2 * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
3 * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice(s), this list of conditions and the following disclaimer as
11 * the first lines of this file unmodified other than the possible
12 * addition of one or more copyright notices.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice(s), this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
19 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
20 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
21 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
24 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
31 * Shared/exclusive locks. This implementation attempts to ensure
32 * deterministic lock granting behavior, so that slocks and xlocks are
35 * Priority propagation will not generally raise the priority of lock holders,
36 * so should not be relied upon in combination with sx locks.
40 #include "opt_hwpmc_hooks.h"
41 #include "opt_no_adaptive_sx.h"
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
46 #include <sys/param.h>
47 #include <sys/systm.h>
49 #include <sys/kernel.h>
52 #include <sys/mutex.h>
54 #include <sys/sched.h>
55 #include <sys/sleepqueue.h>
58 #include <sys/sysctl.h>
60 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
61 #include <machine/cpu.h>
68 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
72 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
75 #include <sys/pmckern.h>
76 PMC_SOFT_DECLARE( , , lock, failed);
79 /* Handy macros for sleep queues. */
80 #define SQ_EXCLUSIVE_QUEUE 0
81 #define SQ_SHARED_QUEUE 1
84 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We
85 * drop Giant anytime we have to sleep or if we adaptively spin.
87 #define GIANT_DECLARE \
89 WITNESS_SAVE_DECL(Giant) \
91 #define GIANT_SAVE() do { \
92 if (mtx_owned(&Giant)) { \
93 WITNESS_SAVE(&Giant.lock_object, Giant); \
94 while (mtx_owned(&Giant)) { \
101 #define GIANT_RESTORE() do { \
102 if (_giantcnt > 0) { \
103 mtx_assert(&Giant, MA_NOTOWNED); \
104 while (_giantcnt--) \
106 WITNESS_RESTORE(&Giant.lock_object, Giant); \
111 * Returns true if an exclusive lock is recursed. It assumes
112 * curthread currently has an exclusive lock.
114 #define sx_recursed(sx) ((sx)->sx_recurse != 0)
116 static void assert_sx(const struct lock_object *lock, int what);
118 static void db_show_sx(const struct lock_object *lock);
120 static void lock_sx(struct lock_object *lock, uintptr_t how);
122 static int owner_sx(const struct lock_object *lock, struct thread **owner);
124 static uintptr_t unlock_sx(struct lock_object *lock);
126 struct lock_class lock_class_sx = {
128 .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
129 .lc_assert = assert_sx,
131 .lc_ddb_show = db_show_sx,
134 .lc_unlock = unlock_sx,
136 .lc_owner = owner_sx,
141 #define _sx_assert(sx, what, file, line)
145 static u_int asx_retries = 10;
146 static u_int asx_loops = 10000;
147 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
148 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
149 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
151 static struct lock_delay_config sx_delay = {
158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_initial, CTLFLAG_RW, &sx_delay.initial,
160 SYSCTL_INT(_debug_sx, OID_AUTO, delay_step, CTLFLAG_RW, &sx_delay.step,
162 SYSCTL_INT(_debug_sx, OID_AUTO, delay_min, CTLFLAG_RW, &sx_delay.min,
164 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
168 sx_delay_sysinit(void *dummy)
171 sx_delay.initial = mp_ncpus * 25;
172 sx_delay.step = (mp_ncpus * 25) / 2;
173 sx_delay.min = mp_ncpus * 5;
174 sx_delay.max = mp_ncpus * 25 * 10;
176 LOCK_DELAY_SYSINIT(sx_delay_sysinit);
180 assert_sx(const struct lock_object *lock, int what)
183 sx_assert((const struct sx *)lock, what);
187 lock_sx(struct lock_object *lock, uintptr_t how)
191 sx = (struct sx *)lock;
199 unlock_sx(struct lock_object *lock)
203 sx = (struct sx *)lock;
204 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
205 if (sx_xlocked(sx)) {
216 owner_sx(const struct lock_object *lock, struct thread **owner)
218 const struct sx *sx = (const struct sx *)lock;
219 uintptr_t x = sx->sx_lock;
221 *owner = (struct thread *)SX_OWNER(x);
222 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
228 sx_sysinit(void *arg)
230 struct sx_args *sargs = arg;
232 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
236 sx_init_flags(struct sx *sx, const char *description, int opts)
240 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
241 SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
242 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
243 ("%s: sx_lock not aligned for %s: %p", __func__, description,
246 flags = LO_SLEEPABLE | LO_UPGRADABLE;
249 if (opts & SX_NOPROFILE)
250 flags |= LO_NOPROFILE;
251 if (!(opts & SX_NOWITNESS))
253 if (opts & SX_RECURSE)
254 flags |= LO_RECURSABLE;
260 flags |= opts & SX_NOADAPTIVE;
261 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
262 sx->sx_lock = SX_LOCK_UNLOCKED;
267 sx_destroy(struct sx *sx)
270 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
271 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
272 sx->sx_lock = SX_LOCK_DESTROYED;
273 lock_destroy(&sx->lock_object);
277 _sx_slock(struct sx *sx, int opts, const char *file, int line)
281 if (SCHEDULER_STOPPED())
