xref: /openbmc/linux/kernel/locking/rtmutex.c (revision d87c25e8)
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 "rtmutex_common.h"
28 
29 #ifndef WW_RT
30 # define build_ww_mutex()	(false)
31 # define ww_container_of(rtm)	NULL
32 
33 static inline int __ww_mutex_add_waiter(struct rt_mutex_waiter *waiter,
34 					struct rt_mutex *lock,
35 					struct ww_acquire_ctx *ww_ctx)
36 {
37 	return 0;
38 }
39 
40 static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
41 					    struct ww_acquire_ctx *ww_ctx)
42 {
43 }
44 
45 static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
46 					  struct ww_acquire_ctx *ww_ctx)
47 {
48 }
49 
50 static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
51 					struct rt_mutex_waiter *waiter,
52 					struct ww_acquire_ctx *ww_ctx)
53 {
54 	return 0;
55 }
56 
57 #else
58 # define build_ww_mutex()	(true)
59 # define ww_container_of(rtm)	container_of(rtm, struct ww_mutex, base)
60 # include "ww_mutex.h"
61 #endif
62 
63 /*
64  * lock->owner state tracking:
65  *
66  * lock->owner holds the task_struct pointer of the owner. Bit 0
67  * is used to keep track of the "lock has waiters" state.
68  *
69  * owner	bit0
70  * NULL		0	lock is free (fast acquire possible)
71  * NULL		1	lock is free and has waiters and the top waiter
72  *				is going to take the lock*
73  * taskpointer	0	lock is held (fast release possible)
74  * taskpointer	1	lock is held and has waiters**
75  *
76  * The fast atomic compare exchange based acquire and release is only
77  * possible when bit 0 of lock->owner is 0.
78  *
79  * (*) It also can be a transitional state when grabbing the lock
80  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
81  * we need to set the bit0 before looking at the lock, and the owner may be
82  * NULL in this small time, hence this can be a transitional state.
83  *
84  * (**) There is a small time when bit 0 is set but there are no
85  * waiters. This can happen when grabbing the lock in the slow path.
86  * To prevent a cmpxchg of the owner releasing the lock, we need to
87  * set this bit before looking at the lock.
88  */
89 
90 static __always_inline void
91 rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
92 {
93 	unsigned long val = (unsigned long)owner;
94 
95 	if (rt_mutex_has_waiters(lock))
96 		val |= RT_MUTEX_HAS_WAITERS;
97 
98 	WRITE_ONCE(lock->owner, (struct task_struct *)val);
99 }
100 
101 static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
102 {
103 	lock->owner = (struct task_struct *)
104 			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
105 }
106 
107 static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
108 {
109 	unsigned long owner, *p = (unsigned long *) &lock->owner;
110 
111 	if (rt_mutex_has_waiters(lock))
112 		return;
113 
114 	/*
115 	 * The rbtree has no waiters enqueued, now make sure that the
116 	 * lock->owner still has the waiters bit set, otherwise the
117 	 * following can happen:
118 	 *
119 	 * CPU 0	CPU 1		CPU2
120 	 * l->owner=T1
121 	 *		rt_mutex_lock(l)
122 	 *		lock(l->lock)
123 	 *		l->owner = T1 | HAS_WAITERS;
124 	 *		enqueue(T2)
125 	 *		boost()
126 	 *		  unlock(l->lock)
127 	 *		block()
128 	 *
129 	 *				rt_mutex_lock(l)
130 	 *				lock(l->lock)
131 	 *				l->owner = T1 | HAS_WAITERS;
132 	 *				enqueue(T3)
133 	 *				boost()
134 	 *				  unlock(l->lock)
135 	 *				block()
136 	 *		signal(->T2)	signal(->T3)
137 	 *		lock(l->lock)
138 	 *		dequeue(T2)
139 	 *		deboost()
140 	 *		  unlock(l->lock)
141 	 *				lock(l->lock)
142 	 *				dequeue(T3)
143 	 *				 ==> wait list is empty
144 	 *				deboost()
145 	 *				 unlock(l->lock)
146 	 *		lock(l->lock)
147 	 *		fixup_rt_mutex_waiters()
148 	 *		  if (wait_list_empty(l) {
149 	 *		    l->owner = owner
150 	 *		    owner = l->owner & ~HAS_WAITERS;
151 	 *		      ==> l->owner = T1
152 	 *		  }
153 	 *				lock(l->lock)
154 	 * rt_mutex_unlock(l)		fixup_rt_mutex_waiters()
155 	 *				  if (wait_list_empty(l) {
156 	 *				    owner = l->owner & ~HAS_WAITERS;
157 	 * cmpxchg(l->owner, T1, NULL)
158 	 *  ===> Success (l->owner = NULL)
159 	 *
160 	 *				    l->owner = owner
161 	 *				      ==> l->owner = T1
162 	 *				  }
163 	 *
164 	 * With the check for the waiter bit in place T3 on CPU2 will not
165 	 * overwrite. All tasks fiddling with the waiters bit are
166 	 * serialized by l->lock, so nothing else can modify the waiters
167 	 * bit. If the bit is set then nothing can change l->owner either
168 	 * so the simple RMW is safe. The cmpxchg() will simply fail if it
169 	 * happens in the middle of the RMW because the waiters bit is
170 	 * still set.
171 	 */
172 	owner = READ_ONCE(*p);
173 	if (owner & RT_MUTEX_HAS_WAITERS)
174 		WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
175 }
176 
177 /*
178  * We can speed up the acquire/release, if there's no debugging state to be
179  * set up.
180  */
181 #ifndef CONFIG_DEBUG_RT_MUTEXES
182 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
183 						     struct task_struct *old,
184 						     struct task_struct *new)
185 {
186 	return try_cmpxchg_acquire(&lock->owner, &old, new);
187 }
188 
189 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
190 						     struct task_struct *old,
191 						     struct task_struct *new)
192 {
193 	return try_cmpxchg_release(&lock->owner, &old, new);
194 }
195 
196 /*
197  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
198  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
199  * relaxed semantics suffice.
