1 // SPDX-License-Identifier: GPL-2.0
2 /* kernel/rwsem.c: R/W semaphores, public implementation
4 * Written by David Howells (dhowells@redhat.com).
5 * Derived from asm-i386/semaphore.h
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 * Rwsem count bit fields re-definition and rwsem rearchitecture by
14 * Waiman Long <longman@redhat.com> and
15 * Peter Zijlstra <peterz@infradead.org>.
18 #include <linux/types.h>
19 #include <linux/kernel.h>
20 #include <linux/sched.h>
21 #include <linux/sched/rt.h>
22 #include <linux/sched/task.h>
23 #include <linux/sched/debug.h>
24 #include <linux/sched/wake_q.h>
25 #include <linux/sched/signal.h>
26 #include <linux/sched/clock.h>
27 #include <linux/export.h>
28 #include <linux/rwsem.h>
29 #include <linux/atomic.h>
30 #include <trace/events/lock.h>
32 #ifndef CONFIG_PREEMPT_RT
33 #include "lock_events.h"
36 * The least significant 2 bits of the owner value has the following
38 * - Bit 0: RWSEM_READER_OWNED - The rwsem is owned by readers
39 * - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
41 * When the rwsem is reader-owned and a spinning writer has timed out,
42 * the nonspinnable bit will be set to disable optimistic spinning.
44 * When a writer acquires a rwsem, it puts its task_struct pointer
45 * into the owner field. It is cleared after an unlock.
47 * When a reader acquires a rwsem, it will also puts its task_struct
48 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
49 * On unlock, the owner field will largely be left untouched. So
50 * for a free or reader-owned rwsem, the owner value may contain
51 * information about the last reader that acquires the rwsem.
53 * That information may be helpful in debugging cases where the system
54 * seems to hang on a reader owned rwsem especially if only one reader
55 * is involved. Ideally we would like to track all the readers that own
56 * a rwsem, but the overhead is simply too big.
58 * A fast path reader optimistic lock stealing is supported when the rwsem
59 * is previously owned by a writer and the following conditions are met:
60 * - rwsem is not currently writer owned
61 * - the handoff isn't set.
63 #define RWSEM_READER_OWNED (1UL << 0)
64 #define RWSEM_NONSPINNABLE (1UL << 1)
65 #define RWSEM_OWNER_FLAGS_MASK (RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
67 #ifdef CONFIG_DEBUG_RWSEMS
68 # define DEBUG_RWSEMS_WARN_ON(c, sem) do { \
69 if (!debug_locks_silent && \
70 WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
71 #c, atomic_long_read(&(sem)->count), \
72 (unsigned long) sem->magic, \
73 atomic_long_read(&(sem)->owner), (long)current, \
74 list_empty(&(sem)->wait_list) ? "" : "not ")) \
78 # define DEBUG_RWSEMS_WARN_ON(c, sem)
82 * On 64-bit architectures, the bit definitions of the count are:
84 * Bit 0 - writer locked bit
85 * Bit 1 - waiters present bit
86 * Bit 2 - lock handoff bit
88 * Bits 8-62 - 55-bit reader count
89 * Bit 63 - read fail bit
91 * On 32-bit architectures, the bit definitions of the count are:
93 * Bit 0 - writer locked bit
94 * Bit 1 - waiters present bit
95 * Bit 2 - lock handoff bit
97 * Bits 8-30 - 23-bit reader count
98 * Bit 31 - read fail bit
100 * It is not likely that the most significant bit (read fail bit) will ever
101 * be set. This guard bit is still checked anyway in the down_read() fastpath
102 * just in case we need to use up more of the reader bits for other purpose
105 * atomic_long_fetch_add() is used to obtain reader lock, whereas
106 * atomic_long_cmpxchg() will be used to obtain writer lock.
108 * There are three places where the lock handoff bit may be set or cleared.
109 * 1) rwsem_mark_wake() for readers -- set, clear
110 * 2) rwsem_try_write_lock() for writers -- set, clear
111 * 3) rwsem_del_waiter() -- clear
113 * For all the above cases, wait_lock will be held. A writer must also
114 * be the first one in the wait_list to be eligible for setting the handoff
115 * bit. So concurrent setting/clearing of handoff bit is not possible.
117 #define RWSEM_WRITER_LOCKED (1UL << 0)
118 #define RWSEM_FLAG_WAITERS (1UL << 1)
119 #define RWSEM_FLAG_HANDOFF (1UL << 2)
120 #define RWSEM_FLAG_READFAIL (1UL << (BITS_PER_LONG - 1))
122 #define RWSEM_READER_SHIFT 8
123 #define RWSEM_READER_BIAS (1UL << RWSEM_READER_SHIFT)
124 #define RWSEM_READER_MASK (~(RWSEM_READER_BIAS - 1))
125 #define RWSEM_WRITER_MASK RWSEM_WRITER_LOCKED
126 #define RWSEM_LOCK_MASK (RWSEM_WRITER_MASK|RWSEM_READER_MASK)
127 #define RWSEM_READ_FAILED_MASK (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
128 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
131 * All writes to owner are protected by WRITE_ONCE() to make sure that
132 * store tearing can't happen as optimistic spinners may read and use
133 * the owner value concurrently without lock. Read from owner, however,
134 * may not need READ_ONCE() as long as the pointer value is only used
135 * for comparison and isn't being dereferenced.
