1 | /* Copyright (C) 2002-2016 Free Software Foundation, Inc. |
2 | This file is part of the GNU C Library. |
3 | Contributed by Ulrich Drepper <drepper@redhat.com>, 2002. |
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 |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the 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; if not, see |
17 | <http://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <assert.h> |
20 | #include <errno.h> |
21 | #include <signal.h> |
22 | #include <stdint.h> |
23 | #include <string.h> |
24 | #include <unistd.h> |
25 | #include <sys/mman.h> |
26 | #include <sys/param.h> |
27 | #include <dl-sysdep.h> |
28 | #include <dl-tls.h> |
29 | #include <tls.h> |
30 | #include <list.h> |
31 | #include <lowlevellock.h> |
32 | #include <futex-internal.h> |
33 | #include <kernel-features.h> |
34 | #include <stack-aliasing.h> |
35 | |
36 | |
37 | #ifndef NEED_SEPARATE_REGISTER_STACK |
38 | |
39 | /* Most architectures have exactly one stack pointer. Some have more. */ |
40 | # define STACK_VARIABLES void *stackaddr = NULL |
41 | |
42 | /* How to pass the values to the 'create_thread' function. */ |
43 | # define STACK_VARIABLES_ARGS stackaddr |
44 | |
45 | /* How to declare function which gets there parameters. */ |
46 | # define STACK_VARIABLES_PARMS void *stackaddr |
47 | |
48 | /* How to declare allocate_stack. */ |
49 | # define ALLOCATE_STACK_PARMS void **stack |
50 | |
51 | /* This is how the function is called. We do it this way to allow |
52 | other variants of the function to have more parameters. */ |
53 | # define ALLOCATE_STACK(attr, pd) allocate_stack (attr, pd, &stackaddr) |
54 | |
55 | #else |
56 | |
57 | /* We need two stacks. The kernel will place them but we have to tell |
58 | the kernel about the size of the reserved address space. */ |
59 | # define STACK_VARIABLES void *stackaddr = NULL; size_t stacksize = 0 |
60 | |
61 | /* How to pass the values to the 'create_thread' function. */ |
62 | # define STACK_VARIABLES_ARGS stackaddr, stacksize |
63 | |
64 | /* How to declare function which gets there parameters. */ |
65 | # define STACK_VARIABLES_PARMS void *stackaddr, size_t stacksize |
66 | |
67 | /* How to declare allocate_stack. */ |
68 | # define ALLOCATE_STACK_PARMS void **stack, size_t *stacksize |
69 | |
70 | /* This is how the function is called. We do it this way to allow |
71 | other variants of the function to have more parameters. */ |
72 | # define ALLOCATE_STACK(attr, pd) \ |
73 | allocate_stack (attr, pd, &stackaddr, &stacksize) |
74 | |
75 | #endif |
76 | |
77 | |
78 | /* Default alignment of stack. */ |
79 | #ifndef STACK_ALIGN |
80 | # define STACK_ALIGN __alignof__ (long double) |
81 | #endif |
82 | |
83 | /* Default value for minimal stack size after allocating thread |
84 | descriptor and guard. */ |
85 | #ifndef MINIMAL_REST_STACK |
86 | # define MINIMAL_REST_STACK 4096 |
87 | #endif |
88 | |
89 | |
90 | /* Newer kernels have the MAP_STACK flag to indicate a mapping is used for |
91 | a stack. Use it when possible. */ |
92 | #ifndef MAP_STACK |
93 | # define MAP_STACK 0 |
94 | #endif |
95 | |
96 | /* This yields the pointer that TLS support code calls the thread pointer. */ |
97 | #if TLS_TCB_AT_TP |
98 | # define TLS_TPADJ(pd) (pd) |
99 | #elif TLS_DTV_AT_TP |
100 | # define TLS_TPADJ(pd) ((struct pthread *)((char *) (pd) + TLS_PRE_TCB_SIZE)) |
101 | #endif |
102 | |
103 | /* Cache handling for not-yet free stacks. */ |
104 | |
105 | /* Maximum size in kB of cache. */ |
106 | static size_t stack_cache_maxsize = 40 * 1024 * 1024; /* 40MiBi by default. */ |
107 | static size_t stack_cache_actsize; |
108 | |
109 | /* Mutex protecting this variable. */ |
110 | static int stack_cache_lock = LLL_LOCK_INITIALIZER; |
111 | |
112 | /* List of queued stack frames. */ |
113 | static LIST_HEAD (stack_cache); |
114 | |
115 | /* List of the stacks in use. */ |
116 | static LIST_HEAD (stack_used); |
117 | |
118 | /* We need to record what list operations we are going to do so that, |
119 | in case of an asynchronous interruption due to a fork() call, we |
120 | can correct for the work. */ |
121 | static uintptr_t in_flight_stack; |
122 | |
123 | /* List of the threads with user provided stacks in use. No need to |
124 | initialize this, since it's done in __pthread_initialize_minimal. */ |
125 | list_t __stack_user __attribute__ ((nocommon)); |
126 | hidden_data_def (__stack_user) |
127 | |
128 | #if COLORING_INCREMENT != 0 |
129 | /* Number of threads created. */ |
130 | static unsigned int nptl_ncreated; |
131 | #endif |
132 | |
133 | |
134 | /* Check whether the stack is still used or not. */ |
135 | #define FREE_P(descr) ((descr)->tid <= 0) |
136 | |
137 | |
138 | static void |
139 | stack_list_del (list_t *elem) |
140 | { |
141 | in_flight_stack = (uintptr_t) elem; |
142 | |
143 | atomic_write_barrier (); |
144 | |
145 | list_del (elem); |
146 | |
147 | atomic_write_barrier (); |
148 | |
149 | in_flight_stack = 0; |
150 | } |
151 | |
152 | |
153 | static void |
154 | stack_list_add (list_t *elem, list_t *list) |
155 | { |
156 | in_flight_stack = (uintptr_t) elem | 1; |
157 | |
158 | atomic_write_barrier (); |
159 | |
160 | list_add (elem, list); |
161 | |
162 | atomic_write_barrier (); |
163 | |
164 | in_flight_stack = 0; |
165 | } |
166 | |
167 | |
168 | /* We create a double linked list of all cache entries. Double linked |
