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 struct task_struct *
93 rt_mutex_owner_encode(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 return (struct task_struct *)val;
101 }
102
103 static __always_inline void
104 rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
105 {
106 /*
107 * lock->wait_lock is held but explicit acquire semantics are needed
108 * for a new lock owner so WRITE_ONCE is insufficient.
109 */
110 xchg_acquire(&lock->owner, rt_mutex_owner_encode(lock, owner));
111 }
112
113 static __always_inline void rt_mutex_clear_owner(struct rt_mutex_base *lock)
114 {
115 /* lock->wait_lock is held so the unlock provides release semantics. */
116 WRITE_ONCE(lock->owner, rt_mutex_owner_encode(lock, NULL));
117 }
118
119 static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
120 {
121 lock->owner = (struct task_struct *)
122 ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
123 }
124
125 static __always_inline void
126 fixup_rt_mutex_waiters(struct rt_mutex_base *lock, bool acquire_lock)
127 {
128 unsigned long owner, *p = (unsigned long *) &lock->owner;
129
130 if (rt_mutex_has_waiters(lock))
131 return;
132
133 /*
134 * The rbtree has no waiters enqueued, now make sure that the
135 * lock->owner still has the waiters bit set, otherwise the
136 * following can happen:
137 *
138 * CPU 0 CPU 1 CPU2
139 * l->owner=T1
140 * rt_mutex_lock(l)
141 * lock(l->lock)
142 * l->owner = T1 | HAS_WAITERS;
143 * enqueue(T2)
144 * boost()
145 * unlock(l->lock)
146 * block()
147 *
148 * rt_mutex_lock(l)
149 * lock(l->lock)
150 * l->owner = T1 | HAS_WAITERS;
151 * enqueue(T3)
152 * boost()
153 * unlock(l->lock)
154 * block()
155 * signal(->T2) signal(->T3)
156 * lock(l->lock)
157 * dequeue(T2)
158 * deboost()
159 * unlock(l->lock)
160 * lock(l->lock)
161 * dequeue(T3)
162 * ==> wait list is empty
163 * deboost()
164 * unlock(l->lock)
165 * lock(l->lock)
166 * fixup_rt_mutex_waiters()
167 * if (wait_list_empty(l) {
168 * l->owner = owner
169 * owner = l->owner & ~HAS_WAITERS;
170 * ==> l->owner = T1
171 * }
172 * lock(l->lock)
173 * rt_mutex_unlock(l) fixup_rt_mutex_waiters()
174 * if (wait_list_empty(l) {
175 * owner = l->owner & ~HAS_WAITERS;
176 * cmpxchg(l->owner, T1, NULL)
177 * ===> Success (l->owner = NULL)
178 *
179 * l->owner = owner
180 * ==> l->owner = T1
181 * }
182 *
183 * With the check for the waiter bit in place T3 on CPU2 will not
184 * overwrite. All tasks fiddling with the waiters bit are
185 * serialized by l->lock, so nothing else can modify the waiters
186 * bit. If the bit is set then nothing can change l->owner either
187 * so the simple RMW is safe. The cmpxchg() will simply fail if it
188 * happens in the middle of the RMW because the waiters bit is
189 * still set.
190 */
191 owner = READ_ONCE(*p);
192 if (owner & RT_MUTEX_HAS_WAITERS) {
193 /*
194 * See rt_mutex_set_owner() and rt_mutex_clear_owner() on
195 * why xchg_acquire() is used for updating owner for
196 * locking and WRITE_ONCE() for unlocking.
197 *
198 * WRITE_ONCE() would work for the acquire case too, but
199 * in case that the lock acquisition failed it might
200 * force other lockers into the slow path unnecessarily.
201 */
202 if (acquire_lock)
203 xchg_acquire(p, owner & ~RT_MUTEX_HAS_WAITERS);
204 else
205 WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
206 }
207 }
208
209 /*
210 * We can speed up the acquire/release, if there's no debugging state to be
211 * set up.
212 */
213 #ifndef CONFIG_DEBUG_RT_MUTEXES
214 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
215 struct task_struct *old,
216 struct task_struct *new)
217 {
218 return try_cmpxchg_acquire(&lock->owner, &old, new);
219 }
220
221 static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
222 {
223 return rt_mutex_cmpxchg_acquire(lock, NULL, current);
224 }
225
226 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
227 struct task_struct *old,
228 struct task_struct *new)
229 {
230 return try_cmpxchg_release(&lock->owner, &old, new);
231 }
232
233 /*
234 * Callers must hold the ->wait_lock -- which is the whole purpose as we force
235 * all future threads that attempt to [Rmw] the lock to the slowpath. As such
236 * relaxed semantics suffice.
237 */
238 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
239 {
240 unsigned long *p = (unsigned long *) &lock->owner;
241 unsigned long owner, new;
242
243 owner = READ_ONCE(*p);
244 do {
245 new = owner | RT_MUTEX_HAS_WAITERS;
246 } while (!try_cmpxchg_relaxed(p, &owner, new));
247
248 /*
249 * The cmpxchg loop above is relaxed to avoid back-to-back ACQUIRE
250 * operations in the event of contention. Ensure the successful
251 * cmpxchg is visible.
252 */
253 smp_mb__after_atomic();
254 }
255
256 /*
257 * Safe fastpath aware unlock:
258 * 1) Clear the waiters bit
259 * 2) Drop lock->wait_lock
260 * 3) Try to unlock the lock with cmpxchg
261 */
262 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
263 unsigned long flags)
264 __releases(lock->wait_lock)
265 {
266 struct task_struct *owner = rt_mutex_owner(lock);
267
268 clear_rt_mutex_waiters(lock);
269 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
270 /*
271 * If a new waiter comes in between the unlock and the cmpxchg
272 * we have two situations:
273 *
274 * unlock(wait_lock);
275 * lock(wait_lock);
276 * cmpxchg(p, owner, 0) == owner
277 * mark_rt_mutex_waiters(lock);
278 * acquire(lock);
279 * or:
280 *
281 * unlock(wait_lock);
282 * lock(wait_lock);
283 * mark_rt_mutex_waiters(lock);
284 *
285 * cmpxchg(p, owner, 0) != owner
286 * enqueue_waiter();
287 * unlock(wait_lock);
288 * lock(wait_lock);
289 * wake waiter();
290 * unlock(wait_lock);
291 * lock(wait_lock);
292 * acquire(lock);
293 */
294 return rt_mutex_cmpxchg_release(lock, owner, NULL);
295 }
296
297 #else
298 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
299 struct task_struct *old,
300 struct task_struct *new)
301 {
302 return false;
303
304 }
305
306 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock);
307
308 static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
309 {
310 /*
311 * With debug enabled rt_mutex_cmpxchg trylock() will always fail.
312 *
313 * Avoid unconditionally taking the slow path by using
314 * rt_mutex_slow_trylock() which is covered by the debug code and can
315 * acquire a non-contended rtmutex.
316 */
317 return rt_mutex_slowtrylock(lock);
318 }
319
320 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
321 struct task_struct *old,
322 struct task_struct *new)
323 {
324 return false;
325 }
326
327 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
328 {
329 lock->owner = (struct task_struct *)
330 ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
331 }
332
333 /*
334 * Simple slow path only version: lock->owner is protected by lock->wait_lock.
