]>
Commit | Line | Data |
---|---|---|
bfff8b1b | 1 | /* Copyright (C) 2002-2017 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 PE |
16 | License along with the GNU C Library; if not, see |
17 | <http://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> |
d2e04918 | 35 | #include "libioP.h" |
76a50749 UD |
36 | |
37 | #include <shlib-compat.h> | |
38 | ||
3a097cc7 RM |
39 | #include <stap-probe.h> |
40 | ||
76a50749 | 41 | |
76a50749 UD |
42 | /* Nozero if debugging mode is enabled. */ |
43 | int __pthread_debug; | |
44 | ||
45 | /* Globally enabled events. */ | |
e965d514 | 46 | static td_thr_events_t __nptl_threads_events __attribute_used__; |
76a50749 UD |
47 | |
48 | /* Pointer to descriptor with the last event. */ | |
e965d514 | 49 | static struct pthread *__nptl_last_event __attribute_used__; |
76a50749 | 50 | |
47202270 UD |
51 | /* Number of threads running. */ |
52 | unsigned int __nptl_nthreads = 1; | |
53 | ||
76a50749 UD |
54 | |
55 | /* Code to allocate and deallocate a stack. */ | |
76a50749 UD |
56 | #include "allocatestack.c" |
57 | ||
f8bf15fe CD |
58 | /* CONCURRENCY NOTES: |
59 | ||
60 | Understanding who is the owner of the 'struct pthread' or 'PD' | |
61 | (refers to the value of the 'struct pthread *pd' function argument) | |
62 | is critically important in determining exactly which operations are | |
63 | allowed and which are not and when, particularly when it comes to the | |
64 | implementation of pthread_create, pthread_join, pthread_detach, and | |
65 | other functions which all operate on PD. | |
66 | ||
67 | The owner of PD is responsible for freeing the final resources | |
68 | associated with PD, and may examine the memory underlying PD at any | |
69 | point in time until it frees it back to the OS or to reuse by the | |
70 | runtime. | |
71 | ||
72 | The thread which calls pthread_create is called the creating thread. | |
73 | The creating thread begins as the owner of PD. | |
74 | ||
75 | During startup the new thread may examine PD in coordination with the | |
76 | owner thread (which may be itself). | |
77 | ||
78 | The four cases of ownership transfer are: | |
79 | ||
80 | (1) Ownership of PD is released to the process (all threads may use it) | |
81 | after the new thread starts in a joinable state | |
82 | i.e. pthread_create returns a usable pthread_t. | |
83 | ||
84 | (2) Ownership of PD is released to the new thread starting in a detached | |
85 | state. | |
86 | ||
87 | (3) Ownership of PD is dynamically released to a running thread via | |
88 | pthread_detach. | |
89 | ||
90 | (4) Ownership of PD is acquired by the thread which calls pthread_join. | |
91 | ||
92 | Implementation notes: | |
93 | ||
94 | The PD->stopped_start and thread_ran variables are used to determine | |
95 | exactly which of the four ownership states we are in and therefore | |
96 | what actions can be taken. For example after (2) we cannot read or | |
97 | write from PD anymore since the thread may no longer exist and the | |
fa17b9c7 CD |
98 | memory may be unmapped. |
99 | ||
100 | It is important to point out that PD->lock is being used both | |
101 | similar to a one-shot semaphore and subsequently as a mutex. The | |
102 | lock is taken in the parent to force the child to wait, and then the | |
103 | child releases the lock. However, this semaphore-like effect is used | |
104 | only for synchronizing the parent and child. After startup the lock | |
105 | is used like a mutex to create a critical section during which a | |
106 | single owner modifies the thread parameters. | |
107 | ||
108 | The most complicated cases happen during thread startup: | |
f8bf15fe CD |
109 | |
110 | (a) If the created thread is in a detached (PTHREAD_CREATE_DETACHED), | |
111 | or joinable (default PTHREAD_CREATE_JOINABLE) state and | |
112 | STOPPED_START is true, then the creating thread has ownership of | |
113 | PD until the PD->lock is released by pthread_create. If any | |
114 | errors occur we are in states (c), (d), or (e) below. | |
115 | ||
116 | (b) If the created thread is in a detached state | |
117 | (PTHREAD_CREATED_DETACHED), and STOPPED_START is false, then the | |
118 | creating thread has ownership of PD until it invokes the OS | |
119 | kernel's thread creation routine. If this routine returns | |
120 | without error, then the created thread owns PD; otherwise, see | |
