2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
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, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Berkeley Software Design Inc's name may not be used to endorse or
15 * promote products derived from this software without specific prior
18 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
31 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35 * Machine independent bits of mutex implementation.
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
41 #include "opt_adaptive_mutexes.h"
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
51 #include <sys/kernel.h>
54 #include <sys/malloc.h>
55 #include <sys/mutex.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sched.h>
61 #include <sys/sysctl.h>
62 #include <sys/turnstile.h>
63 #include <sys/vmmeter.h>
64 #include <sys/lock_profile.h>
66 #include <machine/atomic.h>
67 #include <machine/bus.h>
68 #include <machine/cpu.h>
72 #include <fs/devfs/devfs_int.h>
75 #include <vm/vm_extern.h>
77 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
78 #define ADAPTIVE_MUTEXES
82 #include <sys/pmckern.h>
83 PMC_SOFT_DEFINE( , , lock, failed);
87 * Return the mutex address when the lock cookie address is provided.
88 * This functionality assumes that struct mtx* have a member named mtx_lock.
90 #define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock))
93 * Internal utility macros.
95 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
97 #define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
99 static void assert_mtx(const struct lock_object *lock, int what);
101 static void db_show_mtx(const struct lock_object *lock);
103 static void lock_mtx(struct lock_object *lock, uintptr_t how);
104 static void lock_spin(struct lock_object *lock, uintptr_t how);
106 static int owner_mtx(const struct lock_object *lock,
107 struct thread **owner);
109 static uintptr_t unlock_mtx(struct lock_object *lock);
110 static uintptr_t unlock_spin(struct lock_object *lock);
113 * Lock classes for sleep and spin mutexes.
115 struct lock_class lock_class_mtx_sleep = {
116 .lc_name = "sleep mutex",
117 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
118 .lc_assert = assert_mtx,
120 .lc_ddb_show = db_show_mtx,
123 .lc_unlock = unlock_mtx,
125 .lc_owner = owner_mtx,
128 struct lock_class lock_class_mtx_spin = {
129 .lc_name = "spin mutex",
130 .lc_flags = LC_SPINLOCK | LC_RECURSABLE,
131 .lc_assert = assert_mtx,
133 .lc_ddb_show = db_show_mtx,
135 .lc_lock = lock_spin,
136 .lc_unlock = unlock_spin,
138 .lc_owner = owner_mtx,
142 #ifdef ADAPTIVE_MUTEXES
143 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging");
145 static struct lock_delay_config __read_frequently mtx_delay;
147 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base,
149 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max,
152 LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay);
155 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL,
156 "mtx spin debugging");
158 static struct lock_delay_config __read_frequently mtx_spin_delay;
160 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW,
161 &mtx_spin_delay.base, 0, "");
162 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW,
163 &mtx_spin_delay.max, 0, "");
165 LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay);
168 * System-wide mutexes
170 struct mtx blocked_lock;
171 struct mtx __exclusive_cache_line Giant;
173 static void _mtx_lock_indefinite_check(struct mtx *, struct lock_delay_arg *);
176 assert_mtx(const struct lock_object *lock, int what)
180 * Treat LA_LOCKED as if LA_XLOCKED was asserted.
182 * Some callers of lc_assert uses LA_LOCKED to indicate that either
183 * a shared lock or write lock was held, while other callers uses
184 * the more strict LA_XLOCKED (used as MA_OWNED).
186 * Mutex is the only lock class that can not be shared, as a result,
187 * we can reasonably consider the caller really intends to assert
188 * LA_XLOCKED when they are asserting LA_LOCKED on a mutex object.
190 if (what & LA_LOCKED) {
194 mtx_assert((const struct mtx *)lock, what);
198 lock_mtx(struct lock_object *lock, uintptr_t how)
201 mtx_lock((struct mtx *)lock);
205 lock_spin(struct lock_object *lock, uintptr_t how)
208 panic("spin locks can only use msleep_spin");
212 unlock_mtx(struct lock_object *lock)
216 m = (struct mtx *)lock;
217 mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
223 unlock_spin(struct lock_object *lock)
226 panic("spin locks can only use msleep_spin");
231 owner_mtx(const struct lock_object *lock, struct thread **owner)
236 m = (const struct mtx *)lock;
238 *owner = (struct thread *)(x & ~MTX_FLAGMASK);
239 return (*owner != NULL);
244 * Function versions of the inlined __mtx_* macros. These are used by
245 * modules and can also be called from assembly language if needed.
