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);
550 while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
555 PMC_SOFT_CALL( , , lock, failed);
557 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
561 * If the lock is write locked and the owner is
562 * running on another CPU, spin until the owner stops
563 * running or the state of the lock changes.
566 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
567 if ((x & SX_LOCK_SHARED) == 0) {
569 owner = (struct thread *)x;
570 if (TD_IS_RUNNING(owner)) {
571 if (LOCK_LOG_TEST(&sx->lock_object, 0))
573 "%s: spinning on %p held by %p",
574 __func__, sx, owner);
575 KTR_STATE1(KTR_SCHED, "thread",
576 sched_tdname(curthread), "spinning",
578 sx->lock_object.lo_name);
580 while (SX_OWNER(sx->sx_lock) == x &&
581 TD_IS_RUNNING(owner)) {
587 KTR_STATE0(KTR_SCHED, "thread",
588 sched_tdname(curthread), "running");
591 } else if (SX_SHARERS(x) && spintries < asx_retries) {
592 KTR_STATE1(KTR_SCHED, "thread",
593 sched_tdname(curthread), "spinning",
594 "lockname:\"%s\"", sx->lock_object.lo_name);
597 for (i = 0; i < asx_loops; i++) {
598 if (LOCK_LOG_TEST(&sx->lock_object, 0))
600 "%s: shared spinning on %p with %u and %u",
601 __func__, sx, spintries, i);
603 if ((x & SX_LOCK_SHARED) == 0 ||
611 KTR_STATE0(KTR_SCHED, "thread",
612 sched_tdname(curthread), "running");
619 sleepq_lock(&sx->lock_object);
623 * If the lock was released while spinning on the
624 * sleep queue chain lock, try again.
626 if (x == SX_LOCK_UNLOCKED) {
627 sleepq_release(&sx->lock_object);
633 * The current lock owner might have started executing
634 * on another CPU (or the lock could have changed
635 * owners) while we were waiting on the sleep queue
636 * chain lock. If so, drop the sleep queue lock and try
639 if (!(x & SX_LOCK_SHARED) &&
640 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
641 owner = (struct thread *)SX_OWNER(x);
642 if (TD_IS_RUNNING(owner)) {
643 sleepq_release(&sx->lock_object);
650 * If an exclusive lock was released with both shared
651 * and exclusive waiters and a shared waiter hasn't
652 * woken up and acquired the lock yet, sx_lock will be
653 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
654 * If we see that value, try to acquire it once. Note
655 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
656 * as there are other exclusive waiters still. If we
657 * fail, restart the loop.
659 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
660 if (atomic_cmpset_acq_ptr(&sx->sx_lock,
661 SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
662 tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
663 sleepq_release(&sx->lock_object);
664 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
668 sleepq_release(&sx->lock_object);
673 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
674 * than loop back and retry.
676 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
677 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
678 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
679 sleepq_release(&sx->lock_object);
682 if (LOCK_LOG_TEST(&sx->lock_object, 0))
683 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
688 * Since we have been unable to acquire the exclusive
689 * lock and the exclusive waiters flag is set, we have
692 if (LOCK_LOG_TEST(&sx->lock_object, 0))
693 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
697 sleep_time -= lockstat_nsecs(&sx->lock_object);
700 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
701 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
702 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
703 if (!(opts & SX_INTERRUPTIBLE))
704 sleepq_wait(&sx->lock_object, 0);
706 error = sleepq_wait_sig(&sx->lock_object, 0);
708 sleep_time += lockstat_nsecs(&sx->lock_object);
712 if (LOCK_LOG_TEST(&sx->lock_object, 0))
714 "%s: interruptible sleep by %p suspended by signal",
718 if (LOCK_LOG_TEST(&sx->lock_object, 0))
719 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
723 all_time += lockstat_nsecs(&sx->lock_object);
725 LOCKSTAT_RECORD4(LS_SX_XLOCK_BLOCK, sx, sleep_time,
726 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
727 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
728 if (spin_cnt > sleep_cnt)
729 LOCKSTAT_RECORD4(LS_SX_XLOCK_SPIN, sx, all_time - sleep_time,
730 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
731 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
734 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx,
735 contested, waittime, file, line);
741 * This function represents the so-called 'hard case' for sx_xunlock
742 * operation. All 'easy case' failures are redirected to this. Note
743 * that ideally this would be a static function, but it needs to be
744 * accessible from at least sx.h.
