GNU Linux-libre 4.9.309-gnu1
[releases.git] / kernel / sched / wait.c
1 /*
2  * Generic waiting primitives.
3  *
4  * (C) 2004 Nadia Yvette Chambers, Oracle
5  */
6 #include <linux/init.h>
7 #include <linux/export.h>
8 #include <linux/sched.h>
9 #include <linux/mm.h>
10 #include <linux/wait.h>
11 #include <linux/hash.h>
12 #include <linux/kthread.h>
13 #include <linux/poll.h>
14
15 void __init_waitqueue_head(wait_queue_head_t *q, const char *name, struct lock_class_key *key)
16 {
17         spin_lock_init(&q->lock);
18         lockdep_set_class_and_name(&q->lock, key, name);
19         INIT_LIST_HEAD(&q->task_list);
20 }
21
22 EXPORT_SYMBOL(__init_waitqueue_head);
23
24 void add_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
25 {
26         unsigned long flags;
27
28         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
29         spin_lock_irqsave(&q->lock, flags);
30         __add_wait_queue(q, wait);
31         spin_unlock_irqrestore(&q->lock, flags);
32 }
33 EXPORT_SYMBOL(add_wait_queue);
34
35 void add_wait_queue_exclusive(wait_queue_head_t *q, wait_queue_t *wait)
36 {
37         unsigned long flags;
38
39         wait->flags |= WQ_FLAG_EXCLUSIVE;
40         spin_lock_irqsave(&q->lock, flags);
41         __add_wait_queue_tail(q, wait);
42         spin_unlock_irqrestore(&q->lock, flags);
43 }
44 EXPORT_SYMBOL(add_wait_queue_exclusive);
45
46 void remove_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
47 {
48         unsigned long flags;
49
50         spin_lock_irqsave(&q->lock, flags);
51         __remove_wait_queue(q, wait);
52         spin_unlock_irqrestore(&q->lock, flags);
53 }
54 EXPORT_SYMBOL(remove_wait_queue);
55
56
57 /*
58  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
59  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
60  * number) then we wake all the non-exclusive tasks and one exclusive task.
61  *
62  * There are circumstances in which we can try to wake a task which has already
63  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
64  * zero in this (rare) case, and we handle it by continuing to scan the queue.
65  */
66 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
67                         int nr_exclusive, int wake_flags, void *key)
68 {
69         wait_queue_t *curr, *next;
70
71         list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
72                 unsigned flags = curr->flags;
73
74                 if (curr->func(curr, mode, wake_flags, key) &&
75                                 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
76                         break;
77         }
78 }
79
80 /**
81  * __wake_up - wake up threads blocked on a waitqueue.
82  * @q: the waitqueue
83  * @mode: which threads
84  * @nr_exclusive: how many wake-one or wake-many threads to wake up
85  * @key: is directly passed to the wakeup function
86  *
87  * It may be assumed that this function implies a write memory barrier before
88  * changing the task state if and only if any tasks are woken up.
89  */
90 void __wake_up(wait_queue_head_t *q, unsigned int mode,
91                         int nr_exclusive, void *key)
92 {
93         unsigned long flags;
94
95         spin_lock_irqsave(&q->lock, flags);
96         __wake_up_common(q, mode, nr_exclusive, 0, key);
97         spin_unlock_irqrestore(&q->lock, flags);
98 }
99 EXPORT_SYMBOL(__wake_up);
100
101 /*
102  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
103  */
104 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
105 {
106         __wake_up_common(q, mode, nr, 0, NULL);
107 }
108 EXPORT_SYMBOL_GPL(__wake_up_locked);
109
110 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
111 {
112         __wake_up_common(q, mode, 1, 0, key);
113 }
114 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
115
116 /**
117  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
118  * @q: the waitqueue
119  * @mode: which threads
120  * @nr_exclusive: how many wake-one or wake-many threads to wake up
121  * @key: opaque value to be passed to wakeup targets
122  *
123  * The sync wakeup differs that the waker knows that it will schedule
124  * away soon, so while the target thread will be woken up, it will not
125  * be migrated to another CPU - ie. the two threads are 'synchronized'
126  * with each other. This can prevent needless bouncing between CPUs.
127  *
128  * On UP it can prevent extra preemption.
129  *
130  * It may be assumed that this function implies a write memory barrier before
131  * changing the task state if and only if any tasks are woken up.
