]> git.karo-electronics.de Git - karo-tx-linux.git/blob - kernel/locking/rtmutex.c
Merge branch 'pm-cpufreq'
[karo-tx-linux.git] / kernel / locking / 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/export.h>
15 #include <linux/sched.h>
16 #include <linux/sched/rt.h>
17 #include <linux/sched/deadline.h>
18 #include <linux/timer.h>
19
20 #include "rtmutex_common.h"
21
22 /*
23  * lock->owner state tracking:
24  *
25  * lock->owner holds the task_struct pointer of the owner. Bit 0
26  * is used to keep track of the "lock has waiters" state.
27  *
28  * owner        bit0
29  * NULL         0       lock is free (fast acquire possible)
30  * NULL         1       lock is free and has waiters and the top waiter
31  *                              is going to take the lock*
32  * taskpointer  0       lock is held (fast release possible)
33  * taskpointer  1       lock is held and has waiters**
34  *
35  * The fast atomic compare exchange based acquire and release is only
36  * possible when bit 0 of lock->owner is 0.
37  *
38  * (*) It also can be a transitional state when grabbing the lock
39  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
40  * we need to set the bit0 before looking at the lock, and the owner may be
41  * NULL in this small time, hence this can be a transitional state.
42  *
43  * (**) There is a small time when bit 0 is set but there are no
44  * waiters. This can happen when grabbing the lock in the slow path.
45  * To prevent a cmpxchg of the owner releasing the lock, we need to
46  * set this bit before looking at the lock.
47  */
48
49 static void
50 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
51 {
52         unsigned long val = (unsigned long)owner;
53
54         if (rt_mutex_has_waiters(lock))
55                 val |= RT_MUTEX_HAS_WAITERS;
56
57         lock->owner = (struct task_struct *)val;
58 }
59
60 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
61 {
62         lock->owner = (struct task_struct *)
63                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
64 }
65
66 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
67 {
68         if (!rt_mutex_has_waiters(lock))
69                 clear_rt_mutex_waiters(lock);
70 }
71
72 /*
73  * We can speed up the acquire/release, if the architecture
74  * supports cmpxchg and if there's no debugging state to be set up
75  */
76 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
77 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
78 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
79 {
80         unsigned long owner, *p = (unsigned long *) &lock->owner;
81
82         do {
83                 owner = *p;
84         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
85 }
86
87 /*
88  * Safe fastpath aware unlock:
89  * 1) Clear the waiters bit
90  * 2) Drop lock->wait_lock
91  * 3) Try to unlock the lock with cmpxchg
92  */
93 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
94         __releases(lock->wait_lock)
95 {
96         struct task_struct *owner = rt_mutex_owner(lock);
97
98         clear_rt_mutex_waiters(lock);
99         raw_spin_unlock(&lock->wait_lock);
100         /*
101          * If a new waiter comes in between the unlock and the cmpxchg
102          * we have two situations:
103          *
104          * unlock(wait_lock);
105          *                                      lock(wait_lock);
106          * cmpxchg(p, owner, 0) == owner
107          *                                      mark_rt_mutex_waiters(lock);
108          *                                      acquire(lock);
109          * or:
110          *
111          * unlock(wait_lock);
112          *                                      lock(wait_lock);
113          *                                      mark_rt_mutex_waiters(lock);
114          *
115          * cmpxchg(p, owner, 0) != owner
116          *                                      enqueue_waiter();
117          *                                      unlock(wait_lock);
118          * lock(wait_lock);
119          * wake waiter();
120          * unlock(wait_lock);
121          *                                      lock(wait_lock);
122          *                                      acquire(lock);
123          */
124         return rt_mutex_cmpxchg(lock, owner, NULL);
125 }
126
127 #else
128 # define rt_mutex_cmpxchg(l,c,n)        (0)
129 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
130 {
131         lock->owner = (struct task_struct *)
132                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
133 }
134
135 /*
136  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
137  */
138 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
139         __releases(lock->wait_lock)
140 {
141         lock->owner = NULL;
142         raw_spin_unlock(&lock->wait_lock);
143         return true;
144 }
145 #endif
146
147 static inline int
148 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
149                      struct rt_mutex_waiter *right)
150 {
151         if (left->prio < right->prio)
152                 return 1;
153
154         /*
155          * If both waiters have dl_prio(), we check the deadlines of the
156          * associated tasks.
157          * If left waiter has a dl_prio(), and we didn't return 1 above,
158          * then right waiter has a dl_prio() too.
