2 * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice(s), this list of conditions and the following disclaimer as
10 * the first lines of this file unmodified other than the possible
11 * addition of one or more copyright notices.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice(s), this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
29 #include "opt_adaptive_lockmgrs.h"
31 #include "opt_kdtrace.h"
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 #include <sys/param.h>
38 #include <sys/linker_set.h>
40 #include <sys/lock_profile.h>
41 #include <sys/lockmgr.h>
42 #include <sys/mutex.h>
44 #include <sys/sleepqueue.h>
46 #include <sys/stack.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
51 #include <machine/cpu.h>
57 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
58 (LK_ADAPTIVE | LK_NOSHARE));
59 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
60 ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
62 #define SQ_EXCLUSIVE_QUEUE 0
63 #define SQ_SHARED_QUEUE 1
66 #define _lockmgr_assert(lk, what, file, line)
67 #define TD_LOCKS_INC(td)
68 #define TD_LOCKS_DEC(td)
70 #define TD_LOCKS_INC(td) ((td)->td_locks++)
71 #define TD_LOCKS_DEC(td) ((td)->td_locks--)
73 #define TD_SLOCKS_INC(td) ((td)->td_lk_slocks++)
74 #define TD_SLOCKS_DEC(td) ((td)->td_lk_slocks--)
77 #define STACK_PRINT(lk)
78 #define STACK_SAVE(lk)
79 #define STACK_ZERO(lk)
81 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
82 #define STACK_SAVE(lk) stack_save(&(lk)->lk_stack)
83 #define STACK_ZERO(lk) stack_zero(&(lk)->lk_stack)
86 #define LOCK_LOG2(lk, string, arg1, arg2) \
87 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \
88 CTR2(KTR_LOCK, (string), (arg1), (arg2))
89 #define LOCK_LOG3(lk, string, arg1, arg2, arg3) \
90 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \
91 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
93 #define GIANT_DECLARE \
95 WITNESS_SAVE_DECL(Giant)
96 #define GIANT_RESTORE() do { \
100 WITNESS_RESTORE(&Giant.lock_object, Giant); \
103 #define GIANT_SAVE() do { \
104 if (mtx_owned(&Giant)) { \
105 WITNESS_SAVE(&Giant.lock_object, Giant); \
106 while (mtx_owned(&Giant)) { \
108 mtx_unlock(&Giant); \
113 #define LK_CAN_SHARE(x) \
114 (((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 || \
115 ((x) & LK_EXCLUSIVE_SPINNERS) == 0 || \
116 curthread->td_lk_slocks || (curthread->td_pflags & TDP_DEADLKTREAT)))
117 #define LK_TRYOP(x) \
120 #define LK_CAN_WITNESS(x) \
121 (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
122 #define LK_TRYWIT(x) \
123 (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
125 #define LK_CAN_ADAPT(lk, f) \
126 (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 && \
127 ((f) & LK_SLEEPFAIL) == 0)
129 #define lockmgr_disowned(lk) \
130 (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
132 #define lockmgr_xlocked(lk) \
133 (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
135 static void assert_lockmgr(struct lock_object *lock, int how);
137 static void db_show_lockmgr(struct lock_object *lock);
139 static void lock_lockmgr(struct lock_object *lock, int how);
141 static int owner_lockmgr(struct lock_object *lock, struct thread **owner);
143 static int unlock_lockmgr(struct lock_object *lock);
145 struct lock_class lock_class_lockmgr = {
146 .lc_name = "lockmgr",
147 .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
148 .lc_assert = assert_lockmgr,
150 .lc_ddb_show = db_show_lockmgr,
152 .lc_lock = lock_lockmgr,
153 .lc_unlock = unlock_lockmgr,
155 .lc_owner = owner_lockmgr,
159 #ifdef ADAPTIVE_LOCKMGRS
160 static u_int alk_retries = 10;
161 static u_int alk_loops = 10000;
162 SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging");
163 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
164 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
167 static __inline struct thread *
168 lockmgr_xholder(struct lock *lk)
173 return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
177 * It assumes sleepq_lock held and returns with this one unheld.
178 * It also assumes the generic interlock is sane and previously checked.
