kern_mutex.c revision 285759
1/*-
2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 *    notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 *    notice, this list of conditions and the following disclaimer in the
11 *    documentation and/or other materials provided with the distribution.
12 * 3. Berkeley Software Design Inc's name may not be used to endorse or
13 *    promote products derived from this software without specific prior
14 *    written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER 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
26 * SUCH DAMAGE.
27 *
28 *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29 *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30 */
31
32/*
33 * Machine independent bits of mutex implementation.
34 */
35
36#include <sys/cdefs.h>
37__FBSDID("$FreeBSD: stable/10/sys/kern/kern_mutex.c 285759 2015-07-21 17:16:37Z markj $");
38
39#include "opt_adaptive_mutexes.h"
40#include "opt_ddb.h"
41#include "opt_global.h"
42#include "opt_hwpmc_hooks.h"
43#include "opt_kdtrace.h"
44#include "opt_sched.h"
45
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/bus.h>
49#include <sys/conf.h>
50#include <sys/kdb.h>
51#include <sys/kernel.h>
52#include <sys/ktr.h>
53#include <sys/lock.h>
54#include <sys/malloc.h>
55#include <sys/mutex.h>
56#include <sys/proc.h>
57#include <sys/resourcevar.h>
58#include <sys/sched.h>
59#include <sys/sbuf.h>
60#include <sys/sysctl.h>
61#include <sys/turnstile.h>
62#include <sys/vmmeter.h>
63#include <sys/lock_profile.h>
64
65#include <machine/atomic.h>
66#include <machine/bus.h>
67#include <machine/cpu.h>
68
69#include <ddb/ddb.h>
70
71#include <fs/devfs/devfs_int.h>
72
73#include <vm/vm.h>
74#include <vm/vm_extern.h>
75
76#if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
77#define	ADAPTIVE_MUTEXES
78#endif
79
80#ifdef HWPMC_HOOKS
81#include <sys/pmckern.h>
82PMC_SOFT_DEFINE( , , lock, failed);
83#endif
84
85/*
86 * Return the mutex address when the lock cookie address is provided.
87 * This functionality assumes that struct mtx* have a member named mtx_lock.
88 */
89#define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
90
91/*
92 * Internal utility macros.
93 */
94#define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
95
96#define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
97
98#define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
99
100static void	assert_mtx(const struct lock_object *lock, int what);
101#ifdef DDB
102static void	db_show_mtx(const struct lock_object *lock);
103#endif
104static void	lock_mtx(struct lock_object *lock, uintptr_t how);
105static void	lock_spin(struct lock_object *lock, uintptr_t how);
106#ifdef KDTRACE_HOOKS
107static int	owner_mtx(const struct lock_object *lock,
108		    struct thread **owner);
109#endif
110static uintptr_t unlock_mtx(struct lock_object *lock);
111static uintptr_t unlock_spin(struct lock_object *lock);
112
113/*
114 * Lock classes for sleep and spin mutexes.
115 */
116struct lock_class lock_class_mtx_sleep = {
117	.lc_name = "sleep mutex",
118	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
119	.lc_assert = assert_mtx,
120#ifdef DDB
121	.lc_ddb_show = db_show_mtx,
122#endif
123	.lc_lock = lock_mtx,
124	.lc_unlock = unlock_mtx,
125#ifdef KDTRACE_HOOKS
126	.lc_owner = owner_mtx,
127#endif
128};
129struct lock_class lock_class_mtx_spin = {
130	.lc_name = "spin mutex",
131	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
132	.lc_assert = assert_mtx,
133#ifdef DDB
134	.lc_ddb_show = db_show_mtx,
135#endif
136	.lc_lock = lock_spin,
137	.lc_unlock = unlock_spin,
138#ifdef KDTRACE_HOOKS
139	.lc_owner = owner_mtx,
140#endif
141};
142
143/*
144 * System-wide mutexes
145 */
146struct mtx blocked_lock;
147struct mtx Giant;
148
149void
150assert_mtx(const struct lock_object *lock, int what)
151{
152
153	mtx_assert((const struct mtx *)lock, what);
154}
155
156void
157lock_mtx(struct lock_object *lock, uintptr_t how)
158{
159
160	mtx_lock((struct mtx *)lock);
161}
162
163void
164lock_spin(struct lock_object *lock, uintptr_t how)
165{
166
167	panic("spin locks can only use msleep_spin");
168}
169
170uintptr_t
171unlock_mtx(struct lock_object *lock)
172{
173	struct mtx *m;
174
175	m = (struct mtx *)lock;
176	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
177	mtx_unlock(m);
178	return (0);
179}
180
181uintptr_t
182unlock_spin(struct lock_object *lock)
183{
184
185	panic("spin locks can only use msleep_spin");
186}
187
188#ifdef KDTRACE_HOOKS
189int
190owner_mtx(const struct lock_object *lock, struct thread **owner)
191{
192	const struct mtx *m = (const struct mtx *)lock;
193
194	*owner = mtx_owner(m);
195	return (mtx_unowned(m) == 0);
196}
197#endif
198
199/*
200 * Function versions of the inlined __mtx_* macros.  These are used by
201 * modules and can also be called from assembly language if needed.
