xref: /freebsd/sys/kern/kern_mutex.c (revision 41059135ce931c0f1014a999ffabc6bc470ce856)
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("$FreeBSD$");
38 
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>
81 PMC_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 
97 static void	assert_mtx(const struct lock_object *lock, int what);
98 #ifdef DDB
99 static void	db_show_mtx(const struct lock_object *lock);
100 #endif
101 static void	lock_mtx(struct lock_object *lock, uintptr_t how);
102 static void	lock_spin(struct lock_object *lock, uintptr_t how);
103 #ifdef KDTRACE_HOOKS
104 static int	owner_mtx(const struct lock_object *lock,
105 		    struct thread **owner);
106 #endif
107 static uintptr_t unlock_mtx(struct lock_object *lock);
108 static uintptr_t unlock_spin(struct lock_object *lock);
109 
110 /*
111  * Lock classes for sleep and spin mutexes.
112  */
113 struct 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 };
126 struct 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 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging");
142 
143 static struct lock_delay_config __read_mostly mtx_delay;
144 
145 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base,
146     0, "");
147 SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max,
148     0, "");
149 
150 LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay);
151 #endif
152 
153 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL,
154     "mtx spin debugging");
155 
156 static struct lock_delay_config __read_mostly mtx_spin_delay;
157 
158 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW,
159     &mtx_spin_delay.base, 0, "");
160 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW,
161     &mtx_spin_delay.max, 0, "");
162 
163 LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay);
164 
165 /*
166  * System-wide mutexes
167  */
168 struct mtx blocked_lock;
169 struct mtx Giant;
170 
171 void
172 assert_mtx(const struct lock_object *lock, int what)
173 {
174 
175 	mtx_assert((const struct mtx *)lock, what);
176 }
177 
178 void
179 lock_mtx(struct lock_object *lock, uintptr_t how)
180 {
181 
182 	mtx_lock((struct mtx *)lock);
183 }
184 
185 void
186 lock_spin(struct lock_object *lock, uintptr_t how)
187 {
188 
189 	panic("spin locks can only use msleep_spin");
190 }
191 
192 uintptr_t
193 unlock_mtx(struct lock_object *lock)
194 {
195 	struct mtx *m;
196 
197 	m = (struct mtx *)lock;
198 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
199 	mtx_unlock(m);
200 	return (0);
201 }
202 
203 uintptr_t
204 unlock_spin(struct lock_object *lock)
205 {
206 
207 	panic("spin locks can only use msleep_spin");
208 }
209 
210 #ifdef KDTRACE_HOOKS
211 int
212 owner_mtx(const struct lock_object *lock, struct thread **owner)
213 {
214 	const struct mtx *m;
215 	uintptr_t x;
216 
217 	m = (const struct mtx *)lock;
218 	x = m->mtx_lock;
219 	*owner = (struct thread *)(x & ~MTX_FLAGMASK);
220 	return (x != MTX_UNOWNED);
221 }
222 #endif
223 
224 /*
225  * Function versions of the inlined __mtx_* macros.  These are used by
226  * modules and can also be called from assembly language if needed.
