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.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * Machine independent bits of reader/writer lock implementation.
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
35 #include "opt_hwpmc_hooks.h"
36 #include "opt_no_adaptive_rwlocks.h"
38 #include <sys/param.h>
41 #include <sys/kernel.h>
43 #include <sys/mutex.h>
45 #include <sys/rwlock.h>
46 #include <sys/sched.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
59 #include <sys/pmckern.h>
60 PMC_SOFT_DECLARE( , , lock, failed);
64 * Return the rwlock address when the lock cookie address is provided.
65 * This functionality assumes that struct rwlock* have a member named rw_lock.
67 #define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock))
72 static void db_show_rwlock(const struct lock_object *lock);
74 static void assert_rw(const struct lock_object *lock, int what);
75 static void lock_rw(struct lock_object *lock, uintptr_t how);
77 static int owner_rw(const struct lock_object *lock, struct thread **owner);
79 static uintptr_t unlock_rw(struct lock_object *lock);
81 struct lock_class lock_class_rw = {
83 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
84 .lc_assert = assert_rw,
86 .lc_ddb_show = db_show_rwlock,
89 .lc_unlock = unlock_rw,
95 #ifdef ADAPTIVE_RWLOCKS
96 static int rowner_retries = 10;
97 static int rowner_loops = 10000;
98 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
100 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
101 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
103 static struct lock_delay_config rw_delay = {
110 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_initial, CTLFLAG_RW, &rw_delay.initial,
112 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_step, CTLFLAG_RW, &rw_delay.step,
114 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_min, CTLFLAG_RW, &rw_delay.min,
116 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
120 rw_delay_sysinit(void *dummy)
123 rw_delay.initial = mp_ncpus * 25;
124 rw_delay.step = (mp_ncpus * 25) / 2;
125 rw_delay.min = mp_ncpus * 5;
126 rw_delay.max = mp_ncpus * 25 * 10;
128 LOCK_DELAY_SYSINIT(rw_delay_sysinit);
132 * Return a pointer to the owning thread if the lock is write-locked or
133 * NULL if the lock is unlocked or read-locked.
135 #define rw_wowner(rw) \
136 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \
137 (struct thread *)RW_OWNER((rw)->rw_lock))
140 * Returns if a write owner is recursed. Write ownership is not assured
141 * here and should be previously checked.
143 #define rw_recursed(rw) ((rw)->rw_recurse != 0)
146 * Return true if curthread helds the lock.
148 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread)
151 * Return a pointer to the owning thread for this lock who should receive
152 * any priority lent by threads that block on this lock. Currently this
153 * is identical to rw_wowner().
155 #define rw_owner(rw) rw_wowner(rw)
158 #define __rw_assert(c, what, file, line)
162 assert_rw(const struct lock_object *lock, int what)
165 rw_assert((const struct rwlock *)lock, what);
169 lock_rw(struct lock_object *lock, uintptr_t how)
173 rw = (struct rwlock *)lock;
181 unlock_rw(struct lock_object *lock)
185 rw = (struct rwlock *)lock;
186 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
187 if (rw->rw_lock & RW_LOCK_READ) {
198 owner_rw(const struct lock_object *lock, struct thread **owner)
200 const struct rwlock *rw = (const struct rwlock *)lock;
201 uintptr_t x = rw->rw_lock;
203 *owner = rw_wowner(rw);
204 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) :
210 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
217 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
218 RW_RECURSE | RW_NEW)) == 0);
219 ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
220 ("%s: rw_lock not aligned for %s: %p", __func__, name,
223 flags = LO_UPGRADABLE;
226 if (opts & RW_NOPROFILE)
227 flags |= LO_NOPROFILE;
228 if (!(opts & RW_NOWITNESS))
230 if (opts & RW_RECURSE)
231 flags |= LO_RECURSABLE;
237 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
238 rw->rw_lock = RW_UNLOCKED;
243 _rw_destroy(volatile uintptr_t *c)
249 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
