OSDN Git Service

spi: New support and problem adjustment of SPI rockchip
[uclinux-h8/linux.git] / kernel / locking / rtmutex.c
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 */