159          */
160         if (dl_prio(left->prio))
161                 return (left->task->dl.deadline < right->task->dl.deadline);
162
163         return 0;
164 }
165
166 static void
167 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
168 {
169         struct rb_node **link = &lock->waiters.rb_node;
170         struct rb_node *parent = NULL;
171         struct rt_mutex_waiter *entry;
172         int leftmost = 1;
173
174         while (*link) {
175                 parent = *link;
176                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
177                 if (rt_mutex_waiter_less(waiter, entry)) {
178                         link = &parent->rb_left;
179                 } else {
180                         link = &parent->rb_right;
181                         leftmost = 0;
182                 }
183         }
184
185         if (leftmost)
186                 lock->waiters_leftmost = &waiter->tree_entry;
187
188         rb_link_node(&waiter->tree_entry, parent, link);
189         rb_insert_color(&waiter->tree_entry, &lock->waiters);
190 }
191
192 static void
193 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
194 {
195         if (RB_EMPTY_NODE(&waiter->tree_entry))
196                 return;
197
198         if (lock->waiters_leftmost == &waiter->tree_entry)
199                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
200
201         rb_erase(&waiter->tree_entry, &lock->waiters);
202         RB_CLEAR_NODE(&waiter->tree_entry);
203 }
204
205 static void
206 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
207 {
208         struct rb_node **link = &task->pi_waiters.rb_node;
209         struct rb_node *parent = NULL;
210         struct rt_mutex_waiter *entry;
211         int leftmost = 1;
212
213         while (*link) {
214                 parent = *link;
215                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
216                 if (rt_mutex_waiter_less(waiter, entry)) {
217                         link = &parent->rb_left;
218                 } else {
219                         link = &parent->rb_right;
220                         leftmost = 0;
221                 }
222         }
223
224         if (leftmost)
225                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
226
227         rb_link_node(&waiter->pi_tree_entry, parent, link);
228         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
229 }
230
231 static void
232 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
233 {
234         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
235                 return;
236
237         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
238                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
239
240         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
241         RB_CLEAR_NODE(&waiter->pi_tree_entry);
242 }
243
244 /*
245  * Calculate task priority from the waiter tree priority
246  *
247  * Return task->normal_prio when the waiter tree is empty or when
248  * the waiter is not allowed to do priority boosting
249  */
250 int rt_mutex_getprio(struct task_struct *task)
251 {
252         if (likely(!task_has_pi_waiters(task)))
253                 return task->normal_prio;
254
255         return min(task_top_pi_waiter(task)->prio,
256                    task->normal_prio);
257 }
258
259 struct task_struct *rt_mutex_get_top_task(struct task_struct *task)
260 {
261         if (likely(!task_has_pi_waiters(task)))
262                 return NULL;
263
264         return task_top_pi_waiter(task)->task;
265 }
266
267 /*
268  * Called by sched_setscheduler() to check whether the priority change
269  * is overruled by a possible priority boosting.
270  */
271 int rt_mutex_check_prio(struct task_struct *task, int newprio)
272 {
273         if (!task_has_pi_waiters(task))
274                 return 0;
275
276         return task_top_pi_waiter(task)->task->prio <= newprio;
277 }
278
279 /*
280  * Adjust the priority of a task, after its pi_waiters got modified.
281  *
282  * This can be both boosting and unboosting. task->pi_lock must be held.
283  */
284 static void __rt_mutex_adjust_prio(struct task_struct *task)
285 {
286         int prio = rt_mutex_getprio(task);
287
288         if (task->prio != prio || dl_prio(prio))
289                 rt_mutex_setprio(task, prio);
290 }
291
292 /*
293  * Adjust task priority (undo boosting). Called from the exit path of
294  * rt_mutex_slowunlock() and rt_mutex_slowlock().
295  *
296  * (Note: We do this outside of the protection of lock->wait_lock to
297  * allow the lock to be taken while or before we readjust the priority
298  * of task. We do not use the spin_xx_mutex() variants here as we are
299  * outside of the debug path.)
300  */
301 static void rt_mutex_adjust_prio(struct task_struct *task)
302 {
303         unsigned long flags;
304
305         raw_spin_lock_irqsave(&task->pi_lock, flags);
306         __rt_mutex_adjust_prio(task);
307         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
308 }
309
310 /*
311  * Max number of times we'll walk the boosting chain:
312  */
313 int max_lock_depth = 1024;
314
315 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
316 {
317         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
318 }
319
320 /*
321  * Adjust the priority chain. Also used for deadlock detection.
322  * Decreases task's usage by one - may thus free the task.
323  *
324  * @task:       the task owning the mutex (owner) for which a chain walk is
325  *              probably needed
326  * @deadlock_detect: do we have to carry out deadlock detection?
327  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
328  *              things for a task that has just got its priority adjusted, and
329  *              is waiting on a mutex)
330  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
331  *              we dropped its pi_lock. Is never dereferenced, only used for
332  *              comparison to detect lock chain changes.
333  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
334  *              its priority to the mutex owner (can be NULL in the case
335  *              depicted above or if the top waiter is gone away and we are
336  *              actually deboosting the owner)
337  * @top_task:   the current top waiter
338  *
339  * Returns 0 or -EDEADLK.
340  */
341 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
342                                       int deadlock_detect,
343                                       struct rt_mutex *orig_lock,
344                                       struct rt_mutex *next_lock,
345                                       struct rt_mutex_waiter *orig_waiter,
346                                       struct task_struct *top_task)
347 {
348         struct rt_mutex *lock;
349         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
350         int detect_deadlock, ret = 0, depth = 0;
351         unsigned long flags;
352
353         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
354                                                          deadlock_detect);
355
356         /*
357          * The (de)boosting is a step by step approach with a lot of
358          * pitfalls. We want this to be preemptible and we want hold a
359          * maximum of two locks per step. So we have to check
360          * carefully whether things change under us.
