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_hwpmc_hooks.h"
32 #include "opt_kdtrace.h"
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
41 #include <sys/lock_profile.h>
42 #include <sys/lockmgr.h>
43 #include <sys/mutex.h>
45 #include <sys/sleepqueue.h>
47 #include <sys/stack.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
52 #include <machine/cpu.h>
59 #include <sys/pmckern.h>
60 PMC_SOFT_DECLARE( , , lock, failed);
63 CTASSERT(((LK_ADAPTIVE | LK_NOSHARE) & LO_CLASSFLAGS) ==
64 (LK_ADAPTIVE | LK_NOSHARE));
65 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED &
66 ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS)));
68 #define SQ_EXCLUSIVE_QUEUE 0
69 #define SQ_SHARED_QUEUE 1
72 #define _lockmgr_assert(lk, what, file, line)
73 #define TD_LOCKS_INC(td)
74 #define TD_LOCKS_DEC(td)
76 #define TD_LOCKS_INC(td) ((td)->td_locks++)
77 #define TD_LOCKS_DEC(td) ((td)->td_locks--)
79 #define TD_SLOCKS_INC(td) ((td)->td_lk_slocks++)
80 #define TD_SLOCKS_DEC(td) ((td)->td_lk_slocks--)
83 #define STACK_PRINT(lk)
84 #define STACK_SAVE(lk)
85 #define STACK_ZERO(lk)
87 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack)
88 #define STACK_SAVE(lk) stack_save(&(lk)->lk_stack)
89 #define STACK_ZERO(lk) stack_zero(&(lk)->lk_stack)
92 #define LOCK_LOG2(lk, string, arg1, arg2) \
93 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \
94 CTR2(KTR_LOCK, (string), (arg1), (arg2))
95 #define LOCK_LOG3(lk, string, arg1, arg2, arg3) \
96 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \
97 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3))
99 #define GIANT_DECLARE \
101 WITNESS_SAVE_DECL(Giant)
102 #define GIANT_RESTORE() do { \
106 WITNESS_RESTORE(&Giant.lock_object, Giant); \
109 #define GIANT_SAVE() do { \
110 if (mtx_owned(&Giant)) { \
111 WITNESS_SAVE(&Giant.lock_object, Giant); \
112 while (mtx_owned(&Giant)) { \
114 mtx_unlock(&Giant); \
119 #define LK_CAN_SHARE(x) \
120 (((x) & LK_SHARE) && (((x) & LK_EXCLUSIVE_WAITERS) == 0 || \
121 ((x) & LK_EXCLUSIVE_SPINNERS) == 0 || \
122 curthread->td_lk_slocks || (curthread->td_pflags & TDP_DEADLKTREAT)))
123 #define LK_TRYOP(x) \
126 #define LK_CAN_WITNESS(x) \
127 (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x))
128 #define LK_TRYWIT(x) \
129 (LK_TRYOP(x) ? LOP_TRYLOCK : 0)
131 #define LK_CAN_ADAPT(lk, f) \
132 (((lk)->lock_object.lo_flags & LK_ADAPTIVE) != 0 && \
133 ((f) & LK_SLEEPFAIL) == 0)
135 #define lockmgr_disowned(lk) \
136 (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC)
138 #define lockmgr_xlocked(lk) \
139 (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread)
141 static void assert_lockmgr(const struct lock_object *lock, int how);
143 static void db_show_lockmgr(const struct lock_object *lock);
145 static void lock_lockmgr(struct lock_object *lock, int how);
147 static int owner_lockmgr(const struct lock_object *lock,
148 struct thread **owner);
150 static int unlock_lockmgr(struct lock_object *lock);
152 struct lock_class lock_class_lockmgr = {
153 .lc_name = "lockmgr",
154 .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE,
155 .lc_assert = assert_lockmgr,
157 .lc_ddb_show = db_show_lockmgr,
159 .lc_lock = lock_lockmgr,
160 .lc_unlock = unlock_lockmgr,
162 .lc_owner = owner_lockmgr,
166 #ifdef ADAPTIVE_LOCKMGRS
167 static u_int alk_retries = 10;
168 static u_int alk_loops = 10000;
169 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL,
170 "lockmgr debugging");
171 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, retries, CTLFLAG_RW, &alk_retries, 0, "");
172 SYSCTL_UINT(_debug_lockmgr, OID_AUTO, loops, CTLFLAG_RW, &alk_loops, 0, "");
175 static __inline struct thread *
176 lockmgr_xholder(const struct lock *lk)
181 return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x));
185 * It assumes sleepq_lock held and returns with this one unheld.
186 * It also assumes the generic interlock is sane and previously checked.
