2bbba0175ab23666268d6aca263ffd116275758c
[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 long prepare_to_wait_event(wait_queue_head_t *q, wait_queue_t *wait, int state)
200 {
201         unsigned long flags;
202
203         if (signal_pending_state(state, current))
204                 return -ERESTARTSYS;
205
206         wait->private = current;
207         wait->func = autoremove_wake_function;
208
209         spin_lock_irqsave(&q->lock, flags);
210         if (list_empty(&wait->task_list)) {
211                 if (wait->flags & WQ_FLAG_EXCLUSIVE)
212                         __add_wait_queue_tail(q, wait);
213                 else
214                         __add_wait_queue(q, wait);
215         }
216         set_current_state(state);
217         spin_unlock_irqrestore(&q->lock, flags);
218
219         return 0;
220 }
221 EXPORT_SYMBOL(prepare_to_wait_event);
222
223 /**
224  * finish_wait - clean up after waiting in a queue
225  * @q: waitqueue waited on
226  * @wait: wait descriptor
227  *
228  * Sets current thread back to running state and removes
229  * the wait descriptor from the given waitqueue if still
230  * queued.
231  */
232 void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
233 {
234         unsigned long flags;
235
236         __set_current_state(TASK_RUNNING);
237         /*
238          * We can check for list emptiness outside the lock
239          * IFF:
240          *  - we use the "careful" check that verifies both
241          *    the next and prev pointers, so that there cannot
242          *    be any half-pending updates in progress on other
243          *    CPU's that we haven't seen yet (and that might
244          *    still change the stack area.
245          * and
246          *  - all other users take the lock (ie we can only
247          *    have _one_ other CPU that looks at or modifies
248          *    the list).
249          */
250         if (!list_empty_careful(&wait->task_list)) {
251                 spin_lock_irqsave(&q->lock, flags);
252                 list_del_init(&wait->task_list);
253                 spin_unlock_irqrestore(&q->lock, flags);
254         }
255 }
256 EXPORT_SYMBOL(finish_wait);
257
258 /**
259  * abort_exclusive_wait - abort exclusive waiting in a queue
260  * @q: waitqueue waited on
261  * @wait: wait descriptor
262  * @key: key to identify a wait bit queue or %NULL
263  *
264  * Sets current thread back to running state and removes
265  * the wait descriptor from the given waitqueue if still
266  * queued.
267  *
268  * Wakes up the next waiter if the caller is concurrently
269  * woken up through the queue.
270  *
271  * This prevents waiter starvation where an exclusive waiter
272  * aborts and is woken up concurrently and no one wakes up
273  * the next waiter.
274  */
275 void abort_exclusive_wait(wait_queue_head_t *q, wait_queue_t *wait, void *key)
276 {
277         unsigned long flags;
278
279         __set_current_state(TASK_RUNNING);
280         spin_lock_irqsave(&q->lock, flags);
281         if (!list_empty(&wait->task_list))
282                 list_del_init(&wait->task_list);
283         else if (waitqueue_active(q))
284                 __wake_up_locked_key(q, TASK_NORMAL, key);
285         spin_unlock_irqrestore(&q->lock, flags);
286 }
287 EXPORT_SYMBOL(abort_exclusive_wait);
288
289 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
290 {
291         int ret = default_wake_function(wait, mode, sync, key);
292
293         if (ret)
294                 list_del_init(&wait->task_list);
295         return ret;
296 }
297 EXPORT_SYMBOL(autoremove_wake_function);
298
299 static inline bool is_kthread_should_stop(void)
300 {
301         return (current->flags & PF_KTHREAD) && kthread_should_stop();
302 }
303
304 /*
305  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
306  *
307  * add_wait_queue(&wq, &wait);
308  * for (;;) {
309  *     if (condition)
310  *         break;
311  *
312  *     p->state = mode;                         condition = true;
313  *     smp_mb(); // A                           smp_wmb(); // C
314  *     if (!wait->flags & WQ_FLAG_WOKEN)        wait->flags |= WQ_FLAG_WOKEN;
315  *         schedule()                           try_to_wake_up();
316  *     p->state = TASK_RUNNING;             ~~~~~~~~~~~~~~~~~~
317  *     wait->flags &= ~WQ_FLAG_WOKEN;           condition = true;
318  *     smp_mb() // B                            smp_wmb(); // C
319  *                                              wait->flags |= WQ_FLAG_WOKEN;
320  * }
321  * remove_wait_queue(&wq, &wait);
322  *
323  */
324 long wait_woken(wait_queue_t *wait, unsigned mode, long timeout)
325 {
326         set_current_state(mode); /* A */
327         /*
328          * The above implies an smp_mb(), which matches with the smp_wmb() from
329          * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
330          * also observe all state before the wakeup.
