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_kdtrace.h"
42 #include "opt_no_adaptive_sx.h"
44 #include <sys/cdefs.h>
45 __FBSDID("$FreeBSD$");
47 #include <sys/param.h>
48 #include <sys/systm.h>
52 #include <sys/mutex.h>
54 #include <sys/sched.h>
55 #include <sys/sleepqueue.h>
57 #include <sys/sysctl.h>
59 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
60 #include <machine/cpu.h>
67 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
71 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
74 #include <sys/pmckern.h>
75 PMC_SOFT_DECLARE( , , lock, failed);
78 /* Handy macros for sleep queues. */
79 #define SQ_EXCLUSIVE_QUEUE 0
80 #define SQ_SHARED_QUEUE 1
83 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We
84 * drop Giant anytime we have to sleep or if we adaptively spin.
86 #define GIANT_DECLARE \
88 WITNESS_SAVE_DECL(Giant) \
90 #define GIANT_SAVE() do { \
91 if (mtx_owned(&Giant)) { \
92 WITNESS_SAVE(&Giant.lock_object, Giant); \
93 while (mtx_owned(&Giant)) { \
100 #define GIANT_RESTORE() do { \
101 if (_giantcnt > 0) { \
102 mtx_assert(&Giant, MA_NOTOWNED); \
103 while (_giantcnt--) \
105 WITNESS_RESTORE(&Giant.lock_object, Giant); \
110 * Returns true if an exclusive lock is recursed. It assumes
111 * curthread currently has an exclusive lock.
113 #define sx_recurse lock_object.lo_data
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, "");
153 assert_sx(const struct lock_object *lock, int what)
156 sx_assert((const struct sx *)lock, what);
160 lock_sx(struct lock_object *lock, uintptr_t how)
164 sx = (struct sx *)lock;
172 unlock_sx(struct lock_object *lock)
176 sx = (struct sx *)lock;
177 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
178 if (sx_xlocked(sx)) {
189 owner_sx(const struct lock_object *lock, struct thread **owner)
191 const struct sx *sx = (const struct sx *)lock;
192 uintptr_t x = sx->sx_lock;
194 *owner = (struct thread *)SX_OWNER(x);
195 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
201 sx_sysinit(void *arg)
203 struct sx_args *sargs = arg;
205 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
209 sx_init_flags(struct sx *sx, const char *description, int opts)
213 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
214 SX_NOPROFILE | SX_NOADAPTIVE)) == 0);
215 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
216 ("%s: sx_lock not aligned for %s: %p", __func__, description,
219 flags = LO_SLEEPABLE | LO_UPGRADABLE;
222 if (opts & SX_NOPROFILE)
223 flags |= LO_NOPROFILE;
224 if (!(opts & SX_NOWITNESS))
226 if (opts & SX_RECURSE)
227 flags |= LO_RECURSABLE;
231 flags |= opts & SX_NOADAPTIVE;
232 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
233 sx->sx_lock = SX_LOCK_UNLOCKED;
238 sx_destroy(struct sx *sx)
241 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
242 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
243 sx->sx_lock = SX_LOCK_DESTROYED;
244 lock_destroy(&sx->lock_object);
248 _sx_slock(struct sx *sx, int opts, const char *file, int line)
252 if (SCHEDULER_STOPPED())
254 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
255 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
256 curthread, sx->lock_object.lo_name, file, line));
257 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
258 ("sx_slock() of destroyed sx @ %s:%d", file, line));
259 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
260 error = __sx_slock(sx, opts, file, line);
262 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
263 WITNESS_LOCK(&sx->lock_object, 0, file, line);
264 curthread->td_locks++;
271 sx_try_slock_(struct sx *sx, const char *file, int line)
275 if (SCHEDULER_STOPPED())
278 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
279 ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
280 curthread, sx->lock_object.lo_name, file, line));
284 KASSERT(x != SX_LOCK_DESTROYED,
285 ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
286 if (!(x & SX_LOCK_SHARED))
288 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
289 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
290 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
291 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE,
292 sx, 0, 0, file, line);
293 curthread->td_locks++;
298 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
303 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
307 if (SCHEDULER_STOPPED())
309 KASSERT(kdb_active != 0 || !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 error = __sx_xlock(sx, curthread, opts, file, line);
318 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
320 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
321 curthread->td_locks++;
328 sx_try_xlock_(struct sx *sx, const char *file, int line)
332 if (SCHEDULER_STOPPED())
