xref: /freebsd/sys/kern/kern_mutex.c (revision 35ae9291c2621d66ac66ed4a4996761946ac3e2d)
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_global.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/kdb.h>
50 #include <sys/kernel.h>
51 #include <sys/ktr.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/mutex.h>
55 #include <sys/proc.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sbuf.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 /*
80  * Internal utility macros.
81  */
82 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
83 
84 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
85 
86 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
87 
88 static void	assert_mtx(struct lock_object *lock, int what);
89 #ifdef DDB
90 static void	db_show_mtx(struct lock_object *lock);
91 #endif
92 static void	lock_mtx(struct lock_object *lock, int how);
93 static void	lock_spin(struct lock_object *lock, int how);
94 #ifdef KDTRACE_HOOKS
95 static int	owner_mtx(struct lock_object *lock, struct thread **owner);
96 #endif
97 static int	unlock_mtx(struct lock_object *lock);
98 static int	unlock_spin(struct lock_object *lock);
99 
100 /*
101  * Lock classes for sleep and spin mutexes.
102  */
103 struct lock_class lock_class_mtx_sleep = {
104 	.lc_name = "sleep mutex",
105 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
106 	.lc_assert = assert_mtx,
107 #ifdef DDB
108 	.lc_ddb_show = db_show_mtx,
109 #endif
110 	.lc_lock = lock_mtx,
111 	.lc_unlock = unlock_mtx,
112 #ifdef KDTRACE_HOOKS
113 	.lc_owner = owner_mtx,
114 #endif
115 };
116 struct lock_class lock_class_mtx_spin = {
117 	.lc_name = "spin mutex",
118 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
119 	.lc_assert = assert_mtx,
120 #ifdef DDB
121 	.lc_ddb_show = db_show_mtx,
122 #endif
123 	.lc_lock = lock_spin,
124 	.lc_unlock = unlock_spin,
125 #ifdef KDTRACE_HOOKS
126 	.lc_owner = owner_mtx,
127 #endif
128 };
129 
130 /*
131  * System-wide mutexes
132  */
133 struct mtx blocked_lock;
134 struct mtx Giant;
135 
136 void
137 assert_mtx(struct lock_object *lock, int what)
138 {
139 
140 	mtx_assert((struct mtx *)lock, what);
141 }
142 
143 void
144 lock_mtx(struct lock_object *lock, int how)
145 {
146 
147 	mtx_lock((struct mtx *)lock);
148 }
149 
150 void
151 lock_spin(struct lock_object *lock, int how)
152 {
153 
154 	panic("spin locks can only use msleep_spin");
155 }
156 
157 int
158 unlock_mtx(struct lock_object *lock)
159 {
160 	struct mtx *m;
161 
162 	m = (struct mtx *)lock;
163 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
164 	mtx_unlock(m);
165 	return (0);
166 }
167 
168 int
169 unlock_spin(struct lock_object *lock)
170 {
171 
172 	panic("spin locks can only use msleep_spin");
173 }
174 
175 #ifdef KDTRACE_HOOKS
176 int
177 owner_mtx(struct lock_object *lock, struct thread **owner)
178 {
179 	struct mtx *m = (struct mtx *)lock;
180 
181 	*owner = mtx_owner(m);
182 	return (mtx_unowned(m) == 0);
183 }
184 #endif
185 
186 /*
187  * Function versions of the inlined __mtx_* macros.  These are used by
188  * modules and can also be called from assembly language if needed.