331          */
332         if (!(wait->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
333                 timeout = schedule_timeout(timeout);
334         __set_current_state(TASK_RUNNING);
335
336         /*
337          * The below implies an smp_mb(), it too pairs with the smp_wmb() from
338          * woken_wake_function() such that we must either observe the wait
339          * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
340          * an event.
341          */
342         smp_store_mb(wait->flags, wait->flags & ~WQ_FLAG_WOKEN); /* B */
343
344         return timeout;
345 }
346 EXPORT_SYMBOL(wait_woken);
347
348 int woken_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
349 {
350         /*
351          * Although this function is called under waitqueue lock, LOCK
352          * doesn't imply write barrier and the users expects write
353          * barrier semantics on wakeup functions.  The following
354          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
355          * and is paired with smp_store_mb() in wait_woken().
356          */
357         smp_wmb(); /* C */
358         wait->flags |= WQ_FLAG_WOKEN;
359
360         return default_wake_function(wait, mode, sync, key);
361 }
362 EXPORT_SYMBOL(woken_wake_function);
363
364 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
365 {
366         struct wait_bit_key *key = arg;
367         struct wait_bit_queue *wait_bit
368                 = container_of(wait, struct wait_bit_queue, wait);
369
370         if (wait_bit->key.flags != key->flags ||
371                         wait_bit->key.bit_nr != key->bit_nr ||
372                         test_bit(key->bit_nr, key->flags))
373                 return 0;
374         else
375                 return autoremove_wake_function(wait, mode, sync, key);
376 }
377 EXPORT_SYMBOL(wake_bit_function);
378
379 /*
380  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
381  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
382  * permitted return codes. Nonzero return codes halt waiting and return.
383  */
384 int __sched
385 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
386               wait_bit_action_f *action, unsigned mode)
387 {
388         int ret = 0;
389
390         do {
391                 prepare_to_wait(wq, &q->wait, mode);
392                 if (test_bit(q->key.bit_nr, q->key.flags))
393                         ret = (*action)(&q->key, mode);
394         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
395         finish_wait(wq, &q->wait);
396         return ret;
397 }
398 EXPORT_SYMBOL(__wait_on_bit);
399
400 int __sched out_of_line_wait_on_bit(void *word, int bit,
401                                     wait_bit_action_f *action, unsigned mode)
402 {
403         wait_queue_head_t *wq = bit_waitqueue(word, bit);
404         DEFINE_WAIT_BIT(wait, word, bit);
405
406         return __wait_on_bit(wq, &wait, action, mode);
407 }
408 EXPORT_SYMBOL(out_of_line_wait_on_bit);
409
410 int __sched out_of_line_wait_on_bit_timeout(
411         void *word, int bit, wait_bit_action_f *action,
412         unsigned mode, unsigned long timeout)
413 {
414         wait_queue_head_t *wq = bit_waitqueue(word, bit);
415         DEFINE_WAIT_BIT(wait, word, bit);
416
417         wait.key.timeout = jiffies + timeout;
418         return __wait_on_bit(wq, &wait, action, mode);
419 }
420 EXPORT_SYMBOL_GPL(out_of_line_wait_on_bit_timeout);
421
422 int __sched
423 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
424                         wait_bit_action_f *action, unsigned mode)
425 {
426         do {
427                 int ret;
428
429                 prepare_to_wait_exclusive(wq, &q->wait, mode);
430                 if (!test_bit(q->key.bit_nr, q->key.flags))
431                         continue;
