Merge series "ASoC: meson: tdm fixes" from Jerome Brunet <jbrunet@baylibre.com>:
[linux-2.6-microblaze.git] / kernel / locking / rtmutex.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
4  *
5  * started by Ingo Molnar and Thomas Gleixner.
6  *
7  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
8  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
9  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
10  *  Copyright (C) 2006 Esben Nielsen
11  *
12  *  See Documentation/locking/rt-mutex-design.rst for details.
13  */
14 #include <linux/spinlock.h>
15 #include <linux/export.h>
16 #include <linux/sched/signal.h>
17 #include <linux/sched/rt.h>
18 #include <linux/sched/deadline.h>
19 #include <linux/sched/wake_q.h>
20 #include <linux/sched/debug.h>
21 #include <linux/timer.h>
22
23 #include "rtmutex_common.h"
24
25 /*
26  * lock->owner state tracking:
27  *
28  * lock->owner holds the task_struct pointer of the owner. Bit 0
29  * is used to keep track of the "lock has waiters" state.
30  *
31  * owner        bit0
32  * NULL         0       lock is free (fast acquire possible)
33  * NULL         1       lock is free and has waiters and the top waiter
34  *                              is going to take the lock*
35  * taskpointer  0       lock is held (fast release possible)
36  * taskpointer  1       lock is held and has waiters**
37  *
38  * The fast atomic compare exchange based acquire and release is only
39  * possible when bit 0 of lock->owner is 0.
40  *
41  * (*) It also can be a transitional state when grabbing the lock
42  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
43  * we need to set the bit0 before looking at the lock, and the owner may be
44  * NULL in this small time, hence this can be a transitional state.
45  *
46  * (**) There is a small time when bit 0 is set but there are no
47  * waiters. This can happen when grabbing the lock in the slow path.
48  * To prevent a cmpxchg of the owner releasing the lock, we need to
49  * set this bit before looking at the lock.
50  */
51
52 static void
53 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
54 {
55         unsigned long val = (unsigned long)owner;
56
57         if (rt_mutex_has_waiters(lock))
58                 val |= RT_MUTEX_HAS_WAITERS;
59
60         WRITE_ONCE(lock->owner, (struct task_struct *)val);
61 }
62
63 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
64 {
65         lock->owner = (struct task_struct *)
66                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
67 }
68
69 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
70 {
71         unsigned long owner, *p = (unsigned long *) &lock->owner;
72
73         if (rt_mutex_has_waiters(lock))
74                 return;
75
76         /*
77          * The rbtree has no waiters enqueued, now make sure that the
78          * lock->owner still has the waiters bit set, otherwise the
79          * following can happen:
80          *
81          * CPU 0        CPU 1           CPU2
82          * l->owner=T1
83          *              rt_mutex_lock(l)
84          *              lock(l->lock)
85          *              l->owner = T1 | HAS_WAITERS;
86          *              enqueue(T2)
87          *              boost()
88          *                unlock(l->lock)
89          *              block()
90          *
91          *                              rt_mutex_lock(l)
92          *                              lock(l->lock)
93          *                              l->owner = T1 | HAS_WAITERS;
94          *                              enqueue(T3)
95          *                              boost()
96          *                                unlock(l->lock)
97          *                              block()
98          *              signal(->T2)    signal(->T3)
99          *              lock(l->lock)
100          *              dequeue(T2)
101          *              deboost()
102          *                unlock(l->lock)
103          *                              lock(l->lock)
104          *                              dequeue(T3)
105          *                               ==> wait list is empty
106          *                              deboost()
107          *                               unlock(l->lock)
108          *              lock(l->lock)
109          *              fixup_rt_mutex_waiters()
110          *                if (wait_list_empty(l) {
111          *                  l->owner = owner
112          *                  owner = l->owner & ~HAS_WAITERS;
113          *                    ==> l->owner = T1
114          *                }
115          *                              lock(l->lock)
116          * rt_mutex_unlock(l)           fixup_rt_mutex_waiters()
117          *                                if (wait_list_empty(l) {
118          *                                  owner = l->owner & ~HAS_WAITERS;
119          * cmpxchg(l->owner, T1, NULL)
120          *  ===> Success (l->owner = NULL)
121          *
122          *                                  l->owner = owner
123          *                                    ==> l->owner = T1
124          *                                }
125          *
126          * With the check for the waiter bit in place T3 on CPU2 will not
127          * overwrite. All tasks fiddling with the waiters bit are
128          * serialized by l->lock, so nothing else can modify the waiters
129          * bit. If the bit is set then nothing can change l->owner either
130          * so the simple RMW is safe. The cmpxchg() will simply fail if it
131          * happens in the middle of the RMW because the waiters bit is
132          * still set.
133          */
134         owner = READ_ONCE(*p);
135         if (owner & RT_MUTEX_HAS_WAITERS)
136                 WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
137 }
138
139 /*
140  * We can speed up the acquire/release, if there's no debugging state to be
141  * set up.
142  */
143 #ifndef CONFIG_DEBUG_RT_MUTEXES
144 # define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
145 # define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
146
147 /*
148  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
149  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
150  * relaxed semantics suffice.
151  */
152 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
153 {
154         unsigned long owner, *p = (unsigned long *) &lock->owner;
155
156         do {
157                 owner = *p;
158         } while (cmpxchg_relaxed(p, owner,
159                                  owner | RT_MUTEX_HAS_WAITERS) != owner);
160 }
161
162 /*
163  * Safe fastpath aware unlock:
164  * 1) Clear the waiters bit
165  * 2) Drop lock->wait_lock
166  * 3) Try to unlock the lock with cmpxchg
167  */
168 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
169                                         unsigned long flags)
170         __releases(lock->wait_lock)
171 {
172         struct task_struct *owner = rt_mutex_owner(lock);
173
174         clear_rt_mutex_waiters(lock);
175         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
176         /*
177          * If a new waiter comes in between the unlock and the cmpxchg
178          * we have two situations:
179          *
180          * unlock(wait_lock);
181          *                                      lock(wait_lock);
182          * cmpxchg(p, owner, 0) == owner
183          *                                      mark_rt_mutex_waiters(lock);
184          *                                      acquire(lock);
185          * or:
186          *
187          * unlock(wait_lock);
188          *                                      lock(wait_lock);
189          *                                      mark_rt_mutex_waiters(lock);
190          *
191          * cmpxchg(p, owner, 0) != owner
192          *                                      enqueue_waiter();
193          *                                      unlock(wait_lock);
194          * lock(wait_lock);
195          * wake waiter();
196          * unlock(wait_lock);
197          *                                      lock(wait_lock);
198          *                                      acquire(lock);
199          */
200         return rt_mutex_cmpxchg_release(lock, owner, NULL);
201 }
202
203 #else
204 # define rt_mutex_cmpxchg_acquire(l,c,n)        (0)
205 # define rt_mutex_cmpxchg_release(l,c,n)        (0)
206
207 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
208 {
209         lock->owner = (struct task_struct *)
210                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
211 }
212
213 /*
214  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
215  */
216 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
217                                         unsigned long flags)
218         __releases(lock->wait_lock)
219 {
220         lock->owner = NULL;
221         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
222         return true;
223 }
224 #endif
225
226 /*
227  * Only use with rt_mutex_waiter_{less,equal}()
228  */
229 #define task_to_waiter(p)       \
230         &(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
231
232 static inline int
233 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
234                      struct rt_mutex_waiter *right)
235 {
236         if (left->prio < right->prio)
237                 return 1;
238
239         /*
240          * If both waiters have dl_prio(), we check the deadlines of the
241          * associated tasks.