202 */
203void
204__mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
205{
206	struct mtx *m;
207
208	if (SCHEDULER_STOPPED())
209		return;
210
211	m = mtxlock2mtx(c);
212
213	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
214	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
215	    curthread, m->lock_object.lo_name, file, line));
216	KASSERT(m->mtx_lock != MTX_DESTROYED,
217	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
218	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
219	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
220	    file, line));
221	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
222	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
223
224	__mtx_lock(m, curthread, opts, file, line);
225	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
226	    line);
227	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
228	    file, line);
229	curthread->td_locks++;
230}
231
232void
233__mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
234{
235	struct mtx *m;
236
237	if (SCHEDULER_STOPPED())
238		return;
239
240	m = mtxlock2mtx(c);
241
242	KASSERT(m->mtx_lock != MTX_DESTROYED,
243	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
244	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
245	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
246	    file, line));
247	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
248	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
249	    line);
250	mtx_assert(m, MA_OWNED);
251
252	if (m->mtx_recurse == 0)
253		LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m);
254	__mtx_unlock(m, curthread, opts, file, line);
255	curthread->td_locks--;
256}
257
258void
259__mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
260    int line)
261{
262	struct mtx *m;
263
264	if (SCHEDULER_STOPPED())
265		return;
266
267	m = mtxlock2mtx(c);
268
269	KASSERT(m->mtx_lock != MTX_DESTROYED,
270	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
271	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
272	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
273	    m->lock_object.lo_name, file, line));
274	if (mtx_owned(m))
275		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
276		    (opts & MTX_RECURSE) != 0,
277	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
278		    m->lock_object.lo_name, file, line));
279	opts &= ~MTX_RECURSE;
280	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
281	    file, line, NULL);
282	__mtx_lock_spin(m, curthread, opts, file, line);
283	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
284	    line);
285	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
286}
287
288void
289__mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
290    int line)
291{
292	struct mtx *m;
293
294	if (SCHEDULER_STOPPED())
295		return;
296
297	m = mtxlock2mtx(c);
298
299	KASSERT(m->mtx_lock != MTX_DESTROYED,
300	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
301	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
302	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
303	    m->lock_object.lo_name, file, line));
304	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
305	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
306	    line);
307	mtx_assert(m, MA_OWNED);
308
309	__mtx_unlock_spin(m);
310}
311
312/*
313 * The important part of mtx_trylock{,_flags}()
314 * Tries to acquire lock `m.'  If this function is called on a mutex that
315 * is already owned, it will recursively acquire the lock.
