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/mutex-design.txt.
19 */
20#include <linux/mutex.h>
21#include <linux/ww_mutex.h>
22#include <linux/sched.h>
23#include <linux/sched/rt.h>
24#include <linux/export.h>
25#include <linux/spinlock.h>
26#include <linux/interrupt.h>
27#include <linux/debug_locks.h>
28#include "mcs_spinlock.h"
29
30/*
31 * In the DEBUG case we are using the "NULL fastpath" for mutexes,
32 * which forces all calls into the slowpath:
33 */
34#ifdef CONFIG_DEBUG_MUTEXES
35# include "mutex-debug.h"
36# include <asm-generic/mutex-null.h>
37/*
38 * Must be 0 for the debug case so we do not do the unlock outside of the
39 * wait_lock region. debug_mutex_unlock() will do the actual unlock in this
40 * case.
41 */
42# undef __mutex_slowpath_needs_to_unlock
43# define __mutex_slowpath_needs_to_unlock() 0
44#else
45# include "mutex.h"
46# include <asm/mutex.h>
47#endif
48
49/*
50 * A negative mutex count indicates that waiters are sleeping waiting for the
51 * mutex.
52 */
53#define MUTEX_SHOW_NO_WAITER(mutex) (atomic_read(&(mutex)->count) >= 0)
54
55void
56__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
57{
58 atomic_set(&lock->count, 1);
59 spin_lock_init(&lock->wait_lock);
60 INIT_LIST_HEAD(&lock->wait_list);
61 mutex_clear_owner(lock);
62#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
63 lock->osq = NULL;
64#endif
65
66 debug_mutex_init(lock, name, key);
67}
68
69EXPORT_SYMBOL(__mutex_init);
70
71#ifndef CONFIG_DEBUG_LOCK_ALLOC
72/*
73 * We split the mutex lock/unlock logic into separate fastpath and
74 * slowpath functions, to reduce the register pressure on the fastpath.
75 * We also put the fastpath first in the kernel image, to make sure the
76 * branch is predicted by the CPU as default-untaken.
77 */
78__visible void __sched __mutex_lock_slowpath(atomic_t *lock_count);
79
80/**
81 * mutex_lock - acquire the mutex
82 * @lock: the mutex to be acquired
83 *
84 * Lock the mutex exclusively for this task. If the mutex is not
85 * available right now, it will sleep until it can get it.
86 *
87 * The mutex must later on be released by the same task that
88 * acquired it. Recursive locking is not allowed. The task
89 * may not exit without first unlocking the mutex. Also, kernel
90 * memory where the mutex resides mutex must not be freed with
91 * the mutex still locked. The mutex must first be initialized
92 * (or statically defined) before it can be locked. memset()-ing
93 * the mutex to 0 is not allowed.
94 *
95 * ( The CONFIG_DEBUG_MUTEXES .config option turns on debugging
96 * checks that will enforce the restrictions and will also do
97 * deadlock debugging. )
98 *
99 * This function is similar to (but not equivalent to) down().
100 */
101void __sched mutex_lock(struct mutex *lock)
102{
103 might_sleep();
104 /*
105 * The locking fastpath is the 1->0 transition from
106 * 'unlocked' into 'locked' state.
107 */
108 __mutex_fastpath_lock(&lock->count, __mutex_lock_slowpath);
109 mutex_set_owner(lock);
110}
111
112EXPORT_SYMBOL(mutex_lock);
113#endif
114
115#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
116/*
117 * In order to avoid a stampede of mutex spinners from acquiring the mutex
118 * more or less simultaneously, the spinners need to acquire a MCS lock
119 * first before spinning on the owner field.
120 *
121 */
122
123/*
124 * Mutex spinning code migrated from kernel/sched/core.c
125 */
126
127static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
128{
129 if (lock->owner != owner)
130 return false;
131
132 /*
133 * Ensure we emit the owner->on_cpu, dereference _after_ checking
134 * lock->owner still matches owner, if that fails, owner might
135 * point to free()d memory, if it still matches, the rcu_read_lock()
136 * ensures the memory stays valid.
137 */
138 barrier();
139
140 return owner->on_cpu;
141}
142
143/*
144 * Look out! "owner" is an entirely speculative pointer
145 * access and not reliable.
146 */
147static noinline
148int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
149{
150 rcu_read_lock();
151 while (owner_running(lock, owner)) {
152 if (need_resched())
153 break;
154
155 arch_mutex_cpu_relax();
156 }
157 rcu_read_unlock();
158
159 /*
160 * We break out the loop above on need_resched() and when the
161 * owner changed, which is a sign for heavy contention. Return
162 * success only when lock->owner is NULL.
163 */
164 return lock->owner == NULL;
165}
166
167/*
168 * Initial check for entering the mutex spinning loop
169 */
170static inline int mutex_can_spin_on_owner(struct mutex *lock)
171{
172 struct task_struct *owner;
173 int retval = 1;
174
175 if (need_resched())
176 return 0;
177
178 rcu_read_lock();
179 owner = ACCESS_ONCE(lock->owner);
180 if (owner)
181 retval = owner->on_cpu;
182 rcu_read_unlock();
183 /*
184 * if lock->owner is not set, the mutex owner may have just acquired
185 * it and not set the owner yet or the mutex has been released.
186 */
187 return retval;
188}
189#endif
190
191__visible __used noinline
192void __sched __mutex_unlock_slowpath(atomic_t *lock_count);
193
194/**
195 * mutex_unlock - release the mutex
196 * @lock: the mutex to be released
197 *
198 * Unlock a mutex that has been locked by this task previously.
199 *
200 * This function must not be used in interrupt context. Unlocking
201 * of a not locked mutex is not allowed.
202 *
203 * This function is similar to (but not equivalent to) up().
204 */
205void __sched mutex_unlock(struct mutex *lock)
206{
207 /*
208 * The unlocking fastpath is the 0->1 transition from 'locked'
209 * into 'unlocked' state:
210 */
211#ifndef CONFIG_DEBUG_MUTEXES
212 /*
213 * When debugging is enabled we must not clear the owner before time,
214 * the slow path will always be taken, and that clears the owner field
215 * after verifying that it was indeed current.
216 */
217 mutex_clear_owner(lock);
218#endif
219 __mutex_fastpath_unlock(&lock->count, __mutex_unlock_slowpath);
220}
221
222EXPORT_SYMBOL(mutex_unlock);
223
224/**
225 * ww_mutex_unlock - release the w/w mutex
226 * @lock: the mutex to be released
227 *
228 * Unlock a mutex that has been locked by this task previously with any of the
229 * ww_mutex_lock* functions (with or without an acquire context). It is
230 * forbidden to release the locks after releasing the acquire context.
231 *
232 * This function must not be used in interrupt context. Unlocking
233 * of a unlocked mutex is not allowed.
234 */
235void __sched ww_mutex_unlock(struct ww_mutex *lock)
236{
237 /*
238 * The unlocking fastpath is the 0->1 transition from 'locked'
239 * into 'unlocked' state:
240 */
241 if (lock->ctx) {
242#ifdef CONFIG_DEBUG_MUTEXES
243 DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
244#endif
245 if (lock->ctx->acquired > 0)
246 lock->ctx->acquired--;
247 lock->ctx = NULL;
248 }
249
250#ifndef CONFIG_DEBUG_MUTEXES
251 /*
252 * When debugging is enabled we must not clear the owner before time,
253 * the slow path will always be taken, and that clears the owner field
254 * after verifying that it was indeed current.
