1 // SPDX-License-Identifier: GPL-2.0-only
3 * RT-Mutexes: simple blocking mutual exclusion locks with PI support
5 * started by Ingo Molnar and Thomas Gleixner.
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
12 * Copyright (C) 2008 Novell, Inc., Gregory Haskins, Sven Dietrich,
14 * Adaptive Spinlocks simplification:
15 * Copyright (C) 2008 Red Hat, Inc., Steven Rostedt <srostedt@redhat.com>
17 * See Documentation/locking/rt-mutex-design.rst for details.
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>
27 #include <trace/events/lock.h>
29 #include "rtmutex_common.h"
32 # define build_ww_mutex() (false)
33 # define ww_container_of(rtm) NULL
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)
42 static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
43 struct ww_acquire_ctx *ww_ctx)
47 static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
48 struct ww_acquire_ctx *ww_ctx)
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)
60 # define build_ww_mutex() (true)
61 # define ww_container_of(rtm) container_of(rtm, struct ww_mutex, base)
62 # include "ww_mutex.h"
66 * lock->owner state tracking:
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.
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**
78 * The fast atomic compare exchange based acquire and release is only
79 * possible when bit 0 of lock->owner is 0.
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.
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.
92 static __always_inline struct task_struct *
93 rt_mutex_owner_encode(struct rt_mutex_base *lock, struct task_struct *owner)
95 unsigned long val = (unsigned long)owner;
97 if (rt_mutex_has_waiters(lock))
98 val |= RT_MUTEX_HAS_WAITERS;
100 return (struct task_struct *)val;
103 static __always_inline void
104 rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
107 * lock->wait_lock is held but explicit acquire semantics are needed
108 * for a new lock owner so WRITE_ONCE is insufficient.
110 xchg_acquire(&lock->owner, rt_mutex_owner_encode(lock, owner));
113 static __always_inline void rt_mutex_clear_owner(struct rt_mutex_base *lock)
115 /* lock->wait_lock is held so the unlock provides release semantics. */
116 WRITE_ONCE(lock->owner, rt_mutex_owner_encode(lock, NULL));
119 static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
121 lock->owner = (struct task_struct *)
122 ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
125 static __always_inline void
126 fixup_rt_mutex_waiters(struct rt_mutex_base *lock, bool acquire_lock)
128 unsigned long owner, *p = (unsigned long *) &lock->owner;
130 if (rt_mutex_has_waiters(lock))
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:
142 * l->owner = T1 | HAS_WAITERS;
150 * l->owner = T1 | HAS_WAITERS;
155 * signal(->T2) signal(->T3)
162 * ==> wait list is empty
166 * fixup_rt_mutex_waiters()
167 * if (wait_list_empty(l) {
169 * owner = l->owner & ~HAS_WAITERS;
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)
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
191 owner = READ_ONCE(*p);
192 if (owner & RT_MUTEX_HAS_WAITERS) {
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.
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.
203 xchg_acquire(p, owner & ~RT_MUTEX_HAS_WAITERS);
205 WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
210 * We can speed up the acquire/release, if there's no debugging state to be
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)
218 return try_cmpxchg_acquire(&lock->owner, &old, new);
221 static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
223 return rt_mutex_cmpxchg_acquire(lock, NULL, current);
226 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
227 struct task_struct *old,
228 struct task_struct *new)
230 return try_cmpxchg_release(&lock->owner, &old, new);
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.
238 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
240 unsigned long *p = (unsigned long *) &lock->owner;
241 unsigned long owner, new;
243 owner = READ_ONCE(*p);
245 new = owner | RT_MUTEX_HAS_WAITERS;
246 } while (!try_cmpxchg_relaxed(p, &owner, new));
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.