316 */
317int
318_mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
319{
320	struct mtx *m;
321#ifdef LOCK_PROFILING
322	uint64_t waittime = 0;
323	int contested = 0;
324#endif
325	int rval;
326
327	if (SCHEDULER_STOPPED())
328		return (1);
329
330	m = mtxlock2mtx(c);
331
332	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
333	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
334	    curthread, m->lock_object.lo_name, file, line));
335	KASSERT(m->mtx_lock != MTX_DESTROYED,
336	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
337	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
338	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
339	    file, line));
340
341	if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
342	    (opts & MTX_RECURSE) != 0)) {
343		m->mtx_recurse++;
344		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
345		rval = 1;
346	} else
347		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
348	opts &= ~MTX_RECURSE;
349
350	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
351	if (rval) {
352		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
353		    file, line);
354		curthread->td_locks++;
355		if (m->mtx_recurse == 0)
356			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
357			    m, contested, waittime, file, line);
358
359	}
360
361	return (rval);
362}
363
364/*
365 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
366 *
367 * We call this if the lock is either contested (i.e. we need to go to
368 * sleep waiting for it), or if we need to recurse on it.
369 */
370void
371__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
372    const char *file, int line)
373{
374	struct mtx *m;
375	struct turnstile *ts;
376	uintptr_t v;
377#ifdef ADAPTIVE_MUTEXES
378	volatile struct thread *owner;
379#endif
380#ifdef KTR
381	int cont_logged = 0;
382#endif
383#ifdef LOCK_PROFILING
384	int contested = 0;
385	uint64_t waittime = 0;
386#endif
387#ifdef KDTRACE_HOOKS
388	uint64_t spin_cnt = 0;
389	uint64_t sleep_cnt = 0;
390	int64_t sleep_time = 0;
391	int64_t all_time = 0;
392#endif
393
394	if (SCHEDULER_STOPPED())
395		return;
396
397	m = mtxlock2mtx(c);
398
399	if (mtx_owned(m)) {
400		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
401		    (opts & MTX_RECURSE) != 0,
402	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
403		    m->lock_object.lo_name, file, line));
404		opts &= ~MTX_RECURSE;
405		m->mtx_recurse++;
406		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
407		if (LOCK_LOG_TEST(&m->lock_object, opts))
408			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
409		return;
410	}
411	opts &= ~MTX_RECURSE;
412
413#ifdef HWPMC_HOOKS
414	PMC_SOFT_CALL( , , lock, failed);
415#endif
416	lock_profile_obtain_lock_failed(&m->lock_object,
417		    &contested, &waittime);
418	if (LOCK_LOG_TEST(&m->lock_object, opts))
419		CTR4(KTR_LOCK,
420		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
421		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
422#ifdef KDTRACE_HOOKS
423	all_time -= lockstat_nsecs(&m->lock_object);
424#endif
425
426	while (!_mtx_obtain_lock(m, tid)) {
427#ifdef KDTRACE_HOOKS
428		spin_cnt++;
429#endif
430#ifdef ADAPTIVE_MUTEXES
431		/*
432		 * If the owner is running on another CPU, spin until the
433		 * owner stops running or the state of the lock changes.
434		 */
435		v = m->mtx_lock;
436		if (v != MTX_UNOWNED) {
437			owner = (struct thread *)(v & ~MTX_FLAGMASK);
438			if (TD_IS_RUNNING(owner)) {
439				if (LOCK_LOG_TEST(&m->lock_object, 0))
440					CTR3(KTR_LOCK,
441					    "%s: spinning on %p held by %p",
442					    __func__, m, owner);
443				KTR_STATE1(KTR_SCHED, "thread",
444				    sched_tdname((struct thread *)tid),
445				    "spinning", "lockname:\"%s\"",
446				    m->lock_object.lo_name);
447				while (mtx_owner(m) == owner &&
448				    TD_IS_RUNNING(owner)) {
449					cpu_spinwait();
450#ifdef KDTRACE_HOOKS
451					spin_cnt++;
452#endif
453				}
454				KTR_STATE0(KTR_SCHED, "thread",
455				    sched_tdname((struct thread *)tid),
456				    "running");
457				continue;
458			}
459		}
460#endif
461
462		ts = turnstile_trywait(&m->lock_object);
463		v = m->mtx_lock;
464
465		/*
466		 * Check if the lock has been released while spinning for
467		 * the turnstile chain lock.
