1 | /* Copyright (C) 2002-2020 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 | <https://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 | #include <sys/single_threaded.h> |
38 | |
39 | #include <shlib-compat.h> |
40 | |
41 | #include <stap-probe.h> |
42 | |
43 | |
44 | /* Nozero if debugging mode is enabled. */ |
45 | int __pthread_debug; |
46 | |
47 | /* Globally enabled events. */ |
48 | static td_thr_events_t __nptl_threads_events __attribute_used__; |
49 | |
50 | /* Pointer to descriptor with the last event. */ |
51 | static struct pthread *__nptl_last_event __attribute_used__; |
52 | |
53 | /* Number of threads running. */ |
54 | unsigned int __nptl_nthreads = 1; |
55 | |
56 | |
57 | /* Code to allocate and deallocate a stack. */ |
58 | #include "allocatestack.c" |
59 | |
60 | /* CONCURRENCY NOTES: |
61 | |
62 | Understanding who is the owner of the 'struct pthread' or 'PD' |
63 | (refers to the value of the 'struct pthread *pd' function argument) |
64 | is critically important in determining exactly which operations are |
65 | allowed and which are not and when, particularly when it comes to the |
66 | implementation of pthread_create, pthread_join, pthread_detach, and |
67 | other functions which all operate on PD. |
68 | |
69 | The owner of PD is responsible for freeing the final resources |
70 | associated with PD, and may examine the memory underlying PD at any |
71 | point in time until it frees it back to the OS or to reuse by the |
72 | runtime. |
73 | |
74 | The thread which calls pthread_create is called the creating thread. |
75 | The creating thread begins as the owner of PD. |
76 | |
77 | During startup the new thread may examine PD in coordination with the |
78 | owner thread (which may be itself). |
79 | |
80 | The four cases of ownership transfer are: |
81 | |
82 | (1) Ownership of PD is released to the process (all threads may use it) |
83 | after the new thread starts in a joinable state |
84 | i.e. pthread_create returns a usable pthread_t. |
85 | |
86 | (2) Ownership of PD is released to the new thread starting in a detached |
87 | state. |
88 | |
89 | (3) Ownership of PD is dynamically released to a running thread via |
90 | pthread_detach. |
91 | |
92 | (4) Ownership of PD is acquired by the thread which calls pthread_join. |
93 | |
94 | Implementation notes: |
95 | |
96 | The PD->stopped_start and thread_ran variables are used to determine |
97 | exactly which of the four ownership states we are in and therefore |
98 | what actions can be taken. For example after (2) we cannot read or |
99 | write from PD anymore since the thread may no longer exist and the |
100 | memory may be unmapped. |
101 | |
102 | It is important to point out that PD->lock is being used both |
103 | similar to a one-shot semaphore and subsequently as a mutex. The |
104 | lock is taken in the parent to force the child to wait, and then the |
105 | child releases the lock. However, this semaphore-like effect is used |
106 | only for synchronizing the parent and child. After startup the lock |
107 | is used like a mutex to create a critical section during which a |
108 | single owner modifies the thread parameters. |
109 | |
110 | The most complicated cases happen during thread startup: |
111 | |
112 | (a) If the created thread is in a detached (PTHREAD_CREATE_DETACHED), |
113 | or joinable (default PTHREAD_CREATE_JOINABLE) state and |
114 | STOPPED_START is true, then the creating thread has ownership of |
115 | PD until the PD->lock is released by pthread_create. If any |
116 | errors occur we are in states (c), (d), or (e) below. |
117 | |
118 | (b) If the created thread is in a detached state |
119 | (PTHREAD_CREATED_DETACHED), and STOPPED_START is false, then the |
120 | creating thread has ownership of PD until it invokes the OS |
121 | kernel's thread creation routine. If this routine returns |
122 | without error, then the created thread owns PD; otherwise, see |
123 | (c) and (e) below. |
124 | |
125 | (c) If the detached thread setup failed and THREAD_RAN is true, then |
126 | the creating thread releases ownership to the new thread by |
127 | sending a cancellation signal. All threads set THREAD_RAN to |
128 | true as quickly as possible after returning from the OS kernel's |
129 | thread creation routine. |
130 | |