335 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
336 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
337 curthread, sx->lock_object.lo_name, file, line));
338 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
339 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
341 if (sx_xlocked(sx) &&
342 (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
344 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
347 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
348 (uintptr_t)curthread);
349 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
351 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
353 if (!sx_recursed(sx))
354 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE,
355 sx, 0, 0, file, line);
356 curthread->td_locks++;
363 _sx_sunlock(struct sx *sx, const char *file, int line)
366 if (SCHEDULER_STOPPED())
368 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
369 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
370 _sx_assert(sx, SA_SLOCKED, file, line);
371 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
372 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
373 __sx_sunlock(sx, file, line);
374 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_SUNLOCK_RELEASE, sx);
375 curthread->td_locks--;
379 _sx_xunlock(struct sx *sx, const char *file, int line)
382 if (SCHEDULER_STOPPED())
384 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
385 ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
386 _sx_assert(sx, SA_XLOCKED, file, line);
387 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
388 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
390 if (!sx_recursed(sx))
391 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_XUNLOCK_RELEASE, sx);
392 __sx_xunlock(sx, curthread, file, line);
393 curthread->td_locks--;
397 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
398 * This will only succeed if this thread holds a single shared lock.
399 * Return 1 if if the upgrade succeed, 0 otherwise.
402 sx_try_upgrade_(struct sx *sx, const char *file, int line)
407 if (SCHEDULER_STOPPED())
410 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
411 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
412 _sx_assert(sx, SA_SLOCKED, file, line);
415 * Try to switch from one shared lock to an exclusive lock. We need
416 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
417 * we will wake up the exclusive waiters when we drop the lock.
419 x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
420 success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
421 (uintptr_t)curthread | x);
422 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
424 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
426 LOCKSTAT_RECORD0(LS_SX_TRYUPGRADE_UPGRADE, sx);
432 * Downgrade an unrecursed exclusive lock into a single shared lock.
435 sx_downgrade_(struct sx *sx, const char *file, int line)
440 if (SCHEDULER_STOPPED())
443 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
444 ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
445 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
448 panic("downgrade of a recursed lock");
451 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
454 * Try to switch from an exclusive lock with no shared waiters
455 * to one sharer with no shared waiters. If there are
456 * exclusive waiters, we don't need to lock the sleep queue so
457 * long as we preserve the flag. We do one quick try and if
458 * that fails we grab the sleepq lock to keep the flags from
459 * changing and do it the slow way.
461 * We have to lock the sleep queue if there are shared waiters
462 * so we can wake them up.
465 if (!(x & SX_LOCK_SHARED_WAITERS) &&
466 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
467 (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
468 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
473 * Lock the sleep queue so we can read the waiters bits
474 * without any races and wakeup any shared waiters.
476 sleepq_lock(&sx->lock_object);
479 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
480 * shared lock. If there are any shared waiters, wake them up.
484 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
485 (x & SX_LOCK_EXCLUSIVE_WAITERS));
486 if (x & SX_LOCK_SHARED_WAITERS)
487 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
489 sleepq_release(&sx->lock_object);
491 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
492 LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx);
499 * This function represents the so-called 'hard case' for sx_xlock
500 * operation. All 'easy case' failures are redirected to this. Note
501 * that ideally this would be a static function, but it needs to be
502 * accessible from at least sx.h.
