projects/linux·blob v3.10-rc6 · refs

642 lines · 16.8 KiB

1/*
2 * kernel/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 <[email protected]>
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/sched.h>
22#include <linux/sched/rt.h>
23#include <linux/export.h>
24#include <linux/spinlock.h>
25#include <linux/interrupt.h>
26#include <linux/debug_locks.h>
27
28/*
29 * In the DEBUG case we are using the "NULL fastpath" for mutexes,
30 * which forces all calls into the slowpath:
31 */
32#ifdef CONFIG_DEBUG_MUTEXES
33# include "mutex-debug.h"
34# include <asm-generic/mutex-null.h>
35#else
36# include "mutex.h"
37# include <asm/mutex.h>
38#endif
39
40/*
41 * A negative mutex count indicates that waiters are sleeping waiting for the
42 * mutex.
43 */
44#define MUTEX_SHOW_NO_WAITER(mutex) (atomic_read(&(mutex)->count) >= 0)
45
46void
47__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
48{
49 atomic_set(&lock->count, 1);
50 spin_lock_init(&lock->wait_lock);
51 INIT_LIST_HEAD(&lock->wait_list);
52 mutex_clear_owner(lock);
53#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
54 lock->spin_mlock = NULL;
55#endif
56
57 debug_mutex_init(lock, name, key);
58}
59
60EXPORT_SYMBOL(__mutex_init);
61
62#ifndef CONFIG_DEBUG_LOCK_ALLOC
63/*
64 * We split the mutex lock/unlock logic into separate fastpath and
65 * slowpath functions, to reduce the register pressure on the fastpath.
66 * We also put the fastpath first in the kernel image, to make sure the
67 * branch is predicted by the CPU as default-untaken.
68 */
69static __used noinline void __sched
70__mutex_lock_slowpath(atomic_t *lock_count);
71
72/**
73 * mutex_lock - acquire the mutex
74 * @lock: the mutex to be acquired
75 *
76 * Lock the mutex exclusively for this task. If the mutex is not
77 * available right now, it will sleep until it can get it.
78 *
79 * The mutex must later on be released by the same task that
80 * acquired it. Recursive locking is not allowed. The task
81 * may not exit without first unlocking the mutex. Also, kernel
82 * memory where the mutex resides mutex must not be freed with
83 * the mutex still locked. The mutex must first be initialized
84 * (or statically defined) before it can be locked. memset()-ing
85 * the mutex to 0 is not allowed.
86 *
87 * ( The CONFIG_DEBUG_MUTEXES .config option turns on debugging
88 * checks that will enforce the restrictions and will also do
89 * deadlock debugging. )
90 *
91 * This function is similar to (but not equivalent to) down().
92 */
93void __sched mutex_lock(struct mutex *lock)
94{
95 might_sleep();
96 /*
97 * The locking fastpath is the 1->0 transition from
98 * 'unlocked' into 'locked' state.
99 */
100 __mutex_fastpath_lock(&lock->count, __mutex_lock_slowpath);
101 mutex_set_owner(lock);
102}
103
104EXPORT_SYMBOL(mutex_lock);
105#endif
106
107#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
108/*
109 * In order to avoid a stampede of mutex spinners from acquiring the mutex
110 * more or less simultaneously, the spinners need to acquire a MCS lock
111 * first before spinning on the owner field.
112 *
113 * We don't inline mspin_lock() so that perf can correctly account for the
114 * time spent in this lock function.
