2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
5 * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice(s), this list of conditions and the following disclaimer as
13 * the first lines of this file unmodified other than the possible
14 * addition of one or more copyright notices.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice(s), this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
20 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
23 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
24 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
25 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
26 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
33 * Shared/exclusive locks. This implementation attempts to ensure
34 * deterministic lock granting behavior, so that slocks and xlocks are
37 * Priority propagation will not generally raise the priority of lock holders,
38 * so should not be relied upon in combination with sx locks.
42 #include "opt_hwpmc_hooks.h"
43 #include "opt_no_adaptive_sx.h"
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
48 #include <sys/param.h>
49 #include <sys/systm.h>
51 #include <sys/kernel.h>
54 #include <sys/mutex.h>
56 #include <sys/sched.h>
57 #include <sys/sleepqueue.h>
60 #include <sys/sysctl.h>
62 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
63 #include <machine/cpu.h>
70 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
74 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
77 #include <sys/pmckern.h>
78 PMC_SOFT_DECLARE( , , lock, failed);
81 /* Handy macros for sleep queues. */
82 #define SQ_EXCLUSIVE_QUEUE 0
83 #define SQ_SHARED_QUEUE 1
86 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We
87 * drop Giant anytime we have to sleep or if we adaptively spin.
89 #define GIANT_DECLARE \
91 WITNESS_SAVE_DECL(Giant) \
93 #define GIANT_SAVE(work) do { \
94 if (__predict_false(mtx_owned(&Giant))) { \
96 WITNESS_SAVE(&Giant.lock_object, Giant); \
97 while (mtx_owned(&Giant)) { \
104 #define GIANT_RESTORE() do { \
105 if (_giantcnt > 0) { \
106 mtx_assert(&Giant, MA_NOTOWNED); \
107 while (_giantcnt--) \
109 WITNESS_RESTORE(&Giant.lock_object, Giant); \
114 * Returns true if an exclusive lock is recursed. It assumes
115 * curthread currently has an exclusive lock.
117 #define sx_recursed(sx) ((sx)->sx_recurse != 0)
119 static void assert_sx(const struct lock_object *lock, int what);
121 static void db_show_sx(const struct lock_object *lock);
123 static void lock_sx(struct lock_object *lock, uintptr_t how);
125 static int owner_sx(const struct lock_object *lock, struct thread **owner);
127 static uintptr_t unlock_sx(struct lock_object *lock);
129 struct lock_class lock_class_sx = {
131 .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
132 .lc_assert = assert_sx,
134 .lc_ddb_show = db_show_sx,
137 .lc_unlock = unlock_sx,
139 .lc_owner = owner_sx,
144 #define _sx_assert(sx, what, file, line)
148 static __read_frequently u_int asx_retries = 10;
149 static __read_frequently u_int asx_loops = 10000;
150 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
151 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
152 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
154 static struct lock_delay_config __read_frequently sx_delay;
156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
161 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
165 assert_sx(const struct lock_object *lock, int what)
168 sx_assert((const struct sx *)lock, what);
172 lock_sx(struct lock_object *lock, uintptr_t how)
176 sx = (struct sx *)lock;
184 unlock_sx(struct lock_object *lock)
188 sx = (struct sx *)lock;
189 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
190 if (sx_xlocked(sx)) {
201 owner_sx(const struct lock_object *lock, struct thread **owner)
206 sx = (const struct sx *)lock;
209 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
210 ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
215 sx_sysinit(void *arg)
217 struct sx_args *sargs = arg;
219 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
223 sx_init_flags(struct sx *sx, const char *description, int opts)
227 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
228 SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
229 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
230 ("%s: sx_lock not aligned for %s: %p", __func__, description,
233 flags = LO_SLEEPABLE | LO_UPGRADABLE;
236 if (opts & SX_NOPROFILE)
237 flags |= LO_NOPROFILE;
238 if (!(opts & SX_NOWITNESS))
240 if (opts & SX_RECURSE)
241 flags |= LO_RECURSABLE;
247 flags |= opts & SX_NOADAPTIVE;
248 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
249 sx->sx_lock = SX_LOCK_UNLOCKED;
254 sx_destroy(struct sx *sx)
257 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
258 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
259 sx->sx_lock = SX_LOCK_DESTROYED;
260 lock_destroy(&sx->lock_object);
264 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
268 if (SCHEDULER_STOPPED())
271 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
