]> git.ipfire.org Git - thirdparty/glibc.git/blame - nptl/pthread_create.c
Update copyright dates with scripts/update-copyrights.
[thirdparty/glibc.git] / nptl / pthread_create.c
CommitLineData
d614a753 1/* Copyright (C) 2002-2020 Free Software Foundation, Inc.
76a50749
UD
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
59ba27a6 16 License along with the GNU C Library; if not, see
5a82c748 17 <https://www.gnu.org/licenses/>. */
76a50749 18
fd5bdc09 19#include <ctype.h>
76a50749
UD
20#include <errno.h>
21#include <stdbool.h>
22#include <stdlib.h>
23#include <string.h>
e054f494 24#include <stdint.h>
76a50749
UD
25#include "pthreadP.h"
26#include <hp-timing.h>
27#include <ldsodefs.h>
3e4fc359 28#include <atomic.h>
12d7ca07 29#include <libc-internal.h>
0e9d6240 30#include <resolv.h>
f8de5057 31#include <kernel-features.h>
e0db6517 32#include <exit-thread.h>
f214ff74 33#include <default-sched.h>
a2f0363f 34#include <futex-internal.h>
ebff9c5c 35#include <tls-setup.h>
d2e04918 36#include "libioP.h"
76a50749
UD
37
38#include <shlib-compat.h>
39
3a097cc7
RM
40#include <stap-probe.h>
41
76a50749 42
76a50749
UD
43/* Nozero if debugging mode is enabled. */
44int __pthread_debug;
45
46/* Globally enabled events. */
e965d514 47static td_thr_events_t __nptl_threads_events __attribute_used__;
76a50749
UD
48
49/* Pointer to descriptor with the last event. */
e965d514 50static struct pthread *__nptl_last_event __attribute_used__;
76a50749 51
47202270
UD
52/* Number of threads running. */
53unsigned int __nptl_nthreads = 1;
54
76a50749
UD
55
56/* Code to allocate and deallocate a stack. */
76a50749
UD
57#include "allocatestack.c"
58
f8bf15fe
CD
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
fa17b9c7
CD
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:
f8bf15fe
CD
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:
32fed10f
RM
175 START_THREAD_DEFN and START_THREAD_SELF (see below).
176
f8bf15fe
CD
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.
32fed10f
RM
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
f8bf15fe 200 canceled before calling user code (*PD->start_routine). */
32fed10f 201static int create_thread (struct pthread *pd, const struct pthread_attr *attr,
f8bf15fe 202 bool *stopped_start, STACK_VARIABLES_PARMS,
32fed10f
RM
203 bool *thread_ran);
204
8dea90aa 205#include <createthread.c>
76a50749
UD
206
207
76a50749 208struct pthread *
9dd346ff 209__find_in_stack_list (struct pthread *pd)
76a50749
UD
210{
211 list_t *entry;
212 struct pthread *result = NULL;
213
e51deae7 214 lll_lock (stack_cache_lock, LLL_PRIVATE);
76a50749
UD
215
216 list_for_each (entry, &stack_used)
217 {
218 struct pthread *curp;
219
d4f64e1a 220 curp = list_entry (entry, struct pthread, list);
76a50749
UD
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
d4f64e1a 233 curp = list_entry (entry, struct pthread, list);
76a50749
UD
234 if (curp == pd)
235 {
236 result = curp;
237 break;
238 }
239 }
240
e51deae7 241 lll_unlock (stack_cache_lock, LLL_PRIVATE);
76a50749
UD
242
243 return result;
244}
245
246
247/* Deallocate POSIX thread-local-storage. */
3fa21fd8
UD
248void
249attribute_hidden
250__nptl_deallocate_tsd (void)
76a50749 251{
877e51b2
UD
252 struct pthread *self = THREAD_SELF;
253
76a50749
UD
254 /* Maybe no data was ever allocated. This happens often so we have
255 a flag for this. */