248 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
255 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
256 !TD_IS_IDLETHREAD(curthread),
257 ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
258 curthread, m->lock_object.lo_name, file, line));
259 KASSERT(m->mtx_lock != MTX_DESTROYED,
260 ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
261 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
262 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
264 WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
265 LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
267 tid = (uintptr_t)curthread;
269 if (!_mtx_obtain_lock_fetch(m, &v, tid))
270 _mtx_lock_sleep(m, v, opts, file, line);
272 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
273 m, 0, 0, file, line);
274 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
276 WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
278 TD_LOCKS_INC(curthread);
282 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
288 KASSERT(m->mtx_lock != MTX_DESTROYED,
289 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
290 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
291 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
293 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
294 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
296 mtx_assert(m, MA_OWNED);
298 #ifdef LOCK_PROFILING
299 __mtx_unlock_sleep(c, (uintptr_t)curthread, opts, file, line);
301 __mtx_unlock(m, curthread, opts, file, line);
303 TD_LOCKS_DEC(curthread);
307 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
317 KASSERT(m->mtx_lock != MTX_DESTROYED,
318 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
319 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
320 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
321 m->lock_object.lo_name, file, line));
323 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
324 (opts & MTX_RECURSE) != 0,
325 ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
326 m->lock_object.lo_name, file, line));
327 opts &= ~MTX_RECURSE;
328 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
332 tid = (uintptr_t)curthread;
334 if (!_mtx_obtain_lock_fetch(m, &v, tid))
335 _mtx_lock_spin(m, v, opts, file, line);
337 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire,
338 m, 0, 0, file, line);
340 __mtx_lock_spin(m, curthread, opts, file, line);
342 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
344 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
348 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
353 if (SCHEDULER_STOPPED())
358 KASSERT(m->mtx_lock != MTX_DESTROYED,
359 ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line));
360 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
361 ("mtx_trylock_spin() of sleep mutex %s @ %s:%d",
362 m->lock_object.lo_name, file, line));
363 KASSERT((opts & MTX_RECURSE) == 0,
364 ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n",
365 m->lock_object.lo_name, file, line));
366 if (__mtx_trylock_spin(m, curthread, opts, file, line)) {
367 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line);
368 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
371 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line);
376 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
383 KASSERT(m->mtx_lock != MTX_DESTROYED,
384 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
385 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
386 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
387 m->lock_object.lo_name, file, line));
388 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
389 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
391 mtx_assert(m, MA_OWNED);
393 __mtx_unlock_spin(m);
397 * The important part of mtx_trylock{,_flags}()
398 * Tries to acquire lock `m.' If this function is called on a mutex that
399 * is already owned, it will recursively acquire the lock.
402 _mtx_trylock_flags_int(struct mtx *m, int opts LOCK_FILE_LINE_ARG_DEF)
406 #ifdef LOCK_PROFILING
407 uint64_t waittime = 0;
415 if (SCHEDULER_STOPPED_TD(td))
418 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
419 ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
420 curthread, m->lock_object.lo_name, file, line));
421 KASSERT(m->mtx_lock != MTX_DESTROYED,
422 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
423 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
424 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
431 if (_mtx_obtain_lock_fetch(m, &v, tid))
433 if (v == MTX_UNOWNED)
436 ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
437 (opts & MTX_RECURSE) != 0)) {
439 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
447 opts &= ~MTX_RECURSE;
449 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
451 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
453 TD_LOCKS_INC(curthread);
455 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
456 m, contested, waittime, file, line);
463 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
468 return (_mtx_trylock_flags_int(m, opts LOCK_FILE_LINE_ARG));
472 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
474 * We call this if the lock is either contested (i.e. we need to go to
475 * sleep waiting for it), or if we need to recurse on it.