335 */
336 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
337 unsigned long flags)
338 __releases(lock->wait_lock)
339 {
340 lock->owner = NULL;
341 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
342 return true;
343 }
344 #endif
345
346 static __always_inline int __waiter_prio(struct task_struct *task)
347 {
348 int prio = task->prio;
349
350 if (!rt_prio(prio))
351 return DEFAULT_PRIO;
352
353 return prio;
354 }
355
356 /*
357 * Update the waiter->tree copy of the sort keys.
358 */
359 static __always_inline void
360 waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
361 {
362 lockdep_assert_held(&waiter->lock->wait_lock);
363 lockdep_assert(RB_EMPTY_NODE(&waiter->tree.entry));
364
365 waiter->tree.prio = __waiter_prio(task);
366 waiter->tree.deadline = task->dl.deadline;
367 }
368
369 /*
370 * Update the waiter->pi_tree copy of the sort keys (from the tree copy).
371 */
372 static __always_inline void
373 waiter_clone_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
374 {
375 lockdep_assert_held(&waiter->lock->wait_lock);
376 lockdep_assert_held(&task->pi_lock);
377 lockdep_assert(RB_EMPTY_NODE(&waiter->pi_tree.entry));
378
379 waiter->pi_tree.prio = waiter->tree.prio;
380 waiter->pi_tree.deadline = waiter->tree.deadline;
381 }
382
383 /*
384 * Only use with rt_waiter_node_{less,equal}()
385 */
386 #define task_to_waiter_node(p) \
387 &(struct rt_waiter_node){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
388 #define task_to_waiter(p) \
389 &(struct rt_mutex_waiter){ .tree = *task_to_waiter_node(p) }
390
391 static __always_inline int rt_waiter_node_less(struct rt_waiter_node *left,
392 struct rt_waiter_node *right)
393 {
394 if (left->prio < right->prio)
395 return 1;
396
397 /*
398 * If both waiters have dl_prio(), we check the deadlines of the
399 * associated tasks.
400 * If left waiter has a dl_prio(), and we didn't return 1 above,
401 * then right waiter has a dl_prio() too.
402 */
403 if (dl_prio(left->prio))
404 return dl_time_before(left->deadline, right->deadline);
405
406 return 0;
407 }
408
409 static __always_inline int rt_waiter_node_equal(struct rt_waiter_node *left,
410 struct rt_waiter_node *right)
411 {
412 if (left->prio != right->prio)
413 return 0;
414
415 /*
416 * If both waiters have dl_prio(), we check the deadlines of the
417 * associated tasks.
418 * If left waiter has a dl_prio(), and we didn't return 0 above,
419 * then right waiter has a dl_prio() too.
420 */
421 if (dl_prio(left->prio))
422 return left->deadline == right->deadline;
423
424 return 1;
425 }
426
427 static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
428 struct rt_mutex_waiter *top_waiter)
429 {
430 if (rt_waiter_node_less(&waiter->tree, &top_waiter->tree))
431 return true;
432
433 #ifdef RT_MUTEX_BUILD_SPINLOCKS
434 /*
435 * Note that RT tasks are excluded from same priority (lateral)
436 * steals to prevent the introduction of an unbounded latency.
437 */
438 if (rt_prio(waiter->tree.prio) || dl_prio(waiter->tree.prio))
439 return false;
440
441 return rt_waiter_node_equal(&waiter->tree, &top_waiter->tree);
442 #else
443 return false;
444 #endif
445 }
446
447 #define __node_2_waiter(node) \
448 rb_entry((node), struct rt_mutex_waiter, tree.entry)
449
450 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
451 {
452 struct rt_mutex_waiter *aw = __node_2_waiter(a);
453 struct rt_mutex_waiter *bw = __node_2_waiter(b);
454
455 if (rt_waiter_node_less(&aw->tree, &bw->tree))
456 return 1;
457
458 if (!build_ww_mutex())
459 return 0;
460
461 if (rt_waiter_node_less(&bw->tree, &aw->tree))
462 return 0;
463
464 /* NOTE: relies on waiter->ww_ctx being set before insertion */
465 if (aw->ww_ctx) {
466 if (!bw->ww_ctx)
467 return 1;
468
469 return (signed long)(aw->ww_ctx->stamp -
470 bw->ww_ctx->stamp) < 0;
471 }
472
473 return 0;
474 }
475
476 static __always_inline void
477 rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
478 {
479 lockdep_assert_held(&lock->wait_lock);
480
481 rb_add_cached(&waiter->tree.entry, &lock->waiters, __waiter_less);
482 }
483
484 static __always_inline void
485 rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
486 {
487 lockdep_assert_held(&lock->wait_lock);
488
489 if (RB_EMPTY_NODE(&waiter->tree.entry))
490 return;
491
492 rb_erase_cached(&waiter->tree.entry, &lock->waiters);
493 RB_CLEAR_NODE(&waiter->tree.entry);
494 }
495
496 #define __node_2_rt_node(node) \
497 rb_entry((node), struct rt_waiter_node, entry)
498
499 static __always_inline bool __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
500 {
501 return rt_waiter_node_less(__node_2_rt_node(a), __node_2_rt_node(b));
502 }
503
504 static __always_inline void
505 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
506 {
507 lockdep_assert_held(&task->pi_lock);
508
509 rb_add_cached(&waiter->pi_tree.entry, &task->pi_waiters, __pi_waiter_less);
510 }
511
512 static __always_inline void
513 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
514 {
515 lockdep_assert_held(&task->pi_lock);
516
517 if (RB_EMPTY_NODE(&waiter->pi_tree.entry))
518 return;
519
520 rb_erase_cached(&waiter->pi_tree.entry, &task->pi_waiters);
521 RB_CLEAR_NODE(&waiter->pi_tree.entry);
522 }
523
524 static __always_inline void rt_mutex_adjust_prio(struct rt_mutex_base *lock,
525 struct task_struct *p)
526 {
527 struct task_struct *pi_task = NULL;
528
529 lockdep_assert_held(&lock->wait_lock);
530 lockdep_assert(rt_mutex_owner(lock) == p);
531 lockdep_assert_held(&p->pi_lock);
532
533 if (task_has_pi_waiters(p))
534 pi_task = task_top_pi_waiter(p)->task;
535
536 rt_mutex_setprio(p, pi_task);
537 }
538
539 /* RT mutex specific wake_q wrappers */
540 static __always_inline void rt_mutex_wake_q_add_task(struct rt_wake_q_head *wqh,
541 struct task_struct *task,
542 unsigned int wake_state)
543 {
544 if (IS_ENABLED(CONFIG_PREEMPT_RT) && wake_state == TASK_RTLOCK_WAIT) {
545 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
546 WARN_ON_ONCE(wqh->rtlock_task);
547 get_task_struct(task);
548 wqh->rtlock_task = task;
549 } else {
550 wake_q_add(&wqh->head, task);
551 }
552 }
553
554 static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
555 struct rt_mutex_waiter *w)
556 {
557 rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
558 }
559
560 static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
561 {
562 if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
563 wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
564 put_task_struct(wqh->rtlock_task);
565 wqh->rtlock_task = NULL;
566 }
567
568 if (!wake_q_empty(&wqh->head))
569 wake_up_q(&wqh->head);
570
571 /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
572 preempt_enable();
573 }
574
575 /*
576 * Deadlock detection is conditional:
577 *
578 * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
579 * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
580 *
581 * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
582 * conducted independent of the detect argument.
