xref: /openbmc/linux/kernel/locking/rtmutex.c (revision 0aa7be05)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
4  *
5  * started by Ingo Molnar and Thomas Gleixner.
6  *
7  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
8  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
9  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
10  *  Copyright (C) 2006 Esben Nielsen
11  * Adaptive Spinlocks:
12  *  Copyright (C) 2008 Novell, Inc., Gregory Haskins, Sven Dietrich,
13  *				     and Peter Morreale,
14  * Adaptive Spinlocks simplification:
15  *  Copyright (C) 2008 Red Hat, Inc., Steven Rostedt <srostedt@redhat.com>
16  *
17  *  See Documentation/locking/rt-mutex-design.rst for details.
18  */
19 #include <linux/sched.h>
20 #include <linux/sched/debug.h>
21 #include <linux/sched/deadline.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/rt.h>
24 #include <linux/sched/wake_q.h>
25 #include <linux/ww_mutex.h>
26 
27 #include <trace/events/lock.h>
28 
29 #include "rtmutex_common.h"
30 
31 #ifndef WW_RT
32 # define build_ww_mutex()	(false)
33 # define ww_container_of(rtm)	NULL
34 
35 static inline int __ww_mutex_add_waiter(struct rt_mutex_waiter *waiter,
36 					struct rt_mutex *lock,
37 					struct ww_acquire_ctx *ww_ctx)
38 {
39 	return 0;
40 }
41 
42 static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
43 					    struct ww_acquire_ctx *ww_ctx)
44 {
45 }
46 
47 static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
48 					  struct ww_acquire_ctx *ww_ctx)
49 {
50 }
51 
52 static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
53 					struct rt_mutex_waiter *waiter,
54 					struct ww_acquire_ctx *ww_ctx)
55 {
56 	return 0;
57 }
58 
59 #else
60 # define build_ww_mutex()	(true)
61 # define ww_container_of(rtm)	container_of(rtm, struct ww_mutex, base)
62 # include "ww_mutex.h"
63 #endif
64 
65 /*
66  * lock->owner state tracking:
67  *
68  * lock->owner holds the task_struct pointer of the owner. Bit 0
69  * is used to keep track of the "lock has waiters" state.
70  *
71  * owner	bit0
72  * NULL		0	lock is free (fast acquire possible)
73  * NULL		1	lock is free and has waiters and the top waiter
74  *				is going to take the lock*
75  * taskpointer	0	lock is held (fast release possible)
76  * taskpointer	1	lock is held and has waiters**
77  *
78  * The fast atomic compare exchange based acquire and release is only
79  * possible when bit 0 of lock->owner is 0.
80  *
81  * (*) It also can be a transitional state when grabbing the lock
82  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
83  * we need to set the bit0 before looking at the lock, and the owner may be
84  * NULL in this small time, hence this can be a transitional state.
85  *
86  * (**) There is a small time when bit 0 is set but there are no
87  * waiters. This can happen when grabbing the lock in the slow path.
88  * To prevent a cmpxchg of the owner releasing the lock, we need to
89  * set this bit before looking at the lock.
90  */
91 
92 static __always_inline void
93 rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
94 {
95 	unsigned long val = (unsigned long)owner;
96 
97 	if (rt_mutex_has_waiters(lock))
98 		val |= RT_MUTEX_HAS_WAITERS;
99 
100 	WRITE_ONCE(lock->owner, (struct task_struct *)val);
101 }
102 
103 static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
104 {
105 	lock->owner = (struct task_struct *)
106 			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
107 }
108 
109 static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
110 {
111 	unsigned long owner, *p = (unsigned long *) &lock->owner;
112 
113 	if (rt_mutex_has_waiters(lock))
114 		return;
115 
116 	/*
117 	 * The rbtree has no waiters enqueued, now make sure that the
118 	 * lock->owner still has the waiters bit set, otherwise the
119 	 * following can happen:
120 	 *
121 	 * CPU 0	CPU 1		CPU2
122 	 * l->owner=T1
123 	 *		rt_mutex_lock(l)
124 	 *		lock(l->lock)
125 	 *		l->owner = T1 | HAS_WAITERS;
126 	 *		enqueue(T2)
127 	 *		boost()
128 	 *		  unlock(l->lock)
129 	 *		block()
130 	 *
131 	 *				rt_mutex_lock(l)
132 	 *				lock(l->lock)
133 	 *				l->owner = T1 | HAS_WAITERS;
134 	 *				enqueue(T3)
135 	 *				boost()
136 	 *				  unlock(l->lock)
137 	 *				block()
138 	 *		signal(->T2)	signal(->T3)
139 	 *		lock(l->lock)
140 	 *		dequeue(T2)
141 	 *		deboost()
142 	 *		  unlock(l->lock)
143 	 *				lock(l->lock)
144 	 *				dequeue(T3)
145 	 *				 ==> wait list is empty
146 	 *				deboost()
147 	 *				 unlock(l->lock)
148 	 *		lock(l->lock)
149 	 *		fixup_rt_mutex_waiters()
150 	 *		  if (wait_list_empty(l) {
151 	 *		    l->owner = owner
152 	 *		    owner = l->owner & ~HAS_WAITERS;
153 	 *		      ==> l->owner = T1
154 	 *		  }
155 	 *				lock(l->lock)
156 	 * rt_mutex_unlock(l)		fixup_rt_mutex_waiters()
157 	 *				  if (wait_list_empty(l) {
158 	 *				    owner = l->owner & ~HAS_WAITERS;
159 	 * cmpxchg(l->owner, T1, NULL)
160 	 *  ===> Success (l->owner = NULL)
161 	 *
162 	 *				    l->owner = owner
163 	 *				      ==> l->owner = T1
164 	 *				  }
165 	 *
166 	 * With the check for the waiter bit in place T3 on CPU2 will not
167 	 * overwrite. All tasks fiddling with the waiters bit are
168 	 * serialized by l->lock, so nothing else can modify the waiters
169 	 * bit. If the bit is set then nothing can change l->owner either
170 	 * so the simple RMW is safe. The cmpxchg() will simply fail if it
171 	 * happens in the middle of the RMW because the waiters bit is
172 	 * still set.
173 	 */
174 	owner = READ_ONCE(*p);
175 	if (owner & RT_MUTEX_HAS_WAITERS)
176 		WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
177 }
178 
179 /*
180  * We can speed up the acquire/release, if there's no debugging state to be
181  * set up.
182  */
183 #ifndef CONFIG_DEBUG_RT_MUTEXES
184 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
185 						     struct task_struct *old,
186 						     struct task_struct *new)
187 {
188 	return try_cmpxchg_acquire(&lock->owner, &old, new);
189 }
190 
191 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
192 						     struct task_struct *old,
193 						     struct task_struct *new)
194 {
195 	return try_cmpxchg_release(&lock->owner, &old, new);
196 }
197 
198 /*
199  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
200  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
201  * relaxed semantics suffice.
