1/*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
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
16 *    written permission.
17 *
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
28 * SUCH DAMAGE.
29 *
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 $
32 */
33
34/*
35 * Machine independent bits of mutex implementation.
36 */
37
38#include <sys/cdefs.h>
39#include "opt_adaptive_mutexes.h"
40#include "opt_ddb.h"
41#include "opt_hwpmc_hooks.h"
42#include "opt_sched.h"
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/bus.h>
47#include <sys/conf.h>
48#include <sys/kdb.h>
49#include <sys/kernel.h>
50#include <sys/ktr.h>
51#include <sys/lock.h>
52#include <sys/malloc.h>
53#include <sys/mutex.h>
54#include <sys/proc.h>
55#include <sys/resourcevar.h>
56#include <sys/sched.h>
57#include <sys/sbuf.h>
58#include <sys/smp.h>
59#include <sys/sysctl.h>
60#include <sys/turnstile.h>
61#include <sys/vmmeter.h>
62#include <sys/lock_profile.h>
63
64#include <machine/atomic.h>
65#include <machine/bus.h>
66#include <machine/cpu.h>
67
68#include <ddb/ddb.h>
69
70#include <fs/devfs/devfs_int.h>
71
72#include <vm/vm.h>
73#include <vm/vm_extern.h>
74
75#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
76#define	ADAPTIVE_MUTEXES
77#endif
78
79#ifdef HWPMC_HOOKS
80#include <sys/pmckern.h>
81PMC_SOFT_DEFINE( , , lock, failed);
82#endif
83
84/*
85 * Return the mutex address when the lock cookie address is provided.
86 * This functionality assumes that struct mtx* have a member named mtx_lock.
87 */
88#define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
89
90/*
91 * Internal utility macros.
92 */
93#define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
94
95#define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
96
97static void	assert_mtx(const struct lock_object *lock, int what);
98#ifdef DDB
99static void	db_show_mtx(const struct lock_object *lock);
100#endif
101static void	lock_mtx(struct lock_object *lock, uintptr_t how);
102static void	lock_spin(struct lock_object *lock, uintptr_t how);
103#ifdef KDTRACE_HOOKS
104static int	owner_mtx(const struct lock_object *lock,
105		    struct thread **owner);
106#endif
107static uintptr_t unlock_mtx(struct lock_object *lock);
108static uintptr_t unlock_spin(struct lock_object *lock);
109
110/*
111 * Lock classes for sleep and spin mutexes.
112 */
113struct lock_class lock_class_mtx_sleep = {
114	.lc_name = "sleep mutex",
115	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
116	.lc_assert = assert_mtx,
117#ifdef DDB
118	.lc_ddb_show = db_show_mtx,
119#endif
120	.lc_lock = lock_mtx,
121	.lc_unlock = unlock_mtx,
122#ifdef KDTRACE_HOOKS
123	.lc_owner = owner_mtx,
124#endif
125};
126struct lock_class lock_class_mtx_spin = {
127	.lc_name = "spin mutex",
128	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
129	.lc_assert = assert_mtx,
130#ifdef DDB
131	.lc_ddb_show = db_show_mtx,
132#endif
133	.lc_lock = lock_spin,
134	.lc_unlock = unlock_spin,
135#ifdef KDTRACE_HOOKS
136	.lc_owner = owner_mtx,
137#endif
138};
139
140#ifdef ADAPTIVE_MUTEXES
141#ifdef MUTEX_CUSTOM_BACKOFF
142static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
143    "mtx debugging");
144
145static struct lock_delay_config __read_frequently mtx_delay;
146
147SYSCTL_U16(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base,
148    0, "");
149SYSCTL_U16(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max,
150    0, "");
151
152LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay);
153#else
154#define mtx_delay	locks_delay
155#endif
156#endif
157
158#ifdef MUTEX_SPIN_CUSTOM_BACKOFF
159static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin,
160    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
161    "mtx spin debugging");
162
163static struct lock_delay_config __read_frequently mtx_spin_delay;
164
165SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW,
166    &mtx_spin_delay.base, 0, "");
167SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW,
168    &mtx_spin_delay.max, 0, "");
169
170LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay);
171#else
172#define mtx_spin_delay	locks_delay
173#endif
174
175/*
176 * System-wide mutexes
177 */
178struct mtx blocked_lock;
179struct mtx __exclusive_cache_line Giant;
180
181static void _mtx_lock_indefinite_check(struct mtx *, struct lock_delay_arg *);
182
183void
184assert_mtx(const struct lock_object *lock, int what)
185{
186
187	/*
188	 * Treat LA_LOCKED as if LA_XLOCKED was asserted.
189	 *
190	 * Some callers of lc_assert uses LA_LOCKED to indicate that either
191	 * a shared lock or write lock was held, while other callers uses
192	 * the more strict LA_XLOCKED (used as MA_OWNED).