227  */
228 void
229 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
230 {
231 	struct mtx *m;
232 	uintptr_t tid, v;
233 
234 	m = mtxlock2mtx(c);
235 
236 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
237 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
238 	    curthread, m->lock_object.lo_name, file, line));
239 	KASSERT(m->mtx_lock != MTX_DESTROYED,
240 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
241 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
242 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
243 	    file, line));
244 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
245 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
246 
247 	tid = (uintptr_t)curthread;
248 	v = MTX_UNOWNED;
249 	if (!_mtx_obtain_lock_fetch(m, &v, tid))
250 		_mtx_lock_sleep(m, v, tid, opts, file, line);
251 	else
252 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
253 		    m, 0, 0, file, line);
254 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
255 	    line);
256 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
257 	    file, line);
258 	TD_LOCKS_INC(curthread);
259 }
260 
261 void
262 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
263 {
264 	struct mtx *m;
265 
266 	m = mtxlock2mtx(c);
267 
268 	KASSERT(m->mtx_lock != MTX_DESTROYED,
269 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
270 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
271 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
272 	    file, line));
273 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
274 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
275 	    line);
276 	mtx_assert(m, MA_OWNED);
277 
278 #ifdef LOCK_PROFILING
279 	__mtx_unlock_sleep(c, opts, file, line);
280 #else
281 	__mtx_unlock(m, curthread, opts, file, line);
282 #endif
283 	TD_LOCKS_DEC(curthread);
284 }
285 
286 void
287 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
288     int line)
289 {
290 	struct mtx *m;
291 
292 	if (SCHEDULER_STOPPED())
293 		return;
294 
295 	m = mtxlock2mtx(c);
296 
297 	KASSERT(m->mtx_lock != MTX_DESTROYED,
298 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
299 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
300 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
301 	    m->lock_object.lo_name, file, line));
302 	if (mtx_owned(m))
303 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
304 		    (opts & MTX_RECURSE) != 0,
305 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
306 		    m->lock_object.lo_name, file, line));
307 	opts &= ~MTX_RECURSE;
308 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
309 	    file, line, NULL);
310 	__mtx_lock_spin(m, curthread, opts, file, line);
311 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
312 	    line);
313 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
314 }
315 
316 int
317 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
318     int line)
319 {
320 	struct mtx *m;
321 
322 	if (SCHEDULER_STOPPED())
323 		return (1);
324 
325 	m = mtxlock2mtx(c);
326 
327 	KASSERT(m->mtx_lock != MTX_DESTROYED,
328 	    ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line));
329 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
330 	    ("mtx_trylock_spin() of sleep mutex %s @ %s:%d",
331 	    m->lock_object.lo_name, file, line));
332 	KASSERT((opts & MTX_RECURSE) == 0,
333 	    ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n",
334 	    m->lock_object.lo_name, file, line));
335 	if (__mtx_trylock_spin(m, curthread, opts, file, line)) {
336 		LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line);
337 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
338 		return (1);
339 	}
340 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line);
341 	return (0);
342 }
343 
344 void
345 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
346     int line)
347 {
348 	struct mtx *m;
349 
350 	if (SCHEDULER_STOPPED())
351 		return;
352 
353 	m = mtxlock2mtx(c);
354 
355 	KASSERT(m->mtx_lock != MTX_DESTROYED,
356 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
357 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
358 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
359 	    m->lock_object.lo_name, file, line));
360 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
361 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
362 	    line);
363 	mtx_assert(m, MA_OWNED);
364 
365 	__mtx_unlock_spin(m);
366 }
367 
368 /*
369  * The important part of mtx_trylock{,_flags}()
370  * Tries to acquire lock `m.'  If this function is called on a mutex that
371  * is already owned, it will recursively acquire the lock.
372  */
373 int
374 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
375 {
376 	struct mtx *m;
377 	struct thread *td;
378 	uintptr_t tid, v;
379 #ifdef LOCK_PROFILING
380 	uint64_t waittime = 0;
381 	int contested = 0;
382 #endif
383 	int rval;
384 	bool recursed;
385 
386 	td = curthread;
387 	tid = (uintptr_t)td;
388 	if (SCHEDULER_STOPPED_TD(td))
389 		return (1);
390 
391 	m = mtxlock2mtx(c);
392 
393 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
394 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
395 	    curthread, m->lock_object.lo_name, file, line));
396 	KASSERT(m->mtx_lock != MTX_DESTROYED,
397 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
398 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
399 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
400 	    file, line));
401 
402 	rval = 1;
403 	recursed = false;
404 	v = MTX_UNOWNED;
405 	for (;;) {
406 		if (_mtx_obtain_lock_fetch(m, &v, tid))
407 			break;
408 		if (v == MTX_UNOWNED)
409 			continue;
410 		if (v == tid &&
411 		    ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
412 		    (opts & MTX_RECURSE) != 0)) {
413 				 m->mtx_recurse++;
414 				 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
415 				 recursed = true;
416 				 break;
417 		}
418 		rval = 0;
419 		break;
420 	}
421 
422 	opts &= ~MTX_RECURSE;
423 
424 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
425 	if (rval) {
426 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
427 		    file, line);
428 		TD_LOCKS_INC(curthread);
429 		if (!recursed)
430 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
431 			    m, contested, waittime, file, line);
432 	}
433 
434 	return (rval);
435 }
436 
437 /*
438  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
439  *
440  * We call this if the lock is either contested (i.e. we need to go to
441  * sleep waiting for it), or if we need to recurse on it.
