]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - kernel/sched/wait.c
Merge tag 'mfd-fixes-4.3' of git://git.kernel.org/pub/scm/linux/kernel/git/lee/mfd
[karo-tx-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  * @mode: runstate of the waiter to be woken
263  * @key: key to identify a wait bit queue or %NULL
264  *
265  * Sets current thread back to running state and removes
266  * the wait descriptor from the given waitqueue if still
267  * queued.
268  *
269  * Wakes up the next waiter if the caller is concurrently
270  * woken up through the queue.
271  *
272  * This prevents waiter starvation where an exclusive waiter
273  * aborts and is woken up concurrently and no one wakes up
274  * the next waiter.
275  */
276 void abort_exclusive_wait(wait_queue_head_t *q, wait_queue_t *wait,
277                         unsigned int mode, void *key)
278 {
279         unsigned long flags;
280
281         __set_current_state(TASK_RUNNING);
282         spin_lock_irqsave(&q->lock, flags);
283         if (!list_empty(&wait->task_list))
284                 list_del_init(&wait->task_list);
285         else if (waitqueue_active(q))
286                 __wake_up_locked_key(q, mode, key);
287         spin_unlock_irqrestore(&q->lock, flags);
288 }
289 EXPORT_SYMBOL(abort_exclusive_wait);
290
291 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
292 {
293         int ret = default_wake_function(wait, mode, sync, key);
294
295         if (ret)
296                 list_del_init(&wait->task_list);
297         return ret;
298 }
299 EXPORT_SYMBOL(autoremove_wake_function);
300
301 static inline bool is_kthread_should_stop(void)
302 {
303         return (current->flags & PF_KTHREAD) && kthread_should_stop();
304 }
305
306 /*
307  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
308  *
309  * add_wait_queue(&wq, &wait);
310  * for (;;) {
311  *     if (condition)
312  *         break;
313  *
314  *     p->state = mode;                         condition = true;
315  *     smp_mb(); // A                           smp_wmb(); // C
316  *     if (!wait->flags & WQ_FLAG_WOKEN)        wait->flags |= WQ_FLAG_WOKEN;
317  *         schedule()                           try_to_wake_up();
318  *     p->state = TASK_RUNNING;             ~~~~~~~~~~~~~~~~~~
319  *     wait->flags &= ~WQ_FLAG_WOKEN;           condition = true;
320  *     smp_mb() // B                            smp_wmb(); // C
321  *                                              wait->flags |= WQ_FLAG_WOKEN;
322  * }
323  * remove_wait_queue(&wq, &wait);
324  *
325  */
326 long wait_woken(wait_queue_t *wait, unsigned mode, long timeout)
327 {
328         set_current_state(mode); /* A */
329         /*
330          * The above implies an smp_mb(), which matches with the smp_wmb() from
331          * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
332          * also observe all state before the wakeup.
333          */
334         if (!(wait->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
335                 timeout = schedule_timeout(timeout);
336         __set_current_state(TASK_RUNNING);
337
338         /*
339          * The below implies an smp_mb(), it too pairs with the smp_wmb() from
340          * woken_wake_function() such that we must either observe the wait
341          * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
342          * an event.
343          */
344         smp_store_mb(wait->flags, wait->flags & ~WQ_FLAG_WOKEN); /* B */
345
346         return timeout;
347 }
348 EXPORT_SYMBOL(wait_woken);
349
350 int woken_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
351 {
352         /*
353          * Although this function is called under waitqueue lock, LOCK
354          * doesn't imply write barrier and the users expects write
355          * barrier semantics on wakeup functions.  The following
356          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
357          * and is paired with smp_store_mb() in wait_woken().
358          */
359         smp_wmb(); /* C */
360         wait->flags |= WQ_FLAG_WOKEN;
361
362         return default_wake_function(wait, mode, sync, key);
363 }
364 EXPORT_SYMBOL(woken_wake_function);
365
366 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
367 {
368         struct wait_bit_key *key = arg;
369         struct wait_bit_queue *wait_bit
370                 = container_of(wait, struct wait_bit_queue, wait);
371
372         if (wait_bit->key.flags != key->flags ||
373                         wait_bit->key.bit_nr != key->bit_nr ||
374                         test_bit(key->bit_nr, key->flags))
375                 return 0;
376         else
377                 return autoremove_wake_function(wait, mode, sync, key);
378 }
379 EXPORT_SYMBOL(wake_bit_function);
380
381 /*
382  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
383  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
384  * permitted return codes. Nonzero return codes halt waiting and return.
