1 // SPDX-License-Identifier: GPL-2.0
2 /* rwsem.c: R/W semaphores: contention handling functions
4 * Written by David Howells (dhowells@redhat.com).
5 * Derived from arch/i386/kernel/semaphore.c
7 * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
8 * and Michel Lespinasse <walken@google.com>
10 * Optimistic spinning by Tim Chen <tim.c.chen@intel.com>
11 * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes.
13 #include <linux/rwsem.h>
14 #include <linux/init.h>
15 #include <linux/export.h>
16 #include <linux/sched/signal.h>
17 #include <linux/sched/rt.h>
18 #include <linux/sched/wake_q.h>
19 #include <linux/sched/debug.h>
20 #include <linux/osq_lock.h>
25 * Guide to the rw_semaphore's count field for common values.
26 * (32-bit case illustrated, similar for 64-bit)
28 * 0x0000000X (1) X readers active or attempting lock, no writer waiting
29 * X = #active_readers + #readers attempting to lock
32 * 0x00000000 rwsem is unlocked, and no one is waiting for the lock or
33 * attempting to read lock or write lock.
35 * 0xffff000X (1) X readers active or attempting lock, with waiters for lock
36 * X = #active readers + # readers attempting lock
37 * (X*ACTIVE_BIAS + WAITING_BIAS)
38 * (2) 1 writer attempting lock, no waiters for lock
39 * X-1 = #active readers + #readers attempting lock
40 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
41 * (3) 1 writer active, no waiters for lock
42 * X-1 = #active readers + #readers attempting lock
43 * ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
45 * 0xffff0001 (1) 1 reader active or attempting lock, waiters for lock
46 * (WAITING_BIAS + ACTIVE_BIAS)
47 * (2) 1 writer active or attempting lock, no waiters for lock
50 * 0xffff0000 (1) There are writers or readers queued but none active
51 * or in the process of attempting lock.
53 * Note: writer can attempt to steal lock for this count by adding
54 * ACTIVE_WRITE_BIAS in cmpxchg and checking the old count
56 * 0xfffe0001 (1) 1 writer active, or attempting lock. Waiters on queue.
57 * (ACTIVE_WRITE_BIAS + WAITING_BIAS)
59 * Note: Readers attempt to lock by adding ACTIVE_BIAS in down_read and checking
60 * the count becomes more than 0 for successful lock acquisition,
61 * i.e. the case where there are only readers or nobody has lock.
62 * (1st and 2nd case above).
64 * Writers attempt to lock by adding ACTIVE_WRITE_BIAS in down_write and
65 * checking the count becomes ACTIVE_WRITE_BIAS for successful lock
66 * acquisition (i.e. nobody else has lock or attempts lock). If
67 * unsuccessful, in rwsem_down_write_failed, we'll check to see if there
68 * are only waiters but none active (5th case above), and attempt to
74 * Initialize an rwsem:
76 void __init_rwsem(struct rw_semaphore *sem, const char *name,
77 struct lock_class_key *key)
79 #ifdef CONFIG_DEBUG_LOCK_ALLOC
81 * Make sure we are not reinitializing a held semaphore:
83 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
84 lockdep_init_map(&sem->dep_map, name, key, 0);
86 atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
87 raw_spin_lock_init(&sem->wait_lock);
88 INIT_LIST_HEAD(&sem->wait_list);
89 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
91 osq_lock_init(&sem->osq);
95 EXPORT_SYMBOL(__init_rwsem);
97 enum rwsem_waiter_type {
98 RWSEM_WAITING_FOR_WRITE,
99 RWSEM_WAITING_FOR_READ
102 struct rwsem_waiter {
103 struct list_head list;
104 struct task_struct *task;
105 enum rwsem_waiter_type type;
108 enum rwsem_wake_type {
109 RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
110 RWSEM_WAKE_READERS, /* Wake readers only */
111 RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
115 * handle the lock release when processes blocked on it that can now run
116 * - if we come here from up_xxxx(), then:
117 * - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
118 * - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
119 * - there must be someone on the queue
120 * - the wait_lock must be held by the caller
121 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
122 * to actually wakeup the blocked task(s) and drop the reference count,
123 * preferably when the wait_lock is released
124 * - woken process blocks are discarded from the list after having task zeroed
125 * - writers are only marked woken if downgrading is false
127 static void __rwsem_mark_wake(struct rw_semaphore *sem,
128 enum rwsem_wake_type wake_type,
129 struct wake_q_head *wake_q)
131 struct rwsem_waiter *waiter, *tmp;
132 long oldcount, woken = 0, adjustment = 0;
133 struct list_head wlist;
136 * Take a peek at the queue head waiter such that we can determine
137 * the wakeup(s) to perform.
