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