132  */
133 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
134                         int nr_exclusive, void *key)
135 {
136         unsigned long flags;
137         int wake_flags = 1; /* XXX WF_SYNC */
138
139         if (unlikely(!q))
140                 return;
141
142         if (unlikely(nr_exclusive != 1))
143                 wake_flags = 0;
144
145         spin_lock_irqsave(&q->lock, flags);
146         __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
147         spin_unlock_irqrestore(&q->lock, flags);
148 }
149 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
150
151 /*
152  * __wake_up_sync - see __wake_up_sync_key()
153  */
154 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
155 {
156         __wake_up_sync_key(q, mode, nr_exclusive, NULL);
157 }
158 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
159
160 void __wake_up_pollfree(wait_queue_head_t *wq_head)
161 {
162         __wake_up(wq_head, TASK_NORMAL, 0, (void *)(POLLHUP | POLLFREE));
163         /* POLLFREE must have cleared the queue. */
164         WARN_ON_ONCE(waitqueue_active(wq_head));
165 }
166
167 /*
168  * Note: we use "set_current_state()" _after_ the wait-queue add,
169  * because we need a memory barrier there on SMP, so that any
170  * wake-function that tests for the wait-queue being active
171  * will be guaranteed to see waitqueue addition _or_ subsequent
172  * tests in this thread will see the wakeup having taken place.
173  *
174  * The spin_unlock() itself is semi-permeable and only protects
175  * one way (it only protects stuff inside the critical region and
176  * stops them from bleeding out - it would still allow subsequent
177  * loads to move into the critical region).
178  */
179 void
180 prepare_to_wait(wait_queue_head_t *q, wait_queue_t *wait, int state)
181 {
182         unsigned long flags;
183
184         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
185         spin_lock_irqsave(&q->lock, flags);
186         if (list_empty(&wait->task_list))
187                 __add_wait_queue(q, wait);
188         set_current_state(state);
189         spin_unlock_irqrestore(&q->lock, flags);
190 }
191 EXPORT_SYMBOL(prepare_to_wait);
192
193 void
194 prepare_to_wait_exclusive(wait_queue_head_t *q, wait_queue_t *wait, int state)
195 {
196         unsigned long flags;
197
198         wait->flags |= WQ_FLAG_EXCLUSIVE;
199         spin_lock_irqsave(&q->lock, flags);
200         if (list_empty(&wait->task_list))
201                 __add_wait_queue_tail(q, wait);
202         set_current_state(state);
203         spin_unlock_irqrestore(&q->lock, flags);
204 }
205 EXPORT_SYMBOL(prepare_to_wait_exclusive);
206
207 void init_wait_entry(wait_queue_t *wait, int flags)
208 {
209         wait->flags = flags;
210         wait->private = current;
211         wait->func = autoremove_wake_function;
212         INIT_LIST_HEAD(&wait->task_list);
213 }
214 EXPORT_SYMBOL(init_wait_entry);
215
216 long prepare_to_wait_event(wait_queue_head_t *q, wait_queue_t *wait, int state)
217 {
218         unsigned long flags;
219         long ret = 0;
220
221         spin_lock_irqsave(&q->lock, flags);
222         if (unlikely(signal_pending_state(state, current))) {
223                 /*
224                  * Exclusive waiter must not fail if it was selected by wakeup,
225                  * it should "consume" the condition we were waiting for.
226                  *
227                  * The caller will recheck the condition and return success if
228                  * we were already woken up, we can not miss the event because
229                  * wakeup locks/unlocks the same q->lock.
230                  *
231                  * But we need to ensure that set-condition + wakeup after that
232                  * can't see us, it should wake up another exclusive waiter if
233                  * we fail.
234                  */
235                 list_del_init(&wait->task_list);
236                 ret = -ERESTARTSYS;
237         } else {
238                 if (list_empty(&wait->task_list)) {
239                         if (wait->flags & WQ_FLAG_EXCLUSIVE)
240                                 __add_wait_queue_tail(q, wait);
241                         else
242                                 __add_wait_queue(q, wait);
243                 }
244                 set_current_state(state);
245         }
246         spin_unlock_irqrestore(&q->lock, flags);
247
248         return ret;
249 }
250 EXPORT_SYMBOL(prepare_to_wait_event);
251
252 /**
253  * finish_wait - clean up after waiting in a queue
254  * @q: waitqueue waited on
255  * @wait: wait descriptor
256  *
257  * Sets current thread back to running state and removes
258  * the wait descriptor from the given waitqueue if still
259  * queued.