505 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
510 volatile struct thread *owner;
511 u_int i, spintries = 0;
514 #ifdef LOCK_PROFILING
515 uint64_t waittime = 0;
521 uint64_t spin_cnt = 0;
522 uint64_t sleep_cnt = 0;
523 int64_t sleep_time = 0;
524 int64_t all_time = 0;
527 if (SCHEDULER_STOPPED())
530 /* If we already hold an exclusive lock, then recurse. */
531 if (sx_xlocked(sx)) {
532 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
533 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
534 sx->lock_object.lo_name, file, line));
536 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
537 if (LOCK_LOG_TEST(&sx->lock_object, 0))
538 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
542 if (LOCK_LOG_TEST(&sx->lock_object, 0))
543 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
544 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
547 all_time -= lockstat_nsecs(&sx->lock_object);
551 if (sx->sx_lock == SX_LOCK_UNLOCKED &&
552 atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
558 PMC_SOFT_CALL( , , lock, failed);
560 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
564 * If the lock is write locked and the owner is
565 * running on another CPU, spin until the owner stops
566 * running or the state of the lock changes.
569 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
570 if ((x & SX_LOCK_SHARED) == 0) {
572 owner = (struct thread *)x;
573 if (TD_IS_RUNNING(owner)) {
574 if (LOCK_LOG_TEST(&sx->lock_object, 0))
576 "%s: spinning on %p held by %p",
577 __func__, sx, owner);
578 KTR_STATE1(KTR_SCHED, "thread",
579 sched_tdname(curthread), "spinning",
581 sx->lock_object.lo_name);
583 while (SX_OWNER(sx->sx_lock) == x &&
584 TD_IS_RUNNING(owner)) {
590 KTR_STATE0(KTR_SCHED, "thread",
591 sched_tdname(curthread), "running");
594 } else if (SX_SHARERS(x) && spintries < asx_retries) {
595 KTR_STATE1(KTR_SCHED, "thread",
596 sched_tdname(curthread), "spinning",
597 "lockname:\"%s\"", sx->lock_object.lo_name);
600 for (i = 0; i < asx_loops; i++) {
601 if (LOCK_LOG_TEST(&sx->lock_object, 0))
603 "%s: shared spinning on %p with %u and %u",
604 __func__, sx, spintries, i);
606 if ((x & SX_LOCK_SHARED) == 0 ||
614 KTR_STATE0(KTR_SCHED, "thread",
615 sched_tdname(curthread), "running");
622 sleepq_lock(&sx->lock_object);
626 * If the lock was released while spinning on the
627 * sleep queue chain lock, try again.
629 if (x == SX_LOCK_UNLOCKED) {
630 sleepq_release(&sx->lock_object);
636 * The current lock owner might have started executing
637 * on another CPU (or the lock could have changed
638 * owners) while we were waiting on the sleep queue
639 * chain lock. If so, drop the sleep queue lock and try
642 if (!(x & SX_LOCK_SHARED) &&
643 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
644 owner = (struct thread *)SX_OWNER(x);
645 if (TD_IS_RUNNING(owner)) {
646 sleepq_release(&sx->lock_object);
653 * If an exclusive lock was released with both shared
654 * and exclusive waiters and a shared waiter hasn't
655 * woken up and acquired the lock yet, sx_lock will be
656 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
657 * If we see that value, try to acquire it once. Note
658 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
659 * as there are other exclusive waiters still. If we
660 * fail, restart the loop.
662 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
663 if (atomic_cmpset_acq_ptr(&sx->sx_lock,
664 SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
665 tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
666 sleepq_release(&sx->lock_object);
667 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
671 sleepq_release(&sx->lock_object);
676 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
677 * than loop back and retry.
679 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
680 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
681 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
682 sleepq_release(&sx->lock_object);
685 if (LOCK_LOG_TEST(&sx->lock_object, 0))
686 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
691 * Since we have been unable to acquire the exclusive
692 * lock and the exclusive waiters flag is set, we have
695 if (LOCK_LOG_TEST(&sx->lock_object, 0))
696 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
700 sleep_time -= lockstat_nsecs(&sx->lock_object);
703 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
704 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
705 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
706 if (!(opts & SX_INTERRUPTIBLE))
707 sleepq_wait(&sx->lock_object, 0);
709 error = sleepq_wait_sig(&sx->lock_object, 0);
711 sleep_time += lockstat_nsecs(&sx->lock_object);
715 if (LOCK_LOG_TEST(&sx->lock_object, 0))
717 "%s: interruptible sleep by %p suspended by signal",
721 if (LOCK_LOG_TEST(&sx->lock_object, 0))
722 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
726 all_time += lockstat_nsecs(&sx->lock_object);
728 LOCKSTAT_RECORD4(LS_SX_XLOCK_BLOCK, sx, sleep_time,
729 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
730 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
731 if (spin_cnt > sleep_cnt)
732 LOCKSTAT_RECORD4(LS_SX_XLOCK_SPIN, sx, all_time - sleep_time,
733 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
734 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
737 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx,
738 contested, waittime, file, line);
744 * This function represents the so-called 'hard case' for sx_xunlock
745 * operation. All 'easy case' failures are redirected to this. Note
746 * that ideally this would be a static function, but it needs to be
747 * accessible from at least sx.h.
