xref: /openbmc/linux/kernel/locking/rtmutex.c (revision 4da722ca)
1 /*
2  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
3  *
4  * started by Ingo Molnar and Thomas Gleixner.
5  *
6  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
8  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
9  *  Copyright (C) 2006 Esben Nielsen
10  *
11  *  See Documentation/locking/rt-mutex-design.txt for details.
12  */
13 #include <linux/spinlock.h>
14 #include <linux/export.h>
15 #include <linux/sched/signal.h>
16 #include <linux/sched/rt.h>
17 #include <linux/sched/deadline.h>
18 #include <linux/sched/wake_q.h>
19 #include <linux/sched/debug.h>
20 #include <linux/timer.h>
21 
22 #include "rtmutex_common.h"
23 
24 /*
25  * lock->owner state tracking:
26  *
27  * lock->owner holds the task_struct pointer of the owner. Bit 0
28  * is used to keep track of the "lock has waiters" state.
29  *
30  * owner	bit0
31  * NULL		0	lock is free (fast acquire possible)
32  * NULL		1	lock is free and has waiters and the top waiter
33  *				is going to take the lock*
34  * taskpointer	0	lock is held (fast release possible)
35  * taskpointer	1	lock is held and has waiters**
36  *
37  * The fast atomic compare exchange based acquire and release is only
38  * possible when bit 0 of lock->owner is 0.
39  *
40  * (*) It also can be a transitional state when grabbing the lock
41  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
42  * we need to set the bit0 before looking at the lock, and the owner may be
43  * NULL in this small time, hence this can be a transitional state.
44  *
45  * (**) There is a small time when bit 0 is set but there are no
46  * waiters. This can happen when grabbing the lock in the slow path.
47  * To prevent a cmpxchg of the owner releasing the lock, we need to
48  * set this bit before looking at the lock.
49  */
50 
51 static void
52 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
53 {
54 	unsigned long val = (unsigned long)owner;
55 
56 	if (rt_mutex_has_waiters(lock))
57 		val |= RT_MUTEX_HAS_WAITERS;
58 
59 	lock->owner = (struct task_struct *)val;
60 }
61 
62 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
63 {
64 	lock->owner = (struct task_struct *)
65 			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
66 }
67 
68 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
69 {
70 	unsigned long owner, *p = (unsigned long *) &lock->owner;
71 
72 	if (rt_mutex_has_waiters(lock))
73 		return;
74 
75 	/*
76 	 * The rbtree has no waiters enqueued, now make sure that the
77 	 * lock->owner still has the waiters bit set, otherwise the
78 	 * following can happen:
79 	 *
80 	 * CPU 0	CPU 1		CPU2
81 	 * l->owner=T1
82 	 *		rt_mutex_lock(l)
83 	 *		lock(l->lock)
84 	 *		l->owner = T1 | HAS_WAITERS;
85 	 *		enqueue(T2)
86 	 *		boost()
87 	 *		  unlock(l->lock)
88 	 *		block()
89 	 *
90 	 *				rt_mutex_lock(l)
91 	 *				lock(l->lock)
92 	 *				l->owner = T1 | HAS_WAITERS;
93 	 *				enqueue(T3)
94 	 *				boost()
95 	 *				  unlock(l->lock)
96 	 *				block()
97 	 *		signal(->T2)	signal(->T3)
98 	 *		lock(l->lock)
99 	 *		dequeue(T2)
100 	 *		deboost()
101 	 *		  unlock(l->lock)
102 	 *				lock(l->lock)
103 	 *				dequeue(T3)
104 	 *				 ==> wait list is empty
105 	 *				deboost()
106 	 *				 unlock(l->lock)
107 	 *		lock(l->lock)
108 	 *		fixup_rt_mutex_waiters()
109 	 *		  if (wait_list_empty(l) {
110 	 *		    l->owner = owner
111 	 *		    owner = l->owner & ~HAS_WAITERS;
112 	 *		      ==> l->owner = T1
113 	 *		  }
114 	 *				lock(l->lock)
115 	 * rt_mutex_unlock(l)		fixup_rt_mutex_waiters()
116 	 *				  if (wait_list_empty(l) {
117 	 *				    owner = l->owner & ~HAS_WAITERS;
118 	 * cmpxchg(l->owner, T1, NULL)
119 	 *  ===> Success (l->owner = NULL)
120 	 *
121 	 *				    l->owner = owner
122 	 *				      ==> l->owner = T1
123 	 *				  }
124 	 *
125 	 * With the check for the waiter bit in place T3 on CPU2 will not
126 	 * overwrite. All tasks fiddling with the waiters bit are
127 	 * serialized by l->lock, so nothing else can modify the waiters
128 	 * bit. If the bit is set then nothing can change l->owner either
129 	 * so the simple RMW is safe. The cmpxchg() will simply fail if it
130 	 * happens in the middle of the RMW because the waiters bit is
131 	 * still set.
132 	 */
133 	owner = READ_ONCE(*p);
134 	if (owner & RT_MUTEX_HAS_WAITERS)
135 		WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
136 }
137 
138 /*
139  * We can speed up the acquire/release, if there's no debugging state to be
140  * set up.
141  */
142 #ifndef CONFIG_DEBUG_RT_MUTEXES
143 # define rt_mutex_cmpxchg_relaxed(l,c,n) (cmpxchg_relaxed(&l->owner, c, n) == c)
144 # define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
145 # define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
146 
147 /*
148  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
149  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
150  * relaxed semantics suffice.
151  */
152 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
153 {
154 	unsigned long owner, *p = (unsigned long *) &lock->owner;
155 
156 	do {
157 		owner = *p;
158 	} while (cmpxchg_relaxed(p, owner,
159 				 owner | RT_MUTEX_HAS_WAITERS) != owner);
160 }
161 
162 /*
163  * Safe fastpath aware unlock:
164  * 1) Clear the waiters bit
165  * 2) Drop lock->wait_lock
166  * 3) Try to unlock the lock with cmpxchg
167  */
168 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
169 					unsigned long flags)
170 	__releases(lock->wait_lock)
171 {
172 	struct task_struct *owner = rt_mutex_owner(lock);
173 
174 	clear_rt_mutex_waiters(lock);
175 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
176 	/*
177 	 * If a new waiter comes in between the unlock and the cmpxchg
178 	 * we have two situations:
179 	 *
180 	 * unlock(wait_lock);
181 	 *					lock(wait_lock);
182 	 * cmpxchg(p, owner, 0) == owner
183 	 *					mark_rt_mutex_waiters(lock);
184 	 *					acquire(lock);
185 	 * or:
186 	 *
187 	 * unlock(wait_lock);
188 	 *					lock(wait_lock);
189 	 *					mark_rt_mutex_waiters(lock);
190 	 *
191 	 * cmpxchg(p, owner, 0) != owner
192 	 *					enqueue_waiter();
193 	 *					unlock(wait_lock);
194 	 * lock(wait_lock);
195 	 * wake waiter();
196 	 * unlock(wait_lock);
197 	 *					lock(wait_lock);
198 	 *					acquire(lock);
199 	 */
200 	return rt_mutex_cmpxchg_release(lock, owner, NULL);
201 }
202 
203 #else
204 # define rt_mutex_cmpxchg_relaxed(l,c,n)	(0)
205 # define rt_mutex_cmpxchg_acquire(l,c,n)	(0)
206 # define rt_mutex_cmpxchg_release(l,c,n)	(0)
207 
208 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
209 {
210 	lock->owner = (struct task_struct *)
211 			((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
212 }
213 
214 /*
215  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
216  */
217 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
218 					unsigned long flags)
219 	__releases(lock->wait_lock)
220 {
221 	lock->owner = NULL;
222 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
223 	return true;
224 }
225 #endif
226 
227 /*
228  * Only use with rt_mutex_waiter_{less,equal}()
229  */
230 #define task_to_waiter(p)	\
231 	&(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
232 
233 static inline int
234 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
235 		     struct rt_mutex_waiter *right)
236 {
237 	if (left->prio < right->prio)
238 		return 1;
239 
240 	/*
241 	 * If both waiters have dl_prio(), we check the deadlines of the
242 	 * associated tasks.
