Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/locking/mutex.c
4 *
5 * Mutexes: blocking mutual exclusion locks
6 *
7 * Started by Ingo Molnar:
8 *
9 * Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
10 *
11 * Many thanks to Arjan van de Ven, Thomas Gleixner, Steven Rostedt and
12 * David Howells for suggestions and improvements.
13 *
14 * - Adaptive spinning for mutexes by Peter Zijlstra. (Ported to mainline
15 * from the -rt tree, where it was originally implemented for rtmutexes
16 * by Steven Rostedt, based on work by Gregory Haskins, Peter Morreale
17 * and Sven Dietrich.
18 *
19 * Also see Documentation/locking/mutex-design.rst.
20 */
21#include <linux/mutex.h>
22#include <linux/ww_mutex.h>
23#include <linux/sched/signal.h>
24#include <linux/sched/rt.h>
25#include <linux/sched/wake_q.h>
26#include <linux/sched/debug.h>
27#include <linux/export.h>
28#include <linux/spinlock.h>
29#include <linux/interrupt.h>
30#include <linux/debug_locks.h>
31#include <linux/osq_lock.h>
32
33#ifdef CONFIG_DEBUG_MUTEXES
34# include "mutex-debug.h"
35#else
36# include "mutex.h"
37#endif
38
39void
40__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
41{
42 atomic_long_set(&lock->owner, 0);
43 spin_lock_init(&lock->wait_lock);
44 INIT_LIST_HEAD(&lock->wait_list);
45#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
46 osq_lock_init(&lock->osq);
47#endif
48
49 debug_mutex_init(lock, name, key);
50}
51EXPORT_SYMBOL(__mutex_init);
52
53/*
54 * @owner: contains: 'struct task_struct *' to the current lock owner,
55 * NULL means not owned. Since task_struct pointers are aligned at
56 * at least L1_CACHE_BYTES, we have low bits to store extra state.
57 *
58 * Bit0 indicates a non-empty waiter list; unlock must issue a wakeup.
59 * Bit1 indicates unlock needs to hand the lock to the top-waiter
60 * Bit2 indicates handoff has been done and we're waiting for pickup.
61 */
62#define MUTEX_FLAG_WAITERS 0x01
63#define MUTEX_FLAG_HANDOFF 0x02
64#define MUTEX_FLAG_PICKUP 0x04
65
66#define MUTEX_FLAGS 0x07
67
68/*
69 * Internal helper function; C doesn't allow us to hide it :/
70 *
71 * DO NOT USE (outside of mutex code).
72 */
73static inline struct task_struct *__mutex_owner(struct mutex *lock)
74{
75 return (struct task_struct *)(atomic_long_read(&lock->owner) & ~MUTEX_FLAGS);
76}
77
78static inline struct task_struct *__owner_task(unsigned long owner)
79{
80 return (struct task_struct *)(owner & ~MUTEX_FLAGS);
81}
82
83bool mutex_is_locked(struct mutex *lock)
84{
85 return __mutex_owner(lock) != NULL;
86}
87EXPORT_SYMBOL(mutex_is_locked);
88
89__must_check enum mutex_trylock_recursive_enum
90mutex_trylock_recursive(struct mutex *lock)
91{
92 if (unlikely(__mutex_owner(lock) == current))
93 return MUTEX_TRYLOCK_RECURSIVE;
94
95 return mutex_trylock(lock);
96}
97EXPORT_SYMBOL(mutex_trylock_recursive);
98
99static inline unsigned long __owner_flags(unsigned long owner)
100{
101 return owner & MUTEX_FLAGS;
102}
103
104/*
105 * Trylock variant that retuns the owning task on failure.
106 */
107static inline struct task_struct *__mutex_trylock_or_owner(struct mutex *lock)
108{
109 unsigned long owner, curr = (unsigned long)current;
110
111 owner = atomic_long_read(&lock->owner);
112 for (;;) { /* must loop, can race against a flag */
113 unsigned long old, flags = __owner_flags(owner);
114 unsigned long task = owner & ~MUTEX_FLAGS;
115
116 if (task) {
117 if (likely(task != curr))
118 break;
119
120 if (likely(!(flags & MUTEX_FLAG_PICKUP)))
121 break;
122
123 flags &= ~MUTEX_FLAG_PICKUP;
124 } else {
125#ifdef CONFIG_DEBUG_MUTEXES
126 DEBUG_LOCKS_WARN_ON(flags & MUTEX_FLAG_PICKUP);
127#endif
128 }
129
130 /*
131 * We set the HANDOFF bit, we must make sure it doesn't live
132 * past the point where we acquire it. This would be possible
133 * if we (accidentally) set the bit on an unlocked mutex.
134 */
135 flags &= ~MUTEX_FLAG_HANDOFF;
136
137 old = atomic_long_cmpxchg_acquire(&lock->owner, owner, curr | flags);
138 if (old == owner)
139 return NULL;
140
141 owner = old;
142 }
143
144 return __owner_task(owner);
145}
146
147/*
148 * Actual trylock that will work on any unlocked state.
149 */
150static inline bool __mutex_trylock(struct mutex *lock)
151{
152 return !__mutex_trylock_or_owner(lock);
153}
154
155#ifndef CONFIG_DEBUG_LOCK_ALLOC
156/*
157 * Lockdep annotations are contained to the slow paths for simplicity.
158 * There is nothing that would stop spreading the lockdep annotations outwards
159 * except more code.
160 */
161
162/*
163 * Optimistic trylock that only works in the uncontended case. Make sure to
164 * follow with a __mutex_trylock() before failing.
165 */
166static __always_inline bool __mutex_trylock_fast(struct mutex *lock)
167{
168 unsigned long curr = (unsigned long)current;
169 unsigned long zero = 0UL;
170
171 if (atomic_long_try_cmpxchg_acquire(&lock->owner, &zero, curr))
172 return true;
173
174 return false;
175}
176
177static __always_inline bool __mutex_unlock_fast(struct mutex *lock)
178{
179 unsigned long curr = (unsigned long)current;
180
181 if (atomic_long_cmpxchg_release(&lock->owner, curr, 0UL) == curr)
182 return true;
183
184 return false;
185}
186#endif
187
188static inline void __mutex_set_flag(struct mutex *lock, unsigned long flag)
189{
190 atomic_long_or(flag, &lock->owner);
191}
192
193static inline void __mutex_clear_flag(struct mutex *lock, unsigned long flag)
194{
195 atomic_long_andnot(flag, &lock->owner);
196}
197
198static inline bool __mutex_waiter_is_first(struct mutex *lock, struct mutex_waiter *waiter)
199{
200 return list_first_entry(&lock->wait_list, struct mutex_waiter, list) == waiter;
201}
202
203/*
204 * Add @waiter to a given location in the lock wait_list and set the
205 * FLAG_WAITERS flag if it's the first waiter.
206 */
207static void __sched
208__mutex_add_waiter(struct mutex *lock, struct mutex_waiter *waiter,
209 struct list_head *list)
210{
211 debug_mutex_add_waiter(lock, waiter, current);
212
213 list_add_tail(&waiter->list, list);
214 if (__mutex_waiter_is_first(lock, waiter))
215 __mutex_set_flag(lock, MUTEX_FLAG_WAITERS);
216}
217
218/*
219 * Give up ownership to a specific task, when @task = NULL, this is equivalent
220 * to a regular unlock. Sets PICKUP on a handoff, clears HANDOF, preserves
221 * WAITERS. Provides RELEASE semantics like a regular unlock, the
222 * __mutex_trylock() provides a matching ACQUIRE semantics for the handoff.
223 */
224static void __mutex_handoff(struct mutex *lock, struct task_struct *task)
225{
226 unsigned long owner = atomic_long_read(&lock->owner);
227
228 for (;;) {
229 unsigned long old, new;
230
231#ifdef CONFIG_DEBUG_MUTEXES
232 DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
233 DEBUG_LOCKS_WARN_ON(owner & MUTEX_FLAG_PICKUP);
234#endif
235
236 new = (owner & MUTEX_FLAG_WAITERS);
237 new |= (unsigned long)task;
238 if (task)
239 new |= MUTEX_FLAG_PICKUP;
240
241 old = atomic_long_cmpxchg_release(&lock->owner, owner, new);
242 if (old == owner)
243 break;
244
245 owner = old;
246 }
247}
248
249#ifndef CONFIG_DEBUG_LOCK_ALLOC
250/*
251 * We split the mutex lock/unlock logic into separate fastpath and
252 * slowpath functions, to reduce the register pressure on the fastpath.
253 * We also put the fastpath first in the kernel image, to make sure the
254 * branch is predicted by the CPU as default-untaken.
255 */
256static void __sched __mutex_lock_slowpath(struct mutex *lock);
257
258/**
259 * mutex_lock - acquire the mutex
260 * @lock: the mutex to be acquired
261 *
262 * Lock the mutex exclusively for this task. If the mutex is not
263 * available right now, it will sleep until it can get it.
264 *
265 * The mutex must later on be released by the same task that
266 * acquired it. Recursive locking is not allowed. The task
267 * may not exit without first unlocking the mutex. Also, kernel
268 * memory where the mutex resides must not be freed with
269 * the mutex still locked. The mutex must first be initialized
270 * (or statically defined) before it can be locked. memset()-ing
271 * the mutex to 0 is not allowed.
272 *
273 * (The CONFIG_DEBUG_MUTEXES .config option turns on debugging
274 * checks that will enforce the restrictions and will also do
275 * deadlock debugging)
276 *
277 * This function is similar to (but not equivalent to) down().
278 */
279void __sched mutex_lock(struct mutex *lock)
280{
281 might_sleep();
282
283 if (!__mutex_trylock_fast(lock))
284 __mutex_lock_slowpath(lock);
285}
286EXPORT_SYMBOL(mutex_lock);
287#endif
288
289/*
290 * Wait-Die:
291 * The newer transactions are killed when:
292 * It (the new transaction) makes a request for a lock being held
293 * by an older transaction.
294 *
295 * Wound-Wait:
296 * The newer transactions are wounded when:
297 * An older transaction makes a request for a lock being held by
298 * the newer transaction.
299 */
300
301/*
302 * Associate the ww_mutex @ww with the context @ww_ctx under which we acquired
303 * it.
304 */
305static __always_inline void
306ww_mutex_lock_acquired(struct ww_mutex *ww, struct ww_acquire_ctx *ww_ctx)
307{
308#ifdef CONFIG_DEBUG_MUTEXES
309 /*
310 * If this WARN_ON triggers, you used ww_mutex_lock to acquire,
311 * but released with a normal mutex_unlock in this call.
312 *
313 * This should never happen, always use ww_mutex_unlock.