131 | (d) If the joinable thread setup failed and THREAD_RAN is true, then |
132 | then the creating thread retains ownership of PD and must cleanup |
133 | state. Ownership cannot be released to the process via the |
134 | return of pthread_create since a non-zero result entails PD is |
135 | undefined and therefore cannot be joined to free the resources. |
136 | We privately call pthread_join on the thread to finish handling |
137 | the resource shutdown (Or at least we should, see bug 19511). |
138 | |
139 | (e) If the thread creation failed and THREAD_RAN is false, then the |
140 | creating thread retains ownership of PD and must cleanup state. |
141 | No waiting for the new thread is required because it never |
142 | started. |
143 | |
144 | The nptl_db interface: |
145 | |
146 | The interface with nptl_db requires that we enqueue PD into a linked |
147 | list and then call a function which the debugger will trap. The PD |
148 | will then be dequeued and control returned to the thread. The caller |
149 | at the time must have ownership of PD and such ownership remains |
150 | after control returns to thread. The enqueued PD is removed from the |
151 | linked list by the nptl_db callback td_thr_event_getmsg. The debugger |
152 | must ensure that the thread does not resume execution, otherwise |
153 | ownership of PD may be lost and examining PD will not be possible. |
154 | |
155 | Note that the GNU Debugger as of (December 10th 2015) commit |
156 | c2c2a31fdb228d41ce3db62b268efea04bd39c18 no longer uses |
157 | td_thr_event_getmsg and several other related nptl_db interfaces. The |
158 | principal reason for this is that nptl_db does not support non-stop |
159 | mode where other threads can run concurrently and modify runtime |
160 | structures currently in use by the debugger and the nptl_db |
161 | interface. |
162 | |
163 | Axioms: |
164 | |
165 | * The create_thread function can never set stopped_start to false. |
166 | * The created thread can read stopped_start but never write to it. |
167 | * The variable thread_ran is set some time after the OS thread |
168 | creation routine returns, how much time after the thread is created |
169 | is unspecified, but it should be as quickly as possible. |
170 | |
171 | */ |
172 | |
173 | /* CREATE THREAD NOTES: |
174 | |
175 | createthread.c defines the create_thread function, and two macros: |
176 | START_THREAD_DEFN and START_THREAD_SELF (see below). |
177 | |
178 | create_thread must initialize PD->stopped_start. It should be true |
179 | if the STOPPED_START parameter is true, or if create_thread needs the |
180 | new thread to synchronize at startup for some other implementation |
181 | reason. If STOPPED_START will be true, then create_thread is obliged |
182 | to lock PD->lock before starting the thread. Then pthread_create |
183 | unlocks PD->lock which synchronizes-with START_THREAD_DEFN in the |
184 | child thread which does an acquire/release of PD->lock as the last |
185 | action before calling the user entry point. The goal of all of this |
186 | is to ensure that the required initial thread attributes are applied |
187 | (by the creating thread) before the new thread runs user code. Note |
188 | that the the functions pthread_getschedparam, pthread_setschedparam, |
189 | pthread_setschedprio, __pthread_tpp_change_priority, and |
190 | __pthread_current_priority reuse the same lock, PD->lock, for a |
191 | similar purpose e.g. synchronizing the setting of similar thread |
192 | attributes. These functions are never called before the thread is |
193 | created, so don't participate in startup syncronization, but given |
194 | that the lock is present already and in the unlocked state, reusing |
195 | it saves space. |
196 | |
197 | The return value is zero for success or an errno code for failure. |
198 | If the return value is ENOMEM, that will be translated to EAGAIN, |
199 | so create_thread need not do that. On failure, *THREAD_RAN should |
200 | be set to true iff the thread actually started up and then got |
201 | canceled before calling user code (*PD->start_routine). */ |
202 | static int create_thread (struct pthread *pd, const struct pthread_attr *attr, |
203 | bool *stopped_start, STACK_VARIABLES_PARMS, |
204 | bool *thread_ran); |
205 | |
206 | #include <createthread.c> |