243 	 * If left waiter has a dl_prio(), and we didn't return 1 above,
244 	 * then right waiter has a dl_prio() too.
245 	 */
246 	if (dl_prio(left->prio))
247 		return dl_time_before(left->deadline, right->deadline);
248 
249 	return 0;
250 }
251 
252 static inline int
253 rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
254 		      struct rt_mutex_waiter *right)
255 {
256 	if (left->prio != right->prio)
257 		return 0;
258 
259 	/*
260 	 * If both waiters have dl_prio(), we check the deadlines of the
261 	 * associated tasks.
262 	 * If left waiter has a dl_prio(), and we didn't return 0 above,
263 	 * then right waiter has a dl_prio() too.
264 	 */
265 	if (dl_prio(left->prio))
266 		return left->deadline == right->deadline;
267 
268 	return 1;
269 }
270 
271 static void
272 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
273 {
274 	struct rb_node **link = &lock->waiters.rb_node;
275 	struct rb_node *parent = NULL;
276 	struct rt_mutex_waiter *entry;
277 	int leftmost = 1;
278 
279 	while (*link) {
280 		parent = *link;
281 		entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
282 		if (rt_mutex_waiter_less(waiter, entry)) {
283 			link = &parent->rb_left;
284 		} else {
285 			link = &parent->rb_right;
286 			leftmost = 0;
287 		}
288 	}
289 
290 	if (leftmost)
291 		lock->waiters_leftmost = &waiter->tree_entry;
292 
293 	rb_link_node(&waiter->tree_entry, parent, link);
294 	rb_insert_color(&waiter->tree_entry, &lock->waiters);
295 }
296 
297 static void
298 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
299 {
300 	if (RB_EMPTY_NODE(&waiter->tree_entry))
301 		return;
302 
303 	if (lock->waiters_leftmost == &waiter->tree_entry)
304 		lock->waiters_leftmost = rb_next(&waiter->tree_entry);
305 
306 	rb_erase(&waiter->tree_entry, &lock->waiters);
307 	RB_CLEAR_NODE(&waiter->tree_entry);
308 }
309 
310 static void
311 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
312 {
313 	struct rb_node **link = &task->pi_waiters.rb_node;
314 	struct rb_node *parent = NULL;
315 	struct rt_mutex_waiter *entry;
316 	int leftmost = 1;
317 
318 	while (*link) {
319 		parent = *link;
320 		entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
321 		if (rt_mutex_waiter_less(waiter, entry)) {
322 			link = &parent->rb_left;
323 		} else {
324 			link = &parent->rb_right;
325 			leftmost = 0;
326 		}
327 	}
328 
329 	if (leftmost)
330 		task->pi_waiters_leftmost = &waiter->pi_tree_entry;
331 
332 	rb_link_node(&waiter->pi_tree_entry, parent, link);
333 	rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
334 }
335 
336 static void
337 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
338 {
339 	if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
340 		return;
341 
342 	if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
343 		task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
344 
345 	rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
346 	RB_CLEAR_NODE(&waiter->pi_tree_entry);
347 }
348 
349 static void rt_mutex_adjust_prio(struct task_struct *p)
350 {
351 	struct task_struct *pi_task = NULL;
352 
353 	lockdep_assert_held(&p->pi_lock);
354 
355 	if (task_has_pi_waiters(p))
356 		pi_task = task_top_pi_waiter(p)->task;
357 
358 	rt_mutex_setprio(p, pi_task);
359 }
360 
361 /*
362  * Deadlock detection is conditional:
363  *
364  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
365  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
366  *
367  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
368  * conducted independent of the detect argument.
369  *
370  * If the waiter argument is NULL this indicates the deboost path and
371  * deadlock detection is disabled independent of the detect argument
372  * and the config settings.
373  */
374 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
375 					  enum rtmutex_chainwalk chwalk)
376 {
377 	/*
378 	 * This is just a wrapper function for the following call,
379 	 * because debug_rt_mutex_detect_deadlock() smells like a magic
380 	 * debug feature and I wanted to keep the cond function in the
381 	 * main source file along with the comments instead of having
382 	 * two of the same in the headers.
383 	 */
384 	return debug_rt_mutex_detect_deadlock(waiter, chwalk);
385 }
386 
387 /*
388  * Max number of times we'll walk the boosting chain:
389  */
390 int max_lock_depth = 1024;
391 
392 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
393 {
394 	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
395 }
396 
397 /*
398  * Adjust the priority chain. Also used for deadlock detection.
399  * Decreases task's usage by one - may thus free the task.
400  *
401  * @task:	the task owning the mutex (owner) for which a chain walk is
402  *		probably needed
403  * @chwalk:	do we have to carry out deadlock detection?
404  * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
405  *		things for a task that has just got its priority adjusted, and
406  *		is waiting on a mutex)
407  * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
408  *		we dropped its pi_lock. Is never dereferenced, only used for
409  *		comparison to detect lock chain changes.
410  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
411  *		its priority to the mutex owner (can be NULL in the case
412  *		depicted above or if the top waiter is gone away and we are
413  *		actually deboosting the owner)
414  * @top_task:	the current top waiter
415  *
416  * Returns 0 or -EDEADLK.
417  *
418  * Chain walk basics and protection scope
419  *
420  * [R] refcount on task
421  * [P] task->pi_lock held
422  * [L] rtmutex->wait_lock held
423  *
424  * Step	Description				Protected by
425  *	function arguments:
426  *	@task					[R]
427  *	@orig_lock if != NULL			@top_task is blocked on it
428  *	@next_lock				Unprotected. Cannot be
429  *						dereferenced. Only used for
430  *						comparison.
431  *	@orig_waiter if != NULL			@top_task is blocked on it
432  *	@top_task				current, or in case of proxy
433  *						locking protected by calling
434  *						code
435  *	again:
436  *	  loop_sanity_check();
437  *	retry:
438  * [1]	  lock(task->pi_lock);			[R] acquire [P]
439  * [2]	  waiter = task->pi_blocked_on;		[P]
440  * [3]	  check_exit_conditions_1();		[P]
441  * [4]	  lock = waiter->lock;			[P]
442  * [5]	  if (!try_lock(lock->wait_lock)) {	[P] try to acquire [L]
443  *	    unlock(task->pi_lock);		release [P]
444  *	    goto retry;
445  *	  }
446  * [6]	  check_exit_conditions_2();		[P] + [L]
447  * [7]	  requeue_lock_waiter(lock, waiter);	[P] + [L]
448  * [8]	  unlock(task->pi_lock);		release [P]
449  *	  put_task_struct(task);		release [R]
450  * [9]	  check_exit_conditions_3();		[L]
451  * [10]	  task = owner(lock);			[L]
452  *	  get_task_struct(task);		[L] acquire [R]
453  *	  lock(task->pi_lock);			[L] acquire [P]
454  * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
455  * [12]	  check_exit_conditions_4();		[P] + [L]
456  * [13]	  unlock(task->pi_lock);		release [P]
457  *	  unlock(lock->wait_lock);		release [L]
458  *	  goto again;
459  */
460 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
461 				      enum rtmutex_chainwalk chwalk,
462 				      struct rt_mutex *orig_lock,
463 				      struct rt_mutex *next_lock,
464 				      struct rt_mutex_waiter *orig_waiter,
465 				      struct task_struct *top_task)
466 {
467 	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
468 	struct rt_mutex_waiter *prerequeue_top_waiter;
469 	int ret = 0, depth = 0;
470 	struct rt_mutex *lock;
471 	bool detect_deadlock;
472 	bool requeue = true;
473 
474 	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
475 
476 	/*
477 	 * The (de)boosting is a step by step approach with a lot of
478 	 * pitfalls. We want this to be preemptible and we want hold a
479 	 * maximum of two locks per step. So we have to check
480 	 * carefully whether things change under us.