250 KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
251 rw->rw_lock = RW_DESTROYED;
252 lock_destroy(&rw->lock_object);
256 rw_sysinit(void *arg)
258 struct rw_args *args = arg;
260 rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
264 rw_sysinit_flags(void *arg)
266 struct rw_args_flags *args = arg;
268 rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
273 _rw_wowned(const volatile uintptr_t *c)
276 return (rw_wowner(rwlock2rw(c)) == curthread);
280 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
284 if (SCHEDULER_STOPPED())
289 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
290 ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
291 curthread, rw->lock_object.lo_name, file, line));
292 KASSERT(rw->rw_lock != RW_DESTROYED,
293 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
294 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
296 __rw_wlock(rw, curthread, file, line);
297 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
298 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
299 TD_LOCKS_INC(curthread);
303 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
308 if (SCHEDULER_STOPPED())
313 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
314 ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
315 curthread, rw->lock_object.lo_name, file, line));
316 KASSERT(rw->rw_lock != RW_DESTROYED,
317 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
319 if (rw_wlocked(rw) &&
320 (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
324 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
325 (uintptr_t)curthread);
327 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
329 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
331 if (!rw_recursed(rw))
332 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
333 rw, 0, 0, file, line, LOCKSTAT_WRITER);
334 TD_LOCKS_INC(curthread);
340 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
344 if (SCHEDULER_STOPPED())
349 KASSERT(rw->rw_lock != RW_DESTROYED,
350 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
351 __rw_assert(c, RA_WLOCKED, file, line);
352 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
353 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
355 __rw_wunlock(rw, curthread, file, line);
356 TD_LOCKS_DEC(curthread);
360 * Determines whether a new reader can acquire a lock. Succeeds if the
361 * reader already owns a read lock and the lock is locked for read to
362 * prevent deadlock from reader recursion. Also succeeds if the lock
363 * is unlocked and has no writer waiters or spinners. Failing otherwise
364 * prioritizes writers before readers.
366 #define RW_CAN_READ(_rw) \
367 ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \
368 (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \
372 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
375 struct turnstile *ts;
376 #ifdef ADAPTIVE_RWLOCKS
377 volatile struct thread *owner;
381 #ifdef LOCK_PROFILING
382 uint64_t waittime = 0;
386 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
387 struct lock_delay_arg lda;
392 int64_t sleep_time = 0;
393 int64_t all_time = 0;
396 if (SCHEDULER_STOPPED())
399 #if defined(ADAPTIVE_RWLOCKS)
400 lock_delay_arg_init(&lda, &rw_delay);
401 #elif defined(KDTRACE_HOOKS)
402 lock_delay_arg_init(&lda, NULL);
406 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
407 ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
408 curthread, rw->lock_object.lo_name, file, line));
409 KASSERT(rw->rw_lock != RW_DESTROYED,
410 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
411 KASSERT(rw_wowner(rw) != curthread,
412 ("rw_rlock: wlock already held for %s @ %s:%d",
413 rw->lock_object.lo_name, file, line));
414 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
417 all_time -= lockstat_nsecs(&rw->lock_object);
422 * Handle the easy case. If no other thread has a write
423 * lock, then try to bump up the count of read locks. Note
424 * that we have to preserve the current state of the
425 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
426 * read lock, then rw_lock must have changed, so restart
427 * the loop. Note that this handles the case of a
428 * completely unlocked rwlock since such a lock is encoded
429 * as a read lock with no waiters.
