projects/linux·blob v5.2-rc7 · refs

1925 lines · 51.0 KiB

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 <[email protected]>
8 * Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <[email protected]>
9 * Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
10 * Copyright (C) 2006 Esben Nielsen
11 *
12 * See Documentation/locking/rt-mutex-design.txt 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
52static void
53rt_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
63static 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
69static 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 */
153static 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 */
169static 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
209static 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 */
218static 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
234static inline int
235rt_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
253static inline int
254rt_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
272static void
273rt_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
295static void
296rt_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
305static void
306rt_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
328static void
329rt_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
338static 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 */
363static 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 */
379int max_lock_depth = 1024;
380
381static 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 */
449static 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 = rt_mutex_owner(lock);
632 get_task_struct(task);
633 raw_spin_lock(&task->pi_lock);
634
635 /*
636 * No requeue [11] here. We just do deadlock detection.
637 *
638 * [12] Store whether owner is blocked
639 * itself. Decision is made after dropping the locks
640 */
641 next_lock = task_blocked_on_lock(task);
642 /*
643 * Get the top waiter for the next iteration
644 */
645 top_waiter = rt_mutex_top_waiter(lock);
646
647 /* [13] Drop locks */
648 raw_spin_unlock(&task->pi_lock);
649 raw_spin_unlock_irq(&lock->wait_lock);
650
651 /* If owner is not blocked, end of chain. */
652 if (!next_lock)
653 goto out_put_task;
654 goto again;
655 }
656
657 /*
658 * Store the current top waiter before doing the requeue
659 * operation on @lock. We need it for the boost/deboost
660 * decision below.
661 */
662 prerequeue_top_waiter = rt_mutex_top_waiter(lock);
663
664 /* [7] Requeue the waiter in the lock waiter tree. */
665 rt_mutex_dequeue(lock, waiter);
666
667 /*
668 * Update the waiter prio fields now that we're dequeued.
669 *
670 * These values can have changed through either:
671 *
672 * sys_sched_set_scheduler() / sys_sched_setattr()
673 *
674 * or
675 *
676 * DL CBS enforcement advancing the effective deadline.
677 *
678 * Even though pi_waiters also uses these fields, and that tree is only
679 * updated in [11], we can do this here, since we hold [L], which
680 * serializes all pi_waiters access and rb_erase() does not care about
681 * the values of the node being removed.
682 */
683 waiter->prio = task->prio;
684 waiter->deadline = task->dl.deadline;
685
686 rt_mutex_enqueue(lock, waiter);
687
688 /* [8] Release the task */
689 raw_spin_unlock(&task->pi_lock);
690 put_task_struct(task);
691
692 /*
693 * [9] check_exit_conditions_3 protected by lock->wait_lock.
694 *
695 * We must abort the chain walk if there is no lock owner even
696 * in the dead lock detection case, as we have nothing to
697 * follow here. This is the end of the chain we are walking.
698 */
699 if (!rt_mutex_owner(lock)) {
700 /*
701 * If the requeue [7] above changed the top waiter,
702 * then we need to wake the new top waiter up to try
703 * to get the lock.
704 */
705 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
706 wake_up_process(rt_mutex_top_waiter(lock)->task);
707 raw_spin_unlock_irq(&lock->wait_lock);
708 return 0;
709 }
710
711 /* [10] Grab the next task, i.e. the owner of @lock */
712 task = rt_mutex_owner(lock);
713 get_task_struct(task);
714 raw_spin_lock(&task->pi_lock);
715
716 /* [11] requeue the pi waiters if necessary */
717 if (waiter == rt_mutex_top_waiter(lock)) {
718 /*
719 * The waiter became the new top (highest priority)
720 * waiter on the lock. Replace the previous top waiter
721 * in the owner tasks pi waiters tree with this waiter
722 * and adjust the priority of the owner.
723 */
724 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
725 rt_mutex_enqueue_pi(task, waiter);
726 rt_mutex_adjust_prio(task);
727
728 } else if (prerequeue_top_waiter == waiter) {
729 /*
730 * The waiter was the top waiter on the lock, but is
731 * no longer the top prority waiter. Replace waiter in
732 * the owner tasks pi waiters tree with the new top
733 * (highest priority) waiter and adjust the priority
734 * of the owner.
735 * The new top waiter is stored in @waiter so that
736 * @waiter == @top_waiter evaluates to true below and
737 * we continue to deboost the rest of the chain.
738 */
739 rt_mutex_dequeue_pi(task, waiter);
740 waiter = rt_mutex_top_waiter(lock);
741 rt_mutex_enqueue_pi(task, waiter);
742 rt_mutex_adjust_prio(task);
743 } else {
744 /*
745 * Nothing changed. No need to do any priority
746 * adjustment.
747 */
748 }
749
750 /*
751 * [12] check_exit_conditions_4() protected by task->pi_lock
752 * and lock->wait_lock. The actual decisions are made after we
753 * dropped the locks.
754 *
755 * Check whether the task which owns the current lock is pi
756 * blocked itself. If yes we store a pointer to the lock for
757 * the lock chain change detection above. After we dropped
758 * task->pi_lock next_lock cannot be dereferenced anymore.