255 */
256 mutex_clear_owner(&lock->base);
257#endif
258 __mutex_fastpath_unlock(&lock->base.count, __mutex_unlock_slowpath);
259}
260EXPORT_SYMBOL(ww_mutex_unlock);
261
262static inline int __sched
263__mutex_lock_check_stamp(struct mutex *lock, struct ww_acquire_ctx *ctx)
264{
265 struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
266 struct ww_acquire_ctx *hold_ctx = ACCESS_ONCE(ww->ctx);
267
268 if (!hold_ctx)
269 return 0;
270
271 if (unlikely(ctx == hold_ctx))
272 return -EALREADY;
273
274 if (ctx->stamp - hold_ctx->stamp <= LONG_MAX &&
275 (ctx->stamp != hold_ctx->stamp || ctx > hold_ctx)) {
276#ifdef CONFIG_DEBUG_MUTEXES
277 DEBUG_LOCKS_WARN_ON(ctx->contending_lock);
278 ctx->contending_lock = ww;
279#endif
280 return -EDEADLK;
281 }
282
283 return 0;
284}
285
286static __always_inline void ww_mutex_lock_acquired(struct ww_mutex *ww,
287 struct ww_acquire_ctx *ww_ctx)
288{
289#ifdef CONFIG_DEBUG_MUTEXES
290 /*
291 * If this WARN_ON triggers, you used ww_mutex_lock to acquire,
292 * but released with a normal mutex_unlock in this call.
293 *
294 * This should never happen, always use ww_mutex_unlock.
295 */
296 DEBUG_LOCKS_WARN_ON(ww->ctx);
297
298 /*
299 * Not quite done after calling ww_acquire_done() ?
300 */
301 DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
302
303 if (ww_ctx->contending_lock) {
304 /*
305 * After -EDEADLK you tried to
306 * acquire a different ww_mutex? Bad!
307 */
308 DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
309
310 /*
311 * You called ww_mutex_lock after receiving -EDEADLK,
312 * but 'forgot' to unlock everything else first?
313 */
314 DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
315 ww_ctx->contending_lock = NULL;
316 }
317
318 /*
319 * Naughty, using a different class will lead to undefined behavior!
320 */
321 DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
322#endif
323 ww_ctx->acquired++;
324}
325
326/*
327 * after acquiring lock with fastpath or when we lost out in contested
328 * slowpath, set ctx and wake up any waiters so they can recheck.
329 *
330 * This function is never called when CONFIG_DEBUG_LOCK_ALLOC is set,
331 * as the fastpath and opportunistic spinning are disabled in that case.
332 */
333static __always_inline void
334ww_mutex_set_context_fastpath(struct ww_mutex *lock,
335 struct ww_acquire_ctx *ctx)
336{
337 unsigned long flags;
338 struct mutex_waiter *cur;
339
340 ww_mutex_lock_acquired(lock, ctx);
341
342 lock->ctx = ctx;
343
344 /*
345 * The lock->ctx update should be visible on all cores before
346 * the atomic read is done, otherwise contended waiters might be
347 * missed. The contended waiters will either see ww_ctx == NULL
348 * and keep spinning, or it will acquire wait_lock, add itself
349 * to waiter list and sleep.
350 */
351 smp_mb(); /* ^^^ */
352
353 /*
354 * Check if lock is contended, if not there is nobody to wake up
355 */
356 if (likely(atomic_read(&lock->base.count) == 0))
357 return;
358
359 /*
360 * Uh oh, we raced in fastpath, wake up everyone in this case,
361 * so they can see the new lock->ctx.
362 */
363 spin_lock_mutex(&lock->base.wait_lock, flags);
364 list_for_each_entry(cur, &lock->base.wait_list, list) {
365 debug_mutex_wake_waiter(&lock->base, cur);
366 wake_up_process(cur->task);
367 }
368 spin_unlock_mutex(&lock->base.wait_lock, flags);
369}
370
371/*
372 * Lock a mutex (possibly interruptible), slowpath:
373 */
374static __always_inline int __sched
375__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
376 struct lockdep_map *nest_lock, unsigned long ip,
377 struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
378{
379 struct task_struct *task = current;
380 struct mutex_waiter waiter;
381 unsigned long flags;
382 int ret;
383
384 preempt_disable();
385 mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
386
387#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
388 /*
389 * Optimistic spinning.
390 *
391 * We try to spin for acquisition when we find that there are no
392 * pending waiters and the lock owner is currently running on a
393 * (different) CPU.
394 *
395 * The rationale is that if the lock owner is running, it is likely to
396 * release the lock soon.
397 *
398 * Since this needs the lock owner, and this mutex implementation
399 * doesn't track the owner atomically in the lock field, we need to
400 * track it non-atomically.
401 *
402 * We can't do this for DEBUG_MUTEXES because that relies on wait_lock
403 * to serialize everything.
404 *
405 * The mutex spinners are queued up using MCS lock so that only one
406 * spinner can compete for the mutex. However, if mutex spinning isn't
407 * going to happen, there is no point in going through the lock/unlock
408 * overhead.
409 */
410 if (!mutex_can_spin_on_owner(lock))
411 goto slowpath;
412
413 if (!osq_lock(&lock->osq))
414 goto slowpath;
415
416 for (;;) {
417 struct task_struct *owner;
418
419 if (use_ww_ctx && ww_ctx->acquired > 0) {
420 struct ww_mutex *ww;
421
422 ww = container_of(lock, struct ww_mutex, base);
423 /*
424 * If ww->ctx is set the contents are undefined, only
425 * by acquiring wait_lock there is a guarantee that
426 * they are not invalid when reading.
427 *
428 * As such, when deadlock detection needs to be
429 * performed the optimistic spinning cannot be done.
430 */
431 if (ACCESS_ONCE(ww->ctx))
432 break;
433 }
434
435 /*
436 * If there's an owner, wait for it to either
437 * release the lock or go to sleep.
438 */
439 owner = ACCESS_ONCE(lock->owner);
440 if (owner && !mutex_spin_on_owner(lock, owner))
441 break;
442
443 if ((atomic_read(&lock->count) == 1) &&
444 (atomic_cmpxchg(&lock->count, 1, 0) == 1)) {
445 lock_acquired(&lock->dep_map, ip);
446 if (use_ww_ctx) {
447 struct ww_mutex *ww;
448 ww = container_of(lock, struct ww_mutex, base);
449
450 ww_mutex_set_context_fastpath(ww, ww_ctx);
451 }
452
453 mutex_set_owner(lock);
454 osq_unlock(&lock->osq);
455 preempt_enable();
456 return 0;
457 }
458
459 /*
460 * When there's no owner, we might have preempted between the
461 * owner acquiring the lock and setting the owner field. If
462 * we're an RT task that will live-lock because we won't let
463 * the owner complete.
464 */
465 if (!owner && (need_resched() || rt_task(task)))
466 break;
467
468 /*
469 * The cpu_relax() call is a compiler barrier which forces
470 * everything in this loop to be re-loaded. We don't need
471 * memory barriers as we'll eventually observe the right
472 * values at the cost of a few extra spins.
473 */
474 arch_mutex_cpu_relax();
475 }
476 osq_unlock(&lock->osq);
477slowpath:
478 /*
479 * If we fell out of the spin path because of need_resched(),
480 * reschedule now, before we try-lock the mutex. This avoids getting
481 * scheduled out right after we obtained the mutex.