479 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, int opts, const char *file,
483 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v)
488 struct turnstile *ts;
490 struct thread *owner;
491 #ifdef LOCK_PROFILING
493 uint64_t waittime = 0;
495 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
496 struct lock_delay_arg lda;
500 int64_t sleep_time = 0;
501 int64_t all_time = 0;
503 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
504 int doing_lockprof = 0;
512 if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
513 while (v == MTX_UNOWNED) {
514 if (_mtx_obtain_lock_fetch(m, &v, tid))
518 all_time -= lockstat_nsecs(&m->lock_object);
521 #ifdef LOCK_PROFILING
525 if (SCHEDULER_STOPPED_TD(td))
528 #if defined(ADAPTIVE_MUTEXES)
529 lock_delay_arg_init(&lda, &mtx_delay);
530 #elif defined(KDTRACE_HOOKS)
531 lock_delay_arg_init(&lda, NULL);
534 if (__predict_false(v == MTX_UNOWNED))
535 v = MTX_READ_VALUE(m);
537 if (__predict_false(lv_mtx_owner(v) == td)) {
538 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
539 (opts & MTX_RECURSE) != 0,
540 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
541 m->lock_object.lo_name, file, line));
543 opts &= ~MTX_RECURSE;
546 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
547 if (LOCK_LOG_TEST(&m->lock_object, opts))
548 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
552 opts &= ~MTX_RECURSE;
556 PMC_SOFT_CALL( , , lock, failed);
558 lock_profile_obtain_lock_failed(&m->lock_object,
559 &contested, &waittime);
560 if (LOCK_LOG_TEST(&m->lock_object, opts))
562 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
563 m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
566 if (v == MTX_UNOWNED) {
567 if (_mtx_obtain_lock_fetch(m, &v, tid))
574 #ifdef ADAPTIVE_MUTEXES
576 * If the owner is running on another CPU, spin until the
577 * owner stops running or the state of the lock changes.
579 owner = lv_mtx_owner(v);
580 if (TD_IS_RUNNING(owner)) {
581 if (LOCK_LOG_TEST(&m->lock_object, 0))
583 "%s: spinning on %p held by %p",
585 KTR_STATE1(KTR_SCHED, "thread",
586 sched_tdname((struct thread *)tid),
587 "spinning", "lockname:\"%s\"",
588 m->lock_object.lo_name);
591 v = MTX_READ_VALUE(m);
592 owner = lv_mtx_owner(v);
593 } while (v != MTX_UNOWNED && TD_IS_RUNNING(owner));
594 KTR_STATE0(KTR_SCHED, "thread",
595 sched_tdname((struct thread *)tid),
601 ts = turnstile_trywait(&m->lock_object);
602 v = MTX_READ_VALUE(m);
606 * Check if the lock has been released while spinning for
607 * the turnstile chain lock.
609 if (v == MTX_UNOWNED) {
610 turnstile_cancel(ts);
614 #ifdef ADAPTIVE_MUTEXES
616 * The current lock owner might have started executing
617 * on another CPU (or the lock could have changed
618 * owners) while we were waiting on the turnstile
619 * chain lock. If so, drop the turnstile lock and try
622 owner = lv_mtx_owner(v);
623 if (TD_IS_RUNNING(owner)) {
624 turnstile_cancel(ts);
630 * If the mutex isn't already contested and a failure occurs
631 * setting the contested bit, the mutex was either released
632 * or the state of the MTX_RECURSED bit changed.
634 if ((v & MTX_CONTESTED) == 0 &&
635 !atomic_fcmpset_ptr(&m->mtx_lock, &v, v | MTX_CONTESTED)) {
636 goto retry_turnstile;
640 * We definitely must sleep for this lock.
642 mtx_assert(m, MA_NOTOWNED);
645 * Block on the turnstile.
648 sleep_time -= lockstat_nsecs(&m->lock_object);
650 #ifndef ADAPTIVE_MUTEXES
651 owner = mtx_owner(m);
653 MPASS(owner == mtx_owner(m));
654 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
656 sleep_time += lockstat_nsecs(&m->lock_object);
659 v = MTX_READ_VALUE(m);
661 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
662 if (__predict_true(!doing_lockprof))
666 all_time += lockstat_nsecs(&m->lock_object);
668 LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
671 * Only record the loops spinning and not sleeping.