137 static inline void rwsem_set_owner(struct rw_semaphore *sem)
139 atomic_long_set(&sem->owner, (long)current);
142 static inline void rwsem_clear_owner(struct rw_semaphore *sem)
144 atomic_long_set(&sem->owner, 0);
148 * Test the flags in the owner field.
150 static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
152 return atomic_long_read(&sem->owner) & flags;
156 * The task_struct pointer of the last owning reader will be left in
159 * Note that the owner value just indicates the task has owned the rwsem
160 * previously, it may not be the real owner or one of the real owners
161 * anymore when that field is examined, so take it with a grain of salt.
163 * The reader non-spinnable bit is preserved.
165 static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
166 struct task_struct *owner)
168 unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
169 (atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE);
171 atomic_long_set(&sem->owner, val);
174 static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
176 __rwsem_set_reader_owned(sem, current);
180 * Return true if the rwsem is owned by a reader.
182 static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
184 #ifdef CONFIG_DEBUG_RWSEMS
186 * Check the count to see if it is write-locked.
188 long count = atomic_long_read(&sem->count);
190 if (count & RWSEM_WRITER_MASK)
193 return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
196 #ifdef CONFIG_DEBUG_RWSEMS
198 * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
199 * is a task pointer in owner of a reader-owned rwsem, it will be the
200 * real owner or one of the real owners. The only exception is when the
201 * unlock is done by up_read_non_owner().
203 static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
205 unsigned long val = atomic_long_read(&sem->owner);
207 while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
208 if (atomic_long_try_cmpxchg(&sem->owner, &val,
209 val & RWSEM_OWNER_FLAGS_MASK))
214 static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
220 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
221 * remains set. Otherwise, the operation will be aborted.
223 static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
225 unsigned long owner = atomic_long_read(&sem->owner);
228 if (!(owner & RWSEM_READER_OWNED))
230 if (owner & RWSEM_NONSPINNABLE)
232 } while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
233 owner | RWSEM_NONSPINNABLE));
236 static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
238 *cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
240 if (WARN_ON_ONCE(*cntp < 0))
241 rwsem_set_nonspinnable(sem);
243 if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
244 rwsem_set_reader_owned(sem);
251 static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
253 long tmp = RWSEM_UNLOCKED_VALUE;
255 if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
256 rwsem_set_owner(sem);
264 * Return just the real task structure pointer of the owner
266 static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
268 return (struct task_struct *)
269 (atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
273 * Return the real task structure pointer of the owner and the embedded
274 * flags in the owner. pflags must be non-NULL.
276 static inline struct task_struct *
277 rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
279 unsigned long owner = atomic_long_read(&sem->owner);
281 *pflags = owner & RWSEM_OWNER_FLAGS_MASK;
282 return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
286 * Guide to the rw_semaphore's count field.
288 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
291 * The lock is owned by readers when
292 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
293 * (2) some of the reader bits are set in count, and
294 * (3) the owner field has RWSEM_READ_OWNED bit set.
296 * Having some reader bits set is not enough to guarantee a readers owned
297 * lock as the readers may be in the process of backing out from the count
298 * and a writer has just released the lock. So another writer may steal
299 * the lock immediately after that.
303 * Initialize an rwsem:
305 void __init_rwsem(struct rw_semaphore *sem, const char *name,
306 struct lock_class_key *key)
308 #ifdef CONFIG_DEBUG_LOCK_ALLOC
310 * Make sure we are not reinitializing a held semaphore:
312 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
313 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
315 #ifdef CONFIG_DEBUG_RWSEMS
318 atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
319 raw_spin_lock_init(&sem->wait_lock);
320 INIT_LIST_HEAD(&sem->wait_list);
321 atomic_long_set(&sem->owner, 0L);
322 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
323 osq_lock_init(&sem->osq);
326 EXPORT_SYMBOL(__init_rwsem);
328 enum rwsem_waiter_type {
329 RWSEM_WAITING_FOR_WRITE,
330 RWSEM_WAITING_FOR_READ
333 struct rwsem_waiter {
334 struct list_head list;
335 struct task_struct *task;
336 enum rwsem_waiter_type type;
337 unsigned long timeout;
339 /* Writer only, not initialized in reader */
342 #define rwsem_first_waiter(sem) \
343 list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
345 enum rwsem_wake_type {
346 RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
347 RWSEM_WAKE_READERS, /* Wake readers only */
348 RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
352 * The typical HZ value is either 250 or 1000. So set the minimum waiting
353 * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
354 * queue before initiating the handoff protocol.
356 #define RWSEM_WAIT_TIMEOUT DIV_ROUND_UP(HZ, 250)
359 * Magic number to batch-wakeup waiting readers, even when writers are
360 * also present in the queue. This both limits the amount of work the
361 * waking thread must do and also prevents any potential counter overflow,
364 #define MAX_READERS_WAKEUP 0x100
367 rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
369 lockdep_assert_held(&sem->wait_lock);
370 list_add_tail(&waiter->list, &sem->wait_list);
371 /* caller will set RWSEM_FLAG_WAITERS */
375 * Remove a waiter from the wait_list and clear flags.
