locking/lockdep: Remove unused @nested argument from lock_release()
[sfrench/cifs-2.6.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         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_relaxed(l,c,n) (cmpxchg_relaxed(&l->owner, c, n) == c)
145 # define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
146 # define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
147
148 /*
149  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
150  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
151  * relaxed semantics suffice.
152  */
153 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
154 {
155         unsigned long owner, *p = (unsigned long *) &lock->owner;
156
157         do {
158                 owner = *p;
159         } while (cmpxchg_relaxed(p, owner,
160                                  owner | RT_MUTEX_HAS_WAITERS) != owner);
161 }
162
163 /*
164  * Safe fastpath aware unlock:
165  * 1) Clear the waiters bit
166  * 2) Drop lock->wait_lock
167  * 3) Try to unlock the lock with cmpxchg
168  */
169 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
170                                         unsigned long flags)
171         __releases(lock->wait_lock)
172 {
173         struct task_struct *owner = rt_mutex_owner(lock);
174
175         clear_rt_mutex_waiters(lock);
176         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
177         /*
178          * If a new waiter comes in between the unlock and the cmpxchg
179          * we have two situations:
180          *
181          * unlock(wait_lock);
182          *                                      lock(wait_lock);
183          * cmpxchg(p, owner, 0) == owner
184          *                                      mark_rt_mutex_waiters(lock);
185          *                                      acquire(lock);
186          * or:
187          *
188          * unlock(wait_lock);
189          *                                      lock(wait_lock);
190          *                                      mark_rt_mutex_waiters(lock);
191          *
192          * cmpxchg(p, owner, 0) != owner
193          *                                      enqueue_waiter();
194          *                                      unlock(wait_lock);
195          * lock(wait_lock);
196          * wake waiter();
197          * unlock(wait_lock);
198          *                                      lock(wait_lock);
199          *                                      acquire(lock);
200          */
201         return rt_mutex_cmpxchg_release(lock, owner, NULL);
202 }
203
204 #else
205 # define rt_mutex_cmpxchg_relaxed(l,c,n)        (0)
206 # define rt_mutex_cmpxchg_acquire(l,c,n)        (0)
207 # define rt_mutex_cmpxchg_release(l,c,n)        (0)
208
209 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
210 {
211         lock->owner = (struct task_struct *)
212                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
213 }
214
215 /*
216  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
217  */
218 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
219                                         unsigned long flags)
220         __releases(lock->wait_lock)
221 {
222         lock->owner = NULL;
223         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
224         return true;
225 }
226 #endif
227
228 /*
229  * Only use with rt_mutex_waiter_{less,equal}()
230  */
231 #define task_to_waiter(p)       \
232         &(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
233
234 static inline int
235 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
236                      struct rt_mutex_waiter *right)
237 {
238         if (left->prio < right->prio)
239                 return 1;
240
241         /*
242          * If both waiters have dl_prio(), we check the deadlines of the
243          * associated tasks.
244          * If left waiter has a dl_prio(), and we didn't return 1 above,
245          * then right waiter has a dl_prio() too.
246          */
247         if (dl_prio(left->prio))
248                 return dl_time_before(left->deadline, right->deadline);
249
250         return 0;
251 }
252
253 static inline int
254 rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
255                       struct rt_mutex_waiter *right)
256 {
257         if (left->prio != right->prio)
258                 return 0;
259
260         /*
261          * If both waiters have dl_prio(), we check the deadlines of the
262          * associated tasks.
263          * If left waiter has a dl_prio(), and we didn't return 0 above,
264          * then right waiter has a dl_prio() too.
265          */
266         if (dl_prio(left->prio))
267                 return left->deadline == right->deadline;
268
269         return 1;
270 }
271
272 static void
273 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
274 {
275         struct rb_node **link = &lock->waiters.rb_root.rb_node;
276         struct rb_node *parent = NULL;
277         struct rt_mutex_waiter *entry;
278         bool leftmost = true;
279
280         while (*link) {
281                 parent = *link;
282                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
283                 if (rt_mutex_waiter_less(waiter, entry)) {
284                         link = &parent->rb_left;
285                 } else {
286                         link = &parent->rb_right;
287                         leftmost = false;
288                 }
289         }
290
291         rb_link_node(&waiter->tree_entry, parent, link);
292         rb_insert_color_cached(&waiter->tree_entry, &lock->waiters, leftmost);
293 }
294
295 static void
296 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
297 {
298         if (RB_EMPTY_NODE(&waiter->tree_entry))
299                 return;
300
301         rb_erase_cached(&waiter->tree_entry, &lock->waiters);
302         RB_CLEAR_NODE(&waiter->tree_entry);
303 }
304
305 static void
306 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
307 {
308         struct rb_node **link = &task->pi_waiters.rb_root.rb_node;
309         struct rb_node *parent = NULL;
310         struct rt_mutex_waiter *entry;
311         bool leftmost = true;
312
313         while (*link) {
314                 parent = *link;
315                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
316                 if (rt_mutex_waiter_less(waiter, entry)) {
317                         link = &parent->rb_left;
318                 } else {
319                         link = &parent->rb_right;
320                         leftmost = false;
321                 }
322         }
323
324         rb_link_node(&waiter->pi_tree_entry, parent, link);
325         rb_insert_color_cached(&waiter->pi_tree_entry, &task->pi_waiters, leftmost);
326 }
327
328 static void
329 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
330 {
331         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
332                 return;
333
334         rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
335         RB_CLEAR_NODE(&waiter->pi_tree_entry);
336 }
337
338 static void rt_mutex_adjust_prio(struct task_struct *p)
339 {
340         struct task_struct *pi_task = NULL;
341
342         lockdep_assert_held(&p->pi_lock);
343
344         if (task_has_pi_waiters(p))
345                 pi_task = task_top_pi_waiter(p)->task;
346
347         rt_mutex_setprio(p, pi_task);
348 }
349
350 /*
351  * Deadlock detection is conditional:
352  *
353  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
354  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
355  *
356  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
357  * conducted independent of the detect argument.
358  *
359  * If the waiter argument is NULL this indicates the deboost path and
360  * deadlock detection is disabled independent of the detect argument
361  * and the config settings.
362  */
363 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
364                                           enum rtmutex_chainwalk chwalk)
365 {
366         /*
367          * This is just a wrapper function for the following call,
368          * because debug_rt_mutex_detect_deadlock() smells like a magic
369          * debug feature and I wanted to keep the cond function in the
370          * main source file along with the comments instead of having
371          * two of the same in the headers.
372          */
373         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
374 }
375
376 /*
377  * Max number of times we'll walk the boosting chain:
378  */
379 int max_lock_depth = 1024;
380
381 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
382 {
383         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
384 }
385
386 /*
387  * Adjust the priority chain. Also used for deadlock detection.