242          * If left waiter has a dl_prio(), and we didn't return 1 above,
243          * then right waiter has a dl_prio() too.
244          */
245         if (dl_prio(left->prio))
246                 return dl_time_before(left->deadline, right->deadline);
247
248         return 0;
249 }
250
251 static inline int
252 rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
253                       struct rt_mutex_waiter *right)
254 {
255         if (left->prio != right->prio)
256                 return 0;
257
258         /*
259          * If both waiters have dl_prio(), we check the deadlines of the
260          * associated tasks.
261          * If left waiter has a dl_prio(), and we didn't return 0 above,
262          * then right waiter has a dl_prio() too.
263          */
264         if (dl_prio(left->prio))
265                 return left->deadline == right->deadline;
266
267         return 1;
268 }
269
270 static void
271 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
272 {
273         struct rb_node **link = &lock->waiters.rb_root.rb_node;
274         struct rb_node *parent = NULL;
275         struct rt_mutex_waiter *entry;
276         bool leftmost = true;
277
278         while (*link) {
279                 parent = *link;
280                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
281                 if (rt_mutex_waiter_less(waiter, entry)) {
282                         link = &parent->rb_left;
283                 } else {
284                         link = &parent->rb_right;
285                         leftmost = false;
286                 }
287         }
288
289         rb_link_node(&waiter->tree_entry, parent, link);
290         rb_insert_color_cached(&waiter->tree_entry, &lock->waiters, leftmost);
291 }
292
293 static void
294 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
295 {
296         if (RB_EMPTY_NODE(&waiter->tree_entry))
297                 return;
298
299         rb_erase_cached(&waiter->tree_entry, &lock->waiters);
300         RB_CLEAR_NODE(&waiter->tree_entry);
301 }
302
303 static void
304 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
305 {
306         struct rb_node **link = &task->pi_waiters.rb_root.rb_node;
307         struct rb_node *parent = NULL;
308         struct rt_mutex_waiter *entry;
309         bool leftmost = true;
310
311         while (*link) {
312                 parent = *link;
313                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
314                 if (rt_mutex_waiter_less(waiter, entry)) {
315                         link = &parent->rb_left;
316                 } else {
317                         link = &parent->rb_right;
318                         leftmost = false;
319                 }
320         }
321
322         rb_link_node(&waiter->pi_tree_entry, parent, link);
323         rb_insert_color_cached(&waiter->pi_tree_entry, &task->pi_waiters, leftmost);
324 }
325
326 static void
327 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
328 {
329         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
330                 return;
331
332         rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
333         RB_CLEAR_NODE(&waiter->pi_tree_entry);
334 }
335
336 static void rt_mutex_adjust_prio(struct task_struct *p)
337 {
338         struct task_struct *pi_task = NULL;
339
340         lockdep_assert_held(&p->pi_lock);
341
342         if (task_has_pi_waiters(p))
343                 pi_task = task_top_pi_waiter(p)->task;
344
345         rt_mutex_setprio(p, pi_task);
346 }
347
348 /*
349  * Deadlock detection is conditional:
350  *
351  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
352  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
353  *
354  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
355  * conducted independent of the detect argument.
356  *
357  * If the waiter argument is NULL this indicates the deboost path and
358  * deadlock detection is disabled independent of the detect argument
359  * and the config settings.
360  */
361 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
362                                           enum rtmutex_chainwalk chwalk)
363 {
364         /*
365          * This is just a wrapper function for the following call,
366          * because debug_rt_mutex_detect_deadlock() smells like a magic
367          * debug feature and I wanted to keep the cond function in the
368          * main source file along with the comments instead of having
369          * two of the same in the headers.
370          */
371         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
372 }
373
374 /*
375  * Max number of times we'll walk the boosting chain:
376  */
377 int max_lock_depth = 1024;
378
379 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
380 {
381         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
382 }
383
384 /*
385  * Adjust the priority chain. Also used for deadlock detection.
386  * Decreases task's usage by one - may thus free the task.
387  *
388  * @task:       the task owning the mutex (owner) for which a chain walk is
389  *              probably needed
390  * @chwalk:     do we have to carry out deadlock detection?
391  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
392  *              things for a task that has just got its priority adjusted, and
393  *              is waiting on a mutex)
394  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
395  *              we dropped its pi_lock. Is never dereferenced, only used for
396  *              comparison to detect lock chain changes.
397  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
398  *              its priority to the mutex owner (can be NULL in the case
399  *              depicted above or if the top waiter is gone away and we are
400  *              actually deboosting the owner)
401  * @top_task:   the current top waiter
402  *
403  * Returns 0 or -EDEADLK.
404  *
405  * Chain walk basics and protection scope
406  *
407  * [R] refcount on task
408  * [P] task->pi_lock held
409  * [L] rtmutex->wait_lock held
410  *
411  * Step Description                             Protected by
412  *      function arguments:
413  *      @task                                   [R]
414  *      @orig_lock if != NULL                   @top_task is blocked on it
415  *      @next_lock                              Unprotected. Cannot be
416  *                                              dereferenced. Only used for
417  *                                              comparison.
418  *      @orig_waiter if != NULL                 @top_task is blocked on it
419  *      @top_task                               current, or in case of proxy
420  *                                              locking protected by calling
421  *                                              code
422  *      again:
423  *        loop_sanity_check();
424  *      retry:
425  * [1]    lock(task->pi_lock);                  [R] acquire [P]
426  * [2]    waiter = task->pi_blocked_on;         [P]
427  * [3]    check_exit_conditions_1();            [P]
428  * [4]    lock = waiter->lock;                  [P]
429  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
430  *          unlock(task->pi_lock);              release [P]
431  *          goto retry;
432  *        }
433  * [6]    check_exit_conditions_2();            [P] + [L]
434  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
435  * [8]    unlock(task->pi_lock);                release [P]
436  *        put_task_struct(task);                release [R]
437  * [9]    check_exit_conditions_3();            [L]
438  * [10]   task = owner(lock);                   [L]
439  *        get_task_struct(task);                [L] acquire [R]
440  *        lock(task->pi_lock);                  [L] acquire [P]
441  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
442  * [12]   check_exit_conditions_4();            [P] + [L]
443  * [13]   unlock(task->pi_lock);                release [P]
444  *        unlock(lock->wait_lock);              release [L]
445  *        goto again;
446  */
447 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
448                                       enum rtmutex_chainwalk chwalk,
449                                       struct rt_mutex *orig_lock,
450                                       struct rt_mutex *next_lock,
451                                       struct rt_mutex_waiter *orig_waiter,
452                                       struct task_struct *top_task)
453 {
454         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
455         struct rt_mutex_waiter *prerequeue_top_waiter;
456         int ret = 0, depth = 0;
457         struct rt_mutex *lock;
458         bool detect_deadlock;
459         bool requeue = true;
460
461         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
462
463         /*
464          * The (de)boosting is a step by step approach with a lot of
465          * pitfalls. We want this to be preemptible and we want hold a
466          * maximum of two locks per step. So we have to check
467          * carefully whether things change under us.
468          */
469  again:
470         /*
471          * We limit the lock chain length for each invocation.