202  */
203 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
204 {
205 	unsigned long owner, *p = (unsigned long *) &lock->owner;
206 
207 	do {
208 		owner = *p;
209 	} while (cmpxchg_relaxed(p, owner,
210 				 owner | RT_MUTEX_HAS_WAITERS) != owner);
211 }
212 
213 /*
214  * Safe fastpath aware unlock:
215  * 1) Clear the waiters bit
216  * 2) Drop lock->wait_lock
217  * 3) Try to unlock the lock with cmpxchg
218  */
219 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
220 						 unsigned long flags)
221 	__releases(lock->wait_lock)
222 {
223 	struct task_struct *owner = rt_mutex_owner(lock);
224 
225 	clear_rt_mutex_waiters(lock);
226 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
227 	/*
228 	 * If a new waiter comes in between the unlock and the cmpxchg
229 	 * we have two situations:
230 	 *
231 	 * unlock(wait_lock);
232 	 *					lock(wait_lock);
233 	 * cmpxchg(p, owner, 0) == owner
234 	 *					mark_rt_mutex_waiters(lock);
235 	 *					acquire(lock);
236 	 * or:
237 	 *
238 	 * unlock(wait_lock);
239 	 *					lock(wait_lock);
240 	 *					mark_rt_mutex_waiters(lock);
241 	 *
242 	 * cmpxchg(p, owner, 0) != owner
243 	 *					enqueue_waiter();
244 	 *					unlock(wait_lock);
245 	 * lock(wait_lock);
246 	 * wake waiter();
247 	 * unlock(wait_lock);
248 	 *					lock(wait_lock);
249 	 *					acquire(lock);
250 	 */
251 	return rt_mutex_cmpxchg_release(lock, owner, NULL);
252 }
253 
254 #else
255 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
256 						     struct task_struct *old,
257 						     struct task_struct *new)
258 {
259 	return false;
260 
261 }
262 
263 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
264 						     struct task_struct *old,
265 						     struct task_struct *new)
266 {
267 	return false;
268 }
269 
270 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
271 {
272 	lock->owner = (struct task_struct *)
273 			((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
274 }
275 
276 /*
277  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
278  */
279 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
280 						 unsigned long flags)
281 	__releases(lock->wait_lock)
282 {
283 	lock->owner = NULL;
284 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
285 	return true;
286 }
287 #endif
288 
289 static __always_inline int __waiter_prio(struct task_struct *task)
290 {
291 	int prio = task->prio;
292 
293 	if (!rt_prio(prio))
294 		return DEFAULT_PRIO;
295 
296 	return prio;
297 }
298 
299 static __always_inline void
300 waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
301 {
302 	waiter->prio = __waiter_prio(task);
303 	waiter->deadline = task->dl.deadline;
304 }
305 
306 /*
307  * Only use with rt_mutex_waiter_{less,equal}()
308  */
309 #define task_to_waiter(p)	\
310 	&(struct rt_mutex_waiter){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
311 
312 static __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
313 						struct rt_mutex_waiter *right)
314 {
315 	if (left->prio < right->prio)
316 		return 1;
317 
318 	/*
319 	 * If both waiters have dl_prio(), we check the deadlines of the
320 	 * associated tasks.
321 	 * If left waiter has a dl_prio(), and we didn't return 1 above,
322 	 * then right waiter has a dl_prio() too.
323 	 */
324 	if (dl_prio(left->prio))
325 		return dl_time_before(left->deadline, right->deadline);
326 
327 	return 0;
328 }
329 
330 static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
331 						 struct rt_mutex_waiter *right)
332 {
333 	if (left->prio != right->prio)
334 		return 0;
335 
336 	/*
337 	 * If both waiters have dl_prio(), we check the deadlines of the
338 	 * associated tasks.
339 	 * If left waiter has a dl_prio(), and we didn't return 0 above,
340 	 * then right waiter has a dl_prio() too.
341 	 */
342 	if (dl_prio(left->prio))
343 		return left->deadline == right->deadline;
344 
345 	return 1;
346 }
347 
348 static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
349 				  struct rt_mutex_waiter *top_waiter)
350 {
351 	if (rt_mutex_waiter_less(waiter, top_waiter))
352 		return true;
353 
354 #ifdef RT_MUTEX_BUILD_SPINLOCKS
355 	/*
356 	 * Note that RT tasks are excluded from same priority (lateral)
357 	 * steals to prevent the introduction of an unbounded latency.
358 	 */
359 	if (rt_prio(waiter->prio) || dl_prio(waiter->prio))
360 		return false;
361 
362 	return rt_mutex_waiter_equal(waiter, top_waiter);
363 #else
364 	return false;
365 #endif
366 }
367 
368 #define __node_2_waiter(node) \
369 	rb_entry((node), struct rt_mutex_waiter, tree_entry)
370 
371 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
372 {
373 	struct rt_mutex_waiter *aw = __node_2_waiter(a);
374 	struct rt_mutex_waiter *bw = __node_2_waiter(b);
375 
376 	if (rt_mutex_waiter_less(aw, bw))
377 		return 1;
378 
379 	if (!build_ww_mutex())
380 		return 0;
381 
382 	if (rt_mutex_waiter_less(bw, aw))
383 		return 0;
384 
385 	/* NOTE: relies on waiter->ww_ctx being set before insertion */
386 	if (aw->ww_ctx) {
387 		if (!bw->ww_ctx)
388 			return 1;
389 
390 		return (signed long)(aw->ww_ctx->stamp -
391 				     bw->ww_ctx->stamp) < 0;
392 	}
393 
394 	return 0;
395 }
396 
397 static __always_inline void
398 rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
399 {
400 	rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
401 }
402 
403 static __always_inline void
404 rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
405 {
406 	if (RB_EMPTY_NODE(&waiter->tree_entry))
407 		return;
408 
409 	rb_erase_cached(&waiter->tree_entry, &lock->waiters);
410 	RB_CLEAR_NODE(&waiter->tree_entry);
411 }
412 
413 #define __node_2_pi_waiter(node) \
414 	rb_entry((node), struct rt_mutex_waiter, pi_tree_entry)
415 
416 static __always_inline bool
417 __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
418 {
419 	return rt_mutex_waiter_less(__node_2_pi_waiter(a), __node_2_pi_waiter(b));
420 }
421 
422 static __always_inline void
423 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
424 {
425 	rb_add_cached(&waiter->pi_tree_entry, &task->pi_waiters, __pi_waiter_less);
426 }
427 
428 static __always_inline void
429 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
430 {
431 	if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
432 		return;
433 
434 	rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
435 	RB_CLEAR_NODE(&waiter->pi_tree_entry);
436 }
437 
438 static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
439 {
440 	struct task_struct *pi_task = NULL;
441 
442 	lockdep_assert_held(&p->pi_lock);
443 
444 	if (task_has_pi_waiters(p))
445 		pi_task = task_top_pi_waiter(p)->task;
446 
447 	rt_mutex_setprio(p, pi_task);
448 }
449 
450 /* RT mutex specific wake_q wrappers */
451 static __always_inline void rt_mutex_wake_q_add_task(struct rt_wake_q_head *wqh,
452 						     struct task_struct *task,
453 						     unsigned int wake_state)
454 {
455 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wake_state == TASK_RTLOCK_WAIT) {
456 		if (IS_ENABLED(CONFIG_PROVE_LOCKING))
457 			WARN_ON_ONCE(wqh->rtlock_task);
458 		get_task_struct(task);
459 		wqh->rtlock_task = task;
460 	} else {
461 		wake_q_add(&wqh->head, task);
462 	}
463 }
464 
465 static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
466 						struct rt_mutex_waiter *w)
467 {
468 	rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
469 }
470 
471 static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
472 {
473 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
474 		wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
475 		put_task_struct(wqh->rtlock_task);
476 		wqh->rtlock_task = NULL;
477 	}
478 
479 	if (!wake_q_empty(&wqh->head))
480 		wake_up_q(&wqh->head);
481 
482 	/* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
483 	preempt_enable();
484 }
485 
486 /*
487  * Deadlock detection is conditional:
488  *
489  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
490  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
491  *
492  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
493  * conducted independent of the detect argument.
