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