482 */
483 if (need_resched())
484 schedule_preempt_disabled();
485#endif
486 spin_lock_mutex(&lock->wait_lock, flags);
487
488 /* once more, can we acquire the lock? */
489 if (MUTEX_SHOW_NO_WAITER(lock) && (atomic_xchg(&lock->count, 0) == 1))
490 goto skip_wait;
491
492 debug_mutex_lock_common(lock, &waiter);
493 debug_mutex_add_waiter(lock, &waiter, task_thread_info(task));
494
495 /* add waiting tasks to the end of the waitqueue (FIFO): */
496 list_add_tail(&waiter.list, &lock->wait_list);
497 waiter.task = task;
498
499 lock_contended(&lock->dep_map, ip);
500
501 for (;;) {
502 /*
503 * Lets try to take the lock again - this is needed even if
504 * we get here for the first time (shortly after failing to
505 * acquire the lock), to make sure that we get a wakeup once
506 * it's unlocked. Later on, if we sleep, this is the
507 * operation that gives us the lock. We xchg it to -1, so
508 * that when we release the lock, we properly wake up the
509 * other waiters:
510 */
511 if (MUTEX_SHOW_NO_WAITER(lock) &&
512 (atomic_xchg(&lock->count, -1) == 1))
513 break;
514
515 /*
516 * got a signal? (This code gets eliminated in the
517 * TASK_UNINTERRUPTIBLE case.)
518 */
519 if (unlikely(signal_pending_state(state, task))) {
520 ret = -EINTR;
521 goto err;
522 }
523
524 if (use_ww_ctx && ww_ctx->acquired > 0) {
525 ret = __mutex_lock_check_stamp(lock, ww_ctx);
526 if (ret)
527 goto err;
528 }
529
530 __set_task_state(task, state);
531
532 /* didn't get the lock, go to sleep: */
533 spin_unlock_mutex(&lock->wait_lock, flags);
534 schedule_preempt_disabled();
535 spin_lock_mutex(&lock->wait_lock, flags);
536 }
537 mutex_remove_waiter(lock, &waiter, current_thread_info());
538 /* set it to 0 if there are no waiters left: */
539 if (likely(list_empty(&lock->wait_list)))
540 atomic_set(&lock->count, 0);
541 debug_mutex_free_waiter(&waiter);
542
543skip_wait:
544 /* got the lock - cleanup and rejoice! */
545 lock_acquired(&lock->dep_map, ip);
546 mutex_set_owner(lock);
547
548 if (use_ww_ctx) {
549 struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
550 struct mutex_waiter *cur;
551
552 /*
553 * This branch gets optimized out for the common case,
554 * and is only important for ww_mutex_lock.
555 */
556 ww_mutex_lock_acquired(ww, ww_ctx);
557 ww->ctx = ww_ctx;
558
559 /*
560 * Give any possible sleeping processes the chance to wake up,
561 * so they can recheck if they have to back off.
562 */
563 list_for_each_entry(cur, &lock->wait_list, list) {
564 debug_mutex_wake_waiter(lock, cur);
565 wake_up_process(cur->task);
566 }
567 }
568
569 spin_unlock_mutex(&lock->wait_lock, flags);
570 preempt_enable();
571 return 0;
572
573err:
574 mutex_remove_waiter(lock, &waiter, task_thread_info(task));
575 spin_unlock_mutex(&lock->wait_lock, flags);
576 debug_mutex_free_waiter(&waiter);
577 mutex_release(&lock->dep_map, 1, ip);
578 preempt_enable();
579 return ret;
580}
581
582#ifdef CONFIG_DEBUG_LOCK_ALLOC
583void __sched
584mutex_lock_nested(struct mutex *lock, unsigned int subclass)
585{
586 might_sleep();
587 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
588 subclass, NULL, _RET_IP_, NULL, 0);
589}
590
591EXPORT_SYMBOL_GPL(mutex_lock_nested);
592
593void __sched
594_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
595{
596 might_sleep();
597 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
598 0, nest, _RET_IP_, NULL, 0);
599}
600
601EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
602
603int __sched
604mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
605{
606 might_sleep();
607 return __mutex_lock_common(lock, TASK_KILLABLE,
608 subclass, NULL, _RET_IP_, NULL, 0);
609}
610EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
611
612int __sched
613mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
614{
615 might_sleep();
616 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE,
617 subclass, NULL, _RET_IP_, NULL, 0);
618}
619
620EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
621
622static inline int
623ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
624{
625#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
626 unsigned tmp;
627
628 if (ctx->deadlock_inject_countdown-- == 0) {
629 tmp = ctx->deadlock_inject_interval;
630 if (tmp > UINT_MAX/4)
631 tmp = UINT_MAX;
632 else
633 tmp = tmp*2 + tmp + tmp/2;
634
635 ctx->deadlock_inject_interval = tmp;
636 ctx->deadlock_inject_countdown = tmp;
637 ctx->contending_lock = lock;
638
639 ww_mutex_unlock(lock);
640
641 return -EDEADLK;
642 }
643#endif
644
645 return 0;
646}
647
648int __sched
649__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
650{
651 int ret;
652
653 might_sleep();
654 ret = __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE,
655 0, &ctx->dep_map, _RET_IP_, ctx, 1);
656 if (!ret && ctx->acquired > 1)
657 return ww_mutex_deadlock_injection(lock, ctx);
658
659 return ret;
660}
661EXPORT_SYMBOL_GPL(__ww_mutex_lock);
662
663int __sched
664__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
665{
666 int ret;
667
668 might_sleep();
669 ret = __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE,
670 0, &ctx->dep_map, _RET_IP_, ctx, 1);
671
672 if (!ret && ctx->acquired > 1)
673 return ww_mutex_deadlock_injection(lock, ctx);
674
675 return ret;
676}
677EXPORT_SYMBOL_GPL(__ww_mutex_lock_interruptible);
678
679#endif
680
681/*
682 * Release the lock, slowpath:
683 */
684static inline void
685__mutex_unlock_common_slowpath(atomic_t *lock_count, int nested)
686{
687 struct mutex *lock = container_of(lock_count, struct mutex, count);
688 unsigned long flags;
689
690 /*
691 * some architectures leave the lock unlocked in the fastpath failure
692 * case, others need to leave it locked. In the later case we have to
693 * unlock it here
694 */
695 if (__mutex_slowpath_needs_to_unlock())
696 atomic_set(&lock->count, 1);
697
698 spin_lock_mutex(&lock->wait_lock, flags);
699 mutex_release(&lock->dep_map, nested, _RET_IP_);
700 debug_mutex_unlock(lock);
701
702 if (!list_empty(&lock->wait_list)) {
703 /* get the first entry from the wait-list: */
704 struct mutex_waiter *waiter =
705 list_entry(lock->wait_list.next,
706 struct mutex_waiter, list);
707
708 debug_mutex_wake_waiter(lock, waiter);
709
710 wake_up_process(waiter->task);
711 }
712
713 spin_unlock_mutex(&lock->wait_lock, flags);
714}
715
716/*
717 * Release the lock, slowpath:
718 */
719__visible void
720__mutex_unlock_slowpath(atomic_t *lock_count)
721{
722 __mutex_unlock_common_slowpath(lock_count, 1);
723}
724
725#ifndef CONFIG_DEBUG_LOCK_ALLOC
726/*
727 * Here come the less common (and hence less performance-critical) APIs:
728 * mutex_lock_interruptible() and mutex_trylock().
729 */
730static noinline int __sched
731__mutex_lock_killable_slowpath(struct mutex *lock);
732
733static noinline int __sched
734__mutex_lock_interruptible_slowpath(struct mutex *lock);
735
736/**
737 * mutex_lock_interruptible - acquire the mutex, interruptible
738 * @lock: the mutex to be acquired
739 *
740 * Lock the mutex like mutex_lock(), and return 0 if the mutex has
741 * been acquired or sleep until the mutex becomes available. If a
742 * signal arrives while waiting for the lock then this function
743 * returns -EINTR.