583 *
584 * If the waiter argument is NULL this indicates the deboost path and
585 * deadlock detection is disabled independent of the detect argument
586 * and the config settings.
587 */
588 static __always_inline bool
589 rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
590 enum rtmutex_chainwalk chwalk)
591 {
592 if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
593 return waiter != NULL;
594 return chwalk == RT_MUTEX_FULL_CHAINWALK;
595 }
596
597 static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
598 {
599 return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
600 }
601
602 /*
603 * Adjust the priority chain. Also used for deadlock detection.
604 * Decreases task's usage by one - may thus free the task.
605 *
606 * @task: the task owning the mutex (owner) for which a chain walk is
607 * probably needed
608 * @chwalk: do we have to carry out deadlock detection?
609 * @orig_lock: the mutex (can be NULL if we are walking the chain to recheck
610 * things for a task that has just got its priority adjusted, and
611 * is waiting on a mutex)
612 * @next_lock: the mutex on which the owner of @orig_lock was blocked before
613 * we dropped its pi_lock. Is never dereferenced, only used for
614 * comparison to detect lock chain changes.
615 * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
616 * its priority to the mutex owner (can be NULL in the case
617 * depicted above or if the top waiter is gone away and we are
618 * actually deboosting the owner)
619 * @top_task: the current top waiter
620 *
621 * Returns 0 or -EDEADLK.
622 *
623 * Chain walk basics and protection scope
624 *
625 * [R] refcount on task
626 * [Pn] task->pi_lock held
627 * [L] rtmutex->wait_lock held
628 *
629 * Normal locking order:
630 *
631 * rtmutex->wait_lock
632 * task->pi_lock
633 *
634 * Step Description Protected by
635 * function arguments:
636 * @task [R]
637 * @orig_lock if != NULL @top_task is blocked on it
638 * @next_lock Unprotected. Cannot be
639 * dereferenced. Only used for
640 * comparison.
641 * @orig_waiter if != NULL @top_task is blocked on it
642 * @top_task current, or in case of proxy
643 * locking protected by calling
644 * code
645 * again:
646 * loop_sanity_check();
647 * retry:
648 * [1] lock(task->pi_lock); [R] acquire [P1]
649 * [2] waiter = task->pi_blocked_on; [P1]
650 * [3] check_exit_conditions_1(); [P1]
651 * [4] lock = waiter->lock; [P1]
652 * [5] if (!try_lock(lock->wait_lock)) { [P1] try to acquire [L]
653 * unlock(task->pi_lock); release [P1]
654 * goto retry;
655 * }
656 * [6] check_exit_conditions_2(); [P1] + [L]
657 * [7] requeue_lock_waiter(lock, waiter); [P1] + [L]
658 * [8] unlock(task->pi_lock); release [P1]
659 * put_task_struct(task); release [R]
660 * [9] check_exit_conditions_3(); [L]
661 * [10] task = owner(lock); [L]
662 * get_task_struct(task); [L] acquire [R]
663 * lock(task->pi_lock); [L] acquire [P2]
664 * [11] requeue_pi_waiter(tsk, waiters(lock));[P2] + [L]
665 * [12] check_exit_conditions_4(); [P2] + [L]
666 * [13] unlock(task->pi_lock); release [P2]
667 * unlock(lock->wait_lock); release [L]
668 * goto again;
669 *
670 * Where P1 is the blocking task and P2 is the lock owner; going up one step
671 * the owner becomes the next blocked task etc..
672 *
673 *
674 */
675 static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
676 enum rtmutex_chainwalk chwalk,
677 struct rt_mutex_base *orig_lock,
678 struct rt_mutex_base *next_lock,
679 struct rt_mutex_waiter *orig_waiter,
680 struct task_struct *top_task)
681 {
682 struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
683 struct rt_mutex_waiter *prerequeue_top_waiter;
684 int ret = 0, depth = 0;
685 struct rt_mutex_base *lock;
686 bool detect_deadlock;
687 bool requeue = true;
688
689 detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
690
691 /*
692 * The (de)boosting is a step by step approach with a lot of
693 * pitfalls. We want this to be preemptible and we want hold a
694 * maximum of two locks per step. So we have to check
695 * carefully whether things change under us.
696 */
697 again:
698 /*
699 * We limit the lock chain length for each invocation.
700 */
701 if (++depth > max_lock_depth) {
702 static int prev_max;
703
704 /*
705 * Print this only once. If the admin changes the limit,
706 * print a new message when reaching the limit again.
707 */
708 if (prev_max != max_lock_depth) {
709 prev_max = max_lock_depth;
710 printk(KERN_WARNING "Maximum lock depth %d reached "
711 "task: %s (%d)\n", max_lock_depth,
712 top_task->comm, task_pid_nr(top_task));
713 }
714 put_task_struct(task);
715
716 return -EDEADLK;
717 }
718
719 /*
720 * We are fully preemptible here and only hold the refcount on
721 * @task. So everything can have changed under us since the
722 * caller or our own code below (goto retry/again) dropped all
723 * locks.
724 */
725 retry:
726 /*
727 * [1] Task cannot go away as we did a get_task() before !
728 */
729 raw_spin_lock_irq(&task->pi_lock);
730
731 /*
732 * [2] Get the waiter on which @task is blocked on.
733 */
734 waiter = task->pi_blocked_on;
735
736 /*
737 * [3] check_exit_conditions_1() protected by task->pi_lock.
738 */
739
740 /*
741 * Check whether the end of the boosting chain has been
742 * reached or the state of the chain has changed while we
743 * dropped the locks.
744 */
745 if (!waiter)
746 goto out_unlock_pi;
747
748 /*
749 * Check the orig_waiter state. After we dropped the locks,
750 * the previous owner of the lock might have released the lock.
751 */
752 if (orig_waiter && !rt_mutex_owner(orig_lock))
753 goto out_unlock_pi;
754
755 /*
756 * We dropped all locks after taking a refcount on @task, so
757 * the task might have moved on in the lock chain or even left
758 * the chain completely and blocks now on an unrelated lock or
759 * on @orig_lock.
760 *
761 * We stored the lock on which @task was blocked in @next_lock,
762 * so we can detect the chain change.
763 */
764 if (next_lock != waiter->lock)
765 goto out_unlock_pi;
766
767 /*
768 * There could be 'spurious' loops in the lock graph due to ww_mutex,
769 * consider:
770 *
771 * P1: A, ww_A, ww_B
772 * P2: ww_B, ww_A
773 * P3: A
774 *
775 * P3 should not return -EDEADLK because it gets trapped in the cycle
776 * created by P1 and P2 (which will resolve -- and runs into
777 * max_lock_depth above). Therefore disable detect_deadlock such that
778 * the below termination condition can trigger once all relevant tasks
779 * are boosted.
