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/init.h>
14 #include <linux/export.h>
15 #include <linux/sched.h>
16 #include <linux/sched/rt.h>
17 #include <linux/sched/wake_q.h>
18 #include <linux/osq_lock.h>
23 * Guide to the rw_semaphore's count field for common values.
24 * (32-bit case illustrated, similar for 64-bit)
26 * 0x0000000X (1) X readers active or attempting lock, no writer waiting
27 * X = #active_readers + #readers attempting to lock
30 * 0x00000000 rwsem is unlocked, and no one is waiting for the lock or
31 * attempting to read lock or write lock.
33 * 0xffff000X (1) X readers active or attempting lock, with waiters for lock
34 * X = #active readers + # readers attempting lock
35 * (X*ACTIVE_BIAS + WAITING_BIAS)
36 * (2) 1 writer attempting lock, no waiters for lock
37 * X-1 = #active readers + #readers attempting lock
38 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
39 * (3) 1 writer active, no waiters for lock
40 * X-1 = #active readers + #readers attempting lock
41 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
43 * 0xffff0001 (1) 1 reader active or attempting lock, waiters for lock
44 * (WAITING_BIAS + ACTIVE_BIAS)
45 * (2) 1 writer active or attempting lock, no waiters for lock
48 * 0xffff0000 (1) There are writers or readers queued but none active
49 * or in the process of attempting lock.
51 * Note: writer can attempt to steal lock for this count by adding
52 * ACTIVE_WRITE_BIAS in cmpxchg and checking the old count
54 * 0xfffe0001 (1) 1 writer active, or attempting lock. Waiters on queue.
55 * (ACTIVE_WRITE_BIAS + WAITING_BIAS)
57 * Note: Readers attempt to lock by adding ACTIVE_BIAS in down_read and checking
58 * the count becomes more than 0 for successful lock acquisition,
59 * i.e. the case where there are only readers or nobody has lock.
60 * (1st and 2nd case above).
62 * Writers attempt to lock by adding ACTIVE_WRITE_BIAS in down_write and
63 * checking the count becomes ACTIVE_WRITE_BIAS for successful lock
64 * acquisition (i.e. nobody else has lock or attempts lock). If
65 * unsuccessful, in rwsem_down_write_failed, we'll check to see if there
66 * are only waiters but none active (5th case above), and attempt to
72 * Initialize an rwsem:
74 void __init_rwsem(struct rw_semaphore *sem, const char *name,
75 struct lock_class_key *key)
77 #ifdef CONFIG_DEBUG_LOCK_ALLOC
79 * Make sure we are not reinitializing a held semaphore:
81 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
82 lockdep_init_map(&sem->dep_map, name, key, 0);
84 atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
85 raw_spin_lock_init(&sem->wait_lock);
86 INIT_LIST_HEAD(&sem->wait_list);
87 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
89 osq_lock_init(&sem->osq);
93 EXPORT_SYMBOL(__init_rwsem);
95 enum rwsem_waiter_type {
96 RWSEM_WAITING_FOR_WRITE,
97 RWSEM_WAITING_FOR_READ
100 struct rwsem_waiter {
101 struct list_head list;
102 struct task_struct *task;
103 enum rwsem_waiter_type type;
106 enum rwsem_wake_type {
107 RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
108 RWSEM_WAKE_READERS, /* Wake readers only */
109 RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
113 * handle the lock release when processes blocked on it that can now run
114 * - if we come here from up_xxxx(), then:
115 * - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
116 * - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
117 * - there must be someone on the queue
118 * - the wait_lock must be held by the caller
119 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
120 * to actually wakeup the blocked task(s) and drop the reference count,
121 * preferably when the wait_lock is released
122 * - woken process blocks are discarded from the list after having task zeroed
123 * - writers are only marked woken if downgrading is false
125 static void __rwsem_mark_wake(struct rw_semaphore *sem,
126 enum rwsem_wake_type wake_type,
127 struct wake_q_head *wake_q)
129 struct rwsem_waiter *waiter, *tmp;
130 long oldcount, woken = 0, adjustment = 0;
133 * Take a peek at the queue head waiter such that we can determine
134 * the wakeup(s) to perform.
