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;
568 if (SCHEDULER_STOPPED())
571 #if defined(ADAPTIVE_SX)
572 lock_delay_arg_init(&lda, &sx_delay);
573 #elif defined(KDTRACE_HOOKS)
574 lock_delay_arg_init(&lda, NULL);
577 if (__predict_false(x == SX_LOCK_UNLOCKED))
578 x = SX_READ_VALUE(sx);
580 /* If we already hold an exclusive lock, then recurse. */
581 if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
582 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
583 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
584 sx->lock_object.lo_name, file, line));
586 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
587 if (LOCK_LOG_TEST(&sx->lock_object, 0))
588 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
592 if (LOCK_LOG_TEST(&sx->lock_object, 0))
593 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
594 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
597 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
601 PMC_SOFT_CALL( , , lock, failed);
603 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
606 #ifdef LOCK_PROFILING
609 #elif defined(KDTRACE_HOOKS)
610 extra_work = lockstat_enabled;
611 if (__predict_false(extra_work)) {
612 all_time -= lockstat_nsecs(&sx->lock_object);
617 GIANT_SAVE(extra_work);
621 if (x == SX_LOCK_UNLOCKED) {
622 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
627 GIANT_SAVE(extra_work);
633 if (__predict_false(!adaptive))
636 * If the lock is write locked and the owner is
637 * running on another CPU, spin until the owner stops
638 * running or the state of the lock changes.
640 if ((x & SX_LOCK_SHARED) == 0) {
641 sleep_reason = WRITER;
642 owner = lv_sx_owner(x);
643 if (!TD_IS_RUNNING(owner))
645 if (LOCK_LOG_TEST(&sx->lock_object, 0))
646 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
647 __func__, sx, owner);
648 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
649 "spinning", "lockname:\"%s\"",
650 sx->lock_object.lo_name);
653 x = SX_READ_VALUE(sx);
654 owner = lv_sx_owner(x);
655 } while (owner != NULL && TD_IS_RUNNING(owner));
656 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
659 } else if (SX_SHARERS(x) > 0) {
660 sleep_reason = READERS;
661 if (spintries == asx_retries)
664 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
665 "spinning", "lockname:\"%s\"",
666 sx->lock_object.lo_name);
667 for (i = 0; i < asx_loops; i += n) {
670 x = SX_READ_VALUE(sx);
671 if ((x & SX_LOCK_SHARED) == 0 ||
678 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
685 sleepq_lock(&sx->lock_object);
686 x = SX_READ_VALUE(sx);
690 * If the lock was released while spinning on the
691 * sleep queue chain lock, try again.
693 if (x == SX_LOCK_UNLOCKED) {
694 sleepq_release(&sx->lock_object);
700 * The current lock owner might have started executing
701 * on another CPU (or the lock could have changed
702 * owners) while we were waiting on the sleep queue
703 * chain lock. If so, drop the sleep queue lock and try
707 if (!(x & SX_LOCK_SHARED)) {
708 owner = (struct thread *)SX_OWNER(x);
709 if (TD_IS_RUNNING(owner)) {
710 sleepq_release(&sx->lock_object);
713 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
714 sleepq_release(&sx->lock_object);
721 * If an exclusive lock was released with both shared
722 * and exclusive waiters and a shared waiter hasn't
723 * woken up and acquired the lock yet, sx_lock will be
724 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
725 * If we see that value, try to acquire it once. Note
726 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
727 * as there are other exclusive waiters still. If we
728 * fail, restart the loop.
730 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
731 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
732 tid | SX_LOCK_EXCLUSIVE_WAITERS))
734 sleepq_release(&sx->lock_object);
735 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
741 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
742 * than loop back and retry.
744 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
745 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
746 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
749 if (LOCK_LOG_TEST(&sx->lock_object, 0))
750 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
755 * Since we have been unable to acquire the exclusive
756 * lock and the exclusive waiters flag is set, we have
759 if (LOCK_LOG_TEST(&sx->lock_object, 0))
760 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
764 sleep_time -= lockstat_nsecs(&sx->lock_object);
766 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
767 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
768 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
769 if (!(opts & SX_INTERRUPTIBLE))
770 sleepq_wait(&sx->lock_object, 0);
772 error = sleepq_wait_sig(&sx->lock_object, 0);
774 sleep_time += lockstat_nsecs(&sx->lock_object);
778 if (LOCK_LOG_TEST(&sx->lock_object, 0))
780 "%s: interruptible sleep by %p suspended by signal",
784 if (LOCK_LOG_TEST(&sx->lock_object, 0))
785 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
787 x = SX_READ_VALUE(sx);
789 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
790 if (__predict_true(!extra_work))
794 all_time += lockstat_nsecs(&sx->lock_object);
796 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
797 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
798 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
799 if (lda.spin_cnt > sleep_cnt)
800 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
801 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
802 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
805 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
806 contested, waittime, file, line, LOCKSTAT_WRITER);
812 * This function represents the so-called 'hard case' for sx_xunlock
813 * operation. All 'easy case' failures are redirected to this. Note
814 * that ideally this would be a static function, but it needs to be
815 * accessible from at least sx.h.