481 	 */
482  again:
483 	/*
484 	 * We limit the lock chain length for each invocation.
485 	 */
486 	if (++depth > max_lock_depth) {
487 		static int prev_max;
488 
489 		/*
490 		 * Print this only once. If the admin changes the limit,
491 		 * print a new message when reaching the limit again.
492 		 */
493 		if (prev_max != max_lock_depth) {
494 			prev_max = max_lock_depth;
495 			printk(KERN_WARNING "Maximum lock depth %d reached "
496 			       "task: %s (%d)\n", max_lock_depth,
497 			       top_task->comm, task_pid_nr(top_task));
498 		}
499 		put_task_struct(task);
500 
501 		return -EDEADLK;
502 	}
503 
504 	/*
505 	 * We are fully preemptible here and only hold the refcount on
506 	 * @task. So everything can have changed under us since the
507 	 * caller or our own code below (goto retry/again) dropped all
508 	 * locks.
509 	 */
510  retry:
511 	/*
512 	 * [1] Task cannot go away as we did a get_task() before !
513 	 */
514 	raw_spin_lock_irq(&task->pi_lock);
515 
516 	/*
517 	 * [2] Get the waiter on which @task is blocked on.
518 	 */
519 	waiter = task->pi_blocked_on;
520 
521 	/*
522 	 * [3] check_exit_conditions_1() protected by task->pi_lock.
523 	 */
524 
525 	/*
526 	 * Check whether the end of the boosting chain has been
527 	 * reached or the state of the chain has changed while we
528 	 * dropped the locks.
529 	 */
530 	if (!waiter)
531 		goto out_unlock_pi;
532 
533 	/*
534 	 * Check the orig_waiter state. After we dropped the locks,
535 	 * the previous owner of the lock might have released the lock.
536 	 */
537 	if (orig_waiter && !rt_mutex_owner(orig_lock))
538 		goto out_unlock_pi;
539 
540 	/*
541 	 * We dropped all locks after taking a refcount on @task, so
542 	 * the task might have moved on in the lock chain or even left
543 	 * the chain completely and blocks now on an unrelated lock or
544 	 * on @orig_lock.
545 	 *
546 	 * We stored the lock on which @task was blocked in @next_lock,
547 	 * so we can detect the chain change.
548 	 */
549 	if (next_lock != waiter->lock)
550 		goto out_unlock_pi;
551 
552 	/*
553 	 * Drop out, when the task has no waiters. Note,
554 	 * top_waiter can be NULL, when we are in the deboosting
555 	 * mode!
556 	 */
557 	if (top_waiter) {
558 		if (!task_has_pi_waiters(task))
559 			goto out_unlock_pi;
560 		/*
561 		 * If deadlock detection is off, we stop here if we
562 		 * are not the top pi waiter of the task. If deadlock
563 		 * detection is enabled we continue, but stop the
564 		 * requeueing in the chain walk.
565 		 */
566 		if (top_waiter != task_top_pi_waiter(task)) {
567 			if (!detect_deadlock)
568 				goto out_unlock_pi;
569 			else
570 				requeue = false;
571 		}
572 	}
573 
574 	/*
575 	 * If the waiter priority is the same as the task priority
576 	 * then there is no further priority adjustment necessary.  If
577 	 * deadlock detection is off, we stop the chain walk. If its
578 	 * enabled we continue, but stop the requeueing in the chain
579 	 * walk.
580 	 */
581 	if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
582 		if (!detect_deadlock)
583 			goto out_unlock_pi;
584 		else
585 			requeue = false;
586 	}
587 
588 	/*
589 	 * [4] Get the next lock
590 	 */
591 	lock = waiter->lock;
592 	/*
593 	 * [5] We need to trylock here as we are holding task->pi_lock,
594 	 * which is the reverse lock order versus the other rtmutex
595 	 * operations.
596 	 */
597 	if (!raw_spin_trylock(&lock->wait_lock)) {
598 		raw_spin_unlock_irq(&task->pi_lock);
599 		cpu_relax();
600 		goto retry;
601 	}
602 
603 	/*
604 	 * [6] check_exit_conditions_2() protected by task->pi_lock and
605 	 * lock->wait_lock.
606 	 *
607 	 * Deadlock detection. If the lock is the same as the original
608 	 * lock which caused us to walk the lock chain or if the
609 	 * current lock is owned by the task which initiated the chain
610 	 * walk, we detected a deadlock.
611 	 */
612 	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
613 		debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
614 		raw_spin_unlock(&lock->wait_lock);
615 		ret = -EDEADLK;
616 		goto out_unlock_pi;
617 	}
618 
619 	/*
620 	 * If we just follow the lock chain for deadlock detection, no
621 	 * need to do all the requeue operations. To avoid a truckload
622 	 * of conditionals around the various places below, just do the
623 	 * minimum chain walk checks.
624 	 */
625 	if (!requeue) {
626 		/*
627 		 * No requeue[7] here. Just release @task [8]
628 		 */
629 		raw_spin_unlock(&task->pi_lock);
630 		put_task_struct(task);
631 
632 		/*
633 		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
634 		 * If there is no owner of the lock, end of chain.
635 		 */
636 		if (!rt_mutex_owner(lock)) {
637 			raw_spin_unlock_irq(&lock->wait_lock);
638 			return 0;
639 		}
640 
641 		/* [10] Grab the next task, i.e. owner of @lock */
642 		task = rt_mutex_owner(lock);
643 		get_task_struct(task);
644 		raw_spin_lock(&task->pi_lock);
645 
646 		/*
647 		 * No requeue [11] here. We just do deadlock detection.
648 		 *
649 		 * [12] Store whether owner is blocked
650 		 * itself. Decision is made after dropping the locks
651 		 */
652 		next_lock = task_blocked_on_lock(task);
653 		/*
654 		 * Get the top waiter for the next iteration
655 		 */
656 		top_waiter = rt_mutex_top_waiter(lock);
657 
658 		/* [13] Drop locks */
659 		raw_spin_unlock(&task->pi_lock);
660 		raw_spin_unlock_irq(&lock->wait_lock);
661 
662 		/* If owner is not blocked, end of chain. */
663 		if (!next_lock)
664 			goto out_put_task;
665 		goto again;
666 	}
667 
668 	/*
669 	 * Store the current top waiter before doing the requeue
670 	 * operation on @lock. We need it for the boost/deboost
671 	 * decision below.
672 	 */
673 	prerequeue_top_waiter = rt_mutex_top_waiter(lock);
674 
675 	/* [7] Requeue the waiter in the lock waiter tree. */
676 	rt_mutex_dequeue(lock, waiter);
677 
678 	/*
679 	 * Update the waiter prio fields now that we're dequeued.
680 	 *
681 	 * These values can have changed through either:
682 	 *
683 	 *   sys_sched_set_scheduler() / sys_sched_setattr()
684 	 *
685 	 * or
686 	 *
687 	 *   DL CBS enforcement advancing the effective deadline.
688 	 *
689 	 * Even though pi_waiters also uses these fields, and that tree is only
690 	 * updated in [11], we can do this here, since we hold [L], which
691 	 * serializes all pi_waiters access and rb_erase() does not care about
692 	 * the values of the node being removed.
