Revert "futex_requeue_pi optimization"
[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  *  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         /*
216          * Check the orig_waiter state. After we dropped the locks,
217          * the previous owner of the lock might have released the lock
218          * and made us the pending owner:
219          */
220         if (orig_waiter && !orig_waiter->task)
221                 goto out_unlock_pi;
222
223         /*
224          * Drop out, when the task has no waiters. Note,
225          * top_waiter can be NULL, when we are in the deboosting
226          * mode!
227          */
228         if (top_waiter && (!task_has_pi_waiters(task) ||
229                            top_waiter != task_top_pi_waiter(task)))
230                 goto out_unlock_pi;
231
232         /*
233          * When deadlock detection is off then we check, if further
234          * priority adjustment is necessary.
235          */
236         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
237                 goto out_unlock_pi;
238
239         lock = waiter->lock;
240         if (!spin_trylock(&lock->wait_lock)) {
241                 spin_unlock_irqrestore(&task->pi_lock, flags);
242                 cpu_relax();
243                 goto retry;
244         }
245
246         /* Deadlock detection */
247         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
248                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
249                 spin_unlock(&lock->wait_lock);
250                 ret = deadlock_detect ? -EDEADLK : 0;
251                 goto out_unlock_pi;
252         }
253
254         top_waiter = rt_mutex_top_waiter(lock);
255
256         /* Requeue the waiter */
257         plist_del(&waiter->list_entry, &lock->wait_list);
258         waiter->list_entry.prio = task->prio;
259         plist_add(&waiter->list_entry, &lock->wait_list);
260
261         /* Release the task */
262         spin_unlock_irqrestore(&task->pi_lock, flags);
263         put_task_struct(task);
264
265         /* Grab the next task */
266         task = rt_mutex_owner(lock);
267         get_task_struct(task);
268         spin_lock_irqsave(&task->pi_lock, flags);
269
270         if (waiter == rt_mutex_top_waiter(lock)) {
271                 /* Boost the owner */
272                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
273                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
274                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
275                 __rt_mutex_adjust_prio(task);
276
277         } else if (top_waiter == waiter) {
278                 /* Deboost the owner */
279                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
280                 waiter = rt_mutex_top_waiter(lock);
281                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
282                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
283                 __rt_mutex_adjust_prio(task);
284         }
285
286         spin_unlock_irqrestore(&task->pi_lock, flags);
287
288         top_waiter = rt_mutex_top_waiter(lock);
289         spin_unlock(&lock->wait_lock);
290
291         if (!detect_deadlock && waiter != top_waiter)
292                 goto out_put_task;
293
294         goto again;
295
296  out_unlock_pi:
297         spin_unlock_irqrestore(&task->pi_lock, flags);
298  out_put_task:
299         put_task_struct(task);
300
301         return ret;
302 }
303
304 /*
305  * Optimization: check if we can steal the lock from the
306  * assigned pending owner [which might not have taken the
307  * lock yet]:
308  */
309 static inline int try_to_steal_lock(struct rt_mutex *lock)
310 {
311         struct task_struct *pendowner = rt_mutex_owner(lock);
312         struct rt_mutex_waiter *next;
313         unsigned long flags;
314
315         if (!rt_mutex_owner_pending(lock))
316                 return 0;
317
318         if (pendowner == current)
319                 return 1;
320
321         spin_lock_irqsave(&pendowner->pi_lock, flags);
322         if (current->prio >= pendowner->prio) {
323                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
324                 return 0;
325         }
326
327         /*
328          * Check if a waiter is enqueued on the pending owners
329          * pi_waiters list. Remove it and readjust pending owners
330          * priority.
331          */
332         if (likely(!rt_mutex_has_waiters(lock))) {
333                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
334                 return 1;
335         }
336
337         /* No chain handling, pending owner is not blocked on anything: */
338         next = rt_mutex_top_waiter(lock);
339         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
340         __rt_mutex_adjust_prio(pendowner);
341         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
342
343         /*
344          * We are going to steal the lock and a waiter was
345          * enqueued on the pending owners pi_waiters queue. So
346          * we have to enqueue this waiter into
347          * current->pi_waiters list. This covers the case,
348          * where current is boosted because it holds another
349          * lock and gets unboosted because the booster is
350          * interrupted, so we would delay a waiter with higher
351          * priority as current->normal_prio.
