1 /* rwsem.c: R/W semaphores: contention handling functions
3 * Written by David Howells (dhowells@redhat.com).
4 * Derived from arch/i386/kernel/semaphore.c
6 * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
7 * and Michel Lespinasse <walken@google.com>
9 * Optimistic spinning by Tim Chen <tim.c.chen@intel.com>
10 * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes.
12 #include <linux/rwsem.h>
13 #include <linux/sched.h>
14 #include <linux/init.h>
15 #include <linux/export.h>
16 #include <linux/sched/rt.h>
17 #include <linux/osq_lock.h>
22 * Guide to the rw_semaphore's count field for common values.
23 * (32-bit case illustrated, similar for 64-bit)
25 * 0x0000000X (1) X readers active or attempting lock, no writer waiting
26 * X = #active_readers + #readers attempting to lock
29 * 0x00000000 rwsem is unlocked, and no one is waiting for the lock or
30 * attempting to read lock or write lock.
32 * 0xffff000X (1) X readers active or attempting lock, with waiters for lock
33 * X = #active readers + # readers attempting lock
34 * (X*ACTIVE_BIAS + WAITING_BIAS)
35 * (2) 1 writer attempting lock, no waiters for lock
36 * X-1 = #active readers + #readers attempting lock
37 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
38 * (3) 1 writer active, no waiters for lock
39 * X-1 = #active readers + #readers attempting lock
40 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
42 * 0xffff0001 (1) 1 reader active or attempting lock, waiters for lock
43 * (WAITING_BIAS + ACTIVE_BIAS)
44 * (2) 1 writer active or attempting lock, no waiters for lock
47 * 0xffff0000 (1) There are writers or readers queued but none active
48 * or in the process of attempting lock.
50 * Note: writer can attempt to steal lock for this count by adding
51 * ACTIVE_WRITE_BIAS in cmpxchg and checking the old count
53 * 0xfffe0001 (1) 1 writer active, or attempting lock. Waiters on queue.
54 * (ACTIVE_WRITE_BIAS + WAITING_BIAS)
56 * Note: Readers attempt to lock by adding ACTIVE_BIAS in down_read and checking
57 * the count becomes more than 0 for successful lock acquisition,
58 * i.e. the case where there are only readers or nobody has lock.
59 * (1st and 2nd case above).
61 * Writers attempt to lock by adding ACTIVE_WRITE_BIAS in down_write and
62 * checking the count becomes ACTIVE_WRITE_BIAS for successful lock
63 * acquisition (i.e. nobody else has lock or attempts lock). If
64 * unsuccessful, in rwsem_down_write_failed, we'll check to see if there
65 * are only waiters but none active (5th case above), and attempt to
71 * Initialize an rwsem:
73 void __init_rwsem(struct rw_semaphore
*sem
, const char *name
,
74 struct lock_class_key
*key
)
76 #ifdef CONFIG_DEBUG_LOCK_ALLOC
78 * Make sure we are not reinitializing a held semaphore:
80 debug_check_no_locks_freed((void *)sem
, sizeof(*sem
));
81 lockdep_init_map(&sem
->dep_map
, name
, key
, 0);
83 atomic_long_set(&sem
->count
, RWSEM_UNLOCKED_VALUE
);
84 raw_spin_lock_init(&sem
->wait_lock
);
85 INIT_LIST_HEAD(&sem
->wait_list
);
86 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
88 osq_lock_init(&sem
->osq
);
92 EXPORT_SYMBOL(__init_rwsem
);
94 enum rwsem_waiter_type
{
95 RWSEM_WAITING_FOR_WRITE
,
96 RWSEM_WAITING_FOR_READ
100 struct list_head list
;
101 struct task_struct
*task
;
102 enum rwsem_waiter_type type
;
105 enum rwsem_wake_type
{
106 RWSEM_WAKE_ANY
, /* Wake whatever's at head of wait list */
107 RWSEM_WAKE_READERS
, /* Wake readers only */
108 RWSEM_WAKE_READ_OWNED
/* Waker thread holds the read lock */
112 * handle the lock release when processes blocked on it that can now run
113 * - if we come here from up_xxxx(), then:
114 * - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
115 * - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
116 * - there must be someone on the queue
117 * - the wait_lock must be held by the caller
118 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
119 * to actually wakeup the blocked task(s) and drop the reference count,
120 * preferably when the wait_lock is released
121 * - woken process blocks are discarded from the list after having task zeroed
122 * - writers are only marked woken if downgrading is false
124 static void __rwsem_mark_wake(struct rw_semaphore
*sem
,
125 enum rwsem_wake_type wake_type
,
126 struct wake_q_head
*wake_q
)
128 struct rwsem_waiter
*waiter
, *tmp
;
129 long oldcount
, woken
= 0, adjustment
= 0;
132 * Take a peek at the queue head waiter such that we can determine
133 * the wakeup(s) to perform.