283 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
284 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
285 curthread, sx->lock_object.lo_name, file, line));
286 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
287 ("sx_slock() of destroyed sx @ %s:%d", file, line));
288 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
289 error = __sx_slock(sx, opts, file, line);
291 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
292 WITNESS_LOCK(&sx->lock_object, 0, file, line);
293 TD_LOCKS_INC(curthread);
300 sx_try_slock_(struct sx *sx, const char *file, int line)
304 if (SCHEDULER_STOPPED())
307 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
308 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
309 curthread, sx->lock_object.lo_name, file, line));
313 KASSERT(x != SX_LOCK_DESTROYED,
314 ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
315 if (!(x & SX_LOCK_SHARED))
317 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
318 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
319 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
320 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
321 sx, 0, 0, file, line, LOCKSTAT_READER);
322 TD_LOCKS_INC(curthread);
327 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
332 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
336 if (SCHEDULER_STOPPED())
338 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
339 ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
340 curthread, sx->lock_object.lo_name, file, line));
341 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
342 ("sx_xlock() of destroyed sx @ %s:%d", file, line));
343 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
345 error = __sx_xlock(sx, curthread, opts, file, line);
347 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
349 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
350 TD_LOCKS_INC(curthread);
357 sx_try_xlock_(struct sx *sx, const char *file, int line)
361 if (SCHEDULER_STOPPED())
364 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
365 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
366 curthread, sx->lock_object.lo_name, file, line));
367 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
368 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
370 if (sx_xlocked(sx) &&
371 (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
373 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
376 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
377 (uintptr_t)curthread);
378 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
380 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
382 if (!sx_recursed(sx))
383 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
384 sx, 0, 0, file, line, LOCKSTAT_WRITER);
385 TD_LOCKS_INC(curthread);
392 _sx_sunlock(struct sx *sx, const char *file, int line)
395 if (SCHEDULER_STOPPED())
397 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
398 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
399 _sx_assert(sx, SA_SLOCKED, file, line);
400 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
401 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
402 __sx_sunlock(sx, file, line);
403 TD_LOCKS_DEC(curthread);
407 _sx_xunlock(struct sx *sx, const char *file, int line)
410 if (SCHEDULER_STOPPED())
412 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
413 ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
414 _sx_assert(sx, SA_XLOCKED, file, line);
415 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
416 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
418 __sx_xunlock(sx, curthread, file, line);
419 TD_LOCKS_DEC(curthread);
423 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
424 * This will only succeed if this thread holds a single shared lock.
425 * Return 1 if if the upgrade succeed, 0 otherwise.
428 sx_try_upgrade_(struct sx *sx, const char *file, int line)
433 if (SCHEDULER_STOPPED())
436 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
437 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
438 _sx_assert(sx, SA_SLOCKED, file, line);
441 * Try to switch from one shared lock to an exclusive lock. We need
442 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
443 * we will wake up the exclusive waiters when we drop the lock.
445 x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
446 success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
447 (uintptr_t)curthread | x);
448 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
450 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
452 LOCKSTAT_RECORD0(sx__upgrade, sx);
458 * Downgrade an unrecursed exclusive lock into a single shared lock.
461 sx_downgrade_(struct sx *sx, const char *file, int line)
466 if (SCHEDULER_STOPPED())
469 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
470 ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
471 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
474 panic("downgrade of a recursed lock");
477 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
480 * Try to switch from an exclusive lock with no shared waiters
481 * to one sharer with no shared waiters. If there are
482 * exclusive waiters, we don't need to lock the sleep queue so
483 * long as we preserve the flag. We do one quick try and if
484 * that fails we grab the sleepq lock to keep the flags from
485 * changing and do it the slow way.