673 if (lda.spin_cnt > sleep_cnt)
674 LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
677 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
678 waittime, file, line);
683 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
685 * This is only called if we need to actually spin for the lock. Recursion
690 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, int opts,
691 const char *file, int line)
694 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v)
698 struct lock_delay_arg lda;
700 #ifdef LOCK_PROFILING
702 uint64_t waittime = 0;
705 int64_t spin_time = 0;
707 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
708 int doing_lockprof = 0;
711 tid = (uintptr_t)curthread;
715 if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
716 while (v == MTX_UNOWNED) {
717 if (_mtx_obtain_lock_fetch(m, &v, tid))
721 spin_time -= lockstat_nsecs(&m->lock_object);
724 #ifdef LOCK_PROFILING
728 if (__predict_false(v == MTX_UNOWNED))
729 v = MTX_READ_VALUE(m);
731 if (__predict_false(v == tid)) {
736 if (SCHEDULER_STOPPED())
739 lock_delay_arg_init(&lda, &mtx_spin_delay);
741 if (LOCK_LOG_TEST(&m->lock_object, opts))
742 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
743 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
744 "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
747 PMC_SOFT_CALL( , , lock, failed);
749 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
752 if (v == MTX_UNOWNED) {
753 if (_mtx_obtain_lock_fetch(m, &v, tid))
757 /* Give interrupts a chance while we spin. */
760 if (__predict_true(lda.spin_cnt < 10000000)) {
763 _mtx_lock_indefinite_check(m, &lda);
765 v = MTX_READ_VALUE(m);
766 } while (v != MTX_UNOWNED);
770 if (LOCK_LOG_TEST(&m->lock_object, opts))
771 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
772 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
775 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
776 if (__predict_true(!doing_lockprof))
780 spin_time += lockstat_nsecs(&m->lock_object);
781 if (lda.spin_cnt != 0)
782 LOCKSTAT_RECORD1(spin__spin, m, spin_time);
785 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
786 contested, waittime, file, line);
792 thread_lock_validate(struct mtx *m, int opts, const char *file, int line)
795 KASSERT(m->mtx_lock != MTX_DESTROYED,
796 ("thread_lock() of destroyed mutex @ %s:%d", file, line));
797 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
798 ("thread_lock() of sleep mutex %s @ %s:%d",
799 m->lock_object.lo_name, file, line));
801 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
802 ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
803 m->lock_object.lo_name, file, line));
804 WITNESS_CHECKORDER(&m->lock_object,
805 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
808 #define thread_lock_validate(m, opts, file, line) do { } while (0)
811 #ifndef LOCK_PROFILING
814 _thread_lock(struct thread *td, int opts, const char *file, int line)
817 _thread_lock(struct thread *td)
823 tid = (uintptr_t)curthread;
825 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(spin__acquire)))
829 thread_lock_validate(m, 0, file, line);
830 v = MTX_READ_VALUE(m);
831 if (__predict_true(v == MTX_UNOWNED)) {
832 if (__predict_false(!_mtx_obtain_lock(m, tid)))
833 goto slowpath_unlocked;
834 } else if (v == tid) {
837 goto slowpath_unlocked;
838 if (__predict_true(m == td->td_lock)) {
839 WITNESS_LOCK(&m->lock_object, LOP_EXCLUSIVE, file, line);
842 MPASS(m->mtx_recurse == 0);
843 _mtx_release_lock_quick(m);
848 thread_lock_flags_(td, opts, file, line);
850 thread_lock_flags_(td, 0, 0, 0);
856 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
860 struct lock_delay_arg lda;
861 #ifdef LOCK_PROFILING
863 uint64_t waittime = 0;
866 int64_t spin_time = 0;
868 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
869 int doing_lockprof = 1;
872 tid = (uintptr_t)curthread;
874 if (SCHEDULER_STOPPED()) {
876 * Ensure that spinlock sections are balanced even when the
877 * scheduler is stopped, since we may otherwise inadvertently
878 * re-enable interrupts while dumping core.