468		 */
469		if (v == MTX_UNOWNED) {
470			turnstile_cancel(ts);
471			continue;
472		}
473
474#ifdef ADAPTIVE_MUTEXES
475		/*
476		 * The current lock owner might have started executing
477		 * on another CPU (or the lock could have changed
478		 * owners) while we were waiting on the turnstile
479		 * chain lock.  If so, drop the turnstile lock and try
480		 * again.
481		 */
482		owner = (struct thread *)(v & ~MTX_FLAGMASK);
483		if (TD_IS_RUNNING(owner)) {
484			turnstile_cancel(ts);
485			continue;
486		}
487#endif
488
489		/*
490		 * If the mutex isn't already contested and a failure occurs
491		 * setting the contested bit, the mutex was either released
492		 * or the state of the MTX_RECURSED bit changed.
493		 */
494		if ((v & MTX_CONTESTED) == 0 &&
495		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
496			turnstile_cancel(ts);
497			continue;
498		}
499
500		/*
501		 * We definitely must sleep for this lock.
502		 */
503		mtx_assert(m, MA_NOTOWNED);
504
505#ifdef KTR
506		if (!cont_logged) {
507			CTR6(KTR_CONTENTION,
508			    "contention: %p at %s:%d wants %s, taken by %s:%d",
509			    (void *)tid, file, line, m->lock_object.lo_name,
510			    WITNESS_FILE(&m->lock_object),
511			    WITNESS_LINE(&m->lock_object));
512			cont_logged = 1;
513		}
514#endif
515
516		/*
517		 * Block on the turnstile.
518		 */
519#ifdef KDTRACE_HOOKS
520		sleep_time -= lockstat_nsecs(&m->lock_object);
521#endif
522		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
523#ifdef KDTRACE_HOOKS
524		sleep_time += lockstat_nsecs(&m->lock_object);
525		sleep_cnt++;
526#endif
527	}
528#ifdef KDTRACE_HOOKS
529	all_time += lockstat_nsecs(&m->lock_object);
530#endif
531#ifdef KTR
532	if (cont_logged) {
533		CTR4(KTR_CONTENTION,
534		    "contention end: %s acquired by %p at %s:%d",
535		    m->lock_object.lo_name, (void *)tid, file, line);
536	}
537#endif
538	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
539	    waittime, file, line);
540#ifdef KDTRACE_HOOKS
541	if (sleep_time)
542		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
543
544	/*
545	 * Only record the loops spinning and not sleeping.
546	 */
547	if (spin_cnt > sleep_cnt)
548		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (all_time - sleep_time));
549#endif
550}
551
552static void
553_mtx_lock_spin_failed(struct mtx *m)
554{
555	struct thread *td;
556
557	td = mtx_owner(m);
558
559	/* If the mutex is unlocked, try again. */
560	if (td == NULL)
561		return;
562
563	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
564	    m, m->lock_object.lo_name, td, td->td_tid);
565#ifdef WITNESS
566	witness_display_spinlock(&m->lock_object, td, printf);
567#endif
568	panic("spin lock held too long");
569}
570
571#ifdef SMP
572/*
573 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
574 *
575 * This is only called if we need to actually spin for the lock. Recursion
576 * is handled inline.