260  */
261 void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
262 {
263         unsigned long flags;
264
265         __set_current_state(TASK_RUNNING);
266         /*
267          * We can check for list emptiness outside the lock
268          * IFF:
269          *  - we use the "careful" check that verifies both
270          *    the next and prev pointers, so that there cannot
271          *    be any half-pending updates in progress on other
272          *    CPU's that we haven't seen yet (and that might
273          *    still change the stack area.
274          * and
275          *  - all other users take the lock (ie we can only
276          *    have _one_ other CPU that looks at or modifies
277          *    the list).
278          */
279         if (!list_empty_careful(&wait->task_list)) {
280                 spin_lock_irqsave(&q->lock, flags);
281                 list_del_init(&wait->task_list);
282                 spin_unlock_irqrestore(&q->lock, flags);
283         }
284 }
285 EXPORT_SYMBOL(finish_wait);
286
287 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
288 {
289         int ret = default_wake_function(wait, mode, sync, key);
290
291         if (ret)
292                 list_del_init(&wait->task_list);
293         return ret;
294 }
295 EXPORT_SYMBOL(autoremove_wake_function);
296
297 static inline bool is_kthread_should_stop(void)
298 {
299         return (current->flags & PF_KTHREAD) && kthread_should_stop();
300 }
301
302 /*
303  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
304  *
305  * add_wait_queue(&wq, &wait);
306  * for (;;) {
307  *     if (condition)
308  *         break;
309  *
310  *     p->state = mode;                         condition = true;
311  *     smp_mb(); // A                           smp_wmb(); // C
312  *     if (!wait->flags & WQ_FLAG_WOKEN)        wait->flags |= WQ_FLAG_WOKEN;
313  *         schedule()                           try_to_wake_up();
314  *     p->state = TASK_RUNNING;             ~~~~~~~~~~~~~~~~~~
315  *     wait->flags &= ~WQ_FLAG_WOKEN;           condition = true;
316  *     smp_mb() // B                            smp_wmb(); // C
317  *                                              wait->flags |= WQ_FLAG_WOKEN;
318  * }
319  * remove_wait_queue(&wq, &wait);
320  *
321  */
322 long wait_woken(wait_queue_t *wait, unsigned mode, long timeout)
323 {
324         set_current_state(mode); /* A */
325         /*
326          * The above implies an smp_mb(), which matches with the smp_wmb() from
327          * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
328          * also observe all state before the wakeup.
329          */
330         if (!(wait->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
331                 timeout = schedule_timeout(timeout);
332         __set_current_state(TASK_RUNNING);
333
334         /*
335          * The below implies an smp_mb(), it too pairs with the smp_wmb() from
336          * woken_wake_function() such that we must either observe the wait
337          * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
338          * an event.
339          */
340         smp_store_mb(wait->flags, wait->flags & ~WQ_FLAG_WOKEN); /* B */
341
342         return timeout;
343 }
344 EXPORT_SYMBOL(wait_woken);
345
346 int woken_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
347 {
348         /*
349          * Although this function is called under waitqueue lock, LOCK
350          * doesn't imply write barrier and the users expects write
351          * barrier semantics on wakeup functions.  The following
352          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
353          * and is paired with smp_store_mb() in wait_woken().
354          */
355         smp_wmb(); /* C */
356         wait->flags |= WQ_FLAG_WOKEN;
357
358         return default_wake_function(wait, mode, sync, key);
359 }
360 EXPORT_SYMBOL(woken_wake_function);
361
362 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
363 {
364         struct wait_bit_key *key = arg;
365         struct wait_bit_queue *wait_bit
366                 = container_of(wait, struct wait_bit_queue, wait);
367
368         if (wait_bit->key.flags != key->flags ||
369                         wait_bit->key.bit_nr != key->bit_nr ||
370                         test_bit(key->bit_nr, key->flags))
371                 return 0;
372         else
373                 return autoremove_wake_function(wait, mode, sync, key);
374 }
375 EXPORT_SYMBOL(wake_bit_function);
376
377 /*
378  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
379  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
380  * permitted return codes. Nonzero return codes halt waiting and return.
