1/* Copyright (C) 2003-2018 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Martin Schwidefsky <schwidefsky@de.ibm.com>, 2003.
4
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
9
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
14
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <http://www.gnu.org/licenses/>. */
18
19#include <endian.h>
20#include <errno.h>
21#include <sysdep.h>
22#include <futex-internal.h>
23#include <pthread.h>
24#include <pthreadP.h>
25#include <sys/time.h>
26#include <atomic.h>
27#include <stdint.h>
28#include <stdbool.h>
29
30#include <shlib-compat.h>
31#include <stap-probe.h>
32#include <time.h>
33
34#include "pthread_cond_common.c"
35
36
37struct _condvar_cleanup_buffer
38{
39 uint64_t wseq;
40 pthread_cond_t *cond;
41 pthread_mutex_t *mutex;
42 int private;
43};
44
45
46/* Decrease the waiter reference count. */
47static void
48__condvar_confirm_wakeup (pthread_cond_t *cond, int private)
49{
50 /* If destruction is pending (i.e., the wake-request flag is nonzero) and we
51 are the last waiter (prior value of __wrefs was 1 << 3), then wake any
52 threads waiting in pthread_cond_destroy. Release MO to synchronize with
53 these threads. Don't bother clearing the wake-up request flag. */
54 if ((atomic_fetch_add_release (&cond->__data.__wrefs, -8) >> 2) == 3)
55 futex_wake (&cond->__data.__wrefs, INT_MAX, private);
56}
57
58
59/* Cancel waiting after having registered as a waiter previously. SEQ is our
60 position and G is our group index.
61 The goal of cancellation is to make our group smaller if that is still
62 possible. If we are in a closed group, this is not possible anymore; in
63 this case, we need to send a replacement signal for the one we effectively
64 consumed because the signal should have gotten consumed by another waiter
65 instead; we must not both cancel waiting and consume a signal.
66
67 Must not be called while still holding a reference on the group.
68
69 Returns true iff we consumed a signal.
70
71 On some kind of timeouts, we may be able to pretend that a signal we
72 effectively consumed happened before the timeout (i.e., similarly to first
73 spinning on signals before actually checking whether the timeout has
74 passed already). Doing this would allow us to skip sending a replacement
75 signal, but this case might happen rarely because the end of the timeout
76 must race with someone else sending a signal. Therefore, we don't bother
77 trying to optimize this. */
78static void
79__condvar_cancel_waiting (pthread_cond_t *cond, uint64_t seq, unsigned int g,
80 int private)
81{
82 bool consumed_signal = false;
83
84 /* No deadlock with group switching is possible here because we have do
85 not hold a reference on the group. */
86 __condvar_acquire_lock (cond, private);
87
88 uint64_t g1_start = __condvar_load_g1_start_relaxed (cond) >> 1;
89 if (g1_start > seq)
90 {
91 /* Our group is closed, so someone provided enough signals for it.
92 Thus, we effectively consumed a signal. */
93 consumed_signal = true;
94 }
95 else
96 {
97 if (g1_start + __condvar_get_orig_size (cond) <= seq)
98 {
99 /* We are in the current G2 and thus cannot have consumed a signal.
100 Reduce its effective size or handle overflow. Remember that in
101 G2, unsigned int size is zero or a negative value. */
102 if (cond->__data.__g_size[g] + __PTHREAD_COND_MAX_GROUP_SIZE > 0)
103 {
104 cond->__data.__g_size[g]--;
105 }
106 else
107 {
108 /* Cancellations would overflow the maximum group size. Just
109 wake up everyone spuriously to create a clean state. This
110 also means we do not consume a signal someone else sent. */
111 __condvar_release_lock (cond, private);
112 __pthread_cond_broadcast (cond);
113 return;
114 }
115 }
116 else
117 {
118 /* We are in current G1. If the group's size is zero, someone put
119 a signal in the group that nobody else but us can consume. */
120 if (cond->__data.__g_size[g] == 0)
121 consumed_signal = true;
122 else
123 {
124 /* Otherwise, we decrease the size of the group. This is
125 equivalent to atomically putting in a signal just for us and
126 consuming it right away. We do not consume a signal sent
127 by someone else. We also cannot have consumed a futex
128 wake-up because if we were cancelled or timed out in a futex
129 call, the futex will wake another waiter. */
130 cond->__data.__g_size[g]--;
131 }
132 }
133 }
134
135 __condvar_release_lock (cond, private);
136
137 if (consumed_signal)
138 {
139 /* We effectively consumed a signal even though we didn't want to.