169 | because this allows removing entries from the end. */ |
170 | |
171 | |
172 | /* Get a stack frame from the cache. We have to match by size since |
173 | some blocks might be too small or far too large. */ |
174 | static struct pthread * |
175 | get_cached_stack (size_t *sizep, void **memp) |
176 | { |
177 | size_t size = *sizep; |
178 | struct pthread *result = NULL; |
179 | list_t *entry; |
180 | |
181 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
182 | |
183 | /* Search the cache for a matching entry. We search for the |
184 | smallest stack which has at least the required size. Note that |
185 | in normal situations the size of all allocated stacks is the |
186 | same. As the very least there are only a few different sizes. |
187 | Therefore this loop will exit early most of the time with an |
188 | exact match. */ |
189 | list_for_each (entry, &stack_cache) |
190 | { |
191 | struct pthread *curr; |
192 | |
193 | curr = list_entry (entry, struct pthread, list); |
194 | if (FREE_P (curr) && curr->stackblock_size >= size) |
195 | { |
196 | if (curr->stackblock_size == size) |
197 | { |
198 | result = curr; |
199 | break; |
200 | } |
201 | |
202 | if (result == NULL |
203 | || result->stackblock_size > curr->stackblock_size) |
204 | result = curr; |
205 | } |
206 | } |
207 | |
208 | if (__builtin_expect (result == NULL, 0) |
209 | /* Make sure the size difference is not too excessive. In that |
210 | case we do not use the block. */ |
211 | || __builtin_expect (result->stackblock_size > 4 * size, 0)) |
212 | { |
213 | /* Release the lock. */ |
214 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
215 | |
216 | return NULL; |
217 | } |
218 | |
219 | /* Don't allow setxid until cloned. */ |
220 | result->setxid_futex = -1; |
221 | |
222 | /* Dequeue the entry. */ |
223 | stack_list_del (&result->list); |
224 | |
225 | /* And add to the list of stacks in use. */ |
226 | stack_list_add (&result->list, &stack_used); |
227 | |
228 | /* And decrease the cache size. */ |
229 | stack_cache_actsize -= result->stackblock_size; |
230 | |
231 | /* Release the lock early. */ |
232 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
233 | |
234 | /* Report size and location of the stack to the caller. */ |
235 | *sizep = result->stackblock_size; |
236 | *memp = result->stackblock; |
237 | |
238 | /* Cancellation handling is back to the default. */ |
239 | result->cancelhandling = 0; |
240 | result->cleanup = NULL; |
241 | |
242 | /* No pending event. */ |
243 | result->nextevent = NULL; |
244 | |
245 | /* Clear the DTV. */ |
246 | dtv_t *dtv = GET_DTV (TLS_TPADJ (result)); |
247 | for (size_t cnt = 0; cnt < dtv[-1].counter; ++cnt) |
248 | if (! dtv[1 + cnt].pointer.is_static |
249 | && dtv[1 + cnt].pointer.val != TLS_DTV_UNALLOCATED) |
250 | free (dtv[1 + cnt].pointer.val); |
251 | memset (dtv, '\0', (dtv[-1].counter + 1) * sizeof (dtv_t)); |
252 | |
253 | /* Re-initialize the TLS. */ |
254 | _dl_allocate_tls_init (TLS_TPADJ (result)); |
255 | |
256 | return result; |
257 | } |
258 | |
259 | |
260 | /* Free stacks until cache size is lower than LIMIT. */ |
261 | void |
262 | __free_stacks (size_t limit) |
263 | { |
264 | /* We reduce the size of the cache. Remove the last entries until |
265 | the size is below the limit. */ |
266 | list_t *entry; |
267 | list_t *prev; |
268 | |
269 | /* Search from the end of the list. */ |
270 | list_for_each_prev_safe (entry, prev, &stack_cache) |
271 | { |
272 | struct pthread *curr; |
273 | |
274 | curr = list_entry (entry, struct pthread, list); |
275 | if (FREE_P (curr)) |
276 | { |
277 | /* Unlink the block. */ |
278 | stack_list_del (entry); |
279 | |
280 | /* Account for the freed memory. */ |
281 | stack_cache_actsize -= curr->stackblock_size; |
282 | |
283 | /* Free the memory associated with the ELF TLS. */ |
284 | _dl_deallocate_tls (TLS_TPADJ (curr), false); |
285 | |
286 | /* Remove this block. This should never fail. If it does |
287 | something is really wrong. */ |
288 | if (munmap (curr->stackblock, curr->stackblock_size) != 0) |
289 | abort (); |
290 | |
291 | /* Maybe we have freed enough. */ |
292 | if (stack_cache_actsize <= limit) |
293 | break; |
294 | } |
295 | } |
296 | } |
297 | |
298 | |
299 | /* Add a stack frame which is not used anymore to the stack. Must be |
300 | called with the cache lock held. */ |
301 | static inline void |
302 | __attribute ((always_inline)) |
303 | queue_stack (struct pthread *stack) |
304 | { |
305 | /* We unconditionally add the stack to the list. The memory may |
306 | still be in use but it will not be reused until the kernel marks |
307 | the stack as not used anymore. */ |
308 | stack_list_add (&stack->list, &stack_cache); |
309 | |
310 | stack_cache_actsize += stack->stackblock_size; |
311 | if (__glibc_unlikely (stack_cache_actsize > stack_cache_maxsize)) |
312 | __free_stacks (stack_cache_maxsize); |
313 | } |
314 | |
315 | |
316 | static int |
317 | internal_function |
318 | change_stack_perm (struct pthread *pd |
319 | #ifdef NEED_SEPARATE_REGISTER_STACK |
320 | , size_t pagemask |
321 | #endif |
322 | ) |
323 | { |
324 | #ifdef NEED_SEPARATE_REGISTER_STACK |
325 | void *stack = (pd->stackblock |
326 | + (((((pd->stackblock_size - pd->guardsize) / 2) |
327 | & pagemask) + pd->guardsize) & pagemask)); |
