1 | /* Copyright (C) 2002-2018 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 <ctype.h> |
20 | #include <errno.h> |
21 | #include <stdbool.h> |
22 | #include <stdlib.h> |
23 | #include <string.h> |
24 | #include <stdint.h> |
25 | #include "pthreadP.h" |
26 | #include <hp-timing.h> |
27 | #include <ldsodefs.h> |
28 | #include <atomic.h> |
29 | #include <libc-internal.h> |
30 | #include <resolv.h> |
31 | #include <kernel-features.h> |
32 | #include <exit-thread.h> |
33 | #include <default-sched.h> |
34 | #include <futex-internal.h> |
35 | #include <tls-setup.h> |
36 | #include "libioP.h" |
37 | |
38 | #include <shlib-compat.h> |
39 | |
40 | #include <stap-probe.h> |
41 | |
42 | |
43 | /* Nozero if debugging mode is enabled. */ |
44 | int __pthread_debug; |
45 | |
46 | /* Globally enabled events. */ |
47 | static td_thr_events_t __nptl_threads_events __attribute_used__; |
48 | |
49 | /* Pointer to descriptor with the last event. */ |
50 | static struct pthread *__nptl_last_event __attribute_used__; |
51 | |
52 | /* Number of threads running. */ |
53 | unsigned int __nptl_nthreads = 1; |
54 | |
55 | |
56 | /* Code to allocate and deallocate a stack. */ |
57 | #include "allocatestack.c" |
58 | |
59 | /* CONCURRENCY NOTES: |
60 | |
61 | Understanding who is the owner of the 'struct pthread' or 'PD' |
62 | (refers to the value of the 'struct pthread *pd' function argument) |
63 | is critically important in determining exactly which operations are |
64 | allowed and which are not and when, particularly when it comes to the |
65 | implementation of pthread_create, pthread_join, pthread_detach, and |
66 | other functions which all operate on PD. |
67 | |
68 | The owner of PD is responsible for freeing the final resources |
69 | associated with PD, and may examine the memory underlying PD at any |
70 | point in time until it frees it back to the OS or to reuse by the |
71 | runtime. |
72 | |
73 | The thread which calls pthread_create is called the creating thread. |
74 | The creating thread begins as the owner of PD. |
75 | |
76 | During startup the new thread may examine PD in coordination with the |
77 | owner thread (which may be itself). |
78 | |
79 | The four cases of ownership transfer are: |
80 | |
81 | (1) Ownership of PD is released to the process (all threads may use it) |
82 | after the new thread starts in a joinable state |
83 | i.e. pthread_create returns a usable pthread_t. |
84 | |
85 | (2) Ownership of PD is released to the new thread starting in a detached |
86 | state. |
87 | |
88 | (3) Ownership of PD is dynamically released to a running thread via |
89 | pthread_detach. |
90 | |
91 | (4) Ownership of PD is acquired by the thread which calls pthread_join. |
92 | |
93 | Implementation notes: |
94 | |
95 | The PD->stopped_start and thread_ran variables are used to determine |
96 | exactly which of the four ownership states we are in and therefore |
97 | what actions can be taken. For example after (2) we cannot read or |
98 | write from PD anymore since the thread may no longer exist and the |
99 | memory may be unmapped. |
100 | |
101 | It is important to point out that PD->lock is being used both |
102 | similar to a one-shot semaphore and subsequently as a mutex. The |
103 | lock is taken in the parent to force the child to wait, and then the |
104 | child releases the lock. However, this semaphore-like effect is used |
105 | only for synchronizing the parent and child. After startup the lock |
106 | is used like a mutex to create a critical section during which a |
107 | single owner modifies the thread parameters. |
108 | |
109 | The most complicated cases happen during thread startup: |
110 | |
111 | (a) If the created thread is in a detached (PTHREAD_CREATE_DETACHED), |
112 | or joinable (default PTHREAD_CREATE_JOINABLE) state and |
113 | STOPPED_START is true, then the creating thread has ownership of |
114 | PD until the PD->lock is released by pthread_create. If any |
115 | errors occur we are in states (c), (d), or (e) below. |
116 | |
117 | (b) If the created thread is in a detached state |
118 | (PTHREAD_CREATED_DETACHED), and STOPPED_START is false, then the |
119 | creating thread has ownership of PD until it invokes the OS |
120 | kernel's thread creation routine. If this routine returns |
121 | without error, then the created thread owns PD; otherwise, see |
122 | (c) and (e) below. |
123 | |
124 | (c) If the detached thread setup failed and THREAD_RAN is true, then |