377 * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
378 * this function. Modify with care.
380 * Return: true if wait_list isn't empty and false otherwise
383 rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
385 lockdep_assert_held(&sem->wait_lock);
386 list_del(&waiter->list);
387 if (likely(!list_empty(&sem->wait_list)))
390 atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
395 * handle the lock release when processes blocked on it that can now run
396 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
398 * - there must be someone on the queue
399 * - the wait_lock must be held by the caller
400 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
401 * to actually wakeup the blocked task(s) and drop the reference count,
402 * preferably when the wait_lock is released
403 * - woken process blocks are discarded from the list after having task zeroed
404 * - writers are only marked woken if downgrading is false
406 * Implies rwsem_del_waiter() for all woken readers.
408 static void rwsem_mark_wake(struct rw_semaphore *sem,
409 enum rwsem_wake_type wake_type,
410 struct wake_q_head *wake_q)
412 struct rwsem_waiter *waiter, *tmp;
413 long oldcount, woken = 0, adjustment = 0;
414 struct list_head wlist;
416 lockdep_assert_held(&sem->wait_lock);
419 * Take a peek at the queue head waiter such that we can determine
420 * the wakeup(s) to perform.
422 waiter = rwsem_first_waiter(sem);
424 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
425 if (wake_type == RWSEM_WAKE_ANY) {
427 * Mark writer at the front of the queue for wakeup.
428 * Until the task is actually later awoken later by
429 * the caller, other writers are able to steal it.
430 * Readers, on the other hand, will block as they
431 * will notice the queued writer.
433 wake_q_add(wake_q, waiter->task);
434 lockevent_inc(rwsem_wake_writer);
441 * No reader wakeup if there are too many of them already.
443 if (unlikely(atomic_long_read(&sem->count) < 0))
447 * Writers might steal the lock before we grant it to the next reader.
448 * We prefer to do the first reader grant before counting readers
449 * so we can bail out early if a writer stole the lock.
451 if (wake_type != RWSEM_WAKE_READ_OWNED) {
452 struct task_struct *owner;
454 adjustment = RWSEM_READER_BIAS;
455 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
456 if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
458 * When we've been waiting "too" long (for writers
459 * to give up the lock), request a HANDOFF to
462 if (!(oldcount & RWSEM_FLAG_HANDOFF) &&
463 time_after(jiffies, waiter->timeout)) {
464 adjustment -= RWSEM_FLAG_HANDOFF;
465 lockevent_inc(rwsem_rlock_handoff);
468 atomic_long_add(-adjustment, &sem->count);
472 * Set it to reader-owned to give spinners an early
473 * indication that readers now have the lock.
474 * The reader nonspinnable bit seen at slowpath entry of
475 * the reader is copied over.
477 owner = waiter->task;
478 __rwsem_set_reader_owned(sem, owner);
482 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
483 * queue. We know that the woken will be at least 1 as we accounted
484 * for above. Note we increment the 'active part' of the count by the
485 * number of readers before waking any processes up.
487 * This is an adaptation of the phase-fair R/W locks where at the
488 * reader phase (first waiter is a reader), all readers are eligible
489 * to acquire the lock at the same time irrespective of their order
490 * in the queue. The writers acquire the lock according to their
491 * order in the queue.
493 * We have to do wakeup in 2 passes to prevent the possibility that
494 * the reader count may be decremented before it is incremented. It
495 * is because the to-be-woken waiter may not have slept yet. So it
496 * may see waiter->task got cleared, finish its critical section and
497 * do an unlock before the reader count increment.
499 * 1) Collect the read-waiters in a separate list, count them and
500 * fully increment the reader count in rwsem.
501 * 2) For each waiters in the new list, clear waiter->task and
502 * put them into wake_q to be woken up later.
504 INIT_LIST_HEAD(&wlist);
505 list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
506 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
510 list_move_tail(&waiter->list, &wlist);
513 * Limit # of readers that can be woken up per wakeup call.
515 if (unlikely(woken >= MAX_READERS_WAKEUP))
519 adjustment = woken * RWSEM_READER_BIAS - adjustment;
520 lockevent_cond_inc(rwsem_wake_reader, woken);
522 oldcount = atomic_long_read(&sem->count);
523 if (list_empty(&sem->wait_list)) {
525 * Combined with list_move_tail() above, this implies
526 * rwsem_del_waiter().
528 adjustment -= RWSEM_FLAG_WAITERS;
529 if (oldcount & RWSEM_FLAG_HANDOFF)
530 adjustment -= RWSEM_FLAG_HANDOFF;
533 * When we've woken a reader, we no longer need to force
534 * writers to give up the lock and we can clear HANDOFF.
536 if (oldcount & RWSEM_FLAG_HANDOFF)
537 adjustment -= RWSEM_FLAG_HANDOFF;
541 atomic_long_add(adjustment, &sem->count);
544 list_for_each_entry_safe(waiter, tmp, &wlist, list) {
545 struct task_struct *tsk;
548 get_task_struct(tsk);
551 * Ensure calling get_task_struct() before setting the reader
552 * waiter to nil such that rwsem_down_read_slowpath() cannot
553 * race with do_exit() by always holding a reference count
554 * to the task to wakeup.