877e51b2 256 if (THREAD_GETMEM (self, specific_used))
76a50749
UD
257 {
258 size_t round;
6b4686a5 259 size_t cnt;
76a50749 260
6b4686a5
UD
261 round = 0;
262 do
76a50749 263 {
76a50749
UD
264 size_t idx;
265
c5acd3d7 266 /* So far no new nonzero data entry. */
877e51b2 267 THREAD_SETMEM (self, specific_used, false);
c5acd3d7 268
76a50749 269 for (cnt = idx = 0; cnt < PTHREAD_KEY_1STLEVEL_SIZE; ++cnt)
6b4686a5
UD
270 {
271 struct pthread_key_data *level2;
272
877e51b2 273 level2 = THREAD_GETMEM_NC (self, specific, cnt);
6b4686a5
UD
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
877e51b2
UD
284 if (data != NULL)
285 {
286 /* Always clear the data. */
287 level2[inner].data = NULL;
288
6b4686a5
UD
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. */
877e51b2
UD
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);
6b4686a5
UD
302 }
303 }
304 }
305 else
306 idx += PTHREAD_KEY_1STLEVEL_SIZE;
307 }
877e51b2
UD
308
309 if (THREAD_GETMEM (self, specific_used) == 0)
310 /* No data has been modified. */
311 goto just_free;
76a50749 312 }
877e51b2
UD
313 /* We only repeat the process a fixed number of times. */
314 while (__builtin_expect (++round < PTHREAD_DESTRUCTOR_ITERATIONS, 0));
76a50749 315
877e51b2
UD
316 /* Just clear the memory of the first block for reuse. */
317 memset (&THREAD_SELF->specific_1stblock, '\0',
318 sizeof (self->specific_1stblock));
6b4686a5 319
877e51b2 320 just_free:
6b4686a5
UD
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
877e51b2 326 level2 = THREAD_GETMEM_NC (self, specific, cnt);
6b4686a5
UD
327 if (level2 != NULL)
328 {
329 /* The first block is allocated as part of the thread
330 descriptor. */
331 free (level2);
877e51b2 332 THREAD_SETMEM_NC (self, specific, cnt, NULL);
6b4686a5
UD
333 }
334 }
335
877e51b2 336 THREAD_SETMEM (self, specific_used, false);
76a50749
UD
337 }
338}
339
340
341/* Deallocate a thread's stack after optionally making sure the thread
342 descriptor is still valid. */
343void
344__free_tcb (struct pthread *pd)
345{
346 /* The thread is exiting now. */
ba25bb0f
UD
347 if (__builtin_expect (atomic_bit_test_set (&pd->cancelhandling,
348 TERMINATED_BIT) == 0, 1))
76a50749
UD
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
f17efcb4 356 /* Free TPP data. */
a1ffb40e 357 if (__glibc_unlikely (pd->tpp != NULL))
f17efcb4
UD
358 {
359 struct priority_protection_data *tpp = pd->tpp;
360
361 pd->tpp = NULL;
362 free (tpp);
363 }
364
76a50749
UD
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
32fed10f
RM
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. */
378START_THREAD_DEFN
76a50749 379{
32fed10f 380 struct pthread *pd = START_THREAD_SELF;
76a50749 381
0e9d6240
UD
382 /* Initialize resolver state pointer. */
383 __resp = &pd->res;
384
fd5bdc09
UD
385 /* Initialize pointers to locale data. */
386 __ctype_init ();
387
66f1b8ee 388 /* Allow setxid from now onwards. */
a1ffb40e 389 if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) == -2))
a2f0363f 390 futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE);
66f1b8ee 391
0f6699ea
UD
392#ifdef __NR_set_robust_list
393# ifndef __ASSUME_SET_ROBUST_LIST
394 if (__set_robust_list_avail >= 0)
395# endif
396 {
397 INTERNAL_SYSCALL_DECL (err);
398 /* This call should never fail because the initial call in init.c
399 succeeded. */
400 INTERNAL_SYSCALL (set_robust_list, err, 2, &pd->robust_head,