388  * Decreases task's usage by one - may thus free the task.
389  *
390  * @task:       the task owning the mutex (owner) for which a chain walk is
391  *              probably needed
392  * @chwalk:     do we have to carry out deadlock detection?
393  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
394  *              things for a task that has just got its priority adjusted, and
395  *              is waiting on a mutex)
396  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
397  *              we dropped its pi_lock. Is never dereferenced, only used for
398  *              comparison to detect lock chain changes.
399  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
400  *              its priority to the mutex owner (can be NULL in the case
401  *              depicted above or if the top waiter is gone away and we are
402  *              actually deboosting the owner)
403  * @top_task:   the current top waiter
404  *
405  * Returns 0 or -EDEADLK.
406  *
407  * Chain walk basics and protection scope
408  *
409  * [R] refcount on task
410  * [P] task->pi_lock held
411  * [L] rtmutex->wait_lock held
412  *
413  * Step Description                             Protected by
414  *      function arguments:
415  *      @task                                   [R]
416  *      @orig_lock if != NULL                   @top_task is blocked on it
417  *      @next_lock                              Unprotected. Cannot be
418  *                                              dereferenced. Only used for
419  *                                              comparison.
420  *      @orig_waiter if != NULL                 @top_task is blocked on it
421  *      @top_task                               current, or in case of proxy
422  *                                              locking protected by calling
423  *                                              code
424  *      again:
425  *        loop_sanity_check();
426  *      retry:
427  * [1]    lock(task->pi_lock);                  [R] acquire [P]
428  * [2]    waiter = task->pi_blocked_on;         [P]
429  * [3]    check_exit_conditions_1();            [P]
430  * [4]    lock = waiter->lock;                  [P]
431  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
432  *          unlock(task->pi_lock);              release [P]
433  *          goto retry;
434  *        }
435  * [6]    check_exit_conditions_2();            [P] + [L]
436  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
437  * [8]    unlock(task->pi_lock);                release [P]
438  *        put_task_struct(task);                release [R]
439  * [9]    check_exit_conditions_3();            [L]
440  * [10]   task = owner(lock);                   [L]
441  *        get_task_struct(task);                [L] acquire [R]
442  *        lock(task->pi_lock);                  [L] acquire [P]
443  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
444  * [12]   check_exit_conditions_4();            [P] + [L]
445  * [13]   unlock(task->pi_lock);                release [P]
446  *        unlock(lock->wait_lock);              release [L]
447  *        goto again;
448  */
449 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
450                                       enum rtmutex_chainwalk chwalk,
451                                       struct rt_mutex *orig_lock,
452                                       struct rt_mutex *next_lock,
453                                       struct rt_mutex_waiter *orig_waiter,
454                                       struct task_struct *top_task)
455 {
456         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
457         struct rt_mutex_waiter *prerequeue_top_waiter;
458         int ret = 0, depth = 0;
459         struct rt_mutex *lock;
460         bool detect_deadlock;
461         bool requeue = true;
462
463         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
464
465         /*
466          * The (de)boosting is a step by step approach with a lot of
467          * pitfalls. We want this to be preemptible and we want hold a
468          * maximum of two locks per step. So we have to check
469          * carefully whether things change under us.
470          */
471  again:
472         /*
473          * We limit the lock chain length for each invocation.
474          */
475         if (++depth > max_lock_depth) {
476                 static int prev_max;
477
478                 /*
479                  * Print this only once. If the admin changes the limit,
480                  * print a new message when reaching the limit again.
481                  */
482                 if (prev_max != max_lock_depth) {
483                         prev_max = max_lock_depth;
484                         printk(KERN_WARNING "Maximum lock depth %d reached "
485                                "task: %s (%d)\n", max_lock_depth,
486                                top_task->comm, task_pid_nr(top_task));
487                 }
488                 put_task_struct(task);
489
490                 return -EDEADLK;
491         }
492
493         /*
494          * We are fully preemptible here and only hold the refcount on
495          * @task. So everything can have changed under us since the
496          * caller or our own code below (goto retry/again) dropped all
497          * locks.
498          */
499  retry:
500         /*
501          * [1] Task cannot go away as we did a get_task() before !
502          */
503         raw_spin_lock_irq(&task->pi_lock);
504
505         /*
506          * [2] Get the waiter on which @task is blocked on.
507          */
508         waiter = task->pi_blocked_on;
509
510         /*
511          * [3] check_exit_conditions_1() protected by task->pi_lock.
512          */
513
514         /*
515          * Check whether the end of the boosting chain has been
516          * reached or the state of the chain has changed while we
517          * dropped the locks.
518          */
519         if (!waiter)
520                 goto out_unlock_pi;
521
522         /*
523          * Check the orig_waiter state. After we dropped the locks,
524          * the previous owner of the lock might have released the lock.
525          */
526         if (orig_waiter && !rt_mutex_owner(orig_lock))
527                 goto out_unlock_pi;
528
529         /*
530          * We dropped all locks after taking a refcount on @task, so
531          * the task might have moved on in the lock chain or even left
532          * the chain completely and blocks now on an unrelated lock or
533          * on @orig_lock.
534          *
535          * We stored the lock on which @task was blocked in @next_lock,
536          * so we can detect the chain change.
537          */
538         if (next_lock != waiter->lock)
539                 goto out_unlock_pi;
540
541         /*
542          * Drop out, when the task has no waiters. Note,
543          * top_waiter can be NULL, when we are in the deboosting
544          * mode!
545          */
546         if (top_waiter) {
547                 if (!task_has_pi_waiters(task))
548                         goto out_unlock_pi;
549                 /*
550                  * If deadlock detection is off, we stop here if we
551                  * are not the top pi waiter of the task. If deadlock
552                  * detection is enabled we continue, but stop the
553                  * requeueing in the chain walk.
554                  */
555                 if (top_waiter != task_top_pi_waiter(task)) {
556                         if (!detect_deadlock)
557                                 goto out_unlock_pi;
558                         else
559                                 requeue = false;
560                 }
561         }
562
563         /*
564          * If the waiter priority is the same as the task priority
565          * then there is no further priority adjustment necessary.  If
566          * deadlock detection is off, we stop the chain walk. If its
567          * enabled we continue, but stop the requeueing in the chain
568          * walk.
569          */
570         if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
571                 if (!detect_deadlock)
572                         goto out_unlock_pi;
573                 else
574                         requeue = false;
575         }
576
577         /*
578          * [4] Get the next lock
579          */
580         lock = waiter->lock;
581         /*
582          * [5] We need to trylock here as we are holding task->pi_lock,
583          * which is the reverse lock order versus the other rtmutex
584          * operations.
