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;
149 static __read_frequently u_int asx_loops;
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,
162 sx_lock_delay_init(void *arg __unused)
165 lock_delay_default_init(&sx_delay);
167 asx_loops = max(10000, sx_delay.max);
169 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
173 assert_sx(const struct lock_object *lock, int what)
176 sx_assert((const struct sx *)lock, what);
180 lock_sx(struct lock_object *lock, uintptr_t how)
184 sx = (struct sx *)lock;
192 unlock_sx(struct lock_object *lock)
196 sx = (struct sx *)lock;
197 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
198 if (sx_xlocked(sx)) {
209 owner_sx(const struct lock_object *lock, struct thread **owner)
214 sx = (const struct sx *)lock;
217 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
218 ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
223 sx_sysinit(void *arg)
225 struct sx_args *sargs = arg;
227 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
231 sx_init_flags(struct sx *sx, const char *description, int opts)
235 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
236 SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
237 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
238 ("%s: sx_lock not aligned for %s: %p", __func__, description,
241 flags = LO_SLEEPABLE | LO_UPGRADABLE;
244 if (opts & SX_NOPROFILE)
245 flags |= LO_NOPROFILE;
246 if (!(opts & SX_NOWITNESS))
248 if (opts & SX_RECURSE)
249 flags |= LO_RECURSABLE;
255 flags |= opts & SX_NOADAPTIVE;
256 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
257 sx->sx_lock = SX_LOCK_UNLOCKED;
262 sx_destroy(struct sx *sx)
265 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
266 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
267 sx->sx_lock = SX_LOCK_DESTROYED;
268 lock_destroy(&sx->lock_object);
272 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
276 if (SCHEDULER_STOPPED())
279 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
280 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
281 curthread, sx->lock_object.lo_name, file, line));
285 KASSERT(x != SX_LOCK_DESTROYED,
286 ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
287 if (!(x & SX_LOCK_SHARED))
289 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
290 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
291 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
292 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
293 sx, 0, 0, file, line, LOCKSTAT_READER);
294 TD_LOCKS_INC(curthread);
295 curthread->td_sx_slocks++;
300 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
305 sx_try_slock_(struct sx *sx, const char *file, int line)
308 return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
312 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
317 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
318 !TD_IS_IDLETHREAD(curthread),
319 ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
320 curthread, sx->lock_object.lo_name, file, line));
321 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
322 ("sx_xlock() of destroyed sx @ %s:%d", file, line));
323 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
325 tid = (uintptr_t)curthread;
326 x = SX_LOCK_UNLOCKED;
327 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
328 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
330 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
331 0, 0, file, line, LOCKSTAT_WRITER);
333 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
335 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
336 TD_LOCKS_INC(curthread);
343 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
352 if (SCHEDULER_STOPPED_TD(td))
355 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
356 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
357 curthread, sx->lock_object.lo_name, file, line));
358 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
359 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
363 x = SX_LOCK_UNLOCKED;
365 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
367 if (x == SX_LOCK_UNLOCKED)
369 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
371 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
378 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
380 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
383 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
384 sx, 0, 0, file, line, LOCKSTAT_WRITER);
385 TD_LOCKS_INC(curthread);
392 sx_try_xlock_(struct sx *sx, const char *file, int line)
395 return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
399 _sx_xunlock(struct sx *sx, const char *file, int line)
402 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
403 ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
404 _sx_assert(sx, SA_XLOCKED, file, line);
405 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
406 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
409 _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
411 __sx_xunlock(sx, curthread, file, line);
413 TD_LOCKS_DEC(curthread);
417 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
418 * This will only succeed if this thread holds a single shared lock.
419 * Return 1 if if the upgrade succeed, 0 otherwise.
422 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
428 if (SCHEDULER_STOPPED())
431 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
432 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
433 _sx_assert(sx, SA_SLOCKED, file, line);
436 * Try to switch from one shared lock to an exclusive lock. We need
437 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
438 * we will wake up the exclusive waiters when we drop the lock.
441 x = SX_READ_VALUE(sx);
443 if (SX_SHARERS(x) > 1)
445 waiters = (x & SX_LOCK_WAITERS);
446 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
447 (uintptr_t)curthread | waiters)) {
452 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
454 curthread->td_sx_slocks--;
455 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
457 LOCKSTAT_RECORD0(sx__upgrade, sx);
463 sx_try_upgrade_(struct sx *sx, const char *file, int line)
466 return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
470 * Downgrade an unrecursed exclusive lock into a single shared lock.