401 sizeof (struct robust_list_head));
402 }
403#endif
404
b051fc44
UD
405 /* If the parent was running cancellation handlers while creating
406 the thread the new thread inherited the signal mask. Reset the
407 cancellation signal mask. */
a1ffb40e 408 if (__glibc_unlikely (pd->parent_cancelhandling & CANCELING_BITMASK))
b051fc44
UD
409 {
410 INTERNAL_SYSCALL_DECL (err);
411 sigset_t mask;
412 __sigemptyset (&mask);
413 __sigaddset (&mask, SIGCANCEL);
414 (void) INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_UNBLOCK, &mask,
415 NULL, _NSIG / 8);
416 }
417
76a50749
UD
418 /* This is where the try/finally block should be created. For
419 compilers without that support we do use setjmp. */
877e51b2
UD
420 struct pthread_unwind_buf unwind_buf;
421
d6cc1829
L
422 int not_first_call;
423 not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf);
424
425 /* No previous handlers. NB: This must be done after setjmp since the
426 private space in the unwind jump buffer may overlap space used by
427 setjmp to store extra architecture-specific information which is
428 never used by the cancellation-specific __libc_unwind_longjmp.
429
430 The private space is allowed to overlap because the unwinder never
431 has to return through any of the jumped-to call frames, and thus
432 only a minimum amount of saved data need be stored, and for example,
433 need not include the process signal mask information. This is all
434 an optimization to reduce stack usage when pushing cancellation
435 handlers. */
877e51b2
UD
436 unwind_buf.priv.data.prev = NULL;
437 unwind_buf.priv.data.cleanup = NULL;
438
a1ffb40e 439 if (__glibc_likely (! not_first_call))
76a50749 440 {
877e51b2
UD
441 /* Store the new cleanup handler info. */
442 THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf);
443
f8bf15fe
CD
444 /* We are either in (a) or (b), and in either case we either own
445 PD already (2) or are about to own PD (1), and so our only
446 restriction would be that we can't free PD until we know we
447 have ownership (see CONCURRENCY NOTES above). */
a1ffb40e 448 if (__glibc_unlikely (pd->stopped_start))
5f66b766
UD
449 {
450 int oldtype = CANCEL_ASYNC ();
362038b0 451
5f66b766 452 /* Get the lock the parent locked to force synchronization. */
e51deae7 453 lll_lock (pd->lock, LLL_PRIVATE);
f8bf15fe
CD
454
455 /* We have ownership of PD now. */
456
5f66b766 457 /* And give it up right away. */
e51deae7 458 lll_unlock (pd->lock, LLL_PRIVATE);
362038b0 459
5f66b766
UD
460 CANCEL_RESET (oldtype);
461 }
362038b0 462
3a097cc7
RM
463 LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg);
464
76a50749 465 /* Run the code the user provided. */
ce7528f6
AZ
466 void *ret;
467 if (pd->c11)
468 {
469 /* The function pointer of the c11 thread start is cast to an incorrect
470 type on __pthread_create_2_1 call, however it is casted back to correct
471 one so the call behavior is well-defined (it is assumed that pointers
472 to void are able to represent all values of int. */
473 int (*start)(void*) = (int (*) (void*)) pd->start_routine;
474 ret = (void*) (uintptr_t) start (pd->arg);
475 }
476 else
477 ret = pd->start_routine (pd->arg);
478 THREAD_SETMEM (pd, result, ret);
76a50749
UD
479 }
480
ba384f6e 481 /* Call destructors for the thread_local TLS variables. */
e57b0c61
RM
482#ifndef SHARED
483 if (&__call_tls_dtors != NULL)