115 */
116struct mspin_node {
117 struct mspin_node *next ;
118 int locked; /* 1 if lock acquired */
119};
120#define MLOCK(mutex) ((struct mspin_node **)&((mutex)->spin_mlock))
121
122static noinline
123void mspin_lock(struct mspin_node **lock, struct mspin_node *node)
124{
125 struct mspin_node *prev;
126
127 /* Init node */
128 node->locked = 0;
129 node->next = NULL;
130
131 prev = xchg(lock, node);
132 if (likely(prev == NULL)) {
133 /* Lock acquired */
134 node->locked = 1;
135 return;
136 }
137 ACCESS_ONCE(prev->next) = node;
138 smp_wmb();
139 /* Wait until the lock holder passes the lock down */
140 while (!ACCESS_ONCE(node->locked))
141 arch_mutex_cpu_relax();
142}
143
144static void mspin_unlock(struct mspin_node **lock, struct mspin_node *node)
145{
146 struct mspin_node *next = ACCESS_ONCE(node->next);
147
148 if (likely(!next)) {
149 /*
150 * Release the lock by setting it to NULL
151 */
152 if (cmpxchg(lock, node, NULL) == node)
153 return;
154 /* Wait until the next pointer is set */
155 while (!(next = ACCESS_ONCE(node->next)))
156 arch_mutex_cpu_relax();
157 }
158 ACCESS_ONCE(next->locked) = 1;
159 smp_wmb();
160}
161
162/*
163 * Mutex spinning code migrated from kernel/sched/core.c
164 */
165
166static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
167{
168 if (lock->owner != owner)
169 return false;
170
171 /*
172 * Ensure we emit the owner->on_cpu, dereference _after_ checking
173 * lock->owner still matches owner, if that fails, owner might
174 * point to free()d memory, if it still matches, the rcu_read_lock()
175 * ensures the memory stays valid.
176 */
177 barrier();
178
179 return owner->on_cpu;
180}
181
182/*
183 * Look out! "owner" is an entirely speculative pointer
184 * access and not reliable.
185 */
186static noinline
187int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
188{
189 rcu_read_lock();
190 while (owner_running(lock, owner)) {
191 if (need_resched())
192 break;
193
194 arch_mutex_cpu_relax();
195 }
196 rcu_read_unlock();
197
198 /*
199 * We break out the loop above on need_resched() and when the
200 * owner changed, which is a sign for heavy contention. Return
201 * success only when lock->owner is NULL.
202 */
203 return lock->owner == NULL;
204}
205
206/*
207 * Initial check for entering the mutex spinning loop
208 */
209static inline int mutex_can_spin_on_owner(struct mutex *lock)
210{
211 int retval = 1;
212
213 rcu_read_lock();
214 if (lock->owner)
215 retval = lock->owner->on_cpu;
216 rcu_read_unlock();
217 /*
218 * if lock->owner is not set, the mutex owner may have just acquired
219 * it and not set the owner yet or the mutex has been released.
220 */
221 return retval;
222}
223#endif
224
225static __used noinline void __sched __mutex_unlock_slowpath(atomic_t *lock_count);
226
227/**
228 * mutex_unlock - release the mutex
229 * @lock: the mutex to be released
230 *
231 * Unlock a mutex that has been locked by this task previously.
232 *
233 * This function must not be used in interrupt context. Unlocking
234 * of a not locked mutex is not allowed.
235 *
236 * This function is similar to (but not equivalent to) up().
237 */
238void __sched mutex_unlock(struct mutex *lock)
239{
240 /*
241 * The unlocking fastpath is the 0->1 transition from 'locked'
242 * into 'unlocked' state:
243 */
244#ifndef CONFIG_DEBUG_MUTEXES
245 /*
246 * When debugging is enabled we must not clear the owner before time,
247 * the slow path will always be taken, and that clears the owner field
248 * after verifying that it was indeed current.
249 */
250 mutex_clear_owner(lock);
251#endif
252 __mutex_fastpath_unlock(&lock->count, __mutex_unlock_slowpath);
253}
254
255EXPORT_SYMBOL(mutex_unlock);
256
257/*
258 * Lock a mutex (possibly interruptible), slowpath:
259 */
260static inline int __sched
261__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
262 struct lockdep_map *nest_lock, unsigned long ip)
263{
264 struct task_struct *task = current;
265 struct mutex_waiter waiter;
266 unsigned long flags;
267
268 preempt_disable();
269 mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
270
271#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
272 /*
273 * Optimistic spinning.