487 * We have to lock the sleep queue if there are shared waiters
488 * so we can wake them up.
491 if (!(x & SX_LOCK_SHARED_WAITERS) &&
492 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
493 (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
494 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
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);
517 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
518 LOCKSTAT_RECORD0(sx__downgrade, sx);
525 * This function represents the so-called 'hard case' for sx_xlock
526 * operation. All 'easy case' failures are redirected to this. Note
527 * that ideally this would be a static function, but it needs to be
528 * accessible from at least sx.h.
531 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
536 volatile struct thread *owner;
537 u_int i, spintries = 0;
540 #ifdef LOCK_PROFILING
541 uint64_t waittime = 0;
545 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
546 struct lock_delay_arg lda;
551 int64_t sleep_time = 0;
552 int64_t all_time = 0;
555 if (SCHEDULER_STOPPED())
558 #if defined(ADAPTIVE_SX)
559 lock_delay_arg_init(&lda, &sx_delay);
560 #elif defined(KDTRACE_HOOKS)
561 lock_delay_arg_init(&lda, NULL);
564 /* If we already hold an exclusive lock, then recurse. */
565 if (sx_xlocked(sx)) {
566 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
567 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
568 sx->lock_object.lo_name, file, line));
570 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
571 if (LOCK_LOG_TEST(&sx->lock_object, 0))
572 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
576 if (LOCK_LOG_TEST(&sx->lock_object, 0))
577 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
578 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
581 all_time -= lockstat_nsecs(&sx->lock_object);
585 if (sx->sx_lock == SX_LOCK_UNLOCKED &&
586 atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
592 PMC_SOFT_CALL( , , lock, failed);
594 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
598 * If the lock is write locked and the owner is
599 * running on another CPU, spin until the owner stops
600 * running or the state of the lock changes.
603 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
604 if ((x & SX_LOCK_SHARED) == 0) {
606 owner = (struct thread *)x;
607 if (TD_IS_RUNNING(owner)) {
608 if (LOCK_LOG_TEST(&sx->lock_object, 0))
610 "%s: spinning on %p held by %p",
611 __func__, sx, owner);
612 KTR_STATE1(KTR_SCHED, "thread",
613 sched_tdname(curthread), "spinning",
615 sx->lock_object.lo_name);
617 while (SX_OWNER(sx->sx_lock) == x &&
618 TD_IS_RUNNING(owner))
620 KTR_STATE0(KTR_SCHED, "thread",
621 sched_tdname(curthread), "running");
624 } else if (SX_SHARERS(x) && spintries < asx_retries) {
625 KTR_STATE1(KTR_SCHED, "thread",
626 sched_tdname(curthread), "spinning",
627 "lockname:\"%s\"", sx->lock_object.lo_name);
630 for (i = 0; i < asx_loops; i++) {
631 if (LOCK_LOG_TEST(&sx->lock_object, 0))
633 "%s: shared spinning on %p with %u and %u",
634 __func__, sx, spintries, i);
636 if ((x & SX_LOCK_SHARED) == 0 ||
644 KTR_STATE0(KTR_SCHED, "thread",
645 sched_tdname(curthread), "running");
652 sleepq_lock(&sx->lock_object);
656 * If the lock was released while spinning on the
657 * sleep queue chain lock, try again.
659 if (x == SX_LOCK_UNLOCKED) {
660 sleepq_release(&sx->lock_object);
666 * The current lock owner might have started executing
667 * on another CPU (or the lock could have changed
668 * owners) while we were waiting on the sleep queue
669 * chain lock. If so, drop the sleep queue lock and try
672 if (!(x & SX_LOCK_SHARED) &&
673 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
674 owner = (struct thread *)SX_OWNER(x);
675 if (TD_IS_RUNNING(owner)) {
676 sleepq_release(&sx->lock_object);
683 * If an exclusive lock was released with both shared
684 * and exclusive waiters and a shared waiter hasn't
685 * woken up and acquired the lock yet, sx_lock will be
686 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
687 * If we see that value, try to acquire it once. Note
688 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
689 * as there are other exclusive waiters still. If we
690 * fail, restart the loop.
692 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
693 if (atomic_cmpset_acq_ptr(&sx->sx_lock,
694 SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
695 tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
696 sleepq_release(&sx->lock_object);
697 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
701 sleepq_release(&sx->lock_object);
706 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
707 * than loop back and retry.