759 */
760 next_lock = task_blocked_on_lock(task);
761 /*
762 * Store the top waiter of @lock for the end of chain walk
763 * decision below.
764 */
765 top_waiter = rt_mutex_top_waiter(lock);
766
767 /* [13] Drop the locks */
768 raw_spin_unlock(&task->pi_lock);
769 raw_spin_unlock_irq(&lock->wait_lock);
770
771 /*
772 * Make the actual exit decisions [12], based on the stored
773 * values.
774 *
775 * We reached the end of the lock chain. Stop right here. No
776 * point to go back just to figure that out.
777 */
778 if (!next_lock)
779 goto out_put_task;
780
781 /*
782 * If the current waiter is not the top waiter on the lock,
783 * then we can stop the chain walk here if we are not in full
784 * deadlock detection mode.
785 */
786 if (!detect_deadlock && waiter != top_waiter)
787 goto out_put_task;
788
789 goto again;
790
791 out_unlock_pi:
792 raw_spin_unlock_irq(&task->pi_lock);
793 out_put_task:
794 put_task_struct(task);
795
796 return ret;
797}
798
799/*
800 * Try to take an rt-mutex
801 *
802 * Must be called with lock->wait_lock held and interrupts disabled
803 *
804 * @lock: The lock to be acquired.
805 * @task: The task which wants to acquire the lock
806 * @waiter: The waiter that is queued to the lock's wait tree if the
807 * callsite called task_blocked_on_lock(), otherwise NULL
808 */
809static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
810 struct rt_mutex_waiter *waiter)
811{
812 lockdep_assert_held(&lock->wait_lock);
813
814 /*
815 * Before testing whether we can acquire @lock, we set the
816 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
817 * other tasks which try to modify @lock into the slow path
818 * and they serialize on @lock->wait_lock.
819 *
820 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
821 * as explained at the top of this file if and only if:
822 *
823 * - There is a lock owner. The caller must fixup the
824 * transient state if it does a trylock or leaves the lock
825 * function due to a signal or timeout.
826 *
827 * - @task acquires the lock and there are no other
828 * waiters. This is undone in rt_mutex_set_owner(@task) at
829 * the end of this function.
830 */
831 mark_rt_mutex_waiters(lock);
832
833 /*
834 * If @lock has an owner, give up.
835 */
836 if (rt_mutex_owner(lock))
837 return 0;
838
839 /*
840 * If @waiter != NULL, @task has already enqueued the waiter
841 * into @lock waiter tree. If @waiter == NULL then this is a
842 * trylock attempt.
843 */
844 if (waiter) {
845 /*
846 * If waiter is not the highest priority waiter of
847 * @lock, give up.
848 */
849 if (waiter != rt_mutex_top_waiter(lock))
850 return 0;
851
852 /*
853 * We can acquire the lock. Remove the waiter from the
854 * lock waiters tree.
855 */
856 rt_mutex_dequeue(lock, waiter);
857
858 } else {
859 /*
860 * If the lock has waiters already we check whether @task is
861 * eligible to take over the lock.
862 *
863 * If there are no other waiters, @task can acquire
864 * the lock. @task->pi_blocked_on is NULL, so it does
865 * not need to be dequeued.
866 */
867 if (rt_mutex_has_waiters(lock)) {
868 /*
869 * If @task->prio is greater than or equal to
870 * the top waiter priority (kernel view),
871 * @task lost.
872 */
873 if (!rt_mutex_waiter_less(task_to_waiter(task),
874 rt_mutex_top_waiter(lock)))
875 return 0;
876
877 /*
878 * The current top waiter stays enqueued. We
879 * don't have to change anything in the lock
880 * waiters order.
881 */
882 } else {
883 /*
884 * No waiters. Take the lock without the
885 * pi_lock dance.@task->pi_blocked_on is NULL
886 * and we have no waiters to enqueue in @task
887 * pi waiters tree.
888 */
889 goto takeit;
890 }
891 }
892
893 /*
894 * Clear @task->pi_blocked_on. Requires protection by
895 * @task->pi_lock. Redundant operation for the @waiter == NULL
896 * case, but conditionals are more expensive than a redundant
897 * store.
898 */
899 raw_spin_lock(&task->pi_lock);
900 task->pi_blocked_on = NULL;
901 /*
902 * Finish the lock acquisition. @task is the new owner. If
903 * other waiters exist we have to insert the highest priority
904 * waiter into @task->pi_waiters tree.
905 */
906 if (rt_mutex_has_waiters(lock))
907 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
908 raw_spin_unlock(&task->pi_lock);
909
910takeit:
911 /* We got the lock. */
912 debug_rt_mutex_lock(lock);
913
914 /*
915 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
916 * are still waiters or clears it.
917 */
918 rt_mutex_set_owner(lock, task);
919
920 return 1;
921}
922
923/*
924 * Task blocks on lock.