472          */
473         if (++depth > max_lock_depth) {
474                 static int prev_max;
475
476                 /*
477                  * Print this only once. If the admin changes the limit,
478                  * print a new message when reaching the limit again.
479                  */
480                 if (prev_max != max_lock_depth) {
481                         prev_max = max_lock_depth;
482                         printk(KERN_WARNING "Maximum lock depth %d reached "
483                                "task: %s (%d)\n", max_lock_depth,
484                                top_task->comm, task_pid_nr(top_task));
485                 }
486                 put_task_struct(task);
487
488                 return -EDEADLK;
489         }
490
491         /*
492          * We are fully preemptible here and only hold the refcount on
493          * @task. So everything can have changed under us since the
494          * caller or our own code below (goto retry/again) dropped all
495          * locks.
496          */
497  retry:
498         /*
499          * [1] Task cannot go away as we did a get_task() before !
500          */
501         raw_spin_lock_irq(&task->pi_lock);
502
503         /*
504          * [2] Get the waiter on which @task is blocked on.
505          */
506         waiter = task->pi_blocked_on;
507
508         /*
509          * [3] check_exit_conditions_1() protected by task->pi_lock.
510          */
511
512         /*
513          * Check whether the end of the boosting chain has been
514          * reached or the state of the chain has changed while we
515          * dropped the locks.
516          */
517         if (!waiter)
518                 goto out_unlock_pi;
519
520         /*
521          * Check the orig_waiter state. After we dropped the locks,
522          * the previous owner of the lock might have released the lock.
523          */
524         if (orig_waiter && !rt_mutex_owner(orig_lock))
525                 goto out_unlock_pi;
526
527         /*
528          * We dropped all locks after taking a refcount on @task, so
529          * the task might have moved on in the lock chain or even left
530          * the chain completely and blocks now on an unrelated lock or
531          * on @orig_lock.
532          *
533          * We stored the lock on which @task was blocked in @next_lock,
534          * so we can detect the chain change.
535          */
536         if (next_lock != waiter->lock)
537                 goto out_unlock_pi;
538
539         /*
540          * Drop out, when the task has no waiters. Note,
541          * top_waiter can be NULL, when we are in the deboosting
542          * mode!
543          */
544         if (top_waiter) {
545                 if (!task_has_pi_waiters(task))
546                         goto out_unlock_pi;
547                 /*
548                  * If deadlock detection is off, we stop here if we
549                  * are not the top pi waiter of the task. If deadlock
550                  * detection is enabled we continue, but stop the
551                  * requeueing in the chain walk.
552                  */
553                 if (top_waiter != task_top_pi_waiter(task)) {
554                         if (!detect_deadlock)
555                                 goto out_unlock_pi;
556                         else
557                                 requeue = false;
558                 }
559         }
560
561         /*
562          * If the waiter priority is the same as the task priority
563          * then there is no further priority adjustment necessary.  If
564          * deadlock detection is off, we stop the chain walk. If its
565          * enabled we continue, but stop the requeueing in the chain
566          * walk.
567          */
568         if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
569                 if (!detect_deadlock)
570                         goto out_unlock_pi;
571                 else
572                         requeue = false;
573         }
574
575         /*
576          * [4] Get the next lock
577          */
578         lock = waiter->lock;
579         /*
580          * [5] We need to trylock here as we are holding task->pi_lock,
581          * which is the reverse lock order versus the other rtmutex
582          * operations.
583          */
584         if (!raw_spin_trylock(&lock->wait_lock)) {
585                 raw_spin_unlock_irq(&task->pi_lock);
586                 cpu_relax();
587                 goto retry;
588         }
589
590         /*
591          * [6] check_exit_conditions_2() protected by task->pi_lock and
592          * lock->wait_lock.
593          *
594          * Deadlock detection. If the lock is the same as the original
595          * lock which caused us to walk the lock chain or if the
596          * current lock is owned by the task which initiated the chain
597          * walk, we detected a deadlock.
598          */
599         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
600                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
601                 raw_spin_unlock(&lock->wait_lock);
602                 ret = -EDEADLK;
603                 goto out_unlock_pi;
604         }
605
606         /*
607          * If we just follow the lock chain for deadlock detection, no
608          * need to do all the requeue operations. To avoid a truckload
609          * of conditionals around the various places below, just do the
610          * minimum chain walk checks.
611          */
612         if (!requeue) {
613                 /*
614                  * No requeue[7] here. Just release @task [8]
615                  */
616                 raw_spin_unlock(&task->pi_lock);
617                 put_task_struct(task);
618
619                 /*
620                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
621                  * If there is no owner of the lock, end of chain.
622                  */
623                 if (!rt_mutex_owner(lock)) {
624                         raw_spin_unlock_irq(&lock->wait_lock);
625                         return 0;
626                 }
627
628                 /* [10] Grab the next task, i.e. owner of @lock */
629                 task = get_task_struct(rt_mutex_owner(lock));
630                 raw_spin_lock(&task->pi_lock);
631
632                 /*
633                  * No requeue [11] here. We just do deadlock detection.
634                  *
635                  * [12] Store whether owner is blocked
636                  * itself. Decision is made after dropping the locks
637                  */
638                 next_lock = task_blocked_on_lock(task);
639                 /*
640                  * Get the top waiter for the next iteration
641                  */
642                 top_waiter = rt_mutex_top_waiter(lock);
643
644                 /* [13] Drop locks */
645                 raw_spin_unlock(&task->pi_lock);
646                 raw_spin_unlock_irq(&lock->wait_lock);
647
648                 /* If owner is not blocked, end of chain. */
649                 if (!next_lock)
650                         goto out_put_task;
651                 goto again;
652         }
653
654         /*
655          * Store the current top waiter before doing the requeue
656          * operation on @lock. We need it for the boost/deboost
657          * decision below.
658          */
659         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
660
661         /* [7] Requeue the waiter in the lock waiter tree. */
662         rt_mutex_dequeue(lock, waiter);
663
664         /*
665          * Update the waiter prio fields now that we're dequeued.
666          *
667          * These values can have changed through either:
668          *
669          *   sys_sched_set_scheduler() / sys_sched_setattr()
670          *
671          * or
672          *
673          *   DL CBS enforcement advancing the effective deadline.
674          *
675          * Even though pi_waiters also uses these fields, and that tree is only
676          * updated in [11], we can do this here, since we hold [L], which
677          * serializes all pi_waiters access and rb_erase() does not care about
678          * the values of the node being removed.
679          */
680         waiter->prio = task->prio;
681         waiter->deadline = task->dl.deadline;
682
683         rt_mutex_enqueue(lock, waiter);
684
685         /* [8] Release the task */
686         raw_spin_unlock(&task->pi_lock);
687         put_task_struct(task);
688
689         /*
690          * [9] check_exit_conditions_3 protected by lock->wait_lock.
691          *
692          * We must abort the chain walk if there is no lock owner even
693          * in the dead lock detection case, as we have nothing to
694          * follow here. This is the end of the chain we are walking.
695          */
696         if (!rt_mutex_owner(lock)) {
697                 /*
698                  * If the requeue [7] above changed the top waiter,
699                  * then we need to wake the new top waiter up to try
700                  * to get the lock.