193	 *
194	 * Mutex is the only lock class that can not be shared, as a result,
195	 * we can reasonably consider the caller really intends to assert
196	 * LA_XLOCKED when they are asserting LA_LOCKED on a mutex object.
197	 */
198	if (what & LA_LOCKED) {
199		what &= ~LA_LOCKED;
200		what |= LA_XLOCKED;
201	}
202	mtx_assert((const struct mtx *)lock, what);
203}
204
205void
206lock_mtx(struct lock_object *lock, uintptr_t how)
207{
208
209	mtx_lock((struct mtx *)lock);
210}
211
212void
213lock_spin(struct lock_object *lock, uintptr_t how)
214{
215
216	mtx_lock_spin((struct mtx *)lock);
217}
218
219uintptr_t
220unlock_mtx(struct lock_object *lock)
221{
222	struct mtx *m;
223
224	m = (struct mtx *)lock;
225	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
226	mtx_unlock(m);
227	return (0);
228}
229
230uintptr_t
231unlock_spin(struct lock_object *lock)
232{
233	struct mtx *m;
234
235	m = (struct mtx *)lock;
236	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
237	mtx_unlock_spin(m);
238	return (0);
239}
240
241#ifdef KDTRACE_HOOKS
242int
243owner_mtx(const struct lock_object *lock, struct thread **owner)
244{
245	const struct mtx *m;
246	uintptr_t x;
247
248	m = (const struct mtx *)lock;
249	x = m->mtx_lock;
250	*owner = (struct thread *)(x & ~MTX_FLAGMASK);
251	return (*owner != NULL);
252}
253#endif
254
255/*
256 * Function versions of the inlined __mtx_* macros.  These are used by
257 * modules and can also be called from assembly language if needed.
258 */
259void
260__mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
261{
262	struct mtx *m;
263	uintptr_t tid, v;
264
265	m = mtxlock2mtx(c);
266
267	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
268	    !TD_IS_IDLETHREAD(curthread),
269	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
270	    curthread, m->lock_object.lo_name, file, line));
271	KASSERT(m->mtx_lock != MTX_DESTROYED,
272	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
273	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
274	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
275	    file, line));
276	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
277	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
278
279	tid = (uintptr_t)curthread;
280	v = MTX_UNOWNED;
281	if (!_mtx_obtain_lock_fetch(m, &v, tid))
282		_mtx_lock_sleep(m, v, opts, file, line);
283	else
284		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
285		    m, 0, 0, file, line);
286	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
287	    line);
288	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
289	    file, line);
290	TD_LOCKS_INC(curthread);
291}
292
293void
294__mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
295{
296	struct mtx *m;
297
298	m = mtxlock2mtx(c);
299
300	KASSERT(m->mtx_lock != MTX_DESTROYED,
301	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
302	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
303	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
304	    file, line));
305	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
306	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
307	    line);
308	mtx_assert(m, MA_OWNED);
309
310#ifdef LOCK_PROFILING
311	__mtx_unlock_sleep(c, (uintptr_t)curthread, opts, file, line);
312#else
313	__mtx_unlock(m, curthread, opts, file, line);
314#endif
315	TD_LOCKS_DEC(curthread);
316}
317
318void
319__mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
320    int line)
321{
322	struct mtx *m;
323#ifdef SMP
324	uintptr_t tid, v;
325#endif
326
327	m = mtxlock2mtx(c);
328
329	KASSERT(m->mtx_lock != MTX_DESTROYED,
330	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
331	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
332	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
333	    m->lock_object.lo_name, file, line));
334	if (mtx_owned(m))
335		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
336		    (opts & MTX_RECURSE) != 0,
337	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
338		    m->lock_object.lo_name, file, line));
339	opts &= ~MTX_RECURSE;
340	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
341	    file, line, NULL);
342#ifdef SMP
343	spinlock_enter();
344	tid = (uintptr_t)curthread;
345	v = MTX_UNOWNED;
346	if (!_mtx_obtain_lock_fetch(m, &v, tid))
347		_mtx_lock_spin(m, v, opts, file, line);
348	else
349		LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire,
350		    m, 0, 0, file, line);
351#else
352	__mtx_lock_spin(m, curthread, opts, file, line);
353#endif
354	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
355	    line);
356	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
357}
358
359int
360__mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
361    int line)
362{
363	struct mtx *m;
364
365	if (SCHEDULER_STOPPED())
366		return (1);
367
368	m = mtxlock2mtx(c);
369
370	KASSERT(m->mtx_lock != MTX_DESTROYED,
371	    ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line));
372	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
373	    ("mtx_trylock_spin() of sleep mutex %s @ %s:%d",
374	    m->lock_object.lo_name, file, line));
375	KASSERT((opts & MTX_RECURSE) == 0,
376	    ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n",
377	    m->lock_object.lo_name, file, line));
378	if (__mtx_trylock_spin(m, curthread, opts, file, line)) {
379		LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line);
380		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
381		return (1);
382	}
383	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line);
384	return (0);
385}
386
387void
388__mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
389    int line)
390{
391	struct mtx *m;
392
393	m = mtxlock2mtx(c);
394
395	KASSERT(m->mtx_lock != MTX_DESTROYED,
396	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
397	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
398	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
399	    m->lock_object.lo_name, file, line));
400	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
401	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
402	    line);
403	mtx_assert(m, MA_OWNED);
404
405	__mtx_unlock_spin(m);
406}
407
408/*
409 * The important part of mtx_trylock{,_flags}()
410 * Tries to acquire lock `m.'  If this function is called on a mutex that
411 * is already owned, it will recursively acquire the lock.