272 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
273 curthread, sx->lock_object.lo_name, file, line));
277 KASSERT(x != SX_LOCK_DESTROYED,
278 ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
279 if (!(x & SX_LOCK_SHARED))
281 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
282 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
283 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
284 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
285 sx, 0, 0, file, line, LOCKSTAT_READER);
286 TD_LOCKS_INC(curthread);
291 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
296 sx_try_slock_(struct sx *sx, const char *file, int line)
299 return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
303 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
308 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
309 !TD_IS_IDLETHREAD(curthread),
310 ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
311 curthread, sx->lock_object.lo_name, file, line));
312 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
313 ("sx_xlock() of destroyed sx @ %s:%d", file, line));
314 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
316 tid = (uintptr_t)curthread;
317 x = SX_LOCK_UNLOCKED;
318 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
319 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
321 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
322 0, 0, file, line, LOCKSTAT_WRITER);
324 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
326 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
327 TD_LOCKS_INC(curthread);
334 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
343 if (SCHEDULER_STOPPED_TD(td))
346 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
347 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
348 curthread, sx->lock_object.lo_name, file, line));
349 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
350 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
354 x = SX_LOCK_UNLOCKED;
356 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
358 if (x == SX_LOCK_UNLOCKED)
360 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
362 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
369 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
371 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
374 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
375 sx, 0, 0, file, line, LOCKSTAT_WRITER);
376 TD_LOCKS_INC(curthread);
383 sx_try_xlock_(struct sx *sx, const char *file, int line)
386 return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
390 _sx_xunlock(struct sx *sx, const char *file, int line)
393 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
394 ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
395 _sx_assert(sx, SA_XLOCKED, file, line);
396 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
397 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
400 _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
402 __sx_xunlock(sx, curthread, file, line);
404 TD_LOCKS_DEC(curthread);
408 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
409 * This will only succeed if this thread holds a single shared lock.
410 * Return 1 if if the upgrade succeed, 0 otherwise.
413 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
419 if (SCHEDULER_STOPPED())
422 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
423 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
424 _sx_assert(sx, SA_SLOCKED, file, line);
427 * Try to switch from one shared lock to an exclusive lock. We need
428 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
429 * we will wake up the exclusive waiters when we drop the lock.
432 x = SX_READ_VALUE(sx);
434 if (SX_SHARERS(x) > 1)
436 waiters = (x & SX_LOCK_EXCLUSIVE_WAITERS);
437 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
438 (uintptr_t)curthread | waiters)) {
443 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
445 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
447 LOCKSTAT_RECORD0(sx__upgrade, sx);
453 sx_try_upgrade_(struct sx *sx, const char *file, int line)
456 return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
460 * Downgrade an unrecursed exclusive lock into a single shared lock.
463 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
468 if (SCHEDULER_STOPPED())
471 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
472 ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
473 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
476 panic("downgrade of a recursed lock");
479 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
482 * Try to switch from an exclusive lock with no shared waiters
483 * to one sharer with no shared waiters. If there are
484 * exclusive waiters, we don't need to lock the sleep queue so
485 * long as we preserve the flag. We do one quick try and if
486 * that fails we grab the sleepq lock to keep the flags from
487 * changing and do it the slow way.
489 * We have to lock the sleep queue if there are shared waiters
490 * so we can wake them up.