780 *
781 * Even when we start with ww_mutex we can disable deadlock detection,
782 * since we would supress a ww_mutex induced deadlock at [6] anyway.
783 * Supressing it here however is not sufficient since we might still
784 * hit [6] due to adjustment driven iteration.
785 *
786 * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
787 * utterly fail to report it; lockdep should.
788 */
789 if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
790 detect_deadlock = false;
791
792 /*
793 * Drop out, when the task has no waiters. Note,
794 * top_waiter can be NULL, when we are in the deboosting
795 * mode!
796 */
797 if (top_waiter) {
798 if (!task_has_pi_waiters(task))
799 goto out_unlock_pi;
800 /*
801 * If deadlock detection is off, we stop here if we
802 * are not the top pi waiter of the task. If deadlock
803 * detection is enabled we continue, but stop the
804 * requeueing in the chain walk.
805 */
806 if (top_waiter != task_top_pi_waiter(task)) {
807 if (!detect_deadlock)
808 goto out_unlock_pi;
809 else
810 requeue = false;
811 }
812 }
813
814 /*
815 * If the waiter priority is the same as the task priority
816 * then there is no further priority adjustment necessary. If
817 * deadlock detection is off, we stop the chain walk. If its
818 * enabled we continue, but stop the requeueing in the chain
819 * walk.
820 */
821 if (rt_waiter_node_equal(&waiter->tree, task_to_waiter_node(task))) {
822 if (!detect_deadlock)
823 goto out_unlock_pi;
824 else
825 requeue = false;
826 }
827
828 /*
829 * [4] Get the next lock; per holding task->pi_lock we can't unblock
830 * and guarantee @lock's existence.
831 */
832 lock = waiter->lock;
833 /*
834 * [5] We need to trylock here as we are holding task->pi_lock,
835 * which is the reverse lock order versus the other rtmutex
836 * operations.
837 *
838 * Per the above, holding task->pi_lock guarantees lock exists, so
839 * inverting this lock order is infeasible from a life-time
840 * perspective.
841 */
842 if (!raw_spin_trylock(&lock->wait_lock)) {
843 raw_spin_unlock_irq(&task->pi_lock);
844 cpu_relax();
845 goto retry;
846 }
847
848 /*
849 * [6] check_exit_conditions_2() protected by task->pi_lock and
850 * lock->wait_lock.
851 *
852 * Deadlock detection. If the lock is the same as the original
853 * lock which caused us to walk the lock chain or if the
854 * current lock is owned by the task which initiated the chain
855 * walk, we detected a deadlock.
856 */
857 if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
858 ret = -EDEADLK;
859
860 /*
861 * When the deadlock is due to ww_mutex; also see above. Don't
862 * report the deadlock and instead let the ww_mutex wound/die
863 * logic pick which of the contending threads gets -EDEADLK.
864 *
865 * NOTE: assumes the cycle only contains a single ww_class; any
866 * other configuration and we fail to report; also, see
867 * lockdep.
868 */
869 if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
870 ret = 0;
871
872 raw_spin_unlock(&lock->wait_lock);
873 goto out_unlock_pi;
874 }
875
876 /*
877 * If we just follow the lock chain for deadlock detection, no
878 * need to do all the requeue operations. To avoid a truckload
879 * of conditionals around the various places below, just do the
880 * minimum chain walk checks.
881 */
882 if (!requeue) {
883 /*
884 * No requeue[7] here. Just release @task [8]
885 */
886 raw_spin_unlock(&task->pi_lock);
887 put_task_struct(task);
888
889 /*
890 * [9] check_exit_conditions_3 protected by lock->wait_lock.
891 * If there is no owner of the lock, end of chain.
892 */
893 if (!rt_mutex_owner(lock)) {
894 raw_spin_unlock_irq(&lock->wait_lock);
895 return 0;
896 }
897
898 /* [10] Grab the next task, i.e. owner of @lock */
899 task = get_task_struct(rt_mutex_owner(lock));
900 raw_spin_lock(&task->pi_lock);
901
902 /*
903 * No requeue [11] here. We just do deadlock detection.
904 *
905 * [12] Store whether owner is blocked
906 * itself. Decision is made after dropping the locks
907 */
908 next_lock = task_blocked_on_lock(task);
909 /*
910 * Get the top waiter for the next iteration
911 */
912 top_waiter = rt_mutex_top_waiter(lock);
913
914 /* [13] Drop locks */
915 raw_spin_unlock(&task->pi_lock);
916 raw_spin_unlock_irq(&lock->wait_lock);
917
918 /* If owner is not blocked, end of chain. */
919 if (!next_lock)
920 goto out_put_task;
921 goto again;
922 }
923
924 /*
925 * Store the current top waiter before doing the requeue
926 * operation on @lock. We need it for the boost/deboost
927 * decision below.
928 */
929 prerequeue_top_waiter = rt_mutex_top_waiter(lock);
930
931 /* [7] Requeue the waiter in the lock waiter tree. */
932 rt_mutex_dequeue(lock, waiter);
933
934 /*
935 * Update the waiter prio fields now that we're dequeued.
936 *
937 * These values can have changed through either:
938 *
939 * sys_sched_set_scheduler() / sys_sched_setattr()
940 *
941 * or
942 *
943 * DL CBS enforcement advancing the effective deadline.
944 */
945 waiter_update_prio(waiter, task);
946
947 rt_mutex_enqueue(lock, waiter);
948
949 /*
950 * [8] Release the (blocking) task in preparation for
951 * taking the owner task in [10].
952 *
953 * Since we hold lock->waiter_lock, task cannot unblock, even if we
954 * release task->pi_lock.
955 */
956 raw_spin_unlock(&task->pi_lock);
957 put_task_struct(task);
958
959 /*
960 * [9] check_exit_conditions_3 protected by lock->wait_lock.
961 *
962 * We must abort the chain walk if there is no lock owner even
963 * in the dead lock detection case, as we have nothing to
964 * follow here. This is the end of the chain we are walking.
965 */
966 if (!rt_mutex_owner(lock)) {
967 /*
968 * If the requeue [7] above changed the top waiter,
969 * then we need to wake the new top waiter up to try
970 * to get the lock.
971 */
972 top_waiter = rt_mutex_top_waiter(lock);
973 if (prerequeue_top_waiter != top_waiter)
974 wake_up_state(top_waiter->task, top_waiter->wake_state);
975 raw_spin_unlock_irq(&lock->wait_lock);
976 return 0;
977 }
978
979 /*
980 * [10] Grab the next task, i.e. the owner of @lock
981 *
982 * Per holding lock->wait_lock and checking for !owner above, there
983 * must be an owner and it cannot go away.
984 */
985 task = get_task_struct(rt_mutex_owner(lock));
986 raw_spin_lock(&task->pi_lock);
987
988 /* [11] requeue the pi waiters if necessary */
989 if (waiter == rt_mutex_top_waiter(lock)) {
990 /*
991 * The waiter became the new top (highest priority)
992 * waiter on the lock. Replace the previous top waiter
993 * in the owner tasks pi waiters tree with this waiter
994 * and adjust the priority of the owner.
