xref: /freebsd/sys/kern/kern_mutex.c (revision 276da39af92f48350aa01091a2b8b3e735217eea)
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/sysctl.h>
59 #include <sys/turnstile.h>
60 #include <sys/vmmeter.h>
61 #include <sys/lock_profile.h>
62 
63 #include <machine/atomic.h>
64 #include <machine/bus.h>
65 #include <machine/cpu.h>
66 
67 #include <ddb/ddb.h>
68 
69 #include <fs/devfs/devfs_int.h>
70 
71 #include <vm/vm.h>
72 #include <vm/vm_extern.h>
73 
74 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
75 #define	ADAPTIVE_MUTEXES
76 #endif
77 
78 #ifdef HWPMC_HOOKS
79 #include <sys/pmckern.h>
80 PMC_SOFT_DEFINE( , , lock, failed);
81 #endif
82 
83 /*
84  * Return the mutex address when the lock cookie address is provided.
85  * This functionality assumes that struct mtx* have a member named mtx_lock.
86  */
87 #define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
88 
89 /*
90  * Internal utility macros.
91  */
92 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
93 
94 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
95 
96 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
97 
98 static void	assert_mtx(const struct lock_object *lock, int what);
99 #ifdef DDB
100 static void	db_show_mtx(const struct lock_object *lock);
101 #endif
102 static void	lock_mtx(struct lock_object *lock, uintptr_t how);
103 static void	lock_spin(struct lock_object *lock, uintptr_t how);
104 #ifdef KDTRACE_HOOKS
105 static int	owner_mtx(const struct lock_object *lock,
106 		    struct thread **owner);
107 #endif
108 static uintptr_t unlock_mtx(struct lock_object *lock);
109 static uintptr_t unlock_spin(struct lock_object *lock);
110 
111 /*
112  * Lock classes for sleep and spin mutexes.
113  */
114 struct lock_class lock_class_mtx_sleep = {
115 	.lc_name = "sleep mutex",
116 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
117 	.lc_assert = assert_mtx,
118 #ifdef DDB
119 	.lc_ddb_show = db_show_mtx,
120 #endif
121 	.lc_lock = lock_mtx,
122 	.lc_unlock = unlock_mtx,
123 #ifdef KDTRACE_HOOKS
124 	.lc_owner = owner_mtx,
125 #endif
126 };
127 struct lock_class lock_class_mtx_spin = {
128 	.lc_name = "spin mutex",
129 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
130 	.lc_assert = assert_mtx,
131 #ifdef DDB
132 	.lc_ddb_show = db_show_mtx,
133 #endif
134 	.lc_lock = lock_spin,
135 	.lc_unlock = unlock_spin,
136 #ifdef KDTRACE_HOOKS
137 	.lc_owner = owner_mtx,
138 #endif
139 };
140 
141 /*
142  * System-wide mutexes
143  */
144 struct mtx blocked_lock;
145 struct mtx Giant;
146 
147 void
148 assert_mtx(const struct lock_object *lock, int what)
149 {
150 
151 	mtx_assert((const struct mtx *)lock, what);
152 }
153 
154 void
155 lock_mtx(struct lock_object *lock, uintptr_t how)
156 {
157 
158 	mtx_lock((struct mtx *)lock);
159 }
160 
161 void
162 lock_spin(struct lock_object *lock, uintptr_t how)
163 {
164 
165 	panic("spin locks can only use msleep_spin");
166 }
167 
168 uintptr_t
169 unlock_mtx(struct lock_object *lock)
170 {
171 	struct mtx *m;
172 
173 	m = (struct mtx *)lock;
174 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
175 	mtx_unlock(m);
176 	return (0);
177 }
178 
179 uintptr_t
180 unlock_spin(struct lock_object *lock)
181 {
182 
183 	panic("spin locks can only use msleep_spin");
184 }
185 
186 #ifdef KDTRACE_HOOKS
187 int
188 owner_mtx(const struct lock_object *lock, struct thread **owner)
189 {
190 	const struct mtx *m = (const struct mtx *)lock;
191 
192 	*owner = mtx_owner(m);
193 	return (mtx_unowned(m) == 0);
194 }
195 #endif
196 
197 /*
198  * Function versions of the inlined __mtx_* macros.  These are used by
199  * modules and can also be called from assembly language if needed.