140 Therefore, we need to send a replacement signal.
141 If we would want to optimize this, we could do what
142 pthread_cond_signal does right in the critical section above. */
143 __pthread_cond_signal (cond);
144 }
145}
146
147/* Wake up any signalers that might be waiting. */
148static void
149__condvar_dec_grefs (pthread_cond_t *cond, unsigned int g, int private)
150{
151 /* Release MO to synchronize-with the acquire load in
152 __condvar_quiesce_and_switch_g1. */
153 if (atomic_fetch_add_release (cond->__data.__g_refs + g, -2) == 3)
154 {
155 /* Clear the wake-up request flag before waking up. We do not need more
156 than relaxed MO and it doesn't matter if we apply this for an aliased
157 group because we wake all futex waiters right after clearing the
158 flag. */
159 atomic_fetch_and_relaxed (cond->__data.__g_refs + g, ~(unsigned int) 1);
160 futex_wake (cond->__data.__g_refs + g, INT_MAX, private);
161 }
162}
163
164/* Clean-up for cancellation of waiters waiting for normal signals. We cancel
165 our registration as a waiter, confirm we have woken up, and re-acquire the
166 mutex. */
167static void
168__condvar_cleanup_waiting (void *arg)
169{
170 struct _condvar_cleanup_buffer *cbuffer =
171 (struct _condvar_cleanup_buffer *) arg;
172 pthread_cond_t *cond = cbuffer->cond;
173 unsigned g = cbuffer->wseq & 1;
174
175 __condvar_dec_grefs (cond, g, cbuffer->private);
176
177 __condvar_cancel_waiting (cond, cbuffer->wseq >> 1, g, cbuffer->private);
178 /* FIXME With the current cancellation implementation, it is possible that
179 a thread is cancelled after it has returned from a syscall. This could
180 result in a cancelled waiter consuming a futex wake-up that is then
181 causing another waiter in the same group to not wake up. To work around
182 this issue until we have fixed cancellation, just add a futex wake-up
183 conservatively. */
184 futex_wake (cond->__data.__g_signals + g, 1, cbuffer->private);
185
186 __condvar_confirm_wakeup (cond, cbuffer->private);
187
188 /* XXX If locking the mutex fails, should we just stop execution? This
189 might be better than silently ignoring the error. */
190 __pthread_mutex_cond_lock (cbuffer->mutex);
191}
192
193/* This condvar implementation guarantees that all calls to signal and
194 broadcast and all of the three virtually atomic parts of each call to wait
195 (i.e., (1) releasing the mutex and blocking, (2) unblocking, and (3) re-
196 acquiring the mutex) happen in some total order that is consistent with the
197 happens-before relations in the calling program. However, this order does
198 not necessarily result in additional happens-before relations being
199 established (which aligns well with spurious wake-ups being allowed).
200
201 All waiters acquire a certain position in a 64b waiter sequence (__wseq).
202 This sequence determines which waiters are allowed to consume signals.
203 A broadcast is equal to sending as many signals as are unblocked waiters.
204 When a signal arrives, it samples the current value of __wseq with a
205 relaxed-MO load (i.e., the position the next waiter would get). (This is
206 sufficient because it is consistent with happens-before; the caller can
207 enforce stronger ordering constraints by calling signal while holding the
208 mutex.) Only waiters with a position less than the __wseq value observed
209 by the signal are eligible to consume this signal.
210
211 This would be straight-forward to implement if waiters would just spin but
212 we need to let them block using futexes. Futexes give no guarantee of
213 waking in FIFO order, so we cannot reliably wake eligible waiters if we
214 just use a single futex. Also, futex words are 32b in size, but we need
215 to distinguish more than 1<<32 states because we need to represent the
216 order of wake-up (and thus which waiters are eligible to consume signals);
217 blocking in a futex is not atomic with a waiter determining its position in
218 the waiter sequence, so we need the futex word to reliably notify waiters
219 that they should not attempt to block anymore because they have been
220 already signaled in the meantime. While an ABA issue on a 32b value will
221 be rare, ignoring it when we are aware of it is not the right thing to do
222 either.