352          *
353          * Note: in the rare case of a SCHED_OTHER task changing
354          * its priority and thus stealing the lock, next->task
355          * might be current:
356          */
357         if (likely(next->task != current)) {
358                 spin_lock_irqsave(&current->pi_lock, flags);
359                 plist_add(&next->pi_list_entry, &current->pi_waiters);
360                 __rt_mutex_adjust_prio(current);
361                 spin_unlock_irqrestore(&current->pi_lock, flags);
362         }
363         return 1;
364 }
365
366 /*
367  * Try to take an rt-mutex
368  *
369  * This fails
370  * - when the lock has a real owner
371  * - when a different pending owner exists and has higher priority than current
372  *
373  * Must be called with lock->wait_lock held.
374  */
375 static int try_to_take_rt_mutex(struct rt_mutex *lock)
376 {
377         /*
378          * We have to be careful here if the atomic speedups are
379          * enabled, such that, when
380          *  - no other waiter is on the lock
381          *  - the lock has been released since we did the cmpxchg
382          * the lock can be released or taken while we are doing the
383          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
384          *
385          * The atomic acquire/release aware variant of
386          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
387          * the WAITERS bit, the atomic release / acquire can not
388          * happen anymore and lock->wait_lock protects us from the
389          * non-atomic case.
390          *
391          * Note, that this might set lock->owner =
392          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
393          * any more. This is fixed up when we take the ownership.
394          * This is the transitional state explained at the top of this file.
395          */
396         mark_rt_mutex_waiters(lock);
397
398         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
399                 return 0;
400
401         /* We got the lock. */
402         debug_rt_mutex_lock(lock);
403
404         rt_mutex_set_owner(lock, current, 0);
405
406         rt_mutex_deadlock_account_lock(lock, current);
407
408         return 1;
409 }
410
411 /*
412  * Task blocks on lock.
413  *
414  * Prepare waiter and propagate pi chain
415  *
416  * This must be called with lock->wait_lock held.
417  */
418 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
419                                    struct rt_mutex_waiter *waiter,
420                                    int detect_deadlock)
421 {
422         struct task_struct *owner = rt_mutex_owner(lock);
423         struct rt_mutex_waiter *top_waiter = waiter;
424         unsigned long flags;
425         int chain_walk = 0, res;
426
427         spin_lock_irqsave(&current->pi_lock, flags);
428         __rt_mutex_adjust_prio(current);
429         waiter->task = current;
430         waiter->lock = lock;
431         plist_node_init(&waiter->list_entry, current->prio);
432         plist_node_init(&waiter->pi_list_entry, current->prio);
433
434         /* Get the top priority waiter on the lock */
435         if (rt_mutex_has_waiters(lock))
436                 top_waiter = rt_mutex_top_waiter(lock);
437         plist_add(&waiter->list_entry, &lock->wait_list);
438
439         current->pi_blocked_on = waiter;
440
441         spin_unlock_irqrestore(&current->pi_lock, flags);
442
443         if (waiter == rt_mutex_top_waiter(lock)) {
444                 spin_lock_irqsave(&owner->pi_lock, flags);
445                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
446                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
447
448                 __rt_mutex_adjust_prio(owner);
449                 if (owner->pi_blocked_on)
450                         chain_walk = 1;
451                 spin_unlock_irqrestore(&owner->pi_lock, flags);
452         }
453         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
454                 chain_walk = 1;
455
456         if (!chain_walk)
457                 return 0;
458
459         /*
460          * The owner can't disappear while holding a lock,
461          * so the owner struct is protected by wait_lock.
462          * Gets dropped in rt_mutex_adjust_prio_chain()!