361          */
362  again:
363         if (++depth > max_lock_depth) {
364                 static int prev_max;
365
366                 /*
367                  * Print this only once. If the admin changes the limit,
368                  * print a new message when reaching the limit again.
369                  */
370                 if (prev_max != max_lock_depth) {
371                         prev_max = max_lock_depth;
372                         printk(KERN_WARNING "Maximum lock depth %d reached "
373                                "task: %s (%d)\n", max_lock_depth,
374                                top_task->comm, task_pid_nr(top_task));
375                 }
376                 put_task_struct(task);
377
378                 return -EDEADLK;
379         }
380  retry:
381         /*
382          * Task can not go away as we did a get_task() before !
383          */
384         raw_spin_lock_irqsave(&task->pi_lock, flags);
385
386         waiter = task->pi_blocked_on;
387         /*
388          * Check whether the end of the boosting chain has been
389          * reached or the state of the chain has changed while we
390          * dropped the locks.
391          */
392         if (!waiter)
393                 goto out_unlock_pi;
394
395         /*
396          * Check the orig_waiter state. After we dropped the locks,
397          * the previous owner of the lock might have released the lock.
398          */
399         if (orig_waiter && !rt_mutex_owner(orig_lock))
400                 goto out_unlock_pi;
401
402         /*
403          * We dropped all locks after taking a refcount on @task, so
404          * the task might have moved on in the lock chain or even left
405          * the chain completely and blocks now on an unrelated lock or
406          * on @orig_lock.
407          *
408          * We stored the lock on which @task was blocked in @next_lock,
409          * so we can detect the chain change.
410          */
411         if (next_lock != waiter->lock)
412                 goto out_unlock_pi;
413
414         /*
415          * Drop out, when the task has no waiters. Note,
416          * top_waiter can be NULL, when we are in the deboosting
417          * mode!
418          */
419         if (top_waiter) {
420                 if (!task_has_pi_waiters(task))
421                         goto out_unlock_pi;
422                 /*
423                  * If deadlock detection is off, we stop here if we
424                  * are not the top pi waiter of the task.
425                  */
426                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
427                         goto out_unlock_pi;
428         }
429
430         /*
431          * When deadlock detection is off then we check, if further
432          * priority adjustment is necessary.
433          */
434         if (!detect_deadlock && waiter->prio == task->prio)
435                 goto out_unlock_pi;
436
437         lock = waiter->lock;
438         if (!raw_spin_trylock(&lock->wait_lock)) {
439                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
440                 cpu_relax();
441                 goto retry;
442         }
443
444         /*
445          * Deadlock detection. If the lock is the same as the original
446          * lock which caused us to walk the lock chain or if the
447          * current lock is owned by the task which initiated the chain
448          * walk, we detected a deadlock.
449          */
450         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
451                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
452                 raw_spin_unlock(&lock->wait_lock);
453                 ret = -EDEADLK;
454                 goto out_unlock_pi;
455         }
456
457         top_waiter = rt_mutex_top_waiter(lock);
458
459         /* Requeue the waiter */
460         rt_mutex_dequeue(lock, waiter);
461         waiter->prio = task->prio;
462         rt_mutex_enqueue(lock, waiter);
463
464         /* Release the task */
465         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
466         if (!rt_mutex_owner(lock)) {
467                 /*
468                  * If the requeue above changed the top waiter, then we need
469                  * to wake the new top waiter up to try to get the lock.
470                  */
471
472                 if (top_waiter != rt_mutex_top_waiter(lock))
473                         wake_up_process(rt_mutex_top_waiter(lock)->task);
474                 raw_spin_unlock(&lock->wait_lock);
475                 goto out_put_task;
476         }
477         put_task_struct(task);
478
479         /* Grab the next task */
480         task = rt_mutex_owner(lock);
481         get_task_struct(task);
482         raw_spin_lock_irqsave(&task->pi_lock, flags);
483
484         if (waiter == rt_mutex_top_waiter(lock)) {
485                 /* Boost the owner */
486                 rt_mutex_dequeue_pi(task, top_waiter);
487                 rt_mutex_enqueue_pi(task, waiter);
488                 __rt_mutex_adjust_prio(task);
489
490         } else if (top_waiter == waiter) {
491                 /* Deboost the owner */
492                 rt_mutex_dequeue_pi(task, waiter);
493                 waiter = rt_mutex_top_waiter(lock);
494                 rt_mutex_enqueue_pi(task, waiter);
495                 __rt_mutex_adjust_prio(task);
496         }
497
498         /*
499          * Check whether the task which owns the current lock is pi
500          * blocked itself. If yes we store a pointer to the lock for
501          * the lock chain change detection above. After we dropped
502          * task->pi_lock next_lock cannot be dereferenced anymore.
503          */
504         next_lock = task_blocked_on_lock(task);
505
506         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
507
508         top_waiter = rt_mutex_top_waiter(lock);
509         raw_spin_unlock(&lock->wait_lock);
510
511         /*
512          * We reached the end of the lock chain. Stop right here. No
513          * point to go back just to figure that out.