925 *
926 * Prepare waiter and propagate pi chain
927 *
928 * This must be called with lock->wait_lock held and interrupts disabled
929 */
930static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
931 struct rt_mutex_waiter *waiter,
932 struct task_struct *task,
933 enum rtmutex_chainwalk chwalk)
934{
935 struct task_struct *owner = rt_mutex_owner(lock);
936 struct rt_mutex_waiter *top_waiter = waiter;
937 struct rt_mutex *next_lock;
938 int chain_walk = 0, res;
939
940 lockdep_assert_held(&lock->wait_lock);
941
942 /*
943 * Early deadlock detection. We really don't want the task to
944 * enqueue on itself just to untangle the mess later. It's not
945 * only an optimization. We drop the locks, so another waiter
946 * can come in before the chain walk detects the deadlock. So
947 * the other will detect the deadlock and return -EDEADLOCK,
948 * which is wrong, as the other waiter is not in a deadlock
949 * situation.
950 */
951 if (owner == task)
952 return -EDEADLK;
953
954 raw_spin_lock(&task->pi_lock);
955 waiter->task = task;
956 waiter->lock = lock;
957 waiter->prio = task->prio;
958 waiter->deadline = task->dl.deadline;
959
960 /* Get the top priority waiter on the lock */
961 if (rt_mutex_has_waiters(lock))
962 top_waiter = rt_mutex_top_waiter(lock);
963 rt_mutex_enqueue(lock, waiter);
964
965 task->pi_blocked_on = waiter;
966
967 raw_spin_unlock(&task->pi_lock);
968
969 if (!owner)
970 return 0;
971
972 raw_spin_lock(&owner->pi_lock);
973 if (waiter == rt_mutex_top_waiter(lock)) {
974 rt_mutex_dequeue_pi(owner, top_waiter);
975 rt_mutex_enqueue_pi(owner, waiter);
976
977 rt_mutex_adjust_prio(owner);
978 if (owner->pi_blocked_on)
979 chain_walk = 1;
980 } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
981 chain_walk = 1;
982 }
983
984 /* Store the lock on which owner is blocked or NULL */
985 next_lock = task_blocked_on_lock(owner);
986
987 raw_spin_unlock(&owner->pi_lock);
988 /*
989 * Even if full deadlock detection is on, if the owner is not
990 * blocked itself, we can avoid finding this out in the chain
991 * walk.
992 */
993 if (!chain_walk || !next_lock)
994 return 0;
995
996 /*
997 * The owner can't disappear while holding a lock,
998 * so the owner struct is protected by wait_lock.
999 * Gets dropped in rt_mutex_adjust_prio_chain()!
1000 */
1001 get_task_struct(owner);
1002
1003 raw_spin_unlock_irq(&lock->wait_lock);
1004
1005 res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1006 next_lock, waiter, task);
1007
1008 raw_spin_lock_irq(&lock->wait_lock);
1009
1010 return res;
1011}
1012
1013/*
1014 * Remove the top waiter from the current tasks pi waiter tree and
1015 * queue it up.
1016 *
1017 * Called with lock->wait_lock held and interrupts disabled.
1018 */
1019static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1020 struct rt_mutex *lock)
1021{
1022 struct rt_mutex_waiter *waiter;
1023
1024 raw_spin_lock(&current->pi_lock);
1025
1026 waiter = rt_mutex_top_waiter(lock);
1027
1028 /*
1029 * Remove it from current->pi_waiters and deboost.
1030 *
1031 * We must in fact deboost here in order to ensure we call
1032 * rt_mutex_setprio() to update p->pi_top_task before the
1033 * task unblocks.
1034 */
1035 rt_mutex_dequeue_pi(current, waiter);
1036 rt_mutex_adjust_prio(current);
1037
1038 /*
1039 * As we are waking up the top waiter, and the waiter stays
1040 * queued on the lock until it gets the lock, this lock
1041 * obviously has waiters. Just set the bit here and this has
1042 * the added benefit of forcing all new tasks into the
1043 * slow path making sure no task of lower priority than
1044 * the top waiter can steal this lock.
1045 */
1046 lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1047
1048 /*
1049 * We deboosted before waking the top waiter task such that we don't
1050 * run two tasks with the 'same' priority (and ensure the
1051 * p->pi_top_task pointer points to a blocked task). This however can
1052 * lead to priority inversion if we would get preempted after the
1053 * deboost but before waking our donor task, hence the preempt_disable()
1054 * before unlock.
1055 *
1056 * Pairs with preempt_enable() in rt_mutex_postunlock();
1057 */
1058 preempt_disable();
1059 wake_q_add(wake_q, waiter->task);
1060 raw_spin_unlock(&current->pi_lock);
1061}
1062
1063/*
1064 * Remove a waiter from a lock and give up
1065 *
1066 * Must be called with lock->wait_lock held and interrupts disabled. I must
1067 * have just failed to try_to_take_rt_mutex().