121 | (c) and (e) below. | |
122 | ||
123 | (c) If the detached thread setup failed and THREAD_RAN is true, then | |
124 | the creating thread releases ownership to the new thread by | |
125 | sending a cancellation signal. All threads set THREAD_RAN to | |
126 | true as quickly as possible after returning from the OS kernel's | |
127 | thread creation routine. | |
128 | ||
129 | (d) If the joinable thread setup failed and THREAD_RAN is true, then | |
130 | then the creating thread retains ownership of PD and must cleanup | |
131 | state. Ownership cannot be released to the process via the | |
132 | return of pthread_create since a non-zero result entails PD is | |
133 | undefined and therefore cannot be joined to free the resources. | |
134 | We privately call pthread_join on the thread to finish handling | |
135 | the resource shutdown (Or at least we should, see bug 19511). | |
136 | ||
137 | (e) If the thread creation failed and THREAD_RAN is false, then the | |
138 | creating thread retains ownership of PD and must cleanup state. | |
139 | No waiting for the new thread is required because it never | |
140 | started. | |
141 | ||
142 | The nptl_db interface: | |
143 | ||
144 | The interface with nptl_db requires that we enqueue PD into a linked | |
145 | list and then call a function which the debugger will trap. The PD | |
146 | will then be dequeued and control returned to the thread. The caller | |
147 | at the time must have ownership of PD and such ownership remains | |
148 | after control returns to thread. The enqueued PD is removed from the | |
149 | linked list by the nptl_db callback td_thr_event_getmsg. The debugger | |
150 | must ensure that the thread does not resume execution, otherwise | |
151 | ownership of PD may be lost and examining PD will not be possible. | |
152 | ||
153 | Note that the GNU Debugger as of (December 10th 2015) commit | |
154 | c2c2a31fdb228d41ce3db62b268efea04bd39c18 no longer uses | |
155 | td_thr_event_getmsg and several other related nptl_db interfaces. The | |
156 | principal reason for this is that nptl_db does not support non-stop | |
157 | mode where other threads can run concurrently and modify runtime | |
158 | structures currently in use by the debugger and the nptl_db | |
159 | interface. | |
160 | ||
161 | Axioms: | |
162 | ||
163 | * The create_thread function can never set stopped_start to false. | |
164 | * The created thread can read stopped_start but never write to it. | |
165 | * The variable thread_ran is set some time after the OS thread | |
166 | creation routine returns, how much time after the thread is created | |
167 | is unspecified, but it should be as quickly as possible. | |
168 | ||
169 | */ | |
170 | ||
171 | /* CREATE THREAD NOTES: | |
172 | ||
173 | createthread.c defines the create_thread function, and two macros: | |
32fed10f RM |
174 | START_THREAD_DEFN and START_THREAD_SELF (see below). |
175 | ||
f8bf15fe CD |
176 | create_thread must initialize PD->stopped_start. It should be true |
177 | if the STOPPED_START parameter is true, or if create_thread needs the | |
178 | new thread to synchronize at startup for some other implementation | |
179 | reason. If STOPPED_START will be true, then create_thread is obliged | |
180 | to lock PD->lock before starting the thread. Then pthread_create | |
181 | unlocks PD->lock which synchronizes-with START_THREAD_DEFN in the | |
182 | child thread which does an acquire/release of PD->lock as the last | |
183 | action before calling the user entry point. The goal of all of this | |
184 | is to ensure that the required initial thread attributes are applied | |
185 | (by the creating thread) before the new thread runs user code. Note | |
186 | that the the functions pthread_getschedparam, pthread_setschedparam, | |
187 | pthread_setschedprio, __pthread_tpp_change_priority, and | |
188 | __pthread_current_priority reuse the same lock, PD->lock, for a | |
189 | similar purpose e.g. synchronizing the setting of similar thread | |
190 | attributes. These functions are never called before the thread is | |
191 | created, so don't participate in startup syncronization, but given | |
192 | that the lock is present already and in the unlocked state, reusing | |
193 | it saves space. | |
32fed10f RM |
194 | |
195 | The return value is zero for success or an errno code for failure. | |