223
224 Therefore, we use a 64b counter to represent the waiter sequence (on
225 architectures which only support 32b atomics, we use a few bits less).
226 To deal with the blocking using futexes, we maintain two groups of waiters:
227 * Group G1 consists of waiters that are all eligible to consume signals;
228 incoming signals will always signal waiters in this group until all
229 waiters in G1 have been signaled.
230 * Group G2 consists of waiters that arrive when a G1 is present and still
231 contains waiters that have not been signaled. When all waiters in G1
232 are signaled and a new signal arrives, the new signal will convert G2
233 into the new G1 and create a new G2 for future waiters.
234
235 We cannot allocate new memory because of process-shared condvars, so we
236 have just two slots of groups that change their role between G1 and G2.
237 Each has a separate futex word, a number of signals available for
238 consumption, a size (number of waiters in the group that have not been
239 signaled), and a reference count.
240
241 The group reference count is used to maintain the number of waiters that
242 are using the group's futex. Before a group can change its role, the
243 reference count must show that no waiters are using the futex anymore; this
244 prevents ABA issues on the futex word.
245
246 To represent which intervals in the waiter sequence the groups cover (and
247 thus also which group slot contains G1 or G2), we use a 64b counter to
248 designate the start position of G1 (inclusive), and a single bit in the
249 waiter sequence counter to represent which group slot currently contains
250 G2. This allows us to switch group roles atomically wrt. waiters obtaining
251 a position in the waiter sequence. The G1 start position allows waiters to
252 figure out whether they are in a group that has already been completely
253 signaled (i.e., if the current G1 starts at a later position that the
254 waiter's position). Waiters cannot determine whether they are currently
255 in G2 or G1 -- but they do not have too because all they are interested in
256 is whether there are available signals, and they always start in G2 (whose
257 group slot they know because of the bit in the waiter sequence. Signalers
258 will simply fill the right group until it is completely signaled and can
259 be closed (they do not switch group roles until they really have to to
260 decrease the likelihood of having to wait for waiters still holding a
261 reference on the now-closed G1).
262
263 Signalers maintain the initial size of G1 to be able to determine where
264 G2 starts (G2 is always open-ended until it becomes G1). They track the
265 remaining size of a group; when waiters cancel waiting (due to PThreads
266 cancellation or timeouts), they will decrease this remaining size as well.
267
268 To implement condvar destruction requirements (i.e., that
269 pthread_cond_destroy can be called as soon as all waiters have been
270 signaled), waiters increment a reference count before starting to wait and
271 decrement it after they stopped waiting but right before they acquire the
272 mutex associated with the condvar.
273
274 pthread_cond_t thus consists of the following (bits that are used for
275 flags and are not part of the primary value of each field but necessary
276 to make some things atomic or because there was no space for them
277 elsewhere in the data structure):
278
279 __wseq: Waiter sequence counter
280 * LSB is index of current G2.
281 * Waiters fetch-add while having acquire the mutex associated with the
282 condvar. Signalers load it and fetch-xor it concurrently.
283 __g1_start: Starting position of G1 (inclusive)
284 * LSB is index of current G2.
285 * Modified by signalers while having acquired the condvar-internal lock
286 and observed concurrently by waiters.
287 __g1_orig_size: Initial size of G1
288 * The two least-significant bits represent the condvar-internal lock.
289 * Only accessed while having acquired the condvar-internal lock.
290 __wrefs: Waiter reference counter.
291 * Bit 2 is true if waiters should run futex_wake when they remove the
292 last reference. pthread_cond_destroy uses this as futex word.
293 * Bit 1 is the clock ID (0 == CLOCK_REALTIME, 1 == CLOCK_MONOTONIC).
294 * Bit 0 is true iff this is a process-shared condvar.
295 * Simple reference count used by both waiters and pthread_cond_destroy.
296 (If the format of __wrefs is changed, update nptl_lock_constants.pysym
297 and the pretty printers.)