314 */
315 DEBUG_LOCKS_WARN_ON(ww->ctx);
316
317 /*
318 * Not quite done after calling ww_acquire_done() ?
319 */
320 DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
321
322 if (ww_ctx->contending_lock) {
323 /*
324 * After -EDEADLK you tried to
325 * acquire a different ww_mutex? Bad!
326 */
327 DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
328
329 /*
330 * You called ww_mutex_lock after receiving -EDEADLK,
331 * but 'forgot' to unlock everything else first?
332 */
333 DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
334 ww_ctx->contending_lock = NULL;
335 }
336
337 /*
338 * Naughty, using a different class will lead to undefined behavior!
339 */
340 DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
341#endif
342 ww_ctx->acquired++;
343 ww->ctx = ww_ctx;
344}
345
346/*
347 * Determine if context @a is 'after' context @b. IOW, @a is a younger
348 * transaction than @b and depending on algorithm either needs to wait for
349 * @b or die.
350 */
351static inline bool __sched
352__ww_ctx_stamp_after(struct ww_acquire_ctx *a, struct ww_acquire_ctx *b)
353{
354
355 return (signed long)(a->stamp - b->stamp) > 0;
356}
357
358/*
359 * Wait-Die; wake a younger waiter context (when locks held) such that it can
360 * die.
361 *
362 * Among waiters with context, only the first one can have other locks acquired
363 * already (ctx->acquired > 0), because __ww_mutex_add_waiter() and
364 * __ww_mutex_check_kill() wake any but the earliest context.
365 */
366static bool __sched
367__ww_mutex_die(struct mutex *lock, struct mutex_waiter *waiter,
368 struct ww_acquire_ctx *ww_ctx)
369{
370 if (!ww_ctx->is_wait_die)
371 return false;
372
373 if (waiter->ww_ctx->acquired > 0 &&
374 __ww_ctx_stamp_after(waiter->ww_ctx, ww_ctx)) {
375 debug_mutex_wake_waiter(lock, waiter);
376 wake_up_process(waiter->task);
377 }
378
379 return true;
380}
381
382/*
383 * Wound-Wait; wound a younger @hold_ctx if it holds the lock.
384 *
385 * Wound the lock holder if there are waiters with older transactions than
386 * the lock holders. Even if multiple waiters may wound the lock holder,
387 * it's sufficient that only one does.
388 */
389static bool __ww_mutex_wound(struct mutex *lock,
390 struct ww_acquire_ctx *ww_ctx,
391 struct ww_acquire_ctx *hold_ctx)
392{
393 struct task_struct *owner = __mutex_owner(lock);
394
395 lockdep_assert_held(&lock->wait_lock);
396
397 /*
398 * Possible through __ww_mutex_add_waiter() when we race with
399 * ww_mutex_set_context_fastpath(). In that case we'll get here again
400 * through __ww_mutex_check_waiters().
401 */
402 if (!hold_ctx)
403 return false;
404
405 /*
406 * Can have !owner because of __mutex_unlock_slowpath(), but if owner,
407 * it cannot go away because we'll have FLAG_WAITERS set and hold
408 * wait_lock.
409 */
410 if (!owner)
411 return false;
412
413 if (ww_ctx->acquired > 0 && __ww_ctx_stamp_after(hold_ctx, ww_ctx)) {
414 hold_ctx->wounded = 1;
415
416 /*
417 * wake_up_process() paired with set_current_state()
418 * inserts sufficient barriers to make sure @owner either sees
419 * it's wounded in __ww_mutex_check_kill() or has a
420 * wakeup pending to re-read the wounded state.
421 */
422 if (owner != current)
423 wake_up_process(owner);
424
425 return true;
426 }
427
428 return false;
429}
430
431/*
432 * We just acquired @lock under @ww_ctx, if there are later contexts waiting
433 * behind us on the wait-list, check if they need to die, or wound us.
434 *
435 * See __ww_mutex_add_waiter() for the list-order construction; basically the
436 * list is ordered by stamp, smallest (oldest) first.
437 *
438 * This relies on never mixing wait-die/wound-wait on the same wait-list;
439 * which is currently ensured by that being a ww_class property.
440 *
441 * The current task must not be on the wait list.
442 */
443static void __sched
444__ww_mutex_check_waiters(struct mutex *lock, struct ww_acquire_ctx *ww_ctx)
445{
446 struct mutex_waiter *cur;
447
448 lockdep_assert_held(&lock->wait_lock);
449
450 list_for_each_entry(cur, &lock->wait_list, list) {
451 if (!cur->ww_ctx)
452 continue;
453
454 if (__ww_mutex_die(lock, cur, ww_ctx) ||
455 __ww_mutex_wound(lock, cur->ww_ctx, ww_ctx))
456 break;
457 }
458}
459
460/*
461 * After acquiring lock with fastpath, where we do not hold wait_lock, set ctx
462 * and wake up any waiters so they can recheck.
463 */
464static __always_inline void
465ww_mutex_set_context_fastpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
466{
467 ww_mutex_lock_acquired(lock, ctx);
468
469 /*
470 * The lock->ctx update should be visible on all cores before
471 * the WAITERS check is done, otherwise contended waiters might be
472 * missed. The contended waiters will either see ww_ctx == NULL
473 * and keep spinning, or it will acquire wait_lock, add itself
474 * to waiter list and sleep.
475 */
476 smp_mb(); /* See comments above and below. */
477
478 /*
479 * [W] ww->ctx = ctx [W] MUTEX_FLAG_WAITERS
480 * MB MB
481 * [R] MUTEX_FLAG_WAITERS [R] ww->ctx
482 *
483 * The memory barrier above pairs with the memory barrier in
484 * __ww_mutex_add_waiter() and makes sure we either observe ww->ctx
485 * and/or !empty list.
486 */
487 if (likely(!(atomic_long_read(&lock->base.owner) & MUTEX_FLAG_WAITERS)))
488 return;
489
490 /*
491 * Uh oh, we raced in fastpath, check if any of the waiters need to
492 * die or wound us.
493 */
494 spin_lock(&lock->base.wait_lock);
495 __ww_mutex_check_waiters(&lock->base, ctx);
496 spin_unlock(&lock->base.wait_lock);
497}
498
499#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
500
501static inline
502bool ww_mutex_spin_on_owner(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
503 struct mutex_waiter *waiter)
504{
505 struct ww_mutex *ww;
506
507 ww = container_of(lock, struct ww_mutex, base);
508
509 /*
510 * If ww->ctx is set the contents are undefined, only
511 * by acquiring wait_lock there is a guarantee that
512 * they are not invalid when reading.
513 *
514 * As such, when deadlock detection needs to be
515 * performed the optimistic spinning cannot be done.
516 *
517 * Check this in every inner iteration because we may
518 * be racing against another thread's ww_mutex_lock.
519 */
520 if (ww_ctx->acquired > 0 && READ_ONCE(ww->ctx))
521 return false;
522
523 /*
524 * If we aren't on the wait list yet, cancel the spin
525 * if there are waiters. We want to avoid stealing the
526 * lock from a waiter with an earlier stamp, since the
527 * other thread may already own a lock that we also
528 * need.
529 */
530 if (!waiter && (atomic_long_read(&lock->owner) & MUTEX_FLAG_WAITERS))
531 return false;
532
533 /*
534 * Similarly, stop spinning if we are no longer the
535 * first waiter.
536 */
537 if (waiter && !__mutex_waiter_is_first(lock, waiter))
538 return false;
539
540 return true;
541}
542
543/*
544 * Look out! "owner" is an entirely speculative pointer access and not
545 * reliable.
546 *
547 * "noinline" so that this function shows up on perf profiles.
548 */
549static noinline
550bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
551 struct ww_acquire_ctx *ww_ctx, struct mutex_waiter *waiter)
552{
553 bool ret = true;
554
555 rcu_read_lock();
556 while (__mutex_owner(lock) == owner) {
557 /*
558 * Ensure we emit the owner->on_cpu, dereference _after_
559 * checking lock->owner still matches owner. If that fails,
560 * owner might point to freed memory. If it still matches,
561 * the rcu_read_lock() ensures the memory stays valid.
562 */
563 barrier();
564
565 /*
566 * Use vcpu_is_preempted to detect lock holder preemption issue.
567 */
568 if (!owner->on_cpu || need_resched() ||
569 vcpu_is_preempted(task_cpu(owner))) {
570 ret = false;
571 break;
572 }
573
574 if (ww_ctx && !ww_mutex_spin_on_owner(lock, ww_ctx, waiter)) {
575 ret = false;
576 break;
577 }
578
579 cpu_relax();
580 }
581 rcu_read_unlock();
582
583 return ret;
584}
585
586/*
587 * Initial check for entering the mutex spinning loop
588 */
589static inline int mutex_can_spin_on_owner(struct mutex *lock)
590{
591 struct task_struct *owner;
592 int retval = 1;
593
594 if (need_resched())
595 return 0;
596
597 rcu_read_lock();
598 owner = __mutex_owner(lock);
599
600 /*
601 * As lock holder preemption issue, we both skip spinning if task is not
602 * on cpu or its cpu is preempted
603 */
604 if (owner)
605 retval = owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
606 rcu_read_unlock();
607
608 /*
609 * If lock->owner is not set, the mutex has been released. Return true
610 * such that we'll trylock in the spin path, which is a faster option
611 * than the blocking slow path.
612 */
613 return retval;
614}
615
616/*
617 * Optimistic spinning.
618 *
619 * We try to spin for acquisition when we find that the lock owner
620 * is currently running on a (different) CPU and while we don't
621 * need to reschedule. The rationale is that if the lock owner is
622 * running, it is likely to release the lock soon.
623 *
624 * The mutex spinners are queued up using MCS lock so that only one
625 * spinner can compete for the mutex. However, if mutex spinning isn't
626 * going to happen, there is no point in going through the lock/unlock
627 * overhead.
628 *
629 * Returns true when the lock was taken, otherwise false, indicating
630 * that we need to jump to the slowpath and sleep.
631 *
632 * The waiter flag is set to true if the spinner is a waiter in the wait
633 * queue. The waiter-spinner will spin on the lock directly and concurrently
634 * with the spinner at the head of the OSQ, if present, until the owner is
635 * changed to itself.
636 */
637static __always_inline bool
638mutex_optimistic_spin(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
639 const bool use_ww_ctx, struct mutex_waiter *waiter)
640{
641 if (!waiter) {
642 /*
643 * The purpose of the mutex_can_spin_on_owner() function is
644 * to eliminate the overhead of osq_lock() and osq_unlock()
645 * in case spinning isn't possible. As a waiter-spinner
646 * is not going to take OSQ lock anyway, there is no need
647 * to call mutex_can_spin_on_owner().