196 | If the return value is ENOMEM, that will be translated to EAGAIN, | |
197 | so create_thread need not do that. On failure, *THREAD_RAN should | |
198 | be set to true iff the thread actually started up and then got | |
f8bf15fe | 199 | canceled before calling user code (*PD->start_routine). */ |
32fed10f | 200 | static int create_thread (struct pthread *pd, const struct pthread_attr *attr, |
f8bf15fe | 201 | bool *stopped_start, STACK_VARIABLES_PARMS, |
32fed10f RM |
202 | bool *thread_ran); |
203 | ||
8dea90aa | 204 | #include <createthread.c> |
76a50749 UD |
205 | |
206 | ||
76a50749 | 207 | struct pthread * |
90491dc4 | 208 | internal_function |
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 |
248 | void |
249 | attribute_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. */ | |
343 | void | |
90491dc4 | 344 | internal_function |
76a50749 UD |
345 | __free_tcb (struct pthread *pd) |
346 | { | |
347 | /* The thread is exiting now. */ | |
ba25bb0f UD |
348 | if (__builtin_expect (atomic_bit_test_set (&pd->cancelhandling, |
349 | TERMINATED_BIT) == 0, 1)) | |
76a50749 UD |
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 | ||
f17efcb4 | 357 | /* Free TPP data. */ |
a1ffb40e | 358 | if (__glibc_unlikely (pd->tpp != NULL)) |
f17efcb4 UD |
359 | { |
360 | struct priority_protection_data *tpp = pd->tpp; | |
361 | ||
362 | pd->tpp = NULL; | |
363 | free (tpp); | |
364 | } | |
365 | ||
76a50749 UD |
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 | ||
32fed10f RM |
374 | /* Local function to start thread and handle cleanup. |
375 | createthread.c defines the macro START_THREAD_DEFN to the | |
376 | declaration that its create_thread function will refer to, and | |
377 | START_THREAD_SELF to the expression to optimally deliver the new | |
378 | thread's THREAD_SELF value. */ | |
379 | START_THREAD_DEFN | |
76a50749 | 380 | { |
32fed10f | 381 | struct pthread *pd = START_THREAD_SELF; |
76a50749 UD |
382 | |
383 | #if HP_TIMING_AVAIL | |
384 | /* Remember the time when the thread was started. */ | |
385 | hp_timing_t now; | |
386 | HP_TIMING_NOW (now); | |
387 | THREAD_SETMEM (pd, cpuclock_offset, now); | |
388 | #endif | |
389 | ||
0e9d6240 UD |
390 | /* Initialize resolver state pointer. */ |
391 | __resp = &pd->res; | |
392 | ||
fd5bdc09 UD |
393 | /* Initialize pointers to locale data. */ |
394 | __ctype_init (); | |
395 | ||
66f1b8ee | 396 | /* Allow setxid from now onwards. */ |
a1ffb40e | 397 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) == -2)) |
a2f0363f | 398 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
66f1b8ee | 399 | |
0f6699ea UD |
400 | #ifdef __NR_set_robust_list |
401 | # ifndef __ASSUME_SET_ROBUST_LIST | |
402 | if (__set_robust_list_avail >= 0) | |
403 | # endif | |
404 | { | |
405 | INTERNAL_SYSCALL_DECL (err); | |
406 | /* This call should never fail because the initial call in init.c | |
407 | succeeded. */ | |
408 | INTERNAL_SYSCALL (set_robust_list, err, 2, &pd->robust_head, | |
409 | sizeof (struct robust_list_head)); | |
410 | } | |
411 | #endif | |
412 | ||
327ae257 | 413 | #ifdef SIGCANCEL |
b051fc44 UD |
414 | /* If the parent was running cancellation handlers while creating |
415 | the thread the new thread inherited the signal mask. Reset the | |
416 | cancellation signal mask. */ | |
a1ffb40e | 417 | if (__glibc_unlikely (pd->parent_cancelhandling & CANCELING_BITMASK)) |
b051fc44 UD |
418 | { |
419 | INTERNAL_SYSCALL_DECL (err); | |
420 | sigset_t mask; | |
421 | __sigemptyset (&mask); | |
422 | __sigaddset (&mask, SIGCANCEL); | |
423 | (void) INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_UNBLOCK, &mask, | |
424 | NULL, _NSIG / 8); | |
425 | } | |
327ae257 | 426 | #endif |
b051fc44 | 427 | |
76a50749 UD |
428 | /* This is where the try/finally block should be created. For |
429 | compilers without that support we do use setjmp. */ | |
877e51b2 UD |
430 | struct pthread_unwind_buf unwind_buf; |
431 | ||
432 | /* No previous handlers. */ | |
433 | unwind_buf.priv.data.prev = NULL; | |
434 | unwind_buf.priv.data.cleanup = NULL; | |
435 | ||
436 | int not_first_call; | |
437 | not_first_call = setjmp ((struct __jmp_buf_tag *) unwind_buf.cancel_jmp_buf); | |
a1ffb40e | 438 | if (__glibc_likely (! not_first_call)) |
76a50749 | 439 | { |
877e51b2 UD |
440 | /* Store the new cleanup handler info. */ |
