Merge tag 'for-5.14-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux
[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 __always_inline 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 __always_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 __always_inline 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 __always_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 __always_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 __always_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 __always_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 __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
233                                                 struct rt_mutex_waiter *right)
234 {
235         if (left->prio < right->prio)
236                 return 1;
237
238         /*
239          * If both waiters have dl_prio(), we check the deadlines of the
240          * associated tasks.
241          * If left waiter has a dl_prio(), and we didn't return 1 above,
242          * then right waiter has a dl_prio() too.
243          */
244         if (dl_prio(left->prio))
245                 return dl_time_before(left->deadline, right->deadline);
246
247         return 0;
248 }
249
250 static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
251                                                  struct rt_mutex_waiter *right)
252 {
253         if (left->prio != right->prio)
254                 return 0;
255
256         /*
257          * If both waiters have dl_prio(), we check the deadlines of the
258          * associated tasks.
259          * If left waiter has a dl_prio(), and we didn't return 0 above,
260          * then right waiter has a dl_prio() too.
261          */
262         if (dl_prio(left->prio))
263                 return left->deadline == right->deadline;
264
265         return 1;
266 }
267
268 #define __node_2_waiter(node) \
269         rb_entry((node), struct rt_mutex_waiter, tree_entry)
270
271 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
272 {
273         return rt_mutex_waiter_less(__node_2_waiter(a), __node_2_waiter(b));
274 }
275
276 static __always_inline void
277 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
278 {
279         rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
280 }
281
282 static __always_inline void
283 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
284 {
285         if (RB_EMPTY_NODE(&waiter->tree_entry))
286                 return;
287
288         rb_erase_cached(&waiter->tree_entry, &lock->waiters);
289         RB_CLEAR_NODE(&waiter->tree_entry);
290 }
291
292 #define __node_2_pi_waiter(node) \
293         rb_entry((node), struct rt_mutex_waiter, pi_tree_entry)
294
295 static __always_inline bool
296 __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
297 {
298         return rt_mutex_waiter_less(__node_2_pi_waiter(a), __node_2_pi_waiter(b));
299 }
300
301 static __always_inline void
302 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
303 {
304         rb_add_cached(&waiter->pi_tree_entry, &task->pi_waiters, __pi_waiter_less);
305 }
306
307 static __always_inline void
308 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
309 {
310         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
311                 return;
312
313         rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
314         RB_CLEAR_NODE(&waiter->pi_tree_entry);
315 }
316
317 static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
318 {
319         struct task_struct *pi_task = NULL;
320
321         lockdep_assert_held(&p->pi_lock);
322
323         if (task_has_pi_waiters(p))
324                 pi_task = task_top_pi_waiter(p)->task;
325
326         rt_mutex_setprio(p, pi_task);
327 }
328
329 /*
330  * Deadlock detection is conditional:
331  *
332  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
333  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
334  *
335  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
336  * conducted independent of the detect argument.
337  *
338  * If the waiter argument is NULL this indicates the deboost path and
339  * deadlock detection is disabled independent of the detect argument
340  * and the config settings.
341  */
342 static __always_inline bool
343 rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
344                               enum rtmutex_chainwalk chwalk)
345 {
346         if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEX))
347                 return waiter != NULL;
348         return chwalk == RT_MUTEX_FULL_CHAINWALK;
349 }
350
351 /*
352  * Max number of times we'll walk the boosting chain:
353  */
354 int max_lock_depth = 1024;
355
356 static __always_inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
357 {
358         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
359 }
360
361 /*
362  * Adjust the priority chain. Also used for deadlock detection.
363  * Decreases task's usage by one - may thus free the task.
364  *
365  * @task:       the task owning the mutex (owner) for which a chain walk is
366  *              probably needed
367  * @chwalk:     do we have to carry out deadlock detection?
368  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
369  *              things for a task that has just got its priority adjusted, and
370  *              is waiting on a mutex)
371  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
372  *              we dropped its pi_lock. Is never dereferenced, only used for
373  *              comparison to detect lock chain changes.
374  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
375  *              its priority to the mutex owner (can be NULL in the case
376  *              depicted above or if the top waiter is gone away and we are
377  *              actually deboosting the owner)
378  * @top_task:   the current top waiter
379  *
380  * Returns 0 or -EDEADLK.
381  *
382  * Chain walk basics and protection scope
383  *
384  * [R] refcount on task
385  * [P] task->pi_lock held
386  * [L] rtmutex->wait_lock held
387  *
388  * Step Description                             Protected by
389  *      function arguments:
390  *      @task                                   [R]
391  *      @orig_lock if != NULL                   @top_task is blocked on it
392  *      @next_lock                              Unprotected. Cannot be
393  *                                              dereferenced. Only used for
394  *                                              comparison.
395  *      @orig_waiter if != NULL                 @top_task is blocked on it
396  *      @top_task                               current, or in case of proxy
397  *                                              locking protected by calling
398  *                                              code
399  *      again:
400  *        loop_sanity_check();
401  *      retry:
402  * [1]    lock(task->pi_lock);                  [R] acquire [P]
403  * [2]    waiter = task->pi_blocked_on;         [P]
404  * [3]    check_exit_conditions_1();            [P]
405  * [4]    lock = waiter->lock;                  [P]
406  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
407  *          unlock(task->pi_lock);              release [P]
408  *          goto retry;
409  *        }
410  * [6]    check_exit_conditions_2();            [P] + [L]
411  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
412  * [8]    unlock(task->pi_lock);                release [P]
413  *        put_task_struct(task);                release [R]
414  * [9]    check_exit_conditions_3();            [L]
415  * [10]   task = owner(lock);                   [L]
416  *        get_task_struct(task);                [L] acquire [R]
417  *        lock(task->pi_lock);                  [L] acquire [P]
418  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
419  * [12]   check_exit_conditions_4();            [P] + [L]
420  * [13]   unlock(task->pi_lock);                release [P]
421  *        unlock(lock->wait_lock);              release [L]
422  *        goto again;
423  */
424 static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
425                                               enum rtmutex_chainwalk chwalk,
426                                               struct rt_mutex *orig_lock,
427                                               struct rt_mutex *next_lock,
428                                               struct rt_mutex_waiter *orig_waiter,
429                                               struct task_struct *top_task)
430 {
431         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
432         struct rt_mutex_waiter *prerequeue_top_waiter;
433         int ret = 0, depth = 0;
434         struct rt_mutex *lock;
435         bool detect_deadlock;
436         bool requeue = true;
437
438         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
439
440         /*
441          * The (de)boosting is a step by step approach with a lot of
442          * pitfalls. We want this to be preemptible and we want hold a
443          * maximum of two locks per step. So we have to check
444          * carefully whether things change under us.