381  */
382 int __sched
383 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
384               wait_bit_action_f *action, unsigned mode)
385 {
386         int ret = 0;
387
388         do {
389                 prepare_to_wait(wq, &q->wait, mode);
390                 if (test_bit(q->key.bit_nr, q->key.flags))
391                         ret = (*action)(&q->key, mode);
392         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
393         finish_wait(wq, &q->wait);
394         return ret;
395 }
396 EXPORT_SYMBOL(__wait_on_bit);
397
398 int __sched out_of_line_wait_on_bit(void *word, int bit,
399                                     wait_bit_action_f *action, unsigned mode)
400 {
401         wait_queue_head_t *wq = bit_waitqueue(word, bit);
402         DEFINE_WAIT_BIT(wait, word, bit);
403
404         return __wait_on_bit(wq, &wait, action, mode);
405 }
406 EXPORT_SYMBOL(out_of_line_wait_on_bit);
407
408 int __sched out_of_line_wait_on_bit_timeout(
409         void *word, int bit, wait_bit_action_f *action,
410         unsigned mode, unsigned long timeout)
411 {
412         wait_queue_head_t *wq = bit_waitqueue(word, bit);
413         DEFINE_WAIT_BIT(wait, word, bit);
414
415         wait.key.timeout = jiffies + timeout;
416         return __wait_on_bit(wq, &wait, action, mode);
417 }
418 EXPORT_SYMBOL_GPL(out_of_line_wait_on_bit_timeout);
419
420 int __sched
421 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
422                         wait_bit_action_f *action, unsigned mode)
423 {
424         int ret = 0;
425
426         for (;;) {
427                 prepare_to_wait_exclusive(wq, &q->wait, mode);
428                 if (test_bit(q->key.bit_nr, q->key.flags)) {
429                         ret = action(&q->key, mode);
430                         /*
431                          * See the comment in prepare_to_wait_event().
432                          * finish_wait() does not necessarily takes wq->lock,
433                          * but test_and_set_bit() implies mb() which pairs with
434                          * smp_mb__after_atomic() before wake_up_page().
435                          */
436                         if (ret)
437                                 finish_wait(wq, &q->wait);
438                 }
439                 if (!test_and_set_bit(q->key.bit_nr, q->key.flags)) {
440                         if (!ret)
441                                 finish_wait(wq, &q->wait);
442                         return 0;
443                 } else if (ret) {
444                         return ret;
445                 }
446         }
447 }
448 EXPORT_SYMBOL(__wait_on_bit_lock);
449
450 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
451                                          wait_bit_action_f *action, unsigned mode)
452 {
453         wait_queue_head_t *wq = bit_waitqueue(word, bit);
454         DEFINE_WAIT_BIT(wait, word, bit);
455
456         return __wait_on_bit_lock(wq, &wait, action, mode);
457 }
458 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
459
460 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
461 {
462         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
463         if (waitqueue_active(wq))
464                 __wake_up(wq, TASK_NORMAL, 1, &key);
465 }
466 EXPORT_SYMBOL(__wake_up_bit);
467
468 /**
469  * wake_up_bit - wake up a waiter on a bit
470  * @word: the word being waited on, a kernel virtual address
471  * @bit: the bit of the word being waited on
472  *
473  * There is a standard hashed waitqueue table for generic use. This
474  * is the part of the hashtable's accessor API that wakes up waiters
475  * on a bit. For instance, if one were to have waiters on a bitflag,
476  * one would call wake_up_bit() after clearing the bit.
477  *
478  * In order for this to function properly, as it uses waitqueue_active()
479  * internally, some kind of memory barrier must be done prior to calling
480  * this. Typically, this will be smp_mb__after_atomic(), but in some
481  * cases where bitflags are manipulated non-atomically under a lock, one
482  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
483  * because spin_unlock() does not guarantee a memory barrier.
484  */
485 void wake_up_bit(void *word, int bit)
486 {
487         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
488 }
489 EXPORT_SYMBOL(wake_up_bit);
490
491 /*
492  * Manipulate the atomic_t address to produce a better bit waitqueue table hash
493  * index (we're keying off bit -1, but that would produce a horrible hash
494  * value).