514          */
515         if (!next_lock)
516                 goto out_put_task;
517
518         if (!detect_deadlock && waiter != top_waiter)
519                 goto out_put_task;
520
521         goto again;
522
523  out_unlock_pi:
524         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
525  out_put_task:
526         put_task_struct(task);
527
528         return ret;
529 }
530
531 /*
532  * Try to take an rt-mutex
533  *
534  * Must be called with lock->wait_lock held.
535  *
536  * @lock:   the lock to be acquired.
537  * @task:   the task which wants to acquire the lock
538  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
539  */
540 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
541                 struct rt_mutex_waiter *waiter)
542 {
543         /*
544          * We have to be careful here if the atomic speedups are
545          * enabled, such that, when
546          *  - no other waiter is on the lock
547          *  - the lock has been released since we did the cmpxchg
548          * the lock can be released or taken while we are doing the
549          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
550          *
551          * The atomic acquire/release aware variant of
552          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
553          * the WAITERS bit, the atomic release / acquire can not
554          * happen anymore and lock->wait_lock protects us from the
555          * non-atomic case.
556          *
557          * Note, that this might set lock->owner =
558          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
559          * any more. This is fixed up when we take the ownership.
560          * This is the transitional state explained at the top of this file.
561          */
562         mark_rt_mutex_waiters(lock);
563
564         if (rt_mutex_owner(lock))
565                 return 0;
566
567         /*
568          * It will get the lock because of one of these conditions:
569          * 1) there is no waiter
570          * 2) higher priority than waiters
571          * 3) it is top waiter
572          */
573         if (rt_mutex_has_waiters(lock)) {
574                 if (task->prio >= rt_mutex_top_waiter(lock)->prio) {
575                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
576                                 return 0;
577                 }
578         }
579
580         if (waiter || rt_mutex_has_waiters(lock)) {
581                 unsigned long flags;
582                 struct rt_mutex_waiter *top;
583
584                 raw_spin_lock_irqsave(&task->pi_lock, flags);
585
586                 /* remove the queued waiter. */
587                 if (waiter) {
588                         rt_mutex_dequeue(lock, waiter);
589                         task->pi_blocked_on = NULL;
590                 }
591
592                 /*
593                  * We have to enqueue the top waiter(if it exists) into
594                  * task->pi_waiters list.
595                  */
596                 if (rt_mutex_has_waiters(lock)) {
597                         top = rt_mutex_top_waiter(lock);
598                         rt_mutex_enqueue_pi(task, top);
599                 }
600                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
601         }
602
603         /* We got the lock. */
604         debug_rt_mutex_lock(lock);
605
606         rt_mutex_set_owner(lock, task);
607
608         rt_mutex_deadlock_account_lock(lock, task);
609
610         return 1;
611 }
612
613 /*
614  * Task blocks on lock.
615  *
616  * Prepare waiter and propagate pi chain
617  *
618  * This must be called with lock->wait_lock held.
619  */
620 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
621                                    struct rt_mutex_waiter *waiter,
622                                    struct task_struct *task,
623                                    int detect_deadlock)
624 {
625         struct task_struct *owner = rt_mutex_owner(lock);
626         struct rt_mutex_waiter *top_waiter = waiter;
627         struct rt_mutex *next_lock;
628         int chain_walk = 0, res;
629         unsigned long flags;
630
631         /*
632          * Early deadlock detection. We really don't want the task to
633          * enqueue on itself just to untangle the mess later. It's not
634          * only an optimization. We drop the locks, so another waiter
635          * can come in before the chain walk detects the deadlock. So
636          * the other will detect the deadlock and return -EDEADLOCK,
637          * which is wrong, as the other waiter is not in a deadlock
638          * situation.
639          */
640         if (owner == task)
641                 return -EDEADLK;
642
643         raw_spin_lock_irqsave(&task->pi_lock, flags);
644         __rt_mutex_adjust_prio(task);
645         waiter->task = task;
646         waiter->lock = lock;
647         waiter->prio = task->prio;
648
649         /* Get the top priority waiter on the lock */
650         if (rt_mutex_has_waiters(lock))
651                 top_waiter = rt_mutex_top_waiter(lock);
652         rt_mutex_enqueue(lock, waiter);
653
654         task->pi_blocked_on = waiter;
655
656         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
657
658         if (!owner)
659                 return 0;
660
661         raw_spin_lock_irqsave(&owner->pi_lock, flags);
662         if (waiter == rt_mutex_top_waiter(lock)) {
663                 rt_mutex_dequeue_pi(owner, top_waiter);
664                 rt_mutex_enqueue_pi(owner, waiter);
665
666                 __rt_mutex_adjust_prio(owner);
667                 if (owner->pi_blocked_on)
668                         chain_walk = 1;
669         } else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
670                 chain_walk = 1;
671         }
672
673         /* Store the lock on which owner is blocked or NULL */
674         next_lock = task_blocked_on_lock(owner);
675
676         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
677         /*
678          * Even if full deadlock detection is on, if the owner is not
679          * blocked itself, we can avoid finding this out in the chain
680          * walk.