412 */
413int
414_mtx_trylock_flags_int(struct mtx *m, int opts LOCK_FILE_LINE_ARG_DEF)
415{
416	struct thread *td;
417	uintptr_t tid, v;
418#ifdef LOCK_PROFILING
419	uint64_t waittime = 0;
420	int contested = 0;
421#endif
422	int rval;
423	bool recursed;
424
425	td = curthread;
426	tid = (uintptr_t)td;
427	if (SCHEDULER_STOPPED())
428		return (1);
429
430	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
431	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
432	    curthread, m->lock_object.lo_name, file, line));
433	KASSERT(m->mtx_lock != MTX_DESTROYED,
434	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
435	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
436	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
437	    file, line));
438
439	rval = 1;
440	recursed = false;
441	v = MTX_UNOWNED;
442	for (;;) {
443		if (_mtx_obtain_lock_fetch(m, &v, tid))
444			break;
445		if (v == MTX_UNOWNED)
446			continue;
447		if (v == tid &&
448		    ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
449		    (opts & MTX_RECURSE) != 0)) {
450			m->mtx_recurse++;
451			atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
452			recursed = true;
453			break;
454		}
455		rval = 0;
456		break;
457	}
458
459	opts &= ~MTX_RECURSE;
460
461	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
462	if (rval) {
463		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
464		    file, line);
465		TD_LOCKS_INC(curthread);
466		if (!recursed)
467			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
468			    m, contested, waittime, file, line);
469	}
470
471	return (rval);
472}
473
474int
475_mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
476{
477	struct mtx *m;
478
479	m = mtxlock2mtx(c);
480	return (_mtx_trylock_flags_int(m, opts LOCK_FILE_LINE_ARG));
481}
482
483/*
484 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
485 *
486 * We call this if the lock is either contested (i.e. we need to go to
487 * sleep waiting for it), or if we need to recurse on it.
488 */
489#if LOCK_DEBUG > 0
490void
491__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, int opts, const char *file,
492    int line)
493#else
494void
495__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v)
496#endif
497{
498	struct thread *td;
499	struct mtx *m;
500	struct turnstile *ts;
501	uintptr_t tid;
502	struct thread *owner;
503#ifdef LOCK_PROFILING
504	int contested = 0;
505	uint64_t waittime = 0;
506#endif
507#if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
508	struct lock_delay_arg lda;
509#endif
510#ifdef KDTRACE_HOOKS
511	u_int sleep_cnt = 0;
512	int64_t sleep_time = 0;
513	int64_t all_time = 0;
514#endif
515#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
516	int doing_lockprof = 0;
517#endif
518
519	td = curthread;
520	tid = (uintptr_t)td;
521	m = mtxlock2mtx(c);
522
523#ifdef KDTRACE_HOOKS
524	if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
525		while (v == MTX_UNOWNED) {
526			if (_mtx_obtain_lock_fetch(m, &v, tid))
527				goto out_lockstat;
528		}
529		doing_lockprof = 1;
530		all_time -= lockstat_nsecs(&m->lock_object);
531	}
532#endif
533#ifdef LOCK_PROFILING
534	doing_lockprof = 1;
535#endif
536
537	if (SCHEDULER_STOPPED())
538		return;
539
540	if (__predict_false(v == MTX_UNOWNED))
541		v = MTX_READ_VALUE(m);
542
543	if (__predict_false(lv_mtx_owner(v) == td)) {
544		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
545		    (opts & MTX_RECURSE) != 0,
546	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
547		    m->lock_object.lo_name, file, line));
548#if LOCK_DEBUG > 0
549		opts &= ~MTX_RECURSE;
550#endif
551		m->mtx_recurse++;
552		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
553		if (LOCK_LOG_TEST(&m->lock_object, opts))
554			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
555		return;
556	}
557#if LOCK_DEBUG > 0
558	opts &= ~MTX_RECURSE;
559#endif
560
561#if defined(ADAPTIVE_MUTEXES)
562	lock_delay_arg_init(&lda, &mtx_delay);
563#elif defined(KDTRACE_HOOKS)
564	lock_delay_arg_init_noadapt(&lda);
565#endif
566
567#ifdef HWPMC_HOOKS
568	PMC_SOFT_CALL( , , lock, failed);
569#endif
570	lock_profile_obtain_lock_failed(&m->lock_object, false,
571		    &contested, &waittime);
572	if (LOCK_LOG_TEST(&m->lock_object, opts))
573		CTR4(KTR_LOCK,
574		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
575		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
576
577	THREAD_CONTENDS_ON_LOCK(&m->lock_object);
578
579	for (;;) {
580		if (v == MTX_UNOWNED) {
581			if (_mtx_obtain_lock_fetch(m, &v, tid))
582				break;
583			continue;
584		}
585#ifdef KDTRACE_HOOKS
586		lda.spin_cnt++;
587#endif
588#ifdef ADAPTIVE_MUTEXES
589		/*
590		 * If the owner is running on another CPU, spin until the
591		 * owner stops running or the state of the lock changes.
