2 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. Neither the name of the author nor the names of any co-contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * Machine independent bits of reader/writer lock implementation.
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_kdtrace.h"
39 #include "opt_no_adaptive_rwlocks.h"
41 #include <sys/param.h>
43 #include <sys/kernel.h>
45 #include <sys/mutex.h>
47 #include <sys/rwlock.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 #include <sys/turnstile.h>
52 #include <machine/cpu.h>
54 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
55 #define ADAPTIVE_RWLOCKS
58 #ifdef ADAPTIVE_RWLOCKS
59 static int rowner_retries = 10;
60 static int rowner_loops = 10000;
61 SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging");
62 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
63 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
69 static void db_show_rwlock(struct lock_object *lock);
71 static void assert_rw(struct lock_object *lock, int what);
72 static void lock_rw(struct lock_object *lock, int how);
74 static int owner_rw(struct lock_object *lock, struct thread **owner);
76 static int unlock_rw(struct lock_object *lock);
78 struct lock_class lock_class_rw = {
80 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
81 .lc_assert = assert_rw,
83 .lc_ddb_show = db_show_rwlock,
86 .lc_unlock = unlock_rw,
93 * Return a pointer to the owning thread if the lock is write-locked or
94 * NULL if the lock is unlocked or read-locked.
96 #define rw_wowner(rw) \
97 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \
98 (struct thread *)RW_OWNER((rw)->rw_lock))
101 * Returns if a write owner is recursed. Write ownership is not assured
102 * here and should be previously checked.
104 #define rw_recursed(rw) ((rw)->rw_recurse != 0)
107 * Return true if curthread helds the lock.
109 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread)
112 * Return a pointer to the owning thread for this lock who should receive
113 * any priority lent by threads that block on this lock. Currently this
114 * is identical to rw_wowner().
116 #define rw_owner(rw) rw_wowner(rw)
119 #define _rw_assert(rw, what, file, line)
123 assert_rw(struct lock_object *lock, int what)
126 rw_assert((struct rwlock *)lock, what);
130 lock_rw(struct lock_object *lock, int how)
134 rw = (struct rwlock *)lock;
142 unlock_rw(struct lock_object *lock)
146 rw = (struct rwlock *)lock;
147 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
148 if (rw->rw_lock & RW_LOCK_READ) {
159 owner_rw(struct lock_object *lock, struct thread **owner)
161 struct rwlock *rw = (struct rwlock *)lock;
162 uintptr_t x = rw->rw_lock;
164 *owner = rw_wowner(rw);
165 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) :
171 rw_init_flags(struct rwlock *rw, const char *name, int opts)
175 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
178 flags = LO_UPGRADABLE;
181 if (opts & RW_NOPROFILE)
182 flags |= LO_NOPROFILE;
183 if (!(opts & RW_NOWITNESS))
185 if (opts & RW_RECURSE)
186 flags |= LO_RECURSABLE;
190 rw->rw_lock = RW_UNLOCKED;
192 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
196 rw_destroy(struct rwlock *rw)
199 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked"));
200 KASSERT(rw->rw_recurse == 0, ("rw lock still recursed"));
201 rw->rw_lock = RW_DESTROYED;
202 lock_destroy(&rw->lock_object);
206 rw_sysinit(void *arg)
208 struct rw_args *args = arg;
210 rw_init(args->ra_rw, args->ra_desc);
214 rw_sysinit_flags(void *arg)
216 struct rw_args_flags *args = arg;
218 rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags);
222 rw_wowned(struct rwlock *rw)
225 return (rw_wowner(rw) == curthread);
229 _rw_wlock(struct rwlock *rw, const char *file, int line)
232 MPASS(curthread != NULL);
233 KASSERT(rw->rw_lock != RW_DESTROYED,
234 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
235 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
237 __rw_wlock(rw, curthread, file, line);
238 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
239 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
240 curthread->td_locks++;
244 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
248 KASSERT(rw->rw_lock != RW_DESTROYED,
249 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
251 if (rw_wlocked(rw) &&
252 (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
256 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
257 (uintptr_t)curthread);
259 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
261 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
263 curthread->td_locks++;
269 _rw_wunlock(struct rwlock *rw, const char *file, int line)
272 MPASS(curthread != NULL);
273 KASSERT(rw->rw_lock != RW_DESTROYED,
274 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
275 _rw_assert(rw, RA_WLOCKED, file, line);
276 curthread->td_locks--;
277 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
278 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
280 if (!rw_recursed(rw))
281 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
282 __rw_wunlock(rw, curthread, file, line);
285 * Determines whether a new reader can acquire a lock. Succeeds if the
286 * reader already owns a read lock and the lock is locked for read to
287 * prevent deadlock from reader recursion. Also succeeds if the lock
288 * is unlocked and has no writer waiters or spinners. Failing otherwise
289 * prioritizes writers before readers.