577 */
578void
579_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
580    const char *file, int line)
581{
582	struct mtx *m;
583	int i = 0;
584#ifdef LOCK_PROFILING
585	int contested = 0;
586	uint64_t waittime = 0;
587#endif
588#ifdef KDTRACE_HOOKS
589	int64_t spin_time = 0;
590#endif
591
592	if (SCHEDULER_STOPPED())
593		return;
594
595	m = mtxlock2mtx(c);
596
597	if (LOCK_LOG_TEST(&m->lock_object, opts))
598		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
599	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
600	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
601
602#ifdef HWPMC_HOOKS
603	PMC_SOFT_CALL( , , lock, failed);
604#endif
605	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
606#ifdef KDTRACE_HOOKS
607	spin_time -= lockstat_nsecs(&m->lock_object);
608#endif
609	while (!_mtx_obtain_lock(m, tid)) {
610
611		/* Give interrupts a chance while we spin. */
612		spinlock_exit();
613		while (m->mtx_lock != MTX_UNOWNED) {
614			if (i++ < 10000000) {
615				cpu_spinwait();
616				continue;
617			}
618			if (i < 60000000 || kdb_active || panicstr != NULL)
619				DELAY(1);
620			else
621				_mtx_lock_spin_failed(m);
622			cpu_spinwait();
623		}
624		spinlock_enter();
625	}
626#ifdef KDTRACE_HOOKS
627	spin_time += lockstat_nsecs(&m->lock_object);
628#endif
629
630	if (LOCK_LOG_TEST(&m->lock_object, opts))
631		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
632	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
633	    "running");
634
635	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
636	    contested, waittime, (file), (line));
637#ifdef KDTRACE_HOOKS
638	if (spin_time != 0)
639		LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, spin_time);
640#endif
641}
642#endif /* SMP */
643
644void
645thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
646{
647	struct mtx *m;
648	uintptr_t tid;
649	int i;
650#ifdef LOCK_PROFILING
651	int contested = 0;
652	uint64_t waittime = 0;
653#endif
654#ifdef KDTRACE_HOOKS
655	int64_t spin_time = 0;
656#endif
657
658	i = 0;
659	tid = (uintptr_t)curthread;
660
661	if (SCHEDULER_STOPPED())
662		return;
663
664#ifdef KDTRACE_HOOKS
665	spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
666#endif
667	for (;;) {
668retry:
669		spinlock_enter();
670		m = td->td_lock;
671		KASSERT(m->mtx_lock != MTX_DESTROYED,
672		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
673		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
674		    ("thread_lock() of sleep mutex %s @ %s:%d",
675		    m->lock_object.lo_name, file, line));
676		if (mtx_owned(m))
677			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
678	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
679			    m->lock_object.lo_name, file, line));
680		WITNESS_CHECKORDER(&m->lock_object,
681		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
682		while (!_mtx_obtain_lock(m, tid)) {
683			if (m->mtx_lock == tid) {
684				m->mtx_recurse++;
685				break;
686			}
687#ifdef HWPMC_HOOKS
688			PMC_SOFT_CALL( , , lock, failed);
689#endif
690			lock_profile_obtain_lock_failed(&m->lock_object,
691			    &contested, &waittime);
692			/* Give interrupts a chance while we spin. */
693			spinlock_exit();
694			while (m->mtx_lock != MTX_UNOWNED) {
695				if (i++ < 10000000)
696					cpu_spinwait();
697				else if (i < 60000000 ||
698				    kdb_active || panicstr != NULL)
699					DELAY(1);
700				else
701					_mtx_lock_spin_failed(m);
702				cpu_spinwait();
703				if (m != td->td_lock)
704					goto retry;
705			}
706			spinlock_enter();
707		}
708		if (m == td->td_lock)
709			break;
710		__mtx_unlock_spin(m);	/* does spinlock_exit() */
711	}
712#ifdef KDTRACE_HOOKS
713	spin_time += lockstat_nsecs(&m->lock_object);
714#endif
715	if (m->mtx_recurse == 0)
716		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
717		    m, contested, waittime, (file), (line));
718	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
719	    line);
720	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
721	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_time);
722}
723
724struct mtx *
725thread_lock_block(struct thread *td)
726{
727	struct mtx *lock;
728
729	THREAD_LOCK_ASSERT(td, MA_OWNED);
730	lock = td->td_lock;
731	td->td_lock = &blocked_lock;
732	mtx_unlock_spin(lock);
733
734	return (lock);
735}
736
737void
738thread_lock_unblock(struct thread *td, struct mtx *new)
739{
740	mtx_assert(new, MA_OWNED);
741	MPASS(td->td_lock == &blocked_lock);
742	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
743}
744
745void
746thread_lock_set(struct thread *td, struct mtx *new)
747{
748	struct mtx *lock;
749
750	mtx_assert(new, MA_OWNED);
751	THREAD_LOCK_ASSERT(td, MA_OWNED);
752	lock = td->td_lock;
753	td->td_lock = new;
754	mtx_unlock_spin(lock);
755}
756
757/*
758 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
759 *
760 * We are only called here if the lock is recursed or contested (i.e. we
761 * need to wake up a blocked thread).