207 | |
208 | |
209 | struct pthread * |
210 | __find_in_stack_list (struct pthread *pd) |
211 | { |
212 | list_t *entry; |
213 | struct pthread *result = NULL; |
214 | |
215 | lll_lock (stack_cache_lock, LLL_PRIVATE); |
216 | |
217 | list_for_each (entry, &stack_used) |
218 | { |
219 | struct pthread *curp; |
220 | |
221 | curp = list_entry (entry, struct pthread, list); |
222 | if (curp == pd) |
223 | { |
224 | result = curp; |
225 | break; |
226 | } |
227 | } |
228 | |
229 | if (result == NULL) |
230 | list_for_each (entry, &__stack_user) |
231 | { |
232 | struct pthread *curp; |
233 | |
234 | curp = list_entry (entry, struct pthread, list); |
235 | if (curp == pd) |
236 | { |
237 | result = curp; |
238 | break; |
239 | } |
240 | } |
241 | |
242 | lll_unlock (stack_cache_lock, LLL_PRIVATE); |
243 | |
244 | return result; |
245 | } |
246 | |
247 | |
248 | /* Deallocate POSIX thread-local-storage. */ |
249 | void |
250 | attribute_hidden |
251 | __nptl_deallocate_tsd (void) |
252 | { |
253 | struct pthread *self = THREAD_SELF; |
254 | |
255 | /* Maybe no data was ever allocated. This happens often so we have |
256 | a flag for this. */ |
257 | if (THREAD_GETMEM (self, specific_used)) |
258 | { |
259 | size_t round; |
260 | size_t cnt; |
261 | |
262 | round = 0; |
263 | do |
264 | { |
265 | size_t idx; |
266 | |
267 | /* So far no new nonzero data entry. */ |
268 | THREAD_SETMEM (self, specific_used, false); |
269 | |
270 | for (cnt = idx = 0; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt) |
271 | { |
272 | struct pthread_key_data *level2; |
273 | |
274 | level2 = THREAD_GETMEM_NC (self, specific, cnt); |
275 | |
276 | if (level2 != NULL) |
277 | { |
278 | size_t inner; |
279 | |
280 | for (inner = 0; inner < PTHREAD_KEY_2NDLEVEL_SIZE; |
281 | ++inner, ++idx) |
282 | { |
283 | void *data = level2[inner].data; |
284 | |
285 | if (data != NULL) |
286 | { |
287 | /* Always clear the data. */ |
288 | level2[inner].data = NULL; |
289 | |
290 | /* Make sure the data corresponds to a valid |
291 | key. This test fails if the key was |
292 | deallocated and also if it was |
293 | re-allocated. It is the user's |
294 | responsibility to free the memory in this |
295 | case. */ |
296 | if (level2[inner].seq |
297 | == __pthread_keys[idx].seq |
298 | /* It is not necessary to register a destructor |
299 | function. */ |
300 | && __pthread_keys[idx].destr != NULL) |
301 | /* Call the user-provided destructor. */ |
302 | __pthread_keys[idx].destr (data); |
303 | } |
304 | } |
305 | } |
306 | else |
307 | idx += PTHREAD_KEY_1STLEVEL_SIZE; |
308 | } |
309 | |
310 | if (THREAD_GETMEM (self, specific_used) == 0) |
311 | /* No data has been modified. */ |
312 | goto just_free; |
313 | } |
314 | /* We only repeat the process a fixed number of times. */ |
315 | while (__builtin_expect (++round < PTHREAD_DESTRUCTOR_ITERATIONS, 0)); |
316 | |
317 | /* Just clear the memory of the first block for reuse. */ |
318 | memset (&THREAD_SELF->specific_1stblock, '\0', |
319 | sizeof (self->specific_1stblock)); |
320 | |
321 | just_free: |
322 | /* Free the memory for the other blocks. */ |
323 | for (cnt = 1; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt) |
324 | { |
325 | struct pthread_key_data *level2; |
326 | |
327 | level2 = THREAD_GETMEM_NC (self, specific, cnt); |
328 | if (level2 != NULL) |
329 | { |
330 | /* The first block is allocated as part of the thread |
331 | descriptor. */ |
332 | free (level2); |
333 | THREAD_SETMEM_NC (self, specific, cnt, NULL); |
334 | } |
335 | } |
336 | |
337 | THREAD_SETMEM (self, specific_used, false); |
338 | } |
339 | } |
340 | |
341 | |
342 | /* Deallocate a thread's stack after optionally making sure the thread |
343 | descriptor is still valid. */ |
344 | void |
345 | __free_tcb (struct pthread *pd) |
346 | { |
347 | /* The thread is exiting now. */ |
348 | if (__builtin_expect (atomic_bit_test_set (&pd->cancelhandling, |
349 | TERMINATED_BIT) == 0, 1)) |
350 | { |
351 | /* Remove the descriptor from the list. */ |
352 | if (DEBUGGING_P && __find_in_stack_list (pd) == NULL) |
353 | /* Something is really wrong. The descriptor for a still |
354 | running thread is gone. */ |
355 | abort (); |
356 | |
357 | /* Free TPP data. */ |
358 | if (__glibc_unlikely (pd->tpp != NULL)) |
359 | { |
