RDMA/cma: Document rdma_destroy_id() function
[linux-2.6] / kernel / rtmutex.c
1 /*
2  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
3  *
4  * started by Ingo Molnar and Thomas Gleixner.
5  *
6  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
8  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
9  *  Copyright (C) 2006 Esben Nielsen
10  *
11  *  See Documentation/rt-mutex-design.txt for details.
12  */
13 #include <linux/spinlock.h>
14 #include <linux/module.h>
15 #include <linux/sched.h>
16 #include <linux/timer.h>
17
18 #include "rtmutex_common.h"
19
20 #ifdef CONFIG_DEBUG_RT_MUTEXES
21 # include "rtmutex-debug.h"
22 #else
23 # include "rtmutex.h"
24 #endif
25
26 /*
27  * lock->owner state tracking:
28  *
29  * lock->owner holds the task_struct pointer of the owner. Bit 0 and 1
30  * are used to keep track of the "owner is pending" and "lock has
31  * waiters" state.
32  *
33  * owner        bit1    bit0
34  * NULL         0       0       lock is free (fast acquire possible)
35  * NULL         0       1       invalid state
36  * NULL         1       0       Transitional State*
37  * NULL         1       1       invalid state
38  * taskpointer  0       0       lock is held (fast release possible)
39  * taskpointer  0       1       task is pending owner
40  * taskpointer  1       0       lock is held and has waiters
41  * taskpointer  1       1       task is pending owner and lock has more waiters
42  *
43  * Pending ownership is assigned to the top (highest priority)
44  * waiter of the lock, when the lock is released. The thread is woken
45  * up and can now take the lock. Until the lock is taken (bit 0
46  * cleared) a competing higher priority thread can steal the lock
47  * which puts the woken up thread back on the waiters list.
48  *
49  * The fast atomic compare exchange based acquire and release is only
50  * possible when bit 0 and 1 of lock->owner are 0.
51  *
52  * (*) There's a small time where the owner can be NULL and the
53  * "lock has waiters" bit is set.  This can happen when grabbing the lock.
54  * To prevent a cmpxchg of the owner releasing the lock, we need to set this
55  * bit before looking at the lock, hence the reason this is a transitional
56  * state.
57  */
58
59 static void
60 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner,
61                    unsigned long mask)
62 {
63         unsigned long val = (unsigned long)owner | mask;
64
65         if (rt_mutex_has_waiters(lock))
66                 val |= RT_MUTEX_HAS_WAITERS;
67
68         lock->owner = (struct task_struct *)val;
69 }
70
71 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
72 {
73         lock->owner = (struct task_struct *)
74                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
75 }
76
77 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
78 {
79         if (!rt_mutex_has_waiters(lock))
80                 clear_rt_mutex_waiters(lock);
81 }
82
83 /*
84  * We can speed up the acquire/release, if the architecture
85  * supports cmpxchg and if there's no debugging state to be set up
86  */
87 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
88 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
89 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
90 {
91         unsigned long owner, *p = (unsigned long *) &lock->owner;
92
93         do {
94                 owner = *p;
95         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
96 }
97 #else
98 # define rt_mutex_cmpxchg(l,c,n)        (0)
99 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
100 {
101         lock->owner = (struct task_struct *)
102                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
103 }
104 #endif
105
106 /*
107  * Calculate task priority from the waiter list priority
108  *
109  * Return task->normal_prio when the waiter list is empty or when
110  * the waiter is not allowed to do priority boosting
111  */
112 int rt_mutex_getprio(struct task_struct *task)
113 {
114         if (likely(!task_has_pi_waiters(task)))
115                 return task->normal_prio;
116
117         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
118                    task->normal_prio);
119 }
120
121 /*
122  * Adjust the priority of a task, after its pi_waiters got modified.
123  *
124  * This can be both boosting and unboosting. task->pi_lock must be held.
125  */
126 static void __rt_mutex_adjust_prio(struct task_struct *task)
127 {
128         int prio = rt_mutex_getprio(task);
129
130         if (task->prio != prio)
131                 rt_mutex_setprio(task, prio);
132 }
133
134 /*
135  * Adjust task priority (undo boosting). Called from the exit path of
136  * rt_mutex_slowunlock() and rt_mutex_slowlock().