442  */
443 #if LOCK_DEBUG > 0
444 void
445 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, int opts,
446     const char *file, int line)
447 #else
448 void
449 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, uintptr_t tid)
450 #endif
451 {
452 	struct mtx *m;
453 	struct turnstile *ts;
454 #ifdef ADAPTIVE_MUTEXES
455 	volatile struct thread *owner;
456 #endif
457 #ifdef KTR
458 	int cont_logged = 0;
459 #endif
460 #ifdef LOCK_PROFILING
461 	int contested = 0;
462 	uint64_t waittime = 0;
463 #endif
464 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
465 	struct lock_delay_arg lda;
466 #endif
467 #ifdef KDTRACE_HOOKS
468 	u_int sleep_cnt = 0;
469 	int64_t sleep_time = 0;
470 	int64_t all_time = 0;
471 #endif
472 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
473 	int doing_lockprof;
474 #endif
475 
476 	if (SCHEDULER_STOPPED())
477 		return;
478 
479 #if defined(ADAPTIVE_MUTEXES)
480 	lock_delay_arg_init(&lda, &mtx_delay);
481 #elif defined(KDTRACE_HOOKS)
482 	lock_delay_arg_init(&lda, NULL);
483 #endif
484 	m = mtxlock2mtx(c);
485 	if (__predict_false(v == MTX_UNOWNED))
486 		v = MTX_READ_VALUE(m);
487 
488 	if (__predict_false(lv_mtx_owner(v) == (struct thread *)tid)) {
489 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
490 		    (opts & MTX_RECURSE) != 0,
491 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
492 		    m->lock_object.lo_name, file, line));
493 #if LOCK_DEBUG > 0
494 		opts &= ~MTX_RECURSE;
495 #endif
496 		m->mtx_recurse++;
497 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
498 		if (LOCK_LOG_TEST(&m->lock_object, opts))
499 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
500 		return;
501 	}
502 #if LOCK_DEBUG > 0
503 	opts &= ~MTX_RECURSE;
504 #endif
505 
506 #ifdef HWPMC_HOOKS
507 	PMC_SOFT_CALL( , , lock, failed);
508 #endif
509 	lock_profile_obtain_lock_failed(&m->lock_object,
510 		    &contested, &waittime);
511 	if (LOCK_LOG_TEST(&m->lock_object, opts))
512 		CTR4(KTR_LOCK,
513 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
514 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
515 #ifdef LOCK_PROFILING
516 	doing_lockprof = 1;
517 #elif defined(KDTRACE_HOOKS)
518 	doing_lockprof = lockstat_enabled;
519 	if (__predict_false(doing_lockprof))
520 		all_time -= lockstat_nsecs(&m->lock_object);
521 #endif
522 
523 	for (;;) {
524 		if (v == MTX_UNOWNED) {
525 			if (_mtx_obtain_lock_fetch(m, &v, tid))
526 				break;
527 			continue;
528 		}
529 #ifdef KDTRACE_HOOKS
530 		lda.spin_cnt++;
531 #endif
532 #ifdef ADAPTIVE_MUTEXES
533 		/*
534 		 * If the owner is running on another CPU, spin until the
535 		 * owner stops running or the state of the lock changes.