648 */
649 if (!mutex_can_spin_on_owner(lock))
650 goto fail;
651
652 /*
653 * In order to avoid a stampede of mutex spinners trying to
654 * acquire the mutex all at once, the spinners need to take a
655 * MCS (queued) lock first before spinning on the owner field.
656 */
657 if (!osq_lock(&lock->osq))
658 goto fail;
659 }
660
661 for (;;) {
662 struct task_struct *owner;
663
664 /* Try to acquire the mutex... */
665 owner = __mutex_trylock_or_owner(lock);
666 if (!owner)
667 break;
668
669 /*
670 * There's an owner, wait for it to either
671 * release the lock or go to sleep.
672 */
673 if (!mutex_spin_on_owner(lock, owner, ww_ctx, waiter))
674 goto fail_unlock;
675
676 /*
677 * The cpu_relax() call is a compiler barrier which forces
678 * everything in this loop to be re-loaded. We don't need
679 * memory barriers as we'll eventually observe the right
680 * values at the cost of a few extra spins.
681 */
682 cpu_relax();
683 }
684
685 if (!waiter)
686 osq_unlock(&lock->osq);
687
688 return true;
689
690
691fail_unlock:
692 if (!waiter)
693 osq_unlock(&lock->osq);
694
695fail:
696 /*
697 * If we fell out of the spin path because of need_resched(),
698 * reschedule now, before we try-lock the mutex. This avoids getting
699 * scheduled out right after we obtained the mutex.
700 */
701 if (need_resched()) {
702 /*
703 * We _should_ have TASK_RUNNING here, but just in case
704 * we do not, make it so, otherwise we might get stuck.
705 */
706 __set_current_state(TASK_RUNNING);
707 schedule_preempt_disabled();
708 }
709
710 return false;
711}
712#else
713static __always_inline bool
714mutex_optimistic_spin(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
715 const bool use_ww_ctx, struct mutex_waiter *waiter)
716{
717 return false;
718}
719#endif
720
721static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip);
722
723/**
724 * mutex_unlock - release the mutex
725 * @lock: the mutex to be released
726 *
727 * Unlock a mutex that has been locked by this task previously.
728 *
729 * This function must not be used in interrupt context. Unlocking
730 * of a not locked mutex is not allowed.
731 *
732 * This function is similar to (but not equivalent to) up().
733 */
734void __sched mutex_unlock(struct mutex *lock)
735{
736#ifndef CONFIG_DEBUG_LOCK_ALLOC
737 if (__mutex_unlock_fast(lock))
738 return;
739#endif
740 __mutex_unlock_slowpath(lock, _RET_IP_);
741}
742EXPORT_SYMBOL(mutex_unlock);
743
744/**
745 * ww_mutex_unlock - release the w/w mutex
746 * @lock: the mutex to be released
747 *
748 * Unlock a mutex that has been locked by this task previously with any of the
749 * ww_mutex_lock* functions (with or without an acquire context). It is
750 * forbidden to release the locks after releasing the acquire context.
751 *
752 * This function must not be used in interrupt context. Unlocking
753 * of a unlocked mutex is not allowed.
754 */
755void __sched ww_mutex_unlock(struct ww_mutex *lock)
756{
757 /*
758 * The unlocking fastpath is the 0->1 transition from 'locked'
759 * into 'unlocked' state:
760 */
761 if (lock->ctx) {
762#ifdef CONFIG_DEBUG_MUTEXES
763 DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
764#endif
765 if (lock->ctx->acquired > 0)
766 lock->ctx->acquired--;
767 lock->ctx = NULL;
768 }
769
770 mutex_unlock(&lock->base);
771}
772EXPORT_SYMBOL(ww_mutex_unlock);
773
774
775static __always_inline int __sched
776__ww_mutex_kill(struct mutex *lock, struct ww_acquire_ctx *ww_ctx)
777{
778 if (ww_ctx->acquired > 0) {
779#ifdef CONFIG_DEBUG_MUTEXES
780 struct ww_mutex *ww;
781
782 ww = container_of(lock, struct ww_mutex, base);
783 DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock);
784 ww_ctx->contending_lock = ww;
785#endif
786 return -EDEADLK;
787 }
788
789 return 0;
790}
791
792
793/*
794 * Check the wound condition for the current lock acquire.
795 *
796 * Wound-Wait: If we're wounded, kill ourself.
797 *
798 * Wait-Die: If we're trying to acquire a lock already held by an older
799 * context, kill ourselves.
800 *
801 * Since __ww_mutex_add_waiter() orders the wait-list on stamp, we only have to
802 * look at waiters before us in the wait-list.
803 */
804static inline int __sched
805__ww_mutex_check_kill(struct mutex *lock, struct mutex_waiter *waiter,
806 struct ww_acquire_ctx *ctx)
807{
808 struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
809 struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
810 struct mutex_waiter *cur;
811
812 if (ctx->acquired == 0)
813 return 0;
814
815 if (!ctx->is_wait_die) {
816 if (ctx->wounded)
817 return __ww_mutex_kill(lock, ctx);
818
819 return 0;
820 }
821
822 if (hold_ctx && __ww_ctx_stamp_after(ctx, hold_ctx))
823 return __ww_mutex_kill(lock, ctx);
824
825 /*
826 * If there is a waiter in front of us that has a context, then its
827 * stamp is earlier than ours and we must kill ourself.
828 */
829 cur = waiter;
830 list_for_each_entry_continue_reverse(cur, &lock->wait_list, list) {
831 if (!cur->ww_ctx)
832 continue;
833
834 return __ww_mutex_kill(lock, ctx);
835 }
836
837 return 0;
838}
839
840/*
841 * Add @waiter to the wait-list, keep the wait-list ordered by stamp, smallest
842 * first. Such that older contexts are preferred to acquire the lock over
843 * younger contexts.
844 *
845 * Waiters without context are interspersed in FIFO order.
846 *
847 * Furthermore, for Wait-Die kill ourself immediately when possible (there are
848 * older contexts already waiting) to avoid unnecessary waiting and for
849 * Wound-Wait ensure we wound the owning context when it is younger.
850 */
851static inline int __sched
852__ww_mutex_add_waiter(struct mutex_waiter *waiter,
853 struct mutex *lock,
854 struct ww_acquire_ctx *ww_ctx)
855{
856 struct mutex_waiter *cur;
857 struct list_head *pos;
858 bool is_wait_die;
859
860 if (!ww_ctx) {
861 __mutex_add_waiter(lock, waiter, &lock->wait_list);
862 return 0;
863 }
864
865 is_wait_die = ww_ctx->is_wait_die;
866
867 /*
868 * Add the waiter before the first waiter with a higher stamp.
869 * Waiters without a context are skipped to avoid starving
870 * them. Wait-Die waiters may die here. Wound-Wait waiters
871 * never die here, but they are sorted in stamp order and
872 * may wound the lock holder.
873 */
874 pos = &lock->wait_list;
875 list_for_each_entry_reverse(cur, &lock->wait_list, list) {
876 if (!cur->ww_ctx)
877 continue;
878
879 if (__ww_ctx_stamp_after(ww_ctx, cur->ww_ctx)) {
880 /*
881 * Wait-Die: if we find an older context waiting, there
882 * is no point in queueing behind it, as we'd have to
883 * die the moment it would acquire the lock.
884 */
885 if (is_wait_die) {
886 int ret = __ww_mutex_kill(lock, ww_ctx);
887
888 if (ret)
889 return ret;
890 }
891
892 break;
893 }
894
895 pos = &cur->list;
896
897 /* Wait-Die: ensure younger waiters die. */
898 __ww_mutex_die(lock, cur, ww_ctx);
899 }
900
901 __mutex_add_waiter(lock, waiter, pos);
902
903 /*
904 * Wound-Wait: if we're blocking on a mutex owned by a younger context,
905 * wound that such that we might proceed.
906 */
907 if (!is_wait_die) {
908 struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
909
910 /*
911 * See ww_mutex_set_context_fastpath(). Orders setting
912 * MUTEX_FLAG_WAITERS vs the ww->ctx load,
913 * such that either we or the fastpath will wound @ww->ctx.
914 */
915 smp_mb();
916 __ww_mutex_wound(lock, ww_ctx, ww->ctx);
917 }
918
919 return 0;
920}
921
922/*
923 * Lock a mutex (possibly interruptible), slowpath:
924 */
925static __always_inline int __sched
926__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
927 struct lockdep_map *nest_lock, unsigned long ip,
928 struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
929{
930 struct mutex_waiter waiter;
931 bool first = false;
932 struct ww_mutex *ww;
933 int ret;
934
935 might_sleep();
936
937#ifdef CONFIG_DEBUG_MUTEXES
938 DEBUG_LOCKS_WARN_ON(lock->magic != lock);
939#endif
940
941 ww = container_of(lock, struct ww_mutex, base);
942 if (use_ww_ctx && ww_ctx) {
943 if (unlikely(ww_ctx == READ_ONCE(ww->ctx)))
944 return -EALREADY;
945
946 /*
947 * Reset the wounded flag after a kill. No other process can
948 * race and wound us here since they can't have a valid owner
949 * pointer if we don't have any locks held.
950 */
951 if (ww_ctx->acquired == 0)
952 ww_ctx->wounded = 0;
953 }
954
955 preempt_disable();
956 mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
957
958 if (__mutex_trylock(lock) ||
959 mutex_optimistic_spin(lock, ww_ctx, use_ww_ctx, NULL)) {
960 /* got the lock, yay! */
961 lock_acquired(&lock->dep_map, ip);
962 if (use_ww_ctx && ww_ctx)
963 ww_mutex_set_context_fastpath(ww, ww_ctx);
964 preempt_enable();
965 return 0;
966 }
967
968 spin_lock(&lock->wait_lock);
969 /*
970 * After waiting to acquire the wait_lock, try again.
971 */
972 if (__mutex_trylock(lock)) {
973 if (use_ww_ctx && ww_ctx)
974 __ww_mutex_check_waiters(lock, ww_ctx);
975
976 goto skip_wait;
977 }
978
979 debug_mutex_lock_common(lock, &waiter);
980
981 lock_contended(&lock->dep_map, ip);
982
983 if (!use_ww_ctx) {
984 /* add waiting tasks to the end of the waitqueue (FIFO): */
985 __mutex_add_waiter(lock, &waiter, &lock->wait_list);
986
987
988#ifdef CONFIG_DEBUG_MUTEXES
989 waiter.ww_ctx = MUTEX_POISON_WW_CTX;
990#endif
991 } else {
992 /*
993 * Add in stamp order, waking up waiters that must kill
994 * themselves.