179 * If LK_INTERLOCK is specified the interlock is not reacquired after the
183 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
184 const char *wmesg, int pri, int timo, int queue)
187 struct lock_class *class;
190 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
191 catch = pri & PCATCH;
195 LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
196 (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
198 if (flags & LK_INTERLOCK)
199 class->lc_unlock(ilk);
200 if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
203 sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
204 SLEEPQ_INTERRUPTIBLE : 0), queue);
205 if ((flags & LK_TIMELOCK) && timo)
206 sleepq_set_timeout(&lk->lock_object, timo);
209 * Decisional switch for real sleeping.
211 if ((flags & LK_TIMELOCK) && timo && catch)
212 error = sleepq_timedwait_sig(&lk->lock_object, pri);
213 else if ((flags & LK_TIMELOCK) && timo)
214 error = sleepq_timedwait(&lk->lock_object, pri);
216 error = sleepq_wait_sig(&lk->lock_object, pri);
218 sleepq_wait(&lk->lock_object, pri);
220 if ((flags & LK_SLEEPFAIL) && error == 0)
227 wakeupshlk(struct lock *lk, const char *file, int line)
231 int queue, wakeup_swapper;
233 TD_LOCKS_DEC(curthread);
234 TD_SLOCKS_DEC(curthread);
235 WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
236 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
243 * If there is more than one shared lock held, just drop one
246 if (LK_SHARERS(x) > 1) {
247 if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
254 * If there are not waiters on the exclusive queue, drop the
257 if ((x & LK_ALL_WAITERS) == 0) {
258 MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
260 if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
266 * We should have a sharer with waiters, so enter the hard
267 * path in order to handle wakeups correctly.
269 sleepq_lock(&lk->lock_object);
270 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
274 * If the lock has exclusive waiters, give them preference in
275 * order to avoid deadlock with shared runners up.
276 * If interruptible sleeps left the exclusive queue empty
277 * avoid a starvation for the threads sleeping on the shared
278 * queue by giving them precedence and cleaning up the
279 * exclusive waiters bit anyway.
280 * Please note that lk_exslpfail count may be lying about
281 * the real number of waiters with the LK_SLEEPFAIL flag on
282 * because they may be used in conjuction with interruptible
283 * sleeps so lk_exslpfail might be considered an 'upper limit'
284 * bound, including the edge cases.
286 realexslp = sleepq_sleepcnt(&lk->lock_object,
288 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
289 if (lk->lk_exslpfail < realexslp) {
290 lk->lk_exslpfail = 0;
291 queue = SQ_EXCLUSIVE_QUEUE;
292 v |= (x & LK_SHARED_WAITERS);
294 lk->lk_exslpfail = 0;
296 "%s: %p has only LK_SLEEPFAIL sleepers",
299 "%s: %p waking up threads on the exclusive queue",
302 sleepq_broadcast(&lk->lock_object,
303 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
304 queue = SQ_SHARED_QUEUE;
310 * Exclusive waiters sleeping with LK_SLEEPFAIL on
311 * and using interruptible sleeps/timeout may have
312 * left spourious lk_exslpfail counts on, so clean
315 lk->lk_exslpfail = 0;
316 queue = SQ_SHARED_QUEUE;
319 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
321 sleepq_release(&lk->lock_object);
324 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
325 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
327 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
329 sleepq_release(&lk->lock_object);
333 lock_profile_release_lock(&lk->lock_object);
334 return (wakeup_swapper);
338 assert_lockmgr(struct lock_object *lock, int what)
341 panic("lockmgr locks do not support assertions");
345 lock_lockmgr(struct lock_object *lock, int how)
348 panic("lockmgr locks do not support sleep interlocking");
352 unlock_lockmgr(struct lock_object *lock)
355 panic("lockmgr locks do not support sleep interlocking");
360 owner_lockmgr(struct lock_object *lock, struct thread **owner)
363 panic("lockmgr locks do not support owner inquiring");
368 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
372 MPASS((flags & ~LK_INIT_MASK) == 0);
373 ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
374 ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
377 iflags = LO_SLEEPABLE | LO_UPGRADABLE;
378 if (flags & LK_CANRECURSE)
379 iflags |= LO_RECURSABLE;
380 if ((flags & LK_NODUP) == 0)
382 if (flags & LK_NOPROFILE)
383 iflags |= LO_NOPROFILE;
384 if ((flags & LK_NOWITNESS) == 0)
385 iflags |= LO_WITNESS;
386 if (flags & LK_QUIET)
388 iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
390 lk->lk_lock = LK_UNLOCKED;
392 lk->lk_exslpfail = 0;
395 lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
400 * XXX: Gross hacks to manipulate external lock flags after
401 * initialization. Used for certain vnode and buf locks.