360 | struct priority_protection_data *tpp = pd->tpp; |
361 | |
362 | pd->tpp = NULL; |
363 | free (tpp); |
364 | } |
365 | |
366 | /* Queue the stack memory block for reuse and exit the process. The |
367 | kernel will signal via writing to the address returned by |
368 | QUEUE-STACK when the stack is available. */ |
369 | __deallocate_stack (pd); |
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 | /* Initialize resolver state pointer. */ |
383 | __resp = &pd->res; |
384 | |
385 | /* Initialize pointers to locale data. */ |
386 | __ctype_init (); |
387 | |
388 | #ifndef __ASSUME_SET_ROBUST_LIST |
389 | if (__set_robust_list_avail >= 0) |
390 | #endif |
391 | { |
392 | /* This call should never fail because the initial call in init.c |
393 | succeeded. */ |
394 | INTERNAL_SYSCALL_CALL (set_robust_list, &pd->robust_head, |
395 | sizeof (struct robust_list_head)); |
396 | } |
397 | |
398 | /* This is where the try/finally block should be created. For |
399 | compilers without that support we do use setjmp. */ |
400 | struct pthread_unwind_buf unwind_buf; |
401 | |
402 | int not_first_call; |
403 | not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf); |
404 | |
405 | /* No previous handlers. NB: This must be done after setjmp since the |
406 | private space in the unwind jump buffer may overlap space used by |
407 | setjmp to store extra architecture-specific information which is |
408 | never used by the cancellation-specific __libc_unwind_longjmp. |
409 | |
410 | The private space is allowed to overlap because the unwinder never |
411 | has to return through any of the jumped-to call frames, and thus |
412 | only a minimum amount of saved data need be stored, and for example, |
413 | need not include the process signal mask information. This is all |
414 | an optimization to reduce stack usage when pushing cancellation |
415 | handlers. */ |
416 | unwind_buf.priv.data.prev = NULL; |
417 | unwind_buf.priv.data.cleanup = NULL; |
418 | |
419 | __libc_signal_restore_set (&pd->sigmask); |
420 | |
421 | /* Allow setxid from now onwards. */ |
422 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) == -2)) |
423 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
424 | |
425 | if (__glibc_likely (! not_first_call)) |
426 | { |
427 | /* Store the new cleanup handler info. */ |
428 | THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf); |
429 | |
430 | /* We are either in (a) or (b), and in either case we either own |
431 | PD already (2) or are about to own PD (1), and so our only |
432 | restriction would be that we can't free PD until we know we |
433 | have ownership (see CONCURRENCY NOTES above). */ |
434 | if (__glibc_unlikely (pd->stopped_start)) |
435 | { |
436 | int oldtype = CANCEL_ASYNC (); |
437 | |
438 | /* Get the lock the parent locked to force synchronization. */ |
439 | lll_lock (pd->lock, LLL_PRIVATE); |
440 | |
441 | /* We have ownership of PD now. */ |
442 | |
443 | /* And give it up right away. */ |
444 | lll_unlock (pd->lock, LLL_PRIVATE); |
445 | |
446 | CANCEL_RESET (oldtype); |
447 | } |
448 | |
449 | LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg); |
450 | |
451 | /* Run the code the user provided. */ |
452 | void *ret; |
453 | if (pd->c11) |
454 | { |
455 | /* The function pointer of the c11 thread start is cast to an incorrect |
456 | type on __pthread_create_2_1 call, however it is casted back to correct |
457 | one so the call behavior is well-defined (it is assumed that pointers |
458 | to void are able to represent all values of int. */ |
459 | int (*start)(void*) = (int (*) (void*)) pd->start_routine; |
460 | ret = (void*) (uintptr_t) start (pd->arg); |
461 | } |
462 | else |
463 | ret = pd->start_routine (pd->arg); |
464 | THREAD_SETMEM (pd, result, ret); |
465 | } |
466 | |
467 | /* Call destructors for the thread_local TLS variables. */ |
468 | #ifndef SHARED |
469 | if (&__call_tls_dtors != NULL) |
470 | #endif |
471 | __call_tls_dtors (); |
472 | |
473 | /* Run the destructor for the thread-local data. */ |
474 | __nptl_deallocate_tsd (); |
475 | |
476 | /* Clean up any state libc stored in thread-local variables. */ |
477 | __libc_thread_freeres (); |
478 | |
479 | /* If this is the last thread we terminate the process now. We |
480 | do not notify the debugger, it might just irritate it if there |