137  *
138  * (Note: We do this outside of the protection of lock->wait_lock to
139  * allow the lock to be taken while or before we readjust the priority
140  * of task. We do not use the spin_xx_mutex() variants here as we are
141  * outside of the debug path.)
142  */
143 static void rt_mutex_adjust_prio(struct task_struct *task)
144 {
145         unsigned long flags;
146
147         spin_lock_irqsave(&task->pi_lock, flags);
148         __rt_mutex_adjust_prio(task);
149         spin_unlock_irqrestore(&task->pi_lock, flags);
150 }
151
152 /*
153  * Max number of times we'll walk the boosting chain:
154  */
155 int max_lock_depth = 1024;
156
157 /*
158  * Adjust the priority chain. Also used for deadlock detection.
159  * Decreases task's usage by one - may thus free the task.
160  * Returns 0 or -EDEADLK.
161  */
162 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
163                                       int deadlock_detect,
164                                       struct rt_mutex *orig_lock,
165                                       struct rt_mutex_waiter *orig_waiter,
166                                       struct task_struct *top_task)
167 {
168         struct rt_mutex *lock;
169         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
170         int detect_deadlock, ret = 0, depth = 0;
171         unsigned long flags;
172
173         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
174                                                          deadlock_detect);
175
176         /*
177          * The (de)boosting is a step by step approach with a lot of
178          * pitfalls. We want this to be preemptible and we want hold a
179          * maximum of two locks per step. So we have to check
180          * carefully whether things change under us.
181          */
182  again:
183         if (++depth > max_lock_depth) {
184                 static int prev_max;
185
186                 /*
187                  * Print this only once. If the admin changes the limit,
188                  * print a new message when reaching the limit again.
189                  */
190                 if (prev_max != max_lock_depth) {
191                         prev_max = max_lock_depth;
192                         printk(KERN_WARNING "Maximum lock depth %d reached "
193                                "task: %s (%d)\n", max_lock_depth,
194                                top_task->comm, top_task->pid);
195                 }
196                 put_task_struct(task);
197
198                 return deadlock_detect ? -EDEADLK : 0;
199         }
200  retry:
201         /*
202          * Task can not go away as we did a get_task() before !
203          */
204         spin_lock_irqsave(&task->pi_lock, flags);
205
206         waiter = task->pi_blocked_on;
207         /*
208          * Check whether the end of the boosting chain has been
209          * reached or the state of the chain has changed while we
210          * dropped the locks.
211          */
212         if (!waiter || !waiter->task)
213                 goto out_unlock_pi;
214
215         if (top_waiter && (!task_has_pi_waiters(task) ||
216                            top_waiter != task_top_pi_waiter(task)))
217                 goto out_unlock_pi;
218
219         /*
220          * When deadlock detection is off then we check, if further
221          * priority adjustment is necessary.
222          */
223         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
224                 goto out_unlock_pi;
225
226         lock = waiter->lock;
227         if (!spin_trylock(&lock->wait_lock)) {
228                 spin_unlock_irqrestore(&task->pi_lock, flags);
229                 cpu_relax();
230                 goto retry;
231         }
232
233         /* Deadlock detection */
234         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
235                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
236                 spin_unlock(&lock->wait_lock);
237                 ret = deadlock_detect ? -EDEADLK : 0;
238                 goto out_unlock_pi;
239         }
240
241         top_waiter = rt_mutex_top_waiter(lock);
242
243         /* Requeue the waiter */
244         plist_del(&waiter->list_entry, &lock->wait_list);
245         waiter->list_entry.prio = task->prio;
246         plist_add(&waiter->list_entry, &lock->wait_list);
247
248         /* Release the task */
249         spin_unlock_irqrestore(&task->pi_lock, flags);
250         put_task_struct(task);
251
252         /* Grab the next task */
253         task = rt_mutex_owner(lock);
254         spin_lock_irqsave(&task->pi_lock, flags);
255
256         if (waiter == rt_mutex_top_waiter(lock)) {
257                 /* Boost the owner */
258                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
259                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
260                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
261                 __rt_mutex_adjust_prio(task);
262
263         } else if (top_waiter == waiter) {
264                 /* Deboost the owner */