473 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
478 if (SCHEDULER_STOPPED())
481 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
482 ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
483 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
486 panic("downgrade of a recursed lock");
489 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
492 * Try to switch from an exclusive lock with no shared waiters
493 * to one sharer with no shared waiters. If there are
494 * exclusive waiters, we don't need to lock the sleep queue so
495 * long as we preserve the flag. We do one quick try and if
496 * that fails we grab the sleepq lock to keep the flags from
497 * changing and do it the slow way.
499 * We have to lock the sleep queue if there are shared waiters
500 * so we can wake them up.
503 if (!(x & SX_LOCK_SHARED_WAITERS) &&
504 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
505 (x & SX_LOCK_EXCLUSIVE_WAITERS)))
509 * Lock the sleep queue so we can read the waiters bits
510 * without any races and wakeup any shared waiters.
512 sleepq_lock(&sx->lock_object);
515 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
516 * shared lock. If there are any shared waiters, wake them up.
520 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
521 (x & SX_LOCK_EXCLUSIVE_WAITERS));
522 if (x & SX_LOCK_SHARED_WAITERS)
523 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
525 sleepq_release(&sx->lock_object);
531 curthread->td_sx_slocks++;
532 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
533 LOCKSTAT_RECORD0(sx__downgrade, sx);
537 sx_downgrade_(struct sx *sx, const char *file, int line)
540 sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
545 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
548 if (x & SX_LOCK_WRITE_SPINNER)
552 *in_critical = false;
557 #define sx_drop_critical(x, in_critical, extra_work) do { } while(0)
561 * This function represents the so-called 'hard case' for sx_xlock
562 * operation. All 'easy case' failures are redirected to this. Note
563 * that ideally this would be a static function, but it needs to be
564 * accessible from at least sx.h.
567 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
572 volatile struct thread *owner;
573 u_int i, n, spintries = 0;
574 enum { READERS, WRITER } sleep_reason = READERS;
576 bool in_critical = false;
578 #ifdef LOCK_PROFILING
579 uint64_t waittime = 0;
583 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
584 struct lock_delay_arg lda;
588 int64_t sleep_time = 0;
589 int64_t all_time = 0;
591 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
593 int doing_lockprof = 0;
597 tid = (uintptr_t)curthread;
600 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
601 while (x == SX_LOCK_UNLOCKED) {
602 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
607 all_time -= lockstat_nsecs(&sx->lock_object);
611 #ifdef LOCK_PROFILING
617 if (SCHEDULER_STOPPED())
620 #if defined(ADAPTIVE_SX)
621 lock_delay_arg_init(&lda, &sx_delay);
622 #elif defined(KDTRACE_HOOKS)
623 lock_delay_arg_init(&lda, NULL);
626 if (__predict_false(x == SX_LOCK_UNLOCKED))
627 x = SX_READ_VALUE(sx);
629 /* If we already hold an exclusive lock, then recurse. */
630 if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
631 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
632 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
633 sx->lock_object.lo_name, file, line));
635 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
636 if (LOCK_LOG_TEST(&sx->lock_object, 0))
637 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
641 if (LOCK_LOG_TEST(&sx->lock_object, 0))
642 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
643 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
646 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
650 PMC_SOFT_CALL( , , lock, failed);
652 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
656 GIANT_SAVE(extra_work);
660 if (x == SX_LOCK_UNLOCKED) {
661 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
666 GIANT_SAVE(extra_work);
672 if (__predict_false(!adaptive))
675 * If the lock is write locked and the owner is
676 * running on another CPU, spin until the owner stops
677 * running or the state of the lock changes.
679 if ((x & SX_LOCK_SHARED) == 0) {
680 sx_drop_critical(x, &in_critical, &extra_work);
681 sleep_reason = WRITER;
682 owner = lv_sx_owner(x);
683 if (!TD_IS_RUNNING(owner))
685 if (LOCK_LOG_TEST(&sx->lock_object, 0))
686 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
687 __func__, sx, owner);
688 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
689 "spinning", "lockname:\"%s\"",
690 sx->lock_object.lo_name);
693 x = SX_READ_VALUE(sx);
694 owner = lv_sx_owner(x);
695 } while (owner != NULL && TD_IS_RUNNING(owner));
696 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
699 } else if (SX_SHARERS(x) > 0) {
700 sleep_reason = READERS;
701 if (spintries == asx_retries)
703 if (!(x & SX_LOCK_WRITE_SPINNER)) {
709 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
710 x | SX_LOCK_WRITE_SPINNER)) {
718 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
719 "spinning", "lockname:\"%s\"",
720 sx->lock_object.lo_name);
722 for (i = 0; i < asx_loops; i += n) {
724 x = SX_READ_VALUE(sx);
725 if (!(x & SX_LOCK_WRITE_SPINNER))
727 if (!(x & SX_LOCK_SHARED))
736 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
743 sleepq_lock(&sx->lock_object);
744 x = SX_READ_VALUE(sx);
748 * If the lock was released while spinning on the
749 * sleep queue chain lock, try again.