481 | is no thread left. */ |
482 | if (__glibc_unlikely (atomic_decrement_and_test (&__nptl_nthreads))) |
483 | /* This was the last thread. */ |
484 | exit (0); |
485 | |
486 | /* Report the death of the thread if this is wanted. */ |
487 | if (__glibc_unlikely (pd->report_events)) |
488 | { |
489 | /* See whether TD_DEATH is in any of the mask. */ |
490 | const int idx = __td_eventword (TD_DEATH); |
491 | const uint32_t mask = __td_eventmask (TD_DEATH); |
492 | |
493 | if ((mask & (__nptl_threads_events.event_bits[idx] |
494 | | pd->eventbuf.eventmask.event_bits[idx])) != 0) |
495 | { |
496 | /* Yep, we have to signal the death. Add the descriptor to |
497 | the list but only if it is not already on it. */ |
498 | if (pd->nextevent == NULL) |
499 | { |
500 | pd->eventbuf.eventnum = TD_DEATH; |
501 | pd->eventbuf.eventdata = pd; |
502 | |
503 | do |
504 | pd->nextevent = __nptl_last_event; |
505 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, |
506 | pd, pd->nextevent)); |
507 | } |
508 | |
509 | /* Now call the function which signals the event. See |
510 | CONCURRENCY NOTES for the nptl_db interface comments. */ |
511 | __nptl_death_event (); |
512 | } |
513 | } |
514 | |
515 | /* The thread is exiting now. Don't set this bit until after we've hit |
516 | the event-reporting breakpoint, so that td_thr_get_info on us while at |
517 | the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE. */ |
518 | atomic_bit_set (&pd->cancelhandling, EXITING_BIT); |
519 | |
520 | #ifndef __ASSUME_SET_ROBUST_LIST |
521 | /* If this thread has any robust mutexes locked, handle them now. */ |
522 | # if __PTHREAD_MUTEX_HAVE_PREV |
523 | void *robust = pd->robust_head.list; |
524 | # else |
525 | __pthread_slist_t *robust = pd->robust_list.__next; |
526 | # endif |
527 | /* We let the kernel do the notification if it is able to do so. |
528 | If we have to do it here there for sure are no PI mutexes involved |
529 | since the kernel support for them is even more recent. */ |
530 | if (__set_robust_list_avail < 0 |
531 | && __builtin_expect (robust != (void *) &pd->robust_head, 0)) |
532 | { |
533 | do |
534 | { |
535 | struct __pthread_mutex_s *this = (struct __pthread_mutex_s *) |
536 | ((char *) robust - offsetof (struct __pthread_mutex_s, |
537 | __list.__next)); |
538 | robust = *((void **) robust); |
539 | |
540 | # if __PTHREAD_MUTEX_HAVE_PREV |
541 | this->__list.__prev = NULL; |
542 | # endif |
543 | this->__list.__next = NULL; |
544 | |
545 | atomic_or (&this->__lock, FUTEX_OWNER_DIED); |
546 | futex_wake ((unsigned int *) &this->__lock, 1, |
547 | /* XYZ */ FUTEX_SHARED); |
548 | } |
549 | while (robust != (void *) &pd->robust_head); |
550 | } |
551 | #endif |
552 | |
553 | if (!pd->user_stack) |
554 | advise_stack_range (pd->stackblock, pd->stackblock_size, (uintptr_t) pd, |
555 | pd->guardsize); |
556 | |
557 | if (__glibc_unlikely (pd->cancelhandling & SETXID_BITMASK)) |
558 | { |
559 | /* Some other thread might call any of the setXid functions and expect |
560 | us to reply. In this case wait until we did that. */ |
561 | do |
562 | /* XXX This differs from the typical futex_wait_simple pattern in that |
563 | the futex_wait condition (setxid_futex) is different from the |
564 | condition used in the surrounding loop (cancelhandling). We need |
565 | to check and document why this is correct. */ |
566 | futex_wait_simple (&pd->setxid_futex, 0, FUTEX_PRIVATE); |
567 | while (pd->cancelhandling & SETXID_BITMASK); |
568 | |
569 | /* Reset the value so that the stack can be reused. */ |
570 | pd->setxid_futex = 0; |
571 | } |
572 | |
573 | /* If the thread is detached free the TCB. */ |
574 | if (IS_DETACHED (pd)) |
575 | /* Free the TCB. */ |
576 | __free_tcb (pd); |
577 | |
578 | /* We cannot call '_exit' here. '_exit' will terminate the process. |
579 | |
580 | The 'exit' implementation in the kernel will signal when the |
581 | process is really dead since 'clone' got passed the CLONE_CHILD_CLEARTID |
582 | flag. The 'tid' field in the TCB will be set to zero. |
583 | |
584 | The exit code is zero since in case all threads exit by calling |
585 | 'pthread_exit' the exit status must be 0 (zero). */ |
586 | __exit_thread (); |
587 | |
588 | /* NOTREACHED */ |
589 | } |
590 | |
591 | |
592 | /* Return true iff obliged to report TD_CREATE events. */ |