709 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
710 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
711 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
712 sleepq_release(&sx->lock_object);
715 if (LOCK_LOG_TEST(&sx->lock_object, 0))
716 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
721 * Since we have been unable to acquire the exclusive
722 * lock and the exclusive waiters flag is set, we have
725 if (LOCK_LOG_TEST(&sx->lock_object, 0))
726 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
730 sleep_time -= lockstat_nsecs(&sx->lock_object);
733 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
734 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
735 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
736 if (!(opts & SX_INTERRUPTIBLE))
737 sleepq_wait(&sx->lock_object, 0);
739 error = sleepq_wait_sig(&sx->lock_object, 0);
741 sleep_time += lockstat_nsecs(&sx->lock_object);
745 if (LOCK_LOG_TEST(&sx->lock_object, 0))
747 "%s: interruptible sleep by %p suspended by signal",
751 if (LOCK_LOG_TEST(&sx->lock_object, 0))
752 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
756 all_time += lockstat_nsecs(&sx->lock_object);
758 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
759 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
760 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
761 if (lda.spin_cnt > sleep_cnt)
762 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
763 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
764 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
767 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
768 contested, waittime, file, line, LOCKSTAT_WRITER);
774 * This function represents the so-called 'hard case' for sx_xunlock
775 * operation. All 'easy case' failures are redirected to this. Note
776 * that ideally this would be a static function, but it needs to be
777 * accessible from at least sx.h.
780 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
783 int queue, wakeup_swapper;
785 if (SCHEDULER_STOPPED())
788 MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
790 /* If the lock is recursed, then unrecurse one level. */
791 if (sx_xlocked(sx) && sx_recursed(sx)) {
792 if ((--sx->sx_recurse) == 0)
793 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
794 if (LOCK_LOG_TEST(&sx->lock_object, 0))
795 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
798 MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
799 SX_LOCK_EXCLUSIVE_WAITERS));
800 if (LOCK_LOG_TEST(&sx->lock_object, 0))
801 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
803 sleepq_lock(&sx->lock_object);
804 x = SX_LOCK_UNLOCKED;
807 * The wake up algorithm here is quite simple and probably not
808 * ideal. It gives precedence to shared waiters if they are
809 * present. For this condition, we have to preserve the
810 * state of the exclusive waiters flag.
811 * If interruptible sleeps left the shared queue empty avoid a
812 * starvation for the threads sleeping on the exclusive queue by giving
813 * them precedence and cleaning up the shared waiters bit anyway.
815 if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
816 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
817 queue = SQ_SHARED_QUEUE;
818 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
820 queue = SQ_EXCLUSIVE_QUEUE;
822 /* Wake up all the waiters for the specific queue. */
823 if (LOCK_LOG_TEST(&sx->lock_object, 0))
824 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
825 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
827 atomic_store_rel_ptr(&sx->sx_lock, x);
828 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
830 sleepq_release(&sx->lock_object);
836 * This function represents the so-called 'hard case' for sx_slock
837 * operation. All 'easy case' failures are redirected to this. Note
838 * that ideally this would be a static function, but it needs to be
839 * accessible from at least sx.h.
842 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
846 volatile struct thread *owner;
848 #ifdef LOCK_PROFILING
849 uint64_t waittime = 0;
854 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
855 struct lock_delay_arg lda;
860 int64_t sleep_time = 0;
861 int64_t all_time = 0;
864 if (SCHEDULER_STOPPED())
867 #if defined(ADAPTIVE_SX)
868 lock_delay_arg_init(&lda, &sx_delay);
869 #elif defined(KDTRACE_HOOKS)
870 lock_delay_arg_init(&lda, NULL);
874 all_time -= lockstat_nsecs(&sx->lock_object);
878 * As with rwlocks, we don't make any attempt to try to block
879 * shared locks once there is an exclusive waiter.
888 * If no other thread has an exclusive lock then try to bump up
889 * the count of sharers. Since we have to preserve the state
890 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
891 * shared lock loop back and retry.