494  *
495  * If the waiter argument is NULL this indicates the deboost path and
496  * deadlock detection is disabled independent of the detect argument
497  * and the config settings.
498  */
499 static __always_inline bool
500 rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
501 			      enum rtmutex_chainwalk chwalk)
502 {
503 	if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
504 		return waiter != NULL;
505 	return chwalk == RT_MUTEX_FULL_CHAINWALK;
506 }
507 
508 static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
509 {
510 	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
511 }
512 
513 /*
514  * Adjust the priority chain. Also used for deadlock detection.
515  * Decreases task's usage by one - may thus free the task.
516  *
517  * @task:	the task owning the mutex (owner) for which a chain walk is
518  *		probably needed
519  * @chwalk:	do we have to carry out deadlock detection?
520  * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
521  *		things for a task that has just got its priority adjusted, and
522  *		is waiting on a mutex)
523  * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
524  *		we dropped its pi_lock. Is never dereferenced, only used for
525  *		comparison to detect lock chain changes.
526  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
527  *		its priority to the mutex owner (can be NULL in the case
528  *		depicted above or if the top waiter is gone away and we are
529  *		actually deboosting the owner)
530  * @top_task:	the current top waiter
531  *
532  * Returns 0 or -EDEADLK.
533  *
534  * Chain walk basics and protection scope
535  *
536  * [R] refcount on task
537  * [P] task->pi_lock held
538  * [L] rtmutex->wait_lock held
539  *
540  * Step	Description				Protected by
541  *	function arguments:
542  *	@task					[R]
543  *	@orig_lock if != NULL			@top_task is blocked on it
544  *	@next_lock				Unprotected. Cannot be
545  *						dereferenced. Only used for
546  *						comparison.
547  *	@orig_waiter if != NULL			@top_task is blocked on it
548  *	@top_task				current, or in case of proxy
549  *						locking protected by calling
550  *						code
551  *	again:
552  *	  loop_sanity_check();
553  *	retry:
554  * [1]	  lock(task->pi_lock);			[R] acquire [P]
555  * [2]	  waiter = task->pi_blocked_on;		[P]
556  * [3]	  check_exit_conditions_1();		[P]
557  * [4]	  lock = waiter->lock;			[P]
558  * [5]	  if (!try_lock(lock->wait_lock)) {	[P] try to acquire [L]
559  *	    unlock(task->pi_lock);		release [P]
560  *	    goto retry;
561  *	  }
562  * [6]	  check_exit_conditions_2();		[P] + [L]
563  * [7]	  requeue_lock_waiter(lock, waiter);	[P] + [L]
564  * [8]	  unlock(task->pi_lock);		release [P]
565  *	  put_task_struct(task);		release [R]
566  * [9]	  check_exit_conditions_3();		[L]
567  * [10]	  task = owner(lock);			[L]
568  *	  get_task_struct(task);		[L] acquire [R]
569  *	  lock(task->pi_lock);			[L] acquire [P]
570  * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
571  * [12]	  check_exit_conditions_4();		[P] + [L]
572  * [13]	  unlock(task->pi_lock);		release [P]
573  *	  unlock(lock->wait_lock);		release [L]
574  *	  goto again;
575  */
576 static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
577 					      enum rtmutex_chainwalk chwalk,
578 					      struct rt_mutex_base *orig_lock,
579 					      struct rt_mutex_base *next_lock,
580 					      struct rt_mutex_waiter *orig_waiter,
581 					      struct task_struct *top_task)
582 {
583 	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
584 	struct rt_mutex_waiter *prerequeue_top_waiter;
585 	int ret = 0, depth = 0;
586 	struct rt_mutex_base *lock;
587 	bool detect_deadlock;
588 	bool requeue = true;
589 
590 	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
591 
592 	/*
593 	 * The (de)boosting is a step by step approach with a lot of
594 	 * pitfalls. We want this to be preemptible and we want hold a
595 	 * maximum of two locks per step. So we have to check
596 	 * carefully whether things change under us.
597 	 */
598  again:
599 	/*
600 	 * We limit the lock chain length for each invocation.
601 	 */
602 	if (++depth > max_lock_depth) {
603 		static int prev_max;
604 
605 		/*
606 		 * Print this only once. If the admin changes the limit,
607 		 * print a new message when reaching the limit again.
608 		 */
609 		if (prev_max != max_lock_depth) {
610 			prev_max = max_lock_depth;
611 			printk(KERN_WARNING "Maximum lock depth %d reached "
612 			       "task: %s (%d)\n", max_lock_depth,
613 			       top_task->comm, task_pid_nr(top_task));
614 		}
615 		put_task_struct(task);
616 
617 		return -EDEADLK;
618 	}
619 
620 	/*
621 	 * We are fully preemptible here and only hold the refcount on
622 	 * @task. So everything can have changed under us since the
623 	 * caller or our own code below (goto retry/again) dropped all
624 	 * locks.
625 	 */
626  retry:
627 	/*
628 	 * [1] Task cannot go away as we did a get_task() before !
629 	 */
630 	raw_spin_lock_irq(&task->pi_lock);
631 
632 	/*
633 	 * [2] Get the waiter on which @task is blocked on.
634 	 */
635 	waiter = task->pi_blocked_on;
636 
637 	/*
638 	 * [3] check_exit_conditions_1() protected by task->pi_lock.
639 	 */
640 
641 	/*
642 	 * Check whether the end of the boosting chain has been
643 	 * reached or the state of the chain has changed while we
644 	 * dropped the locks.
645 	 */
646 	if (!waiter)
647 		goto out_unlock_pi;
648 
649 	/*
650 	 * Check the orig_waiter state. After we dropped the locks,
651 	 * the previous owner of the lock might have released the lock.
652 	 */
653 	if (orig_waiter && !rt_mutex_owner(orig_lock))
654 		goto out_unlock_pi;
655 
656 	/*
657 	 * We dropped all locks after taking a refcount on @task, so
658 	 * the task might have moved on in the lock chain or even left
659 	 * the chain completely and blocks now on an unrelated lock or
660 	 * on @orig_lock.
661 	 *
662 	 * We stored the lock on which @task was blocked in @next_lock,
663 	 * so we can detect the chain change.
664 	 */
665 	if (next_lock != waiter->lock)
666 		goto out_unlock_pi;
667 
668 	/*
669 	 * There could be 'spurious' loops in the lock graph due to ww_mutex,
670 	 * consider:
671 	 *
672 	 *   P1: A, ww_A, ww_B
673 	 *   P2: ww_B, ww_A
674 	 *   P3: A
675 	 *
676 	 * P3 should not return -EDEADLK because it gets trapped in the cycle
677 	 * created by P1 and P2 (which will resolve -- and runs into
678 	 * max_lock_depth above). Therefore disable detect_deadlock such that
679 	 * the below termination condition can trigger once all relevant tasks
680 	 * are boosted.