136 waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list);
138 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
139 if (wake_type == RWSEM_WAKE_ANY) {
141 * Mark writer at the front of the queue for wakeup.
142 * Until the task is actually later awoken later by
143 * the caller, other writers are able to steal it.
144 * Readers, on the other hand, will block as they
145 * will notice the queued writer.
147 wake_q_add(wake_q, waiter->task);
154 * Writers might steal the lock before we grant it to the next reader.
155 * We prefer to do the first reader grant before counting readers
156 * so we can bail out early if a writer stole the lock.
158 if (wake_type != RWSEM_WAKE_READ_OWNED) {
159 adjustment = RWSEM_ACTIVE_READ_BIAS;
161 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
162 if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
164 * If the count is still less than RWSEM_WAITING_BIAS
165 * after removing the adjustment, it is assumed that
166 * a writer has stolen the lock. We have to undo our
169 if (atomic_long_add_return(-adjustment, &sem->count) <
173 /* Last active locker left. Retry waking readers. */
174 goto try_reader_grant;
177 * It is not really necessary to set it to reader-owned here,
178 * but it gives the spinners an early indication that the
179 * readers now have the lock.
181 rwsem_set_reader_owned(sem);
185 * Grant an infinite number of read locks to the readers at the front
186 * of the queue. We know that woken will be at least 1 as we accounted
187 * for above. Note we increment the 'active part' of the count by the
188 * number of readers before waking any processes up.
190 list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
191 struct task_struct *tsk;
193 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
199 wake_q_add(wake_q, tsk);
200 list_del(&waiter->list);
202 * Ensure that the last operation is setting the reader
203 * waiter to nil such that rwsem_down_read_failed() cannot
204 * race with do_exit() by always holding a reference count
205 * to the task to wakeup.
207 smp_store_release(&waiter->task, NULL);
210 adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
211 if (list_empty(&sem->wait_list)) {
212 /* hit end of list above */
213 adjustment -= RWSEM_WAITING_BIAS;
217 atomic_long_add(adjustment, &sem->count);
221 * Wait for the read lock to be granted
224 struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
226 long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
227 struct rwsem_waiter waiter;
228 DEFINE_WAKE_Q(wake_q);
230 waiter.task = current;
231 waiter.type = RWSEM_WAITING_FOR_READ;
233 raw_spin_lock_irq(&sem->wait_lock);
234 if (list_empty(&sem->wait_list))
235 adjustment += RWSEM_WAITING_BIAS;
236 list_add_tail(&waiter.list, &sem->wait_list);
238 /* we're now waiting on the lock, but no longer actively locking */
239 count = atomic_long_add_return(adjustment, &sem->count);
242 * If there are no active locks, wake the front queued process(es).
244 * If there are no writers and we are first in the queue,
245 * wake our own waiter to join the existing active readers !
247 if (count == RWSEM_WAITING_BIAS ||
248 (count > RWSEM_WAITING_BIAS &&
249 adjustment != -RWSEM_ACTIVE_READ_BIAS))
250 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
252 raw_spin_unlock_irq(&sem->wait_lock);
255 /* wait to be given the lock */
257 set_current_state(TASK_UNINTERRUPTIBLE);
263 __set_current_state(TASK_RUNNING);
266 EXPORT_SYMBOL(rwsem_down_read_failed);
269 * This function must be called with the sem->wait_lock held to prevent
270 * race conditions between checking the rwsem wait list and setting the
271 * sem->count accordingly.
273 static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
276 * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS.
278 if (count != RWSEM_WAITING_BIAS)
282 * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there
283 * are other tasks on the wait list, we need to add on WAITING_BIAS.