189  */
190 void
191 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
192 {
193 
194 	MPASS(curthread != NULL);
195 	KASSERT(m->mtx_lock != MTX_DESTROYED,
196 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
197 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
198 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
199 	    file, line));
200 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
201 	    file, line, NULL);
202 
203 	_get_sleep_lock(m, curthread, opts, file, line);
204 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
205 	    line);
206 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
207 	curthread->td_locks++;
208 }
209 
210 void
211 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
212 {
213 	MPASS(curthread != NULL);
214 	KASSERT(m->mtx_lock != MTX_DESTROYED,
215 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
216 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
217 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
218 	    file, line));
219 	curthread->td_locks--;
220 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
221 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
222 	    line);
223 	mtx_assert(m, MA_OWNED);
224 
225 	if (m->mtx_recurse == 0)
226 		LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m);
227 	_rel_sleep_lock(m, curthread, opts, file, line);
228 }
229 
230 void
231 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
232 {
233 
234 	MPASS(curthread != NULL);
235 	KASSERT(m->mtx_lock != MTX_DESTROYED,
236 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
237 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
238 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
239 	    m->lock_object.lo_name, file, line));
240 	if (mtx_owned(m))
241 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
242 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
243 		    m->lock_object.lo_name, file, line));
244 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
245 	    file, line, NULL);
246 	_get_spin_lock(m, curthread, opts, file, line);
247 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
248 	    line);
249 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
250 }
251 
252 void
253 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
254 {
255 
256 	MPASS(curthread != NULL);
257 	KASSERT(m->mtx_lock != MTX_DESTROYED,
258 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
259 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
260 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
261 	    m->lock_object.lo_name, file, line));
262 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
263 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
264 	    line);
265 	mtx_assert(m, MA_OWNED);
266 
267 	_rel_spin_lock(m);
268 }
269 
270 /*
271  * The important part of mtx_trylock{,_flags}()
272  * Tries to acquire lock `m.'  If this function is called on a mutex that
273  * is already owned, it will recursively acquire the lock.
274  */
275 int
276 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
277 {
278 #ifdef LOCK_PROFILING
279 	uint64_t waittime = 0;
280 	int contested = 0;
281 #endif
282 	int rval;
283 
284 	MPASS(curthread != NULL);
285 	KASSERT(m->mtx_lock != MTX_DESTROYED,
286 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
287 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
288 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
289 	    file, line));
290 
291 	if (mtx_owned(m) && (m->lock_object.lo_flags & LO_RECURSABLE) != 0) {
292 		m->mtx_recurse++;
293 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
294 		rval = 1;
295 	} else
296 		rval = _obtain_lock(m, (uintptr_t)curthread);
297 
298 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
299 	if (rval) {
300 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
301 		    file, line);
302 		curthread->td_locks++;
303 		if (m->mtx_recurse == 0)
304 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
305 			    m, contested, waittime, file, line);
306 
307 	}
308 
309 	return (rval);
310 }
311 
312 /*
313  * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
314  *
315  * We call this if the lock is either contested (i.e. we need to go to
316  * sleep waiting for it), or if we need to recurse on it.
317  */
318 void
319 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
320     int line)
321 {
322 	struct turnstile *ts;
323 	uintptr_t v;
324 #ifdef ADAPTIVE_MUTEXES
325 	volatile struct thread *owner;
326 #endif
327 #ifdef KTR
328 	int cont_logged = 0;
329 #endif
330 #ifdef LOCK_PROFILING
331 	int contested = 0;
332 	uint64_t waittime = 0;
333 #endif
334 #ifdef KDTRACE_HOOKS
335 	uint64_t spin_cnt = 0;
336 	uint64_t sleep_cnt = 0;
337 	int64_t sleep_time = 0;
338 #endif
339 
340 	if (mtx_owned(m)) {
341 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
342 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
343 		    m->lock_object.lo_name, file, line));
344 		m->mtx_recurse++;
345 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
346 		if (LOCK_LOG_TEST(&m->lock_object, opts))
347 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
348 		return;
349 	}
350 
351 	lock_profile_obtain_lock_failed(&m->lock_object,
352 		    &contested, &waittime);
353 	if (LOCK_LOG_TEST(&m->lock_object, opts))
354 		CTR4(KTR_LOCK,
355 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
356 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
357 
358 	while (!_obtain_lock(m, tid)) {
359 #ifdef KDTRACE_HOOKS
360 		spin_cnt++;
361 #endif
362 #ifdef ADAPTIVE_MUTEXES
363 		/*
364 		 * If the owner is running on another CPU, spin until the
365 		 * owner stops running or the state of the lock changes.