139 waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list);
141 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
142 if (wake_type == RWSEM_WAKE_ANY) {
144 * Mark writer at the front of the queue for wakeup.
145 * Until the task is actually later awoken later by
146 * the caller, other writers are able to steal it.
147 * Readers, on the other hand, will block as they
148 * will notice the queued writer.
150 wake_q_add(wake_q, waiter->task);
157 * Writers might steal the lock before we grant it to the next reader.
158 * We prefer to do the first reader grant before counting readers
159 * so we can bail out early if a writer stole the lock.
161 if (wake_type != RWSEM_WAKE_READ_OWNED) {
162 adjustment = RWSEM_ACTIVE_READ_BIAS;
164 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
165 if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
167 * If the count is still less than RWSEM_WAITING_BIAS
168 * after removing the adjustment, it is assumed that
169 * a writer has stolen the lock. We have to undo our
172 if (atomic_long_add_return(-adjustment, &sem->count) <
176 /* Last active locker left. Retry waking readers. */
177 goto try_reader_grant;
180 * It is not really necessary to set it to reader-owned here,
181 * but it gives the spinners an early indication that the
182 * readers now have the lock.
184 rwsem_set_reader_owned(sem);
188 * Grant an infinite number of read locks to the readers at the front
189 * of the queue. We know that woken will be at least 1 as we accounted
190 * for above. Note we increment the 'active part' of the count by the
191 * number of readers before waking any processes up.
193 * We have to do wakeup in 2 passes to prevent the possibility that
194 * the reader count may be decremented before it is incremented. It
195 * is because the to-be-woken waiter may not have slept yet. So it
196 * may see waiter->task got cleared, finish its critical section and
197 * do an unlock before the reader count increment.
199 * 1) Collect the read-waiters in a separate list, count them and
200 * fully increment the reader count in rwsem.
201 * 2) For each waiters in the new list, clear waiter->task and
202 * put them into wake_q to be woken up later.
204 list_for_each_entry(waiter, &sem->wait_list, list) {
205 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
210 list_cut_before(&wlist, &sem->wait_list, &waiter->list);
212 adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
213 if (list_empty(&sem->wait_list)) {
214 /* hit end of list above */
215 adjustment -= RWSEM_WAITING_BIAS;
219 atomic_long_add(adjustment, &sem->count);
222 list_for_each_entry_safe(waiter, tmp, &wlist, list) {
223 struct task_struct *tsk;
226 get_task_struct(tsk);
229 * Ensure calling get_task_struct() before setting the reader
230 * waiter to nil such that rwsem_down_read_failed() cannot
231 * race with do_exit() by always holding a reference count
232 * to the task to wakeup.
234 smp_store_release(&waiter->task, NULL);
236 * Ensure issuing the wakeup (either by us or someone else)
237 * after setting the reader waiter to nil.
239 wake_q_add(wake_q, tsk);
240 /* wake_q_add() already take the task ref */
241 put_task_struct(tsk);
246 * Wait for the read lock to be granted
248 static inline struct rw_semaphore __sched *
249 __rwsem_down_read_failed_common(struct rw_semaphore *sem, int state)
251 long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
252 struct rwsem_waiter waiter;
253 DEFINE_WAKE_Q(wake_q);
255 waiter.task = current;
256 waiter.type = RWSEM_WAITING_FOR_READ;
258 raw_spin_lock_irq(&sem->wait_lock);
259 if (list_empty(&sem->wait_list))
260 adjustment += RWSEM_WAITING_BIAS;
261 list_add_tail(&waiter.list, &sem->wait_list);
263 /* we're now waiting on the lock, but no longer actively locking */
264 count = atomic_long_add_return(adjustment, &sem->count);
267 * If there are no active locks, wake the front queued process(es).
269 * If there are no writers and we are first in the queue,
270 * wake our own waiter to join the existing active readers !
272 if (count == RWSEM_WAITING_BIAS ||
273 (count > RWSEM_WAITING_BIAS &&
274 adjustment != -RWSEM_ACTIVE_READ_BIAS))
275 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
277 raw_spin_unlock_irq(&sem->wait_lock);
280 /* wait to be given the lock */
282 set_current_state(state);
285 if (signal_pending_state(state, current)) {
286 raw_spin_lock_irq(&sem->wait_lock);
289 raw_spin_unlock_irq(&sem->wait_lock);
295 __set_current_state(TASK_RUNNING);
298 list_del(&waiter.list);
299 if (list_empty(&sem->wait_list))
300 atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count);
301 raw_spin_unlock_irq(&sem->wait_lock);
302 __set_current_state(TASK_RUNNING);
303 return ERR_PTR(-EINTR);
306 __visible struct rw_semaphore * __sched
307 rwsem_down_read_failed(struct rw_semaphore *sem)
309 return __rwsem_down_read_failed_common(sem, TASK_UNINTERRUPTIBLE);
311 EXPORT_SYMBOL(rwsem_down_read_failed);
313 __visible struct rw_semaphore * __sched
314 rwsem_down_read_failed_killable(struct rw_semaphore *sem)
316 return __rwsem_down_read_failed_common(sem, TASK_KILLABLE);
318 EXPORT_SYMBOL(rwsem_down_read_failed_killable);
321 * This function must be called with the sem->wait_lock held to prevent
322 * race conditions between checking the rwsem wait list and setting the
323 * sem->count accordingly.