536 		 */
537 		owner = lv_mtx_owner(v);
538 		if (TD_IS_RUNNING(owner)) {
539 			if (LOCK_LOG_TEST(&m->lock_object, 0))
540 				CTR3(KTR_LOCK,
541 				    "%s: spinning on %p held by %p",
542 				    __func__, m, owner);
543 			KTR_STATE1(KTR_SCHED, "thread",
544 			    sched_tdname((struct thread *)tid),
545 			    "spinning", "lockname:\"%s\"",
546 			    m->lock_object.lo_name);
547 			do {
548 				lock_delay(&lda);
549 				v = MTX_READ_VALUE(m);
550 				owner = lv_mtx_owner(v);
551 			} while (v != MTX_UNOWNED && TD_IS_RUNNING(owner));
552 			KTR_STATE0(KTR_SCHED, "thread",
553 			    sched_tdname((struct thread *)tid),
554 			    "running");
555 			continue;
556 		}
557 #endif
558 
559 		ts = turnstile_trywait(&m->lock_object);
560 		v = MTX_READ_VALUE(m);
561 
562 		/*
563 		 * Check if the lock has been released while spinning for
564 		 * the turnstile chain lock.
565 		 */
566 		if (v == MTX_UNOWNED) {
567 			turnstile_cancel(ts);
568 			continue;
569 		}
570 
571 #ifdef ADAPTIVE_MUTEXES
572 		/*
573 		 * The current lock owner might have started executing
574 		 * on another CPU (or the lock could have changed
575 		 * owners) while we were waiting on the turnstile
576 		 * chain lock.  If so, drop the turnstile lock and try
577 		 * again.
578 		 */
579 		owner = lv_mtx_owner(v);
580 		if (TD_IS_RUNNING(owner)) {
581 			turnstile_cancel(ts);
582 			continue;
583 		}
584 #endif
585 
586 		/*
587 		 * If the mutex isn't already contested and a failure occurs
588 		 * setting the contested bit, the mutex was either released
589 		 * or the state of the MTX_RECURSED bit changed.
590 		 */
591 		if ((v & MTX_CONTESTED) == 0 &&
592 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
593 			turnstile_cancel(ts);
594 			v = MTX_READ_VALUE(m);
595 			continue;
596 		}
597 
598 		/*
599 		 * We definitely must sleep for this lock.
600 		 */
601 		mtx_assert(m, MA_NOTOWNED);
602 
603 #ifdef KTR
604 		if (!cont_logged) {
605 			CTR6(KTR_CONTENTION,
606 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
607 			    (void *)tid, file, line, m->lock_object.lo_name,
608 			    WITNESS_FILE(&m->lock_object),
609 			    WITNESS_LINE(&m->lock_object));
610 			cont_logged = 1;
611 		}
612 #endif
613 
614 		/*
615 		 * Block on the turnstile.
616 		 */
617 #ifdef KDTRACE_HOOKS
618 		sleep_time -= lockstat_nsecs(&m->lock_object);
619 #endif
620 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
621 #ifdef KDTRACE_HOOKS
622 		sleep_time += lockstat_nsecs(&m->lock_object);
623 		sleep_cnt++;
624 #endif
625 		v = MTX_READ_VALUE(m);
626 	}
627 #ifdef KTR
628 	if (cont_logged) {
629 		CTR4(KTR_CONTENTION,
630 		    "contention end: %s acquired by %p at %s:%d",
631 		    m->lock_object.lo_name, (void *)tid, file, line);
632 	}
633 #endif
634 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
635 	if (__predict_true(!doing_lockprof))
636 		return;
637 #endif
638 #ifdef KDTRACE_HOOKS
639 	all_time += lockstat_nsecs(&m->lock_object);
640 #endif
641 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
642 	    waittime, file, line);
643 #ifdef KDTRACE_HOOKS
644 	if (sleep_time)
645 		LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
646 
647 	/*
648 	 * Only record the loops spinning and not sleeping.
649 	 */
650 	if (lda.spin_cnt > sleep_cnt)
651 		LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
652 #endif
653 }
654 
655 static void
656 _mtx_lock_spin_failed(struct mtx *m)
657 {
658 	struct thread *td;
659 
660 	td = mtx_owner(m);
661 
662 	/* If the mutex is unlocked, try again. */
663 	if (td == NULL)
664 		return;
665 
666 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
667 	    m, m->lock_object.lo_name, td, td->td_tid);
668 #ifdef WITNESS
669 	witness_display_spinlock(&m->lock_object, td, printf);
670 #endif
671 	panic("spin lock held too long");
672 }
673 
674 #ifdef SMP
675 /*
676  * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
677  *
678  * This is only called if we need to actually spin for the lock. Recursion
679  * is handled inline.