441 | THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf); | |
442 | ||
f8bf15fe CD |
443 | /* We are either in (a) or (b), and in either case we either own |
444 | PD already (2) or are about to own PD (1), and so our only | |
445 | restriction would be that we can't free PD until we know we | |
446 | have ownership (see CONCURRENCY NOTES above). */ | |
a1ffb40e | 447 | if (__glibc_unlikely (pd->stopped_start)) |
5f66b766 UD |
448 | { |
449 | int oldtype = CANCEL_ASYNC (); | |
362038b0 | 450 | |
5f66b766 | 451 | /* Get the lock the parent locked to force synchronization. */ |
e51deae7 | 452 | lll_lock (pd->lock, LLL_PRIVATE); |
f8bf15fe CD |
453 | |
454 | /* We have ownership of PD now. */ | |
455 | ||
5f66b766 | 456 | /* And give it up right away. */ |
e51deae7 | 457 | lll_unlock (pd->lock, LLL_PRIVATE); |
362038b0 | 458 | |
5f66b766 UD |
459 | CANCEL_RESET (oldtype); |
460 | } | |
362038b0 | 461 | |
3a097cc7 RM |
462 | LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg); |
463 | ||
76a50749 | 464 | /* Run the code the user provided. */ |
cc775edf | 465 | THREAD_SETMEM (pd, result, pd->start_routine (pd->arg)); |
76a50749 UD |
466 | } |
467 | ||
ba384f6e | 468 | /* Call destructors for the thread_local TLS variables. */ |
e57b0c61 RM |
469 | #ifndef SHARED |
470 | if (&__call_tls_dtors != NULL) | |
471 | #endif | |
472 | __call_tls_dtors (); | |
ba384f6e | 473 | |
6b4686a5 | 474 | /* Run the destructor for the thread-local data. */ |
3fa21fd8 | 475 | __nptl_deallocate_tsd (); |
6b4686a5 | 476 | |
12d7ca07 RM |
477 | /* Clean up any state libc stored in thread-local variables. */ |
478 | __libc_thread_freeres (); | |
76a50749 | 479 | |
47202270 UD |
480 | /* If this is the last thread we terminate the process now. We |
481 | do not notify the debugger, it might just irritate it if there | |
482 | is no thread left. */ | |
a1ffb40e | 483 | if (__glibc_unlikely (atomic_decrement_and_test (&__nptl_nthreads))) |
47202270 UD |
484 | /* This was the last thread. */ |
485 | exit (0); | |
486 | ||
76a50749 | 487 | /* Report the death of the thread if this is wanted. */ |
a1ffb40e | 488 | if (__glibc_unlikely (pd->report_events)) |
76a50749 UD |
489 | { |
490 | /* See whether TD_DEATH is in any of the mask. */ | |
491 | const int idx = __td_eventword (TD_DEATH); | |
492 | const uint32_t mask = __td_eventmask (TD_DEATH); | |
493 | ||
494 | if ((mask & (__nptl_threads_events.event_bits[idx] | |
495 | | pd->eventbuf.eventmask.event_bits[idx])) != 0) | |
496 | { | |
497 | /* Yep, we have to signal the death. Add the descriptor to | |
498 | the list but only if it is not already on it. */ | |
499 | if (pd->nextevent == NULL) | |
500 | { | |
501 | pd->eventbuf.eventnum = TD_DEATH; | |
502 | pd->eventbuf.eventdata = pd; | |
503 | ||
504 | do | |
505 | pd->nextevent = __nptl_last_event; | |
5a3ab2fc UD |
506 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, |
507 | pd, pd->nextevent)); | |
76a50749 UD |
508 | } |
509 | ||
f8bf15fe CD |
510 | /* Now call the function which signals the event. See |
511 | CONCURRENCY NOTES for the nptl_db interface comments. */ | |
76a50749 UD |
512 | __nptl_death_event (); |
513 | } | |
514 | } | |
515 | ||
6461e577 RM |
516 | /* The thread is exiting now. Don't set this bit until after we've hit |
517 | the event-reporting breakpoint, so that td_thr_get_info on us while at | |
518 | the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE. */ | |
519 | atomic_bit_set (&pd->cancelhandling, EXITING_BIT); | |
76a50749 | 520 | |
0f6699ea | 521 | #ifndef __ASSUME_SET_ROBUST_LIST |
1bcfb5a5 | 522 | /* If this thread has any robust mutexes locked, handle them now. */ |
c252ec15 | 523 | # ifdef __PTHREAD_MUTEX_HAVE_PREV |
0f6699ea UD |
524 | void *robust = pd->robust_head.list; |
525 | # else | |
b007ce7c | 526 | __pthread_slist_t *robust = pd->robust_list.__next; |
0f6699ea | 527 | # endif |
df47504c UD |
528 | /* We let the kernel do the notification if it is able to do so. |
529 | If we have to do it here there for sure are no PI mutexes involved | |
530 | since the kernel support for them is even more recent. */ | |
0f6699ea | 531 | if (__set_robust_list_avail < 0 |
df47504c | 532 | && __builtin_expect (robust != (void *) &pd->robust_head, 0)) |
1bcfb5a5 UD |
533 | { |
534 | do | |
535 | { | |