253 smp_mb__after_atomic();
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
262 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
264 __releases(lock->wait_lock)
266 struct task_struct *owner = rt_mutex_owner(lock);
268 clear_rt_mutex_waiters(lock);
269 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
271 * If a new waiter comes in between the unlock and the cmpxchg
272 * we have two situations:
276 * cmpxchg(p, owner, 0) == owner
277 * mark_rt_mutex_waiters(lock);
283 * mark_rt_mutex_waiters(lock);
285 * cmpxchg(p, owner, 0) != owner
294 return rt_mutex_cmpxchg_release(lock, owner, NULL);
298 static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
299 struct task_struct *old,
300 struct task_struct *new)
306 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock);
308 static __always_inline bool rt_mutex_try_acquire(struct rt_mutex_base *lock)
311 * With debug enabled rt_mutex_cmpxchg trylock() will always fail.
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.
317 return rt_mutex_slowtrylock(lock);
320 static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
321 struct task_struct *old,
322 struct task_struct *new)
327 static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
329 lock->owner = (struct task_struct *)
330 ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
334 * Simple slow path only version: lock->owner is protected by lock->wait_lock.
336 static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
338 __releases(lock->wait_lock)
341 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
346 static __always_inline int __waiter_prio(struct task_struct *task)
348 int prio = task->prio;
357 * Update the waiter->tree copy of the sort keys.
359 static __always_inline void
360 waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
362 lockdep_assert_held(&waiter->lock->wait_lock);
363 lockdep_assert(RB_EMPTY_NODE(&waiter->tree.entry));
365 waiter->tree.prio = __waiter_prio(task);
366 waiter->tree.deadline = task->dl.deadline;
370 * Update the waiter->pi_tree copy of the sort keys (from the tree copy).
372 static __always_inline void
373 waiter_clone_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
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));
379 waiter->pi_tree.prio = waiter->tree.prio;
380 waiter->pi_tree.deadline = waiter->tree.deadline;
384 * Only use with rt_waiter_node_{less,equal}()
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) }
391 static __always_inline int rt_waiter_node_less(struct rt_waiter_node *left,
392 struct rt_waiter_node *right)
394 if (left->prio < right->prio)
398 * If both waiters have dl_prio(), we check the deadlines of the
400 * If left waiter has a dl_prio(), and we didn't return 1 above,
401 * then right waiter has a dl_prio() too.
403 if (dl_prio(left->prio))
404 return dl_time_before(left->deadline, right->deadline);
409 static __always_inline int rt_waiter_node_equal(struct rt_waiter_node *left,
410 struct rt_waiter_node *right)
412 if (left->prio != right->prio)
416 * If both waiters have dl_prio(), we check the deadlines of the
418 * If left waiter has a dl_prio(), and we didn't return 0 above,
419 * then right waiter has a dl_prio() too.
421 if (dl_prio(left->prio))
422 return left->deadline == right->deadline;
427 static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
428 struct rt_mutex_waiter *top_waiter)
430 if (rt_waiter_node_less(&waiter->tree, &top_waiter->tree))
433 #ifdef RT_MUTEX_BUILD_SPINLOCKS
435 * Note that RT tasks are excluded from same priority (lateral)
436 * steals to prevent the introduction of an unbounded latency.
