1/* Copyright (C) 2002-2023 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
8
9 The GNU C Library is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 Lesser General Public License for more details.
13
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library; if not, see
16 <https://www.gnu.org/licenses/>. */
17
18#include <ctype.h>
19#include <errno.h>
20#include <stdbool.h>
21#include <stdlib.h>
22#include <string.h>
23#include <stdint.h>
24#include "pthreadP.h"
25#include <hp-timing.h>
26#include <ldsodefs.h>
27#include <atomic.h>
28#include <libc-diag.h>
29#include <libc-internal.h>
30#include <resolv.h>
31#include <kernel-features.h>
32#include <default-sched.h>
33#include <futex-internal.h>
34#include <tls-setup.h>
35#include <rseq-internal.h>
36#include "libioP.h"
37#include <sys/single_threaded.h>
38#include <version.h>
39#include <clone_internal.h>
40#include <futex-internal.h>
41
42#include <shlib-compat.h>
43
44#include <stap-probe.h>
45
46
47/* Globally enabled events. */
48extern td_thr_events_t __nptl_threads_events;
49libc_hidden_proto (__nptl_threads_events)
50td_thr_events_t __nptl_threads_events;
51libc_hidden_data_def (__nptl_threads_events)
52
53/* Pointer to descriptor with the last event. */
54extern struct pthread *__nptl_last_event;
55libc_hidden_proto (__nptl_last_event)
56struct pthread *__nptl_last_event;
57libc_hidden_data_def (__nptl_last_event)
58
59#ifdef SHARED
60/* This variable is used to access _rtld_global from libthread_db. If
61 GDB loads libpthread before ld.so, it is not possible to resolve
62 _rtld_global directly during libpthread initialization. */
63struct rtld_global *__nptl_rtld_global = &_rtld_global;
64#endif
65
66/* Version of the library, used in libthread_db to detect mismatches. */
67const char __nptl_version[] = VERSION;
68
69/* This performs the initialization necessary when going from
70 single-threaded to multi-threaded mode for the first time. */
71static void
72late_init (void)
73{
74 struct sigaction sa;
75 __sigemptyset (&sa.sa_mask);
76
77 /* Install the handle to change the threads' uid/gid. Use
78 SA_ONSTACK because the signal may be sent to threads that are
79 running with custom stacks. (This is less likely for
80 SIGCANCEL.) */
81 sa.sa_sigaction = __nptl_setxid_sighandler;
82 sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTART;
83 (void) __libc_sigaction (SIGSETXID, &sa, NULL);
84
85 /* The parent process might have left the signals blocked. Just in
86 case, unblock it. We reuse the signal mask in the sigaction
87 structure. It is already cleared. */
88 __sigaddset (&sa.sa_mask, SIGCANCEL);
89 __sigaddset (&sa.sa_mask, SIGSETXID);
90 INTERNAL_SYSCALL_CALL (rt_sigprocmask, SIG_UNBLOCK, &sa.sa_mask,
91 NULL, __NSIG_BYTES);
92}
93
94/* Code to allocate and deallocate a stack. */
95#include "allocatestack.c"
96
97/* CONCURRENCY NOTES:
98
99 Understanding who is the owner of the 'struct pthread' or 'PD'
100 (refers to the value of the 'struct pthread *pd' function argument)
101 is critically important in determining exactly which operations are
102 allowed and which are not and when, particularly when it comes to the
103 implementation of pthread_create, pthread_join, pthread_detach, and
104 other functions which all operate on PD.
105
106 The owner of PD is responsible for freeing the final resources
107 associated with PD, and may examine the memory underlying PD at any
108 point in time until it frees it back to the OS or to reuse by the
109 runtime.
110
111 The thread which calls pthread_create is called the creating thread.
112 The creating thread begins as the owner of PD.
113
114 During startup the new thread may examine PD in coordination with the
115 owner thread (which may be itself).
116
117 The four cases of ownership transfer are:
118
119 (1) Ownership of PD is released to the process (all threads may use it)
120 after the new thread starts in a joinable state
121 i.e. pthread_create returns a usable pthread_t.
122
123 (2) Ownership of PD is released to the new thread starting in a detached
124 state.
125
126 (3) Ownership of PD is dynamically released to a running thread via
127 pthread_detach.
128
129 (4) Ownership of PD is acquired by the thread which calls pthread_join.
130
131 Implementation notes:
132
133 The PD->stopped_start and thread_ran variables are used to determine
134 exactly which of the four ownership states we are in and therefore
135 what actions can be taken. For example after (2) we cannot read or
136 write from PD anymore since the thread may no longer exist and the
137 memory may be unmapped.