818 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
821 int queue, wakeup_swapper;
823 if (SCHEDULER_STOPPED())
826 tid = (uintptr_t)curthread;
828 if (__predict_false(x == tid))
829 x = SX_READ_VALUE(sx);
831 MPASS(!(x & SX_LOCK_SHARED));
833 if (__predict_false(x & SX_LOCK_RECURSED)) {
834 /* The lock is recursed, unrecurse one level. */
835 if ((--sx->sx_recurse) == 0)
836 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
837 if (LOCK_LOG_TEST(&sx->lock_object, 0))
838 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
842 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
844 atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
847 if (LOCK_LOG_TEST(&sx->lock_object, 0))
848 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
850 sleepq_lock(&sx->lock_object);
851 x = SX_READ_VALUE(sx);
852 MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
855 * The wake up algorithm here is quite simple and probably not
856 * ideal. It gives precedence to shared waiters if they are
857 * present. For this condition, we have to preserve the
858 * state of the exclusive waiters flag.
859 * If interruptible sleeps left the shared queue empty avoid a
860 * starvation for the threads sleeping on the exclusive queue by giving
861 * them precedence and cleaning up the shared waiters bit anyway.
863 setx = SX_LOCK_UNLOCKED;
864 queue = SQ_EXCLUSIVE_QUEUE;
865 if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
866 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
867 queue = SQ_SHARED_QUEUE;
868 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
870 atomic_store_rel_ptr(&sx->sx_lock, setx);
872 /* Wake up all the waiters for the specific queue. */
873 if (LOCK_LOG_TEST(&sx->lock_object, 0))
874 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
875 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
878 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
880 sleepq_release(&sx->lock_object);
885 static bool __always_inline
886 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
890 * If no other thread has an exclusive lock then try to bump up
891 * the count of sharers. Since we have to preserve the state
892 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
893 * shared lock loop back and retry.
895 while (*xp & SX_LOCK_SHARED) {
896 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
897 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
898 *xp + SX_ONE_SHARER)) {
899 if (LOCK_LOG_TEST(&sx->lock_object, 0))
900 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
901 __func__, sx, (void *)*xp,
902 (void *)(*xp + SX_ONE_SHARER));
909 static int __noinline
910 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
914 volatile struct thread *owner;
917 #ifdef LOCK_PROFILING
918 uint64_t waittime = 0;
922 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
923 struct lock_delay_arg lda;
927 int64_t sleep_time = 0;
928 int64_t all_time = 0;
930 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
935 if (SCHEDULER_STOPPED())
938 #if defined(ADAPTIVE_SX)
939 lock_delay_arg_init(&lda, &sx_delay);
940 #elif defined(KDTRACE_HOOKS)
941 lock_delay_arg_init(&lda, NULL);
945 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
949 PMC_SOFT_CALL( , , lock, failed);
951 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
954 #ifdef LOCK_PROFILING
957 #elif defined(KDTRACE_HOOKS)
958 extra_work = lockstat_enabled;
959 if (__predict_false(extra_work)) {
960 all_time -= lockstat_nsecs(&sx->lock_object);
965 GIANT_SAVE(extra_work);
969 * As with rwlocks, we don't make any attempt to try to block
970 * shared locks once there is an exclusive waiter.
973 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
976 GIANT_SAVE(extra_work);
983 if (__predict_false(!adaptive))
986 * If the owner is running on another CPU, spin until
987 * the owner stops running or the state of the lock
990 owner = lv_sx_owner(x);
991 if (TD_IS_RUNNING(owner)) {
992 if (LOCK_LOG_TEST(&sx->lock_object, 0))
994 "%s: spinning on %p held by %p",
995 __func__, sx, owner);
996 KTR_STATE1(KTR_SCHED, "thread",
997 sched_tdname(curthread), "spinning",
998 "lockname:\"%s\"", sx->lock_object.lo_name);
1001 x = SX_READ_VALUE(sx);
1002 owner = lv_sx_owner(x);
1003 } while (owner != NULL && TD_IS_RUNNING(owner));
1004 KTR_STATE0(KTR_SCHED, "thread",
1005 sched_tdname(curthread), "running");
1012 * Some other thread already has an exclusive lock, so
1013 * start the process of blocking.