200  */
201 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
202 {
203 	unsigned long owner, *p = (unsigned long *) &lock->owner;
204 
205 	do {
206 		owner = *p;
207 	} while (cmpxchg_relaxed(p, owner,
208 				 owner | RT_MUTEX_HAS_WAITERS) != owner);
209 }
210 
211 /*
212  * Safe fastpath aware unlock:
213  * 1) Clear the waiters bit
214  * 2) Drop lock->wait_lock
215  * 3) Try to unlock the lock with cmpxchg
216  */
217 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
218 						 unsigned long flags)
219 	__releases(lock->wait_lock)
220 {
221 	struct task_struct *owner = rt_mutex_owner(lock);
222 
223 	clear_rt_mutex_waiters(lock);
224 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
225 	/*
226 	 * If a new waiter comes in between the unlock and the cmpxchg
227 	 * we have two situations:
228 	 *
229 	 * unlock(wait_lock);
230 	 *					lock(wait_lock);
231 	 * cmpxchg(p, owner, 0) == owner
232 	 *					mark_rt_mutex_waiters(lock);
233 	 *					acquire(lock);
234 	 * or:
235 	 *
236 	 * unlock(wait_lock);
237 	 *					lock(wait_lock);
238 	 *					mark_rt_mutex_waiters(lock);
239 	 *
240 	 * cmpxchg(p, owner, 0) != owner
241 	 *					enqueue_waiter();
242 	 *					unlock(wait_lock);
243 	 * lock(wait_lock);
244 	 * wake waiter();
245 	 * unlock(wait_lock);
246 	 *					lock(wait_lock);
247 	 *					acquire(lock);
248 	 */
249 	return rt_mutex_cmpxchg_release(lock, owner, NULL);
250 }
251 
252 #else
253 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
254 						     struct task_struct *old,
255 						     struct task_struct *new)
256 {
257 	return false;
258 
259 }
260 
261 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
262 						     struct task_struct *old,
263 						     struct task_struct *new)
264 {
265 	return false;
266 }
267 
268 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
269 {
270 	lock->owner = (struct task_struct *)
271 			((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
272 }
273 
274 /*
275  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
276  */
277 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
278 						 unsigned long flags)
279 	__releases(lock->wait_lock)
280 {
281 	lock->owner = NULL;
282 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
283 	return true;
284 }
285 #endif
286 
287 static __always_inline int __waiter_prio(struct task_struct *task)
288 {
289 	int prio = task->prio;
290 
291 	if (!rt_prio(prio))
292 		return DEFAULT_PRIO;
293 
294 	return prio;
295 }
296 
297 static __always_inline void
298 waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
299 {
300 	waiter->prio = __waiter_prio(task);
301 	waiter->deadline = task->dl.deadline;
302 }
303 
304 /*
305  * Only use with rt_mutex_waiter_{less,equal}()
306  */
307 #define task_to_waiter(p)	\
308 	&(struct rt_mutex_waiter){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
309 
310 static __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
311 						struct rt_mutex_waiter *right)
312 {
313 	if (left->prio < right->prio)
314 		return 1;
315 
316 	/*
317 	 * If both waiters have dl_prio(), we check the deadlines of the
318 	 * associated tasks.
319 	 * If left waiter has a dl_prio(), and we didn't return 1 above,
320 	 * then right waiter has a dl_prio() too.
321 	 */
322 	if (dl_prio(left->prio))
323 		return dl_time_before(left->deadline, right->deadline);
324 
325 	return 0;
326 }
327 
328 static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
329 						 struct rt_mutex_waiter *right)
330 {
331 	if (left->prio != right->prio)
332 		return 0;
333 
334 	/*
335 	 * If both waiters have dl_prio(), we check the deadlines of the
336 	 * associated tasks.
337 	 * If left waiter has a dl_prio(), and we didn't return 0 above,
338 	 * then right waiter has a dl_prio() too.
339 	 */
340 	if (dl_prio(left->prio))
341 		return left->deadline == right->deadline;
342 
343 	return 1;
344 }
345 
346 static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
347 				  struct rt_mutex_waiter *top_waiter)
348 {
349 	if (rt_mutex_waiter_less(waiter, top_waiter))
350 		return true;
351 
352 #ifdef RT_MUTEX_BUILD_SPINLOCKS
353 	/*
354 	 * Note that RT tasks are excluded from same priority (lateral)
355 	 * steals to prevent the introduction of an unbounded latency.
356 	 */
357 	if (rt_prio(waiter->prio) || dl_prio(waiter->prio))
358 		return false;
359 
360 	return rt_mutex_waiter_equal(waiter, top_waiter);
361 #else
362 	return false;
363 #endif
364 }
365 
366 #define __node_2_waiter(node) \
367 	rb_entry((node), struct rt_mutex_waiter, tree_entry)
368 
369 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
370 {
371 	struct rt_mutex_waiter *aw = __node_2_waiter(a);
372 	struct rt_mutex_waiter *bw = __node_2_waiter(b);
373 
374 	if (rt_mutex_waiter_less(aw, bw))
375 		return 1;
376 
377 	if (!build_ww_mutex())
378 		return 0;
379 
380 	if (rt_mutex_waiter_less(bw, aw))
381 		return 0;
382 
383 	/* NOTE: relies on waiter->ww_ctx being set before insertion */
384 	if (aw->ww_ctx) {
385 		if (!bw->ww_ctx)
386 			return 1;
387 
388 		return (signed long)(aw->ww_ctx->stamp -
389 				     bw->ww_ctx->stamp) < 0;
390 	}
391 
392 	return 0;
393 }
394 
395 static __always_inline void
396 rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
397 {
398 	rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
399 }
400 
401 static __always_inline void
402 rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
403 {
404 	if (RB_EMPTY_NODE(&waiter->tree_entry))
405 		return;
406 
407 	rb_erase_cached(&waiter->tree_entry, &lock->waiters);
408 	RB_CLEAR_NODE(&waiter->tree_entry);
409 }
410 
411 #define __node_2_pi_waiter(node) \
412 	rb_entry((node), struct rt_mutex_waiter, pi_tree_entry)
413 
414 static __always_inline bool
415 __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
416 {
417 	return rt_mutex_waiter_less(__node_2_pi_waiter(a), __node_2_pi_waiter(b));
418 }
419 
420 static __always_inline void
421 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
422 {
423 	rb_add_cached(&waiter->pi_tree_entry, &task->pi_waiters, __pi_waiter_less);
424 }
425 
426 static __always_inline void
427 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
428 {
429 	if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
430 		return;
431 
432 	rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
433 	RB_CLEAR_NODE(&waiter->pi_tree_entry);
434 }
435 
436 static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
437 {
438 	struct task_struct *pi_task = NULL;
439 
440 	lockdep_assert_held(&p->pi_lock);
441 
442 	if (task_has_pi_waiters(p))
443 		pi_task = task_top_pi_waiter(p)->task;
444 
445 	rt_mutex_setprio(p, pi_task);
446 }
447 
448 /* RT mutex specific wake_q wrappers */
449 static __always_inline void rt_mutex_wake_q_add_task(struct rt_wake_q_head *wqh,
450 						     struct task_struct *task,
451 						     unsigned int wake_state)
452 {
453 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wake_state == TASK_RTLOCK_WAIT) {
454 		if (IS_ENABLED(CONFIG_PROVE_LOCKING))
455 			WARN_ON_ONCE(wqh->rtlock_task);
456 		get_task_struct(task);
457 		wqh->rtlock_task = task;
458 	} else {
459 		wake_q_add(&wqh->head, task);
460 	}
461 }
462 
463 static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
464 						struct rt_mutex_waiter *w)
465 {
466 	rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
467 }
468 
469 static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
470 {
471 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
472 		wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
473 		put_task_struct(wqh->rtlock_task);
474 		wqh->rtlock_task = NULL;
475 	}
476 
477 	if (!wake_q_empty(&wqh->head))
478 		wake_up_q(&wqh->head);
479 
480 	/* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
481 	preempt_enable();
482 }
483 
484 /*
485  * Deadlock detection is conditional:
486  *
487  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
488  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
489  *
490  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
491  * conducted independent of the detect argument.
492  *
493  * If the waiter argument is NULL this indicates the deboost path and
494  * deadlock detection is disabled independent of the detect argument
495  * and the config settings.