445          */
446  again:
447         /*
448          * We limit the lock chain length for each invocation.
449          */
450         if (++depth > max_lock_depth) {
451                 static int prev_max;
452
453                 /*
454                  * Print this only once. If the admin changes the limit,
455                  * print a new message when reaching the limit again.
456                  */
457                 if (prev_max != max_lock_depth) {
458                         prev_max = max_lock_depth;
459                         printk(KERN_WARNING "Maximum lock depth %d reached "
460                                "task: %s (%d)\n", max_lock_depth,
461                                top_task->comm, task_pid_nr(top_task));
462                 }
463                 put_task_struct(task);
464
465                 return -EDEADLK;
466         }
467
468         /*
469          * We are fully preemptible here and only hold the refcount on
470          * @task. So everything can have changed under us since the
471          * caller or our own code below (goto retry/again) dropped all
472          * locks.
473          */
474  retry:
475         /*
476          * [1] Task cannot go away as we did a get_task() before !
477          */
478         raw_spin_lock_irq(&task->pi_lock);
479
480         /*
481          * [2] Get the waiter on which @task is blocked on.
482          */
483         waiter = task->pi_blocked_on;
484
485         /*
486          * [3] check_exit_conditions_1() protected by task->pi_lock.
487          */
488
489         /*
490          * Check whether the end of the boosting chain has been
491          * reached or the state of the chain has changed while we
492          * dropped the locks.
493          */
494         if (!waiter)
495                 goto out_unlock_pi;
496
497         /*
498          * Check the orig_waiter state. After we dropped the locks,
499          * the previous owner of the lock might have released the lock.
500          */
501         if (orig_waiter && !rt_mutex_owner(orig_lock))
502                 goto out_unlock_pi;
503
504         /*
505          * We dropped all locks after taking a refcount on @task, so
506          * the task might have moved on in the lock chain or even left
507          * the chain completely and blocks now on an unrelated lock or
508          * on @orig_lock.
509          *
510          * We stored the lock on which @task was blocked in @next_lock,
511          * so we can detect the chain change.
512          */
513         if (next_lock != waiter->lock)
514                 goto out_unlock_pi;
515
516         /*
517          * Drop out, when the task has no waiters. Note,
518          * top_waiter can be NULL, when we are in the deboosting
519          * mode!
520          */
521         if (top_waiter) {
522                 if (!task_has_pi_waiters(task))
523                         goto out_unlock_pi;
524                 /*
525                  * If deadlock detection is off, we stop here if we
526                  * are not the top pi waiter of the task. If deadlock
527                  * detection is enabled we continue, but stop the
528                  * requeueing in the chain walk.
529                  */
530                 if (top_waiter != task_top_pi_waiter(task)) {
531                         if (!detect_deadlock)
532                                 goto out_unlock_pi;
533                         else
534                                 requeue = false;
535                 }
536         }
537
538         /*
539          * If the waiter priority is the same as the task priority
540          * then there is no further priority adjustment necessary.  If
541          * deadlock detection is off, we stop the chain walk. If its
542          * enabled we continue, but stop the requeueing in the chain
543          * walk.
544          */
545         if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
546                 if (!detect_deadlock)
547                         goto out_unlock_pi;
548                 else
549                         requeue = false;
550         }
551
552         /*
553          * [4] Get the next lock
554          */
555         lock = waiter->lock;
556         /*
557          * [5] We need to trylock here as we are holding task->pi_lock,
558          * which is the reverse lock order versus the other rtmutex
559          * operations.
560          */
561         if (!raw_spin_trylock(&lock->wait_lock)) {
562                 raw_spin_unlock_irq(&task->pi_lock);
563                 cpu_relax();
564                 goto retry;
565         }
566
567         /*
568          * [6] check_exit_conditions_2() protected by task->pi_lock and
569          * lock->wait_lock.
570          *
571          * Deadlock detection. If the lock is the same as the original
572          * lock which caused us to walk the lock chain or if the
573          * current lock is owned by the task which initiated the chain
574          * walk, we detected a deadlock.
575          */
576         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
577                 raw_spin_unlock(&lock->wait_lock);
578                 ret = -EDEADLK;
579                 goto out_unlock_pi;
580         }
581
582         /*
583          * If we just follow the lock chain for deadlock detection, no
584          * need to do all the requeue operations. To avoid a truckload
585          * of conditionals around the various places below, just do the
586          * minimum chain walk checks.
587          */
588         if (!requeue) {
589                 /*
590                  * No requeue[7] here. Just release @task [8]
591                  */
592                 raw_spin_unlock(&task->pi_lock);
593                 put_task_struct(task);
594
595                 /*
596                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
597                  * If there is no owner of the lock, end of chain.
598                  */
599                 if (!rt_mutex_owner(lock)) {
600                         raw_spin_unlock_irq(&lock->wait_lock);
601                         return 0;
602                 }
603
604                 /* [10] Grab the next task, i.e. owner of @lock */
605                 task = get_task_struct(rt_mutex_owner(lock));
606                 raw_spin_lock(&task->pi_lock);
607
608                 /*
609                  * No requeue [11] here. We just do deadlock detection.