493 if (!(x & SX_LOCK_SHARED_WAITERS) &&
494 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
495 (x & SX_LOCK_EXCLUSIVE_WAITERS)))
499 * Lock the sleep queue so we can read the waiters bits
500 * without any races and wakeup any shared waiters.
502 sleepq_lock(&sx->lock_object);
505 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
506 * shared lock. If there are any shared waiters, wake them up.
510 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
511 (x & SX_LOCK_EXCLUSIVE_WAITERS));
512 if (x & SX_LOCK_SHARED_WAITERS)
513 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
515 sleepq_release(&sx->lock_object);
521 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
522 LOCKSTAT_RECORD0(sx__downgrade, sx);
526 sx_downgrade_(struct sx *sx, const char *file, int line)
529 sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
533 * This function represents the so-called 'hard case' for sx_xlock
534 * operation. All 'easy case' failures are redirected to this. Note
535 * that ideally this would be a static function, but it needs to be
536 * accessible from at least sx.h.
539 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
544 volatile struct thread *owner;
545 u_int i, n, spintries = 0;
546 enum { READERS, WRITER } sleep_reason;
549 #ifdef LOCK_PROFILING
550 uint64_t waittime = 0;
554 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
555 struct lock_delay_arg lda;
559 int64_t sleep_time = 0;
560 int64_t all_time = 0;
562 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
567 tid = (uintptr_t)curthread;
570 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
571 while (x == SX_LOCK_UNLOCKED) {
572 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
576 all_time -= lockstat_nsecs(&sx->lock_object);
580 #ifdef LOCK_PROFILING
585 if (SCHEDULER_STOPPED())
588 #if defined(ADAPTIVE_SX)
589 lock_delay_arg_init(&lda, &sx_delay);
590 #elif defined(KDTRACE_HOOKS)
591 lock_delay_arg_init(&lda, NULL);
594 if (__predict_false(x == SX_LOCK_UNLOCKED))
595 x = SX_READ_VALUE(sx);
597 /* If we already hold an exclusive lock, then recurse. */
598 if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
599 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
600 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
601 sx->lock_object.lo_name, file, line));
603 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
604 if (LOCK_LOG_TEST(&sx->lock_object, 0))
605 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
609 if (LOCK_LOG_TEST(&sx->lock_object, 0))
610 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
611 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
614 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
618 PMC_SOFT_CALL( , , lock, failed);
620 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
624 GIANT_SAVE(extra_work);
628 if (x == SX_LOCK_UNLOCKED) {
629 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
634 GIANT_SAVE(extra_work);
640 if (__predict_false(!adaptive))
643 * If the lock is write locked and the owner is
644 * running on another CPU, spin until the owner stops
645 * running or the state of the lock changes.
647 if ((x & SX_LOCK_SHARED) == 0) {
648 sleep_reason = WRITER;
649 owner = lv_sx_owner(x);
650 if (!TD_IS_RUNNING(owner))
652 if (LOCK_LOG_TEST(&sx->lock_object, 0))
653 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
654 __func__, sx, owner);
655 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
656 "spinning", "lockname:\"%s\"",
657 sx->lock_object.lo_name);
660 x = SX_READ_VALUE(sx);
661 owner = lv_sx_owner(x);
662 } while (owner != NULL && TD_IS_RUNNING(owner));
663 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
666 } else if (SX_SHARERS(x) > 0) {
667 sleep_reason = READERS;
668 if (spintries == asx_retries)
671 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
672 "spinning", "lockname:\"%s\"",
673 sx->lock_object.lo_name);
674 for (i = 0; i < asx_loops; i += n) {
677 x = SX_READ_VALUE(sx);
678 if ((x & SX_LOCK_SHARED) == 0 ||
685 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
692 sleepq_lock(&sx->lock_object);
693 x = SX_READ_VALUE(sx);
697 * If the lock was released while spinning on the
698 * sleep queue chain lock, try again.