496  */
497 static __always_inline bool
498 rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
499 			      enum rtmutex_chainwalk chwalk)
500 {
501 	if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
502 		return waiter != NULL;
503 	return chwalk == RT_MUTEX_FULL_CHAINWALK;
504 }
505 
506 static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
507 {
508 	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
509 }
510 
511 /*
512  * Adjust the priority chain. Also used for deadlock detection.
513  * Decreases task's usage by one - may thus free the task.
514  *
515  * @task:	the task owning the mutex (owner) for which a chain walk is
516  *		probably needed
517  * @chwalk:	do we have to carry out deadlock detection?
518  * @orig_lock:	the mutex (can be NULL if we are walking the chain to recheck
519  *		things for a task that has just got its priority adjusted, and
520  *		is waiting on a mutex)
521  * @next_lock:	the mutex on which the owner of @orig_lock was blocked before
522  *		we dropped its pi_lock. Is never dereferenced, only used for
523  *		comparison to detect lock chain changes.
524  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
525  *		its priority to the mutex owner (can be NULL in the case
526  *		depicted above or if the top waiter is gone away and we are
527  *		actually deboosting the owner)
528  * @top_task:	the current top waiter
529  *
530  * Returns 0 or -EDEADLK.
531  *
532  * Chain walk basics and protection scope
533  *
534  * [R] refcount on task
535  * [P] task->pi_lock held
536  * [L] rtmutex->wait_lock held
537  *
538  * Step	Description				Protected by
539  *	function arguments:
540  *	@task					[R]
541  *	@orig_lock if != NULL			@top_task is blocked on it
542  *	@next_lock				Unprotected. Cannot be
543  *						dereferenced. Only used for
544  *						comparison.
545  *	@orig_waiter if != NULL			@top_task is blocked on it
546  *	@top_task				current, or in case of proxy
547  *						locking protected by calling
548  *						code
549  *	again:
550  *	  loop_sanity_check();
551  *	retry:
552  * [1]	  lock(task->pi_lock);			[R] acquire [P]
553  * [2]	  waiter = task->pi_blocked_on;		[P]
554  * [3]	  check_exit_conditions_1();		[P]
555  * [4]	  lock = waiter->lock;			[P]
556  * [5]	  if (!try_lock(lock->wait_lock)) {	[P] try to acquire [L]
557  *	    unlock(task->pi_lock);		release [P]
558  *	    goto retry;
559  *	  }
560  * [6]	  check_exit_conditions_2();		[P] + [L]
561  * [7]	  requeue_lock_waiter(lock, waiter);	[P] + [L]
562  * [8]	  unlock(task->pi_lock);		release [P]
563  *	  put_task_struct(task);		release [R]
564  * [9]	  check_exit_conditions_3();		[L]
565  * [10]	  task = owner(lock);			[L]
566  *	  get_task_struct(task);		[L] acquire [R]
567  *	  lock(task->pi_lock);			[L] acquire [P]
568  * [11]	  requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
569  * [12]	  check_exit_conditions_4();		[P] + [L]
570  * [13]	  unlock(task->pi_lock);		release [P]
571  *	  unlock(lock->wait_lock);		release [L]
572  *	  goto again;
573  */
574 static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
575 					      enum rtmutex_chainwalk chwalk,
576 					      struct rt_mutex_base *orig_lock,
577 					      struct rt_mutex_base *next_lock,
578 					      struct rt_mutex_waiter *orig_waiter,
579 					      struct task_struct *top_task)
580 {
581 	struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
582 	struct rt_mutex_waiter *prerequeue_top_waiter;
583 	int ret = 0, depth = 0;
584 	struct rt_mutex_base *lock;
585 	bool detect_deadlock;
586 	bool requeue = true;
587 
588 	detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
589 
590 	/*
591 	 * The (de)boosting is a step by step approach with a lot of
592 	 * pitfalls. We want this to be preemptible and we want hold a
593 	 * maximum of two locks per step. So we have to check
594 	 * carefully whether things change under us.
595 	 */
596  again:
597 	/*
598 	 * We limit the lock chain length for each invocation.
599 	 */
600 	if (++depth > max_lock_depth) {
601 		static int prev_max;
602 
603 		/*
604 		 * Print this only once. If the admin changes the limit,
605 		 * print a new message when reaching the limit again.
606 		 */
607 		if (prev_max != max_lock_depth) {
608 			prev_max = max_lock_depth;
609 			printk(KERN_WARNING "Maximum lock depth %d reached "
610 			       "task: %s (%d)\n", max_lock_depth,
611 			       top_task->comm, task_pid_nr(top_task));
612 		}
613 		put_task_struct(task);
614 
615 		return -EDEADLK;
616 	}
617 
618 	/*
619 	 * We are fully preemptible here and only hold the refcount on
620 	 * @task. So everything can have changed under us since the
621 	 * caller or our own code below (goto retry/again) dropped all
622 	 * locks.
623 	 */
624  retry:
625 	/*
626 	 * [1] Task cannot go away as we did a get_task() before !
627 	 */
628 	raw_spin_lock_irq(&task->pi_lock);
629 
630 	/*
631 	 * [2] Get the waiter on which @task is blocked on.
632 	 */
633 	waiter = task->pi_blocked_on;
634 
635 	/*
636 	 * [3] check_exit_conditions_1() protected by task->pi_lock.
637 	 */
638 
639 	/*
640 	 * Check whether the end of the boosting chain has been
641 	 * reached or the state of the chain has changed while we
642 	 * dropped the locks.
643 	 */
644 	if (!waiter)
645 		goto out_unlock_pi;
646 
647 	/*
648 	 * Check the orig_waiter state. After we dropped the locks,
649 	 * the previous owner of the lock might have released the lock.
650 	 */
651 	if (orig_waiter && !rt_mutex_owner(orig_lock))
652 		goto out_unlock_pi;
653 
654 	/*
655 	 * We dropped all locks after taking a refcount on @task, so
656 	 * the task might have moved on in the lock chain or even left
657 	 * the chain completely and blocks now on an unrelated lock or
658 	 * on @orig_lock.
659 	 *
660 	 * We stored the lock on which @task was blocked in @next_lock,
661 	 * so we can detect the chain change.
662 	 */
663 	if (next_lock != waiter->lock)
664 		goto out_unlock_pi;
665 
666 	/*
667 	 * There could be 'spurious' loops in the lock graph due to ww_mutex,
668 	 * consider:
669 	 *
670 	 *   P1: A, ww_A, ww_B
671 	 *   P2: ww_B, ww_A
672 	 *   P3: A
673 	 *
674 	 * P3 should not return -EDEADLK because it gets trapped in the cycle
675 	 * created by P1 and P2 (which will resolve -- and runs into
676 	 * max_lock_depth above). Therefore disable detect_deadlock such that
677 	 * the below termination condition can trigger once all relevant tasks
678 	 * are boosted.
679 	 *
680 	 * Even when we start with ww_mutex we can disable deadlock detection,
681 	 * since we would supress a ww_mutex induced deadlock at [6] anyway.
682 	 * Supressing it here however is not sufficient since we might still
683 	 * hit [6] due to adjustment driven iteration.
684 	 *
685 	 * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
686 	 * utterly fail to report it; lockdep should.