995 */
996 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
997 waiter_clone_prio(waiter, task);
998 rt_mutex_enqueue_pi(task, waiter);
999 rt_mutex_adjust_prio(lock, task);
1000
1001 } else if (prerequeue_top_waiter == waiter) {
1002 /*
1003 * The waiter was the top waiter on the lock, but is
1004 * no longer the top priority waiter. Replace waiter in
1005 * the owner tasks pi waiters tree with the new top
1006 * (highest priority) waiter and adjust the priority
1007 * of the owner.
1008 * The new top waiter is stored in @waiter so that
1009 * @waiter == @top_waiter evaluates to true below and
1010 * we continue to deboost the rest of the chain.
1011 */
1012 rt_mutex_dequeue_pi(task, waiter);
1013 waiter = rt_mutex_top_waiter(lock);
1014 waiter_clone_prio(waiter, task);
1015 rt_mutex_enqueue_pi(task, waiter);
1016 rt_mutex_adjust_prio(lock, task);
1017 } else {
1018 /*
1019 * Nothing changed. No need to do any priority
1020 * adjustment.
1021 */
1022 }
1023
1024 /*
1025 * [12] check_exit_conditions_4() protected by task->pi_lock
1026 * and lock->wait_lock. The actual decisions are made after we
1027 * dropped the locks.
1028 *
1029 * Check whether the task which owns the current lock is pi
1030 * blocked itself. If yes we store a pointer to the lock for
1031 * the lock chain change detection above. After we dropped
1032 * task->pi_lock next_lock cannot be dereferenced anymore.
1033 */
1034 next_lock = task_blocked_on_lock(task);
1035 /*
1036 * Store the top waiter of @lock for the end of chain walk
1037 * decision below.
1038 */
1039 top_waiter = rt_mutex_top_waiter(lock);
1040
1041 /* [13] Drop the locks */
1042 raw_spin_unlock(&task->pi_lock);
1043 raw_spin_unlock_irq(&lock->wait_lock);
1044
1045 /*
1046 * Make the actual exit decisions [12], based on the stored
1047 * values.
1048 *
1049 * We reached the end of the lock chain. Stop right here. No
1050 * point to go back just to figure that out.
1051 */
1052 if (!next_lock)
1053 goto out_put_task;
1054
1055 /*
1056 * If the current waiter is not the top waiter on the lock,
1057 * then we can stop the chain walk here if we are not in full
1058 * deadlock detection mode.
1059 */
1060 if (!detect_deadlock && waiter != top_waiter)
1061 goto out_put_task;
1062
1063 goto again;
1064
1065 out_unlock_pi:
1066 raw_spin_unlock_irq(&task->pi_lock);
1067 out_put_task:
1068 put_task_struct(task);
1069
1070 return ret;
1071 }
1072
1073 /*
1074 * Try to take an rt-mutex
1075 *
1076 * Must be called with lock->wait_lock held and interrupts disabled
1077 *
1078 * @lock: The lock to be acquired.
1079 * @task: The task which wants to acquire the lock
1080 * @waiter: The waiter that is queued to the lock's wait tree if the
1081 * callsite called task_blocked_on_lock(), otherwise NULL
1082 */
1083 static int __sched
1084 try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
1085 struct rt_mutex_waiter *waiter)
1086 {
1087 lockdep_assert_held(&lock->wait_lock);
1088
1089 /*
1090 * Before testing whether we can acquire @lock, we set the
1091 * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
1092 * other tasks which try to modify @lock into the slow path
1093 * and they serialize on @lock->wait_lock.
1094 *
1095 * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
1096 * as explained at the top of this file if and only if:
1097 *
1098 * - There is a lock owner. The caller must fixup the
1099 * transient state if it does a trylock or leaves the lock
1100 * function due to a signal or timeout.
1101 *
1102 * - @task acquires the lock and there are no other
1103 * waiters. This is undone in rt_mutex_set_owner(@task) at
1104 * the end of this function.
1105 */
1106 mark_rt_mutex_waiters(lock);
1107
1108 /*
1109 * If @lock has an owner, give up.
1110 */
1111 if (rt_mutex_owner(lock))
1112 return 0;
1113
1114 /*
1115 * If @waiter != NULL, @task has already enqueued the waiter
1116 * into @lock waiter tree. If @waiter == NULL then this is a
1117 * trylock attempt.
1118 */
1119 if (waiter) {
1120 struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
1121
1122 /*
1123 * If waiter is the highest priority waiter of @lock,
1124 * or allowed to steal it, take it over.
1125 */
1126 if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1127 /*
1128 * We can acquire the lock. Remove the waiter from the
1129 * lock waiters tree.
1130 */
1131 rt_mutex_dequeue(lock, waiter);
1132 } else {
1133 return 0;
1134 }
1135 } else {
1136 /*
1137 * If the lock has waiters already we check whether @task is
1138 * eligible to take over the lock.
1139 *
1140 * If there are no other waiters, @task can acquire
1141 * the lock. @task->pi_blocked_on is NULL, so it does
1142 * not need to be dequeued.
1143 */
1144 if (rt_mutex_has_waiters(lock)) {
1145 /* Check whether the trylock can steal it. */
1146 if (!rt_mutex_steal(task_to_waiter(task),
1147 rt_mutex_top_waiter(lock)))
1148 return 0;
1149
1150 /*
1151 * The current top waiter stays enqueued. We
1152 * don't have to change anything in the lock
1153 * waiters order.
1154 */
1155 } else {
1156 /*
1157 * No waiters. Take the lock without the
1158 * pi_lock dance.@task->pi_blocked_on is NULL
1159 * and we have no waiters to enqueue in @task
1160 * pi waiters tree.
1161 */
1162 goto takeit;
1163 }
1164 }
1165
1166 /*
1167 * Clear @task->pi_blocked_on. Requires protection by
1168 * @task->pi_lock. Redundant operation for the @waiter == NULL
1169 * case, but conditionals are more expensive than a redundant
1170 * store.
1171 */
1172 raw_spin_lock(&task->pi_lock);
1173 task->pi_blocked_on = NULL;
1174 /*
1175 * Finish the lock acquisition. @task is the new owner. If
1176 * other waiters exist we have to insert the highest priority
1177 * waiter into @task->pi_waiters tree.
1178 */
1179 if (rt_mutex_has_waiters(lock))
1180 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
1181 raw_spin_unlock(&task->pi_lock);
1182
1183 takeit:
1184 /*
1185 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1186 * are still waiters or clears it.
1187 */
1188 rt_mutex_set_owner(lock, task);
1189
1190 return 1;
1191 }
1192
1193 /*
1194 * Task blocks on lock.