291 #define RW_CAN_READ(_rw) \
292 ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \
293 (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \
297 _rw_rlock(struct rwlock *rw, const char *file, int line)
299 struct turnstile *ts;
300 #ifdef ADAPTIVE_RWLOCKS
301 volatile struct thread *owner;
305 #ifdef LOCK_PROFILING
306 uint64_t waittime = 0;
311 uint64_t spin_cnt = 0;
312 uint64_t sleep_cnt = 0;
313 int64_t sleep_time = 0;
316 KASSERT(rw->rw_lock != RW_DESTROYED,
317 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
318 KASSERT(rw_wowner(rw) != curthread,
319 ("%s (%s): wlock already held @ %s:%d", __func__,
320 rw->lock_object.lo_name, file, line));
321 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
328 * Handle the easy case. If no other thread has a write
329 * lock, then try to bump up the count of read locks. Note
330 * that we have to preserve the current state of the
331 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
332 * read lock, then rw_lock must have changed, so restart
333 * the loop. Note that this handles the case of a
334 * completely unlocked rwlock since such a lock is encoded
335 * as a read lock with no waiters.
338 if (RW_CAN_READ(v)) {
340 * The RW_LOCK_READ_WAITERS flag should only be set
341 * if the lock has been unlocked and write waiters
344 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
345 v + RW_ONE_READER)) {
346 if (LOCK_LOG_TEST(&rw->lock_object, 0))
348 "%s: %p succeed %p -> %p", __func__,
350 (void *)(v + RW_ONE_READER));
355 lock_profile_obtain_lock_failed(&rw->lock_object,
356 &contested, &waittime);
358 #ifdef ADAPTIVE_RWLOCKS
360 * If the owner is running on another CPU, spin until
361 * the owner stops running or the state of the lock
364 if ((v & RW_LOCK_READ) == 0) {
365 owner = (struct thread *)RW_OWNER(v);
366 if (TD_IS_RUNNING(owner)) {
367 if (LOCK_LOG_TEST(&rw->lock_object, 0))
369 "%s: spinning on %p held by %p",
370 __func__, rw, owner);
371 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
372 owner && TD_IS_RUNNING(owner)) {
380 } else if (spintries < rowner_retries) {
382 for (i = 0; i < rowner_loops; i++) {
384 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
388 if (i != rowner_loops)
394 * Okay, now it's the hard case. Some other thread already
395 * has a write lock or there are write waiters present,
396 * acquire the turnstile lock so we can begin the process
399 ts = turnstile_trywait(&rw->lock_object);
402 * The lock might have been released while we spun, so
403 * recheck its state and restart the loop if needed.
406 if (RW_CAN_READ(v)) {
407 turnstile_cancel(ts);
411 #ifdef ADAPTIVE_RWLOCKS
413 * The current lock owner might have started executing
414 * on another CPU (or the lock could have changed
415 * owners) while we were waiting on the turnstile
416 * chain lock. If so, drop the turnstile lock and try
419 if ((v & RW_LOCK_READ) == 0) {
420 owner = (struct thread *)RW_OWNER(v);
421 if (TD_IS_RUNNING(owner)) {
422 turnstile_cancel(ts);
429 * The lock is held in write mode or it already has waiters.
431 MPASS(!RW_CAN_READ(v));
434 * If the RW_LOCK_READ_WAITERS flag is already set, then
435 * we can go ahead and block. If it is not set then try
436 * to set it. If we fail to set it drop the turnstile
437 * lock and restart the loop.