747 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
750 int queue, wakeup_swapper;
752 if (SCHEDULER_STOPPED())
755 MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
757 /* If the lock is recursed, then unrecurse one level. */
758 if (sx_xlocked(sx) && sx_recursed(sx)) {
759 if ((--sx->sx_recurse) == 0)
760 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
761 if (LOCK_LOG_TEST(&sx->lock_object, 0))
762 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
765 MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
766 SX_LOCK_EXCLUSIVE_WAITERS));
767 if (LOCK_LOG_TEST(&sx->lock_object, 0))
768 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
770 sleepq_lock(&sx->lock_object);
771 x = SX_LOCK_UNLOCKED;
774 * The wake up algorithm here is quite simple and probably not
775 * ideal. It gives precedence to shared waiters if they are
776 * present. For this condition, we have to preserve the
777 * state of the exclusive waiters flag.
778 * If interruptible sleeps left the shared queue empty avoid a
779 * starvation for the threads sleeping on the exclusive queue by giving
780 * them precedence and cleaning up the shared waiters bit anyway.
782 if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
783 sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
784 queue = SQ_SHARED_QUEUE;
785 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
787 queue = SQ_EXCLUSIVE_QUEUE;
789 /* Wake up all the waiters for the specific queue. */
790 if (LOCK_LOG_TEST(&sx->lock_object, 0))
791 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
792 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
794 atomic_store_rel_ptr(&sx->sx_lock, x);
795 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
797 sleepq_release(&sx->lock_object);
803 * This function represents the so-called 'hard case' for sx_slock
804 * operation. All 'easy case' failures are redirected to this. Note
805 * that ideally this would be a static function, but it needs to be
806 * accessible from at least sx.h.
809 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
813 volatile struct thread *owner;
815 #ifdef LOCK_PROFILING
816 uint64_t waittime = 0;
823 uint64_t spin_cnt = 0;
824 uint64_t sleep_cnt = 0;
825 int64_t sleep_time = 0;
826 int64_t all_time = 0;
829 if (SCHEDULER_STOPPED())
834 all_time -= lockstat_nsecs(&sx->lock_object);
838 * As with rwlocks, we don't make any attempt to try to block
839 * shared locks once there is an exclusive waiter.
848 * If no other thread has an exclusive lock then try to bump up
849 * the count of sharers. Since we have to preserve the state
850 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
851 * shared lock loop back and retry.
853 if (x & SX_LOCK_SHARED) {
854 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
855 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
856 x + SX_ONE_SHARER)) {
857 if (LOCK_LOG_TEST(&sx->lock_object, 0))
859 "%s: %p succeed %p -> %p", __func__,
861 (void *)(x + SX_ONE_SHARER));
867 PMC_SOFT_CALL( , , lock, failed);
869 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
874 * If the owner is running on another CPU, spin until
875 * the owner stops running or the state of the lock
878 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
880 owner = (struct thread *)x;
881 if (TD_IS_RUNNING(owner)) {
882 if (LOCK_LOG_TEST(&sx->lock_object, 0))
884 "%s: spinning on %p held by %p",
885 __func__, sx, owner);
886 KTR_STATE1(KTR_SCHED, "thread",
887 sched_tdname(curthread), "spinning",
888 "lockname:\"%s\"", sx->lock_object.lo_name);
890 while (SX_OWNER(sx->sx_lock) == x &&
891 TD_IS_RUNNING(owner)) {
897 KTR_STATE0(KTR_SCHED, "thread",
898 sched_tdname(curthread), "running");
905 * Some other thread already has an exclusive lock, so
906 * start the process of blocking.
908 sleepq_lock(&sx->lock_object);
912 * The lock could have been released while we spun.