995 */
996 ret = __ww_mutex_add_waiter(&waiter, lock, ww_ctx);
997 if (ret)
998 goto err_early_kill;
999
1000 waiter.ww_ctx = ww_ctx;
1001 }
1002
1003 waiter.task = current;
1004
1005 set_current_state(state);
1006 for (;;) {
1007 /*
1008 * Once we hold wait_lock, we're serialized against
1009 * mutex_unlock() handing the lock off to us, do a trylock
1010 * before testing the error conditions to make sure we pick up
1011 * the handoff.
1012 */
1013 if (__mutex_trylock(lock))
1014 goto acquired;
1015
1016 /*
1017 * Check for signals and kill conditions while holding
1018 * wait_lock. This ensures the lock cancellation is ordered
1019 * against mutex_unlock() and wake-ups do not go missing.
1020 */
1021 if (signal_pending_state(state, current)) {
1022 ret = -EINTR;
1023 goto err;
1024 }
1025
1026 if (use_ww_ctx && ww_ctx) {
1027 ret = __ww_mutex_check_kill(lock, &waiter, ww_ctx);
1028 if (ret)
1029 goto err;
1030 }
1031
1032 spin_unlock(&lock->wait_lock);
1033 schedule_preempt_disabled();
1034
1035 /*
1036 * ww_mutex needs to always recheck its position since its waiter
1037 * list is not FIFO ordered.
1038 */
1039 if ((use_ww_ctx && ww_ctx) || !first) {
1040 first = __mutex_waiter_is_first(lock, &waiter);
1041 if (first)
1042 __mutex_set_flag(lock, MUTEX_FLAG_HANDOFF);
1043 }
1044
1045 set_current_state(state);
1046 /*
1047 * Here we order against unlock; we must either see it change
1048 * state back to RUNNING and fall through the next schedule(),
1049 * or we must see its unlock and acquire.
1050 */
1051 if (__mutex_trylock(lock) ||
1052 (first && mutex_optimistic_spin(lock, ww_ctx, use_ww_ctx, &waiter)))
1053 break;
1054
1055 spin_lock(&lock->wait_lock);
1056 }
1057 spin_lock(&lock->wait_lock);
1058acquired:
1059 __set_current_state(TASK_RUNNING);
1060
1061 if (use_ww_ctx && ww_ctx) {
1062 /*
1063 * Wound-Wait; we stole the lock (!first_waiter), check the
1064 * waiters as anyone might want to wound us.
1065 */
1066 if (!ww_ctx->is_wait_die &&
1067 !__mutex_waiter_is_first(lock, &waiter))
1068 __ww_mutex_check_waiters(lock, ww_ctx);
1069 }
1070
1071 mutex_remove_waiter(lock, &waiter, current);
1072 if (likely(list_empty(&lock->wait_list)))
1073 __mutex_clear_flag(lock, MUTEX_FLAGS);
1074
1075 debug_mutex_free_waiter(&waiter);
1076
1077skip_wait:
1078 /* got the lock - cleanup and rejoice! */
1079 lock_acquired(&lock->dep_map, ip);
1080
1081 if (use_ww_ctx && ww_ctx)
1082 ww_mutex_lock_acquired(ww, ww_ctx);
1083
1084 spin_unlock(&lock->wait_lock);
1085 preempt_enable();
1086 return 0;
1087
1088err:
1089 __set_current_state(TASK_RUNNING);
1090 mutex_remove_waiter(lock, &waiter, current);
1091err_early_kill:
1092 spin_unlock(&lock->wait_lock);
1093 debug_mutex_free_waiter(&waiter);
1094 mutex_release(&lock->dep_map, 1, ip);
1095 preempt_enable();
1096 return ret;
1097}
1098
1099static int __sched
1100__mutex_lock(struct mutex *lock, long state, unsigned int subclass,
1101 struct lockdep_map *nest_lock, unsigned long ip)
1102{
1103 return __mutex_lock_common(lock, state, subclass, nest_lock, ip, NULL, false);
1104}
1105
1106static int __sched
1107__ww_mutex_lock(struct mutex *lock, long state, unsigned int subclass,
1108 struct lockdep_map *nest_lock, unsigned long ip,
1109 struct ww_acquire_ctx *ww_ctx)
1110{
1111 return __mutex_lock_common(lock, state, subclass, nest_lock, ip, ww_ctx, true);
1112}
1113
1114#ifdef CONFIG_DEBUG_LOCK_ALLOC
1115void __sched
1116mutex_lock_nested(struct mutex *lock, unsigned int subclass)
1117{
1118 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, subclass, NULL, _RET_IP_);
1119}
1120
1121EXPORT_SYMBOL_GPL(mutex_lock_nested);
1122
1123void __sched
1124_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
1125{
1126 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0, nest, _RET_IP_);
1127}
1128EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
1129
1130int __sched
1131mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
1132{
1133 return __mutex_lock(lock, TASK_KILLABLE, subclass, NULL, _RET_IP_);
1134}
1135EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
1136
1137int __sched
1138mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
1139{
1140 return __mutex_lock(lock, TASK_INTERRUPTIBLE, subclass, NULL, _RET_IP_);
1141}
1142EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
1143
1144void __sched
1145mutex_lock_io_nested(struct mutex *lock, unsigned int subclass)
1146{
1147 int token;
1148
1149 might_sleep();
1150
1151 token = io_schedule_prepare();
1152 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
1153 subclass, NULL, _RET_IP_, NULL, 0);
1154 io_schedule_finish(token);
1155}
1156EXPORT_SYMBOL_GPL(mutex_lock_io_nested);
1157
1158static inline int
1159ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1160{
1161#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
1162 unsigned tmp;
1163
1164 if (ctx->deadlock_inject_countdown-- == 0) {
1165 tmp = ctx->deadlock_inject_interval;
1166 if (tmp > UINT_MAX/4)
1167 tmp = UINT_MAX;
1168 else
1169 tmp = tmp*2 + tmp + tmp/2;
1170
1171 ctx->deadlock_inject_interval = tmp;
1172 ctx->deadlock_inject_countdown = tmp;
1173 ctx->contending_lock = lock;
1174
1175 ww_mutex_unlock(lock);
1176
1177 return -EDEADLK;
1178 }
1179#endif
1180
1181 return 0;
1182}
1183
1184int __sched
1185ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1186{
1187 int ret;
1188
1189 might_sleep();
1190 ret = __ww_mutex_lock(&lock->base, TASK_UNINTERRUPTIBLE,
1191 0, ctx ? &ctx->dep_map : NULL, _RET_IP_,
1192 ctx);
1193 if (!ret && ctx && ctx->acquired > 1)
1194 return ww_mutex_deadlock_injection(lock, ctx);
1195
1196 return ret;
1197}
1198EXPORT_SYMBOL_GPL(ww_mutex_lock);
1199
1200int __sched
1201ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1202{
1203 int ret;
1204
1205 might_sleep();
1206 ret = __ww_mutex_lock(&lock->base, TASK_INTERRUPTIBLE,
1207 0, ctx ? &ctx->dep_map : NULL, _RET_IP_,
1208 ctx);
1209
1210 if (!ret && ctx && ctx->acquired > 1)
1211 return ww_mutex_deadlock_injection(lock, ctx);
1212
1213 return ret;
1214}
1215EXPORT_SYMBOL_GPL(ww_mutex_lock_interruptible);
1216
1217#endif
1218
1219/*
1220 * Release the lock, slowpath:
1221 */
1222static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip)
1223{
1224 struct task_struct *next = NULL;
1225 DEFINE_WAKE_Q(wake_q);
1226 unsigned long owner;
1227
1228 mutex_release(&lock->dep_map, 1, ip);
1229
1230 /*
1231 * Release the lock before (potentially) taking the spinlock such that
1232 * other contenders can get on with things ASAP.
1233 *
1234 * Except when HANDOFF, in that case we must not clear the owner field,
1235 * but instead set it to the top waiter.
1236 */
1237 owner = atomic_long_read(&lock->owner);
1238 for (;;) {
1239 unsigned long old;
1240
1241#ifdef CONFIG_DEBUG_MUTEXES
1242 DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
1243 DEBUG_LOCKS_WARN_ON(owner & MUTEX_FLAG_PICKUP);
1244#endif
1245
1246 if (owner & MUTEX_FLAG_HANDOFF)
1247 break;
1248
1249 old = atomic_long_cmpxchg_release(&lock->owner, owner,
1250 __owner_flags(owner));
1251 if (old == owner) {
1252 if (owner & MUTEX_FLAG_WAITERS)
1253 break;
1254
1255 return;
1256 }
1257
1258 owner = old;
1259 }
1260
1261 spin_lock(&lock->wait_lock);
1262 debug_mutex_unlock(lock);
1263 if (!list_empty(&lock->wait_list)) {
1264 /* get the first entry from the wait-list: */
1265 struct mutex_waiter *waiter =
1266 list_first_entry(&lock->wait_list,
1267 struct mutex_waiter, list);
1268
1269 next = waiter->task;
1270
1271 debug_mutex_wake_waiter(lock, waiter);
1272 wake_q_add(&wake_q, next);
1273 }
1274
1275 if (owner & MUTEX_FLAG_HANDOFF)
1276 __mutex_handoff(lock, next);
1277
1278 spin_unlock(&lock->wait_lock);
1279
1280 wake_up_q(&wake_q);
1281}
1282
1283#ifndef CONFIG_DEBUG_LOCK_ALLOC
1284/*
1285 * Here come the less common (and hence less performance-critical) APIs:
1286 * mutex_lock_interruptible() and mutex_trylock().
1287 */
1288static noinline int __sched
1289__mutex_lock_killable_slowpath(struct mutex *lock);
1290
1291static noinline int __sched
1292__mutex_lock_interruptible_slowpath(struct mutex *lock);
1293
1294/**
1295 * mutex_lock_interruptible() - Acquire the mutex, interruptible by signals.
1296 * @lock: The mutex to be acquired.
1297 *
1298 * Lock the mutex like mutex_lock(). If a signal is delivered while the
1299 * process is sleeping, this function will return without acquiring the
1300 * mutex.
1301 *
1302 * Context: Process context.
1303 * Return: 0 if the lock was successfully acquired or %-EINTR if a
1304 * signal arrived.
1305 */
1306int __sched mutex_lock_interruptible(struct mutex *lock)
1307{
1308 might_sleep();
1309
1310 if (__mutex_trylock_fast(lock))
1311 return 0;
1312
1313 return __mutex_lock_interruptible_slowpath(lock);
1314}
1315
1316EXPORT_SYMBOL(mutex_lock_interruptible);
1317
1318/**
1319 * mutex_lock_killable() - Acquire the mutex, interruptible by fatal signals.