744 *
745 * This function is similar to (but not equivalent to) down_interruptible().
746 */
747int __sched mutex_lock_interruptible(struct mutex *lock)
748{
749 int ret;
750
751 might_sleep();
752 ret = __mutex_fastpath_lock_retval(&lock->count);
753 if (likely(!ret)) {
754 mutex_set_owner(lock);
755 return 0;
756 } else
757 return __mutex_lock_interruptible_slowpath(lock);
758}
759
760EXPORT_SYMBOL(mutex_lock_interruptible);
761
762int __sched mutex_lock_killable(struct mutex *lock)
763{
764 int ret;
765
766 might_sleep();
767 ret = __mutex_fastpath_lock_retval(&lock->count);
768 if (likely(!ret)) {
769 mutex_set_owner(lock);
770 return 0;
771 } else
772 return __mutex_lock_killable_slowpath(lock);
773}
774EXPORT_SYMBOL(mutex_lock_killable);
775
776__visible void __sched
777__mutex_lock_slowpath(atomic_t *lock_count)
778{
779 struct mutex *lock = container_of(lock_count, struct mutex, count);
780
781 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0,
782 NULL, _RET_IP_, NULL, 0);
783}
784
785static noinline int __sched
786__mutex_lock_killable_slowpath(struct mutex *lock)
787{
788 return __mutex_lock_common(lock, TASK_KILLABLE, 0,
789 NULL, _RET_IP_, NULL, 0);
790}
791
792static noinline int __sched
793__mutex_lock_interruptible_slowpath(struct mutex *lock)
794{
795 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0,
796 NULL, _RET_IP_, NULL, 0);
797}
798
799static noinline int __sched
800__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
801{
802 return __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE, 0,
803 NULL, _RET_IP_, ctx, 1);
804}
805
806static noinline int __sched
807__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
808 struct ww_acquire_ctx *ctx)
809{
810 return __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE, 0,
811 NULL, _RET_IP_, ctx, 1);
812}
813
814#endif
815
816/*
817 * Spinlock based trylock, we take the spinlock and check whether we
818 * can get the lock:
819 */
820static inline int __mutex_trylock_slowpath(atomic_t *lock_count)
821{
822 struct mutex *lock = container_of(lock_count, struct mutex, count);
823 unsigned long flags;
824 int prev;
825
826 spin_lock_mutex(&lock->wait_lock, flags);
827
828 prev = atomic_xchg(&lock->count, -1);
829 if (likely(prev == 1)) {
830 mutex_set_owner(lock);
831 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
832 }
833
834 /* Set it back to 0 if there are no waiters: */
835 if (likely(list_empty(&lock->wait_list)))
836 atomic_set(&lock->count, 0);
837
838 spin_unlock_mutex(&lock->wait_lock, flags);
839
840 return prev == 1;
841}
842
843/**
844 * mutex_trylock - try to acquire the mutex, without waiting
845 * @lock: the mutex to be acquired
846 *
847 * Try to acquire the mutex atomically. Returns 1 if the mutex
848 * has been acquired successfully, and 0 on contention.
849 *
850 * NOTE: this function follows the spin_trylock() convention, so
851 * it is negated from the down_trylock() return values! Be careful
852 * about this when converting semaphore users to mutexes.
853 *
854 * This function must not be used in interrupt context. The
855 * mutex must be released by the same task that acquired it.
856 */
857int __sched mutex_trylock(struct mutex *lock)
858{
859 int ret;
860
861 ret = __mutex_fastpath_trylock(&lock->count, __mutex_trylock_slowpath);
862 if (ret)
863 mutex_set_owner(lock);
864
865 return ret;
866}
867EXPORT_SYMBOL(mutex_trylock);
868
869#ifndef CONFIG_DEBUG_LOCK_ALLOC
870int __sched
871__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
872{
873 int ret;
874
875 might_sleep();
876
877 ret = __mutex_fastpath_lock_retval(&lock->base.count);
878
879 if (likely(!ret)) {
880 ww_mutex_set_context_fastpath(lock, ctx);
881 mutex_set_owner(&lock->base);
882 } else
883 ret = __ww_mutex_lock_slowpath(lock, ctx);
884 return ret;
885}
886EXPORT_SYMBOL(__ww_mutex_lock);
887
888int __sched
889__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
890{
891 int ret;
892
893 might_sleep();
894
895 ret = __mutex_fastpath_lock_retval(&lock->base.count);
896
897 if (likely(!ret)) {
898 ww_mutex_set_context_fastpath(lock, ctx);
899 mutex_set_owner(&lock->base);
900 } else
901 ret = __ww_mutex_lock_interruptible_slowpath(lock, ctx);
902 return ret;
903}
904EXPORT_SYMBOL(__ww_mutex_lock_interruptible);
905
906#endif
907
908/**
909 * atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
910 * @cnt: the atomic which we are to dec
911 * @lock: the mutex to return holding if we dec to 0
912 *
913 * return true and hold lock if we dec to 0, return false otherwise
914 */
915int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
916{
917 /* dec if we can't possibly hit 0 */
918 if (atomic_add_unless(cnt, -1, 1))
919 return 0;
920 /* we might hit 0, so take the lock */
921 mutex_lock(lock);
922 if (!atomic_dec_and_test(cnt)) {
923 /* when we actually did the dec, we didn't hit 0 */
924 mutex_unlock(lock);
925 return 0;
926 }
927 /* we hit 0, and we hold the lock */
928 return 1;
929}
930EXPORT_SYMBOL(atomic_dec_and_mutex_lock);
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.h>
23#include <linux/sched/rt.h>
24#include <linux/export.h>
25#include <linux/spinlock.h>
26#include <linux/interrupt.h>
27#include <linux/debug_locks.h>
28#include <linux/osq_lock.h>
29
30#ifdef CONFIG_DEBUG_MUTEXES
31# include "mutex-debug.h"
32#else
33# include "mutex.h"
34#endif
35
36void
37__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
38{
39 atomic_long_set(&lock->owner, 0);
40 spin_lock_init(&lock->wait_lock);
41 INIT_LIST_HEAD(&lock->wait_list);
42#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
43 osq_lock_init(&lock->osq);
44#endif
45
46 debug_mutex_init(lock, name, key);
47}
48EXPORT_SYMBOL(__mutex_init);
49
50/*
51 * @owner: contains: 'struct task_struct *' to the current lock owner,
52 * NULL means not owned. Since task_struct pointers are aligned at
53 * ARCH_MIN_TASKALIGN (which is at least sizeof(void *)), we have low
54 * bits to store extra state.
55 *
56 * Bit0 indicates a non-empty waiter list; unlock must issue a wakeup.
57 * Bit1 indicates unlock needs to hand the lock to the top-waiter
58 */
59#define MUTEX_FLAG_WAITERS 0x01
60#define MUTEX_FLAG_HANDOFF 0x02
61
62#define MUTEX_FLAGS 0x03
63
64static inline struct task_struct *__owner_task(unsigned long owner)
65{
66 return (struct task_struct *)(owner & ~MUTEX_FLAGS);
67}
68
69static inline unsigned long __owner_flags(unsigned long owner)
70{
71 return owner & MUTEX_FLAGS;
72}
73
74/*
75 * Actual trylock that will work on any unlocked state.
76 *
77 * When setting the owner field, we must preserve the low flag bits.
78 *
79 * Be careful with @handoff, only set that in a wait-loop (where you set
80 * HANDOFF) to avoid recursive lock attempts.