1068 */
1069static void remove_waiter(struct rt_mutex *lock,
1070 struct rt_mutex_waiter *waiter)
1071{
1072 bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1073 struct task_struct *owner = rt_mutex_owner(lock);
1074 struct rt_mutex *next_lock;
1075
1076 lockdep_assert_held(&lock->wait_lock);
1077
1078 raw_spin_lock(&current->pi_lock);
1079 rt_mutex_dequeue(lock, waiter);
1080 current->pi_blocked_on = NULL;
1081 raw_spin_unlock(&current->pi_lock);
1082
1083 /*
1084 * Only update priority if the waiter was the highest priority
1085 * waiter of the lock and there is an owner to update.
1086 */
1087 if (!owner || !is_top_waiter)
1088 return;
1089
1090 raw_spin_lock(&owner->pi_lock);
1091
1092 rt_mutex_dequeue_pi(owner, waiter);
1093
1094 if (rt_mutex_has_waiters(lock))
1095 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1096
1097 rt_mutex_adjust_prio(owner);
1098
1099 /* Store the lock on which owner is blocked or NULL */
1100 next_lock = task_blocked_on_lock(owner);
1101
1102 raw_spin_unlock(&owner->pi_lock);
1103
1104 /*
1105 * Don't walk the chain, if the owner task is not blocked
1106 * itself.
1107 */
1108 if (!next_lock)
1109 return;
1110
1111 /* gets dropped in rt_mutex_adjust_prio_chain()! */
1112 get_task_struct(owner);
1113
1114 raw_spin_unlock_irq(&lock->wait_lock);
1115
1116 rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1117 next_lock, NULL, current);
1118
1119 raw_spin_lock_irq(&lock->wait_lock);
1120}
1121
1122/*
1123 * Recheck the pi chain, in case we got a priority setting
1124 *
1125 * Called from sched_setscheduler
1126 */
1127void rt_mutex_adjust_pi(struct task_struct *task)
1128{
1129 struct rt_mutex_waiter *waiter;
1130 struct rt_mutex *next_lock;
1131 unsigned long flags;
1132
1133 raw_spin_lock_irqsave(&task->pi_lock, flags);
1134
1135 waiter = task->pi_blocked_on;
1136 if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1137 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1138 return;
1139 }
1140 next_lock = waiter->lock;
1141 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1142
1143 /* gets dropped in rt_mutex_adjust_prio_chain()! */
1144 get_task_struct(task);
1145
1146 rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1147 next_lock, NULL, task);
1148}
1149
1150void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1151{
1152 debug_rt_mutex_init_waiter(waiter);
1153 RB_CLEAR_NODE(&waiter->pi_tree_entry);
1154 RB_CLEAR_NODE(&waiter->tree_entry);
1155 waiter->task = NULL;
1156}
1157
1158/**
1159 * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1160 * @lock: the rt_mutex to take
1161 * @state: the state the task should block in (TASK_INTERRUPTIBLE
1162 * or TASK_UNINTERRUPTIBLE)
1163 * @timeout: the pre-initialized and started timer, or NULL for none
1164 * @waiter: the pre-initialized rt_mutex_waiter
1165 *
1166 * Must be called with lock->wait_lock held and interrupts disabled
1167 */
1168static int __sched
1169__rt_mutex_slowlock(struct rt_mutex *lock, int state,
1170 struct hrtimer_sleeper *timeout,
1171 struct rt_mutex_waiter *waiter)
1172{
1173 int ret = 0;
1174
1175 for (;;) {
1176 /* Try to acquire the lock: */
1177 if (try_to_take_rt_mutex(lock, current, waiter))
1178 break;
1179
1180 /*
1181 * TASK_INTERRUPTIBLE checks for signals and
1182 * timeout. Ignored otherwise.
1183 */
1184 if (likely(state == TASK_INTERRUPTIBLE)) {
1185 /* Signal pending? */
1186 if (signal_pending(current))
1187 ret = -EINTR;
1188 if (timeout && !timeout->task)
1189 ret = -ETIMEDOUT;
1190 if (ret)
1191 break;
1192 }
1193
1194 raw_spin_unlock_irq(&lock->wait_lock);
1195
1196 debug_rt_mutex_print_deadlock(waiter);
1197
1198 schedule();
1199
1200 raw_spin_lock_irq(&lock->wait_lock);
1201 set_current_state(state);
1202 }
1203
1204 __set_current_state(TASK_RUNNING);
1205 return ret;
1206}
1207
1208static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1209 struct rt_mutex_waiter *w)
1210{
1211 /*
1212 * If the result is not -EDEADLOCK or the caller requested
1213 * deadlock detection, nothing to do here.
1214 */
1215 if (res != -EDEADLOCK || detect_deadlock)
1216 return;
1217
1218 /*
1219 * Yell lowdly and stop the task right here.
1220 */
1221 rt_mutex_print_deadlock(w);
1222 while (1) {
1223 set_current_state(TASK_INTERRUPTIBLE);
1224 schedule();
1225 }
1226}
1227
1228/*
1229 * Slow path lock function:
1230 */
1231static int __sched
1232rt_mutex_slowlock(struct rt_mutex *lock, int state,
1233 struct hrtimer_sleeper *timeout,
1234 enum rtmutex_chainwalk chwalk)
1235{
1236 struct rt_mutex_waiter waiter;
1237 unsigned long flags;
1238 int ret = 0;
1239
1240 rt_mutex_init_waiter(&waiter);
1241
1242 /*
1243 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1244 * be called in early boot if the cmpxchg() fast path is disabled
1245 * (debug, no architecture support). In this case we will acquire the
1246 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1247 * enable interrupts in that early boot case. So we need to use the
1248 * irqsave/restore variants.