274 *
275 * We try to spin for acquisition when we find that there are no
276 * pending waiters and the lock owner is currently running on a
277 * (different) CPU.
278 *
279 * The rationale is that if the lock owner is running, it is likely to
280 * release the lock soon.
281 *
282 * Since this needs the lock owner, and this mutex implementation
283 * doesn't track the owner atomically in the lock field, we need to
284 * track it non-atomically.
285 *
286 * We can't do this for DEBUG_MUTEXES because that relies on wait_lock
287 * to serialize everything.
288 *
289 * The mutex spinners are queued up using MCS lock so that only one
290 * spinner can compete for the mutex. However, if mutex spinning isn't
291 * going to happen, there is no point in going through the lock/unlock
292 * overhead.
293 */
294 if (!mutex_can_spin_on_owner(lock))
295 goto slowpath;
296
297 for (;;) {
298 struct task_struct *owner;
299 struct mspin_node node;
300
301 /*
302 * If there's an owner, wait for it to either
303 * release the lock or go to sleep.
304 */
305 mspin_lock(MLOCK(lock), &node);
306 owner = ACCESS_ONCE(lock->owner);
307 if (owner && !mutex_spin_on_owner(lock, owner)) {
308 mspin_unlock(MLOCK(lock), &node);
309 break;
310 }
311
312 if ((atomic_read(&lock->count) == 1) &&
313 (atomic_cmpxchg(&lock->count, 1, 0) == 1)) {
314 lock_acquired(&lock->dep_map, ip);
315 mutex_set_owner(lock);
316 mspin_unlock(MLOCK(lock), &node);
317 preempt_enable();
318 return 0;
319 }
320 mspin_unlock(MLOCK(lock), &node);
321
322 /*
323 * When there's no owner, we might have preempted between the
324 * owner acquiring the lock and setting the owner field. If
325 * we're an RT task that will live-lock because we won't let
326 * the owner complete.
327 */
328 if (!owner && (need_resched() || rt_task(task)))
329 break;
330
331 /*
332 * The cpu_relax() call is a compiler barrier which forces
333 * everything in this loop to be re-loaded. We don't need
334 * memory barriers as we'll eventually observe the right
335 * values at the cost of a few extra spins.
336 */
337 arch_mutex_cpu_relax();
338 }
339slowpath:
340#endif
341 spin_lock_mutex(&lock->wait_lock, flags);
342
343 debug_mutex_lock_common(lock, &waiter);
344 debug_mutex_add_waiter(lock, &waiter, task_thread_info(task));
345
346 /* add waiting tasks to the end of the waitqueue (FIFO): */
347 list_add_tail(&waiter.list, &lock->wait_list);
348 waiter.task = task;
349
350 if (MUTEX_SHOW_NO_WAITER(lock) && (atomic_xchg(&lock->count, -1) == 1))
351 goto done;
352
353 lock_contended(&lock->dep_map, ip);
354
355 for (;;) {
356 /*
357 * Lets try to take the lock again - this is needed even if
358 * we get here for the first time (shortly after failing to
359 * acquire the lock), to make sure that we get a wakeup once
360 * it's unlocked. Later on, if we sleep, this is the
361 * operation that gives us the lock. We xchg it to -1, so
362 * that when we release the lock, we properly wake up the
363 * other waiters:
364 */
365 if (MUTEX_SHOW_NO_WAITER(lock) &&
366 (atomic_xchg(&lock->count, -1) == 1))
367 break;
368
369 /*
370 * got a signal? (This code gets eliminated in the
371 * TASK_UNINTERRUPTIBLE case.)