585          */
586         if (!raw_spin_trylock(&lock->wait_lock)) {
587                 raw_spin_unlock_irq(&task->pi_lock);
588                 cpu_relax();
589                 goto retry;
590         }
591
592         /*
593          * [6] check_exit_conditions_2() protected by task->pi_lock and
594          * lock->wait_lock.
595          *
596          * Deadlock detection. If the lock is the same as the original
597          * lock which caused us to walk the lock chain or if the
598          * current lock is owned by the task which initiated the chain
599          * walk, we detected a deadlock.
600          */
601         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
602                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
603                 raw_spin_unlock(&lock->wait_lock);
604                 ret = -EDEADLK;
605                 goto out_unlock_pi;
606         }
607
608         /*
609          * If we just follow the lock chain for deadlock detection, no
610          * need to do all the requeue operations. To avoid a truckload
611          * of conditionals around the various places below, just do the
612          * minimum chain walk checks.
613          */
614         if (!requeue) {
615                 /*
616                  * No requeue[7] here. Just release @task [8]
617                  */
618                 raw_spin_unlock(&task->pi_lock);
619                 put_task_struct(task);
620
621                 /*
622                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
623                  * If there is no owner of the lock, end of chain.
624                  */
625                 if (!rt_mutex_owner(lock)) {
626                         raw_spin_unlock_irq(&lock->wait_lock);
627                         return 0;
628                 }
629
630                 /* [10] Grab the next task, i.e. owner of @lock */
631                 task = get_task_struct(rt_mutex_owner(lock));
632                 raw_spin_lock(&task->pi_lock);
633
634                 /*
635                  * No requeue [11] here. We just do deadlock detection.
636                  *
637                  * [12] Store whether owner is blocked
638                  * itself. Decision is made after dropping the locks
639                  */
640                 next_lock = task_blocked_on_lock(task);
641                 /*
642                  * Get the top waiter for the next iteration
643                  */
644                 top_waiter = rt_mutex_top_waiter(lock);
645
646                 /* [13] Drop locks */
647                 raw_spin_unlock(&task->pi_lock);
648                 raw_spin_unlock_irq(&lock->wait_lock);
649
650                 /* If owner is not blocked, end of chain. */
651                 if (!next_lock)
652                         goto out_put_task;
653                 goto again;
654         }
655
656         /*
657          * Store the current top waiter before doing the requeue
658          * operation on @lock. We need it for the boost/deboost
659          * decision below.
660          */
661         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
662
663         /* [7] Requeue the waiter in the lock waiter tree. */
664         rt_mutex_dequeue(lock, waiter);
665
666         /*
667          * Update the waiter prio fields now that we're dequeued.
668          *
669          * These values can have changed through either:
670          *
671          *   sys_sched_set_scheduler() / sys_sched_setattr()
672          *
673          * or
674          *
675          *   DL CBS enforcement advancing the effective deadline.
676          *
677          * Even though pi_waiters also uses these fields, and that tree is only
678          * updated in [11], we can do this here, since we hold [L], which
679          * serializes all pi_waiters access and rb_erase() does not care about
680          * the values of the node being removed.
681          */
682         waiter->prio = task->prio;
683         waiter->deadline = task->dl.deadline;
684
685         rt_mutex_enqueue(lock, waiter);
686
687         /* [8] Release the task */
688         raw_spin_unlock(&task->pi_lock);
689         put_task_struct(task);
690
691         /*
692          * [9] check_exit_conditions_3 protected by lock->wait_lock.
693          *
694          * We must abort the chain walk if there is no lock owner even
695          * in the dead lock detection case, as we have nothing to
696          * follow here. This is the end of the chain we are walking.
697          */
698         if (!rt_mutex_owner(lock)) {
699                 /*
700                  * If the requeue [7] above changed the top waiter,
701                  * then we need to wake the new top waiter up to try
702                  * to get the lock.
703                  */
704                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
705                         wake_up_process(rt_mutex_top_waiter(lock)->task);
706                 raw_spin_unlock_irq(&lock->wait_lock);
707                 return 0;
708         }
709
710         /* [10] Grab the next task, i.e. the owner of @lock */
711         task = get_task_struct(rt_mutex_owner(lock));
712         raw_spin_lock(&task->pi_lock);
713
714         /* [11] requeue the pi waiters if necessary */
715         if (waiter == rt_mutex_top_waiter(lock)) {
716                 /*
717                  * The waiter became the new top (highest priority)
718                  * waiter on the lock. Replace the previous top waiter
719                  * in the owner tasks pi waiters tree with this waiter
720                  * and adjust the priority of the owner.
721                  */
722                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
723                 rt_mutex_enqueue_pi(task, waiter);
724                 rt_mutex_adjust_prio(task);
725
726         } else if (prerequeue_top_waiter == waiter) {
727                 /*
728                  * The waiter was the top waiter on the lock, but is
729                  * no longer the top prority waiter. Replace waiter in
730                  * the owner tasks pi waiters tree with the new top
731                  * (highest priority) waiter and adjust the priority
732                  * of the owner.
733                  * The new top waiter is stored in @waiter so that
734                  * @waiter == @top_waiter evaluates to true below and
735                  * we continue to deboost the rest of the chain.
736                  */
737                 rt_mutex_dequeue_pi(task, waiter);
738                 waiter = rt_mutex_top_waiter(lock);
739                 rt_mutex_enqueue_pi(task, waiter);
740                 rt_mutex_adjust_prio(task);
741         } else {
742                 /*
743                  * Nothing changed. No need to do any priority
744                  * adjustment.
745                  */
746         }
747
748         /*
749          * [12] check_exit_conditions_4() protected by task->pi_lock
750          * and lock->wait_lock. The actual decisions are made after we
751          * dropped the locks.
752          *
753          * Check whether the task which owns the current lock is pi
754          * blocked itself. If yes we store a pointer to the lock for
755          * the lock chain change detection above. After we dropped
756          * task->pi_lock next_lock cannot be dereferenced anymore.
757          */
758         next_lock = task_blocked_on_lock(task);
759         /*
760          * Store the top waiter of @lock for the end of chain walk
761          * decision below.
762          */
763         top_waiter = rt_mutex_top_waiter(lock);
764
765         /* [13] Drop the locks */
766         raw_spin_unlock(&task->pi_lock);
767         raw_spin_unlock_irq(&lock->wait_lock);
768
769         /*
770          * Make the actual exit decisions [12], based on the stored
771          * values.
772          *
773          * We reached the end of the lock chain. Stop right here. No
774          * point to go back just to figure that out.
775          */
776         if (!next_lock)
777                 goto out_put_task;
778
779         /*
780          * If the current waiter is not the top waiter on the lock,
781          * then we can stop the chain walk here if we are not in full
782          * deadlock detection mode.