432 if (RW_CAN_READ(v)) {
434 * The RW_LOCK_READ_WAITERS flag should only be set
435 * if the lock has been unlocked and write waiters
438 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
439 v + RW_ONE_READER)) {
440 if (LOCK_LOG_TEST(&rw->lock_object, 0))
442 "%s: %p succeed %p -> %p", __func__,
444 (void *)(v + RW_ONE_READER));
453 PMC_SOFT_CALL( , , lock, failed);
455 lock_profile_obtain_lock_failed(&rw->lock_object,
456 &contested, &waittime);
458 #ifdef ADAPTIVE_RWLOCKS
460 * If the owner is running on another CPU, spin until
461 * the owner stops running or the state of the lock
464 if ((v & RW_LOCK_READ) == 0) {
465 owner = (struct thread *)RW_OWNER(v);
466 if (TD_IS_RUNNING(owner)) {
467 if (LOCK_LOG_TEST(&rw->lock_object, 0))
469 "%s: spinning on %p held by %p",
470 __func__, rw, owner);
471 KTR_STATE1(KTR_SCHED, "thread",
472 sched_tdname(curthread), "spinning",
473 "lockname:\"%s\"", rw->lock_object.lo_name);
474 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
475 owner && TD_IS_RUNNING(owner))
477 KTR_STATE0(KTR_SCHED, "thread",
478 sched_tdname(curthread), "running");
481 } else if (spintries < rowner_retries) {
483 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
484 "spinning", "lockname:\"%s\"",
485 rw->lock_object.lo_name);
486 for (i = 0; i < rowner_loops; i++) {
488 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
493 lda.spin_cnt += rowner_loops - i;
495 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
497 if (i != rowner_loops)
503 * Okay, now it's the hard case. Some other thread already
504 * has a write lock or there are write waiters present,
505 * acquire the turnstile lock so we can begin the process
508 ts = turnstile_trywait(&rw->lock_object);
511 * The lock might have been released while we spun, so
512 * recheck its state and restart the loop if needed.
515 if (RW_CAN_READ(v)) {
516 turnstile_cancel(ts);
520 #ifdef ADAPTIVE_RWLOCKS
522 * The current lock owner might have started executing
523 * on another CPU (or the lock could have changed
524 * owners) while we were waiting on the turnstile
525 * chain lock. If so, drop the turnstile lock and try
528 if ((v & RW_LOCK_READ) == 0) {
529 owner = (struct thread *)RW_OWNER(v);
530 if (TD_IS_RUNNING(owner)) {
531 turnstile_cancel(ts);
538 * The lock is held in write mode or it already has waiters.
540 MPASS(!RW_CAN_READ(v));
543 * If the RW_LOCK_READ_WAITERS flag is already set, then
544 * we can go ahead and block. If it is not set then try
545 * to set it. If we fail to set it drop the turnstile
546 * lock and restart the loop.
548 if (!(v & RW_LOCK_READ_WAITERS)) {
549 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
550 v | RW_LOCK_READ_WAITERS)) {
551 turnstile_cancel(ts);
554 if (LOCK_LOG_TEST(&rw->lock_object, 0))
555 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
560 * We were unable to acquire the lock and the read waiters
561 * flag is set, so we must block on the turnstile.
563 if (LOCK_LOG_TEST(&rw->lock_object, 0))
564 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
567 sleep_time -= lockstat_nsecs(&rw->lock_object);
569 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
571 sleep_time += lockstat_nsecs(&rw->lock_object);
574 if (LOCK_LOG_TEST(&rw->lock_object, 0))
575 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
579 all_time += lockstat_nsecs(&rw->lock_object);
581 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
582 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
583 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
585 /* Record only the loops spinning and not sleeping. */
586 if (lda.spin_cnt > sleep_cnt)
587 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
588 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
589 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
592 * TODO: acquire "owner of record" here. Here be turnstile dragons
593 * however. turnstiles don't like owners changing between calls to
594 * turnstile_wait() currently.
596 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
597 waittime, file, line, LOCKSTAT_READER);
598 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
599 WITNESS_LOCK(&rw->lock_object, 0, file, line);
600 TD_LOCKS_INC(curthread);
601 curthread->td_rw_rlocks++;
605 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
610 if (SCHEDULER_STOPPED())
615 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
616 ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
617 curthread, rw->lock_object.lo_name, file, line));
621 KASSERT(rw->rw_lock != RW_DESTROYED,
622 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
623 if (!(x & RW_LOCK_READ))
625 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
626 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
628 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
629 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
630 rw, 0, 0, file, line, LOCKSTAT_READER);
631 TD_LOCKS_INC(curthread);
632 curthread->td_rw_rlocks++;
637 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
642 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
645 struct turnstile *ts;
646 uintptr_t x, v, queue;
648 if (SCHEDULER_STOPPED())
653 KASSERT(rw->rw_lock != RW_DESTROYED,
654 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
655 __rw_assert(c, RA_RLOCKED, file, line);
656 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
657 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
659 /* TODO: drop "owner of record" here. */
663 * See if there is more than one read lock held. If so,
664 * just drop one and return.