298 For each of the two groups, we have:
299 __g_refs: Futex waiter reference count.
300 * LSB is true if waiters should run futex_wake when they remove the
301 last reference.
302 * Reference count used by waiters concurrently with signalers that have
303 acquired the condvar-internal lock.
304 __g_signals: The number of signals that can still be consumed.
305 * Used as a futex word by waiters. Used concurrently by waiters and
306 signalers.
307 * LSB is true iff this group has been completely signaled (i.e., it is
308 closed).
309 __g_size: Waiters remaining in this group (i.e., which have not been
310 signaled yet.
311 * Accessed by signalers and waiters that cancel waiting (both do so only
312 when having acquired the condvar-internal lock.
313 * The size of G2 is always zero because it cannot be determined until
314 the group becomes G1.
315 * Although this is of unsigned type, we rely on using unsigned overflow
316 rules to make this hold effectively negative values too (in
317 particular, when waiters in G2 cancel waiting).
318
319 A PTHREAD_COND_INITIALIZER condvar has all fields set to zero, which yields
320 a condvar that has G2 starting at position 0 and a G1 that is closed.
321
322 Because waiters do not claim ownership of a group right when obtaining a
323 position in __wseq but only reference count the group when using futexes
324 to block, it can happen that a group gets closed before a waiter can
325 increment the reference count. Therefore, waiters have to check whether
326 their group is already closed using __g1_start. They also have to perform
327 this check when spinning when trying to grab a signal from __g_signals.
328 Note that for these checks, using relaxed MO to load __g1_start is
329 sufficient because if a waiter can see a sufficiently large value, it could
330 have also consume a signal in the waiters group.
331
332 Waiters try to grab a signal from __g_signals without holding a reference
333 count, which can lead to stealing a signal from a more recent group after
334 their own group was already closed. They cannot always detect whether they
335 in fact did because they do not know when they stole, but they can
336 conservatively add a signal back to the group they stole from; if they
337 did so unnecessarily, all that happens is a spurious wake-up. To make this
338 even less likely, __g1_start contains the index of the current g2 too,
339 which allows waiters to check if there aliasing on the group slots; if
340 there wasn't, they didn't steal from the current G1, which means that the
341 G1 they stole from must have been already closed and they do not need to
342 fix anything.
343
344 It is essential that the last field in pthread_cond_t is __g_signals[1]:
345 The previous condvar used a pointer-sized field in pthread_cond_t, so a
346 PTHREAD_COND_INITIALIZER from that condvar implementation might only
347 initialize 4 bytes to zero instead of the 8 bytes we need (i.e., 44 bytes
348 in total instead of the 48 we need). __g_signals[1] is not accessed before
349 the first group switch (G2 starts at index 0), which will set its value to
350 zero after a harmless fetch-or whose return value is ignored. This
351 effectively completes initialization.
352
353
354 Limitations:
355 * This condvar isn't designed to allow for more than
356 __PTHREAD_COND_MAX_GROUP_SIZE * (1 << 31) calls to __pthread_cond_wait.
357 * More than __PTHREAD_COND_MAX_GROUP_SIZE concurrent waiters are not
358 supported.
359 * Beyond what is allowed as errors by POSIX or documented, we can also
360 return the following errors:
361 * EPERM if MUTEX is a recursive mutex and the caller doesn't own it.
362 * EOWNERDEAD or ENOTRECOVERABLE when using robust mutexes. Unlike
363 for other errors, this can happen when we re-acquire the mutex; this
364 isn't allowed by POSIX (which requires all errors to virtually happen
365 before we release the mutex or change the condvar state), but there's
366 nothing we can do really.
367 * When using PTHREAD_MUTEX_PP_* mutexes, we can also return all errors
368 returned by __pthread_tpp_change_priority. We will already have
369 released the mutex in such cases, so the caller cannot expect to own
370 MUTEX.
371
372 Other notes:
373 * Instead of the normal mutex unlock / lock functions, we use
374 __pthread_mutex_unlock_usercnt(m, 0) / __pthread_mutex_cond_lock(m)
375 because those will not change the mutex-internal users count, so that it
376 can be detected when a condvar is still associated with a particular
377 mutex because there is a waiter blocked on this condvar using this mutex.