783          */
784         if (!detect_deadlock && waiter != top_waiter)
785                 goto out_put_task;
786
787         goto again;
788
789  out_unlock_pi:
790         raw_spin_unlock_irq(&task->pi_lock);
791  out_put_task:
792         put_task_struct(task);
793
794         return ret;
795 }
796
797 /*
798  * Try to take an rt-mutex
799  *
800  * Must be called with lock->wait_lock held and interrupts disabled
801  *
802  * @lock:   The lock to be acquired.
803  * @task:   The task which wants to acquire the lock
804  * @waiter: The waiter that is queued to the lock's wait tree if the
805  *          callsite called task_blocked_on_lock(), otherwise NULL
806  */
807 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
808                                 struct rt_mutex_waiter *waiter)
809 {
810         lockdep_assert_held(&lock->wait_lock);
811
812         /*
813          * Before testing whether we can acquire @lock, we set the
814          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
815          * other tasks which try to modify @lock into the slow path
816          * and they serialize on @lock->wait_lock.
817          *
818          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
819          * as explained at the top of this file if and only if:
820          *
821          * - There is a lock owner. The caller must fixup the
822          *   transient state if it does a trylock or leaves the lock
823          *   function due to a signal or timeout.
824          *
825          * - @task acquires the lock and there are no other
826          *   waiters. This is undone in rt_mutex_set_owner(@task) at
827          *   the end of this function.
828          */
829         mark_rt_mutex_waiters(lock);
830
831         /*
832          * If @lock has an owner, give up.
833          */
834         if (rt_mutex_owner(lock))
835                 return 0;
836
837         /*
838          * If @waiter != NULL, @task has already enqueued the waiter
839          * into @lock waiter tree. If @waiter == NULL then this is a
840          * trylock attempt.
841          */
842         if (waiter) {
843                 /*
844                  * If waiter is not the highest priority waiter of
845                  * @lock, give up.
846                  */
847                 if (waiter != rt_mutex_top_waiter(lock))
848                         return 0;
849
850                 /*
851                  * We can acquire the lock. Remove the waiter from the
852                  * lock waiters tree.
853                  */
854                 rt_mutex_dequeue(lock, waiter);
855
856         } else {
857                 /*
858                  * If the lock has waiters already we check whether @task is
859                  * eligible to take over the lock.
860                  *
861                  * If there are no other waiters, @task can acquire
862                  * the lock.  @task->pi_blocked_on is NULL, so it does
863                  * not need to be dequeued.
864                  */
865                 if (rt_mutex_has_waiters(lock)) {
866                         /*
867                          * If @task->prio is greater than or equal to
868                          * the top waiter priority (kernel view),
869                          * @task lost.
870                          */
871                         if (!rt_mutex_waiter_less(task_to_waiter(task),
872                                                   rt_mutex_top_waiter(lock)))
873                                 return 0;
874
875                         /*
876                          * The current top waiter stays enqueued. We
877                          * don't have to change anything in the lock
878                          * waiters order.
879                          */
880                 } else {
881                         /*
882                          * No waiters. Take the lock without the
883                          * pi_lock dance.@task->pi_blocked_on is NULL
884                          * and we have no waiters to enqueue in @task
885                          * pi waiters tree.
886                          */
887                         goto takeit;
888                 }
889         }
890
891         /*
892          * Clear @task->pi_blocked_on. Requires protection by
893          * @task->pi_lock. Redundant operation for the @waiter == NULL
894          * case, but conditionals are more expensive than a redundant
895          * store.
896          */
897         raw_spin_lock(&task->pi_lock);
898         task->pi_blocked_on = NULL;
899         /*
900          * Finish the lock acquisition. @task is the new owner. If
901          * other waiters exist we have to insert the highest priority
902          * waiter into @task->pi_waiters tree.
903          */
904         if (rt_mutex_has_waiters(lock))
905                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
906         raw_spin_unlock(&task->pi_lock);
907
908 takeit:
909         /* We got the lock. */
910         debug_rt_mutex_lock(lock);
911
912         /*
913          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
914          * are still waiters or clears it.
915          */
916         rt_mutex_set_owner(lock, task);
917
918         return 1;
919 }
920
921 /*
922  * Task blocks on lock.
923  *
924  * Prepare waiter and propagate pi chain
925  *
926  * This must be called with lock->wait_lock held and interrupts disabled
927  */
928 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
929                                    struct rt_mutex_waiter *waiter,
930                                    struct task_struct *task,
931                                    enum rtmutex_chainwalk chwalk)
932 {
933         struct task_struct *owner = rt_mutex_owner(lock);
934         struct rt_mutex_waiter *top_waiter = waiter;
935         struct rt_mutex *next_lock;
936         int chain_walk = 0, res;
937
938         lockdep_assert_held(&lock->wait_lock);
939
940         /*
941          * Early deadlock detection. We really don't want the task to
942          * enqueue on itself just to untangle the mess later. It's not
943          * only an optimization. We drop the locks, so another waiter
944          * can come in before the chain walk detects the deadlock. So
945          * the other will detect the deadlock and return -EDEADLOCK,
946          * which is wrong, as the other waiter is not in a deadlock
947          * situation.
948          */
949         if (owner == task)
950                 return -EDEADLK;
951
952         raw_spin_lock(&task->pi_lock);
953         waiter->task = task;
954         waiter->lock = lock;
955         waiter->prio = task->prio;
956         waiter->deadline = task->dl.deadline;
957
958         /* Get the top priority waiter on the lock */
959         if (rt_mutex_has_waiters(lock))
960                 top_waiter = rt_mutex_top_waiter(lock);
961         rt_mutex_enqueue(lock, waiter);
962
963         task->pi_blocked_on = waiter;
964
965         raw_spin_unlock(&task->pi_lock);
966
967         if (!owner)
968                 return 0;
969
970         raw_spin_lock(&owner->pi_lock);
971         if (waiter == rt_mutex_top_waiter(lock)) {
972                 rt_mutex_dequeue_pi(owner, top_waiter);
973                 rt_mutex_enqueue_pi(owner, waiter);
974
975                 rt_mutex_adjust_prio(owner);
976                 if (owner->pi_blocked_on)
977                         chain_walk = 1;
978         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
979                 chain_walk = 1;
980         }
981
982         /* Store the lock on which owner is blocked or NULL */
983         next_lock = task_blocked_on_lock(owner);
984
985         raw_spin_unlock(&owner->pi_lock);
986         /*
987          * Even if full deadlock detection is on, if the owner is not
988          * blocked itself, we can avoid finding this out in the chain
989          * walk.
990          */
991         if (!chain_walk || !next_lock)
992                 return 0;
993
994         /*
995          * The owner can't disappear while holding a lock,
996          * so the owner struct is protected by wait_lock.
997          * Gets dropped in rt_mutex_adjust_prio_chain()!