610                  *
611                  * [12] Store whether owner is blocked
612                  * itself. Decision is made after dropping the locks
613                  */
614                 next_lock = task_blocked_on_lock(task);
615                 /*
616                  * Get the top waiter for the next iteration
617                  */
618                 top_waiter = rt_mutex_top_waiter(lock);
619
620                 /* [13] Drop locks */
621                 raw_spin_unlock(&task->pi_lock);
622                 raw_spin_unlock_irq(&lock->wait_lock);
623
624                 /* If owner is not blocked, end of chain. */
625                 if (!next_lock)
626                         goto out_put_task;
627                 goto again;
628         }
629
630         /*
631          * Store the current top waiter before doing the requeue
632          * operation on @lock. We need it for the boost/deboost
633          * decision below.
634          */
635         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
636
637         /* [7] Requeue the waiter in the lock waiter tree. */
638         rt_mutex_dequeue(lock, waiter);
639
640         /*
641          * Update the waiter prio fields now that we're dequeued.
642          *
643          * These values can have changed through either:
644          *
645          *   sys_sched_set_scheduler() / sys_sched_setattr()
646          *
647          * or
648          *
649          *   DL CBS enforcement advancing the effective deadline.
650          *
651          * Even though pi_waiters also uses these fields, and that tree is only
652          * updated in [11], we can do this here, since we hold [L], which
653          * serializes all pi_waiters access and rb_erase() does not care about
654          * the values of the node being removed.
655          */
656         waiter->prio = task->prio;
657         waiter->deadline = task->dl.deadline;
658
659         rt_mutex_enqueue(lock, waiter);
660
661         /* [8] Release the task */
662         raw_spin_unlock(&task->pi_lock);
663         put_task_struct(task);
664
665         /*
666          * [9] check_exit_conditions_3 protected by lock->wait_lock.
667          *
668          * We must abort the chain walk if there is no lock owner even
669          * in the dead lock detection case, as we have nothing to
670          * follow here. This is the end of the chain we are walking.
671          */
672         if (!rt_mutex_owner(lock)) {
673                 /*
674                  * If the requeue [7] above changed the top waiter,
675                  * then we need to wake the new top waiter up to try
676                  * to get the lock.
677                  */
678                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
679                         wake_up_process(rt_mutex_top_waiter(lock)->task);
680                 raw_spin_unlock_irq(&lock->wait_lock);
681                 return 0;
682         }
683
684         /* [10] Grab the next task, i.e. the owner of @lock */
685         task = get_task_struct(rt_mutex_owner(lock));
686         raw_spin_lock(&task->pi_lock);
687
688         /* [11] requeue the pi waiters if necessary */
689         if (waiter == rt_mutex_top_waiter(lock)) {
690                 /*
691                  * The waiter became the new top (highest priority)
692                  * waiter on the lock. Replace the previous top waiter
693                  * in the owner tasks pi waiters tree with this waiter
694                  * and adjust the priority of the owner.
695                  */
696                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
697                 rt_mutex_enqueue_pi(task, waiter);
698                 rt_mutex_adjust_prio(task);
699
700         } else if (prerequeue_top_waiter == waiter) {
701                 /*
702                  * The waiter was the top waiter on the lock, but is
703                  * no longer the top priority waiter. Replace waiter in
704                  * the owner tasks pi waiters tree with the new top
705                  * (highest priority) waiter and adjust the priority
706                  * of the owner.
707                  * The new top waiter is stored in @waiter so that
708                  * @waiter == @top_waiter evaluates to true below and
709                  * we continue to deboost the rest of the chain.
710                  */
711                 rt_mutex_dequeue_pi(task, waiter);
712                 waiter = rt_mutex_top_waiter(lock);
713                 rt_mutex_enqueue_pi(task, waiter);
714                 rt_mutex_adjust_prio(task);
715         } else {
716                 /*
717                  * Nothing changed. No need to do any priority
718                  * adjustment.
719                  */
720         }
721
722         /*
723          * [12] check_exit_conditions_4() protected by task->pi_lock
724          * and lock->wait_lock. The actual decisions are made after we
725          * dropped the locks.
726          *
727          * Check whether the task which owns the current lock is pi
728          * blocked itself. If yes we store a pointer to the lock for
729          * the lock chain change detection above. After we dropped
730          * task->pi_lock next_lock cannot be dereferenced anymore.
731          */
732         next_lock = task_blocked_on_lock(task);
733         /*
734          * Store the top waiter of @lock for the end of chain walk
735          * decision below.
736          */
737         top_waiter = rt_mutex_top_waiter(lock);
738
739         /* [13] Drop the locks */
740         raw_spin_unlock(&task->pi_lock);
741         raw_spin_unlock_irq(&lock->wait_lock);
742
743         /*
744          * Make the actual exit decisions [12], based on the stored
745          * values.
746          *
747          * We reached the end of the lock chain. Stop right here. No
748          * point to go back just to figure that out.
749          */
750         if (!next_lock)
751                 goto out_put_task;
752
753         /*
754          * If the current waiter is not the top waiter on the lock,
755          * then we can stop the chain walk here if we are not in full
756          * deadlock detection mode.
757          */
758         if (!detect_deadlock && waiter != top_waiter)
759                 goto out_put_task;
760
761         goto again;
762
763  out_unlock_pi:
764         raw_spin_unlock_irq(&task->pi_lock);
765  out_put_task:
766         put_task_struct(task);
767
768         return ret;
769 }
770
771 /*
772  * Try to take an rt-mutex
773  *
774  * Must be called with lock->wait_lock held and interrupts disabled
775  *
776  * @lock:   The lock to be acquired.
777  * @task:   The task which wants to acquire the lock
778  * @waiter: The waiter that is queued to the lock's wait tree if the
779  *          callsite called task_blocked_on_lock(), otherwise NULL
780  */
781 static int __sched
782 try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
783                      struct rt_mutex_waiter *waiter)
784 {
785         lockdep_assert_held(&lock->wait_lock);
786
787         /*
788          * Before testing whether we can acquire @lock, we set the
789          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
790          * other tasks which try to modify @lock into the slow path
791          * and they serialize on @lock->wait_lock.
792          *
793          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
794          * as explained at the top of this file if and only if:
795          *
796          * - There is a lock owner. The caller must fixup the
797          *   transient state if it does a trylock or leaves the lock
798          *   function due to a signal or timeout.