187 * If LK_INTERLOCK is specified the interlock is not reacquired after the
191 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk,
192 const char *wmesg, int pri, int timo, int queue)
195 struct lock_class *class;
198 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
199 catch = pri & PCATCH;
203 LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk,
204 (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared");
206 if (flags & LK_INTERLOCK)
207 class->lc_unlock(ilk);
208 if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0)
211 sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ?
212 SLEEPQ_INTERRUPTIBLE : 0), queue);
213 if ((flags & LK_TIMELOCK) && timo)
214 sleepq_set_timeout(&lk->lock_object, timo);
217 * Decisional switch for real sleeping.
219 if ((flags & LK_TIMELOCK) && timo && catch)
220 error = sleepq_timedwait_sig(&lk->lock_object, pri);
221 else if ((flags & LK_TIMELOCK) && timo)
222 error = sleepq_timedwait(&lk->lock_object, pri);
224 error = sleepq_wait_sig(&lk->lock_object, pri);
226 sleepq_wait(&lk->lock_object, pri);
228 if ((flags & LK_SLEEPFAIL) && error == 0)
235 wakeupshlk(struct lock *lk, const char *file, int line)
239 int queue, wakeup_swapper;
241 WITNESS_UNLOCK(&lk->lock_object, 0, file, line);
242 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line);
249 * If there is more than one shared lock held, just drop one
252 if (LK_SHARERS(x) > 1) {
253 if (atomic_cmpset_rel_ptr(&lk->lk_lock, x,
260 * If there are not waiters on the exclusive queue, drop the
263 if ((x & LK_ALL_WAITERS) == 0) {
264 MPASS((x & ~LK_EXCLUSIVE_SPINNERS) ==
266 if (atomic_cmpset_rel_ptr(&lk->lk_lock, x, LK_UNLOCKED))
272 * We should have a sharer with waiters, so enter the hard
273 * path in order to handle wakeups correctly.
275 sleepq_lock(&lk->lock_object);
276 x = lk->lk_lock & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
280 * If the lock has exclusive waiters, give them preference in
281 * order to avoid deadlock with shared runners up.
282 * If interruptible sleeps left the exclusive queue empty
283 * avoid a starvation for the threads sleeping on the shared
284 * queue by giving them precedence and cleaning up the
285 * exclusive waiters bit anyway.
286 * Please note that lk_exslpfail count may be lying about
287 * the real number of waiters with the LK_SLEEPFAIL flag on
288 * because they may be used in conjuction with interruptible
289 * sleeps so lk_exslpfail might be considered an 'upper limit'
290 * bound, including the edge cases.
292 realexslp = sleepq_sleepcnt(&lk->lock_object,
294 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
295 if (lk->lk_exslpfail < realexslp) {
296 lk->lk_exslpfail = 0;
297 queue = SQ_EXCLUSIVE_QUEUE;
298 v |= (x & LK_SHARED_WAITERS);
300 lk->lk_exslpfail = 0;
302 "%s: %p has only LK_SLEEPFAIL sleepers",
305 "%s: %p waking up threads on the exclusive queue",
308 sleepq_broadcast(&lk->lock_object,
309 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
310 queue = SQ_SHARED_QUEUE;
316 * Exclusive waiters sleeping with LK_SLEEPFAIL on
317 * and using interruptible sleeps/timeout may have
318 * left spourious lk_exslpfail counts on, so clean
321 lk->lk_exslpfail = 0;
322 queue = SQ_SHARED_QUEUE;
325 if (!atomic_cmpset_rel_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x,
327 sleepq_release(&lk->lock_object);
330 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue",
331 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
333 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK,
335 sleepq_release(&lk->lock_object);
339 lock_profile_release_lock(&lk->lock_object);
340 TD_LOCKS_DEC(curthread);
341 TD_SLOCKS_DEC(curthread);
342 return (wakeup_swapper);
346 assert_lockmgr(const struct lock_object *lock, int what)
349 panic("lockmgr locks do not support assertions");
353 lock_lockmgr(struct lock_object *lock, int how)
356 panic("lockmgr locks do not support sleep interlocking");
360 unlock_lockmgr(struct lock_object *lock)
363 panic("lockmgr locks do not support sleep interlocking");
368 owner_lockmgr(const struct lock_object *lock, struct thread **owner)
371 panic("lockmgr locks do not support owner inquiring");
376 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags)
380 MPASS((flags & ~LK_INIT_MASK) == 0);
381 ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock,
382 ("%s: lockmgr not aligned for %s: %p", __func__, wmesg,
385 iflags = LO_SLEEPABLE | LO_UPGRADABLE;
386 if (flags & LK_CANRECURSE)
387 iflags |= LO_RECURSABLE;
388 if ((flags & LK_NODUP) == 0)
390 if (flags & LK_NOPROFILE)
391 iflags |= LO_NOPROFILE;
392 if ((flags & LK_NOWITNESS) == 0)
393 iflags |= LO_WITNESS;
394 if (flags & LK_QUIET)
396 if (flags & LK_IS_VNODE)
397 iflags |= LO_IS_VNODE;
398 iflags |= flags & (LK_ADAPTIVE | LK_NOSHARE);
400 lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags);
401 lk->lk_lock = LK_UNLOCKED;
403 lk->lk_exslpfail = 0;
410 * XXX: Gross hacks to manipulate external lock flags after
411 * initialization. Used for certain vnode and buf locks.