438 if (rt_prio(waiter->tree.prio) || dl_prio(waiter->tree.prio))
441 return rt_waiter_node_equal(&waiter->tree, &top_waiter->tree);
447 #define __node_2_waiter(node) \
448 rb_entry((node), struct rt_mutex_waiter, tree.entry)
450 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
452 struct rt_mutex_waiter *aw = __node_2_waiter(a);
453 struct rt_mutex_waiter *bw = __node_2_waiter(b);
455 if (rt_waiter_node_less(&aw->tree, &bw->tree))
458 if (!build_ww_mutex())
461 if (rt_waiter_node_less(&bw->tree, &aw->tree))
464 /* NOTE: relies on waiter->ww_ctx being set before insertion */
469 return (signed long)(aw->ww_ctx->stamp -
470 bw->ww_ctx->stamp) < 0;
476 static __always_inline void
477 rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
479 lockdep_assert_held(&lock->wait_lock);
481 rb_add_cached(&waiter->tree.entry, &lock->waiters, __waiter_less);
484 static __always_inline void
485 rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
487 lockdep_assert_held(&lock->wait_lock);
489 if (RB_EMPTY_NODE(&waiter->tree.entry))
492 rb_erase_cached(&waiter->tree.entry, &lock->waiters);
493 RB_CLEAR_NODE(&waiter->tree.entry);
496 #define __node_2_rt_node(node) \
497 rb_entry((node), struct rt_waiter_node, entry)
499 static __always_inline bool __pi_waiter_less(struct rb_node *a, const struct rb_node *b)
501 return rt_waiter_node_less(__node_2_rt_node(a), __node_2_rt_node(b));
504 static __always_inline void
505 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
507 lockdep_assert_held(&task->pi_lock);
509 rb_add_cached(&waiter->pi_tree.entry, &task->pi_waiters, __pi_waiter_less);
512 static __always_inline void
513 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
515 lockdep_assert_held(&task->pi_lock);
517 if (RB_EMPTY_NODE(&waiter->pi_tree.entry))
520 rb_erase_cached(&waiter->pi_tree.entry, &task->pi_waiters);
521 RB_CLEAR_NODE(&waiter->pi_tree.entry);
524 static __always_inline void rt_mutex_adjust_prio(struct rt_mutex_base *lock,
525 struct task_struct *p)
527 struct task_struct *pi_task = NULL;
529 lockdep_assert_held(&lock->wait_lock);
530 lockdep_assert(rt_mutex_owner(lock) == p);
531 lockdep_assert_held(&p->pi_lock);
533 if (task_has_pi_waiters(p))
534 pi_task = task_top_pi_waiter(p)->task;
536 rt_mutex_setprio(p, pi_task);
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)
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;
550 wake_q_add(&wqh->head, task);
554 static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
555 struct rt_mutex_waiter *w)
557 rt_mutex_wake_q_add_task(wqh, w->task, w->wake_state);
560 static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
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;
568 if (!wake_q_empty(&wqh->head))
569 wake_up_q(&wqh->head);
571 /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
576 * Deadlock detection is conditional:
578 * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
579 * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
581 * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
582 * conducted independent of the detect argument.
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.
588 static __always_inline bool
589 rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
590 enum rtmutex_chainwalk chwalk)
592 if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
593 return waiter != NULL;
594 return chwalk == RT_MUTEX_FULL_CHAINWALK;
597 static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
599 return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
603 * Adjust the priority chain. Also used for deadlock detection.
604 * Decreases task's usage by one - may thus free the task.
606 * @task: the task owning the mutex (owner) for which a chain walk is
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
621 * Returns 0 or -EDEADLK.
623 * Chain walk basics and protection scope
625 * [R] refcount on task
626 * [Pn] task->pi_lock held
627 * [L] rtmutex->wait_lock held
629 * Normal locking order:
634 * Step Description Protected by
635 * function arguments:
637 * @orig_lock if != NULL @top_task is blocked on it
638 * @next_lock Unprotected. Cannot be
639 * dereferenced. Only used for
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
646 * loop_sanity_check();
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]
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]
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..
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)
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;
689 detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
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.
699 * We limit the lock chain length for each invocation.
701 if (++depth > max_lock_depth) {
705 * Print this only once. If the admin changes the limit,
706 * print a new message when reaching the limit again.
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));
714 put_task_struct(task);
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
727 * [1] Task cannot go away as we did a get_task() before !
729 raw_spin_lock_irq(&task->pi_lock);
732 * [2] Get the waiter on which @task is blocked on.
734 waiter = task->pi_blocked_on;
737 * [3] check_exit_conditions_1() protected by task->pi_lock.
741 * Check whether the end of the boosting chain has been
742 * reached or the state of the chain has changed while we
749 * Check the orig_waiter state. After we dropped the locks,
750 * the previous owner of the lock might have released the lock.
752 if (orig_waiter && !rt_mutex_owner(orig_lock))
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
761 * We stored the lock on which @task was blocked in @next_lock,
762 * so we can detect the chain change.
764 if (next_lock != waiter->lock)
768 * There could be 'spurious' loops in the lock graph due to ww_mutex,
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
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.
786 * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
787 * utterly fail to report it; lockdep should.