799          *
800          * - @task acquires the lock and there are no other
801          *   waiters. This is undone in rt_mutex_set_owner(@task) at
802          *   the end of this function.
803          */
804         mark_rt_mutex_waiters(lock);
805
806         /*
807          * If @lock has an owner, give up.
808          */
809         if (rt_mutex_owner(lock))
810                 return 0;
811
812         /*
813          * If @waiter != NULL, @task has already enqueued the waiter
814          * into @lock waiter tree. If @waiter == NULL then this is a
815          * trylock attempt.
816          */
817         if (waiter) {
818                 /*
819                  * If waiter is not the highest priority waiter of
820                  * @lock, give up.
821                  */
822                 if (waiter != rt_mutex_top_waiter(lock))
823                         return 0;
824
825                 /*
826                  * We can acquire the lock. Remove the waiter from the
827                  * lock waiters tree.
828                  */
829                 rt_mutex_dequeue(lock, waiter);
830
831         } else {
832                 /*
833                  * If the lock has waiters already we check whether @task is
834                  * eligible to take over the lock.
835                  *
836                  * If there are no other waiters, @task can acquire
837                  * the lock.  @task->pi_blocked_on is NULL, so it does
838                  * not need to be dequeued.
839                  */
840                 if (rt_mutex_has_waiters(lock)) {
841                         /*
842                          * If @task->prio is greater than or equal to
843                          * the top waiter priority (kernel view),
844                          * @task lost.
845                          */
846                         if (!rt_mutex_waiter_less(task_to_waiter(task),
847                                                   rt_mutex_top_waiter(lock)))
848                                 return 0;
849
850                         /*
851                          * The current top waiter stays enqueued. We
852                          * don't have to change anything in the lock
853                          * waiters order.
854                          */
855                 } else {
856                         /*
857                          * No waiters. Take the lock without the
858                          * pi_lock dance.@task->pi_blocked_on is NULL
859                          * and we have no waiters to enqueue in @task
860                          * pi waiters tree.
861                          */
862                         goto takeit;
863                 }
864         }
865
866         /*
867          * Clear @task->pi_blocked_on. Requires protection by
868          * @task->pi_lock. Redundant operation for the @waiter == NULL
869          * case, but conditionals are more expensive than a redundant
870          * store.
871          */
872         raw_spin_lock(&task->pi_lock);
873         task->pi_blocked_on = NULL;
874         /*
875          * Finish the lock acquisition. @task is the new owner. If
876          * other waiters exist we have to insert the highest priority
877          * waiter into @task->pi_waiters tree.
878          */
879         if (rt_mutex_has_waiters(lock))
880                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
881         raw_spin_unlock(&task->pi_lock);
882
883 takeit:
884         /*
885          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
886          * are still waiters or clears it.
887          */
888         rt_mutex_set_owner(lock, task);
889
890         return 1;
891 }
892
893 /*
894  * Task blocks on lock.
895  *
896  * Prepare waiter and propagate pi chain
897  *
898  * This must be called with lock->wait_lock held and interrupts disabled
899  */
900 static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
901                                            struct rt_mutex_waiter *waiter,
902                                            struct task_struct *task,
903                                            enum rtmutex_chainwalk chwalk)
904 {
905         struct task_struct *owner = rt_mutex_owner(lock);
906         struct rt_mutex_waiter *top_waiter = waiter;
907         struct rt_mutex *next_lock;
908         int chain_walk = 0, res;
909
910         lockdep_assert_held(&lock->wait_lock);
911
912         /*
913          * Early deadlock detection. We really don't want the task to
914          * enqueue on itself just to untangle the mess later. It's not
915          * only an optimization. We drop the locks, so another waiter
916          * can come in before the chain walk detects the deadlock. So
917          * the other will detect the deadlock and return -EDEADLOCK,
918          * which is wrong, as the other waiter is not in a deadlock
919          * situation.
920          */
921         if (owner == task)
922                 return -EDEADLK;
923
924         raw_spin_lock(&task->pi_lock);
925         waiter->task = task;
926         waiter->lock = lock;
927         waiter->prio = task->prio;
928         waiter->deadline = task->dl.deadline;
929
930         /* Get the top priority waiter on the lock */
931         if (rt_mutex_has_waiters(lock))
932                 top_waiter = rt_mutex_top_waiter(lock);
933         rt_mutex_enqueue(lock, waiter);
934
935         task->pi_blocked_on = waiter;
936
937         raw_spin_unlock(&task->pi_lock);
938
939         if (!owner)
940                 return 0;
941
942         raw_spin_lock(&owner->pi_lock);
943         if (waiter == rt_mutex_top_waiter(lock)) {
944                 rt_mutex_dequeue_pi(owner, top_waiter);
945                 rt_mutex_enqueue_pi(owner, waiter);
946
947                 rt_mutex_adjust_prio(owner);
948                 if (owner->pi_blocked_on)
949                         chain_walk = 1;
950         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
951                 chain_walk = 1;
952         }
953
954         /* Store the lock on which owner is blocked or NULL */
955         next_lock = task_blocked_on_lock(owner);
956
957         raw_spin_unlock(&owner->pi_lock);
958         /*
959          * Even if full deadlock detection is on, if the owner is not
960          * blocked itself, we can avoid finding this out in the chain
961          * walk.
962          */
963         if (!chain_walk || !next_lock)
964                 return 0;
965
966         /*
967          * The owner can't disappear while holding a lock,
968          * so the owner struct is protected by wait_lock.
969          * Gets dropped in rt_mutex_adjust_prio_chain()!
970          */
971         get_task_struct(owner);
972
973         raw_spin_unlock_irq(&lock->wait_lock);
974
975         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
976                                          next_lock, waiter, task);
977
978         raw_spin_lock_irq(&lock->wait_lock);
979
980         return res;
981 }
982
983 /*
984  * Remove the top waiter from the current tasks pi waiter tree and
985  * queue it up.
986  *
987  * Called with lock->wait_lock held and interrupts disabled.