432                 ret = action(&q->key, mode);
433                 if (!ret)
434                         continue;
435                 abort_exclusive_wait(wq, &q->wait, &q->key);
436                 return ret;
437         } while (test_and_set_bit(q->key.bit_nr, q->key.flags));
438         finish_wait(wq, &q->wait);
439         return 0;
440 }
441 EXPORT_SYMBOL(__wait_on_bit_lock);
442
443 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
444                                          wait_bit_action_f *action, unsigned mode)
445 {
446         wait_queue_head_t *wq = bit_waitqueue(word, bit);
447         DEFINE_WAIT_BIT(wait, word, bit);
448
449         return __wait_on_bit_lock(wq, &wait, action, mode);
450 }
451 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
452
453 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
454 {
455         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
456         if (waitqueue_active(wq))
457                 __wake_up(wq, TASK_NORMAL, 1, &key);
458 }
459 EXPORT_SYMBOL(__wake_up_bit);
460
461 /**
462  * wake_up_bit - wake up a waiter on a bit
463  * @word: the word being waited on, a kernel virtual address
464  * @bit: the bit of the word being waited on
465  *
466  * There is a standard hashed waitqueue table for generic use. This
467  * is the part of the hashtable's accessor API that wakes up waiters
468  * on a bit. For instance, if one were to have waiters on a bitflag,
469  * one would call wake_up_bit() after clearing the bit.
470  *
471  * In order for this to function properly, as it uses waitqueue_active()
472  * internally, some kind of memory barrier must be done prior to calling
473  * this. Typically, this will be smp_mb__after_atomic(), but in some
474  * cases where bitflags are manipulated non-atomically under a lock, one
475  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
476  * because spin_unlock() does not guarantee a memory barrier.
477  */
478 void wake_up_bit(void *word, int bit)
479 {
480         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
481 }
482 EXPORT_SYMBOL(wake_up_bit);
483
484 wait_queue_head_t *bit_waitqueue(void *word, int bit)
485 {
486         const int shift = BITS_PER_LONG == 32 ? 5 : 6;
487         const struct zone *zone = page_zone(virt_to_page(word));
488         unsigned long val = (unsigned long)word << shift | bit;
489
490         return &zone->wait_table[hash_long(val, zone->wait_table_bits)];
491 }
492 EXPORT_SYMBOL(bit_waitqueue);
493
494 /*
495  * Manipulate the atomic_t address to produce a better bit waitqueue table hash
496  * index (we're keying off bit -1, but that would produce a horrible hash
497  * value).
498  */
499 static inline wait_queue_head_t *atomic_t_waitqueue(atomic_t *p)
500 {
501         if (BITS_PER_LONG == 64) {
502                 unsigned long q = (unsigned long)p;
503                 return bit_waitqueue((void *)(q & ~1), q & 1);
504         }
505         return bit_waitqueue(p, 0);
506 }
507
508 static int wake_atomic_t_function(wait_queue_t *wait, unsigned mode, int sync,
509                                   void *arg)
510 {
511         struct wait_bit_key *key = arg;
512         struct wait_bit_queue *wait_bit
513                 = container_of(wait, struct wait_bit_queue, wait);
514         atomic_t *val = key->flags;
515
516         if (wait_bit->key.flags != key->flags ||
517             wait_bit->key.bit_nr != key->bit_nr ||
518             atomic_read(val) != 0)
519                 return 0;
520         return autoremove_wake_function(wait, mode, sync, key);
521 }
522
523 /*
524  * To allow interruptible waiting and asynchronous (i.e. nonblocking) waiting,
525  * the actions of __wait_on_atomic_t() are permitted return codes.  Nonzero
526  * return codes halt waiting and return.