265                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
266                 waiter = rt_mutex_top_waiter(lock);
267                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
268                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
269                 __rt_mutex_adjust_prio(task);
270         }
271
272         get_task_struct(task);
273         spin_unlock_irqrestore(&task->pi_lock, flags);
274
275         top_waiter = rt_mutex_top_waiter(lock);
276         spin_unlock(&lock->wait_lock);
277
278         if (!detect_deadlock && waiter != top_waiter)
279                 goto out_put_task;
280
281         goto again;
282
283  out_unlock_pi:
284         spin_unlock_irqrestore(&task->pi_lock, flags);
285  out_put_task:
286         put_task_struct(task);
287
288         return ret;
289 }
290
291 /*
292  * Optimization: check if we can steal the lock from the
293  * assigned pending owner [which might not have taken the
294  * lock yet]:
295  */
296 static inline int try_to_steal_lock(struct rt_mutex *lock)
297 {
298         struct task_struct *pendowner = rt_mutex_owner(lock);
299         struct rt_mutex_waiter *next;
300         unsigned long flags;
301
302         if (!rt_mutex_owner_pending(lock))
303                 return 0;
304
305         if (pendowner == current)
306                 return 1;
307
308         spin_lock_irqsave(&pendowner->pi_lock, flags);
309         if (current->prio >= pendowner->prio) {
310                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
311                 return 0;
312         }
313
314         /*
315          * Check if a waiter is enqueued on the pending owners
316          * pi_waiters list. Remove it and readjust pending owners
317          * priority.
318          */
319         if (likely(!rt_mutex_has_waiters(lock))) {
320                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
321                 return 1;
322         }
323
324         /* No chain handling, pending owner is not blocked on anything: */
325         next = rt_mutex_top_waiter(lock);
326         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
327         __rt_mutex_adjust_prio(pendowner);
328         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
329
330         /*
331          * We are going to steal the lock and a waiter was
332          * enqueued on the pending owners pi_waiters queue. So
333          * we have to enqueue this waiter into
334          * current->pi_waiters list. This covers the case,
335          * where current is boosted because it holds another
336          * lock and gets unboosted because the booster is
337          * interrupted, so we would delay a waiter with higher
338          * priority as current->normal_prio.
339          *
340          * Note: in the rare case of a SCHED_OTHER task changing
341          * its priority and thus stealing the lock, next->task
342          * might be current:
343          */
344         if (likely(next->task != current)) {
345                 spin_lock_irqsave(&current->pi_lock, flags);
346                 plist_add(&next->pi_list_entry, &current->pi_waiters);
347                 __rt_mutex_adjust_prio(current);
348                 spin_unlock_irqrestore(&current->pi_lock, flags);
349         }
350         return 1;
351 }
352
353 /*
354  * Try to take an rt-mutex
355  *
356  * This fails
357  * - when the lock has a real owner
358  * - when a different pending owner exists and has higher priority than current
359  *
360  * Must be called with lock->wait_lock held.
361  */
362 static int try_to_take_rt_mutex(struct rt_mutex *lock)
363 {
364         /*
365          * We have to be careful here if the atomic speedups are
366          * enabled, such that, when
367          *  - no other waiter is on the lock
368          *  - the lock has been released since we did the cmpxchg
369          * the lock can be released or taken while we are doing the
370          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
371          *
372          * The atomic acquire/release aware variant of
373          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
374          * the WAITERS bit, the atomic release / acquire can not
375          * happen anymore and lock->wait_lock protects us from the
376          * non-atomic case.
377          *
378          * Note, that this might set lock->owner =
379          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
380          * any more. This is fixed up when we take the ownership.
381          * This is the transitional state explained at the top of this file.
382          */
383         mark_rt_mutex_waiters(lock);
384
385         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
386                 return 0;
387
388         /* We got the lock. */
389         debug_rt_mutex_lock(lock);
390
391         rt_mutex_set_owner(lock, current, 0);
392
393         rt_mutex_deadlock_account_lock(lock, current);
394
395         return 1;
396 }
397
398 /*
399  * Task blocks on lock.
400  *
401  * Prepare waiter and propagate pi chain
402  *
403  * This must be called with lock->wait_lock held.