328 | size_t len = pd->stackblock + pd->stackblock_size - stack; |
329 | #elif _STACK_GROWS_DOWN |
330 | void *stack = pd->stackblock + pd->guardsize; |
331 | size_t len = pd->stackblock_size - pd->guardsize; |
332 | #elif _STACK_GROWS_UP |
333 | void *stack = pd->stackblock; |
334 | size_t len = (uintptr_t) pd - pd->guardsize - (uintptr_t) pd->stackblock; |
335 | #else |
336 | # error "Define either _STACK_GROWS_DOWN or _STACK_GROWS_UP" |
337 | #endif |
338 | if (mprotect (stack, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0) |
339 | return errno; |
340 | |
341 | return 0; |
342 | } |
343 | |
344 | |
345 | /* Returns a usable stack for a new thread either by allocating a |
346 | new stack or reusing a cached stack of sufficient size. |
347 | ATTR must be non-NULL and point to a valid pthread_attr. |
348 | PDP must be non-NULL. */ |
349 | static int |
350 | allocate_stack (const struct pthread_attr *attr, struct pthread **pdp, |
351 | ALLOCATE_STACK_PARMS) |
352 | { |
353 | struct pthread *pd; |
354 | size_t size; |
355 | size_t pagesize_m1 = __getpagesize () - 1; |
356 | |
357 | assert (powerof2 (pagesize_m1 + 1)); |
358 | assert (TCB_ALIGNMENT >= STACK_ALIGN); |
359 | |
360 | /* Get the stack size from the attribute if it is set. Otherwise we |
361 | use the default we determined at start time. */ |
362 | if (attr->stacksize != 0) |
363 | size = attr->stacksize; |
364 | else |
365 | { |
366 | lll_lock (__default_pthread_attr_lock, LLL_PRIVATE); |
367 | size = __default_pthread_attr.stacksize; |
368 | lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE); |
369 | } |
370 | |
371 | /* Get memory for the stack. */ |
372 | if (__glibc_unlikely (attr->flags & ATTR_FLAG_STACKADDR)) |
373 | { |
374 | uintptr_t adj; |
375 | |
376 | /* If the user also specified the size of the stack make sure it |
377 | is large enough. */ |
378 | if (attr->stacksize != 0 |
379 | && attr->stacksize < (__static_tls_size + MINIMAL_REST_STACK)) |
380 | return EINVAL; |
381 | |
382 | /* Adjust stack size for alignment of the TLS block. */ |
383 | #if TLS_TCB_AT_TP |
384 | adj = ((uintptr_t) attr->stackaddr - TLS_TCB_SIZE) |
385 | & __static_tls_align_m1; |
386 | assert (size > adj + TLS_TCB_SIZE); |
387 | #elif TLS_DTV_AT_TP |
388 | adj = ((uintptr_t) attr->stackaddr - __static_tls_size) |
389 | & __static_tls_align_m1; |
390 | assert (size > adj); |
391 | #endif |
392 | |
393 | /* The user provided some memory. Let's hope it matches the |
394 | size... We do not allocate guard pages if the user provided |
395 | the stack. It is the user's responsibility to do this if it |
396 | is wanted. */ |
397 | #if TLS_TCB_AT_TP |
398 | pd = (struct pthread *) ((uintptr_t) attr->stackaddr |
399 | - TLS_TCB_SIZE - adj); |
400 | #elif TLS_DTV_AT_TP |
401 | pd = (struct pthread *) (((uintptr_t) attr->stackaddr |
402 | - __static_tls_size - adj) |
403 | - TLS_PRE_TCB_SIZE); |
404 | #endif |
405 | |
406 | /* The user provided stack memory needs to be cleared. */ |
407 | memset (pd, '\0', sizeof (struct pthread)); |
408 | |
409 | /* The first TSD block is included in the TCB. */ |
410 | pd->specific[0] = pd->specific_1stblock; |
411 | |
412 | /* Remember the stack-related values. */ |
413 | pd->stackblock = (char *) attr->stackaddr - size; |
414 | pd->stackblock_size = size; |
415 | |
416 | /* This is a user-provided stack. It will not be queued in the |
417 | stack cache nor will the memory (except the TLS memory) be freed. */ |
418 | pd->user_stack = true; |
419 | |
420 | /* This is at least the second thread. */ |
421 | pd->header.multiple_threads = 1; |
422 | #ifndef TLS_MULTIPLE_THREADS_IN_TCB |
423 | __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1; |
424 | #endif |
425 | |
426 | #ifndef __ASSUME_PRIVATE_FUTEX |
427 | /* The thread must know when private futexes are supported. */ |
428 | pd->header.private_futex = THREAD_GETMEM (THREAD_SELF, |
429 | header.private_futex); |
430 | #endif |
431 | |
432 | #ifdef NEED_DL_SYSINFO |
433 | SETUP_THREAD_SYSINFO (pd); |
434 | #endif |
435 | |
436 | /* The process ID is also the same as that of the caller. */ |
437 | pd->pid = THREAD_GETMEM (THREAD_SELF, pid); |
438 | |
439 | /* Don't allow setxid until cloned. */ |
440 | pd->setxid_futex = -1; |
441 | |
442 | /* Allocate the DTV for this thread. */ |
443 | if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL) |
444 | { |
445 | /* Something went wrong. */ |
446 | assert (errno == ENOMEM); |
447 | return errno; |
448 | } |
449 | |
450 | |
451 | /* Prepare to modify global data. */ |
452 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
453 | |
454 | /* And add to the list of stacks in use. */ |
455 | list_add (&pd->list, &__stack_user); |
456 | |
457 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
458 | } |
459 | else |
460 | { |
461 | /* Allocate some anonymous memory. If possible use the cache. */ |
462 | size_t guardsize; |
463 | size_t reqsize; |
464 | void *mem; |
465 | const int prot = (PROT_READ | PROT_WRITE |
466 | | ((GL(dl_stack_flags) & PF_X) ? PROT_EXEC : 0)); |
467 | |
468 | #if COLORING_INCREMENT != 0 |
469 | /* Add one more page for stack coloring. Don't do it for stacks |
470 | with 16 times pagesize or larger. This might just cause |
471 | unnecessary misalignment. */ |
472 | if (size <= 16 * pagesize_m1) |
473 | size += pagesize_m1 + 1; |
474 | #endif |
475 | |