527  */
528 static __sched
529 int __wait_on_atomic_t(wait_queue_head_t *wq, struct wait_bit_queue *q,
530                        int (*action)(atomic_t *), unsigned mode)
531 {
532         atomic_t *val;
533         int ret = 0;
534
535         do {
536                 prepare_to_wait(wq, &q->wait, mode);
537                 val = q->key.flags;
538                 if (atomic_read(val) == 0)
539                         break;
540                 ret = (*action)(val);
541         } while (!ret && atomic_read(val) != 0);
542         finish_wait(wq, &q->wait);
543         return ret;
544 }
545
546 #define DEFINE_WAIT_ATOMIC_T(name, p)                                   \
547         struct wait_bit_queue name = {                                  \
548                 .key = __WAIT_ATOMIC_T_KEY_INITIALIZER(p),              \
549                 .wait   = {                                             \
550                         .private        = current,                      \
551                         .func           = wake_atomic_t_function,       \
552                         .task_list      =                               \
553                                 LIST_HEAD_INIT((name).wait.task_list),  \
554                 },                                                      \
555         }
556
557 __sched int out_of_line_wait_on_atomic_t(atomic_t *p, int (*action)(atomic_t *),
558                                          unsigned mode)
559 {
560         wait_queue_head_t *wq = atomic_t_waitqueue(p);
561         DEFINE_WAIT_ATOMIC_T(wait, p);
562
563         return __wait_on_atomic_t(wq, &wait, action, mode);
564 }
565 EXPORT_SYMBOL(out_of_line_wait_on_atomic_t);
566
567 /**
568  * wake_up_atomic_t - Wake up a waiter on a atomic_t
569  * @p: The atomic_t being waited on, a kernel virtual address
570  *
571  * Wake up anyone waiting for the atomic_t to go to zero.
572  *
573  * Abuse the bit-waker function and its waitqueue hash table set (the atomic_t
574  * check is done by the waiter's wake function, not the by the waker itself).
575  */
576 void wake_up_atomic_t(atomic_t *p)
577 {
578         __wake_up_bit(atomic_t_waitqueue(p), p, WAIT_ATOMIC_T_BIT_NR);
579 }
580 EXPORT_SYMBOL(wake_up_atomic_t);
581
582 __sched int bit_wait(struct wait_bit_key *word, int mode)
583 {
584         schedule();
585         if (signal_pending_state(mode, current))
586                 return -EINTR;
587         return 0;
588 }
589 EXPORT_SYMBOL(bit_wait);
590
591 __sched int bit_wait_io(struct wait_bit_key *word, int mode)
592 {
593         io_schedule();
594         if (signal_pending_state(mode, current))
595                 return -EINTR;
596         return 0;
597 }
598 EXPORT_SYMBOL(bit_wait_io);
599
600 __sched int bit_wait_timeout(struct wait_bit_key *word, int mode)
601 {
602         unsigned long now = READ_ONCE(jiffies);
603         if (time_after_eq(now, word->timeout))
604                 return -EAGAIN;
605         schedule_timeout(word->timeout - now);
606         if (signal_pending_state(mode, current))
607                 return -EINTR;
608         return 0;
609 }
610 EXPORT_SYMBOL_GPL(bit_wait_timeout);
611
612 __sched int bit_wait_io_timeout(struct wait_bit_key *word, int mode)
613 {
614         unsigned long now = READ_ONCE(jiffies);
615         if (time_after_eq(now, word->timeout))
616                 return -EAGAIN;
617         io_schedule_timeout(word->timeout - now);
618         if (signal_pending_state(mode, current))
619                 return -EINTR;
620         return 0;
621 }
622 EXPORT_SYMBOL_GPL(bit_wait_io_timeout);