701                  */
702                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
703                         wake_up_process(rt_mutex_top_waiter(lock)->task);
704                 raw_spin_unlock_irq(&lock->wait_lock);
705                 return 0;
706         }
707
708         /* [10] Grab the next task, i.e. the owner of @lock */
709         task = get_task_struct(rt_mutex_owner(lock));
710         raw_spin_lock(&task->pi_lock);
711
712         /* [11] requeue the pi waiters if necessary */
713         if (waiter == rt_mutex_top_waiter(lock)) {
714                 /*
715                  * The waiter became the new top (highest priority)
716                  * waiter on the lock. Replace the previous top waiter
717                  * in the owner tasks pi waiters tree with this waiter
718                  * and adjust the priority of the owner.
719                  */
720                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
721                 rt_mutex_enqueue_pi(task, waiter);
722                 rt_mutex_adjust_prio(task);
723
724         } else if (prerequeue_top_waiter == waiter) {
725                 /*
726                  * The waiter was the top waiter on the lock, but is
727                  * no longer the top prority waiter. Replace waiter in
728                  * the owner tasks pi waiters tree with the new top
729                  * (highest priority) waiter and adjust the priority
730                  * of the owner.
731                  * The new top waiter is stored in @waiter so that
732                  * @waiter == @top_waiter evaluates to true below and
733                  * we continue to deboost the rest of the chain.
734                  */
735                 rt_mutex_dequeue_pi(task, waiter);
736                 waiter = rt_mutex_top_waiter(lock);
737                 rt_mutex_enqueue_pi(task, waiter);
738                 rt_mutex_adjust_prio(task);
739         } else {
740                 /*
741                  * Nothing changed. No need to do any priority
742                  * adjustment.
743                  */
744         }
745
746         /*
747          * [12] check_exit_conditions_4() protected by task->pi_lock
748          * and lock->wait_lock. The actual decisions are made after we
749          * dropped the locks.
750          *
751          * Check whether the task which owns the current lock is pi
752          * blocked itself. If yes we store a pointer to the lock for
753          * the lock chain change detection above. After we dropped
754          * task->pi_lock next_lock cannot be dereferenced anymore.
755          */
756         next_lock = task_blocked_on_lock(task);
757         /*
758          * Store the top waiter of @lock for the end of chain walk
759          * decision below.
760          */
761         top_waiter = rt_mutex_top_waiter(lock);
762
763         /* [13] Drop the locks */
764         raw_spin_unlock(&task->pi_lock);
765         raw_spin_unlock_irq(&lock->wait_lock);
766
767         /*
768          * Make the actual exit decisions [12], based on the stored
769          * values.
770          *
771          * We reached the end of the lock chain. Stop right here. No
772          * point to go back just to figure that out.
773          */
774         if (!next_lock)
775                 goto out_put_task;
776
777         /*
778          * If the current waiter is not the top waiter on the lock,
779          * then we can stop the chain walk here if we are not in full
780          * deadlock detection mode.
781          */
782         if (!detect_deadlock && waiter != top_waiter)
783                 goto out_put_task;
784
785         goto again;
786
787  out_unlock_pi:
788         raw_spin_unlock_irq(&task->pi_lock);
789  out_put_task:
790         put_task_struct(task);
791
792         return ret;
793 }
794
795 /*
796  * Try to take an rt-mutex
797  *
798  * Must be called with lock->wait_lock held and interrupts disabled
799  *
800  * @lock:   The lock to be acquired.
801  * @task:   The task which wants to acquire the lock
802  * @waiter: The waiter that is queued to the lock's wait tree if the
803  *          callsite called task_blocked_on_lock(), otherwise NULL
804  */
805 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
806                                 struct rt_mutex_waiter *waiter)
807 {
808         lockdep_assert_held(&lock->wait_lock);
809
810         /*
811          * Before testing whether we can acquire @lock, we set the
812          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
813          * other tasks which try to modify @lock into the slow path
814          * and they serialize on @lock->wait_lock.
815          *
816          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
817          * as explained at the top of this file if and only if:
818          *
819          * - There is a lock owner. The caller must fixup the
820          *   transient state if it does a trylock or leaves the lock
821          *   function due to a signal or timeout.
822          *
823          * - @task acquires the lock and there are no other
824          *   waiters. This is undone in rt_mutex_set_owner(@task) at
825          *   the end of this function.
826          */
827         mark_rt_mutex_waiters(lock);
828
829         /*
830          * If @lock has an owner, give up.
831          */
832         if (rt_mutex_owner(lock))
833                 return 0;
834
835         /*
836          * If @waiter != NULL, @task has already enqueued the waiter
837          * into @lock waiter tree. If @waiter == NULL then this is a
838          * trylock attempt.
839          */
840         if (waiter) {
841                 /*
842                  * If waiter is not the highest priority waiter of
843                  * @lock, give up.
844                  */
845                 if (waiter != rt_mutex_top_waiter(lock))
846                         return 0;
847
848                 /*
849                  * We can acquire the lock. Remove the waiter from the
850                  * lock waiters tree.
851                  */
852                 rt_mutex_dequeue(lock, waiter);
853
854         } else {
855                 /*
856                  * If the lock has waiters already we check whether @task is
857                  * eligible to take over the lock.
858                  *
859                  * If there are no other waiters, @task can acquire
860                  * the lock.  @task->pi_blocked_on is NULL, so it does
861                  * not need to be dequeued.
862                  */
863                 if (rt_mutex_has_waiters(lock)) {
864                         /*
865                          * If @task->prio is greater than or equal to
866                          * the top waiter priority (kernel view),
867                          * @task lost.
868                          */
869                         if (!rt_mutex_waiter_less(task_to_waiter(task),
870                                                   rt_mutex_top_waiter(lock)))
871                                 return 0;
872
873                         /*
874                          * The current top waiter stays enqueued. We
875                          * don't have to change anything in the lock
876                          * waiters order.
877                          */
878                 } else {
879                         /*
880                          * No waiters. Take the lock without the
881                          * pi_lock dance.@task->pi_blocked_on is NULL
882                          * and we have no waiters to enqueue in @task
883                          * pi waiters tree.
884                          */
885                         goto takeit;
886                 }
887         }
888
889         /*
890          * Clear @task->pi_blocked_on. Requires protection by
891          * @task->pi_lock. Redundant operation for the @waiter == NULL
892          * case, but conditionals are more expensive than a redundant
893          * store.
894          */
895         raw_spin_lock(&task->pi_lock);
896         task->pi_blocked_on = NULL;
897         /*
898          * Finish the lock acquisition. @task is the new owner. If
899          * other waiters exist we have to insert the highest priority
900          * waiter into @task->pi_waiters tree.
901          */
902         if (rt_mutex_has_waiters(lock))
903                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
904         raw_spin_unlock(&task->pi_lock);
905
906 takeit:
907         /* We got the lock. */
908         debug_rt_mutex_lock(lock);
909
910         /*
911          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
912          * are still waiters or clears it.
913          */
914         rt_mutex_set_owner(lock, task);
915
916         return 1;
917 }
918
919 /*
920  * Task blocks on lock.