81 */
82static inline bool __mutex_trylock(struct mutex *lock, const bool handoff)
83{
84 unsigned long owner, curr = (unsigned long)current;
85
86 owner = atomic_long_read(&lock->owner);
87 for (;;) { /* must loop, can race against a flag */
88 unsigned long old, flags = __owner_flags(owner);
89
90 if (__owner_task(owner)) {
91 if (handoff && unlikely(__owner_task(owner) == current)) {
92 /*
93 * Provide ACQUIRE semantics for the lock-handoff.
94 *
95 * We cannot easily use load-acquire here, since
96 * the actual load is a failed cmpxchg, which
97 * doesn't imply any barriers.
98 *
99 * Also, this is a fairly unlikely scenario, and
100 * this contains the cost.
101 */
102 smp_mb(); /* ACQUIRE */
103 return true;
104 }
105
106 return false;
107 }
108
109 /*
110 * We set the HANDOFF bit, we must make sure it doesn't live
111 * past the point where we acquire it. This would be possible
112 * if we (accidentally) set the bit on an unlocked mutex.
113 */
114 if (handoff)
115 flags &= ~MUTEX_FLAG_HANDOFF;
116
117 old = atomic_long_cmpxchg_acquire(&lock->owner, owner, curr | flags);
118 if (old == owner)
119 return true;
120
121 owner = old;
122 }
123}
124
125#ifndef CONFIG_DEBUG_LOCK_ALLOC
126/*
127 * Lockdep annotations are contained to the slow paths for simplicity.
128 * There is nothing that would stop spreading the lockdep annotations outwards
129 * except more code.
130 */
131
132/*
133 * Optimistic trylock that only works in the uncontended case. Make sure to
134 * follow with a __mutex_trylock() before failing.
135 */
136static __always_inline bool __mutex_trylock_fast(struct mutex *lock)
137{
138 unsigned long curr = (unsigned long)current;
139
140 if (!atomic_long_cmpxchg_acquire(&lock->owner, 0UL, curr))
141 return true;
142
143 return false;
144}
145
146static __always_inline bool __mutex_unlock_fast(struct mutex *lock)
147{
148 unsigned long curr = (unsigned long)current;
149
150 if (atomic_long_cmpxchg_release(&lock->owner, curr, 0UL) == curr)
151 return true;
152
153 return false;
154}
155#endif
156
157static inline void __mutex_set_flag(struct mutex *lock, unsigned long flag)
158{
159 atomic_long_or(flag, &lock->owner);
160}
161
162static inline void __mutex_clear_flag(struct mutex *lock, unsigned long flag)
163{
164 atomic_long_andnot(flag, &lock->owner);
165}
166
167static inline bool __mutex_waiter_is_first(struct mutex *lock, struct mutex_waiter *waiter)
168{
169 return list_first_entry(&lock->wait_list, struct mutex_waiter, list) == waiter;
170}
171
172/*
173 * Give up ownership to a specific task, when @task = NULL, this is equivalent
174 * to a regular unlock. Clears HANDOFF, preserves WAITERS. Provides RELEASE
175 * semantics like a regular unlock, the __mutex_trylock() provides matching
176 * ACQUIRE semantics for the handoff.
177 */
178static void __mutex_handoff(struct mutex *lock, struct task_struct *task)
179{
180 unsigned long owner = atomic_long_read(&lock->owner);
181
182 for (;;) {
183 unsigned long old, new;
184
185#ifdef CONFIG_DEBUG_MUTEXES
186 DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
187#endif
188
189 new = (owner & MUTEX_FLAG_WAITERS);
190 new |= (unsigned long)task;
191
192 old = atomic_long_cmpxchg_release(&lock->owner, owner, new);
193 if (old == owner)
194 break;
195
196 owner = old;
197 }
198}
199
200#ifndef CONFIG_DEBUG_LOCK_ALLOC
201/*
202 * We split the mutex lock/unlock logic into separate fastpath and
203 * slowpath functions, to reduce the register pressure on the fastpath.
204 * We also put the fastpath first in the kernel image, to make sure the
205 * branch is predicted by the CPU as default-untaken.
206 */
207static void __sched __mutex_lock_slowpath(struct mutex *lock);
208
209/**
210 * mutex_lock - acquire the mutex
211 * @lock: the mutex to be acquired
212 *
213 * Lock the mutex exclusively for this task. If the mutex is not
214 * available right now, it will sleep until it can get it.
215 *
216 * The mutex must later on be released by the same task that
217 * acquired it. Recursive locking is not allowed. The task
218 * may not exit without first unlocking the mutex. Also, kernel
219 * memory where the mutex resides must not be freed with
220 * the mutex still locked. The mutex must first be initialized
221 * (or statically defined) before it can be locked. memset()-ing
222 * the mutex to 0 is not allowed.
223 *
224 * ( The CONFIG_DEBUG_MUTEXES .config option turns on debugging
225 * checks that will enforce the restrictions and will also do
226 * deadlock debugging. )
227 *
228 * This function is similar to (but not equivalent to) down().
229 */
230void __sched mutex_lock(struct mutex *lock)
231{
232 might_sleep();
233
234 if (!__mutex_trylock_fast(lock))
235 __mutex_lock_slowpath(lock);
236}
237EXPORT_SYMBOL(mutex_lock);
238#endif
239
240static __always_inline void ww_mutex_lock_acquired(struct ww_mutex *ww,
241 struct ww_acquire_ctx *ww_ctx)
242{
243#ifdef CONFIG_DEBUG_MUTEXES
244 /*
245 * If this WARN_ON triggers, you used ww_mutex_lock to acquire,
246 * but released with a normal mutex_unlock in this call.
247 *
248 * This should never happen, always use ww_mutex_unlock.
249 */
250 DEBUG_LOCKS_WARN_ON(ww->ctx);
251
252 /*
253 * Not quite done after calling ww_acquire_done() ?
254 */
255 DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
256
257 if (ww_ctx->contending_lock) {
258 /*
259 * After -EDEADLK you tried to
260 * acquire a different ww_mutex? Bad!
261 */
262 DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
263
264 /*
265 * You called ww_mutex_lock after receiving -EDEADLK,
266 * but 'forgot' to unlock everything else first?
267 */
268 DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
269 ww_ctx->contending_lock = NULL;
270 }
271
272 /*
273 * Naughty, using a different class will lead to undefined behavior!
274 */
275 DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
276#endif
277 ww_ctx->acquired++;
278}
279
280/*
281 * After acquiring lock with fastpath or when we lost out in contested
282 * slowpath, set ctx and wake up any waiters so they can recheck.
283 */
284static __always_inline void
285ww_mutex_set_context_fastpath(struct ww_mutex *lock,
286 struct ww_acquire_ctx *ctx)
287{
288 unsigned long flags;
289 struct mutex_waiter *cur;
290
291 ww_mutex_lock_acquired(lock, ctx);
292
293 lock->ctx = ctx;
294
295 /*
296 * The lock->ctx update should be visible on all cores before
297 * the atomic read is done, otherwise contended waiters might be
298 * missed. The contended waiters will either see ww_ctx == NULL
299 * and keep spinning, or it will acquire wait_lock, add itself
300 * to waiter list and sleep.
301 */
302 smp_mb(); /* ^^^ */
303
304 /*
305 * Check if lock is contended, if not there is nobody to wake up
306 */
307 if (likely(!(atomic_long_read(&lock->base.owner) & MUTEX_FLAG_WAITERS)))
308 return;
309
310 /*
311 * Uh oh, we raced in fastpath, wake up everyone in this case,
312 * so they can see the new lock->ctx.