789 if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
790 detect_deadlock = false;
793 * Drop out, when the task has no waiters. Note,
794 * top_waiter can be NULL, when we are in the deboosting
798 if (!task_has_pi_waiters(task))
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.
806 if (top_waiter != task_top_pi_waiter(task)) {
807 if (!detect_deadlock)
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
821 if (rt_waiter_node_equal(&waiter->tree, task_to_waiter_node(task))) {
822 if (!detect_deadlock)
829 * [4] Get the next lock; per holding task->pi_lock we can't unblock
830 * and guarantee @lock's existence.
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
838 * Per the above, holding task->pi_lock guarantees lock exists, so
839 * inverting this lock order is infeasible from a life-time
842 if (!raw_spin_trylock(&lock->wait_lock)) {
843 raw_spin_unlock_irq(&task->pi_lock);
849 * [6] check_exit_conditions_2() protected by task->pi_lock and
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.
857 if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
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.
865 * NOTE: assumes the cycle only contains a single ww_class; any
866 * other configuration and we fail to report; also, see
869 if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
872 raw_spin_unlock(&lock->wait_lock);
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.
884 * No requeue[7] here. Just release @task [8]
886 raw_spin_unlock(&task->pi_lock);
887 put_task_struct(task);
890 * [9] check_exit_conditions_3 protected by lock->wait_lock.
891 * If there is no owner of the lock, end of chain.
893 if (!rt_mutex_owner(lock)) {
894 raw_spin_unlock_irq(&lock->wait_lock);
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);
903 * No requeue [11] here. We just do deadlock detection.
905 * [12] Store whether owner is blocked
906 * itself. Decision is made after dropping the locks
908 next_lock = task_blocked_on_lock(task);
910 * Get the top waiter for the next iteration
912 top_waiter = rt_mutex_top_waiter(lock);
914 /* [13] Drop locks */
915 raw_spin_unlock(&task->pi_lock);
916 raw_spin_unlock_irq(&lock->wait_lock);
918 /* If owner is not blocked, end of chain. */
925 * Store the current top waiter before doing the requeue
926 * operation on @lock. We need it for the boost/deboost
929 prerequeue_top_waiter = rt_mutex_top_waiter(lock);
931 /* [7] Requeue the waiter in the lock waiter tree. */
932 rt_mutex_dequeue(lock, waiter);
935 * Update the waiter prio fields now that we're dequeued.
937 * These values can have changed through either:
939 * sys_sched_set_scheduler() / sys_sched_setattr()
943 * DL CBS enforcement advancing the effective deadline.
945 waiter_update_prio(waiter, task);
947 rt_mutex_enqueue(lock, waiter);
950 * [8] Release the (blocking) task in preparation for
951 * taking the owner task in [10].
953 * Since we hold lock->waiter_lock, task cannot unblock, even if we
954 * release task->pi_lock.
956 raw_spin_unlock(&task->pi_lock);
957 put_task_struct(task);
960 * [9] check_exit_conditions_3 protected by lock->wait_lock.
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.
966 if (!rt_mutex_owner(lock)) {
968 * If the requeue [7] above changed the top waiter,
969 * then we need to wake the new top waiter up to try
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);
980 * [10] Grab the next task, i.e. the owner of @lock
982 * Per holding lock->wait_lock and checking for !owner above, there
983 * must be an owner and it cannot go away.
985 task = get_task_struct(rt_mutex_owner(lock));
986 raw_spin_lock(&task->pi_lock);
988 /* [11] requeue the pi waiters if necessary */
989 if (waiter == rt_mutex_top_waiter(lock)) {
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.
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);
1001 } else if (prerequeue_top_waiter == waiter) {
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
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.
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);
1019 * Nothing changed. No need to do any priority
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.
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.
1034 next_lock = task_blocked_on_lock(task);
1036 * Store the top waiter of @lock for the end of chain walk
1039 top_waiter = rt_mutex_top_waiter(lock);
1041 /* [13] Drop the locks */
1042 raw_spin_unlock(&task->pi_lock);
1043 raw_spin_unlock_irq(&lock->wait_lock);
1046 * Make the actual exit decisions [12], based on the stored
1049 * We reached the end of the lock chain. Stop right here. No
1050 * point to go back just to figure that out.