681          */
682         if (!chain_walk || !next_lock)
683                 return 0;
684
685         /*
686          * The owner can't disappear while holding a lock,
687          * so the owner struct is protected by wait_lock.
688          * Gets dropped in rt_mutex_adjust_prio_chain()!
689          */
690         get_task_struct(owner);
691
692         raw_spin_unlock(&lock->wait_lock);
693
694         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock,
695                                          next_lock, waiter, task);
696
697         raw_spin_lock(&lock->wait_lock);
698
699         return res;
700 }
701
702 /*
703  * Wake up the next waiter on the lock.
704  *
705  * Remove the top waiter from the current tasks pi waiter list and
706  * wake it up.
707  *
708  * Called with lock->wait_lock held.
709  */
710 static void wakeup_next_waiter(struct rt_mutex *lock)
711 {
712         struct rt_mutex_waiter *waiter;
713         unsigned long flags;
714
715         raw_spin_lock_irqsave(&current->pi_lock, flags);
716
717         waiter = rt_mutex_top_waiter(lock);
718
719         /*
720          * Remove it from current->pi_waiters. We do not adjust a
721          * possible priority boost right now. We execute wakeup in the
722          * boosted mode and go back to normal after releasing
723          * lock->wait_lock.
724          */
725         rt_mutex_dequeue_pi(current, waiter);
726
727         /*
728          * As we are waking up the top waiter, and the waiter stays
729          * queued on the lock until it gets the lock, this lock
730          * obviously has waiters. Just set the bit here and this has
731          * the added benefit of forcing all new tasks into the
732          * slow path making sure no task of lower priority than
733          * the top waiter can steal this lock.
734          */
735         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
736
737         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
738
739         /*
740          * It's safe to dereference waiter as it cannot go away as
741          * long as we hold lock->wait_lock. The waiter task needs to
742          * acquire it in order to dequeue the waiter.
743          */
744         wake_up_process(waiter->task);
745 }
746
747 /*
748  * Remove a waiter from a lock and give up
749  *
750  * Must be called with lock->wait_lock held and
751  * have just failed to try_to_take_rt_mutex().
752  */
753 static void remove_waiter(struct rt_mutex *lock,
754                           struct rt_mutex_waiter *waiter)
755 {
756         int first = (waiter == rt_mutex_top_waiter(lock));
757         struct task_struct *owner = rt_mutex_owner(lock);
758         struct rt_mutex *next_lock = NULL;
759         unsigned long flags;
760
761         raw_spin_lock_irqsave(&current->pi_lock, flags);
762         rt_mutex_dequeue(lock, waiter);
763         current->pi_blocked_on = NULL;
764         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
765
766         if (!owner)
767                 return;
768
769         if (first) {
770
771                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
772
773                 rt_mutex_dequeue_pi(owner, waiter);
774
775                 if (rt_mutex_has_waiters(lock)) {
776                         struct rt_mutex_waiter *next;
777
778                         next = rt_mutex_top_waiter(lock);
779                         rt_mutex_enqueue_pi(owner, next);
780                 }
781                 __rt_mutex_adjust_prio(owner);
782
783                 /* Store the lock on which owner is blocked or NULL */
784                 next_lock = task_blocked_on_lock(owner);
785
786                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
787         }
788
789         if (!next_lock)
790                 return;
791
792         /* gets dropped in rt_mutex_adjust_prio_chain()! */
793         get_task_struct(owner);
794
795         raw_spin_unlock(&lock->wait_lock);
796
797         rt_mutex_adjust_prio_chain(owner, 0, lock, next_lock, NULL, current);
798
799         raw_spin_lock(&lock->wait_lock);
800 }
801
802 /*
803  * Recheck the pi chain, in case we got a priority setting
804  *
805  * Called from sched_setscheduler
806  */
807 void rt_mutex_adjust_pi(struct task_struct *task)
808 {
809         struct rt_mutex_waiter *waiter;
810         struct rt_mutex *next_lock;
811         unsigned long flags;
812
813         raw_spin_lock_irqsave(&task->pi_lock, flags);
814
815         waiter = task->pi_blocked_on;
816         if (!waiter || (waiter->prio == task->prio &&
817                         !dl_prio(task->prio))) {
818                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
819                 return;
820         }
821         next_lock = waiter->lock;
822         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
823
824         /* gets dropped in rt_mutex_adjust_prio_chain()! */
825         get_task_struct(task);
826
827         rt_mutex_adjust_prio_chain(task, 0, NULL, next_lock, NULL, task);
828 }
829
830 /**
831  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
832  * @lock:                the rt_mutex to take
833  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
834  *                       or TASK_UNINTERRUPTIBLE)
835  * @timeout:             the pre-initialized and started timer, or NULL for none
836  * @waiter:              the pre-initialized rt_mutex_waiter
837  *
838  * lock->wait_lock must be held by the caller.
839  */
840 static int __sched
841 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
842                     struct hrtimer_sleeper *timeout,
843                     struct rt_mutex_waiter *waiter)
844 {
845         int ret = 0;
846
847         for (;;) {
848                 /* Try to acquire the lock: */
849                 if (try_to_take_rt_mutex(lock, current, waiter))
850                         break;
851
852                 /*
853                  * TASK_INTERRUPTIBLE checks for signals and
854                  * timeout. Ignored otherwise.