385  */
386 int __sched
387 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
388               wait_bit_action_f *action, unsigned mode)
389 {
390         int ret = 0;
391
392         do {
393                 prepare_to_wait(wq, &q->wait, mode);
394                 if (test_bit(q->key.bit_nr, q->key.flags))
395                         ret = (*action)(&q->key);
396         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
397         finish_wait(wq, &q->wait);
398         return ret;
399 }
400 EXPORT_SYMBOL(__wait_on_bit);
401
402 int __sched out_of_line_wait_on_bit(void *word, int bit,
403                                     wait_bit_action_f *action, unsigned mode)
404 {
405         wait_queue_head_t *wq = bit_waitqueue(word, bit);
406         DEFINE_WAIT_BIT(wait, word, bit);
407
408         return __wait_on_bit(wq, &wait, action, mode);
409 }
410 EXPORT_SYMBOL(out_of_line_wait_on_bit);
411
412 int __sched out_of_line_wait_on_bit_timeout(
413         void *word, int bit, wait_bit_action_f *action,
414         unsigned mode, unsigned long timeout)
415 {
416         wait_queue_head_t *wq = bit_waitqueue(word, bit);
417         DEFINE_WAIT_BIT(wait, word, bit);
418
419         wait.key.timeout = jiffies + timeout;
420         return __wait_on_bit(wq, &wait, action, mode);
421 }
422 EXPORT_SYMBOL_GPL(out_of_line_wait_on_bit_timeout);
423
424 int __sched
425 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
426                         wait_bit_action_f *action, unsigned mode)
427 {
428         do {
429                 int ret;
430
431                 prepare_to_wait_exclusive(wq, &q->wait, mode);
432                 if (!test_bit(q->key.bit_nr, q->key.flags))
433                         continue;
434                 ret = action(&q->key);
435                 if (!ret)
436                         continue;
437                 abort_exclusive_wait(wq, &q->wait, mode, &q->key);
438                 return ret;
439         } while (test_and_set_bit(q->key.bit_nr, q->key.flags));
440         finish_wait(wq, &q->wait);
441         return 0;
442 }
443 EXPORT_SYMBOL(__wait_on_bit_lock);
444
445 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
446                                          wait_bit_action_f *action, unsigned mode)
447 {
448         wait_queue_head_t *wq = bit_waitqueue(word, bit);
449         DEFINE_WAIT_BIT(wait, word, bit);
450
451         return __wait_on_bit_lock(wq, &wait, action, mode);
452 }
453 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
454
455 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
456 {
457         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
458         if (waitqueue_active(wq))
459                 __wake_up(wq, TASK_NORMAL, 1, &key);
460 }
461 EXPORT_SYMBOL(__wake_up_bit);
462
463 /**
464  * wake_up_bit - wake up a waiter on a bit
465  * @word: the word being waited on, a kernel virtual address
466  * @bit: the bit of the word being waited on
467  *
468  * There is a standard hashed waitqueue table for generic use. This
469  * is the part of the hashtable's accessor API that wakes up waiters
470  * on a bit. For instance, if one were to have waiters on a bitflag,
471  * one would call wake_up_bit() after clearing the bit.
472  *
473  * In order for this to function properly, as it uses waitqueue_active()
474  * internally, some kind of memory barrier must be done prior to calling
475  * this. Typically, this will be smp_mb__after_atomic(), but in some
476  * cases where bitflags are manipulated non-atomically under a lock, one
477  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
478  * because spin_unlock() does not guarantee a memory barrier.
479  */
480 void wake_up_bit(void *word, int bit)
481 {
482         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
483 }
484 EXPORT_SYMBOL(wake_up_bit);
485
486 wait_queue_head_t *bit_waitqueue(void *word, int bit)
487 {
488         const int shift = BITS_PER_LONG == 32 ? 5 : 6;
489         const struct zone *zone = page_zone(virt_to_page(word));
490         unsigned long val = (unsigned long)word << shift | bit;
491
492         return &zone->wait_table[hash_long(val, zone->wait_table_bits)];
493 }
494 EXPORT_SYMBOL(bit_waitqueue);
495
496 /*
497  * Manipulate the atomic_t address to produce a better bit waitqueue table hash
498  * index (we're keying off bit -1, but that would produce a horrible hash
499  * value).