1249 */
1250 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1251
1252 /* Try to acquire the lock again: */
1253 if (try_to_take_rt_mutex(lock, current, NULL)) {
1254 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1255 return 0;
1256 }
1257
1258 set_current_state(state);
1259
1260 /* Setup the timer, when timeout != NULL */
1261 if (unlikely(timeout))
1262 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1263
1264 ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1265
1266 if (likely(!ret))
1267 /* sleep on the mutex */
1268 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1269
1270 if (unlikely(ret)) {
1271 __set_current_state(TASK_RUNNING);
1272 remove_waiter(lock, &waiter);
1273 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1274 }
1275
1276 /*
1277 * try_to_take_rt_mutex() sets the waiter bit
1278 * unconditionally. We might have to fix that up.
1279 */
1280 fixup_rt_mutex_waiters(lock);
1281
1282 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1283
1284 /* Remove pending timer: */
1285 if (unlikely(timeout))
1286 hrtimer_cancel(&timeout->timer);
1287
1288 debug_rt_mutex_free_waiter(&waiter);
1289
1290 return ret;
1291}
1292
1293static inline int __rt_mutex_slowtrylock(struct rt_mutex *lock)
1294{
1295 int ret = try_to_take_rt_mutex(lock, current, NULL);
1296
1297 /*
1298 * try_to_take_rt_mutex() sets the lock waiters bit
1299 * unconditionally. Clean this up.
1300 */
1301 fixup_rt_mutex_waiters(lock);
1302
1303 return ret;
1304}
1305
1306/*
1307 * Slow path try-lock function:
1308 */
1309static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1310{
1311 unsigned long flags;
1312 int ret;
1313
1314 /*
1315 * If the lock already has an owner we fail to get the lock.
1316 * This can be done without taking the @lock->wait_lock as
1317 * it is only being read, and this is a trylock anyway.
1318 */
1319 if (rt_mutex_owner(lock))
1320 return 0;
1321
1322 /*
1323 * The mutex has currently no owner. Lock the wait lock and try to
1324 * acquire the lock. We use irqsave here to support early boot calls.
1325 */
1326 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1327
1328 ret = __rt_mutex_slowtrylock(lock);
1329
1330 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1331
1332 return ret;
1333}
1334
1335/*
1336 * Slow path to release a rt-mutex.
1337 *
1338 * Return whether the current task needs to call rt_mutex_postunlock().
1339 */
1340static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1341 struct wake_q_head *wake_q)
1342{
1343 unsigned long flags;
1344
1345 /* irqsave required to support early boot calls */
1346 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1347
1348 debug_rt_mutex_unlock(lock);
1349
1350 /*
1351 * We must be careful here if the fast path is enabled. If we
1352 * have no waiters queued we cannot set owner to NULL here
1353 * because of:
1354 *
1355 * foo->lock->owner = NULL;
1356 * rtmutex_lock(foo->lock); <- fast path
1357 * free = atomic_dec_and_test(foo->refcnt);
1358 * rtmutex_unlock(foo->lock); <- fast path
1359 * if (free)
1360 * kfree(foo);
1361 * raw_spin_unlock(foo->lock->wait_lock);
1362 *
1363 * So for the fastpath enabled kernel:
1364 *
1365 * Nothing can set the waiters bit as long as we hold
1366 * lock->wait_lock. So we do the following sequence:
1367 *
1368 * owner = rt_mutex_owner(lock);
1369 * clear_rt_mutex_waiters(lock);
1370 * raw_spin_unlock(&lock->wait_lock);
1371 * if (cmpxchg(&lock->owner, owner, 0) == owner)
1372 * return;
1373 * goto retry;
1374 *
1375 * The fastpath disabled variant is simple as all access to
1376 * lock->owner is serialized by lock->wait_lock:
1377 *
1378 * lock->owner = NULL;
1379 * raw_spin_unlock(&lock->wait_lock);
1380 */
1381 while (!rt_mutex_has_waiters(lock)) {
1382 /* Drops lock->wait_lock ! */
1383 if (unlock_rt_mutex_safe(lock, flags) == true)
1384 return false;
1385 /* Relock the rtmutex and try again */
1386 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1387 }
1388
1389 /*
1390 * The wakeup next waiter path does not suffer from the above
1391 * race. See the comments there.
1392 *
1393 * Queue the next waiter for wakeup once we release the wait_lock.
1394 */
1395 mark_wakeup_next_waiter(wake_q, lock);
1396 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1397
1398 return true; /* call rt_mutex_postunlock() */
1399}
1400
1401/*
1402 * debug aware fast / slowpath lock,trylock,unlock
1403 *
1404 * The atomic acquire/release ops are compiled away, when either the
1405 * architecture does not support cmpxchg or when debugging is enabled.