404 lockallowshare(struct lock *lk)
407 lockmgr_assert(lk, KA_XLOCKED);
408 lk->lock_object.lo_flags &= ~LK_NOSHARE;
412 lockallowrecurse(struct lock *lk)
415 lockmgr_assert(lk, KA_XLOCKED);
416 lk->lock_object.lo_flags |= LO_RECURSABLE;
420 lockdisablerecurse(struct lock *lk)
423 lockmgr_assert(lk, KA_XLOCKED);
424 lk->lock_object.lo_flags &= ~LO_RECURSABLE;
428 lockdestroy(struct lock *lk)
431 KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
432 KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
433 KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
434 lock_destroy(&lk->lock_object);
438 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
439 const char *wmesg, int pri, int timo, const char *file, int line)
442 struct lock_class *class;
446 int error, ipri, itimo, queue, wakeup_swapper;
447 #ifdef LOCK_PROFILING
448 uint64_t waittime = 0;
451 #ifdef ADAPTIVE_LOCKMGRS
452 volatile struct thread *owner;
453 u_int i, spintries = 0;
457 tid = (uintptr_t)curthread;
458 op = (flags & LK_TYPE_MASK);
459 iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
460 ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
461 itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
463 MPASS((flags & ~LK_TOTAL_MASK) == 0);
464 KASSERT((op & (op - 1)) == 0,
465 ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
466 KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
467 (op != LK_DOWNGRADE && op != LK_RELEASE),
468 ("%s: Invalid flags in regard of the operation desired @ %s:%d",
469 __func__, file, line));
470 KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
471 ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
472 __func__, file, line));
474 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
475 if (panicstr != NULL) {
476 if (flags & LK_INTERLOCK)
477 class->lc_unlock(ilk);
481 if (op == LK_SHARED && (lk->lock_object.lo_flags & LK_NOSHARE))
487 if (LK_CAN_WITNESS(flags))
488 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
494 * If no other thread has an exclusive lock, or
495 * no exclusive waiter is present, bump the count of
496 * sharers. Since we have to preserve the state of
497 * waiters, if we fail to acquire the shared lock
498 * loop back and retry.
500 if (LK_CAN_SHARE(x)) {
501 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
506 lock_profile_obtain_lock_failed(&lk->lock_object,
507 &contested, &waittime);
510 * If the lock is already held by curthread in
511 * exclusive way avoid a deadlock.
513 if (LK_HOLDER(x) == tid) {
515 "%s: %p already held in exclusive mode",
522 * If the lock is expected to not sleep just give up
525 if (LK_TRYOP(flags)) {
526 LOCK_LOG2(lk, "%s: %p fails the try operation",
532 #ifdef ADAPTIVE_LOCKMGRS
534 * If the owner is running on another CPU, spin until
535 * the owner stops running or the state of the lock
536 * changes. We need a double-state handle here
537 * because for a failed acquisition the lock can be
538 * either held in exclusive mode or shared mode
539 * (for the writer starvation avoidance technique).