404  */
405 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
406                                    struct rt_mutex_waiter *waiter,
407                                    int detect_deadlock)
408 {
409         struct task_struct *owner = rt_mutex_owner(lock);
410         struct rt_mutex_waiter *top_waiter = waiter;
411         unsigned long flags;
412         int boost = 0, res;
413
414         spin_lock_irqsave(&current->pi_lock, flags);
415         __rt_mutex_adjust_prio(current);
416         waiter->task = current;
417         waiter->lock = lock;
418         plist_node_init(&waiter->list_entry, current->prio);
419         plist_node_init(&waiter->pi_list_entry, current->prio);
420
421         /* Get the top priority waiter on the lock */
422         if (rt_mutex_has_waiters(lock))
423                 top_waiter = rt_mutex_top_waiter(lock);
424         plist_add(&waiter->list_entry, &lock->wait_list);
425
426         current->pi_blocked_on = waiter;
427
428         spin_unlock_irqrestore(&current->pi_lock, flags);
429
430         if (waiter == rt_mutex_top_waiter(lock)) {
431                 spin_lock_irqsave(&owner->pi_lock, flags);
432                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
433                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
434
435                 __rt_mutex_adjust_prio(owner);
436                 if (owner->pi_blocked_on) {
437                         boost = 1;
438                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
439                         get_task_struct(owner);
440                 }
441                 spin_unlock_irqrestore(&owner->pi_lock, flags);
442         }
443         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
444                 spin_lock_irqsave(&owner->pi_lock, flags);
445                 if (owner->pi_blocked_on) {
446                         boost = 1;
447                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
448                         get_task_struct(owner);
449                 }
450                 spin_unlock_irqrestore(&owner->pi_lock, flags);
451         }
452         if (!boost)
453                 return 0;
454
455         spin_unlock(&lock->wait_lock);
456
457         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
458                                          current);
459
460         spin_lock(&lock->wait_lock);
461
462         return res;
463 }
464
465 /*
466  * Wake up the next waiter on the lock.
467  *
468  * Remove the top waiter from the current tasks waiter list and from
469  * the lock waiter list. Set it as pending owner. Then wake it up.
470  *
471  * Called with lock->wait_lock held.
472  */
473 static void wakeup_next_waiter(struct rt_mutex *lock)
474 {
475         struct rt_mutex_waiter *waiter;
476         struct task_struct *pendowner;
477         unsigned long flags;
478
479         spin_lock_irqsave(&current->pi_lock, flags);
480
481         waiter = rt_mutex_top_waiter(lock);
482         plist_del(&waiter->list_entry, &lock->wait_list);
483
484         /*
485          * Remove it from current->pi_waiters. We do not adjust a
486          * possible priority boost right now. We execute wakeup in the
487          * boosted mode and go back to normal after releasing
488          * lock->wait_lock.
489          */
490         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
491         pendowner = waiter->task;
492         waiter->task = NULL;
493
494         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
495
496         spin_unlock_irqrestore(&current->pi_lock, flags);
497
498         /*
499          * Clear the pi_blocked_on variable and enqueue a possible
500          * waiter into the pi_waiters list of the pending owner. This
501          * prevents that in case the pending owner gets unboosted a
502          * waiter with higher priority than pending-owner->normal_prio
503          * is blocked on the unboosted (pending) owner.