125 | the creating thread releases ownership to the new thread by |
126 | sending a cancellation signal. All threads set THREAD_RAN to |
127 | true as quickly as possible after returning from the OS kernel's |
128 | thread creation routine. |
129 | |
130 | (d) If the joinable thread setup failed and THREAD_RAN is true, then |
131 | then the creating thread retains ownership of PD and must cleanup |
132 | state. Ownership cannot be released to the process via the |
133 | return of pthread_create since a non-zero result entails PD is |
134 | undefined and therefore cannot be joined to free the resources. |
135 | We privately call pthread_join on the thread to finish handling |
136 | the resource shutdown (Or at least we should, see bug 19511). |
137 | |
138 | (e) If the thread creation failed and THREAD_RAN is false, then the |
139 | creating thread retains ownership of PD and must cleanup state. |
140 | No waiting for the new thread is required because it never |
141 | started. |
142 | |
143 | The nptl_db interface: |
144 | |
145 | The interface with nptl_db requires that we enqueue PD into a linked |
146 | list and then call a function which the debugger will trap. The PD |
147 | will then be dequeued and control returned to the thread. The caller |
148 | at the time must have ownership of PD and such ownership remains |
149 | after control returns to thread. The enqueued PD is removed from the |
150 | linked list by the nptl_db callback td_thr_event_getmsg. The debugger |
151 | must ensure that the thread does not resume execution, otherwise |
152 | ownership of PD may be lost and examining PD will not be possible. |
153 | |
154 | Note that the GNU Debugger as of (December 10th 2015) commit |
155 | c2c2a31fdb228d41ce3db62b268efea04bd39c18 no longer uses |
156 | td_thr_event_getmsg and several other related nptl_db interfaces. The |
157 | principal reason for this is that nptl_db does not support non-stop |
158 | mode where other threads can run concurrently and modify runtime |
159 | structures currently in use by the debugger and the nptl_db |
160 | interface. |
161 | |
162 | Axioms: |
163 | |
164 | * The create_thread function can never set stopped_start to false. |
165 | * The created thread can read stopped_start but never write to it. |
166 | * The variable thread_ran is set some time after the OS thread |
167 | creation routine returns, how much time after the thread is created |
168 | is unspecified, but it should be as quickly as possible. |
169 | |
170 | */ |
171 | |
172 | /* CREATE THREAD NOTES: |
173 | |
174 | createthread.c defines the create_thread function, and two macros: |
175 | START_THREAD_DEFN and START_THREAD_SELF (see below). |
176 | |
177 | create_thread must initialize PD->stopped_start. It should be true |
178 | if the STOPPED_START parameter is true, or if create_thread needs the |
179 | new thread to synchronize at startup for some other implementation |
180 | reason. If STOPPED_START will be true, then create_thread is obliged |
181 | to lock PD->lock before starting the thread. Then pthread_create |
182 | unlocks PD->lock which synchronizes-with START_THREAD_DEFN in the |
183 | child thread which does an acquire/release of PD->lock as the last |
184 | action before calling the user entry point. The goal of all of this |
185 | is to ensure that the required initial thread attributes are applied |
186 | (by the creating thread) before the new thread runs user code. Note |
187 | that the the functions pthread_getschedparam, pthread_setschedparam, |
188 | pthread_setschedprio, __pthread_tpp_change_priority, and |
189 | __pthread_current_priority reuse the same lock, PD->lock, for a |
190 | similar purpose e.g. synchronizing the setting of similar thread |
191 | attributes. These functions are never called before the thread is |
192 | created, so don't participate in startup syncronization, but given |
193 | that the lock is present already and in the unlocked state, reusing |
194 | it saves space. |
195 | |
196 | The return value is zero for success or an errno code for failure. |
197 | If the return value is ENOMEM, that will be translated to EAGAIN, |
198 | so create_thread need not do that. On failure, *THREAD_RAN should |
199 | be set to true iff the thread actually started up and then got |
200 | canceled before calling user code (*PD->start_routine). */ |
201 | static int create_thread (struct pthread *pd, const struct pthread_attr *attr, |