200  */
201 void
202 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
203 {
204 	struct mtx *m;
205 
206 	if (SCHEDULER_STOPPED())
207 		return;
208 
209 	m = mtxlock2mtx(c);
210 
211 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
212 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
213 	    curthread, m->lock_object.lo_name, file, line));
214 	KASSERT(m->mtx_lock != MTX_DESTROYED,
215 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
216 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
217 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
218 	    file, line));
219 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
220 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
221 
222 	__mtx_lock(m, curthread, opts, file, line);
223 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
224 	    line);
225 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
226 	    file, line);
227 	curthread->td_locks++;
228 }
229 
230 void
231 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
232 {
233 	struct mtx *m;
234 
235 	if (SCHEDULER_STOPPED())
236 		return;
237 
238 	m = mtxlock2mtx(c);
239 
240 	KASSERT(m->mtx_lock != MTX_DESTROYED,
241 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
242 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
243 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
244 	    file, line));
245 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
246 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
247 	    line);
248 	mtx_assert(m, MA_OWNED);
249 
250 	__mtx_unlock(m, curthread, opts, file, line);
251 	curthread->td_locks--;
252 }
253 
254 void
255 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
256     int line)
257 {
258 	struct mtx *m;
259 
260 	if (SCHEDULER_STOPPED())
261 		return;
262 
263 	m = mtxlock2mtx(c);
264 
265 	KASSERT(m->mtx_lock != MTX_DESTROYED,
266 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
267 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
268 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
269 	    m->lock_object.lo_name, file, line));
270 	if (mtx_owned(m))
271 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
272 		    (opts & MTX_RECURSE) != 0,
273 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
274 		    m->lock_object.lo_name, file, line));
275 	opts &= ~MTX_RECURSE;
276 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
277 	    file, line, NULL);
278 	__mtx_lock_spin(m, curthread, opts, file, line);
279 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
280 	    line);
281 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
282 }
283 
284 void
285 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
286     int line)
287 {
288 	struct mtx *m;
289 
290 	if (SCHEDULER_STOPPED())
291 		return;
292 
293 	m = mtxlock2mtx(c);
294 
295 	KASSERT(m->mtx_lock != MTX_DESTROYED,
296 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
297 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
298 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
299 	    m->lock_object.lo_name, file, line));
300 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
301 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
302 	    line);
303 	mtx_assert(m, MA_OWNED);
304 
305 	__mtx_unlock_spin(m);
306 }
307 
308 /*
309  * The important part of mtx_trylock{,_flags}()
310  * Tries to acquire lock `m.'  If this function is called on a mutex that
311  * is already owned, it will recursively acquire the lock.
312  */
313 int
314 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
315 {
316 	struct mtx *m;
317 #ifdef LOCK_PROFILING
318 	uint64_t waittime = 0;
319 	int contested = 0;
320 #endif
321 	int rval;
322 
323 	if (SCHEDULER_STOPPED())
324 		return (1);
325 
326 	m = mtxlock2mtx(c);
327 
328 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
329 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
330 	    curthread, m->lock_object.lo_name, file, line));
331 	KASSERT(m->mtx_lock != MTX_DESTROYED,
332 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
333 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
334 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
335 	    file, line));
336 
337 	if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
338 	    (opts & MTX_RECURSE) != 0)) {
339 		m->mtx_recurse++;
340 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
341 		rval = 1;
342 	} else
343 		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
344 	opts &= ~MTX_RECURSE;
345 
346 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
347 	if (rval) {
348 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
349 		    file, line);
350 		curthread->td_locks++;
351 		if (m->mtx_recurse == 0)
352 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
353 			    m, contested, waittime, file, line);
354 
355 	}
356 
357 	return (rval);
358 }
359 
360 /*
361  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
362  *
363  * We call this if the lock is either contested (i.e. we need to go to
364  * sleep waiting for it), or if we need to recurse on it.