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.
1060 if (!detect_deadlock && waiter != top_waiter)
1066 raw_spin_unlock_irq(&task->pi_lock);
1068 put_task_struct(task);
1074 * Try to take an rt-mutex
1076 * Must be called with lock->wait_lock held and interrupts disabled
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
1084 try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
1085 struct rt_mutex_waiter *waiter)
1087 lockdep_assert_held(&lock->wait_lock);
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.
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:
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.
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.
1106 mark_rt_mutex_waiters(lock);
1109 * If @lock has an owner, give up.
1111 if (rt_mutex_owner(lock))
1115 * If @waiter != NULL, @task has already enqueued the waiter
1116 * into @lock waiter tree. If @waiter == NULL then this is a
1120 struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
1123 * If waiter is the highest priority waiter of @lock,
1124 * or allowed to steal it, take it over.
1126 if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1128 * We can acquire the lock. Remove the waiter from the
1129 * lock waiters tree.
1131 rt_mutex_dequeue(lock, waiter);
1137 * If the lock has waiters already we check whether @task is
1138 * eligible to take over the lock.
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.
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)))
1151 * The current top waiter stays enqueued. We
1152 * don't have to change anything in the lock
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
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
1172 raw_spin_lock(&task->pi_lock);
1173 task->pi_blocked_on = NULL;
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.
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);
1185 * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1186 * are still waiters or clears it.
1188 rt_mutex_set_owner(lock, task);
1194 * Task blocks on lock.
1196 * Prepare waiter and propagate pi chain
1198 * This must be called with lock->wait_lock held and interrupts disabled
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)
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;
1211 lockdep_assert_held(&lock->wait_lock);
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
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].
1225 if (owner == task && !(build_ww_mutex() && ww_ctx))
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);
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);
1239 task->pi_blocked_on = waiter;
1241 raw_spin_unlock(&task->pi_lock);
1243 if (build_ww_mutex() && ww_ctx) {
1244 struct rt_mutex *rtm;
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);
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);
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);
1266 rt_mutex_adjust_prio(lock, owner);
1267 if (owner->pi_blocked_on)
1269 } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1273 /* Store the lock on which owner is blocked or NULL */
1274 next_lock = task_blocked_on_lock(owner);
1276 raw_spin_unlock(&owner->pi_lock);
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
1282 if (!chain_walk || !next_lock)
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()!
1290 get_task_struct(owner);
1292 raw_spin_unlock_irq(&lock->wait_lock);
1294 res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1295 next_lock, waiter, task);
1297 raw_spin_lock_irq(&lock->wait_lock);
1303 * Remove the top waiter from the current tasks pi waiter tree and
1306 * Called with lock->wait_lock held and interrupts disabled.
1308 static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1309 struct rt_mutex_base *lock)
1311 struct rt_mutex_waiter *waiter;
1313 lockdep_assert_held(&lock->wait_lock);
1315 raw_spin_lock(¤t->pi_lock);
1317 waiter = rt_mutex_top_waiter(lock);
1320 * Remove it from current->pi_waiters and deboost.
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
1326 rt_mutex_dequeue_pi(current, waiter);
1327 rt_mutex_adjust_prio(lock, current);
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.
1337 lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
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()
1347 * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1350 rt_mutex_wake_q_add(wqh, waiter);
1351 raw_spin_unlock(¤t->pi_lock);
1354 static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1356 int ret = try_to_take_rt_mutex(lock, current, NULL);
1359 * try_to_take_rt_mutex() sets the lock waiters bit
1360 * unconditionally. Clean this up.
1362 fixup_rt_mutex_waiters(lock, true);
1368 * Slow path try-lock function:
1370 static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1372 unsigned long flags;
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.
1380 if (rt_mutex_owner(lock))
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.
1387 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1389 ret = __rt_mutex_slowtrylock(lock);
1391 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1396 static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
1398 if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1401 return rt_mutex_slowtrylock(lock);
1405 * Slow path to release a rt-mutex.