541 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
542 LK_HOLDER(x) != LK_KERNPROC) {
543 owner = (struct thread *)LK_HOLDER(x);
544 if (LOCK_LOG_TEST(&lk->lock_object, 0))
546 "%s: spinning on %p held by %p",
547 __func__, lk, owner);
550 * If we are holding also an interlock drop it
551 * in order to avoid a deadlock if the lockmgr
552 * owner is adaptively spinning on the
555 if (flags & LK_INTERLOCK) {
556 class->lc_unlock(ilk);
557 flags &= ~LK_INTERLOCK;
560 while (LK_HOLDER(lk->lk_lock) ==
561 (uintptr_t)owner && TD_IS_RUNNING(owner))
565 } else if (LK_CAN_ADAPT(lk, flags) &&
566 (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
567 spintries < alk_retries) {
568 if (flags & LK_INTERLOCK) {
569 class->lc_unlock(ilk);
570 flags &= ~LK_INTERLOCK;
574 for (i = 0; i < alk_loops; i++) {
575 if (LOCK_LOG_TEST(&lk->lock_object, 0))
577 "%s: shared spinning on %p with %u and %u",
578 __func__, lk, spintries, i);
580 if ((x & LK_SHARE) == 0 ||
581 LK_CAN_SHARE(x) != 0)
592 * Acquire the sleepqueue chain lock because we
593 * probabilly will need to manipulate waiters flags.
595 sleepq_lock(&lk->lock_object);
599 * if the lock can be acquired in shared mode, try
602 if (LK_CAN_SHARE(x)) {
603 sleepq_release(&lk->lock_object);
607 #ifdef ADAPTIVE_LOCKMGRS
609 * The current lock owner might have started executing
610 * on another CPU (or the lock could have changed
611 * owner) while we were waiting on the turnstile
612 * chain lock. If so, drop the turnstile lock and try
615 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
616 LK_HOLDER(x) != LK_KERNPROC) {
617 owner = (struct thread *)LK_HOLDER(x);
618 if (TD_IS_RUNNING(owner)) {
619 sleepq_release(&lk->lock_object);
626 * Try to set the LK_SHARED_WAITERS flag. If we fail,
627 * loop back and retry.
629 if ((x & LK_SHARED_WAITERS) == 0) {
630 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
631 x | LK_SHARED_WAITERS)) {
632 sleepq_release(&lk->lock_object);
635 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
640 * As far as we have been unable to acquire the
641 * shared lock and the shared waiters flag is set,
644 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
646 flags &= ~LK_INTERLOCK;
649 "%s: interrupted sleep for %p with %d",
650 __func__, lk, error);
653 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
657 lock_profile_obtain_lock_success(&lk->lock_object,
658 contested, waittime, file, line);
659 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
661 WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
663 TD_LOCKS_INC(curthread);
664 TD_SLOCKS_INC(curthread);
669 _lockmgr_assert(lk, KA_SLOCKED, file, line);
671 x = v & LK_ALL_WAITERS;
672 v &= LK_EXCLUSIVE_SPINNERS;
675 * Try to switch from one shared lock to an exclusive one.
676 * We need to preserve waiters flags during the operation.
678 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
680 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
682 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
683 LK_TRYWIT(flags), file, line);
684 TD_SLOCKS_DEC(curthread);
689 * We have been unable to succeed in upgrading, so just
690 * give up the shared lock.
692 wakeup_swapper |= wakeupshlk(lk, file, line);
696 if (LK_CAN_WITNESS(flags))
697 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
698 LOP_EXCLUSIVE, file, line, ilk);
701 * If curthread already holds the lock and this one is
702 * allowed to recurse, simply recurse on it.
704 if (lockmgr_xlocked(lk)) {
705 if ((flags & LK_CANRECURSE) == 0 &&
706 (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
709 * If the lock is expected to not panic just
710 * give up and return.