556 smp_store_release(&waiter->task, NULL);
558 * Ensure issuing the wakeup (either by us or someone else)
559 * after setting the reader waiter to nil.
561 wake_q_add_safe(wake_q, tsk);
566 * Remove a waiter and try to wake up other waiters in the wait queue
567 * This function is called from the out_nolock path of both the reader and
568 * writer slowpaths with wait_lock held. It releases the wait_lock and
569 * optionally wake up waiters before it returns.
572 rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
573 struct wake_q_head *wake_q)
574 __releases(&sem->wait_lock)
576 bool first = rwsem_first_waiter(sem) == waiter;
581 * If the wait_list isn't empty and the waiter to be deleted is
582 * the first waiter, we wake up the remaining waiters as they may
583 * be eligible to acquire or spin on the lock.
585 if (rwsem_del_waiter(sem, waiter) && first)
586 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, wake_q);
587 raw_spin_unlock_irq(&sem->wait_lock);
588 if (!wake_q_empty(wake_q))
593 * This function must be called with the sem->wait_lock held to prevent
594 * race conditions between checking the rwsem wait list and setting the
595 * sem->count accordingly.
597 * Implies rwsem_del_waiter() on success.
599 static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
600 struct rwsem_waiter *waiter)
602 bool first = rwsem_first_waiter(sem) == waiter;
605 lockdep_assert_held(&sem->wait_lock);
607 count = atomic_long_read(&sem->count);
609 bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
615 /* First waiter inherits a previously set handoff bit */
616 waiter->handoff_set = true;
621 if (count & RWSEM_LOCK_MASK) {
622 if (has_handoff || (!rt_task(waiter->task) &&
623 !time_after(jiffies, waiter->timeout)))
626 new |= RWSEM_FLAG_HANDOFF;
628 new |= RWSEM_WRITER_LOCKED;
629 new &= ~RWSEM_FLAG_HANDOFF;
631 if (list_is_singular(&sem->wait_list))
632 new &= ~RWSEM_FLAG_WAITERS;
634 } while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
637 * We have either acquired the lock with handoff bit cleared or
638 * set the handoff bit.
640 if (new & RWSEM_FLAG_HANDOFF) {
641 waiter->handoff_set = true;
642 lockevent_inc(rwsem_wlock_handoff);
647 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
650 list_del(&waiter->list);
651 rwsem_set_owner(sem);
656 * The rwsem_spin_on_owner() function returns the following 4 values
657 * depending on the lock owner state.
658 * OWNER_NULL : owner is currently NULL
659 * OWNER_WRITER: when owner changes and is a writer
660 * OWNER_READER: when owner changes and the new owner may be a reader.
661 * OWNER_NONSPINNABLE:
662 * when optimistic spinning has to stop because either the
663 * owner stops running, is unknown, or its timeslice has
668 OWNER_WRITER = 1 << 1,
669 OWNER_READER = 1 << 2,
670 OWNER_NONSPINNABLE = 1 << 3,
673 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
675 * Try to acquire write lock before the writer has been put on wait queue.
677 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
679 long count = atomic_long_read(&sem->count);
681 while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
682 if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
683 count | RWSEM_WRITER_LOCKED)) {
684 rwsem_set_owner(sem);
685 lockevent_inc(rwsem_opt_lock);
692 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
694 struct task_struct *owner;
698 if (need_resched()) {
699 lockevent_inc(rwsem_opt_fail);
705 * Disable preemption is equal to the RCU read-side crital section,
706 * thus the task_strcut structure won't go away.
708 owner = rwsem_owner_flags(sem, &flags);
710 * Don't check the read-owner as the entry may be stale.
712 if ((flags & RWSEM_NONSPINNABLE) ||
713 (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
717 lockevent_cond_inc(rwsem_opt_fail, !ret);
721 #define OWNER_SPINNABLE (OWNER_NULL | OWNER_WRITER | OWNER_READER)
723 static inline enum owner_state
724 rwsem_owner_state(struct task_struct *owner, unsigned long flags)
726 if (flags & RWSEM_NONSPINNABLE)
727 return OWNER_NONSPINNABLE;
729 if (flags & RWSEM_READER_OWNED)
732 return owner ? OWNER_WRITER : OWNER_NULL;
735 static noinline enum owner_state
736 rwsem_spin_on_owner(struct rw_semaphore *sem)
738 struct task_struct *new, *owner;
739 unsigned long flags, new_flags;
740 enum owner_state state;
742 lockdep_assert_preemption_disabled();
744 owner = rwsem_owner_flags(sem, &flags);
745 state = rwsem_owner_state(owner, flags);
746 if (state != OWNER_WRITER)
751 * When a waiting writer set the handoff flag, it may spin
752 * on the owner as well. Once that writer acquires the lock,
753 * we can spin on it. So we don't need to quit even when the
754 * handoff bit is set.