484#endif
485 __call_tls_dtors ();
ba384f6e 486
6b4686a5 487 /* Run the destructor for the thread-local data. */
3fa21fd8 488 __nptl_deallocate_tsd ();
6b4686a5 489
12d7ca07
RM
490 /* Clean up any state libc stored in thread-local variables. */
491 __libc_thread_freeres ();
76a50749 492
47202270
UD
493 /* If this is the last thread we terminate the process now. We
494 do not notify the debugger, it might just irritate it if there
495 is no thread left. */
a1ffb40e 496 if (__glibc_unlikely (atomic_decrement_and_test (&__nptl_nthreads)))
47202270
UD
497 /* This was the last thread. */
498 exit (0);
499
76a50749 500 /* Report the death of the thread if this is wanted. */
a1ffb40e 501 if (__glibc_unlikely (pd->report_events))
76a50749
UD
502 {
503 /* See whether TD_DEATH is in any of the mask. */
504 const int idx = __td_eventword (TD_DEATH);
505 const uint32_t mask = __td_eventmask (TD_DEATH);
506
507 if ((mask & (__nptl_threads_events.event_bits[idx]
508 | pd->eventbuf.eventmask.event_bits[idx])) != 0)
509 {
510 /* Yep, we have to signal the death. Add the descriptor to
511 the list but only if it is not already on it. */
512 if (pd->nextevent == NULL)
513 {
514 pd->eventbuf.eventnum = TD_DEATH;
515 pd->eventbuf.eventdata = pd;
516
517 do
518 pd->nextevent = __nptl_last_event;
5a3ab2fc
UD
519 while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event,
520 pd, pd->nextevent));
76a50749
UD
521 }
522
f8bf15fe
CD
523 /* Now call the function which signals the event. See
524 CONCURRENCY NOTES for the nptl_db interface comments. */
76a50749
UD
525 __nptl_death_event ();
526 }
527 }
528
6461e577
RM
529 /* The thread is exiting now. Don't set this bit until after we've hit
530 the event-reporting breakpoint, so that td_thr_get_info on us while at
531 the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE. */
532 atomic_bit_set (&pd->cancelhandling, EXITING_BIT);
76a50749 533
0f6699ea 534#ifndef __ASSUME_SET_ROBUST_LIST
1bcfb5a5 535 /* If this thread has any robust mutexes locked, handle them now. */
06be6368 536# if __PTHREAD_MUTEX_HAVE_PREV
0f6699ea
UD
537 void *robust = pd->robust_head.list;
538# else
b007ce7c 539 __pthread_slist_t *robust = pd->robust_list.__next;
0f6699ea 540# endif
df47504c
UD
541 /* We let the kernel do the notification if it is able to do so.
542 If we have to do it here there for sure are no PI mutexes involved
543 since the kernel support for them is even more recent. */
0f6699ea 544 if (__set_robust_list_avail < 0
df47504c 545 && __builtin_expect (robust != (void *) &pd->robust_head, 0))
1bcfb5a5
UD
546 {
547 do
548 {
b007ce7c 549 struct __pthread_mutex_s *this = (struct __pthread_mutex_s *)
0f6699ea
UD
550 ((char *) robust - offsetof (struct __pthread_mutex_s,
551 __list.__next));
552 robust = *((void **) robust);
d804f5df 553
06be6368 554# if __PTHREAD_MUTEX_HAVE_PREV
b007ce7c 555 this->__list.__prev = NULL;
0f6699ea
UD
556# endif
557 this->__list.__next = NULL;
1bcfb5a5 558
c0c6bac9 559 atomic_or (&this->__lock, FUTEX_OWNER_DIED);
a2f0363f
TR
560 futex_wake ((unsigned int *) &this->__lock, 1,
561 /* XYZ */ FUTEX_SHARED);
1bcfb5a5 562 }
df47504c 563 while (robust != (void *) &pd->robust_head);
1bcfb5a5 564 }
0f6699ea 565#endif
1bcfb5a5 566
01b87c65
AZ
567 advise_stack_range (pd->stackblock, pd->stackblock_size, (uintptr_t) pd,
568 pd->guardsize);
b42a214c 569
76a50749