855                  */
856                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
857                         /* Signal pending? */
858                         if (signal_pending(current))
859                                 ret = -EINTR;
860                         if (timeout && !timeout->task)
861                                 ret = -ETIMEDOUT;
862                         if (ret)
863                                 break;
864                 }
865
866                 raw_spin_unlock(&lock->wait_lock);
867
868                 debug_rt_mutex_print_deadlock(waiter);
869
870                 schedule_rt_mutex(lock);
871
872                 raw_spin_lock(&lock->wait_lock);
873                 set_current_state(state);
874         }
875
876         return ret;
877 }
878
879 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
880                                      struct rt_mutex_waiter *w)
881 {
882         /*
883          * If the result is not -EDEADLOCK or the caller requested
884          * deadlock detection, nothing to do here.
885          */
886         if (res != -EDEADLOCK || detect_deadlock)
887                 return;
888
889         /*
890          * Yell lowdly and stop the task right here.
891          */
892         rt_mutex_print_deadlock(w);
893         while (1) {
894                 set_current_state(TASK_INTERRUPTIBLE);
895                 schedule();
896         }
897 }
898
899 /*
900  * Slow path lock function:
901  */
902 static int __sched
903 rt_mutex_slowlock(struct rt_mutex *lock, int state,
904                   struct hrtimer_sleeper *timeout,
905                   int detect_deadlock)
906 {
907         struct rt_mutex_waiter waiter;
908         int ret = 0;
909
910         debug_rt_mutex_init_waiter(&waiter);
911         RB_CLEAR_NODE(&waiter.pi_tree_entry);
912         RB_CLEAR_NODE(&waiter.tree_entry);
913
914         raw_spin_lock(&lock->wait_lock);
915
916         /* Try to acquire the lock again: */
917         if (try_to_take_rt_mutex(lock, current, NULL)) {
918                 raw_spin_unlock(&lock->wait_lock);
919                 return 0;
920         }
921
922         set_current_state(state);
923
924         /* Setup the timer, when timeout != NULL */
925         if (unlikely(timeout)) {
926                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
927                 if (!hrtimer_active(&timeout->timer))
928                         timeout->task = NULL;
929         }
930
931         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
932
933         if (likely(!ret))
934                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
935
936         set_current_state(TASK_RUNNING);
937
938         if (unlikely(ret)) {
939                 remove_waiter(lock, &waiter);
940                 rt_mutex_handle_deadlock(ret, detect_deadlock, &waiter);
941         }
942
943         /*
944          * try_to_take_rt_mutex() sets the waiter bit
945          * unconditionally. We might have to fix that up.
946          */
947         fixup_rt_mutex_waiters(lock);
948
949         raw_spin_unlock(&lock->wait_lock);
950
951         /* Remove pending timer: */
952         if (unlikely(timeout))
953                 hrtimer_cancel(&timeout->timer);
954
955         debug_rt_mutex_free_waiter(&waiter);
956
957         return ret;
958 }
959
960 /*
961  * Slow path try-lock function:
962  */
963 static inline int
964 rt_mutex_slowtrylock(struct rt_mutex *lock)
965 {
966         int ret = 0;
967
968         raw_spin_lock(&lock->wait_lock);
969
970         if (likely(rt_mutex_owner(lock) != current)) {
971
972                 ret = try_to_take_rt_mutex(lock, current, NULL);
973                 /*
974                  * try_to_take_rt_mutex() sets the lock waiters
975                  * bit unconditionally. Clean this up.
976                  */
977                 fixup_rt_mutex_waiters(lock);
978         }
979
980         raw_spin_unlock(&lock->wait_lock);
981
982         return ret;
983 }
984
985 /*
986  * Slow path to release a rt-mutex:
987  */
988 static void __sched
989 rt_mutex_slowunlock(struct rt_mutex *lock)
990 {
991         raw_spin_lock(&lock->wait_lock);
992
993         debug_rt_mutex_unlock(lock);
994
995         rt_mutex_deadlock_account_unlock(current);
996
997         /*
998          * We must be careful here if the fast path is enabled. If we
999          * have no waiters queued we cannot set owner to NULL here
1000          * because of:
1001          *
1002          * foo->lock->owner = NULL;
1003          *                      rtmutex_lock(foo->lock);   <- fast path
1004          *                      free = atomic_dec_and_test(foo->refcnt);
1005          *                      rtmutex_unlock(foo->lock); <- fast path
1006          *                      if (free)
1007          *                              kfree(foo);
1008          * raw_spin_unlock(foo->lock->wait_lock);
1009          *
1010          * So for the fastpath enabled kernel:
1011          *
1012          * Nothing can set the waiters bit as long as we hold
1013          * lock->wait_lock. So we do the following sequence:
1014          *
1015          *      owner = rt_mutex_owner(lock);
1016          *      clear_rt_mutex_waiters(lock);
1017          *      raw_spin_unlock(&lock->wait_lock);
1018          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1019          *              return;
1020          *      goto retry;
1021          *
1022          * The fastpath disabled variant is simple as all access to
1023          * lock->owner is serialized by lock->wait_lock:
1024          *
1025          *      lock->owner = NULL;
1026          *      raw_spin_unlock(&lock->wait_lock);
1027          */
1028         while (!rt_mutex_has_waiters(lock)) {
1029                 /* Drops lock->wait_lock ! */
1030                 if (unlock_rt_mutex_safe(lock) == true)
1031                         return;
1032                 /* Relock the rtmutex and try again */
1033                 raw_spin_lock(&lock->wait_lock);
1034         }
1035
1036         /*
1037          * The wakeup next waiter path does not suffer from the above
1038          * race. See the comments there.
