Merge branch 'acpi-ec'
[linux-drm-fsl-dcu.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/locking/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  * Deadlock detection is conditional:
312  *
313  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
314  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
315  *
316  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
317  * conducted independent of the detect argument.
318  *
319  * If the waiter argument is NULL this indicates the deboost path and
320  * deadlock detection is disabled independent of the detect argument
321  * and the config settings.
322  */
323 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
324                                           enum rtmutex_chainwalk chwalk)
325 {
326         /*
327          * This is just a wrapper function for the following call,
328          * because debug_rt_mutex_detect_deadlock() smells like a magic
329          * debug feature and I wanted to keep the cond function in the
330          * main source file along with the comments instead of having
331          * two of the same in the headers.
332          */
333         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
334 }
335
336 /*
337  * Max number of times we'll walk the boosting chain:
338  */
339 int max_lock_depth = 1024;
340
341 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
342 {
343         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
344 }
345
346 /*
347  * Adjust the priority chain. Also used for deadlock detection.
348  * Decreases task's usage by one - may thus free the task.
349  *
350  * @task:       the task owning the mutex (owner) for which a chain walk is
351  *              probably needed
352  * @deadlock_detect: do we have to carry out deadlock detection?
353  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
354  *              things for a task that has just got its priority adjusted, and
355  *              is waiting on a mutex)
356  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
357  *              we dropped its pi_lock. Is never dereferenced, only used for
358  *              comparison to detect lock chain changes.
359  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
360  *              its priority to the mutex owner (can be NULL in the case
361  *              depicted above or if the top waiter is gone away and we are
362  *              actually deboosting the owner)
363  * @top_task:   the current top waiter
364  *
365  * Returns 0 or -EDEADLK.
366  *
367  * Chain walk basics and protection scope
368  *
369  * [R] refcount on task
370  * [P] task->pi_lock held
371  * [L] rtmutex->wait_lock held
372  *
373  * Step Description                             Protected by
374  *      function arguments:
375  *      @task                                   [R]
376  *      @orig_lock if != NULL                   @top_task is blocked on it
377  *      @next_lock                              Unprotected. Cannot be
378  *                                              dereferenced. Only used for
379  *                                              comparison.
380  *      @orig_waiter if != NULL                 @top_task is blocked on it
381  *      @top_task                               current, or in case of proxy
382  *                                              locking protected by calling
383  *                                              code
384  *      again:
385  *        loop_sanity_check();
386  *      retry:
387  * [1]    lock(task->pi_lock);                  [R] acquire [P]
388  * [2]    waiter = task->pi_blocked_on;         [P]
389  * [3]    check_exit_conditions_1();            [P]
390  * [4]    lock = waiter->lock;                  [P]
391  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
392  *          unlock(task->pi_lock);              release [P]
393  *          goto retry;
394  *        }
395  * [6]    check_exit_conditions_2();            [P] + [L]
396  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
397  * [8]    unlock(task->pi_lock);                release [P]
398  *        put_task_struct(task);                release [R]
399  * [9]    check_exit_conditions_3();            [L]
400  * [10]   task = owner(lock);                   [L]
401  *        get_task_struct(task);                [L] acquire [R]
402  *        lock(task->pi_lock);                  [L] acquire [P]
403  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
404  * [12]   check_exit_conditions_4();            [P] + [L]
405  * [13]   unlock(task->pi_lock);                release [P]
406  *        unlock(lock->wait_lock);              release [L]
407  *        goto again;
408  */
409 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
410                                       enum rtmutex_chainwalk chwalk,
411                                       struct rt_mutex *orig_lock,
412                                       struct rt_mutex *next_lock,
413                                       struct rt_mutex_waiter *orig_waiter,
414                                       struct task_struct *top_task)
415 {
416         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
417         struct rt_mutex_waiter *prerequeue_top_waiter;
418         int ret = 0, depth = 0;
419         struct rt_mutex *lock;
420         bool detect_deadlock;
421         unsigned long flags;
422         bool requeue = true;
423
424         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
425
426         /*
427          * The (de)boosting is a step by step approach with a lot of
428          * pitfalls. We want this to be preemptible and we want hold a
429          * maximum of two locks per step. So we have to check
430          * carefully whether things change under us.
431          */
432  again:
433         /*
434          * We limit the lock chain length for each invocation.
435          */
436         if (++depth > max_lock_depth) {
437                 static int prev_max;
438
439                 /*
440                  * Print this only once. If the admin changes the limit,
441                  * print a new message when reaching the limit again.
442                  */
443                 if (prev_max != max_lock_depth) {
444                         prev_max = max_lock_depth;
445                         printk(KERN_WARNING "Maximum lock depth %d reached "
446                                "task: %s (%d)\n", max_lock_depth,
447                                top_task->comm, task_pid_nr(top_task));
448                 }
449                 put_task_struct(task);
450
451                 return -EDEADLK;
452         }
453
454         /*
455          * We are fully preemptible here and only hold the refcount on
456          * @task. So everything can have changed under us since the
457          * caller or our own code below (goto retry/again) dropped all
458          * locks.
459          */
460  retry:
461         /*
462          * [1] Task cannot go away as we did a get_task() before !