414 lockallowshare(struct lock *lk)
417 lockmgr_assert(lk, KA_XLOCKED);
418 lk->lock_object.lo_flags &= ~LK_NOSHARE;
422 lockallowrecurse(struct lock *lk)
425 lockmgr_assert(lk, KA_XLOCKED);
426 lk->lock_object.lo_flags |= LO_RECURSABLE;
430 lockdisablerecurse(struct lock *lk)
433 lockmgr_assert(lk, KA_XLOCKED);
434 lk->lock_object.lo_flags &= ~LO_RECURSABLE;
438 lockdestroy(struct lock *lk)
441 KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held"));
442 KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed"));
443 KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters"));
444 lock_destroy(&lk->lock_object);
448 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
449 const char *wmesg, int pri, int timo, const char *file, int line)
452 struct lock_class *class;
456 int error, ipri, itimo, queue, wakeup_swapper;
457 #ifdef LOCK_PROFILING
458 uint64_t waittime = 0;
461 #ifdef ADAPTIVE_LOCKMGRS
462 volatile struct thread *owner;
463 u_int i, spintries = 0;
467 tid = (uintptr_t)curthread;
468 op = (flags & LK_TYPE_MASK);
469 iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg;
470 ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri;
471 itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo;
473 MPASS((flags & ~LK_TOTAL_MASK) == 0);
474 KASSERT((op & (op - 1)) == 0,
475 ("%s: Invalid requested operation @ %s:%d", __func__, file, line));
476 KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 ||
477 (op != LK_DOWNGRADE && op != LK_RELEASE),
478 ("%s: Invalid flags in regard of the operation desired @ %s:%d",
479 __func__, file, line));
480 KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL,
481 ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d",
482 __func__, file, line));
483 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
484 ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread,
485 lk->lock_object.lo_name, file, line));
487 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL;
488 if (panicstr != NULL) {
489 if (flags & LK_INTERLOCK)
490 class->lc_unlock(ilk);
494 if (lk->lock_object.lo_flags & LK_NOSHARE) {
501 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED,
503 if (flags & LK_INTERLOCK)
504 class->lc_unlock(ilk);
512 if (LK_CAN_WITNESS(flags))
513 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
514 file, line, flags & LK_INTERLOCK ? ilk : NULL);
519 * If no other thread has an exclusive lock, or
520 * no exclusive waiter is present, bump the count of
521 * sharers. Since we have to preserve the state of
522 * waiters, if we fail to acquire the shared lock
523 * loop back and retry.
525 if (LK_CAN_SHARE(x)) {
526 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
532 PMC_SOFT_CALL( , , lock, failed);
534 lock_profile_obtain_lock_failed(&lk->lock_object,
535 &contested, &waittime);
538 * If the lock is already held by curthread in
539 * exclusive way avoid a deadlock.
541 if (LK_HOLDER(x) == tid) {
543 "%s: %p already held in exclusive mode",
550 * If the lock is expected to not sleep just give up
553 if (LK_TRYOP(flags)) {
554 LOCK_LOG2(lk, "%s: %p fails the try operation",
560 #ifdef ADAPTIVE_LOCKMGRS
562 * If the owner is running on another CPU, spin until
563 * the owner stops running or the state of the lock
564 * changes. We need a double-state handle here
565 * because for a failed acquisition the lock can be
566 * either held in exclusive mode or shared mode
567 * (for the writer starvation avoidance technique).