762 */
763void
764__mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
765{
766	struct mtx *m;
767	struct turnstile *ts;
768
769	if (SCHEDULER_STOPPED())
770		return;
771
772	m = mtxlock2mtx(c);
773
774	if (mtx_recursed(m)) {
775		if (--(m->mtx_recurse) == 0)
776			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
777		if (LOCK_LOG_TEST(&m->lock_object, opts))
778			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
779		return;
780	}
781
782	/*
783	 * We have to lock the chain before the turnstile so this turnstile
784	 * can be removed from the hash list if it is empty.
785	 */
786	turnstile_chain_lock(&m->lock_object);
787	ts = turnstile_lookup(&m->lock_object);
788	if (LOCK_LOG_TEST(&m->lock_object, opts))
789		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
790	MPASS(ts != NULL);
791	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
792	_mtx_release_lock_quick(m);
793
794	/*
795	 * This turnstile is now no longer associated with the mutex.  We can
796	 * unlock the chain lock so a new turnstile may take it's place.
797	 */
798	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
799	turnstile_chain_unlock(&m->lock_object);
800}
801
802/*
803 * All the unlocking of MTX_SPIN locks is done inline.
804 * See the __mtx_unlock_spin() macro for the details.
805 */
806
807/*
808 * The backing function for the INVARIANTS-enabled mtx_assert()
809 */
810#ifdef INVARIANT_SUPPORT
811void
812__mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
813{
814	const struct mtx *m;
815
816	if (panicstr != NULL || dumping)
817		return;
818
819	m = mtxlock2mtx(c);
820
821	switch (what) {
822	case MA_OWNED:
823	case MA_OWNED | MA_RECURSED:
824	case MA_OWNED | MA_NOTRECURSED:
825		if (!mtx_owned(m))
826			panic("mutex %s not owned at %s:%d",
827			    m->lock_object.lo_name, file, line);
828		if (mtx_recursed(m)) {
829			if ((what & MA_NOTRECURSED) != 0)
830				panic("mutex %s recursed at %s:%d",
831				    m->lock_object.lo_name, file, line);
832		} else if ((what & MA_RECURSED) != 0) {
833			panic("mutex %s unrecursed at %s:%d",
834			    m->lock_object.lo_name, file, line);
835		}
836		break;
837	case MA_NOTOWNED:
838		if (mtx_owned(m))
839			panic("mutex %s owned at %s:%d",
840			    m->lock_object.lo_name, file, line);
841		break;
842	default:
843		panic("unknown mtx_assert at %s:%d", file, line);
844	}
845}
846#endif
847
848/*
849 * The MUTEX_DEBUG-enabled mtx_validate()
850 *
851 * Most of these checks have been moved off into the LO_INITIALIZED flag
852 * maintained by the witness code.
853 */
854#ifdef MUTEX_DEBUG
855
856void	mtx_validate(struct mtx *);
857
858void
859mtx_validate(struct mtx *m)
860{
861
862/*
863 * XXX: When kernacc() does not require Giant we can reenable this check
864 */
865#ifdef notyet
866	/*
867	 * Can't call kernacc() from early init386(), especially when
868	 * initializing Giant mutex, because some stuff in kernacc()
869	 * requires Giant itself.
870	 */
871	if (!cold)
872		if (!kernacc((caddr_t)m, sizeof(m),
873		    VM_PROT_READ | VM_PROT_WRITE))
874			panic("Can't read and write to mutex %p", m);
875#endif
876}
877#endif
878
879/*
880 * General init routine used by the MTX_SYSINIT() macro.
881 */
882void
883mtx_sysinit(void *arg)
884{
885	struct mtx_args *margs = arg;
886
887	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
888	    margs->ma_opts);
889}
890
891/*
892 * Mutex initialization routine; initialize lock `m' of type contained in
893 * `opts' with options contained in `opts' and name `name.'  The optional
894 * lock type `type' is used as a general lock category name for use with
895 * witness.