372 */
373 if (unlikely(signal_pending_state(state, task))) {
374 mutex_remove_waiter(lock, &waiter,
375 task_thread_info(task));
376 mutex_release(&lock->dep_map, 1, ip);
377 spin_unlock_mutex(&lock->wait_lock, flags);
378
379 debug_mutex_free_waiter(&waiter);
380 preempt_enable();
381 return -EINTR;
382 }
383 __set_task_state(task, state);
384
385 /* didn't get the lock, go to sleep: */
386 spin_unlock_mutex(&lock->wait_lock, flags);
387 schedule_preempt_disabled();
388 spin_lock_mutex(&lock->wait_lock, flags);
389 }
390
391done:
392 lock_acquired(&lock->dep_map, ip);
393 /* got the lock - rejoice! */
394 mutex_remove_waiter(lock, &waiter, current_thread_info());
395 mutex_set_owner(lock);
396
397 /* set it to 0 if there are no waiters left: */
398 if (likely(list_empty(&lock->wait_list)))
399 atomic_set(&lock->count, 0);
400
401 spin_unlock_mutex(&lock->wait_lock, flags);
402
403 debug_mutex_free_waiter(&waiter);
404 preempt_enable();
405
406 return 0;
407}
408
409#ifdef CONFIG_DEBUG_LOCK_ALLOC
410void __sched
411mutex_lock_nested(struct mutex *lock, unsigned int subclass)
412{
413 might_sleep();
414 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, subclass, NULL, _RET_IP_);
415}
416
417EXPORT_SYMBOL_GPL(mutex_lock_nested);
418
419void __sched
420_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
421{
422 might_sleep();
423 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0, nest, _RET_IP_);
424}
425
426EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
427
428int __sched
429mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
430{
431 might_sleep();
432 return __mutex_lock_common(lock, TASK_KILLABLE, subclass, NULL, _RET_IP_);
433}
434EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
435
436int __sched
437mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
438{
439 might_sleep();
440 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE,
441 subclass, NULL, _RET_IP_);
442}
443
444EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
445#endif
446
447/*
448 * Release the lock, slowpath:
449 */
450static inline void
451__mutex_unlock_common_slowpath(atomic_t *lock_count, int nested)
452{
453 struct mutex *lock = container_of(lock_count, struct mutex, count);
454 unsigned long flags;
455
456 spin_lock_mutex(&lock->wait_lock, flags);
457 mutex_release(&lock->dep_map, nested, _RET_IP_);
458 debug_mutex_unlock(lock);
459
460 /*
461 * some architectures leave the lock unlocked in the fastpath failure
462 * case, others need to leave it locked. In the later case we have to
463 * unlock it here
464 */
465 if (__mutex_slowpath_needs_to_unlock())
466 atomic_set(&lock->count, 1);
467
468 if (!list_empty(&lock->wait_list)) {
469 /* get the first entry from the wait-list: */
470 struct mutex_waiter *waiter =
471 list_entry(lock->wait_list.next,
472 struct mutex_waiter, list);
473
474 debug_mutex_wake_waiter(lock, waiter);
475
476 wake_up_process(waiter->task);
477 }
478
479 spin_unlock_mutex(&lock->wait_lock, flags);
480}
481
482/*
483 * Release the lock, slowpath:
484 */
485static __used noinline void
486__mutex_unlock_slowpath(atomic_t *lock_count)
487{
488 __mutex_unlock_common_slowpath(lock_count, 1);
489}
490
491#ifndef CONFIG_DEBUG_LOCK_ALLOC
492/*
493 * Here come the less common (and hence less performance-critical) APIs:
494 * mutex_lock_interruptible() and mutex_trylock().
495 */
496static noinline int __sched
497__mutex_lock_killable_slowpath(atomic_t *lock_count);
498
499static noinline int __sched
500__mutex_lock_interruptible_slowpath(atomic_t *lock_count);
501
502/**
503 * mutex_lock_interruptible - acquire the mutex, interruptible
504 * @lock: the mutex to be acquired
505 *
506 * Lock the mutex like mutex_lock(), and return 0 if the mutex has
507 * been acquired or sleep until the mutex becomes available. If a
508 * signal arrives while waiting for the lock then this function
509 * returns -EINTR.
510 *
511 * This function is similar to (but not equivalent to) down_interruptible().