921  *
922  * Prepare waiter and propagate pi chain
923  *
924  * This must be called with lock->wait_lock held and interrupts disabled
925  */
926 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
927                                    struct rt_mutex_waiter *waiter,
928                                    struct task_struct *task,
929                                    enum rtmutex_chainwalk chwalk)
930 {
931         struct task_struct *owner = rt_mutex_owner(lock);
932         struct rt_mutex_waiter *top_waiter = waiter;
933         struct rt_mutex *next_lock;
934         int chain_walk = 0, res;
935
936         lockdep_assert_held(&lock->wait_lock);
937
938         /*
939          * Early deadlock detection. We really don't want the task to
940          * enqueue on itself just to untangle the mess later. It's not
941          * only an optimization. We drop the locks, so another waiter
942          * can come in before the chain walk detects the deadlock. So
943          * the other will detect the deadlock and return -EDEADLOCK,
944          * which is wrong, as the other waiter is not in a deadlock
945          * situation.
946          */
947         if (owner == task)
948                 return -EDEADLK;
949
950         raw_spin_lock(&task->pi_lock);
951         waiter->task = task;
952         waiter->lock = lock;
953         waiter->prio = task->prio;
954         waiter->deadline = task->dl.deadline;
955
956         /* Get the top priority waiter on the lock */
957         if (rt_mutex_has_waiters(lock))
958                 top_waiter = rt_mutex_top_waiter(lock);
959         rt_mutex_enqueue(lock, waiter);
960
961         task->pi_blocked_on = waiter;
962
963         raw_spin_unlock(&task->pi_lock);
964
965         if (!owner)
966                 return 0;
967
968         raw_spin_lock(&owner->pi_lock);
969         if (waiter == rt_mutex_top_waiter(lock)) {
970                 rt_mutex_dequeue_pi(owner, top_waiter);
971                 rt_mutex_enqueue_pi(owner, waiter);
972
973                 rt_mutex_adjust_prio(owner);
974                 if (owner->pi_blocked_on)
975                         chain_walk = 1;
976         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
977                 chain_walk = 1;
978         }
979
980         /* Store the lock on which owner is blocked or NULL */
981         next_lock = task_blocked_on_lock(owner);
982
983         raw_spin_unlock(&owner->pi_lock);
984         /*
985          * Even if full deadlock detection is on, if the owner is not
986          * blocked itself, we can avoid finding this out in the chain
987          * walk.
988          */
989         if (!chain_walk || !next_lock)
990                 return 0;
991
992         /*
993          * The owner can't disappear while holding a lock,
994          * so the owner struct is protected by wait_lock.
995          * Gets dropped in rt_mutex_adjust_prio_chain()!
996          */
997         get_task_struct(owner);
998
999         raw_spin_unlock_irq(&lock->wait_lock);
1000
1001         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1002                                          next_lock, waiter, task);
1003
1004         raw_spin_lock_irq(&lock->wait_lock);
1005
1006         return res;
1007 }
1008
1009 /*
1010  * Remove the top waiter from the current tasks pi waiter tree and
1011  * queue it up.
1012  *
1013  * Called with lock->wait_lock held and interrupts disabled.
1014  */
1015 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1016                                     struct rt_mutex *lock)
1017 {
1018         struct rt_mutex_waiter *waiter;
1019
1020         raw_spin_lock(&current->pi_lock);
1021
1022         waiter = rt_mutex_top_waiter(lock);
1023
1024         /*
1025          * Remove it from current->pi_waiters and deboost.
1026          *
1027          * We must in fact deboost here in order to ensure we call
1028          * rt_mutex_setprio() to update p->pi_top_task before the
1029          * task unblocks.
1030          */
1031         rt_mutex_dequeue_pi(current, waiter);
1032         rt_mutex_adjust_prio(current);
1033
1034         /*
1035          * As we are waking up the top waiter, and the waiter stays
1036          * queued on the lock until it gets the lock, this lock
1037          * obviously has waiters. Just set the bit here and this has
1038          * the added benefit of forcing all new tasks into the
1039          * slow path making sure no task of lower priority than
1040          * the top waiter can steal this lock.
1041          */
1042         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1043
1044         /*
1045          * We deboosted before waking the top waiter task such that we don't
1046          * run two tasks with the 'same' priority (and ensure the
1047          * p->pi_top_task pointer points to a blocked task). This however can
1048          * lead to priority inversion if we would get preempted after the
1049          * deboost but before waking our donor task, hence the preempt_disable()
1050          * before unlock.
1051          *
1052          * Pairs with preempt_enable() in rt_mutex_postunlock();
1053          */
1054         preempt_disable();
1055         wake_q_add(wake_q, waiter->task);
1056         raw_spin_unlock(&current->pi_lock);
1057 }
1058
1059 /*
1060  * Remove a waiter from a lock and give up
1061  *
1062  * Must be called with lock->wait_lock held and interrupts disabled. I must
1063  * have just failed to try_to_take_rt_mutex().
1064  */
1065 static void remove_waiter(struct rt_mutex *lock,
1066                           struct rt_mutex_waiter *waiter)
1067 {
1068         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1069         struct task_struct *owner = rt_mutex_owner(lock);
1070         struct rt_mutex *next_lock;
1071
1072         lockdep_assert_held(&lock->wait_lock);
1073
1074         raw_spin_lock(&current->pi_lock);
1075         rt_mutex_dequeue(lock, waiter);
1076         current->pi_blocked_on = NULL;
1077         raw_spin_unlock(&current->pi_lock);
1078
1079         /*
1080          * Only update priority if the waiter was the highest priority
1081          * waiter of the lock and there is an owner to update.
1082          */
1083         if (!owner || !is_top_waiter)
1084                 return;
1085
1086         raw_spin_lock(&owner->pi_lock);
1087
1088         rt_mutex_dequeue_pi(owner, waiter);
1089
1090         if (rt_mutex_has_waiters(lock))
1091                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1092
1093         rt_mutex_adjust_prio(owner);
1094
1095         /* Store the lock on which owner is blocked or NULL */
1096         next_lock = task_blocked_on_lock(owner);
1097
1098         raw_spin_unlock(&owner->pi_lock);
1099
1100         /*
1101          * Don't walk the chain, if the owner task is not blocked
1102          * itself.
1103          */
1104         if (!next_lock)
1105                 return;
1106
1107         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1108         get_task_struct(owner);
1109
1110         raw_spin_unlock_irq(&lock->wait_lock);
1111
1112         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1113                                    next_lock, NULL, current);
1114
1115         raw_spin_lock_irq(&lock->wait_lock);
1116 }
1117
1118 /*
1119  * Recheck the pi chain, in case we got a priority setting
1120  *
1121  * Called from sched_setscheduler
1122  */
1123 void rt_mutex_adjust_pi(struct task_struct *task)
1124 {
1125         struct rt_mutex_waiter *waiter;
1126         struct rt_mutex *next_lock;
1127         unsigned long flags;
1128
1129         raw_spin_lock_irqsave(&task->pi_lock, flags);
1130
1131         waiter = task->pi_blocked_on;
1132         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1133                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1134                 return;
1135         }
1136         next_lock = waiter->lock;
1137         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1138
1139         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1140         get_task_struct(task);
1141
1142         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1143                                    next_lock, NULL, task);
1144 }
1145
1146 void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1147 {
1148         debug_rt_mutex_init_waiter(waiter);
1149         RB_CLEAR_NODE(&waiter->pi_tree_entry);
1150         RB_CLEAR_NODE(&waiter->tree_entry);
1151         waiter->task = NULL;
1152 }
1153
1154 /**
1155  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1156  * @lock:                the rt_mutex to take
1157  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1158  *                       or TASK_UNINTERRUPTIBLE)
1159  * @timeout:             the pre-initialized and started timer, or NULL for none
1160  * @waiter:              the pre-initialized rt_mutex_waiter
1161  *
1162  * Must be called with lock->wait_lock held and interrupts disabled
1163  */
1164 static int __sched
1165 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1166                     struct hrtimer_sleeper *timeout,
1167                     struct rt_mutex_waiter *waiter)
1168 {
1169         int ret = 0;
1170
1171         for (;;) {
1172                 /* Try to acquire the lock: */
1173                 if (try_to_take_rt_mutex(lock, current, waiter))
1174                         break;
1175
1176                 /*
1177                  * TASK_INTERRUPTIBLE checks for signals and
1178                  * timeout. Ignored otherwise.