667 if (RW_READERS(x) > 1) {
668 if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
669 x - RW_ONE_READER)) {
670 if (LOCK_LOG_TEST(&rw->lock_object, 0))
672 "%s: %p succeeded %p -> %p",
673 __func__, rw, (void *)x,
674 (void *)(x - RW_ONE_READER));
680 * If there aren't any waiters for a write lock, then try
681 * to drop it quickly.
683 if (!(x & RW_LOCK_WAITERS)) {
684 MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
686 if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
688 if (LOCK_LOG_TEST(&rw->lock_object, 0))
689 CTR2(KTR_LOCK, "%s: %p last succeeded",
696 * Ok, we know we have waiters and we think we are the
697 * last reader, so grab the turnstile lock.
699 turnstile_chain_lock(&rw->lock_object);
700 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
701 MPASS(v & RW_LOCK_WAITERS);
704 * Try to drop our lock leaving the lock in a unlocked
707 * If you wanted to do explicit lock handoff you'd have to
708 * do it here. You'd also want to use turnstile_signal()
709 * and you'd have to handle the race where a higher
710 * priority thread blocks on the write lock before the
711 * thread you wakeup actually runs and have the new thread
712 * "steal" the lock. For now it's a lot simpler to just
713 * wakeup all of the waiters.
715 * As above, if we fail, then another thread might have
716 * acquired a read lock, so drop the turnstile lock and
720 if (v & RW_LOCK_WRITE_WAITERS) {
721 queue = TS_EXCLUSIVE_QUEUE;
722 x |= (v & RW_LOCK_READ_WAITERS);
724 queue = TS_SHARED_QUEUE;
725 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
727 turnstile_chain_unlock(&rw->lock_object);
730 if (LOCK_LOG_TEST(&rw->lock_object, 0))
731 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
735 * Ok. The lock is released and all that's left is to
736 * wake up the waiters. Note that the lock might not be
737 * free anymore, but in that case the writers will just
738 * block again if they run before the new lock holder(s)
741 ts = turnstile_lookup(&rw->lock_object);
743 turnstile_broadcast(ts, queue);
744 turnstile_unpend(ts, TS_SHARED_LOCK);
745 turnstile_chain_unlock(&rw->lock_object);
748 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
749 TD_LOCKS_DEC(curthread);
750 curthread->td_rw_rlocks--;
754 * This function is called when we are unable to obtain a write lock on the
755 * first try. This means that at least one other thread holds either a
756 * read or write lock.
759 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
763 struct turnstile *ts;
764 #ifdef ADAPTIVE_RWLOCKS
765 volatile struct thread *owner;
770 #ifdef LOCK_PROFILING
771 uint64_t waittime = 0;
774 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
775 struct lock_delay_arg lda;
780 int64_t sleep_time = 0;
781 int64_t all_time = 0;
784 if (SCHEDULER_STOPPED())
787 #if defined(ADAPTIVE_RWLOCKS)
788 lock_delay_arg_init(&lda, &rw_delay);
789 #elif defined(KDTRACE_HOOKS)
790 lock_delay_arg_init(&lda, NULL);
794 if (rw_wlocked(rw)) {
795 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
796 ("%s: recursing but non-recursive rw %s @ %s:%d\n",
797 __func__, rw->lock_object.lo_name, file, line));
799 if (LOCK_LOG_TEST(&rw->lock_object, 0))
800 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
804 if (LOCK_LOG_TEST(&rw->lock_object, 0))
805 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
806 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
809 all_time -= lockstat_nsecs(&rw->lock_object);
813 if (rw->rw_lock == RW_UNLOCKED && _rw_write_lock(rw, tid))
819 PMC_SOFT_CALL( , , lock, failed);
821 lock_profile_obtain_lock_failed(&rw->lock_object,
822 &contested, &waittime);
823 #ifdef ADAPTIVE_RWLOCKS
825 * If the lock is write locked and the owner is
826 * running on another CPU, spin until the owner stops
827 * running or the state of the lock changes.