1039          */
1040         wakeup_next_waiter(lock);
1041
1042         raw_spin_unlock(&lock->wait_lock);
1043
1044         /* Undo pi boosting if necessary: */
1045         rt_mutex_adjust_prio(current);
1046 }
1047
1048 /*
1049  * debug aware fast / slowpath lock,trylock,unlock
1050  *
1051  * The atomic acquire/release ops are compiled away, when either the
1052  * architecture does not support cmpxchg or when debugging is enabled.
1053  */
1054 static inline int
1055 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1056                   int detect_deadlock,
1057                   int (*slowfn)(struct rt_mutex *lock, int state,
1058                                 struct hrtimer_sleeper *timeout,
1059                                 int detect_deadlock))
1060 {
1061         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1062                 rt_mutex_deadlock_account_lock(lock, current);
1063                 return 0;
1064         } else
1065                 return slowfn(lock, state, NULL, detect_deadlock);
1066 }
1067
1068 static inline int
1069 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1070                         struct hrtimer_sleeper *timeout, int detect_deadlock,
1071                         int (*slowfn)(struct rt_mutex *lock, int state,
1072                                       struct hrtimer_sleeper *timeout,
1073                                       int detect_deadlock))
1074 {
1075         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1076                 rt_mutex_deadlock_account_lock(lock, current);
1077                 return 0;
1078         } else
1079                 return slowfn(lock, state, timeout, detect_deadlock);
1080 }
1081
1082 static inline int
1083 rt_mutex_fasttrylock(struct rt_mutex *lock,
1084                      int (*slowfn)(struct rt_mutex *lock))
1085 {
1086         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1087                 rt_mutex_deadlock_account_lock(lock, current);
1088                 return 1;
1089         }
1090         return slowfn(lock);
1091 }
1092
1093 static inline void
1094 rt_mutex_fastunlock(struct rt_mutex *lock,
1095                     void (*slowfn)(struct rt_mutex *lock))
1096 {
1097         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
1098                 rt_mutex_deadlock_account_unlock(current);
1099         else
1100                 slowfn(lock);
1101 }
1102
1103 /**
1104  * rt_mutex_lock - lock a rt_mutex
1105  *
1106  * @lock: the rt_mutex to be locked
1107  */
1108 void __sched rt_mutex_lock(struct rt_mutex *lock)
1109 {
1110         might_sleep();
1111
1112         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
1113 }
1114 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1115
1116 /**
1117  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1118  *
1119  * @lock:               the rt_mutex to be locked
1120  * @detect_deadlock:    deadlock detection on/off
1121  *
1122  * Returns:
1123  *  0           on success
1124  * -EINTR       when interrupted by a signal
1125  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1126  */
1127 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
1128                                                  int detect_deadlock)
1129 {
1130         might_sleep();
1131
1132         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
1133                                  detect_deadlock, rt_mutex_slowlock);
1134 }
1135 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1136
1137 /**
1138  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1139  *                      the timeout structure is provided
1140  *                      by the caller
1141  *
1142  * @lock:               the rt_mutex to be locked
1143  * @timeout:            timeout structure or NULL (no timeout)
1144  * @detect_deadlock:    deadlock detection on/off
1145  *
1146  * Returns:
1147  *  0           on success
1148  * -EINTR       when interrupted by a signal
1149  * -ETIMEDOUT   when the timeout expired
1150  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1151  */
1152 int
1153 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
1154                     int detect_deadlock)
1155 {
1156         might_sleep();
1157
1158         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1159                                        detect_deadlock, rt_mutex_slowlock);
1160 }
1161 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1162
1163 /**
1164  * rt_mutex_trylock - try to lock a rt_mutex
1165  *
1166  * @lock:       the rt_mutex to be locked
1167  *
1168  * Returns 1 on success and 0 on contention
1169  */
1170 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1171 {
1172         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1173 }
1174 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1175
1176 /**
1177  * rt_mutex_unlock - unlock a rt_mutex
1178  *
1179  * @lock: the rt_mutex to be unlocked
1180  */
1181 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1182 {
1183         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1184 }
1185 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1186
1187 /**
1188  * rt_mutex_destroy - mark a mutex unusable
1189  * @lock: the mutex to be destroyed
1190  *
1191  * This function marks the mutex uninitialized, and any subsequent
1192  * use of the mutex is forbidden. The mutex must not be locked when
1193  * this function is called.