476 | /* Adjust the stack size for alignment. */ |
477 | size &= ~__static_tls_align_m1; |
478 | assert (size != 0); |
479 | |
480 | /* Make sure the size of the stack is enough for the guard and |
481 | eventually the thread descriptor. */ |
482 | guardsize = (attr->guardsize + pagesize_m1) & ~pagesize_m1; |
483 | if (__builtin_expect (size < ((guardsize + __static_tls_size |
484 | + MINIMAL_REST_STACK + pagesize_m1) |
485 | & ~pagesize_m1), |
486 | 0)) |
487 | /* The stack is too small (or the guard too large). */ |
488 | return EINVAL; |
489 | |
490 | /* Try to get a stack from the cache. */ |
491 | reqsize = size; |
492 | pd = get_cached_stack (&size, &mem); |
493 | if (pd == NULL) |
494 | { |
495 | /* To avoid aliasing effects on a larger scale than pages we |
496 | adjust the allocated stack size if necessary. This way |
497 | allocations directly following each other will not have |
498 | aliasing problems. */ |
499 | #if MULTI_PAGE_ALIASING != 0 |
500 | if ((size % MULTI_PAGE_ALIASING) == 0) |
501 | size += pagesize_m1 + 1; |
502 | #endif |
503 | |
504 | mem = mmap (NULL, size, prot, |
505 | MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0); |
506 | |
507 | if (__glibc_unlikely (mem == MAP_FAILED)) |
508 | return errno; |
509 | |
510 | /* SIZE is guaranteed to be greater than zero. |
511 | So we can never get a null pointer back from mmap. */ |
512 | assert (mem != NULL); |
513 | |
514 | #if COLORING_INCREMENT != 0 |
515 | /* Atomically increment NCREATED. */ |
516 | unsigned int ncreated = atomic_increment_val (&nptl_ncreated); |
517 | |
518 | /* We chose the offset for coloring by incrementing it for |
519 | every new thread by a fixed amount. The offset used |
520 | module the page size. Even if coloring would be better |
521 | relative to higher alignment values it makes no sense to |
522 | do it since the mmap() interface does not allow us to |
523 | specify any alignment for the returned memory block. */ |
524 | size_t coloring = (ncreated * COLORING_INCREMENT) & pagesize_m1; |
525 | |
526 | /* Make sure the coloring offsets does not disturb the alignment |
527 | of the TCB and static TLS block. */ |
528 | if (__glibc_unlikely ((coloring & __static_tls_align_m1) != 0)) |
529 | coloring = (((coloring + __static_tls_align_m1) |
530 | & ~(__static_tls_align_m1)) |
531 | & ~pagesize_m1); |
532 | #else |
533 | /* Unless specified we do not make any adjustments. */ |
534 | # define coloring 0 |
535 | #endif |
536 | |
537 | /* Place the thread descriptor at the end of the stack. */ |
538 | #if TLS_TCB_AT_TP |
539 | pd = (struct pthread *) ((char *) mem + size - coloring) - 1; |
540 | #elif TLS_DTV_AT_TP |
541 | pd = (struct pthread *) ((((uintptr_t) mem + size - coloring |
542 | - __static_tls_size) |
543 | & ~__static_tls_align_m1) |
544 | - TLS_PRE_TCB_SIZE); |
545 | #endif |
546 | |
547 | /* Remember the stack-related values. */ |
548 | pd->stackblock = mem; |
549 | pd->stackblock_size = size; |
550 | |
551 | /* We allocated the first block thread-specific data array. |
552 | This address will not change for the lifetime of this |
553 | descriptor. */ |
554 | pd->specific[0] = pd->specific_1stblock; |
555 | |
556 | /* This is at least the second thread. */ |
557 | pd->header.multiple_threads = 1; |
558 | #ifndef TLS_MULTIPLE_THREADS_IN_TCB |
559 | __pthread_multiple_threads = *__libc_multiple_threads_ptr = 1; |
560 | #endif |
561 | |
562 | #ifndef __ASSUME_PRIVATE_FUTEX |
563 | /* The thread must know when private futexes are supported. */ |
564 | pd->header.private_futex = THREAD_GETMEM (THREAD_SELF, |
565 | header.private_futex); |
566 | #endif |
567 | |
568 | #ifdef NEED_DL_SYSINFO |
569 | SETUP_THREAD_SYSINFO (pd); |
570 | #endif |
571 | |
572 | /* Don't allow setxid until cloned. */ |
573 | pd->setxid_futex = -1; |
574 | |
575 | /* The process ID is also the same as that of the caller. */ |
576 | pd->pid = THREAD_GETMEM (THREAD_SELF, pid); |
577 | |
578 | /* Allocate the DTV for this thread. */ |
579 | if (_dl_allocate_tls (TLS_TPADJ (pd)) == NULL) |
580 | { |
581 | /* Something went wrong. */ |
582 | assert (errno == ENOMEM); |
583 | |
584 | /* Free the stack memory we just allocated. */ |
585 | (void) munmap (mem, size); |
586 | |
587 | return errno; |
588 | } |
589 | |
590 | |
591 | /* Prepare to modify global data. */ |
592 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
593 | |
594 | /* And add to the list of stacks in use. */ |
595 | stack_list_add (&pd->list, &stack_used); |
596 | |
597 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
598 | |
599 | |
600 | /* There might have been a race. Another thread might have |
601 | caused the stacks to get exec permission while this new |
602 | stack was prepared. Detect if this was possible and |
603 | change the permission if necessary. */ |
604 | if (__builtin_expect ((GL(dl_stack_flags) & PF_X) != 0 |
605 | && (prot & PROT_EXEC) == 0, 0)) |
606 | { |
607 | int err = change_stack_perm (pd |
608 | #ifdef NEED_SEPARATE_REGISTER_STACK |
609 | , ~pagesize_m1 |
610 | #endif |
611 | ); |
612 | if (err != 0) |
613 | { |
614 | /* Free the stack memory we just allocated. */ |
615 | (void) munmap (mem, size); |
616 | |
617 | return err; |
618 | } |
619 | } |
620 | |
621 | |
622 | /* Note that all of the stack and the thread descriptor is |
623 | zeroed. This means we do not have to initialize fields |