463          */
464         raw_spin_lock_irqsave(&task->pi_lock, flags);
465
466         /*
467          * [2] Get the waiter on which @task is blocked on.
468          */
469         waiter = task->pi_blocked_on;
470
471         /*
472          * [3] check_exit_conditions_1() protected by task->pi_lock.
473          */
474
475         /*
476          * Check whether the end of the boosting chain has been
477          * reached or the state of the chain has changed while we
478          * dropped the locks.
479          */
480         if (!waiter)
481                 goto out_unlock_pi;
482
483         /*
484          * Check the orig_waiter state. After we dropped the locks,
485          * the previous owner of the lock might have released the lock.
486          */
487         if (orig_waiter && !rt_mutex_owner(orig_lock))
488                 goto out_unlock_pi;
489
490         /*
491          * We dropped all locks after taking a refcount on @task, so
492          * the task might have moved on in the lock chain or even left
493          * the chain completely and blocks now on an unrelated lock or
494          * on @orig_lock.
495          *
496          * We stored the lock on which @task was blocked in @next_lock,
497          * so we can detect the chain change.
498          */
499         if (next_lock != waiter->lock)
500                 goto out_unlock_pi;
501
502         /*
503          * Drop out, when the task has no waiters. Note,
504          * top_waiter can be NULL, when we are in the deboosting
505          * mode!
506          */
507         if (top_waiter) {
508                 if (!task_has_pi_waiters(task))
509                         goto out_unlock_pi;
510                 /*
511                  * If deadlock detection is off, we stop here if we
512                  * are not the top pi waiter of the task. If deadlock
513                  * detection is enabled we continue, but stop the
514                  * requeueing in the chain walk.
515                  */
516                 if (top_waiter != task_top_pi_waiter(task)) {
517                         if (!detect_deadlock)
518                                 goto out_unlock_pi;
519                         else
520                                 requeue = false;
521                 }
522         }
523
524         /*
525          * If the waiter priority is the same as the task priority
526          * then there is no further priority adjustment necessary.  If
527          * deadlock detection is off, we stop the chain walk. If its
528          * enabled we continue, but stop the requeueing in the chain
529          * walk.
530          */
531         if (waiter->prio == task->prio) {
532                 if (!detect_deadlock)
533                         goto out_unlock_pi;
534                 else
535                         requeue = false;
536         }
537
538         /*
539          * [4] Get the next lock
540          */
541         lock = waiter->lock;
542         /*
543          * [5] We need to trylock here as we are holding task->pi_lock,
544          * which is the reverse lock order versus the other rtmutex
545          * operations.
546          */
547         if (!raw_spin_trylock(&lock->wait_lock)) {
548                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
549                 cpu_relax();
550                 goto retry;
551         }
552
553         /*
554          * [6] check_exit_conditions_2() protected by task->pi_lock and
555          * lock->wait_lock.
556          *
557          * Deadlock detection. If the lock is the same as the original
558          * lock which caused us to walk the lock chain or if the
559          * current lock is owned by the task which initiated the chain
560          * walk, we detected a deadlock.
561          */
562         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
563                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
564                 raw_spin_unlock(&lock->wait_lock);
565                 ret = -EDEADLK;
566                 goto out_unlock_pi;
567         }
568
569         /*
570          * If we just follow the lock chain for deadlock detection, no
571          * need to do all the requeue operations. To avoid a truckload
572          * of conditionals around the various places below, just do the
573          * minimum chain walk checks.
574          */
575         if (!requeue) {
576                 /*
577                  * No requeue[7] here. Just release @task [8]
578                  */
579                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
580                 put_task_struct(task);
581
582                 /*
583                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
584                  * If there is no owner of the lock, end of chain.
585                  */
586                 if (!rt_mutex_owner(lock)) {
587                         raw_spin_unlock(&lock->wait_lock);
588                         return 0;
589                 }
590
591                 /* [10] Grab the next task, i.e. owner of @lock */
592                 task = rt_mutex_owner(lock);
593                 get_task_struct(task);
594                 raw_spin_lock_irqsave(&task->pi_lock, flags);
595
596                 /*
597                  * No requeue [11] here. We just do deadlock detection.
598                  *
599                  * [12] Store whether owner is blocked
600                  * itself. Decision is made after dropping the locks
601                  */
602                 next_lock = task_blocked_on_lock(task);
603                 /*
604                  * Get the top waiter for the next iteration
605                  */
606                 top_waiter = rt_mutex_top_waiter(lock);
607
608                 /* [13] Drop locks */
609                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
610                 raw_spin_unlock(&lock->wait_lock);
611
612                 /* If owner is not blocked, end of chain. */
613                 if (!next_lock)
614                         goto out_put_task;
615                 goto again;
616         }
617
618         /*
619          * Store the current top waiter before doing the requeue
620          * operation on @lock. We need it for the boost/deboost
621          * decision below.
622          */
623         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
624
625         /* [7] Requeue the waiter in the lock waiter list. */
626         rt_mutex_dequeue(lock, waiter);
627         waiter->prio = task->prio;
628         rt_mutex_enqueue(lock, waiter);
629
630         /* [8] Release the task */
631         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
632         put_task_struct(task);
633
634         /*
635          * [9] check_exit_conditions_3 protected by lock->wait_lock.