439 if (!(v & RW_LOCK_READ_WAITERS)) {
440 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
441 v | RW_LOCK_READ_WAITERS)) {
442 turnstile_cancel(ts);
445 if (LOCK_LOG_TEST(&rw->lock_object, 0))
446 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
451 * We were unable to acquire the lock and the read waiters
452 * flag is set, so we must block on the turnstile.
454 if (LOCK_LOG_TEST(&rw->lock_object, 0))
455 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
458 sleep_time -= lockstat_nsecs();
460 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
462 sleep_time += lockstat_nsecs();
465 if (LOCK_LOG_TEST(&rw->lock_object, 0))
466 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
471 * TODO: acquire "owner of record" here. Here be turnstile dragons
472 * however. turnstiles don't like owners changing between calls to
473 * turnstile_wait() currently.
475 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
476 waittime, file, line);
477 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
478 WITNESS_LOCK(&rw->lock_object, 0, file, line);
479 curthread->td_locks++;
480 curthread->td_rw_rlocks++;
483 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
486 * Record only the loops spinning and not sleeping.
488 if (spin_cnt > sleep_cnt)
489 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
494 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
500 KASSERT(rw->rw_lock != RW_DESTROYED,
501 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
502 if (!(x & RW_LOCK_READ))
504 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
505 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
507 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
508 curthread->td_locks++;
509 curthread->td_rw_rlocks++;
514 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
519 _rw_runlock(struct rwlock *rw, const char *file, int line)
521 struct turnstile *ts;
522 uintptr_t x, v, queue;
524 KASSERT(rw->rw_lock != RW_DESTROYED,
525 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
526 _rw_assert(rw, RA_RLOCKED, file, line);
527 curthread->td_locks--;
528 curthread->td_rw_rlocks--;
529 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
530 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
532 /* TODO: drop "owner of record" here. */
536 * See if there is more than one read lock held. If so,
537 * just drop one and return.
540 if (RW_READERS(x) > 1) {
541 if (atomic_cmpset_ptr(&rw->rw_lock, x,
542 x - RW_ONE_READER)) {
543 if (LOCK_LOG_TEST(&rw->lock_object, 0))
545 "%s: %p succeeded %p -> %p",
546 __func__, rw, (void *)x,
547 (void *)(x - RW_ONE_READER));
553 * If there aren't any waiters for a write lock, then try
554 * to drop it quickly.
556 if (!(x & RW_LOCK_WAITERS)) {
557 MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
559 if (atomic_cmpset_ptr(&rw->rw_lock, x, RW_UNLOCKED)) {
560 if (LOCK_LOG_TEST(&rw->lock_object, 0))
561 CTR2(KTR_LOCK, "%s: %p last succeeded",
568 * Ok, we know we have waiters and we think we are the
569 * last reader, so grab the turnstile lock.
571 turnstile_chain_lock(&rw->lock_object);
572 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
573 MPASS(v & RW_LOCK_WAITERS);
576 * Try to drop our lock leaving the lock in a unlocked
579 * If you wanted to do explicit lock handoff you'd have to
580 * do it here. You'd also want to use turnstile_signal()
581 * and you'd have to handle the race where a higher
582 * priority thread blocks on the write lock before the
583 * thread you wakeup actually runs and have the new thread
584 * "steal" the lock. For now it's a lot simpler to just
585 * wakeup all of the waiters.
587 * As above, if we fail, then another thread might have
588 * acquired a read lock, so drop the turnstile lock and
592 if (v & RW_LOCK_WRITE_WAITERS) {
593 queue = TS_EXCLUSIVE_QUEUE;
594 x |= (v & RW_LOCK_READ_WAITERS);
596 queue = TS_SHARED_QUEUE;
597 if (!atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
599 turnstile_chain_unlock(&rw->lock_object);
602 if (LOCK_LOG_TEST(&rw->lock_object, 0))
603 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
607 * Ok. The lock is released and all that's left is to
608 * wake up the waiters. Note that the lock might not be
609 * free anymore, but in that case the writers will just
610 * block again if they run before the new lock holder(s)
613 ts = turnstile_lookup(&rw->lock_object);
615 turnstile_broadcast(ts, queue);
616 turnstile_unpend(ts, TS_SHARED_LOCK);
617 turnstile_chain_unlock(&rw->lock_object);
620 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
624 * This function is called when we are unable to obtain a write lock on the
625 * first try. This means that at least one other thread holds either a
626 * read or write lock.