624 | with initial value zero. This is specifically true for |
625 | the 'tid' field which is always set back to zero once the |
626 | stack is not used anymore and for the 'guardsize' field |
627 | which will be read next. */ |
628 | } |
629 | |
630 | /* Create or resize the guard area if necessary. */ |
631 | if (__glibc_unlikely (guardsize > pd->guardsize)) |
632 | { |
633 | #ifdef NEED_SEPARATE_REGISTER_STACK |
634 | char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1); |
635 | #elif _STACK_GROWS_DOWN |
636 | char *guard = mem; |
637 | # elif _STACK_GROWS_UP |
638 | char *guard = (char *) (((uintptr_t) pd - guardsize) & ~pagesize_m1); |
639 | #endif |
640 | if (mprotect (guard, guardsize, PROT_NONE) != 0) |
641 | { |
642 | mprot_error: |
643 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
644 | |
645 | /* Remove the thread from the list. */ |
646 | stack_list_del (&pd->list); |
647 | |
648 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
649 | |
650 | /* Get rid of the TLS block we allocated. */ |
651 | _dl_deallocate_tls (TLS_TPADJ (pd), false); |
652 | |
653 | /* Free the stack memory regardless of whether the size |
654 | of the cache is over the limit or not. If this piece |
655 | of memory caused problems we better do not use it |
656 | anymore. Uh, and we ignore possible errors. There |
657 | is nothing we could do. */ |
658 | (void) munmap (mem, size); |
659 | |
660 | return errno; |
661 | } |
662 | |
663 | pd->guardsize = guardsize; |
664 | } |
665 | else if (__builtin_expect (pd->guardsize - guardsize > size - reqsize, |
666 | 0)) |
667 | { |
668 | /* The old guard area is too large. */ |
669 | |
670 | #ifdef NEED_SEPARATE_REGISTER_STACK |
671 | char *guard = mem + (((size - guardsize) / 2) & ~pagesize_m1); |
672 | char *oldguard = mem + (((size - pd->guardsize) / 2) & ~pagesize_m1); |
673 | |
674 | if (oldguard < guard |
675 | && mprotect (oldguard, guard - oldguard, prot) != 0) |
676 | goto mprot_error; |
677 | |
678 | if (mprotect (guard + guardsize, |
679 | oldguard + pd->guardsize - guard - guardsize, |
680 | prot) != 0) |
681 | goto mprot_error; |
682 | #elif _STACK_GROWS_DOWN |
683 | if (mprotect ((char *) mem + guardsize, pd->guardsize - guardsize, |
684 | prot) != 0) |
685 | goto mprot_error; |
686 | #elif _STACK_GROWS_UP |
687 | if (mprotect ((char *) pd - pd->guardsize, |
688 | pd->guardsize - guardsize, prot) != 0) |
689 | goto mprot_error; |
690 | #endif |
691 | |
692 | pd->guardsize = guardsize; |
693 | } |
694 | /* The pthread_getattr_np() calls need to get passed the size |
695 | requested in the attribute, regardless of how large the |
696 | actually used guardsize is. */ |
697 | pd->reported_guardsize = guardsize; |
698 | } |
699 | |
700 | /* Initialize the lock. We have to do this unconditionally since the |
701 | stillborn thread could be canceled while the lock is taken. */ |
702 | pd->lock = LLL_LOCK_INITIALIZER; |
703 | |
704 | /* The robust mutex lists also need to be initialized |
705 | unconditionally because the cleanup for the previous stack owner |
706 | might have happened in the kernel. */ |
707 | pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock) |
708 | - offsetof (pthread_mutex_t, |
709 | __data.__list.__next)); |
710 | pd->robust_head.list_op_pending = NULL; |
711 | #ifdef __PTHREAD_MUTEX_HAVE_PREV |
712 | pd->robust_prev = &pd->robust_head; |
713 | #endif |
714 | pd->robust_head.list = &pd->robust_head; |
715 | |
716 | /* We place the thread descriptor at the end of the stack. */ |
717 | *pdp = pd; |
718 | |
719 | #if _STACK_GROWS_DOWN |
720 | void *stacktop; |
721 | |
722 | # if TLS_TCB_AT_TP |
723 | /* The stack begins before the TCB and the static TLS block. */ |
724 | stacktop = ((char *) (pd + 1) - __static_tls_size); |
725 | # elif TLS_DTV_AT_TP |
726 | stacktop = (char *) (pd - 1); |
727 | # endif |
728 | |
729 | # ifdef NEED_SEPARATE_REGISTER_STACK |
730 | *stack = pd->stackblock; |
731 | *stacksize = stacktop - *stack; |
732 | # else |
733 | *stack = stacktop; |
734 | # endif |
735 | #else |
736 | *stack = pd->stackblock; |
737 | assert (*stack > 0); |
738 | #endif |
739 | |
740 | return 0; |
741 | } |
742 | |
743 | |
744 | void |
745 | internal_function |
746 | __deallocate_stack (struct pthread *pd) |
747 | { |
748 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
749 | |
750 | /* Remove the thread from the list of threads with user defined |
751 | stacks. */ |
752 | stack_list_del (&pd->list); |
753 | |
754 | /* Not much to do. Just free the mmap()ed memory. Note that we do |
755 | not reset the 'used' flag in the 'tid' field. This is done by |
756 | the kernel. If no thread has been created yet this field is |
757 | still zero. */ |
758 | if (__glibc_likely (! pd->user_stack)) |
759 | (void) queue_stack (pd); |
760 | else |
761 | /* Free the memory associated with the ELF TLS. */ |
762 | _dl_deallocate_tls (TLS_TPADJ (pd), false); |
763 | |
764 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
765 | } |
766 | |
767 | |
768 | int |
769 | internal_function |
770 | __make_stacks_executable (void **stack_endp) |
771 | { |
772 | /* First the main thread's stack. */ |
773 | int err = _dl_make_stack_executable (stack_endp); |
774 | if (err != 0) |
775 | return err; |
776 | |
777 | #ifdef NEED_SEPARATE_REGISTER_STACK |
778 | const size_t pagemask = ~(__getpagesize () - 1); |