681 	 *
682 	 * Even when we start with ww_mutex we can disable deadlock detection,
683 	 * since we would supress a ww_mutex induced deadlock at [6] anyway.
684 	 * Supressing it here however is not sufficient since we might still
685 	 * hit [6] due to adjustment driven iteration.
686 	 *
687 	 * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
688 	 * utterly fail to report it; lockdep should.
689 	 */
690 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
691 		detect_deadlock = false;
692 
693 	/*
694 	 * Drop out, when the task has no waiters. Note,
695 	 * top_waiter can be NULL, when we are in the deboosting
696 	 * mode!
697 	 */
698 	if (top_waiter) {
699 		if (!task_has_pi_waiters(task))
700 			goto out_unlock_pi;
701 		/*
702 		 * If deadlock detection is off, we stop here if we
703 		 * are not the top pi waiter of the task. If deadlock
704 		 * detection is enabled we continue, but stop the
705 		 * requeueing in the chain walk.
706 		 */
707 		if (top_waiter != task_top_pi_waiter(task)) {
708 			if (!detect_deadlock)
709 				goto out_unlock_pi;
710 			else
711 				requeue = false;
712 		}
713 	}
714 
715 	/*
716 	 * If the waiter priority is the same as the task priority
717 	 * then there is no further priority adjustment necessary.  If
718 	 * deadlock detection is off, we stop the chain walk. If its
719 	 * enabled we continue, but stop the requeueing in the chain
720 	 * walk.
721 	 */
722 	if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
723 		if (!detect_deadlock)
724 			goto out_unlock_pi;
725 		else
726 			requeue = false;
727 	}
728 
729 	/*
730 	 * [4] Get the next lock
731 	 */
732 	lock = waiter->lock;
733 	/*
734 	 * [5] We need to trylock here as we are holding task->pi_lock,
735 	 * which is the reverse lock order versus the other rtmutex
736 	 * operations.
737 	 */
738 	if (!raw_spin_trylock(&lock->wait_lock)) {
739 		raw_spin_unlock_irq(&task->pi_lock);
740 		cpu_relax();
741 		goto retry;
742 	}
743 
744 	/*
745 	 * [6] check_exit_conditions_2() protected by task->pi_lock and
746 	 * lock->wait_lock.
747 	 *
748 	 * Deadlock detection. If the lock is the same as the original
749 	 * lock which caused us to walk the lock chain or if the
750 	 * current lock is owned by the task which initiated the chain
751 	 * walk, we detected a deadlock.
752 	 */
753 	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
754 		ret = -EDEADLK;
755 
756 		/*
757 		 * When the deadlock is due to ww_mutex; also see above. Don't
758 		 * report the deadlock and instead let the ww_mutex wound/die
759 		 * logic pick which of the contending threads gets -EDEADLK.
760 		 *
761 		 * NOTE: assumes the cycle only contains a single ww_class; any
762 		 * other configuration and we fail to report; also, see
763 		 * lockdep.
764 		 */
765 		if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
766 			ret = 0;
767 
768 		raw_spin_unlock(&lock->wait_lock);
769 		goto out_unlock_pi;
770 	}
771 
772 	/*
773 	 * If we just follow the lock chain for deadlock detection, no
774 	 * need to do all the requeue operations. To avoid a truckload
775 	 * of conditionals around the various places below, just do the
776 	 * minimum chain walk checks.
777 	 */
778 	if (!requeue) {
779 		/*
780 		 * No requeue[7] here. Just release @task [8]
781 		 */
782 		raw_spin_unlock(&task->pi_lock);
783 		put_task_struct(task);
784 
785 		/*
786 		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
787 		 * If there is no owner of the lock, end of chain.
788 		 */
789 		if (!rt_mutex_owner(lock)) {
790 			raw_spin_unlock_irq(&lock->wait_lock);
791 			return 0;
792 		}
793 
794 		/* [10] Grab the next task, i.e. owner of @lock */
795 		task = get_task_struct(rt_mutex_owner(lock));
796 		raw_spin_lock(&task->pi_lock);
797 
798 		/*
799 		 * No requeue [11] here. We just do deadlock detection.
800 		 *
801 		 * [12] Store whether owner is blocked
802 		 * itself. Decision is made after dropping the locks
803 		 */
804 		next_lock = task_blocked_on_lock(task);
805 		/*
806 		 * Get the top waiter for the next iteration
807 		 */
808 		top_waiter = rt_mutex_top_waiter(lock);
809 
810 		/* [13] Drop locks */
811 		raw_spin_unlock(&task->pi_lock);
812 		raw_spin_unlock_irq(&lock->wait_lock);
813 
814 		/* If owner is not blocked, end of chain. */
815 		if (!next_lock)
816 			goto out_put_task;
817 		goto again;
818 	}
819 
820 	/*
821 	 * Store the current top waiter before doing the requeue
822 	 * operation on @lock. We need it for the boost/deboost
823 	 * decision below.
824 	 */
825 	prerequeue_top_waiter = rt_mutex_top_waiter(lock);
826 
827 	/* [7] Requeue the waiter in the lock waiter tree. */
828 	rt_mutex_dequeue(lock, waiter);
829 
830 	/*
831 	 * Update the waiter prio fields now that we're dequeued.
832 	 *
833 	 * These values can have changed through either:
834 	 *
835 	 *   sys_sched_set_scheduler() / sys_sched_setattr()
836 	 *
837 	 * or
838 	 *
839 	 *   DL CBS enforcement advancing the effective deadline.
840 	 *
841 	 * Even though pi_waiters also uses these fields, and that tree is only
842 	 * updated in [11], we can do this here, since we hold [L], which
843 	 * serializes all pi_waiters access and rb_erase() does not care about
844 	 * the values of the node being removed.
845 	 */
846 	waiter_update_prio(waiter, task);
847 
848 	rt_mutex_enqueue(lock, waiter);
849 
850 	/* [8] Release the task */
851 	raw_spin_unlock(&task->pi_lock);
852 	put_task_struct(task);
853 
854 	/*
855 	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
856 	 *
857 	 * We must abort the chain walk if there is no lock owner even
858 	 * in the dead lock detection case, as we have nothing to
859 	 * follow here. This is the end of the chain we are walking.
860 	 */
861 	if (!rt_mutex_owner(lock)) {
862 		/*
863 		 * If the requeue [7] above changed the top waiter,
864 		 * then we need to wake the new top waiter up to try
865 		 * to get the lock.
866 		 */
867 		if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
868 			wake_up_state(waiter->task, waiter->wake_state);
869 		raw_spin_unlock_irq(&lock->wait_lock);
870 		return 0;
871 	}
872 
873 	/* [10] Grab the next task, i.e. the owner of @lock */
874 	task = get_task_struct(rt_mutex_owner(lock));
875 	raw_spin_lock(&task->pi_lock);
876 
877 	/* [11] requeue the pi waiters if necessary */
878 	if (waiter == rt_mutex_top_waiter(lock)) {
879 		/*
880 		 * The waiter became the new top (highest priority)
881 		 * waiter on the lock. Replace the previous top waiter
882 		 * in the owner tasks pi waiters tree with this waiter
883 		 * and adjust the priority of the owner.