504          */
505         spin_lock_irqsave(&pendowner->pi_lock, flags);
506
507         WARN_ON(!pendowner->pi_blocked_on);
508         WARN_ON(pendowner->pi_blocked_on != waiter);
509         WARN_ON(pendowner->pi_blocked_on->lock != lock);
510
511         pendowner->pi_blocked_on = NULL;
512
513         if (rt_mutex_has_waiters(lock)) {
514                 struct rt_mutex_waiter *next;
515
516                 next = rt_mutex_top_waiter(lock);
517                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
518         }
519         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
520
521         wake_up_process(pendowner);
522 }
523
524 /*
525  * Remove a waiter from a lock
526  *
527  * Must be called with lock->wait_lock held
528  */
529 static void remove_waiter(struct rt_mutex *lock,
530                           struct rt_mutex_waiter *waiter)
531 {
532         int first = (waiter == rt_mutex_top_waiter(lock));
533         struct task_struct *owner = rt_mutex_owner(lock);
534         unsigned long flags;
535         int boost = 0;
536
537         spin_lock_irqsave(&current->pi_lock, flags);
538         plist_del(&waiter->list_entry, &lock->wait_list);
539         waiter->task = NULL;
540         current->pi_blocked_on = NULL;
541         spin_unlock_irqrestore(&current->pi_lock, flags);
542
543         if (first && owner != current) {
544
545                 spin_lock_irqsave(&owner->pi_lock, flags);
546
547                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
548
549                 if (rt_mutex_has_waiters(lock)) {
550                         struct rt_mutex_waiter *next;
551
552                         next = rt_mutex_top_waiter(lock);
553                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
554                 }
555                 __rt_mutex_adjust_prio(owner);
556
557                 if (owner->pi_blocked_on) {
558                         boost = 1;
559                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
560                         get_task_struct(owner);
561                 }
562                 spin_unlock_irqrestore(&owner->pi_lock, flags);
563         }
564
565         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
566
567         if (!boost)
568                 return;
569
570         spin_unlock(&lock->wait_lock);
571
572         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
573
574         spin_lock(&lock->wait_lock);
575 }
576
577 /*
578  * Recheck the pi chain, in case we got a priority setting
579  *
580  * Called from sched_setscheduler
581  */
582 void rt_mutex_adjust_pi(struct task_struct *task)
583 {
584         struct rt_mutex_waiter *waiter;
585         unsigned long flags;
586
587         spin_lock_irqsave(&task->pi_lock, flags);
588
589         waiter = task->pi_blocked_on;
590         if (!waiter || waiter->list_entry.prio == task->prio) {
591                 spin_unlock_irqrestore(&task->pi_lock, flags);
592                 return;
593         }
594
595         /* gets dropped in rt_mutex_adjust_prio_chain()! */
596         get_task_struct(task);
597         spin_unlock_irqrestore(&task->pi_lock, flags);
598
599         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
600 }
601
602 /*
603  * Slow path lock function:
604  */
605 static int __sched
606 rt_mutex_slowlock(struct rt_mutex *lock, int state,
607                   struct hrtimer_sleeper *timeout,
608                   int detect_deadlock)
609 {
610         struct rt_mutex_waiter waiter;
611         int ret = 0;
612
613         debug_rt_mutex_init_waiter(&waiter);
614         waiter.task = NULL;
615
616         spin_lock(&lock->wait_lock);
617
618         /* Try to acquire the lock again: */
619         if (try_to_take_rt_mutex(lock)) {
620                 spin_unlock(&lock->wait_lock);
621                 return 0;
622         }
623
624         set_current_state(state);
625
626         /* Setup the timer, when timeout != NULL */
627         if (unlikely(timeout))
628                 hrtimer_start(&timeout->timer, timeout->timer.expires,
629                               HRTIMER_ABS);
630
631         for (;;) {
632                 /* Try to acquire the lock: */
633                 if (try_to_take_rt_mutex(lock))
634                         break;
635
636                 /*
637                  * TASK_INTERRUPTIBLE checks for signals and
638                  * timeout. Ignored otherwise.
639                  */
640                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
641                         /* Signal pending? */
642                         if (signal_pending(current))
643                                 ret = -EINTR;
644                         if (timeout && !timeout->task)
645                                 ret = -ETIMEDOUT;
646                         if (ret)
647                                 break;
648                 }
649
650                 /*
651                  * waiter.task is NULL the first time we come here and
652                  * when we have been woken up by the previous owner
653                  * but the lock got stolen by a higher prio task.
654                  */
655                 if (!waiter.task) {
656                         ret = task_blocks_on_rt_mutex(lock, &waiter,
657                                                       detect_deadlock);
658                         /*
659                          * If we got woken up by the owner then start loop
660                          * all over without going into schedule to try
661                          * to get the lock now:
662                          */
663                         if (unlikely(!waiter.task))
664                                 continue;
665
666                         if (unlikely(ret))
667                                 break;
668                 }
669
670                 spin_unlock(&lock->wait_lock);
671
672                 debug_rt_mutex_print_deadlock(&waiter);
673
674                 if (waiter.task)
675                         schedule_rt_mutex(lock);
676
677                 spin_lock(&lock->wait_lock);
678                 set_current_state(state);
679         }
680
681         set_current_state(TASK_RUNNING);
682
683         if (unlikely(waiter.task))
684                 remove_waiter(lock, &waiter);
685
686         /*
687          * try_to_take_rt_mutex() sets the waiter bit
688          * unconditionally. We might have to fix that up.