313 */
314 spin_lock_mutex(&lock->base.wait_lock, flags);
315 list_for_each_entry(cur, &lock->base.wait_list, list) {
316 debug_mutex_wake_waiter(&lock->base, cur);
317 wake_up_process(cur->task);
318 }
319 spin_unlock_mutex(&lock->base.wait_lock, flags);
320}
321
322/*
323 * After acquiring lock in the slowpath set ctx and wake up any
324 * waiters so they can recheck.
325 *
326 * Callers must hold the mutex wait_lock.
327 */
328static __always_inline void
329ww_mutex_set_context_slowpath(struct ww_mutex *lock,
330 struct ww_acquire_ctx *ctx)
331{
332 struct mutex_waiter *cur;
333
334 ww_mutex_lock_acquired(lock, ctx);
335 lock->ctx = ctx;
336
337 /*
338 * Give any possible sleeping processes the chance to wake up,
339 * so they can recheck if they have to back off.
340 */
341 list_for_each_entry(cur, &lock->base.wait_list, list) {
342 debug_mutex_wake_waiter(&lock->base, cur);
343 wake_up_process(cur->task);
344 }
345}
346
347#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
348/*
349 * Look out! "owner" is an entirely speculative pointer
350 * access and not reliable.
351 */
352static noinline
353bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
354{
355 bool ret = true;
356
357 rcu_read_lock();
358 while (__mutex_owner(lock) == owner) {
359 /*
360 * Ensure we emit the owner->on_cpu, dereference _after_
361 * checking lock->owner still matches owner. If that fails,
362 * owner might point to freed memory. If it still matches,
363 * the rcu_read_lock() ensures the memory stays valid.
364 */
365 barrier();
366
367 /*
368 * Use vcpu_is_preempted to detect lock holder preemption issue.
369 */
370 if (!owner->on_cpu || need_resched() ||
371 vcpu_is_preempted(task_cpu(owner))) {
372 ret = false;
373 break;
374 }
375
376 cpu_relax();
377 }
378 rcu_read_unlock();
379
380 return ret;
381}
382
383/*
384 * Initial check for entering the mutex spinning loop
385 */
386static inline int mutex_can_spin_on_owner(struct mutex *lock)
387{
388 struct task_struct *owner;
389 int retval = 1;
390
391 if (need_resched())
392 return 0;
393
394 rcu_read_lock();
395 owner = __mutex_owner(lock);
396
397 /*
398 * As lock holder preemption issue, we both skip spinning if task is not
399 * on cpu or its cpu is preempted
400 */
401 if (owner)
402 retval = owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
403 rcu_read_unlock();
404
405 /*
406 * If lock->owner is not set, the mutex has been released. Return true
407 * such that we'll trylock in the spin path, which is a faster option
408 * than the blocking slow path.
409 */
410 return retval;
411}
412
413/*
414 * Optimistic spinning.
415 *
416 * We try to spin for acquisition when we find that the lock owner
417 * is currently running on a (different) CPU and while we don't
418 * need to reschedule. The rationale is that if the lock owner is
419 * running, it is likely to release the lock soon.
420 *
421 * The mutex spinners are queued up using MCS lock so that only one
422 * spinner can compete for the mutex. However, if mutex spinning isn't
423 * going to happen, there is no point in going through the lock/unlock
424 * overhead.
425 *
426 * Returns true when the lock was taken, otherwise false, indicating
427 * that we need to jump to the slowpath and sleep.
428 *
429 * The waiter flag is set to true if the spinner is a waiter in the wait
430 * queue. The waiter-spinner will spin on the lock directly and concurrently
431 * with the spinner at the head of the OSQ, if present, until the owner is
432 * changed to itself.
433 */
434static bool mutex_optimistic_spin(struct mutex *lock,
435 struct ww_acquire_ctx *ww_ctx,
436 const bool use_ww_ctx, const bool waiter)
437{
438 struct task_struct *task = current;
439
440 if (!waiter) {
441 /*
442 * The purpose of the mutex_can_spin_on_owner() function is
443 * to eliminate the overhead of osq_lock() and osq_unlock()
444 * in case spinning isn't possible. As a waiter-spinner
445 * is not going to take OSQ lock anyway, there is no need
446 * to call mutex_can_spin_on_owner().
447 */
448 if (!mutex_can_spin_on_owner(lock))
449 goto fail;
450
451 /*
452 * In order to avoid a stampede of mutex spinners trying to
453 * acquire the mutex all at once, the spinners need to take a
454 * MCS (queued) lock first before spinning on the owner field.
455 */
456 if (!osq_lock(&lock->osq))
457 goto fail;
458 }
459
460 for (;;) {
461 struct task_struct *owner;
462
463 if (use_ww_ctx && ww_ctx->acquired > 0) {
464 struct ww_mutex *ww;
465
466 ww = container_of(lock, struct ww_mutex, base);
467 /*
468 * If ww->ctx is set the contents are undefined, only
469 * by acquiring wait_lock there is a guarantee that
470 * they are not invalid when reading.
471 *
472 * As such, when deadlock detection needs to be
473 * performed the optimistic spinning cannot be done.
474 */
475 if (READ_ONCE(ww->ctx))
476 goto fail_unlock;
477 }
478
479 /*
480 * If there's an owner, wait for it to either
481 * release the lock or go to sleep.
482 */
483 owner = __mutex_owner(lock);
484 if (owner) {
485 if (waiter && owner == task) {
486 smp_mb(); /* ACQUIRE */
487 break;
488 }
489
490 if (!mutex_spin_on_owner(lock, owner))
491 goto fail_unlock;
492 }
493
494 /* Try to acquire the mutex if it is unlocked. */
495 if (__mutex_trylock(lock, waiter))
496 break;
497
498 /*
499 * The cpu_relax() call is a compiler barrier which forces
500 * everything in this loop to be re-loaded. We don't need
501 * memory barriers as we'll eventually observe the right
502 * values at the cost of a few extra spins.
503 */
504 cpu_relax();
505 }
506
507 if (!waiter)
508 osq_unlock(&lock->osq);
509
510 return true;
511
512
513fail_unlock:
514 if (!waiter)
515 osq_unlock(&lock->osq);
516
517fail:
518 /*
519 * If we fell out of the spin path because of need_resched(),
520 * reschedule now, before we try-lock the mutex. This avoids getting
521 * scheduled out right after we obtained the mutex.
522 */
523 if (need_resched()) {
524 /*
525 * We _should_ have TASK_RUNNING here, but just in case
526 * we do not, make it so, otherwise we might get stuck.
527 */
528 __set_current_state(TASK_RUNNING);
529 schedule_preempt_disabled();
530 }
531
532 return false;
533}
534#else
535static bool mutex_optimistic_spin(struct mutex *lock,
536 struct ww_acquire_ctx *ww_ctx,
537 const bool use_ww_ctx, const bool waiter)
538{
539 return false;
540}
541#endif
542
543static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip);
544
545/**
546 * mutex_unlock - release the mutex
547 * @lock: the mutex to be released
548 *
549 * Unlock a mutex that has been locked by this task previously.
550 *
551 * This function must not be used in interrupt context. Unlocking
552 * of a not locked mutex is not allowed.
553 *
554 * This function is similar to (but not equivalent to) up().
555 */
556void __sched mutex_unlock(struct mutex *lock)
557{
558#ifndef CONFIG_DEBUG_LOCK_ALLOC
559 if (__mutex_unlock_fast(lock))
560 return;
561#endif
562 __mutex_unlock_slowpath(lock, _RET_IP_);
563}
564EXPORT_SYMBOL(mutex_unlock);
565
566/**
567 * ww_mutex_unlock - release the w/w mutex
568 * @lock: the mutex to be released
569 *
570 * Unlock a mutex that has been locked by this task previously with any of the
571 * ww_mutex_lock* functions (with or without an acquire context). It is
572 * forbidden to release the locks after releasing the acquire context.