285 count = list_is_singular(&sem->wait_list) ?
286 RWSEM_ACTIVE_WRITE_BIAS :
287 RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS;
289 if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count)
290 == RWSEM_WAITING_BIAS) {
291 rwsem_set_owner(sem);
298 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
300 * Try to acquire write lock before the writer has been put on wait queue.
302 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
304 long old, count = atomic_long_read(&sem->count);
307 if (!(count == 0 || count == RWSEM_WAITING_BIAS))
310 old = atomic_long_cmpxchg_acquire(&sem->count, count,
311 count + RWSEM_ACTIVE_WRITE_BIAS);
313 rwsem_set_owner(sem);
321 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
323 struct task_struct *owner;
330 owner = READ_ONCE(sem->owner);
331 if (!rwsem_owner_is_writer(owner)) {
333 * Don't spin if the rwsem is readers owned.
335 ret = !rwsem_owner_is_reader(owner);
340 * As lock holder preemption issue, we both skip spinning if task is not
341 * on cpu or its cpu is preempted
343 ret = owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
350 * Return true only if we can still spin on the owner field of the rwsem.
352 static noinline bool rwsem_spin_on_owner(struct rw_semaphore *sem)
354 struct task_struct *owner = READ_ONCE(sem->owner);
356 if (!rwsem_owner_is_writer(owner))
360 while (sem->owner == owner) {
362 * Ensure we emit the owner->on_cpu, dereference _after_
363 * checking sem->owner still matches owner, if that fails,
364 * owner might point to free()d memory, if it still matches,
365 * the rcu_read_lock() ensures the memory stays valid.
370 * abort spinning when need_resched or owner is not running or
371 * owner's cpu is preempted.
373 if (!owner->on_cpu || need_resched() ||
374 vcpu_is_preempted(task_cpu(owner))) {
384 * If there is a new owner or the owner is not set, we continue
387 return !rwsem_owner_is_reader(READ_ONCE(sem->owner));
390 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
396 /* sem->wait_lock should not be held when doing optimistic spinning */
397 if (!rwsem_can_spin_on_owner(sem))
400 if (!osq_lock(&sem->osq))
404 * Optimistically spin on the owner field and attempt to acquire the
405 * lock whenever the owner changes. Spinning will be stopped when:
406 * 1) the owning writer isn't running; or
407 * 2) readers own the lock as we can't determine if they are
408 * actively running or not.
410 while (rwsem_spin_on_owner(sem)) {
412 * Try to acquire the lock
414 if (rwsem_try_write_lock_unqueued(sem)) {
420 * When there's no owner, we might have preempted between the
421 * owner acquiring the lock and setting the owner field. If
422 * we're an RT task that will live-lock because we won't let
423 * the owner complete.
425 if (!sem->owner && (need_resched() || rt_task(current)))
429 * The cpu_relax() call is a compiler barrier which forces
430 * everything in this loop to be re-loaded. We don't need
431 * memory barriers as we'll eventually observe the right
432 * values at the cost of a few extra spins.
436 osq_unlock(&sem->osq);
443 * Return true if the rwsem has active spinner
445 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
447 return osq_is_locked(&sem->osq);
451 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
456 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
463 * Wait until we successfully acquire the write lock
465 static inline struct rw_semaphore *
466 __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state)
469 bool waiting = true; /* any queued threads before us */
470 struct rwsem_waiter waiter;
471 struct rw_semaphore *ret = sem;
472 DEFINE_WAKE_Q(wake_q);
474 /* undo write bias from down_write operation, stop active locking */
475 count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count);
477 /* do optimistic spinning and steal lock if possible */
478 if (rwsem_optimistic_spin(sem))
482 * Optimistic spinning failed, proceed to the slowpath
483 * and block until we can acquire the sem.