138
139 It is important to point out that PD->lock is being used both
140 similar to a one-shot semaphore and subsequently as a mutex. The
141 lock is taken in the parent to force the child to wait, and then the
142 child releases the lock. However, this semaphore-like effect is used
143 only for synchronizing the parent and child. After startup the lock
144 is used like a mutex to create a critical section during which a
145 single owner modifies the thread parameters.
146
147 The most complicated cases happen during thread startup:
148
149 (a) If the created thread is in a detached (PTHREAD_CREATE_DETACHED),
150 or joinable (default PTHREAD_CREATE_JOINABLE) state and
151 STOPPED_START is true, then the creating thread has ownership of
152 PD until the PD->lock is released by pthread_create. If any
153 errors occur we are in states (c) or (d) below.
154
155 (b) If the created thread is in a detached state
156 (PTHREAD_CREATED_DETACHED), and STOPPED_START is false, then the
157 creating thread has ownership of PD until it invokes the OS
158 kernel's thread creation routine. If this routine returns
159 without error, then the created thread owns PD; otherwise, see
160 (c) or (d) below.
161
162 (c) If either a joinable or detached thread setup failed and THREAD_RAN
163 is true, then the creating thread releases ownership to the new thread,
164 the created thread sees the failed setup through PD->setup_failed
165 member, releases the PD ownership, and exits. The creating thread will
166 be responsible for cleanup the allocated resources. The THREAD_RAN is
167 local to creating thread and indicate whether thread creation or setup
168 has failed.
169
170 (d) If the thread creation failed and THREAD_RAN is false (meaning
171 ARCH_CLONE has failed), then the creating thread retains ownership
172 of PD and must cleanup he allocated resource. No waiting for the new
173 thread is required because it never started.
174
175 The nptl_db interface:
176
177 The interface with nptl_db requires that we enqueue PD into a linked
178 list and then call a function which the debugger will trap. The PD
179 will then be dequeued and control returned to the thread. The caller
180 at the time must have ownership of PD and such ownership remains
181 after control returns to thread. The enqueued PD is removed from the
182 linked list by the nptl_db callback td_thr_event_getmsg. The debugger
183 must ensure that the thread does not resume execution, otherwise
184 ownership of PD may be lost and examining PD will not be possible.
185
186 Note that the GNU Debugger as of (December 10th 2015) commit
187 c2c2a31fdb228d41ce3db62b268efea04bd39c18 no longer uses
188 td_thr_event_getmsg and several other related nptl_db interfaces. The
189 principal reason for this is that nptl_db does not support non-stop
190 mode where other threads can run concurrently and modify runtime
191 structures currently in use by the debugger and the nptl_db
192 interface.
193
194 Axioms:
195
196 * The create_thread function can never set stopped_start to false.
197 * The created thread can read stopped_start but never write to it.
198 * The variable thread_ran is set some time after the OS thread
199 creation routine returns, how much time after the thread is created
200 is unspecified, but it should be as quickly as possible.
201
202*/
203
204/* CREATE THREAD NOTES:
205
206 create_thread must initialize PD->stopped_start. It should be true
207 if the STOPPED_START parameter is true, or if create_thread needs the
208 new thread to synchronize at startup for some other implementation
209 reason. If STOPPED_START will be true, then create_thread is obliged
210 to lock PD->lock before starting the thread. Then pthread_create
211 unlocks PD->lock which synchronizes-with create_thread in the
212 child thread which does an acquire/release of PD->lock as the last
213 action before calling the user entry point. The goal of all of this
214 is to ensure that the required initial thread attributes are applied
215 (by the creating thread) before the new thread runs user code. Note
216 that the the functions pthread_getschedparam, pthread_setschedparam,
217 pthread_setschedprio, __pthread_tpp_change_priority, and
218 __pthread_current_priority reuse the same lock, PD->lock, for a
219 similar purpose e.g. synchronizing the setting of similar thread
220 attributes. These functions are never called before the thread is
221 created, so don't participate in startup syncronization, but given
222 that the lock is present already and in the unlocked state, reusing
223 it saves space.