365  */
366 void
367 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
368     const char *file, int line)
369 {
370 	struct mtx *m;
371 	struct turnstile *ts;
372 	uintptr_t v;
373 #ifdef ADAPTIVE_MUTEXES
374 	volatile struct thread *owner;
375 #endif
376 #ifdef KTR
377 	int cont_logged = 0;
378 #endif
379 #ifdef LOCK_PROFILING
380 	int contested = 0;
381 	uint64_t waittime = 0;
382 #endif
383 #ifdef KDTRACE_HOOKS
384 	uint64_t spin_cnt = 0;
385 	uint64_t sleep_cnt = 0;
386 	int64_t sleep_time = 0;
387 	int64_t all_time = 0;
388 #endif
389 
390 	if (SCHEDULER_STOPPED())
391 		return;
392 
393 	m = mtxlock2mtx(c);
394 
395 	if (mtx_owned(m)) {
396 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
397 		    (opts & MTX_RECURSE) != 0,
398 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
399 		    m->lock_object.lo_name, file, line));
400 		opts &= ~MTX_RECURSE;
401 		m->mtx_recurse++;
402 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
403 		if (LOCK_LOG_TEST(&m->lock_object, opts))
404 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
405 		return;
406 	}
407 	opts &= ~MTX_RECURSE;
408 
409 #ifdef HWPMC_HOOKS
410 	PMC_SOFT_CALL( , , lock, failed);
411 #endif
412 	lock_profile_obtain_lock_failed(&m->lock_object,
413 		    &contested, &waittime);
414 	if (LOCK_LOG_TEST(&m->lock_object, opts))
415 		CTR4(KTR_LOCK,
416 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
417 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
418 #ifdef KDTRACE_HOOKS
419 	all_time -= lockstat_nsecs();
420 #endif
421 
422 	while (!_mtx_obtain_lock(m, tid)) {
423 #ifdef KDTRACE_HOOKS
424 		spin_cnt++;
425 #endif
426 #ifdef ADAPTIVE_MUTEXES
427 		/*
428 		 * If the owner is running on another CPU, spin until the
429 		 * owner stops running or the state of the lock changes.
430 		 */
431 		v = m->mtx_lock;
432 		if (v != MTX_UNOWNED) {
433 			owner = (struct thread *)(v & ~MTX_FLAGMASK);
434 			if (TD_IS_RUNNING(owner)) {
435 				if (LOCK_LOG_TEST(&m->lock_object, 0))
436 					CTR3(KTR_LOCK,
437 					    "%s: spinning on %p held by %p",
438 					    __func__, m, owner);
439 				KTR_STATE1(KTR_SCHED, "thread",
440 				    sched_tdname((struct thread *)tid),
441 				    "spinning", "lockname:\"%s\"",
442 				    m->lock_object.lo_name);
443 				while (mtx_owner(m) == owner &&
444 				    TD_IS_RUNNING(owner)) {
445 					cpu_spinwait();
446 #ifdef KDTRACE_HOOKS
447 					spin_cnt++;
448 #endif
449 				}
450 				KTR_STATE0(KTR_SCHED, "thread",
451 				    sched_tdname((struct thread *)tid),
452 				    "running");
453 				continue;
454 			}
455 		}
456 #endif
457 
458 		ts = turnstile_trywait(&m->lock_object);
459 		v = m->mtx_lock;
460 
461 		/*
462 		 * Check if the lock has been released while spinning for
463 		 * the turnstile chain lock.
464 		 */
465 		if (v == MTX_UNOWNED) {
466 			turnstile_cancel(ts);
467 			continue;
468 		}
469 
470 #ifdef ADAPTIVE_MUTEXES
471 		/*
472 		 * The current lock owner might have started executing
473 		 * on another CPU (or the lock could have changed
474 		 * owners) while we were waiting on the turnstile
475 		 * chain lock.  If so, drop the turnstile lock and try
476 		 * again.
477 		 */
478 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
479 		if (TD_IS_RUNNING(owner)) {
480 			turnstile_cancel(ts);
481 			continue;
482 		}
483 #endif
484 
485 		/*
486 		 * If the mutex isn't already contested and a failure occurs
487 		 * setting the contested bit, the mutex was either released
488 		 * or the state of the MTX_RECURSED bit changed.