896 */
897void
898_mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
899{
900	struct mtx *m;
901	struct lock_class *class;
902	int flags;
903
904	m = mtxlock2mtx(c);
905
906	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
907		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
908	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
909	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
910	    &m->mtx_lock));
911
912#ifdef MUTEX_DEBUG
913	/* Diagnostic and error correction */
914	mtx_validate(m);
915#endif
916
917	/* Determine lock class and lock flags. */
918	if (opts & MTX_SPIN)
919		class = &lock_class_mtx_spin;
920	else
921		class = &lock_class_mtx_sleep;
922	flags = 0;
923	if (opts & MTX_QUIET)
924		flags |= LO_QUIET;
925	if (opts & MTX_RECURSE)
926		flags |= LO_RECURSABLE;
927	if ((opts & MTX_NOWITNESS) == 0)
928		flags |= LO_WITNESS;
929	if (opts & MTX_DUPOK)
930		flags |= LO_DUPOK;
931	if (opts & MTX_NOPROFILE)
932		flags |= LO_NOPROFILE;
933
934	/* Initialize mutex. */
935	lock_init(&m->lock_object, class, name, type, flags);
936
937	m->mtx_lock = MTX_UNOWNED;
938	m->mtx_recurse = 0;
939}
940
941/*
942 * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
943 * passed in as a flag here because if the corresponding mtx_init() was
944 * called with MTX_QUIET set, then it will already be set in the mutex's
945 * flags.
946 */
947void
948_mtx_destroy(volatile uintptr_t *c)
949{
950	struct mtx *m;
951
952	m = mtxlock2mtx(c);
953
954	if (!mtx_owned(m))
955		MPASS(mtx_unowned(m));
956	else {
957		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
958
959		/* Perform the non-mtx related part of mtx_unlock_spin(). */
960		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
961			spinlock_exit();
962		else
963			curthread->td_locks--;
964
965		lock_profile_release_lock(&m->lock_object);
966		/* Tell witness this isn't locked to make it happy. */
967		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
968		    __LINE__);
969	}
970
971	m->mtx_lock = MTX_DESTROYED;
972	lock_destroy(&m->lock_object);
973}
974
975/*
976 * Intialize the mutex code and system mutexes.  This is called from the MD
977 * startup code prior to mi_startup().  The per-CPU data space needs to be
978 * setup before this is called.
979 */
980void
981mutex_init(void)
982{
983
984	/* Setup turnstiles so that sleep mutexes work. */
985	init_turnstiles();
986
987	/*
988	 * Initialize mutexes.
989	 */
990	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
991	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
992	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
993	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
994	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
995	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
996	mtx_lock(&Giant);
997}
998
999#ifdef DDB
1000void
1001db_show_mtx(const struct lock_object *lock)
1002{
1003	struct thread *td;
1004	const struct mtx *m;
1005
1006	m = (const struct mtx *)lock;
1007
1008	db_printf(" flags: {");
1009	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1010		db_printf("SPIN");
1011	else
1012		db_printf("DEF");
1013	if (m->lock_object.lo_flags & LO_RECURSABLE)
1014		db_printf(", RECURSE");
1015	if (m->lock_object.lo_flags & LO_DUPOK)
1016		db_printf(", DUPOK");
1017	db_printf("}\n");
1018	db_printf(" state: {");
1019	if (mtx_unowned(m))
1020		db_printf("UNOWNED");
1021	else if (mtx_destroyed(m))
1022		db_printf("DESTROYED");
1023	else {
1024		db_printf("OWNED");
1025		if (m->mtx_lock & MTX_CONTESTED)
1026			db_printf(", CONTESTED");
1027		if (m->mtx_lock & MTX_RECURSED)
1028			db_printf(", RECURSED");
1029	}
1030	db_printf("}\n");
1031	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1032		td = mtx_owner(m);
1033		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1034		    td->td_tid, td->td_proc->p_pid, td->td_name);
1035		if (mtx_recursed(m))
1036			db_printf(" recursed: %d\n", m->mtx_recurse);
1037	}
1038}
1039#endif
1040