593 | static bool |
594 | report_thread_creation (struct pthread *pd) |
595 | { |
596 | if (__glibc_unlikely (THREAD_GETMEM (THREAD_SELF, report_events))) |
597 | { |
598 | /* The parent thread is supposed to report events. |
599 | Check whether the TD_CREATE event is needed, too. */ |
600 | const size_t idx = __td_eventword (TD_CREATE); |
601 | const uint32_t mask = __td_eventmask (TD_CREATE); |
602 | |
603 | return ((mask & (__nptl_threads_events.event_bits[idx] |
604 | | pd->eventbuf.eventmask.event_bits[idx])) != 0); |
605 | } |
606 | return false; |
607 | } |
608 | |
609 | |
610 | int |
611 | __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr, |
612 | void *(*start_routine) (void *), void *arg) |
613 | { |
614 | STACK_VARIABLES; |
615 | |
616 | /* Avoid a data race in the multi-threaded case. */ |
617 | if (__libc_single_threaded) |
618 | __libc_single_threaded = 0; |
619 | |
620 | const struct pthread_attr *iattr = (struct pthread_attr *) attr; |
621 | union pthread_attr_transparent default_attr; |
622 | bool destroy_default_attr = false; |
623 | bool c11 = (attr == ATTR_C11_THREAD); |
624 | if (iattr == NULL || c11) |
625 | { |
626 | int ret = __pthread_getattr_default_np (&default_attr.external); |
627 | if (ret != 0) |
628 | return ret; |
629 | destroy_default_attr = true; |
630 | iattr = &default_attr.internal; |
631 | } |
632 | |
633 | struct pthread *pd = NULL; |
634 | int err = ALLOCATE_STACK (iattr, &pd); |
635 | int retval = 0; |
636 | |
637 | if (__glibc_unlikely (err != 0)) |
638 | /* Something went wrong. Maybe a parameter of the attributes is |
639 | invalid or we could not allocate memory. Note we have to |
640 | translate error codes. */ |
641 | { |
642 | retval = err == ENOMEM ? EAGAIN : err; |
643 | goto out; |
644 | } |
645 | |
646 | |
647 | /* Initialize the TCB. All initializations with zero should be |
648 | performed in 'get_cached_stack'. This way we avoid doing this if |
649 | the stack freshly allocated with 'mmap'. */ |
650 | |
651 | #if TLS_TCB_AT_TP |
652 | /* Reference to the TCB itself. */ |
653 | pd->header.self = pd; |
654 | |
655 | /* Self-reference for TLS. */ |
656 | pd->header.tcb = pd; |
657 | #endif |
658 | |
659 | /* Store the address of the start routine and the parameter. Since |
660 | we do not start the function directly the stillborn thread will |
661 | get the information from its thread descriptor. */ |
662 | pd->start_routine = start_routine; |
663 | pd->arg = arg; |
664 | pd->c11 = c11; |
665 | |
666 | /* Copy the thread attribute flags. */ |
667 | struct pthread *self = THREAD_SELF; |
668 | pd->flags = ((iattr->flags & ~(ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
669 | | (self->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))); |
670 | |
671 | /* Initialize the field for the ID of the thread which is waiting |
672 | for us. This is a self-reference in case the thread is created |
673 | detached. */ |
674 | pd->joinid = iattr->flags & ATTR_FLAG_DETACHSTATE ? pd : NULL; |
675 | |
676 | /* The debug events are inherited from the parent. */ |
677 | pd->eventbuf = self->eventbuf; |
678 | |
679 | |
680 | /* Copy the parent's scheduling parameters. The flags will say what |
681 | is valid and what is not. */ |
682 | pd->schedpolicy = self->schedpolicy; |
683 | pd->schedparam = self->schedparam; |
684 | |
685 | /* Copy the stack guard canary. */ |
686 | #ifdef THREAD_COPY_STACK_GUARD |
687 | THREAD_COPY_STACK_GUARD (pd); |
688 | #endif |
689 | |
690 | /* Copy the pointer guard value. */ |
691 | #ifdef THREAD_COPY_POINTER_GUARD |
692 | THREAD_COPY_POINTER_GUARD (pd); |
693 | #endif |
694 | |
695 | /* Setup tcbhead. */ |
696 | tls_setup_tcbhead (pd); |
697 | |
698 | /* Verify the sysinfo bits were copied in allocate_stack if needed. */ |
699 | #ifdef NEED_DL_SYSINFO |
700 | CHECK_THREAD_SYSINFO (pd); |
701 | #endif |
702 | |
703 | /* Determine scheduling parameters for the thread. */ |
704 | if (__builtin_expect ((iattr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0, 0) |
705 | && (iattr->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) != 0) |
706 | { |
707 | /* Use the scheduling parameters the user provided. */ |
708 | if (iattr->flags & ATTR_FLAG_POLICY_SET) |
709 | { |
710 | pd->schedpolicy = iattr->schedpolicy; |
711 | pd->flags |= ATTR_FLAG_POLICY_SET; |