569 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
570 LK_HOLDER(x) != LK_KERNPROC) {
571 owner = (struct thread *)LK_HOLDER(x);
572 if (LOCK_LOG_TEST(&lk->lock_object, 0))
574 "%s: spinning on %p held by %p",
575 __func__, lk, owner);
578 * If we are holding also an interlock drop it
579 * in order to avoid a deadlock if the lockmgr
580 * owner is adaptively spinning on the
583 if (flags & LK_INTERLOCK) {
584 class->lc_unlock(ilk);
585 flags &= ~LK_INTERLOCK;
588 while (LK_HOLDER(lk->lk_lock) ==
589 (uintptr_t)owner && TD_IS_RUNNING(owner))
593 } else if (LK_CAN_ADAPT(lk, flags) &&
594 (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
595 spintries < alk_retries) {
596 if (flags & LK_INTERLOCK) {
597 class->lc_unlock(ilk);
598 flags &= ~LK_INTERLOCK;
602 for (i = 0; i < alk_loops; i++) {
603 if (LOCK_LOG_TEST(&lk->lock_object, 0))
605 "%s: shared spinning on %p with %u and %u",
606 __func__, lk, spintries, i);
608 if ((x & LK_SHARE) == 0 ||
609 LK_CAN_SHARE(x) != 0)
620 * Acquire the sleepqueue chain lock because we
621 * probabilly will need to manipulate waiters flags.
623 sleepq_lock(&lk->lock_object);
627 * if the lock can be acquired in shared mode, try
630 if (LK_CAN_SHARE(x)) {
631 sleepq_release(&lk->lock_object);
635 #ifdef ADAPTIVE_LOCKMGRS
637 * The current lock owner might have started executing
638 * on another CPU (or the lock could have changed
639 * owner) while we were waiting on the turnstile
640 * chain lock. If so, drop the turnstile lock and try
643 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
644 LK_HOLDER(x) != LK_KERNPROC) {
645 owner = (struct thread *)LK_HOLDER(x);
646 if (TD_IS_RUNNING(owner)) {
647 sleepq_release(&lk->lock_object);
654 * Try to set the LK_SHARED_WAITERS flag. If we fail,
655 * loop back and retry.
657 if ((x & LK_SHARED_WAITERS) == 0) {
658 if (!atomic_cmpset_acq_ptr(&lk->lk_lock, x,
659 x | LK_SHARED_WAITERS)) {
660 sleepq_release(&lk->lock_object);
663 LOCK_LOG2(lk, "%s: %p set shared waiters flag",
668 * As far as we have been unable to acquire the
669 * shared lock and the shared waiters flag is set,
672 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
674 flags &= ~LK_INTERLOCK;
677 "%s: interrupted sleep for %p with %d",
678 __func__, lk, error);
681 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
685 lock_profile_obtain_lock_success(&lk->lock_object,
686 contested, waittime, file, line);
687 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file,
689 WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file,
691 TD_LOCKS_INC(curthread);
692 TD_SLOCKS_INC(curthread);
697 _lockmgr_assert(lk, KA_SLOCKED, file, line);
699 x = v & LK_ALL_WAITERS;
700 v &= LK_EXCLUSIVE_SPINNERS;
703 * Try to switch from one shared lock to an exclusive one.
704 * We need to preserve waiters flags during the operation.
706 if (atomic_cmpset_ptr(&lk->lk_lock, LK_SHARERS_LOCK(1) | x | v,
708 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file,
710 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE |
711 LK_TRYWIT(flags), file, line);
712 TD_SLOCKS_DEC(curthread);
717 * We have been unable to succeed in upgrading, so just
718 * give up the shared lock.
720 wakeup_swapper |= wakeupshlk(lk, file, line);
724 if (LK_CAN_WITNESS(flags))
725 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
726 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
730 * If curthread already holds the lock and this one is
731 * allowed to recurse, simply recurse on it.
733 if (lockmgr_xlocked(lk)) {
734 if ((flags & LK_CANRECURSE) == 0 &&
735 (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) {
738 * If the lock is expected to not panic just
739 * give up and return.