884 		 */
885 		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
886 		rt_mutex_enqueue_pi(task, waiter);
887 		rt_mutex_adjust_prio(task);
888 
889 	} else if (prerequeue_top_waiter == waiter) {
890 		/*
891 		 * The waiter was the top waiter on the lock, but is
892 		 * no longer the top priority waiter. Replace waiter in
893 		 * the owner tasks pi waiters tree with the new top
894 		 * (highest priority) waiter and adjust the priority
895 		 * of the owner.
896 		 * The new top waiter is stored in @waiter so that
897 		 * @waiter == @top_waiter evaluates to true below and
898 		 * we continue to deboost the rest of the chain.
899 		 */
900 		rt_mutex_dequeue_pi(task, waiter);
901 		waiter = rt_mutex_top_waiter(lock);
902 		rt_mutex_enqueue_pi(task, waiter);
903 		rt_mutex_adjust_prio(task);
904 	} else {
905 		/*
906 		 * Nothing changed. No need to do any priority
907 		 * adjustment.
908 		 */
909 	}
910 
911 	/*
912 	 * [12] check_exit_conditions_4() protected by task->pi_lock
913 	 * and lock->wait_lock. The actual decisions are made after we
914 	 * dropped the locks.
915 	 *
916 	 * Check whether the task which owns the current lock is pi
917 	 * blocked itself. If yes we store a pointer to the lock for
918 	 * the lock chain change detection above. After we dropped
919 	 * task->pi_lock next_lock cannot be dereferenced anymore.
920 	 */
921 	next_lock = task_blocked_on_lock(task);
922 	/*
923 	 * Store the top waiter of @lock for the end of chain walk
924 	 * decision below.
925 	 */
926 	top_waiter = rt_mutex_top_waiter(lock);
927 
928 	/* [13] Drop the locks */
929 	raw_spin_unlock(&task->pi_lock);
930 	raw_spin_unlock_irq(&lock->wait_lock);
931 
932 	/*
933 	 * Make the actual exit decisions [12], based on the stored
934 	 * values.
935 	 *
936 	 * We reached the end of the lock chain. Stop right here. No
937 	 * point to go back just to figure that out.
938 	 */
939 	if (!next_lock)
940 		goto out_put_task;
941 
942 	/*
943 	 * If the current waiter is not the top waiter on the lock,
944 	 * then we can stop the chain walk here if we are not in full
945 	 * deadlock detection mode.
946 	 */
947 	if (!detect_deadlock && waiter != top_waiter)
948 		goto out_put_task;
949 
950 	goto again;
951 
952  out_unlock_pi:
953 	raw_spin_unlock_irq(&task->pi_lock);
954  out_put_task:
955 	put_task_struct(task);
956 
957 	return ret;
958 }
959 
960 /*
961  * Try to take an rt-mutex
962  *
963  * Must be called with lock->wait_lock held and interrupts disabled
964  *
965  * @lock:   The lock to be acquired.
966  * @task:   The task which wants to acquire the lock
967  * @waiter: The waiter that is queued to the lock's wait tree if the
968  *	    callsite called task_blocked_on_lock(), otherwise NULL
969  */
970 static int __sched
971 try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
972 		     struct rt_mutex_waiter *waiter)
973 {
974 	lockdep_assert_held(&lock->wait_lock);
975 
976 	/*
977 	 * Before testing whether we can acquire @lock, we set the
978 	 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
979 	 * other tasks which try to modify @lock into the slow path
980 	 * and they serialize on @lock->wait_lock.
981 	 *
982 	 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
983 	 * as explained at the top of this file if and only if:
984 	 *
985 	 * - There is a lock owner. The caller must fixup the
986 	 *   transient state if it does a trylock or leaves the lock
987 	 *   function due to a signal or timeout.
988 	 *
989 	 * - @task acquires the lock and there are no other
990 	 *   waiters. This is undone in rt_mutex_set_owner(@task) at
991 	 *   the end of this function.
992 	 */
993 	mark_rt_mutex_waiters(lock);
994 
995 	/*
996 	 * If @lock has an owner, give up.
997 	 */
998 	if (rt_mutex_owner(lock))
999 		return 0;
1000 
1001 	/*
1002 	 * If @waiter != NULL, @task has already enqueued the waiter
1003 	 * into @lock waiter tree. If @waiter == NULL then this is a
1004 	 * trylock attempt.
1005 	 */
1006 	if (waiter) {
1007 		struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
1008 
1009 		/*
1010 		 * If waiter is the highest priority waiter of @lock,
1011 		 * or allowed to steal it, take it over.
1012 		 */
1013 		if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1014 			/*
1015 			 * We can acquire the lock. Remove the waiter from the
1016 			 * lock waiters tree.
1017 			 */
1018 			rt_mutex_dequeue(lock, waiter);
1019 		} else {
1020 			return 0;
1021 		}
1022 	} else {
1023 		/*
1024 		 * If the lock has waiters already we check whether @task is
1025 		 * eligible to take over the lock.
1026 		 *
1027 		 * If there are no other waiters, @task can acquire
1028 		 * the lock.  @task->pi_blocked_on is NULL, so it does
1029 		 * not need to be dequeued.
1030 		 */
1031 		if (rt_mutex_has_waiters(lock)) {
1032 			/* Check whether the trylock can steal it. */
1033 			if (!rt_mutex_steal(task_to_waiter(task),
1034 					    rt_mutex_top_waiter(lock)))
1035 				return 0;
1036 
1037 			/*
1038 			 * The current top waiter stays enqueued. We
1039 			 * don't have to change anything in the lock
1040 			 * waiters order.
1041 			 */
1042 		} else {
1043 			/*
1044 			 * No waiters. Take the lock without the
1045 			 * pi_lock dance.@task->pi_blocked_on is NULL
1046 			 * and we have no waiters to enqueue in @task
1047 			 * pi waiters tree.
1048 			 */
1049 			goto takeit;
1050 		}
1051 	}
1052 
1053 	/*
1054 	 * Clear @task->pi_blocked_on. Requires protection by
1055 	 * @task->pi_lock. Redundant operation for the @waiter == NULL
1056 	 * case, but conditionals are more expensive than a redundant
1057 	 * store.
1058 	 */
1059 	raw_spin_lock(&task->pi_lock);
1060 	task->pi_blocked_on = NULL;
1061 	/*
1062 	 * Finish the lock acquisition. @task is the new owner. If
1063 	 * other waiters exist we have to insert the highest priority
1064 	 * waiter into @task->pi_waiters tree.
1065 	 */
1066 	if (rt_mutex_has_waiters(lock))
1067 		rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
1068 	raw_spin_unlock(&task->pi_lock);
1069 
1070 takeit:
1071 	/*
1072 	 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1073 	 * are still waiters or clears it.
1074 	 */
1075 	rt_mutex_set_owner(lock, task);
1076 
1077 	return 1;
1078 }
1079 
1080 /*
1081  * Task blocks on lock.