693 	 */
694 	waiter->prio = task->prio;
695 	waiter->deadline = task->dl.deadline;
696 
697 	rt_mutex_enqueue(lock, waiter);
698 
699 	/* [8] Release the task */
700 	raw_spin_unlock(&task->pi_lock);
701 	put_task_struct(task);
702 
703 	/*
704 	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
705 	 *
706 	 * We must abort the chain walk if there is no lock owner even
707 	 * in the dead lock detection case, as we have nothing to
708 	 * follow here. This is the end of the chain we are walking.
709 	 */
710 	if (!rt_mutex_owner(lock)) {
711 		/*
712 		 * If the requeue [7] above changed the top waiter,
713 		 * then we need to wake the new top waiter up to try
714 		 * to get the lock.
715 		 */
716 		if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
717 			wake_up_process(rt_mutex_top_waiter(lock)->task);
718 		raw_spin_unlock_irq(&lock->wait_lock);
719 		return 0;
720 	}
721 
722 	/* [10] Grab the next task, i.e. the owner of @lock */
723 	task = rt_mutex_owner(lock);
724 	get_task_struct(task);
725 	raw_spin_lock(&task->pi_lock);
726 
727 	/* [11] requeue the pi waiters if necessary */
728 	if (waiter == rt_mutex_top_waiter(lock)) {
729 		/*
730 		 * The waiter became the new top (highest priority)
731 		 * waiter on the lock. Replace the previous top waiter
732 		 * in the owner tasks pi waiters tree with this waiter
733 		 * and adjust the priority of the owner.
734 		 */
735 		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
736 		rt_mutex_enqueue_pi(task, waiter);
737 		rt_mutex_adjust_prio(task);
738 
739 	} else if (prerequeue_top_waiter == waiter) {
740 		/*
741 		 * The waiter was the top waiter on the lock, but is
742 		 * no longer the top prority waiter. Replace waiter in
743 		 * the owner tasks pi waiters tree with the new top
744 		 * (highest priority) waiter and adjust the priority
745 		 * of the owner.
746 		 * The new top waiter is stored in @waiter so that
747 		 * @waiter == @top_waiter evaluates to true below and
748 		 * we continue to deboost the rest of the chain.
749 		 */
750 		rt_mutex_dequeue_pi(task, waiter);
751 		waiter = rt_mutex_top_waiter(lock);
752 		rt_mutex_enqueue_pi(task, waiter);
753 		rt_mutex_adjust_prio(task);
754 	} else {
755 		/*
756 		 * Nothing changed. No need to do any priority
757 		 * adjustment.
758 		 */
759 	}
760 
761 	/*
762 	 * [12] check_exit_conditions_4() protected by task->pi_lock
763 	 * and lock->wait_lock. The actual decisions are made after we
764 	 * dropped the locks.
765 	 *
766 	 * Check whether the task which owns the current lock is pi
767 	 * blocked itself. If yes we store a pointer to the lock for
768 	 * the lock chain change detection above. After we dropped
769 	 * task->pi_lock next_lock cannot be dereferenced anymore.
770 	 */
771 	next_lock = task_blocked_on_lock(task);
772 	/*
773 	 * Store the top waiter of @lock for the end of chain walk
774 	 * decision below.
775 	 */
776 	top_waiter = rt_mutex_top_waiter(lock);
777 
778 	/* [13] Drop the locks */
779 	raw_spin_unlock(&task->pi_lock);
780 	raw_spin_unlock_irq(&lock->wait_lock);
781 
782 	/*
783 	 * Make the actual exit decisions [12], based on the stored
784 	 * values.
785 	 *
786 	 * We reached the end of the lock chain. Stop right here. No
787 	 * point to go back just to figure that out.
788 	 */
789 	if (!next_lock)
790 		goto out_put_task;
791 
792 	/*
793 	 * If the current waiter is not the top waiter on the lock,
794 	 * then we can stop the chain walk here if we are not in full
795 	 * deadlock detection mode.
796 	 */
797 	if (!detect_deadlock && waiter != top_waiter)
798 		goto out_put_task;
799 
800 	goto again;
801 
802  out_unlock_pi:
803 	raw_spin_unlock_irq(&task->pi_lock);
804  out_put_task:
805 	put_task_struct(task);
806 
807 	return ret;
808 }
809 
810 /*
811  * Try to take an rt-mutex
812  *
813  * Must be called with lock->wait_lock held and interrupts disabled
814  *
815  * @lock:   The lock to be acquired.
816  * @task:   The task which wants to acquire the lock
817  * @waiter: The waiter that is queued to the lock's wait tree if the
818  *	    callsite called task_blocked_on_lock(), otherwise NULL
819  */
820 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
821 				struct rt_mutex_waiter *waiter)
822 {
823 	lockdep_assert_held(&lock->wait_lock);
824 
825 	/*
826 	 * Before testing whether we can acquire @lock, we set the
827 	 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
828 	 * other tasks which try to modify @lock into the slow path
829 	 * and they serialize on @lock->wait_lock.
830 	 *
831 	 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
832 	 * as explained at the top of this file if and only if:
833 	 *
834 	 * - There is a lock owner. The caller must fixup the
835 	 *   transient state if it does a trylock or leaves the lock
836 	 *   function due to a signal or timeout.
837 	 *
838 	 * - @task acquires the lock and there are no other
839 	 *   waiters. This is undone in rt_mutex_set_owner(@task) at
840 	 *   the end of this function.
841 	 */
842 	mark_rt_mutex_waiters(lock);
843 
844 	/*
845 	 * If @lock has an owner, give up.
846 	 */
847 	if (rt_mutex_owner(lock))
848 		return 0;
849 
850 	/*
851 	 * If @waiter != NULL, @task has already enqueued the waiter
852 	 * into @lock waiter tree. If @waiter == NULL then this is a
853 	 * trylock attempt.
854 	 */
855 	if (waiter) {
856 		/*
857 		 * If waiter is not the highest priority waiter of
858 		 * @lock, give up.
859 		 */
860 		if (waiter != rt_mutex_top_waiter(lock))
861 			return 0;
862 
863 		/*
864 		 * We can acquire the lock. Remove the waiter from the
865 		 * lock waiters tree.
866 		 */
867 		rt_mutex_dequeue(lock, waiter);
868 
869 	} else {
870 		/*
871 		 * If the lock has waiters already we check whether @task is
872 		 * eligible to take over the lock.
873 		 *
874 		 * If there are no other waiters, @task can acquire
875 		 * the lock.  @task->pi_blocked_on is NULL, so it does
876 		 * not need to be dequeued.
877 		 */
878 		if (rt_mutex_has_waiters(lock)) {
879 			/*
880 			 * If @task->prio is greater than or equal to
881 			 * the top waiter priority (kernel view),
882 			 * @task lost.
883 			 */
884 			if (!rt_mutex_waiter_less(task_to_waiter(task),
885 						  rt_mutex_top_waiter(lock)))
886 				return 0;
887 
888 			/*
889 			 * The current top waiter stays enqueued. We
890 			 * don't have to change anything in the lock
891 			 * waiters order.
892 			 */
893 		} else {
894 			/*
895 			 * No waiters. Take the lock without the
896 			 * pi_lock dance.@task->pi_blocked_on is NULL
897 			 * and we have no waiters to enqueue in @task
898 			 * pi waiters tree.
899 			 */
900 			goto takeit;
901 		}
902 	}
903 
904 	/*
905 	 * Clear @task->pi_blocked_on. Requires protection by
906 	 * @task->pi_lock. Redundant operation for the @waiter == NULL
907 	 * case, but conditionals are more expensive than a redundant
908 	 * store.