884 lock_delay_arg_init(&lda, &mtx_spin_delay);
887 PMC_SOFT_CALL( , , lock, failed);
890 #ifdef LOCK_PROFILING
892 #elif defined(KDTRACE_HOOKS)
893 doing_lockprof = lockstat_enabled;
894 if (__predict_false(doing_lockprof))
895 spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
902 thread_lock_validate(m, opts, file, line);
903 v = MTX_READ_VALUE(m);
905 if (v == MTX_UNOWNED) {
906 if (_mtx_obtain_lock_fetch(m, &v, tid))
912 MPASS(m == td->td_lock);
915 lock_profile_obtain_lock_failed(&m->lock_object,
916 &contested, &waittime);
917 /* Give interrupts a chance while we spin. */
920 if (__predict_true(lda.spin_cnt < 10000000)) {
923 _mtx_lock_indefinite_check(m, &lda);
925 if (m != td->td_lock) {
929 v = MTX_READ_VALUE(m);
930 } while (v != MTX_UNOWNED);
933 if (m == td->td_lock)
935 MPASS(m->mtx_recurse == 0);
936 _mtx_release_lock_quick(m);
938 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
940 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
942 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
943 if (__predict_true(!doing_lockprof))
947 spin_time += lockstat_nsecs(&m->lock_object);
949 if (m->mtx_recurse == 0)
950 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
951 contested, waittime, file, line);
953 if (lda.spin_cnt != 0)
954 LOCKSTAT_RECORD1(thread__spin, m, spin_time);
959 thread_lock_block(struct thread *td)
963 THREAD_LOCK_ASSERT(td, MA_OWNED);
965 td->td_lock = &blocked_lock;
966 mtx_unlock_spin(lock);
972 thread_lock_unblock(struct thread *td, struct mtx *new)
974 mtx_assert(new, MA_OWNED);
975 MPASS(td->td_lock == &blocked_lock);
976 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
980 thread_lock_set(struct thread *td, struct mtx *new)
984 mtx_assert(new, MA_OWNED);
985 THREAD_LOCK_ASSERT(td, MA_OWNED);
988 mtx_unlock_spin(lock);
992 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
994 * We are only called here if the lock is recursed, contested (i.e. we
995 * need to wake up a blocked thread) or lockstat probe is active.
999 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v, int opts,
1000 const char *file, int line)
1003 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v)
1007 struct turnstile *ts;
1010 if (SCHEDULER_STOPPED())
1013 tid = (uintptr_t)curthread;
1016 if (__predict_false(v == tid))
1017 v = MTX_READ_VALUE(m);
1019 if (__predict_false(v & MTX_RECURSED)) {
1020 if (--(m->mtx_recurse) == 0)
1021 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
1022 if (LOCK_LOG_TEST(&m->lock_object, opts))
1023 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
1027 LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m);
1028 if (v == tid && _mtx_release_lock(m, tid))
1032 * We have to lock the chain before the turnstile so this turnstile
1033 * can be removed from the hash list if it is empty.
1035 turnstile_chain_lock(&m->lock_object);
1036 _mtx_release_lock_quick(m);
1037 ts = turnstile_lookup(&m->lock_object);
1039 if (LOCK_LOG_TEST(&m->lock_object, opts))
1040 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
1041 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
1044 * This turnstile is now no longer associated with the mutex. We can
1045 * unlock the chain lock so a new turnstile may take it's place.
1047 turnstile_unpend(ts);
1048 turnstile_chain_unlock(&m->lock_object);
1052 * All the unlocking of MTX_SPIN locks is done inline.
1053 * See the __mtx_unlock_spin() macro for the details.
1057 * The backing function for the INVARIANTS-enabled mtx_assert()
1059 #ifdef INVARIANT_SUPPORT
1061 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1063 const struct mtx *m;
1065 if (panicstr != NULL || dumping || SCHEDULER_STOPPED())
1072 case MA_OWNED | MA_RECURSED:
1073 case MA_OWNED | MA_NOTRECURSED:
1075 panic("mutex %s not owned at %s:%d",
1076 m->lock_object.lo_name, file, line);
1077 if (mtx_recursed(m)) {
1078 if ((what & MA_NOTRECURSED) != 0)
1079 panic("mutex %s recursed at %s:%d",
1080 m->lock_object.lo_name, file, line);
1081 } else if ((what & MA_RECURSED) != 0) {
1082 panic("mutex %s unrecursed at %s:%d",
1083 m->lock_object.lo_name, file, line);
1088 panic("mutex %s owned at %s:%d",
1089 m->lock_object.lo_name, file, line);
1092 panic("unknown mtx_assert at %s:%d", file, line);
1098 * General init routine used by the MTX_SYSINIT() macro.