489 		 */
490 		if ((v & MTX_CONTESTED) == 0 &&
491 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
492 			turnstile_cancel(ts);
493 			continue;
494 		}
495 
496 		/*
497 		 * We definitely must sleep for this lock.
498 		 */
499 		mtx_assert(m, MA_NOTOWNED);
500 
501 #ifdef KTR
502 		if (!cont_logged) {
503 			CTR6(KTR_CONTENTION,
504 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
505 			    (void *)tid, file, line, m->lock_object.lo_name,
506 			    WITNESS_FILE(&m->lock_object),
507 			    WITNESS_LINE(&m->lock_object));
508 			cont_logged = 1;
509 		}
510 #endif
511 
512 		/*
513 		 * Block on the turnstile.
514 		 */
515 #ifdef KDTRACE_HOOKS
516 		sleep_time -= lockstat_nsecs();
517 #endif
518 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
519 #ifdef KDTRACE_HOOKS
520 		sleep_time += lockstat_nsecs();
521 		sleep_cnt++;
522 #endif
523 	}
524 #ifdef KDTRACE_HOOKS
525 	all_time += lockstat_nsecs();
526 #endif
527 #ifdef KTR
528 	if (cont_logged) {
529 		CTR4(KTR_CONTENTION,
530 		    "contention end: %s acquired by %p at %s:%d",
531 		    m->lock_object.lo_name, (void *)tid, file, line);
532 	}
533 #endif
534 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
535 	    waittime, file, line);
536 #ifdef KDTRACE_HOOKS
537 	if (sleep_time)
538 		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
539 
540 	/*
541 	 * Only record the loops spinning and not sleeping.
542 	 */
543 	if (spin_cnt > sleep_cnt)
544 		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (all_time - sleep_time));
545 #endif
546 }
547 
548 static void
549 _mtx_lock_spin_failed(struct mtx *m)
550 {
551 	struct thread *td;
552 
553 	td = mtx_owner(m);
554 
555 	/* If the mutex is unlocked, try again. */
556 	if (td == NULL)
557 		return;
558 
559 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
560 	    m, m->lock_object.lo_name, td, td->td_tid);
561 #ifdef WITNESS
562 	witness_display_spinlock(&m->lock_object, td, printf);
563 #endif
564 	panic("spin lock held too long");
565 }
566 
567 #ifdef SMP
568 /*
569  * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
570  *
571  * This is only called if we need to actually spin for the lock. Recursion
572  * is handled inline.
573  */
574 void
575 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
576     const char *file, int line)
577 {
578 	struct mtx *m;
579 	int i = 0;
580 #ifdef LOCK_PROFILING
581 	int contested = 0;
582 	uint64_t waittime = 0;
583 #endif
584 #ifdef KDTRACE_HOOKS
585 	int64_t spin_time = 0;
586 #endif
587 
588 	if (SCHEDULER_STOPPED())
589 		return;
590 
591 	m = mtxlock2mtx(c);
592 
593 	if (LOCK_LOG_TEST(&m->lock_object, opts))
594 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
595 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
596 	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
597 
598 #ifdef HWPMC_HOOKS
599 	PMC_SOFT_CALL( , , lock, failed);
600 #endif
601 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
602 #ifdef KDTRACE_HOOKS
603 	spin_time -= lockstat_nsecs();
604 #endif
605 	while (!_mtx_obtain_lock(m, tid)) {
606 
607 		/* Give interrupts a chance while we spin. */
608 		spinlock_exit();
609 		while (m->mtx_lock != MTX_UNOWNED) {
610 			if (i++ < 10000000) {
611 				cpu_spinwait();
612 				continue;
613 			}
614 			if (i < 60000000 || kdb_active || panicstr != NULL)
615 				DELAY(1);
616 			else
617 				_mtx_lock_spin_failed(m);
618 			cpu_spinwait();
619 		}
620 		spinlock_enter();
621 	}
622 #ifdef KDTRACE_HOOKS
623 	spin_time += lockstat_nsecs();
624 #endif
625 
626 	if (LOCK_LOG_TEST(&m->lock_object, opts))