202 | bool *stopped_start, STACK_VARIABLES_PARMS, |
203 | bool *thread_ran); |
204 | |
205 | #include <createthread.c> |
206 | |
207 | |
208 | struct pthread * |
209 | __find_in_stack_list (struct pthread *pd) |
210 | { |
211 | list_t *entry; |
212 | struct pthread *result = NULL; |
213 | |
214 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
215 | |
216 | list_for_each (entry, &stack_used) |
217 | { |
218 | struct pthread *curp; |
219 | |
220 | curp = list_entry (entry, struct pthread, list); |
221 | if (curp == pd) |
222 | { |
223 | result = curp; |
224 | break; |
225 | } |
226 | } |
227 | |
228 | if (result == NULL) |
229 | list_for_each (entry, &__stack_user) |
230 | { |
231 | struct pthread *curp; |
232 | |
233 | curp = list_entry (entry, struct pthread, list); |
234 | if (curp == pd) |
235 | { |
236 | result = curp; |
237 | break; |
238 | } |
239 | } |
240 | |
241 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
242 | |
243 | return result; |
244 | } |
245 | |
246 | |
247 | /* Deallocate POSIX thread-local-storage. */ |
248 | void |
249 | attribute_hidden |
250 | __nptl_deallocate_tsd (void) |
251 | { |
252 | struct pthread *self = THREAD_SELF; |
253 | |
254 | /* Maybe no data was ever allocated. This happens often so we have |
255 | a flag for this. */ |
256 | if (THREAD_GETMEM (self, specific_used)) |
257 | { |
258 | size_t round; |
259 | size_t cnt; |
260 | |
261 | round = 0; |
262 | do |
263 | { |
264 | size_t idx; |
265 | |
266 | /* So far no new nonzero data entry. */ |
267 | THREAD_SETMEM (self, specific_used, false); |
268 | |
269 | for (cnt = idx = 0; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt) |
270 | { |
271 | struct pthread_key_data *level2; |
272 | |
273 | level2 = THREAD_GETMEM_NC (self, specific, cnt); |
274 | |
275 | if (level2 != NULL) |
276 | { |
277 | size_t inner; |
278 | |
279 | for (inner = 0; inner < PTHREAD_KEY_2NDLEVEL_SIZE; |
280 | ++inner, ++idx) |
281 | { |
282 | void *data = level2[inner].data; |
283 | |
284 | if (data != NULL) |
285 | { |
286 | /* Always clear the data. */ |
287 | level2[inner].data = NULL; |
288 | |
289 | /* Make sure the data corresponds to a valid |
290 | key. This test fails if the key was |
291 | deallocated and also if it was |
292 | re-allocated. It is the user's |
293 | responsibility to free the memory in this |
294 | case. */ |
295 | if (level2[inner].seq |
296 | == __pthread_keys[idx].seq |
297 | /* It is not necessary to register a destructor |
298 | function. */ |
299 | && __pthread_keys[idx].destr != NULL) |
300 | /* Call the user-provided destructor. */ |
301 | __pthread_keys[idx].destr (data); |
302 | } |
303 | } |
304 | } |
305 | else |
306 | idx += PTHREAD_KEY_1STLEVEL_SIZE; |
307 | } |
308 | |
309 | if (THREAD_GETMEM (self, specific_used) == 0) |
310 | /* No data has been modified. */ |
311 | goto just_free; |
312 | } |
313 | /* We only repeat the process a fixed number of times. */ |
314 | while (__builtin_expect (++round < PTHREAD_DESTRUCTOR_ITERATIONS, 0)); |
315 | |
316 | /* Just clear the memory of the first block for reuse. */ |
317 | memset (&THREAD_SELF->specific_1stblock, '\0', |
318 | sizeof (self->specific_1stblock)); |
319 | |
320 | just_free: |
321 | /* Free the memory for the other blocks. */ |
322 | for (cnt = 1; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt) |
323 | { |
324 | struct pthread_key_data *level2; |
325 | |
326 | level2 = THREAD_GETMEM_NC (self, specific, cnt); |
327 | if (level2 != NULL) |
328 | { |
329 | /* The first block is allocated as part of the thread |
330 | descriptor. */ |
331 | free (level2); |
332 | THREAD_SETMEM_NC (self, specific, cnt, NULL); |
333 | } |
334 | } |
335 | |
336 | THREAD_SETMEM (self, specific_used, false); |
337 | } |
338 | } |
339 | |
340 | |
341 | /* Deallocate a thread's stack after optionally making sure the thread |
342 | descriptor is still valid. */ |
343 | void |
344 | __free_tcb (struct pthread *pd) |
345 | { |
346 | /* The thread is exiting now. */ |
347 | if (__builtin_expect (atomic_bit_test_set (&pd->cancelhandling, |
348 | TERMINATED_BIT) == 0, 1)) |
349 | { |
350 | /* Remove the descriptor from the list. */ |
351 | if (DEBUGGING_P && __find_in_stack_list (pd) == NULL) |
352 | /* Something is really wrong. The descriptor for a still |
353 | running thread is gone. */ |
354 | abort (); |
355 | |
356 | /* Free TPP data. */ |
357 | if (__glibc_unlikely (pd->tpp != NULL)) |