UD
570 /* If the thread is detached free the TCB. */
571 if (IS_DETACHED (pd))
572 /* Free the TCB. */
573 __free_tcb (pd);
a1ffb40e 574 else if (__glibc_unlikely (pd->cancelhandling & SETXID_BITMASK))
dff9a7a1
UD
575 {
576 /* Some other thread might call any of the setXid functions and expect
577 us to reply. In this case wait until we did that. */
578 do
a2f0363f
TR
579 /* XXX This differs from the typical futex_wait_simple pattern in that
580 the futex_wait condition (setxid_futex) is different from the
581 condition used in the surrounding loop (cancelhandling). We need
582 to check and document why this is correct. */
583 futex_wait_simple (&pd->setxid_futex, 0, FUTEX_PRIVATE);
dff9a7a1
UD
584 while (pd->cancelhandling & SETXID_BITMASK);
585
586 /* Reset the value so that the stack can be reused. */
587 pd->setxid_futex = 0;
588 }
76a50749
UD
589
590 /* We cannot call '_exit' here. '_exit' will terminate the process.
591
592 The 'exit' implementation in the kernel will signal when the
adcdc775 593 process is really dead since 'clone' got passed the CLONE_CHILD_CLEARTID
76a50749
UD
594 flag. The 'tid' field in the TCB will be set to zero.
595
596 The exit code is zero since in case all threads exit by calling
597 'pthread_exit' the exit status must be 0 (zero). */
e0db6517 598 __exit_thread ();
76a50749
UD
599
600 /* NOTREACHED */
32fed10f
RM
601}
602
603
604/* Return true iff obliged to report TD_CREATE events. */
605static bool
606report_thread_creation (struct pthread *pd)
607{
608 if (__glibc_unlikely (THREAD_GETMEM (THREAD_SELF, report_events)))
609 {
610 /* The parent thread is supposed to report events.
611 Check whether the TD_CREATE event is needed, too. */
612 const size_t idx = __td_eventword (TD_CREATE);
613 const uint32_t mask = __td_eventmask (TD_CREATE);
614
615 return ((mask & (__nptl_threads_events.event_bits[idx]
616 | pd->eventbuf.eventmask.event_bits[idx])) != 0);
617 }
618 return false;
76a50749
UD
619}
620
621
76a50749 622int
80d9be81
JM
623__pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr,
624 void *(*start_routine) (void *), void *arg)
76a50749
UD
625{
626 STACK_VARIABLES;
76a50749 627
1e6da2b0 628 const struct pthread_attr *iattr = (struct pthread_attr *) attr;
61dd6208
SP
629 struct pthread_attr default_attr;
630 bool free_cpuset = false;
ce7528f6
AZ
631 bool c11 = (attr == ATTR_C11_THREAD);
632 if (iattr == NULL || c11)
61dd6208
SP
633 {
634 lll_lock (__default_pthread_attr_lock, LLL_PRIVATE);
635 default_attr = __default_pthread_attr;
636 size_t cpusetsize = default_attr.cpusetsize;
637 if (cpusetsize > 0)
638 {
639 cpu_set_t *cpuset;
640 if (__glibc_likely (__libc_use_alloca (cpusetsize)))
641 cpuset = __alloca (cpusetsize);
642 else
643 {
644 cpuset = malloc (cpusetsize);
645 if (cpuset == NULL)
646 {
647 lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
648 return ENOMEM;
649 }
650 free_cpuset = true;
651 }
652 memcpy (cpuset, default_attr.cpuset, cpusetsize);
653 default_attr.cpuset = cpuset;
654 }
655 lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
656 iattr = &default_attr;
657 }
76a50749 658
dff9a7a1 659 struct pthread *pd = NULL;
1e6da2b0 660 int err = ALLOCATE_STACK (iattr, &pd);
61dd6208
SP
661 int retval = 0;
662
a1ffb40e 663 if (__glibc_unlikely (err != 0))
76a50749 664 /* Something went wrong. Maybe a parameter of the attributes is