909 	 */
910 	raw_spin_lock(&task->pi_lock);
911 	task->pi_blocked_on = NULL;
912 	/*
913 	 * Finish the lock acquisition. @task is the new owner. If
914 	 * other waiters exist we have to insert the highest priority
915 	 * waiter into @task->pi_waiters tree.
916 	 */
917 	if (rt_mutex_has_waiters(lock))
918 		rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
919 	raw_spin_unlock(&task->pi_lock);
920 
921 takeit:
922 	/* We got the lock. */
923 	debug_rt_mutex_lock(lock);
924 
925 	/*
926 	 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
927 	 * are still waiters or clears it.
928 	 */
929 	rt_mutex_set_owner(lock, task);
930 
931 	return 1;
932 }
933 
934 /*
935  * Task blocks on lock.
936  *
937  * Prepare waiter and propagate pi chain
938  *
939  * This must be called with lock->wait_lock held and interrupts disabled
940  */
941 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
942 				   struct rt_mutex_waiter *waiter,
943 				   struct task_struct *task,
944 				   enum rtmutex_chainwalk chwalk)
945 {
946 	struct task_struct *owner = rt_mutex_owner(lock);
947 	struct rt_mutex_waiter *top_waiter = waiter;
948 	struct rt_mutex *next_lock;
949 	int chain_walk = 0, res;
950 
951 	lockdep_assert_held(&lock->wait_lock);
952 
953 	/*
954 	 * Early deadlock detection. We really don't want the task to
955 	 * enqueue on itself just to untangle the mess later. It's not
956 	 * only an optimization. We drop the locks, so another waiter
957 	 * can come in before the chain walk detects the deadlock. So
958 	 * the other will detect the deadlock and return -EDEADLOCK,
959 	 * which is wrong, as the other waiter is not in a deadlock
960 	 * situation.
961 	 */
962 	if (owner == task)
963 		return -EDEADLK;
964 
965 	raw_spin_lock(&task->pi_lock);
966 	rt_mutex_adjust_prio(task);
967 	waiter->task = task;
968 	waiter->lock = lock;
969 	waiter->prio = task->prio;
970 	waiter->deadline = task->dl.deadline;
971 
972 	/* Get the top priority waiter on the lock */
973 	if (rt_mutex_has_waiters(lock))
974 		top_waiter = rt_mutex_top_waiter(lock);
975 	rt_mutex_enqueue(lock, waiter);
976 
977 	task->pi_blocked_on = waiter;
978 
979 	raw_spin_unlock(&task->pi_lock);
980 
981 	if (!owner)
982 		return 0;
983 
984 	raw_spin_lock(&owner->pi_lock);
985 	if (waiter == rt_mutex_top_waiter(lock)) {
986 		rt_mutex_dequeue_pi(owner, top_waiter);
987 		rt_mutex_enqueue_pi(owner, waiter);
988 
989 		rt_mutex_adjust_prio(owner);
990 		if (owner->pi_blocked_on)
991 			chain_walk = 1;
992 	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
993 		chain_walk = 1;
994 	}
995 
996 	/* Store the lock on which owner is blocked or NULL */
997 	next_lock = task_blocked_on_lock(owner);
998 
999 	raw_spin_unlock(&owner->pi_lock);
1000 	/*
1001 	 * Even if full deadlock detection is on, if the owner is not
1002 	 * blocked itself, we can avoid finding this out in the chain
1003 	 * walk.
1004 	 */
1005 	if (!chain_walk || !next_lock)
1006 		return 0;
1007 
1008 	/*
1009 	 * The owner can't disappear while holding a lock,
1010 	 * so the owner struct is protected by wait_lock.
1011 	 * Gets dropped in rt_mutex_adjust_prio_chain()!
1012 	 */
1013 	get_task_struct(owner);
1014 
1015 	raw_spin_unlock_irq(&lock->wait_lock);
1016 
1017 	res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1018 					 next_lock, waiter, task);
1019 
1020 	raw_spin_lock_irq(&lock->wait_lock);
1021 
1022 	return res;
1023 }
1024 
1025 /*
1026  * Remove the top waiter from the current tasks pi waiter tree and
1027  * queue it up.
1028  *
1029  * Called with lock->wait_lock held and interrupts disabled.
1030  */
1031 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1032 				    struct rt_mutex *lock)
1033 {
1034 	struct rt_mutex_waiter *waiter;
1035 
1036 	raw_spin_lock(&current->pi_lock);
1037 
1038 	waiter = rt_mutex_top_waiter(lock);
1039 
1040 	/*
1041 	 * Remove it from current->pi_waiters and deboost.
1042 	 *
1043 	 * We must in fact deboost here in order to ensure we call
1044 	 * rt_mutex_setprio() to update p->pi_top_task before the
1045 	 * task unblocks.
1046 	 */
1047 	rt_mutex_dequeue_pi(current, waiter);
1048 	rt_mutex_adjust_prio(current);
1049 
1050 	/*
1051 	 * As we are waking up the top waiter, and the waiter stays
1052 	 * queued on the lock until it gets the lock, this lock
1053 	 * obviously has waiters. Just set the bit here and this has
1054 	 * the added benefit of forcing all new tasks into the
1055 	 * slow path making sure no task of lower priority than
1056 	 * the top waiter can steal this lock.
1057 	 */
1058 	lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1059 
1060 	/*
1061 	 * We deboosted before waking the top waiter task such that we don't
1062 	 * run two tasks with the 'same' priority (and ensure the
1063 	 * p->pi_top_task pointer points to a blocked task). This however can
1064 	 * lead to priority inversion if we would get preempted after the
1065 	 * deboost but before waking our donor task, hence the preempt_disable()
1066 	 * before unlock.
1067 	 *
1068 	 * Pairs with preempt_enable() in rt_mutex_postunlock();
1069 	 */
1070 	preempt_disable();
1071 	wake_q_add(wake_q, waiter->task);
1072 	raw_spin_unlock(&current->pi_lock);
1073 }
1074 
1075 /*
1076  * Remove a waiter from a lock and give up
1077  *
1078  * Must be called with lock->wait_lock held and interrupts disabled. I must
1079  * have just failed to try_to_take_rt_mutex().
1080  */
1081 static void remove_waiter(struct rt_mutex *lock,
1082 			  struct rt_mutex_waiter *waiter)
1083 {
1084 	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1085 	struct task_struct *owner = rt_mutex_owner(lock);
1086 	struct rt_mutex *next_lock;
1087 
1088 	lockdep_assert_held(&lock->wait_lock);
1089 
1090 	raw_spin_lock(&current->pi_lock);
1091 	rt_mutex_dequeue(lock, waiter);
1092 	current->pi_blocked_on = NULL;
1093 	raw_spin_unlock(&current->pi_lock);
1094 
1095 	/*
1096 	 * Only update priority if the waiter was the highest priority
1097 	 * waiter of the lock and there is an owner to update.
1098 	 */
1099 	if (!owner || !is_top_waiter)
1100 		return;
1101 
1102 	raw_spin_lock(&owner->pi_lock);
1103 
1104 	rt_mutex_dequeue_pi(owner, waiter);
1105 
1106 	if (rt_mutex_has_waiters(lock))
1107 		rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1108 
1109 	rt_mutex_adjust_prio(owner);
1110 
1111 	/* Store the lock on which owner is blocked or NULL */
1112 	next_lock = task_blocked_on_lock(owner);
1113 
1114 	raw_spin_unlock(&owner->pi_lock);
1115 
1116 	/*
1117 	 * Don't walk the chain, if the owner task is not blocked
1118 	 * itself.