1015 sleepq_lock(&sx->lock_object);
1016 x = SX_READ_VALUE(sx);
1019 * The lock could have been released while we spun.
1020 * In this case loop back and retry.
1022 if (x & SX_LOCK_SHARED) {
1023 sleepq_release(&sx->lock_object);
1029 * If the owner is running on another CPU, spin until
1030 * the owner stops running or the state of the lock
1033 if (!(x & SX_LOCK_SHARED) && adaptive) {
1034 owner = (struct thread *)SX_OWNER(x);
1035 if (TD_IS_RUNNING(owner)) {
1036 sleepq_release(&sx->lock_object);
1037 x = SX_READ_VALUE(sx);
1044 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
1045 * fail to set it drop the sleep queue lock and loop
1048 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1049 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1050 x | SX_LOCK_SHARED_WAITERS))
1052 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1053 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1058 * Since we have been unable to acquire the shared lock,
1061 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1062 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1065 #ifdef KDTRACE_HOOKS
1066 sleep_time -= lockstat_nsecs(&sx->lock_object);
1068 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1069 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1070 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1071 if (!(opts & SX_INTERRUPTIBLE))
1072 sleepq_wait(&sx->lock_object, 0);
1074 error = sleepq_wait_sig(&sx->lock_object, 0);
1075 #ifdef KDTRACE_HOOKS
1076 sleep_time += lockstat_nsecs(&sx->lock_object);
1080 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1082 "%s: interruptible sleep by %p suspended by signal",
1086 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1087 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1089 x = SX_READ_VALUE(sx);
1091 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1092 if (__predict_true(!extra_work))
1095 #ifdef KDTRACE_HOOKS
1096 all_time += lockstat_nsecs(&sx->lock_object);
1098 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1099 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1100 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1101 if (lda.spin_cnt > sleep_cnt)
1102 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1103 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1104 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1107 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1108 contested, waittime, file, line, LOCKSTAT_READER);
1115 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1120 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1121 !TD_IS_IDLETHREAD(curthread),
1122 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1123 curthread, sx->lock_object.lo_name, file, line));
1124 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1125 ("sx_slock() of destroyed sx @ %s:%d", file, line));
1126 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1129 x = SX_READ_VALUE(sx);
1130 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1131 !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1132 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1134 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1137 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1138 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1139 TD_LOCKS_INC(curthread);
1145 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1148 return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1151 static bool __always_inline
1152 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1157 * We should never have sharers while at least one thread
1158 * holds a shared lock.
1160 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1161 ("%s: waiting sharers", __func__));
1164 * See if there is more than one shared lock held. If
1165 * so, just drop one and return.
1167 if (SX_SHARERS(*xp) > 1) {
1168 if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1169 *xp - SX_ONE_SHARER)) {
1170 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1172 "%s: %p succeeded %p -> %p",
1173 __func__, sx, (void *)*xp,
1174 (void *)(*xp - SX_ONE_SHARER));
1181 * If there aren't any waiters for an exclusive lock,
1182 * then try to drop it quickly.
1184 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1185 MPASS(*xp == SX_SHARERS_LOCK(1));
1186 *xp = SX_SHARERS_LOCK(1);
1187 if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1188 xp, SX_LOCK_UNLOCKED)) {
1189 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1190 CTR2(KTR_LOCK, "%s: %p last succeeded",
1201 static void __noinline
1202 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1204 int wakeup_swapper = 0;
1207 if (SCHEDULER_STOPPED())
1210 if (_sx_sunlock_try(sx, &x))
1214 * At this point, there should just be one sharer with
1215 * exclusive waiters.
1217 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1219 sleepq_lock(&sx->lock_object);
1220 x = SX_READ_VALUE(sx);
1222 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1223 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1224 if (_sx_sunlock_try(sx, &x))
1228 * Wake up semantic here is quite simple:
1229 * Just wake up all the exclusive waiters.
1230 * Note that the state of the lock could have changed,
1231 * so if it fails loop back and retry.