135 waiter
= list_first_entry(&sem
->wait_list
, struct rwsem_waiter
, list
);
137 if (waiter
->type
== RWSEM_WAITING_FOR_WRITE
) {
138 if (wake_type
== RWSEM_WAKE_ANY
) {
140 * Mark writer at the front of the queue for wakeup.
141 * Until the task is actually later awoken later by
142 * the caller, other writers are able to steal it.
143 * Readers, on the other hand, will block as they
144 * will notice the queued writer.
146 wake_q_add(wake_q
, waiter
->task
);
153 * Writers might steal the lock before we grant it to the next reader.
154 * We prefer to do the first reader grant before counting readers
155 * so we can bail out early if a writer stole the lock.
157 if (wake_type
!= RWSEM_WAKE_READ_OWNED
) {
158 adjustment
= RWSEM_ACTIVE_READ_BIAS
;
160 oldcount
= atomic_long_fetch_add(adjustment
, &sem
->count
);
161 if (unlikely(oldcount
< RWSEM_WAITING_BIAS
)) {
163 * If the count is still less than RWSEM_WAITING_BIAS
164 * after removing the adjustment, it is assumed that
165 * a writer has stolen the lock. We have to undo our
168 if (atomic_long_add_return(-adjustment
, &sem
->count
) <
172 /* Last active locker left. Retry waking readers. */
173 goto try_reader_grant
;
176 * It is not really necessary to set it to reader-owned here,
177 * but it gives the spinners an early indication that the
178 * readers now have the lock.
180 rwsem_set_reader_owned(sem
);
184 * Grant an infinite number of read locks to the readers at the front
185 * of the queue. We know that woken will be at least 1 as we accounted
186 * for above. Note we increment the 'active part' of the count by the
187 * number of readers before waking any processes up.
189 list_for_each_entry_safe(waiter
, tmp
, &sem
->wait_list
, list
) {
190 struct task_struct
*tsk
;
192 if (waiter
->type
== RWSEM_WAITING_FOR_WRITE
)
198 wake_q_add(wake_q
, tsk
);
199 list_del(&waiter
->list
);
201 * Ensure that the last operation is setting the reader
202 * waiter to nil such that rwsem_down_read_failed() cannot
203 * race with do_exit() by always holding a reference count
204 * to the task to wakeup.
206 smp_store_release(&waiter
->task
, NULL
);
209 adjustment
= woken
* RWSEM_ACTIVE_READ_BIAS
- adjustment
;
210 if (list_empty(&sem
->wait_list
)) {
211 /* hit end of list above */
212 adjustment
-= RWSEM_WAITING_BIAS
;
216 atomic_long_add(adjustment
, &sem
->count
);
220 * Wait for the read lock to be granted
223 struct rw_semaphore __sched
*rwsem_down_read_failed(struct rw_semaphore
*sem
)
225 long count
, adjustment
= -RWSEM_ACTIVE_READ_BIAS
;
226 struct rwsem_waiter waiter
;
227 DEFINE_WAKE_Q(wake_q
);
229 waiter
.task
= current
;
230 waiter
.type
= RWSEM_WAITING_FOR_READ
;
232 raw_spin_lock_irq(&sem
->wait_lock
);
233 if (list_empty(&sem
->wait_list
))
234 adjustment
+= RWSEM_WAITING_BIAS
;
235 list_add_tail(&waiter
.list
, &sem
->wait_list
);
237 /* we're now waiting on the lock, but no longer actively locking */
238 count
= atomic_long_add_return(adjustment
, &sem
->count
);
241 * If there are no active locks, wake the front queued process(es).