751 if (x == SX_LOCK_UNLOCKED) {
752 sleepq_release(&sx->lock_object);
753 sx_drop_critical(x, &in_critical, &extra_work);
759 * The current lock owner might have started executing
760 * on another CPU (or the lock could have changed
761 * owners) while we were waiting on the sleep queue
762 * chain lock. If so, drop the sleep queue lock and try
766 if (!(x & SX_LOCK_SHARED)) {
767 owner = (struct thread *)SX_OWNER(x);
768 if (TD_IS_RUNNING(owner)) {
769 sleepq_release(&sx->lock_object);
770 sx_drop_critical(x, &in_critical,
774 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
775 sleepq_release(&sx->lock_object);
776 sx_drop_critical(x, &in_critical, &extra_work);
783 * If an exclusive lock was released with both shared
784 * and exclusive waiters and a shared waiter hasn't
785 * woken up and acquired the lock yet, sx_lock will be
786 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
787 * If we see that value, try to acquire it once. Note
788 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
789 * as there are other exclusive waiters still. If we
790 * fail, restart the loop.
792 setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
793 if ((x & ~setx) == SX_LOCK_SHARED) {
794 setx &= ~SX_LOCK_WRITE_SPINNER;
795 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
797 sleepq_release(&sx->lock_object);
798 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
805 * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
806 * It is an invariant that when the bit is set, there is
807 * a writer ready to grab the lock. Thus clear the bit since
808 * we are going to sleep.
811 if ((x & SX_LOCK_WRITE_SPINNER) ||
812 !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
813 setx = x & ~SX_LOCK_WRITE_SPINNER;
814 setx |= SX_LOCK_EXCLUSIVE_WAITERS;
815 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
825 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
826 * than loop back and retry.
828 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
829 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
830 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
833 if (LOCK_LOG_TEST(&sx->lock_object, 0))
834 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
842 * Since we have been unable to acquire the exclusive
843 * lock and the exclusive waiters flag is set, we have
846 if (LOCK_LOG_TEST(&sx->lock_object, 0))
847 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
851 sleep_time -= lockstat_nsecs(&sx->lock_object);
853 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
854 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
855 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
856 if (!(opts & SX_INTERRUPTIBLE))
857 sleepq_wait(&sx->lock_object, 0);
859 error = sleepq_wait_sig(&sx->lock_object, 0);
861 sleep_time += lockstat_nsecs(&sx->lock_object);
865 if (LOCK_LOG_TEST(&sx->lock_object, 0))
867 "%s: interruptible sleep by %p suspended by signal",
871 if (LOCK_LOG_TEST(&sx->lock_object, 0))
872 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
874 x = SX_READ_VALUE(sx);
876 if (__predict_true(!extra_work))
883 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
884 if (__predict_true(!doing_lockprof))
888 all_time += lockstat_nsecs(&sx->lock_object);
890 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
891 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
892 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
893 if (lda.spin_cnt > sleep_cnt)
894 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
895 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
896 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
900 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
901 contested, waittime, file, line, LOCKSTAT_WRITER);
906 * This function represents the so-called 'hard case' for sx_xunlock
907 * operation. All 'easy case' failures are redirected to this. Note
908 * that ideally this would be a static function, but it needs to be
909 * accessible from at least sx.h.