495  */
496 static inline wait_queue_head_t *atomic_t_waitqueue(atomic_t *p)
497 {
498         if (BITS_PER_LONG == 64) {
499                 unsigned long q = (unsigned long)p;
500                 return bit_waitqueue((void *)(q & ~1), q & 1);
501         }
502         return bit_waitqueue(p, 0);
503 }
504
505 static int wake_atomic_t_function(wait_queue_t *wait, unsigned mode, int sync,
506                                   void *arg)
507 {
508         struct wait_bit_key *key = arg;
509         struct wait_bit_queue *wait_bit
510                 = container_of(wait, struct wait_bit_queue, wait);
511         atomic_t *val = key->flags;
512
513         if (wait_bit->key.flags != key->flags ||
514             wait_bit->key.bit_nr != key->bit_nr ||
515             atomic_read(val) != 0)
516                 return 0;
517         return autoremove_wake_function(wait, mode, sync, key);
518 }
519
520 /*
521  * To allow interruptible waiting and asynchronous (i.e. nonblocking) waiting,
522  * the actions of __wait_on_atomic_t() are permitted return codes.  Nonzero
523  * return codes halt waiting and return.
524  */
525 static __sched
526 int __wait_on_atomic_t(wait_queue_head_t *wq, struct wait_bit_queue *q,
527                        int (*action)(atomic_t *), unsigned mode)
528 {
529         atomic_t *val;
530         int ret = 0;
531
532         do {
533                 prepare_to_wait(wq, &q->wait, mode);
534                 val = q->key.flags;
535                 if (atomic_read(val) == 0)
536                         break;
537                 ret = (*action)(val);
538         } while (!ret && atomic_read(val) != 0);
539         finish_wait(wq, &q->wait);
540         return ret;
541 }
542
543 #define DEFINE_WAIT_ATOMIC_T(name, p)                                   \
544         struct wait_bit_queue name = {                                  \
545                 .key = __WAIT_ATOMIC_T_KEY_INITIALIZER(p),              \
546                 .wait   = {                                             \
547                         .private        = current,                      \
548                         .func           = wake_atomic_t_function,       \
549                         .task_list      =                               \
550                                 LIST_HEAD_INIT((name).wait.task_list),  \
551                 },                                                      \
552         }
553
554 __sched int out_of_line_wait_on_atomic_t(atomic_t *p, int (*action)(atomic_t *),
555                                          unsigned mode)
556 {
557         wait_queue_head_t *wq = atomic_t_waitqueue(p);
558         DEFINE_WAIT_ATOMIC_T(wait, p);
559
560         return __wait_on_atomic_t(wq, &wait, action, mode);
561 }
562 EXPORT_SYMBOL(out_of_line_wait_on_atomic_t);
563
564 /**
565  * wake_up_atomic_t - Wake up a waiter on a atomic_t
566  * @p: The atomic_t being waited on, a kernel virtual address
567  *
568  * Wake up anyone waiting for the atomic_t to go to zero.
569  *
570  * Abuse the bit-waker function and its waitqueue hash table set (the atomic_t
571  * check is done by the waiter's wake function, not the by the waker itself).
572  */
573 void wake_up_atomic_t(atomic_t *p)
574 {
575         __wake_up_bit(atomic_t_waitqueue(p), p, WAIT_ATOMIC_T_BIT_NR);
576 }
577 EXPORT_SYMBOL(wake_up_atomic_t);
578
579 __sched int bit_wait(struct wait_bit_key *word, int mode)
580 {
581         schedule();
582         if (signal_pending_state(mode, current))
583                 return -EINTR;
584         return 0;
585 }
586 EXPORT_SYMBOL(bit_wait);
587
588 __sched int bit_wait_io(struct wait_bit_key *word, int mode)
589 {
590         io_schedule();
591         if (signal_pending_state(mode, current))
592                 return -EINTR;
593         return 0;
594 }
595 EXPORT_SYMBOL(bit_wait_io);
596
597 __sched int bit_wait_timeout(struct wait_bit_key *word, int mode)
598 {
599         unsigned long now = READ_ONCE(jiffies);
600         if (time_after_eq(now, word->timeout))
601                 return -EAGAIN;
602         schedule_timeout(word->timeout - now);
603         if (signal_pending_state(mode, current))
604                 return -EINTR;
605         return 0;
606 }
607 EXPORT_SYMBOL_GPL(bit_wait_timeout);
608
609 __sched int bit_wait_io_timeout(struct wait_bit_key *word, int mode)
610 {
611         unsigned long now = READ_ONCE(jiffies);
612         if (time_after_eq(now, word->timeout))
613                 return -EAGAIN;
614         io_schedule_timeout(word->timeout - now);
615         if (signal_pending_state(mode, current))
616                 return -EINTR;
617         return 0;
618 }
619 EXPORT_SYMBOL_GPL(bit_wait_io_timeout);