1195 *
1196 * Prepare waiter and propagate pi chain
1197 *
1198 * This must be called with lock->wait_lock held and interrupts disabled
1199 */
1200 static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
1201 struct rt_mutex_waiter *waiter,
1202 struct task_struct *task,
1203 struct ww_acquire_ctx *ww_ctx,
1204 enum rtmutex_chainwalk chwalk)
1205 {
1206 struct task_struct *owner = rt_mutex_owner(lock);
1207 struct rt_mutex_waiter *top_waiter = waiter;
1208 struct rt_mutex_base *next_lock;
1209 int chain_walk = 0, res;
1210
1211 lockdep_assert_held(&lock->wait_lock);
1212
1213 /*
1214 * Early deadlock detection. We really don't want the task to
1215 * enqueue on itself just to untangle the mess later. It's not
1216 * only an optimization. We drop the locks, so another waiter
1217 * can come in before the chain walk detects the deadlock. So
1218 * the other will detect the deadlock and return -EDEADLOCK,
1219 * which is wrong, as the other waiter is not in a deadlock
1220 * situation.
1221 *
1222 * Except for ww_mutex, in that case the chain walk must already deal
1223 * with spurious cycles, see the comments at [3] and [6].
1224 */
1225 if (owner == task && !(build_ww_mutex() && ww_ctx))
1226 return -EDEADLK;
1227
1228 raw_spin_lock(&task->pi_lock);
1229 waiter->task = task;
1230 waiter->lock = lock;
1231 waiter_update_prio(waiter, task);
1232 waiter_clone_prio(waiter, task);
1233
1234 /* Get the top priority waiter on the lock */
1235 if (rt_mutex_has_waiters(lock))
1236 top_waiter = rt_mutex_top_waiter(lock);
1237 rt_mutex_enqueue(lock, waiter);
1238
1239 task->pi_blocked_on = waiter;
1240
1241 raw_spin_unlock(&task->pi_lock);
1242
1243 if (build_ww_mutex() && ww_ctx) {
1244 struct rt_mutex *rtm;
1245
1246 /* Check whether the waiter should back out immediately */
1247 rtm = container_of(lock, struct rt_mutex, rtmutex);
1248 res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx);
1249 if (res) {
1250 raw_spin_lock(&task->pi_lock);
1251 rt_mutex_dequeue(lock, waiter);
1252 task->pi_blocked_on = NULL;
1253 raw_spin_unlock(&task->pi_lock);
1254 return res;
1255 }
1256 }
1257
1258 if (!owner)
1259 return 0;
1260
1261 raw_spin_lock(&owner->pi_lock);
1262 if (waiter == rt_mutex_top_waiter(lock)) {
1263 rt_mutex_dequeue_pi(owner, top_waiter);
1264 rt_mutex_enqueue_pi(owner, waiter);
1265
1266 rt_mutex_adjust_prio(lock, owner);
1267 if (owner->pi_blocked_on)
1268 chain_walk = 1;
1269 } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1270 chain_walk = 1;
1271 }
1272
1273 /* Store the lock on which owner is blocked or NULL */
1274 next_lock = task_blocked_on_lock(owner);
1275
1276 raw_spin_unlock(&owner->pi_lock);
1277 /*
1278 * Even if full deadlock detection is on, if the owner is not
1279 * blocked itself, we can avoid finding this out in the chain
1280 * walk.
1281 */
1282 if (!chain_walk || !next_lock)
1283 return 0;
1284
1285 /*
1286 * The owner can't disappear while holding a lock,
1287 * so the owner struct is protected by wait_lock.
1288 * Gets dropped in rt_mutex_adjust_prio_chain()!
1289 */
1290 get_task_struct(owner);
1291
1292 raw_spin_unlock_irq(&lock->wait_lock);
1293
1294 res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1295 next_lock, waiter, task);
1296
1297 raw_spin_lock_irq(&lock->wait_lock);
1298
1299 return res;
1300 }
1301
1302 /*
1303 * Remove the top waiter from the current tasks pi waiter tree and
1304 * queue it up.
1305 *
1306 * Called with lock->wait_lock held and interrupts disabled.
1307 */
1308 static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1309 struct rt_mutex_base *lock)
1310 {
1311 struct rt_mutex_waiter *waiter;
1312
1313 lockdep_assert_held(&lock->wait_lock);
1314
1315 raw_spin_lock(¤t->pi_lock);
1316
1317 waiter = rt_mutex_top_waiter(lock);
1318
1319 /*
1320 * Remove it from current->pi_waiters and deboost.
1321 *
1322 * We must in fact deboost here in order to ensure we call
1323 * rt_mutex_setprio() to update p->pi_top_task before the
1324 * task unblocks.
1325 */
1326 rt_mutex_dequeue_pi(current, waiter);
1327 rt_mutex_adjust_prio(lock, current);
1328
1329 /*
1330 * As we are waking up the top waiter, and the waiter stays
1331 * queued on the lock until it gets the lock, this lock
1332 * obviously has waiters. Just set the bit here and this has
1333 * the added benefit of forcing all new tasks into the
1334 * slow path making sure no task of lower priority than
1335 * the top waiter can steal this lock.
1336 */
1337 lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1338
1339 /*
1340 * We deboosted before waking the top waiter task such that we don't
1341 * run two tasks with the 'same' priority (and ensure the
1342 * p->pi_top_task pointer points to a blocked task). This however can
1343 * lead to priority inversion if we would get preempted after the
1344 * deboost but before waking our donor task, hence the preempt_disable()
1345 * before unlock.
1346 *
1347 * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1348 */
1349 preempt_disable();
1350 rt_mutex_wake_q_add(wqh, waiter);
1351 raw_spin_unlock(¤t->pi_lock);
1352 }
1353
1354 static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1355 {
1356 int ret = try_to_take_rt_mutex(lock, current, NULL);
1357
1358 /*
1359 * try_to_take_rt_mutex() sets the lock waiters bit
1360 * unconditionally. Clean this up.
1361 */
1362 fixup_rt_mutex_waiters(lock, true);
1363
1364 return ret;
1365 }
1366
1367 /*
1368 * Slow path try-lock function:
1369 */
1370 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1371 {
1372 unsigned long flags;
1373 int ret;
1374
1375 /*
1376 * If the lock already has an owner we fail to get the lock.
1377 * This can be done without taking the @lock->wait_lock as
1378 * it is only being read, and this is a trylock anyway.
1379 */
1380 if (rt_mutex_owner(lock))
1381 return 0;
1382
1383 /*
1384 * The mutex has currently no owner. Lock the wait lock and try to
1385 * acquire the lock. We use irqsave here to support early boot calls.
1386 */
1387 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1388
1389 ret = __rt_mutex_slowtrylock(lock);
1390
1391 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1392
1393 return ret;
1394 }
1395
1396 static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
1397 {
1398 if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1399 return 1;
1400
1401 return rt_mutex_slowtrylock(lock);
1402 }
1403
1404 /*
1405 * Slow path to release a rt-mutex.