680  */
681 void
682 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, uintptr_t tid,
683     int opts, const char *file, int line)
684 {
685 	struct mtx *m;
686 	struct lock_delay_arg lda;
687 #ifdef LOCK_PROFILING
688 	int contested = 0;
689 	uint64_t waittime = 0;
690 #endif
691 #ifdef KDTRACE_HOOKS
692 	int64_t spin_time = 0;
693 #endif
694 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
695 	int doing_lockprof;
696 #endif
697 
698 	if (SCHEDULER_STOPPED())
699 		return;
700 
701 	lock_delay_arg_init(&lda, &mtx_spin_delay);
702 	m = mtxlock2mtx(c);
703 
704 	if (__predict_false(v == MTX_UNOWNED))
705 		v = MTX_READ_VALUE(m);
706 
707 	if (__predict_false(v == tid)) {
708 		m->mtx_recurse++;
709 		return;
710 	}
711 
712 	if (LOCK_LOG_TEST(&m->lock_object, opts))
713 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
714 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
715 	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
716 
717 #ifdef HWPMC_HOOKS
718 	PMC_SOFT_CALL( , , lock, failed);
719 #endif
720 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
721 #ifdef LOCK_PROFILING
722 	doing_lockprof = 1;
723 #elif defined(KDTRACE_HOOKS)
724 	doing_lockprof = lockstat_enabled;
725 	if (__predict_false(doing_lockprof))
726 		spin_time -= lockstat_nsecs(&m->lock_object);
727 #endif
728 	for (;;) {
729 		if (v == MTX_UNOWNED) {
730 			if (_mtx_obtain_lock_fetch(m, &v, tid))
731 				break;
732 			continue;
733 		}
734 		/* Give interrupts a chance while we spin. */
735 		spinlock_exit();
736 		do {
737 			if (lda.spin_cnt < 10000000) {
738 				lock_delay(&lda);
739 			} else {
740 				lda.spin_cnt++;
741 				if (lda.spin_cnt < 60000000 || kdb_active ||
742 				    panicstr != NULL)
743 					DELAY(1);
744 				else
745 					_mtx_lock_spin_failed(m);
746 				cpu_spinwait();
747 			}
748 			v = MTX_READ_VALUE(m);
749 		} while (v != MTX_UNOWNED);
750 		spinlock_enter();
751 	}
752 
753 	if (LOCK_LOG_TEST(&m->lock_object, opts))
754 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
755 	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
756 	    "running");
757 
758 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
759 	if (__predict_true(!doing_lockprof))
760 		return;
761 #endif
762 #ifdef KDTRACE_HOOKS
763 	spin_time += lockstat_nsecs(&m->lock_object);
764 #endif
765 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
766 	    contested, waittime, file, line);
767 #ifdef KDTRACE_HOOKS
768 	if (spin_time != 0)
769 		LOCKSTAT_RECORD1(spin__spin, m, spin_time);
770 #endif
771 }
772 #endif /* SMP */
773 
774 void
775 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
776 {
777 	struct mtx *m;
778 	uintptr_t tid, v;
779 	struct lock_delay_arg lda;
780 #ifdef LOCK_PROFILING
781 	int contested = 0;
782 	uint64_t waittime = 0;
783 #endif
784 #ifdef KDTRACE_HOOKS
785 	int64_t spin_time = 0;
786 #endif
787 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
788 	int doing_lockprof = 1;
789 #endif
790 
791 	tid = (uintptr_t)curthread;
792 
793 	if (SCHEDULER_STOPPED()) {
794 		/*
795 		 * Ensure that spinlock sections are balanced even when the
796 		 * scheduler is stopped, since we may otherwise inadvertently
797 		 * re-enable interrupts while dumping core.