e988dba9
JL
665 invalid or we could not allocate memory. Note we have to
666 translate error codes. */
61dd6208
SP
667 {
668 retval = err == ENOMEM ? EAGAIN : err;
669 goto out;
670 }
76a50749
UD
671
672
673 /* Initialize the TCB. All initializations with zero should be
674 performed in 'get_cached_stack'. This way we avoid doing this if
675 the stack freshly allocated with 'mmap'. */
676
d7329d4b 677#if TLS_TCB_AT_TP
76a50749 678 /* Reference to the TCB itself. */
55c11fbd 679 pd->header.self = pd;
76a50749 680
d4f64e1a 681 /* Self-reference for TLS. */
55c11fbd 682 pd->header.tcb = pd;
76a50749
UD
683#endif
684
685 /* Store the address of the start routine and the parameter. Since
686 we do not start the function directly the stillborn thread will
687 get the information from its thread descriptor. */
688 pd->start_routine = start_routine;
689 pd->arg = arg;
ce7528f6 690 pd->c11 = c11;
76a50749
UD
691
692 /* Copy the thread attribute flags. */
14ffbc83
UD
693 struct pthread *self = THREAD_SELF;
694 pd->flags = ((iattr->flags & ~(ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
695 | (self->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)));
76a50749
UD
696
697 /* Initialize the field for the ID of the thread which is waiting
698 for us. This is a self-reference in case the thread is created
699 detached. */
700 pd->joinid = iattr->flags & ATTR_FLAG_DETACHSTATE ? pd : NULL;
701
702 /* The debug events are inherited from the parent. */
14ffbc83
UD
703 pd->eventbuf = self->eventbuf;
704
76a50749 705
14ffbc83
UD
706 /* Copy the parent's scheduling parameters. The flags will say what
707 is valid and what is not. */
708 pd->schedpolicy = self->schedpolicy;
709 pd->schedparam = self->schedparam;
76a50749 710
35f1e827
UD
711 /* Copy the stack guard canary. */
712#ifdef THREAD_COPY_STACK_GUARD
713 THREAD_COPY_STACK_GUARD (pd);
714#endif
715
827b7087
UD
716 /* Copy the pointer guard value. */
717#ifdef THREAD_COPY_POINTER_GUARD
718 THREAD_COPY_POINTER_GUARD (pd);
719#endif
720
ebff9c5c
L
721 /* Setup tcbhead. */
722 tls_setup_tcbhead (pd);
723
32fed10f
RM
724 /* Verify the sysinfo bits were copied in allocate_stack if needed. */
725#ifdef NEED_DL_SYSINFO
726 CHECK_THREAD_SYSINFO (pd);
727#endif
728
729 /* Inform start_thread (above) about cancellation state that might
730 translate into inherited signal state. */
731 pd->parent_cancelhandling = THREAD_GETMEM (THREAD_SELF, cancelhandling);
732
14ffbc83 733 /* Determine scheduling parameters for the thread. */
61dd6208 734 if (__builtin_expect ((iattr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0, 0)
14ffbc83 735 && (iattr->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) != 0)
76a50749 736 {
14ffbc83
UD
737 /* Use the scheduling parameters the user provided. */
738 if (iattr->flags & ATTR_FLAG_POLICY_SET)
33cd1f74
RM
739 {
740 pd->schedpolicy = iattr->schedpolicy;
741 pd->flags |= ATTR_FLAG_POLICY_SET;
742 }
14ffbc83 743 if (iattr->flags & ATTR_FLAG_SCHED_SET)
33cd1f74
RM
744 {
745 /* The values were validated in pthread_attr_setschedparam. */
746 pd->schedparam = iattr->schedparam;
747 pd->flags |= ATTR_FLAG_SCHED_SET;
748 }
f214ff74
RM
749
750 if ((pd->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
751 != (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
752 collect_default_sched (pd);
76a50749
UD
753 }
754
d2e04918
SN