1406 */
1407 static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1408 {
1409 DEFINE_RT_WAKE_Q(wqh);
1410 unsigned long flags;
1411
1412 /* irqsave required to support early boot calls */
1413 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1414
1415 debug_rt_mutex_unlock(lock);
1416
1417 /*
1418 * We must be careful here if the fast path is enabled. If we
1419 * have no waiters queued we cannot set owner to NULL here
1420 * because of:
1421 *
1422 * foo->lock->owner = NULL;
1423 * rtmutex_lock(foo->lock); <- fast path
1424 * free = atomic_dec_and_test(foo->refcnt);
1425 * rtmutex_unlock(foo->lock); <- fast path
1426 * if (free)
1427 * kfree(foo);
1428 * raw_spin_unlock(foo->lock->wait_lock);
1429 *
1430 * So for the fastpath enabled kernel:
1431 *
1432 * Nothing can set the waiters bit as long as we hold
1433 * lock->wait_lock. So we do the following sequence:
1434 *
1435 * owner = rt_mutex_owner(lock);
1436 * clear_rt_mutex_waiters(lock);
1437 * raw_spin_unlock(&lock->wait_lock);
1438 * if (cmpxchg(&lock->owner, owner, 0) == owner)
1439 * return;
1440 * goto retry;
1441 *
1442 * The fastpath disabled variant is simple as all access to
1443 * lock->owner is serialized by lock->wait_lock:
1444 *
1445 * lock->owner = NULL;
1446 * raw_spin_unlock(&lock->wait_lock);
1447 */
1448 while (!rt_mutex_has_waiters(lock)) {
1449 /* Drops lock->wait_lock ! */
1450 if (unlock_rt_mutex_safe(lock, flags) == true)
1451 return;
1452 /* Relock the rtmutex and try again */
1453 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1454 }
1455
1456 /*
1457 * The wakeup next waiter path does not suffer from the above
1458 * race. See the comments there.
1459 *
1460 * Queue the next waiter for wakeup once we release the wait_lock.
1461 */
1462 mark_wakeup_next_waiter(&wqh, lock);
1463 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1464
1465 rt_mutex_wake_up_q(&wqh);
1466 }
1467
1468 static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
1469 {
1470 if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1471 return;
1472
1473 rt_mutex_slowunlock(lock);
1474 }
1475
1476 #ifdef CONFIG_SMP
1477 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1478 struct rt_mutex_waiter *waiter,
1479 struct task_struct *owner)
1480 {
1481 bool res = true;
1482
1483 rcu_read_lock();
1484 for (;;) {
1485 /* If owner changed, trylock again. */
1486 if (owner != rt_mutex_owner(lock))
1487 break;
1488 /*
1489 * Ensure that @owner is dereferenced after checking that
1490 * the lock owner still matches @owner. If that fails,
1491 * @owner might point to freed memory. If it still matches,
1492 * the rcu_read_lock() ensures the memory stays valid.
1493 */
1494 barrier();
1495 /*
1496 * Stop spinning when:
1497 * - the lock owner has been scheduled out
1498 * - current is not longer the top waiter
1499 * - current is requested to reschedule (redundant
1500 * for CONFIG_PREEMPT_RCU=y)
1501 * - the VCPU on which owner runs is preempted
1502 */
1503 if (!owner_on_cpu(owner) || need_resched() ||
1504 !rt_mutex_waiter_is_top_waiter(lock, waiter)) {
1505 res = false;
1506 break;
1507 }
1508 cpu_relax();
1509 }
1510 rcu_read_unlock();
1511 return res;
1512 }
1513 #else
1514 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1515 struct rt_mutex_waiter *waiter,
1516 struct task_struct *owner)
1517 {
1518 return false;
1519 }
1520 #endif
1521
1522 #ifdef RT_MUTEX_BUILD_MUTEX
1523 /*
1524 * Functions required for:
1525 * - rtmutex, futex on all kernels
1526 * - mutex and rwsem substitutions on RT kernels
1527 */
1528
1529 /*
1530 * Remove a waiter from a lock and give up
1531 *
1532 * Must be called with lock->wait_lock held and interrupts disabled. It must
1533 * have just failed to try_to_take_rt_mutex().
1534 */
1535 static void __sched remove_waiter(struct rt_mutex_base *lock,
1536 struct rt_mutex_waiter *waiter)
1537 {
1538 bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1539 struct task_struct *owner = rt_mutex_owner(lock);
1540 struct rt_mutex_base *next_lock;
1541
1542 lockdep_assert_held(&lock->wait_lock);
1543
1544 raw_spin_lock(¤t->pi_lock);
1545 rt_mutex_dequeue(lock, waiter);
1546 current->pi_blocked_on = NULL;
1547 raw_spin_unlock(¤t->pi_lock);
1548
1549 /*
1550 * Only update priority if the waiter was the highest priority
1551 * waiter of the lock and there is an owner to update.
1552 */
1553 if (!owner || !is_top_waiter)
1554 return;
1555
1556 raw_spin_lock(&owner->pi_lock);
1557
1558 rt_mutex_dequeue_pi(owner, waiter);
1559
1560 if (rt_mutex_has_waiters(lock))
1561 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1562
1563 rt_mutex_adjust_prio(lock, owner);
1564
1565 /* Store the lock on which owner is blocked or NULL */
1566 next_lock = task_blocked_on_lock(owner);
1567
1568 raw_spin_unlock(&owner->pi_lock);
1569
1570 /*
1571 * Don't walk the chain, if the owner task is not blocked
1572 * itself.
1573 */
1574 if (!next_lock)
1575 return;
1576
1577 /* gets dropped in rt_mutex_adjust_prio_chain()! */
1578 get_task_struct(owner);
1579
1580 raw_spin_unlock_irq(&lock->wait_lock);
1581
1582 rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1583 next_lock, NULL, current);
1584
1585 raw_spin_lock_irq(&lock->wait_lock);
1586 }
1587
1588 /**
1589 * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
1590 * @lock: the rt_mutex to take
1591 * @ww_ctx: WW mutex context pointer
1592 * @state: the state the task should block in (TASK_INTERRUPTIBLE
1593 * or TASK_UNINTERRUPTIBLE)
1594 * @timeout: the pre-initialized and started timer, or NULL for none
1595 * @waiter: the pre-initialized rt_mutex_waiter
1596 *
1597 * Must be called with lock->wait_lock held and interrupts disabled
1598 */
1599 static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1600 struct ww_acquire_ctx *ww_ctx,
1601 unsigned int state,
1602 struct hrtimer_sleeper *timeout,
1603 struct rt_mutex_waiter *waiter)
1604 {
1605 struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1606 struct task_struct *owner;
1607 int ret = 0;
1608
1609 for (;;) {
1610 /* Try to acquire the lock: */
1611 if (try_to_take_rt_mutex(lock, current, waiter))
1612 break;
1613
1614 if (timeout && !timeout->task) {
1615 ret = -ETIMEDOUT;
1616 break;
1617 }
1618 if (signal_pending_state(state, current)) {
1619 ret = -EINTR;
1620 break;
1621 }
1622
1623 if (build_ww_mutex() && ww_ctx) {
1624 ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1625 if (ret)
1626 break;
1627 }
1628
1629 if (waiter == rt_mutex_top_waiter(lock))
1630 owner = rt_mutex_owner(lock);
1631 else
1632 owner = NULL;
1633 raw_spin_unlock_irq(&lock->wait_lock);
1634
1635 if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1636 rt_mutex_schedule();
1637
1638 raw_spin_lock_irq(&lock->wait_lock);
1639 set_current_state(state);
1640 }
1641
1642 __set_current_state(TASK_RUNNING);
1643 return ret;
1644 }
1645
1646 static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1647 struct rt_mutex_waiter *w)
1648 {
1649 /*
1650 * If the result is not -EDEADLOCK or the caller requested
1651 * deadlock detection, nothing to do here.