998          */
999         get_task_struct(owner);
1000
1001         raw_spin_unlock_irq(&lock->wait_lock);
1002
1003         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1004                                          next_lock, waiter, task);
1005
1006         raw_spin_lock_irq(&lock->wait_lock);
1007
1008         return res;
1009 }
1010
1011 /*
1012  * Remove the top waiter from the current tasks pi waiter tree and
1013  * queue it up.
1014  *
1015  * Called with lock->wait_lock held and interrupts disabled.
1016  */
1017 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1018                                     struct rt_mutex *lock)
1019 {
1020         struct rt_mutex_waiter *waiter;
1021
1022         raw_spin_lock(&current->pi_lock);
1023
1024         waiter = rt_mutex_top_waiter(lock);
1025
1026         /*
1027          * Remove it from current->pi_waiters and deboost.
1028          *
1029          * We must in fact deboost here in order to ensure we call
1030          * rt_mutex_setprio() to update p->pi_top_task before the
1031          * task unblocks.
1032          */
1033         rt_mutex_dequeue_pi(current, waiter);
1034         rt_mutex_adjust_prio(current);
1035
1036         /*
1037          * As we are waking up the top waiter, and the waiter stays
1038          * queued on the lock until it gets the lock, this lock
1039          * obviously has waiters. Just set the bit here and this has
1040          * the added benefit of forcing all new tasks into the
1041          * slow path making sure no task of lower priority than
1042          * the top waiter can steal this lock.
1043          */
1044         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1045
1046         /*
1047          * We deboosted before waking the top waiter task such that we don't
1048          * run two tasks with the 'same' priority (and ensure the
1049          * p->pi_top_task pointer points to a blocked task). This however can
1050          * lead to priority inversion if we would get preempted after the
1051          * deboost but before waking our donor task, hence the preempt_disable()
1052          * before unlock.
1053          *
1054          * Pairs with preempt_enable() in rt_mutex_postunlock();
1055          */
1056         preempt_disable();
1057         wake_q_add(wake_q, waiter->task);
1058         raw_spin_unlock(&current->pi_lock);
1059 }
1060
1061 /*
1062  * Remove a waiter from a lock and give up
1063  *
1064  * Must be called with lock->wait_lock held and interrupts disabled. I must
1065  * have just failed to try_to_take_rt_mutex().
1066  */
1067 static void remove_waiter(struct rt_mutex *lock,
1068                           struct rt_mutex_waiter *waiter)
1069 {
1070         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1071         struct task_struct *owner = rt_mutex_owner(lock);
1072         struct rt_mutex *next_lock;
1073
1074         lockdep_assert_held(&lock->wait_lock);
1075
1076         raw_spin_lock(&current->pi_lock);
1077         rt_mutex_dequeue(lock, waiter);
1078         current->pi_blocked_on = NULL;
1079         raw_spin_unlock(&current->pi_lock);
1080
1081         /*
1082          * Only update priority if the waiter was the highest priority
1083          * waiter of the lock and there is an owner to update.
1084          */
1085         if (!owner || !is_top_waiter)
1086                 return;
1087
1088         raw_spin_lock(&owner->pi_lock);
1089
1090         rt_mutex_dequeue_pi(owner, waiter);
1091
1092         if (rt_mutex_has_waiters(lock))
1093                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1094
1095         rt_mutex_adjust_prio(owner);
1096
1097         /* Store the lock on which owner is blocked or NULL */
1098         next_lock = task_blocked_on_lock(owner);
1099
1100         raw_spin_unlock(&owner->pi_lock);
1101
1102         /*
1103          * Don't walk the chain, if the owner task is not blocked
1104          * itself.
1105          */
1106         if (!next_lock)
1107                 return;
1108
1109         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1110         get_task_struct(owner);
1111
1112         raw_spin_unlock_irq(&lock->wait_lock);
1113
1114         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1115                                    next_lock, NULL, current);
1116
1117         raw_spin_lock_irq(&lock->wait_lock);
1118 }
1119
1120 /*
1121  * Recheck the pi chain, in case we got a priority setting
1122  *
1123  * Called from sched_setscheduler
1124  */
1125 void rt_mutex_adjust_pi(struct task_struct *task)
1126 {
1127         struct rt_mutex_waiter *waiter;
1128         struct rt_mutex *next_lock;
1129         unsigned long flags;
1130
1131         raw_spin_lock_irqsave(&task->pi_lock, flags);
1132
1133         waiter = task->pi_blocked_on;
1134         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1135                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1136                 return;
1137         }
1138         next_lock = waiter->lock;
1139         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1140
1141         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1142         get_task_struct(task);
1143
1144         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1145                                    next_lock, NULL, task);
1146 }
1147
1148 void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1149 {
1150         debug_rt_mutex_init_waiter(waiter);
1151         RB_CLEAR_NODE(&waiter->pi_tree_entry);
1152         RB_CLEAR_NODE(&waiter->tree_entry);
1153         waiter->task = NULL;
1154 }
1155
1156 /**
1157  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1158  * @lock:                the rt_mutex to take
1159  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1160  *                       or TASK_UNINTERRUPTIBLE)
1161  * @timeout:             the pre-initialized and started timer, or NULL for none
1162  * @waiter:              the pre-initialized rt_mutex_waiter
1163  *
1164  * Must be called with lock->wait_lock held and interrupts disabled
1165  */
1166 static int __sched
1167 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1168                     struct hrtimer_sleeper *timeout,
1169                     struct rt_mutex_waiter *waiter)
1170 {
1171         int ret = 0;
1172
1173         for (;;) {
1174                 /* Try to acquire the lock: */
1175                 if (try_to_take_rt_mutex(lock, current, waiter))
1176                         break;
1177
1178                 /*
1179                  * TASK_INTERRUPTIBLE checks for signals and
1180                  * timeout. Ignored otherwise.
1181                  */
1182                 if (likely(state == TASK_INTERRUPTIBLE)) {
1183                         /* Signal pending? */
1184                         if (signal_pending(current))
1185                                 ret = -EINTR;
1186                         if (timeout && !timeout->task)
1187                                 ret = -ETIMEDOUT;
1188                         if (ret)
1189                                 break;
1190                 }
1191
1192                 raw_spin_unlock_irq(&lock->wait_lock);
1193
1194                 debug_rt_mutex_print_deadlock(waiter);
1195
1196                 schedule();
1197
1198                 raw_spin_lock_irq(&lock->wait_lock);
1199                 set_current_state(state);
1200         }
1201
1202         __set_current_state(TASK_RUNNING);
1203         return ret;
1204 }
1205
1206 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1207                                      struct rt_mutex_waiter *w)
1208 {
1209         /*
1210          * If the result is not -EDEADLOCK or the caller requested
1211          * deadlock detection, nothing to do here.
1212          */
1213         if (res != -EDEADLOCK || detect_deadlock)
1214                 return;
1215
1216         /*
1217          * Yell lowdly and stop the task right here.