687 	 */
688 	if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
689 		detect_deadlock = false;
690 
691 	/*
692 	 * Drop out, when the task has no waiters. Note,
693 	 * top_waiter can be NULL, when we are in the deboosting
694 	 * mode!
695 	 */
696 	if (top_waiter) {
697 		if (!task_has_pi_waiters(task))
698 			goto out_unlock_pi;
699 		/*
700 		 * If deadlock detection is off, we stop here if we
701 		 * are not the top pi waiter of the task. If deadlock
702 		 * detection is enabled we continue, but stop the
703 		 * requeueing in the chain walk.
704 		 */
705 		if (top_waiter != task_top_pi_waiter(task)) {
706 			if (!detect_deadlock)
707 				goto out_unlock_pi;
708 			else
709 				requeue = false;
710 		}
711 	}
712 
713 	/*
714 	 * If the waiter priority is the same as the task priority
715 	 * then there is no further priority adjustment necessary.  If
716 	 * deadlock detection is off, we stop the chain walk. If its
717 	 * enabled we continue, but stop the requeueing in the chain
718 	 * walk.
719 	 */
720 	if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
721 		if (!detect_deadlock)
722 			goto out_unlock_pi;
723 		else
724 			requeue = false;
725 	}
726 
727 	/*
728 	 * [4] Get the next lock
729 	 */
730 	lock = waiter->lock;
731 	/*
732 	 * [5] We need to trylock here as we are holding task->pi_lock,
733 	 * which is the reverse lock order versus the other rtmutex
734 	 * operations.
735 	 */
736 	if (!raw_spin_trylock(&lock->wait_lock)) {
737 		raw_spin_unlock_irq(&task->pi_lock);
738 		cpu_relax();
739 		goto retry;
740 	}
741 
742 	/*
743 	 * [6] check_exit_conditions_2() protected by task->pi_lock and
744 	 * lock->wait_lock.
745 	 *
746 	 * Deadlock detection. If the lock is the same as the original
747 	 * lock which caused us to walk the lock chain or if the
748 	 * current lock is owned by the task which initiated the chain
749 	 * walk, we detected a deadlock.
750 	 */
751 	if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
752 		ret = -EDEADLK;
753 
754 		/*
755 		 * When the deadlock is due to ww_mutex; also see above. Don't
756 		 * report the deadlock and instead let the ww_mutex wound/die
757 		 * logic pick which of the contending threads gets -EDEADLK.
758 		 *
759 		 * NOTE: assumes the cycle only contains a single ww_class; any
760 		 * other configuration and we fail to report; also, see
761 		 * lockdep.
762 		 */
763 		if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
764 			ret = 0;
765 
766 		raw_spin_unlock(&lock->wait_lock);
767 		goto out_unlock_pi;
768 	}
769 
770 	/*
771 	 * If we just follow the lock chain for deadlock detection, no
772 	 * need to do all the requeue operations. To avoid a truckload
773 	 * of conditionals around the various places below, just do the
774 	 * minimum chain walk checks.
775 	 */
776 	if (!requeue) {
777 		/*
778 		 * No requeue[7] here. Just release @task [8]
779 		 */
780 		raw_spin_unlock(&task->pi_lock);
781 		put_task_struct(task);
782 
783 		/*
784 		 * [9] check_exit_conditions_3 protected by lock->wait_lock.
785 		 * If there is no owner of the lock, end of chain.
786 		 */
787 		if (!rt_mutex_owner(lock)) {
788 			raw_spin_unlock_irq(&lock->wait_lock);
789 			return 0;
790 		}
791 
792 		/* [10] Grab the next task, i.e. owner of @lock */
793 		task = get_task_struct(rt_mutex_owner(lock));
794 		raw_spin_lock(&task->pi_lock);
795 
796 		/*
797 		 * No requeue [11] here. We just do deadlock detection.
798 		 *
799 		 * [12] Store whether owner is blocked
800 		 * itself. Decision is made after dropping the locks
801 		 */
802 		next_lock = task_blocked_on_lock(task);
803 		/*
804 		 * Get the top waiter for the next iteration
805 		 */
806 		top_waiter = rt_mutex_top_waiter(lock);
807 
808 		/* [13] Drop locks */
809 		raw_spin_unlock(&task->pi_lock);
810 		raw_spin_unlock_irq(&lock->wait_lock);
811 
812 		/* If owner is not blocked, end of chain. */
813 		if (!next_lock)
814 			goto out_put_task;
815 		goto again;
816 	}
817 
818 	/*
819 	 * Store the current top waiter before doing the requeue
820 	 * operation on @lock. We need it for the boost/deboost
821 	 * decision below.
822 	 */
823 	prerequeue_top_waiter = rt_mutex_top_waiter(lock);
824 
825 	/* [7] Requeue the waiter in the lock waiter tree. */
826 	rt_mutex_dequeue(lock, waiter);
827 
828 	/*
829 	 * Update the waiter prio fields now that we're dequeued.
830 	 *
831 	 * These values can have changed through either:
832 	 *
833 	 *   sys_sched_set_scheduler() / sys_sched_setattr()
834 	 *
835 	 * or
836 	 *
837 	 *   DL CBS enforcement advancing the effective deadline.
838 	 *
839 	 * Even though pi_waiters also uses these fields, and that tree is only
840 	 * updated in [11], we can do this here, since we hold [L], which
841 	 * serializes all pi_waiters access and rb_erase() does not care about
842 	 * the values of the node being removed.
843 	 */
844 	waiter_update_prio(waiter, task);
845 
846 	rt_mutex_enqueue(lock, waiter);
847 
848 	/* [8] Release the task */
849 	raw_spin_unlock(&task->pi_lock);
850 	put_task_struct(task);
851 
852 	/*
853 	 * [9] check_exit_conditions_3 protected by lock->wait_lock.
854 	 *
855 	 * We must abort the chain walk if there is no lock owner even
856 	 * in the dead lock detection case, as we have nothing to
857 	 * follow here. This is the end of the chain we are walking.
858 	 */
859 	if (!rt_mutex_owner(lock)) {
860 		/*
861 		 * If the requeue [7] above changed the top waiter,
862 		 * then we need to wake the new top waiter up to try
863 		 * to get the lock.
864 		 */
865 		if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
866 			wake_up_state(waiter->task, waiter->wake_state);
867 		raw_spin_unlock_irq(&lock->wait_lock);
868 		return 0;
869 	}
870 
871 	/* [10] Grab the next task, i.e. the owner of @lock */
872 	task = get_task_struct(rt_mutex_owner(lock));
873 	raw_spin_lock(&task->pi_lock);
874 
875 	/* [11] requeue the pi waiters if necessary */
876 	if (waiter == rt_mutex_top_waiter(lock)) {
877 		/*
878 		 * The waiter became the new top (highest priority)
879 		 * waiter on the lock. Replace the previous top waiter
880 		 * in the owner tasks pi waiters tree with this waiter
881 		 * and adjust the priority of the owner.
882 		 */
883 		rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
884 		rt_mutex_enqueue_pi(task, waiter);
885 		rt_mutex_adjust_prio(task);
886 
887 	} else if (prerequeue_top_waiter == waiter) {
888 		/*
889 		 * The waiter was the top waiter on the lock, but is
890 		 * no longer the top priority waiter. Replace waiter in
891 		 * the owner tasks pi waiters tree with the new top
892 		 * (highest priority) waiter and adjust the priority
893 		 * of the owner.