912 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
915 int queue, wakeup_swapper;
917 if (SCHEDULER_STOPPED())
920 tid = (uintptr_t)curthread;
922 if (__predict_false(x == tid))
923 x = SX_READ_VALUE(sx);
925 MPASS(!(x & SX_LOCK_SHARED));
927 if (__predict_false(x & SX_LOCK_RECURSED)) {
928 /* The lock is recursed, unrecurse one level. */
929 if ((--sx->sx_recurse) == 0)
930 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
931 if (LOCK_LOG_TEST(&sx->lock_object, 0))
932 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
936 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
938 atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
941 if (LOCK_LOG_TEST(&sx->lock_object, 0))
942 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
944 sleepq_lock(&sx->lock_object);
945 x = SX_READ_VALUE(sx);
946 MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
949 * The wake up algorithm here is quite simple and probably not
950 * ideal. It gives precedence to shared waiters if they are
951 * present. For this condition, we have to preserve the
952 * state of the exclusive waiters flag.
953 * If interruptible sleeps left the shared queue empty avoid a
954 * starvation for the threads sleeping on the exclusive queue by giving
955 * them precedence and cleaning up the shared waiters bit anyway.
957 setx = SX_LOCK_UNLOCKED;
958 queue = SQ_SHARED_QUEUE;
959 if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
960 sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
961 queue = SQ_EXCLUSIVE_QUEUE;
962 setx |= (x & SX_LOCK_SHARED_WAITERS);
964 atomic_store_rel_ptr(&sx->sx_lock, setx);
966 /* Wake up all the waiters for the specific queue. */
967 if (LOCK_LOG_TEST(&sx->lock_object, 0))
968 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
969 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
972 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
974 sleepq_release(&sx->lock_object);
979 static bool __always_inline
980 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
983 if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
986 if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
991 static bool __always_inline
992 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
993 LOCK_FILE_LINE_ARG_DEF)
997 * If no other thread has an exclusive lock then try to bump up
998 * the count of sharers. Since we have to preserve the state
999 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
1000 * shared lock loop back and retry.
1002 while (__sx_can_read(td, *xp, fp)) {
1003 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1004 *xp + SX_ONE_SHARER)) {
1005 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1006 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1007 __func__, sx, (void *)*xp,
1008 (void *)(*xp + SX_ONE_SHARER));
1016 static int __noinline
1017 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1022 volatile struct thread *owner;
1023 u_int i, n, spintries = 0;
1026 #ifdef LOCK_PROFILING
1027 uint64_t waittime = 0;
1031 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1032 struct lock_delay_arg lda;
1034 #ifdef KDTRACE_HOOKS
1035 u_int sleep_cnt = 0;
1036 int64_t sleep_time = 0;
1037 int64_t all_time = 0;
1039 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1040 uintptr_t state = 0;
1046 #ifdef KDTRACE_HOOKS
1047 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1048 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1051 all_time -= lockstat_nsecs(&sx->lock_object);
1055 #ifdef LOCK_PROFILING
1060 if (SCHEDULER_STOPPED())
1063 #if defined(ADAPTIVE_SX)
1064 lock_delay_arg_init(&lda, &sx_delay);
1065 #elif defined(KDTRACE_HOOKS)
1066 lock_delay_arg_init(&lda, NULL);
1070 adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
1074 PMC_SOFT_CALL( , , lock, failed);
1076 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
1080 GIANT_SAVE(extra_work);
1084 * As with rwlocks, we don't make any attempt to try to block
1085 * shared locks once there is an exclusive waiter.
1088 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1091 GIANT_SAVE(extra_work);
1093 #ifdef KDTRACE_HOOKS
1098 if (__predict_false(!adaptive))
1102 * If the owner is running on another CPU, spin until
1103 * the owner stops running or the state of the lock
1106 if ((x & SX_LOCK_SHARED) == 0) {
1107 owner = lv_sx_owner(x);
1108 if (TD_IS_RUNNING(owner)) {
1109 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1111 "%s: spinning on %p held by %p",
1112 __func__, sx, owner);
1113 KTR_STATE1(KTR_SCHED, "thread",
1114 sched_tdname(curthread), "spinning",
1115 "lockname:\"%s\"", sx->lock_object.lo_name);
1118 x = SX_READ_VALUE(sx);
1119 owner = lv_sx_owner(x);
1120 } while (owner != NULL && TD_IS_RUNNING(owner));
1121 KTR_STATE0(KTR_SCHED, "thread",
1122 sched_tdname(curthread), "running");
1126 if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1127 MPASS(!__sx_can_read(td, x, false));
1129 x = SX_READ_VALUE(sx);
1132 if (spintries < asx_retries) {
1133 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1134 "spinning", "lockname:\"%s\"",
1135 sx->lock_object.lo_name);
1137 for (i = 0; i < asx_loops; i += n) {
1139 x = SX_READ_VALUE(sx);
1140 if (!(x & SX_LOCK_SHARED))
1145 if (__sx_can_read(td, x, false))
1148 #ifdef KDTRACE_HOOKS
1151 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1161 * Some other thread already has an exclusive lock, so
1162 * start the process of blocking.