700 if (x == SX_LOCK_UNLOCKED) {
701 sleepq_release(&sx->lock_object);
707 * The current lock owner might have started executing
708 * on another CPU (or the lock could have changed
709 * owners) while we were waiting on the sleep queue
710 * chain lock. If so, drop the sleep queue lock and try
714 if (!(x & SX_LOCK_SHARED)) {
715 owner = (struct thread *)SX_OWNER(x);
716 if (TD_IS_RUNNING(owner)) {
717 sleepq_release(&sx->lock_object);
720 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
721 sleepq_release(&sx->lock_object);
728 * If an exclusive lock was released with both shared
729 * and exclusive waiters and a shared waiter hasn't
730 * woken up and acquired the lock yet, sx_lock will be
731 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
732 * If we see that value, try to acquire it once. Note
733 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
734 * as there are other exclusive waiters still. If we
735 * fail, restart the loop.
737 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
738 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
739 tid | SX_LOCK_EXCLUSIVE_WAITERS))
741 sleepq_release(&sx->lock_object);
742 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
748 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
749 * than loop back and retry.
751 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
752 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
753 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
756 if (LOCK_LOG_TEST(&sx->lock_object, 0))
757 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
762 * Since we have been unable to acquire the exclusive
763 * lock and the exclusive waiters flag is set, we have
766 if (LOCK_LOG_TEST(&sx->lock_object, 0))
767 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
771 sleep_time -= lockstat_nsecs(&sx->lock_object);
773 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
774 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
775 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
776 if (!(opts & SX_INTERRUPTIBLE))
777 sleepq_wait(&sx->lock_object, 0);
779 error = sleepq_wait_sig(&sx->lock_object, 0);
781 sleep_time += lockstat_nsecs(&sx->lock_object);
785 if (LOCK_LOG_TEST(&sx->lock_object, 0))
787 "%s: interruptible sleep by %p suspended by signal",
791 if (LOCK_LOG_TEST(&sx->lock_object, 0))
792 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
794 x = SX_READ_VALUE(sx);
796 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
797 if (__predict_true(!extra_work))
801 all_time += lockstat_nsecs(&sx->lock_object);
803 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
804 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
805 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
806 if (lda.spin_cnt > sleep_cnt)
807 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
808 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
809 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
813 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
814 contested, waittime, file, line, LOCKSTAT_WRITER);
820 * This function represents the so-called 'hard case' for sx_xunlock
821 * operation. All 'easy case' failures are redirected to this. Note
822 * that ideally this would be a static function, but it needs to be
823 * accessible from at least sx.h.
826 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
829 int queue, wakeup_swapper;
831 if (SCHEDULER_STOPPED())
834 tid = (uintptr_t)curthread;
836 if (__predict_false(x == tid))
837 x = SX_READ_VALUE(sx);
839 MPASS(!(x & SX_LOCK_SHARED));
841 if (__predict_false(x & SX_LOCK_RECURSED)) {
842 /* The lock is recursed, unrecurse one level. */
843 if ((--sx->sx_recurse) == 0)
844 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
845 if (LOCK_LOG_TEST(&sx->lock_object, 0))
846 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
850 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
852 atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
855 if (LOCK_LOG_TEST(&sx->lock_object, 0))
856 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
858 sleepq_lock(&sx->lock_object);
859 x = SX_READ_VALUE(sx);
860 MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
863 * The wake up algorithm here is quite simple and probably not
864 * ideal. It gives precedence to shared waiters if they are
865 * present. For this condition, we have to preserve the
866 * state of the exclusive waiters flag.
867 * If interruptible sleeps left the shared queue empty avoid a
868 * starvation for the threads sleeping on the exclusive queue by giving
869 * them precedence and cleaning up the shared waiters bit anyway.