366 		 */
367 		v = m->mtx_lock;
368 		if (v != MTX_UNOWNED) {
369 			owner = (struct thread *)(v & ~MTX_FLAGMASK);
370 			if (TD_IS_RUNNING(owner)) {
371 				if (LOCK_LOG_TEST(&m->lock_object, 0))
372 					CTR3(KTR_LOCK,
373 					    "%s: spinning on %p held by %p",
374 					    __func__, m, owner);
375 				while (mtx_owner(m) == owner &&
376 				    TD_IS_RUNNING(owner)) {
377 					cpu_spinwait();
378 #ifdef KDTRACE_HOOKS
379 					spin_cnt++;
380 #endif
381 				}
382 				continue;
383 			}
384 		}
385 #endif
386 
387 		ts = turnstile_trywait(&m->lock_object);
388 		v = m->mtx_lock;
389 
390 		/*
391 		 * Check if the lock has been released while spinning for
392 		 * the turnstile chain lock.
393 		 */
394 		if (v == MTX_UNOWNED) {
395 			turnstile_cancel(ts);
396 			continue;
397 		}
398 
399 #ifdef ADAPTIVE_MUTEXES
400 		/*
401 		 * The current lock owner might have started executing
402 		 * on another CPU (or the lock could have changed
403 		 * owners) while we were waiting on the turnstile
404 		 * chain lock.  If so, drop the turnstile lock and try
405 		 * again.
406 		 */
407 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
408 		if (TD_IS_RUNNING(owner)) {
409 			turnstile_cancel(ts);
410 			continue;
411 		}
412 #endif
413 
414 		/*
415 		 * If the mutex isn't already contested and a failure occurs
416 		 * setting the contested bit, the mutex was either released
417 		 * or the state of the MTX_RECURSED bit changed.
418 		 */
419 		if ((v & MTX_CONTESTED) == 0 &&
420 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
421 			turnstile_cancel(ts);
422 			continue;
423 		}
424 
425 		/*
426 		 * We definitely must sleep for this lock.
427 		 */
428 		mtx_assert(m, MA_NOTOWNED);
429 
430 #ifdef KTR
431 		if (!cont_logged) {
432 			CTR6(KTR_CONTENTION,
433 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
434 			    (void *)tid, file, line, m->lock_object.lo_name,
435 			    WITNESS_FILE(&m->lock_object),
436 			    WITNESS_LINE(&m->lock_object));
437 			cont_logged = 1;
438 		}
439 #endif
440 
441 		/*
442 		 * Block on the turnstile.
443 		 */
444 #ifdef KDTRACE_HOOKS
445 		sleep_time -= lockstat_nsecs();
446 #endif
447 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
448 #ifdef KDTRACE_HOOKS
449 		sleep_time += lockstat_nsecs();
450 		sleep_cnt++;
451 #endif
452 	}
453 #ifdef KTR
454 	if (cont_logged) {
455 		CTR4(KTR_CONTENTION,
456 		    "contention end: %s acquired by %p at %s:%d",
457 		    m->lock_object.lo_name, (void *)tid, file, line);
458 	}
459 #endif
460 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
461 	    waittime, file, line);
462 #ifdef KDTRACE_HOOKS
463 	if (sleep_time)
464 		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
465 
466 	/*
467 	 * Only record the loops spinning and not sleeping.
468 	 */
469 	if (spin_cnt > sleep_cnt)
470 		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt));
471 #endif
472 }
473 
474 static void
475 _mtx_lock_spin_failed(struct mtx *m)
476 {
477 	struct thread *td;
478 
479 	td = mtx_owner(m);
480 
481 	/* If the mutex is unlocked, try again. */
482 	if (td == NULL)
483 		return;
484 
485 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
486 	    m, m->lock_object.lo_name, td, td->td_tid);
487 #ifdef WITNESS
488 	witness_display_spinlock(&m->lock_object, td);
489 #endif
490 	panic("spin lock held too long");
491 }
492 
493 #ifdef SMP
494 /*
495  * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
496  *
497  * This is only called if we need to actually spin for the lock. Recursion
498  * is handled inline.
