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