224
225 The return value is zero for success or an errno code for failure.
226 If the return value is ENOMEM, that will be translated to EAGAIN,
227 so create_thread need not do that. On failure, *THREAD_RAN should
228 be set to true iff the thread actually started up but before calling
229 the user code (*PD->start_routine). */
230
231static int _Noreturn start_thread (void *arg);
232
233static int create_thread (struct pthread *pd, const struct pthread_attr *attr,
234 bool *stopped_start, void *stackaddr,
235 size_t stacksize, bool *thread_ran)
236{
237 /* Determine whether the newly created threads has to be started
238 stopped since we have to set the scheduling parameters or set the
239 affinity. */
240 bool need_setaffinity = (attr != NULL && attr->extension != NULL
241 && attr->extension->cpuset != 0);
242 if (attr != NULL
243 && (__glibc_unlikely (need_setaffinity)
244 || __glibc_unlikely ((attr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0)))
245 *stopped_start = true;
246
247 pd->stopped_start = *stopped_start;
248 if (__glibc_unlikely (*stopped_start))
249 lll_lock (pd->lock, LLL_PRIVATE);
250
251 /* We rely heavily on various flags the CLONE function understands:
252
253 CLONE_VM, CLONE_FS, CLONE_FILES
254 These flags select semantics with shared address space and
255 file descriptors according to what POSIX requires.
256
257 CLONE_SIGHAND, CLONE_THREAD
258 This flag selects the POSIX signal semantics and various
259 other kinds of sharing (itimers, POSIX timers, etc.).
260
261 CLONE_SETTLS
262 The sixth parameter to CLONE determines the TLS area for the
263 new thread.
264
265 CLONE_PARENT_SETTID
266 The kernels writes the thread ID of the newly created thread
267 into the location pointed to by the fifth parameters to CLONE.
268
269 Note that it would be semantically equivalent to use
270 CLONE_CHILD_SETTID but it is be more expensive in the kernel.
271
272 CLONE_CHILD_CLEARTID
273 The kernels clears the thread ID of a thread that has called
274 sys_exit() in the location pointed to by the seventh parameter
275 to CLONE.
276
277 The termination signal is chosen to be zero which means no signal
278 is sent. */
279 const int clone_flags = (CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SYSVSEM
280 | CLONE_SIGHAND | CLONE_THREAD
281 | CLONE_SETTLS | CLONE_PARENT_SETTID
282 | CLONE_CHILD_CLEARTID
283 | 0);
284
285 TLS_DEFINE_INIT_TP (tp, pd);
286
287 struct clone_args args =
288 {
289 .flags = clone_flags,
290 .pidfd = (uintptr_t) &pd->tid,
291 .parent_tid = (uintptr_t) &pd->tid,
292 .child_tid = (uintptr_t) &pd->tid,
293 .stack = (uintptr_t) stackaddr,
294 .stack_size = stacksize,
295 .tls = (uintptr_t) tp,
296 };
297 int ret = __clone_internal (&args, &start_thread, pd);
298 if (__glibc_unlikely (ret == -1))
299 return errno;
300
301 /* It's started now, so if we fail below, we'll have to let it clean itself
302 up. */
303 *thread_ran = true;
304
305 /* Now we have the possibility to set scheduling parameters etc. */
306 if (attr != NULL)
307 {
308 /* Set the affinity mask if necessary. */
309 if (need_setaffinity)
310 {
311 assert (*stopped_start);
312
313 int res = INTERNAL_SYSCALL_CALL (sched_setaffinity, pd->tid,
314 attr->extension->cpusetsize,
315 attr->extension->cpuset);
316 if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (res)))
317 return INTERNAL_SYSCALL_ERRNO (res);
318 }
319
320 /* Set the scheduling parameters. */
321 if ((attr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0)
322 {
323 assert (*stopped_start);
324
325 int res = INTERNAL_SYSCALL_CALL (sched_setscheduler, pd->tid,
326 pd->schedpolicy, &pd->schedparam);
327 if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (res)))
328 return INTERNAL_SYSCALL_ERRNO (res);
329 }
330 }
331
332 return 0;
333}
334
335/* Local function to start thread and handle cleanup. */
336static int _Noreturn
337start_thread (void *arg)
338{
339 struct pthread *pd = arg;
340
341 /* We are either in (a) or (b), and in either case we either own PD already
342 (2) or are about to own PD (1), and so our only restriction would be that
343 we can't free PD until we know we have ownership (see CONCURRENCY NOTES
344 above). */
345 if (pd->stopped_start)
346 {
347 bool setup_failed = false;
348
349 /* Get the lock the parent locked to force synchronization. */
350 lll_lock (pd->lock, LLL_PRIVATE);
351
352 /* We have ownership of PD now, for detached threads with setup failure
353 we set it as joinable so the creating thread could synchronous join
354 and free any resource prior return to the pthread_create caller. */