1320 * @lock: The mutex to be acquired.
1321 *
1322 * Lock the mutex like mutex_lock(). If a signal which will be fatal to
1323 * the current process is delivered while the process is sleeping, this
1324 * function will return without acquiring the mutex.
1325 *
1326 * Context: Process context.
1327 * Return: 0 if the lock was successfully acquired or %-EINTR if a
1328 * fatal signal arrived.
1329 */
1330int __sched mutex_lock_killable(struct mutex *lock)
1331{
1332 might_sleep();
1333
1334 if (__mutex_trylock_fast(lock))
1335 return 0;
1336
1337 return __mutex_lock_killable_slowpath(lock);
1338}
1339EXPORT_SYMBOL(mutex_lock_killable);
1340
1341/**
1342 * mutex_lock_io() - Acquire the mutex and mark the process as waiting for I/O
1343 * @lock: The mutex to be acquired.
1344 *
1345 * Lock the mutex like mutex_lock(). While the task is waiting for this
1346 * mutex, it will be accounted as being in the IO wait state by the
1347 * scheduler.
1348 *
1349 * Context: Process context.
1350 */
1351void __sched mutex_lock_io(struct mutex *lock)
1352{
1353 int token;
1354
1355 token = io_schedule_prepare();
1356 mutex_lock(lock);
1357 io_schedule_finish(token);
1358}
1359EXPORT_SYMBOL_GPL(mutex_lock_io);
1360
1361static noinline void __sched
1362__mutex_lock_slowpath(struct mutex *lock)
1363{
1364 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0, NULL, _RET_IP_);
1365}
1366
1367static noinline int __sched
1368__mutex_lock_killable_slowpath(struct mutex *lock)
1369{
1370 return __mutex_lock(lock, TASK_KILLABLE, 0, NULL, _RET_IP_);
1371}
1372
1373static noinline int __sched
1374__mutex_lock_interruptible_slowpath(struct mutex *lock)
1375{
1376 return __mutex_lock(lock, TASK_INTERRUPTIBLE, 0, NULL, _RET_IP_);
1377}
1378
1379static noinline int __sched
1380__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1381{
1382 return __ww_mutex_lock(&lock->base, TASK_UNINTERRUPTIBLE, 0, NULL,
1383 _RET_IP_, ctx);
1384}
1385
1386static noinline int __sched
1387__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
1388 struct ww_acquire_ctx *ctx)
1389{
1390 return __ww_mutex_lock(&lock->base, TASK_INTERRUPTIBLE, 0, NULL,
1391 _RET_IP_, ctx);
1392}
1393
1394#endif
1395
1396/**
1397 * mutex_trylock - try to acquire the mutex, without waiting
1398 * @lock: the mutex to be acquired
1399 *
1400 * Try to acquire the mutex atomically. Returns 1 if the mutex
1401 * has been acquired successfully, and 0 on contention.
1402 *
1403 * NOTE: this function follows the spin_trylock() convention, so
1404 * it is negated from the down_trylock() return values! Be careful
1405 * about this when converting semaphore users to mutexes.
1406 *
1407 * This function must not be used in interrupt context. The
1408 * mutex must be released by the same task that acquired it.
1409 */
1410int __sched mutex_trylock(struct mutex *lock)
1411{
1412 bool locked;
1413
1414#ifdef CONFIG_DEBUG_MUTEXES
1415 DEBUG_LOCKS_WARN_ON(lock->magic != lock);
1416#endif
1417
1418 locked = __mutex_trylock(lock);
1419 if (locked)
1420 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1421
1422 return locked;
1423}
1424EXPORT_SYMBOL(mutex_trylock);
1425
1426#ifndef CONFIG_DEBUG_LOCK_ALLOC
1427int __sched
1428ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1429{
1430 might_sleep();
1431
1432 if (__mutex_trylock_fast(&lock->base)) {
1433 if (ctx)
1434 ww_mutex_set_context_fastpath(lock, ctx);
1435 return 0;
1436 }
1437
1438 return __ww_mutex_lock_slowpath(lock, ctx);
1439}
1440EXPORT_SYMBOL(ww_mutex_lock);
1441
1442int __sched
1443ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1444{
1445 might_sleep();
1446
1447 if (__mutex_trylock_fast(&lock->base)) {
1448 if (ctx)
1449 ww_mutex_set_context_fastpath(lock, ctx);
1450 return 0;
1451 }
1452
1453 return __ww_mutex_lock_interruptible_slowpath(lock, ctx);
1454}
1455EXPORT_SYMBOL(ww_mutex_lock_interruptible);
1456
1457#endif
1458
1459/**
1460 * atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
1461 * @cnt: the atomic which we are to dec
1462 * @lock: the mutex to return holding if we dec to 0
1463 *
1464 * return true and hold lock if we dec to 0, return false otherwise
1465 */
1466int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
1467{
1468 /* dec if we can't possibly hit 0 */
1469 if (atomic_add_unless(cnt, -1, 1))
1470 return 0;
1471 /* we might hit 0, so take the lock */
1472 mutex_lock(lock);
1473 if (!atomic_dec_and_test(cnt)) {
1474 /* when we actually did the dec, we didn't hit 0 */
1475 mutex_unlock(lock);
1476 return 0;
1477 }
1478 /* we hit 0, and we hold the lock */
1479 return 1;
1480}
1481EXPORT_SYMBOL(atomic_dec_and_mutex_lock);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/locking/mutex.c
4 *
5 * Mutexes: blocking mutual exclusion locks
6 *
7 * Started by Ingo Molnar:
8 *
9 * Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
10 *
11 * Many thanks to Arjan van de Ven, Thomas Gleixner, Steven Rostedt and
12 * David Howells for suggestions and improvements.
13 *
14 * - Adaptive spinning for mutexes by Peter Zijlstra. (Ported to mainline
15 * from the -rt tree, where it was originally implemented for rtmutexes
16 * by Steven Rostedt, based on work by Gregory Haskins, Peter Morreale
17 * and Sven Dietrich.
18 *
19 * Also see Documentation/locking/mutex-design.rst.
20 */
21#include <linux/mutex.h>
22#include <linux/ww_mutex.h>
23#include <linux/sched/signal.h>
24#include <linux/sched/rt.h>
25#include <linux/sched/wake_q.h>
26#include <linux/sched/debug.h>
27#include <linux/export.h>
28#include <linux/spinlock.h>
29#include <linux/interrupt.h>
30#include <linux/debug_locks.h>
31#include <linux/osq_lock.h>
32
33#define CREATE_TRACE_POINTS
34#include <trace/events/lock.h>
35
36#ifndef CONFIG_PREEMPT_RT
37#include "mutex.h"
38
39#ifdef CONFIG_DEBUG_MUTEXES
40# define MUTEX_WARN_ON(cond) DEBUG_LOCKS_WARN_ON(cond)
41#else
42# define MUTEX_WARN_ON(cond)
43#endif
44
45void
46__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
47{
48 atomic_long_set(&lock->owner, 0);
49 raw_spin_lock_init(&lock->wait_lock);
50 INIT_LIST_HEAD(&lock->wait_list);
51#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
52 osq_lock_init(&lock->osq);
53#endif
54
55 debug_mutex_init(lock, name, key);
56}
57EXPORT_SYMBOL(__mutex_init);
58
59/*
60 * @owner: contains: 'struct task_struct *' to the current lock owner,
61 * NULL means not owned. Since task_struct pointers are aligned at
62 * at least L1_CACHE_BYTES, we have low bits to store extra state.
63 *
64 * Bit0 indicates a non-empty waiter list; unlock must issue a wakeup.
65 * Bit1 indicates unlock needs to hand the lock to the top-waiter
66 * Bit2 indicates handoff has been done and we're waiting for pickup.
67 */
68#define MUTEX_FLAG_WAITERS 0x01
69#define MUTEX_FLAG_HANDOFF 0x02
70#define MUTEX_FLAG_PICKUP 0x04
71
72#define MUTEX_FLAGS 0x07
73
74/*
75 * Internal helper function; C doesn't allow us to hide it :/
76 *
77 * DO NOT USE (outside of mutex code).
78 */
79static inline struct task_struct *__mutex_owner(struct mutex *lock)
80{
81 return (struct task_struct *)(atomic_long_read(&lock->owner) & ~MUTEX_FLAGS);
82}
83
84static inline struct task_struct *__owner_task(unsigned long owner)
85{
86 return (struct task_struct *)(owner & ~MUTEX_FLAGS);
87}
88
89bool mutex_is_locked(struct mutex *lock)
90{
91 return __mutex_owner(lock) != NULL;
92}
93EXPORT_SYMBOL(mutex_is_locked);
94
95static inline unsigned long __owner_flags(unsigned long owner)
96{
97 return owner & MUTEX_FLAGS;
98}
99
100/*
101 * Returns: __mutex_owner(lock) on failure or NULL on success.
102 */
103static inline struct task_struct *__mutex_trylock_common(struct mutex *lock, bool handoff)
104{
105 unsigned long owner, curr = (unsigned long)current;
106
107 owner = atomic_long_read(&lock->owner);
108 for (;;) { /* must loop, can race against a flag */
109 unsigned long flags = __owner_flags(owner);
110 unsigned long task = owner & ~MUTEX_FLAGS;
111
112 if (task) {
113 if (flags & MUTEX_FLAG_PICKUP) {
114 if (task != curr)
115 break;
116 flags &= ~MUTEX_FLAG_PICKUP;
117 } else if (handoff) {
118 if (flags & MUTEX_FLAG_HANDOFF)
119 break;
120 flags |= MUTEX_FLAG_HANDOFF;
121 } else {
122 break;
123 }
124 } else {
125 MUTEX_WARN_ON(flags & (MUTEX_FLAG_HANDOFF | MUTEX_FLAG_PICKUP));
126 task = curr;
127 }
128
129 if (atomic_long_try_cmpxchg_acquire(&lock->owner, &owner, task | flags)) {
130 if (task == curr)
131 return NULL;
132 break;
133 }
134 }
135
136 return __owner_task(owner);
137}
138
139/*
140 * Trylock or set HANDOFF
141 */
142static inline bool __mutex_trylock_or_handoff(struct mutex *lock, bool handoff)
143{
144 return !__mutex_trylock_common(lock, handoff);
145}
146
147/*
148 * Actual trylock that will work on any unlocked state.
149 */
150static inline bool __mutex_trylock(struct mutex *lock)
151{
152 return !__mutex_trylock_common(lock, false);
153}
154
155#ifndef CONFIG_DEBUG_LOCK_ALLOC
156/*
157 * Lockdep annotations are contained to the slow paths for simplicity.
158 * There is nothing that would stop spreading the lockdep annotations outwards
159 * except more code.