750 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
753 int queue, wakeup_swapper;
755 if (SCHEDULER_STOPPED())
758 MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
760 /* If the lock is recursed, then unrecurse one level. */
761 if (sx_xlocked(sx) && sx_recursed(sx)) {
762 if ((--sx->sx_recurse) == 0)
763 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
764 if (LOCK_LOG_TEST(&sx->lock_object, 0))
765 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
768 MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
769 SX_LOCK_EXCLUSIVE_WAITERS));
770 if (LOCK_LOG_TEST(&sx->lock_object, 0))
771 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
773 sleepq_lock(&sx->lock_object);
774 x = SX_LOCK_UNLOCKED;
777 * The wake up algorithm here is quite simple and probably not
778 * ideal. It gives precedence to shared waiters if they are
779 * present. For this condition, we have to preserve the
780 * state of the exclusive waiters flag.
781 * If interruptible sleeps left the shared queue empty avoid a
782 * starvation for the threads sleeping on the exclusive queue by giving
783 * them precedence and cleaning up the shared waiters bit anyway.
785 if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
786 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
787 queue = SQ_SHARED_QUEUE;
788 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
790 queue = SQ_EXCLUSIVE_QUEUE;
792 /* Wake up all the waiters for the specific queue. */
793 if (LOCK_LOG_TEST(&sx->lock_object, 0))
794 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
795 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
797 atomic_store_rel_ptr(&sx->sx_lock, x);
798 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
800 sleepq_release(&sx->lock_object);
806 * This function represents the so-called 'hard case' for sx_slock
807 * operation. All 'easy case' failures are redirected to this. Note
808 * that ideally this would be a static function, but it needs to be
809 * accessible from at least sx.h.
812 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
816 volatile struct thread *owner;
818 #ifdef LOCK_PROFILING
819 uint64_t waittime = 0;
826 uint64_t spin_cnt = 0;
827 uint64_t sleep_cnt = 0;
828 int64_t sleep_time = 0;
829 int64_t all_time = 0;
832 if (SCHEDULER_STOPPED())
837 all_time -= lockstat_nsecs(&sx->lock_object);
841 * As with rwlocks, we don't make any attempt to try to block
842 * shared locks once there is an exclusive waiter.
851 * If no other thread has an exclusive lock then try to bump up
852 * the count of sharers. Since we have to preserve the state
853 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
854 * shared lock loop back and retry.
856 if (x & SX_LOCK_SHARED) {
857 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
858 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
859 x + SX_ONE_SHARER)) {
860 if (LOCK_LOG_TEST(&sx->lock_object, 0))
862 "%s: %p succeed %p -> %p", __func__,
864 (void *)(x + SX_ONE_SHARER));
870 PMC_SOFT_CALL( , , lock, failed);
872 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
877 * If the owner is running on another CPU, spin until
878 * the owner stops running or the state of the lock
881 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
883 owner = (struct thread *)x;
884 if (TD_IS_RUNNING(owner)) {
885 if (LOCK_LOG_TEST(&sx->lock_object, 0))
887 "%s: spinning on %p held by %p",
888 __func__, sx, owner);
889 KTR_STATE1(KTR_SCHED, "thread",
890 sched_tdname(curthread), "spinning",
891 "lockname:\"%s\"", sx->lock_object.lo_name);
893 while (SX_OWNER(sx->sx_lock) == x &&
894 TD_IS_RUNNING(owner)) {
900 KTR_STATE0(KTR_SCHED, "thread",
901 sched_tdname(curthread), "running");
908 * Some other thread already has an exclusive lock, so
909 * start the process of blocking.