512 */
513int __sched mutex_lock_interruptible(struct mutex *lock)
514{
515 int ret;
516
517 might_sleep();
518 ret = __mutex_fastpath_lock_retval
519 (&lock->count, __mutex_lock_interruptible_slowpath);
520 if (!ret)
521 mutex_set_owner(lock);
522
523 return ret;
524}
525
526EXPORT_SYMBOL(mutex_lock_interruptible);
527
528int __sched mutex_lock_killable(struct mutex *lock)
529{
530 int ret;
531
532 might_sleep();
533 ret = __mutex_fastpath_lock_retval
534 (&lock->count, __mutex_lock_killable_slowpath);
535 if (!ret)
536 mutex_set_owner(lock);
537
538 return ret;
539}
540EXPORT_SYMBOL(mutex_lock_killable);
541
542static __used noinline void __sched
543__mutex_lock_slowpath(atomic_t *lock_count)
544{
545 struct mutex *lock = container_of(lock_count, struct mutex, count);
546
547 __mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0, NULL, _RET_IP_);
548}
549
550static noinline int __sched
551__mutex_lock_killable_slowpath(atomic_t *lock_count)
552{
553 struct mutex *lock = container_of(lock_count, struct mutex, count);
554
555 return __mutex_lock_common(lock, TASK_KILLABLE, 0, NULL, _RET_IP_);
556}
557
558static noinline int __sched
559__mutex_lock_interruptible_slowpath(atomic_t *lock_count)
560{
561 struct mutex *lock = container_of(lock_count, struct mutex, count);
562
563 return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0, NULL, _RET_IP_);
564}
565#endif
566
567/*
568 * Spinlock based trylock, we take the spinlock and check whether we
569 * can get the lock:
570 */
571static inline int __mutex_trylock_slowpath(atomic_t *lock_count)
572{
573 struct mutex *lock = container_of(lock_count, struct mutex, count);
574 unsigned long flags;
575 int prev;
576
577 spin_lock_mutex(&lock->wait_lock, flags);
578
579 prev = atomic_xchg(&lock->count, -1);
580 if (likely(prev == 1)) {
581 mutex_set_owner(lock);
582 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
583 }
584
585 /* Set it back to 0 if there are no waiters: */
586 if (likely(list_empty(&lock->wait_list)))
587 atomic_set(&lock->count, 0);
588
589 spin_unlock_mutex(&lock->wait_lock, flags);
590
591 return prev == 1;
592}
593
594/**
595 * mutex_trylock - try to acquire the mutex, without waiting
596 * @lock: the mutex to be acquired
597 *
598 * Try to acquire the mutex atomically. Returns 1 if the mutex
599 * has been acquired successfully, and 0 on contention.
600 *
601 * NOTE: this function follows the spin_trylock() convention, so
602 * it is negated from the down_trylock() return values! Be careful
603 * about this when converting semaphore users to mutexes.
604 *
605 * This function must not be used in interrupt context. The
606 * mutex must be released by the same task that acquired it.
607 */
608int __sched mutex_trylock(struct mutex *lock)
609{
610 int ret;
611
612 ret = __mutex_fastpath_trylock(&lock->count, __mutex_trylock_slowpath);
613 if (ret)
614 mutex_set_owner(lock);
615
616 return ret;
617}
618EXPORT_SYMBOL(mutex_trylock);
619
620/**
621 * atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
622 * @cnt: the atomic which we are to dec
623 * @lock: the mutex to return holding if we dec to 0
624 *
625 * return true and hold lock if we dec to 0, return false otherwise
626 */
627int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
628{
629 /* dec if we can't possibly hit 0 */
630 if (atomic_add_unless(cnt, -1, 1))
631 return 0;
632 /* we might hit 0, so take the lock */
633 mutex_lock(lock);
634 if (!atomic_dec_and_test(cnt)) {
635 /* when we actually did the dec, we didn't hit 0 */
636 mutex_unlock(lock);
637 return 0;
638 }
639 /* we hit 0, and we hold the lock */
640 return 1;
641}
642EXPORT_SYMBOL(atomic_dec_and_mutex_lock);