1406 */
1407static inline int
1408rt_mutex_fastlock(struct rt_mutex *lock, int state,
1409 int (*slowfn)(struct rt_mutex *lock, int state,
1410 struct hrtimer_sleeper *timeout,
1411 enum rtmutex_chainwalk chwalk))
1412{
1413 if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1414 return 0;
1415
1416 return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1417}
1418
1419static inline int
1420rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1421 struct hrtimer_sleeper *timeout,
1422 enum rtmutex_chainwalk chwalk,
1423 int (*slowfn)(struct rt_mutex *lock, int state,
1424 struct hrtimer_sleeper *timeout,
1425 enum rtmutex_chainwalk chwalk))
1426{
1427 if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1428 likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1429 return 0;
1430
1431 return slowfn(lock, state, timeout, chwalk);
1432}
1433
1434static inline int
1435rt_mutex_fasttrylock(struct rt_mutex *lock,
1436 int (*slowfn)(struct rt_mutex *lock))
1437{
1438 if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1439 return 1;
1440
1441 return slowfn(lock);
1442}
1443
1444/*
1445 * Performs the wakeup of the the top-waiter and re-enables preemption.
1446 */
1447void rt_mutex_postunlock(struct wake_q_head *wake_q)
1448{
1449 wake_up_q(wake_q);
1450
1451 /* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1452 preempt_enable();
1453}
1454
1455static inline void
1456rt_mutex_fastunlock(struct rt_mutex *lock,
1457 bool (*slowfn)(struct rt_mutex *lock,
1458 struct wake_q_head *wqh))
1459{
1460 DEFINE_WAKE_Q(wake_q);
1461
1462 if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1463 return;
1464
1465 if (slowfn(lock, &wake_q))
1466 rt_mutex_postunlock(&wake_q);
1467}
1468
1469static inline void __rt_mutex_lock(struct rt_mutex *lock, unsigned int subclass)
1470{
1471 might_sleep();
1472
1473 mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
1474 rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1475}
1476
1477#ifdef CONFIG_DEBUG_LOCK_ALLOC
1478/**
1479 * rt_mutex_lock_nested - lock a rt_mutex
1480 *
1481 * @lock: the rt_mutex to be locked
1482 * @subclass: the lockdep subclass
1483 */
1484void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
1485{
1486 __rt_mutex_lock(lock, subclass);
1487}
1488EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
1489
1490#else /* !CONFIG_DEBUG_LOCK_ALLOC */
1491
1492/**
1493 * rt_mutex_lock - lock a rt_mutex
1494 *
1495 * @lock: the rt_mutex to be locked
1496 */
1497void __sched rt_mutex_lock(struct rt_mutex *lock)
1498{
1499 __rt_mutex_lock(lock, 0);
1500}
1501EXPORT_SYMBOL_GPL(rt_mutex_lock);
1502#endif
1503
1504/**
1505 * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1506 *
1507 * @lock: the rt_mutex to be locked
1508 *
1509 * Returns:
1510 * 0 on success
1511 * -EINTR when interrupted by a signal
1512 */
1513int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1514{
1515 int ret;
1516
1517 might_sleep();
1518
1519 mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1520 ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1521 if (ret)
1522 mutex_release(&lock->dep_map, 1, _RET_IP_);
1523
1524 return ret;
1525}
1526EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1527
1528/*
1529 * Futex variant, must not use fastpath.
1530 */
1531int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1532{
1533 return rt_mutex_slowtrylock(lock);
1534}
1535
1536int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1537{
1538 return __rt_mutex_slowtrylock(lock);
1539}
1540
1541/**
1542 * rt_mutex_timed_lock - lock a rt_mutex interruptible
1543 * the timeout structure is provided
1544 * by the caller
1545 *
1546 * @lock: the rt_mutex to be locked
1547 * @timeout: timeout structure or NULL (no timeout)
1548 *
1549 * Returns:
1550 * 0 on success
1551 * -EINTR when interrupted by a signal
1552 * -ETIMEDOUT when the timeout expired
1553 */
1554int
1555rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1556{
1557 int ret;
1558
1559 might_sleep();
1560
1561 mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1562 ret = rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1563 RT_MUTEX_MIN_CHAINWALK,
1564 rt_mutex_slowlock);
1565 if (ret)
1566 mutex_release(&lock->dep_map, 1, _RET_IP_);
1567
1568 return ret;
1569}
1570EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1571
1572/**
1573 * rt_mutex_trylock - try to lock a rt_mutex
1574 *
1575 * @lock: the rt_mutex to be locked
1576 *
1577 * This function can only be called in thread context. It's safe to
1578 * call it from atomic regions, but not from hard interrupt or soft
1579 * interrupt context.