355 setup_failed = pd->setup_failed == 1;
356 if (setup_failed)
357 pd->joinid = NULL;
358
359 /* And give it up right away. */
360 lll_unlock (pd->lock, LLL_PRIVATE);
361
362 if (setup_failed)
363 goto out;
364 }
365
366 /* Initialize resolver state pointer. */
367 __resp = &pd->res;
368
369 /* Initialize pointers to locale data. */
370 __ctype_init ();
371
372 /* Register rseq TLS to the kernel. */
373 {
374 bool do_rseq = THREAD_GETMEM (pd, flags) & ATTR_FLAG_DO_RSEQ;
375 if (!rseq_register_current_thread (pd, do_rseq) && do_rseq)
376 __libc_fatal ("Fatal glibc error: rseq registration failed\n");
377 }
378
379#ifndef __ASSUME_SET_ROBUST_LIST
380 if (__nptl_set_robust_list_avail)
381#endif
382 {
383 /* This call should never fail because the initial call in init.c
384 succeeded. */
385 INTERNAL_SYSCALL_CALL (set_robust_list, &pd->robust_head,
386 sizeof (struct robust_list_head));
387 }
388
389 /* This is where the try/finally block should be created. For
390 compilers without that support we do use setjmp. */
391 struct pthread_unwind_buf unwind_buf;
392
393 int not_first_call;
394 DIAG_PUSH_NEEDS_COMMENT;
395#if __GNUC_PREREQ (7, 0)
396 /* This call results in a -Wstringop-overflow warning because struct
397 pthread_unwind_buf is smaller than jmp_buf. setjmp and longjmp
398 do not use anything beyond the common prefix (they never access
399 the saved signal mask), so that is a false positive. */
400 DIAG_IGNORE_NEEDS_COMMENT (11, "-Wstringop-overflow=");
401#endif
402 not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf);
403 DIAG_POP_NEEDS_COMMENT;
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 /* Allow setxid from now onwards. */
420 if (__glibc_unlikely (atomic_exchange_acquire (&pd->setxid_futex, 0) == -2))
421 futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE);
422
423 if (__glibc_likely (! not_first_call))
424 {
425 /* Store the new cleanup handler info. */
426 THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf);
427
428 internal_signal_restore_set (&pd->sigmask);
429
430 LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg);
431
432 /* Run the code the user provided. */
433 void *ret;
434 if (pd->c11)
435 {
436 /* The function pointer of the c11 thread start is cast to an incorrect
437 type on __pthread_create_2_1 call, however it is casted back to correct
438 one so the call behavior is well-defined (it is assumed that pointers
439 to void are able to represent all values of int. */
440 int (*start)(void*) = (int (*) (void*)) pd->start_routine;
441 ret = (void*) (uintptr_t) start (pd->arg);
442 }
443 else
444 ret = pd->start_routine (pd->arg);
445 THREAD_SETMEM (pd, result, ret);
446 }
447
448 /* Call destructors for the thread_local TLS variables. */
449#ifndef SHARED
450 if (&__call_tls_dtors != NULL)
451#endif
452 __call_tls_dtors ();
453
454 /* Run the destructor for the thread-local data. */
455 __nptl_deallocate_tsd ();
456
457 /* Clean up any state libc stored in thread-local variables. */
458 __libc_thread_freeres ();
459
460 /* Report the death of the thread if this is wanted. */
461 if (__glibc_unlikely (pd->report_events))
462 {
463 /* See whether TD_DEATH is in any of the mask. */
464 const int idx = __td_eventword (TD_DEATH);
465 const uint32_t mask = __td_eventmask (TD_DEATH);
466
467 if ((mask & (__nptl_threads_events.event_bits[idx]
468 | pd->eventbuf.eventmask.event_bits[idx])) != 0)
469 {
470 /* Yep, we have to signal the death. Add the descriptor to
471 the list but only if it is not already on it. */
472 if (pd->nextevent == NULL)
473 {
474 pd->eventbuf.eventnum = TD_DEATH;
475 pd->eventbuf.eventdata = pd;
476
477 do
478 pd->nextevent = __nptl_last_event;
479 while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event,
480 pd, pd->nextevent));
481 }
482
483 /* Now call the function which signals the event. See
484 CONCURRENCY NOTES for the nptl_db interface comments. */
485 __nptl_death_event ();
486 }
487 }
488
489 /* The thread is exiting now. Don't set this bit until after we've hit
490 the event-reporting breakpoint, so that td_thr_get_info on us while at
491 the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE. */
492 atomic_fetch_or_relaxed (&pd->cancelhandling, EXITING_BITMASK);
493
494 if (__glibc_unlikely (atomic_fetch_add_relaxed (&__nptl_nthreads, -1) == 1))
495 /* This was the last thread. */
496 exit (0);
497
498 /* This prevents sending a signal from this thread to itself during
499 its final stages. This must come after the exit call above
500 because atexit handlers must not run with signals blocked.