1233 setx = x - SX_ONE_SHARER;
1234 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1235 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1237 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1238 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1239 "exclusive queue", __func__, sx);
1240 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1241 0, SQ_EXCLUSIVE_QUEUE);
1244 sleepq_release(&sx->lock_object);
1248 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1252 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1256 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1257 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1258 _sx_assert(sx, SA_SLOCKED, file, line);
1259 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1260 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1262 x = SX_READ_VALUE(sx);
1263 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1264 !_sx_sunlock_try(sx, &x)))
1265 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1267 lock_profile_release_lock(&sx->lock_object);
1269 TD_LOCKS_DEC(curthread);
1273 _sx_sunlock(struct sx *sx, const char *file, int line)
1276 _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1279 #ifdef INVARIANT_SUPPORT
1285 * In the non-WITNESS case, sx_assert() can only detect that at least
1286 * *some* thread owns an slock, but it cannot guarantee that *this*
1287 * thread owns an slock.
1290 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1296 if (panicstr != NULL)
1300 case SA_SLOCKED | SA_NOTRECURSED:
1301 case SA_SLOCKED | SA_RECURSED:
1307 case SA_LOCKED | SA_NOTRECURSED:
1308 case SA_LOCKED | SA_RECURSED:
1310 witness_assert(&sx->lock_object, what, file, line);
1313 * If some other thread has an exclusive lock or we
1314 * have one and are asserting a shared lock, fail.
1315 * Also, if no one has a lock at all, fail.
1317 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1318 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1319 sx_xholder(sx) != curthread)))
1320 panic("Lock %s not %slocked @ %s:%d\n",
1321 sx->lock_object.lo_name, slocked ? "share " : "",
1324 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1325 if (sx_recursed(sx)) {
1326 if (what & SA_NOTRECURSED)
1327 panic("Lock %s recursed @ %s:%d\n",
1328 sx->lock_object.lo_name, file,
1330 } else if (what & SA_RECURSED)
1331 panic("Lock %s not recursed @ %s:%d\n",
1332 sx->lock_object.lo_name, file, line);
1337 case SA_XLOCKED | SA_NOTRECURSED:
1338 case SA_XLOCKED | SA_RECURSED:
1339 if (sx_xholder(sx) != curthread)
1340 panic("Lock %s not exclusively locked @ %s:%d\n",
1341 sx->lock_object.lo_name, file, line);
1342 if (sx_recursed(sx)) {
1343 if (what & SA_NOTRECURSED)
1344 panic("Lock %s recursed @ %s:%d\n",
1345 sx->lock_object.lo_name, file, line);
1346 } else if (what & SA_RECURSED)
1347 panic("Lock %s not recursed @ %s:%d\n",
1348 sx->lock_object.lo_name, file, line);
1352 witness_assert(&sx->lock_object, what, file, line);
1355 * If we hold an exclusve lock fail. We can't
1356 * reliably check to see if we hold a shared lock or
1359 if (sx_xholder(sx) == curthread)
1360 panic("Lock %s exclusively locked @ %s:%d\n",
1361 sx->lock_object.lo_name, file, line);
1365 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1369 #endif /* INVARIANT_SUPPORT */
1373 db_show_sx(const struct lock_object *lock)
1376 const struct sx *sx;
1378 sx = (const struct sx *)lock;
1380 db_printf(" state: ");
1381 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1382 db_printf("UNLOCKED\n");
1383 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1384 db_printf("DESTROYED\n");
1386 } else if (sx->sx_lock & SX_LOCK_SHARED)
1387 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1389 td = sx_xholder(sx);
1390 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1391 td->td_tid, td->td_proc->p_pid, td->td_name);
1392 if (sx_recursed(sx))
1393 db_printf(" recursed: %d\n", sx->sx_recurse);
1396 db_printf(" waiters: ");
1397 switch(sx->sx_lock &
1398 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1399 case SX_LOCK_SHARED_WAITERS:
1400 db_printf("shared\n");
1402 case SX_LOCK_EXCLUSIVE_WAITERS:
1403 db_printf("exclusive\n");
1405 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1406 db_printf("exclusive and shared\n");
1409 db_printf("none\n");
1414 * Check to see if a thread that is blocked on a sleep queue is actually
1415 * blocked on an sx lock. If so, output some details and return true.
1416 * If the lock has an exclusive owner, return that in *ownerp.
1419 sx_chain(struct thread *td, struct thread **ownerp)
1424 * Check to see if this thread is blocked on an sx lock.
1425 * First, we check the lock class. If that is ok, then we
1426 * compare the lock name against the wait message.
1429 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1430 sx->lock_object.lo_name != td->td_wmesg)
1433 /* We think we have an sx lock, so output some details. */
1434 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1435 *ownerp = sx_xholder(sx);
1436 if (sx->sx_lock & SX_LOCK_SHARED)
1437 db_printf("SLOCK (count %ju)\n",
1438 (uintmax_t)SX_SHARERS(sx->sx_lock));
1440 db_printf("XLOCK\n");