1218          */
1219         rt_mutex_print_deadlock(w);
1220         while (1) {
1221                 set_current_state(TASK_INTERRUPTIBLE);
1222                 schedule();
1223         }
1224 }
1225
1226 /*
1227  * Slow path lock function:
1228  */
1229 static int __sched
1230 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1231                   struct hrtimer_sleeper *timeout,
1232                   enum rtmutex_chainwalk chwalk)
1233 {
1234         struct rt_mutex_waiter waiter;
1235         unsigned long flags;
1236         int ret = 0;
1237
1238         rt_mutex_init_waiter(&waiter);
1239
1240         /*
1241          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1242          * be called in early boot if the cmpxchg() fast path is disabled
1243          * (debug, no architecture support). In this case we will acquire the
1244          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1245          * enable interrupts in that early boot case. So we need to use the
1246          * irqsave/restore variants.
1247          */
1248         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1249
1250         /* Try to acquire the lock again: */
1251         if (try_to_take_rt_mutex(lock, current, NULL)) {
1252                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1253                 return 0;
1254         }
1255
1256         set_current_state(state);
1257
1258         /* Setup the timer, when timeout != NULL */
1259         if (unlikely(timeout))
1260                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1261
1262         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1263
1264         if (likely(!ret))
1265                 /* sleep on the mutex */
1266                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1267
1268         if (unlikely(ret)) {
1269                 __set_current_state(TASK_RUNNING);
1270                 remove_waiter(lock, &waiter);
1271                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1272         }
1273
1274         /*
1275          * try_to_take_rt_mutex() sets the waiter bit
1276          * unconditionally. We might have to fix that up.
1277          */
1278         fixup_rt_mutex_waiters(lock);
1279
1280         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1281
1282         /* Remove pending timer: */
1283         if (unlikely(timeout))
1284                 hrtimer_cancel(&timeout->timer);
1285
1286         debug_rt_mutex_free_waiter(&waiter);
1287
1288         return ret;
1289 }
1290
1291 static inline int __rt_mutex_slowtrylock(struct rt_mutex *lock)
1292 {
1293         int ret = try_to_take_rt_mutex(lock, current, NULL);
1294
1295         /*
1296          * try_to_take_rt_mutex() sets the lock waiters bit
1297          * unconditionally. Clean this up.
1298          */
1299         fixup_rt_mutex_waiters(lock);
1300
1301         return ret;
1302 }
1303
1304 /*
1305  * Slow path try-lock function:
1306  */
1307 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1308 {
1309         unsigned long flags;
1310         int ret;
1311
1312         /*
1313          * If the lock already has an owner we fail to get the lock.
1314          * This can be done without taking the @lock->wait_lock as
1315          * it is only being read, and this is a trylock anyway.
1316          */
1317         if (rt_mutex_owner(lock))
1318                 return 0;
1319
1320         /*
1321          * The mutex has currently no owner. Lock the wait lock and try to
1322          * acquire the lock. We use irqsave here to support early boot calls.
1323          */
1324         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1325
1326         ret = __rt_mutex_slowtrylock(lock);
1327
1328         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1329
1330         return ret;
1331 }
1332
1333 /*
1334  * Slow path to release a rt-mutex.
1335  *
1336  * Return whether the current task needs to call rt_mutex_postunlock().
1337  */
1338 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1339                                         struct wake_q_head *wake_q)
1340 {
1341         unsigned long flags;
1342
1343         /* irqsave required to support early boot calls */
1344         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1345
1346         debug_rt_mutex_unlock(lock);
1347
1348         /*
1349          * We must be careful here if the fast path is enabled. If we
1350          * have no waiters queued we cannot set owner to NULL here
1351          * because of:
1352          *
1353          * foo->lock->owner = NULL;
1354          *                      rtmutex_lock(foo->lock);   <- fast path
1355          *                      free = atomic_dec_and_test(foo->refcnt);
1356          *                      rtmutex_unlock(foo->lock); <- fast path
1357          *                      if (free)
1358          *                              kfree(foo);
1359          * raw_spin_unlock(foo->lock->wait_lock);
1360          *
1361          * So for the fastpath enabled kernel:
1362          *
1363          * Nothing can set the waiters bit as long as we hold
1364          * lock->wait_lock. So we do the following sequence:
1365          *
1366          *      owner = rt_mutex_owner(lock);
1367          *      clear_rt_mutex_waiters(lock);
1368          *      raw_spin_unlock(&lock->wait_lock);
1369          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1370          *              return;
1371          *      goto retry;
1372          *
1373          * The fastpath disabled variant is simple as all access to
1374          * lock->owner is serialized by lock->wait_lock:
1375          *
1376          *      lock->owner = NULL;
1377          *      raw_spin_unlock(&lock->wait_lock);
1378          */
1379         while (!rt_mutex_has_waiters(lock)) {
1380                 /* Drops lock->wait_lock ! */
1381                 if (unlock_rt_mutex_safe(lock, flags) == true)
1382                         return false;
1383                 /* Relock the rtmutex and try again */
1384                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1385         }
1386
1387         /*
1388          * The wakeup next waiter path does not suffer from the above
1389          * race. See the comments there.
1390          *
1391          * Queue the next waiter for wakeup once we release the wait_lock.
1392          */
1393         mark_wakeup_next_waiter(wake_q, lock);
1394         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1395
1396         return true; /* call rt_mutex_postunlock() */
1397 }
1398
1399 /*
1400  * debug aware fast / slowpath lock,trylock,unlock
1401  *
1402  * The atomic acquire/release ops are compiled away, when either the
1403  * architecture does not support cmpxchg or when debugging is enabled.
1404  */
1405 static inline int
1406 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1407                   int (*slowfn)(struct rt_mutex *lock, int state,
1408                                 struct hrtimer_sleeper *timeout,
1409                                 enum rtmutex_chainwalk chwalk))
1410 {
1411         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1412                 return 0;
1413
1414         return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1415 }
1416
1417 static inline int
1418 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1419                         struct hrtimer_sleeper *timeout,
1420                         enum rtmutex_chainwalk chwalk,
1421                         int (*slowfn)(struct rt_mutex *lock, int state,
1422                                       struct hrtimer_sleeper *timeout,
1423                                       enum rtmutex_chainwalk chwalk))
1424 {
1425         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1426             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1427                 return 0;
1428
1429         return slowfn(lock, state, timeout, chwalk);
1430 }
1431
1432 static inline int
1433 rt_mutex_fasttrylock(struct rt_mutex *lock,
1434                      int (*slowfn)(struct rt_mutex *lock))
1435 {
1436         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1437                 return 1;
1438
1439         return slowfn(lock);
1440 }
1441
1442 /*
1443  * Performs the wakeup of the the top-waiter and re-enables preemption.