798 		 */
799 		spinlock_enter();
800 		return;
801 	}
802 
803 	lock_delay_arg_init(&lda, &mtx_spin_delay);
804 
805 #ifdef LOCK_PROFILING
806 	doing_lockprof = 1;
807 #elif defined(KDTRACE_HOOKS)
808 	doing_lockprof = lockstat_enabled;
809 	if (__predict_false(doing_lockprof))
810 		spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
811 #endif
812 	for (;;) {
813 retry:
814 		v = MTX_UNOWNED;
815 		spinlock_enter();
816 		m = td->td_lock;
817 		KASSERT(m->mtx_lock != MTX_DESTROYED,
818 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
819 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
820 		    ("thread_lock() of sleep mutex %s @ %s:%d",
821 		    m->lock_object.lo_name, file, line));
822 		if (mtx_owned(m))
823 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
824 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
825 			    m->lock_object.lo_name, file, line));
826 		WITNESS_CHECKORDER(&m->lock_object,
827 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
828 		for (;;) {
829 			if (_mtx_obtain_lock_fetch(m, &v, tid))
830 				break;
831 			if (v == MTX_UNOWNED)
832 				continue;
833 			if (v == tid) {
834 				m->mtx_recurse++;
835 				break;
836 			}
837 #ifdef HWPMC_HOOKS
838 			PMC_SOFT_CALL( , , lock, failed);
839 #endif
840 			lock_profile_obtain_lock_failed(&m->lock_object,
841 			    &contested, &waittime);
842 			/* Give interrupts a chance while we spin. */
843 			spinlock_exit();
844 			do {
845 				if (lda.spin_cnt < 10000000) {
846 					lock_delay(&lda);
847 				} else {
848 					lda.spin_cnt++;
849 					if (lda.spin_cnt < 60000000 ||
850 					    kdb_active || panicstr != NULL)
851 						DELAY(1);
852 					else
853 						_mtx_lock_spin_failed(m);
854 					cpu_spinwait();
855 				}
856 				if (m != td->td_lock)
857 					goto retry;
858 				v = MTX_READ_VALUE(m);
859 			} while (v != MTX_UNOWNED);
860 			spinlock_enter();
861 		}
862 		if (m == td->td_lock)
863 			break;
864 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
865 	}
866 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
867 	    line);
868 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
869 
870 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
871 	if (__predict_true(!doing_lockprof))
872 		return;
873 #endif
874 #ifdef KDTRACE_HOOKS
875 	spin_time += lockstat_nsecs(&m->lock_object);
876 #endif
877 	if (m->mtx_recurse == 0)
878 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
879 		    contested, waittime, file, line);
880 #ifdef KDTRACE_HOOKS
881 	if (spin_time != 0)
882 		LOCKSTAT_RECORD1(thread__spin, m, spin_time);
883 #endif
884 }
885 
886 struct mtx *
887 thread_lock_block(struct thread *td)
888 {
889 	struct mtx *lock;
890 
891 	THREAD_LOCK_ASSERT(td, MA_OWNED);
892 	lock = td->td_lock;
893 	td->td_lock = &blocked_lock;
894 	mtx_unlock_spin(lock);
895 
896 	return (lock);
897 }
898 
899 void
900 thread_lock_unblock(struct thread *td, struct mtx *new)
901 {
902 	mtx_assert(new, MA_OWNED);
903 	MPASS(td->td_lock == &blocked_lock);
904 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
905 }
906 
907 void
908 thread_lock_set(struct thread *td, struct mtx *new)
909 {
910 	struct mtx *lock;
911 
912 	mtx_assert(new, MA_OWNED);
913 	THREAD_LOCK_ASSERT(td, MA_OWNED);
914 	lock = td->td_lock;
915 	td->td_lock = new;
916 	mtx_unlock_spin(lock);
917 }
918 
919 /*
920  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
921  *
922  * We are only called here if the lock is recursed, contested (i.e. we
923  * need to wake up a blocked thread) or lockstat probe is active.
