Merge tag 'driver-core-4.9-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git...
[cascardo/linux.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
14 void __init_waitqueue_head(wait_queue_head_t *q, const char *name, struct lock_class_key *key)
15 {
16         spin_lock_init(&q->lock);
17         lockdep_set_class_and_name(&q->lock, key, name);
18         INIT_LIST_HEAD(&q->task_list);
19 }
20
21 EXPORT_SYMBOL(__init_waitqueue_head);
22
23 void add_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
24 {
25         unsigned long flags;
26
27         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
28         spin_lock_irqsave(&q->lock, flags);
29         __add_wait_queue(q, wait);
30         spin_unlock_irqrestore(&q->lock, flags);
31 }
32 EXPORT_SYMBOL(add_wait_queue);
33
34 void add_wait_queue_exclusive(wait_queue_head_t *q, wait_queue_t *wait)
35 {
36         unsigned long flags;
37
38         wait->flags |= WQ_FLAG_EXCLUSIVE;
39         spin_lock_irqsave(&q->lock, flags);
40         __add_wait_queue_tail(q, wait);
41         spin_unlock_irqrestore(&q->lock, flags);
42 }
43 EXPORT_SYMBOL(add_wait_queue_exclusive);
44
45 void remove_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
46 {
47         unsigned long flags;
48
49         spin_lock_irqsave(&q->lock, flags);
50         __remove_wait_queue(q, wait);
51         spin_unlock_irqrestore(&q->lock, flags);
52 }
53 EXPORT_SYMBOL(remove_wait_queue);
54
55
56 /*
57  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
58  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
59  * number) then we wake all the non-exclusive tasks and one exclusive task.
60  *
61  * There are circumstances in which we can try to wake a task which has already
62  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
63  * zero in this (rare) case, and we handle it by continuing to scan the queue.
64  */
65 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
66                         int nr_exclusive, int wake_flags, void *key)
67 {
68         wait_queue_t *curr, *next;
69
70         list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
71                 unsigned flags = curr->flags;
72
73                 if (curr->func(curr, mode, wake_flags, key) &&
74                                 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
75                         break;
76         }
77 }
78
79 /**
80  * __wake_up - wake up threads blocked on a waitqueue.
81  * @q: the waitqueue
82  * @mode: which threads
83  * @nr_exclusive: how many wake-one or wake-many threads to wake up
84  * @key: is directly passed to the wakeup function
85  *
86  * It may be assumed that this function implies a write memory barrier before
87  * changing the task state if and only if any tasks are woken up.
88  */
89 void __wake_up(wait_queue_head_t *q, unsigned int mode,
90                         int nr_exclusive, void *key)
91 {
92         unsigned long flags;
93
94         spin_lock_irqsave(&q->lock, flags);
95         __wake_up_common(q, mode, nr_exclusive, 0, key);
96         spin_unlock_irqrestore(&q->lock, flags);
97 }
98 EXPORT_SYMBOL(__wake_up);
99
100 /*
101  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
102  */
103 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
104 {
105         __wake_up_common(q, mode, nr, 0, NULL);
106 }
107 EXPORT_SYMBOL_GPL(__wake_up_locked);
108
109 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
110 {
111         __wake_up_common(q, mode, 1, 0, key);
112 }
113 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
114
115 /**
116  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
117  * @q: the waitqueue
118  * @mode: which threads
119  * @nr_exclusive: how many wake-one or wake-many threads to wake up
120  * @key: opaque value to be passed to wakeup targets
121  *
122  * The sync wakeup differs that the waker knows that it will schedule
123  * away soon, so while the target thread will be woken up, it will not
124  * be migrated to another CPU - ie. the two threads are 'synchronized'
125  * with each other. This can prevent needless bouncing between CPUs.
126  *
127  * On UP it can prevent extra preemption.
128  *
129  * It may be assumed that this function implies a write memory barrier before
130  * changing the task state if and only if any tasks are woken up.