1580 *
1581 * Returns 1 on success and 0 on contention
1582 */
1583int __sched rt_mutex_trylock(struct rt_mutex *lock)
1584{
1585 int ret;
1586
1587 if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1588 return 0;
1589
1590 ret = rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1591 if (ret)
1592 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1593
1594 return ret;
1595}
1596EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1597
1598/**
1599 * rt_mutex_unlock - unlock a rt_mutex
1600 *
1601 * @lock: the rt_mutex to be unlocked
1602 */
1603void __sched rt_mutex_unlock(struct rt_mutex *lock)
1604{
1605 mutex_release(&lock->dep_map, 1, _RET_IP_);
1606 rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1607}
1608EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1609
1610/**
1611 * Futex variant, that since futex variants do not use the fast-path, can be
1612 * simple and will not need to retry.
1613 */
1614bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1615 struct wake_q_head *wake_q)
1616{
1617 lockdep_assert_held(&lock->wait_lock);
1618
1619 debug_rt_mutex_unlock(lock);
1620
1621 if (!rt_mutex_has_waiters(lock)) {
1622 lock->owner = NULL;
1623 return false; /* done */
1624 }
1625
1626 /*
1627 * We've already deboosted, mark_wakeup_next_waiter() will
1628 * retain preempt_disabled when we drop the wait_lock, to
1629 * avoid inversion prior to the wakeup. preempt_disable()
1630 * therein pairs with rt_mutex_postunlock().
1631 */
1632 mark_wakeup_next_waiter(wake_q, lock);
1633
1634 return true; /* call postunlock() */
1635}
1636
1637void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1638{
1639 DEFINE_WAKE_Q(wake_q);
1640 unsigned long flags;
1641 bool postunlock;
1642
1643 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1644 postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1645 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1646
1647 if (postunlock)
1648 rt_mutex_postunlock(&wake_q);
1649}
1650
1651/**
1652 * rt_mutex_destroy - mark a mutex unusable
1653 * @lock: the mutex to be destroyed
1654 *
1655 * This function marks the mutex uninitialized, and any subsequent
1656 * use of the mutex is forbidden. The mutex must not be locked when
1657 * this function is called.
1658 */
1659void rt_mutex_destroy(struct rt_mutex *lock)
1660{
1661 WARN_ON(rt_mutex_is_locked(lock));
1662#ifdef CONFIG_DEBUG_RT_MUTEXES
1663 lock->magic = NULL;
1664#endif
1665}
1666EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1667
1668/**
1669 * __rt_mutex_init - initialize the rt lock
1670 *
1671 * @lock: the rt lock to be initialized
1672 *
1673 * Initialize the rt lock to unlocked state.
1674 *
1675 * Initializing of a locked rt lock is not allowed
1676 */
1677void __rt_mutex_init(struct rt_mutex *lock, const char *name,
1678 struct lock_class_key *key)
1679{
1680 lock->owner = NULL;
1681 raw_spin_lock_init(&lock->wait_lock);
1682 lock->waiters = RB_ROOT_CACHED;
1683
1684 if (name && key)
1685 debug_rt_mutex_init(lock, name, key);
1686}
1687EXPORT_SYMBOL_GPL(__rt_mutex_init);
1688
1689/**
1690 * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1691 * proxy owner
1692 *
1693 * @lock: the rt_mutex to be locked
1694 * @proxy_owner:the task to set as owner
1695 *
1696 * No locking. Caller has to do serializing itself
1697 *
1698 * Special API call for PI-futex support. This initializes the rtmutex and
1699 * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1700 * possible at this point because the pi_state which contains the rtmutex
1701 * is not yet visible to other tasks.
1702 */
1703void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1704 struct task_struct *proxy_owner)
1705{
1706 __rt_mutex_init(lock, NULL, NULL);
1707 debug_rt_mutex_proxy_lock(lock, proxy_owner);
1708 rt_mutex_set_owner(lock, proxy_owner);
1709}
1710
1711/**
1712 * rt_mutex_proxy_unlock - release a lock on behalf of owner
1713 *
1714 * @lock: the rt_mutex to be locked
1715 *
1716 * No locking. Caller has to do serializing itself
1717 *
1718 * Special API call for PI-futex support. This merrily cleans up the rtmutex
1719 * (debugging) state. Concurrent operations on this rt_mutex are not
1720 * possible because it belongs to the pi_state which is about to be freed
1721 * and it is not longer visible to other tasks.
1722 */
1723void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1724 struct task_struct *proxy_owner)
1725{
1726 debug_rt_mutex_proxy_unlock(lock);
1727 rt_mutex_set_owner(lock, NULL);
1728}
1729
1730/**
1731 * __rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1732 * @lock: the rt_mutex to take
1733 * @waiter: the pre-initialized rt_mutex_waiter
1734 * @task: the task to prepare
1735 *
1736 * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1737 * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1738 *
1739 * NOTE: does _NOT_ remove the @waiter on failure; must either call
1740 * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
1741 *
1742 * Returns:
1743 * 0 - task blocked on lock
1744 * 1 - acquired the lock for task, caller should wake it up
1745 * <0 - error
1746 *
1747 * Special API call for PI-futex support.