893 if (x & SX_LOCK_SHARED) {
894 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
895 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
896 x + SX_ONE_SHARER)) {
897 if (LOCK_LOG_TEST(&sx->lock_object, 0))
899 "%s: %p succeed %p -> %p", __func__,
901 (void *)(x + SX_ONE_SHARER));
907 PMC_SOFT_CALL( , , lock, failed);
909 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
914 * If the owner is running on another CPU, spin until
915 * the owner stops running or the state of the lock
918 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
920 owner = (struct thread *)x;
921 if (TD_IS_RUNNING(owner)) {
922 if (LOCK_LOG_TEST(&sx->lock_object, 0))
924 "%s: spinning on %p held by %p",
925 __func__, sx, owner);
926 KTR_STATE1(KTR_SCHED, "thread",
927 sched_tdname(curthread), "spinning",
928 "lockname:\"%s\"", sx->lock_object.lo_name);
930 while (SX_OWNER(sx->sx_lock) == x &&
931 TD_IS_RUNNING(owner))
933 KTR_STATE0(KTR_SCHED, "thread",
934 sched_tdname(curthread), "running");
941 * Some other thread already has an exclusive lock, so
942 * start the process of blocking.
944 sleepq_lock(&sx->lock_object);
948 * The lock could have been released while we spun.
949 * In this case loop back and retry.
951 if (x & SX_LOCK_SHARED) {
952 sleepq_release(&sx->lock_object);
958 * If the owner is running on another CPU, spin until
959 * the owner stops running or the state of the lock
962 if (!(x & SX_LOCK_SHARED) &&
963 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
964 owner = (struct thread *)SX_OWNER(x);
965 if (TD_IS_RUNNING(owner)) {
966 sleepq_release(&sx->lock_object);
973 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
974 * fail to set it drop the sleep queue lock and loop
977 if (!(x & SX_LOCK_SHARED_WAITERS)) {
978 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
979 x | SX_LOCK_SHARED_WAITERS)) {
980 sleepq_release(&sx->lock_object);
983 if (LOCK_LOG_TEST(&sx->lock_object, 0))
984 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
989 * Since we have been unable to acquire the shared lock,
992 if (LOCK_LOG_TEST(&sx->lock_object, 0))
993 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
997 sleep_time -= lockstat_nsecs(&sx->lock_object);
1000 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1001 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1002 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1003 if (!(opts & SX_INTERRUPTIBLE))
1004 sleepq_wait(&sx->lock_object, 0);
1006 error = sleepq_wait_sig(&sx->lock_object, 0);
1007 #ifdef KDTRACE_HOOKS
1008 sleep_time += lockstat_nsecs(&sx->lock_object);
1012 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1014 "%s: interruptible sleep by %p suspended by signal",
1018 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1019 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1022 #ifdef KDTRACE_HOOKS
1023 all_time += lockstat_nsecs(&sx->lock_object);
1025 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1026 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1027 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1028 if (lda.spin_cnt > sleep_cnt)
1029 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1030 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1031 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1034 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1035 contested, waittime, file, line, LOCKSTAT_READER);
1041 * This function represents the so-called 'hard case' for sx_sunlock
1042 * operation. All 'easy case' failures are redirected to this. Note
1043 * that ideally this would be a static function, but it needs to be
1044 * accessible from at least sx.h.
1047 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1052 if (SCHEDULER_STOPPED())
1059 * We should never have sharers while at least one thread
1060 * holds a shared lock.
1062 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1063 ("%s: waiting sharers", __func__));
1066 * See if there is more than one shared lock held. If
1067 * so, just drop one and return.
1069 if (SX_SHARERS(x) > 1) {
1070 if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1071 x - SX_ONE_SHARER)) {
1072 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1074 "%s: %p succeeded %p -> %p",
1075 __func__, sx, (void *)x,
1076 (void *)(x - SX_ONE_SHARER));
1083 * If there aren't any waiters for an exclusive lock,
1084 * then try to drop it quickly.
1086 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1087 MPASS(x == SX_SHARERS_LOCK(1));
1088 if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1089 SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1090 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1091 CTR2(KTR_LOCK, "%s: %p last succeeded",
1099 * At this point, there should just be one sharer with
1100 * exclusive waiters.
1102 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1104 sleepq_lock(&sx->lock_object);
1107 * Wake up semantic here is quite simple:
1108 * Just wake up all the exclusive waiters.
1109 * Note that the state of the lock could have changed,
1110 * so if it fails loop back and retry.