131  */
132 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
133                         int nr_exclusive, void *key)
134 {
135         unsigned long flags;
136         int wake_flags = 1; /* XXX WF_SYNC */
137
138         if (unlikely(!q))
139                 return;
140
141         if (unlikely(nr_exclusive != 1))
142                 wake_flags = 0;
143
144         spin_lock_irqsave(&q->lock, flags);
145         __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
146         spin_unlock_irqrestore(&q->lock, flags);
147 }
148 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
149
150 /*
151  * __wake_up_sync - see __wake_up_sync_key()
152  */
153 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
154 {
155         __wake_up_sync_key(q, mode, nr_exclusive, NULL);
156 }
157 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
158
159 /*
160  * Note: we use "set_current_state()" _after_ the wait-queue add,
161  * because we need a memory barrier there on SMP, so that any
162  * wake-function that tests for the wait-queue being active
163  * will be guaranteed to see waitqueue addition _or_ subsequent
164  * tests in this thread will see the wakeup having taken place.
165  *
166  * The spin_unlock() itself is semi-permeable and only protects
167  * one way (it only protects stuff inside the critical region and
168  * stops them from bleeding out - it would still allow subsequent
169  * loads to move into the critical region).
170  */
171 void
172 prepare_to_wait(wait_queue_head_t *q, wait_queue_t *wait, int state)
173 {
174         unsigned long flags;
175
176         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
177         spin_lock_irqsave(&q->lock, flags);
178         if (list_empty(&wait->task_list))
179                 __add_wait_queue(q, wait);
180         set_current_state(state);
181         spin_unlock_irqrestore(&q->lock, flags);
182 }
183 EXPORT_SYMBOL(prepare_to_wait);
184
185 void
186 prepare_to_wait_exclusive(wait_queue_head_t *q, wait_queue_t *wait, int state)
187 {
188         unsigned long flags;
189
190         wait->flags |= WQ_FLAG_EXCLUSIVE;
191         spin_lock_irqsave(&q->lock, flags);
192         if (list_empty(&wait->task_list))
193                 __add_wait_queue_tail(q, wait);
194         set_current_state(state);
195         spin_unlock_irqrestore(&q->lock, flags);
196 }
197 EXPORT_SYMBOL(prepare_to_wait_exclusive);
198
199 void init_wait_entry(wait_queue_t *wait, int flags)
200 {
201         wait->flags = flags;
202         wait->private = current;
203         wait->func = autoremove_wake_function;
204         INIT_LIST_HEAD(&wait->task_list);
205 }
206 EXPORT_SYMBOL(init_wait_entry);
207
208 long prepare_to_wait_event(wait_queue_head_t *q, wait_queue_t *wait, int state)
209 {
210         unsigned long flags;
211         long ret = 0;
212
213         spin_lock_irqsave(&q->lock, flags);
214         if (unlikely(signal_pending_state(state, current))) {
215                 /*
216                  * Exclusive waiter must not fail if it was selected by wakeup,
217                  * it should "consume" the condition we were waiting for.
218                  *
219                  * The caller will recheck the condition and return success if
220                  * we were already woken up, we can not miss the event because
221                  * wakeup locks/unlocks the same q->lock.
222                  *
223                  * But we need to ensure that set-condition + wakeup after that
224                  * can't see us, it should wake up another exclusive waiter if
225                  * we fail.
226                  */
227                 list_del_init(&wait->task_list);
228                 ret = -ERESTARTSYS;
229         } else {
230                 if (list_empty(&wait->task_list)) {
231                         if (wait->flags & WQ_FLAG_EXCLUSIVE)
232                                 __add_wait_queue_tail(q, wait);
233                         else
234                                 __add_wait_queue(q, wait);
235                 }
236                 set_current_state(state);
237         }
238         spin_unlock_irqrestore(&q->lock, flags);
239
240         return ret;
241 }
242 EXPORT_SYMBOL(prepare_to_wait_event);
243
244 /**
245  * finish_wait - clean up after waiting in a queue
246  * @q: waitqueue waited on
247  * @wait: wait descriptor
248  *
249  * Sets current thread back to running state and removes
250  * the wait descriptor from the given waitqueue if still
251  * queued.
252  */
253 void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
254 {
255         unsigned long flags;
256
257         __set_current_state(TASK_RUNNING);
258         /*
259          * We can check for list emptiness outside the lock
260          * IFF:
261          *  - we use the "careful" check that verifies both
262          *    the next and prev pointers, so that there cannot
263          *    be any half-pending updates in progress on other
264          *    CPU's that we haven't seen yet (and that might
265          *    still change the stack area.