629 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
631 struct turnstile *ts;
632 #ifdef ADAPTIVE_RWLOCKS
633 volatile struct thread *owner;
638 #ifdef LOCK_PROFILING
639 uint64_t waittime = 0;
643 uint64_t spin_cnt = 0;
644 uint64_t sleep_cnt = 0;
645 int64_t sleep_time = 0;
648 if (rw_wlocked(rw)) {
649 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
650 ("%s: recursing but non-recursive rw %s @ %s:%d\n",
651 __func__, rw->lock_object.lo_name, file, line));
653 if (LOCK_LOG_TEST(&rw->lock_object, 0))
654 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
658 if (LOCK_LOG_TEST(&rw->lock_object, 0))
659 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
660 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
662 while (!_rw_write_lock(rw, tid)) {
666 lock_profile_obtain_lock_failed(&rw->lock_object,
667 &contested, &waittime);
668 #ifdef ADAPTIVE_RWLOCKS
670 * If the lock is write locked and the owner is
671 * running on another CPU, spin until the owner stops
672 * running or the state of the lock changes.
675 owner = (struct thread *)RW_OWNER(v);
676 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
677 if (LOCK_LOG_TEST(&rw->lock_object, 0))
678 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
679 __func__, rw, owner);
680 while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
681 TD_IS_RUNNING(owner)) {
689 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
690 spintries < rowner_retries) {
691 if (!(v & RW_LOCK_WRITE_SPINNER)) {
692 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
693 v | RW_LOCK_WRITE_SPINNER)) {
698 for (i = 0; i < rowner_loops; i++) {
699 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
704 spin_cnt += rowner_loops - i;
706 if (i != rowner_loops)
710 ts = turnstile_trywait(&rw->lock_object);
713 #ifdef ADAPTIVE_RWLOCKS
715 * The current lock owner might have started executing
716 * on another CPU (or the lock could have changed
717 * owners) while we were waiting on the turnstile
718 * chain lock. If so, drop the turnstile lock and try
721 if (!(v & RW_LOCK_READ)) {
722 owner = (struct thread *)RW_OWNER(v);
723 if (TD_IS_RUNNING(owner)) {
724 turnstile_cancel(ts);
730 * Check for the waiters flags about this rwlock.
731 * If the lock was released, without maintain any pending
732 * waiters queue, simply try to acquire it.
733 * If a pending waiters queue is present, claim the lock
734 * ownership and maintain the pending queue.
736 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
737 if ((v & ~x) == RW_UNLOCKED) {
738 x &= ~RW_LOCK_WRITE_SPINNER;
739 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
743 turnstile_cancel(ts);
746 turnstile_cancel(ts);
750 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
751 * set it. If we fail to set it, then loop back and try
754 if (!(v & RW_LOCK_WRITE_WAITERS)) {
755 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
756 v | RW_LOCK_WRITE_WAITERS)) {
757 turnstile_cancel(ts);
760 if (LOCK_LOG_TEST(&rw->lock_object, 0))
761 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
765 * We were unable to acquire the lock and the write waiters
766 * flag is set, so we must block on the turnstile.
768 if (LOCK_LOG_TEST(&rw->lock_object, 0))
769 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
772 sleep_time -= lockstat_nsecs();
774 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
776 sleep_time += lockstat_nsecs();
779 if (LOCK_LOG_TEST(&rw->lock_object, 0))
780 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
782 #ifdef ADAPTIVE_RWLOCKS
786 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
787 waittime, file, line);
790 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
793 * Record only the loops spinning and not sleeping.
795 if (spin_cnt > sleep_cnt)
796 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
801 * This function is called if the first try at releasing a write lock failed.
802 * This means that one of the 2 waiter bits must be set indicating that at
803 * least one thread is waiting on this lock.