924  */
925 #if LOCK_DEBUG > 0
926 void
927 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
928 #else
929 void
930 __mtx_unlock_sleep(volatile uintptr_t *c)
931 #endif
932 {
933 	struct mtx *m;
934 	struct turnstile *ts;
935 	uintptr_t tid, v;
936 
937 	if (SCHEDULER_STOPPED())
938 		return;
939 
940 	tid = (uintptr_t)curthread;
941 	m = mtxlock2mtx(c);
942 	v = MTX_READ_VALUE(m);
943 
944 	if (v & MTX_RECURSED) {
945 		if (--(m->mtx_recurse) == 0)
946 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
947 		if (LOCK_LOG_TEST(&m->lock_object, opts))
948 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
949 		return;
950 	}
951 
952 	LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m);
953 	if (v == tid && _mtx_release_lock(m, tid))
954 		return;
955 
956 	/*
957 	 * We have to lock the chain before the turnstile so this turnstile
958 	 * can be removed from the hash list if it is empty.
959 	 */
960 	turnstile_chain_lock(&m->lock_object);
961 	ts = turnstile_lookup(&m->lock_object);
962 	if (LOCK_LOG_TEST(&m->lock_object, opts))
963 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
964 	MPASS(ts != NULL);
965 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
966 	_mtx_release_lock_quick(m);
967 
968 	/*
969 	 * This turnstile is now no longer associated with the mutex.  We can
970 	 * unlock the chain lock so a new turnstile may take it's place.
971 	 */
972 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
973 	turnstile_chain_unlock(&m->lock_object);
974 }
975 
976 /*
977  * All the unlocking of MTX_SPIN locks is done inline.
978  * See the __mtx_unlock_spin() macro for the details.
979  */
980 
981 /*
982  * The backing function for the INVARIANTS-enabled mtx_assert()
983  */
984 #ifdef INVARIANT_SUPPORT
985 void
986 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
987 {
988 	const struct mtx *m;
989 
990 	if (panicstr != NULL || dumping || SCHEDULER_STOPPED())
991 		return;
992 
993 	m = mtxlock2mtx(c);
994 
995 	switch (what) {
996 	case MA_OWNED:
997 	case MA_OWNED | MA_RECURSED:
998 	case MA_OWNED | MA_NOTRECURSED:
999 		if (!mtx_owned(m))
1000 			panic("mutex %s not owned at %s:%d",
1001 			    m->lock_object.lo_name, file, line);
1002 		if (mtx_recursed(m)) {
1003 			if ((what & MA_NOTRECURSED) != 0)
1004 				panic("mutex %s recursed at %s:%d",
1005 				    m->lock_object.lo_name, file, line);
1006 		} else if ((what & MA_RECURSED) != 0) {
1007 			panic("mutex %s unrecursed at %s:%d",
1008 			    m->lock_object.lo_name, file, line);
1009 		}
1010 		break;
1011 	case MA_NOTOWNED:
1012 		if (mtx_owned(m))
1013 			panic("mutex %s owned at %s:%d",
1014 			    m->lock_object.lo_name, file, line);
1015 		break;
1016 	default:
1017 		panic("unknown mtx_assert at %s:%d", file, line);
1018 	}
1019 }
1020 #endif
1021 
1022 /*
1023  * General init routine used by the MTX_SYSINIT() macro.
1024  */
1025 void
1026 mtx_sysinit(void *arg)
1027 {
1028 	struct mtx_args *margs = arg;
1029 
1030 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
1031 	    margs->ma_opts);
1032 }
1033 
1034 /*
1035  * Mutex initialization routine; initialize lock `m' of type contained in
1036  * `opts' with options contained in `opts' and name `name.'  The optional
1037  * lock type `type' is used as a general lock category name for use with
1038  * witness.