266          * and
267          *  - all other users take the lock (ie we can only
268          *    have _one_ other CPU that looks at or modifies
269          *    the list).
270          */
271         if (!list_empty_careful(&wait->task_list)) {
272                 spin_lock_irqsave(&q->lock, flags);
273                 list_del_init(&wait->task_list);
274                 spin_unlock_irqrestore(&q->lock, flags);
275         }
276 }
277 EXPORT_SYMBOL(finish_wait);
278
279 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
280 {
281         int ret = default_wake_function(wait, mode, sync, key);
282
283         if (ret)
284                 list_del_init(&wait->task_list);
285         return ret;
286 }
287 EXPORT_SYMBOL(autoremove_wake_function);
288
289 static inline bool is_kthread_should_stop(void)
290 {
291         return (current->flags & PF_KTHREAD) && kthread_should_stop();
292 }
293
294 /*
295  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
296  *
297  * add_wait_queue(&wq, &wait);
298  * for (;;) {
299  *     if (condition)
300  *         break;
301  *
302  *     p->state = mode;                         condition = true;
303  *     smp_mb(); // A                           smp_wmb(); // C
304  *     if (!wait->flags & WQ_FLAG_WOKEN)        wait->flags |= WQ_FLAG_WOKEN;
305  *         schedule()                           try_to_wake_up();
306  *     p->state = TASK_RUNNING;             ~~~~~~~~~~~~~~~~~~
307  *     wait->flags &= ~WQ_FLAG_WOKEN;           condition = true;
308  *     smp_mb() // B                            smp_wmb(); // C
309  *                                              wait->flags |= WQ_FLAG_WOKEN;
310  * }
311  * remove_wait_queue(&wq, &wait);
312  *
313  */
314 long wait_woken(wait_queue_t *wait, unsigned mode, long timeout)
315 {
316         set_current_state(mode); /* A */
317         /*
318          * The above implies an smp_mb(), which matches with the smp_wmb() from
319          * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
320          * also observe all state before the wakeup.
321          */
322         if (!(wait->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
323                 timeout = schedule_timeout(timeout);
324         __set_current_state(TASK_RUNNING);
325
326         /*
327          * The below implies an smp_mb(), it too pairs with the smp_wmb() from
328          * woken_wake_function() such that we must either observe the wait
329          * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
330          * an event.
331          */
332         smp_store_mb(wait->flags, wait->flags & ~WQ_FLAG_WOKEN); /* B */
333
334         return timeout;
335 }
336 EXPORT_SYMBOL(wait_woken);
337
338 int woken_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
339 {
340         /*
341          * Although this function is called under waitqueue lock, LOCK
342          * doesn't imply write barrier and the users expects write
343          * barrier semantics on wakeup functions.  The following
344          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
345          * and is paired with smp_store_mb() in wait_woken().
346          */
347         smp_wmb(); /* C */
348         wait->flags |= WQ_FLAG_WOKEN;
349
350         return default_wake_function(wait, mode, sync, key);
351 }
352 EXPORT_SYMBOL(woken_wake_function);
353
354 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
355 {
356         struct wait_bit_key *key = arg;
357         struct wait_bit_queue *wait_bit
358                 = container_of(wait, struct wait_bit_queue, wait);
359
360         if (wait_bit->key.flags != key->flags ||
361                         wait_bit->key.bit_nr != key->bit_nr ||
362                         test_bit(key->bit_nr, key->flags))
363                 return 0;
364         else
365                 return autoremove_wake_function(wait, mode, sync, key);
366 }
367 EXPORT_SYMBOL(wake_bit_function);
368
369 /*
370  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
371  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
372  * permitted return codes. Nonzero return codes halt waiting and return.
