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 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 spinlock must be held by the caller
118 * - woken process blocks are discarded from the list after having task zeroed
119 * - writers are only woken if downgrading is false
121 static struct rw_semaphore *
122 __rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
124 struct rwsem_waiter *waiter;
125 struct task_struct *tsk;
126 struct list_head *next;
127 long oldcount, woken, loop, adjustment;
129 waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
130 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
131 if (wake_type == RWSEM_WAKE_ANY)
132 /* Wake writer at the front of the queue, but do not
133 * grant it the lock yet as we want other writers
134 * to be able to steal it. Readers, on the other hand,
135 * will block as they will notice the queued writer.
137 wake_up_process(waiter->task);
141 /* Writers might steal the lock before we grant it to the next reader.
142 * We prefer to do the first reader grant before counting readers
143 * so we can bail out early if a writer stole the lock.
146 if (wake_type != RWSEM_WAKE_READ_OWNED) {
147 adjustment = RWSEM_ACTIVE_READ_BIAS;
149 oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
150 if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
151 /* A writer stole the lock. Undo our reader grant. */
152 if (rwsem_atomic_update(-adjustment, sem) &
155 /* Last active locker left. Retry waking readers. */
156 goto try_reader_grant;
160 /* Grant an infinite number of read locks to the readers at the front
161 * of the queue. Note we increment the 'active part' of the count by
162 * the number of readers before waking any processes up.
168 if (waiter->list.next == &sem->wait_list)
171 waiter = list_entry(waiter->list.next,
172 struct rwsem_waiter, list);
174 } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
176 adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
177 if (waiter->type != RWSEM_WAITING_FOR_WRITE)
178 /* hit end of list above */
179 adjustment -= RWSEM_WAITING_BIAS;
182 rwsem_atomic_add(adjustment, sem);
184 next = sem->wait_list.next;
187 waiter = list_entry(next, struct rwsem_waiter, list);
188 next = waiter->list.next;
191 * Make sure we do not wakeup the next reader before
192 * setting the nil condition to grant the next reader;
193 * otherwise we could miss the wakeup on the other
194 * side and end up sleeping again. See the pairing
195 * in rwsem_down_read_failed().
199 wake_up_process(tsk);
200 put_task_struct(tsk);
203 sem->wait_list.next = next;
204 next->prev = &sem->wait_list;
211 * Wait for the read lock to be granted
214 struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
216 long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
217 struct rwsem_waiter waiter;
218 struct task_struct *tsk = current;
220 /* set up my own style of waitqueue */
222 waiter.type = RWSEM_WAITING_FOR_READ;
223 get_task_struct(tsk);
225 raw_spin_lock_irq(&sem->wait_lock);
226 if (list_empty(&sem->wait_list))
227 adjustment += RWSEM_WAITING_BIAS;
228 list_add_tail(&waiter.list, &sem->wait_list);
230 /* we're now waiting on the lock, but no longer actively locking */
231 count = rwsem_atomic_update(adjustment, sem);
233 /* If there are no active locks, wake the front queued process(es).
235 * If there are no writers and we are first in the queue,
236 * wake our own waiter to join the existing active readers !
238 if (count == RWSEM_WAITING_BIAS ||
239 (count > RWSEM_WAITING_BIAS &&
240 adjustment != -RWSEM_ACTIVE_READ_BIAS))
241 sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
243 raw_spin_unlock_irq(&sem->wait_lock);
245 /* wait to be given the lock */
247 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
253 __set_task_state(tsk, TASK_RUNNING);
256 EXPORT_SYMBOL(rwsem_down_read_failed);
258 static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
261 * Try acquiring the write lock. Check count first in order
262 * to reduce unnecessary expensive cmpxchg() operations.
264 if (count == RWSEM_WAITING_BIAS &&
265 cmpxchg(&sem->count, RWSEM_WAITING_BIAS,
266 RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_WAITING_BIAS) {
267 if (!list_is_singular(&sem->wait_list))
268 rwsem_atomic_update(RWSEM_WAITING_BIAS, sem);
269 rwsem_set_owner(sem);
276 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
278 * Try to acquire write lock before the writer has been put on wait queue.
280 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
282 long old, count = ACCESS_ONCE(sem->count);
285 if (!(count == 0 || count == RWSEM_WAITING_BIAS))
288 old = cmpxchg(&sem->count, count, count + RWSEM_ACTIVE_WRITE_BIAS);
290 rwsem_set_owner(sem);
298 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
300 struct task_struct *owner;
307 owner = ACCESS_ONCE(sem->owner);
309 on_cpu = owner->on_cpu;
313 * If sem->owner is not set, yet we have just recently entered the
314 * slowpath, then there is a possibility reader(s) may have the lock.