988  */
989 static void __sched mark_wakeup_next_waiter(struct wake_q_head *wake_q,
990                                             struct rt_mutex *lock)
991 {
992         struct rt_mutex_waiter *waiter;
993
994         raw_spin_lock(&current->pi_lock);
995
996         waiter = rt_mutex_top_waiter(lock);
997
998         /*
999          * Remove it from current->pi_waiters and deboost.
1000          *
1001          * We must in fact deboost here in order to ensure we call
1002          * rt_mutex_setprio() to update p->pi_top_task before the
1003          * task unblocks.
1004          */
1005         rt_mutex_dequeue_pi(current, waiter);
1006         rt_mutex_adjust_prio(current);
1007
1008         /*
1009          * As we are waking up the top waiter, and the waiter stays
1010          * queued on the lock until it gets the lock, this lock
1011          * obviously has waiters. Just set the bit here and this has
1012          * the added benefit of forcing all new tasks into the
1013          * slow path making sure no task of lower priority than
1014          * the top waiter can steal this lock.
1015          */
1016         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1017
1018         /*
1019          * We deboosted before waking the top waiter task such that we don't
1020          * run two tasks with the 'same' priority (and ensure the
1021          * p->pi_top_task pointer points to a blocked task). This however can
1022          * lead to priority inversion if we would get preempted after the
1023          * deboost but before waking our donor task, hence the preempt_disable()
1024          * before unlock.
1025          *
1026          * Pairs with preempt_enable() in rt_mutex_postunlock();
1027          */
1028         preempt_disable();
1029         wake_q_add(wake_q, waiter->task);
1030         raw_spin_unlock(&current->pi_lock);
1031 }
1032
1033 /*
1034  * Remove a waiter from a lock and give up
1035  *
1036  * Must be called with lock->wait_lock held and interrupts disabled. I must
1037  * have just failed to try_to_take_rt_mutex().
1038  */
1039 static void __sched remove_waiter(struct rt_mutex *lock,
1040                                   struct rt_mutex_waiter *waiter)
1041 {
1042         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1043         struct task_struct *owner = rt_mutex_owner(lock);
1044         struct rt_mutex *next_lock;
1045
1046         lockdep_assert_held(&lock->wait_lock);
1047
1048         raw_spin_lock(&current->pi_lock);
1049         rt_mutex_dequeue(lock, waiter);
1050         current->pi_blocked_on = NULL;
1051         raw_spin_unlock(&current->pi_lock);
1052
1053         /*
1054          * Only update priority if the waiter was the highest priority
1055          * waiter of the lock and there is an owner to update.
1056          */
1057         if (!owner || !is_top_waiter)
1058                 return;
1059
1060         raw_spin_lock(&owner->pi_lock);
1061
1062         rt_mutex_dequeue_pi(owner, waiter);
1063
1064         if (rt_mutex_has_waiters(lock))
1065                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1066
1067         rt_mutex_adjust_prio(owner);
1068
1069         /* Store the lock on which owner is blocked or NULL */
1070         next_lock = task_blocked_on_lock(owner);
1071
1072         raw_spin_unlock(&owner->pi_lock);
1073
1074         /*
1075          * Don't walk the chain, if the owner task is not blocked
1076          * itself.
1077          */
1078         if (!next_lock)
1079                 return;
1080
1081         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1082         get_task_struct(owner);
1083
1084         raw_spin_unlock_irq(&lock->wait_lock);
1085
1086         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1087                                    next_lock, NULL, current);
1088
1089         raw_spin_lock_irq(&lock->wait_lock);
1090 }
1091
1092 /*
1093  * Recheck the pi chain, in case we got a priority setting
1094  *
1095  * Called from sched_setscheduler
1096  */
1097 void __sched rt_mutex_adjust_pi(struct task_struct *task)
1098 {
1099         struct rt_mutex_waiter *waiter;
1100         struct rt_mutex *next_lock;
1101         unsigned long flags;
1102
1103         raw_spin_lock_irqsave(&task->pi_lock, flags);
1104
1105         waiter = task->pi_blocked_on;
1106         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1107                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1108                 return;
1109         }
1110         next_lock = waiter->lock;
1111         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1112
1113         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1114         get_task_struct(task);
1115
1116         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1117                                    next_lock, NULL, task);
1118 }
1119
1120 void __sched rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1121 {
1122         debug_rt_mutex_init_waiter(waiter);
1123         RB_CLEAR_NODE(&waiter->pi_tree_entry);
1124         RB_CLEAR_NODE(&waiter->tree_entry);
1125         waiter->task = NULL;
1126 }
1127
1128 /**
1129  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1130  * @lock:                the rt_mutex to take
1131  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1132  *                       or TASK_UNINTERRUPTIBLE)
1133  * @timeout:             the pre-initialized and started timer, or NULL for none
1134  * @waiter:              the pre-initialized rt_mutex_waiter
1135  *
1136  * Must be called with lock->wait_lock held and interrupts disabled
1137  */
1138 static int __sched __rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
1139                                        struct hrtimer_sleeper *timeout,
1140                                        struct rt_mutex_waiter *waiter)
1141 {
1142         int ret = 0;
1143
1144         for (;;) {
1145                 /* Try to acquire the lock: */
1146                 if (try_to_take_rt_mutex(lock, current, waiter))
1147                         break;
1148
1149                 if (timeout && !timeout->task) {
1150                         ret = -ETIMEDOUT;
1151                         break;
1152                 }
1153                 if (signal_pending_state(state, current)) {
1154                         ret = -EINTR;
1155                         break;
1156                 }
1157
1158                 raw_spin_unlock_irq(&lock->wait_lock);
1159
1160                 schedule();
1161
1162                 raw_spin_lock_irq(&lock->wait_lock);
1163                 set_current_state(state);
1164         }
1165
1166         __set_current_state(TASK_RUNNING);
1167         return ret;
1168 }
1169
1170 static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1171                                              struct rt_mutex_waiter *w)
1172 {
1173         /*
1174          * If the result is not -EDEADLOCK or the caller requested
1175          * deadlock detection, nothing to do here.
1176          */
1177         if (res != -EDEADLOCK || detect_deadlock)
1178                 return;
1179
1180         /*
1181          * Yell loudly and stop the task right here.