830 owner = (struct thread *)RW_OWNER(v);
831 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
832 if (LOCK_LOG_TEST(&rw->lock_object, 0))
833 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
834 __func__, rw, owner);
835 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
836 "spinning", "lockname:\"%s\"",
837 rw->lock_object.lo_name);
838 while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
839 TD_IS_RUNNING(owner))
841 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
845 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
846 spintries < rowner_retries) {
847 if (!(v & RW_LOCK_WRITE_SPINNER)) {
848 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
849 v | RW_LOCK_WRITE_SPINNER)) {
854 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
855 "spinning", "lockname:\"%s\"",
856 rw->lock_object.lo_name);
857 for (i = 0; i < rowner_loops; i++) {
858 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
862 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
865 lda.spin_cnt += rowner_loops - i;
867 if (i != rowner_loops)
871 ts = turnstile_trywait(&rw->lock_object);
874 #ifdef ADAPTIVE_RWLOCKS
876 * The current lock owner might have started executing
877 * on another CPU (or the lock could have changed
878 * owners) while we were waiting on the turnstile
879 * chain lock. If so, drop the turnstile lock and try
882 if (!(v & RW_LOCK_READ)) {
883 owner = (struct thread *)RW_OWNER(v);
884 if (TD_IS_RUNNING(owner)) {
885 turnstile_cancel(ts);
891 * Check for the waiters flags about this rwlock.
892 * If the lock was released, without maintain any pending
893 * waiters queue, simply try to acquire it.
894 * If a pending waiters queue is present, claim the lock
895 * ownership and maintain the pending queue.
897 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
898 if ((v & ~x) == RW_UNLOCKED) {
899 x &= ~RW_LOCK_WRITE_SPINNER;
900 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
904 turnstile_cancel(ts);
907 turnstile_cancel(ts);
911 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
912 * set it. If we fail to set it, then loop back and try
915 if (!(v & RW_LOCK_WRITE_WAITERS)) {
916 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
917 v | RW_LOCK_WRITE_WAITERS)) {
918 turnstile_cancel(ts);
921 if (LOCK_LOG_TEST(&rw->lock_object, 0))
922 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
926 * We were unable to acquire the lock and the write waiters
927 * flag is set, so we must block on the turnstile.
929 if (LOCK_LOG_TEST(&rw->lock_object, 0))
930 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
933 sleep_time -= lockstat_nsecs(&rw->lock_object);
935 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
937 sleep_time += lockstat_nsecs(&rw->lock_object);
940 if (LOCK_LOG_TEST(&rw->lock_object, 0))
941 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
943 #ifdef ADAPTIVE_RWLOCKS
948 all_time += lockstat_nsecs(&rw->lock_object);
950 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
951 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
952 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
954 /* Record only the loops spinning and not sleeping. */
955 if (lda.spin_cnt > sleep_cnt)
956 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
957 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
958 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
960 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
961 waittime, file, line, LOCKSTAT_WRITER);
965 * This function is called if the first try at releasing a write lock failed.
966 * This means that one of the 2 waiter bits must be set indicating that at
967 * least one thread is waiting on this lock.
970 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
974 struct turnstile *ts;
978 if (SCHEDULER_STOPPED())
983 if (rw_wlocked(rw) && rw_recursed(rw)) {
985 if (LOCK_LOG_TEST(&rw->lock_object, 0))
986 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
990 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
991 ("%s: neither of the waiter flags are set", __func__));
993 if (LOCK_LOG_TEST(&rw->lock_object, 0))
994 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
996 turnstile_chain_lock(&rw->lock_object);
997 ts = turnstile_lookup(&rw->lock_object);
1001 * Use the same algo as sx locks for now. Prefer waking up shared
1002 * waiters if we have any over writers. This is probably not ideal.