871 setx = SX_LOCK_UNLOCKED;
872 queue = SQ_EXCLUSIVE_QUEUE;
873 if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
874 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
875 queue = SQ_SHARED_QUEUE;
876 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
878 atomic_store_rel_ptr(&sx->sx_lock, setx);
880 /* Wake up all the waiters for the specific queue. */
881 if (LOCK_LOG_TEST(&sx->lock_object, 0))
882 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
883 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
886 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
888 sleepq_release(&sx->lock_object);
893 static bool __always_inline
894 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
898 * If no other thread has an exclusive lock then try to bump up
899 * the count of sharers. Since we have to preserve the state
900 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
901 * shared lock loop back and retry.
903 while (*xp & SX_LOCK_SHARED) {
904 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
905 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
906 *xp + SX_ONE_SHARER)) {
907 if (LOCK_LOG_TEST(&sx->lock_object, 0))
908 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
909 __func__, sx, (void *)*xp,
910 (void *)(*xp + SX_ONE_SHARER));
917 static int __noinline
918 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
922 volatile struct thread *owner;
925 #ifdef LOCK_PROFILING
926 uint64_t waittime = 0;
930 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
931 struct lock_delay_arg lda;
935 int64_t sleep_time = 0;
936 int64_t all_time = 0;
938 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
944 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
945 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
948 all_time -= lockstat_nsecs(&sx->lock_object);
952 #ifdef LOCK_PROFILING
957 if (SCHEDULER_STOPPED())
960 #if defined(ADAPTIVE_SX)
961 lock_delay_arg_init(&lda, &sx_delay);
962 #elif defined(KDTRACE_HOOKS)
963 lock_delay_arg_init(&lda, NULL);
967 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
971 PMC_SOFT_CALL( , , lock, failed);
973 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
977 GIANT_SAVE(extra_work);
981 * As with rwlocks, we don't make any attempt to try to block
982 * shared locks once there is an exclusive waiter.
985 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
988 GIANT_SAVE(extra_work);
995 if (__predict_false(!adaptive))
998 * If the owner is running on another CPU, spin until
999 * the owner stops running or the state of the lock
1002 owner = lv_sx_owner(x);
1003 if (TD_IS_RUNNING(owner)) {
1004 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1006 "%s: spinning on %p held by %p",
1007 __func__, sx, owner);
1008 KTR_STATE1(KTR_SCHED, "thread",
1009 sched_tdname(curthread), "spinning",
1010 "lockname:\"%s\"", sx->lock_object.lo_name);
1013 x = SX_READ_VALUE(sx);
1014 owner = lv_sx_owner(x);
1015 } while (owner != NULL && TD_IS_RUNNING(owner));
1016 KTR_STATE0(KTR_SCHED, "thread",
1017 sched_tdname(curthread), "running");
1024 * Some other thread already has an exclusive lock, so
1025 * start the process of blocking.
1027 sleepq_lock(&sx->lock_object);
1028 x = SX_READ_VALUE(sx);
1031 * The lock could have been released while we spun.
1032 * In this case loop back and retry.