373  */
374 int __sched
375 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
376               wait_bit_action_f *action, unsigned mode)
377 {
378         int ret = 0;
379
380         do {
381                 prepare_to_wait(wq, &q->wait, mode);
382                 if (test_bit(q->key.bit_nr, q->key.flags))
383                         ret = (*action)(&q->key, mode);
384         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
385         finish_wait(wq, &q->wait);
386         return ret;
387 }
388 EXPORT_SYMBOL(__wait_on_bit);
389
390 int __sched out_of_line_wait_on_bit(void *word, int bit,
391                                     wait_bit_action_f *action, unsigned mode)
392 {
393         wait_queue_head_t *wq = bit_waitqueue(word, bit);
394         DEFINE_WAIT_BIT(wait, word, bit);
395
396         return __wait_on_bit(wq, &wait, action, mode);
397 }
398 EXPORT_SYMBOL(out_of_line_wait_on_bit);
399
400 int __sched out_of_line_wait_on_bit_timeout(
401         void *word, int bit, wait_bit_action_f *action,
402         unsigned mode, unsigned long timeout)
403 {
404         wait_queue_head_t *wq = bit_waitqueue(word, bit);
405         DEFINE_WAIT_BIT(wait, word, bit);
406
407         wait.key.timeout = jiffies + timeout;
408         return __wait_on_bit(wq, &wait, action, mode);
409 }
410 EXPORT_SYMBOL_GPL(out_of_line_wait_on_bit_timeout);
411
412 int __sched
413 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
414                         wait_bit_action_f *action, unsigned mode)
415 {
416         int ret = 0;
417
418         for (;;) {
419                 prepare_to_wait_exclusive(wq, &q->wait, mode);
420                 if (test_bit(q->key.bit_nr, q->key.flags)) {
421                         ret = action(&q->key, mode);
422                         /*
423                          * See the comment in prepare_to_wait_event().
424                          * finish_wait() does not necessarily takes wq->lock,
425                          * but test_and_set_bit() implies mb() which pairs with
426                          * smp_mb__after_atomic() before wake_up_page().
427                          */
428                         if (ret)
429                                 finish_wait(wq, &q->wait);
430                 }
431                 if (!test_and_set_bit(q->key.bit_nr, q->key.flags)) {
432                         if (!ret)
433                                 finish_wait(wq, &q->wait);
434                         return 0;
435                 } else if (ret) {
436                         return ret;
437                 }
438         }
439 }
440 EXPORT_SYMBOL(__wait_on_bit_lock);
441
442 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
443                                          wait_bit_action_f *action, unsigned mode)
444 {
445         wait_queue_head_t *wq = bit_waitqueue(word, bit);
446         DEFINE_WAIT_BIT(wait, word, bit);
447
448         return __wait_on_bit_lock(wq, &wait, action, mode);
449 }
450 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
451
452 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
453 {
454         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
455         if (waitqueue_active(wq))
456                 __wake_up(wq, TASK_NORMAL, 1, &key);
457 }
458 EXPORT_SYMBOL(__wake_up_bit);
459
460 /**
461  * wake_up_bit - wake up a waiter on a bit
462  * @word: the word being waited on, a kernel virtual address
463  * @bit: the bit of the word being waited on
464  *
465  * There is a standard hashed waitqueue table for generic use. This
466  * is the part of the hashtable's accessor API that wakes up waiters
467  * on a bit. For instance, if one were to have waiters on a bitflag,
468  * one would call wake_up_bit() after clearing the bit.
469  *
470  * In order for this to function properly, as it uses waitqueue_active()
471  * internally, some kind of memory barrier must be done prior to calling
472  * this. Typically, this will be smp_mb__after_atomic(), but in some
473  * cases where bitflags are manipulated non-atomically under a lock, one
474  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
475  * because spin_unlock() does not guarantee a memory barrier.
476  */
477 void wake_up_bit(void *word, int bit)
478 {
479         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
480 }
481 EXPORT_SYMBOL(wake_up_bit);
482
483 /*
484  * Manipulate the atomic_t address to produce a better bit waitqueue table hash
485  * index (we're keying off bit -1, but that would produce a horrible hash
486  * value).
487  */
488 static inline wait_queue_head_t *atomic_t_waitqueue(atomic_t *p)
489 {
490         if (BITS_PER_LONG == 64) {
491                 unsigned long q = (unsigned long)p;
492                 return bit_waitqueue((void *)(q & ~1), q & 1);
493         }
494         return bit_waitqueue(p, 0);
495 }
496
497 static int wake_atomic_t_function(wait_queue_t *wait, unsigned mode, int sync,
498                                   void *arg)
499 {
500         struct wait_bit_key *key = arg;
501         struct wait_bit_queue *wait_bit
502                 = container_of(wait, struct wait_bit_queue, wait);
503         atomic_t *val = key->flags;
504
505         if (wait_bit->key.flags != key->flags ||
506             wait_bit->key.bit_nr != key->bit_nr ||
507             atomic_read(val) != 0)
508                 return 0;
509         return autoremove_wake_function(wait, mode, sync, key);
510 }
511
512 /*
513  * To allow interruptible waiting and asynchronous (i.e. nonblocking) waiting,
514  * the actions of __wait_on_atomic_t() are permitted return codes.  Nonzero
515  * return codes halt waiting and return.