1164 sleepq_lock(&sx->lock_object);
1165 x = SX_READ_VALUE(sx);
1167 if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1168 __sx_can_read(td, x, false)) {
1169 sleepq_release(&sx->lock_object);
1175 * If the owner is running on another CPU, spin until
1176 * the owner stops running or the state of the lock
1179 if (!(x & SX_LOCK_SHARED) && adaptive) {
1180 owner = (struct thread *)SX_OWNER(x);
1181 if (TD_IS_RUNNING(owner)) {
1182 sleepq_release(&sx->lock_object);
1183 x = SX_READ_VALUE(sx);
1190 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
1191 * fail to set it drop the sleep queue lock and loop
1194 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1195 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1196 x | SX_LOCK_SHARED_WAITERS))
1198 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1199 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1204 * Since we have been unable to acquire the shared lock,
1207 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1208 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1211 #ifdef KDTRACE_HOOKS
1212 sleep_time -= lockstat_nsecs(&sx->lock_object);
1214 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1215 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1216 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1217 if (!(opts & SX_INTERRUPTIBLE))
1218 sleepq_wait(&sx->lock_object, 0);
1220 error = sleepq_wait_sig(&sx->lock_object, 0);
1221 #ifdef KDTRACE_HOOKS
1222 sleep_time += lockstat_nsecs(&sx->lock_object);
1226 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1228 "%s: interruptible sleep by %p suspended by signal",
1232 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1233 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1235 x = SX_READ_VALUE(sx);
1237 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1238 if (__predict_true(!extra_work))
1241 #ifdef KDTRACE_HOOKS
1242 all_time += lockstat_nsecs(&sx->lock_object);
1244 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1245 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1246 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1247 if (lda.spin_cnt > sleep_cnt)
1248 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1249 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1250 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1254 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1255 contested, waittime, file, line, LOCKSTAT_READER);
1262 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1268 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1269 !TD_IS_IDLETHREAD(curthread),
1270 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1271 curthread, sx->lock_object.lo_name, file, line));
1272 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1273 ("sx_slock() of destroyed sx @ %s:%d", file, line));
1274 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1278 x = SX_READ_VALUE(sx);
1279 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1280 !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1281 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1283 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1286 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1287 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1288 TD_LOCKS_INC(curthread);
1294 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1297 return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1300 static bool __always_inline
1301 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1305 if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1306 if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1307 *xp - SX_ONE_SHARER)) {
1308 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1310 "%s: %p succeeded %p -> %p",
1311 __func__, sx, (void *)*xp,
1312 (void *)(*xp - SX_ONE_SHARER));
1323 static void __noinline
1324 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1325 LOCK_FILE_LINE_ARG_DEF)
1327 int wakeup_swapper = 0;
1328 uintptr_t setx, queue;
1330 if (SCHEDULER_STOPPED())
1333 if (_sx_sunlock_try(sx, td, &x))
1336 sleepq_lock(&sx->lock_object);
1337 x = SX_READ_VALUE(sx);
1339 if (_sx_sunlock_try(sx, td, &x))
1343 * Wake up semantic here is quite simple:
1344 * Just wake up all the exclusive waiters.
1345 * Note that the state of the lock could have changed,
1346 * so if it fails loop back and retry.
1348 setx = SX_LOCK_UNLOCKED;
1349 queue = SQ_SHARED_QUEUE;
1350 if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1351 setx |= (x & SX_LOCK_SHARED_WAITERS);
1352 queue = SQ_EXCLUSIVE_QUEUE;
1354 setx |= (x & SX_LOCK_WRITE_SPINNER);
1355 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1357 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1358 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1359 "exclusive queue", __func__, sx);
1360 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1365 sleepq_release(&sx->lock_object);
1369 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1373 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1378 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1379 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1380 _sx_assert(sx, SA_SLOCKED, file, line);
1381 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1382 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1385 x = SX_READ_VALUE(sx);
1386 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1387 !_sx_sunlock_try(sx, td, &x)))
1388 _sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1390 lock_profile_release_lock(&sx->lock_object);
1392 TD_LOCKS_DEC(curthread);
1396 _sx_sunlock(struct sx *sx, const char *file, int line)
1399 _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1402 #ifdef INVARIANT_SUPPORT
1408 * In the non-WITNESS case, sx_assert() can only detect that at least
1409 * *some* thread owns an slock, but it cannot guarantee that *this*
1410 * thread owns an slock.