378*/
379static __always_inline int
380__pthread_cond_wait_common (pthread_cond_t *cond, pthread_mutex_t *mutex,
381 const struct timespec *abstime)
382{
383 const int maxspin = 0;
384 int err;
385 int result = 0;
386
387 LIBC_PROBE (cond_wait, 2, cond, mutex);
388
389 /* Acquire a position (SEQ) in the waiter sequence (WSEQ). We use an
390 atomic operation because signals and broadcasts may update the group
391 switch without acquiring the mutex. We do not need release MO here
392 because we do not need to establish any happens-before relation with
393 signalers (see __pthread_cond_signal); modification order alone
394 establishes a total order of waiters/signals. We do need acquire MO
395 to synchronize with group reinitialization in
396 __condvar_quiesce_and_switch_g1. */
397 uint64_t wseq = __condvar_fetch_add_wseq_acquire (cond, 2);
398 /* Find our group's index. We always go into what was G2 when we acquired
399 our position. */
400 unsigned int g = wseq & 1;
401 uint64_t seq = wseq >> 1;
402
403 /* Increase the waiter reference count. Relaxed MO is sufficient because
404 we only need to synchronize when decrementing the reference count. */
405 unsigned int flags = atomic_fetch_add_relaxed (&cond->__data.__wrefs, 8);
406 int private = __condvar_get_private (flags);
407
408 /* Now that we are registered as a waiter, we can release the mutex.
409 Waiting on the condvar must be atomic with releasing the mutex, so if
410 the mutex is used to establish a happens-before relation with any
411 signaler, the waiter must be visible to the latter; thus, we release the
412 mutex after registering as waiter.
413 If releasing the mutex fails, we just cancel our registration as a
414 waiter and confirm that we have woken up. */
415 err = __pthread_mutex_unlock_usercnt (mutex, 0);
416 if (__glibc_unlikely (err != 0))
417 {
418 __condvar_cancel_waiting (cond, seq, g, private);
419 __condvar_confirm_wakeup (cond, private);
420 return err;
421 }
422
423 /* Now wait until a signal is available in our group or it is closed.
424 Acquire MO so that if we observe a value of zero written after group
425 switching in __condvar_quiesce_and_switch_g1, we synchronize with that
426 store and will see the prior update of __g1_start done while switching
427 groups too. */
428 unsigned int signals = atomic_load_acquire (cond->__data.__g_signals + g);
429
430 do
431 {
432 while (1)
433 {
434 /* Spin-wait first.
435 Note that spinning first without checking whether a timeout
436 passed might lead to what looks like a spurious wake-up even
437 though we should return ETIMEDOUT (e.g., if the caller provides
438 an absolute timeout that is clearly in the past). However,
439 (1) spurious wake-ups are allowed, (2) it seems unlikely that a
440 user will (ab)use pthread_cond_wait as a check for whether a
441 point in time is in the past, and (3) spinning first without
442 having to compare against the current time seems to be the right
443 choice from a performance perspective for most use cases. */
444 unsigned int spin = maxspin;
445 while (signals == 0 && spin > 0)
446 {
447 /* Check that we are not spinning on a group that's already
448 closed. */
449 if (seq < (__condvar_load_g1_start_relaxed (cond) >> 1))
450 goto done;
451
452 /* TODO Back off. */
453
454 /* Reload signals. See above for MO. */
455 signals = atomic_load_acquire (cond->__data.__g_signals + g);
456 spin--;
457 }
458
459 /* If our group will be closed as indicated by the flag on signals,
460 don't bother grabbing a signal. */
461 if (signals & 1)
462 goto done;
463
464 /* If there is an available signal, don't block. */
465 if (signals != 0)
466 break;
467
468 /* No signals available after spinning, so prepare to block.
469 We first acquire a group reference and use acquire MO for that so
470 that we synchronize with the dummy read-modify-write in
471 __condvar_quiesce_and_switch_g1 if we read from that. In turn,
472 in this case this will make us see the closed flag on __g_signals
473 that designates a concurrent attempt to reuse the group's slot.
474 We use acquire MO for the __g_signals check to make the
475 __g1_start check work (see spinning above).