627 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
628 	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
629 	    "running");
630 
631 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
632 	    contested, waittime, (file), (line));
633 	LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, spin_time);
634 }
635 #endif /* SMP */
636 
637 void
638 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
639 {
640 	struct mtx *m;
641 	uintptr_t tid;
642 	int i;
643 #ifdef LOCK_PROFILING
644 	int contested = 0;
645 	uint64_t waittime = 0;
646 #endif
647 #ifdef KDTRACE_HOOKS
648 	int64_t spin_time = 0;
649 #endif
650 
651 	i = 0;
652 	tid = (uintptr_t)curthread;
653 
654 	if (SCHEDULER_STOPPED())
655 		return;
656 
657 #ifdef KDTRACE_HOOKS
658 	spin_time -= lockstat_nsecs();
659 #endif
660 	for (;;) {
661 retry:
662 		spinlock_enter();
663 		m = td->td_lock;
664 		KASSERT(m->mtx_lock != MTX_DESTROYED,
665 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
666 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
667 		    ("thread_lock() of sleep mutex %s @ %s:%d",
668 		    m->lock_object.lo_name, file, line));
669 		if (mtx_owned(m))
670 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
671 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
672 			    m->lock_object.lo_name, file, line));
673 		WITNESS_CHECKORDER(&m->lock_object,
674 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
675 		while (!_mtx_obtain_lock(m, tid)) {
676 			if (m->mtx_lock == tid) {
677 				m->mtx_recurse++;
678 				break;
679 			}
680 #ifdef HWPMC_HOOKS
681 			PMC_SOFT_CALL( , , lock, failed);
682 #endif
683 			lock_profile_obtain_lock_failed(&m->lock_object,
684 			    &contested, &waittime);
685 			/* Give interrupts a chance while we spin. */
686 			spinlock_exit();
687 			while (m->mtx_lock != MTX_UNOWNED) {
688 				if (i++ < 10000000)
689 					cpu_spinwait();
690 				else if (i < 60000000 ||
691 				    kdb_active || panicstr != NULL)
692 					DELAY(1);
693 				else
694 					_mtx_lock_spin_failed(m);
695 				cpu_spinwait();
696 				if (m != td->td_lock)
697 					goto retry;
698 			}
699 			spinlock_enter();
700 		}
701 		if (m == td->td_lock)
702 			break;
703 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
704 	}
705 #ifdef KDTRACE_HOOKS
706 	spin_time += lockstat_nsecs();
707 #endif
708 	if (m->mtx_recurse == 0)
709 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
710 		    m, contested, waittime, (file), (line));
711 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
712 	    line);
713 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
714 	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_time);
715 }
716 
717 struct mtx *
718 thread_lock_block(struct thread *td)
719 {
720 	struct mtx *lock;
721 
722 	THREAD_LOCK_ASSERT(td, MA_OWNED);
723 	lock = td->td_lock;
724 	td->td_lock = &blocked_lock;
725 	mtx_unlock_spin(lock);
726 
727 	return (lock);
728 }
729 
730 void
731 thread_lock_unblock(struct thread *td, struct mtx *new)
732 {
733 	mtx_assert(new, MA_OWNED);
734 	MPASS(td->td_lock == &blocked_lock);
735 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
736 }
737 
738 void
739 thread_lock_set(struct thread *td, struct mtx *new)
740 {
741 	struct mtx *lock;
742 
743 	mtx_assert(new, MA_OWNED);
744 	THREAD_LOCK_ASSERT(td, MA_OWNED);
745 	lock = td->td_lock;
746 	td->td_lock = new;
747 	mtx_unlock_spin(lock);
748 }
749 
750 /*
751  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
752  *
753  * We are only called here if the lock is recursed or contested (i.e. we
754  * need to wake up a blocked thread).