592		 */
593		owner = lv_mtx_owner(v);
594		if (TD_IS_RUNNING(owner)) {
595			if (LOCK_LOG_TEST(&m->lock_object, 0))
596				CTR3(KTR_LOCK,
597				    "%s: spinning on %p held by %p",
598				    __func__, m, owner);
599			KTR_STATE1(KTR_SCHED, "thread",
600			    sched_tdname((struct thread *)tid),
601			    "spinning", "lockname:\"%s\"",
602			    m->lock_object.lo_name);
603			do {
604				lock_delay(&lda);
605				v = MTX_READ_VALUE(m);
606				owner = lv_mtx_owner(v);
607			} while (v != MTX_UNOWNED && TD_IS_RUNNING(owner));
608			KTR_STATE0(KTR_SCHED, "thread",
609			    sched_tdname((struct thread *)tid),
610			    "running");
611			continue;
612		}
613#endif
614
615		ts = turnstile_trywait(&m->lock_object);
616		v = MTX_READ_VALUE(m);
617retry_turnstile:
618
619		/*
620		 * Check if the lock has been released while spinning for
621		 * the turnstile chain lock.
622		 */
623		if (v == MTX_UNOWNED) {
624			turnstile_cancel(ts);
625			continue;
626		}
627
628#ifdef ADAPTIVE_MUTEXES
629		/*
630		 * The current lock owner might have started executing
631		 * on another CPU (or the lock could have changed
632		 * owners) while we were waiting on the turnstile
633		 * chain lock.  If so, drop the turnstile lock and try
634		 * again.
635		 */
636		owner = lv_mtx_owner(v);
637		if (TD_IS_RUNNING(owner)) {
638			turnstile_cancel(ts);
639			continue;
640		}
641#endif
642
643		/*
644		 * If the mutex isn't already contested and a failure occurs
645		 * setting the contested bit, the mutex was either released
646		 * or the state of the MTX_RECURSED bit changed.
647		 */
648		if ((v & MTX_CONTESTED) == 0 &&
649		    !atomic_fcmpset_ptr(&m->mtx_lock, &v, v | MTX_CONTESTED)) {
650			goto retry_turnstile;
651		}
652
653		/*
654		 * We definitely must sleep for this lock.
655		 */
656		mtx_assert(m, MA_NOTOWNED);
657
658		/*
659		 * Block on the turnstile.
660		 */
661#ifdef KDTRACE_HOOKS
662		sleep_time -= lockstat_nsecs(&m->lock_object);
663#endif
664#ifndef ADAPTIVE_MUTEXES
665		owner = mtx_owner(m);
666#endif
667		MPASS(owner == mtx_owner(m));
668		turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
669#ifdef KDTRACE_HOOKS
670		sleep_time += lockstat_nsecs(&m->lock_object);
671		sleep_cnt++;
672#endif
673		v = MTX_READ_VALUE(m);
674	}
675	THREAD_CONTENTION_DONE(&m->lock_object);
676#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
677	if (__predict_true(!doing_lockprof))
678		return;
679#endif
680#ifdef KDTRACE_HOOKS
681	all_time += lockstat_nsecs(&m->lock_object);
682	if (sleep_time)
683		LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
684
685	/*
686	 * Only record the loops spinning and not sleeping.
687	 */
688	if (lda.spin_cnt > sleep_cnt)
689		LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
690out_lockstat:
691#endif
692	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
693	    waittime, file, line);
694}
695
696#ifdef SMP
697/*
698 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
699 *
700 * This is only called if we need to actually spin for the lock. Recursion
701 * is handled inline.