741 if (LK_TRYOP(flags)) {
743 "%s: %p fails the try operation",
748 if (flags & LK_INTERLOCK)
749 class->lc_unlock(ilk);
750 panic("%s: recursing on non recursive lockmgr %s @ %s:%d\n",
751 __func__, iwmesg, file, line);
754 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk);
755 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
756 lk->lk_recurse, file, line);
757 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
758 LK_TRYWIT(flags), file, line);
759 TD_LOCKS_INC(curthread);
763 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
766 PMC_SOFT_CALL( , , lock, failed);
768 lock_profile_obtain_lock_failed(&lk->lock_object,
769 &contested, &waittime);
772 * If the lock is expected to not sleep just give up
775 if (LK_TRYOP(flags)) {
776 LOCK_LOG2(lk, "%s: %p fails the try operation",
782 #ifdef ADAPTIVE_LOCKMGRS
784 * If the owner is running on another CPU, spin until
785 * the owner stops running or the state of the lock
789 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
790 LK_HOLDER(x) != LK_KERNPROC) {
791 owner = (struct thread *)LK_HOLDER(x);
792 if (LOCK_LOG_TEST(&lk->lock_object, 0))
794 "%s: spinning on %p held by %p",
795 __func__, lk, owner);
798 * If we are holding also an interlock drop it
799 * in order to avoid a deadlock if the lockmgr
800 * owner is adaptively spinning on the
803 if (flags & LK_INTERLOCK) {
804 class->lc_unlock(ilk);
805 flags &= ~LK_INTERLOCK;
808 while (LK_HOLDER(lk->lk_lock) ==
809 (uintptr_t)owner && TD_IS_RUNNING(owner))
813 } else if (LK_CAN_ADAPT(lk, flags) &&
814 (x & LK_SHARE) != 0 && LK_SHARERS(x) &&
815 spintries < alk_retries) {
816 if ((x & LK_EXCLUSIVE_SPINNERS) == 0 &&
817 !atomic_cmpset_ptr(&lk->lk_lock, x,
818 x | LK_EXCLUSIVE_SPINNERS))
820 if (flags & LK_INTERLOCK) {
821 class->lc_unlock(ilk);
822 flags &= ~LK_INTERLOCK;
826 for (i = 0; i < alk_loops; i++) {
827 if (LOCK_LOG_TEST(&lk->lock_object, 0))
829 "%s: shared spinning on %p with %u and %u",
830 __func__, lk, spintries, i);
832 LK_EXCLUSIVE_SPINNERS) == 0)
843 * Acquire the sleepqueue chain lock because we
844 * probabilly will need to manipulate waiters flags.
846 sleepq_lock(&lk->lock_object);
850 * if the lock has been released while we spun on
851 * the sleepqueue chain lock just try again.
853 if (x == LK_UNLOCKED) {
854 sleepq_release(&lk->lock_object);
858 #ifdef ADAPTIVE_LOCKMGRS
860 * The current lock owner might have started executing
861 * on another CPU (or the lock could have changed
862 * owner) while we were waiting on the turnstile
863 * chain lock. If so, drop the turnstile lock and try
866 if (LK_CAN_ADAPT(lk, flags) && (x & LK_SHARE) == 0 &&
867 LK_HOLDER(x) != LK_KERNPROC) {
868 owner = (struct thread *)LK_HOLDER(x);
869 if (TD_IS_RUNNING(owner)) {
870 sleepq_release(&lk->lock_object);
877 * The lock can be in the state where there is a
878 * pending queue of waiters, but still no owner.
879 * This happens when the lock is contested and an
880 * owner is going to claim the lock.
881 * If curthread is the one successfully acquiring it
882 * claim lock ownership and return, preserving waiters
885 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
886 if ((x & ~v) == LK_UNLOCKED) {
887 v &= ~LK_EXCLUSIVE_SPINNERS;
888 if (atomic_cmpset_acq_ptr(&lk->lk_lock, x,
890 sleepq_release(&lk->lock_object);
892 "%s: %p claimed by a new writer",
896 sleepq_release(&lk->lock_object);
901 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we
902 * fail, loop back and retry.
904 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
905 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
906 x | LK_EXCLUSIVE_WAITERS)) {
907 sleepq_release(&lk->lock_object);
910 LOCK_LOG2(lk, "%s: %p set excl waiters flag",
915 * As far as we have been unable to acquire the
916 * exclusive lock and the exclusive waiters flag
917 * is set, we will sleep.
919 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo,
921 flags &= ~LK_INTERLOCK;
924 "%s: interrupted sleep for %p with %d",
925 __func__, lk, error);
928 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
932 lock_profile_obtain_lock_success(&lk->lock_object,
933 contested, waittime, file, line);
934 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0,
935 lk->lk_recurse, file, line);
936 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
937 LK_TRYWIT(flags), file, line);
938 TD_LOCKS_INC(curthread);
943 _lockmgr_assert(lk, KA_XLOCKED, file, line);
944 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line);
945 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line);
948 * Panic if the lock is recursed.
950 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
951 if (flags & LK_INTERLOCK)
952 class->lc_unlock(ilk);
953 panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n",
954 __func__, iwmesg, file, line);
956 TD_SLOCKS_INC(curthread);
959 * In order to preserve waiters flags, just spin.
963 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
965 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
966 LK_SHARERS_LOCK(1) | x))
972 _lockmgr_assert(lk, KA_LOCKED, file, line);
975 if ((x & LK_SHARE) == 0) {
978 * As first option, treact the lock as if it has not
980 * Fix-up the tid var if the lock has been disowned.
982 if (LK_HOLDER(x) == LK_KERNPROC)
985 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE,
987 TD_LOCKS_DEC(curthread);
989 LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0,
990 lk->lk_recurse, file, line);
993 * The lock is held in exclusive mode.