636          *
637          * We must abort the chain walk if there is no lock owner even
638          * in the dead lock detection case, as we have nothing to
639          * follow here. This is the end of the chain we are walking.
640          */
641         if (!rt_mutex_owner(lock)) {
642                 /*
643                  * If the requeue [7] above changed the top waiter,
644                  * then we need to wake the new top waiter up to try
645                  * to get the lock.
646                  */
647                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
648                         wake_up_process(rt_mutex_top_waiter(lock)->task);
649                 raw_spin_unlock(&lock->wait_lock);
650                 return 0;
651         }
652
653         /* [10] Grab the next task, i.e. the owner of @lock */
654         task = rt_mutex_owner(lock);
655         get_task_struct(task);
656         raw_spin_lock_irqsave(&task->pi_lock, flags);
657
658         /* [11] requeue the pi waiters if necessary */
659         if (waiter == rt_mutex_top_waiter(lock)) {
660                 /*
661                  * The waiter became the new top (highest priority)
662                  * waiter on the lock. Replace the previous top waiter
663                  * in the owner tasks pi waiters list with this waiter
664                  * and adjust the priority of the owner.
665                  */
666                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
667                 rt_mutex_enqueue_pi(task, waiter);
668                 __rt_mutex_adjust_prio(task);
669
670         } else if (prerequeue_top_waiter == waiter) {
671                 /*
672                  * The waiter was the top waiter on the lock, but is
673                  * no longer the top prority waiter. Replace waiter in
674                  * the owner tasks pi waiters list with the new top
675                  * (highest priority) waiter and adjust the priority
676                  * of the owner.
677                  * The new top waiter is stored in @waiter so that
678                  * @waiter == @top_waiter evaluates to true below and
679                  * we continue to deboost the rest of the chain.
680                  */
681                 rt_mutex_dequeue_pi(task, waiter);
682                 waiter = rt_mutex_top_waiter(lock);
683                 rt_mutex_enqueue_pi(task, waiter);
684                 __rt_mutex_adjust_prio(task);
685         } else {
686                 /*
687                  * Nothing changed. No need to do any priority
688                  * adjustment.
689                  */
690         }
691
692         /*
693          * [12] check_exit_conditions_4() protected by task->pi_lock
694          * and lock->wait_lock. The actual decisions are made after we
695          * dropped the locks.
696          *
697          * Check whether the task which owns the current lock is pi
698          * blocked itself. If yes we store a pointer to the lock for
699          * the lock chain change detection above. After we dropped
700          * task->pi_lock next_lock cannot be dereferenced anymore.
701          */
702         next_lock = task_blocked_on_lock(task);
703         /*
704          * Store the top waiter of @lock for the end of chain walk
705          * decision below.
706          */
707         top_waiter = rt_mutex_top_waiter(lock);
708
709         /* [13] Drop the locks */
710         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
711         raw_spin_unlock(&lock->wait_lock);
712
713         /*
714          * Make the actual exit decisions [12], based on the stored
715          * values.
716          *
717          * We reached the end of the lock chain. Stop right here. No
718          * point to go back just to figure that out.
719          */
720         if (!next_lock)
721                 goto out_put_task;
722
723         /*
724          * If the current waiter is not the top waiter on the lock,
725          * then we can stop the chain walk here if we are not in full
726          * deadlock detection mode.
727          */
728         if (!detect_deadlock && waiter != top_waiter)
729                 goto out_put_task;
730
731         goto again;
732
733  out_unlock_pi:
734         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
735  out_put_task:
736         put_task_struct(task);
737
738         return ret;
739 }
740
741 /*
742  * Try to take an rt-mutex
743  *
744  * Must be called with lock->wait_lock held.
745  *
746  * @lock:   The lock to be acquired.
747  * @task:   The task which wants to acquire the lock
748  * @waiter: The waiter that is queued to the lock's wait list if the
749  *          callsite called task_blocked_on_lock(), otherwise NULL
750  */
751 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
752                                 struct rt_mutex_waiter *waiter)
753 {
754         unsigned long flags;
755
756         /*
757          * Before testing whether we can acquire @lock, we set the
758          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
759          * other tasks which try to modify @lock into the slow path
760          * and they serialize on @lock->wait_lock.
761          *
762          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
763          * as explained at the top of this file if and only if:
764          *
765          * - There is a lock owner. The caller must fixup the
766          *   transient state if it does a trylock or leaves the lock
767          *   function due to a signal or timeout.
768          *
769          * - @task acquires the lock and there are no other
770          *   waiters. This is undone in rt_mutex_set_owner(@task) at
771          *   the end of this function.
772          */
773         mark_rt_mutex_waiters(lock);
774
775         /*
776          * If @lock has an owner, give up.
777          */
778         if (rt_mutex_owner(lock))
779                 return 0;
780
781         /*
782          * If @waiter != NULL, @task has already enqueued the waiter
783          * into @lock waiter list. If @waiter == NULL then this is a
784          * trylock attempt.
785          */
786         if (waiter) {
787                 /*
788                  * If waiter is not the highest priority waiter of
789                  * @lock, give up.