755 if (__glibc_unlikely (__nptl_nthreads == 1))
756 _IO_enable_locks ();
757
76a50749
UD
758 /* Pass the descriptor to the caller. */
759 *newthread = (pthread_t) pd;
760
5acf7263
RM
761 LIBC_PROBE (pthread_create, 4, newthread, attr, start_routine, arg);
762
32fed10f
RM
763 /* One more thread. We cannot have the thread do this itself, since it
764 might exist but not have been scheduled yet by the time we've returned
765 and need to check the value to behave correctly. We must do it before
766 creating the thread, in case it does get scheduled first and then
767 might mistakenly think it was the only thread. In the failure case,
768 we momentarily store a false value; this doesn't matter because there
769 is no kosher thing a signal handler interrupting us right here can do
770 that cares whether the thread count is correct. */
771 atomic_increment (&__nptl_nthreads);
772
f8bf15fe
CD
773 /* Our local value of stopped_start and thread_ran can be accessed at
774 any time. The PD->stopped_start may only be accessed if we have
775 ownership of PD (see CONCURRENCY NOTES above). */
776 bool stopped_start = false; bool thread_ran = false;
32fed10f 777
76a50749 778 /* Start the thread. */
32fed10f
RM
779 if (__glibc_unlikely (report_thread_creation (pd)))
780 {
f8bf15fe
CD
781 stopped_start = true;
782
783 /* We always create the thread stopped at startup so we can
784 notify the debugger. */
785 retval = create_thread (pd, iattr, &stopped_start,
786 STACK_VARIABLES_ARGS, &thread_ran);
32fed10f
RM
787 if (retval == 0)
788 {
f8bf15fe
CD
789 /* We retain ownership of PD until (a) (see CONCURRENCY NOTES
790 above). */
791
792 /* Assert stopped_start is true in both our local copy and the
793 PD copy. */
794 assert (stopped_start);
32fed10f
RM
795 assert (pd->stopped_start);
796
797 /* Now fill in the information about the new thread in
798 the newly created thread's data structure. We cannot let
799 the new thread do this since we don't know whether it was
800 already scheduled when we send the event. */
801 pd->eventbuf.eventnum = TD_CREATE;
802 pd->eventbuf.eventdata = pd;
803
804 /* Enqueue the descriptor. */
805 do
806 pd->nextevent = __nptl_last_event;
807 while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event,
808 pd, pd->nextevent)
809 != 0);
810
f8bf15fe
CD
811 /* Now call the function which signals the event. See
812 CONCURRENCY NOTES for the nptl_db interface comments. */
32fed10f
RM
813 __nptl_create_event ();
814 }
815 }
816 else
f8bf15fe
CD
817 retval = create_thread (pd, iattr, &stopped_start,
818 STACK_VARIABLES_ARGS, &thread_ran);
32fed10f
RM
819
820 if (__glibc_unlikely (retval != 0))
821 {
32fed10f 822 if (thread_ran)
f8bf15fe
CD
823 /* State (c) or (d) and we may not have PD ownership (see
824 CONCURRENCY NOTES above). We can assert that STOPPED_START
825 must have been true because thread creation didn't fail, but
826 thread attribute setting did. */
827 /* See bug 19511 which explains why doing nothing here is a
828 resource leak for a joinable thread. */
829 assert (stopped_start);
32fed10f
RM
830 else
831 {
f8bf15fe
CD
832 /* State (e) and we have ownership of PD (see CONCURRENCY
833 NOTES above). */
834
32fed10f
RM
835 /* Oops, we lied for a second. */
836 atomic_decrement (&__nptl_nthreads);
837
838 /* Perhaps a thread wants to change the IDs and is waiting for this
839 stillborn thread. */