501
502 Do not block SIGSETXID. The setxid handshake below expects the
503 signal to be delivered. (SIGSETXID cannot run application code,
504 nor does it use pthread_kill.) Reuse the pd->sigmask space for
505 computing the signal mask, to save stack space. */
506 internal_sigfillset (&pd->sigmask);
507 internal_sigdelset (&pd->sigmask, SIGSETXID);
508 INTERNAL_SYSCALL_CALL (rt_sigprocmask, SIG_BLOCK, &pd->sigmask, NULL,
509 __NSIG_BYTES);
510
511 /* Tell __pthread_kill_internal that this thread is about to exit.
512 If there is a __pthread_kill_internal in progress, this delays
513 the thread exit until the signal has been queued by the kernel
514 (so that the TID used to send it remains valid). */
515 __libc_lock_lock (pd->exit_lock);
516 pd->exiting = true;
517 __libc_lock_unlock (pd->exit_lock);
518
519#ifndef __ASSUME_SET_ROBUST_LIST
520 /* If this thread has any robust mutexes locked, handle them now. */
521# if __PTHREAD_MUTEX_HAVE_PREV
522 void *robust = pd->robust_head.list;
523# else
524 __pthread_slist_t *robust = pd->robust_list.__next;
525# endif
526 /* We let the kernel do the notification if it is able to do so.
527 If we have to do it here there for sure are no PI mutexes involved
528 since the kernel support for them is even more recent. */
529 if (!__nptl_set_robust_list_avail
530 && __builtin_expect (robust != (void *) &pd->robust_head, 0))
531 {
532 do
533 {
534 struct __pthread_mutex_s *this = (struct __pthread_mutex_s *)
535 ((char *) robust - offsetof (struct __pthread_mutex_s,
536 __list.__next));
537 robust = *((void **) robust);
538
539# if __PTHREAD_MUTEX_HAVE_PREV
540 this->__list.__prev = NULL;
541# endif
542 this->__list.__next = NULL;
543
544 atomic_fetch_or_acquire (&this->__lock, FUTEX_OWNER_DIED);
545 futex_wake ((unsigned int *) &this->__lock, 1,
546 /* XYZ */ FUTEX_SHARED);
547 }
548 while (robust != (void *) &pd->robust_head);
549 }
550#endif
551
552 if (!pd->user_stack)
553 advise_stack_range (pd->stackblock, pd->stackblock_size, (uintptr_t) pd,
554 pd->guardsize);
555
556 if (__glibc_unlikely (pd->cancelhandling & SETXID_BITMASK))
557 {
558 /* Some other thread might call any of the setXid functions and expect
559 us to reply. In this case wait until we did that. */
560 do
561 /* XXX This differs from the typical futex_wait_simple pattern in that
562 the futex_wait condition (setxid_futex) is different from the
563 condition used in the surrounding loop (cancelhandling). We need
564 to check and document why this is correct. */
565 futex_wait_simple (&pd->setxid_futex, 0, FUTEX_PRIVATE);
566 while (pd->cancelhandling & SETXID_BITMASK);
567
568 /* Reset the value so that the stack can be reused. */
569 pd->setxid_futex = 0;
570 }
571
572 /* If the thread is detached free the TCB. */
573 if (IS_DETACHED (pd))
574 /* Free the TCB. */
575 __nptl_free_tcb (pd);
576
577out:
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 rseq TLS is still registered at this point. Rely on implicit
585 unregistration performed by the kernel on thread teardown. This is not a
586 problem because the rseq TLS lives on the stack, and the stack outlives
587 the thread. If TCB allocation is ever changed, additional steps may be
588 required, such as performing explicit rseq unregistration before
589 reclaiming the rseq TLS area memory. It is NOT sufficient to block
590 signals because the kernel may write to the rseq area even without
591 signals.
592
593 The exit code is zero since in case all threads exit by calling
594 'pthread_exit' the exit status must be 0 (zero). */
595 while (1)
596 INTERNAL_SYSCALL_CALL (exit, 0);
597
598 /* NOTREACHED */
599}
600
601
602/* Return true iff obliged to report TD_CREATE events. */
603static bool
604report_thread_creation (struct pthread *pd)
605{
606 if (__glibc_unlikely (THREAD_GETMEM (THREAD_SELF, report_events)))
607 {
608 /* The parent thread is supposed to report events.