790                  */
791                 if (waiter != rt_mutex_top_waiter(lock))
792                         return 0;
793
794                 /*
795                  * We can acquire the lock. Remove the waiter from the
796                  * lock waiters list.
797                  */
798                 rt_mutex_dequeue(lock, waiter);
799
800         } else {
801                 /*
802                  * If the lock has waiters already we check whether @task is
803                  * eligible to take over the lock.
804                  *
805                  * If there are no other waiters, @task can acquire
806                  * the lock.  @task->pi_blocked_on is NULL, so it does
807                  * not need to be dequeued.
808                  */
809                 if (rt_mutex_has_waiters(lock)) {
810                         /*
811                          * If @task->prio is greater than or equal to
812                          * the top waiter priority (kernel view),
813                          * @task lost.
814                          */
815                         if (task->prio >= rt_mutex_top_waiter(lock)->prio)
816                                 return 0;
817
818                         /*
819                          * The current top waiter stays enqueued. We
820                          * don't have to change anything in the lock
821                          * waiters order.
822                          */
823                 } else {
824                         /*
825                          * No waiters. Take the lock without the
826                          * pi_lock dance.@task->pi_blocked_on is NULL
827                          * and we have no waiters to enqueue in @task
828                          * pi waiters list.
829                          */
830                         goto takeit;
831                 }
832         }
833
834         /*
835          * Clear @task->pi_blocked_on. Requires protection by
836          * @task->pi_lock. Redundant operation for the @waiter == NULL
837          * case, but conditionals are more expensive than a redundant
838          * store.
839          */
840         raw_spin_lock_irqsave(&task->pi_lock, flags);
841         task->pi_blocked_on = NULL;
842         /*
843          * Finish the lock acquisition. @task is the new owner. If
844          * other waiters exist we have to insert the highest priority
845          * waiter into @task->pi_waiters list.
846          */
847         if (rt_mutex_has_waiters(lock))
848                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
849         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
850
851 takeit:
852         /* We got the lock. */
853         debug_rt_mutex_lock(lock);
854
855         /*
856          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
857          * are still waiters or clears it.
858          */
859         rt_mutex_set_owner(lock, task);
860
861         rt_mutex_deadlock_account_lock(lock, task);
862
863         return 1;
864 }
865
866 /*
867  * Task blocks on lock.
868  *
869  * Prepare waiter and propagate pi chain
870  *
871  * This must be called with lock->wait_lock held.
872  */
873 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
874                                    struct rt_mutex_waiter *waiter,
875                                    struct task_struct *task,
876                                    enum rtmutex_chainwalk chwalk)
877 {
878         struct task_struct *owner = rt_mutex_owner(lock);
879         struct rt_mutex_waiter *top_waiter = waiter;
880         struct rt_mutex *next_lock;
881         int chain_walk = 0, res;
882         unsigned long flags;
883
884         /*
885          * Early deadlock detection. We really don't want the task to
886          * enqueue on itself just to untangle the mess later. It's not
887          * only an optimization. We drop the locks, so another waiter
888          * can come in before the chain walk detects the deadlock. So
889          * the other will detect the deadlock and return -EDEADLOCK,
890          * which is wrong, as the other waiter is not in a deadlock
891          * situation.
892          */
893         if (owner == task)
894                 return -EDEADLK;
895
896         raw_spin_lock_irqsave(&task->pi_lock, flags);
897         __rt_mutex_adjust_prio(task);
898         waiter->task = task;
899         waiter->lock = lock;
900         waiter->prio = task->prio;
901
902         /* Get the top priority waiter on the lock */
903         if (rt_mutex_has_waiters(lock))
904                 top_waiter = rt_mutex_top_waiter(lock);
905         rt_mutex_enqueue(lock, waiter);
906
907         task->pi_blocked_on = waiter;
908
909         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
910
911         if (!owner)
912                 return 0;
913
914         raw_spin_lock_irqsave(&owner->pi_lock, flags);
915         if (waiter == rt_mutex_top_waiter(lock)) {
916                 rt_mutex_dequeue_pi(owner, top_waiter);
917                 rt_mutex_enqueue_pi(owner, waiter);
918
919                 __rt_mutex_adjust_prio(owner);
920                 if (owner->pi_blocked_on)
921                         chain_walk = 1;
922         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
923                 chain_walk = 1;
924         }
925
926         /* Store the lock on which owner is blocked or NULL */
927         next_lock = task_blocked_on_lock(owner);
928
929         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
930         /*
931          * Even if full deadlock detection is on, if the owner is not
932          * blocked itself, we can avoid finding this out in the chain
933          * walk.
934          */
935         if (!chain_walk || !next_lock)
936                 return 0;
937
938         /*
939          * The owner can't disappear while holding a lock,
940          * so the owner struct is protected by wait_lock.
941          * Gets dropped in rt_mutex_adjust_prio_chain()!
942          */
943         get_task_struct(owner);
944
945         raw_spin_unlock(&lock->wait_lock);
946
947         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
948                                          next_lock, waiter, task);
949
950         raw_spin_lock(&lock->wait_lock);
951
952         return res;
953 }
954
955 /*
956  * Wake up the next waiter on the lock.