463          */
464         get_task_struct(owner);
465
466         spin_unlock(&lock->wait_lock);
467
468         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
469                                          current);
470
471         spin_lock(&lock->wait_lock);
472
473         return res;
474 }
475
476 /*
477  * Wake up the next waiter on the lock.
478  *
479  * Remove the top waiter from the current tasks waiter list and from
480  * the lock waiter list. Set it as pending owner. Then wake it up.
481  *
482  * Called with lock->wait_lock held.
483  */
484 static void wakeup_next_waiter(struct rt_mutex *lock)
485 {
486         struct rt_mutex_waiter *waiter;
487         struct task_struct *pendowner;
488         unsigned long flags;
489
490         spin_lock_irqsave(&current->pi_lock, flags);
491
492         waiter = rt_mutex_top_waiter(lock);
493         plist_del(&waiter->list_entry, &lock->wait_list);
494
495         /*
496          * Remove it from current->pi_waiters. We do not adjust a
497          * possible priority boost right now. We execute wakeup in the
498          * boosted mode and go back to normal after releasing
499          * lock->wait_lock.
500          */
501         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
502         pendowner = waiter->task;
503         waiter->task = NULL;
504
505         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
506
507         spin_unlock_irqrestore(&current->pi_lock, flags);
508
509         /*
510          * Clear the pi_blocked_on variable and enqueue a possible
511          * waiter into the pi_waiters list of the pending owner. This
512          * prevents that in case the pending owner gets unboosted a
513          * waiter with higher priority than pending-owner->normal_prio
514          * is blocked on the unboosted (pending) owner.
515          */
516         spin_lock_irqsave(&pendowner->pi_lock, flags);
517
518         WARN_ON(!pendowner->pi_blocked_on);
519         WARN_ON(pendowner->pi_blocked_on != waiter);
520         WARN_ON(pendowner->pi_blocked_on->lock != lock);
521
522         pendowner->pi_blocked_on = NULL;
523
524         if (rt_mutex_has_waiters(lock)) {
525                 struct rt_mutex_waiter *next;
526
527                 next = rt_mutex_top_waiter(lock);
528                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
529         }
530         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
531
532         wake_up_process(pendowner);
533 }
534
535 /*
536  * Remove a waiter from a lock
537  *
538  * Must be called with lock->wait_lock held
539  */
540 static void remove_waiter(struct rt_mutex *lock,
541                           struct rt_mutex_waiter *waiter)
542 {
543         int first = (waiter == rt_mutex_top_waiter(lock));
544         struct task_struct *owner = rt_mutex_owner(lock);
545         unsigned long flags;
546         int chain_walk = 0;
547
548         spin_lock_irqsave(&current->pi_lock, flags);
549         plist_del(&waiter->list_entry, &lock->wait_list);
550         waiter->task = NULL;
551         current->pi_blocked_on = NULL;
552         spin_unlock_irqrestore(&current->pi_lock, flags);
553
554         if (first && owner != current) {
555
556                 spin_lock_irqsave(&owner->pi_lock, flags);
557
558                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
559
560                 if (rt_mutex_has_waiters(lock)) {
561                         struct rt_mutex_waiter *next;
562
563                         next = rt_mutex_top_waiter(lock);
564                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
565                 }
566                 __rt_mutex_adjust_prio(owner);
567
568                 if (owner->pi_blocked_on)
569                         chain_walk = 1;
570
571                 spin_unlock_irqrestore(&owner->pi_lock, flags);
572         }
573
574         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
575
576         if (!chain_walk)
577                 return;
578
579         /* gets dropped in rt_mutex_adjust_prio_chain()! */
580         get_task_struct(owner);
581
582         spin_unlock(&lock->wait_lock);
583
584         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
585
586         spin_lock(&lock->wait_lock);
587 }
588
589 /*
590  * Recheck the pi chain, in case we got a priority setting
591  *
592  * Called from sched_setscheduler
593  */
594 void rt_mutex_adjust_pi(struct task_struct *task)
595 {
596         struct rt_mutex_waiter *waiter;
597         unsigned long flags;
598
599         spin_lock_irqsave(&task->pi_lock, flags);