1082  *
1083  * Prepare waiter and propagate pi chain
1084  *
1085  * This must be called with lock->wait_lock held and interrupts disabled
1086  */
1087 static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
1088 					   struct rt_mutex_waiter *waiter,
1089 					   struct task_struct *task,
1090 					   struct ww_acquire_ctx *ww_ctx,
1091 					   enum rtmutex_chainwalk chwalk)
1092 {
1093 	struct task_struct *owner = rt_mutex_owner(lock);
1094 	struct rt_mutex_waiter *top_waiter = waiter;
1095 	struct rt_mutex_base *next_lock;
1096 	int chain_walk = 0, res;
1097 
1098 	lockdep_assert_held(&lock->wait_lock);
1099 
1100 	/*
1101 	 * Early deadlock detection. We really don't want the task to
1102 	 * enqueue on itself just to untangle the mess later. It's not
1103 	 * only an optimization. We drop the locks, so another waiter
1104 	 * can come in before the chain walk detects the deadlock. So
1105 	 * the other will detect the deadlock and return -EDEADLOCK,
1106 	 * which is wrong, as the other waiter is not in a deadlock
1107 	 * situation.
1108 	 *
1109 	 * Except for ww_mutex, in that case the chain walk must already deal
1110 	 * with spurious cycles, see the comments at [3] and [6].
1111 	 */
1112 	if (owner == task && !(build_ww_mutex() && ww_ctx))
1113 		return -EDEADLK;
1114 
1115 	raw_spin_lock(&task->pi_lock);
1116 	waiter->task = task;
1117 	waiter->lock = lock;
1118 	waiter_update_prio(waiter, task);
1119 
1120 	/* Get the top priority waiter on the lock */
1121 	if (rt_mutex_has_waiters(lock))
1122 		top_waiter = rt_mutex_top_waiter(lock);
1123 	rt_mutex_enqueue(lock, waiter);
1124 
1125 	task->pi_blocked_on = waiter;
1126 
1127 	raw_spin_unlock(&task->pi_lock);
1128 
1129 	if (build_ww_mutex() && ww_ctx) {
1130 		struct rt_mutex *rtm;
1131 
1132 		/* Check whether the waiter should back out immediately */
1133 		rtm = container_of(lock, struct rt_mutex, rtmutex);
1134 		res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx);
1135 		if (res) {
1136 			raw_spin_lock(&task->pi_lock);
1137 			rt_mutex_dequeue(lock, waiter);
1138 			task->pi_blocked_on = NULL;
1139 			raw_spin_unlock(&task->pi_lock);
1140 			return res;
1141 		}
1142 	}
1143 
1144 	if (!owner)
1145 		return 0;
1146 
1147 	raw_spin_lock(&owner->pi_lock);
1148 	if (waiter == rt_mutex_top_waiter(lock)) {
1149 		rt_mutex_dequeue_pi(owner, top_waiter);
1150 		rt_mutex_enqueue_pi(owner, waiter);
1151 
1152 		rt_mutex_adjust_prio(owner);
1153 		if (owner->pi_blocked_on)
1154 			chain_walk = 1;
1155 	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1156 		chain_walk = 1;
1157 	}
1158 
1159 	/* Store the lock on which owner is blocked or NULL */
1160 	next_lock = task_blocked_on_lock(owner);
1161 
1162 	raw_spin_unlock(&owner->pi_lock);
1163 	/*
1164 	 * Even if full deadlock detection is on, if the owner is not
1165 	 * blocked itself, we can avoid finding this out in the chain
1166 	 * walk.
1167 	 */
1168 	if (!chain_walk || !next_lock)
1169 		return 0;
1170 
1171 	/*
1172 	 * The owner can't disappear while holding a lock,
1173 	 * so the owner struct is protected by wait_lock.
1174 	 * Gets dropped in rt_mutex_adjust_prio_chain()!
1175 	 */
1176 	get_task_struct(owner);
1177 
1178 	raw_spin_unlock_irq(&lock->wait_lock);
1179 
1180 	res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1181 					 next_lock, waiter, task);
1182 
1183 	raw_spin_lock_irq(&lock->wait_lock);
1184 
1185 	return res;
1186 }
1187 
1188 /*
1189  * Remove the top waiter from the current tasks pi waiter tree and
1190  * queue it up.
1191  *
1192  * Called with lock->wait_lock held and interrupts disabled.
1193  */
1194 static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1195 					    struct rt_mutex_base *lock)
1196 {
1197 	struct rt_mutex_waiter *waiter;
1198 
1199 	raw_spin_lock(&current->pi_lock);
1200 
1201 	waiter = rt_mutex_top_waiter(lock);
1202 
1203 	/*
1204 	 * Remove it from current->pi_waiters and deboost.
1205 	 *
1206 	 * We must in fact deboost here in order to ensure we call
1207 	 * rt_mutex_setprio() to update p->pi_top_task before the
1208 	 * task unblocks.
1209 	 */
1210 	rt_mutex_dequeue_pi(current, waiter);
1211 	rt_mutex_adjust_prio(current);
1212 
1213 	/*
1214 	 * As we are waking up the top waiter, and the waiter stays
1215 	 * queued on the lock until it gets the lock, this lock
1216 	 * obviously has waiters. Just set the bit here and this has
1217 	 * the added benefit of forcing all new tasks into the
1218 	 * slow path making sure no task of lower priority than
1219 	 * the top waiter can steal this lock.
1220 	 */
1221 	lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1222 
1223 	/*
1224 	 * We deboosted before waking the top waiter task such that we don't
1225 	 * run two tasks with the 'same' priority (and ensure the
1226 	 * p->pi_top_task pointer points to a blocked task). This however can
1227 	 * lead to priority inversion if we would get preempted after the
1228 	 * deboost but before waking our donor task, hence the preempt_disable()
1229 	 * before unlock.
1230 	 *
1231 	 * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1232 	 */
1233 	preempt_disable();
1234 	rt_mutex_wake_q_add(wqh, waiter);
1235 	raw_spin_unlock(&current->pi_lock);
1236 }
1237 
1238 static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1239 {
1240 	int ret = try_to_take_rt_mutex(lock, current, NULL);
1241 
1242 	/*
1243 	 * try_to_take_rt_mutex() sets the lock waiters bit
1244 	 * unconditionally. Clean this up.
1245 	 */
1246 	fixup_rt_mutex_waiters(lock);
1247 
1248 	return ret;
1249 }
1250 
1251 /*
1252  * Slow path try-lock function:
1253  */
1254 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1255 {
1256 	unsigned long flags;
1257 	int ret;
1258 
1259 	/*
1260 	 * If the lock already has an owner we fail to get the lock.
1261 	 * This can be done without taking the @lock->wait_lock as
1262 	 * it is only being read, and this is a trylock anyway.
1263 	 */
1264 	if (rt_mutex_owner(lock))
1265 		return 0;
1266 
1267 	/*
1268 	 * The mutex has currently no owner. Lock the wait lock and try to
1269 	 * acquire the lock. We use irqsave here to support early boot calls.
1270 	 */
1271 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1272 
1273 	ret = __rt_mutex_slowtrylock(lock);
1274 
1275 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1276 
1277 	return ret;
1278 }
1279 
1280 static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
1281 {
1282 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1283 		return 1;
1284 
1285 	return rt_mutex_slowtrylock(lock);
1286 }
1287 
1288 /*
1289  * Slow path to release a rt-mutex.