957  *
958  * Remove the top waiter from the current tasks pi waiter list and
959  * wake it up.
960  *
961  * Called with lock->wait_lock held.
962  */
963 static void wakeup_next_waiter(struct rt_mutex *lock)
964 {
965         struct rt_mutex_waiter *waiter;
966         unsigned long flags;
967
968         raw_spin_lock_irqsave(&current->pi_lock, flags);
969
970         waiter = rt_mutex_top_waiter(lock);
971
972         /*
973          * Remove it from current->pi_waiters. We do not adjust a
974          * possible priority boost right now. We execute wakeup in the
975          * boosted mode and go back to normal after releasing
976          * lock->wait_lock.
977          */
978         rt_mutex_dequeue_pi(current, waiter);
979
980         /*
981          * As we are waking up the top waiter, and the waiter stays
982          * queued on the lock until it gets the lock, this lock
983          * obviously has waiters. Just set the bit here and this has
984          * the added benefit of forcing all new tasks into the
985          * slow path making sure no task of lower priority than
986          * the top waiter can steal this lock.
987          */
988         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
989
990         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
991
992         /*
993          * It's safe to dereference waiter as it cannot go away as
994          * long as we hold lock->wait_lock. The waiter task needs to
995          * acquire it in order to dequeue the waiter.
996          */
997         wake_up_process(waiter->task);
998 }
999
1000 /*
1001  * Remove a waiter from a lock and give up
1002  *
1003  * Must be called with lock->wait_lock held and
1004  * have just failed to try_to_take_rt_mutex().
1005  */
1006 static void remove_waiter(struct rt_mutex *lock,
1007                           struct rt_mutex_waiter *waiter)
1008 {
1009         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1010         struct task_struct *owner = rt_mutex_owner(lock);
1011         struct rt_mutex *next_lock;
1012         unsigned long flags;
1013
1014         raw_spin_lock_irqsave(&current->pi_lock, flags);
1015         rt_mutex_dequeue(lock, waiter);
1016         current->pi_blocked_on = NULL;
1017         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1018
1019         /*
1020          * Only update priority if the waiter was the highest priority
1021          * waiter of the lock and there is an owner to update.
1022          */
1023         if (!owner || !is_top_waiter)
1024                 return;
1025
1026         raw_spin_lock_irqsave(&owner->pi_lock, flags);
1027
1028         rt_mutex_dequeue_pi(owner, waiter);
1029
1030         if (rt_mutex_has_waiters(lock))
1031                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1032
1033         __rt_mutex_adjust_prio(owner);
1034
1035         /* Store the lock on which owner is blocked or NULL */
1036         next_lock = task_blocked_on_lock(owner);
1037
1038         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
1039
1040         /*
1041          * Don't walk the chain, if the owner task is not blocked
1042          * itself.
1043          */
1044         if (!next_lock)
1045                 return;
1046
1047         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1048         get_task_struct(owner);
1049
1050         raw_spin_unlock(&lock->wait_lock);
1051
1052         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1053                                    next_lock, NULL, current);
1054
1055         raw_spin_lock(&lock->wait_lock);
1056 }
1057
1058 /*
1059  * Recheck the pi chain, in case we got a priority setting
1060  *
1061  * Called from sched_setscheduler
1062  */
1063 void rt_mutex_adjust_pi(struct task_struct *task)
1064 {
1065         struct rt_mutex_waiter *waiter;
1066         struct rt_mutex *next_lock;
1067         unsigned long flags;
1068
1069         raw_spin_lock_irqsave(&task->pi_lock, flags);
1070
1071         waiter = task->pi_blocked_on;
1072         if (!waiter || (waiter->prio == task->prio &&
1073                         !dl_prio(task->prio))) {
1074                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1075                 return;
1076         }
1077         next_lock = waiter->lock;
1078         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1079
1080         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1081         get_task_struct(task);
1082
1083         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1084                                    next_lock, NULL, task);
1085 }
1086
1087 /**
1088  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1089  * @lock:                the rt_mutex to take
1090  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1091  *                       or TASK_UNINTERRUPTIBLE)
1092  * @timeout:             the pre-initialized and started timer, or NULL for none
1093  * @waiter:              the pre-initialized rt_mutex_waiter
1094  *
1095  * lock->wait_lock must be held by the caller.
1096  */
1097 static int __sched
1098 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1099                     struct hrtimer_sleeper *timeout,
1100                     struct rt_mutex_waiter *waiter)
1101 {
1102         int ret = 0;
1103
1104         for (;;) {
1105                 /* Try to acquire the lock: */
1106                 if (try_to_take_rt_mutex(lock, current, waiter))
1107                         break;
1108
1109                 /*
1110                  * TASK_INTERRUPTIBLE checks for signals and
1111                  * timeout. Ignored otherwise.