609 Check whether the TD_CREATE event is needed, too. */
610 const size_t idx = __td_eventword (TD_CREATE);
611 const uint32_t mask = __td_eventmask (TD_CREATE);
612
613 return ((mask & (__nptl_threads_events.event_bits[idx]
614 | pd->eventbuf.eventmask.event_bits[idx])) != 0);
615 }
616 return false;
617}
618
619
620int
621__pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr,
622 void *(*start_routine) (void *), void *arg)
623{
624 void *stackaddr = NULL;
625 size_t stacksize = 0;
626
627 /* Avoid a data race in the multi-threaded case, and call the
628 deferred initialization only once. */
629 if (__libc_single_threaded_internal)
630 {
631 late_init ();
632 __libc_single_threaded_internal = 0;
633 /* __libc_single_threaded can be accessed through copy relocations, so
634 it requires to update the external copy. */
635 __libc_single_threaded = 0;
636 }
637
638 const struct pthread_attr *iattr = (struct pthread_attr *) attr;
639 union pthread_attr_transparent default_attr;
640 bool destroy_default_attr = false;
641 bool c11 = (attr == ATTR_C11_THREAD);
642 if (iattr == NULL || c11)
643 {
644 int ret = __pthread_getattr_default_np (&default_attr.external);
645 if (ret != 0)
646 return ret;
647 destroy_default_attr = true;
648 iattr = &default_attr.internal;
649 }
650
651 struct pthread *pd = NULL;
652 int err = allocate_stack (iattr, &pd, &stackaddr, &stacksize);
653 int retval = 0;
654
655 if (__glibc_unlikely (err != 0))
656 /* Something went wrong. Maybe a parameter of the attributes is
657 invalid or we could not allocate memory. Note we have to
658 translate error codes. */
659 {
660 retval = err == ENOMEM ? EAGAIN : err;
661 goto out;
662 }
663
664
665 /* Initialize the TCB. All initializations with zero should be
666 performed in 'get_cached_stack'. This way we avoid doing this if
667 the stack freshly allocated with 'mmap'. */
668
669#if TLS_TCB_AT_TP
670 /* Reference to the TCB itself. */
671 pd->header.self = pd;
672
673 /* Self-reference for TLS. */
674 pd->header.tcb = pd;
675#endif
676
677 /* Store the address of the start routine and the parameter. Since
678 we do not start the function directly the stillborn thread will
679 get the information from its thread descriptor. */
680 pd->start_routine = start_routine;
681 pd->arg = arg;
682 pd->c11 = c11;
683
684 /* Copy the thread attribute flags. */
685 struct pthread *self = THREAD_SELF;
686 pd->flags = ((iattr->flags & ~(ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
687 | (self->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)));
688
689 /* Inherit rseq registration state. Without seccomp filters, rseq
690 registration will either always fail or always succeed. */
691 if ((int) THREAD_GETMEM_VOLATILE (self, rseq_area.cpu_id) >= 0)
692 pd->flags |= ATTR_FLAG_DO_RSEQ;
693
694 /* Initialize the field for the ID of the thread which is waiting
695 for us. This is a self-reference in case the thread is created
696 detached. */
697 pd->joinid = iattr->flags & ATTR_FLAG_DETACHSTATE ? pd : NULL;
698
699 /* The debug events are inherited from the parent. */
700 pd->eventbuf = self->eventbuf;
701
702
703 /* Copy the parent's scheduling parameters. The flags will say what
704 is valid and what is not. */
705 pd->schedpolicy = self->schedpolicy;
706 pd->schedparam = self->schedparam;
707
708 /* Copy the stack guard canary. */
709#ifdef THREAD_COPY_STACK_GUARD
710 THREAD_COPY_STACK_GUARD (pd);
711#endif
712
713 /* Copy the pointer guard value. */
714#ifdef THREAD_COPY_POINTER_GUARD
715 THREAD_COPY_POINTER_GUARD (pd);
716#endif
717
718 /* Setup tcbhead. */
719 tls_setup_tcbhead (pd);
720
721 /* Verify the sysinfo bits were copied in allocate_stack if needed. */
722#ifdef NEED_DL_SYSINFO
723 CHECK_THREAD_SYSINFO (pd);
724#endif
725
726 /* Determine scheduling parameters for the thread. */
727 if (__builtin_expect ((iattr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0, 0)
728 && (iattr->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) != 0)
729 {
730 /* Use the scheduling parameters the user provided. */
731 if (iattr->flags & ATTR_FLAG_POLICY_SET)
732 {
733 pd->schedpolicy = iattr->schedpolicy;