b007ce7c | 536 | struct __pthread_mutex_s *this = (struct __pthread_mutex_s *) |
0f6699ea UD |
537 | ((char *) robust - offsetof (struct __pthread_mutex_s, |
538 | __list.__next)); | |
539 | robust = *((void **) robust); | |
d804f5df | 540 | |
0f6699ea | 541 | # ifdef __PTHREAD_MUTEX_HAVE_PREV |
b007ce7c | 542 | this->__list.__prev = NULL; |
0f6699ea UD |
543 | # endif |
544 | this->__list.__next = NULL; | |
1bcfb5a5 | 545 | |
c0c6bac9 | 546 | atomic_or (&this->__lock, FUTEX_OWNER_DIED); |
a2f0363f TR |
547 | futex_wake ((unsigned int *) &this->__lock, 1, |
548 | /* XYZ */ FUTEX_SHARED); | |
1bcfb5a5 | 549 | } |
df47504c | 550 | while (robust != (void *) &pd->robust_head); |
1bcfb5a5 | 551 | } |
0f6699ea | 552 | #endif |
1bcfb5a5 | 553 | |
b42a214c UD |
554 | /* Mark the memory of the stack as usable to the kernel. We free |
555 | everything except for the space used for the TCB itself. */ | |
556 | size_t pagesize_m1 = __getpagesize () - 1; | |
557 | #ifdef _STACK_GROWS_DOWN | |
558 | char *sp = CURRENT_STACK_FRAME; | |
559 | size_t freesize = (sp - (char *) pd->stackblock) & ~pagesize_m1; | |
b42a214c UD |
560 | assert (freesize < pd->stackblock_size); |
561 | if (freesize > PTHREAD_STACK_MIN) | |
9043e228 | 562 | __madvise (pd->stackblock, freesize - PTHREAD_STACK_MIN, MADV_DONTNEED); |
d615a473 CD |
563 | #else |
564 | /* Page aligned start of memory to free (higher than or equal | |
565 | to current sp plus the minimum stack size). */ | |
566 | void *freeblock = (void*)((size_t)(CURRENT_STACK_FRAME | |
567 | + PTHREAD_STACK_MIN | |
568 | + pagesize_m1) | |
569 | & ~pagesize_m1); | |
570 | char *free_end = (char *) (((uintptr_t) pd - pd->guardsize) & ~pagesize_m1); | |
571 | /* Is there any space to free? */ | |
572 | if (free_end > (char *)freeblock) | |
573 | { | |
574 | size_t freesize = (size_t)(free_end - (char *)freeblock); | |
575 | assert (freesize < pd->stackblock_size); | |
576 | __madvise (freeblock, freesize, MADV_DONTNEED); | |
577 | } | |
578 | #endif | |
b42a214c | 579 | |
76a50749 UD |
580 | /* If the thread is detached free the TCB. */ |
581 | if (IS_DETACHED (pd)) | |
582 | /* Free the TCB. */ | |
583 | __free_tcb (pd); | |
a1ffb40e | 584 | else if (__glibc_unlikely (pd->cancelhandling & SETXID_BITMASK)) |
dff9a7a1 UD |
585 | { |
586 | /* Some other thread might call any of the setXid functions and expect | |
587 | us to reply. In this case wait until we did that. */ | |
588 | do | |
a2f0363f TR |
589 | /* XXX This differs from the typical futex_wait_simple pattern in that |
590 | the futex_wait condition (setxid_futex) is different from the | |
591 | condition used in the surrounding loop (cancelhandling). We need | |
592 | to check and document why this is correct. */ | |
593 | futex_wait_simple (&pd->setxid_futex, 0, FUTEX_PRIVATE); | |
dff9a7a1 UD |
594 | while (pd->cancelhandling & SETXID_BITMASK); |
595 | ||
596 | /* Reset the value so that the stack can be reused. */ | |
597 | pd->setxid_futex = 0; | |
598 | } | |
76a50749 UD |
599 | |
600 | /* We cannot call '_exit' here. '_exit' will terminate the process. | |
601 | ||
602 | The 'exit' implementation in the kernel will signal when the | |
adcdc775 | 603 | process is really dead since 'clone' got passed the CLONE_CHILD_CLEARTID |
76a50749 UD |
604 | flag. The 'tid' field in the TCB will be set to zero. |
605 | ||
606 | The exit code is zero since in case all threads exit by calling | |
607 | 'pthread_exit' the exit status must be 0 (zero). */ | |
e0db6517 | 608 | __exit_thread (); |
76a50749 UD |
609 | |
610 | /* NOTREACHED */ | |
32fed10f RM |
611 | } |
612 | ||
613 | ||
614 | /* Return true iff obliged to report TD_CREATE events. */ | |
615 | static bool | |
616 | report_thread_creation (struct pthread *pd) | |
617 | { | |
618 | if (__glibc_unlikely (THREAD_GETMEM (THREAD_SELF, report_events))) | |
619 | { | |
620 | /* The parent thread is supposed to report events. | |
621 | Check whether the TD_CREATE event is needed, too. */ | |
622 | const size_t idx = __td_eventword (TD_CREATE); | |
623 | const uint32_t mask = __td_eventmask (TD_CREATE); | |
624 | ||
625 | return ((mask & (__nptl_threads_events.event_bits[idx] | |
626 | | pd->eventbuf.eventmask.event_bits[idx])) != 0); | |
627 | } | |
628 | return false; | |
76a50749 UD |
629 | } |
630 | ||
631 | ||
76a50749 | 632 | int |
80d9be81 JM |
633 | __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr, |