911 sleepq_lock(&sx->lock_object);
915 * The lock could have been released while we spun.
916 * In this case loop back and retry.
918 if (x & SX_LOCK_SHARED) {
919 sleepq_release(&sx->lock_object);
925 * If the owner is running on another CPU, spin until
926 * the owner stops running or the state of the lock
929 if (!(x & SX_LOCK_SHARED) &&
930 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
931 owner = (struct thread *)SX_OWNER(x);
932 if (TD_IS_RUNNING(owner)) {
933 sleepq_release(&sx->lock_object);
940 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
941 * fail to set it drop the sleep queue lock and loop
944 if (!(x & SX_LOCK_SHARED_WAITERS)) {
945 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
946 x | SX_LOCK_SHARED_WAITERS)) {
947 sleepq_release(&sx->lock_object);
950 if (LOCK_LOG_TEST(&sx->lock_object, 0))
951 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
956 * Since we have been unable to acquire the shared lock,
959 if (LOCK_LOG_TEST(&sx->lock_object, 0))
960 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
964 sleep_time -= lockstat_nsecs(&sx->lock_object);
967 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
968 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
969 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
970 if (!(opts & SX_INTERRUPTIBLE))
971 sleepq_wait(&sx->lock_object, 0);
973 error = sleepq_wait_sig(&sx->lock_object, 0);
975 sleep_time += lockstat_nsecs(&sx->lock_object);
979 if (LOCK_LOG_TEST(&sx->lock_object, 0))
981 "%s: interruptible sleep by %p suspended by signal",
985 if (LOCK_LOG_TEST(&sx->lock_object, 0))
986 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
990 all_time += lockstat_nsecs(&sx->lock_object);
992 LOCKSTAT_RECORD4(LS_SX_SLOCK_BLOCK, sx, sleep_time,
993 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
994 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
995 if (spin_cnt > sleep_cnt)
996 LOCKSTAT_RECORD4(LS_SX_SLOCK_SPIN, sx, all_time - sleep_time,
997 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
998 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1001 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx,
1002 contested, waittime, file, line);
1008 * This function represents the so-called 'hard case' for sx_sunlock
1009 * operation. All 'easy case' failures are redirected to this. Note
1010 * that ideally this would be a static function, but it needs to be
1011 * accessible from at least sx.h.
1014 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1019 if (SCHEDULER_STOPPED())
1026 * We should never have sharers while at least one thread
1027 * holds a shared lock.
1029 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1030 ("%s: waiting sharers", __func__));
1033 * See if there is more than one shared lock held. If
1034 * so, just drop one and return.
1036 if (SX_SHARERS(x) > 1) {
1037 if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1038 x - SX_ONE_SHARER)) {
1039 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1041 "%s: %p succeeded %p -> %p",
1042 __func__, sx, (void *)x,
1043 (void *)(x - SX_ONE_SHARER));
1050 * If there aren't any waiters for an exclusive lock,
1051 * then try to drop it quickly.
1053 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1054 MPASS(x == SX_SHARERS_LOCK(1));
1055 if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1056 SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1057 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1058 CTR2(KTR_LOCK, "%s: %p last succeeded",
1066 * At this point, there should just be one sharer with
1067 * exclusive waiters.
1069 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1071 sleepq_lock(&sx->lock_object);
1074 * Wake up semantic here is quite simple:
1075 * Just wake up all the exclusive waiters.
1076 * Note that the state of the lock could have changed,
1077 * so if it fails loop back and retry.
1079 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1080 SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1081 SX_LOCK_UNLOCKED)) {
1082 sleepq_release(&sx->lock_object);
1085 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1086 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1087 "exclusive queue", __func__, sx);
1088 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1089 0, SQ_EXCLUSIVE_QUEUE);
1090 sleepq_release(&sx->lock_object);
1097 #ifdef INVARIANT_SUPPORT
1103 * In the non-WITNESS case, sx_assert() can only detect that at least
1104 * *some* thread owns an slock, but it cannot guarantee that *this*
1105 * thread owns an slock.