358 | { |
359 | struct priority_protection_data *tpp = pd->tpp; |
360 | |
361 | pd->tpp = NULL; |
362 | free (tpp); |
363 | } |
364 | |
365 | /* Queue the stack memory block for reuse and exit the process. The |
366 | kernel will signal via writing to the address returned by |
367 | QUEUE-STACK when the stack is available. */ |
368 | __deallocate_stack (pd); |
369 | } |
370 | } |
371 | |
372 | |
373 | /* Local function to start thread and handle cleanup. |
374 | createthread.c defines the macro START_THREAD_DEFN to the |
375 | declaration that its create_thread function will refer to, and |
376 | START_THREAD_SELF to the expression to optimally deliver the new |
377 | thread's THREAD_SELF value. */ |
378 | START_THREAD_DEFN |
379 | { |
380 | struct pthread *pd = START_THREAD_SELF; |
381 | |
382 | #if HP_TIMING_AVAIL |
383 | /* Remember the time when the thread was started. */ |
384 | hp_timing_t now; |
385 | HP_TIMING_NOW (now); |
386 | THREAD_SETMEM (pd, cpuclock_offset, now); |
387 | #endif |
388 | |
389 | /* Initialize resolver state pointer. */ |
390 | __resp = &pd->res; |
391 | |
392 | /* Initialize pointers to locale data. */ |
393 | __ctype_init (); |
394 | |
395 | /* Allow setxid from now onwards. */ |
396 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) == -2)) |
397 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
398 | |
399 | #ifdef __NR_set_robust_list |
400 | # ifndef __ASSUME_SET_ROBUST_LIST |
401 | if (__set_robust_list_avail >= 0) |
402 | # endif |
403 | { |
404 | INTERNAL_SYSCALL_DECL (err); |
405 | /* This call should never fail because the initial call in init.c |
406 | succeeded. */ |
407 | INTERNAL_SYSCALL (set_robust_list, err, 2, &pd->robust_head, |
408 | sizeof (struct robust_list_head)); |
409 | } |
410 | #endif |
411 | |
412 | #ifdef SIGCANCEL |
413 | /* If the parent was running cancellation handlers while creating |
414 | the thread the new thread inherited the signal mask. Reset the |
415 | cancellation signal mask. */ |
416 | if (__glibc_unlikely (pd->parent_cancelhandling & CANCELING_BITMASK)) |
417 | { |
418 | INTERNAL_SYSCALL_DECL (err); |
419 | sigset_t mask; |
420 | __sigemptyset (&mask); |
421 | __sigaddset (&mask, SIGCANCEL); |
422 | (void) INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_UNBLOCK, &mask, |
423 | NULL, _NSIG / 8); |
424 | } |
425 | #endif |
426 | |
427 | /* This is where the try/finally block should be created. For |
428 | compilers without that support we do use setjmp. */ |
429 | struct pthread_unwind_buf unwind_buf; |
430 | |
431 | int not_first_call; |
432 | not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf); |
433 | |
434 | /* No previous handlers. NB: This must be done after setjmp since the |
435 | private space in the unwind jump buffer may overlap space used by |
436 | setjmp to store extra architecture-specific information which is |
437 | never used by the cancellation-specific __libc_unwind_longjmp. |
438 | |
439 | The private space is allowed to overlap because the unwinder never |
440 | has to return through any of the jumped-to call frames, and thus |
441 | only a minimum amount of saved data need be stored, and for example, |
442 | need not include the process signal mask information. This is all |
443 | an optimization to reduce stack usage when pushing cancellation |
444 | handlers. */ |
445 | unwind_buf.priv.data.prev = NULL; |
446 | unwind_buf.priv.data.cleanup = NULL; |
447 | |
448 | if (__glibc_likely (! not_first_call)) |
449 | { |
450 | /* Store the new cleanup handler info. */ |
451 | THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf); |
452 | |
453 | /* We are either in (a) or (b), and in either case we either own |
454 | PD already (2) or are about to own PD (1), and so our only |
455 | restriction would be that we can't free PD until we know we |
456 | have ownership (see CONCURRENCY NOTES above). */ |
457 | if (__glibc_unlikely (pd->stopped_start)) |
458 | { |
459 | int oldtype = CANCEL_ASYNC (); |
460 | |
461 | /* Get the lock the parent locked to force synchronization. */ |
462 | lll_lock (pd->lock, LLL_PRIVATE); |
463 | |
464 | /* We have ownership of PD now. */ |
465 | |
466 | /* And give it up right away. */ |
467 | lll_unlock (pd->lock, LLL_PRIVATE); |
468 | |
469 | CANCEL_RESET (oldtype); |
470 | } |
471 | |
472 | LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg); |
473 | |
474 | /* Run the code the user provided. */ |
475 | void *ret; |