634 | void *(*start_routine) (void *), void *arg) | |
76a50749 UD |
635 | { |
636 | STACK_VARIABLES; | |
76a50749 | 637 | |
1e6da2b0 | 638 | const struct pthread_attr *iattr = (struct pthread_attr *) attr; |
61dd6208 SP |
639 | struct pthread_attr default_attr; |
640 | bool free_cpuset = false; | |
76a50749 | 641 | if (iattr == NULL) |
61dd6208 SP |
642 | { |
643 | lll_lock (__default_pthread_attr_lock, LLL_PRIVATE); | |
644 | default_attr = __default_pthread_attr; | |
645 | size_t cpusetsize = default_attr.cpusetsize; | |
646 | if (cpusetsize > 0) | |
647 | { | |
648 | cpu_set_t *cpuset; | |
649 | if (__glibc_likely (__libc_use_alloca (cpusetsize))) | |
650 | cpuset = __alloca (cpusetsize); | |
651 | else | |
652 | { | |
653 | cpuset = malloc (cpusetsize); | |
654 | if (cpuset == NULL) | |
655 | { | |
656 | lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE); | |
657 | return ENOMEM; | |
658 | } | |
659 | free_cpuset = true; | |
660 | } | |
661 | memcpy (cpuset, default_attr.cpuset, cpusetsize); | |
662 | default_attr.cpuset = cpuset; | |
663 | } | |
664 | lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE); | |
665 | iattr = &default_attr; | |
666 | } | |
76a50749 | 667 | |
dff9a7a1 | 668 | struct pthread *pd = NULL; |
1e6da2b0 | 669 | int err = ALLOCATE_STACK (iattr, &pd); |
61dd6208 SP |
670 | int retval = 0; |
671 | ||
a1ffb40e | 672 | if (__glibc_unlikely (err != 0)) |
76a50749 | 673 | /* Something went wrong. Maybe a parameter of the attributes is |
e988dba9 JL |
674 | invalid or we could not allocate memory. Note we have to |
675 | translate error codes. */ | |
61dd6208 SP |
676 | { |
677 | retval = err == ENOMEM ? EAGAIN : err; | |
678 | goto out; | |
679 | } | |
76a50749 UD |
680 | |
681 | ||
682 | /* Initialize the TCB. All initializations with zero should be | |
683 | performed in 'get_cached_stack'. This way we avoid doing this if | |
684 | the stack freshly allocated with 'mmap'. */ | |
685 | ||
d7329d4b | 686 | #if TLS_TCB_AT_TP |
76a50749 | 687 | /* Reference to the TCB itself. */ |
55c11fbd | 688 | pd->header.self = pd; |
76a50749 | 689 | |
d4f64e1a | 690 | /* Self-reference for TLS. */ |
55c11fbd | 691 | pd->header.tcb = pd; |
76a50749 UD |
692 | #endif |
693 | ||
694 | /* Store the address of the start routine and the parameter. Since | |
695 | we do not start the function directly the stillborn thread will | |
696 | get the information from its thread descriptor. */ | |
697 | pd->start_routine = start_routine; | |
698 | pd->arg = arg; | |
699 | ||
700 | /* Copy the thread attribute flags. */ | |
14ffbc83 UD |
701 | struct pthread *self = THREAD_SELF; |
702 | pd->flags = ((iattr->flags & ~(ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) | |
703 | | (self->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET))); | |
76a50749 UD |
704 | |
705 | /* Initialize the field for the ID of the thread which is waiting | |
706 | for us. This is a self-reference in case the thread is created | |
707 | detached. */ | |
708 | pd->joinid = iattr->flags & ATTR_FLAG_DETACHSTATE ? pd : NULL; | |
709 | ||
710 | /* The debug events are inherited from the parent. */ | |
14ffbc83 UD |
711 | pd->eventbuf = self->eventbuf; |
712 | ||
76a50749 | 713 | |
14ffbc83 UD |
714 | /* Copy the parent's scheduling parameters. The flags will say what |
715 | is valid and what is not. */ | |
716 | pd->schedpolicy = self->schedpolicy; | |
717 | pd->schedparam = self->schedparam; | |
76a50749 | 718 | |
35f1e827 UD |
719 | /* Copy the stack guard canary. */ |
720 | #ifdef THREAD_COPY_STACK_GUARD | |
721 | THREAD_COPY_STACK_GUARD (pd); | |
722 | #endif | |
723 | ||
827b7087 UD |
724 | /* Copy the pointer guard value. */ |
725 | #ifdef THREAD_COPY_POINTER_GUARD | |
726 | THREAD_COPY_POINTER_GUARD (pd); | |
727 | #endif | |
728 | ||
32fed10f RM |
729 | /* Verify the sysinfo bits were copied in allocate_stack if needed. */ |
730 | #ifdef NEED_DL_SYSINFO | |
731 | CHECK_THREAD_SYSINFO (pd); | |
732 | #endif | |
733 | ||
734 | /* Inform start_thread (above) about cancellation state that might | |
735 | translate into inherited signal state. */ | |
736 | pd->parent_cancelhandling = THREAD_GETMEM (THREAD_SELF, cancelhandling); | |
737 | ||
14ffbc83 | 738 | /* Determine scheduling parameters for the thread. */ |
61dd6208 | 739 | if (__builtin_expect ((iattr->flags & ATTR_FLAG_NOTINHERITSCHED) != 0, 0) |