689          */
690         fixup_rt_mutex_waiters(lock);
691
692         spin_unlock(&lock->wait_lock);
693
694         /* Remove pending timer: */
695         if (unlikely(timeout))
696                 hrtimer_cancel(&timeout->timer);
697
698         /*
699          * Readjust priority, when we did not get the lock. We might
700          * have been the pending owner and boosted. Since we did not
701          * take the lock, the PI boost has to go.
702          */
703         if (unlikely(ret))
704                 rt_mutex_adjust_prio(current);
705
706         debug_rt_mutex_free_waiter(&waiter);
707
708         return ret;
709 }
710
711 /*
712  * Slow path try-lock function:
713  */
714 static inline int
715 rt_mutex_slowtrylock(struct rt_mutex *lock)
716 {
717         int ret = 0;
718
719         spin_lock(&lock->wait_lock);
720
721         if (likely(rt_mutex_owner(lock) != current)) {
722
723                 ret = try_to_take_rt_mutex(lock);
724                 /*
725                  * try_to_take_rt_mutex() sets the lock waiters
726                  * bit unconditionally. Clean this up.
727                  */
728                 fixup_rt_mutex_waiters(lock);
729         }
730
731         spin_unlock(&lock->wait_lock);
732
733         return ret;
734 }
735
736 /*
737  * Slow path to release a rt-mutex:
738  */
739 static void __sched
740 rt_mutex_slowunlock(struct rt_mutex *lock)
741 {
742         spin_lock(&lock->wait_lock);
743
744         debug_rt_mutex_unlock(lock);
745
746         rt_mutex_deadlock_account_unlock(current);
747
748         if (!rt_mutex_has_waiters(lock)) {
749                 lock->owner = NULL;
750                 spin_unlock(&lock->wait_lock);
751                 return;
752         }
753
754         wakeup_next_waiter(lock);
755
756         spin_unlock(&lock->wait_lock);
757
758         /* Undo pi boosting if necessary: */
759         rt_mutex_adjust_prio(current);
760 }
761
762 /*
763  * debug aware fast / slowpath lock,trylock,unlock
764  *
765  * The atomic acquire/release ops are compiled away, when either the
766  * architecture does not support cmpxchg or when debugging is enabled.
767  */
768 static inline int
769 rt_mutex_fastlock(struct rt_mutex *lock, int state,
770                   int detect_deadlock,
771                   int (*slowfn)(struct rt_mutex *lock, int state,
772                                 struct hrtimer_sleeper *timeout,
773                                 int detect_deadlock))
774 {
775         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
776                 rt_mutex_deadlock_account_lock(lock, current);
777                 return 0;
778         } else
779                 return slowfn(lock, state, NULL, detect_deadlock);
780 }
781
782 static inline int
783 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
784                         struct hrtimer_sleeper *timeout, int detect_deadlock,
785                         int (*slowfn)(struct rt_mutex *lock, int state,
786                                       struct hrtimer_sleeper *timeout,
787                                       int detect_deadlock))
788 {
789         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
790                 rt_mutex_deadlock_account_lock(lock, current);
791                 return 0;
792         } else
793                 return slowfn(lock, state, timeout, detect_deadlock);
794 }
795
796 static inline int
797 rt_mutex_fasttrylock(struct rt_mutex *lock,
798                      int (*slowfn)(struct rt_mutex *lock))
799 {
800         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
801                 rt_mutex_deadlock_account_lock(lock, current);
802                 return 1;
803         }
804         return slowfn(lock);
805 }
806
807 static inline void
808 rt_mutex_fastunlock(struct rt_mutex *lock,
809                     void (*slowfn)(struct rt_mutex *lock))
810 {
811         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
812                 rt_mutex_deadlock_account_unlock(current);
813         else
814                 slowfn(lock);
815 }
816
817 /**
818  * rt_mutex_lock - lock a rt_mutex
819  *
820  * @lock: the rt_mutex to be locked
821  */
822 void __sched rt_mutex_lock(struct rt_mutex *lock)
823 {
824         might_sleep();
825
826         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
827 }
828 EXPORT_SYMBOL_GPL(rt_mutex_lock);
829
830 /**
831  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
832  *
833  * @lock:               the rt_mutex to be locked
834  * @detect_deadlock:    deadlock detection on/off
835  *
836  * Returns:
837  *  0           on success
838  * -EINTR       when interrupted by a signal
839  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
840  */
841 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