1652 */
1653 if (res != -EDEADLOCK || detect_deadlock)
1654 return;
1655
1656 if (build_ww_mutex() && w->ww_ctx)
1657 return;
1658
1659 /*
1660 * Yell loudly and stop the task right here.
1661 */
1662 WARN(1, "rtmutex deadlock detected\n");
1663 while (1) {
1664 set_current_state(TASK_INTERRUPTIBLE);
1665 rt_mutex_schedule();
1666 }
1667 }
1668
1669 /**
1670 * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
1671 * @lock: The rtmutex to block lock
1672 * @ww_ctx: WW mutex context pointer
1673 * @state: The task state for sleeping
1674 * @chwalk: Indicator whether full or partial chainwalk is requested
1675 * @waiter: Initializer waiter for blocking
1676 */
1677 static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1678 struct ww_acquire_ctx *ww_ctx,
1679 unsigned int state,
1680 enum rtmutex_chainwalk chwalk,
1681 struct rt_mutex_waiter *waiter)
1682 {
1683 struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1684 struct ww_mutex *ww = ww_container_of(rtm);
1685 int ret;
1686
1687 lockdep_assert_held(&lock->wait_lock);
1688
1689 /* Try to acquire the lock again: */
1690 if (try_to_take_rt_mutex(lock, current, NULL)) {
1691 if (build_ww_mutex() && ww_ctx) {
1692 __ww_mutex_check_waiters(rtm, ww_ctx);
1693 ww_mutex_lock_acquired(ww, ww_ctx);
1694 }
1695 return 0;
1696 }
1697
1698 set_current_state(state);
1699
1700 trace_contention_begin(lock, LCB_F_RT);
1701
1702 ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
1703 if (likely(!ret))
1704 ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
1705
1706 if (likely(!ret)) {
1707 /* acquired the lock */
1708 if (build_ww_mutex() && ww_ctx) {
1709 if (!ww_ctx->is_wait_die)
1710 __ww_mutex_check_waiters(rtm, ww_ctx);
1711 ww_mutex_lock_acquired(ww, ww_ctx);
1712 }
1713 } else {
1714 __set_current_state(TASK_RUNNING);
1715 remove_waiter(lock, waiter);
1716 rt_mutex_handle_deadlock(ret, chwalk, waiter);
1717 }
1718
1719 /*
1720 * try_to_take_rt_mutex() sets the waiter bit
1721 * unconditionally. We might have to fix that up.
1722 */
1723 fixup_rt_mutex_waiters(lock, true);
1724
1725 trace_contention_end(lock, ret);
1726
1727 return ret;
1728 }
1729
1730 static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1731 struct ww_acquire_ctx *ww_ctx,
1732 unsigned int state)
1733 {
1734 struct rt_mutex_waiter waiter;
1735 int ret;
1736
1737 rt_mutex_init_waiter(&waiter);
1738 waiter.ww_ctx = ww_ctx;
1739
1740 ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1741 &waiter);
1742
1743 debug_rt_mutex_free_waiter(&waiter);
1744 return ret;
1745 }
1746
1747 /*
1748 * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
1749 * @lock: The rtmutex to block lock
1750 * @ww_ctx: WW mutex context pointer
1751 * @state: The task state for sleeping
1752 */
1753 static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1754 struct ww_acquire_ctx *ww_ctx,
1755 unsigned int state)
1756 {
1757 unsigned long flags;
1758 int ret;
1759
1760 /*
1761 * Do all pre-schedule work here, before we queue a waiter and invoke
1762 * PI -- any such work that trips on rtlock (PREEMPT_RT spinlock) would
1763 * otherwise recurse back into task_blocks_on_rt_mutex() through
1764 * rtlock_slowlock() and will then enqueue a second waiter for this
1765 * same task and things get really confusing real fast.
1766 */
1767 rt_mutex_pre_schedule();
1768
1769 /*
1770 * Technically we could use raw_spin_[un]lock_irq() here, but this can
1771 * be called in early boot if the cmpxchg() fast path is disabled
1772 * (debug, no architecture support). In this case we will acquire the
1773 * rtmutex with lock->wait_lock held. But we cannot unconditionally
1774 * enable interrupts in that early boot case. So we need to use the
1775 * irqsave/restore variants.
1776 */
1777 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1778 ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state);
1779 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1780 rt_mutex_post_schedule();
1781
1782 return ret;
1783 }
1784
1785 static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1786 unsigned int state)
1787 {
1788 lockdep_assert(!current->pi_blocked_on);
1789
1790 if (likely(rt_mutex_try_acquire(lock)))
1791 return 0;
1792
1793 return rt_mutex_slowlock(lock, NULL, state);
1794 }
1795 #endif /* RT_MUTEX_BUILD_MUTEX */
1796
1797 #ifdef RT_MUTEX_BUILD_SPINLOCKS
1798 /*
1799 * Functions required for spin/rw_lock substitution on RT kernels
1800 */
1801
1802 /**
1803 * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1804 * @lock: The underlying RT mutex
1805 */
1806 static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
1807 {
1808 struct rt_mutex_waiter waiter;
1809 struct task_struct *owner;
1810
1811 lockdep_assert_held(&lock->wait_lock);
1812
1813 if (try_to_take_rt_mutex(lock, current, NULL))
1814 return;
1815
1816 rt_mutex_init_rtlock_waiter(&waiter);
1817
1818 /* Save current state and set state to TASK_RTLOCK_WAIT */
1819 current_save_and_set_rtlock_wait_state();
1820
1821 trace_contention_begin(lock, LCB_F_RT);
1822
1823 task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
1824
1825 for (;;) {
1826 /* Try to acquire the lock again */
1827 if (try_to_take_rt_mutex(lock, current, &waiter))
1828 break;
1829
1830 if (&waiter == rt_mutex_top_waiter(lock))
1831 owner = rt_mutex_owner(lock);
1832 else
1833 owner = NULL;
1834 raw_spin_unlock_irq(&lock->wait_lock);
1835
1836 if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1837 schedule_rtlock();
1838
1839 raw_spin_lock_irq(&lock->wait_lock);
1840 set_current_state(TASK_RTLOCK_WAIT);
1841 }
1842
1843 /* Restore the task state */
1844 current_restore_rtlock_saved_state();
1845
1846 /*
1847 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1848 * We might have to fix that up:
1849 */
1850 fixup_rt_mutex_waiters(lock, true);
1851 debug_rt_mutex_free_waiter(&waiter);
1852
1853 trace_contention_end(lock, 0);
1854 }
1855
1856 static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
1857 {
1858 unsigned long flags;
1859
1860 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1861 rtlock_slowlock_locked(lock);
1862 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1863 }
1864
1865 #endif /* RT_MUTEX_BUILD_SPINLOCKS */
1866
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.