160 */
161
162/*
163 * Optimistic trylock that only works in the uncontended case. Make sure to
164 * follow with a __mutex_trylock() before failing.
165 */
166static __always_inline bool __mutex_trylock_fast(struct mutex *lock)
167{
168 unsigned long curr = (unsigned long)current;
169 unsigned long zero = 0UL;
170
171 if (atomic_long_try_cmpxchg_acquire(&lock->owner, &zero, curr))
172 return true;
173
174 return false;
175}
176
177static __always_inline bool __mutex_unlock_fast(struct mutex *lock)
178{
179 unsigned long curr = (unsigned long)current;
180
181 return atomic_long_try_cmpxchg_release(&lock->owner, &curr, 0UL);
182}
183#endif
184
185static inline void __mutex_set_flag(struct mutex *lock, unsigned long flag)
186{
187 atomic_long_or(flag, &lock->owner);
188}
189
190static inline void __mutex_clear_flag(struct mutex *lock, unsigned long flag)
191{
192 atomic_long_andnot(flag, &lock->owner);
193}
194
195static inline bool __mutex_waiter_is_first(struct mutex *lock, struct mutex_waiter *waiter)
196{
197 return list_first_entry(&lock->wait_list, struct mutex_waiter, list) == waiter;
198}
199
200/*
201 * Add @waiter to a given location in the lock wait_list and set the
202 * FLAG_WAITERS flag if it's the first waiter.
203 */
204static void
205__mutex_add_waiter(struct mutex *lock, struct mutex_waiter *waiter,
206 struct list_head *list)
207{
208 debug_mutex_add_waiter(lock, waiter, current);
209
210 list_add_tail(&waiter->list, list);
211 if (__mutex_waiter_is_first(lock, waiter))
212 __mutex_set_flag(lock, MUTEX_FLAG_WAITERS);
213}
214
215static void
216__mutex_remove_waiter(struct mutex *lock, struct mutex_waiter *waiter)
217{
218 list_del(&waiter->list);
219 if (likely(list_empty(&lock->wait_list)))
220 __mutex_clear_flag(lock, MUTEX_FLAGS);
221
222 debug_mutex_remove_waiter(lock, waiter, current);
223}
224
225/*
226 * Give up ownership to a specific task, when @task = NULL, this is equivalent
227 * to a regular unlock. Sets PICKUP on a handoff, clears HANDOFF, preserves
228 * WAITERS. Provides RELEASE semantics like a regular unlock, the
229 * __mutex_trylock() provides a matching ACQUIRE semantics for the handoff.
230 */
231static void __mutex_handoff(struct mutex *lock, struct task_struct *task)
232{
233 unsigned long owner = atomic_long_read(&lock->owner);
234
235 for (;;) {
236 unsigned long new;
237
238 MUTEX_WARN_ON(__owner_task(owner) != current);
239 MUTEX_WARN_ON(owner & MUTEX_FLAG_PICKUP);
240
241 new = (owner & MUTEX_FLAG_WAITERS);
242 new |= (unsigned long)task;
243 if (task)
244 new |= MUTEX_FLAG_PICKUP;
245
246 if (atomic_long_try_cmpxchg_release(&lock->owner, &owner, new))
247 break;
248 }
249}
250
251#ifndef CONFIG_DEBUG_LOCK_ALLOC
252/*
253 * We split the mutex lock/unlock logic into separate fastpath and
254 * slowpath functions, to reduce the register pressure on the fastpath.
255 * We also put the fastpath first in the kernel image, to make sure the
256 * branch is predicted by the CPU as default-untaken.
257 */
258static void __sched __mutex_lock_slowpath(struct mutex *lock);
259
260/**
261 * mutex_lock - acquire the mutex
262 * @lock: the mutex to be acquired
263 *
264 * Lock the mutex exclusively for this task. If the mutex is not
265 * available right now, it will sleep until it can get it.
266 *
267 * The mutex must later on be released by the same task that
268 * acquired it. Recursive locking is not allowed. The task
269 * may not exit without first unlocking the mutex. Also, kernel
270 * memory where the mutex resides must not be freed with
271 * the mutex still locked. The mutex must first be initialized
272 * (or statically defined) before it can be locked. memset()-ing
273 * the mutex to 0 is not allowed.
274 *
275 * (The CONFIG_DEBUG_MUTEXES .config option turns on debugging
276 * checks that will enforce the restrictions and will also do
277 * deadlock debugging)
278 *
279 * This function is similar to (but not equivalent to) down().
280 */
281void __sched mutex_lock(struct mutex *lock)
282{
283 might_sleep();
284
285 if (!__mutex_trylock_fast(lock))
286 __mutex_lock_slowpath(lock);
287}
288EXPORT_SYMBOL(mutex_lock);
289#endif
290
291#include "ww_mutex.h"
292
293#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
294
295/*
296 * Trylock variant that returns the owning task on failure.
297 */
298static inline struct task_struct *__mutex_trylock_or_owner(struct mutex *lock)
299{
300 return __mutex_trylock_common(lock, false);
301}
302
303static inline
304bool ww_mutex_spin_on_owner(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
305 struct mutex_waiter *waiter)
306{
307 struct ww_mutex *ww;
308
309 ww = container_of(lock, struct ww_mutex, base);
310
311 /*
312 * If ww->ctx is set the contents are undefined, only
313 * by acquiring wait_lock there is a guarantee that
314 * they are not invalid when reading.
315 *
316 * As such, when deadlock detection needs to be
317 * performed the optimistic spinning cannot be done.
318 *
319 * Check this in every inner iteration because we may
320 * be racing against another thread's ww_mutex_lock.
321 */
322 if (ww_ctx->acquired > 0 && READ_ONCE(ww->ctx))
323 return false;
324
325 /*
326 * If we aren't on the wait list yet, cancel the spin
327 * if there are waiters. We want to avoid stealing the
328 * lock from a waiter with an earlier stamp, since the
329 * other thread may already own a lock that we also
330 * need.
331 */
332 if (!waiter && (atomic_long_read(&lock->owner) & MUTEX_FLAG_WAITERS))
333 return false;
334
335 /*
336 * Similarly, stop spinning if we are no longer the
337 * first waiter.
338 */
339 if (waiter && !__mutex_waiter_is_first(lock, waiter))
340 return false;
341
342 return true;
343}
344
345/*
346 * Look out! "owner" is an entirely speculative pointer access and not
347 * reliable.
348 *
349 * "noinline" so that this function shows up on perf profiles.
350 */
351static noinline
352bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
353 struct ww_acquire_ctx *ww_ctx, struct mutex_waiter *waiter)
354{
355 bool ret = true;
356
357 lockdep_assert_preemption_disabled();
358
359 while (__mutex_owner(lock) == owner) {
360 /*
361 * Ensure we emit the owner->on_cpu, dereference _after_
362 * checking lock->owner still matches owner. And we already
363 * disabled preemption which is equal to the RCU read-side
364 * crital section in optimistic spinning code. Thus the
365 * task_strcut structure won't go away during the spinning
366 * period
367 */
368 barrier();
369
370 /*
371 * Use vcpu_is_preempted to detect lock holder preemption issue.
372 */
373 if (!owner_on_cpu(owner) || need_resched()) {
374 ret = false;
375 break;
376 }
377
378 if (ww_ctx && !ww_mutex_spin_on_owner(lock, ww_ctx, waiter)) {
379 ret = false;
380 break;
381 }
382
383 cpu_relax();
384 }
385
386 return ret;
387}
388
389/*
390 * Initial check for entering the mutex spinning loop
391 */
392static inline int mutex_can_spin_on_owner(struct mutex *lock)
393{
394 struct task_struct *owner;
395 int retval = 1;
396
397 lockdep_assert_preemption_disabled();
398
399 if (need_resched())
400 return 0;
401
402 /*
403 * We already disabled preemption which is equal to the RCU read-side
404 * crital section in optimistic spinning code. Thus the task_strcut
405 * structure won't go away during the spinning period.
406 */
407 owner = __mutex_owner(lock);
408 if (owner)
409 retval = owner_on_cpu(owner);
410
411 /*
412 * If lock->owner is not set, the mutex has been released. Return true
413 * such that we'll trylock in the spin path, which is a faster option
414 * than the blocking slow path.
415 */
416 return retval;
417}
418
419/*
420 * Optimistic spinning.
421 *
422 * We try to spin for acquisition when we find that the lock owner
423 * is currently running on a (different) CPU and while we don't
424 * need to reschedule. The rationale is that if the lock owner is
425 * running, it is likely to release the lock soon.
426 *
427 * The mutex spinners are queued up using MCS lock so that only one
428 * spinner can compete for the mutex. However, if mutex spinning isn't
429 * going to happen, there is no point in going through the lock/unlock
430 * overhead.
431 *
432 * Returns true when the lock was taken, otherwise false, indicating
433 * that we need to jump to the slowpath and sleep.
434 *
435 * The waiter flag is set to true if the spinner is a waiter in the wait
436 * queue. The waiter-spinner will spin on the lock directly and concurrently
437 * with the spinner at the head of the OSQ, if present, until the owner is
438 * changed to itself.
439 */
440static __always_inline bool
441mutex_optimistic_spin(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
442 struct mutex_waiter *waiter)
443{
444 if (!waiter) {
445 /*
446 * The purpose of the mutex_can_spin_on_owner() function is
447 * to eliminate the overhead of osq_lock() and osq_unlock()
448 * in case spinning isn't possible. As a waiter-spinner
449 * is not going to take OSQ lock anyway, there is no need
450 * to call mutex_can_spin_on_owner().
451 */
452 if (!mutex_can_spin_on_owner(lock))
453 goto fail;
454
455 /*
456 * In order to avoid a stampede of mutex spinners trying to
457 * acquire the mutex all at once, the spinners need to take a
458 * MCS (queued) lock first before spinning on the owner field.
459 */
460 if (!osq_lock(&lock->osq))
461 goto fail;
462 }
463
464 for (;;) {
465 struct task_struct *owner;
466
467 /* Try to acquire the mutex... */
468 owner = __mutex_trylock_or_owner(lock);
469 if (!owner)
470 break;
471
472 /*
473 * There's an owner, wait for it to either
474 * release the lock or go to sleep.
475 */
476 if (!mutex_spin_on_owner(lock, owner, ww_ctx, waiter))
477 goto fail_unlock;
478
479 /*
480 * The cpu_relax() call is a compiler barrier which forces
481 * everything in this loop to be re-loaded. We don't need
482 * memory barriers as we'll eventually observe the right
483 * values at the cost of a few extra spins.