894 		 * The new top waiter is stored in @waiter so that
895 		 * @waiter == @top_waiter evaluates to true below and
896 		 * we continue to deboost the rest of the chain.
897 		 */
898 		rt_mutex_dequeue_pi(task, waiter);
899 		waiter = rt_mutex_top_waiter(lock);
900 		rt_mutex_enqueue_pi(task, waiter);
901 		rt_mutex_adjust_prio(task);
902 	} else {
903 		/*
904 		 * Nothing changed. No need to do any priority
905 		 * adjustment.
906 		 */
907 	}
908 
909 	/*
910 	 * [12] check_exit_conditions_4() protected by task->pi_lock
911 	 * and lock->wait_lock. The actual decisions are made after we
912 	 * dropped the locks.
913 	 *
914 	 * Check whether the task which owns the current lock is pi
915 	 * blocked itself. If yes we store a pointer to the lock for
916 	 * the lock chain change detection above. After we dropped
917 	 * task->pi_lock next_lock cannot be dereferenced anymore.
918 	 */
919 	next_lock = task_blocked_on_lock(task);
920 	/*
921 	 * Store the top waiter of @lock for the end of chain walk
922 	 * decision below.
923 	 */
924 	top_waiter = rt_mutex_top_waiter(lock);
925 
926 	/* [13] Drop the locks */
927 	raw_spin_unlock(&task->pi_lock);
928 	raw_spin_unlock_irq(&lock->wait_lock);
929 
930 	/*
931 	 * Make the actual exit decisions [12], based on the stored
932 	 * values.
933 	 *
934 	 * We reached the end of the lock chain. Stop right here. No
935 	 * point to go back just to figure that out.
936 	 */
937 	if (!next_lock)
938 		goto out_put_task;
939 
940 	/*
941 	 * If the current waiter is not the top waiter on the lock,
942 	 * then we can stop the chain walk here if we are not in full
943 	 * deadlock detection mode.
944 	 */
945 	if (!detect_deadlock && waiter != top_waiter)
946 		goto out_put_task;
947 
948 	goto again;
949 
950  out_unlock_pi:
951 	raw_spin_unlock_irq(&task->pi_lock);
952  out_put_task:
953 	put_task_struct(task);
954 
955 	return ret;
956 }
957 
958 /*
959  * Try to take an rt-mutex
960  *
961  * Must be called with lock->wait_lock held and interrupts disabled
962  *
963  * @lock:   The lock to be acquired.
964  * @task:   The task which wants to acquire the lock
965  * @waiter: The waiter that is queued to the lock's wait tree if the
966  *	    callsite called task_blocked_on_lock(), otherwise NULL
967  */
968 static int __sched
969 try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
970 		     struct rt_mutex_waiter *waiter)
971 {
972 	lockdep_assert_held(&lock->wait_lock);
973 
974 	/*
975 	 * Before testing whether we can acquire @lock, we set the
976 	 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
977 	 * other tasks which try to modify @lock into the slow path
978 	 * and they serialize on @lock->wait_lock.
979 	 *
980 	 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
981 	 * as explained at the top of this file if and only if:
982 	 *
983 	 * - There is a lock owner. The caller must fixup the
984 	 *   transient state if it does a trylock or leaves the lock
985 	 *   function due to a signal or timeout.
986 	 *
987 	 * - @task acquires the lock and there are no other
988 	 *   waiters. This is undone in rt_mutex_set_owner(@task) at
989 	 *   the end of this function.
990 	 */
991 	mark_rt_mutex_waiters(lock);
992 
993 	/*
994 	 * If @lock has an owner, give up.
995 	 */
996 	if (rt_mutex_owner(lock))
997 		return 0;
998 
999 	/*
1000 	 * If @waiter != NULL, @task has already enqueued the waiter
1001 	 * into @lock waiter tree. If @waiter == NULL then this is a
1002 	 * trylock attempt.
1003 	 */
1004 	if (waiter) {
1005 		struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
1006 
1007 		/*
1008 		 * If waiter is the highest priority waiter of @lock,
1009 		 * or allowed to steal it, take it over.
1010 		 */
1011 		if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1012 			/*
1013 			 * We can acquire the lock. Remove the waiter from the
1014 			 * lock waiters tree.
1015 			 */
1016 			rt_mutex_dequeue(lock, waiter);
1017 		} else {
1018 			return 0;
1019 		}
1020 	} else {
1021 		/*
1022 		 * If the lock has waiters already we check whether @task is
1023 		 * eligible to take over the lock.
1024 		 *
1025 		 * If there are no other waiters, @task can acquire
1026 		 * the lock.  @task->pi_blocked_on is NULL, so it does
1027 		 * not need to be dequeued.
1028 		 */
1029 		if (rt_mutex_has_waiters(lock)) {
1030 			/* Check whether the trylock can steal it. */
1031 			if (!rt_mutex_steal(task_to_waiter(task),
1032 					    rt_mutex_top_waiter(lock)))
1033 				return 0;
1034 
1035 			/*
1036 			 * The current top waiter stays enqueued. We
1037 			 * don't have to change anything in the lock
1038 			 * waiters order.
1039 			 */
1040 		} else {
1041 			/*
1042 			 * No waiters. Take the lock without the
1043 			 * pi_lock dance.@task->pi_blocked_on is NULL
1044 			 * and we have no waiters to enqueue in @task
1045 			 * pi waiters tree.
1046 			 */
1047 			goto takeit;
1048 		}
1049 	}
1050 
1051 	/*
1052 	 * Clear @task->pi_blocked_on. Requires protection by
1053 	 * @task->pi_lock. Redundant operation for the @waiter == NULL
1054 	 * case, but conditionals are more expensive than a redundant
1055 	 * store.
1056 	 */
1057 	raw_spin_lock(&task->pi_lock);
1058 	task->pi_blocked_on = NULL;
1059 	/*
1060 	 * Finish the lock acquisition. @task is the new owner. If
1061 	 * other waiters exist we have to insert the highest priority
1062 	 * waiter into @task->pi_waiters tree.
1063 	 */
1064 	if (rt_mutex_has_waiters(lock))
1065 		rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
1066 	raw_spin_unlock(&task->pi_lock);
1067 
1068 takeit:
1069 	/*
1070 	 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1071 	 * are still waiters or clears it.
1072 	 */
1073 	rt_mutex_set_owner(lock, task);
1074 
1075 	return 1;
1076 }
1077 
1078 /*
1079  * Task blocks on lock.
1080  *
1081  * Prepare waiter and propagate pi chain
1082  *
1083  * This must be called with lock->wait_lock held and interrupts disabled
1084  */
1085 static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
1086 					   struct rt_mutex_waiter *waiter,
1087 					   struct task_struct *task,
1088 					   struct ww_acquire_ctx *ww_ctx,
1089 					   enum rtmutex_chainwalk chwalk)
1090 {
1091 	struct task_struct *owner = rt_mutex_owner(lock);
1092 	struct rt_mutex_waiter *top_waiter = waiter;
1093 	struct rt_mutex_base *next_lock;
1094 	int chain_walk = 0, res;
1095 
1096 	lockdep_assert_held(&lock->wait_lock);
1097 
1098 	/*
1099 	 * Early deadlock detection. We really don't want the task to
1100 	 * enqueue on itself just to untangle the mess later. It's not
1101 	 * only an optimization. We drop the locks, so another waiter
1102 	 * can come in before the chain walk detects the deadlock. So
1103 	 * the other will detect the deadlock and return -EDEADLOCK,
1104 	 * which is wrong, as the other waiter is not in a deadlock
1105 	 * situation.