1179                  */
1180                 if (likely(state == TASK_INTERRUPTIBLE)) {
1181                         /* Signal pending? */
1182                         if (signal_pending(current))
1183                                 ret = -EINTR;
1184                         if (timeout && !timeout->task)
1185                                 ret = -ETIMEDOUT;
1186                         if (ret)
1187                                 break;
1188                 }
1189
1190                 raw_spin_unlock_irq(&lock->wait_lock);
1191
1192                 debug_rt_mutex_print_deadlock(waiter);
1193
1194                 schedule();
1195
1196                 raw_spin_lock_irq(&lock->wait_lock);
1197                 set_current_state(state);
1198         }
1199
1200         __set_current_state(TASK_RUNNING);
1201         return ret;
1202 }
1203
1204 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1205                                      struct rt_mutex_waiter *w)
1206 {
1207         /*
1208          * If the result is not -EDEADLOCK or the caller requested
1209          * deadlock detection, nothing to do here.
1210          */
1211         if (res != -EDEADLOCK || detect_deadlock)
1212                 return;
1213
1214         /*
1215          * Yell lowdly and stop the task right here.
1216          */
1217         rt_mutex_print_deadlock(w);
1218         while (1) {
1219                 set_current_state(TASK_INTERRUPTIBLE);
1220                 schedule();
1221         }
1222 }
1223
1224 /*
1225  * Slow path lock function:
1226  */
1227 static int __sched
1228 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1229                   struct hrtimer_sleeper *timeout,
1230                   enum rtmutex_chainwalk chwalk)
1231 {
1232         struct rt_mutex_waiter waiter;
1233         unsigned long flags;
1234         int ret = 0;
1235
1236         rt_mutex_init_waiter(&waiter);
1237
1238         /*
1239          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1240          * be called in early boot if the cmpxchg() fast path is disabled
1241          * (debug, no architecture support). In this case we will acquire the
1242          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1243          * enable interrupts in that early boot case. So we need to use the
1244          * irqsave/restore variants.
1245          */
1246         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1247
1248         /* Try to acquire the lock again: */
1249         if (try_to_take_rt_mutex(lock, current, NULL)) {
1250                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1251                 return 0;
1252         }
1253
1254         set_current_state(state);
1255
1256         /* Setup the timer, when timeout != NULL */
1257         if (unlikely(timeout))
1258                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1259
1260         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1261
1262         if (likely(!ret))
1263                 /* sleep on the mutex */
1264                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1265
1266         if (unlikely(ret)) {
1267                 __set_current_state(TASK_RUNNING);
1268                 remove_waiter(lock, &waiter);
1269                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1270         }
1271
1272         /*
1273          * try_to_take_rt_mutex() sets the waiter bit
1274          * unconditionally. We might have to fix that up.
1275          */
1276         fixup_rt_mutex_waiters(lock);
1277
1278         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1279
1280         /* Remove pending timer: */
1281         if (unlikely(timeout))
1282                 hrtimer_cancel(&timeout->timer);
1283
1284         debug_rt_mutex_free_waiter(&waiter);
1285
1286         return ret;
1287 }
1288
1289 static inline int __rt_mutex_slowtrylock(struct rt_mutex *lock)
1290 {
1291         int ret = try_to_take_rt_mutex(lock, current, NULL);
1292
1293         /*
1294          * try_to_take_rt_mutex() sets the lock waiters bit
1295          * unconditionally. Clean this up.
1296          */
1297         fixup_rt_mutex_waiters(lock);
1298
1299         return ret;
1300 }
1301
1302 /*
1303  * Slow path try-lock function:
1304  */
1305 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1306 {
1307         unsigned long flags;
1308         int ret;
1309
1310         /*
1311          * If the lock already has an owner we fail to get the lock.
1312          * This can be done without taking the @lock->wait_lock as
1313          * it is only being read, and this is a trylock anyway.
1314          */
1315         if (rt_mutex_owner(lock))
1316                 return 0;
1317
1318         /*
1319          * The mutex has currently no owner. Lock the wait lock and try to
1320          * acquire the lock. We use irqsave here to support early boot calls.
1321          */
1322         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1323
1324         ret = __rt_mutex_slowtrylock(lock);
1325
1326         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1327
1328         return ret;
1329 }
1330
1331 /*
1332  * Slow path to release a rt-mutex.
1333  *
1334  * Return whether the current task needs to call rt_mutex_postunlock().
1335  */
1336 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1337                                         struct wake_q_head *wake_q)
1338 {
1339         unsigned long flags;
1340
1341         /* irqsave required to support early boot calls */
1342         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1343
1344         debug_rt_mutex_unlock(lock);
1345
1346         /*
1347          * We must be careful here if the fast path is enabled. If we
1348          * have no waiters queued we cannot set owner to NULL here
1349          * because of:
1350          *
1351          * foo->lock->owner = NULL;
1352          *                      rtmutex_lock(foo->lock);   <- fast path
1353          *                      free = atomic_dec_and_test(foo->refcnt);
1354          *                      rtmutex_unlock(foo->lock); <- fast path
1355          *                      if (free)
1356          *                              kfree(foo);
1357          * raw_spin_unlock(foo->lock->wait_lock);
1358          *
1359          * So for the fastpath enabled kernel:
1360          *
1361          * Nothing can set the waiters bit as long as we hold
1362          * lock->wait_lock. So we do the following sequence:
1363          *
1364          *      owner = rt_mutex_owner(lock);
1365          *      clear_rt_mutex_waiters(lock);
1366          *      raw_spin_unlock(&lock->wait_lock);
1367          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1368          *              return;
1369          *      goto retry;
1370          *
1371          * The fastpath disabled variant is simple as all access to
1372          * lock->owner is serialized by lock->wait_lock:
1373          *
1374          *      lock->owner = NULL;
1375          *      raw_spin_unlock(&lock->wait_lock);
1376          */
1377         while (!rt_mutex_has_waiters(lock)) {
1378                 /* Drops lock->wait_lock ! */
1379                 if (unlock_rt_mutex_safe(lock, flags) == true)
1380                         return false;
1381                 /* Relock the rtmutex and try again */
1382                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1383         }
1384
1385         /*
1386          * The wakeup next waiter path does not suffer from the above
1387          * race. See the comments there.
1388          *
1389          * Queue the next waiter for wakeup once we release the wait_lock.
1390          */
1391         mark_wakeup_next_waiter(wake_q, lock);
1392         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1393
1394         return true; /* call rt_mutex_postunlock() */
1395 }
1396
1397 /*
1398  * debug aware fast / slowpath lock,trylock,unlock
1399  *
1400  * The atomic acquire/release ops are compiled away, when either the
1401  * architecture does not support cmpxchg or when debugging is enabled.