600
601         waiter = task->pi_blocked_on;
602         if (!waiter || waiter->list_entry.prio == task->prio) {
603                 spin_unlock_irqrestore(&task->pi_lock, flags);
604                 return;
605         }
606
607         spin_unlock_irqrestore(&task->pi_lock, flags);
608
609         /* gets dropped in rt_mutex_adjust_prio_chain()! */
610         get_task_struct(task);
611         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
612 }
613
614 /*
615  * Slow path lock function:
616  */
617 static int __sched
618 rt_mutex_slowlock(struct rt_mutex *lock, int state,
619                   struct hrtimer_sleeper *timeout,
620                   int detect_deadlock)
621 {
622         struct rt_mutex_waiter waiter;
623         int ret = 0;
624
625         debug_rt_mutex_init_waiter(&waiter);
626         waiter.task = NULL;
627
628         spin_lock(&lock->wait_lock);
629
630         /* Try to acquire the lock again: */
631         if (try_to_take_rt_mutex(lock)) {
632                 spin_unlock(&lock->wait_lock);
633                 return 0;
634         }
635
636         set_current_state(state);
637
638         /* Setup the timer, when timeout != NULL */
639         if (unlikely(timeout))
640                 hrtimer_start(&timeout->timer, timeout->timer.expires,
641                               HRTIMER_MODE_ABS);
642
643         for (;;) {
644                 /* Try to acquire the lock: */
645                 if (try_to_take_rt_mutex(lock))
646                         break;
647
648                 /*
649                  * TASK_INTERRUPTIBLE checks for signals and
650                  * timeout. Ignored otherwise.
651                  */
652                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
653                         /* Signal pending? */
654                         if (signal_pending(current))
655                                 ret = -EINTR;
656                         if (timeout && !timeout->task)
657                                 ret = -ETIMEDOUT;
658                         if (ret)
659                                 break;
660                 }
661
662                 /*
663                  * waiter.task is NULL the first time we come here and
664                  * when we have been woken up by the previous owner
665                  * but the lock got stolen by a higher prio task.
666                  */
667                 if (!waiter.task) {
668                         ret = task_blocks_on_rt_mutex(lock, &waiter,
669                                                       detect_deadlock);
670                         /*
671                          * If we got woken up by the owner then start loop
672                          * all over without going into schedule to try
673                          * to get the lock now:
674                          */
675                         if (unlikely(!waiter.task)) {
676                                 /*
677                                  * Reset the return value. We might
678                                  * have returned with -EDEADLK and the
679                                  * owner released the lock while we
680                                  * were walking the pi chain.
681                                  */
682                                 ret = 0;
683                                 continue;
684                         }
685                         if (unlikely(ret))
686                                 break;
687                 }
688
689                 spin_unlock(&lock->wait_lock);
690
691                 debug_rt_mutex_print_deadlock(&waiter);
692
693                 if (waiter.task)
694                         schedule_rt_mutex(lock);
695
696                 spin_lock(&lock->wait_lock);
697                 set_current_state(state);
698         }
699
700         set_current_state(TASK_RUNNING);
701
702         if (unlikely(waiter.task))
703                 remove_waiter(lock, &waiter);
704
705         /*
706          * try_to_take_rt_mutex() sets the waiter bit
707          * unconditionally. We might have to fix that up.
708          */
709         fixup_rt_mutex_waiters(lock);
710
711         spin_unlock(&lock->wait_lock);
712
713         /* Remove pending timer: */
714         if (unlikely(timeout))
715                 hrtimer_cancel(&timeout->timer);
716
717         /*
718          * Readjust priority, when we did not get the lock. We might
719          * have been the pending owner and boosted. Since we did not
720          * take the lock, the PI boost has to go.