756 new = rwsem_owner_flags(sem, &new_flags);
757 if ((new != owner) || (new_flags != flags)) {
758 state = rwsem_owner_state(new, new_flags);
763 * Ensure we emit the owner->on_cpu, dereference _after_
764 * checking sem->owner still matches owner, if that fails,
765 * owner might point to free()d memory, if it still matches,
766 * our spinning context already disabled preemption which is
767 * equal to RCU read-side crital section ensures the memory
772 if (need_resched() || !owner_on_cpu(owner)) {
773 state = OWNER_NONSPINNABLE;
784 * Calculate reader-owned rwsem spinning threshold for writer
786 * The more readers own the rwsem, the longer it will take for them to
787 * wind down and free the rwsem. So the empirical formula used to
788 * determine the actual spinning time limit here is:
790 * Spinning threshold = (10 + nr_readers/2)us
792 * The limit is capped to a maximum of 25us (30 readers). This is just
793 * a heuristic and is subjected to change in the future.
795 static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
797 long count = atomic_long_read(&sem->count);
798 int readers = count >> RWSEM_READER_SHIFT;
803 delta = (20 + readers) * NSEC_PER_USEC / 2;
805 return sched_clock() + delta;
808 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
811 int prev_owner_state = OWNER_NULL;
813 u64 rspin_threshold = 0;
817 /* sem->wait_lock should not be held when doing optimistic spinning */
818 if (!osq_lock(&sem->osq))
822 * Optimistically spin on the owner field and attempt to acquire the
823 * lock whenever the owner changes. Spinning will be stopped when:
824 * 1) the owning writer isn't running; or
825 * 2) readers own the lock and spinning time has exceeded limit.
828 enum owner_state owner_state;
830 owner_state = rwsem_spin_on_owner(sem);
831 if (!(owner_state & OWNER_SPINNABLE))
835 * Try to acquire the lock
837 taken = rwsem_try_write_lock_unqueued(sem);
843 * Time-based reader-owned rwsem optimistic spinning
845 if (owner_state == OWNER_READER) {
847 * Re-initialize rspin_threshold every time when
848 * the owner state changes from non-reader to reader.
849 * This allows a writer to steal the lock in between
850 * 2 reader phases and have the threshold reset at
851 * the beginning of the 2nd reader phase.
853 if (prev_owner_state != OWNER_READER) {
854 if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
856 rspin_threshold = rwsem_rspin_threshold(sem);
861 * Check time threshold once every 16 iterations to
862 * avoid calling sched_clock() too frequently so
863 * as to reduce the average latency between the times
864 * when the lock becomes free and when the spinner
865 * is ready to do a trylock.
867 else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
868 rwsem_set_nonspinnable(sem);
869 lockevent_inc(rwsem_opt_nospin);
875 * An RT task cannot do optimistic spinning if it cannot
876 * be sure the lock holder is running or live-lock may
877 * happen if the current task and the lock holder happen
878 * to run in the same CPU. However, aborting optimistic
879 * spinning while a NULL owner is detected may miss some
880 * opportunity where spinning can continue without causing
883 * There are 2 possible cases where an RT task may be able
884 * to continue spinning.
886 * 1) The lock owner is in the process of releasing the
887 * lock, sem->owner is cleared but the lock has not
889 * 2) The lock was free and owner cleared, but another
890 * task just comes in and acquire the lock before
891 * we try to get it. The new owner may be a spinnable
894 * To take advantage of two scenarios listed above, the RT
895 * task is made to retry one more time to see if it can
896 * acquire the lock or continue spinning on the new owning
897 * writer. Of course, if the time lag is long enough or the
898 * new owner is not a writer or spinnable, the RT task will
901 * If the owner is a writer, the need_resched() check is
902 * done inside rwsem_spin_on_owner(). If the owner is not
903 * a writer, need_resched() check needs to be done here.
905 if (owner_state != OWNER_WRITER) {
908 if (rt_task(current) &&
909 (prev_owner_state != OWNER_WRITER))
912 prev_owner_state = owner_state;
915 * The cpu_relax() call is a compiler barrier which forces
916 * everything in this loop to be re-loaded. We don't need
917 * memory barriers as we'll eventually observe the right
918 * values at the cost of a few extra spins.
922 osq_unlock(&sem->osq);
925 lockevent_cond_inc(rwsem_opt_fail, !taken);
930 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
931 * only be called when the reader count reaches 0.
933 static inline void clear_nonspinnable(struct rw_semaphore *sem)
935 if (unlikely(rwsem_test_oflags(sem, RWSEM_NONSPINNABLE)))
936 atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
940 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
945 static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
950 static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
952 static inline enum owner_state
953 rwsem_spin_on_owner(struct rw_semaphore *sem)
955 return OWNER_NONSPINNABLE;
960 * Prepare to wake up waiter(s) in the wait queue by putting them into the
961 * given wake_q if the rwsem lock owner isn't a writer. If rwsem is likely
962 * reader-owned, wake up read lock waiters in queue front or wake up any
963 * front waiter otherwise.
965 * This is being called from both reader and writer slow paths.