840 if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0)
841 == -2))
a2f0363f 842 futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE);
32fed10f
RM
843
844 /* Free the resources. */
845 __deallocate_stack (pd);
846 }
847
848 /* We have to translate error codes. */
849 if (retval == ENOMEM)
850 retval = EAGAIN;
851 }
852 else
853 {
f8bf15fe
CD
854 /* We don't know if we have PD ownership. Once we check the local
855 stopped_start we'll know if we're in state (a) or (b) (see
856 CONCURRENCY NOTES above). */
857 if (stopped_start)
858 /* State (a), we own PD. The thread blocked on this lock either
859 because we're doing TD_CREATE event reporting, or for some
860 other reason that create_thread chose. Now let it run
861 free. */
32fed10f
RM
862 lll_unlock (pd->lock, LLL_PRIVATE);
863
864 /* We now have for sure more than one thread. The main thread might
865 not yet have the flag set. No need to set the global variable
866 again if this is what we use. */
867 THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1);
868 }
61dd6208
SP
869
870 out:
871 if (__glibc_unlikely (free_cpuset))
872 free (default_attr.cpuset);
873
874 return retval;
76a50749
UD
875}
876versioned_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1);
877
878
879#if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1)
880int
80d9be81
JM
881__pthread_create_2_0 (pthread_t *newthread, const pthread_attr_t *attr,
882 void *(*start_routine) (void *), void *arg)
76a50749
UD
883{
884 /* The ATTR attribute is not really of type `pthread_attr_t *'. It has
885 the old size and access to the new members might crash the program.
886 We convert the struct now. */
887 struct pthread_attr new_attr;
888
889 if (attr != NULL)
890 {
891 struct pthread_attr *iattr = (struct pthread_attr *) attr;
892 size_t ps = __getpagesize ();
893
894 /* Copy values from the user-provided attributes. */
895 new_attr.schedparam = iattr->schedparam;
896 new_attr.schedpolicy = iattr->schedpolicy;
897 new_attr.flags = iattr->flags;
898
899 /* Fill in default values for the fields not present in the old
900 implementation. */
901 new_attr.guardsize = ps;
902 new_attr.stackaddr = NULL;
903 new_attr.stacksize = 0;
ca85ede0 904 new_attr.cpuset = NULL;
76a50749
UD
905
906 /* We will pass this value on to the real implementation. */
907 attr = (pthread_attr_t *) &new_attr;
908 }
909
910 return __pthread_create_2_1 (newthread, attr, start_routine, arg);
911}
912compat_symbol (libpthread, __pthread_create_2_0, pthread_create,
913 GLIBC_2_0);
914#endif
7f08f55a
RM
915\f
916/* Information for libthread_db. */
917
918#include "../nptl_db/db_info.c"
b639d0c9
UD
919\f
920/* If pthread_create is present, libgcc_eh.a and libsupc++.a expects some other POSIX thread
921 functions to be present as well. */
fa872e1b
AZ
922PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_lock)
923PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_trylock)
924PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_unlock)
b639d0c9 925
fa872e1b
AZ
926PTHREAD_STATIC_FN_REQUIRE (__pthread_once)
927PTHREAD_STATIC_FN_REQUIRE (__pthread_cancel)
b639d0c9 928
fa872e1b
AZ
929PTHREAD_STATIC_FN_REQUIRE (__pthread_key_create)
930PTHREAD_STATIC_FN_REQUIRE (__pthread_key_delete)
931PTHREAD_STATIC_FN_REQUIRE (__pthread_setspecific)
932PTHREAD_STATIC_FN_REQUIRE (__pthread_getspecific)