779 | #endif |
780 | |
781 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
782 | |
783 | list_t *runp; |
784 | list_for_each (runp, &stack_used) |
785 | { |
786 | err = change_stack_perm (list_entry (runp, struct pthread, list) |
787 | #ifdef NEED_SEPARATE_REGISTER_STACK |
788 | , pagemask |
789 | #endif |
790 | ); |
791 | if (err != 0) |
792 | break; |
793 | } |
794 | |
795 | /* Also change the permission for the currently unused stacks. This |
796 | might be wasted time but better spend it here than adding a check |
797 | in the fast path. */ |
798 | if (err == 0) |
799 | list_for_each (runp, &stack_cache) |
800 | { |
801 | err = change_stack_perm (list_entry (runp, struct pthread, list) |
802 | #ifdef NEED_SEPARATE_REGISTER_STACK |
803 | , pagemask |
804 | #endif |
805 | ); |
806 | if (err != 0) |
807 | break; |
808 | } |
809 | |
810 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
811 | |
812 | return err; |
813 | } |
814 | |
815 | |
816 | /* In case of a fork() call the memory allocation in the child will be |
817 | the same but only one thread is running. All stacks except that of |
818 | the one running thread are not used anymore. We have to recycle |
819 | them. */ |
820 | void |
821 | __reclaim_stacks (void) |
822 | { |
823 | struct pthread *self = (struct pthread *) THREAD_SELF; |
824 | |
825 | /* No locking necessary. The caller is the only stack in use. But |
826 | we have to be aware that we might have interrupted a list |
827 | operation. */ |
828 | |
829 | if (in_flight_stack != 0) |
830 | { |
831 | bool add_p = in_flight_stack & 1; |
832 | list_t *elem = (list_t *) (in_flight_stack & ~(uintptr_t) 1); |
833 | |
834 | if (add_p) |
835 | { |
836 | /* We always add at the beginning of the list. So in this case we |
837 | only need to check the beginning of these lists to see if the |
838 | pointers at the head of the list are inconsistent. */ |
839 | list_t *l = NULL; |
840 | |
841 | if (stack_used.next->prev != &stack_used) |
842 | l = &stack_used; |
843 | else if (stack_cache.next->prev != &stack_cache) |
844 | l = &stack_cache; |
845 | |
846 | if (l != NULL) |
847 | { |
848 | assert (l->next->prev == elem); |
849 | elem->next = l->next; |
850 | elem->prev = l; |
851 | l->next = elem; |
852 | } |
853 | } |
854 | else |
855 | { |
856 | /* We can simply always replay the delete operation. */ |
857 | elem->next->prev = elem->prev; |
858 | elem->prev->next = elem->next; |
859 | } |
860 | } |
861 | |
862 | /* Mark all stacks except the still running one as free. */ |
863 | list_t *runp; |
864 | list_for_each (runp, &stack_used) |
865 | { |
866 | struct pthread *curp = list_entry (runp, struct pthread, list); |
867 | if (curp != self) |
868 | { |
869 | /* This marks the stack as free. */ |
870 | curp->tid = 0; |
871 | |
872 | /* The PID field must be initialized for the new process. */ |
873 | curp->pid = self->pid; |
874 | |
875 | /* Account for the size of the stack. */ |
876 | stack_cache_actsize += curp->stackblock_size; |
877 | |
878 | if (curp->specific_used) |
879 | { |
880 | /* Clear the thread-specific data. */ |
881 | memset (curp->specific_1stblock, '\0', |
882 | sizeof (curp->specific_1stblock)); |
883 | |
884 | curp->specific_used = false; |
885 | |
886 | for (size_t cnt = 1; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt) |
887 | if (curp->specific[cnt] != NULL) |
888 | { |
889 | memset (curp->specific[cnt], '\0', |
890 | sizeof (curp->specific_1stblock)); |
891 | |
892 | /* We have allocated the block which we do not |
893 | free here so re-set the bit. */ |
894 | curp->specific_used = true; |
895 | } |
896 | } |
897 | } |
898 | } |
899 | |
900 | /* Reset the PIDs in any cached stacks. */ |
901 | list_for_each (runp, &stack_cache) |
902 | { |
903 | struct pthread *curp = list_entry (runp, struct pthread, list); |
904 | curp->pid = self->pid; |
905 | } |
906 | |
907 | /* Add the stack of all running threads to the cache. */ |
908 | list_splice (&stack_used, &stack_cache); |
909 | |
910 | /* Remove the entry for the current thread to from the cache list |
911 | and add it to the list of running threads. Which of the two |
912 | lists is decided by the user_stack flag. */ |
913 | stack_list_del (&self->list); |
914 | |
915 | /* Re-initialize the lists for all the threads. */ |
916 | INIT_LIST_HEAD (&stack_used); |
917 | INIT_LIST_HEAD (&__stack_user); |
918 | |
919 | if (__glibc_unlikely (THREAD_GETMEM (self, user_stack))) |
920 | list_add (&self->list, &__stack_user); |
921 | else |
922 | list_add (&self->list, &stack_used); |
923 | |
924 | /* There is one thread running. */ |
925 | __nptl_nthreads = 1; |
926 | |
927 | in_flight_stack = 0; |
928 | |
929 | /* Initialize locks. */ |
930 | stack_cache_lock = LLL_LOCK_INITIALIZER; |
931 | __default_pthread_attr_lock = LLL_LOCK_INITIALIZER; |
932 | } |
933 | |
934 | |
935 | #if HP_TIMING_AVAIL |
936 | # undef __find_thread_by_id |
937 | /* Find a thread given the thread ID. */ |
938 | attribute_hidden |
939 | struct pthread * |
940 | __find_thread_by_id (pid_t tid) |
941 | { |
942 | struct pthread *result = NULL; |
943 | |
944 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
945 | |
946 | /* Iterate over the list with system-allocated threads first. */ |
947 | list_t *runp; |
948 | list_for_each (runp, &stack_used) |
949 | { |
950 | struct pthread *curp; |