806 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
808 struct turnstile *ts;
812 if (rw_wlocked(rw) && rw_recursed(rw)) {
814 if (LOCK_LOG_TEST(&rw->lock_object, 0))
815 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
819 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
820 ("%s: neither of the waiter flags are set", __func__));
822 if (LOCK_LOG_TEST(&rw->lock_object, 0))
823 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
825 turnstile_chain_lock(&rw->lock_object);
826 ts = turnstile_lookup(&rw->lock_object);
830 * Use the same algo as sx locks for now. Prefer waking up shared
831 * waiters if we have any over writers. This is probably not ideal.
833 * 'v' is the value we are going to write back to rw_lock. If we
834 * have waiters on both queues, we need to preserve the state of
835 * the waiter flag for the queue we don't wake up. For now this is
836 * hardcoded for the algorithm mentioned above.
838 * In the case of both readers and writers waiting we wakeup the
839 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
840 * new writer comes in before a reader it will claim the lock up
841 * above. There is probably a potential priority inversion in
842 * there that could be worked around either by waking both queues
843 * of waiters or doing some complicated lock handoff gymnastics.
846 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
847 queue = TS_EXCLUSIVE_QUEUE;
848 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
850 queue = TS_SHARED_QUEUE;
852 /* Wake up all waiters for the specific queue. */
853 if (LOCK_LOG_TEST(&rw->lock_object, 0))
854 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
855 queue == TS_SHARED_QUEUE ? "read" : "write");
856 turnstile_broadcast(ts, queue);
857 atomic_store_rel_ptr(&rw->rw_lock, v);
858 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
859 turnstile_chain_unlock(&rw->lock_object);
863 * Attempt to do a non-blocking upgrade from a read lock to a write
864 * lock. This will only succeed if this thread holds a single read
865 * lock. Returns true if the upgrade succeeded and false otherwise.
868 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
871 struct turnstile *ts;
874 KASSERT(rw->rw_lock != RW_DESTROYED,
875 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
876 _rw_assert(rw, RA_RLOCKED, file, line);
879 * Attempt to switch from one reader to a writer. If there
880 * are any write waiters, then we will have to lock the
881 * turnstile first to prevent races with another writer
882 * calling turnstile_wait() before we have claimed this
883 * turnstile. So, do the simple case of no waiters first.
885 tid = (uintptr_t)curthread;
889 if (RW_READERS(v) > 1)
891 if (!(v & RW_LOCK_WAITERS)) {
892 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
899 * Ok, we think we have waiters, so lock the turnstile.
901 ts = turnstile_trywait(&rw->lock_object);
903 if (RW_READERS(v) > 1) {
904 turnstile_cancel(ts);
908 * Try to switch from one reader to a writer again. This time
909 * we honor the current state of the waiters flags.
910 * If we obtain the lock with the flags set, then claim
911 * ownership of the turnstile.
913 x = rw->rw_lock & RW_LOCK_WAITERS;
914 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
919 turnstile_cancel(ts);
922 turnstile_cancel(ts);
924 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
926 curthread->td_rw_rlocks--;
927 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
929 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
935 * Downgrade a write lock into a single read lock.
938 _rw_downgrade(struct rwlock *rw, const char *file, int line)
940 struct turnstile *ts;
944 KASSERT(rw->rw_lock != RW_DESTROYED,
945 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
946 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
949 panic("downgrade of a recursed lock");
952 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
955 * Convert from a writer to a single reader. First we handle
956 * the easy case with no waiters. If there are any waiters, we
957 * lock the turnstile and "disown" the lock.
959 tid = (uintptr_t)curthread;
960 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
964 * Ok, we think we have waiters, so lock the turnstile so we can
965 * read the waiter flags without any races.
967 turnstile_chain_lock(&rw->lock_object);
968 v = rw->rw_lock & RW_LOCK_WAITERS;
969 rwait = v & RW_LOCK_READ_WAITERS;
970 wwait = v & RW_LOCK_WRITE_WAITERS;
971 MPASS(rwait | wwait);
974 * Downgrade from a write lock while preserving waiters flag
975 * and give up ownership of the turnstile.