842                                                  int detect_deadlock)
843 {
844         might_sleep();
845
846         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
847                                  detect_deadlock, rt_mutex_slowlock);
848 }
849 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
850
851 /**
852  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
853  *                                     the timeout structure is provided
854  *                                     by the caller
855  *
856  * @lock:               the rt_mutex to be locked
857  * @timeout:            timeout structure or NULL (no timeout)
858  * @detect_deadlock:    deadlock detection on/off
859  *
860  * Returns:
861  *  0           on success
862  * -EINTR       when interrupted by a signal
863  * -ETIMEOUT    when the timeout expired
864  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
865  */
866 int
867 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
868                     int detect_deadlock)
869 {
870         might_sleep();
871
872         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
873                                        detect_deadlock, rt_mutex_slowlock);
874 }
875 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
876
877 /**
878  * rt_mutex_trylock - try to lock a rt_mutex
879  *
880  * @lock:       the rt_mutex to be locked
881  *
882  * Returns 1 on success and 0 on contention
883  */
884 int __sched rt_mutex_trylock(struct rt_mutex *lock)
885 {
886         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
887 }
888 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
889
890 /**
891  * rt_mutex_unlock - unlock a rt_mutex
892  *
893  * @lock: the rt_mutex to be unlocked
894  */
895 void __sched rt_mutex_unlock(struct rt_mutex *lock)
896 {
897         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
898 }
899 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
900
901 /***
902  * rt_mutex_destroy - mark a mutex unusable
903  * @lock: the mutex to be destroyed
904  *
905  * This function marks the mutex uninitialized, and any subsequent
906  * use of the mutex is forbidden. The mutex must not be locked when
907  * this function is called.
908  */
909 void rt_mutex_destroy(struct rt_mutex *lock)
910 {
911         WARN_ON(rt_mutex_is_locked(lock));
912 #ifdef CONFIG_DEBUG_RT_MUTEXES
913         lock->magic = NULL;
914 #endif
915 }
916
917 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
918
919 /**
920  * __rt_mutex_init - initialize the rt lock
921  *
922  * @lock: the rt lock to be initialized
923  *
924  * Initialize the rt lock to unlocked state.
925  *
926  * Initializing of a locked rt lock is not allowed
927  */
928 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
929 {
930         lock->owner = NULL;
931         spin_lock_init(&lock->wait_lock);
932         plist_head_init(&lock->wait_list, &lock->wait_lock);
933
934         debug_rt_mutex_init(lock, name);
935 }
936 EXPORT_SYMBOL_GPL(__rt_mutex_init);
937
938 /**
939  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
940  *                              proxy owner
941  *
942  * @lock:       the rt_mutex to be locked
943  * @proxy_owner:the task to set as owner
944  *
945  * No locking. Caller has to do serializing itself
946  * Special API call for PI-futex support
947  */
948 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
949                                 struct task_struct *proxy_owner)
950 {
951         __rt_mutex_init(lock, NULL);
952         debug_rt_mutex_proxy_lock(lock, proxy_owner);
953         rt_mutex_set_owner(lock, proxy_owner, 0);
954         rt_mutex_deadlock_account_lock(lock, proxy_owner);
955 }
956
957 /**
958  * rt_mutex_proxy_unlock - release a lock on behalf of owner
959  *
960  * @lock:       the rt_mutex to be locked
961  *
962  * No locking. Caller has to do serializing itself
963  * Special API call for PI-futex support
964  */
965 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
966                            struct task_struct *proxy_owner)
967 {
968         debug_rt_mutex_proxy_unlock(lock);
969         rt_mutex_set_owner(lock, NULL, 0);
970         rt_mutex_deadlock_account_unlock(proxy_owner);
971 }
972
973 /**
974  * rt_mutex_next_owner - return the next owner of the lock
975  *
976  * @lock: the rt lock query
977  *
978  * Returns the next owner of the lock or NULL
979  *
980  * Caller has to serialize against other accessors to the lock
981  * itself.
982  *
983  * Special API call for PI-futex support
984  */
985 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
986 {
987         if (!rt_mutex_has_waiters(lock))
988                 return NULL;
989
990         return rt_mutex_top_waiter(lock)->task;
991 }