734 pd->flags |= ATTR_FLAG_POLICY_SET;
735 }
736 if (iattr->flags & ATTR_FLAG_SCHED_SET)
737 {
738 /* The values were validated in pthread_attr_setschedparam. */
739 pd->schedparam = iattr->schedparam;
740 pd->flags |= ATTR_FLAG_SCHED_SET;
741 }
742
743 if ((pd->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
744 != (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))
745 collect_default_sched (pd);
746 }
747
748 if (__glibc_unlikely (__nptl_nthreads == 1))
749 _IO_enable_locks ();
750
751 /* Pass the descriptor to the caller. */
752 *newthread = (pthread_t) pd;
753
754 LIBC_PROBE (pthread_create, 4, newthread, attr, start_routine, arg);
755
756 /* One more thread. We cannot have the thread do this itself, since it
757 might exist but not have been scheduled yet by the time we've returned
758 and need to check the value to behave correctly. We must do it before
759 creating the thread, in case it does get scheduled first and then
760 might mistakenly think it was the only thread. In the failure case,
761 we momentarily store a false value; this doesn't matter because there
762 is no kosher thing a signal handler interrupting us right here can do
763 that cares whether the thread count is correct. */
764 atomic_fetch_add_relaxed (&__nptl_nthreads, 1);
765
766 /* Our local value of stopped_start and thread_ran can be accessed at
767 any time. The PD->stopped_start may only be accessed if we have
768 ownership of PD (see CONCURRENCY NOTES above). */
769 bool stopped_start = false; bool thread_ran = false;
770
771 /* Block all signals, so that the new thread starts out with
772 signals disabled. This avoids race conditions in the thread
773 startup. */
774 internal_sigset_t original_sigmask;
775 internal_signal_block_all (&original_sigmask);
776
777 if (iattr->extension != NULL && iattr->extension->sigmask_set)
778 /* Use the signal mask in the attribute. The internal signals
779 have already been filtered by the public
780 pthread_attr_setsigmask_np interface. */
781 internal_sigset_from_sigset (&pd->sigmask, &iattr->extension->sigmask);
782 else
783 {
784 /* Conceptually, the new thread needs to inherit the signal mask
785 of this thread. Therefore, it needs to restore the saved
786 signal mask of this thread, so save it in the startup
787 information. */
788 pd->sigmask = original_sigmask;
789 /* Reset the cancellation signal mask in case this thread is
790 running cancellation. */
791 internal_sigdelset (&pd->sigmask, SIGCANCEL);
792 }
793
794 /* Start the thread. */
795 if (__glibc_unlikely (report_thread_creation (pd)))
796 {
797 stopped_start = true;
798
799 /* We always create the thread stopped at startup so we can
800 notify the debugger. */
801 retval = create_thread (pd, iattr, &stopped_start, stackaddr,
802 stacksize, &thread_ran);
803 if (retval == 0)
804 {
805 /* We retain ownership of PD until (a) (see CONCURRENCY NOTES
806 above). */
807
808 /* Assert stopped_start is true in both our local copy and the
809 PD copy. */
810 assert (stopped_start);
811 assert (pd->stopped_start);
812
813 /* Now fill in the information about the new thread in
814 the newly created thread's data structure. We cannot let
815 the new thread do this since we don't know whether it was
816 already scheduled when we send the event. */
817 pd->eventbuf.eventnum = TD_CREATE;
818 pd->eventbuf.eventdata = pd;
819
820 /* Enqueue the descriptor. */
821 do
822 pd->nextevent = __nptl_last_event;
823 while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event,
824 pd, pd->nextevent)
825 != 0);
826
827 /* Now call the function which signals the event. See
828 CONCURRENCY NOTES for the nptl_db interface comments. */
829 __nptl_create_event ();
830 }
831 }
832 else
833 retval = create_thread (pd, iattr, &stopped_start, stackaddr,
834 stacksize, &thread_ran);
835
836 /* Return to the previous signal mask, after creating the new
837 thread. */
838 internal_signal_restore_set (&original_sigmask);
839
840 if (__glibc_unlikely (retval != 0))
841 {
842 if (thread_ran)
843 /* State (c) and we not have PD ownership (see CONCURRENCY NOTES