913 * In this case loop back and retry.
915 if (x & SX_LOCK_SHARED) {
916 sleepq_release(&sx->lock_object);
922 * If the owner is running on another CPU, spin until
923 * the owner stops running or the state of the lock
926 if (!(x & SX_LOCK_SHARED) &&
927 (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
928 owner = (struct thread *)SX_OWNER(x);
929 if (TD_IS_RUNNING(owner)) {
930 sleepq_release(&sx->lock_object);
937 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
938 * fail to set it drop the sleep queue lock and loop
941 if (!(x & SX_LOCK_SHARED_WAITERS)) {
942 if (!atomic_cmpset_ptr(&sx->sx_lock, x,
943 x | SX_LOCK_SHARED_WAITERS)) {
944 sleepq_release(&sx->lock_object);
947 if (LOCK_LOG_TEST(&sx->lock_object, 0))
948 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
953 * Since we have been unable to acquire the shared lock,
956 if (LOCK_LOG_TEST(&sx->lock_object, 0))
957 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
961 sleep_time -= lockstat_nsecs(&sx->lock_object);
964 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
965 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
966 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
967 if (!(opts & SX_INTERRUPTIBLE))
968 sleepq_wait(&sx->lock_object, 0);
970 error = sleepq_wait_sig(&sx->lock_object, 0);
972 sleep_time += lockstat_nsecs(&sx->lock_object);
976 if (LOCK_LOG_TEST(&sx->lock_object, 0))
978 "%s: interruptible sleep by %p suspended by signal",
982 if (LOCK_LOG_TEST(&sx->lock_object, 0))
983 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
987 all_time += lockstat_nsecs(&sx->lock_object);
989 LOCKSTAT_RECORD4(LS_SX_SLOCK_BLOCK, sx, sleep_time,
990 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
991 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
992 if (spin_cnt > sleep_cnt)
993 LOCKSTAT_RECORD4(LS_SX_SLOCK_SPIN, sx, all_time - sleep_time,
994 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
995 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
998 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx,
999 contested, waittime, file, line);
1005 * This function represents the so-called 'hard case' for sx_sunlock
1006 * operation. All 'easy case' failures are redirected to this. Note
1007 * that ideally this would be a static function, but it needs to be
1008 * accessible from at least sx.h.
1011 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1016 if (SCHEDULER_STOPPED())
1023 * We should never have sharers while at least one thread
1024 * holds a shared lock.
1026 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1027 ("%s: waiting sharers", __func__));
1030 * See if there is more than one shared lock held. If
1031 * so, just drop one and return.
1033 if (SX_SHARERS(x) > 1) {
1034 if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1035 x - SX_ONE_SHARER)) {
1036 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1038 "%s: %p succeeded %p -> %p",
1039 __func__, sx, (void *)x,
1040 (void *)(x - SX_ONE_SHARER));
1047 * If there aren't any waiters for an exclusive lock,
1048 * then try to drop it quickly.
1050 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1051 MPASS(x == SX_SHARERS_LOCK(1));
1052 if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1053 SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1054 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1055 CTR2(KTR_LOCK, "%s: %p last succeeded",
1063 * At this point, there should just be one sharer with
1064 * exclusive waiters.
1066 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1068 sleepq_lock(&sx->lock_object);
1071 * Wake up semantic here is quite simple:
1072 * Just wake up all the exclusive waiters.
1073 * Note that the state of the lock could have changed,
1074 * so if it fails loop back and retry.
1076 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1077 SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1078 SX_LOCK_UNLOCKED)) {
1079 sleepq_release(&sx->lock_object);
1082 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1083 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1084 "exclusive queue", __func__, sx);
1085 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1086 0, SQ_EXCLUSIVE_QUEUE);
1087 sleepq_release(&sx->lock_object);
1094 #ifdef INVARIANT_SUPPORT
1100 * In the non-WITNESS case, sx_assert() can only detect that at least
1101 * *some* thread owns an slock, but it cannot guarantee that *this*
1102 * thread owns an slock.