994 * If the lock is recursed also, then unrecurse it.
996 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) {
997 LOCK_LOG2(lk, "%s: %p unrecursing", __func__,
1002 if (tid != LK_KERNPROC)
1003 lock_profile_release_lock(&lk->lock_object);
1005 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid,
1009 sleepq_lock(&lk->lock_object);
1014 * If the lock has exclusive waiters, give them
1015 * preference in order to avoid deadlock with
1016 * shared runners up.
1017 * If interruptible sleeps left the exclusive queue
1018 * empty avoid a starvation for the threads sleeping
1019 * on the shared queue by giving them precedence
1020 * and cleaning up the exclusive waiters bit anyway.
1021 * Please note that lk_exslpfail count may be lying
1022 * about the real number of waiters with the
1023 * LK_SLEEPFAIL flag on because they may be used in
1024 * conjuction with interruptible sleeps so
1025 * lk_exslpfail might be considered an 'upper limit'
1026 * bound, including the edge cases.
1028 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1029 realexslp = sleepq_sleepcnt(&lk->lock_object,
1030 SQ_EXCLUSIVE_QUEUE);
1031 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) {
1032 if (lk->lk_exslpfail < realexslp) {
1033 lk->lk_exslpfail = 0;
1034 queue = SQ_EXCLUSIVE_QUEUE;
1035 v |= (x & LK_SHARED_WAITERS);
1037 lk->lk_exslpfail = 0;
1039 "%s: %p has only LK_SLEEPFAIL sleepers",
1042 "%s: %p waking up threads on the exclusive queue",
1045 sleepq_broadcast(&lk->lock_object,
1046 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE);
1047 queue = SQ_SHARED_QUEUE;
1052 * Exclusive waiters sleeping with LK_SLEEPFAIL
1053 * on and using interruptible sleeps/timeout
1054 * may have left spourious lk_exslpfail counts
1055 * on, so clean it up anyway.
1057 lk->lk_exslpfail = 0;
1058 queue = SQ_SHARED_QUEUE;
1062 "%s: %p waking up threads on the %s queue",
1063 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" :
1065 atomic_store_rel_ptr(&lk->lk_lock, v);
1066 wakeup_swapper |= sleepq_broadcast(&lk->lock_object,
1067 SLEEPQ_LK, 0, queue);
1068 sleepq_release(&lk->lock_object);
1071 wakeup_swapper = wakeupshlk(lk, file, line);
1074 if (LK_CAN_WITNESS(flags))
1075 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER |
1076 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ?
1080 * Trying to drain a lock we already own will result in a
1083 if (lockmgr_xlocked(lk)) {
1084 if (flags & LK_INTERLOCK)
1085 class->lc_unlock(ilk);
1086 panic("%s: draining %s with the lock held @ %s:%d\n",
1087 __func__, iwmesg, file, line);
1090 while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
1092 PMC_SOFT_CALL( , , lock, failed);
1094 lock_profile_obtain_lock_failed(&lk->lock_object,
1095 &contested, &waittime);
1098 * If the lock is expected to not sleep just give up
1101 if (LK_TRYOP(flags)) {
1102 LOCK_LOG2(lk, "%s: %p fails the try operation",
1109 * Acquire the sleepqueue chain lock because we
1110 * probabilly will need to manipulate waiters flags.
1112 sleepq_lock(&lk->lock_object);
1116 * if the lock has been released while we spun on
1117 * the sleepqueue chain lock just try again.
1119 if (x == LK_UNLOCKED) {
1120 sleepq_release(&lk->lock_object);
1124 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS);
1125 if ((x & ~v) == LK_UNLOCKED) {
1126 v = (x & ~LK_EXCLUSIVE_SPINNERS);
1129 * If interruptible sleeps left the exclusive
1130 * queue empty avoid a starvation for the
1131 * threads sleeping on the shared queue by
1132 * giving them precedence and cleaning up the
1133 * exclusive waiters bit anyway.
1134 * Please note that lk_exslpfail count may be
1135 * lying about the real number of waiters with
1136 * the LK_SLEEPFAIL flag on because they may
1137 * be used in conjuction with interruptible
1138 * sleeps so lk_exslpfail might be considered
1139 * an 'upper limit' bound, including the edge
1142 if (v & LK_EXCLUSIVE_WAITERS) {
1143 queue = SQ_EXCLUSIVE_QUEUE;
1144 v &= ~LK_EXCLUSIVE_WAITERS;
1148 * Exclusive waiters sleeping with
1149 * LK_SLEEPFAIL on and using
1150 * interruptible sleeps/timeout may
1151 * have left spourious lk_exslpfail
1152 * counts on, so clean it up anyway.