476 | if (pd->c11) |
477 | { |
478 | /* The function pointer of the c11 thread start is cast to an incorrect |
479 | type on __pthread_create_2_1 call, however it is casted back to correct |
480 | one so the call behavior is well-defined (it is assumed that pointers |
481 | to void are able to represent all values of int. */ |
482 | int (*start)(void*) = (int (*) (void*)) pd->start_routine; |
483 | ret = (void*) (uintptr_t) start (pd->arg); |
484 | } |
485 | else |
486 | ret = pd->start_routine (pd->arg); |
487 | THREAD_SETMEM (pd, result, ret); |
488 | } |
489 | |
490 | /* Call destructors for the thread_local TLS variables. */ |
491 | #ifndef SHARED |
492 | if (&__call_tls_dtors != NULL) |
493 | #endif |
494 | __call_tls_dtors (); |
495 | |
496 | /* Run the destructor for the thread-local data. */ |
497 | __nptl_deallocate_tsd (); |
498 | |
499 | /* Clean up any state libc stored in thread-local variables. */ |
500 | __libc_thread_freeres (); |
501 | |
502 | /* If this is the last thread we terminate the process now. We |
503 | do not notify the debugger, it might just irritate it if there |
504 | is no thread left. */ |
505 | if (__glibc_unlikely (atomic_decrement_and_test (&__nptl_nthreads))) |
506 | /* This was the last thread. */ |
507 | exit (0); |
508 | |
509 | /* Report the death of the thread if this is wanted. */ |
510 | if (__glibc_unlikely (pd->report_events)) |
511 | { |
512 | /* See whether TD_DEATH is in any of the mask. */ |
513 | const int idx = __td_eventword (TD_DEATH); |
514 | const uint32_t mask = __td_eventmask (TD_DEATH); |
515 | |
516 | if ((mask & (__nptl_threads_events.event_bits[idx] |
517 | | pd->eventbuf.eventmask.event_bits[idx])) != 0) |
518 | { |
519 | /* Yep, we have to signal the death. Add the descriptor to |
520 | the list but only if it is not already on it. */ |
521 | if (pd->nextevent == NULL) |
522 | { |
523 | pd->eventbuf.eventnum = TD_DEATH; |
524 | pd->eventbuf.eventdata = pd; |
525 | |
526 | do |
527 | pd->nextevent = __nptl_last_event; |
528 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, |
529 | pd, pd->nextevent)); |
530 | } |
531 | |
532 | /* Now call the function which signals the event. See |
533 | CONCURRENCY NOTES for the nptl_db interface comments. */ |
534 | __nptl_death_event (); |
535 | } |
536 | } |
537 | |
538 | /* The thread is exiting now. Don't set this bit until after we've hit |
539 | the event-reporting breakpoint, so that td_thr_get_info on us while at |
540 | the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE. */ |
541 | atomic_bit_set (&pd->cancelhandling, EXITING_BIT); |
542 | |
543 | #ifndef __ASSUME_SET_ROBUST_LIST |
544 | /* If this thread has any robust mutexes locked, handle them now. */ |
545 | # if __PTHREAD_MUTEX_HAVE_PREV |
546 | void *robust = pd->robust_head.list; |
547 | # else |
548 | __pthread_slist_t *robust = pd->robust_list.__next; |
549 | # endif |
550 | /* We let the kernel do the notification if it is able to do so. |
551 | If we have to do it here there for sure are no PI mutexes involved |
552 | since the kernel support for them is even more recent. */ |
553 | if (__set_robust_list_avail < 0 |
554 | && __builtin_expect (robust != (void *) &pd->robust_head, 0)) |
555 | { |
556 | do |
557 | { |
558 | struct __pthread_mutex_s *this = (struct __pthread_mutex_s *) |
559 | ((char *) robust - offsetof (struct __pthread_mutex_s, |
560 | __list.__next)); |
561 | robust = *((void **) robust); |
562 | |
563 | # if __PTHREAD_MUTEX_HAVE_PREV |
564 | this->__list.__prev = NULL; |
565 | # endif |
566 | this->__list.__next = NULL; |
567 | |
568 | atomic_or (&this->__lock, FUTEX_OWNER_DIED); |
569 | futex_wake ((unsigned int *) &this->__lock, 1, |
570 | /* XYZ */ FUTEX_SHARED); |
571 | } |
572 | while (robust != (void *) &pd->robust_head); |
573 | } |
574 | #endif |
575 | |
576 | advise_stack_range (pd->stackblock, pd->stackblock_size, (uintptr_t) pd, |
577 | pd->guardsize); |
578 | |
579 | /* If the thread is detached free the TCB. */ |
580 | if (IS_DETACHED (pd)) |
581 | /* Free the TCB. */ |
582 | __free_tcb (pd); |
583 | else if (__glibc_unlikely (pd->cancelhandling & SETXID_BITMASK)) |
584 | { |
585 | /* Some other thread might call any of the setXid functions and expect |
586 | us to reply. In this case wait until we did that. */ |
587 | do |
588 | /* XXX This differs from the typical futex_wait_simple pattern in that |
589 | the futex_wait condition (setxid_futex) is different from the |