1119 	 */
1120 	if (!next_lock)
1121 		return;
1122 
1123 	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1124 	get_task_struct(owner);
1125 
1126 	raw_spin_unlock_irq(&lock->wait_lock);
1127 
1128 	rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1129 				   next_lock, NULL, current);
1130 
1131 	raw_spin_lock_irq(&lock->wait_lock);
1132 }
1133 
1134 /*
1135  * Recheck the pi chain, in case we got a priority setting
1136  *
1137  * Called from sched_setscheduler
1138  */
1139 void rt_mutex_adjust_pi(struct task_struct *task)
1140 {
1141 	struct rt_mutex_waiter *waiter;
1142 	struct rt_mutex *next_lock;
1143 	unsigned long flags;
1144 
1145 	raw_spin_lock_irqsave(&task->pi_lock, flags);
1146 
1147 	waiter = task->pi_blocked_on;
1148 	if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1149 		raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1150 		return;
1151 	}
1152 	next_lock = waiter->lock;
1153 	raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1154 
1155 	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1156 	get_task_struct(task);
1157 
1158 	rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1159 				   next_lock, NULL, task);
1160 }
1161 
1162 void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1163 {
1164 	debug_rt_mutex_init_waiter(waiter);
1165 	RB_CLEAR_NODE(&waiter->pi_tree_entry);
1166 	RB_CLEAR_NODE(&waiter->tree_entry);
1167 	waiter->task = NULL;
1168 }
1169 
1170 /**
1171  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1172  * @lock:		 the rt_mutex to take
1173  * @state:		 the state the task should block in (TASK_INTERRUPTIBLE
1174  *			 or TASK_UNINTERRUPTIBLE)
1175  * @timeout:		 the pre-initialized and started timer, or NULL for none
1176  * @waiter:		 the pre-initialized rt_mutex_waiter
1177  *
1178  * Must be called with lock->wait_lock held and interrupts disabled
1179  */
1180 static int __sched
1181 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1182 		    struct hrtimer_sleeper *timeout,
1183 		    struct rt_mutex_waiter *waiter)
1184 {
1185 	int ret = 0;
1186 
1187 	for (;;) {
1188 		/* Try to acquire the lock: */
1189 		if (try_to_take_rt_mutex(lock, current, waiter))
1190 			break;
1191 
1192 		/*
1193 		 * TASK_INTERRUPTIBLE checks for signals and
1194 		 * timeout. Ignored otherwise.
1195 		 */
1196 		if (likely(state == TASK_INTERRUPTIBLE)) {
1197 			/* Signal pending? */
1198 			if (signal_pending(current))
1199 				ret = -EINTR;
1200 			if (timeout && !timeout->task)
1201 				ret = -ETIMEDOUT;
1202 			if (ret)
1203 				break;
1204 		}
1205 
1206 		raw_spin_unlock_irq(&lock->wait_lock);
1207 
1208 		debug_rt_mutex_print_deadlock(waiter);
1209 
1210 		schedule();
1211 
1212 		raw_spin_lock_irq(&lock->wait_lock);
1213 		set_current_state(state);
1214 	}
1215 
1216 	__set_current_state(TASK_RUNNING);
1217 	return ret;
1218 }
1219 
1220 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1221 				     struct rt_mutex_waiter *w)
1222 {
1223 	/*
1224 	 * If the result is not -EDEADLOCK or the caller requested
1225 	 * deadlock detection, nothing to do here.
1226 	 */
1227 	if (res != -EDEADLOCK || detect_deadlock)
1228 		return;
1229 
1230 	/*
1231 	 * Yell lowdly and stop the task right here.
1232 	 */
1233 	rt_mutex_print_deadlock(w);
1234 	while (1) {
1235 		set_current_state(TASK_INTERRUPTIBLE);
1236 		schedule();
1237 	}
1238 }
1239 
1240 /*
1241  * Slow path lock function:
1242  */
1243 static int __sched
1244 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1245 		  struct hrtimer_sleeper *timeout,
1246 		  enum rtmutex_chainwalk chwalk)
1247 {
1248 	struct rt_mutex_waiter waiter;
1249 	unsigned long flags;
1250 	int ret = 0;
1251 
1252 	rt_mutex_init_waiter(&waiter);
1253 
1254 	/*
1255 	 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1256 	 * be called in early boot if the cmpxchg() fast path is disabled
1257 	 * (debug, no architecture support). In this case we will acquire the
1258 	 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1259 	 * enable interrupts in that early boot case. So we need to use the
1260 	 * irqsave/restore variants.
1261 	 */
1262 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1263 
1264 	/* Try to acquire the lock again: */
1265 	if (try_to_take_rt_mutex(lock, current, NULL)) {
1266 		raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1267 		return 0;
1268 	}
1269 
1270 	set_current_state(state);
1271 
1272 	/* Setup the timer, when timeout != NULL */
1273 	if (unlikely(timeout))
1274 		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1275 
1276 	ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1277 
1278 	if (likely(!ret))
1279 		/* sleep on the mutex */
1280 		ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1281 
1282 	if (unlikely(ret)) {
1283 		__set_current_state(TASK_RUNNING);
1284 		if (rt_mutex_has_waiters(lock))
1285 			remove_waiter(lock, &waiter);
1286 		rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1287 	}
1288 
1289 	/*
1290 	 * try_to_take_rt_mutex() sets the waiter bit
1291 	 * unconditionally. We might have to fix that up.
1292 	 */
1293 	fixup_rt_mutex_waiters(lock);
1294 
1295 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1296 
1297 	/* Remove pending timer: */
1298 	if (unlikely(timeout))
1299 		hrtimer_cancel(&timeout->timer);
1300 
1301 	debug_rt_mutex_free_waiter(&waiter);
1302 
1303 	return ret;
1304 }
1305 
1306 /*
1307  * Slow path try-lock function:
1308  */
1309 static 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 = try_to_take_rt_mutex(lock, current, NULL);
1329 
1330 	/*
1331 	 * try_to_take_rt_mutex() sets the lock waiters bit
1332 	 * unconditionally. Clean this up.
1333 	 */
1334 	fixup_rt_mutex_waiters(lock);
1335 
1336 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1337 
1338 	return ret;
1339 }
1340 
1341 /*
1342  * Slow path to release a rt-mutex.
1343  *
1344  * Return whether the current task needs to call rt_mutex_postunlock().
1345  */
1346 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1347 					struct wake_q_head *wake_q)
1348 {
1349 	unsigned long flags;
1350 
1351 	/* irqsave required to support early boot calls */
1352 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1353 
1354 	debug_rt_mutex_unlock(lock);
1355 
1356 	/*
1357 	 * We must be careful here if the fast path is enabled. If we
1358 	 * have no waiters queued we cannot set owner to NULL here
1359 	 * because of:
1360 	 *
1361 	 * foo->lock->owner = NULL;
1362 	 *			rtmutex_lock(foo->lock);   <- fast path
1363 	 *			free = atomic_dec_and_test(foo->refcnt);
1364 	 *			rtmutex_unlock(foo->lock); <- fast path
1365 	 *			if (free)
1366 	 *				kfree(foo);
1367 	 * raw_spin_unlock(foo->lock->wait_lock);
1368 	 *
1369 	 * So for the fastpath enabled kernel:
1370 	 *
1371 	 * Nothing can set the waiters bit as long as we hold
1372 	 * lock->wait_lock. So we do the following sequence:
1373 	 *
1374 	 *	owner = rt_mutex_owner(lock);
1375 	 *	clear_rt_mutex_waiters(lock);
1376 	 *	raw_spin_unlock(&lock->wait_lock);
1377 	 *	if (cmpxchg(&lock->owner, owner, 0) == owner)
1378 	 *		return;
1379 	 *	goto retry;
1380 	 *
1381 	 * The fastpath disabled variant is simple as all access to
1382 	 * lock->owner is serialized by lock->wait_lock:
1383 	 *
1384 	 *	lock->owner = NULL;
1385 	 *	raw_spin_unlock(&lock->wait_lock);
1386 	 */
1387 	while (!rt_mutex_has_waiters(lock)) {
1388 		/* Drops lock->wait_lock ! */
1389 		if (unlock_rt_mutex_safe(lock, flags) == true)
1390 			return false;
1391 		/* Relock the rtmutex and try again */
1392 		raw_spin_lock_irqsave(&lock->wait_lock, flags);
1393 	}
1394 
1395 	/*
1396 	 * The wakeup next waiter path does not suffer from the above
1397 	 * race. See the comments there.
