLL_LOCK_INITIALIZER;
494 /* This is at least the second thread. */
495 pd->header.multiple_threads = 1;
496 #ifndef TLS_MULTIPLE_THREADS_IN_TCB
497 __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1;
500 #ifdef NEED_DL_SYSINFO
501 /* Copy the sysinfo value from the parent. */
502 THREAD_SYSINFO(pd) = THREAD_SELF_SYSINFO;
505 /* The process ID is also the same as that of the caller. */
506 pd->pid = THREAD_GETMEM (THREAD_SELF, pid);
508 /* Allocate the DTV for this thread. */
509 if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL)
511 /* Something went wrong. */
514 /* Free the stack memory we just allocated. */
515 (void) munmap (mem, size);
521 /* Prepare to modify global data. */
522 lll_lock (stack_cache_lock);
524 /* And add to the list of stacks in use. */
525 list_add (&pd->list, &stack_used);
527 lll_unlock (stack_cache_lock);
530 /* There might have been a race. Another thread might have
531 caused the stacks to get exec permission while this new
532 stack was prepared. Detect if this was possible and
533 change the permission if necessary. */
534 if (__builtin_expect ((GL(dl_stack_flags) & PF_X) != 0
535 && (prot & PROT_EXEC) == 0, 0))
537 int err = change_stack_perm (pd
538 #ifdef NEED_SEPARATE_REGISTER_STACK
544 /* Free the stack memory we just allocated. */
545 (void) munmap (mem, size);
552 /* Note that all of the stack and the thread descriptor is
553 zeroed. This means we do not have to initialize fields
554 with initial value zero. This is specifically true for
555 the 'tid' field which is always set back to zero once the
556 stack is not used anymore and for the 'guardsize' field
557 which will be read next. */
560 /* Create or resize the guard area if necessary. */
561 if (__builtin_expect (guardsize > pd->guardsize, 0))
563 #ifdef NEED_SEPARATE_REGISTER_STACK
564 char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1);
568 if (mprotect (guard, guardsize, PROT_NONE) != 0)
574 lll_lock (stack_cache_lock);
576 /* Remove the thread from the list. */
577 list_del (&pd->list);
579 lll_unlock (stack_cache_lock);
581 /* Get rid of the TLS block we allocated. */
582 _dl_deallocate_tls (TLS_TPADJ (pd), false);
584 /* Free the stack memory regardless of whether the size
585 of the cache is over the limit or not. If this piece
586 of memory caused problems we better do not use it
587 anymore. Uh, and we ignore possible errors. There
588 is nothing we could do. */
589 (void) munmap (mem, size);
594 pd->guardsize = guardsize;
596 else if (__builtin_expect (pd->guardsize - guardsize > size - reqsize,
599 /* The old guard area is too large. */
601 #ifdef NEED_SEPARATE_REGISTER_STACK
602 char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1);
603 char *oldguard = mem + (((size - pd->guardsize) / 2) & ~pagesize_m1);
606 && mprotect (oldguard, guard - oldguard, prot) != 0)
609 if (mprotect (guard + guardsize,
610 oldguard + pd->guardsize - guard - guardsize,
614 if (mprotect ((char *) mem + guardsize, pd->guardsize - guardsize,
619 pd->guardsize = guardsize;
621 /* The pthread_getattr_np() calls need to get passed the size
622 requested in the attribute, regardless of how large the
623 actually used guardsize is. */
624 pd->reported_guardsize = guardsize;
627 #ifndef __ASSUME_CLONE_STOPPED
628 /* Initialize the lock. We have to do this unconditionally if the
629 CLONE_STOPPED flag is not available since then the stillborn
630 thread could be canceled while the lock is taken. */
631 pd->lock = LLL_LOCK_INITIALIZER;
634 /* We place the thread descriptor at the end of the stack. */
638 /* The stack begins before the TCB and the static TLS block. */
639 stacktop = ((char *) (pd + 1) - __static_tls_size);
641 stacktop = (char *) (pd - 1);
644 #ifdef NEED_SEPARATE_REGISTER_STACK
645 *stack = pd->stackblock;
646 *stacksize = stacktop - *stack;
657 __deallocate_stack (struct pthread *pd)
659 lll_lock (stack_cache_lock);
661 /* Remove the thread from the list of threads with user defined
663 list_del (&pd->list);
665 /* Not much to do. Just free the mmap()ed memory. Note that we do
666 not reset the 'used' flag in the 'tid' field. This is done by
667 the kernel. If no thread has been created yet this field is