1194  */
1195 void rt_mutex_destroy(struct rt_mutex *lock)
1196 {
1197         WARN_ON(rt_mutex_is_locked(lock));
1198 #ifdef CONFIG_DEBUG_RT_MUTEXES
1199         lock->magic = NULL;
1200 #endif
1201 }
1202
1203 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1204
1205 /**
1206  * __rt_mutex_init - initialize the rt lock
1207  *
1208  * @lock: the rt lock to be initialized
1209  *
1210  * Initialize the rt lock to unlocked state.
1211  *
1212  * Initializing of a locked rt lock is not allowed
1213  */
1214 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1215 {
1216         lock->owner = NULL;
1217         raw_spin_lock_init(&lock->wait_lock);
1218         lock->waiters = RB_ROOT;
1219         lock->waiters_leftmost = NULL;
1220
1221         debug_rt_mutex_init(lock, name);
1222 }
1223 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1224
1225 /**
1226  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1227  *                              proxy owner
1228  *
1229  * @lock:       the rt_mutex to be locked
1230  * @proxy_owner:the task to set as owner
1231  *
1232  * No locking. Caller has to do serializing itself
1233  * Special API call for PI-futex support
1234  */
1235 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1236                                 struct task_struct *proxy_owner)
1237 {
1238         __rt_mutex_init(lock, NULL);
1239         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1240         rt_mutex_set_owner(lock, proxy_owner);
1241         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1242 }
1243
1244 /**
1245  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1246  *
1247  * @lock:       the rt_mutex to be locked
1248  *
1249  * No locking. Caller has to do serializing itself
1250  * Special API call for PI-futex support
1251  */
1252 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1253                            struct task_struct *proxy_owner)
1254 {
1255         debug_rt_mutex_proxy_unlock(lock);
1256         rt_mutex_set_owner(lock, NULL);
1257         rt_mutex_deadlock_account_unlock(proxy_owner);
1258 }
1259
1260 /**
1261  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1262  * @lock:               the rt_mutex to take
1263  * @waiter:             the pre-initialized rt_mutex_waiter
1264  * @task:               the task to prepare
1265  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1266  *
1267  * Returns:
1268  *  0 - task blocked on lock
1269  *  1 - acquired the lock for task, caller should wake it up
1270  * <0 - error
1271  *
1272  * Special API call for FUTEX_REQUEUE_PI support.
1273  */
1274 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1275                               struct rt_mutex_waiter *waiter,
1276                               struct task_struct *task, int detect_deadlock)
1277 {
1278         int ret;
1279
1280         raw_spin_lock(&lock->wait_lock);
1281
1282         if (try_to_take_rt_mutex(lock, task, NULL)) {
1283                 raw_spin_unlock(&lock->wait_lock);
1284                 return 1;
1285         }
1286
1287         /* We enforce deadlock detection for futexes */
1288         ret = task_blocks_on_rt_mutex(lock, waiter, task, 1);
1289
1290         if (ret && !rt_mutex_owner(lock)) {
1291                 /*
1292                  * Reset the return value. We might have
1293                  * returned with -EDEADLK and the owner
1294                  * released the lock while we were walking the
1295                  * pi chain.  Let the waiter sort it out.
1296                  */
1297                 ret = 0;
1298         }
1299
1300         if (unlikely(ret))
1301                 remove_waiter(lock, waiter);
1302
1303         raw_spin_unlock(&lock->wait_lock);
1304
1305         debug_rt_mutex_print_deadlock(waiter);
1306
1307         return ret;
1308 }
1309
1310 /**
1311  * rt_mutex_next_owner - return the next owner of the lock
1312  *
1313  * @lock: the rt lock query
1314  *
1315  * Returns the next owner of the lock or NULL
1316  *
1317  * Caller has to serialize against other accessors to the lock
1318  * itself.
1319  *
1320  * Special API call for PI-futex support
1321  */
1322 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1323 {
1324         if (!rt_mutex_has_waiters(lock))
1325                 return NULL;
1326
1327         return rt_mutex_top_waiter(lock)->task;
1328 }
1329
1330 /**
1331  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1332  * @lock:               the rt_mutex we were woken on
1333  * @to:                 the timeout, null if none. hrtimer should already have
1334  *                      been started.
1335  * @waiter:             the pre-initialized rt_mutex_waiter
1336  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1337  *
1338  * Complete the lock acquisition started our behalf by another thread.
1339  *
1340  * Returns:
1341  *  0 - success
1342  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1343  *
1344  * Special API call for PI-futex requeue support
1345  */
1346 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1347                                struct hrtimer_sleeper *to,
1348                                struct rt_mutex_waiter *waiter,
1349                                int detect_deadlock)
1350 {
1351         int ret;
1352
1353         raw_spin_lock(&lock->wait_lock);
1354
1355         set_current_state(TASK_INTERRUPTIBLE);
1356
1357         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1358
1359         set_current_state(TASK_RUNNING);
1360
1361         if (unlikely(ret))
1362                 remove_waiter(lock, waiter);
1363
1364         /*
1365          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1366          * have to fix that up.
1367          */
1368         fixup_rt_mutex_waiters(lock);
1369
1370         raw_spin_unlock(&lock->wait_lock);
1371
1372         return ret;
1373 }