590 | condition used in the surrounding loop (cancelhandling). We need |
591 | to check and document why this is correct. */ |
592 | futex_wait_simple (&pd->setxid_futex, 0, FUTEX_PRIVATE); |
593 | while (pd->cancelhandling & SETXID_BITMASK); |
594 | |
595 | /* Reset the value so that the stack can be reused. */ |
596 | pd->setxid_futex = 0; |
597 | } |
598 | |
599 | /* We cannot call '_exit' here. '_exit' will terminate the process. |
600 | |
601 | The 'exit' implementation in the kernel will signal when the |
602 | process is really dead since 'clone' got passed the CLONE_CHILD_CLEARTID |
603 | flag. The 'tid' field in the TCB will be set to zero. |
604 | |
605 | The exit code is zero since in case all threads exit by calling |
606 | 'pthread_exit' the exit status must be 0 (zero). */ |
607 | __exit_thread (); |
608 | |
609 | /* NOTREACHED */ |
610 | } |
611 | |
612 | |
613 | /* Return true iff obliged to report TD_CREATE events. */ |
614 | static bool |
615 | report_thread_creation (struct pthread *pd) |
616 | { |
617 | if (__glibc_unlikely (THREAD_GETMEM (THREAD_SELF, report_events))) |
618 | { |
619 | /* The parent thread is supposed to report events. |
620 | Check whether the TD_CREATE event is needed, too. */ |
621 | const size_t idx = __td_eventword (TD_CREATE); |
622 | const uint32_t mask = __td_eventmask (TD_CREATE); |
623 | |
624 | return ((mask & (__nptl_threads_events.event_bits[idx] |
625 | | pd->eventbuf.eventmask.event_bits[idx])) != 0); |
626 | } |
627 | return false; |
628 | } |
629 | |
630 | |
631 | int |
632 | __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr, |
633 | void *(*start_routine) (void *), void *arg) |
634 | { |
635 | STACK_VARIABLES; |
636 | |
637 | const struct pthread_attr *iattr = (struct pthread_attr *) attr; |
638 | struct pthread_attr default_attr; |
639 | bool free_cpuset = false; |
640 | bool c11 = (attr == ATTR_C11_THREAD); |
641 | if (iattr == NULL || c11) |
642 | { |
643 | lll_lock (__default_pthread_attr_lock, LLL_PRIVATE); |
644 | default_attr = __default_pthread_attr; |
645 | size_t cpusetsize = default_attr.cpusetsize; |
646 | if (cpusetsize > 0) |
647 | { |
648 | cpu_set_t *cpuset; |
649 | if (__glibc_likely (__libc_use_alloca (cpusetsize))) |
650 | cpuset = __alloca (cpusetsize); |
651 | else |
652 | { |
653 | cpuset = malloc (cpusetsize); |
654 | if (cpuset == NULL) |
655 | { |
656 | lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE); |
657 | return ENOMEM; |
658 | } |
659 | free_cpuset = true; |
660 | } |
661 | memcpy (cpuset, default_attr.cpuset, cpusetsize); |
662 | default_attr.cpuset = cpuset; |
663 | } |
664 | lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE); |
665 | iattr = &default_attr; |
666 | } |
667 | |
668 | struct pthread *pd = NULL; |
669 | int err = ALLOCATE_STACK (iattr, &pd); |
670 | int retval = 0; |
671 | |
672 | if (__glibc_unlikely (err != 0)) |
673 | /* Something went wrong. Maybe a parameter of the attributes is |
674 | invalid or we could not allocate memory. Note we have to |
675 | translate error codes. */ |
676 | { |
677 | retval = err == ENOMEM ? EAGAIN : err; |
678 | goto out; |
679 | } |
680 | |
681 | |
682 | /* Initialize the TCB. All initializations with zero should be |
683 | performed in 'get_cached_stack'. This way we avoid doing this if |
684 | the stack freshly allocated with 'mmap'. */ |
685 | |
686 | #if TLS_TCB_AT_TP |
687 | /* Reference to the TCB itself. */ |
688 | pd->header.self = pd; |
689 | |
690 | /* Self-reference for TLS. */ |
691 | pd->header.tcb = pd; |
692 | #endif |
693 | |
694 | /* Store the address of the start routine and the parameter. Since |
695 | we do not start the function directly the stillborn thread will |
696 | get the information from its thread descriptor. */ |
697 | pd->start_routine = start_routine; |
698 | pd->arg = arg; |
699 | pd->c11 = c11; |
700 | |
701 | /* Copy the thread attribute flags. */ |
702 | struct pthread *self = THREAD_SELF; |
703 | pd->flags = ((iattr->flags & ~(ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
704 | | (self->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))); |
705 | |
706 | /* Initialize the field for the ID of the thread which is waiting |
707 | for us. This is a self-reference in case the thread is created |
708 | detached. */ |
709 | pd->joinid = iattr->flags & ATTR_FLAG_DETACHSTATE ? pd : NULL; |
710 | |
711 | /* The debug events are inherited from the parent. */ |