1444  */
1445 void rt_mutex_postunlock(struct wake_q_head *wake_q)
1446 {
1447         wake_up_q(wake_q);
1448
1449         /* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1450         preempt_enable();
1451 }
1452
1453 static inline void
1454 rt_mutex_fastunlock(struct rt_mutex *lock,
1455                     bool (*slowfn)(struct rt_mutex *lock,
1456                                    struct wake_q_head *wqh))
1457 {
1458         DEFINE_WAKE_Q(wake_q);
1459
1460         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1461                 return;
1462
1463         if (slowfn(lock, &wake_q))
1464                 rt_mutex_postunlock(&wake_q);
1465 }
1466
1467 static inline void __rt_mutex_lock(struct rt_mutex *lock, unsigned int subclass)
1468 {
1469         might_sleep();
1470
1471         mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
1472         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1473 }
1474
1475 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1476 /**
1477  * rt_mutex_lock_nested - lock a rt_mutex
1478  *
1479  * @lock: the rt_mutex to be locked
1480  * @subclass: the lockdep subclass
1481  */
1482 void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
1483 {
1484         __rt_mutex_lock(lock, subclass);
1485 }
1486 EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
1487
1488 #else /* !CONFIG_DEBUG_LOCK_ALLOC */
1489
1490 /**
1491  * rt_mutex_lock - lock a rt_mutex
1492  *
1493  * @lock: the rt_mutex to be locked
1494  */
1495 void __sched rt_mutex_lock(struct rt_mutex *lock)
1496 {
1497         __rt_mutex_lock(lock, 0);
1498 }
1499 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1500 #endif
1501
1502 /**
1503  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1504  *
1505  * @lock:               the rt_mutex to be locked
1506  *
1507  * Returns:
1508  *  0           on success
1509  * -EINTR       when interrupted by a signal
1510  */
1511 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1512 {
1513         int ret;
1514
1515         might_sleep();
1516
1517         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1518         ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1519         if (ret)
1520                 mutex_release(&lock->dep_map, _RET_IP_);
1521
1522         return ret;
1523 }
1524 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1525
1526 /*
1527  * Futex variant, must not use fastpath.
1528  */
1529 int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1530 {
1531         return rt_mutex_slowtrylock(lock);
1532 }
1533
1534 int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1535 {
1536         return __rt_mutex_slowtrylock(lock);
1537 }
1538
1539 /**
1540  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1541  *                      the timeout structure is provided
1542  *                      by the caller
1543  *
1544  * @lock:               the rt_mutex to be locked
1545  * @timeout:            timeout structure or NULL (no timeout)
1546  *
1547  * Returns:
1548  *  0           on success
1549  * -EINTR       when interrupted by a signal
1550  * -ETIMEDOUT   when the timeout expired
1551  */
1552 int
1553 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1554 {
1555         int ret;
1556
1557         might_sleep();
1558
1559         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1560         ret = rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1561                                        RT_MUTEX_MIN_CHAINWALK,
1562                                        rt_mutex_slowlock);
1563         if (ret)
1564                 mutex_release(&lock->dep_map, _RET_IP_);
1565
1566         return ret;
1567 }
1568 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1569
1570 /**
1571  * rt_mutex_trylock - try to lock a rt_mutex
1572  *
1573  * @lock:       the rt_mutex to be locked
1574  *
1575  * This function can only be called in thread context. It's safe to
1576  * call it from atomic regions, but not from hard interrupt or soft
1577  * interrupt context.
1578  *
1579  * Returns 1 on success and 0 on contention
1580  */
1581 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1582 {
1583         int ret;
1584
1585         if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1586                 return 0;
1587
1588         ret = rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1589         if (ret)
1590                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1591
1592         return ret;
1593 }
1594 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1595
1596 /**
1597  * rt_mutex_unlock - unlock a rt_mutex
1598  *
1599  * @lock: the rt_mutex to be unlocked
1600  */
1601 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1602 {
1603         mutex_release(&lock->dep_map, _RET_IP_);
1604         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1605 }
1606 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1607
1608 /**
1609  * Futex variant, that since futex variants do not use the fast-path, can be
1610  * simple and will not need to retry.
1611  */
1612 bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1613                                     struct wake_q_head *wake_q)
1614 {
1615         lockdep_assert_held(&lock->wait_lock);
1616
1617         debug_rt_mutex_unlock(lock);
1618
1619         if (!rt_mutex_has_waiters(lock)) {
1620                 lock->owner = NULL;
1621                 return false; /* done */
1622         }
1623
1624         /*
1625          * We've already deboosted, mark_wakeup_next_waiter() will
1626          * retain preempt_disabled when we drop the wait_lock, to
1627          * avoid inversion prior to the wakeup.  preempt_disable()
1628          * therein pairs with rt_mutex_postunlock().
1629          */
1630         mark_wakeup_next_waiter(wake_q, lock);
1631
1632         return true; /* call postunlock() */
1633 }
1634
1635 void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1636 {
1637         DEFINE_WAKE_Q(wake_q);
1638         unsigned long flags;
1639         bool postunlock;
1640
1641         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1642         postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1643         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1644
1645         if (postunlock)
1646                 rt_mutex_postunlock(&wake_q);
1647 }
1648
1649 /**
1650  * rt_mutex_destroy - mark a mutex unusable
1651  * @lock: the mutex to be destroyed
1652  *
1653  * This function marks the mutex uninitialized, and any subsequent
1654  * use of the mutex is forbidden. The mutex must not be locked when
1655  * this function is called.
1656  */
1657 void rt_mutex_destroy(struct rt_mutex *lock)
1658 {
1659         WARN_ON(rt_mutex_is_locked(lock));
1660 #ifdef CONFIG_DEBUG_RT_MUTEXES
1661         lock->magic = NULL;
1662 #endif
1663 }
1664 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1665
1666 /**
1667  * __rt_mutex_init - initialize the rt lock
1668  *
1669  * @lock: the rt lock to be initialized
1670  *
1671  * Initialize the rt lock to unlocked state.
1672  *
1673  * Initializing of a locked rt lock is not allowed
1674  */
1675 void __rt_mutex_init(struct rt_mutex *lock, const char *name,
1676                      struct lock_class_key *key)
1677 {
1678         lock->owner = NULL;
1679         raw_spin_lock_init(&lock->wait_lock);
1680         lock->waiters = RB_ROOT_CACHED;
1681
1682         if (name && key)
1683                 debug_rt_mutex_init(lock, name, key);
1684 }
1685 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1686
1687 /**
1688  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1689  *                              proxy owner
1690  *
1691  * @lock:       the rt_mutex to be locked
1692  * @proxy_owner:the task to set as owner
1693  *
1694  * No locking. Caller has to do serializing itself
1695  *
1696  * Special API call for PI-futex support. This initializes the rtmutex and
1697  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1698  * possible at this point because the pi_state which contains the rtmutex
1699  * is not yet visible to other tasks.