325 static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
328 * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS.
330 if (count != RWSEM_WAITING_BIAS)
334 * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there
335 * are other tasks on the wait list, we need to add on WAITING_BIAS.
337 count = list_is_singular(&sem->wait_list) ?
338 RWSEM_ACTIVE_WRITE_BIAS :
339 RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS;
341 if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count)
342 == RWSEM_WAITING_BIAS) {
343 rwsem_set_owner(sem);
350 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
352 * Try to acquire write lock before the writer has been put on wait queue.
354 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
356 long old, count = atomic_long_read(&sem->count);
359 if (!(count == 0 || count == RWSEM_WAITING_BIAS))
362 old = atomic_long_cmpxchg_acquire(&sem->count, count,
363 count + RWSEM_ACTIVE_WRITE_BIAS);
365 rwsem_set_owner(sem);
373 static inline bool owner_on_cpu(struct task_struct *owner)
376 * As lock holder preemption issue, we both skip spinning if
377 * task is not on cpu or its cpu is preempted
379 return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
382 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
384 struct task_struct *owner;
387 BUILD_BUG_ON(!rwsem_has_anonymous_owner(RWSEM_OWNER_UNKNOWN));
393 owner = READ_ONCE(sem->owner);
395 ret = is_rwsem_owner_spinnable(owner) &&
403 * Return true only if we can still spin on the owner field of the rwsem.
405 static noinline bool rwsem_spin_on_owner(struct rw_semaphore *sem)
407 struct task_struct *owner = READ_ONCE(sem->owner);
409 if (!is_rwsem_owner_spinnable(owner))
413 while (owner && (READ_ONCE(sem->owner) == owner)) {
415 * Ensure we emit the owner->on_cpu, dereference _after_
416 * checking sem->owner still matches owner, if that fails,
417 * owner might point to free()d memory, if it still matches,
418 * the rcu_read_lock() ensures the memory stays valid.
423 * abort spinning when need_resched or owner is not running or
424 * owner's cpu is preempted.
426 if (need_resched() || !owner_on_cpu(owner)) {
436 * If there is a new owner or the owner is not set, we continue
439 return is_rwsem_owner_spinnable(READ_ONCE(sem->owner));
442 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
448 /* sem->wait_lock should not be held when doing optimistic spinning */
449 if (!rwsem_can_spin_on_owner(sem))
452 if (!osq_lock(&sem->osq))
456 * Optimistically spin on the owner field and attempt to acquire the
457 * lock whenever the owner changes. Spinning will be stopped when:
458 * 1) the owning writer isn't running; or
459 * 2) readers own the lock as we can't determine if they are
460 * actively running or not.
462 while (rwsem_spin_on_owner(sem)) {
464 * Try to acquire the lock
466 if (rwsem_try_write_lock_unqueued(sem)) {
472 * When there's no owner, we might have preempted between the
473 * owner acquiring the lock and setting the owner field. If
474 * we're an RT task that will live-lock because we won't let
475 * the owner complete.
477 if (!sem->owner && (need_resched() || rt_task(current)))
481 * The cpu_relax() call is a compiler barrier which forces
482 * everything in this loop to be re-loaded. We don't need
483 * memory barriers as we'll eventually observe the right
484 * values at the cost of a few extra spins.
488 osq_unlock(&sem->osq);
495 * Return true if the rwsem has active spinner
497 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
499 return osq_is_locked(&sem->osq);
503 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
508 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
515 * Wait until we successfully acquire the write lock
517 static inline struct rw_semaphore *
518 __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state)
521 bool waiting = true; /* any queued threads before us */
522 struct rwsem_waiter waiter;
523 struct rw_semaphore *ret = sem;
524 DEFINE_WAKE_Q(wake_q);
526 /* undo write bias from down_write operation, stop active locking */
527 count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count);
529 /* do optimistic spinning and steal lock if possible */
530 if (rwsem_optimistic_spin(sem))
534 * Optimistic spinning failed, proceed to the slowpath
535 * and block until we can acquire the sem.