1407 static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1409 DEFINE_RT_WAKE_Q(wqh);
1410 unsigned long flags;
1412 /* irqsave required to support early boot calls */
1413 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1415 debug_rt_mutex_unlock(lock);
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
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
1428 * raw_spin_unlock(foo->lock->wait_lock);
1430 * So for the fastpath enabled kernel:
1432 * Nothing can set the waiters bit as long as we hold
1433 * lock->wait_lock. So we do the following sequence:
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)
1442 * The fastpath disabled variant is simple as all access to
1443 * lock->owner is serialized by lock->wait_lock:
1445 * lock->owner = NULL;
1446 * raw_spin_unlock(&lock->wait_lock);
1448 while (!rt_mutex_has_waiters(lock)) {
1449 /* Drops lock->wait_lock ! */
1450 if (unlock_rt_mutex_safe(lock, flags) == true)
1452 /* Relock the rtmutex and try again */
1453 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1457 * The wakeup next waiter path does not suffer from the above
1458 * race. See the comments there.
1460 * Queue the next waiter for wakeup once we release the wait_lock.
1462 mark_wakeup_next_waiter(&wqh, lock);
1463 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1465 rt_mutex_wake_up_q(&wqh);
1468 static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
1470 if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1473 rt_mutex_slowunlock(lock);
1477 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1478 struct rt_mutex_waiter *waiter,
1479 struct task_struct *owner)
1485 /* If owner changed, trylock again. */
1486 if (owner != rt_mutex_owner(lock))
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.
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
1503 if (!owner_on_cpu(owner) || need_resched() ||
1504 !rt_mutex_waiter_is_top_waiter(lock, waiter)) {
1514 static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1515 struct rt_mutex_waiter *waiter,
1516 struct task_struct *owner)
1522 #ifdef RT_MUTEX_BUILD_MUTEX
1524 * Functions required for:
1525 * - rtmutex, futex on all kernels
1526 * - mutex and rwsem substitutions on RT kernels
1530 * Remove a waiter from a lock and give up
1532 * Must be called with lock->wait_lock held and interrupts disabled. It must
1533 * have just failed to try_to_take_rt_mutex().
1535 static void __sched remove_waiter(struct rt_mutex_base *lock,
1536 struct rt_mutex_waiter *waiter)
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;
1542 lockdep_assert_held(&lock->wait_lock);
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);
1550 * Only update priority if the waiter was the highest priority
1551 * waiter of the lock and there is an owner to update.
1553 if (!owner || !is_top_waiter)
1556 raw_spin_lock(&owner->pi_lock);
1558 rt_mutex_dequeue_pi(owner, waiter);
1560 if (rt_mutex_has_waiters(lock))
1561 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1563 rt_mutex_adjust_prio(lock, owner);
1565 /* Store the lock on which owner is blocked or NULL */
1566 next_lock = task_blocked_on_lock(owner);
1568 raw_spin_unlock(&owner->pi_lock);
1571 * Don't walk the chain, if the owner task is not blocked
1577 /* gets dropped in rt_mutex_adjust_prio_chain()! */
1578 get_task_struct(owner);
1580 raw_spin_unlock_irq(&lock->wait_lock);
1582 rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1583 next_lock, NULL, current);
1585 raw_spin_lock_irq(&lock->wait_lock);
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
1597 * Must be called with lock->wait_lock held and interrupts disabled
1599 static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1600 struct ww_acquire_ctx *ww_ctx,
1602 struct hrtimer_sleeper *timeout,
1603 struct rt_mutex_waiter *waiter)
1605 struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1606 struct task_struct *owner;
1610 /* Try to acquire the lock: */
1611 if (try_to_take_rt_mutex(lock, current, waiter))
1614 if (timeout && !timeout->task) {
1618 if (signal_pending_state(state, current)) {
1623 if (build_ww_mutex() && ww_ctx) {
1624 ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1629 if (waiter == rt_mutex_top_waiter(lock))
1630 owner = rt_mutex_owner(lock);
1633 raw_spin_unlock_irq(&lock->wait_lock);
1635 if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1636 rt_mutex_schedule();
1638 raw_spin_lock_irq(&lock->wait_lock);
1639 set_current_state(state);
1642 __set_current_state(TASK_RUNNING);
1646 static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1647 struct rt_mutex_waiter *w)
1650 * If the result is not -EDEADLOCK or the caller requested
1651 * deadlock detection, nothing to do here.