712 if (LK_TRYOP(flags)) {
714 "%s: %p fails the try operation",
719 if (flags & LK_INTERLOCK)
720 class->lc_unlock(ilk);
721 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
722 __func__, iwmesg, file, line);
725 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
726 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
727 lk->lk_recurse, file, line);
728 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
729 LK_TRYWIT(flags), file, line);
730 TD_LOCKS_INC(curthread);
734 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
736 lock_profile_obtain_lock_failed(&lk->lock_object,
737 &contested, &waittime);
740 * If the lock is expected to not sleep just give up
743 if (LK_TRYOP(flags)) {
744 LOCK_LOG2(lk, "%s: %p fails the try operation",
750 #ifdef ADAPTIVE_LOCKMGRS
752 * If the owner is running on another CPU, spin until
753 * the owner stops running or the state of the lock
757 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
758 LK_HOLDER(x) != LK_KERNPROC) {
759 owner = (struct thread *)LK_HOLDER(x);
760 if (LOCK_LOG_TEST(&lk->lock_object, 0))
762 "%s: spinning on %p held by %p",
763 __func__, lk, owner);
766 * If we are holding also an interlock drop it
767 * in order to avoid a deadlock if the lockmgr
768 * owner is adaptively spinning on the
771 if (flags & LK_INTERLOCK) {
772 class->lc_unlock(ilk);
773 flags &= ~LK_INTERLOCK;
776 while (LK_HOLDER(lk->lk_lock) ==
777 (uintptr_t)owner && TD_IS_RUNNING(owner))
781 } else if (LK_CAN_ADAPT(lk, flags) &&
782 (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
783 spintries < alk_retries) {
784 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
785 !atomic_cmpset_ptr(&lk->lk_lock, x,
786 x | LK_EXCLUSIVE_SPINNERS))
788 if (flags & LK_INTERLOCK) {
789 class->lc_unlock(ilk);
790 flags &= ~LK_INTERLOCK;
794 for (i = 0; i < alk_loops; i++) {
795 if (LOCK_LOG_TEST(&lk->lock_object, 0))
797 "%s: shared spinning on %p with %u and %u",
798 __func__, lk, spintries, i);
800 LK_EXCLUSIVE_SPINNERS) == 0)
811 * Acquire the sleepqueue chain lock because we
812 * probabilly will need to manipulate waiters flags.
814 sleepq_lock(&lk->lock_object);
818 * if the lock has been released while we spun on
819 * the sleepqueue chain lock just try again.
821 if (x == LK_UNLOCKED) {
822 sleepq_release(&lk->lock_object);
826 #ifdef ADAPTIVE_LOCKMGRS
828 * The current lock owner might have started executing
829 * on another CPU (or the lock could have changed
830 * owner) while we were waiting on the turnstile
831 * chain lock. If so, drop the turnstile lock and try
834 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
835 LK_HOLDER(x) != LK_KERNPROC) {
836 owner = (struct thread *)LK_HOLDER(x);
837 if (TD_IS_RUNNING(owner)) {
838 sleepq_release(&lk->lock_object);
845 * The lock can be in the state where there is a
846 * pending queue of waiters, but still no owner.
847 * This happens when the lock is contested and an
848 * owner is going to claim the lock.
849 * If curthread is the one successfully acquiring it
850 * claim lock ownership and return, preserving waiters
853 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
854 if ((x & ~v) == LK_UNLOCKED) {
855 v &= ~LK_EXCLUSIVE_SPINNERS;
856 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
858 sleepq_release(&lk->lock_object);
860 "%s: %p claimed by a new writer",
864 sleepq_release(&lk->lock_object);
869 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we
870 * fail, loop back and retry.
872 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
873 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
874 x | LK_EXCLUSIVE_WAITERS)) {
875 sleepq_release(&lk->lock_object);
878 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
883 * As far as we have been unable to acquire the
884 * exclusive lock and the exclusive waiters flag
885 * is set, we will sleep.
887 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
889 flags &= ~LK_INTERLOCK;
892 "%s: interrupted sleep for %p with %d",
893 __func__, lk, error);
896 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
900 lock_profile_obtain_lock_success(&lk->lock_object,
901 contested, waittime, file, line);
902 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
903 lk->lk_recurse, file, line);
904 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
905 LK_TRYWIT(flags), file, line);
906 TD_LOCKS_INC(curthread);
911 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
912 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
913 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
914 TD_SLOCKS_INC(curthread);
917 * In order to preserve waiters flags, just spin.
921 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
923 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
924 LK_SHARERS_LOCK(1) | x))
930 _lockmgr_assert(lk, KA_LOCKED, file, line);
933 if ((x & LK_SHARE) == 0) {
936 * As first option, treact the lock as if it has not
938 * Fix-up the tid var if the lock has been disowned.
940 if (LK_HOLDER(x) == LK_KERNPROC)
943 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
945 TD_LOCKS_DEC(curthread);
947 LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
948 lk->lk_recurse, file, line);
951 * The lock is held in exclusive mode.
952 * If the lock is recursed also, then unrecurse it.