755  */
756 void
757 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
758 {
759 	struct mtx *m;
760 	struct turnstile *ts;
761 
762 	if (SCHEDULER_STOPPED())
763 		return;
764 
765 	m = mtxlock2mtx(c);
766 
767 	if (mtx_recursed(m)) {
768 		if (--(m->mtx_recurse) == 0)
769 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
770 		if (LOCK_LOG_TEST(&m->lock_object, opts))
771 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
772 		return;
773 	}
774 
775 	/*
776 	 * We have to lock the chain before the turnstile so this turnstile
777 	 * can be removed from the hash list if it is empty.
778 	 */
779 	turnstile_chain_lock(&m->lock_object);
780 	ts = turnstile_lookup(&m->lock_object);
781 	if (LOCK_LOG_TEST(&m->lock_object, opts))
782 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
783 	MPASS(ts != NULL);
784 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
785 	_mtx_release_lock_quick(m);
786 
787 	/*
788 	 * This turnstile is now no longer associated with the mutex.  We can
789 	 * unlock the chain lock so a new turnstile may take it's place.
790 	 */
791 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
792 	turnstile_chain_unlock(&m->lock_object);
793 }
794 
795 /*
796  * All the unlocking of MTX_SPIN locks is done inline.
797  * See the __mtx_unlock_spin() macro for the details.
798  */
799 
800 /*
801  * The backing function for the INVARIANTS-enabled mtx_assert()
802  */
803 #ifdef INVARIANT_SUPPORT
804 void
805 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
806 {
807 	const struct mtx *m;
808 
809 	if (panicstr != NULL || dumping)
810 		return;
811 
812 	m = mtxlock2mtx(c);
813 
814 	switch (what) {
815 	case MA_OWNED:
816 	case MA_OWNED | MA_RECURSED:
817 	case MA_OWNED | MA_NOTRECURSED:
818 		if (!mtx_owned(m))
819 			panic("mutex %s not owned at %s:%d",
820 			    m->lock_object.lo_name, file, line);
821 		if (mtx_recursed(m)) {
822 			if ((what & MA_NOTRECURSED) != 0)
823 				panic("mutex %s recursed at %s:%d",
824 				    m->lock_object.lo_name, file, line);
825 		} else if ((what & MA_RECURSED) != 0) {
826 			panic("mutex %s unrecursed at %s:%d",
827 			    m->lock_object.lo_name, file, line);
828 		}
829 		break;
830 	case MA_NOTOWNED:
831 		if (mtx_owned(m))
832 			panic("mutex %s owned at %s:%d",
833 			    m->lock_object.lo_name, file, line);
834 		break;
835 	default:
836 		panic("unknown mtx_assert at %s:%d", file, line);
837 	}
838 }
839 #endif
840 
841 /*
842  * The MUTEX_DEBUG-enabled mtx_validate()
843  *
844  * Most of these checks have been moved off into the LO_INITIALIZED flag
845  * maintained by the witness code.
846  */
847 #ifdef MUTEX_DEBUG
848 
849 void	mtx_validate(struct mtx *);
850 
851 void
852 mtx_validate(struct mtx *m)
853 {
854 
855 /*
856  * XXX: When kernacc() does not require Giant we can reenable this check
857  */
858 #ifdef notyet
859 	/*
860 	 * Can't call kernacc() from early init386(), especially when
861 	 * initializing Giant mutex, because some stuff in kernacc()
862 	 * requires Giant itself.
863 	 */
864 	if (!cold)
865 		if (!kernacc((caddr_t)m, sizeof(m),
866 		    VM_PROT_READ | VM_PROT_WRITE))
867 			panic("Can't read and write to mutex %p", m);
868 #endif
869 }
870 #endif
871 
872 /*
873  * General init routine used by the MTX_SYSINIT() macro.
874  */
875 void
876 mtx_sysinit(void *arg)
877 {
878 	struct mtx_args *margs = arg;
879 
880 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
881 	    margs->ma_opts);
882 }
883 
884 /*
885  * Mutex initialization routine; initialize lock `m' of type contained in
886  * `opts' with options contained in `opts' and name `name.'  The optional
887  * lock type `type' is used as a general lock category name for use with
888  * witness.