243 * If there are no writers and we are first in the queue,
244 * wake our own waiter to join the existing active readers !
246 if (count
== RWSEM_WAITING_BIAS
||
247 (count
> RWSEM_WAITING_BIAS
&&
248 adjustment
!= -RWSEM_ACTIVE_READ_BIAS
))
249 __rwsem_mark_wake(sem
, RWSEM_WAKE_ANY
, &wake_q
);
251 raw_spin_unlock_irq(&sem
->wait_lock
);
254 /* wait to be given the lock */
256 set_current_state(TASK_UNINTERRUPTIBLE
);
262 __set_current_state(TASK_RUNNING
);
265 EXPORT_SYMBOL(rwsem_down_read_failed
);
268 * This function must be called with the sem->wait_lock held to prevent
269 * race conditions between checking the rwsem wait list and setting the
270 * sem->count accordingly.
272 static inline bool rwsem_try_write_lock(long count
, struct rw_semaphore
*sem
)
275 * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS.
277 if (count
!= RWSEM_WAITING_BIAS
)
281 * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there
282 * are other tasks on the wait list, we need to add on WAITING_BIAS.
284 count
= list_is_singular(&sem
->wait_list
) ?
285 RWSEM_ACTIVE_WRITE_BIAS
:
286 RWSEM_ACTIVE_WRITE_BIAS
+ RWSEM_WAITING_BIAS
;
288 if (atomic_long_cmpxchg_acquire(&sem
->count
, RWSEM_WAITING_BIAS
, count
)
289 == RWSEM_WAITING_BIAS
) {
290 rwsem_set_owner(sem
);
297 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
299 * Try to acquire write lock before the writer has been put on wait queue.
301 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore
*sem
)
303 long old
, count
= atomic_long_read(&sem
->count
);
306 if (!(count
== 0 || count
== RWSEM_WAITING_BIAS
))
309 old
= atomic_long_cmpxchg_acquire(&sem
->count
, count
,
310 count
+ RWSEM_ACTIVE_WRITE_BIAS
);
312 rwsem_set_owner(sem
);
320 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore
*sem
)
322 struct task_struct
*owner
;
329 owner
= READ_ONCE(sem
->owner
);
330 if (!rwsem_owner_is_writer(owner
)) {
332 * Don't spin if the rwsem is readers owned.
334 ret
= !rwsem_owner_is_reader(owner
);
339 * As lock holder preemption issue, we both skip spinning if task is not
340 * on cpu or its cpu is preempted
342 ret
= owner
->on_cpu
&& !vcpu_is_preempted(task_cpu(owner
));
349 * Return true only if we can still spin on the owner field of the rwsem.
351 static noinline
bool rwsem_spin_on_owner(struct rw_semaphore
*sem
)
353 struct task_struct
*owner
= READ_ONCE(sem
->owner
);
355 if (!rwsem_owner_is_writer(owner
))
359 while (sem
->owner
== owner
) {
361 * Ensure we emit the owner->on_cpu, dereference _after_
362 * checking sem->owner still matches owner, if that fails,
363 * owner might point to free()d memory, if it still matches,
364 * the rcu_read_lock() ensures the memory stays valid.
369 * abort spinning when need_resched or owner is not running or
370 * owner's cpu is preempted.