967 static inline void rwsem_cond_wake_waiter(struct rw_semaphore *sem, long count,
968 struct wake_q_head *wake_q)
970 enum rwsem_wake_type wake_type;
972 if (count & RWSEM_WRITER_MASK)
975 if (count & RWSEM_READER_MASK) {
976 wake_type = RWSEM_WAKE_READERS;
978 wake_type = RWSEM_WAKE_ANY;
979 clear_nonspinnable(sem);
981 rwsem_mark_wake(sem, wake_type, wake_q);
985 * Wait for the read lock to be granted
987 static struct rw_semaphore __sched *
988 rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
990 long adjustment = -RWSEM_READER_BIAS;
991 long rcnt = (count >> RWSEM_READER_SHIFT);
992 struct rwsem_waiter waiter;
993 DEFINE_WAKE_Q(wake_q);
996 * To prevent a constant stream of readers from starving a sleeping
997 * waiter, don't attempt optimistic lock stealing if the lock is
998 * currently owned by readers.
1000 if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
1001 (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
1005 * Reader optimistic lock stealing.
1007 if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
1008 rwsem_set_reader_owned(sem);
1009 lockevent_inc(rwsem_rlock_steal);
1012 * Wake up other readers in the wait queue if it is
1015 if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
1016 raw_spin_lock_irq(&sem->wait_lock);
1017 if (!list_empty(&sem->wait_list))
1018 rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
1020 raw_spin_unlock_irq(&sem->wait_lock);
1027 waiter.task = current;
1028 waiter.type = RWSEM_WAITING_FOR_READ;
1029 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1031 raw_spin_lock_irq(&sem->wait_lock);
1032 if (list_empty(&sem->wait_list)) {
1034 * In case the wait queue is empty and the lock isn't owned
1035 * by a writer, this reader can exit the slowpath and return
1036 * immediately as its RWSEM_READER_BIAS has already been set
1039 if (!(atomic_long_read(&sem->count) & RWSEM_WRITER_MASK)) {
1040 /* Provide lock ACQUIRE */
1041 smp_acquire__after_ctrl_dep();
1042 raw_spin_unlock_irq(&sem->wait_lock);
1043 rwsem_set_reader_owned(sem);
1044 lockevent_inc(rwsem_rlock_fast);
1047 adjustment += RWSEM_FLAG_WAITERS;
1049 rwsem_add_waiter(sem, &waiter);
1051 /* we're now waiting on the lock, but no longer actively locking */
1052 count = atomic_long_add_return(adjustment, &sem->count);
1054 rwsem_cond_wake_waiter(sem, count, &wake_q);
1055 raw_spin_unlock_irq(&sem->wait_lock);
1057 if (!wake_q_empty(&wake_q))
1060 trace_contention_begin(sem, LCB_F_READ);
1062 /* wait to be given the lock */
1064 set_current_state(state);
1065 if (!smp_load_acquire(&waiter.task)) {
1066 /* Matches rwsem_mark_wake()'s smp_store_release(). */
1069 if (signal_pending_state(state, current)) {
1070 raw_spin_lock_irq(&sem->wait_lock);
1073 raw_spin_unlock_irq(&sem->wait_lock);
1074 /* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1078 lockevent_inc(rwsem_sleep_reader);
1081 __set_current_state(TASK_RUNNING);
1082 lockevent_inc(rwsem_rlock);
1083 trace_contention_end(sem, 0);
1087 rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1088 __set_current_state(TASK_RUNNING);
1089 lockevent_inc(rwsem_rlock_fail);
1090 trace_contention_end(sem, -EINTR);
1091 return ERR_PTR(-EINTR);
1095 * Wait until we successfully acquire the write lock
1097 static struct rw_semaphore __sched *
1098 rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1100 struct rwsem_waiter waiter;
1101 DEFINE_WAKE_Q(wake_q);
1103 /* do optimistic spinning and steal lock if possible */
1104 if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1105 /* rwsem_optimistic_spin() implies ACQUIRE on success */
1110 * Optimistic spinning failed, proceed to the slowpath
1111 * and block until we can acquire the sem.
1113 waiter.task = current;
1114 waiter.type = RWSEM_WAITING_FOR_WRITE;
1115 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1116 waiter.handoff_set = false;
1118 raw_spin_lock_irq(&sem->wait_lock);
1119 rwsem_add_waiter(sem, &waiter);
1121 /* we're now waiting on the lock */
1122 if (rwsem_first_waiter(sem) != &waiter) {
1123 rwsem_cond_wake_waiter(sem, atomic_long_read(&sem->count),
1125 if (!wake_q_empty(&wake_q)) {
1127 * We want to minimize wait_lock hold time especially
1128 * when a large number of readers are to be woken up.
1130 raw_spin_unlock_irq(&sem->wait_lock);
1132 raw_spin_lock_irq(&sem->wait_lock);
1135 atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1138 /* wait until we successfully acquire the lock */
1139 set_current_state(state);
1140 trace_contention_begin(sem, LCB_F_WRITE);
1143 if (rwsem_try_write_lock(sem, &waiter)) {
1144 /* rwsem_try_write_lock() implies ACQUIRE on success */
1148 raw_spin_unlock_irq(&sem->wait_lock);
1150 if (signal_pending_state(state, current))
1154 * After setting the handoff bit and failing to acquire
1155 * the lock, attempt to spin on owner to accelerate lock
1156 * transfer. If the previous owner is a on-cpu writer and it
1157 * has just released the lock, OWNER_NULL will be returned.