1106 	 *
1107 	 * Except for ww_mutex, in that case the chain walk must already deal
1108 	 * with spurious cycles, see the comments at [3] and [6].
1109 	 */
1110 	if (owner == task && !(build_ww_mutex() && ww_ctx))
1111 		return -EDEADLK;
1112 
1113 	raw_spin_lock(&task->pi_lock);
1114 	waiter->task = task;
1115 	waiter->lock = lock;
1116 	waiter_update_prio(waiter, task);
1117 
1118 	/* Get the top priority waiter on the lock */
1119 	if (rt_mutex_has_waiters(lock))
1120 		top_waiter = rt_mutex_top_waiter(lock);
1121 	rt_mutex_enqueue(lock, waiter);
1122 
1123 	task->pi_blocked_on = waiter;
1124 
1125 	raw_spin_unlock(&task->pi_lock);
1126 
1127 	if (build_ww_mutex() && ww_ctx) {
1128 		struct rt_mutex *rtm;
1129 
1130 		/* Check whether the waiter should back out immediately */
1131 		rtm = container_of(lock, struct rt_mutex, rtmutex);
1132 		res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx);
1133 		if (res) {
1134 			raw_spin_lock(&task->pi_lock);
1135 			rt_mutex_dequeue(lock, waiter);
1136 			task->pi_blocked_on = NULL;
1137 			raw_spin_unlock(&task->pi_lock);
1138 			return res;
1139 		}
1140 	}
1141 
1142 	if (!owner)
1143 		return 0;
1144 
1145 	raw_spin_lock(&owner->pi_lock);
1146 	if (waiter == rt_mutex_top_waiter(lock)) {
1147 		rt_mutex_dequeue_pi(owner, top_waiter);
1148 		rt_mutex_enqueue_pi(owner, waiter);
1149 
1150 		rt_mutex_adjust_prio(owner);
1151 		if (owner->pi_blocked_on)
1152 			chain_walk = 1;
1153 	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1154 		chain_walk = 1;
1155 	}
1156 
1157 	/* Store the lock on which owner is blocked or NULL */
1158 	next_lock = task_blocked_on_lock(owner);
1159 
1160 	raw_spin_unlock(&owner->pi_lock);
1161 	/*
1162 	 * Even if full deadlock detection is on, if the owner is not
1163 	 * blocked itself, we can avoid finding this out in the chain
1164 	 * walk.
1165 	 */
1166 	if (!chain_walk || !next_lock)
1167 		return 0;
1168 
1169 	/*
1170 	 * The owner can't disappear while holding a lock,
1171 	 * so the owner struct is protected by wait_lock.
1172 	 * Gets dropped in rt_mutex_adjust_prio_chain()!
1173 	 */
1174 	get_task_struct(owner);
1175 
1176 	raw_spin_unlock_irq(&lock->wait_lock);
1177 
1178 	res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1179 					 next_lock, waiter, task);
1180 
1181 	raw_spin_lock_irq(&lock->wait_lock);
1182 
1183 	return res;
1184 }
1185 
1186 /*
1187  * Remove the top waiter from the current tasks pi waiter tree and
1188  * queue it up.
1189  *
1190  * Called with lock->wait_lock held and interrupts disabled.
1191  */
1192 static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1193 					    struct rt_mutex_base *lock)
1194 {
1195 	struct rt_mutex_waiter *waiter;
1196 
1197 	raw_spin_lock(&current->pi_lock);
1198 
1199 	waiter = rt_mutex_top_waiter(lock);
1200 
1201 	/*
1202 	 * Remove it from current->pi_waiters and deboost.
1203 	 *
1204 	 * We must in fact deboost here in order to ensure we call
1205 	 * rt_mutex_setprio() to update p->pi_top_task before the
1206 	 * task unblocks.
1207 	 */
1208 	rt_mutex_dequeue_pi(current, waiter);
1209 	rt_mutex_adjust_prio(current);
1210 
1211 	/*
1212 	 * As we are waking up the top waiter, and the waiter stays
1213 	 * queued on the lock until it gets the lock, this lock
1214 	 * obviously has waiters. Just set the bit here and this has
1215 	 * the added benefit of forcing all new tasks into the
1216 	 * slow path making sure no task of lower priority than
1217 	 * the top waiter can steal this lock.
1218 	 */
1219 	lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1220 
1221 	/*
1222 	 * We deboosted before waking the top waiter task such that we don't
1223 	 * run two tasks with the 'same' priority (and ensure the
1224 	 * p->pi_top_task pointer points to a blocked task). This however can
1225 	 * lead to priority inversion if we would get preempted after the
1226 	 * deboost but before waking our donor task, hence the preempt_disable()
1227 	 * before unlock.
1228 	 *
1229 	 * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1230 	 */
1231 	preempt_disable();
1232 	rt_mutex_wake_q_add(wqh, waiter);
1233 	raw_spin_unlock(&current->pi_lock);
1234 }
1235 
1236 static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1237 {
1238 	int ret = try_to_take_rt_mutex(lock, current, NULL);
1239 
1240 	/*
1241 	 * try_to_take_rt_mutex() sets the lock waiters bit
1242 	 * unconditionally. Clean this up.
1243 	 */
1244 	fixup_rt_mutex_waiters(lock);
1245 
1246 	return ret;
1247 }
1248 
1249 /*
1250  * Slow path try-lock function:
1251  */
1252 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1253 {
1254 	unsigned long flags;
1255 	int ret;
1256 
1257 	/*
1258 	 * If the lock already has an owner we fail to get the lock.
1259 	 * This can be done without taking the @lock->wait_lock as
1260 	 * it is only being read, and this is a trylock anyway.
1261 	 */
1262 	if (rt_mutex_owner(lock))
1263 		return 0;
1264 
1265 	/*
1266 	 * The mutex has currently no owner. Lock the wait lock and try to
1267 	 * acquire the lock. We use irqsave here to support early boot calls.
1268 	 */
1269 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1270 
1271 	ret = __rt_mutex_slowtrylock(lock);
1272 
1273 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1274 
1275 	return ret;
1276 }
1277 
1278 static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
1279 {
1280 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1281 		return 1;
1282 
1283 	return rt_mutex_slowtrylock(lock);
1284 }
1285 
1286 /*
1287  * Slow path to release a rt-mutex.