1398 	 *
1399 	 * Queue the next waiter for wakeup once we release the wait_lock.
1400 	 */
1401 	mark_wakeup_next_waiter(wake_q, lock);
1402 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1403 
1404 	return true; /* call rt_mutex_postunlock() */
1405 }
1406 
1407 /*
1408  * debug aware fast / slowpath lock,trylock,unlock
1409  *
1410  * The atomic acquire/release ops are compiled away, when either the
1411  * architecture does not support cmpxchg or when debugging is enabled.
1412  */
1413 static inline int
1414 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1415 		  int (*slowfn)(struct rt_mutex *lock, int state,
1416 				struct hrtimer_sleeper *timeout,
1417 				enum rtmutex_chainwalk chwalk))
1418 {
1419 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1420 		return 0;
1421 
1422 	return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1423 }
1424 
1425 static inline int
1426 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1427 			struct hrtimer_sleeper *timeout,
1428 			enum rtmutex_chainwalk chwalk,
1429 			int (*slowfn)(struct rt_mutex *lock, int state,
1430 				      struct hrtimer_sleeper *timeout,
1431 				      enum rtmutex_chainwalk chwalk))
1432 {
1433 	if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1434 	    likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1435 		return 0;
1436 
1437 	return slowfn(lock, state, timeout, chwalk);
1438 }
1439 
1440 static inline int
1441 rt_mutex_fasttrylock(struct rt_mutex *lock,
1442 		     int (*slowfn)(struct rt_mutex *lock))
1443 {
1444 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1445 		return 1;
1446 
1447 	return slowfn(lock);
1448 }
1449 
1450 /*
1451  * Performs the wakeup of the the top-waiter and re-enables preemption.
1452  */
1453 void rt_mutex_postunlock(struct wake_q_head *wake_q)
1454 {
1455 	wake_up_q(wake_q);
1456 
1457 	/* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1458 	preempt_enable();
1459 }
1460 
1461 static inline void
1462 rt_mutex_fastunlock(struct rt_mutex *lock,
1463 		    bool (*slowfn)(struct rt_mutex *lock,
1464 				   struct wake_q_head *wqh))
1465 {
1466 	DEFINE_WAKE_Q(wake_q);
1467 
1468 	if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1469 		return;
1470 
1471 	if (slowfn(lock, &wake_q))
1472 		rt_mutex_postunlock(&wake_q);
1473 }
1474 
1475 /**
1476  * rt_mutex_lock - lock a rt_mutex
1477  *
1478  * @lock: the rt_mutex to be locked
1479  */
1480 void __sched rt_mutex_lock(struct rt_mutex *lock)
1481 {
1482 	might_sleep();
1483 
1484 	mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1485 	rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1486 }
1487 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1488 
1489 /**
1490  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1491  *
1492  * @lock:		the rt_mutex to be locked
1493  *
1494  * Returns:
1495  *  0		on success
1496  * -EINTR	when interrupted by a signal
1497  */
1498 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1499 {
1500 	int ret;
1501 
1502 	might_sleep();
1503 
1504 	mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1505 	ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1506 	if (ret)
1507 		mutex_release(&lock->dep_map, 1, _RET_IP_);
1508 
1509 	return ret;
1510 }
1511 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1512 
1513 /*
1514  * Futex variant, must not use fastpath.
1515  */
1516 int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1517 {
1518 	return rt_mutex_slowtrylock(lock);
1519 }
1520 
1521 /**
1522  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1523  *			the timeout structure is provided
1524  *			by the caller
1525  *
1526  * @lock:		the rt_mutex to be locked
1527  * @timeout:		timeout structure or NULL (no timeout)
1528  *
1529  * Returns:
1530  *  0		on success
1531  * -EINTR	when interrupted by a signal
1532  * -ETIMEDOUT	when the timeout expired
1533  */
1534 int
1535 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1536 {
1537 	int ret;
1538 
1539 	might_sleep();
1540 
1541 	mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1542 	ret = rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1543 				       RT_MUTEX_MIN_CHAINWALK,
1544 				       rt_mutex_slowlock);
1545 	if (ret)
1546 		mutex_release(&lock->dep_map, 1, _RET_IP_);
1547 
1548 	return ret;
1549 }
1550 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1551 
1552 /**
1553  * rt_mutex_trylock - try to lock a rt_mutex
1554  *
1555  * @lock:	the rt_mutex to be locked
1556  *
1557  * This function can only be called in thread context. It's safe to
1558  * call it from atomic regions, but not from hard interrupt or soft
1559  * interrupt context.
1560  *
1561  * Returns 1 on success and 0 on contention
1562  */
1563 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1564 {
1565 	int ret;
1566 
1567 	if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1568 		return 0;
1569 
1570 	ret = rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1571 	if (ret)
1572 		mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1573 
1574 	return ret;
1575 }
1576 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1577 
1578 /**
1579  * rt_mutex_unlock - unlock a rt_mutex
1580  *
1581  * @lock: the rt_mutex to be unlocked
1582  */
1583 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1584 {
1585 	mutex_release(&lock->dep_map, 1, _RET_IP_);
1586 	rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1587 }
1588 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1589 
1590 /**
1591  * Futex variant, that since futex variants do not use the fast-path, can be
1592  * simple and will not need to retry.
1593  */
1594 bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1595 				    struct wake_q_head *wake_q)
1596 {
1597 	lockdep_assert_held(&lock->wait_lock);
1598 
1599 	debug_rt_mutex_unlock(lock);
1600 
1601 	if (!rt_mutex_has_waiters(lock)) {
1602 		lock->owner = NULL;
1603 		return false; /* done */
1604 	}
1605 
1606 	/*
1607 	 * We've already deboosted, mark_wakeup_next_waiter() will
1608 	 * retain preempt_disabled when we drop the wait_lock, to
1609 	 * avoid inversion prior to the wakeup.  preempt_disable()
1610 	 * therein pairs with rt_mutex_postunlock().
1611 	 */
1612 	mark_wakeup_next_waiter(wake_q, lock);
1613 
1614 	return true; /* call postunlock() */
1615 }
1616 
1617 void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1618 {
1619 	DEFINE_WAKE_Q(wake_q);
1620 	bool postunlock;
1621 
1622 	raw_spin_lock_irq(&lock->wait_lock);
1623 	postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1624 	raw_spin_unlock_irq(&lock->wait_lock);
1625 
1626 	if (postunlock)
1627 		rt_mutex_postunlock(&wake_q);
1628 }
1629 
1630 /**
1631  * rt_mutex_destroy - mark a mutex unusable
1632  * @lock: the mutex to be destroyed
1633  *
1634  * This function marks the mutex uninitialized, and any subsequent
1635  * use of the mutex is forbidden. The mutex must not be locked when
1636  * this function is called.