712 | pd->eventbuf = self->eventbuf; |
713 | |
714 | |
715 | /* Copy the parent's scheduling parameters. The flags will say what |
716 | is valid and what is not. */ |
717 | pd->schedpolicy = self->schedpolicy; |
718 | pd->schedparam = self->schedparam; |
719 | |
720 | /* Copy the stack guard canary. */ |
721 | #ifdef THREAD_COPY_STACK_GUARD |
722 | THREAD_COPY_STACK_GUARD (pd); |
723 | #endif |
724 | |
725 | /* Copy the pointer guard value. */ |
726 | #ifdef THREAD_COPY_POINTER_GUARD |
727 | THREAD_COPY_POINTER_GUARD (pd); |
728 | #endif |
729 | |
730 | /* Setup tcbhead. */ |
731 | tls_setup_tcbhead (pd); |
732 | |
733 | /* Verify the sysinfo bits were copied in allocate_stack if needed. */ |
734 | #ifdef NEED_DL_SYSINFO |
735 | CHECK_THREAD_SYSINFO (pd); |
736 | #endif |
737 | |
738 | /* Inform start_thread (above) about cancellation state that might |
739 | translate into inherited signal state. */ |
740 | pd->parent_cancelhandling = THREAD_GETMEM (THREAD_SELF, cancelhandling); |
741 | |
742 | /* Determine scheduling parameters for the thread. */ |
743 | if (__builtin_expect ((iattr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0, 0) |
744 | && (iattr->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) != 0) |
745 | { |
746 | /* Use the scheduling parameters the user provided. */ |
747 | if (iattr->flags & ATTR_FLAG_POLICY_SET) |
748 | { |
749 | pd->schedpolicy = iattr->schedpolicy; |
750 | pd->flags |= ATTR_FLAG_POLICY_SET; |
751 | } |
752 | if (iattr->flags & ATTR_FLAG_SCHED_SET) |
753 | { |
754 | /* The values were validated in pthread_attr_setschedparam. */ |
755 | pd->schedparam = iattr->schedparam; |
756 | pd->flags |= ATTR_FLAG_SCHED_SET; |
757 | } |
758 | |
759 | if ((pd->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
760 | != (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
761 | collect_default_sched (pd); |
762 | } |
763 | |
764 | if (__glibc_unlikely (__nptl_nthreads == 1)) |
765 | _IO_enable_locks (); |
766 | |
767 | /* Pass the descriptor to the caller. */ |
768 | *newthread = (pthread_t) pd; |
769 | |
770 | LIBC_PROBE (pthread_create, 4, newthread, attr, start_routine, arg); |
771 | |
772 | /* One more thread. We cannot have the thread do this itself, since it |
773 | might exist but not have been scheduled yet by the time we've returned |
774 | and need to check the value to behave correctly. We must do it before |
775 | creating the thread, in case it does get scheduled first and then |
776 | might mistakenly think it was the only thread. In the failure case, |
777 | we momentarily store a false value; this doesn't matter because there |
778 | is no kosher thing a signal handler interrupting us right here can do |
779 | that cares whether the thread count is correct. */ |
780 | atomic_increment (&__nptl_nthreads); |
781 | |
782 | /* Our local value of stopped_start and thread_ran can be accessed at |
783 | any time. The PD->stopped_start may only be accessed if we have |
784 | ownership of PD (see CONCURRENCY NOTES above). */ |
785 | bool stopped_start = false; bool thread_ran = false; |
786 | |
787 | /* Start the thread. */ |
788 | if (__glibc_unlikely (report_thread_creation (pd))) |
789 | { |
790 | stopped_start = true; |
791 | |
792 | /* We always create the thread stopped at startup so we can |
793 | notify the debugger. */ |
794 | retval = create_thread (pd, iattr, &stopped_start, |
795 | STACK_VARIABLES_ARGS, &thread_ran); |
796 | if (retval == 0) |
797 | { |
798 | /* We retain ownership of PD until (a) (see CONCURRENCY NOTES |
799 | above). */ |
800 | |
801 | /* Assert stopped_start is true in both our local copy and the |
802 | PD copy. */ |
803 | assert (stopped_start); |
804 | assert (pd->stopped_start); |
805 | |
806 | /* Now fill in the information about the new thread in |
807 | the newly created thread's data structure. We cannot let |
808 | the new thread do this since we don't know whether it was |
809 | already scheduled when we send the event. */ |
810 | pd->eventbuf.eventnum = TD_CREATE; |
811 | pd->eventbuf.eventdata = pd; |
812 | |
813 | /* Enqueue the descriptor. */ |
814 | do |
815 | pd->nextevent = __nptl_last_event; |
816 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, |
817 | pd, pd->nextevent) |
818 | != 0); |
819 | |
820 | /* Now call the function which signals the event. See |
821 | CONCURRENCY NOTES for the nptl_db interface comments. */ |
822 | __nptl_create_event (); |
823 | } |
824 | } |
825 | else |