1748 */
1749int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1750 struct rt_mutex_waiter *waiter,
1751 struct task_struct *task)
1752{
1753 int ret;
1754
1755 lockdep_assert_held(&lock->wait_lock);
1756
1757 if (try_to_take_rt_mutex(lock, task, NULL))
1758 return 1;
1759
1760 /* We enforce deadlock detection for futexes */
1761 ret = task_blocks_on_rt_mutex(lock, waiter, task,
1762 RT_MUTEX_FULL_CHAINWALK);
1763
1764 if (ret && !rt_mutex_owner(lock)) {
1765 /*
1766 * Reset the return value. We might have
1767 * returned with -EDEADLK and the owner
1768 * released the lock while we were walking the
1769 * pi chain. Let the waiter sort it out.
1770 */
1771 ret = 0;
1772 }
1773
1774 debug_rt_mutex_print_deadlock(waiter);
1775
1776 return ret;
1777}
1778
1779/**
1780 * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1781 * @lock: the rt_mutex to take
1782 * @waiter: the pre-initialized rt_mutex_waiter
1783 * @task: the task to prepare
1784 *
1785 * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
1786 * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
1787 *
1788 * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
1789 * on failure.
1790 *
1791 * Returns:
1792 * 0 - task blocked on lock
1793 * 1 - acquired the lock for task, caller should wake it up
1794 * <0 - error
1795 *
1796 * Special API call for PI-futex support.
1797 */
1798int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1799 struct rt_mutex_waiter *waiter,
1800 struct task_struct *task)
1801{
1802 int ret;
1803
1804 raw_spin_lock_irq(&lock->wait_lock);
1805 ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1806 if (unlikely(ret))
1807 remove_waiter(lock, waiter);
1808 raw_spin_unlock_irq(&lock->wait_lock);
1809
1810 return ret;
1811}
1812
1813/**
1814 * rt_mutex_next_owner - return the next owner of the lock
1815 *
1816 * @lock: the rt lock query
1817 *
1818 * Returns the next owner of the lock or NULL
1819 *
1820 * Caller has to serialize against other accessors to the lock
1821 * itself.
1822 *
1823 * Special API call for PI-futex support
1824 */
1825struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1826{
1827 if (!rt_mutex_has_waiters(lock))
1828 return NULL;
1829
1830 return rt_mutex_top_waiter(lock)->task;
1831}
1832
1833/**
1834 * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1835 * @lock: the rt_mutex we were woken on
1836 * @to: the timeout, null if none. hrtimer should already have
1837 * been started.
1838 * @waiter: the pre-initialized rt_mutex_waiter
1839 *
1840 * Wait for the the lock acquisition started on our behalf by
1841 * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1842 * rt_mutex_cleanup_proxy_lock().
1843 *
1844 * Returns:
1845 * 0 - success
1846 * <0 - error, one of -EINTR, -ETIMEDOUT
1847 *
1848 * Special API call for PI-futex support
1849 */
1850int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1851 struct hrtimer_sleeper *to,
1852 struct rt_mutex_waiter *waiter)
1853{
1854 int ret;
1855
1856 raw_spin_lock_irq(&lock->wait_lock);
1857 /* sleep on the mutex */
1858 set_current_state(TASK_INTERRUPTIBLE);
1859 ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1860 /*
1861 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1862 * have to fix that up.
1863 */
1864 fixup_rt_mutex_waiters(lock);
1865 raw_spin_unlock_irq(&lock->wait_lock);
1866
1867 return ret;
1868}
1869
1870/**
1871 * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1872 * @lock: the rt_mutex we were woken on
1873 * @waiter: the pre-initialized rt_mutex_waiter
1874 *
1875 * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
1876 * rt_mutex_wait_proxy_lock().
1877 *
1878 * Unless we acquired the lock; we're still enqueued on the wait-list and can
1879 * in fact still be granted ownership until we're removed. Therefore we can
1880 * find we are in fact the owner and must disregard the
1881 * rt_mutex_wait_proxy_lock() failure.
1882 *
1883 * Returns:
1884 * true - did the cleanup, we done.
1885 * false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1886 * caller should disregards its return value.
1887 *
1888 * Special API call for PI-futex support
1889 */
1890bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1891 struct rt_mutex_waiter *waiter)
1892{
1893 bool cleanup = false;
1894
1895 raw_spin_lock_irq(&lock->wait_lock);
1896 /*
1897 * Do an unconditional try-lock, this deals with the lock stealing
1898 * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1899 * sets a NULL owner.
1900 *
1901 * We're not interested in the return value, because the subsequent
1902 * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1903 * we will own the lock and it will have removed the waiter. If we
1904 * failed the trylock, we're still not owner and we need to remove
1905 * ourselves.
1906 */
1907 try_to_take_rt_mutex(lock, current, waiter);
1908 /*
1909 * Unless we're the owner; we're still enqueued on the wait_list.
1910 * So check if we became owner, if not, take us off the wait_list.
1911 */
1912 if (rt_mutex_owner(lock) != current) {
1913 remove_waiter(lock, waiter);
1914 cleanup = true;
1915 }
1916 /*
1917 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1918 * have to fix that up.
1919 */
1920 fixup_rt_mutex_waiters(lock);
1921
1922 raw_spin_unlock_irq(&lock->wait_lock);
1923
1924 return cleanup;
1925}