1039  */
1040 void
1041 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
1042 {
1043 	struct mtx *m;
1044 	struct lock_class *class;
1045 	int flags;
1046 
1047 	m = mtxlock2mtx(c);
1048 
1049 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
1050 	    MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
1051 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
1052 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
1053 	    &m->mtx_lock));
1054 
1055 	/* Determine lock class and lock flags. */
1056 	if (opts & MTX_SPIN)
1057 		class = &lock_class_mtx_spin;
1058 	else
1059 		class = &lock_class_mtx_sleep;
1060 	flags = 0;
1061 	if (opts & MTX_QUIET)
1062 		flags |= LO_QUIET;
1063 	if (opts & MTX_RECURSE)
1064 		flags |= LO_RECURSABLE;
1065 	if ((opts & MTX_NOWITNESS) == 0)
1066 		flags |= LO_WITNESS;
1067 	if (opts & MTX_DUPOK)
1068 		flags |= LO_DUPOK;
1069 	if (opts & MTX_NOPROFILE)
1070 		flags |= LO_NOPROFILE;
1071 	if (opts & MTX_NEW)
1072 		flags |= LO_NEW;
1073 
1074 	/* Initialize mutex. */
1075 	lock_init(&m->lock_object, class, name, type, flags);
1076 
1077 	m->mtx_lock = MTX_UNOWNED;
1078 	m->mtx_recurse = 0;
1079 }
1080 
1081 /*
1082  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
1083  * passed in as a flag here because if the corresponding mtx_init() was
1084  * called with MTX_QUIET set, then it will already be set in the mutex's
1085  * flags.
1086  */
1087 void
1088 _mtx_destroy(volatile uintptr_t *c)
1089 {
1090 	struct mtx *m;
1091 
1092 	m = mtxlock2mtx(c);
1093 
1094 	if (!mtx_owned(m))
1095 		MPASS(mtx_unowned(m));
1096 	else {
1097 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
1098 
1099 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
1100 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
1101 			spinlock_exit();
1102 		else
1103 			TD_LOCKS_DEC(curthread);
1104 
1105 		lock_profile_release_lock(&m->lock_object);
1106 		/* Tell witness this isn't locked to make it happy. */
1107 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
1108 		    __LINE__);
1109 	}
1110 
1111 	m->mtx_lock = MTX_DESTROYED;
1112 	lock_destroy(&m->lock_object);
1113 }
1114 
1115 /*
1116  * Intialize the mutex code and system mutexes.  This is called from the MD
1117  * startup code prior to mi_startup().  The per-CPU data space needs to be
1118  * setup before this is called.
1119  */
1120 void
1121 mutex_init(void)
1122 {
1123 
1124 	/* Setup turnstiles so that sleep mutexes work. */
1125 	init_turnstiles();
1126 
1127 	/*
1128 	 * Initialize mutexes.
1129 	 */
1130 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
1131 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
1132 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
1133 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
1134 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
1135 	mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
1136 	mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1137 	mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1138 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1139 	mtx_lock(&Giant);
1140 }
1141 
1142 #ifdef DDB
1143 void
1144 db_show_mtx(const struct lock_object *lock)
1145 {
1146 	struct thread *td;
1147 	const struct mtx *m;
1148 
1149 	m = (const struct mtx *)lock;
1150 
1151 	db_printf(" flags: {");
1152 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1153 		db_printf("SPIN");
1154 	else
1155 		db_printf("DEF");
1156 	if (m->lock_object.lo_flags & LO_RECURSABLE)
1157 		db_printf(", RECURSE");
1158 	if (m->lock_object.lo_flags & LO_DUPOK)
1159 		db_printf(", DUPOK");
1160 	db_printf("}\n");
1161 	db_printf(" state: {");
1162 	if (mtx_unowned(m))
1163 		db_printf("UNOWNED");
1164 	else if (mtx_destroyed(m))
1165 		db_printf("DESTROYED");
1166 	else {
1167 		db_printf("OWNED");
1168 		if (m->mtx_lock & MTX_CONTESTED)
1169 			db_printf(", CONTESTED");
1170 		if (m->mtx_lock & MTX_RECURSED)
1171 			db_printf(", RECURSED");
1172 	}
1173 	db_printf("}\n");
1174 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1175 		td = mtx_owner(m);
1176 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1177 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1178 		if (mtx_recursed(m))
1179 			db_printf(" recursed: %d\n", m->mtx_recurse);
1180 	}
1181 }
1182 #endif
1183