1402  */
1403 static inline int
1404 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1405                   int (*slowfn)(struct rt_mutex *lock, int state,
1406                                 struct hrtimer_sleeper *timeout,
1407                                 enum rtmutex_chainwalk chwalk))
1408 {
1409         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1410                 return 0;
1411
1412         return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1413 }
1414
1415 static inline int
1416 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1417                         struct hrtimer_sleeper *timeout,
1418                         enum rtmutex_chainwalk chwalk,
1419                         int (*slowfn)(struct rt_mutex *lock, int state,
1420                                       struct hrtimer_sleeper *timeout,
1421                                       enum rtmutex_chainwalk chwalk))
1422 {
1423         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1424             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1425                 return 0;
1426
1427         return slowfn(lock, state, timeout, chwalk);
1428 }
1429
1430 static inline int
1431 rt_mutex_fasttrylock(struct rt_mutex *lock,
1432                      int (*slowfn)(struct rt_mutex *lock))
1433 {
1434         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1435                 return 1;
1436
1437         return slowfn(lock);
1438 }
1439
1440 /*
1441  * Performs the wakeup of the the top-waiter and re-enables preemption.
1442  */
1443 void rt_mutex_postunlock(struct wake_q_head *wake_q)
1444 {
1445         wake_up_q(wake_q);
1446
1447         /* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1448         preempt_enable();
1449 }
1450
1451 static inline void
1452 rt_mutex_fastunlock(struct rt_mutex *lock,
1453                     bool (*slowfn)(struct rt_mutex *lock,
1454                                    struct wake_q_head *wqh))
1455 {
1456         DEFINE_WAKE_Q(wake_q);
1457
1458         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1459                 return;
1460
1461         if (slowfn(lock, &wake_q))
1462                 rt_mutex_postunlock(&wake_q);
1463 }
1464
1465 static inline void __rt_mutex_lock(struct rt_mutex *lock, unsigned int subclass)
1466 {
1467         might_sleep();
1468
1469         mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
1470         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1471 }
1472
1473 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1474 /**
1475  * rt_mutex_lock_nested - lock a rt_mutex
1476  *
1477  * @lock: the rt_mutex to be locked
1478  * @subclass: the lockdep subclass
1479  */
1480 void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
1481 {
1482         __rt_mutex_lock(lock, subclass);
1483 }
1484 EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
1485
1486 #else /* !CONFIG_DEBUG_LOCK_ALLOC */
1487
1488 /**
1489  * rt_mutex_lock - lock a rt_mutex
1490  *
1491  * @lock: the rt_mutex to be locked
1492  */
1493 void __sched rt_mutex_lock(struct rt_mutex *lock)
1494 {
1495         __rt_mutex_lock(lock, 0);
1496 }
1497 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1498 #endif
1499
1500 /**
1501  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1502  *
1503  * @lock:               the rt_mutex to be locked
1504  *
1505  * Returns:
1506  *  0           on success
1507  * -EINTR       when interrupted by a signal
1508  */
1509 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1510 {
1511         int ret;
1512
1513         might_sleep();
1514
1515         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1516         ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1517         if (ret)
1518                 mutex_release(&lock->dep_map, _RET_IP_);
1519
1520         return ret;
1521 }
1522 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1523
1524 /*
1525  * Futex variant, must not use fastpath.
1526  */
1527 int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1528 {
1529         return rt_mutex_slowtrylock(lock);
1530 }
1531
1532 int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1533 {
1534         return __rt_mutex_slowtrylock(lock);
1535 }
1536
1537 /**
1538  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1539  *                      the timeout structure is provided
1540  *                      by the caller
1541  *
1542  * @lock:               the rt_mutex to be locked
1543  * @timeout:            timeout structure or NULL (no timeout)
1544  *
1545  * Returns:
1546  *  0           on success
1547  * -EINTR       when interrupted by a signal
1548  * -ETIMEDOUT   when the timeout expired
1549  */
1550 int
1551 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1552 {
1553         int ret;
1554
1555         might_sleep();
1556
1557         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1558         ret = rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1559                                        RT_MUTEX_MIN_CHAINWALK,
1560                                        rt_mutex_slowlock);
1561         if (ret)
1562                 mutex_release(&lock->dep_map, _RET_IP_);
1563
1564         return ret;
1565 }
1566 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1567
1568 /**
1569  * rt_mutex_trylock - try to lock a rt_mutex
1570  *
1571  * @lock:       the rt_mutex to be locked
1572  *
1573  * This function can only be called in thread context. It's safe to
1574  * call it from atomic regions, but not from hard interrupt or soft
1575  * interrupt context.
1576  *
1577  * Returns 1 on success and 0 on contention
1578  */
1579 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1580 {
1581         int ret;
1582
1583         if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1584                 return 0;
1585
1586         ret = rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1587         if (ret)
1588                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1589
1590         return ret;
1591 }
1592 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1593
1594 /**
1595  * rt_mutex_unlock - unlock a rt_mutex
1596  *
1597  * @lock: the rt_mutex to be unlocked
1598  */
1599 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1600 {
1601         mutex_release(&lock->dep_map, _RET_IP_);
1602         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1603 }
1604 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1605
1606 /**
1607  * Futex variant, that since futex variants do not use the fast-path, can be
1608  * simple and will not need to retry.
1609  */
1610 bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1611                                     struct wake_q_head *wake_q)
1612 {
1613         lockdep_assert_held(&lock->wait_lock);
1614
1615         debug_rt_mutex_unlock(lock);
1616
1617         if (!rt_mutex_has_waiters(lock)) {
1618                 lock->owner = NULL;
1619                 return false; /* done */
1620         }
1621
1622         /*
1623          * We've already deboosted, mark_wakeup_next_waiter() will
1624          * retain preempt_disabled when we drop the wait_lock, to
1625          * avoid inversion prior to the wakeup.  preempt_disable()
1626          * therein pairs with rt_mutex_postunlock().
1627          */
1628         mark_wakeup_next_waiter(wake_q, lock);
1629
1630         return true; /* call postunlock() */
1631 }
1632
1633 void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1634 {
1635         DEFINE_WAKE_Q(wake_q);
1636         unsigned long flags;
1637         bool postunlock;
1638
1639         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1640         postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1641         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1642
1643         if (postunlock)
1644                 rt_mutex_postunlock(&wake_q);
1645 }
1646
1647 /**
1648  * rt_mutex_destroy - mark a mutex unusable
1649  * @lock: the mutex to be destroyed
1650  *
1651  * This function marks the mutex uninitialized, and any subsequent
1652  * use of the mutex is forbidden. The mutex must not be locked when
1653  * this function is called.
1654  */
1655 void rt_mutex_destroy(struct rt_mutex *lock)
1656 {
1657         WARN_ON(rt_mutex_is_locked(lock));
1658 #ifdef CONFIG_DEBUG_RT_MUTEXES
1659         lock->magic = NULL;
1660 #endif
1661 }
1662 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1663
1664 /**
1665  * __rt_mutex_init - initialize the rt lock
1666  *
1667  * @lock: the rt lock to be initialized
1668  *
1669  * Initialize the rt lock to unlocked state.