573 *
574 * This function must not be used in interrupt context. Unlocking
575 * of a unlocked mutex is not allowed.
576 */
577void __sched ww_mutex_unlock(struct ww_mutex *lock)
578{
579 /*
580 * The unlocking fastpath is the 0->1 transition from 'locked'
581 * into 'unlocked' state:
582 */
583 if (lock->ctx) {
584#ifdef CONFIG_DEBUG_MUTEXES
585 DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
586#endif
587 if (lock->ctx->acquired > 0)
588 lock->ctx->acquired--;
589 lock->ctx = NULL;
590 }
591
592 mutex_unlock(&lock->base);
593}
594EXPORT_SYMBOL(ww_mutex_unlock);
595
596static inline int __sched
597__ww_mutex_lock_check_stamp(struct mutex *lock, struct ww_acquire_ctx *ctx)
598{
599 struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
600 struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
601
602 if (!hold_ctx)
603 return 0;
604
605 if (ctx->stamp - hold_ctx->stamp <= LONG_MAX &&
606 (ctx->stamp != hold_ctx->stamp || ctx > hold_ctx)) {
607#ifdef CONFIG_DEBUG_MUTEXES
608 DEBUG_LOCKS_WARN_ON(ctx->contending_lock);
609 ctx->contending_lock = ww;
610#endif
611 return -EDEADLK;
612 }
613
614 return 0;
615}
616
617/*
618 * Lock a mutex (possibly interruptible), slowpath:
619 */
620static __always_inline int __sched
621__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
622 struct lockdep_map *nest_lock, unsigned long ip,
623 struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
624{
625 struct task_struct *task = current;
626 struct mutex_waiter waiter;
627 unsigned long flags;
628 bool first = false;
629 struct ww_mutex *ww;
630 int ret;
631
632 if (use_ww_ctx) {
633 ww = container_of(lock, struct ww_mutex, base);
634 if (unlikely(ww_ctx == READ_ONCE(ww->ctx)))
635 return -EALREADY;
636 }
637
638 preempt_disable();
639 mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
640
641 if (__mutex_trylock(lock, false) ||
642 mutex_optimistic_spin(lock, ww_ctx, use_ww_ctx, false)) {
643 /* got the lock, yay! */
644 lock_acquired(&lock->dep_map, ip);
645 if (use_ww_ctx)
646 ww_mutex_set_context_fastpath(ww, ww_ctx);
647 preempt_enable();
648 return 0;
649 }
650
651 spin_lock_mutex(&lock->wait_lock, flags);
652 /*
653 * After waiting to acquire the wait_lock, try again.
654 */
655 if (__mutex_trylock(lock, false))
656 goto skip_wait;
657
658 debug_mutex_lock_common(lock, &waiter);
659 debug_mutex_add_waiter(lock, &waiter, task);
660
661 /* add waiting tasks to the end of the waitqueue (FIFO): */
662 list_add_tail(&waiter.list, &lock->wait_list);
663 waiter.task = task;
664
665 if (__mutex_waiter_is_first(lock, &waiter))
666 __mutex_set_flag(lock, MUTEX_FLAG_WAITERS);
667
668 lock_contended(&lock->dep_map, ip);
669
670 set_task_state(task, state);
671 for (;;) {
672 /*
673 * Once we hold wait_lock, we're serialized against
674 * mutex_unlock() handing the lock off to us, do a trylock
675 * before testing the error conditions to make sure we pick up
676 * the handoff.
677 */
678 if (__mutex_trylock(lock, first))
679 goto acquired;
680
681 /*
682 * Check for signals and wound conditions while holding
683 * wait_lock. This ensures the lock cancellation is ordered
684 * against mutex_unlock() and wake-ups do not go missing.
685 */
686 if (unlikely(signal_pending_state(state, task))) {
687 ret = -EINTR;
688 goto err;
689 }
690
691 if (use_ww_ctx && ww_ctx->acquired > 0) {
692 ret = __ww_mutex_lock_check_stamp(lock, ww_ctx);
693 if (ret)
694 goto err;
695 }
696
697 spin_unlock_mutex(&lock->wait_lock, flags);
698 schedule_preempt_disabled();
699
700 if (!first && __mutex_waiter_is_first(lock, &waiter)) {
701 first = true;
702 __mutex_set_flag(lock, MUTEX_FLAG_HANDOFF);
703 }
704
705 set_task_state(task, state);
706 /*
707 * Here we order against unlock; we must either see it change
708 * state back to RUNNING and fall through the next schedule(),
709 * or we must see its unlock and acquire.
710 */
711 if ((first && mutex_optimistic_spin(lock, ww_ctx, use_ww_ctx, true)) ||
712 __mutex_trylock(lock, first))
713 break;
714
715 spin_lock_mutex(&lock->wait_lock, flags);
716 }
717 spin_lock_mutex(&lock->wait_lock, flags);
718acquired:
719 __set_task_state(task, TASK_RUNNING);
720
721 mutex_remove_waiter(lock, &waiter, task);
722 if (likely(list_empty(&lock->wait_list)))
723 __mutex_clear_flag(lock, MUTEX_FLAGS);
724
725 debug_mutex_free_waiter(&waiter);
726
727skip_wait:
728 /* got the lock - cleanup and rejoice! */
729 lock_acquired(&lock->dep_map, ip);
730
731 if (use_ww_ctx)
732 ww_mutex_set_context_slowpath(ww, ww_ctx);
733
734 spin_unlock_mutex(&lock->wait_lock, flags);
735 preempt_enable();
736 return 0;
737
738err:
739 __set_task_state(task, TASK_RUNNING);
740 mutex_remove_waiter(lock, &waiter, task);
741 spin_unlock_mutex(&lock->wait_lock, flags);
742 debug_mutex_free_waiter(&waiter);
743 mutex_release(&lock->dep_map, 1, ip);
744 preempt_enable();
745 return ret;
746}
747
748#ifdef CONFIG_DEBUG_LOCK_ALLOC
749void __sched
750mutex_lock_nested(struct mutex *lock, unsigned int subclass)
751{
752 might_sleep();
753 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
754 subclass, NULL, _RET_IP_, NULL, 0);
755}
756
757EXPORT_SYMBOL_GPL(mutex_lock_nested);
758
759void __sched
760_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
761{
762 might_sleep();
763 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
764 0, nest, _RET_IP_, NULL, 0);
765}
766EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
767
768int __sched
769mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
770{
771 might_sleep();
772 return __mutex_lock_common(lock, TASK_KILLABLE,
773 subclass, NULL, _RET_IP_, NULL, 0);
774}
775EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
776
777int __sched
778mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
779{
780 might_sleep();
781 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE,
782 subclass, NULL, _RET_IP_, NULL, 0);
783}
784EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
785
786static inline int
787ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
788{
789#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
790 unsigned tmp;
791
792 if (ctx->deadlock_inject_countdown-- == 0) {
793 tmp = ctx->deadlock_inject_interval;
794 if (tmp > UINT_MAX/4)
795 tmp = UINT_MAX;
796 else
797 tmp = tmp*2 + tmp + tmp/2;
798
799 ctx->deadlock_inject_interval = tmp;
800 ctx->deadlock_inject_countdown = tmp;
801 ctx->contending_lock = lock;
802
803 ww_mutex_unlock(lock);
804
805 return -EDEADLK;
806 }
807#endif
808
809 return 0;
810}
811
812int __sched
813__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
814{
815 int ret;
816
817 might_sleep();
818 ret = __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE,
819 0, &ctx->dep_map, _RET_IP_, ctx, 1);
820 if (!ret && ctx->acquired > 1)
821 return ww_mutex_deadlock_injection(lock, ctx);
822
823 return ret;
824}
825EXPORT_SYMBOL_GPL(__ww_mutex_lock);
826
827int __sched
828__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
829{
830 int ret;
831
832 might_sleep();
833 ret = __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE,
834 0, &ctx->dep_map, _RET_IP_, ctx, 1);
835
836 if (!ret && ctx->acquired > 1)
837 return ww_mutex_deadlock_injection(lock, ctx);
838
839 return ret;
840}
841EXPORT_SYMBOL_GPL(__ww_mutex_lock_interruptible);
842
843#endif
844
845/*
846 * Release the lock, slowpath:
847 */
848static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip)
849{
850 struct task_struct *next = NULL;
851 unsigned long owner, flags;
852 DEFINE_WAKE_Q(wake_q);
853
854 mutex_release(&lock->dep_map, 1, ip);
855
856 /*
857 * Release the lock before (potentially) taking the spinlock such that
858 * other contenders can get on with things ASAP.