1290  */
1291 static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1292 {
1293 	DEFINE_RT_WAKE_Q(wqh);
1294 	unsigned long flags;
1295 
1296 	/* irqsave required to support early boot calls */
1297 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1298 
1299 	debug_rt_mutex_unlock(lock);
1300 
1301 	/*
1302 	 * We must be careful here if the fast path is enabled. If we
1303 	 * have no waiters queued we cannot set owner to NULL here
1304 	 * because of:
1305 	 *
1306 	 * foo->lock->owner = NULL;
1307 	 *			rtmutex_lock(foo->lock);   <- fast path
1308 	 *			free = atomic_dec_and_test(foo->refcnt);
1309 	 *			rtmutex_unlock(foo->lock); <- fast path
1310 	 *			if (free)
1311 	 *				kfree(foo);
1312 	 * raw_spin_unlock(foo->lock->wait_lock);
1313 	 *
1314 	 * So for the fastpath enabled kernel:
1315 	 *
1316 	 * Nothing can set the waiters bit as long as we hold
1317 	 * lock->wait_lock. So we do the following sequence:
1318 	 *
1319 	 *	owner = rt_mutex_owner(lock);
1320 	 *	clear_rt_mutex_waiters(lock);
1321 	 *	raw_spin_unlock(&lock->wait_lock);
1322 	 *	if (cmpxchg(&lock->owner, owner, 0) == owner)
1323 	 *		return;
1324 	 *	goto retry;
1325 	 *
1326 	 * The fastpath disabled variant is simple as all access to
1327 	 * lock->owner is serialized by lock->wait_lock:
1328 	 *
1329 	 *	lock->owner = NULL;
1330 	 *	raw_spin_unlock(&lock->wait_lock);
1331 	 */
1332 	while (!rt_mutex_has_waiters(lock)) {
1333 		/* Drops lock->wait_lock ! */
1334 		if (unlock_rt_mutex_safe(lock, flags) == true)
1335 			return;
1336 		/* Relock the rtmutex and try again */
1337 		raw_spin_lock_irqsave(&lock->wait_lock, flags);
1338 	}
1339 
1340 	/*
1341 	 * The wakeup next waiter path does not suffer from the above
1342 	 * race. See the comments there.
1343 	 *
1344 	 * Queue the next waiter for wakeup once we release the wait_lock.
1345 	 */
1346 	mark_wakeup_next_waiter(&wqh, lock);
1347 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1348 
1349 	rt_mutex_wake_up_q(&wqh);
1350 }
1351 
1352 static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
1353 {
1354 	if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1355 		return;
1356 
1357 	rt_mutex_slowunlock(lock);
1358 }
1359 
1360 #ifdef CONFIG_SMP
1361 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1362 				  struct rt_mutex_waiter *waiter,
1363 				  struct task_struct *owner)
1364 {
1365 	bool res = true;
1366 
1367 	rcu_read_lock();
1368 	for (;;) {
1369 		/* If owner changed, trylock again. */
1370 		if (owner != rt_mutex_owner(lock))
1371 			break;
1372 		/*
1373 		 * Ensure that @owner is dereferenced after checking that
1374 		 * the lock owner still matches @owner. If that fails,
1375 		 * @owner might point to freed memory. If it still matches,
1376 		 * the rcu_read_lock() ensures the memory stays valid.
1377 		 */
1378 		barrier();
1379 		/*
1380 		 * Stop spinning when:
1381 		 *  - the lock owner has been scheduled out
1382 		 *  - current is not longer the top waiter
1383 		 *  - current is requested to reschedule (redundant
1384 		 *    for CONFIG_PREEMPT_RCU=y)
1385 		 *  - the VCPU on which owner runs is preempted
1386 		 */
1387 		if (!owner_on_cpu(owner) || need_resched() ||
1388 		    !rt_mutex_waiter_is_top_waiter(lock, waiter)) {
1389 			res = false;
1390 			break;
1391 		}
1392 		cpu_relax();
1393 	}
1394 	rcu_read_unlock();
1395 	return res;
1396 }
1397 #else
1398 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1399 				  struct rt_mutex_waiter *waiter,
1400 				  struct task_struct *owner)
1401 {
1402 	return false;
1403 }
1404 #endif
1405 
1406 #ifdef RT_MUTEX_BUILD_MUTEX
1407 /*
1408  * Functions required for:
1409  *	- rtmutex, futex on all kernels
1410  *	- mutex and rwsem substitutions on RT kernels
1411  */
1412 
1413 /*
1414  * Remove a waiter from a lock and give up
1415  *
1416  * Must be called with lock->wait_lock held and interrupts disabled. It must
1417  * have just failed to try_to_take_rt_mutex().
1418  */
1419 static void __sched remove_waiter(struct rt_mutex_base *lock,
1420 				  struct rt_mutex_waiter *waiter)
1421 {
1422 	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1423 	struct task_struct *owner = rt_mutex_owner(lock);
1424 	struct rt_mutex_base *next_lock;
1425 
1426 	lockdep_assert_held(&lock->wait_lock);
1427 
1428 	raw_spin_lock(&current->pi_lock);
1429 	rt_mutex_dequeue(lock, waiter);
1430 	current->pi_blocked_on = NULL;
1431 	raw_spin_unlock(&current->pi_lock);
1432 
1433 	/*
1434 	 * Only update priority if the waiter was the highest priority
1435 	 * waiter of the lock and there is an owner to update.
1436 	 */
1437 	if (!owner || !is_top_waiter)
1438 		return;
1439 
1440 	raw_spin_lock(&owner->pi_lock);
1441 
1442 	rt_mutex_dequeue_pi(owner, waiter);
1443 
1444 	if (rt_mutex_has_waiters(lock))
1445 		rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1446 
1447 	rt_mutex_adjust_prio(owner);
1448 
1449 	/* Store the lock on which owner is blocked or NULL */
1450 	next_lock = task_blocked_on_lock(owner);
1451 
1452 	raw_spin_unlock(&owner->pi_lock);
1453 
1454 	/*
1455 	 * Don't walk the chain, if the owner task is not blocked
1456 	 * itself.
1457 	 */
1458 	if (!next_lock)
1459 		return;
1460 
1461 	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1462 	get_task_struct(owner);
1463 
1464 	raw_spin_unlock_irq(&lock->wait_lock);
1465 
1466 	rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1467 				   next_lock, NULL, current);
1468 
1469 	raw_spin_lock_irq(&lock->wait_lock);
1470 }
1471 
1472 /**
1473  * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
1474  * @lock:		 the rt_mutex to take
1475  * @ww_ctx:		 WW mutex context pointer
1476  * @state:		 the state the task should block in (TASK_INTERRUPTIBLE
1477  *			 or TASK_UNINTERRUPTIBLE)
1478  * @timeout:		 the pre-initialized and started timer, or NULL for none
1479  * @waiter:		 the pre-initialized rt_mutex_waiter
1480  *
1481  * Must be called with lock->wait_lock held and interrupts disabled
1482  */
1483 static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1484 					   struct ww_acquire_ctx *ww_ctx,
1485 					   unsigned int state,
1486 					   struct hrtimer_sleeper *timeout,
1487 					   struct rt_mutex_waiter *waiter)
1488 {
1489 	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1490 	struct task_struct *owner;
1491 	int ret = 0;
1492 
1493 	for (;;) {
1494 		/* Try to acquire the lock: */
1495 		if (try_to_take_rt_mutex(lock, current, waiter))
1496 			break;
1497 
1498 		if (timeout && !timeout->task) {
1499 			ret = -ETIMEDOUT;
1500 			break;
1501 		}
1502 		if (signal_pending_state(state, current)) {
1503 			ret = -EINTR;
1504 			break;
1505 		}
1506 
1507 		if (build_ww_mutex() && ww_ctx) {
1508 			ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1509 			if (ret)
1510 				break;
1511 		}
1512 
1513 		if (waiter == rt_mutex_top_waiter(lock))
1514 			owner = rt_mutex_owner(lock);
1515 		else
1516 			owner = NULL;
1517 		raw_spin_unlock_irq(&lock->wait_lock);
1518 
1519 		if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1520 			schedule();
1521 
1522 		raw_spin_lock_irq(&lock->wait_lock);
1523 		set_current_state(state);
1524 	}
1525 
1526 	__set_current_state(TASK_RUNNING);
1527 	return ret;
1528 }
1529 
1530 static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1531 					     struct rt_mutex_waiter *w)
1532 {
1533 	/*
1534 	 * If the result is not -EDEADLOCK or the caller requested
1535 	 * deadlock detection, nothing to do here.