14ffbc83 | 740 | && (iattr->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) != 0) |
76a50749 | 741 | { |
14ffbc83 UD |
742 | /* Use the scheduling parameters the user provided. */ |
743 | if (iattr->flags & ATTR_FLAG_POLICY_SET) | |
33cd1f74 RM |
744 | { |
745 | pd->schedpolicy = iattr->schedpolicy; | |
746 | pd->flags |= ATTR_FLAG_POLICY_SET; | |
747 | } | |
14ffbc83 | 748 | if (iattr->flags & ATTR_FLAG_SCHED_SET) |
33cd1f74 RM |
749 | { |
750 | /* The values were validated in pthread_attr_setschedparam. */ | |
751 | pd->schedparam = iattr->schedparam; | |
752 | pd->flags |= ATTR_FLAG_SCHED_SET; | |
753 | } | |
f214ff74 RM |
754 | |
755 | if ((pd->flags & (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) | |
756 | != (ATTR_FLAG_SCHED_SET | ATTR_FLAG_POLICY_SET)) | |
757 | collect_default_sched (pd); | |
76a50749 UD |
758 | } |
759 | ||
d2e04918 SN |
760 | if (__glibc_unlikely (__nptl_nthreads == 1)) |
761 | _IO_enable_locks (); | |
762 | ||
76a50749 UD |
763 | /* Pass the descriptor to the caller. */ |
764 | *newthread = (pthread_t) pd; | |
765 | ||
5acf7263 RM |
766 | LIBC_PROBE (pthread_create, 4, newthread, attr, start_routine, arg); |
767 | ||
32fed10f RM |
768 | /* One more thread. We cannot have the thread do this itself, since it |
769 | might exist but not have been scheduled yet by the time we've returned | |
770 | and need to check the value to behave correctly. We must do it before | |
771 | creating the thread, in case it does get scheduled first and then | |
772 | might mistakenly think it was the only thread. In the failure case, | |
773 | we momentarily store a false value; this doesn't matter because there | |
774 | is no kosher thing a signal handler interrupting us right here can do | |
775 | that cares whether the thread count is correct. */ | |
776 | atomic_increment (&__nptl_nthreads); | |
777 | ||
f8bf15fe CD |
778 | /* Our local value of stopped_start and thread_ran can be accessed at |
779 | any time. The PD->stopped_start may only be accessed if we have | |
780 | ownership of PD (see CONCURRENCY NOTES above). */ | |
781 | bool stopped_start = false; bool thread_ran = false; | |
32fed10f | 782 | |
76a50749 | 783 | /* Start the thread. */ |
32fed10f RM |
784 | if (__glibc_unlikely (report_thread_creation (pd))) |
785 | { | |
f8bf15fe CD |
786 | stopped_start = true; |
787 | ||
788 | /* We always create the thread stopped at startup so we can | |
789 | notify the debugger. */ | |
790 | retval = create_thread (pd, iattr, &stopped_start, | |
791 | STACK_VARIABLES_ARGS, &thread_ran); | |
32fed10f RM |
792 | if (retval == 0) |
793 | { | |
f8bf15fe CD |
794 | /* We retain ownership of PD until (a) (see CONCURRENCY NOTES |
795 | above). */ | |
796 | ||
797 | /* Assert stopped_start is true in both our local copy and the | |
798 | PD copy. */ | |
799 | assert (stopped_start); | |
32fed10f RM |
800 | assert (pd->stopped_start); |
801 | ||
802 | /* Now fill in the information about the new thread in | |
803 | the newly created thread's data structure. We cannot let | |
804 | the new thread do this since we don't know whether it was | |
805 | already scheduled when we send the event. */ | |
806 | pd->eventbuf.eventnum = TD_CREATE; | |
807 | pd->eventbuf.eventdata = pd; | |
808 | ||
809 | /* Enqueue the descriptor. */ | |
810 | do | |
811 | pd->nextevent = __nptl_last_event; | |
812 | while (atomic_compare_and_exchange_bool_acq (&__nptl_last_event, | |
813 | pd, pd->nextevent) | |
814 | != 0); | |
815 | ||
f8bf15fe CD |
816 | /* Now call the function which signals the event. See |
817 | CONCURRENCY NOTES for the nptl_db interface comments. */ | |
32fed10f RM |
818 | __nptl_create_event (); |
819 | } | |
820 | } | |
821 | else | |
f8bf15fe CD |
822 | retval = create_thread (pd, iattr, &stopped_start, |
823 | STACK_VARIABLES_ARGS, &thread_ran); | |
32fed10f RM |
824 | |
825 | if (__glibc_unlikely (retval != 0)) | |
826 | { | |
32fed10f | 827 | if (thread_ran) |
f8bf15fe CD |
828 | /* State (c) or (d) and we may not have PD ownership (see |
829 | CONCURRENCY NOTES above). We can assert that STOPPED_START | |
830 | must have been true because thread creation didn't fail, but | |
831 | thread attribute setting did. */ | |
832 | /* See bug 19511 which explains why doing nothing here is a | |