702 */
703#if LOCK_DEBUG > 0
704void
705_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, int opts,
706    const char *file, int line)
707#else
708void
709_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v)
710#endif
711{
712	struct mtx *m;
713	struct lock_delay_arg lda;
714	uintptr_t tid;
715#ifdef LOCK_PROFILING
716	int contested = 0;
717	uint64_t waittime = 0;
718#endif
719#ifdef KDTRACE_HOOKS
720	int64_t spin_time = 0;
721#endif
722#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
723	int doing_lockprof = 0;
724#endif
725
726	tid = (uintptr_t)curthread;
727	m = mtxlock2mtx(c);
728
729#ifdef KDTRACE_HOOKS
730	if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
731		while (v == MTX_UNOWNED) {
732			if (_mtx_obtain_lock_fetch(m, &v, tid))
733				goto out_lockstat;
734		}
735		doing_lockprof = 1;
736		spin_time -= lockstat_nsecs(&m->lock_object);
737	}
738#endif
739#ifdef LOCK_PROFILING
740	doing_lockprof = 1;
741#endif
742
743	if (__predict_false(v == MTX_UNOWNED))
744		v = MTX_READ_VALUE(m);
745
746	if (__predict_false(v == tid)) {
747		m->mtx_recurse++;
748		return;
749	}
750
751	if (SCHEDULER_STOPPED())
752		return;
753
754	if (LOCK_LOG_TEST(&m->lock_object, opts))
755		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
756	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
757	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
758
759	lock_delay_arg_init(&lda, &mtx_spin_delay);
760
761#ifdef HWPMC_HOOKS
762	PMC_SOFT_CALL( , , lock, failed);
763#endif
764	lock_profile_obtain_lock_failed(&m->lock_object, true, &contested, &waittime);
765
766	for (;;) {
767		if (v == MTX_UNOWNED) {
768			if (_mtx_obtain_lock_fetch(m, &v, tid))
769				break;
770			continue;
771		}
772		/* Give interrupts a chance while we spin. */
773		spinlock_exit();
774		do {
775			if (__predict_true(lda.spin_cnt < 10000000)) {
776				lock_delay(&lda);
777			} else {
778				_mtx_lock_indefinite_check(m, &lda);
779			}
780			v = MTX_READ_VALUE(m);
781		} while (v != MTX_UNOWNED);
782		spinlock_enter();
783	}
784
785	if (LOCK_LOG_TEST(&m->lock_object, opts))
786		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
787	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
788	    "running");
789
790#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
791	if (__predict_true(!doing_lockprof))
792		return;
793#endif
794#ifdef KDTRACE_HOOKS
795	spin_time += lockstat_nsecs(&m->lock_object);
796	if (lda.spin_cnt != 0)
797		LOCKSTAT_RECORD1(spin__spin, m, spin_time);
798out_lockstat:
799#endif
800	LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire, m,
801	    contested, waittime, file, line);
802}
803#endif /* SMP */
804
805#ifdef INVARIANTS
806static void
807thread_lock_validate(struct mtx *m, int opts, const char *file, int line)
808{
809
810	KASSERT(m->mtx_lock != MTX_DESTROYED,
811	    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
812	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
813	    ("thread_lock() of sleep mutex %s @ %s:%d",
814	    m->lock_object.lo_name, file, line));
815	KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) == 0,
816	    ("thread_lock: got a recursive mutex %s @ %s:%d\n",
817	    m->lock_object.lo_name, file, line));
818	WITNESS_CHECKORDER(&m->lock_object,
819	    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
820}
821#else
822#define thread_lock_validate(m, opts, file, line) do { } while (0)
823#endif
824
825#ifndef LOCK_PROFILING
826#if LOCK_DEBUG > 0
827void
828_thread_lock(struct thread *td, int opts, const char *file, int line)
829#else
830void
831_thread_lock(struct thread *td)
832#endif
833{
834	struct mtx *m;
835	uintptr_t tid;
836
837	tid = (uintptr_t)curthread;
838
839	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(spin__acquire)))
840		goto slowpath_noirq;
841	spinlock_enter();
842	m = td->td_lock;
843	thread_lock_validate(m, 0, file, line);
844	if (__predict_false(m == &blocked_lock))
845		goto slowpath_unlocked;
846	if (__predict_false(!_mtx_obtain_lock(m, tid)))
847		goto slowpath_unlocked;
848	if (__predict_true(m == td->td_lock)) {
849		WITNESS_LOCK(&m->lock_object, LOP_EXCLUSIVE, file, line);
850		return;
851	}
852	_mtx_release_lock_quick(m);
853slowpath_unlocked:
854	spinlock_exit();
855slowpath_noirq:
856#if LOCK_DEBUG > 0
857	thread_lock_flags_(td, opts, file, line);
858#else
859	thread_lock_flags_(td, 0, 0, 0);
860#endif
861}
862#endif
863
864void
865thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
866{
867	struct mtx *m;
868	uintptr_t tid, v;
869	struct lock_delay_arg lda;
870#ifdef LOCK_PROFILING
871	int contested = 0;
872	uint64_t waittime = 0;
873#endif
874#ifdef KDTRACE_HOOKS
875	int64_t spin_time = 0;
876#endif
877#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
878	int doing_lockprof = 1;
879#endif
880
881	tid = (uintptr_t)curthread;
882
883	if (SCHEDULER_STOPPED()) {
884		/*
885		 * Ensure that spinlock sections are balanced even when the
886		 * scheduler is stopped, since we may otherwise inadvertently
887		 * re-enable interrupts while dumping core.