485 waiter.task = current;
486 waiter.type = RWSEM_WAITING_FOR_WRITE;
488 raw_spin_lock_irq(&sem->wait_lock);
490 /* account for this before adding a new element to the list */
491 if (list_empty(&sem->wait_list))
494 list_add_tail(&waiter.list, &sem->wait_list);
496 /* we're now waiting on the lock, but no longer actively locking */
498 count = atomic_long_read(&sem->count);
501 * If there were already threads queued before us and there are
502 * no active writers, the lock must be read owned; so we try to
503 * wake any read locks that were queued ahead of us.
505 if (count > RWSEM_WAITING_BIAS) {
506 __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
508 * The wakeup is normally called _after_ the wait_lock
509 * is released, but given that we are proactively waking
510 * readers we can deal with the wake_q overhead as it is
511 * similar to releasing and taking the wait_lock again
512 * for attempting rwsem_try_write_lock().
517 * Reinitialize wake_q after use.
519 wake_q_init(&wake_q);
523 count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count);
525 /* wait until we successfully acquire the lock */
526 set_current_state(state);
528 if (rwsem_try_write_lock(count, sem))
530 raw_spin_unlock_irq(&sem->wait_lock);
532 /* Block until there are no active lockers. */
534 if (signal_pending_state(state, current))
538 set_current_state(state);
539 } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK);
541 raw_spin_lock_irq(&sem->wait_lock);
543 __set_current_state(TASK_RUNNING);
544 list_del(&waiter.list);
545 raw_spin_unlock_irq(&sem->wait_lock);
550 __set_current_state(TASK_RUNNING);
551 raw_spin_lock_irq(&sem->wait_lock);
552 list_del(&waiter.list);
553 if (list_empty(&sem->wait_list))
554 atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count);
556 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
557 raw_spin_unlock_irq(&sem->wait_lock);
560 return ERR_PTR(-EINTR);
563 __visible struct rw_semaphore * __sched
564 rwsem_down_write_failed(struct rw_semaphore *sem)
566 return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE);
568 EXPORT_SYMBOL(rwsem_down_write_failed);
570 __visible struct rw_semaphore * __sched
571 rwsem_down_write_failed_killable(struct rw_semaphore *sem)
573 return __rwsem_down_write_failed_common(sem, TASK_KILLABLE);
575 EXPORT_SYMBOL(rwsem_down_write_failed_killable);
578 * handle waking up a waiter on the semaphore
579 * - up_read/up_write has decremented the active part of count if we come here
582 struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
585 DEFINE_WAKE_Q(wake_q);
588 * If a spinner is present, it is not necessary to do the wakeup.
589 * Try to do wakeup only if the trylock succeeds to minimize
590 * spinlock contention which may introduce too much delay in the
593 * spinning writer up_write/up_read caller
594 * --------------- -----------------------
595 * [S] osq_unlock() [L] osq
597 * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
599 * Here, it is important to make sure that there won't be a missed
600 * wakeup while the rwsem is free and the only spinning writer goes
601 * to sleep without taking the rwsem. Even when the spinning writer
602 * is just going to break out of the waiting loop, it will still do
603 * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
604 * rwsem_has_spinner() is true, it will guarantee at least one
605 * trylock attempt on the rwsem later on.
607 if (rwsem_has_spinner(sem)) {
609 * The smp_rmb() here is to make sure that the spinner
610 * state is consulted before reading the wait_lock.
613 if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
617 raw_spin_lock_irqsave(&sem->wait_lock, flags);
620 if (!list_empty(&sem->wait_list))
621 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
623 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
628 EXPORT_SYMBOL(rwsem_wake);
631 * downgrade a write lock into a read lock
632 * - caller incremented waiting part of count and discovered it still negative
633 * - just wake up any readers at the front of the queue
636 struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
639 DEFINE_WAKE_Q(wake_q);
641 raw_spin_lock_irqsave(&sem->wait_lock, flags);
643 if (!list_empty(&sem->wait_list))
644 __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
646 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
651 EXPORT_SYMBOL(rwsem_downgrade_wake);