516  */
517 static __sched
518 int __wait_on_atomic_t(wait_queue_head_t *wq, struct wait_bit_queue *q,
519                        int (*action)(atomic_t *), unsigned mode)
520 {
521         atomic_t *val;
522         int ret = 0;
523
524         do {
525                 prepare_to_wait(wq, &q->wait, mode);
526                 val = q->key.flags;
527                 if (atomic_read(val) == 0)
528                         break;
529                 ret = (*action)(val);
530         } while (!ret && atomic_read(val) != 0);
531         finish_wait(wq, &q->wait);
532         return ret;
533 }
534
535 #define DEFINE_WAIT_ATOMIC_T(name, p)                                   \
536         struct wait_bit_queue name = {                                  \
537                 .key = __WAIT_ATOMIC_T_KEY_INITIALIZER(p),              \
538                 .wait   = {                                             \
539                         .private        = current,                      \
540                         .func           = wake_atomic_t_function,       \
541                         .task_list      =                               \
542                                 LIST_HEAD_INIT((name).wait.task_list),  \
543                 },                                                      \
544         }
545
546 __sched int out_of_line_wait_on_atomic_t(atomic_t *p, int (*action)(atomic_t *),
547                                          unsigned mode)
548 {
549         wait_queue_head_t *wq = atomic_t_waitqueue(p);
550         DEFINE_WAIT_ATOMIC_T(wait, p);
551
552         return __wait_on_atomic_t(wq, &wait, action, mode);
553 }
554 EXPORT_SYMBOL(out_of_line_wait_on_atomic_t);
555
556 /**
557  * wake_up_atomic_t - Wake up a waiter on a atomic_t
558  * @p: The atomic_t being waited on, a kernel virtual address
559  *
560  * Wake up anyone waiting for the atomic_t to go to zero.
561  *
562  * Abuse the bit-waker function and its waitqueue hash table set (the atomic_t
563  * check is done by the waiter's wake function, not the by the waker itself).
564  */
565 void wake_up_atomic_t(atomic_t *p)
566 {
567         __wake_up_bit(atomic_t_waitqueue(p), p, WAIT_ATOMIC_T_BIT_NR);
568 }
569 EXPORT_SYMBOL(wake_up_atomic_t);
570
571 __sched int bit_wait(struct wait_bit_key *word, int mode)
572 {
573         schedule();
574         if (signal_pending_state(mode, current))
575                 return -EINTR;
576         return 0;
577 }
578 EXPORT_SYMBOL(bit_wait);
579
580 __sched int bit_wait_io(struct wait_bit_key *word, int mode)
581 {
582         io_schedule();
583         if (signal_pending_state(mode, current))
584                 return -EINTR;
585         return 0;
586 }
587 EXPORT_SYMBOL(bit_wait_io);
588
589 __sched int bit_wait_timeout(struct wait_bit_key *word, int mode)
590 {
591         unsigned long now = READ_ONCE(jiffies);
592         if (time_after_eq(now, word->timeout))
593                 return -EAGAIN;
594         schedule_timeout(word->timeout - now);
595         if (signal_pending_state(mode, current))
596                 return -EINTR;
597         return 0;
598 }
599 EXPORT_SYMBOL_GPL(bit_wait_timeout);
600
601 __sched int bit_wait_io_timeout(struct wait_bit_key *word, int mode)
602 {
603         unsigned long now = READ_ONCE(jiffies);
604         if (time_after_eq(now, word->timeout))
605                 return -EAGAIN;
606         io_schedule_timeout(word->timeout - now);
607         if (signal_pending_state(mode, current))
608                 return -EINTR;
609         return 0;
610 }
611 EXPORT_SYMBOL_GPL(bit_wait_io_timeout);