833 | resource leak for a joinable thread. */ | |
834 | assert (stopped_start); | |
32fed10f RM |
835 | else |
836 | { | |
f8bf15fe CD |
837 | /* State (e) and we have ownership of PD (see CONCURRENCY |
838 | NOTES above). */ | |
839 | ||
32fed10f RM |
840 | /* Oops, we lied for a second. */ |
841 | atomic_decrement (&__nptl_nthreads); | |
842 | ||
843 | /* Perhaps a thread wants to change the IDs and is waiting for this | |
844 | stillborn thread. */ | |
845 | if (__glibc_unlikely (atomic_exchange_acq (&pd->setxid_futex, 0) | |
846 | == -2)) | |
a2f0363f | 847 | futex_wake (&pd->setxid_futex, 1, FUTEX_PRIVATE); |
32fed10f RM |
848 | |
849 | /* Free the resources. */ | |
850 | __deallocate_stack (pd); | |
851 | } | |
852 | ||
853 | /* We have to translate error codes. */ | |
854 | if (retval == ENOMEM) | |
855 | retval = EAGAIN; | |
856 | } | |
857 | else | |
858 | { | |
f8bf15fe CD |
859 | /* We don't know if we have PD ownership. Once we check the local |
860 | stopped_start we'll know if we're in state (a) or (b) (see | |
861 | CONCURRENCY NOTES above). */ | |
862 | if (stopped_start) | |
863 | /* State (a), we own PD. The thread blocked on this lock either | |
864 | because we're doing TD_CREATE event reporting, or for some | |
865 | other reason that create_thread chose. Now let it run | |
866 | free. */ | |
32fed10f RM |
867 | lll_unlock (pd->lock, LLL_PRIVATE); |
868 | ||
869 | /* We now have for sure more than one thread. The main thread might | |
870 | not yet have the flag set. No need to set the global variable | |
871 | again if this is what we use. */ | |
872 | THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1); | |
873 | } | |
61dd6208 SP |
874 | |
875 | out: | |
876 | if (__glibc_unlikely (free_cpuset)) | |
877 | free (default_attr.cpuset); | |
878 | ||
879 | return retval; | |
76a50749 UD |
880 | } |
881 | versioned_symbol (libpthread, __pthread_create_2_1, pthread_create, GLIBC_2_1); | |
882 | ||
883 | ||
884 | #if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1) | |
885 | int | |
80d9be81 JM |
886 | __pthread_create_2_0 (pthread_t *newthread, const pthread_attr_t *attr, |
887 | void *(*start_routine) (void *), void *arg) | |
76a50749 UD |
888 | { |
889 | /* The ATTR attribute is not really of type `pthread_attr_t *'. It has | |
890 | the old size and access to the new members might crash the program. | |
891 | We convert the struct now. */ | |
892 | struct pthread_attr new_attr; | |
893 | ||
894 | if (attr != NULL) | |
895 | { | |
896 | struct pthread_attr *iattr = (struct pthread_attr *) attr; | |
897 | size_t ps = __getpagesize (); | |
898 | ||
899 | /* Copy values from the user-provided attributes. */ | |
900 | new_attr.schedparam = iattr->schedparam; | |
901 | new_attr.schedpolicy = iattr->schedpolicy; | |
902 | new_attr.flags = iattr->flags; | |
903 | ||
904 | /* Fill in default values for the fields not present in the old | |
905 | implementation. */ | |
906 | new_attr.guardsize = ps; | |
907 | new_attr.stackaddr = NULL; | |
908 | new_attr.stacksize = 0; | |
ca85ede0 | 909 | new_attr.cpuset = NULL; |
76a50749 UD |
910 | |
911 | /* We will pass this value on to the real implementation. */ | |
912 | attr = (pthread_attr_t *) &new_attr; | |
913 | } | |
914 | ||
915 | return __pthread_create_2_1 (newthread, attr, start_routine, arg); | |
916 | } | |
917 | compat_symbol (libpthread, __pthread_create_2_0, pthread_create, | |
918 | GLIBC_2_0); | |
919 | #endif | |
7f08f55a RM |
920 | \f |
921 | /* Information for libthread_db. */ | |
922 | ||
923 | #include "../nptl_db/db_info.c" | |
b639d0c9 UD |
924 | \f |
925 | /* If pthread_create is present, libgcc_eh.a and libsupc++.a expects some other POSIX thread | |
926 | functions to be present as well. */ | |
fa872e1b AZ |
927 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_lock) |
928 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_trylock) | |
929 | PTHREAD_STATIC_FN_REQUIRE (__pthread_mutex_unlock) | |
b639d0c9 | 930 | |
fa872e1b AZ |
931 | PTHREAD_STATIC_FN_REQUIRE (__pthread_once) |
932 | PTHREAD_STATIC_FN_REQUIRE (__pthread_cancel) | |
b639d0c9 | 933 | |
fa872e1b AZ |
934 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_create) |
935 | PTHREAD_STATIC_FN_REQUIRE (__pthread_key_delete) | |
936 | PTHREAD_STATIC_FN_REQUIRE (__pthread_setspecific) | |
937 | PTHREAD_STATIC_FN_REQUIRE (__pthread_getspecific) |