1637  */
1638 void rt_mutex_destroy(struct rt_mutex *lock)
1639 {
1640 	WARN_ON(rt_mutex_is_locked(lock));
1641 #ifdef CONFIG_DEBUG_RT_MUTEXES
1642 	lock->magic = NULL;
1643 #endif
1644 }
1645 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1646 
1647 /**
1648  * __rt_mutex_init - initialize the rt lock
1649  *
1650  * @lock: the rt lock to be initialized
1651  *
1652  * Initialize the rt lock to unlocked state.
1653  *
1654  * Initializing of a locked rt lock is not allowed
1655  */
1656 void __rt_mutex_init(struct rt_mutex *lock, const char *name,
1657 		     struct lock_class_key *key)
1658 {
1659 	lock->owner = NULL;
1660 	raw_spin_lock_init(&lock->wait_lock);
1661 	lock->waiters = RB_ROOT;
1662 	lock->waiters_leftmost = NULL;
1663 
1664 	if (name && key)
1665 		debug_rt_mutex_init(lock, name, key);
1666 }
1667 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1668 
1669 /**
1670  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1671  *				proxy owner
1672  *
1673  * @lock:	the rt_mutex to be locked
1674  * @proxy_owner:the task to set as owner
1675  *
1676  * No locking. Caller has to do serializing itself
1677  *
1678  * Special API call for PI-futex support. This initializes the rtmutex and
1679  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1680  * possible at this point because the pi_state which contains the rtmutex
1681  * is not yet visible to other tasks.
1682  */
1683 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1684 				struct task_struct *proxy_owner)
1685 {
1686 	__rt_mutex_init(lock, NULL, NULL);
1687 	debug_rt_mutex_proxy_lock(lock, proxy_owner);
1688 	rt_mutex_set_owner(lock, proxy_owner);
1689 }
1690 
1691 /**
1692  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1693  *
1694  * @lock:	the rt_mutex to be locked
1695  *
1696  * No locking. Caller has to do serializing itself
1697  *
1698  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1699  * (debugging) state. Concurrent operations on this rt_mutex are not
1700  * possible because it belongs to the pi_state which is about to be freed
1701  * and it is not longer visible to other tasks.
1702  */
1703 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1704 			   struct task_struct *proxy_owner)
1705 {
1706 	debug_rt_mutex_proxy_unlock(lock);
1707 	rt_mutex_set_owner(lock, NULL);
1708 }
1709 
1710 int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1711 			      struct rt_mutex_waiter *waiter,
1712 			      struct task_struct *task)
1713 {
1714 	int ret;
1715 
1716 	if (try_to_take_rt_mutex(lock, task, NULL))
1717 		return 1;
1718 
1719 	/* We enforce deadlock detection for futexes */
1720 	ret = task_blocks_on_rt_mutex(lock, waiter, task,
1721 				      RT_MUTEX_FULL_CHAINWALK);
1722 
1723 	if (ret && !rt_mutex_owner(lock)) {
1724 		/*
1725 		 * Reset the return value. We might have
1726 		 * returned with -EDEADLK and the owner
1727 		 * released the lock while we were walking the
1728 		 * pi chain.  Let the waiter sort it out.
1729 		 */
1730 		ret = 0;
1731 	}
1732 
1733 	if (unlikely(ret))
1734 		remove_waiter(lock, waiter);
1735 
1736 	debug_rt_mutex_print_deadlock(waiter);
1737 
1738 	return ret;
1739 }
1740 
1741 /**
1742  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1743  * @lock:		the rt_mutex to take
1744  * @waiter:		the pre-initialized rt_mutex_waiter
1745  * @task:		the task to prepare
1746  *
1747  * Returns:
1748  *  0 - task blocked on lock
1749  *  1 - acquired the lock for task, caller should wake it up
1750  * <0 - error
1751  *
1752  * Special API call for FUTEX_REQUEUE_PI support.
1753  */
1754 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1755 			      struct rt_mutex_waiter *waiter,
1756 			      struct task_struct *task)
1757 {
1758 	int ret;
1759 
1760 	raw_spin_lock_irq(&lock->wait_lock);
1761 	ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1762 	raw_spin_unlock_irq(&lock->wait_lock);
1763 
1764 	return ret;
1765 }
1766 
1767 /**
1768  * rt_mutex_next_owner - return the next owner of the lock
1769  *
1770  * @lock: the rt lock query
1771  *
1772  * Returns the next owner of the lock or NULL
1773  *
1774  * Caller has to serialize against other accessors to the lock
1775  * itself.
1776  *
1777  * Special API call for PI-futex support
1778  */
1779 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1780 {
1781 	if (!rt_mutex_has_waiters(lock))
1782 		return NULL;
1783 
1784 	return rt_mutex_top_waiter(lock)->task;
1785 }
1786 
1787 /**
1788  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1789  * @lock:		the rt_mutex we were woken on
1790  * @to:			the timeout, null if none. hrtimer should already have
1791  *			been started.
1792  * @waiter:		the pre-initialized rt_mutex_waiter
1793  *
1794  * Wait for the the lock acquisition started on our behalf by
1795  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1796  * rt_mutex_cleanup_proxy_lock().
1797  *
1798  * Returns:
1799  *  0 - success
1800  * <0 - error, one of -EINTR, -ETIMEDOUT
1801  *
1802  * Special API call for PI-futex support
1803  */
1804 int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1805 			       struct hrtimer_sleeper *to,
1806 			       struct rt_mutex_waiter *waiter)
1807 {
1808 	int ret;
1809 
1810 	raw_spin_lock_irq(&lock->wait_lock);
1811 	/* sleep on the mutex */
1812 	set_current_state(TASK_INTERRUPTIBLE);
1813 	ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1814 	/*
1815 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1816 	 * have to fix that up.
1817 	 */
1818 	fixup_rt_mutex_waiters(lock);
1819 	raw_spin_unlock_irq(&lock->wait_lock);
1820 
1821 	return ret;
1822 }
1823 
1824 /**
1825  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1826  * @lock:		the rt_mutex we were woken on
1827  * @waiter:		the pre-initialized rt_mutex_waiter
1828  *
1829  * Attempt to clean up after a failed rt_mutex_wait_proxy_lock().
1830  *
1831  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1832  * in fact still be granted ownership until we're removed. Therefore we can
1833  * find we are in fact the owner and must disregard the
1834  * rt_mutex_wait_proxy_lock() failure.
1835  *
1836  * Returns:
1837  *  true  - did the cleanup, we done.
1838  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1839  *          caller should disregards its return value.
1840  *
1841  * Special API call for PI-futex support
1842  */
1843 bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1844 				 struct rt_mutex_waiter *waiter)
1845 {
1846 	bool cleanup = false;
1847 
1848 	raw_spin_lock_irq(&lock->wait_lock);
1849 	/*
1850 	 * Do an unconditional try-lock, this deals with the lock stealing
1851 	 * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1852 	 * sets a NULL owner.
1853 	 *
1854 	 * We're not interested in the return value, because the subsequent
1855 	 * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1856 	 * we will own the lock and it will have removed the waiter. If we
1857 	 * failed the trylock, we're still not owner and we need to remove
1858 	 * ourselves.
1859 	 */
1860 	try_to_take_rt_mutex(lock, current, waiter);
1861 	/*
1862 	 * Unless we're the owner; we're still enqueued on the wait_list.
1863 	 * So check if we became owner, if not, take us off the wait_list.
1864 	 */
1865 	if (rt_mutex_owner(lock) != current) {
1866 		remove_waiter(lock, waiter);
1867 		cleanup = true;
1868 	}
1869 	/*
1870 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1871 	 * have to fix that up.
1872 	 */
1873 	fixup_rt_mutex_waiters(lock);
1874 
1875 	raw_spin_unlock_irq(&lock->wait_lock);
1876 
1877 	return cleanup;
1878 }
1879