859 *
860 * Except when HANDOFF, in that case we must not clear the owner field,
861 * but instead set it to the top waiter.
862 */
863 owner = atomic_long_read(&lock->owner);
864 for (;;) {
865 unsigned long old;
866
867#ifdef CONFIG_DEBUG_MUTEXES
868 DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
869#endif
870
871 if (owner & MUTEX_FLAG_HANDOFF)
872 break;
873
874 old = atomic_long_cmpxchg_release(&lock->owner, owner,
875 __owner_flags(owner));
876 if (old == owner) {
877 if (owner & MUTEX_FLAG_WAITERS)
878 break;
879
880 return;
881 }
882
883 owner = old;
884 }
885
886 spin_lock_mutex(&lock->wait_lock, flags);
887 debug_mutex_unlock(lock);
888 if (!list_empty(&lock->wait_list)) {
889 /* get the first entry from the wait-list: */
890 struct mutex_waiter *waiter =
891 list_first_entry(&lock->wait_list,
892 struct mutex_waiter, list);
893
894 next = waiter->task;
895
896 debug_mutex_wake_waiter(lock, waiter);
897 wake_q_add(&wake_q, next);
898 }
899
900 if (owner & MUTEX_FLAG_HANDOFF)
901 __mutex_handoff(lock, next);
902
903 spin_unlock_mutex(&lock->wait_lock, flags);
904
905 wake_up_q(&wake_q);
906}
907
908#ifndef CONFIG_DEBUG_LOCK_ALLOC
909/*
910 * Here come the less common (and hence less performance-critical) APIs:
911 * mutex_lock_interruptible() and mutex_trylock().
912 */
913static noinline int __sched
914__mutex_lock_killable_slowpath(struct mutex *lock);
915
916static noinline int __sched
917__mutex_lock_interruptible_slowpath(struct mutex *lock);
918
919/**
920 * mutex_lock_interruptible - acquire the mutex, interruptible
921 * @lock: the mutex to be acquired
922 *
923 * Lock the mutex like mutex_lock(), and return 0 if the mutex has
924 * been acquired or sleep until the mutex becomes available. If a
925 * signal arrives while waiting for the lock then this function
926 * returns -EINTR.
927 *
928 * This function is similar to (but not equivalent to) down_interruptible().
929 */
930int __sched mutex_lock_interruptible(struct mutex *lock)
931{
932 might_sleep();
933
934 if (__mutex_trylock_fast(lock))
935 return 0;
936
937 return __mutex_lock_interruptible_slowpath(lock);
938}
939
940EXPORT_SYMBOL(mutex_lock_interruptible);
941
942int __sched mutex_lock_killable(struct mutex *lock)
943{
944 might_sleep();
945
946 if (__mutex_trylock_fast(lock))
947 return 0;
948
949 return __mutex_lock_killable_slowpath(lock);
950}
951EXPORT_SYMBOL(mutex_lock_killable);
952
953static noinline void __sched
954__mutex_lock_slowpath(struct mutex *lock)
955{
956 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0,
957 NULL, _RET_IP_, NULL, 0);
958}
959
960static noinline int __sched
961__mutex_lock_killable_slowpath(struct mutex *lock)
962{
963 return __mutex_lock_common(lock, TASK_KILLABLE, 0,
964 NULL, _RET_IP_, NULL, 0);
965}
966
967static noinline int __sched
968__mutex_lock_interruptible_slowpath(struct mutex *lock)
969{
970 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0,
971 NULL, _RET_IP_, NULL, 0);
972}
973
974static noinline int __sched
975__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
976{
977 return __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE, 0,
978 NULL, _RET_IP_, ctx, 1);
979}
980
981static noinline int __sched
982__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
983 struct ww_acquire_ctx *ctx)
984{
985 return __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE, 0,
986 NULL, _RET_IP_, ctx, 1);
987}
988
989#endif
990
991/**
992 * mutex_trylock - try to acquire the mutex, without waiting
993 * @lock: the mutex to be acquired
994 *
995 * Try to acquire the mutex atomically. Returns 1 if the mutex
996 * has been acquired successfully, and 0 on contention.
997 *
998 * NOTE: this function follows the spin_trylock() convention, so
999 * it is negated from the down_trylock() return values! Be careful
1000 * about this when converting semaphore users to mutexes.
1001 *
1002 * This function must not be used in interrupt context. The
1003 * mutex must be released by the same task that acquired it.
1004 */
1005int __sched mutex_trylock(struct mutex *lock)
1006{
1007 bool locked = __mutex_trylock(lock, false);
1008
1009 if (locked)
1010 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1011
1012 return locked;
1013}
1014EXPORT_SYMBOL(mutex_trylock);
1015
1016#ifndef CONFIG_DEBUG_LOCK_ALLOC
1017int __sched
1018__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1019{
1020 might_sleep();
1021
1022 if (__mutex_trylock_fast(&lock->base)) {
1023 ww_mutex_set_context_fastpath(lock, ctx);
1024 return 0;
1025 }
1026
1027 return __ww_mutex_lock_slowpath(lock, ctx);
1028}
1029EXPORT_SYMBOL(__ww_mutex_lock);
1030
1031int __sched
1032__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
1033{
1034 might_sleep();
1035
1036 if (__mutex_trylock_fast(&lock->base)) {
1037 ww_mutex_set_context_fastpath(lock, ctx);
1038 return 0;
1039 }
1040
1041 return __ww_mutex_lock_interruptible_slowpath(lock, ctx);
1042}
1043EXPORT_SYMBOL(__ww_mutex_lock_interruptible);
1044
1045#endif
1046
1047/**
1048 * atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
1049 * @cnt: the atomic which we are to dec
1050 * @lock: the mutex to return holding if we dec to 0
1051 *
1052 * return true and hold lock if we dec to 0, return false otherwise
1053 */
1054int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
1055{
1056 /* dec if we can't possibly hit 0 */
1057 if (atomic_add_unless(cnt, -1, 1))
1058 return 0;
1059 /* we might hit 0, so take the lock */
1060 mutex_lock(lock);
1061 if (!atomic_dec_and_test(cnt)) {
1062 /* when we actually did the dec, we didn't hit 0 */
1063 mutex_unlock(lock);
1064 return 0;
1065 }
1066 /* we hit 0, and we hold the lock */
1067 return 1;
1068}
1069EXPORT_SYMBOL(atomic_dec_and_mutex_lock);