712 | } |
713 | if (iattr->flags & ATTR_FLAG_SCHED_SET) |
714 | { |
715 | /* The values were validated in pthread_attr_setschedparam. */ |
716 | pd->schedparam = iattr->schedparam; |
717 | pd->flags |= ATTR_FLAG_SCHED_SET; |
718 | } |
719 | |
720 | if ((pd->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
721 | != (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) |
722 | collect_default_sched (pd); |
723 | } |
724 | |
725 | if (__glibc_unlikely (__nptl_nthreads == 1)) |
726 | _IO_enable_locks (); |
727 | |
728 | /* Pass the descriptor to the caller. */ |
729 | *newthread = (pthread_t) pd; |
730 | |
731 | LIBC_PROBE (pthread_create, 4, newthread, attr, start_routine, arg); |
732 | |
733 | /* One more thread. We cannot have the thread do this itself, since it |
734 | might exist but not have been scheduled yet by the time we've returned |
735 | and need to check the value to behave correctly. We must do it before |
736 | creating the thread, in case it does get scheduled first and then |
737 | might mistakenly think it was the only thread. In the failure case, |
738 | we momentarily store a false value; this doesn't matter because there |
739 | is no kosher thing a signal handler interrupting us right here can do |
740 | that cares whether the thread count is correct. */ |
741 | atomic_increment (&__nptl_nthreads); |
742 | |
743 | /* Our local value of stopped_start and thread_ran can be accessed at |
744 | any time. The PD->stopped_start may only be accessed if we have |
745 | ownership of PD (see CONCURRENCY NOTES above). */ |
746 | bool stopped_start = false; bool thread_ran = false; |
747 | |
748 | /* Block all signals, so that the new thread starts out with |
749 | signals disabled. This avoids race conditions in the thread |
750 | startup. */ |
751 | sigset_t original_sigmask; |
752 | __libc_signal_block_all (&original_sigmask); |
753 | |
754 | if (iattr->extension != NULL && iattr->extension->sigmask_set) |
755 | /* Use the signal mask in the attribute. The internal signals |
756 | have already been filtered by the public |
757 | pthread_attr_setsigmask_np interface. */ |
758 | pd->sigmask = iattr->extension->sigmask; |
759 | else |
760 | { |
761 | /* Conceptually, the new thread needs to inherit the signal mask |
762 | of this thread. Therefore, it needs to restore the saved |
763 | signal mask of this thread, so save it in the startup |
764 | information. */ |
765 | pd->sigmask = original_sigmask; |
766 | |
767 | /* Reset the cancellation signal mask in case this thread is |
768 | running cancellation. */ |
769 | __sigdelset (&pd->sigmask, SIGCANCEL); |
770 | } |
771 | |
772 | /* Start the thread. */ |
773 | if (__glibc_unlikely (report_thread_creation (pd))) |
774 | { |
775 | stopped_start = true; |
776 | |
777 | /* We always create the thread stopped at startup so we can |
778 | notify the debugger. */ |
779 | retval = create_thread (pd, iattr, &stopped_start, |
780 | STACK_VARIABLES_ARGS, &thread_ran); |
781 | if (retval == 0) |
782 | { |
783 | /* We retain ownership of PD until (a) (see CONCURRENCY NOTES |
784 | above). */ |
785 | |
786 | /* Assert stopped_start is true in both our local copy and the |
787 | PD copy. */ |
788 | assert (stopped_start); |
789 | assert (pd->stopped_start); |
790 | |
791 | /* Now fill in the information about the new thread in |
792 | the newly created thread's data structure. We cannot let |
793 | the new thread do this since we don't know whether it was |
794 | already scheduled when we send the event. */ |
795 | pd->eventbuf.eventnum = TD_CREATE; |
796 | pd->eventbuf.eventdata = pd; |
797 | |
798 | /* Enqueue the descriptor. */ |
799 | do |
800 | pd->nextevent = __nptl_last_event; |
801 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, |
802 | pd, pd->nextevent) |
803 | != 0); |
804 | |
805 | /* Now call the function which signals the event. See |
806 | CONCURRENCY NOTES for the nptl_db interface comments. */ |
807 | __nptl_create_event (); |
808 | } |
809 | } |
810 | else |
811 | retval = create_thread (pd, iattr, &stopped_start, |
812 | STACK_VARIABLES_ARGS, &thread_ran); |
813 | |
814 | /* Return to the previous signal mask, after creating the new |
815 | thread. */ |
816 | __libc_signal_restore_set (&original_sigmask); |
817 | |
818 | if (__glibc_unlikely (retval != 0)) |