669 if (__builtin_expect (! pd->user_stack, 1))
670 (void) queue_stack (pd);
672 /* Free the memory associated with the ELF TLS. */
673 _dl_deallocate_tls (TLS_TPADJ (pd), false);
675 lll_unlock (stack_cache_lock);
681 __make_stacks_executable (void **stack_endp)
683 /* First the main thread's stack. */
684 int err = _dl_make_stack_executable (stack_endp);
688 #ifdef NEED_SEPARATE_REGISTER_STACK
689 const size_t pagemask = ~(__getpagesize () - 1);
692 lll_lock (stack_cache_lock);
695 list_for_each (runp, &stack_used)
697 err = change_stack_perm (list_entry (runp, struct pthread, list)
698 #ifdef NEED_SEPARATE_REGISTER_STACK
706 /* Also change the permission for the currently unused stacks. This
707 might be wasted time but better spend it here than adding a check
710 list_for_each (runp, &stack_cache)
712 err = change_stack_perm (list_entry (runp, struct pthread, list)
713 #ifdef NEED_SEPARATE_REGISTER_STACK
721 lll_unlock (stack_cache_lock);
727 /* In case of a fork() call the memory allocation in the child will be
728 the same but only one thread is running. All stacks except that of
729 the one running thread are not used anymore. We have to recycle
732 __reclaim_stacks (void)
734 struct pthread *self = (struct pthread *) THREAD_SELF;
736 /* No locking necessary. The caller is the only stack in use. */
738 /* Mark all stacks except the still running one as free. */
740 list_for_each (runp, &stack_used)
742 struct pthread *curp;
744 curp = list_entry (runp, struct pthread, list);
747 /* This marks the stack as free. */
750 /* The PID field must be initialized for the new process. */
751 curp->pid = self->pid;
753 /* Account for the size of the stack. */
754 stack_cache_actsize += curp->stackblock_size;
758 /* Add the stack of all running threads to the cache. */
759 list_splice (&stack_used, &stack_cache);
761 /* Remove the entry for the current thread to from the cache list
762 and add it to the list of running threads. Which of the two
763 lists is decided by the user_stack flag. */
764 list_del (&self->list);
766 /* Re-initialize the lists for all the threads. */
767 INIT_LIST_HEAD (&stack_used);
768 INIT_LIST_HEAD (&__stack_user);
770 if (__builtin_expect (THREAD_GETMEM (self, user_stack), 0))
771 list_add (&self->list, &__stack_user);
773 list_add (&self->list, &stack_used);
775 /* There is one thread running. */
778 /* Initialize the lock. */
779 stack_cache_lock = LLL_LOCK_INITIALIZER;
784 /* Find a thread given the thread ID. */
787 __find_thread_by_id (pid_t tid)
789 struct pthread *result = NULL;
791 lll_lock (stack_cache_lock);
793 /* Iterate over the list with system-allocated threads first. */
795 list_for_each (runp, &stack_used)
797 struct pthread *curp;
799 curp = list_entry (runp, struct pthread, list);
801 if (curp->tid == tid)
808 /* Now the list with threads using user-allocated stacks. */
809 list_for_each (runp, &__stack_user)
811 struct pthread *curp;
813 curp = list_entry (runp, struct pthread, list);
815 if (curp->tid == tid)
823 lll_unlock (stack_cache_lock);
829 static inline void __attribute__((always_inline))
830 init_one_static_tls (struct pthread *curp, struct link_map *map)
832 dtv_t *dtv = GET_DTV (TLS_TPADJ (curp));
834 void *dest = (char *) curp - map->l_tls_offset;
836 void *dest = (char *) curp + map->l_tls_offset + TLS_PRE_TCB_SIZE;
838 # error "Either TLS_TCB_AT_TP or TLS_DTV_AT_TP must be defined"
841 /* Fill in the DTV slot so that a later LD/GD access will find it. */
842 dtv[map->l_tls_modid].pointer = dest;
844 /* Initialize the memory. */
845 memset (__mempcpy (dest, map->l_tls_initimage, map->l_tls_initimage_size),
846 '\0', map->l_tls_blocksize - map->l_tls_initimage_size);
851 __pthread_init_static_tls (struct link_map *map)
853 lll_lock (stack_cache_lock);
855 /* Iterate over the list with system-allocated threads first. */
857 list_for_each (runp, &stack_used)
858 init_one_static_tls (list_entry (runp, struct pthread, list), map);
860 /* Now the list with threads using user-allocated stacks. */
861 list_for_each (runp, &__stack_user)
862 init_one_static_tls (list_entry (runp, struct pthread, list), map);
864 lll_unlock (stack_cache_lock);