1182          */
1183         WARN(1, "rtmutex deadlock detected\n");
1184         while (1) {
1185                 set_current_state(TASK_INTERRUPTIBLE);
1186                 schedule();
1187         }
1188 }
1189
1190 /*
1191  * Slow path lock function:
1192  */
1193 static int __sched rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
1194                                      struct hrtimer_sleeper *timeout,
1195                                      enum rtmutex_chainwalk chwalk)
1196 {
1197         struct rt_mutex_waiter waiter;
1198         unsigned long flags;
1199         int ret = 0;
1200
1201         rt_mutex_init_waiter(&waiter);
1202
1203         /*
1204          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1205          * be called in early boot if the cmpxchg() fast path is disabled
1206          * (debug, no architecture support). In this case we will acquire the
1207          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1208          * enable interrupts in that early boot case. So we need to use the
1209          * irqsave/restore variants.
1210          */
1211         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1212
1213         /* Try to acquire the lock again: */
1214         if (try_to_take_rt_mutex(lock, current, NULL)) {
1215                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1216                 return 0;
1217         }
1218
1219         set_current_state(state);
1220
1221         /* Setup the timer, when timeout != NULL */
1222         if (unlikely(timeout))
1223                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1224
1225         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1226
1227         if (likely(!ret))
1228                 /* sleep on the mutex */
1229                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1230
1231         if (unlikely(ret)) {
1232                 __set_current_state(TASK_RUNNING);
1233                 remove_waiter(lock, &waiter);
1234                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1235         }
1236
1237         /*
1238          * try_to_take_rt_mutex() sets the waiter bit
1239          * unconditionally. We might have to fix that up.
1240          */
1241         fixup_rt_mutex_waiters(lock);
1242
1243         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1244
1245         /* Remove pending timer: */
1246         if (unlikely(timeout))
1247                 hrtimer_cancel(&timeout->timer);
1248
1249         debug_rt_mutex_free_waiter(&waiter);
1250
1251         return ret;
1252 }
1253
1254 static int __sched __rt_mutex_slowtrylock(struct rt_mutex *lock)
1255 {
1256         int ret = try_to_take_rt_mutex(lock, current, NULL);
1257
1258         /*
1259          * try_to_take_rt_mutex() sets the lock waiters bit
1260          * unconditionally. Clean this up.
1261          */
1262         fixup_rt_mutex_waiters(lock);
1263
1264         return ret;
1265 }
1266
1267 /*
1268  * Slow path try-lock function:
1269  */
1270 static int __sched rt_mutex_slowtrylock(struct rt_mutex *lock)
1271 {
1272         unsigned long flags;
1273         int ret;
1274
1275         /*
1276          * If the lock already has an owner we fail to get the lock.
1277          * This can be done without taking the @lock->wait_lock as
1278          * it is only being read, and this is a trylock anyway.
1279          */
1280         if (rt_mutex_owner(lock))
1281                 return 0;
1282
1283         /*
1284          * The mutex has currently no owner. Lock the wait lock and try to
1285          * acquire the lock. We use irqsave here to support early boot calls.
1286          */
1287         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1288
1289         ret = __rt_mutex_slowtrylock(lock);
1290
1291         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1292
1293         return ret;
1294 }
1295
1296 /*
1297  * Performs the wakeup of the top-waiter and re-enables preemption.
1298  */
1299 void __sched rt_mutex_postunlock(struct wake_q_head *wake_q)
1300 {
1301         wake_up_q(wake_q);
1302
1303         /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
1304         preempt_enable();
1305 }
1306
1307 /*
1308  * Slow path to release a rt-mutex.
1309  *
1310  * Return whether the current task needs to call rt_mutex_postunlock().
1311  */
1312 static void __sched rt_mutex_slowunlock(struct rt_mutex *lock)
1313 {
1314         DEFINE_WAKE_Q(wake_q);
1315         unsigned long flags;
1316
1317         /* irqsave required to support early boot calls */
1318         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1319
1320         debug_rt_mutex_unlock(lock);
1321
1322         /*
1323          * We must be careful here if the fast path is enabled. If we
1324          * have no waiters queued we cannot set owner to NULL here
1325          * because of:
1326          *
1327          * foo->lock->owner = NULL;
1328          *                      rtmutex_lock(foo->lock);   <- fast path
1329          *                      free = atomic_dec_and_test(foo->refcnt);
1330          *                      rtmutex_unlock(foo->lock); <- fast path
1331          *                      if (free)
1332          *                              kfree(foo);
1333          * raw_spin_unlock(foo->lock->wait_lock);
1334          *
1335          * So for the fastpath enabled kernel:
1336          *
1337          * Nothing can set the waiters bit as long as we hold
1338          * lock->wait_lock. So we do the following sequence:
1339          *
1340          *      owner = rt_mutex_owner(lock);
1341          *      clear_rt_mutex_waiters(lock);
1342          *      raw_spin_unlock(&lock->wait_lock);
1343          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1344          *              return;
1345          *      goto retry;
1346          *
1347          * The fastpath disabled variant is simple as all access to
1348          * lock->owner is serialized by lock->wait_lock:
1349          *
1350          *      lock->owner = NULL;
1351          *      raw_spin_unlock(&lock->wait_lock);
1352          */
1353         while (!rt_mutex_has_waiters(lock)) {
1354                 /* Drops lock->wait_lock ! */
1355                 if (unlock_rt_mutex_safe(lock, flags) == true)
1356                         return;
1357                 /* Relock the rtmutex and try again */
1358                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1359         }
1360
1361         /*
1362          * The wakeup next waiter path does not suffer from the above
1363          * race. See the comments there.
1364          *
1365          * Queue the next waiter for wakeup once we release the wait_lock.
1366          */
1367         mark_wakeup_next_waiter(&wake_q, lock);
1368         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1369
1370         rt_mutex_postunlock(&wake_q);
1371 }
1372
1373 /*
1374  * debug aware fast / slowpath lock,trylock,unlock
1375  *
1376  * The atomic acquire/release ops are compiled away, when either the
1377  * architecture does not support cmpxchg or when debugging is enabled.