484 */
485 cpu_relax();
486 }
487
488 if (!waiter)
489 osq_unlock(&lock->osq);
490
491 return true;
492
493
494fail_unlock:
495 if (!waiter)
496 osq_unlock(&lock->osq);
497
498fail:
499 /*
500 * If we fell out of the spin path because of need_resched(),
501 * reschedule now, before we try-lock the mutex. This avoids getting
502 * scheduled out right after we obtained the mutex.
503 */
504 if (need_resched()) {
505 /*
506 * We _should_ have TASK_RUNNING here, but just in case
507 * we do not, make it so, otherwise we might get stuck.
508 */
509 __set_current_state(TASK_RUNNING);
510 schedule_preempt_disabled();
511 }
512
513 return false;
514}
515#else
516static __always_inline bool
517mutex_optimistic_spin(struct mutex *lock, struct ww_acquire_ctx *ww_ctx,
518 struct mutex_waiter *waiter)
519{
520 return false;
521}
522#endif
523
524static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip);
525
526/**
527 * mutex_unlock - release the mutex
528 * @lock: the mutex to be released
529 *
530 * Unlock a mutex that has been locked by this task previously.
531 *
532 * This function must not be used in interrupt context. Unlocking
533 * of a not locked mutex is not allowed.
534 *
535 * The caller must ensure that the mutex stays alive until this function has
536 * returned - mutex_unlock() can NOT directly be used to release an object such
537 * that another concurrent task can free it.
538 * Mutexes are different from spinlocks & refcounts in this aspect.
539 *
540 * This function is similar to (but not equivalent to) up().
541 */
542void __sched mutex_unlock(struct mutex *lock)
543{
544#ifndef CONFIG_DEBUG_LOCK_ALLOC
545 if (__mutex_unlock_fast(lock))
546 return;
547#endif
548 __mutex_unlock_slowpath(lock, _RET_IP_);
549}
550EXPORT_SYMBOL(mutex_unlock);
551
552/**
553 * ww_mutex_unlock - release the w/w mutex
554 * @lock: the mutex to be released
555 *
556 * Unlock a mutex that has been locked by this task previously with any of the
557 * ww_mutex_lock* functions (with or without an acquire context). It is
558 * forbidden to release the locks after releasing the acquire context.
559 *
560 * This function must not be used in interrupt context. Unlocking
561 * of a unlocked mutex is not allowed.
562 */
563void __sched ww_mutex_unlock(struct ww_mutex *lock)
564{
565 __ww_mutex_unlock(lock);
566 mutex_unlock(&lock->base);
567}
568EXPORT_SYMBOL(ww_mutex_unlock);
569
570/*
571 * Lock a mutex (possibly interruptible), slowpath:
572 */
573static __always_inline int __sched
574__mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclass,
575 struct lockdep_map *nest_lock, unsigned long ip,
576 struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
577{
578 struct mutex_waiter waiter;
579 struct ww_mutex *ww;
580 int ret;
581
582 if (!use_ww_ctx)
583 ww_ctx = NULL;
584
585 might_sleep();
586
587 MUTEX_WARN_ON(lock->magic != lock);
588
589 ww = container_of(lock, struct ww_mutex, base);
590 if (ww_ctx) {
591 if (unlikely(ww_ctx == READ_ONCE(ww->ctx)))
592 return -EALREADY;
593
594 /*
595 * Reset the wounded flag after a kill. No other process can
596 * race and wound us here since they can't have a valid owner
597 * pointer if we don't have any locks held.
598 */
599 if (ww_ctx->acquired == 0)
600 ww_ctx->wounded = 0;
601
602#ifdef CONFIG_DEBUG_LOCK_ALLOC
603 nest_lock = &ww_ctx->dep_map;
604#endif
605 }
606
607 preempt_disable();
608 mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
609
610 trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
611 if (__mutex_trylock(lock) ||
612 mutex_optimistic_spin(lock, ww_ctx, NULL)) {
613 /* got the lock, yay! */
614 lock_acquired(&lock->dep_map, ip);
615 if (ww_ctx)
616 ww_mutex_set_context_fastpath(ww, ww_ctx);
617 trace_contention_end(lock, 0);
618 preempt_enable();
619 return 0;
620 }
621
622 raw_spin_lock(&lock->wait_lock);
623 /*
624 * After waiting to acquire the wait_lock, try again.
625 */
626 if (__mutex_trylock(lock)) {
627 if (ww_ctx)
628 __ww_mutex_check_waiters(lock, ww_ctx);
629
630 goto skip_wait;
631 }
632
633 debug_mutex_lock_common(lock, &waiter);
634 waiter.task = current;
635 if (use_ww_ctx)
636 waiter.ww_ctx = ww_ctx;
637
638 lock_contended(&lock->dep_map, ip);
639
640 if (!use_ww_ctx) {
641 /* add waiting tasks to the end of the waitqueue (FIFO): */
642 __mutex_add_waiter(lock, &waiter, &lock->wait_list);
643 } else {
644 /*
645 * Add in stamp order, waking up waiters that must kill
646 * themselves.
647 */
648 ret = __ww_mutex_add_waiter(&waiter, lock, ww_ctx);
649 if (ret)
650 goto err_early_kill;
651 }
652
653 set_current_state(state);
654 trace_contention_begin(lock, LCB_F_MUTEX);
655 for (;;) {
656 bool first;
657
658 /*
659 * Once we hold wait_lock, we're serialized against
660 * mutex_unlock() handing the lock off to us, do a trylock
661 * before testing the error conditions to make sure we pick up
662 * the handoff.
663 */
664 if (__mutex_trylock(lock))
665 goto acquired;
666
667 /*
668 * Check for signals and kill conditions while holding
669 * wait_lock. This ensures the lock cancellation is ordered
670 * against mutex_unlock() and wake-ups do not go missing.
671 */
672 if (signal_pending_state(state, current)) {
673 ret = -EINTR;
674 goto err;
675 }
676
677 if (ww_ctx) {
678 ret = __ww_mutex_check_kill(lock, &waiter, ww_ctx);
679 if (ret)
680 goto err;
681 }
682
683 raw_spin_unlock(&lock->wait_lock);
684 schedule_preempt_disabled();
685
686 first = __mutex_waiter_is_first(lock, &waiter);
687
688 set_current_state(state);
689 /*
690 * Here we order against unlock; we must either see it change
691 * state back to RUNNING and fall through the next schedule(),
692 * or we must see its unlock and acquire.
693 */
694 if (__mutex_trylock_or_handoff(lock, first))
695 break;
696
697 if (first) {
698 trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
699 if (mutex_optimistic_spin(lock, ww_ctx, &waiter))
700 break;
701 trace_contention_begin(lock, LCB_F_MUTEX);
702 }
703
704 raw_spin_lock(&lock->wait_lock);
705 }
706 raw_spin_lock(&lock->wait_lock);
707acquired:
708 __set_current_state(TASK_RUNNING);
709
710 if (ww_ctx) {
711 /*
712 * Wound-Wait; we stole the lock (!first_waiter), check the
713 * waiters as anyone might want to wound us.
714 */
715 if (!ww_ctx->is_wait_die &&
716 !__mutex_waiter_is_first(lock, &waiter))
717 __ww_mutex_check_waiters(lock, ww_ctx);
718 }
719
720 __mutex_remove_waiter(lock, &waiter);
721
722 debug_mutex_free_waiter(&waiter);
723
724skip_wait:
725 /* got the lock - cleanup and rejoice! */
726 lock_acquired(&lock->dep_map, ip);
727 trace_contention_end(lock, 0);
728
729 if (ww_ctx)
730 ww_mutex_lock_acquired(ww, ww_ctx);
731
732 raw_spin_unlock(&lock->wait_lock);
733 preempt_enable();
734 return 0;
735
736err:
737 __set_current_state(TASK_RUNNING);
738 __mutex_remove_waiter(lock, &waiter);
739err_early_kill:
740 trace_contention_end(lock, ret);
741 raw_spin_unlock(&lock->wait_lock);
742 debug_mutex_free_waiter(&waiter);
743 mutex_release(&lock->dep_map, ip);
744 preempt_enable();
745 return ret;
746}
747
748static int __sched
749__mutex_lock(struct mutex *lock, unsigned int state, unsigned int subclass,
750 struct lockdep_map *nest_lock, unsigned long ip)
751{
752 return __mutex_lock_common(lock, state, subclass, nest_lock, ip, NULL, false);
753}
754
755static int __sched
756__ww_mutex_lock(struct mutex *lock, unsigned int state, unsigned int subclass,
757 unsigned long ip, struct ww_acquire_ctx *ww_ctx)
758{
759 return __mutex_lock_common(lock, state, subclass, NULL, ip, ww_ctx, true);
760}
761
762/**
763 * ww_mutex_trylock - tries to acquire the w/w mutex with optional acquire context
764 * @ww: mutex to lock
765 * @ww_ctx: optional w/w acquire context
766 *
767 * Trylocks a mutex with the optional acquire context; no deadlock detection is
768 * possible. Returns 1 if the mutex has been acquired successfully, 0 otherwise.
769 *
770 * Unlike ww_mutex_lock, no deadlock handling is performed. However, if a @ctx is
771 * specified, -EALREADY handling may happen in calls to ww_mutex_trylock.
772 *
773 * A mutex acquired with this function must be released with ww_mutex_unlock.
774 */
775int ww_mutex_trylock(struct ww_mutex *ww, struct ww_acquire_ctx *ww_ctx)
776{
777 if (!ww_ctx)
778 return mutex_trylock(&ww->base);
779
780 MUTEX_WARN_ON(ww->base.magic != &ww->base);
781
782 /*
783 * Reset the wounded flag after a kill. No other process can
784 * race and wound us here, since they can't have a valid owner
785 * pointer if we don't have any locks held.