1670  *
1671  * Initializing of a locked rt lock is not allowed
1672  */
1673 void __rt_mutex_init(struct rt_mutex *lock, const char *name,
1674                      struct lock_class_key *key)
1675 {
1676         lock->owner = NULL;
1677         raw_spin_lock_init(&lock->wait_lock);
1678         lock->waiters = RB_ROOT_CACHED;
1679
1680         if (name && key)
1681                 debug_rt_mutex_init(lock, name, key);
1682 }
1683 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1684
1685 /**
1686  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1687  *                              proxy owner
1688  *
1689  * @lock:       the rt_mutex to be locked
1690  * @proxy_owner:the task to set as owner
1691  *
1692  * No locking. Caller has to do serializing itself
1693  *
1694  * Special API call for PI-futex support. This initializes the rtmutex and
1695  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1696  * possible at this point because the pi_state which contains the rtmutex
1697  * is not yet visible to other tasks.
1698  */
1699 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1700                                 struct task_struct *proxy_owner)
1701 {
1702         __rt_mutex_init(lock, NULL, NULL);
1703         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1704         rt_mutex_set_owner(lock, proxy_owner);
1705 }
1706
1707 /**
1708  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1709  *
1710  * @lock:       the rt_mutex to be locked
1711  *
1712  * No locking. Caller has to do serializing itself
1713  *
1714  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1715  * (debugging) state. Concurrent operations on this rt_mutex are not
1716  * possible because it belongs to the pi_state which is about to be freed
1717  * and it is not longer visible to other tasks.
1718  */
1719 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1720                            struct task_struct *proxy_owner)
1721 {
1722         debug_rt_mutex_proxy_unlock(lock);
1723         rt_mutex_set_owner(lock, NULL);
1724 }
1725
1726 /**
1727  * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1728  * @lock:               the rt_mutex to take
1729  * @waiter:             the pre-initialized rt_mutex_waiter
1730  * @task:               the task to prepare
1731  *
1732  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1733  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1734  *
1735  * NOTE: does _NOT_ remove the @waiter on failure; must either call
1736  * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
1737  *
1738  * Returns:
1739  *  0 - task blocked on lock
1740  *  1 - acquired the lock for task, caller should wake it up
1741  * <0 - error
1742  *
1743  * Special API call for PI-futex support.
1744  */
1745 int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1746                               struct rt_mutex_waiter *waiter,
1747                               struct task_struct *task)
1748 {
1749         int ret;
1750
1751         lockdep_assert_held(&lock->wait_lock);
1752
1753         if (try_to_take_rt_mutex(lock, task, NULL))
1754                 return 1;
1755
1756         /* We enforce deadlock detection for futexes */
1757         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1758                                       RT_MUTEX_FULL_CHAINWALK);
1759
1760         if (ret && !rt_mutex_owner(lock)) {
1761                 /*
1762                  * Reset the return value. We might have
1763                  * returned with -EDEADLK and the owner
1764                  * released the lock while we were walking the
1765                  * pi chain.  Let the waiter sort it out.
1766                  */
1767                 ret = 0;
1768         }
1769
1770         debug_rt_mutex_print_deadlock(waiter);
1771
1772         return ret;
1773 }
1774
1775 /**
1776  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1777  * @lock:               the rt_mutex to take
1778  * @waiter:             the pre-initialized rt_mutex_waiter
1779  * @task:               the task to prepare
1780  *
1781  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1782  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1783  *
1784  * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
1785  * on failure.
1786  *
1787  * Returns:
1788  *  0 - task blocked on lock
1789  *  1 - acquired the lock for task, caller should wake it up
1790  * <0 - error
1791  *
1792  * Special API call for PI-futex support.
1793  */
1794 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1795                               struct rt_mutex_waiter *waiter,
1796                               struct task_struct *task)
1797 {
1798         int ret;
1799
1800         raw_spin_lock_irq(&lock->wait_lock);
1801         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1802         if (unlikely(ret))
1803                 remove_waiter(lock, waiter);
1804         raw_spin_unlock_irq(&lock->wait_lock);
1805
1806         return ret;
1807 }
1808
1809 /**
1810  * rt_mutex_next_owner - return the next owner of the lock
1811  *
1812  * @lock: the rt lock query
1813  *
1814  * Returns the next owner of the lock or NULL
1815  *
1816  * Caller has to serialize against other accessors to the lock
1817  * itself.
1818  *
1819  * Special API call for PI-futex support
1820  */
1821 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1822 {
1823         if (!rt_mutex_has_waiters(lock))
1824                 return NULL;
1825
1826         return rt_mutex_top_waiter(lock)->task;
1827 }
1828
1829 /**
1830  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1831  * @lock:               the rt_mutex we were woken on
1832  * @to:                 the timeout, null if none. hrtimer should already have
1833  *                      been started.
1834  * @waiter:             the pre-initialized rt_mutex_waiter
1835  *
1836  * Wait for the the lock acquisition started on our behalf by
1837  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1838  * rt_mutex_cleanup_proxy_lock().
1839  *
1840  * Returns:
1841  *  0 - success
1842  * <0 - error, one of -EINTR, -ETIMEDOUT
1843  *
1844  * Special API call for PI-futex support
1845  */
1846 int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1847                                struct hrtimer_sleeper *to,
1848                                struct rt_mutex_waiter *waiter)
1849 {
1850         int ret;
1851
1852         raw_spin_lock_irq(&lock->wait_lock);
1853         /* sleep on the mutex */
1854         set_current_state(TASK_INTERRUPTIBLE);
1855         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1856         /*
1857          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1858          * have to fix that up.
1859          */
1860         fixup_rt_mutex_waiters(lock);
1861         raw_spin_unlock_irq(&lock->wait_lock);
1862
1863         return ret;
1864 }
1865
1866 /**
1867  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1868  * @lock:               the rt_mutex we were woken on
1869  * @waiter:             the pre-initialized rt_mutex_waiter
1870  *
1871  * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
1872  * rt_mutex_wait_proxy_lock().
1873  *
1874  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1875  * in fact still be granted ownership until we're removed. Therefore we can
1876  * find we are in fact the owner and must disregard the
1877  * rt_mutex_wait_proxy_lock() failure.
1878  *
1879  * Returns:
1880  *  true  - did the cleanup, we done.
1881  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1882  *          caller should disregards its return value.
1883  *
1884  * Special API call for PI-futex support
1885  */
1886 bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1887                                  struct rt_mutex_waiter *waiter)
1888 {
1889         bool cleanup = false;
1890
1891         raw_spin_lock_irq(&lock->wait_lock);
1892         /*
1893          * Do an unconditional try-lock, this deals with the lock stealing
1894          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1895          * sets a NULL owner.
1896          *
1897          * We're not interested in the return value, because the subsequent
1898          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1899          * we will own the lock and it will have removed the waiter. If we
1900          * failed the trylock, we're still not owner and we need to remove
1901          * ourselves.
1902          */
1903         try_to_take_rt_mutex(lock, current, waiter);
1904         /*
1905          * Unless we're the owner; we're still enqueued on the wait_list.
1906          * So check if we became owner, if not, take us off the wait_list.
1907          */
1908         if (rt_mutex_owner(lock) != current) {
1909                 remove_waiter(lock, waiter);
1910                 cleanup = true;
1911         }
1912         /*
1913          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1914          * have to fix that up.
1915          */
1916         fixup_rt_mutex_waiters(lock);
1917
1918         raw_spin_unlock_irq(&lock->wait_lock);
1919
1920         return cleanup;
1921 }