1378  */
1379 static __always_inline int __rt_mutex_lock(struct rt_mutex *lock, long state,
1380                                            unsigned int subclass)
1381 {
1382         int ret;
1383
1384         might_sleep();
1385         mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
1386
1387         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1388                 return 0;
1389
1390         ret = rt_mutex_slowlock(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1391         if (ret)
1392                 mutex_release(&lock->dep_map, _RET_IP_);
1393         return ret;
1394 }
1395
1396 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1397 /**
1398  * rt_mutex_lock_nested - lock a rt_mutex
1399  *
1400  * @lock: the rt_mutex to be locked
1401  * @subclass: the lockdep subclass
1402  */
1403 void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
1404 {
1405         __rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, subclass);
1406 }
1407 EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
1408
1409 #else /* !CONFIG_DEBUG_LOCK_ALLOC */
1410
1411 /**
1412  * rt_mutex_lock - lock a rt_mutex
1413  *
1414  * @lock: the rt_mutex to be locked
1415  */
1416 void __sched rt_mutex_lock(struct rt_mutex *lock)
1417 {
1418         __rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0);
1419 }
1420 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1421 #endif
1422
1423 /**
1424  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1425  *
1426  * @lock:               the rt_mutex to be locked
1427  *
1428  * Returns:
1429  *  0           on success
1430  * -EINTR       when interrupted by a signal
1431  */
1432 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1433 {
1434         return __rt_mutex_lock(lock, TASK_INTERRUPTIBLE, 0);
1435 }
1436 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1437
1438 /**
1439  * rt_mutex_trylock - try to lock a rt_mutex
1440  *
1441  * @lock:       the rt_mutex to be locked
1442  *
1443  * This function can only be called in thread context. It's safe to call it
1444  * from atomic regions, but not from hard or soft interrupt context.
1445  *
1446  * Returns:
1447  *  1 on success
1448  *  0 on contention
1449  */
1450 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1451 {
1452         int ret;
1453
1454         if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES) && WARN_ON_ONCE(!in_task()))
1455                 return 0;
1456
1457         /*
1458          * No lockdep annotation required because lockdep disables the fast
1459          * path.
1460          */
1461         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1462                 return 1;
1463
1464         ret = rt_mutex_slowtrylock(lock);
1465         if (ret)
1466                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1467
1468         return ret;
1469 }
1470 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1471
1472 /**
1473  * rt_mutex_unlock - unlock a rt_mutex
1474  *
1475  * @lock: the rt_mutex to be unlocked
1476  */
1477 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1478 {
1479         mutex_release(&lock->dep_map, _RET_IP_);
1480         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1481                 return;
1482
1483         rt_mutex_slowunlock(lock);
1484 }
1485 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1486
1487 /*
1488  * Futex variants, must not use fastpath.
1489  */
1490 int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1491 {
1492         return rt_mutex_slowtrylock(lock);
1493 }
1494
1495 int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1496 {
1497         return __rt_mutex_slowtrylock(lock);
1498 }
1499
1500 /**
1501  * __rt_mutex_futex_unlock - Futex variant, that since futex variants
1502  * do not use the fast-path, can be simple and will not need to retry.
1503  *
1504  * @lock:       The rt_mutex to be unlocked
1505  * @wake_q:     The wake queue head from which to get the next lock waiter
1506  */
1507 bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1508                                      struct wake_q_head *wake_q)
1509 {
1510         lockdep_assert_held(&lock->wait_lock);
1511
1512         debug_rt_mutex_unlock(lock);
1513
1514         if (!rt_mutex_has_waiters(lock)) {
1515                 lock->owner = NULL;
1516                 return false; /* done */
1517         }
1518
1519         /*
1520          * We've already deboosted, mark_wakeup_next_waiter() will
1521          * retain preempt_disabled when we drop the wait_lock, to
1522          * avoid inversion prior to the wakeup.  preempt_disable()
1523          * therein pairs with rt_mutex_postunlock().
1524          */
1525         mark_wakeup_next_waiter(wake_q, lock);
1526
1527         return true; /* call postunlock() */
1528 }
1529
1530 void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1531 {
1532         DEFINE_WAKE_Q(wake_q);
1533         unsigned long flags;
1534         bool postunlock;
1535
1536         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1537         postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1538         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1539
1540         if (postunlock)
1541                 rt_mutex_postunlock(&wake_q);
1542 }
1543
1544 /**
1545  * __rt_mutex_init - initialize the rt_mutex
1546  *
1547  * @lock:       The rt_mutex to be initialized
1548  * @name:       The lock name used for debugging
1549  * @key:        The lock class key used for debugging
1550  *
1551  * Initialize the rt_mutex to unlocked state.
1552  *
1553  * Initializing of a locked rt_mutex is not allowed
1554  */
1555 void __sched __rt_mutex_init(struct rt_mutex *lock, const char *name,
1556                      struct lock_class_key *key)
1557 {
1558         debug_check_no_locks_freed((void *)lock, sizeof(*lock));
1559         lockdep_init_map(&lock->dep_map, name, key, 0);
1560
1561         __rt_mutex_basic_init(lock);
1562 }
1563 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1564
1565 /**
1566  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1567  *                              proxy owner
1568  *
1569  * @lock:       the rt_mutex to be locked
1570  * @proxy_owner:the task to set as owner
1571  *
1572  * No locking. Caller has to do serializing itself
1573  *
1574  * Special API call for PI-futex support. This initializes the rtmutex and
1575  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1576  * possible at this point because the pi_state which contains the rtmutex
1577  * is not yet visible to other tasks.
1578  */
1579 void __sched rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1580                                         struct task_struct *proxy_owner)
1581 {
1582         __rt_mutex_basic_init(lock);
1583         rt_mutex_set_owner(lock, proxy_owner);
1584 }
1585
1586 /**
1587  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1588  *
1589  * @lock:       the rt_mutex to be locked
1590  *
1591  * No locking. Caller has to do serializing itself
1592  *
1593  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1594  * (debugging) state. Concurrent operations on this rt_mutex are not
1595  * possible because it belongs to the pi_state which is about to be freed
1596  * and it is not longer visible to other tasks.