315 * To be safe, avoid spinning in these situations.
320 static inline bool owner_running(struct rw_semaphore *sem,
321 struct task_struct *owner)
323 if (sem->owner != owner)
327 * Ensure we emit the owner->on_cpu, dereference _after_ checking
328 * sem->owner still matches owner, if that fails, owner might
329 * point to free()d memory, if it still matches, the rcu_read_lock()
330 * ensures the memory stays valid.
334 return owner->on_cpu;
338 bool rwsem_spin_on_owner(struct rw_semaphore *sem, struct task_struct *owner)
341 while (owner_running(sem, owner)) {
345 cpu_relax_lowlatency();
350 * We break out the loop above on need_resched() or when the
351 * owner changed, which is a sign for heavy contention. Return
352 * success only when sem->owner is NULL.
354 return sem->owner == NULL;
357 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
359 struct task_struct *owner;
364 /* sem->wait_lock should not be held when doing optimistic spinning */
365 if (!rwsem_can_spin_on_owner(sem))
368 if (!osq_lock(&sem->osq))
372 owner = ACCESS_ONCE(sem->owner);
373 if (owner && !rwsem_spin_on_owner(sem, owner))
376 /* wait_lock will be acquired if write_lock is obtained */
377 if (rwsem_try_write_lock_unqueued(sem)) {
383 * When there's no owner, we might have preempted between the
384 * owner acquiring the lock and setting the owner field. If
385 * we're an RT task that will live-lock because we won't let
386 * the owner complete.
388 if (!owner && (need_resched() || rt_task(current)))
392 * The cpu_relax() call is a compiler barrier which forces
393 * everything in this loop to be re-loaded. We don't need
394 * memory barriers as we'll eventually observe the right
395 * values at the cost of a few extra spins.
397 cpu_relax_lowlatency();
399 osq_unlock(&sem->osq);
406 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
413 * Wait until we successfully acquire the write lock
416 struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
419 bool waiting = true; /* any queued threads before us */
420 struct rwsem_waiter waiter;
422 /* undo write bias from down_write operation, stop active locking */
423 count = rwsem_atomic_update(-RWSEM_ACTIVE_WRITE_BIAS, sem);
425 /* do optimistic spinning and steal lock if possible */
426 if (rwsem_optimistic_spin(sem))
430 * Optimistic spinning failed, proceed to the slowpath
431 * and block until we can acquire the sem.
433 waiter.task = current;
434 waiter.type = RWSEM_WAITING_FOR_WRITE;
436 raw_spin_lock_irq(&sem->wait_lock);
438 /* account for this before adding a new element to the list */
439 if (list_empty(&sem->wait_list))
442 list_add_tail(&waiter.list, &sem->wait_list);
444 /* we're now waiting on the lock, but no longer actively locking */
446 count = ACCESS_ONCE(sem->count);
449 * If there were already threads queued before us and there are
450 * no active writers, the lock must be read owned; so we try to
451 * wake any read locks that were queued ahead of us.
453 if (count > RWSEM_WAITING_BIAS)
454 sem = __rwsem_do_wake(sem, RWSEM_WAKE_READERS);
457 count = rwsem_atomic_update(RWSEM_WAITING_BIAS, sem);
459 /* wait until we successfully acquire the lock */
460 set_current_state(TASK_UNINTERRUPTIBLE);
462 if (rwsem_try_write_lock(count, sem))
464 raw_spin_unlock_irq(&sem->wait_lock);
466 /* Block until there are no active lockers. */
469 set_current_state(TASK_UNINTERRUPTIBLE);
470 } while ((count = sem->count) & RWSEM_ACTIVE_MASK);
472 raw_spin_lock_irq(&sem->wait_lock);
474 __set_current_state(TASK_RUNNING);
476 list_del(&waiter.list);
477 raw_spin_unlock_irq(&sem->wait_lock);
481 EXPORT_SYMBOL(rwsem_down_write_failed);
484 * handle waking up a waiter on the semaphore
485 * - up_read/up_write has decremented the active part of count if we come here
488 struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
492 raw_spin_lock_irqsave(&sem->wait_lock, flags);
494 /* do nothing if list empty */
495 if (!list_empty(&sem->wait_list))
496 sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
498 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
502 EXPORT_SYMBOL(rwsem_wake);
505 * downgrade a write lock into a read lock
506 * - caller incremented waiting part of count and discovered it still negative
507 * - just wake up any readers at the front of the queue
510 struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
514 raw_spin_lock_irqsave(&sem->wait_lock, flags);
516 /* do nothing if list empty */
517 if (!list_empty(&sem->wait_list))
518 sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
520 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
524 EXPORT_SYMBOL(rwsem_downgrade_wake);