954 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
955 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
960 if (tid != LK_KERNPROC)
961 lock_profile_release_lock(&lk->lock_object);
963 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
967 sleepq_lock(&lk->lock_object);
972 * If the lock has exclusive waiters, give them
973 * preference in order to avoid deadlock with
975 * If interruptible sleeps left the exclusive queue
976 * empty avoid a starvation for the threads sleeping
977 * on the shared queue by giving them precedence
978 * and cleaning up the exclusive waiters bit anyway.
979 * Please note that lk_exslpfail count may be lying
980 * about the real number of waiters with the
981 * LK_SLEEPFAIL flag on because they may be used in
982 * conjuction with interruptible sleeps so
983 * lk_exslpfail might be considered an 'upper limit'
984 * bound, including the edge cases.
986 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
987 realexslp = sleepq_sleepcnt(&lk->lock_object,
989 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
990 if (lk->lk_exslpfail < realexslp) {
991 lk->lk_exslpfail = 0;
992 queue = SQ_EXCLUSIVE_QUEUE;
993 v |= (x & LK_SHARED_WAITERS);
995 lk->lk_exslpfail = 0;
997 "%s: %p has only LK_SLEEPFAIL sleepers",
1000 "%s: %p waking up threads on the exclusive queue",
1003 sleepq_broadcast(&lk->lock_object,
1004 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1005 queue = SQ_SHARED_QUEUE;
1010 * Exclusive waiters sleeping with LK_SLEEPFAIL
1011 * on and using interruptible sleeps/timeout
1012 * may have left spourious lk_exslpfail counts
1013 * on, so clean it up anyway.
1015 lk->lk_exslpfail = 0;
1016 queue = SQ_SHARED_QUEUE;
1020 "%s: %p waking up threads on the %s queue",
1021 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1023 atomic_store_rel_ptr(&lk->lk_lock, v);
1024 wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1025 SLEEPQ_LK, 0, queue);
1026 sleepq_release(&lk->lock_object);
1029 wakeup_swapper = wakeupshlk(lk, file, line);
1032 if (LK_CAN_WITNESS(flags))
1033 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1034 LOP_EXCLUSIVE, file, line, ilk);
1037 * Trying to drain a lock we already own will result in a
1040 if (lockmgr_xlocked(lk)) {
1041 if (flags & LK_INTERLOCK)
1042 class->lc_unlock(ilk);
1043 panic("%s: draining %s with the lock held @ %s:%d\n",
1044 __func__, iwmesg, file, line);
1047 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1048 lock_profile_obtain_lock_failed(&lk->lock_object,
1049 &contested, &waittime);
1052 * If the lock is expected to not sleep just give up
1055 if (LK_TRYOP(flags)) {
1056 LOCK_LOG2(lk, "%s: %p fails the try operation",
1063 * Acquire the sleepqueue chain lock because we
1064 * probabilly will need to manipulate waiters flags.
1066 sleepq_lock(&lk->lock_object);
1070 * if the lock has been released while we spun on
1071 * the sleepqueue chain lock just try again.
1073 if (x == LK_UNLOCKED) {
1074 sleepq_release(&lk->lock_object);
1078 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1079 if ((x & ~v) == LK_UNLOCKED) {
1080 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1083 * If interruptible sleeps left the exclusive
1084 * queue empty avoid a starvation for the
1085 * threads sleeping on the shared queue by
1086 * giving them precedence and cleaning up the
1087 * exclusive waiters bit anyway.
1088 * Please note that lk_exslpfail count may be
1089 * lying about the real number of waiters with
1090 * the LK_SLEEPFAIL flag on because they may
1091 * be used in conjuction with interruptible
1092 * sleeps so lk_exslpfail might be considered
1093 * an 'upper limit' bound, including the edge
1096 if (v & LK_EXCLUSIVE_WAITERS) {
1097 queue = SQ_EXCLUSIVE_QUEUE;
1098 v &= ~LK_EXCLUSIVE_WAITERS;
1102 * Exclusive waiters sleeping with
1103 * LK_SLEEPFAIL on and using
1104 * interruptible sleeps/timeout may
1105 * have left spourious lk_exslpfail
1106 * counts on, so clean it up anyway.