888		 */
889		spinlock_enter();
890		return;
891	}
892
893	lock_delay_arg_init(&lda, &mtx_spin_delay);
894
895#ifdef HWPMC_HOOKS
896	PMC_SOFT_CALL( , , lock, failed);
897#endif
898
899#ifdef LOCK_PROFILING
900	doing_lockprof = 1;
901#elif defined(KDTRACE_HOOKS)
902	doing_lockprof = lockstat_enabled;
903#endif
904#ifdef KDTRACE_HOOKS
905	if (__predict_false(doing_lockprof))
906		spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
907#endif
908	spinlock_enter();
909
910	for (;;) {
911retry:
912		m = td->td_lock;
913		thread_lock_validate(m, opts, file, line);
914		v = MTX_READ_VALUE(m);
915		for (;;) {
916			if (v == MTX_UNOWNED) {
917				if (_mtx_obtain_lock_fetch(m, &v, tid))
918					break;
919				continue;
920			}
921			MPASS(v != tid);
922			lock_profile_obtain_lock_failed(&m->lock_object, true,
923			    &contested, &waittime);
924			/* Give interrupts a chance while we spin. */
925			spinlock_exit();
926			do {
927				if (__predict_true(lda.spin_cnt < 10000000)) {
928					lock_delay(&lda);
929				} else {
930					_mtx_lock_indefinite_check(m, &lda);
931				}
932				if (m != td->td_lock) {
933					spinlock_enter();
934					goto retry;
935				}
936				v = MTX_READ_VALUE(m);
937			} while (v != MTX_UNOWNED);
938			spinlock_enter();
939		}
940		if (m == td->td_lock)
941			break;
942		_mtx_release_lock_quick(m);
943	}
944	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
945	    line);
946	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
947
948#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
949	if (__predict_true(!doing_lockprof))
950		return;
951#endif
952#ifdef KDTRACE_HOOKS
953	spin_time += lockstat_nsecs(&m->lock_object);
954#endif
955	LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire, m, contested,
956	    waittime, file, line);
957#ifdef KDTRACE_HOOKS
958	if (lda.spin_cnt != 0)
959		LOCKSTAT_RECORD1(thread__spin, m, spin_time);
960#endif
961}
962
963struct mtx *
964thread_lock_block(struct thread *td)
965{
966	struct mtx *lock;
967
968	lock = td->td_lock;
969	mtx_assert(lock, MA_OWNED);
970	td->td_lock = &blocked_lock;
971
972	return (lock);
973}
974
975void
976thread_lock_unblock(struct thread *td, struct mtx *new)
977{
978
979	mtx_assert(new, MA_OWNED);
980	KASSERT(td->td_lock == &blocked_lock,
981	    ("thread %p lock %p not blocked_lock %p",
982	    td, td->td_lock, &blocked_lock));
983	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
984}
985
986void
987thread_lock_block_wait(struct thread *td)
988{
989
990	while (td->td_lock == &blocked_lock)
991		cpu_spinwait();
992
993	/* Acquire fence to be certain that all thread state is visible. */
994	atomic_thread_fence_acq();
995}
996
997void
998thread_lock_set(struct thread *td, struct mtx *new)
999{
1000	struct mtx *lock;
1001
1002	mtx_assert(new, MA_OWNED);
1003	lock = td->td_lock;
1004	mtx_assert(lock, MA_OWNED);
1005	td->td_lock = new;
1006	mtx_unlock_spin(lock);
1007}
1008
1009/*
1010 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
1011 *
1012 * We are only called here if the lock is recursed, contested (i.e. we
1013 * need to wake up a blocked thread) or lockstat probe is active.
1014 */
1015#if LOCK_DEBUG > 0
1016void
1017__mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v, int opts,
1018    const char *file, int line)
1019#else
1020void
1021__mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v)
1022#endif
1023{
1024	struct mtx *m;
1025	struct turnstile *ts;
1026	uintptr_t tid;
1027
1028	if (SCHEDULER_STOPPED())
1029		return;
1030
1031	tid = (uintptr_t)curthread;
1032	m = mtxlock2mtx(c);
1033
1034	if (__predict_false(v == tid))
1035		v = MTX_READ_VALUE(m);
1036
1037	if (__predict_false(v & MTX_RECURSED)) {
1038		if (--(m->mtx_recurse) == 0)
1039			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
1040		if (LOCK_LOG_TEST(&m->lock_object, opts))
1041			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
1042		return;
1043	}
1044
1045	LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m);
1046	if (v == tid && _mtx_release_lock(m, tid))
1047		return;
1048
1049	/*
1050	 * We have to lock the chain before the turnstile so this turnstile
1051	 * can be removed from the hash list if it is empty.