889  */
890 void
891 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
892 {
893 	struct mtx *m;
894 	struct lock_class *class;
895 	int flags;
896 
897 	m = mtxlock2mtx(c);
898 
899 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
900 	    MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
901 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
902 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
903 	    &m->mtx_lock));
904 
905 #ifdef MUTEX_DEBUG
906 	/* Diagnostic and error correction */
907 	mtx_validate(m);
908 #endif
909 
910 	/* Determine lock class and lock flags. */
911 	if (opts & MTX_SPIN)
912 		class = &lock_class_mtx_spin;
913 	else
914 		class = &lock_class_mtx_sleep;
915 	flags = 0;
916 	if (opts & MTX_QUIET)
917 		flags |= LO_QUIET;
918 	if (opts & MTX_RECURSE)
919 		flags |= LO_RECURSABLE;
920 	if ((opts & MTX_NOWITNESS) == 0)
921 		flags |= LO_WITNESS;
922 	if (opts & MTX_DUPOK)
923 		flags |= LO_DUPOK;
924 	if (opts & MTX_NOPROFILE)
925 		flags |= LO_NOPROFILE;
926 	if (opts & MTX_NEW)
927 		flags |= LO_NEW;
928 
929 	/* Initialize mutex. */
930 	lock_init(&m->lock_object, class, name, type, flags);
931 
932 	m->mtx_lock = MTX_UNOWNED;
933 	m->mtx_recurse = 0;
934 }
935 
936 /*
937  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
938  * passed in as a flag here because if the corresponding mtx_init() was
939  * called with MTX_QUIET set, then it will already be set in the mutex's
940  * flags.
941  */
942 void
943 _mtx_destroy(volatile uintptr_t *c)
944 {
945 	struct mtx *m;
946 
947 	m = mtxlock2mtx(c);
948 
949 	if (!mtx_owned(m))
950 		MPASS(mtx_unowned(m));
951 	else {
952 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
953 
954 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
955 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
956 			spinlock_exit();
957 		else
958 			curthread->td_locks--;
959 
960 		lock_profile_release_lock(&m->lock_object);
961 		/* Tell witness this isn't locked to make it happy. */
962 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
963 		    __LINE__);
964 	}
965 
966 	m->mtx_lock = MTX_DESTROYED;
967 	lock_destroy(&m->lock_object);
968 }
969 
970 /*
971  * Intialize the mutex code and system mutexes.  This is called from the MD
972  * startup code prior to mi_startup().  The per-CPU data space needs to be
973  * setup before this is called.
974  */
975 void
976 mutex_init(void)
977 {
978 
979 	/* Setup turnstiles so that sleep mutexes work. */
980 	init_turnstiles();
981 
982 	/*
983 	 * Initialize mutexes.
984 	 */
985 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
986 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
987 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
988 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
989 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
990 	mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
991 	mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
992 	mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
993 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
994 	mtx_lock(&Giant);
995 }
996 
997 #ifdef DDB
998 void
999 db_show_mtx(const struct lock_object *lock)
1000 {
1001 	struct thread *td;
1002 	const struct mtx *m;
1003 
1004 	m = (const struct mtx *)lock;
1005 
1006 	db_printf(" flags: {");
1007 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1008 		db_printf("SPIN");
1009 	else
1010 		db_printf("DEF");
1011 	if (m->lock_object.lo_flags & LO_RECURSABLE)
1012 		db_printf(", RECURSE");
1013 	if (m->lock_object.lo_flags & LO_DUPOK)
1014 		db_printf(", DUPOK");
1015 	db_printf("}\n");
1016 	db_printf(" state: {");
1017 	if (mtx_unowned(m))
1018 		db_printf("UNOWNED");
1019 	else if (mtx_destroyed(m))
1020 		db_printf("DESTROYED");
1021 	else {
1022 		db_printf("OWNED");
1023 		if (m->mtx_lock & MTX_CONTESTED)
1024 			db_printf(", CONTESTED");
1025 		if (m->mtx_lock & MTX_RECURSED)
1026 			db_printf(", RECURSED");
1027 	}
1028 	db_printf("}\n");
1029 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1030 		td = mtx_owner(m);
1031 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1032 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1033 		if (mtx_recursed(m))
1034 			db_printf(" recursed: %d\n", m->mtx_recurse);
1035 	}
1036 }
1037 #endif
1038