1112 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1113 SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1114 SX_LOCK_UNLOCKED)) {
1115 sleepq_release(&sx->lock_object);
1118 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1119 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1120 "exclusive queue", __func__, sx);
1121 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1122 0, SQ_EXCLUSIVE_QUEUE);
1123 sleepq_release(&sx->lock_object);
1130 #ifdef INVARIANT_SUPPORT
1136 * In the non-WITNESS case, sx_assert() can only detect that at least
1137 * *some* thread owns an slock, but it cannot guarantee that *this*
1138 * thread owns an slock.
1141 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1147 if (panicstr != NULL)
1151 case SA_SLOCKED | SA_NOTRECURSED:
1152 case SA_SLOCKED | SA_RECURSED:
1158 case SA_LOCKED | SA_NOTRECURSED:
1159 case SA_LOCKED | SA_RECURSED:
1161 witness_assert(&sx->lock_object, what, file, line);
1164 * If some other thread has an exclusive lock or we
1165 * have one and are asserting a shared lock, fail.
1166 * Also, if no one has a lock at all, fail.
1168 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1169 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1170 sx_xholder(sx) != curthread)))
1171 panic("Lock %s not %slocked @ %s:%d\n",
1172 sx->lock_object.lo_name, slocked ? "share " : "",
1175 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1176 if (sx_recursed(sx)) {
1177 if (what & SA_NOTRECURSED)
1178 panic("Lock %s recursed @ %s:%d\n",
1179 sx->lock_object.lo_name, file,
1181 } else if (what & SA_RECURSED)
1182 panic("Lock %s not recursed @ %s:%d\n",
1183 sx->lock_object.lo_name, file, line);
1188 case SA_XLOCKED | SA_NOTRECURSED:
1189 case SA_XLOCKED | SA_RECURSED:
1190 if (sx_xholder(sx) != curthread)
1191 panic("Lock %s not exclusively locked @ %s:%d\n",
1192 sx->lock_object.lo_name, file, line);
1193 if (sx_recursed(sx)) {
1194 if (what & SA_NOTRECURSED)
1195 panic("Lock %s recursed @ %s:%d\n",
1196 sx->lock_object.lo_name, file, line);
1197 } else if (what & SA_RECURSED)
1198 panic("Lock %s not recursed @ %s:%d\n",
1199 sx->lock_object.lo_name, file, line);
1203 witness_assert(&sx->lock_object, what, file, line);
1206 * If we hold an exclusve lock fail. We can't
1207 * reliably check to see if we hold a shared lock or
1210 if (sx_xholder(sx) == curthread)
1211 panic("Lock %s exclusively locked @ %s:%d\n",
1212 sx->lock_object.lo_name, file, line);
1216 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1220 #endif /* INVARIANT_SUPPORT */
1224 db_show_sx(const struct lock_object *lock)
1227 const struct sx *sx;
1229 sx = (const struct sx *)lock;
1231 db_printf(" state: ");
1232 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1233 db_printf("UNLOCKED\n");
1234 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1235 db_printf("DESTROYED\n");
1237 } else if (sx->sx_lock & SX_LOCK_SHARED)
1238 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1240 td = sx_xholder(sx);
1241 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1242 td->td_tid, td->td_proc->p_pid, td->td_name);
1243 if (sx_recursed(sx))
1244 db_printf(" recursed: %d\n", sx->sx_recurse);
1247 db_printf(" waiters: ");
1248 switch(sx->sx_lock &
1249 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1250 case SX_LOCK_SHARED_WAITERS:
1251 db_printf("shared\n");
1253 case SX_LOCK_EXCLUSIVE_WAITERS:
1254 db_printf("exclusive\n");
1256 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1257 db_printf("exclusive and shared\n");
1260 db_printf("none\n");
1265 * Check to see if a thread that is blocked on a sleep queue is actually
1266 * blocked on an sx lock. If so, output some details and return true.
1267 * If the lock has an exclusive owner, return that in *ownerp.
1270 sx_chain(struct thread *td, struct thread **ownerp)
1275 * Check to see if this thread is blocked on an sx lock.
1276 * First, we check the lock class. If that is ok, then we
1277 * compare the lock name against the wait message.
1280 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1281 sx->lock_object.lo_name != td->td_wmesg)
1284 /* We think we have an sx lock, so output some details. */
1285 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1286 *ownerp = sx_xholder(sx);
1287 if (sx->sx_lock & SX_LOCK_SHARED)
1288 db_printf("SLOCK (count %ju)\n",
1289 (uintmax_t)SX_SHARERS(sx->sx_lock));
1291 db_printf("XLOCK\n");