1034 if (x & SX_LOCK_SHARED) {
1035 sleepq_release(&sx->lock_object);
1041 * If the owner is running on another CPU, spin until
1042 * the owner stops running or the state of the lock
1045 if (!(x & SX_LOCK_SHARED) && adaptive) {
1046 owner = (struct thread *)SX_OWNER(x);
1047 if (TD_IS_RUNNING(owner)) {
1048 sleepq_release(&sx->lock_object);
1049 x = SX_READ_VALUE(sx);
1056 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
1057 * fail to set it drop the sleep queue lock and loop
1060 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1061 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1062 x | SX_LOCK_SHARED_WAITERS))
1064 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1065 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1070 * Since we have been unable to acquire the shared lock,
1073 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1074 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1077 #ifdef KDTRACE_HOOKS
1078 sleep_time -= lockstat_nsecs(&sx->lock_object);
1080 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1081 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1082 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1083 if (!(opts & SX_INTERRUPTIBLE))
1084 sleepq_wait(&sx->lock_object, 0);
1086 error = sleepq_wait_sig(&sx->lock_object, 0);
1087 #ifdef KDTRACE_HOOKS
1088 sleep_time += lockstat_nsecs(&sx->lock_object);
1092 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1094 "%s: interruptible sleep by %p suspended by signal",
1098 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1099 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1101 x = SX_READ_VALUE(sx);
1103 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1104 if (__predict_true(!extra_work))
1107 #ifdef KDTRACE_HOOKS
1108 all_time += lockstat_nsecs(&sx->lock_object);
1110 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1111 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1112 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1113 if (lda.spin_cnt > sleep_cnt)
1114 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1115 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1116 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1120 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1121 contested, waittime, file, line, LOCKSTAT_READER);
1128 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1133 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1134 !TD_IS_IDLETHREAD(curthread),
1135 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1136 curthread, sx->lock_object.lo_name, file, line));
1137 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1138 ("sx_slock() of destroyed sx @ %s:%d", file, line));
1139 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1142 x = SX_READ_VALUE(sx);
1143 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1144 !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1145 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1147 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1150 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1151 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1152 TD_LOCKS_INC(curthread);
1158 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1161 return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1164 static bool __always_inline
1165 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1170 * We should never have sharers while at least one thread
1171 * holds a shared lock.
1173 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1174 ("%s: waiting sharers", __func__));
1177 * See if there is more than one shared lock held. If
1178 * so, just drop one and return.
1180 if (SX_SHARERS(*xp) > 1) {
1181 if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1182 *xp - SX_ONE_SHARER)) {
1183 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1185 "%s: %p succeeded %p -> %p",
1186 __func__, sx, (void *)*xp,
1187 (void *)(*xp - SX_ONE_SHARER));
1194 * If there aren't any waiters for an exclusive lock,
1195 * then try to drop it quickly.
1197 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1198 MPASS(*xp == SX_SHARERS_LOCK(1));
1199 *xp = SX_SHARERS_LOCK(1);
1200 if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1201 xp, SX_LOCK_UNLOCKED)) {
1202 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1203 CTR2(KTR_LOCK, "%s: %p last succeeded",
1214 static void __noinline
1215 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1217 int wakeup_swapper = 0;
1220 if (SCHEDULER_STOPPED())
1223 if (_sx_sunlock_try(sx, &x))
1227 * At this point, there should just be one sharer with
1228 * exclusive waiters.
1230 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1232 sleepq_lock(&sx->lock_object);
1233 x = SX_READ_VALUE(sx);
1235 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1236 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1237 if (_sx_sunlock_try(sx, &x))
1241 * Wake up semantic here is quite simple:
1242 * Just wake up all the exclusive waiters.
1243 * Note that the state of the lock could have changed,
1244 * so if it fails loop back and retry.
1246 setx = x - SX_ONE_SHARER;
1247 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1248 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1250 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1251 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1252 "exclusive queue", __func__, sx);
1253 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1254 0, SQ_EXCLUSIVE_QUEUE);
1257 sleepq_release(&sx->lock_object);
1261 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1265 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1269 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1270 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1271 _sx_assert(sx, SA_SLOCKED, file, line);
1272 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1273 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1275 x = SX_READ_VALUE(sx);
1276 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1277 !_sx_sunlock_try(sx, &x)))
1278 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1280 lock_profile_release_lock(&sx->lock_object);
1282 TD_LOCKS_DEC(curthread);
1286 _sx_sunlock(struct sx *sx, const char *file, int line)
1289 _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1292 #ifdef INVARIANT_SUPPORT
1298 * In the non-WITNESS case, sx_assert() can only detect that at least
1299 * *some* thread owns an slock, but it cannot guarantee that *this*
1300 * thread owns an slock.