1108 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1114 if (panicstr != NULL)
1118 case SA_SLOCKED | SA_NOTRECURSED:
1119 case SA_SLOCKED | SA_RECURSED:
1125 case SA_LOCKED | SA_NOTRECURSED:
1126 case SA_LOCKED | SA_RECURSED:
1128 witness_assert(&sx->lock_object, what, file, line);
1131 * If some other thread has an exclusive lock or we
1132 * have one and are asserting a shared lock, fail.
1133 * Also, if no one has a lock at all, fail.
1135 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1136 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1137 sx_xholder(sx) != curthread)))
1138 panic("Lock %s not %slocked @ %s:%d\n",
1139 sx->lock_object.lo_name, slocked ? "share " : "",
1142 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1143 if (sx_recursed(sx)) {
1144 if (what & SA_NOTRECURSED)
1145 panic("Lock %s recursed @ %s:%d\n",
1146 sx->lock_object.lo_name, file,
1148 } else if (what & SA_RECURSED)
1149 panic("Lock %s not recursed @ %s:%d\n",
1150 sx->lock_object.lo_name, file, line);
1155 case SA_XLOCKED | SA_NOTRECURSED:
1156 case SA_XLOCKED | SA_RECURSED:
1157 if (sx_xholder(sx) != curthread)
1158 panic("Lock %s not exclusively locked @ %s:%d\n",
1159 sx->lock_object.lo_name, file, line);
1160 if (sx_recursed(sx)) {
1161 if (what & SA_NOTRECURSED)
1162 panic("Lock %s recursed @ %s:%d\n",
1163 sx->lock_object.lo_name, file, line);
1164 } else if (what & SA_RECURSED)
1165 panic("Lock %s not recursed @ %s:%d\n",
1166 sx->lock_object.lo_name, file, line);
1170 witness_assert(&sx->lock_object, what, file, line);
1173 * If we hold an exclusve lock fail. We can't
1174 * reliably check to see if we hold a shared lock or
1177 if (sx_xholder(sx) == curthread)
1178 panic("Lock %s exclusively locked @ %s:%d\n",
1179 sx->lock_object.lo_name, file, line);
1183 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1187 #endif /* INVARIANT_SUPPORT */
1191 db_show_sx(const struct lock_object *lock)
1194 const struct sx *sx;
1196 sx = (const struct sx *)lock;
1198 db_printf(" state: ");
1199 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1200 db_printf("UNLOCKED\n");
1201 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1202 db_printf("DESTROYED\n");
1204 } else if (sx->sx_lock & SX_LOCK_SHARED)
1205 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1207 td = sx_xholder(sx);
1208 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1209 td->td_tid, td->td_proc->p_pid, td->td_name);
1210 if (sx_recursed(sx))
1211 db_printf(" recursed: %d\n", sx->sx_recurse);
1214 db_printf(" waiters: ");
1215 switch(sx->sx_lock &
1216 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1217 case SX_LOCK_SHARED_WAITERS:
1218 db_printf("shared\n");
1220 case SX_LOCK_EXCLUSIVE_WAITERS:
1221 db_printf("exclusive\n");
1223 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1224 db_printf("exclusive and shared\n");
1227 db_printf("none\n");
1232 * Check to see if a thread that is blocked on a sleep queue is actually
1233 * blocked on an sx lock. If so, output some details and return true.
1234 * If the lock has an exclusive owner, return that in *ownerp.
1237 sx_chain(struct thread *td, struct thread **ownerp)
1242 * Check to see if this thread is blocked on an sx lock.
1243 * First, we check the lock class. If that is ok, then we
1244 * compare the lock name against the wait message.
1247 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1248 sx->lock_object.lo_name != td->td_wmesg)
1251 /* We think we have an sx lock, so output some details. */
1252 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1253 *ownerp = sx_xholder(sx);
1254 if (sx->sx_lock & SX_LOCK_SHARED)
1255 db_printf("SLOCK (count %ju)\n",
1256 (uintmax_t)SX_SHARERS(sx->sx_lock));
1258 db_printf("XLOCK\n");