721          */
722         if (unlikely(ret))
723                 rt_mutex_adjust_prio(current);
724
725         debug_rt_mutex_free_waiter(&waiter);
726
727         return ret;
728 }
729
730 /*
731  * Slow path try-lock function:
732  */
733 static inline int
734 rt_mutex_slowtrylock(struct rt_mutex *lock)
735 {
736         int ret = 0;
737
738         spin_lock(&lock->wait_lock);
739
740         if (likely(rt_mutex_owner(lock) != current)) {
741
742                 ret = try_to_take_rt_mutex(lock);
743                 /*
744                  * try_to_take_rt_mutex() sets the lock waiters
745                  * bit unconditionally. Clean this up.
746                  */
747                 fixup_rt_mutex_waiters(lock);
748         }
749
750         spin_unlock(&lock->wait_lock);
751
752         return ret;
753 }
754
755 /*
756  * Slow path to release a rt-mutex:
757  */
758 static void __sched
759 rt_mutex_slowunlock(struct rt_mutex *lock)
760 {
761         spin_lock(&lock->wait_lock);
762
763         debug_rt_mutex_unlock(lock);
764
765         rt_mutex_deadlock_account_unlock(current);
766
767         if (!rt_mutex_has_waiters(lock)) {
768                 lock->owner = NULL;
769                 spin_unlock(&lock->wait_lock);
770                 return;
771         }
772
773         wakeup_next_waiter(lock);
774
775         spin_unlock(&lock->wait_lock);
776
777         /* Undo pi boosting if necessary: */
778         rt_mutex_adjust_prio(current);
779 }
780
781 /*
782  * debug aware fast / slowpath lock,trylock,unlock
783  *
784  * The atomic acquire/release ops are compiled away, when either the
785  * architecture does not support cmpxchg or when debugging is enabled.
786  */
787 static inline int
788 rt_mutex_fastlock(struct rt_mutex *lock, int state,
789                   int detect_deadlock,
790                   int (*slowfn)(struct rt_mutex *lock, int state,
791                                 struct hrtimer_sleeper *timeout,
792                                 int detect_deadlock))
793 {
794         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
795                 rt_mutex_deadlock_account_lock(lock, current);
796                 return 0;
797         } else
798                 return slowfn(lock, state, NULL, detect_deadlock);
799 }
800
801 static inline int
802 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
803                         struct hrtimer_sleeper *timeout, int detect_deadlock,
804                         int (*slowfn)(struct rt_mutex *lock, int state,
805                                       struct hrtimer_sleeper *timeout,
806                                       int detect_deadlock))
807 {
808         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
809                 rt_mutex_deadlock_account_lock(lock, current);
810                 return 0;
811         } else
812                 return slowfn(lock, state, timeout, detect_deadlock);
813 }
814
815 static inline int
816 rt_mutex_fasttrylock(struct rt_mutex *lock,
817                      int (*slowfn)(struct rt_mutex *lock))
818 {
819         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
820                 rt_mutex_deadlock_account_lock(lock, current);
821                 return 1;
822         }
823         return slowfn(lock);
824 }
825
826 static inline void
827 rt_mutex_fastunlock(struct rt_mutex *lock,
828                     void (*slowfn)(struct rt_mutex *lock))
829 {
830         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
831                 rt_mutex_deadlock_account_unlock(current);
832         else
833                 slowfn(lock);
834 }
835
836 /**
837  * rt_mutex_lock - lock a rt_mutex
838  *
839  * @lock: the rt_mutex to be locked
840  */
841 void __sched rt_mutex_lock(struct rt_mutex *lock)
842 {
843         might_sleep();
844
845         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
846 }
847 EXPORT_SYMBOL_GPL(rt_mutex_lock);
848
849 /**
850  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
851  *
852  * @lock:               the rt_mutex to be locked
853  * @detect_deadlock:    deadlock detection on/off
854  *
855  * Returns:
856  *  0           on success
857  * -EINTR       when interrupted by a signal
858  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
859  */
860 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
861                                                  int detect_deadlock)
862 {
863         might_sleep();
864
865         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