1413 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1419 if (panicstr != NULL)
1423 case SA_SLOCKED | SA_NOTRECURSED:
1424 case SA_SLOCKED | SA_RECURSED:
1430 case SA_LOCKED | SA_NOTRECURSED:
1431 case SA_LOCKED | SA_RECURSED:
1433 witness_assert(&sx->lock_object, what, file, line);
1436 * If some other thread has an exclusive lock or we
1437 * have one and are asserting a shared lock, fail.
1438 * Also, if no one has a lock at all, fail.
1440 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1441 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1442 sx_xholder(sx) != curthread)))
1443 panic("Lock %s not %slocked @ %s:%d\n",
1444 sx->lock_object.lo_name, slocked ? "share " : "",
1447 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1448 if (sx_recursed(sx)) {
1449 if (what & SA_NOTRECURSED)
1450 panic("Lock %s recursed @ %s:%d\n",
1451 sx->lock_object.lo_name, file,
1453 } else if (what & SA_RECURSED)
1454 panic("Lock %s not recursed @ %s:%d\n",
1455 sx->lock_object.lo_name, file, line);
1460 case SA_XLOCKED | SA_NOTRECURSED:
1461 case SA_XLOCKED | SA_RECURSED:
1462 if (sx_xholder(sx) != curthread)
1463 panic("Lock %s not exclusively locked @ %s:%d\n",
1464 sx->lock_object.lo_name, file, line);
1465 if (sx_recursed(sx)) {
1466 if (what & SA_NOTRECURSED)
1467 panic("Lock %s recursed @ %s:%d\n",
1468 sx->lock_object.lo_name, file, line);
1469 } else if (what & SA_RECURSED)
1470 panic("Lock %s not recursed @ %s:%d\n",
1471 sx->lock_object.lo_name, file, line);
1475 witness_assert(&sx->lock_object, what, file, line);
1478 * If we hold an exclusve lock fail. We can't
1479 * reliably check to see if we hold a shared lock or
1482 if (sx_xholder(sx) == curthread)
1483 panic("Lock %s exclusively locked @ %s:%d\n",
1484 sx->lock_object.lo_name, file, line);
1488 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1492 #endif /* INVARIANT_SUPPORT */
1496 db_show_sx(const struct lock_object *lock)
1499 const struct sx *sx;
1501 sx = (const struct sx *)lock;
1503 db_printf(" state: ");
1504 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1505 db_printf("UNLOCKED\n");
1506 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1507 db_printf("DESTROYED\n");
1509 } else if (sx->sx_lock & SX_LOCK_SHARED)
1510 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1512 td = sx_xholder(sx);
1513 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1514 td->td_tid, td->td_proc->p_pid, td->td_name);
1515 if (sx_recursed(sx))
1516 db_printf(" recursed: %d\n", sx->sx_recurse);
1519 db_printf(" waiters: ");
1520 switch(sx->sx_lock &
1521 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1522 case SX_LOCK_SHARED_WAITERS:
1523 db_printf("shared\n");
1525 case SX_LOCK_EXCLUSIVE_WAITERS:
1526 db_printf("exclusive\n");
1528 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1529 db_printf("exclusive and shared\n");
1532 db_printf("none\n");
1537 * Check to see if a thread that is blocked on a sleep queue is actually
1538 * blocked on an sx lock. If so, output some details and return true.
1539 * If the lock has an exclusive owner, return that in *ownerp.
1542 sx_chain(struct thread *td, struct thread **ownerp)
1547 * Check to see if this thread is blocked on an sx lock.
1548 * First, we check the lock class. If that is ok, then we
1549 * compare the lock name against the wait message.
1552 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1553 sx->lock_object.lo_name != td->td_wmesg)
1556 /* We think we have an sx lock, so output some details. */
1557 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1558 *ownerp = sx_xholder(sx);
1559 if (sx->sx_lock & SX_LOCK_SHARED)
1560 db_printf("SLOCK (count %ju)\n",
1561 (uintmax_t)SX_SHARERS(sx->sx_lock));
1563 db_printf("XLOCK\n");