537 waiter.task = current;
538 waiter.type = RWSEM_WAITING_FOR_WRITE;
540 raw_spin_lock_irq(&sem->wait_lock);
542 /* account for this before adding a new element to the list */
543 if (list_empty(&sem->wait_list))
546 list_add_tail(&waiter.list, &sem->wait_list);
548 /* we're now waiting on the lock, but no longer actively locking */
550 count = atomic_long_read(&sem->count);
553 * If there were already threads queued before us and there are
554 * no active writers, the lock must be read owned; so we try to
555 * wake any read locks that were queued ahead of us.
557 if (count > RWSEM_WAITING_BIAS) {
558 __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
560 * The wakeup is normally called _after_ the wait_lock
561 * is released, but given that we are proactively waking
562 * readers we can deal with the wake_q overhead as it is
563 * similar to releasing and taking the wait_lock again
564 * for attempting rwsem_try_write_lock().
569 * Reinitialize wake_q after use.
571 wake_q_init(&wake_q);
575 count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count);
577 /* wait until we successfully acquire the lock */
578 set_current_state(state);
580 if (rwsem_try_write_lock(count, sem))
582 raw_spin_unlock_irq(&sem->wait_lock);
584 /* Block until there are no active lockers. */
586 if (signal_pending_state(state, current))
590 set_current_state(state);
591 } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK);
593 raw_spin_lock_irq(&sem->wait_lock);
595 __set_current_state(TASK_RUNNING);
596 list_del(&waiter.list);
597 raw_spin_unlock_irq(&sem->wait_lock);
602 __set_current_state(TASK_RUNNING);
603 raw_spin_lock_irq(&sem->wait_lock);
604 list_del(&waiter.list);
605 if (list_empty(&sem->wait_list))
606 atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count);
608 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
609 raw_spin_unlock_irq(&sem->wait_lock);
612 return ERR_PTR(-EINTR);
615 __visible struct rw_semaphore * __sched
616 rwsem_down_write_failed(struct rw_semaphore *sem)
618 return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE);
620 EXPORT_SYMBOL(rwsem_down_write_failed);
622 __visible struct rw_semaphore * __sched
623 rwsem_down_write_failed_killable(struct rw_semaphore *sem)
625 return __rwsem_down_write_failed_common(sem, TASK_KILLABLE);
627 EXPORT_SYMBOL(rwsem_down_write_failed_killable);
630 * handle waking up a waiter on the semaphore
631 * - up_read/up_write has decremented the active part of count if we come here
634 struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
637 DEFINE_WAKE_Q(wake_q);
640 * __rwsem_down_write_failed_common(sem)
641 * rwsem_optimistic_spin(sem)
642 * osq_unlock(sem->osq)
644 * atomic_long_add_return(&sem->count)
649 * if (atomic_long_sub_return_release(&sem->count) < 0)
651 * osq_is_locked(&sem->osq)
653 * And __up_write() must observe !osq_is_locked() when it observes the
654 * atomic_long_add_return() in order to not miss a wakeup.
656 * This boils down to:
658 * [S.rel] X = 1 [RmW] r0 = (Y += 0)
660 * [RmW] Y += 1 [L] r1 = X
662 * exists (r0=1 /\ r1=0)
667 * If a spinner is present, it is not necessary to do the wakeup.
668 * Try to do wakeup only if the trylock succeeds to minimize
669 * spinlock contention which may introduce too much delay in the
672 * spinning writer up_write/up_read caller
673 * --------------- -----------------------
674 * [S] osq_unlock() [L] osq
676 * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
678 * Here, it is important to make sure that there won't be a missed
679 * wakeup while the rwsem is free and the only spinning writer goes
680 * to sleep without taking the rwsem. Even when the spinning writer
681 * is just going to break out of the waiting loop, it will still do
682 * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
683 * rwsem_has_spinner() is true, it will guarantee at least one
684 * trylock attempt on the rwsem later on.
686 if (rwsem_has_spinner(sem)) {
688 * The smp_rmb() here is to make sure that the spinner
689 * state is consulted before reading the wait_lock.
692 if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
696 raw_spin_lock_irqsave(&sem->wait_lock, flags);
699 if (!list_empty(&sem->wait_list))
700 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
702 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
707 EXPORT_SYMBOL(rwsem_wake);
710 * downgrade a write lock into a read lock
711 * - caller incremented waiting part of count and discovered it still negative
712 * - just wake up any readers at the front of the queue
715 struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
718 DEFINE_WAKE_Q(wake_q);
720 raw_spin_lock_irqsave(&sem->wait_lock, flags);
722 if (!list_empty(&sem->wait_list))
723 __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
725 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
730 EXPORT_SYMBOL(rwsem_downgrade_wake);