866                                  detect_deadlock, rt_mutex_slowlock);
867 }
868 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
869
870 /**
871  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
872  *                                     the timeout structure is provided
873  *                                     by the caller
874  *
875  * @lock:               the rt_mutex to be locked
876  * @timeout:            timeout structure or NULL (no timeout)
877  * @detect_deadlock:    deadlock detection on/off
878  *
879  * Returns:
880  *  0           on success
881  * -EINTR       when interrupted by a signal
882  * -ETIMEOUT    when the timeout expired
883  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
884  */
885 int
886 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
887                     int detect_deadlock)
888 {
889         might_sleep();
890
891         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
892                                        detect_deadlock, rt_mutex_slowlock);
893 }
894 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
895
896 /**
897  * rt_mutex_trylock - try to lock a rt_mutex
898  *
899  * @lock:       the rt_mutex to be locked
900  *
901  * Returns 1 on success and 0 on contention
902  */
903 int __sched rt_mutex_trylock(struct rt_mutex *lock)
904 {
905         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
906 }
907 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
908
909 /**
910  * rt_mutex_unlock - unlock a rt_mutex
911  *
912  * @lock: the rt_mutex to be unlocked
913  */
914 void __sched rt_mutex_unlock(struct rt_mutex *lock)
915 {
916         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
917 }
918 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
919
920 /***
921  * rt_mutex_destroy - mark a mutex unusable
922  * @lock: the mutex to be destroyed
923  *
924  * This function marks the mutex uninitialized, and any subsequent
925  * use of the mutex is forbidden. The mutex must not be locked when
926  * this function is called.
927  */
928 void rt_mutex_destroy(struct rt_mutex *lock)
929 {
930         WARN_ON(rt_mutex_is_locked(lock));
931 #ifdef CONFIG_DEBUG_RT_MUTEXES
932         lock->magic = NULL;
933 #endif
934 }
935
936 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
937
938 /**
939  * __rt_mutex_init - initialize the rt lock
940  *
941  * @lock: the rt lock to be initialized
942  *
943  * Initialize the rt lock to unlocked state.
944  *
945  * Initializing of a locked rt lock is not allowed
946  */
947 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
948 {
949         lock->owner = NULL;
950         spin_lock_init(&lock->wait_lock);
951         plist_head_init(&lock->wait_list, &lock->wait_lock);
952
953         debug_rt_mutex_init(lock, name);
954 }
955 EXPORT_SYMBOL_GPL(__rt_mutex_init);
956
957 /**
958  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
959  *                              proxy owner
960  *
961  * @lock:       the rt_mutex to be locked
962  * @proxy_owner:the task to set as owner
963  *
964  * No locking. Caller has to do serializing itself
965  * Special API call for PI-futex support
966  */
967 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
968                                 struct task_struct *proxy_owner)
969 {
970         __rt_mutex_init(lock, NULL);
971         debug_rt_mutex_proxy_lock(lock, proxy_owner);
972         rt_mutex_set_owner(lock, proxy_owner, 0);
973         rt_mutex_deadlock_account_lock(lock, proxy_owner);
974 }
975
976 /**
977  * rt_mutex_proxy_unlock - release a lock on behalf of owner
978  *
979  * @lock:       the rt_mutex to be locked
980  *
981  * No locking. Caller has to do serializing itself
982  * Special API call for PI-futex support
983  */
984 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
985                            struct task_struct *proxy_owner)
986 {
987         debug_rt_mutex_proxy_unlock(lock);
988         rt_mutex_set_owner(lock, NULL, 0);
989         rt_mutex_deadlock_account_unlock(proxy_owner);
990 }
991
992 /**
993  * rt_mutex_next_owner - return the next owner of the lock
994  *
995  * @lock: the rt lock query
996  *
997  * Returns the next owner of the lock or NULL
998  *
999  * Caller has to serialize against other accessors to the lock
1000  * itself.
1001  *
1002  * Special API call for PI-futex support
1003  */
1004 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1005 {
1006         if (!rt_mutex_has_waiters(lock))
1007                 return NULL;
1008
1009         return rt_mutex_top_waiter(lock)->task;
1010 }