1653 if (res != -EDEADLOCK || detect_deadlock)
1656 if (build_ww_mutex() && w->ww_ctx)
1660 * Yell loudly and stop the task right here.
1662 WARN(1, "rtmutex deadlock detected\n");
1664 set_current_state(TASK_INTERRUPTIBLE);
1665 rt_mutex_schedule();
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
1677 static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1678 struct ww_acquire_ctx *ww_ctx,
1680 enum rtmutex_chainwalk chwalk,
1681 struct rt_mutex_waiter *waiter)
1683 struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1684 struct ww_mutex *ww = ww_container_of(rtm);
1687 lockdep_assert_held(&lock->wait_lock);
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);
1698 set_current_state(state);
1700 trace_contention_begin(lock, LCB_F_RT);
1702 ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
1704 ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
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);
1714 __set_current_state(TASK_RUNNING);
1715 remove_waiter(lock, waiter);
1716 rt_mutex_handle_deadlock(ret, chwalk, waiter);
1720 * try_to_take_rt_mutex() sets the waiter bit
1721 * unconditionally. We might have to fix that up.
1723 fixup_rt_mutex_waiters(lock, true);
1725 trace_contention_end(lock, ret);
1730 static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1731 struct ww_acquire_ctx *ww_ctx,
1734 struct rt_mutex_waiter waiter;
1737 rt_mutex_init_waiter(&waiter);
1738 waiter.ww_ctx = ww_ctx;
1740 ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1743 debug_rt_mutex_free_waiter(&waiter);
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
1753 static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1754 struct ww_acquire_ctx *ww_ctx,
1757 unsigned long flags;
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.
1767 rt_mutex_pre_schedule();
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.
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();
1785 static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1788 lockdep_assert(!current->pi_blocked_on);
1790 if (likely(rt_mutex_try_acquire(lock)))
1793 return rt_mutex_slowlock(lock, NULL, state);
1795 #endif /* RT_MUTEX_BUILD_MUTEX */
1797 #ifdef RT_MUTEX_BUILD_SPINLOCKS
1799 * Functions required for spin/rw_lock substitution on RT kernels
1803 * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1804 * @lock: The underlying RT mutex
1806 static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
1808 struct rt_mutex_waiter waiter;
1809 struct task_struct *owner;
1811 lockdep_assert_held(&lock->wait_lock);
1813 if (try_to_take_rt_mutex(lock, current, NULL))
1816 rt_mutex_init_rtlock_waiter(&waiter);
1818 /* Save current state and set state to TASK_RTLOCK_WAIT */
1819 current_save_and_set_rtlock_wait_state();
1821 trace_contention_begin(lock, LCB_F_RT);
1823 task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
1826 /* Try to acquire the lock again */
1827 if (try_to_take_rt_mutex(lock, current, &waiter))
1830 if (&waiter == rt_mutex_top_waiter(lock))
1831 owner = rt_mutex_owner(lock);
1834 raw_spin_unlock_irq(&lock->wait_lock);
1836 if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1839 raw_spin_lock_irq(&lock->wait_lock);
1840 set_current_state(TASK_RTLOCK_WAIT);
1843 /* Restore the task state */
1844 current_restore_rtlock_saved_state();
1847 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1848 * We might have to fix that up:
1850 fixup_rt_mutex_waiters(lock, true);
1851 debug_rt_mutex_free_waiter(&waiter);
1853 trace_contention_end(lock, 0);
1856 static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
1858 unsigned long flags;
1860 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1861 rtlock_slowlock_locked(lock);
1862 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1865 #endif /* RT_MUTEX_BUILD_SPINLOCKS */