372 if (!owner
->on_cpu
|| need_resched() ||
373 vcpu_is_preempted(task_cpu(owner
))) {
383 * If there is a new owner or the owner is not set, we continue
386 return !rwsem_owner_is_reader(READ_ONCE(sem
->owner
));
389 static bool rwsem_optimistic_spin(struct rw_semaphore
*sem
)
395 /* sem->wait_lock should not be held when doing optimistic spinning */
396 if (!rwsem_can_spin_on_owner(sem
))
399 if (!osq_lock(&sem
->osq
))
403 * Optimistically spin on the owner field and attempt to acquire the
404 * lock whenever the owner changes. Spinning will be stopped when:
405 * 1) the owning writer isn't running; or
406 * 2) readers own the lock as we can't determine if they are
407 * actively running or not.
409 while (rwsem_spin_on_owner(sem
)) {
411 * Try to acquire the lock
413 if (rwsem_try_write_lock_unqueued(sem
)) {
419 * When there's no owner, we might have preempted between the
420 * owner acquiring the lock and setting the owner field. If
421 * we're an RT task that will live-lock because we won't let
422 * the owner complete.
424 if (!sem
->owner
&& (need_resched() || rt_task(current
)))
428 * The cpu_relax() call is a compiler barrier which forces
429 * everything in this loop to be re-loaded. We don't need
430 * memory barriers as we'll eventually observe the right
431 * values at the cost of a few extra spins.
435 osq_unlock(&sem
->osq
);
442 * Return true if the rwsem has active spinner
444 static inline bool rwsem_has_spinner(struct rw_semaphore
*sem
)
446 return osq_is_locked(&sem
->osq
);
450 static bool rwsem_optimistic_spin(struct rw_semaphore
*sem
)
455 static inline bool rwsem_has_spinner(struct rw_semaphore
*sem
)
462 * Wait until we successfully acquire the write lock
464 static inline struct rw_semaphore
*
465 __rwsem_down_write_failed_common(struct rw_semaphore
*sem
, int state
)
468 bool waiting
= true; /* any queued threads before us */
469 struct rwsem_waiter waiter
;
470 struct rw_semaphore
*ret
= sem
;
471 DEFINE_WAKE_Q(wake_q
);
473 /* undo write bias from down_write operation, stop active locking */
474 count
= atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS
, &sem
->count
);
476 /* do optimistic spinning and steal lock if possible */
477 if (rwsem_optimistic_spin(sem
))
481 * Optimistic spinning failed, proceed to the slowpath
482 * and block until we can acquire the sem.
484 waiter
.task
= current
;
485 waiter
.type
= RWSEM_WAITING_FOR_WRITE
;
487 raw_spin_lock_irq(&sem
->wait_lock
);
489 /* account for this before adding a new element to the list */
490 if (list_empty(&sem
->wait_list
))
493 list_add_tail(&waiter
.list
, &sem
->wait_list
);
495 /* we're now waiting on the lock, but no longer actively locking */
497 count
= atomic_long_read(&sem
->count
);
500 * If there were already threads queued before us and there are
501 * no active writers, the lock must be read owned; so we try to
502 * wake any read locks that were queued ahead of us.
504 if (count
> RWSEM_WAITING_BIAS
) {
505 __rwsem_mark_wake(sem
, RWSEM_WAKE_READERS
, &wake_q
);
507 * The wakeup is normally called _after_ the wait_lock
508 * is released, but given that we are proactively waking
509 * readers we can deal with the wake_q overhead as it is
510 * similar to releasing and taking the wait_lock again
511 * for attempting rwsem_try_write_lock().
516 * Reinitialize wake_q after use.