1288  */
1289 static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1290 {
1291 	DEFINE_RT_WAKE_Q(wqh);
1292 	unsigned long flags;
1293 
1294 	/* irqsave required to support early boot calls */
1295 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1296 
1297 	debug_rt_mutex_unlock(lock);
1298 
1299 	/*
1300 	 * We must be careful here if the fast path is enabled. If we
1301 	 * have no waiters queued we cannot set owner to NULL here
1302 	 * because of:
1303 	 *
1304 	 * foo->lock->owner = NULL;
1305 	 *			rtmutex_lock(foo->lock);   <- fast path
1306 	 *			free = atomic_dec_and_test(foo->refcnt);
1307 	 *			rtmutex_unlock(foo->lock); <- fast path
1308 	 *			if (free)
1309 	 *				kfree(foo);
1310 	 * raw_spin_unlock(foo->lock->wait_lock);
1311 	 *
1312 	 * So for the fastpath enabled kernel:
1313 	 *
1314 	 * Nothing can set the waiters bit as long as we hold
1315 	 * lock->wait_lock. So we do the following sequence:
1316 	 *
1317 	 *	owner = rt_mutex_owner(lock);
1318 	 *	clear_rt_mutex_waiters(lock);
1319 	 *	raw_spin_unlock(&lock->wait_lock);
1320 	 *	if (cmpxchg(&lock->owner, owner, 0) == owner)
1321 	 *		return;
1322 	 *	goto retry;
1323 	 *
1324 	 * The fastpath disabled variant is simple as all access to
1325 	 * lock->owner is serialized by lock->wait_lock:
1326 	 *
1327 	 *	lock->owner = NULL;
1328 	 *	raw_spin_unlock(&lock->wait_lock);
1329 	 */
1330 	while (!rt_mutex_has_waiters(lock)) {
1331 		/* Drops lock->wait_lock ! */
1332 		if (unlock_rt_mutex_safe(lock, flags) == true)
1333 			return;
1334 		/* Relock the rtmutex and try again */
1335 		raw_spin_lock_irqsave(&lock->wait_lock, flags);
1336 	}
1337 
1338 	/*
1339 	 * The wakeup next waiter path does not suffer from the above
1340 	 * race. See the comments there.
1341 	 *
1342 	 * Queue the next waiter for wakeup once we release the wait_lock.
1343 	 */
1344 	mark_wakeup_next_waiter(&wqh, lock);
1345 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1346 
1347 	rt_mutex_wake_up_q(&wqh);
1348 }
1349 
1350 static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
1351 {
1352 	if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1353 		return;
1354 
1355 	rt_mutex_slowunlock(lock);
1356 }
1357 
1358 #ifdef CONFIG_SMP
1359 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1360 				  struct rt_mutex_waiter *waiter,
1361 				  struct task_struct *owner)
1362 {
1363 	bool res = true;
1364 
1365 	rcu_read_lock();
1366 	for (;;) {
1367 		/* If owner changed, trylock again. */
1368 		if (owner != rt_mutex_owner(lock))
1369 			break;
1370 		/*
1371 		 * Ensure that @owner is dereferenced after checking that
1372 		 * the lock owner still matches @owner. If that fails,
1373 		 * @owner might point to freed memory. If it still matches,
1374 		 * the rcu_read_lock() ensures the memory stays valid.
1375 		 */
1376 		barrier();
1377 		/*
1378 		 * Stop spinning when:
1379 		 *  - the lock owner has been scheduled out
1380 		 *  - current is not longer the top waiter
1381 		 *  - current is requested to reschedule (redundant
1382 		 *    for CONFIG_PREEMPT_RCU=y)
1383 		 *  - the VCPU on which owner runs is preempted
1384 		 */
1385 		if (!owner_on_cpu(owner) || need_resched() ||
1386 		    !rt_mutex_waiter_is_top_waiter(lock, waiter)) {
1387 			res = false;
1388 			break;
1389 		}
1390 		cpu_relax();
1391 	}
1392 	rcu_read_unlock();
1393 	return res;
1394 }
1395 #else
1396 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1397 				  struct rt_mutex_waiter *waiter,
1398 				  struct task_struct *owner)
1399 {
1400 	return false;
1401 }
1402 #endif
1403 
1404 #ifdef RT_MUTEX_BUILD_MUTEX
1405 /*
1406  * Functions required for:
1407  *	- rtmutex, futex on all kernels
1408  *	- mutex and rwsem substitutions on RT kernels
1409  */
1410 
1411 /*
1412  * Remove a waiter from a lock and give up
1413  *
1414  * Must be called with lock->wait_lock held and interrupts disabled. It must
1415  * have just failed to try_to_take_rt_mutex().
1416  */
1417 static void __sched remove_waiter(struct rt_mutex_base *lock,
1418 				  struct rt_mutex_waiter *waiter)
1419 {
1420 	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1421 	struct task_struct *owner = rt_mutex_owner(lock);
1422 	struct rt_mutex_base *next_lock;
1423 
1424 	lockdep_assert_held(&lock->wait_lock);
1425 
1426 	raw_spin_lock(&current->pi_lock);
1427 	rt_mutex_dequeue(lock, waiter);
1428 	current->pi_blocked_on = NULL;
1429 	raw_spin_unlock(&current->pi_lock);
1430 
1431 	/*
1432 	 * Only update priority if the waiter was the highest priority
1433 	 * waiter of the lock and there is an owner to update.
1434 	 */
1435 	if (!owner || !is_top_waiter)
1436 		return;
1437 
1438 	raw_spin_lock(&owner->pi_lock);
1439 
1440 	rt_mutex_dequeue_pi(owner, waiter);
1441 
1442 	if (rt_mutex_has_waiters(lock))
1443 		rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1444 
1445 	rt_mutex_adjust_prio(owner);
1446 
1447 	/* Store the lock on which owner is blocked or NULL */
1448 	next_lock = task_blocked_on_lock(owner);
1449 
1450 	raw_spin_unlock(&owner->pi_lock);
1451 
1452 	/*
1453 	 * Don't walk the chain, if the owner task is not blocked
1454 	 * itself.
1455 	 */
1456 	if (!next_lock)
1457 		return;
1458 
1459 	/* gets dropped in rt_mutex_adjust_prio_chain()! */
1460 	get_task_struct(owner);
1461 
1462 	raw_spin_unlock_irq(&lock->wait_lock);
1463 
1464 	rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1465 				   next_lock, NULL, current);
1466 
1467 	raw_spin_lock_irq(&lock->wait_lock);
1468 }
1469 
1470 /**
1471  * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
1472  * @lock:		 the rt_mutex to take
1473  * @ww_ctx:		 WW mutex context pointer
1474  * @state:		 the state the task should block in (TASK_INTERRUPTIBLE
1475  *			 or TASK_UNINTERRUPTIBLE)
1476  * @timeout:		 the pre-initialized and started timer, or NULL for none
1477  * @waiter:		 the pre-initialized rt_mutex_waiter
1478  *
1479  * Must be called with lock->wait_lock held and interrupts disabled
1480  */
1481 static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1482 					   struct ww_acquire_ctx *ww_ctx,
1483 					   unsigned int state,
1484 					   struct hrtimer_sleeper *timeout,
1485 					   struct rt_mutex_waiter *waiter)
1486 {
1487 	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1488 	struct task_struct *owner;
1489 	int ret = 0;
1490 
1491 	for (;;) {
1492 		/* Try to acquire the lock: */
1493 		if (try_to_take_rt_mutex(lock, current, waiter))
1494 			break;
1495 
1496 		if (timeout && !timeout->task) {
1497 			ret = -ETIMEDOUT;
1498 			break;
1499 		}
1500 		if (signal_pending_state(state, current)) {
1501 			ret = -EINTR;
1502 			break;
1503 		}
1504 
1505 		if (build_ww_mutex() && ww_ctx) {
1506 			ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1507 			if (ret)
1508 				break;
1509 		}
1510 
1511 		if (waiter == rt_mutex_top_waiter(lock))
1512 			owner = rt_mutex_owner(lock);
1513 		else
1514 			owner = NULL;
1515 		raw_spin_unlock_irq(&lock->wait_lock);
1516 
1517 		if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1518 			schedule();
1519 
1520 		raw_spin_lock_irq(&lock->wait_lock);
1521 		set_current_state(state);
1522 	}
1523 
1524 	__set_current_state(TASK_RUNNING);
1525 	return ret;
1526 }
1527 
1528 static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1529 					     struct rt_mutex_waiter *w)
1530 {
1531 	/*
1532 	 * If the result is not -EDEADLOCK or the caller requested
1533 	 * deadlock detection, nothing to do here.