1303 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1309 if (panicstr != NULL)
1313 case SA_SLOCKED | SA_NOTRECURSED:
1314 case SA_SLOCKED | SA_RECURSED:
1320 case SA_LOCKED | SA_NOTRECURSED:
1321 case SA_LOCKED | SA_RECURSED:
1323 witness_assert(&sx->lock_object, what, file, line);
1326 * If some other thread has an exclusive lock or we
1327 * have one and are asserting a shared lock, fail.
1328 * Also, if no one has a lock at all, fail.
1330 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1331 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1332 sx_xholder(sx) != curthread)))
1333 panic("Lock %s not %slocked @ %s:%d\n",
1334 sx->lock_object.lo_name, slocked ? "share " : "",
1337 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1338 if (sx_recursed(sx)) {
1339 if (what & SA_NOTRECURSED)
1340 panic("Lock %s recursed @ %s:%d\n",
1341 sx->lock_object.lo_name, file,
1343 } else if (what & SA_RECURSED)
1344 panic("Lock %s not recursed @ %s:%d\n",
1345 sx->lock_object.lo_name, file, line);
1350 case SA_XLOCKED | SA_NOTRECURSED:
1351 case SA_XLOCKED | SA_RECURSED:
1352 if (sx_xholder(sx) != curthread)
1353 panic("Lock %s not exclusively locked @ %s:%d\n",
1354 sx->lock_object.lo_name, file, line);
1355 if (sx_recursed(sx)) {
1356 if (what & SA_NOTRECURSED)
1357 panic("Lock %s recursed @ %s:%d\n",
1358 sx->lock_object.lo_name, file, line);
1359 } else if (what & SA_RECURSED)
1360 panic("Lock %s not recursed @ %s:%d\n",
1361 sx->lock_object.lo_name, file, line);
1365 witness_assert(&sx->lock_object, what, file, line);
1368 * If we hold an exclusve lock fail. We can't
1369 * reliably check to see if we hold a shared lock or
1372 if (sx_xholder(sx) == curthread)
1373 panic("Lock %s exclusively locked @ %s:%d\n",
1374 sx->lock_object.lo_name, file, line);
1378 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1382 #endif /* INVARIANT_SUPPORT */
1386 db_show_sx(const struct lock_object *lock)
1389 const struct sx *sx;
1391 sx = (const struct sx *)lock;
1393 db_printf(" state: ");
1394 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1395 db_printf("UNLOCKED\n");
1396 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1397 db_printf("DESTROYED\n");
1399 } else if (sx->sx_lock & SX_LOCK_SHARED)
1400 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1402 td = sx_xholder(sx);
1403 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1404 td->td_tid, td->td_proc->p_pid, td->td_name);
1405 if (sx_recursed(sx))
1406 db_printf(" recursed: %d\n", sx->sx_recurse);
1409 db_printf(" waiters: ");
1410 switch(sx->sx_lock &
1411 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1412 case SX_LOCK_SHARED_WAITERS:
1413 db_printf("shared\n");
1415 case SX_LOCK_EXCLUSIVE_WAITERS:
1416 db_printf("exclusive\n");
1418 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1419 db_printf("exclusive and shared\n");
1422 db_printf("none\n");
1427 * Check to see if a thread that is blocked on a sleep queue is actually
1428 * blocked on an sx lock. If so, output some details and return true.
1429 * If the lock has an exclusive owner, return that in *ownerp.
1432 sx_chain(struct thread *td, struct thread **ownerp)
1437 * Check to see if this thread is blocked on an sx lock.
1438 * First, we check the lock class. If that is ok, then we
1439 * compare the lock name against the wait message.
1442 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1443 sx->lock_object.lo_name != td->td_wmesg)
1446 /* We think we have an sx lock, so output some details. */
1447 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1448 *ownerp = sx_xholder(sx);
1449 if (sx->sx_lock & SX_LOCK_SHARED)
1450 db_printf("SLOCK (count %ju)\n",
1451 (uintmax_t)SX_SHARERS(sx->sx_lock));
1453 db_printf("XLOCK\n");