1108 MPASS(v & LK_SHARED_WAITERS);
1109 lk->lk_exslpfail = 0;
1110 queue = SQ_SHARED_QUEUE;
1111 v &= ~LK_SHARED_WAITERS;
1113 if (queue == SQ_EXCLUSIVE_QUEUE) {
1115 sleepq_sleepcnt(&lk->lock_object,
1116 SQ_EXCLUSIVE_QUEUE);
1117 if (lk->lk_exslpfail >= realexslp) {
1118 lk->lk_exslpfail = 0;
1119 queue = SQ_SHARED_QUEUE;
1120 v &= ~LK_SHARED_WAITERS;
1121 if (realexslp != 0) {
1123 "%s: %p has only LK_SLEEPFAIL sleepers",
1126 "%s: %p waking up threads on the exclusive queue",
1132 SQ_EXCLUSIVE_QUEUE);
1135 lk->lk_exslpfail = 0;
1137 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1138 sleepq_release(&lk->lock_object);
1142 "%s: %p waking up all threads on the %s queue",
1143 __func__, lk, queue == SQ_SHARED_QUEUE ?
1144 "shared" : "exclusive");
1145 wakeup_swapper |= sleepq_broadcast(
1146 &lk->lock_object, SLEEPQ_LK, 0, queue);
1149 * If shared waiters have been woken up we need
1150 * to wait for one of them to acquire the lock
1151 * before to set the exclusive waiters in
1152 * order to avoid a deadlock.
1154 if (queue == SQ_SHARED_QUEUE) {
1155 for (v = lk->lk_lock;
1156 (v & LK_SHARE) && !LK_SHARERS(v);
1163 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we
1164 * fail, loop back and retry.
1166 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1167 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1168 x | LK_EXCLUSIVE_WAITERS)) {
1169 sleepq_release(&lk->lock_object);
1172 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1177 * As far as we have been unable to acquire the
1178 * exclusive lock and the exclusive waiters flag
1179 * is set, we will sleep.
1181 if (flags & LK_INTERLOCK) {
1182 class->lc_unlock(ilk);
1183 flags &= ~LK_INTERLOCK;
1186 sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1187 SQ_EXCLUSIVE_QUEUE);
1188 sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1190 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1195 lock_profile_obtain_lock_success(&lk->lock_object,
1196 contested, waittime, file, line);
1197 LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1198 lk->lk_recurse, file, line);
1199 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1200 LK_TRYWIT(flags), file, line);
1201 TD_LOCKS_INC(curthread);
1206 if (flags & LK_INTERLOCK)
1207 class->lc_unlock(ilk);
1208 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1211 if (flags & LK_INTERLOCK)
1212 class->lc_unlock(ilk);
1220 _lockmgr_disown(struct lock *lk, const char *file, int line)
1224 tid = (uintptr_t)curthread;
1225 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line);
1228 * If the owner is already LK_KERNPROC just skip the whole operation.
1230 if (LK_HOLDER(lk->lk_lock) != tid)
1232 lock_profile_release_lock(&lk->lock_object);
1233 LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1234 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1235 TD_LOCKS_DEC(curthread);
1239 * In order to preserve waiters flags, just spin.
1243 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1244 x &= LK_ALL_WAITERS;
1245 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1253 lockmgr_printinfo(struct lock *lk)
1258 if (lk->lk_lock == LK_UNLOCKED)
1259 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1260 else if (lk->lk_lock & LK_SHARE)
1261 printf("lock type %s: SHARED (count %ju)\n",
1262 lk->lock_object.lo_name,
1263 (uintmax_t)LK_SHARERS(lk->lk_lock));
1265 td = lockmgr_xholder(lk);
1266 printf("lock type %s: EXCL by thread %p (pid %d)\n",
1267 lk->lock_object.lo_name, td, td->td_proc->p_pid);
1271 if (x & LK_EXCLUSIVE_WAITERS)
1272 printf(" with exclusive waiters pending\n");
1273 if (x & LK_SHARED_WAITERS)
1274 printf(" with shared waiters pending\n");
1275 if (x & LK_EXCLUSIVE_SPINNERS)
1276 printf(" with exclusive spinners pending\n");
1282 lockstatus(struct lock *lk)
1291 if ((x & LK_SHARE) == 0) {
1292 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1296 } else if (x == LK_UNLOCKED)
1302 #ifdef INVARIANT_SUPPORT
1304 #undef _lockmgr_assert
1308 _lockmgr_assert(struct lock *lk, int what, const char *file, int line)
1312 if (panicstr != NULL)
1316 case KA_SLOCKED | KA_NOTRECURSED:
1317 case KA_SLOCKED | KA_RECURSED:
1320 case KA_LOCKED | KA_NOTRECURSED:
1321 case KA_LOCKED | KA_RECURSED:
1325 * We cannot trust WITNESS if the lock is held in exclusive
1326 * mode and a call to lockmgr_disown() happened.