499  */
500 void
501 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file,
502     int line)
503 {
504 	int i = 0;
505 #ifdef LOCK_PROFILING
506 	int contested = 0;
507 	uint64_t waittime = 0;
508 #endif
509 
510 	if (LOCK_LOG_TEST(&m->lock_object, opts))
511 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
512 
513 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
514 	while (!_obtain_lock(m, tid)) {
515 
516 		/* Give interrupts a chance while we spin. */
517 		spinlock_exit();
518 		while (m->mtx_lock != MTX_UNOWNED) {
519 			if (i++ < 10000000) {
520 				cpu_spinwait();
521 				continue;
522 			}
523 			if (i < 60000000 || kdb_active || panicstr != NULL)
524 				DELAY(1);
525 			else
526 				_mtx_lock_spin_failed(m);
527 			cpu_spinwait();
528 		}
529 		spinlock_enter();
530 	}
531 
532 	if (LOCK_LOG_TEST(&m->lock_object, opts))
533 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
534 
535 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
536 	    contested, waittime, (file), (line));
537 	LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, i);
538 }
539 #endif /* SMP */
540 
541 void
542 _thread_lock_flags(struct thread *td, int opts, const char *file, int line)
543 {
544 	struct mtx *m;
545 	uintptr_t tid;
546 	int i;
547 #ifdef LOCK_PROFILING
548 	int contested = 0;
549 	uint64_t waittime = 0;
550 #endif
551 #ifdef KDTRACE_HOOKS
552 	uint64_t spin_cnt = 0;
553 #endif
554 
555 	i = 0;
556 	tid = (uintptr_t)curthread;
557 	for (;;) {
558 retry:
559 		spinlock_enter();
560 		m = td->td_lock;
561 		KASSERT(m->mtx_lock != MTX_DESTROYED,
562 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
563 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
564 		    ("thread_lock() of sleep mutex %s @ %s:%d",
565 		    m->lock_object.lo_name, file, line));
566 		if (mtx_owned(m))
567 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
568 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
569 			    m->lock_object.lo_name, file, line));
570 		WITNESS_CHECKORDER(&m->lock_object,
571 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
572 		while (!_obtain_lock(m, tid)) {
573 #ifdef KDTRACE_HOOKS
574 			spin_cnt++;
575 #endif
576 			if (m->mtx_lock == tid) {
577 				m->mtx_recurse++;
578 				break;
579 			}
580 			lock_profile_obtain_lock_failed(&m->lock_object,
581 			    &contested, &waittime);
582 			/* Give interrupts a chance while we spin. */
583 			spinlock_exit();
584 			while (m->mtx_lock != MTX_UNOWNED) {
585 				if (i++ < 10000000)
586 					cpu_spinwait();
587 				else if (i < 60000000 ||
588 				    kdb_active || panicstr != NULL)
589 					DELAY(1);
590 				else
591 					_mtx_lock_spin_failed(m);
592 				cpu_spinwait();
593 				if (m != td->td_lock)
594 					goto retry;
595 			}
596 			spinlock_enter();
597 		}
598 		if (m == td->td_lock)
599 			break;
600 		_rel_spin_lock(m);	/* does spinlock_exit() */
601 #ifdef KDTRACE_HOOKS
602 		spin_cnt++;
603 #endif
604 	}
605 	if (m->mtx_recurse == 0)
606 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
607 		    m, contested, waittime, (file), (line));
608 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
609 	    line);
610 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
611 	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_cnt);
612 }
613 
614 struct mtx *
615 thread_lock_block(struct thread *td)
616 {
617 	struct mtx *lock;
618 
619 	spinlock_enter();
620 	THREAD_LOCK_ASSERT(td, MA_OWNED);
621 	lock = td->td_lock;
622 	td->td_lock = &blocked_lock;
623 	mtx_unlock_spin(lock);
624 
625 	return (lock);
626 }
627 
628 void
629 thread_lock_unblock(struct thread *td, struct mtx *new)
630 {
631 	mtx_assert(new, MA_OWNED);
632 	MPASS(td->td_lock == &blocked_lock);
633 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
634 	spinlock_exit();
635 }
636 
637 void
638 thread_lock_set(struct thread *td, struct mtx *new)
639 {
640 	struct mtx *lock;
641 
642 	mtx_assert(new, MA_OWNED);
643 	THREAD_LOCK_ASSERT(td, MA_OWNED);
644 	lock = td->td_lock;
645 	td->td_lock = new;
646 	mtx_unlock_spin(lock);
647 }
648 
649 /*
650  * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
651  *
652  * We are only called here if the lock is recursed or contested (i.e. we
653  * need to wake up a blocked thread).