1052	 */
1053	turnstile_chain_lock(&m->lock_object);
1054	_mtx_release_lock_quick(m);
1055	ts = turnstile_lookup(&m->lock_object);
1056	MPASS(ts != NULL);
1057	if (LOCK_LOG_TEST(&m->lock_object, opts))
1058		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
1059	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
1060
1061	/*
1062	 * This turnstile is now no longer associated with the mutex.  We can
1063	 * unlock the chain lock so a new turnstile may take it's place.
1064	 */
1065	turnstile_unpend(ts);
1066	turnstile_chain_unlock(&m->lock_object);
1067}
1068
1069/*
1070 * All the unlocking of MTX_SPIN locks is done inline.
1071 * See the __mtx_unlock_spin() macro for the details.
1072 */
1073
1074/*
1075 * The backing function for the INVARIANTS-enabled mtx_assert()
1076 */
1077#ifdef INVARIANT_SUPPORT
1078void
1079__mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1080{
1081	const struct mtx *m;
1082
1083	if (KERNEL_PANICKED() || dumping || SCHEDULER_STOPPED())
1084		return;
1085
1086	m = mtxlock2mtx(c);
1087
1088	switch (what) {
1089	case MA_OWNED:
1090	case MA_OWNED | MA_RECURSED:
1091	case MA_OWNED | MA_NOTRECURSED:
1092		if (!mtx_owned(m))
1093			panic("mutex %s not owned at %s:%d",
1094			    m->lock_object.lo_name, file, line);
1095		if (mtx_recursed(m)) {
1096			if ((what & MA_NOTRECURSED) != 0)
1097				panic("mutex %s recursed at %s:%d",
1098				    m->lock_object.lo_name, file, line);
1099		} else if ((what & MA_RECURSED) != 0) {
1100			panic("mutex %s unrecursed at %s:%d",
1101			    m->lock_object.lo_name, file, line);
1102		}
1103		break;
1104	case MA_NOTOWNED:
1105		if (mtx_owned(m))
1106			panic("mutex %s owned at %s:%d",
1107			    m->lock_object.lo_name, file, line);
1108		break;
1109	default:
1110		panic("unknown mtx_assert at %s:%d", file, line);
1111	}
1112}
1113#endif
1114
1115/*
1116 * General init routine used by the MTX_SYSINIT() macro.
1117 */
1118void
1119mtx_sysinit(void *arg)
1120{
1121	struct mtx_args *margs = arg;
1122
1123	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
1124	    margs->ma_opts);
1125}
1126
1127/*
1128 * Mutex initialization routine; initialize lock `m' of type contained in
1129 * `opts' with options contained in `opts' and name `name.'  The optional
1130 * lock type `type' is used as a general lock category name for use with
1131 * witness.
1132 */
1133void
1134_mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
1135{
1136	struct mtx *m;
1137	struct lock_class *class;
1138	int flags;
1139
1140	m = mtxlock2mtx(c);
1141
1142	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
1143	    MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
1144	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
1145	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
1146	    &m->mtx_lock));
1147
1148	/* Determine lock class and lock flags. */
1149	if (opts & MTX_SPIN)
1150		class = &lock_class_mtx_spin;
1151	else
1152		class = &lock_class_mtx_sleep;
1153	flags = 0;
1154	if (opts & MTX_QUIET)
1155		flags |= LO_QUIET;
1156	if (opts & MTX_RECURSE)
1157		flags |= LO_RECURSABLE;
1158	if ((opts & MTX_NOWITNESS) == 0)
1159		flags |= LO_WITNESS;
1160	if (opts & MTX_DUPOK)
1161		flags |= LO_DUPOK;
1162	if (opts & MTX_NOPROFILE)
1163		flags |= LO_NOPROFILE;
1164	if (opts & MTX_NEW)
1165		flags |= LO_NEW;
1166
1167	/* Initialize mutex. */
1168	lock_init(&m->lock_object, class, name, type, flags);
1169
1170	m->mtx_lock = MTX_UNOWNED;
1171	m->mtx_recurse = 0;
1172}
1173
1174/*
1175 * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
1176 * passed in as a flag here because if the corresponding mtx_init() was
1177 * called with MTX_QUIET set, then it will already be set in the mutex's
1178 * flags.
1179 */
1180void
1181_mtx_destroy(volatile uintptr_t *c)
1182{
1183	struct mtx *m;
1184
1185	m = mtxlock2mtx(c);
1186
1187	if (!mtx_owned(m))
1188		MPASS(mtx_unowned(m));
1189	else {
1190		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
1191
1192		/* Perform the non-mtx related part of mtx_unlock_spin(). */
1193		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) {
1194			lock_profile_release_lock(&m->lock_object, true);
1195			spinlock_exit();
1196		} else {
1197			TD_LOCKS_DEC(curthread);
1198			lock_profile_release_lock(&m->lock_object, false);
1199		}
1200
1201		/* Tell witness this isn't locked to make it happy. */
1202		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
1203		    __LINE__);
1204	}
1205
1206	m->mtx_lock = MTX_DESTROYED;
1207	lock_destroy(&m->lock_object);
1208}
1209
1210/*
1211 * Intialize the mutex code and system mutexes.  This is called from the MD
1212 * startup code prior to mi_startup().  The per-CPU data space needs to be
1213 * setup before this is called.