844 above). We can assert that STOPPED_START must have been true
845 because thread creation didn't fail, but thread attribute setting
846 did. */
847 {
848 assert (stopped_start);
849 /* Signal the created thread to release PD ownership and early
850 exit so it could be joined. */
851 pd->setup_failed = 1;
852 lll_unlock (pd->lock, LLL_PRIVATE);
853
854 /* Similar to pthread_join, but since thread creation has failed at
855 startup there is no need to handle all the steps. */
856 pid_t tid;
857 while ((tid = atomic_load_acquire (&pd->tid)) != 0)
858 __futex_abstimed_wait_cancelable64 ((unsigned int *) &pd->tid,
859 tid, 0, NULL, LLL_SHARED);
860 }
861
862 /* State (c) or (d) and we have ownership of PD (see CONCURRENCY
863 NOTES above). */
864
865 /* Oops, we lied for a second. */
866 atomic_fetch_add_relaxed (&__nptl_nthreads, -1);
867
868 /* Free the resources. */
869 __nptl_deallocate_stack (pd);
870
871 /* We have to translate error codes. */
872 if (retval == ENOMEM)
873 retval = EAGAIN;
874 }
875 else
876 {
877 /* We don't know if we have PD ownership. Once we check the local
878 stopped_start we'll know if we're in state (a) or (b) (see
879 CONCURRENCY NOTES above). */
880 if (stopped_start)
881 /* State (a), we own PD. The thread blocked on this lock either
882 because we're doing TD_CREATE event reporting, or for some
883 other reason that create_thread chose. Now let it run
884 free. */
885 lll_unlock (pd->lock, LLL_PRIVATE);
886
887 /* We now have for sure more than one thread. The main thread might
888 not yet have the flag set. No need to set the global variable
889 again if this is what we use. */
890 THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1);
891 }
892
893 out:
894 if (destroy_default_attr)
895 __pthread_attr_destroy (&default_attr.external);
896
897 return retval;
898}
899versioned_symbol (libc, __pthread_create_2_1, pthread_create, GLIBC_2_34);
900libc_hidden_ver (__pthread_create_2_1, __pthread_create)
901#ifndef SHARED
902strong_alias (__pthread_create_2_1, __pthread_create)
903#endif
904
905#if OTHER_SHLIB_COMPAT (libpthread, GLIBC_2_1, GLIBC_2_34)
906compat_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1);
907#endif
908
909#if OTHER_SHLIB_COMPAT (libpthread, GLIBC_2_0, GLIBC_2_1)
910int
911__pthread_create_2_0 (pthread_t *newthread, const pthread_attr_t *attr,
912 void *(*start_routine) (void *), void *arg)
913{
914 /* The ATTR attribute is not really of type `pthread_attr_t *'. It has
915 the old size and access to the new members might crash the program.
916 We convert the struct now. */
917 struct pthread_attr new_attr;
918
919 if (attr != NULL)
920 {
921 struct pthread_attr *iattr = (struct pthread_attr *) attr;
922 size_t ps = __getpagesize ();
923
924 /* Copy values from the user-provided attributes. */
925 new_attr.schedparam = iattr->schedparam;
926 new_attr.schedpolicy = iattr->schedpolicy;
927 new_attr.flags = iattr->flags;
928
929 /* Fill in default values for the fields not present in the old
930 implementation. */
931 new_attr.guardsize = ps;
932 new_attr.stackaddr = NULL;
933 new_attr.stacksize = 0;
934 new_attr.extension = NULL;
935
936 /* We will pass this value on to the real implementation. */
937 attr = (pthread_attr_t *) &new_attr;
938 }
939
940 return __pthread_create_2_1 (newthread, attr, start_routine, arg);
941}
942compat_symbol (libpthread, __pthread_create_2_0, pthread_create,
943 GLIBC_2_0);
944#endif
945
946/* Information for libthread_db. */
947
948#include "../nptl_db/db_info.c"
949
950/* If pthread_create is present, libgcc_eh.a and libsupc++.a expects some other POSIX thread
951 functions to be present as well. */
952PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_lock)
953PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_trylock)
954PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_unlock)
955
956PTHREAD_STATIC_FN_REQUIRE (__pthread_once)
957PTHREAD_STATIC_FN_REQUIRE (__pthread_cancel)
958
959PTHREAD_STATIC_FN_REQUIRE (__pthread_key_create)
960PTHREAD_STATIC_FN_REQUIRE (__pthread_key_delete)
961PTHREAD_STATIC_FN_REQUIRE (__pthread_setspecific)
962PTHREAD_STATIC_FN_REQUIRE (__pthread_getspecific)
963