1158 * In this case, we attempt to acquire the lock again
1161 if (waiter.handoff_set) {
1162 enum owner_state owner_state;
1165 owner_state = rwsem_spin_on_owner(sem);
1168 if (owner_state == OWNER_NULL)
1173 lockevent_inc(rwsem_sleep_writer);
1174 set_current_state(state);
1176 raw_spin_lock_irq(&sem->wait_lock);
1178 __set_current_state(TASK_RUNNING);
1179 raw_spin_unlock_irq(&sem->wait_lock);
1180 lockevent_inc(rwsem_wlock);
1181 trace_contention_end(sem, 0);
1185 __set_current_state(TASK_RUNNING);
1186 raw_spin_lock_irq(&sem->wait_lock);
1187 rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1188 lockevent_inc(rwsem_wlock_fail);
1189 trace_contention_end(sem, -EINTR);
1190 return ERR_PTR(-EINTR);
1194 * handle waking up a waiter on the semaphore
1195 * - up_read/up_write has decremented the active part of count if we come here
1197 static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1199 unsigned long flags;
1200 DEFINE_WAKE_Q(wake_q);
1202 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1204 if (!list_empty(&sem->wait_list))
1205 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1207 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1214 * downgrade a write lock into a read lock
1215 * - caller incremented waiting part of count and discovered it still negative
1216 * - just wake up any readers at the front of the queue
1218 static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1220 unsigned long flags;
1221 DEFINE_WAKE_Q(wake_q);
1223 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1225 if (!list_empty(&sem->wait_list))
1226 rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1228 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1237 static inline int __down_read_common(struct rw_semaphore *sem, int state)
1241 if (!rwsem_read_trylock(sem, &count)) {
1242 if (IS_ERR(rwsem_down_read_slowpath(sem, count, state)))
1244 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1249 static inline void __down_read(struct rw_semaphore *sem)
1251 __down_read_common(sem, TASK_UNINTERRUPTIBLE);
1254 static inline int __down_read_interruptible(struct rw_semaphore *sem)
1256 return __down_read_common(sem, TASK_INTERRUPTIBLE);
1259 static inline int __down_read_killable(struct rw_semaphore *sem)
1261 return __down_read_common(sem, TASK_KILLABLE);
1264 static inline int __down_read_trylock(struct rw_semaphore *sem)
1268 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1270 tmp = atomic_long_read(&sem->count);
1271 while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1272 if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1273 tmp + RWSEM_READER_BIAS)) {
1274 rwsem_set_reader_owned(sem);
1284 static inline int __down_write_common(struct rw_semaphore *sem, int state)
1286 if (unlikely(!rwsem_write_trylock(sem))) {
1287 if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1294 static inline void __down_write(struct rw_semaphore *sem)
1296 __down_write_common(sem, TASK_UNINTERRUPTIBLE);
1299 static inline int __down_write_killable(struct rw_semaphore *sem)
1301 return __down_write_common(sem, TASK_KILLABLE);
1304 static inline int __down_write_trylock(struct rw_semaphore *sem)
1306 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1307 return rwsem_write_trylock(sem);
1311 * unlock after reading
1313 static inline void __up_read(struct rw_semaphore *sem)
1317 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1318 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1320 rwsem_clear_reader_owned(sem);
1321 tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1322 DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1323 if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1324 RWSEM_FLAG_WAITERS)) {
1325 clear_nonspinnable(sem);
1331 * unlock after writing
1333 static inline void __up_write(struct rw_semaphore *sem)
1337 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1339 * sem->owner may differ from current if the ownership is transferred
1340 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1342 DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1343 !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1345 rwsem_clear_owner(sem);
1346 tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1347 if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1352 * downgrade write lock to read lock
1354 static inline void __downgrade_write(struct rw_semaphore *sem)
1359 * When downgrading from exclusive to shared ownership,
1360 * anything inside the write-locked region cannot leak
1361 * into the read side. In contrast, anything in the
1362 * read-locked region is ok to be re-ordered into the
1363 * write side. As such, rely on RELEASE semantics.