826 | retval = create_thread (pd, iattr, &stopped_start, |
827 | STACK_VARIABLES_ARGS, &thread_ran); |
828 | |
829 | if (__glibc_unlikely (retval != 0)) |
830 | { |
831 | if (thread_ran) |
832 | /* State (c) or (d) and we may not have PD ownership (see |
833 | CONCURRENCY NOTES above). We can assert that STOPPED_START |
834 | must have been true because thread creation didn't fail, but |
835 | thread attribute setting did. */ |
836 | /* See bug 19511 which explains why doing nothing here is a |
837 | resource leak for a joinable thread. */ |
838 | assert (stopped_start); |
839 | else |
840 | { |
841 | /* State (e) and we have ownership of PD (see CONCURRENCY |
842 | NOTES above). */ |
843 | |
844 | /* Oops, we lied for a second. */ |
845 | atomic_decrement (&__nptl_nthreads); |
846 | |
847 | /* Perhaps a thread wants to change the IDs and is waiting for this |
848 | stillborn thread. */ |
849 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) |
850 | == -2)) |
851 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
852 | |
853 | /* Free the resources. */ |
854 | __deallocate_stack (pd); |
855 | } |
856 | |
857 | /* We have to translate error codes. */ |
858 | if (retval == ENOMEM) |
859 | retval = EAGAIN; |
860 | } |
861 | else |
862 | { |
863 | /* We don't know if we have PD ownership. Once we check the local |
864 | stopped_start we'll know if we're in state (a) or (b) (see |
865 | CONCURRENCY NOTES above). */ |
866 | if (stopped_start) |
867 | /* State (a), we own PD. The thread blocked on this lock either |
868 | because we're doing TD_CREATE event reporting, or for some |
869 | other reason that create_thread chose. Now let it run |
870 | free. */ |
871 | lll_unlock (pd->lock, LLL_PRIVATE); |
872 | |
873 | /* We now have for sure more than one thread. The main thread might |
874 | not yet have the flag set. No need to set the global variable |
875 | again if this is what we use. */ |
876 | THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1); |
877 | } |
878 | |
879 | out: |
880 | if (__glibc_unlikely (free_cpuset)) |
881 | free (default_attr.cpuset); |
882 | |
883 | return retval; |
884 | } |
885 | versioned_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1); |
886 | |
887 | |
888 | #if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1) |
889 | int |
890 | __pthread_create_2_0 (pthread_t *newthread, const pthread_attr_t *attr, |
891 | void *(*start_routine) (void *), void *arg) |
892 | { |
893 | /* The ATTR attribute is not really of type `pthread_attr_t *'. It has |
894 | the old size and access to the new members might crash the program. |
895 | We convert the struct now. */ |
896 | struct pthread_attr new_attr; |
897 | |
898 | if (attr != NULL) |
899 | { |
900 | struct pthread_attr *iattr = (struct pthread_attr *) attr; |
901 | size_t ps = __getpagesize (); |
902 | |
903 | /* Copy values from the user-provided attributes. */ |
904 | new_attr.schedparam = iattr->schedparam; |
905 | new_attr.schedpolicy = iattr->schedpolicy; |
906 | new_attr.flags = iattr->flags; |
907 | |
908 | /* Fill in default values for the fields not present in the old |
909 | implementation. */ |
910 | new_attr.guardsize = ps; |
911 | new_attr.stackaddr = NULL; |
912 | new_attr.stacksize = 0; |
913 | new_attr.cpuset = NULL; |
914 | |
915 | /* We will pass this value on to the real implementation. */ |
916 | attr = (pthread_attr_t *) &new_attr; |
917 | } |
918 | |
919 | return __pthread_create_2_1 (newthread, attr, start_routine, arg); |
920 | } |
921 | compat_symbol (libpthread, __pthread_create_2_0, pthread_create, |
922 | GLIBC_2_0); |
923 | #endif |
924 | |
925 | /* Information for libthread_db. */ |
926 | |
927 | #include "../nptl_db/db_info.c" |
928 | |
929 | /* If pthread_create is present, libgcc_eh.a and libsupc++.a expects some other POSIX thread |
930 | functions to be present as well. */ |
931 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_lock) |
932 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_trylock) |
933 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_unlock) |
934 | |
935 | PTHREAD_STATIC_FN_REQUIRE (__pthread_once) |
936 | PTHREAD_STATIC_FN_REQUIRE (__pthread_cancel) |
937 | |
938 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_create) |
939 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_delete) |
940 | PTHREAD_STATIC_FN_REQUIRE (__pthread_setspecific) |
941 | PTHREAD_STATIC_FN_REQUIRE (__pthread_getspecific) |
942 | |