1700  */
1701 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1702                                 struct task_struct *proxy_owner)
1703 {
1704         __rt_mutex_init(lock, NULL, NULL);
1705         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1706         rt_mutex_set_owner(lock, proxy_owner);
1707 }
1708
1709 /**
1710  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1711  *
1712  * @lock:       the rt_mutex to be locked
1713  *
1714  * No locking. Caller has to do serializing itself
1715  *
1716  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1717  * (debugging) state. Concurrent operations on this rt_mutex are not
1718  * possible because it belongs to the pi_state which is about to be freed
1719  * and it is not longer visible to other tasks.
1720  */
1721 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1722                            struct task_struct *proxy_owner)
1723 {
1724         debug_rt_mutex_proxy_unlock(lock);
1725         rt_mutex_set_owner(lock, NULL);
1726 }
1727
1728 /**
1729  * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1730  * @lock:               the rt_mutex to take
1731  * @waiter:             the pre-initialized rt_mutex_waiter
1732  * @task:               the task to prepare
1733  *
1734  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1735  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1736  *
1737  * NOTE: does _NOT_ remove the @waiter on failure; must either call
1738  * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
1739  *
1740  * Returns:
1741  *  0 - task blocked on lock
1742  *  1 - acquired the lock for task, caller should wake it up
1743  * <0 - error
1744  *
1745  * Special API call for PI-futex support.
1746  */
1747 int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1748                               struct rt_mutex_waiter *waiter,
1749                               struct task_struct *task)
1750 {
1751         int ret;
1752
1753         lockdep_assert_held(&lock->wait_lock);
1754
1755         if (try_to_take_rt_mutex(lock, task, NULL))
1756                 return 1;
1757
1758         /* We enforce deadlock detection for futexes */
1759         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1760                                       RT_MUTEX_FULL_CHAINWALK);
1761
1762         if (ret && !rt_mutex_owner(lock)) {
1763                 /*
1764                  * Reset the return value. We might have
1765                  * returned with -EDEADLK and the owner
1766                  * released the lock while we were walking the
1767                  * pi chain.  Let the waiter sort it out.
1768                  */
1769                 ret = 0;
1770         }
1771
1772         debug_rt_mutex_print_deadlock(waiter);
1773
1774         return ret;
1775 }
1776
1777 /**
1778  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1779  * @lock:               the rt_mutex to take
1780  * @waiter:             the pre-initialized rt_mutex_waiter
1781  * @task:               the task to prepare
1782  *
1783  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1784  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1785  *
1786  * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
1787  * on failure.
1788  *
1789  * Returns:
1790  *  0 - task blocked on lock
1791  *  1 - acquired the lock for task, caller should wake it up
1792  * <0 - error
1793  *
1794  * Special API call for PI-futex support.
1795  */
1796 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1797                               struct rt_mutex_waiter *waiter,
1798                               struct task_struct *task)
1799 {
1800         int ret;
1801
1802         raw_spin_lock_irq(&lock->wait_lock);
1803         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1804         if (unlikely(ret))
1805                 remove_waiter(lock, waiter);
1806         raw_spin_unlock_irq(&lock->wait_lock);
1807
1808         return ret;
1809 }
1810
1811 /**
1812  * rt_mutex_next_owner - return the next owner of the lock
1813  *
1814  * @lock: the rt lock query
1815  *
1816  * Returns the next owner of the lock or NULL
1817  *
1818  * Caller has to serialize against other accessors to the lock
1819  * itself.
1820  *
1821  * Special API call for PI-futex support
1822  */
1823 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1824 {
1825         if (!rt_mutex_has_waiters(lock))
1826                 return NULL;
1827
1828         return rt_mutex_top_waiter(lock)->task;
1829 }
1830
1831 /**
1832  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1833  * @lock:               the rt_mutex we were woken on
1834  * @to:                 the timeout, null if none. hrtimer should already have
1835  *                      been started.
1836  * @waiter:             the pre-initialized rt_mutex_waiter
1837  *
1838  * Wait for the the lock acquisition started on our behalf by
1839  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1840  * rt_mutex_cleanup_proxy_lock().
1841  *
1842  * Returns:
1843  *  0 - success
1844  * <0 - error, one of -EINTR, -ETIMEDOUT
1845  *
1846  * Special API call for PI-futex support
1847  */
1848 int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1849                                struct hrtimer_sleeper *to,
1850                                struct rt_mutex_waiter *waiter)
1851 {
1852         int ret;
1853
1854         raw_spin_lock_irq(&lock->wait_lock);
1855         /* sleep on the mutex */
1856         set_current_state(TASK_INTERRUPTIBLE);
1857         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1858         /*
1859          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1860          * have to fix that up.
1861          */
1862         fixup_rt_mutex_waiters(lock);
1863         raw_spin_unlock_irq(&lock->wait_lock);
1864
1865         return ret;
1866 }
1867
1868 /**
1869  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1870  * @lock:               the rt_mutex we were woken on
1871  * @waiter:             the pre-initialized rt_mutex_waiter
1872  *
1873  * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
1874  * rt_mutex_wait_proxy_lock().
1875  *
1876  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1877  * in fact still be granted ownership until we're removed. Therefore we can
1878  * find we are in fact the owner and must disregard the
1879  * rt_mutex_wait_proxy_lock() failure.
1880  *
1881  * Returns:
1882  *  true  - did the cleanup, we done.
1883  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1884  *          caller should disregards its return value.
1885  *
1886  * Special API call for PI-futex support
1887  */
1888 bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1889                                  struct rt_mutex_waiter *waiter)
1890 {
1891         bool cleanup = false;
1892
1893         raw_spin_lock_irq(&lock->wait_lock);
1894         /*
1895          * Do an unconditional try-lock, this deals with the lock stealing
1896          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1897          * sets a NULL owner.
1898          *
1899          * We're not interested in the return value, because the subsequent
1900          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1901          * we will own the lock and it will have removed the waiter. If we
1902          * failed the trylock, we're still not owner and we need to remove
1903          * ourselves.
1904          */
1905         try_to_take_rt_mutex(lock, current, waiter);
1906         /*
1907          * Unless we're the owner; we're still enqueued on the wait_list.
1908          * So check if we became owner, if not, take us off the wait_list.
1909          */
1910         if (rt_mutex_owner(lock) != current) {
1911                 remove_waiter(lock, waiter);
1912                 cleanup = true;
1913         }
1914         /*
1915          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1916          * have to fix that up.
1917          */
1918         fixup_rt_mutex_waiters(lock);
1919
1920         raw_spin_unlock_irq(&lock->wait_lock);
1921
1922         return cleanup;
1923 }