1536 	 */
1537 	if (res != -EDEADLOCK || detect_deadlock)
1538 		return;
1539 
1540 	if (build_ww_mutex() && w->ww_ctx)
1541 		return;
1542 
1543 	/*
1544 	 * Yell loudly and stop the task right here.
1545 	 */
1546 	WARN(1, "rtmutex deadlock detected\n");
1547 	while (1) {
1548 		set_current_state(TASK_INTERRUPTIBLE);
1549 		schedule();
1550 	}
1551 }
1552 
1553 /**
1554  * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
1555  * @lock:	The rtmutex to block lock
1556  * @ww_ctx:	WW mutex context pointer
1557  * @state:	The task state for sleeping
1558  * @chwalk:	Indicator whether full or partial chainwalk is requested
1559  * @waiter:	Initializer waiter for blocking
1560  */
1561 static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1562 				       struct ww_acquire_ctx *ww_ctx,
1563 				       unsigned int state,
1564 				       enum rtmutex_chainwalk chwalk,
1565 				       struct rt_mutex_waiter *waiter)
1566 {
1567 	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1568 	struct ww_mutex *ww = ww_container_of(rtm);
1569 	int ret;
1570 
1571 	lockdep_assert_held(&lock->wait_lock);
1572 
1573 	/* Try to acquire the lock again: */
1574 	if (try_to_take_rt_mutex(lock, current, NULL)) {
1575 		if (build_ww_mutex() && ww_ctx) {
1576 			__ww_mutex_check_waiters(rtm, ww_ctx);
1577 			ww_mutex_lock_acquired(ww, ww_ctx);
1578 		}
1579 		return 0;
1580 	}
1581 
1582 	set_current_state(state);
1583 
1584 	trace_contention_begin(lock, LCB_F_RT);
1585 
1586 	ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
1587 	if (likely(!ret))
1588 		ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
1589 
1590 	if (likely(!ret)) {
1591 		/* acquired the lock */
1592 		if (build_ww_mutex() && ww_ctx) {
1593 			if (!ww_ctx->is_wait_die)
1594 				__ww_mutex_check_waiters(rtm, ww_ctx);
1595 			ww_mutex_lock_acquired(ww, ww_ctx);
1596 		}
1597 	} else {
1598 		__set_current_state(TASK_RUNNING);
1599 		remove_waiter(lock, waiter);
1600 		rt_mutex_handle_deadlock(ret, chwalk, waiter);
1601 	}
1602 
1603 	/*
1604 	 * try_to_take_rt_mutex() sets the waiter bit
1605 	 * unconditionally. We might have to fix that up.
1606 	 */
1607 	fixup_rt_mutex_waiters(lock);
1608 
1609 	trace_contention_end(lock, ret);
1610 
1611 	return ret;
1612 }
1613 
1614 static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1615 					     struct ww_acquire_ctx *ww_ctx,
1616 					     unsigned int state)
1617 {
1618 	struct rt_mutex_waiter waiter;
1619 	int ret;
1620 
1621 	rt_mutex_init_waiter(&waiter);
1622 	waiter.ww_ctx = ww_ctx;
1623 
1624 	ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1625 				  &waiter);
1626 
1627 	debug_rt_mutex_free_waiter(&waiter);
1628 	return ret;
1629 }
1630 
1631 /*
1632  * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
1633  * @lock:	The rtmutex to block lock
1634  * @ww_ctx:	WW mutex context pointer
1635  * @state:	The task state for sleeping
1636  */
1637 static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1638 				     struct ww_acquire_ctx *ww_ctx,
1639 				     unsigned int state)
1640 {
1641 	unsigned long flags;
1642 	int ret;
1643 
1644 	/*
1645 	 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1646 	 * be called in early boot if the cmpxchg() fast path is disabled
1647 	 * (debug, no architecture support). In this case we will acquire the
1648 	 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1649 	 * enable interrupts in that early boot case. So we need to use the
1650 	 * irqsave/restore variants.
1651 	 */
1652 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1653 	ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state);
1654 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1655 
1656 	return ret;
1657 }
1658 
1659 static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1660 					   unsigned int state)
1661 {
1662 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1663 		return 0;
1664 
1665 	return rt_mutex_slowlock(lock, NULL, state);
1666 }
1667 #endif /* RT_MUTEX_BUILD_MUTEX */
1668 
1669 #ifdef RT_MUTEX_BUILD_SPINLOCKS
1670 /*
1671  * Functions required for spin/rw_lock substitution on RT kernels
1672  */
1673 
1674 /**
1675  * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1676  * @lock:	The underlying RT mutex
1677  */
1678 static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
1679 {
1680 	struct rt_mutex_waiter waiter;
1681 	struct task_struct *owner;
1682 
1683 	lockdep_assert_held(&lock->wait_lock);
1684 
1685 	if (try_to_take_rt_mutex(lock, current, NULL))
1686 		return;
1687 
1688 	rt_mutex_init_rtlock_waiter(&waiter);
1689 
1690 	/* Save current state and set state to TASK_RTLOCK_WAIT */
1691 	current_save_and_set_rtlock_wait_state();
1692 
1693 	trace_contention_begin(lock, LCB_F_RT);
1694 
1695 	task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
1696 
1697 	for (;;) {
1698 		/* Try to acquire the lock again */
1699 		if (try_to_take_rt_mutex(lock, current, &waiter))
1700 			break;
1701 
1702 		if (&waiter == rt_mutex_top_waiter(lock))
1703 			owner = rt_mutex_owner(lock);
1704 		else
1705 			owner = NULL;
1706 		raw_spin_unlock_irq(&lock->wait_lock);
1707 
1708 		if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1709 			schedule_rtlock();
1710 
1711 		raw_spin_lock_irq(&lock->wait_lock);
1712 		set_current_state(TASK_RTLOCK_WAIT);
1713 	}
1714 
1715 	/* Restore the task state */
1716 	current_restore_rtlock_saved_state();
1717 
1718 	/*
1719 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1720 	 * We might have to fix that up:
1721 	 */
1722 	fixup_rt_mutex_waiters(lock);
1723 	debug_rt_mutex_free_waiter(&waiter);
1724 
1725 	trace_contention_end(lock, 0);
1726 }
1727 
1728 static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
1729 {
1730 	unsigned long flags;
1731 
1732 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1733 	rtlock_slowlock_locked(lock);
1734 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1735 }
1736 
1737 #endif /* RT_MUTEX_BUILD_SPINLOCKS */
1738