1004 * 'v' is the value we are going to write back to rw_lock. If we
1005 * have waiters on both queues, we need to preserve the state of
1006 * the waiter flag for the queue we don't wake up. For now this is
1007 * hardcoded for the algorithm mentioned above.
1009 * In the case of both readers and writers waiting we wakeup the
1010 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
1011 * new writer comes in before a reader it will claim the lock up
1012 * above. There is probably a potential priority inversion in
1013 * there that could be worked around either by waking both queues
1014 * of waiters or doing some complicated lock handoff gymnastics.
1017 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1018 queue = TS_EXCLUSIVE_QUEUE;
1019 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1021 queue = TS_SHARED_QUEUE;
1023 /* Wake up all waiters for the specific queue. */
1024 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1025 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1026 queue == TS_SHARED_QUEUE ? "read" : "write");
1027 turnstile_broadcast(ts, queue);
1028 atomic_store_rel_ptr(&rw->rw_lock, v);
1029 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1030 turnstile_chain_unlock(&rw->lock_object);
1034 * Attempt to do a non-blocking upgrade from a read lock to a write
1035 * lock. This will only succeed if this thread holds a single read
1036 * lock. Returns true if the upgrade succeeded and false otherwise.
1039 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1042 uintptr_t v, x, tid;
1043 struct turnstile *ts;
1046 if (SCHEDULER_STOPPED())
1051 KASSERT(rw->rw_lock != RW_DESTROYED,
1052 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1053 __rw_assert(c, RA_RLOCKED, file, line);
1056 * Attempt to switch from one reader to a writer. If there
1057 * are any write waiters, then we will have to lock the
1058 * turnstile first to prevent races with another writer
1059 * calling turnstile_wait() before we have claimed this
1060 * turnstile. So, do the simple case of no waiters first.
1062 tid = (uintptr_t)curthread;
1066 if (RW_READERS(v) > 1)
1068 if (!(v & RW_LOCK_WAITERS)) {
1069 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1076 * Ok, we think we have waiters, so lock the turnstile.
1078 ts = turnstile_trywait(&rw->lock_object);
1080 if (RW_READERS(v) > 1) {
1081 turnstile_cancel(ts);
1085 * Try to switch from one reader to a writer again. This time
1086 * we honor the current state of the waiters flags.
1087 * If we obtain the lock with the flags set, then claim
1088 * ownership of the turnstile.
1090 x = rw->rw_lock & RW_LOCK_WAITERS;
1091 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1094 turnstile_claim(ts);
1096 turnstile_cancel(ts);
1099 turnstile_cancel(ts);
1101 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1103 curthread->td_rw_rlocks--;
1104 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1106 LOCKSTAT_RECORD0(rw__upgrade, rw);
1112 * Downgrade a write lock into a single read lock.
1115 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1118 struct turnstile *ts;
1122 if (SCHEDULER_STOPPED())
1127 KASSERT(rw->rw_lock != RW_DESTROYED,
1128 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1129 __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1131 if (rw_recursed(rw))
1132 panic("downgrade of a recursed lock");
1135 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1138 * Convert from a writer to a single reader. First we handle
1139 * the easy case with no waiters. If there are any waiters, we
1140 * lock the turnstile and "disown" the lock.
1142 tid = (uintptr_t)curthread;
1143 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1147 * Ok, we think we have waiters, so lock the turnstile so we can
1148 * read the waiter flags without any races.
1150 turnstile_chain_lock(&rw->lock_object);
1151 v = rw->rw_lock & RW_LOCK_WAITERS;
1152 rwait = v & RW_LOCK_READ_WAITERS;
1153 wwait = v & RW_LOCK_WRITE_WAITERS;
1154 MPASS(rwait | wwait);
1157 * Downgrade from a write lock while preserving waiters flag
1158 * and give up ownership of the turnstile.
1160 ts = turnstile_lookup(&rw->lock_object);
1163 v &= ~RW_LOCK_READ_WAITERS;
1164 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1166 * Wake other readers if there are no writers pending. Otherwise they
1167 * won't be able to acquire the lock anyway.