1112                  */
1113                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
1114                         /* Signal pending? */
1115                         if (signal_pending(current))
1116                                 ret = -EINTR;
1117                         if (timeout && !timeout->task)
1118                                 ret = -ETIMEDOUT;
1119                         if (ret)
1120                                 break;
1121                 }
1122
1123                 raw_spin_unlock(&lock->wait_lock);
1124
1125                 debug_rt_mutex_print_deadlock(waiter);
1126
1127                 schedule_rt_mutex(lock);
1128
1129                 raw_spin_lock(&lock->wait_lock);
1130                 set_current_state(state);
1131         }
1132
1133         __set_current_state(TASK_RUNNING);
1134         return ret;
1135 }
1136
1137 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1138                                      struct rt_mutex_waiter *w)
1139 {
1140         /*
1141          * If the result is not -EDEADLOCK or the caller requested
1142          * deadlock detection, nothing to do here.
1143          */
1144         if (res != -EDEADLOCK || detect_deadlock)
1145                 return;
1146
1147         /*
1148          * Yell lowdly and stop the task right here.
1149          */
1150         rt_mutex_print_deadlock(w);
1151         while (1) {
1152                 set_current_state(TASK_INTERRUPTIBLE);
1153                 schedule();
1154         }
1155 }
1156
1157 /*
1158  * Slow path lock function:
1159  */
1160 static int __sched
1161 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1162                   struct hrtimer_sleeper *timeout,
1163                   enum rtmutex_chainwalk chwalk)
1164 {
1165         struct rt_mutex_waiter waiter;
1166         int ret = 0;
1167
1168         debug_rt_mutex_init_waiter(&waiter);
1169         RB_CLEAR_NODE(&waiter.pi_tree_entry);
1170         RB_CLEAR_NODE(&waiter.tree_entry);
1171
1172         raw_spin_lock(&lock->wait_lock);
1173
1174         /* Try to acquire the lock again: */
1175         if (try_to_take_rt_mutex(lock, current, NULL)) {
1176                 raw_spin_unlock(&lock->wait_lock);
1177                 return 0;
1178         }
1179
1180         set_current_state(state);
1181
1182         /* Setup the timer, when timeout != NULL */
1183         if (unlikely(timeout)) {
1184                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1185                 if (!hrtimer_active(&timeout->timer))
1186                         timeout->task = NULL;
1187         }
1188
1189         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1190
1191         if (likely(!ret))
1192                 /* sleep on the mutex */
1193                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1194
1195         if (unlikely(ret)) {
1196                 remove_waiter(lock, &waiter);
1197                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1198         }
1199
1200         /*
1201          * try_to_take_rt_mutex() sets the waiter bit
1202          * unconditionally. We might have to fix that up.
1203          */
1204         fixup_rt_mutex_waiters(lock);
1205
1206         raw_spin_unlock(&lock->wait_lock);
1207
1208         /* Remove pending timer: */
1209         if (unlikely(timeout))
1210                 hrtimer_cancel(&timeout->timer);
1211
1212         debug_rt_mutex_free_waiter(&waiter);
1213
1214         return ret;
1215 }
1216
1217 /*
1218  * Slow path try-lock function:
1219  */
1220 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1221 {
1222         int ret;
1223
1224         /*
1225          * If the lock already has an owner we fail to get the lock.
1226          * This can be done without taking the @lock->wait_lock as
1227          * it is only being read, and this is a trylock anyway.
1228          */
1229         if (rt_mutex_owner(lock))
1230                 return 0;
1231
1232         /*
1233          * The mutex has currently no owner. Lock the wait lock and
1234          * try to acquire the lock.
1235          */
1236         raw_spin_lock(&lock->wait_lock);
1237
1238         ret = try_to_take_rt_mutex(lock, current, NULL);
1239
1240         /*
1241          * try_to_take_rt_mutex() sets the lock waiters bit
1242          * unconditionally. Clean this up.
1243          */
1244         fixup_rt_mutex_waiters(lock);
1245
1246         raw_spin_unlock(&lock->wait_lock);
1247
1248         return ret;
1249 }
1250
1251 /*
1252  * Slow path to release a rt-mutex:
1253  */
1254 static void __sched
1255 rt_mutex_slowunlock(struct rt_mutex *lock)
1256 {
1257         raw_spin_lock(&lock->wait_lock);
1258
1259         debug_rt_mutex_unlock(lock);
1260
1261         rt_mutex_deadlock_account_unlock(current);
1262
1263         /*
1264          * We must be careful here if the fast path is enabled. If we
1265          * have no waiters queued we cannot set owner to NULL here
1266          * because of:
1267          *
1268          * foo->lock->owner = NULL;
1269          *                      rtmutex_lock(foo->lock);   <- fast path
1270          *                      free = atomic_dec_and_test(foo->refcnt);
1271          *                      rtmutex_unlock(foo->lock); <- fast path
1272          *                      if (free)
1273          *                              kfree(foo);
1274          * raw_spin_unlock(foo->lock->wait_lock);
1275          *
1276          * So for the fastpath enabled kernel:
1277          *
1278          * Nothing can set the waiters bit as long as we hold
1279          * lock->wait_lock. So we do the following sequence:
1280          *
1281          *      owner = rt_mutex_owner(lock);
1282          *      clear_rt_mutex_waiters(lock);
1283          *      raw_spin_unlock(&lock->wait_lock);
1284          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1285          *              return;
1286          *      goto retry;
1287          *
1288          * The fastpath disabled variant is simple as all access to
1289          * lock->owner is serialized by lock->wait_lock:
1290          *
1291          *      lock->owner = NULL;
1292          *      raw_spin_unlock(&lock->wait_lock);
1293          */
1294         while (!rt_mutex_has_waiters(lock)) {
1295                 /* Drops lock->wait_lock ! */
1296                 if (unlock_rt_mutex_safe(lock) == true)
1297                         return;
1298                 /* Relock the rtmutex and try again */
1299                 raw_spin_lock(&lock->wait_lock);
1300         }
1301
1302         /*
1303          * The wakeup next waiter path does not suffer from the above
1304          * race. See the comments there.