819 | { |
820 | if (thread_ran) |
821 | /* State (c) or (d) and we may not have PD ownership (see |
822 | CONCURRENCY NOTES above). We can assert that STOPPED_START |
823 | must have been true because thread creation didn't fail, but |
824 | thread attribute setting did. */ |
825 | /* See bug 19511 which explains why doing nothing here is a |
826 | resource leak for a joinable thread. */ |
827 | assert (stopped_start); |
828 | else |
829 | { |
830 | /* State (e) and we have ownership of PD (see CONCURRENCY |
831 | NOTES above). */ |
832 | |
833 | /* Oops, we lied for a second. */ |
834 | atomic_decrement (&__nptl_nthreads); |
835 | |
836 | /* Perhaps a thread wants to change the IDs and is waiting for this |
837 | stillborn thread. */ |
838 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) |
839 | == -2)) |
840 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
841 | |
842 | /* Free the resources. */ |
843 | __deallocate_stack (pd); |
844 | } |
845 | |
846 | /* We have to translate error codes. */ |
847 | if (retval == ENOMEM) |
848 | retval = EAGAIN; |
849 | } |
850 | else |
851 | { |
852 | /* We don't know if we have PD ownership. Once we check the local |
853 | stopped_start we'll know if we're in state (a) or (b) (see |
854 | CONCURRENCY NOTES above). */ |
855 | if (stopped_start) |
856 | /* State (a), we own PD. The thread blocked on this lock either |
857 | because we're doing TD_CREATE event reporting, or for some |
858 | other reason that create_thread chose. Now let it run |
859 | free. */ |
860 | lll_unlock (pd->lock, LLL_PRIVATE); |
861 | |
862 | /* We now have for sure more than one thread. The main thread might |
863 | not yet have the flag set. No need to set the global variable |
864 | again if this is what we use. */ |
865 | THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1); |
866 | } |
867 | |
868 | out: |
869 | if (destroy_default_attr) |
870 | __pthread_attr_destroy (&default_attr.external); |
871 | |
872 | return retval; |
873 | } |
874 | versioned_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1); |
875 | |
876 | |
877 | #if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1) |
878 | int |
879 | __pthread_create_2_0 (pthread_t *newthread, const pthread_attr_t *attr, |
880 | void *(*start_routine) (void *), void *arg) |
881 | { |
882 | /* The ATTR attribute is not really of type `pthread_attr_t *'. It has |
883 | the old size and access to the new members might crash the program. |
884 | We convert the struct now. */ |
885 | struct pthread_attr new_attr; |
886 | |
887 | if (attr != NULL) |
888 | { |
889 | struct pthread_attr *iattr = (struct pthread_attr *) attr; |
890 | size_t ps = __getpagesize (); |
891 | |
892 | /* Copy values from the user-provided attributes. */ |
893 | new_attr.schedparam = iattr->schedparam; |
894 | new_attr.schedpolicy = iattr->schedpolicy; |
895 | new_attr.flags = iattr->flags; |
896 | |
897 | /* Fill in default values for the fields not present in the old |
898 | implementation. */ |
899 | new_attr.guardsize = ps; |
900 | new_attr.stackaddr = NULL; |
901 | new_attr.stacksize = 0; |
902 | new_attr.extension = NULL; |
903 | |
904 | /* We will pass this value on to the real implementation. */ |
905 | attr = (pthread_attr_t *) &new_attr; |
906 | } |
907 | |
908 | return __pthread_create_2_1 (newthread, attr, start_routine, arg); |
909 | } |
910 | compat_symbol (libpthread, __pthread_create_2_0, pthread_create, |
911 | GLIBC_2_0); |
912 | #endif |
913 | |
914 | /* Information for libthread_db. */ |
915 | |
916 | #include "../nptl_db/db_info.c" |
917 | |
918 | /* If pthread_create is present, libgcc_eh.a and libsupc++.a expects some other POSIX thread |
919 | functions to be present as well. */ |
920 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_lock) |
921 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_trylock) |
922 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_unlock) |
923 | |
924 | PTHREAD_STATIC_FN_REQUIRE (__pthread_once) |
925 | PTHREAD_STATIC_FN_REQUIRE (__pthread_cancel) |
926 | |
927 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_create) |
928 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_delete) |
929 | PTHREAD_STATIC_FN_REQUIRE (__pthread_setspecific) |
930 | PTHREAD_STATIC_FN_REQUIRE (__pthread_getspecific) |
931 | |