500  */
501 static inline wait_queue_head_t *atomic_t_waitqueue(atomic_t *p)
502 {
503         if (BITS_PER_LONG == 64) {
504                 unsigned long q = (unsigned long)p;
505                 return bit_waitqueue((void *)(q & ~1), q & 1);
506         }
507         return bit_waitqueue(p, 0);
508 }
509
510 static int wake_atomic_t_function(wait_queue_t *wait, unsigned mode, int sync,
511                                   void *arg)
512 {
513         struct wait_bit_key *key = arg;
514         struct wait_bit_queue *wait_bit
515                 = container_of(wait, struct wait_bit_queue, wait);
516         atomic_t *val = key->flags;
517
518         if (wait_bit->key.flags != key->flags ||
519             wait_bit->key.bit_nr != key->bit_nr ||
520             atomic_read(val) != 0)
521                 return 0;
522         return autoremove_wake_function(wait, mode, sync, key);
523 }
524
525 /*
526  * To allow interruptible waiting and asynchronous (i.e. nonblocking) waiting,
527  * the actions of __wait_on_atomic_t() are permitted return codes.  Nonzero
528  * return codes halt waiting and return.
529  */
530 static __sched
531 int __wait_on_atomic_t(wait_queue_head_t *wq, struct wait_bit_queue *q,
532                        int (*action)(atomic_t *), unsigned mode)
533 {
534         atomic_t *val;
535         int ret = 0;
536
537         do {
538                 prepare_to_wait(wq, &q->wait, mode);
539                 val = q->key.flags;
540                 if (atomic_read(val) == 0)
541                         break;
542                 ret = (*action)(val);
543         } while (!ret && atomic_read(val) != 0);
544         finish_wait(wq, &q->wait);
545         return ret;
546 }
547
548 #define DEFINE_WAIT_ATOMIC_T(name, p)                                   \
549         struct wait_bit_queue name = {                                  \
550                 .key = __WAIT_ATOMIC_T_KEY_INITIALIZER(p),              \
551                 .wait   = {                                             \
552                         .private        = current,                      \
553                         .func           = wake_atomic_t_function,       \
554                         .task_list      =                               \
555                                 LIST_HEAD_INIT((name).wait.task_list),  \
556                 },                                                      \
557         }
558
559 __sched int out_of_line_wait_on_atomic_t(atomic_t *p, int (*action)(atomic_t *),
560                                          unsigned mode)
561 {
562         wait_queue_head_t *wq = atomic_t_waitqueue(p);
563         DEFINE_WAIT_ATOMIC_T(wait, p);
564
565         return __wait_on_atomic_t(wq, &wait, action, mode);
566 }
567 EXPORT_SYMBOL(out_of_line_wait_on_atomic_t);
568
569 /**
570  * wake_up_atomic_t - Wake up a waiter on a atomic_t
571  * @p: The atomic_t being waited on, a kernel virtual address
572  *
573  * Wake up anyone waiting for the atomic_t to go to zero.
574  *
575  * Abuse the bit-waker function and its waitqueue hash table set (the atomic_t
576  * check is done by the waiter's wake function, not the by the waker itself).
577  */
578 void wake_up_atomic_t(atomic_t *p)
579 {
580         __wake_up_bit(atomic_t_waitqueue(p), p, WAIT_ATOMIC_T_BIT_NR);
581 }
582 EXPORT_SYMBOL(wake_up_atomic_t);
583
584 __sched int bit_wait(struct wait_bit_key *word)
585 {
586         if (signal_pending_state(current->state, current))
587                 return 1;
588         schedule();
589         return 0;
590 }
591 EXPORT_SYMBOL(bit_wait);
592
593 __sched int bit_wait_io(struct wait_bit_key *word)
594 {
595         if (signal_pending_state(current->state, current))
596                 return 1;
597         io_schedule();
598         return 0;
599 }
600 EXPORT_SYMBOL(bit_wait_io);
601
602 __sched int bit_wait_timeout(struct wait_bit_key *word)
603 {
604         unsigned long now = READ_ONCE(jiffies);
605         if (signal_pending_state(current->state, current))
606                 return 1;
607         if (time_after_eq(now, word->timeout))
608                 return -EAGAIN;
609         schedule_timeout(word->timeout - now);
610         return 0;
611 }
612 EXPORT_SYMBOL_GPL(bit_wait_timeout);
613
614 __sched int bit_wait_io_timeout(struct wait_bit_key *word)
615 {
616         unsigned long now = READ_ONCE(jiffies);
617         if (signal_pending_state(current->state, current))
618                 return 1;
619         if (time_after_eq(now, word->timeout))
620                 return -EAGAIN;
621         io_schedule_timeout(word->timeout - now);
622         return 0;
623 }
624 EXPORT_SYMBOL_GPL(bit_wait_io_timeout);