476 Note that the group reference acquisition will not mask the
477 release MO when decrementing the reference count because we use
478 an atomic read-modify-write operation and thus extend the release
479 sequence. */
480 atomic_fetch_add_acquire (cond->__data.__g_refs + g, 2);
481 if (((atomic_load_acquire (cond->__data.__g_signals + g) & 1) != 0)
482 || (seq < (__condvar_load_g1_start_relaxed (cond) >> 1)))
483 {
484 /* Our group is closed. Wake up any signalers that might be
485 waiting. */
486 __condvar_dec_grefs (cond, g, private);
487 goto done;
488 }
489
490 // Now block.
491 struct _pthread_cleanup_buffer buffer;
492 struct _condvar_cleanup_buffer cbuffer;
493 cbuffer.wseq = wseq;
494 cbuffer.cond = cond;
495 cbuffer.mutex = mutex;
496 cbuffer.private = private;
497 __pthread_cleanup_push (&buffer, __condvar_cleanup_waiting, &cbuffer);
498
499 if (abstime == NULL)
500 {
501 /* Block without a timeout. */
502 err = futex_wait_cancelable (
503 cond->__data.__g_signals + g, 0, private);
504 }
505 else
506 {
507 /* Block, but with a timeout.
508 Work around the fact that the kernel rejects negative timeout
509 values despite them being valid. */
510 if (__glibc_unlikely (abstime->tv_sec < 0))
511 err = ETIMEDOUT;
512
513 else if ((flags & __PTHREAD_COND_CLOCK_MONOTONIC_MASK) != 0)
514 {
515 /* CLOCK_MONOTONIC is requested. */
516 struct timespec rt;
517 if (__clock_gettime (CLOCK_MONOTONIC, &rt) != 0)
518 __libc_fatal ("clock_gettime does not support "
519 "CLOCK_MONOTONIC");
520 /* Convert the absolute timeout value to a relative
521 timeout. */
522 rt.tv_sec = abstime->tv_sec - rt.tv_sec;
523 rt.tv_nsec = abstime->tv_nsec - rt.tv_nsec;
524 if (rt.tv_nsec < 0)
525 {
526 rt.tv_nsec += 1000000000;
527 --rt.tv_sec;
528 }
529 /* Did we already time out? */
530 if (__glibc_unlikely (rt.tv_sec < 0))
531 err = ETIMEDOUT;
532 else
533 err = futex_reltimed_wait_cancelable
534 (cond->__data.__g_signals + g, 0, &rt, private);
535 }
536 else
537 {
538 /* Use CLOCK_REALTIME. */
539 err = futex_abstimed_wait_cancelable
540 (cond->__data.__g_signals + g, 0, abstime, private);
541 }
542 }
543
544 __pthread_cleanup_pop (&buffer, 0);
545
546 if (__glibc_unlikely (err == ETIMEDOUT))
547 {
548 __condvar_dec_grefs (cond, g, private);
549 /* If we timed out, we effectively cancel waiting. Note that
550 we have decremented __g_refs before cancellation, so that a
551 deadlock between waiting for quiescence of our group in
552 __condvar_quiesce_and_switch_g1 and us trying to acquire
553 the lock during cancellation is not possible. */
554 __condvar_cancel_waiting (cond, seq, g, private);
555 result = ETIMEDOUT;
556 goto done;
557 }
558 else
559 __condvar_dec_grefs (cond, g, private);
560
561 /* Reload signals. See above for MO. */
562 signals = atomic_load_acquire (cond->__data.__g_signals + g);
563 }
564
565 }
566 /* Try to grab a signal. Use acquire MO so that we see an up-to-date value
567 of __g1_start below (see spinning above for a similar case). In
568 particular, if we steal from a more recent group, we will also see a
569 more recent __g1_start below. */
570 while (!atomic_compare_exchange_weak_acquire (cond->__data.__g_signals + g,
571 &signals, signals - 2));
572
573 /* We consumed a signal but we could have consumed from a more recent group
574 that aliased with ours due to being in the same group slot. If this
575 might be the case our group must be closed as visible through
576 __g1_start. */
577 uint64_t g1_start = __condvar_load_g1_start_relaxed (cond);
578 if (seq < (g1_start >> 1))
579 {
580 /* We potentially stole a signal from a more recent group but we do not
581 know which group we really consumed from.