1365 DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1366 tmp = atomic_long_fetch_add_release(
1367 -RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1368 rwsem_set_reader_owned(sem);
1369 if (tmp & RWSEM_FLAG_WAITERS)
1370 rwsem_downgrade_wake(sem);
1373 #else /* !CONFIG_PREEMPT_RT */
1375 #define RT_MUTEX_BUILD_MUTEX
1376 #include "rtmutex.c"
1378 #define rwbase_set_and_save_current_state(state) \
1379 set_current_state(state)
1381 #define rwbase_restore_current_state() \
1382 __set_current_state(TASK_RUNNING)
1384 #define rwbase_rtmutex_lock_state(rtm, state) \
1385 __rt_mutex_lock(rtm, state)
1387 #define rwbase_rtmutex_slowlock_locked(rtm, state) \
1388 __rt_mutex_slowlock_locked(rtm, NULL, state)
1390 #define rwbase_rtmutex_unlock(rtm) \
1391 __rt_mutex_unlock(rtm)
1393 #define rwbase_rtmutex_trylock(rtm) \
1394 __rt_mutex_trylock(rtm)
1396 #define rwbase_signal_pending_state(state, current) \
1397 signal_pending_state(state, current)
1399 #define rwbase_schedule() \
1402 #include "rwbase_rt.c"
1404 void __init_rwsem(struct rw_semaphore *sem, const char *name,
1405 struct lock_class_key *key)
1407 init_rwbase_rt(&(sem)->rwbase);
1409 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1410 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1411 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1414 EXPORT_SYMBOL(__init_rwsem);
1416 static inline void __down_read(struct rw_semaphore *sem)
1418 rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1421 static inline int __down_read_interruptible(struct rw_semaphore *sem)
1423 return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1426 static inline int __down_read_killable(struct rw_semaphore *sem)
1428 return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1431 static inline int __down_read_trylock(struct rw_semaphore *sem)
1433 return rwbase_read_trylock(&sem->rwbase);
1436 static inline void __up_read(struct rw_semaphore *sem)
1438 rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1441 static inline void __sched __down_write(struct rw_semaphore *sem)
1443 rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1446 static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1448 return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1451 static inline int __down_write_trylock(struct rw_semaphore *sem)
1453 return rwbase_write_trylock(&sem->rwbase);
1456 static inline void __up_write(struct rw_semaphore *sem)
1458 rwbase_write_unlock(&sem->rwbase);
1461 static inline void __downgrade_write(struct rw_semaphore *sem)
1463 rwbase_write_downgrade(&sem->rwbase);
1466 /* Debug stubs for the common API */
1467 #define DEBUG_RWSEMS_WARN_ON(c, sem)
1469 static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1470 struct task_struct *owner)
1474 static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1476 int count = atomic_read(&sem->rwbase.readers);
1478 return count < 0 && count != READER_BIAS;
1481 #endif /* CONFIG_PREEMPT_RT */
1486 void __sched down_read(struct rw_semaphore *sem)
1489 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1491 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1493 EXPORT_SYMBOL(down_read);
1495 int __sched down_read_interruptible(struct rw_semaphore *sem)
1498 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1500 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1501 rwsem_release(&sem->dep_map, _RET_IP_);
1507 EXPORT_SYMBOL(down_read_interruptible);
1509 int __sched down_read_killable(struct rw_semaphore *sem)
1512 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1514 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1515 rwsem_release(&sem->dep_map, _RET_IP_);
1521 EXPORT_SYMBOL(down_read_killable);
1524 * trylock for reading -- returns 1 if successful, 0 if contention
1526 int down_read_trylock(struct rw_semaphore *sem)
1528 int ret = __down_read_trylock(sem);
1531 rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1534 EXPORT_SYMBOL(down_read_trylock);
1539 void __sched down_write(struct rw_semaphore *sem)
1542 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1543 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1545 EXPORT_SYMBOL(down_write);
1550 int __sched down_write_killable(struct rw_semaphore *sem)
1553 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1555 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1556 __down_write_killable)) {
1557 rwsem_release(&sem->dep_map, _RET_IP_);
1563 EXPORT_SYMBOL(down_write_killable);
1566 * trylock for writing -- returns 1 if successful, 0 if contention
1568 int down_write_trylock(struct rw_semaphore *sem)
1570 int ret = __down_write_trylock(sem);
1573 rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1577 EXPORT_SYMBOL(down_write_trylock);
1580 * release a read lock
1582 void up_read(struct rw_semaphore *sem)
1584 rwsem_release(&sem->dep_map, _RET_IP_);
1587 EXPORT_SYMBOL(up_read);
1590 * release a write lock
1592 void up_write(struct rw_semaphore *sem)
1594 rwsem_release(&sem->dep_map, _RET_IP_);
1597 EXPORT_SYMBOL(up_write);
1600 * downgrade write lock to read lock
1602 void downgrade_write(struct rw_semaphore *sem)
1604 lock_downgrade(&sem->dep_map, _RET_IP_);
1605 __downgrade_write(sem);
1607 EXPORT_SYMBOL(downgrade_write);
1609 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1611 void down_read_nested(struct rw_semaphore *sem, int subclass)
1614 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1615 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1617 EXPORT_SYMBOL(down_read_nested);
1619 int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1622 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1624 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1625 rwsem_release(&sem->dep_map, _RET_IP_);
1631 EXPORT_SYMBOL(down_read_killable_nested);
1633 void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1636 rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1637 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1639 EXPORT_SYMBOL(_down_write_nest_lock);
1641 void down_read_non_owner(struct rw_semaphore *sem)
1645 __rwsem_set_reader_owned(sem, NULL);
1647 EXPORT_SYMBOL(down_read_non_owner);
1649 void down_write_nested(struct rw_semaphore *sem, int subclass)
1652 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1653 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1655 EXPORT_SYMBOL(down_write_nested);
1657 int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1660 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1662 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1663 __down_write_killable)) {
1664 rwsem_release(&sem->dep_map, _RET_IP_);
1670 EXPORT_SYMBOL(down_write_killable_nested);
1672 void up_read_non_owner(struct rw_semaphore *sem)
1674 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1677 EXPORT_SYMBOL(up_read_non_owner);