1534 	 */
1535 	if (res != -EDEADLOCK || detect_deadlock)
1536 		return;
1537 
1538 	if (build_ww_mutex() && w->ww_ctx)
1539 		return;
1540 
1541 	/*
1542 	 * Yell loudly and stop the task right here.
1543 	 */
1544 	WARN(1, "rtmutex deadlock detected\n");
1545 	while (1) {
1546 		set_current_state(TASK_INTERRUPTIBLE);
1547 		schedule();
1548 	}
1549 }
1550 
1551 /**
1552  * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
1553  * @lock:	The rtmutex to block lock
1554  * @ww_ctx:	WW mutex context pointer
1555  * @state:	The task state for sleeping
1556  * @chwalk:	Indicator whether full or partial chainwalk is requested
1557  * @waiter:	Initializer waiter for blocking
1558  */
1559 static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1560 				       struct ww_acquire_ctx *ww_ctx,
1561 				       unsigned int state,
1562 				       enum rtmutex_chainwalk chwalk,
1563 				       struct rt_mutex_waiter *waiter)
1564 {
1565 	struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1566 	struct ww_mutex *ww = ww_container_of(rtm);
1567 	int ret;
1568 
1569 	lockdep_assert_held(&lock->wait_lock);
1570 
1571 	/* Try to acquire the lock again: */
1572 	if (try_to_take_rt_mutex(lock, current, NULL)) {
1573 		if (build_ww_mutex() && ww_ctx) {
1574 			__ww_mutex_check_waiters(rtm, ww_ctx);
1575 			ww_mutex_lock_acquired(ww, ww_ctx);
1576 		}
1577 		return 0;
1578 	}
1579 
1580 	set_current_state(state);
1581 
1582 	ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
1583 	if (likely(!ret))
1584 		ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
1585 
1586 	if (likely(!ret)) {
1587 		/* acquired the lock */
1588 		if (build_ww_mutex() && ww_ctx) {
1589 			if (!ww_ctx->is_wait_die)
1590 				__ww_mutex_check_waiters(rtm, ww_ctx);
1591 			ww_mutex_lock_acquired(ww, ww_ctx);
1592 		}
1593 	} else {
1594 		__set_current_state(TASK_RUNNING);
1595 		remove_waiter(lock, waiter);
1596 		rt_mutex_handle_deadlock(ret, chwalk, waiter);
1597 	}
1598 
1599 	/*
1600 	 * try_to_take_rt_mutex() sets the waiter bit
1601 	 * unconditionally. We might have to fix that up.
1602 	 */
1603 	fixup_rt_mutex_waiters(lock);
1604 	return ret;
1605 }
1606 
1607 static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1608 					     struct ww_acquire_ctx *ww_ctx,
1609 					     unsigned int state)
1610 {
1611 	struct rt_mutex_waiter waiter;
1612 	int ret;
1613 
1614 	rt_mutex_init_waiter(&waiter);
1615 	waiter.ww_ctx = ww_ctx;
1616 
1617 	ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1618 				  &waiter);
1619 
1620 	debug_rt_mutex_free_waiter(&waiter);
1621 	return ret;
1622 }
1623 
1624 /*
1625  * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
1626  * @lock:	The rtmutex to block lock
1627  * @ww_ctx:	WW mutex context pointer
1628  * @state:	The task state for sleeping
1629  */
1630 static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1631 				     struct ww_acquire_ctx *ww_ctx,
1632 				     unsigned int state)
1633 {
1634 	unsigned long flags;
1635 	int ret;
1636 
1637 	/*
1638 	 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1639 	 * be called in early boot if the cmpxchg() fast path is disabled
1640 	 * (debug, no architecture support). In this case we will acquire the
1641 	 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1642 	 * enable interrupts in that early boot case. So we need to use the
1643 	 * irqsave/restore variants.
1644 	 */
1645 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1646 	ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state);
1647 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1648 
1649 	return ret;
1650 }
1651 
1652 static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1653 					   unsigned int state)
1654 {
1655 	if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1656 		return 0;
1657 
1658 	return rt_mutex_slowlock(lock, NULL, state);
1659 }
1660 #endif /* RT_MUTEX_BUILD_MUTEX */
1661 
1662 #ifdef RT_MUTEX_BUILD_SPINLOCKS
1663 /*
1664  * Functions required for spin/rw_lock substitution on RT kernels
1665  */
1666 
1667 /**
1668  * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1669  * @lock:	The underlying RT mutex
1670  */
1671 static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
1672 {
1673 	struct rt_mutex_waiter waiter;
1674 	struct task_struct *owner;
1675 
1676 	lockdep_assert_held(&lock->wait_lock);
1677 
1678 	if (try_to_take_rt_mutex(lock, current, NULL))
1679 		return;
1680 
1681 	rt_mutex_init_rtlock_waiter(&waiter);
1682 
1683 	/* Save current state and set state to TASK_RTLOCK_WAIT */
1684 	current_save_and_set_rtlock_wait_state();
1685 
1686 	task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
1687 
1688 	for (;;) {
1689 		/* Try to acquire the lock again */
1690 		if (try_to_take_rt_mutex(lock, current, &waiter))
1691 			break;
1692 
1693 		if (&waiter == rt_mutex_top_waiter(lock))
1694 			owner = rt_mutex_owner(lock);
1695 		else
1696 			owner = NULL;
1697 		raw_spin_unlock_irq(&lock->wait_lock);
1698 
1699 		if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1700 			schedule_rtlock();
1701 
1702 		raw_spin_lock_irq(&lock->wait_lock);
1703 		set_current_state(TASK_RTLOCK_WAIT);
1704 	}
1705 
1706 	/* Restore the task state */
1707 	current_restore_rtlock_saved_state();
1708 
1709 	/*
1710 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1711 	 * We might have to fix that up:
1712 	 */
1713 	fixup_rt_mutex_waiters(lock);
1714 	debug_rt_mutex_free_waiter(&waiter);
1715 }
1716 
1717 static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
1718 {
1719 	unsigned long flags;
1720 
1721 	raw_spin_lock_irqsave(&lock->wait_lock, flags);
1722 	rtlock_slowlock_locked(lock);
1723 	raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1724 }
1725 
1726 #endif /* RT_MUTEX_BUILD_SPINLOCKS */
1727