1305          */
1306         wakeup_next_waiter(lock);
1307
1308         raw_spin_unlock(&lock->wait_lock);
1309
1310         /* Undo pi boosting if necessary: */
1311         rt_mutex_adjust_prio(current);
1312 }
1313
1314 /*
1315  * debug aware fast / slowpath lock,trylock,unlock
1316  *
1317  * The atomic acquire/release ops are compiled away, when either the
1318  * architecture does not support cmpxchg or when debugging is enabled.
1319  */
1320 static inline int
1321 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1322                   int (*slowfn)(struct rt_mutex *lock, int state,
1323                                 struct hrtimer_sleeper *timeout,
1324                                 enum rtmutex_chainwalk chwalk))
1325 {
1326         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1327                 rt_mutex_deadlock_account_lock(lock, current);
1328                 return 0;
1329         } else
1330                 return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1331 }
1332
1333 static inline int
1334 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1335                         struct hrtimer_sleeper *timeout,
1336                         enum rtmutex_chainwalk chwalk,
1337                         int (*slowfn)(struct rt_mutex *lock, int state,
1338                                       struct hrtimer_sleeper *timeout,
1339                                       enum rtmutex_chainwalk chwalk))
1340 {
1341         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1342             likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1343                 rt_mutex_deadlock_account_lock(lock, current);
1344                 return 0;
1345         } else
1346                 return slowfn(lock, state, timeout, chwalk);
1347 }
1348
1349 static inline int
1350 rt_mutex_fasttrylock(struct rt_mutex *lock,
1351                      int (*slowfn)(struct rt_mutex *lock))
1352 {
1353         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1354                 rt_mutex_deadlock_account_lock(lock, current);
1355                 return 1;
1356         }
1357         return slowfn(lock);
1358 }
1359
1360 static inline void
1361 rt_mutex_fastunlock(struct rt_mutex *lock,
1362                     void (*slowfn)(struct rt_mutex *lock))
1363 {
1364         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
1365                 rt_mutex_deadlock_account_unlock(current);
1366         else
1367                 slowfn(lock);
1368 }
1369
1370 /**
1371  * rt_mutex_lock - lock a rt_mutex
1372  *
1373  * @lock: the rt_mutex to be locked
1374  */
1375 void __sched rt_mutex_lock(struct rt_mutex *lock)
1376 {
1377         might_sleep();
1378
1379         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1380 }
1381 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1382
1383 /**
1384  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1385  *
1386  * @lock:               the rt_mutex to be locked
1387  *
1388  * Returns:
1389  *  0           on success
1390  * -EINTR       when interrupted by a signal
1391  */
1392 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1393 {
1394         might_sleep();
1395
1396         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1397 }
1398 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1399
1400 /*
1401  * Futex variant with full deadlock detection.
1402  */
1403 int rt_mutex_timed_futex_lock(struct rt_mutex *lock,
1404                               struct hrtimer_sleeper *timeout)
1405 {
1406         might_sleep();
1407
1408         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1409                                        RT_MUTEX_FULL_CHAINWALK,
1410                                        rt_mutex_slowlock);
1411 }
1412
1413 /**
1414  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1415  *                      the timeout structure is provided
1416  *                      by the caller
1417  *
1418  * @lock:               the rt_mutex to be locked
1419  * @timeout:            timeout structure or NULL (no timeout)
1420  *
1421  * Returns:
1422  *  0           on success
1423  * -EINTR       when interrupted by a signal
1424  * -ETIMEDOUT   when the timeout expired
1425  */
1426 int
1427 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1428 {
1429         might_sleep();
1430
1431         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1432                                        RT_MUTEX_MIN_CHAINWALK,
1433                                        rt_mutex_slowlock);
1434 }
1435 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1436
1437 /**
1438  * rt_mutex_trylock - try to lock a rt_mutex
1439  *
1440  * @lock:       the rt_mutex to be locked
1441  *
1442  * Returns 1 on success and 0 on contention
1443  */
1444 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1445 {
1446         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1447 }
1448 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1449
1450 /**
1451  * rt_mutex_unlock - unlock a rt_mutex
1452  *
1453  * @lock: the rt_mutex to be unlocked
1454  */
1455 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1456 {
1457         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1458 }
1459 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1460
1461 /**
1462  * rt_mutex_destroy - mark a mutex unusable
1463  * @lock: the mutex to be destroyed
1464  *
1465  * This function marks the mutex uninitialized, and any subsequent
1466  * use of the mutex is forbidden. The mutex must not be locked when
1467  * this function is called.