1105 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1111 if (panicstr != NULL)
1115 case SA_SLOCKED | SA_NOTRECURSED:
1116 case SA_SLOCKED | SA_RECURSED:
1122 case SA_LOCKED | SA_NOTRECURSED:
1123 case SA_LOCKED | SA_RECURSED:
1125 witness_assert(&sx->lock_object, what, file, line);
1128 * If some other thread has an exclusive lock or we
1129 * have one and are asserting a shared lock, fail.
1130 * Also, if no one has a lock at all, fail.
1132 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1133 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1134 sx_xholder(sx) != curthread)))
1135 panic("Lock %s not %slocked @ %s:%d\n",
1136 sx->lock_object.lo_name, slocked ? "share " : "",
1139 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1140 if (sx_recursed(sx)) {
1141 if (what & SA_NOTRECURSED)
1142 panic("Lock %s recursed @ %s:%d\n",
1143 sx->lock_object.lo_name, file,
1145 } else if (what & SA_RECURSED)
1146 panic("Lock %s not recursed @ %s:%d\n",
1147 sx->lock_object.lo_name, file, line);
1152 case SA_XLOCKED | SA_NOTRECURSED:
1153 case SA_XLOCKED | SA_RECURSED:
1154 if (sx_xholder(sx) != curthread)
1155 panic("Lock %s not exclusively locked @ %s:%d\n",
1156 sx->lock_object.lo_name, file, line);
1157 if (sx_recursed(sx)) {
1158 if (what & SA_NOTRECURSED)
1159 panic("Lock %s recursed @ %s:%d\n",
1160 sx->lock_object.lo_name, file, line);
1161 } else if (what & SA_RECURSED)
1162 panic("Lock %s not recursed @ %s:%d\n",
1163 sx->lock_object.lo_name, file, line);
1167 witness_assert(&sx->lock_object, what, file, line);
1170 * If we hold an exclusve lock fail. We can't
1171 * reliably check to see if we hold a shared lock or
1174 if (sx_xholder(sx) == curthread)
1175 panic("Lock %s exclusively locked @ %s:%d\n",
1176 sx->lock_object.lo_name, file, line);
1180 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1184 #endif /* INVARIANT_SUPPORT */
1188 db_show_sx(const struct lock_object *lock)
1191 const struct sx *sx;
1193 sx = (const struct sx *)lock;
1195 db_printf(" state: ");
1196 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1197 db_printf("UNLOCKED\n");
1198 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1199 db_printf("DESTROYED\n");
1201 } else if (sx->sx_lock & SX_LOCK_SHARED)
1202 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1204 td = sx_xholder(sx);
1205 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1206 td->td_tid, td->td_proc->p_pid, td->td_name);
1207 if (sx_recursed(sx))
1208 db_printf(" recursed: %d\n", sx->sx_recurse);
1211 db_printf(" waiters: ");
1212 switch(sx->sx_lock &
1213 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1214 case SX_LOCK_SHARED_WAITERS:
1215 db_printf("shared\n");
1217 case SX_LOCK_EXCLUSIVE_WAITERS:
1218 db_printf("exclusive\n");
1220 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1221 db_printf("exclusive and shared\n");
1224 db_printf("none\n");
1229 * Check to see if a thread that is blocked on a sleep queue is actually
1230 * blocked on an sx lock. If so, output some details and return true.
1231 * If the lock has an exclusive owner, return that in *ownerp.
1234 sx_chain(struct thread *td, struct thread **ownerp)
1239 * Check to see if this thread is blocked on an sx lock.
1240 * First, we check the lock class. If that is ok, then we
1241 * compare the lock name against the wait message.
1244 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1245 sx->lock_object.lo_name != td->td_wmesg)
1248 /* We think we have an sx lock, so output some details. */
1249 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1250 *ownerp = sx_xholder(sx);
1251 if (sx->sx_lock & SX_LOCK_SHARED)
1252 db_printf("SLOCK (count %ju)\n",
1253 (uintmax_t)SX_SHARERS(sx->sx_lock));
1255 db_printf("XLOCK\n");