1214 */
1215void
1216mutex_init(void)
1217{
1218
1219	/* Setup turnstiles so that sleep mutexes work. */
1220	init_turnstiles();
1221
1222	/*
1223	 * Initialize mutexes.
1224	 */
1225	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
1226	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
1227	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
1228	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
1229	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
1230	mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
1231	mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1232	mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1233	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1234	mtx_lock(&Giant);
1235}
1236
1237static void __noinline
1238_mtx_lock_indefinite_check(struct mtx *m, struct lock_delay_arg *ldap)
1239{
1240	struct thread *td;
1241
1242	ldap->spin_cnt++;
1243	if (ldap->spin_cnt < 60000000 || kdb_active || KERNEL_PANICKED())
1244		cpu_lock_delay();
1245	else {
1246		td = mtx_owner(m);
1247
1248		/* If the mutex is unlocked, try again. */
1249		if (td == NULL)
1250			return;
1251
1252		printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
1253		    m, m->lock_object.lo_name, td, td->td_tid);
1254#ifdef WITNESS
1255		witness_display_spinlock(&m->lock_object, td, printf);
1256#endif
1257		panic("spin lock held too long");
1258	}
1259	cpu_spinwait();
1260}
1261
1262void
1263mtx_spin_wait_unlocked(struct mtx *m)
1264{
1265	struct lock_delay_arg lda;
1266
1267	KASSERT(m->mtx_lock != MTX_DESTROYED,
1268	    ("%s() of destroyed mutex %p", __func__, m));
1269	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
1270	    ("%s() of sleep mutex %p (%s)", __func__, m,
1271	    m->lock_object.lo_name));
1272	KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1273	    m->lock_object.lo_name));
1274
1275	lda.spin_cnt = 0;
1276
1277	while (atomic_load_acq_ptr(&m->mtx_lock) != MTX_UNOWNED) {
1278		if (__predict_true(lda.spin_cnt < 10000000)) {
1279			cpu_spinwait();
1280			lda.spin_cnt++;
1281		} else {
1282			_mtx_lock_indefinite_check(m, &lda);
1283		}
1284	}
1285}
1286
1287void
1288mtx_wait_unlocked(struct mtx *m)
1289{
1290	struct thread *owner;
1291	uintptr_t v;
1292
1293	KASSERT(m->mtx_lock != MTX_DESTROYED,
1294	    ("%s() of destroyed mutex %p", __func__, m));
1295	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
1296	    ("%s() not a sleep mutex %p (%s)", __func__, m,
1297	    m->lock_object.lo_name));
1298	KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1299	    m->lock_object.lo_name));
1300
1301	for (;;) {
1302		v = atomic_load_acq_ptr(&m->mtx_lock);
1303		if (v == MTX_UNOWNED) {
1304			break;
1305		}
1306		owner = lv_mtx_owner(v);
1307		if (!TD_IS_RUNNING(owner)) {
1308			mtx_lock(m);
1309			mtx_unlock(m);
1310			break;
1311		}
1312		cpu_spinwait();
1313	}
1314}
1315
1316#ifdef DDB
1317void
1318db_show_mtx(const struct lock_object *lock)
1319{
1320	struct thread *td;
1321	const struct mtx *m;
1322
1323	m = (const struct mtx *)lock;
1324
1325	db_printf(" flags: {");
1326	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1327		db_printf("SPIN");
1328	else
1329		db_printf("DEF");
1330	if (m->lock_object.lo_flags & LO_RECURSABLE)
1331		db_printf(", RECURSE");
1332	if (m->lock_object.lo_flags & LO_DUPOK)
1333		db_printf(", DUPOK");
1334	db_printf("}\n");
1335	db_printf(" state: {");
1336	if (mtx_unowned(m))
1337		db_printf("UNOWNED");
1338	else if (mtx_destroyed(m))
1339		db_printf("DESTROYED");
1340	else {
1341		db_printf("OWNED");
1342		if (m->mtx_lock & MTX_CONTESTED)
1343			db_printf(", CONTESTED");
1344		if (m->mtx_lock & MTX_RECURSED)
1345			db_printf(", RECURSED");
1346	}
1347	db_printf("}\n");
1348	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1349		td = mtx_owner(m);
1350		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1351		    td->td_tid, td->td_proc->p_pid, td->td_name);
1352		if (mtx_recursed(m))
1353			db_printf(" recursed: %d\n", m->mtx_recurse);
1354	}
1355}
1356#endif
1357