654  */
655 void
656 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
657 {
658 	struct turnstile *ts;
659 
660 	if (mtx_recursed(m)) {
661 		if (--(m->mtx_recurse) == 0)
662 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
663 		if (LOCK_LOG_TEST(&m->lock_object, opts))
664 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
665 		return;
666 	}
667 
668 	/*
669 	 * We have to lock the chain before the turnstile so this turnstile
670 	 * can be removed from the hash list if it is empty.
671 	 */
672 	turnstile_chain_lock(&m->lock_object);
673 	ts = turnstile_lookup(&m->lock_object);
674 	if (LOCK_LOG_TEST(&m->lock_object, opts))
675 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
676 	MPASS(ts != NULL);
677 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
678 	_release_lock_quick(m);
679 
680 	/*
681 	 * This turnstile is now no longer associated with the mutex.  We can
682 	 * unlock the chain lock so a new turnstile may take it's place.
683 	 */
684 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
685 	turnstile_chain_unlock(&m->lock_object);
686 }
687 
688 /*
689  * All the unlocking of MTX_SPIN locks is done inline.
690  * See the _rel_spin_lock() macro for the details.
691  */
692 
693 /*
694  * The backing function for the INVARIANTS-enabled mtx_assert()
695  */
696 #ifdef INVARIANT_SUPPORT
697 void
698 _mtx_assert(struct mtx *m, int what, const char *file, int line)
699 {
700 
701 	if (panicstr != NULL || dumping)
702 		return;
703 	switch (what) {
704 	case MA_OWNED:
705 	case MA_OWNED | MA_RECURSED:
706 	case MA_OWNED | MA_NOTRECURSED:
707 		if (!mtx_owned(m))
708 			panic("mutex %s not owned at %s:%d",
709 			    m->lock_object.lo_name, file, line);
710 		if (mtx_recursed(m)) {
711 			if ((what & MA_NOTRECURSED) != 0)
712 				panic("mutex %s recursed at %s:%d",
713 				    m->lock_object.lo_name, file, line);
714 		} else if ((what & MA_RECURSED) != 0) {
715 			panic("mutex %s unrecursed at %s:%d",
716 			    m->lock_object.lo_name, file, line);
717 		}
718 		break;
719 	case MA_NOTOWNED:
720 		if (mtx_owned(m))
721 			panic("mutex %s owned at %s:%d",
722 			    m->lock_object.lo_name, file, line);
723 		break;
724 	default:
725 		panic("unknown mtx_assert at %s:%d", file, line);
726 	}
727 }
728 #endif
729 
730 /*
731  * The MUTEX_DEBUG-enabled mtx_validate()
732  *
733  * Most of these checks have been moved off into the LO_INITIALIZED flag
734  * maintained by the witness code.
735  */
736 #ifdef MUTEX_DEBUG
737 
738 void	mtx_validate(struct mtx *);
739 
740 void
741 mtx_validate(struct mtx *m)
742 {
743 
744 /*
745  * XXX: When kernacc() does not require Giant we can reenable this check
746  */
747 #ifdef notyet
748 	/*
749 	 * Can't call kernacc() from early init386(), especially when
750 	 * initializing Giant mutex, because some stuff in kernacc()
751 	 * requires Giant itself.
752 	 */
753 	if (!cold)
754 		if (!kernacc((caddr_t)m, sizeof(m),
755 		    VM_PROT_READ | VM_PROT_WRITE))
756 			panic("Can't read and write to mutex %p", m);
757 #endif
758 }
759 #endif
760 
761 /*
762  * General init routine used by the MTX_SYSINIT() macro.
763  */
764 void
765 mtx_sysinit(void *arg)
766 {
767 	struct mtx_args *margs = arg;
768 
769 	mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
770 }
771 
772 /*
773  * Mutex initialization routine; initialize lock `m' of type contained in
774  * `opts' with options contained in `opts' and name `name.'  The optional
775  * lock type `type' is used as a general lock category name for use with
776  * witness.