1154 MPASS(v & LK_SHARED_WAITERS);
1155 lk->lk_exslpfail = 0;
1156 queue = SQ_SHARED_QUEUE;
1157 v &= ~LK_SHARED_WAITERS;
1159 if (queue == SQ_EXCLUSIVE_QUEUE) {
1161 sleepq_sleepcnt(&lk->lock_object,
1162 SQ_EXCLUSIVE_QUEUE);
1163 if (lk->lk_exslpfail >= realexslp) {
1164 lk->lk_exslpfail = 0;
1165 queue = SQ_SHARED_QUEUE;
1166 v &= ~LK_SHARED_WAITERS;
1167 if (realexslp != 0) {
1169 "%s: %p has only LK_SLEEPFAIL sleepers",
1172 "%s: %p waking up threads on the exclusive queue",
1178 SQ_EXCLUSIVE_QUEUE);
1181 lk->lk_exslpfail = 0;
1183 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) {
1184 sleepq_release(&lk->lock_object);
1188 "%s: %p waking up all threads on the %s queue",
1189 __func__, lk, queue == SQ_SHARED_QUEUE ?
1190 "shared" : "exclusive");
1191 wakeup_swapper |= sleepq_broadcast(
1192 &lk->lock_object, SLEEPQ_LK, 0, queue);
1195 * If shared waiters have been woken up we need
1196 * to wait for one of them to acquire the lock
1197 * before to set the exclusive waiters in
1198 * order to avoid a deadlock.
1200 if (queue == SQ_SHARED_QUEUE) {
1201 for (v = lk->lk_lock;
1202 (v & LK_SHARE) && !LK_SHARERS(v);
1209 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we
1210 * fail, loop back and retry.
1212 if ((x & LK_EXCLUSIVE_WAITERS) == 0) {
1213 if (!atomic_cmpset_ptr(&lk->lk_lock, x,
1214 x | LK_EXCLUSIVE_WAITERS)) {
1215 sleepq_release(&lk->lock_object);
1218 LOCK_LOG2(lk, "%s: %p set drain waiters flag",
1223 * As far as we have been unable to acquire the
1224 * exclusive lock and the exclusive waiters flag
1225 * is set, we will sleep.
1227 if (flags & LK_INTERLOCK) {
1228 class->lc_unlock(ilk);
1229 flags &= ~LK_INTERLOCK;
1232 sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK,
1233 SQ_EXCLUSIVE_QUEUE);
1234 sleepq_wait(&lk->lock_object, ipri & PRIMASK);
1236 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
1241 lock_profile_obtain_lock_success(&lk->lock_object,
1242 contested, waittime, file, line);
1243 LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0,
1244 lk->lk_recurse, file, line);
1245 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE |
1246 LK_TRYWIT(flags), file, line);
1247 TD_LOCKS_INC(curthread);
1252 if (flags & LK_INTERLOCK)
1253 class->lc_unlock(ilk);
1254 panic("%s: unknown lockmgr request 0x%x\n", __func__, op);
1257 if (flags & LK_INTERLOCK)
1258 class->lc_unlock(ilk);
1266 _lockmgr_disown(struct lock *lk, const char *file, int line)
1270 if (SCHEDULER_STOPPED())
1273 tid = (uintptr_t)curthread;
1274 _lockmgr_assert(lk, KA_XLOCKED, file, line);
1277 * Panic if the lock is recursed.
1279 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk))
1280 panic("%s: disown a recursed lockmgr @ %s:%d\n",
1281 __func__, file, line);
1284 * If the owner is already LK_KERNPROC just skip the whole operation.
1286 if (LK_HOLDER(lk->lk_lock) != tid)
1288 lock_profile_release_lock(&lk->lock_object);
1289 LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line);
1290 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line);
1291 TD_LOCKS_DEC(curthread);
1295 * In order to preserve waiters flags, just spin.