1327 * Workaround this skipping the check if the lock is held in
1328 * exclusive mode even for the KA_LOCKED case.
1330 if (slocked || (lk->lk_lock & LK_SHARE)) {
1331 witness_assert(&lk->lock_object, what, file, line);
1335 if (lk->lk_lock == LK_UNLOCKED ||
1336 ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1337 (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1338 panic("Lock %s not %slocked @ %s:%d\n",
1339 lk->lock_object.lo_name, slocked ? "share" : "",
1342 if ((lk->lk_lock & LK_SHARE) == 0) {
1343 if (lockmgr_recursed(lk)) {
1344 if (what & KA_NOTRECURSED)
1345 panic("Lock %s recursed @ %s:%d\n",
1346 lk->lock_object.lo_name, file,
1348 } else if (what & KA_RECURSED)
1349 panic("Lock %s not recursed @ %s:%d\n",
1350 lk->lock_object.lo_name, file, line);
1354 case KA_XLOCKED | KA_NOTRECURSED:
1355 case KA_XLOCKED | KA_RECURSED:
1356 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1357 panic("Lock %s not exclusively locked @ %s:%d\n",
1358 lk->lock_object.lo_name, file, line);
1359 if (lockmgr_recursed(lk)) {
1360 if (what & KA_NOTRECURSED)
1361 panic("Lock %s recursed @ %s:%d\n",
1362 lk->lock_object.lo_name, file, line);
1363 } else if (what & KA_RECURSED)
1364 panic("Lock %s not recursed @ %s:%d\n",
1365 lk->lock_object.lo_name, file, line);
1368 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1369 panic("Lock %s exclusively locked @ %s:%d\n",
1370 lk->lock_object.lo_name, file, line);
1373 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1381 lockmgr_chain(struct thread *td, struct thread **ownerp)
1387 if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1389 db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1390 if (lk->lk_lock & LK_SHARE)
1391 db_printf("SHARED (count %ju)\n",
1392 (uintmax_t)LK_SHARERS(lk->lk_lock));
1394 db_printf("EXCL\n");
1395 *ownerp = lockmgr_xholder(lk);
1401 db_show_lockmgr(struct lock_object *lock)
1406 lk = (struct lock *)lock;
1408 db_printf(" state: ");
1409 if (lk->lk_lock == LK_UNLOCKED)
1410 db_printf("UNLOCKED\n");
1411 else if (lk->lk_lock & LK_SHARE)
1412 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1414 td = lockmgr_xholder(lk);
1415 if (td == (struct thread *)LK_KERNPROC)
1416 db_printf("XLOCK: LK_KERNPROC\n");
1418 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1419 td->td_tid, td->td_proc->p_pid,
1420 td->td_proc->p_comm);
1421 if (lockmgr_recursed(lk))
1422 db_printf(" recursed: %d\n", lk->lk_recurse);
1424 db_printf(" waiters: ");
1425 switch (lk->lk_lock & LK_ALL_WAITERS) {
1426 case LK_SHARED_WAITERS:
1427 db_printf("shared\n");
1429 case LK_EXCLUSIVE_WAITERS:
1430 db_printf("exclusive\n");
1432 case LK_ALL_WAITERS:
1433 db_printf("shared and exclusive\n");
1436 db_printf("none\n");
1438 db_printf(" spinners: ");
1439 if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1440 db_printf("exclusive\n");
1442 db_printf("none\n");