1101 mtx_sysinit(void *arg)
1103 struct mtx_args *margs = arg;
1105 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
1110 * Mutex initialization routine; initialize lock `m' of type contained in
1111 * `opts' with options contained in `opts' and name `name.' The optional
1112 * lock type `type' is used as a general lock category name for use with
1116 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
1119 struct lock_class *class;
1124 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
1125 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
1126 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
1127 ("%s: mtx_lock not aligned for %s: %p", __func__, name,
1130 /* Determine lock class and lock flags. */
1131 if (opts & MTX_SPIN)
1132 class = &lock_class_mtx_spin;
1134 class = &lock_class_mtx_sleep;
1136 if (opts & MTX_QUIET)
1138 if (opts & MTX_RECURSE)
1139 flags |= LO_RECURSABLE;
1140 if ((opts & MTX_NOWITNESS) == 0)
1141 flags |= LO_WITNESS;
1142 if (opts & MTX_DUPOK)
1144 if (opts & MTX_NOPROFILE)
1145 flags |= LO_NOPROFILE;
1149 /* Initialize mutex. */
1150 lock_init(&m->lock_object, class, name, type, flags);
1152 m->mtx_lock = MTX_UNOWNED;
1157 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
1158 * passed in as a flag here because if the corresponding mtx_init() was
1159 * called with MTX_QUIET set, then it will already be set in the mutex's
1163 _mtx_destroy(volatile uintptr_t *c)
1170 MPASS(mtx_unowned(m));
1172 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
1174 /* Perform the non-mtx related part of mtx_unlock_spin(). */
1175 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
1178 TD_LOCKS_DEC(curthread);
1180 lock_profile_release_lock(&m->lock_object);
1181 /* Tell witness this isn't locked to make it happy. */
1182 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
1186 m->mtx_lock = MTX_DESTROYED;
1187 lock_destroy(&m->lock_object);
1191 * Intialize the mutex code and system mutexes. This is called from the MD
1192 * startup code prior to mi_startup(). The per-CPU data space needs to be
1193 * setup before this is called.
1199 /* Setup turnstiles so that sleep mutexes work. */
1203 * Initialize mutexes.
1205 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
1206 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
1207 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */
1208 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
1209 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
1210 mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
1211 mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1212 mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1213 mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1217 static void __noinline
1218 _mtx_lock_indefinite_check(struct mtx *m, struct lock_delay_arg *ldap)
1223 if (ldap->spin_cnt < 60000000 || kdb_active || panicstr != NULL)
1228 /* If the mutex is unlocked, try again. */
1232 printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
1233 m, m->lock_object.lo_name, td, td->td_tid);
1235 witness_display_spinlock(&m->lock_object, td, printf);
1237 panic("spin lock held too long");
1243 mtx_spin_wait_unlocked(struct mtx *m)
1245 struct lock_delay_arg lda;
1247 KASSERT(m->mtx_lock != MTX_DESTROYED,
1248 ("%s() of destroyed mutex %p", __func__, m));
1249 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
1250 ("%s() of sleep mutex %p (%s)", __func__, m,
1251 m->lock_object.lo_name));
1252 KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1253 m->lock_object.lo_name));
1257 while (atomic_load_acq_ptr(&m->mtx_lock) != MTX_UNOWNED) {
1258 if (__predict_true(lda.spin_cnt < 10000000)) {
1262 _mtx_lock_indefinite_check(m, &lda);
1269 db_show_mtx(const struct lock_object *lock)
1272 const struct mtx *m;
1274 m = (const struct mtx *)lock;
1276 db_printf(" flags: {");
1277 if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1281 if (m->lock_object.lo_flags & LO_RECURSABLE)
1282 db_printf(", RECURSE");
1283 if (m->lock_object.lo_flags & LO_DUPOK)
1284 db_printf(", DUPOK");
1286 db_printf(" state: {");
1288 db_printf("UNOWNED");
1289 else if (mtx_destroyed(m))
1290 db_printf("DESTROYED");
1293 if (m->mtx_lock & MTX_CONTESTED)
1294 db_printf(", CONTESTED");
1295 if (m->mtx_lock & MTX_RECURSED)
1296 db_printf(", RECURSED");
1299 if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1301 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1302 td->td_tid, td->td_proc->p_pid, td->td_name);
1303 if (mtx_recursed(m))
1304 db_printf(" recursed: %d\n", m->mtx_recurse);