1597  */
1598 void __sched rt_mutex_proxy_unlock(struct rt_mutex *lock)
1599 {
1600         debug_rt_mutex_proxy_unlock(lock);
1601         rt_mutex_set_owner(lock, NULL);
1602 }
1603
1604 /**
1605  * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1606  * @lock:               the rt_mutex to take
1607  * @waiter:             the pre-initialized rt_mutex_waiter
1608  * @task:               the task to prepare
1609  *
1610  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1611  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1612  *
1613  * NOTE: does _NOT_ remove the @waiter on failure; must either call
1614  * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
1615  *
1616  * Returns:
1617  *  0 - task blocked on lock
1618  *  1 - acquired the lock for task, caller should wake it up
1619  * <0 - error
1620  *
1621  * Special API call for PI-futex support.
1622  */
1623 int __sched __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1624                                         struct rt_mutex_waiter *waiter,
1625                                         struct task_struct *task)
1626 {
1627         int ret;
1628
1629         lockdep_assert_held(&lock->wait_lock);
1630
1631         if (try_to_take_rt_mutex(lock, task, NULL))
1632                 return 1;
1633
1634         /* We enforce deadlock detection for futexes */
1635         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1636                                       RT_MUTEX_FULL_CHAINWALK);
1637
1638         if (ret && !rt_mutex_owner(lock)) {
1639                 /*
1640                  * Reset the return value. We might have
1641                  * returned with -EDEADLK and the owner
1642                  * released the lock while we were walking the
1643                  * pi chain.  Let the waiter sort it out.
1644                  */
1645                 ret = 0;
1646         }
1647
1648         return ret;
1649 }
1650
1651 /**
1652  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1653  * @lock:               the rt_mutex to take
1654  * @waiter:             the pre-initialized rt_mutex_waiter
1655  * @task:               the task to prepare
1656  *
1657  * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1658  * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1659  *
1660  * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
1661  * on failure.
1662  *
1663  * Returns:
1664  *  0 - task blocked on lock
1665  *  1 - acquired the lock for task, caller should wake it up
1666  * <0 - error
1667  *
1668  * Special API call for PI-futex support.
1669  */
1670 int __sched rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1671                                       struct rt_mutex_waiter *waiter,
1672                                       struct task_struct *task)
1673 {
1674         int ret;
1675
1676         raw_spin_lock_irq(&lock->wait_lock);
1677         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1678         if (unlikely(ret))
1679                 remove_waiter(lock, waiter);
1680         raw_spin_unlock_irq(&lock->wait_lock);
1681
1682         return ret;
1683 }
1684
1685 /**
1686  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1687  * @lock:               the rt_mutex we were woken on
1688  * @to:                 the timeout, null if none. hrtimer should already have
1689  *                      been started.
1690  * @waiter:             the pre-initialized rt_mutex_waiter
1691  *
1692  * Wait for the lock acquisition started on our behalf by
1693  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1694  * rt_mutex_cleanup_proxy_lock().
1695  *
1696  * Returns:
1697  *  0 - success
1698  * <0 - error, one of -EINTR, -ETIMEDOUT
1699  *
1700  * Special API call for PI-futex support
1701  */
1702 int __sched rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1703                                      struct hrtimer_sleeper *to,
1704                                      struct rt_mutex_waiter *waiter)
1705 {
1706         int ret;
1707
1708         raw_spin_lock_irq(&lock->wait_lock);
1709         /* sleep on the mutex */
1710         set_current_state(TASK_INTERRUPTIBLE);
1711         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1712         /*
1713          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1714          * have to fix that up.
1715          */
1716         fixup_rt_mutex_waiters(lock);
1717         raw_spin_unlock_irq(&lock->wait_lock);
1718
1719         return ret;
1720 }
1721
1722 /**
1723  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1724  * @lock:               the rt_mutex we were woken on
1725  * @waiter:             the pre-initialized rt_mutex_waiter
1726  *
1727  * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
1728  * rt_mutex_wait_proxy_lock().
1729  *
1730  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1731  * in fact still be granted ownership until we're removed. Therefore we can
1732  * find we are in fact the owner and must disregard the
1733  * rt_mutex_wait_proxy_lock() failure.
1734  *
1735  * Returns:
1736  *  true  - did the cleanup, we done.
1737  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1738  *          caller should disregards its return value.
1739  *
1740  * Special API call for PI-futex support
1741  */
1742 bool __sched rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1743                                          struct rt_mutex_waiter *waiter)
1744 {
1745         bool cleanup = false;
1746
1747         raw_spin_lock_irq(&lock->wait_lock);
1748         /*
1749          * Do an unconditional try-lock, this deals with the lock stealing
1750          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1751          * sets a NULL owner.
1752          *
1753          * We're not interested in the return value, because the subsequent
1754          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1755          * we will own the lock and it will have removed the waiter. If we
1756          * failed the trylock, we're still not owner and we need to remove
1757          * ourselves.
1758          */
1759         try_to_take_rt_mutex(lock, current, waiter);
1760         /*
1761          * Unless we're the owner; we're still enqueued on the wait_list.
1762          * So check if we became owner, if not, take us off the wait_list.
1763          */
1764         if (rt_mutex_owner(lock) != current) {
1765                 remove_waiter(lock, waiter);
1766                 cleanup = true;
1767         }
1768         /*
1769          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1770          * have to fix that up.
1771          */
1772         fixup_rt_mutex_waiters(lock);
1773
1774         raw_spin_unlock_irq(&lock->wait_lock);
1775
1776         return cleanup;
1777 }
1778
1779 #ifdef CONFIG_DEBUG_RT_MUTEXES
1780 void rt_mutex_debug_task_free(struct task_struct *task)
1781 {
1782         DEBUG_LOCKS_WARN_ON(!RB_EMPTY_ROOT(&task->pi_waiters.rb_root));
1783         DEBUG_LOCKS_WARN_ON(task->pi_blocked_on);
1784 }
1785 #endif