1169 if (rwait && !wwait) {
1170 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1171 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1173 turnstile_disown(ts);
1174 turnstile_chain_unlock(&rw->lock_object);
1176 curthread->td_rw_rlocks++;
1177 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1178 LOCKSTAT_RECORD0(rw__downgrade, rw);
1181 #ifdef INVARIANT_SUPPORT
1187 * In the non-WITNESS case, rw_assert() can only detect that at least
1188 * *some* thread owns an rlock, but it cannot guarantee that *this*
1189 * thread owns an rlock.
1192 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1194 const struct rwlock *rw;
1196 if (panicstr != NULL)
1203 case RA_LOCKED | RA_RECURSED:
1204 case RA_LOCKED | RA_NOTRECURSED:
1206 case RA_RLOCKED | RA_RECURSED:
1207 case RA_RLOCKED | RA_NOTRECURSED:
1209 witness_assert(&rw->lock_object, what, file, line);
1212 * If some other thread has a write lock or we have one
1213 * and are asserting a read lock, fail. Also, if no one
1214 * has a lock at all, fail.
1216 if (rw->rw_lock == RW_UNLOCKED ||
1217 (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1218 rw_wowner(rw) != curthread)))
1219 panic("Lock %s not %slocked @ %s:%d\n",
1220 rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1221 "read " : "", file, line);
1223 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1224 if (rw_recursed(rw)) {
1225 if (what & RA_NOTRECURSED)
1226 panic("Lock %s recursed @ %s:%d\n",
1227 rw->lock_object.lo_name, file,
1229 } else if (what & RA_RECURSED)
1230 panic("Lock %s not recursed @ %s:%d\n",
1231 rw->lock_object.lo_name, file, line);
1236 case RA_WLOCKED | RA_RECURSED:
1237 case RA_WLOCKED | RA_NOTRECURSED:
1238 if (rw_wowner(rw) != curthread)
1239 panic("Lock %s not exclusively locked @ %s:%d\n",
1240 rw->lock_object.lo_name, file, line);
1241 if (rw_recursed(rw)) {
1242 if (what & RA_NOTRECURSED)
1243 panic("Lock %s recursed @ %s:%d\n",
1244 rw->lock_object.lo_name, file, line);
1245 } else if (what & RA_RECURSED)
1246 panic("Lock %s not recursed @ %s:%d\n",
1247 rw->lock_object.lo_name, file, line);
1251 witness_assert(&rw->lock_object, what, file, line);
1254 * If we hold a write lock fail. We can't reliably check
1255 * to see if we hold a read lock or not.
1257 if (rw_wowner(rw) == curthread)
1258 panic("Lock %s exclusively locked @ %s:%d\n",
1259 rw->lock_object.lo_name, file, line);
1263 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1267 #endif /* INVARIANT_SUPPORT */
1271 db_show_rwlock(const struct lock_object *lock)
1273 const struct rwlock *rw;
1276 rw = (const struct rwlock *)lock;
1278 db_printf(" state: ");
1279 if (rw->rw_lock == RW_UNLOCKED)
1280 db_printf("UNLOCKED\n");
1281 else if (rw->rw_lock == RW_DESTROYED) {
1282 db_printf("DESTROYED\n");
1284 } else if (rw->rw_lock & RW_LOCK_READ)
1285 db_printf("RLOCK: %ju locks\n",
1286 (uintmax_t)(RW_READERS(rw->rw_lock)));
1289 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1290 td->td_tid, td->td_proc->p_pid, td->td_name);
1291 if (rw_recursed(rw))
1292 db_printf(" recursed: %u\n", rw->rw_recurse);
1294 db_printf(" waiters: ");
1295 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1296 case RW_LOCK_READ_WAITERS:
1297 db_printf("readers\n");
1299 case RW_LOCK_WRITE_WAITERS:
1300 db_printf("writers\n");
1302 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1303 db_printf("readers and writers\n");
1306 db_printf("none\n");