518 wake_q_init(&wake_q
);
522 count
= atomic_long_add_return(RWSEM_WAITING_BIAS
, &sem
->count
);
524 /* wait until we successfully acquire the lock */
525 set_current_state(state
);
527 if (rwsem_try_write_lock(count
, sem
))
529 raw_spin_unlock_irq(&sem
->wait_lock
);
531 /* Block until there are no active lockers. */
533 if (signal_pending_state(state
, current
))
537 set_current_state(state
);
538 } while ((count
= atomic_long_read(&sem
->count
)) & RWSEM_ACTIVE_MASK
);
540 raw_spin_lock_irq(&sem
->wait_lock
);
542 __set_current_state(TASK_RUNNING
);
543 list_del(&waiter
.list
);
544 raw_spin_unlock_irq(&sem
->wait_lock
);
549 __set_current_state(TASK_RUNNING
);
550 raw_spin_lock_irq(&sem
->wait_lock
);
551 list_del(&waiter
.list
);
552 if (list_empty(&sem
->wait_list
))
553 atomic_long_add(-RWSEM_WAITING_BIAS
, &sem
->count
);
555 __rwsem_mark_wake(sem
, RWSEM_WAKE_ANY
, &wake_q
);
556 raw_spin_unlock_irq(&sem
->wait_lock
);
559 return ERR_PTR(-EINTR
);
562 __visible
struct rw_semaphore
* __sched
563 rwsem_down_write_failed(struct rw_semaphore
*sem
)
565 return __rwsem_down_write_failed_common(sem
, TASK_UNINTERRUPTIBLE
);
567 EXPORT_SYMBOL(rwsem_down_write_failed
);
569 __visible
struct rw_semaphore
* __sched
570 rwsem_down_write_failed_killable(struct rw_semaphore
*sem
)
572 return __rwsem_down_write_failed_common(sem
, TASK_KILLABLE
);
574 EXPORT_SYMBOL(rwsem_down_write_failed_killable
);
577 * handle waking up a waiter on the semaphore
578 * - up_read/up_write has decremented the active part of count if we come here
581 struct rw_semaphore
*rwsem_wake(struct rw_semaphore
*sem
)
584 DEFINE_WAKE_Q(wake_q
);
587 * If a spinner is present, it is not necessary to do the wakeup.
588 * Try to do wakeup only if the trylock succeeds to minimize
589 * spinlock contention which may introduce too much delay in the
592 * spinning writer up_write/up_read caller
593 * --------------- -----------------------
594 * [S] osq_unlock() [L] osq
596 * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
598 * Here, it is important to make sure that there won't be a missed
599 * wakeup while the rwsem is free and the only spinning writer goes
600 * to sleep without taking the rwsem. Even when the spinning writer
601 * is just going to break out of the waiting loop, it will still do
602 * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
603 * rwsem_has_spinner() is true, it will guarantee at least one
604 * trylock attempt on the rwsem later on.
606 if (rwsem_has_spinner(sem
)) {
608 * The smp_rmb() here is to make sure that the spinner
609 * state is consulted before reading the wait_lock.
612 if (!raw_spin_trylock_irqsave(&sem
->wait_lock
, flags
))
616 raw_spin_lock_irqsave(&sem
->wait_lock
, flags
);
619 if (!list_empty(&sem
->wait_list
))
620 __rwsem_mark_wake(sem
, RWSEM_WAKE_ANY
, &wake_q
);
622 raw_spin_unlock_irqrestore(&sem
->wait_lock
, flags
);
627 EXPORT_SYMBOL(rwsem_wake
);
630 * downgrade a write lock into a read lock
631 * - caller incremented waiting part of count and discovered it still negative
632 * - just wake up any readers at the front of the queue
635 struct rw_semaphore
*rwsem_downgrade_wake(struct rw_semaphore
*sem
)
638 DEFINE_WAKE_Q(wake_q
);
640 raw_spin_lock_irqsave(&sem
->wait_lock
, flags
);
642 if (!list_empty(&sem
->wait_list
))
643 __rwsem_mark_wake(sem
, RWSEM_WAKE_READ_OWNED
, &wake_q
);
645 raw_spin_unlock_irqrestore(&sem
->wait_lock
, flags
);
650 EXPORT_SYMBOL(rwsem_downgrade_wake
);