786 */
787 if (ww_ctx->acquired == 0)
788 ww_ctx->wounded = 0;
789
790 if (__mutex_trylock(&ww->base)) {
791 ww_mutex_set_context_fastpath(ww, ww_ctx);
792 mutex_acquire_nest(&ww->base.dep_map, 0, 1, &ww_ctx->dep_map, _RET_IP_);
793 return 1;
794 }
795
796 return 0;
797}
798EXPORT_SYMBOL(ww_mutex_trylock);
799
800#ifdef CONFIG_DEBUG_LOCK_ALLOC
801void __sched
802mutex_lock_nested(struct mutex *lock, unsigned int subclass)
803{
804 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, subclass, NULL, _RET_IP_);
805}
806
807EXPORT_SYMBOL_GPL(mutex_lock_nested);
808
809void __sched
810_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
811{
812 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0, nest, _RET_IP_);
813}
814EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
815
816int __sched
817mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
818{
819 return __mutex_lock(lock, TASK_KILLABLE, subclass, NULL, _RET_IP_);
820}
821EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
822
823int __sched
824mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
825{
826 return __mutex_lock(lock, TASK_INTERRUPTIBLE, subclass, NULL, _RET_IP_);
827}
828EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
829
830void __sched
831mutex_lock_io_nested(struct mutex *lock, unsigned int subclass)
832{
833 int token;
834
835 might_sleep();
836
837 token = io_schedule_prepare();
838 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
839 subclass, NULL, _RET_IP_, NULL, 0);
840 io_schedule_finish(token);
841}
842EXPORT_SYMBOL_GPL(mutex_lock_io_nested);
843
844static inline int
845ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
846{
847#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
848 unsigned tmp;
849
850 if (ctx->deadlock_inject_countdown-- == 0) {
851 tmp = ctx->deadlock_inject_interval;
852 if (tmp > UINT_MAX/4)
853 tmp = UINT_MAX;
854 else
855 tmp = tmp*2 + tmp + tmp/2;
856
857 ctx->deadlock_inject_interval = tmp;
858 ctx->deadlock_inject_countdown = tmp;
859 ctx->contending_lock = lock;
860
861 ww_mutex_unlock(lock);
862
863 return -EDEADLK;
864 }
865#endif
866
867 return 0;
868}
869
870int __sched
871ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
872{
873 int ret;
874
875 might_sleep();
876 ret = __ww_mutex_lock(&lock->base, TASK_UNINTERRUPTIBLE,
877 0, _RET_IP_, ctx);
878 if (!ret && ctx && ctx->acquired > 1)
879 return ww_mutex_deadlock_injection(lock, ctx);
880
881 return ret;
882}
883EXPORT_SYMBOL_GPL(ww_mutex_lock);
884
885int __sched
886ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
887{
888 int ret;
889
890 might_sleep();
891 ret = __ww_mutex_lock(&lock->base, TASK_INTERRUPTIBLE,
892 0, _RET_IP_, ctx);
893
894 if (!ret && ctx && ctx->acquired > 1)
895 return ww_mutex_deadlock_injection(lock, ctx);
896
897 return ret;
898}
899EXPORT_SYMBOL_GPL(ww_mutex_lock_interruptible);
900
901#endif
902
903/*
904 * Release the lock, slowpath:
905 */
906static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip)
907{
908 struct task_struct *next = NULL;
909 DEFINE_WAKE_Q(wake_q);
910 unsigned long owner;
911
912 mutex_release(&lock->dep_map, ip);
913
914 /*
915 * Release the lock before (potentially) taking the spinlock such that
916 * other contenders can get on with things ASAP.
917 *
918 * Except when HANDOFF, in that case we must not clear the owner field,
919 * but instead set it to the top waiter.
920 */
921 owner = atomic_long_read(&lock->owner);
922 for (;;) {
923 MUTEX_WARN_ON(__owner_task(owner) != current);
924 MUTEX_WARN_ON(owner & MUTEX_FLAG_PICKUP);
925
926 if (owner & MUTEX_FLAG_HANDOFF)
927 break;
928
929 if (atomic_long_try_cmpxchg_release(&lock->owner, &owner, __owner_flags(owner))) {
930 if (owner & MUTEX_FLAG_WAITERS)
931 break;
932
933 return;
934 }
935 }
936
937 raw_spin_lock(&lock->wait_lock);
938 debug_mutex_unlock(lock);
939 if (!list_empty(&lock->wait_list)) {
940 /* get the first entry from the wait-list: */
941 struct mutex_waiter *waiter =
942 list_first_entry(&lock->wait_list,
943 struct mutex_waiter, list);
944
945 next = waiter->task;
946
947 debug_mutex_wake_waiter(lock, waiter);
948 wake_q_add(&wake_q, next);
949 }
950
951 if (owner & MUTEX_FLAG_HANDOFF)
952 __mutex_handoff(lock, next);
953
954 raw_spin_unlock(&lock->wait_lock);
955
956 wake_up_q(&wake_q);
957}
958
959#ifndef CONFIG_DEBUG_LOCK_ALLOC
960/*
961 * Here come the less common (and hence less performance-critical) APIs:
962 * mutex_lock_interruptible() and mutex_trylock().
963 */
964static noinline int __sched
965__mutex_lock_killable_slowpath(struct mutex *lock);
966
967static noinline int __sched
968__mutex_lock_interruptible_slowpath(struct mutex *lock);
969
970/**
971 * mutex_lock_interruptible() - Acquire the mutex, interruptible by signals.
972 * @lock: The mutex to be acquired.
973 *
974 * Lock the mutex like mutex_lock(). If a signal is delivered while the
975 * process is sleeping, this function will return without acquiring the
976 * mutex.
977 *
978 * Context: Process context.
979 * Return: 0 if the lock was successfully acquired or %-EINTR if a
980 * signal arrived.
981 */
982int __sched mutex_lock_interruptible(struct mutex *lock)
983{
984 might_sleep();
985
986 if (__mutex_trylock_fast(lock))
987 return 0;
988
989 return __mutex_lock_interruptible_slowpath(lock);
990}
991
992EXPORT_SYMBOL(mutex_lock_interruptible);
993
994/**
995 * mutex_lock_killable() - Acquire the mutex, interruptible by fatal signals.
996 * @lock: The mutex to be acquired.
997 *
998 * Lock the mutex like mutex_lock(). If a signal which will be fatal to
999 * the current process is delivered while the process is sleeping, this
1000 * function will return without acquiring the mutex.
1001 *
1002 * Context: Process context.
1003 * Return: 0 if the lock was successfully acquired or %-EINTR if a
1004 * fatal signal arrived.
1005 */
1006int __sched mutex_lock_killable(struct mutex *lock)
1007{
1008 might_sleep();
1009
1010 if (__mutex_trylock_fast(lock))
1011 return 0;
1012
1013 return __mutex_lock_killable_slowpath(lock);
1014}
1015EXPORT_SYMBOL(mutex_lock_killable);
1016
1017/**
1018 * mutex_lock_io() - Acquire the mutex and mark the process as waiting for I/O
1019 * @lock: The mutex to be acquired.
1020 *
1021 * Lock the mutex like mutex_lock(). While the task is waiting for this
1022 * mutex, it will be accounted as being in the IO wait state by the
1023 * scheduler.
1024 *
1025 * Context: Process context.
1026 */
1027void __sched mutex_lock_io(struct mutex *lock)
1028{
1029 int token;
1030
1031 token = io_schedule_prepare();
1032 mutex_lock(lock);
1033 io_schedule_finish(token);
1034}
1035EXPORT_SYMBOL_GPL(mutex_lock_io);
1036
1037static noinline void __sched
1038__mutex_lock_slowpath(struct mutex *lock)
1039{
1040 __mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0, NULL, _RET_IP_);
1041}
1042
1043static noinline int __sched
1044__mutex_lock_killable_slowpath(struct mutex *lock)
1045{
1046 return __mutex_lock(lock, TASK_KILLABLE, 0, NULL, _RET_IP_);
1047}
1048
1049static noinline int __sched
1050__mutex_lock_interruptible_slowpath(struct mutex *lock)
1051{
1052 return __mutex_lock(lock, TASK_INTERRUPTIBLE, 0, NULL, _RET_IP_);
1053}
1054
1055static noinline int __sched
1056__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1057{
1058 return __ww_mutex_lock(&lock->base, TASK_UNINTERRUPTIBLE, 0,
1059 _RET_IP_, ctx);
1060}
1061
1062static noinline int __sched
1063__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
1064 struct ww_acquire_ctx *ctx)
1065{
1066 return __ww_mutex_lock(&lock->base, TASK_INTERRUPTIBLE, 0,
1067 _RET_IP_, ctx);
1068}
1069
1070#endif
1071
1072/**
1073 * mutex_trylock - try to acquire the mutex, without waiting
1074 * @lock: the mutex to be acquired
1075 *
1076 * Try to acquire the mutex atomically. Returns 1 if the mutex
1077 * has been acquired successfully, and 0 on contention.
1078 *
1079 * NOTE: this function follows the spin_trylock() convention, so
1080 * it is negated from the down_trylock() return values! Be careful
1081 * about this when converting semaphore users to mutexes.
1082 *
1083 * This function must not be used in interrupt context. The
1084 * mutex must be released by the same task that acquired it.
1085 */
1086int __sched mutex_trylock(struct mutex *lock)
1087{
1088 bool locked;
1089
1090 MUTEX_WARN_ON(lock->magic != lock);
1091
1092 locked = __mutex_trylock(lock);
1093 if (locked)
1094 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1095
1096 return locked;
1097}
1098EXPORT_SYMBOL(mutex_trylock);
1099
1100#ifndef CONFIG_DEBUG_LOCK_ALLOC
1101int __sched
1102ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1103{
1104 might_sleep();
1105
1106 if (__mutex_trylock_fast(&lock->base)) {
1107 if (ctx)
1108 ww_mutex_set_context_fastpath(lock, ctx);
1109 return 0;
1110 }
1111
1112 return __ww_mutex_lock_slowpath(lock, ctx);
1113}
1114EXPORT_SYMBOL(ww_mutex_lock);
1115
1116int __sched
1117ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1118{
1119 might_sleep();
1120
1121 if (__mutex_trylock_fast(&lock->base)) {
1122 if (ctx)
1123 ww_mutex_set_context_fastpath(lock, ctx);
1124 return 0;
1125 }
1126
1127 return __ww_mutex_lock_interruptible_slowpath(lock, ctx);
1128}
1129EXPORT_SYMBOL(ww_mutex_lock_interruptible);
1130
1131#endif /* !CONFIG_DEBUG_LOCK_ALLOC */
1132#endif /* !CONFIG_PREEMPT_RT */
1133
1134EXPORT_TRACEPOINT_SYMBOL_GPL(contention_begin);
1135EXPORT_TRACEPOINT_SYMBOL_GPL(contention_end);
1136
1137/**
1138 * atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
1139 * @cnt: the atomic which we are to dec
1140 * @lock: the mutex to return holding if we dec to 0
1141 *
1142 * return true and hold lock if we dec to 0, return false otherwise
1143 */
1144int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
1145{
1146 /* dec if we can't possibly hit 0 */
1147 if (atomic_add_unless(cnt, -1, 1))
1148 return 0;
1149 /* we might hit 0, so take the lock */
1150 mutex_lock(lock);
1151 if (!atomic_dec_and_test(cnt)) {
1152 /* when we actually did the dec, we didn't hit 0 */
1153 mutex_unlock(lock);
1154 return 0;
1155 }
1156 /* we hit 0, and we hold the lock */
1157 return 1;
1158}
1159EXPORT_SYMBOL(atomic_dec_and_mutex_lock);