1468  */
1469 void rt_mutex_destroy(struct rt_mutex *lock)
1470 {
1471         WARN_ON(rt_mutex_is_locked(lock));
1472 #ifdef CONFIG_DEBUG_RT_MUTEXES
1473         lock->magic = NULL;
1474 #endif
1475 }
1476
1477 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1478
1479 /**
1480  * __rt_mutex_init - initialize the rt lock
1481  *
1482  * @lock: the rt lock to be initialized
1483  *
1484  * Initialize the rt lock to unlocked state.
1485  *
1486  * Initializing of a locked rt lock is not allowed
1487  */
1488 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1489 {
1490         lock->owner = NULL;
1491         raw_spin_lock_init(&lock->wait_lock);
1492         lock->waiters = RB_ROOT;
1493         lock->waiters_leftmost = NULL;
1494
1495         debug_rt_mutex_init(lock, name);
1496 }
1497 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1498
1499 /**
1500  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1501  *                              proxy owner
1502  *
1503  * @lock:       the rt_mutex to be locked
1504  * @proxy_owner:the task to set as owner
1505  *
1506  * No locking. Caller has to do serializing itself
1507  * Special API call for PI-futex support
1508  */
1509 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1510                                 struct task_struct *proxy_owner)
1511 {
1512         __rt_mutex_init(lock, NULL);
1513         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1514         rt_mutex_set_owner(lock, proxy_owner);
1515         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1516 }
1517
1518 /**
1519  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1520  *
1521  * @lock:       the rt_mutex to be locked
1522  *
1523  * No locking. Caller has to do serializing itself
1524  * Special API call for PI-futex support
1525  */
1526 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1527                            struct task_struct *proxy_owner)
1528 {
1529         debug_rt_mutex_proxy_unlock(lock);
1530         rt_mutex_set_owner(lock, NULL);
1531         rt_mutex_deadlock_account_unlock(proxy_owner);
1532 }
1533
1534 /**
1535  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1536  * @lock:               the rt_mutex to take
1537  * @waiter:             the pre-initialized rt_mutex_waiter
1538  * @task:               the task to prepare
1539  *
1540  * Returns:
1541  *  0 - task blocked on lock
1542  *  1 - acquired the lock for task, caller should wake it up
1543  * <0 - error
1544  *
1545  * Special API call for FUTEX_REQUEUE_PI support.
1546  */
1547 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1548                               struct rt_mutex_waiter *waiter,
1549                               struct task_struct *task)
1550 {
1551         int ret;
1552
1553         raw_spin_lock(&lock->wait_lock);
1554
1555         if (try_to_take_rt_mutex(lock, task, NULL)) {
1556                 raw_spin_unlock(&lock->wait_lock);
1557                 return 1;
1558         }
1559
1560         /* We enforce deadlock detection for futexes */
1561         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1562                                       RT_MUTEX_FULL_CHAINWALK);
1563
1564         if (ret && !rt_mutex_owner(lock)) {
1565                 /*
1566                  * Reset the return value. We might have
1567                  * returned with -EDEADLK and the owner
1568                  * released the lock while we were walking the
1569                  * pi chain.  Let the waiter sort it out.
1570                  */
1571                 ret = 0;
1572         }
1573
1574         if (unlikely(ret))
1575                 remove_waiter(lock, waiter);
1576
1577         raw_spin_unlock(&lock->wait_lock);
1578
1579         debug_rt_mutex_print_deadlock(waiter);
1580
1581         return ret;
1582 }
1583
1584 /**
1585  * rt_mutex_next_owner - return the next owner of the lock
1586  *
1587  * @lock: the rt lock query
1588  *
1589  * Returns the next owner of the lock or NULL
1590  *
1591  * Caller has to serialize against other accessors to the lock
1592  * itself.
1593  *
1594  * Special API call for PI-futex support
1595  */
1596 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1597 {
1598         if (!rt_mutex_has_waiters(lock))
1599                 return NULL;
1600
1601         return rt_mutex_top_waiter(lock)->task;
1602 }
1603
1604 /**
1605  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1606  * @lock:               the rt_mutex we were woken on
1607  * @to:                 the timeout, null if none. hrtimer should already have
1608  *                      been started.
1609  * @waiter:             the pre-initialized rt_mutex_waiter
1610  *
1611  * Complete the lock acquisition started our behalf by another thread.
1612  *
1613  * Returns:
1614  *  0 - success
1615  * <0 - error, one of -EINTR, -ETIMEDOUT
1616  *
1617  * Special API call for PI-futex requeue support
1618  */
1619 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1620                                struct hrtimer_sleeper *to,
1621                                struct rt_mutex_waiter *waiter)
1622 {
1623         int ret;
1624
1625         raw_spin_lock(&lock->wait_lock);
1626
1627         set_current_state(TASK_INTERRUPTIBLE);
1628
1629         /* sleep on the mutex */
1630         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1631
1632         if (unlikely(ret))
1633                 remove_waiter(lock, waiter);
1634
1635         /*
1636          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1637          * have to fix that up.
1638          */
1639         fixup_rt_mutex_waiters(lock);
1640
1641         raw_spin_unlock(&lock->wait_lock);
1642
1643         return ret;
1644 }