951 | |
952 | curp = list_entry (runp, struct pthread, list); |
953 | |
954 | if (curp->tid == tid) |
955 | { |
956 | result = curp; |
957 | goto out; |
958 | } |
959 | } |
960 | |
961 | /* Now the list with threads using user-allocated stacks. */ |
962 | list_for_each (runp, &__stack_user) |
963 | { |
964 | struct pthread *curp; |
965 | |
966 | curp = list_entry (runp, struct pthread, list); |
967 | |
968 | if (curp->tid == tid) |
969 | { |
970 | result = curp; |
971 | goto out; |
972 | } |
973 | } |
974 | |
975 | out: |
976 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
977 | |
978 | return result; |
979 | } |
980 | #endif |
981 | |
982 | |
983 | #ifdef SIGSETXID |
984 | static void |
985 | internal_function |
986 | setxid_mark_thread (struct xid_command *cmdp, struct pthread *t) |
987 | { |
988 | int ch; |
989 | |
990 | /* Wait until this thread is cloned. */ |
991 | if (t->setxid_futex == -1 |
992 | && ! atomic_compare_and_exchange_bool_acq (&t->setxid_futex, -2, -1)) |
993 | do |
994 | futex_wait_simple (&t->setxid_futex, -2, FUTEX_PRIVATE); |
995 | while (t->setxid_futex == -2); |
996 | |
997 | /* Don't let the thread exit before the setxid handler runs. */ |
998 | t->setxid_futex = 0; |
999 | |
1000 | do |
1001 | { |
1002 | ch = t->cancelhandling; |
1003 | |
1004 | /* If the thread is exiting right now, ignore it. */ |
1005 | if ((ch & EXITING_BITMASK) != 0) |
1006 | { |
1007 | /* Release the futex if there is no other setxid in |
1008 | progress. */ |
1009 | if ((ch & SETXID_BITMASK) == 0) |
1010 | { |
1011 | t->setxid_futex = 1; |
1012 | futex_wake (&t->setxid_futex, 1, FUTEX_PRIVATE); |
1013 | } |
1014 | return; |
1015 | } |
1016 | } |
1017 | while (atomic_compare_and_exchange_bool_acq (&t->cancelhandling, |
1018 | ch | SETXID_BITMASK, ch)); |
1019 | } |
1020 | |
1021 | |
1022 | static void |
1023 | internal_function |
1024 | setxid_unmark_thread (struct xid_command *cmdp, struct pthread *t) |
1025 | { |
1026 | int ch; |
1027 | |
1028 | do |
1029 | { |
1030 | ch = t->cancelhandling; |
1031 | if ((ch & SETXID_BITMASK) == 0) |
1032 | return; |
1033 | } |
1034 | while (atomic_compare_and_exchange_bool_acq (&t->cancelhandling, |
1035 | ch & ~SETXID_BITMASK, ch)); |
1036 | |
1037 | /* Release the futex just in case. */ |
1038 | t->setxid_futex = 1; |
1039 | futex_wake (&t->setxid_futex, 1, FUTEX_PRIVATE); |
1040 | } |
1041 | |
1042 | |
1043 | static int |
1044 | internal_function |
1045 | setxid_signal_thread (struct xid_command *cmdp, struct pthread *t) |
1046 | { |
1047 | if ((t->cancelhandling & SETXID_BITMASK) == 0) |
1048 | return 0; |
1049 | |
1050 | int val; |
1051 | INTERNAL_SYSCALL_DECL (err); |
1052 | val = INTERNAL_SYSCALL (tgkill, err, 3, THREAD_GETMEM (THREAD_SELF, pid), |
1053 | t->tid, SIGSETXID); |
1054 | |
1055 | /* If this failed, it must have had not started yet or else exited. */ |
1056 | if (!INTERNAL_SYSCALL_ERROR_P (val, err)) |
1057 | { |
1058 | atomic_increment (&cmdp->cntr); |
1059 | return 1; |
1060 | } |
1061 | else |
1062 | return 0; |
1063 | } |
1064 | |
1065 | /* Check for consistency across set*id system call results. The abort |
1066 | should not happen as long as all privileges changes happen through |
1067 | the glibc wrappers. ERROR must be 0 (no error) or an errno |
1068 | code. */ |
1069 | void |
1070 | attribute_hidden |
1071 | __nptl_setxid_error (struct xid_command *cmdp, int error) |
1072 | { |
1073 | do |
1074 | { |
1075 | int olderror = cmdp->error; |
1076 | if (olderror == error) |
1077 | break; |
1078 | if (olderror != -1) |
1079 | /* Mismatch between current and previous results. */ |
1080 | abort (); |
1081 | } |
1082 | while (atomic_compare_and_exchange_bool_acq (&cmdp->error, error, -1)); |
1083 | } |
1084 | |
1085 | int |
1086 | attribute_hidden |
1087 | __nptl_setxid (struct xid_command *cmdp) |
1088 | { |
1089 | int signalled; |
1090 | int result; |
1091 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
1092 | |
1093 | __xidcmd = cmdp; |
1094 | cmdp->cntr = 0; |
1095 | cmdp->error = -1; |
1096 | |
1097 | struct pthread *self = THREAD_SELF; |
1098 | |
1099 | /* Iterate over the list with system-allocated threads first. */ |
1100 | list_t *runp; |
1101 | list_for_each (runp, &stack_used) |
1102 | { |
1103 | struct pthread *t = list_entry (runp, struct pthread, list); |
1104 | if (t == self) |
1105 | continue; |
1106 | |
1107 | setxid_mark_thread (cmdp, t); |
1108 | } |
1109 | |
1110 | /* Now the list with threads using user-allocated stacks. */ |
1111 | list_for_each (runp, &__stack_user) |
1112 | { |
1113 | struct pthread *t = list_entry (runp, struct pthread, list); |
1114 | if (t == self) |
1115 | continue; |
1116 | |
1117 | setxid_mark_thread (cmdp, t); |
1118 | } |
1119 | |
1120 | /* Iterate until we don't succeed in signalling anyone. That means |
1121 | we have gotten all running threads, and their children will be |
1122 | automatically correct once started. */ |
1123 | do |
1124 | { |
1125 | signalled = 0; |
1126 | |
1127 | list_for_each (runp, &stack_used) |
1128 | { |
1129 | struct pthread *t = list_entry (runp, struct pthread, list); |
1130 | if (t == self) |
1131 | continue; |
1132 | |
1133 | signalled += setxid_signal_thread (cmdp, t); |
1134 | } |
1135 | |
1136 | list_for_each (runp, &__stack_user) |
1137 | { |
1138 | struct pthread *t = list_entry (runp, struct pthread, list); |