1299 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0);
1300 x &= LK_ALL_WAITERS;
1301 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x,
1309 lockmgr_printinfo(const struct lock *lk)
1314 if (lk->lk_lock == LK_UNLOCKED)
1315 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name);
1316 else if (lk->lk_lock & LK_SHARE)
1317 printf("lock type %s: SHARED (count %ju)\n",
1318 lk->lock_object.lo_name,
1319 (uintmax_t)LK_SHARERS(lk->lk_lock));
1321 td = lockmgr_xholder(lk);
1322 printf("lock type %s: EXCL by thread %p "
1323 "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td,
1324 td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid);
1328 if (x & LK_EXCLUSIVE_WAITERS)
1329 printf(" with exclusive waiters pending\n");
1330 if (x & LK_SHARED_WAITERS)
1331 printf(" with shared waiters pending\n");
1332 if (x & LK_EXCLUSIVE_SPINNERS)
1333 printf(" with exclusive spinners pending\n");
1339 lockstatus(const struct lock *lk)
1348 if ((x & LK_SHARE) == 0) {
1349 if (v == (uintptr_t)curthread || v == LK_KERNPROC)
1353 } else if (x == LK_UNLOCKED)
1359 #ifdef INVARIANT_SUPPORT
1361 FEATURE(invariant_support,
1362 "Support for modules compiled with INVARIANTS option");
1365 #undef _lockmgr_assert
1369 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line)
1373 if (panicstr != NULL)
1377 case KA_SLOCKED | KA_NOTRECURSED:
1378 case KA_SLOCKED | KA_RECURSED:
1381 case KA_LOCKED | KA_NOTRECURSED:
1382 case KA_LOCKED | KA_RECURSED:
1386 * We cannot trust WITNESS if the lock is held in exclusive
1387 * mode and a call to lockmgr_disown() happened.
1388 * Workaround this skipping the check if the lock is held in
1389 * exclusive mode even for the KA_LOCKED case.
1391 if (slocked || (lk->lk_lock & LK_SHARE)) {
1392 witness_assert(&lk->lock_object, what, file, line);
1396 if (lk->lk_lock == LK_UNLOCKED ||
1397 ((lk->lk_lock & LK_SHARE) == 0 && (slocked ||
1398 (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)))))
1399 panic("Lock %s not %slocked @ %s:%d\n",
1400 lk->lock_object.lo_name, slocked ? "share" : "",
1403 if ((lk->lk_lock & LK_SHARE) == 0) {
1404 if (lockmgr_recursed(lk)) {
1405 if (what & KA_NOTRECURSED)
1406 panic("Lock %s recursed @ %s:%d\n",
1407 lk->lock_object.lo_name, file,
1409 } else if (what & KA_RECURSED)
1410 panic("Lock %s not recursed @ %s:%d\n",
1411 lk->lock_object.lo_name, file, line);
1415 case KA_XLOCKED | KA_NOTRECURSED:
1416 case KA_XLOCKED | KA_RECURSED:
1417 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))
1418 panic("Lock %s not exclusively locked @ %s:%d\n",
1419 lk->lock_object.lo_name, file, line);
1420 if (lockmgr_recursed(lk)) {
1421 if (what & KA_NOTRECURSED)
1422 panic("Lock %s recursed @ %s:%d\n",
1423 lk->lock_object.lo_name, file, line);
1424 } else if (what & KA_RECURSED)
1425 panic("Lock %s not recursed @ %s:%d\n",
1426 lk->lock_object.lo_name, file, line);
1429 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk))
1430 panic("Lock %s exclusively locked @ %s:%d\n",
1431 lk->lock_object.lo_name, file, line);
1434 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file,
1442 lockmgr_chain(struct thread *td, struct thread **ownerp)
1448 if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr)
1450 db_printf("blocked on lockmgr %s", lk->lock_object.lo_name);
1451 if (lk->lk_lock & LK_SHARE)
1452 db_printf("SHARED (count %ju)\n",
1453 (uintmax_t)LK_SHARERS(lk->lk_lock));
1455 db_printf("EXCL\n");
1456 *ownerp = lockmgr_xholder(lk);
1462 db_show_lockmgr(const struct lock_object *lock)
1465 const struct lock *lk;
1467 lk = (const struct lock *)lock;
1469 db_printf(" state: ");
1470 if (lk->lk_lock == LK_UNLOCKED)
1471 db_printf("UNLOCKED\n");
1472 else if (lk->lk_lock & LK_SHARE)
1473 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock));
1475 td = lockmgr_xholder(lk);
1476 if (td == (struct thread *)LK_KERNPROC)
1477 db_printf("XLOCK: LK_KERNPROC\n");
1479 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1480 td->td_tid, td->td_proc->p_pid,
1481 td->td_proc->p_comm);
1482 if (lockmgr_recursed(lk))
1483 db_printf(" recursed: %d\n", lk->lk_recurse);
1485 db_printf(" waiters: ");
1486 switch (lk->lk_lock & LK_ALL_WAITERS) {
1487 case LK_SHARED_WAITERS:
1488 db_printf("shared\n");
1490 case LK_EXCLUSIVE_WAITERS:
1491 db_printf("exclusive\n");
1493 case LK_ALL_WAITERS:
1494 db_printf("shared and exclusive\n");
1497 db_printf("none\n");
1499 db_printf(" spinners: ");
1500 if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS)
1501 db_printf("exclusive\n");
1503 db_printf("none\n");