777  */
778 void
779 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
780 {
781 	struct lock_class *class;
782 	int flags;
783 
784 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
785 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
786 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
787 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
788 	    &m->mtx_lock));
789 
790 #ifdef MUTEX_DEBUG
791 	/* Diagnostic and error correction */
792 	mtx_validate(m);
793 #endif
794 
795 	/* Determine lock class and lock flags. */
796 	if (opts & MTX_SPIN)
797 		class = &lock_class_mtx_spin;
798 	else
799 		class = &lock_class_mtx_sleep;
800 	flags = 0;
801 	if (opts & MTX_QUIET)
802 		flags |= LO_QUIET;
803 	if (opts & MTX_RECURSE)
804 		flags |= LO_RECURSABLE;
805 	if ((opts & MTX_NOWITNESS) == 0)
806 		flags |= LO_WITNESS;
807 	if (opts & MTX_DUPOK)
808 		flags |= LO_DUPOK;
809 	if (opts & MTX_NOPROFILE)
810 		flags |= LO_NOPROFILE;
811 
812 	/* Initialize mutex. */
813 	m->mtx_lock = MTX_UNOWNED;
814 	m->mtx_recurse = 0;
815 
816 	lock_init(&m->lock_object, class, name, type, flags);
817 }
818 
819 /*
820  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
821  * passed in as a flag here because if the corresponding mtx_init() was
822  * called with MTX_QUIET set, then it will already be set in the mutex's
823  * flags.
824  */
825 void
826 mtx_destroy(struct mtx *m)
827 {
828 
829 	if (!mtx_owned(m))
830 		MPASS(mtx_unowned(m));
831 	else {
832 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
833 
834 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
835 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
836 			spinlock_exit();
837 		else
838 			curthread->td_locks--;
839 
840 		lock_profile_release_lock(&m->lock_object);
841 		/* Tell witness this isn't locked to make it happy. */
842 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
843 		    __LINE__);
844 	}
845 
846 	m->mtx_lock = MTX_DESTROYED;
847 	lock_destroy(&m->lock_object);
848 }
849 
850 /*
851  * Intialize the mutex code and system mutexes.  This is called from the MD
852  * startup code prior to mi_startup().  The per-CPU data space needs to be
853  * setup before this is called.
854  */
855 void
856 mutex_init(void)
857 {
858 
859 	/* Setup turnstiles so that sleep mutexes work. */
860 	init_turnstiles();
861 
862 	/*
863 	 * Initialize mutexes.
864 	 */
865 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
866 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
867 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
868 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
869 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
870 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
871 	mtx_lock(&Giant);
872 }
873 
874 #ifdef DDB
875 void
876 db_show_mtx(struct lock_object *lock)
877 {
878 	struct thread *td;
879 	struct mtx *m;
880 
881 	m = (struct mtx *)lock;
882 
883 	db_printf(" flags: {");
884 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
885 		db_printf("SPIN");
886 	else
887 		db_printf("DEF");
888 	if (m->lock_object.lo_flags & LO_RECURSABLE)
889 		db_printf(", RECURSE");
890 	if (m->lock_object.lo_flags & LO_DUPOK)
891 		db_printf(", DUPOK");
892 	db_printf("}\n");
893 	db_printf(" state: {");
894 	if (mtx_unowned(m))
895 		db_printf("UNOWNED");
896 	else if (mtx_destroyed(m))
897 		db_printf("DESTROYED");
898 	else {
899 		db_printf("OWNED");
900 		if (m->mtx_lock & MTX_CONTESTED)
901 			db_printf(", CONTESTED");
902 		if (m->mtx_lock & MTX_RECURSED)
903 			db_printf(", RECURSED");
904 	}
905 	db_printf("}\n");
906 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
907 		td = mtx_owner(m);
908 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
909 		    td->td_tid, td->td_proc->p_pid, td->td_name);
910 		if (mtx_recursed(m))
911 			db_printf(" recursed: %d\n", m->mtx_recurse);
912 	}
913 }
914 #endif
915