1139 | if (t == self) |
1140 | continue; |
1141 | |
1142 | signalled += setxid_signal_thread (cmdp, t); |
1143 | } |
1144 | |
1145 | int cur = cmdp->cntr; |
1146 | while (cur != 0) |
1147 | { |
1148 | futex_wait_simple ((unsigned int *) &cmdp->cntr, cur, |
1149 | FUTEX_PRIVATE); |
1150 | cur = cmdp->cntr; |
1151 | } |
1152 | } |
1153 | while (signalled != 0); |
1154 | |
1155 | /* Clean up flags, so that no thread blocks during exit waiting |
1156 | for a signal which will never come. */ |
1157 | list_for_each (runp, &stack_used) |
1158 | { |
1159 | struct pthread *t = list_entry (runp, struct pthread, list); |
1160 | if (t == self) |
1161 | continue; |
1162 | |
1163 | setxid_unmark_thread (cmdp, t); |
1164 | } |
1165 | |
1166 | list_for_each (runp, &__stack_user) |
1167 | { |
1168 | struct pthread *t = list_entry (runp, struct pthread, list); |
1169 | if (t == self) |
1170 | continue; |
1171 | |
1172 | setxid_unmark_thread (cmdp, t); |
1173 | } |
1174 | |
1175 | /* This must be last, otherwise the current thread might not have |
1176 | permissions to send SIGSETXID syscall to the other threads. */ |
1177 | INTERNAL_SYSCALL_DECL (err); |
1178 | result = INTERNAL_SYSCALL_NCS (cmdp->syscall_no, err, 3, |
1179 | cmdp->id[0], cmdp->id[1], cmdp->id[2]); |
1180 | int error = 0; |
1181 | if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (result, err))) |
1182 | { |
1183 | error = INTERNAL_SYSCALL_ERRNO (result, err); |
1184 | __set_errno (error); |
1185 | result = -1; |
1186 | } |
1187 | __nptl_setxid_error (cmdp, error); |
1188 | |
1189 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
1190 | return result; |
1191 | } |
1192 | #endif /* SIGSETXID. */ |
1193 | |
1194 | |
1195 | static inline void __attribute__((always_inline)) |
1196 | init_one_static_tls (struct pthread *curp, struct link_map *map) |
1197 | { |
1198 | # if TLS_TCB_AT_TP |
1199 | void *dest = (char *) curp - map->l_tls_offset; |
1200 | # elif TLS_DTV_AT_TP |
1201 | void *dest = (char *) curp + map->l_tls_offset + TLS_PRE_TCB_SIZE; |
1202 | # else |
1203 | # error "Either TLS_TCB_AT_TP or TLS_DTV_AT_TP must be defined" |
1204 | # endif |
1205 | |
1206 | /* We cannot delay the initialization of the Static TLS area, since |
1207 | it can be accessed with LE or IE, but since the DTV is only used |
1208 | by GD and LD, we can delay its update to avoid a race. */ |
1209 | memset (__mempcpy (dest, map->l_tls_initimage, map->l_tls_initimage_size), |
1210 | '\0', map->l_tls_blocksize - map->l_tls_initimage_size); |
1211 | } |
1212 | |
1213 | void |
1214 | attribute_hidden |
1215 | __pthread_init_static_tls (struct link_map *map) |
1216 | { |
1217 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
1218 | |
1219 | /* Iterate over the list with system-allocated threads first. */ |
1220 | list_t *runp; |
1221 | list_for_each (runp, &stack_used) |
1222 | init_one_static_tls (list_entry (runp, struct pthread, list), map); |
1223 | |
1224 | /* Now the list with threads using user-allocated stacks. */ |
1225 | list_for_each (runp, &__stack_user) |
1226 | init_one_static_tls (list_entry (runp, struct pthread, list), map); |
1227 | |
1228 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
1229 | } |
1230 | |
1231 | |
1232 | void |
1233 | attribute_hidden |
1234 | __wait_lookup_done (void) |
1235 | { |
1236 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
1237 | |
1238 | struct pthread *self = THREAD_SELF; |
1239 | |
1240 | /* Iterate over the list with system-allocated threads first. */ |
1241 | list_t *runp; |
1242 | list_for_each (runp, &stack_used) |
1243 | { |
1244 | struct pthread *t = list_entry (runp, struct pthread, list); |
1245 | if (t == self || t->header.gscope_flag == THREAD_GSCOPE_FLAG_UNUSED) |
1246 | continue; |
1247 | |
1248 | int *const gscope_flagp = &t->header.gscope_flag; |
1249 | |
1250 | /* We have to wait until this thread is done with the global |
1251 | scope. First tell the thread that we are waiting and |
1252 | possibly have to be woken. */ |
1253 | if (atomic_compare_and_exchange_bool_acq (gscope_flagp, |
1254 | THREAD_GSCOPE_FLAG_WAIT, |
1255 | THREAD_GSCOPE_FLAG_USED)) |
1256 | continue; |
1257 | |
1258 | do |
1259 | futex_wait_simple ((unsigned int *) gscope_flagp, |
1260 | THREAD_GSCOPE_FLAG_WAIT, FUTEX_PRIVATE); |
1261 | while (*gscope_flagp == THREAD_GSCOPE_FLAG_WAIT); |
1262 | } |
1263 | |
1264 | /* Now the list with threads using user-allocated stacks. */ |
1265 | list_for_each (runp, &__stack_user) |
1266 | { |
1267 | struct pthread *t = list_entry (runp, struct pthread, list); |
1268 | if (t == self || t->header.gscope_flag == THREAD_GSCOPE_FLAG_UNUSED) |
1269 | continue; |
1270 | |
1271 | int *const gscope_flagp = &t->header.gscope_flag; |
1272 | |
1273 | /* We have to wait until this thread is done with the global |
1274 | scope. First tell the thread that we are waiting and |
1275 | possibly have to be woken. */ |
1276 | if (atomic_compare_and_exchange_bool_acq (gscope_flagp, |
1277 | THREAD_GSCOPE_FLAG_WAIT, |
1278 | THREAD_GSCOPE_FLAG_USED)) |
1279 | continue; |
1280 | |
1281 | do |
1282 | futex_wait_simple ((unsigned int *) gscope_flagp, |
1283 | THREAD_GSCOPE_FLAG_WAIT, FUTEX_PRIVATE); |
1284 | while (*gscope_flagp == THREAD_GSCOPE_FLAG_WAIT); |
1285 | } |
1286 | |
1287 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
1288 | } |
1289 | |