582 We do not care about groups older than current G1 because they are
583 closed; we could have stolen from these, but then we just add a
584 spurious wake-up for the current groups.
585 We will never steal a signal from current G2 that was really intended
586 for G2 because G2 never receives signals (until it becomes G1). We
587 could have stolen a signal from G2 that was conservatively added by a
588 previous waiter that also thought it stole a signal -- but given that
589 that signal was added unnecessarily, it's not a problem if we steal
590 it.
591 Thus, the remaining case is that we could have stolen from the current
592 G1, where "current" means the __g1_start value we observed. However,
593 if the current G1 does not have the same slot index as we do, we did
594 not steal from it and do not need to undo that. This is the reason
595 for putting a bit with G2's index into__g1_start as well. */
596 if (((g1_start & 1) ^ 1) == g)
597 {
598 /* We have to conservatively undo our potential mistake of stealing
599 a signal. We can stop trying to do that when the current G1
600 changes because other spinning waiters will notice this too and
601 __condvar_quiesce_and_switch_g1 has checked that there are no
602 futex waiters anymore before switching G1.
603 Relaxed MO is fine for the __g1_start load because we need to
604 merely be able to observe this fact and not have to observe
605 something else as well.
606 ??? Would it help to spin for a little while to see whether the
607 current G1 gets closed? This might be worthwhile if the group is
608 small or close to being closed. */
609 unsigned int s = atomic_load_relaxed (cond->__data.__g_signals + g);
610 while (__condvar_load_g1_start_relaxed (cond) == g1_start)
611 {
612 /* Try to add a signal. We don't need to acquire the lock
613 because at worst we can cause a spurious wake-up. If the
614 group is in the process of being closed (LSB is true), this
615 has an effect similar to us adding a signal. */
616 if (((s & 1) != 0)
617 || atomic_compare_exchange_weak_relaxed
618 (cond->__data.__g_signals + g, &s, s + 2))
619 {
620 /* If we added a signal, we also need to add a wake-up on
621 the futex. We also need to do that if we skipped adding
622 a signal because the group is being closed because
623 while __condvar_quiesce_and_switch_g1 could have closed
624 the group, it might stil be waiting for futex waiters to
625 leave (and one of those waiters might be the one we stole
626 the signal from, which cause it to block using the
627 futex). */
628 futex_wake (cond->__data.__g_signals + g, 1, private);
629 break;
630 }
631 /* TODO Back off. */
632 }
633 }
634 }
635
636 done:
637
638 /* Confirm that we have been woken. We do that before acquiring the mutex
639 to allow for execution of pthread_cond_destroy while having acquired the
640 mutex. */
641 __condvar_confirm_wakeup (cond, private);
642
643 /* Woken up; now re-acquire the mutex. If this doesn't fail, return RESULT,
644 which is set to ETIMEDOUT if a timeout occured, or zero otherwise. */
645 err = __pthread_mutex_cond_lock (mutex);
646 /* XXX Abort on errors that are disallowed by POSIX? */
647 return (err != 0) ? err : result;
648}
649
650
651/* See __pthread_cond_wait_common. */
652int
653__pthread_cond_wait (pthread_cond_t *cond, pthread_mutex_t *mutex)
654{
655 return __pthread_cond_wait_common (cond, mutex, NULL);
656}
657
658/* See __pthread_cond_wait_common. */
659int
660__pthread_cond_timedwait (pthread_cond_t *cond, pthread_mutex_t *mutex,
661 const struct timespec *abstime)
662{
663 /* Check parameter validity. This should also tell the compiler that
664 it can assume that abstime is not NULL. */
665 if (abstime->tv_nsec < 0 || abstime->tv_nsec >= 1000000000)
666 return EINVAL;
667 return __pthread_cond_wait_common (cond, mutex, abstime);
668}
669
670versioned_symbol (libpthread, __pthread_cond_wait, pthread_cond_wait,
671 GLIBC_2_3_2);
672versioned_symbol (libpthread, __pthread_cond_timedwait, pthread_cond_timedwait,
673 GLIBC_2_3_2);
674