977 ts = turnstile_lookup(&rw->lock_object);
980 v &= ~RW_LOCK_READ_WAITERS;
981 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
983 * Wake other readers if there are no writers pending. Otherwise they
984 * won't be able to acquire the lock anyway.
986 if (rwait && !wwait) {
987 turnstile_broadcast(ts, TS_SHARED_QUEUE);
988 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
990 turnstile_disown(ts);
991 turnstile_chain_unlock(&rw->lock_object);
993 curthread->td_rw_rlocks++;
994 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
995 LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
998 #ifdef INVARIANT_SUPPORT
1004 * In the non-WITNESS case, rw_assert() can only detect that at least
1005 * *some* thread owns an rlock, but it cannot guarantee that *this*
1006 * thread owns an rlock.
1009 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
1012 if (panicstr != NULL)
1016 case RA_LOCKED | RA_RECURSED:
1017 case RA_LOCKED | RA_NOTRECURSED:
1020 witness_assert(&rw->lock_object, what, file, line);
1023 * If some other thread has a write lock or we have one
1024 * and are asserting a read lock, fail. Also, if no one
1025 * has a lock at all, fail.
1027 if (rw->rw_lock == RW_UNLOCKED ||
1028 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED ||
1029 rw_wowner(rw) != curthread)))
1030 panic("Lock %s not %slocked @ %s:%d\n",
1031 rw->lock_object.lo_name, (what == RA_RLOCKED) ?
1032 "read " : "", file, line);
1034 if (!(rw->rw_lock & RW_LOCK_READ)) {
1035 if (rw_recursed(rw)) {
1036 if (what & RA_NOTRECURSED)
1037 panic("Lock %s recursed @ %s:%d\n",
1038 rw->lock_object.lo_name, file,
1040 } else if (what & RA_RECURSED)
1041 panic("Lock %s not recursed @ %s:%d\n",
1042 rw->lock_object.lo_name, file, line);
1047 case RA_WLOCKED | RA_RECURSED:
1048 case RA_WLOCKED | RA_NOTRECURSED:
1049 if (rw_wowner(rw) != curthread)
1050 panic("Lock %s not exclusively locked @ %s:%d\n",
1051 rw->lock_object.lo_name, file, line);
1052 if (rw_recursed(rw)) {
1053 if (what & RA_NOTRECURSED)
1054 panic("Lock %s recursed @ %s:%d\n",
1055 rw->lock_object.lo_name, file, line);
1056 } else if (what & RA_RECURSED)
1057 panic("Lock %s not recursed @ %s:%d\n",
1058 rw->lock_object.lo_name, file, line);
1062 witness_assert(&rw->lock_object, what, file, line);
1065 * If we hold a write lock fail. We can't reliably check
1066 * to see if we hold a read lock or not.
1068 if (rw_wowner(rw) == curthread)
1069 panic("Lock %s exclusively locked @ %s:%d\n",
1070 rw->lock_object.lo_name, file, line);
1074 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1078 #endif /* INVARIANT_SUPPORT */
1082 db_show_rwlock(struct lock_object *lock)
1087 rw = (struct rwlock *)lock;
1089 db_printf(" state: ");
1090 if (rw->rw_lock == RW_UNLOCKED)
1091 db_printf("UNLOCKED\n");
1092 else if (rw->rw_lock == RW_DESTROYED) {
1093 db_printf("DESTROYED\n");
1095 } else if (rw->rw_lock & RW_LOCK_READ)
1096 db_printf("RLOCK: %ju locks\n",
1097 (uintmax_t)(RW_READERS(rw->rw_lock)));
1100 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1101 td->td_tid, td->td_proc->p_pid, td->td_name);
1102 if (rw_recursed(rw))
1103 db_printf(" recursed: %u\n", rw->rw_recurse);
1105 db_printf(" waiters: ");
1106 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1107 case RW_LOCK_READ_WAITERS:
1108 db_printf("readers\n");
1110 case RW_LOCK_WRITE_WAITERS:
1111 db_printf("writers\n");
1113 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1114 db_printf("readers and writers\n");
1117 db_printf("none\n");