xref: /freebsd/sys/kern/kern_mutex.c (revision 09d325677d53a12c79a43664ff29871e92247629)
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_hwpmc_hooks.h"
43 #include "opt_kdtrace.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/bus.h>
49 #include <sys/conf.h>
50 #include <sys/kdb.h>
51 #include <sys/kernel.h>
52 #include <sys/ktr.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mutex.h>
56 #include <sys/proc.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sched.h>
59 #include <sys/sbuf.h>
60 #include <sys/sysctl.h>
61 #include <sys/turnstile.h>
62 #include <sys/vmmeter.h>
63 #include <sys/lock_profile.h>
64 
65 #include <machine/atomic.h>
66 #include <machine/bus.h>
67 #include <machine/cpu.h>
68 
69 #include <ddb/ddb.h>
70 
71 #include <fs/devfs/devfs_int.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_extern.h>
75 
76 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
77 #define	ADAPTIVE_MUTEXES
78 #endif
79 
80 #ifdef HWPMC_HOOKS
81 #include <sys/pmckern.h>
82 PMC_SOFT_DEFINE( , , lock, failed);
83 #endif
84 
85 /*
86  * Return the mutex address when the lock cookie address is provided.
87  * This functionality assumes that struct mtx* have a member named mtx_lock.
88  */
89 #define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
90 
91 /*
92  * Internal utility macros.
93  */
94 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
95 
96 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
97 
98 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
99 
100 static void	assert_mtx(const struct lock_object *lock, int what);
101 #ifdef DDB
102 static void	db_show_mtx(const struct lock_object *lock);
103 #endif
104 static void	lock_mtx(struct lock_object *lock, uintptr_t how);
105 static void	lock_spin(struct lock_object *lock, uintptr_t how);
106 #ifdef KDTRACE_HOOKS
107 static int	owner_mtx(const struct lock_object *lock,
108 		    struct thread **owner);
109 #endif
110 static uintptr_t unlock_mtx(struct lock_object *lock);
111 static uintptr_t unlock_spin(struct lock_object *lock);
112 
113 /*
114  * Lock classes for sleep and spin mutexes.
115  */
116 struct lock_class lock_class_mtx_sleep = {
117 	.lc_name = "sleep mutex",
118 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
119 	.lc_assert = assert_mtx,
120 #ifdef DDB
121 	.lc_ddb_show = db_show_mtx,
122 #endif
123 	.lc_lock = lock_mtx,
124 	.lc_unlock = unlock_mtx,
125 #ifdef KDTRACE_HOOKS
126 	.lc_owner = owner_mtx,
127 #endif
128 };
129 struct lock_class lock_class_mtx_spin = {
130 	.lc_name = "spin mutex",
131 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
132 	.lc_assert = assert_mtx,
133 #ifdef DDB
134 	.lc_ddb_show = db_show_mtx,
135 #endif
136 	.lc_lock = lock_spin,
137 	.lc_unlock = unlock_spin,
138 #ifdef KDTRACE_HOOKS
139 	.lc_owner = owner_mtx,
140 #endif
141 };
142 
143 /*
144  * System-wide mutexes
145  */
146 struct mtx blocked_lock;
147 struct mtx Giant;
148 
149 void
150 assert_mtx(const struct lock_object *lock, int what)
151 {
152 
153 	mtx_assert((const struct mtx *)lock, what);
154 }
155 
156 void
157 lock_mtx(struct lock_object *lock, uintptr_t how)
158 {
159 
160 	mtx_lock((struct mtx *)lock);
161 }
162 
163 void
164 lock_spin(struct lock_object *lock, uintptr_t how)
165 {
166 
167 	panic("spin locks can only use msleep_spin");
168 }
169 
170 uintptr_t
171 unlock_mtx(struct lock_object *lock)
172 {
173 	struct mtx *m;
174 
175 	m = (struct mtx *)lock;
176 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
177 	mtx_unlock(m);
178 	return (0);
179 }
180 
181 uintptr_t
182 unlock_spin(struct lock_object *lock)
183 {
184 
185 	panic("spin locks can only use msleep_spin");
186 }
187 
188 #ifdef KDTRACE_HOOKS
189 int
190 owner_mtx(const struct lock_object *lock, struct thread **owner)
191 {
192 	const struct mtx *m = (const struct mtx *)lock;
193 
194 	*owner = mtx_owner(m);
195 	return (mtx_unowned(m) == 0);
196 }
197 #endif
198 
199 /*
200  * Function versions of the inlined __mtx_* macros.  These are used by
201  * modules and can also be called from assembly language if needed.
202  */
203 void
204 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
205 {
206 	struct mtx *m;
207 
208 	if (SCHEDULER_STOPPED())
209 		return;
210 
211 	m = mtxlock2mtx(c);
212 
213 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
214 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
215 	    curthread, m->lock_object.lo_name, file, line));
216 	KASSERT(m->mtx_lock != MTX_DESTROYED,
217 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
218 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
219 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
220 	    file, line));
221 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
222 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
223 
224 	__mtx_lock(m, curthread, opts, file, line);
225 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
226 	    line);
227 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
228 	    file, line);
229 	curthread->td_locks++;
230 }
231 
232 void
233 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
234 {
235 	struct mtx *m;
236 
237 	if (SCHEDULER_STOPPED())
238 		return;
239 
240 	m = mtxlock2mtx(c);
241 
242 	KASSERT(m->mtx_lock != MTX_DESTROYED,
243 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
244 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
245 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
246 	    file, line));
247 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
248 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
249 	    line);
250 	mtx_assert(m, MA_OWNED);
251 
252 	if (m->mtx_recurse == 0)
253 		LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m);
254 	__mtx_unlock(m, curthread, opts, file, line);
255 	curthread->td_locks--;
256 }
257 
258 void
259 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
260     int line)
261 {
262 	struct mtx *m;
263 
264 	if (SCHEDULER_STOPPED())
265 		return;
266 
267 	m = mtxlock2mtx(c);
268 
269 	KASSERT(m->mtx_lock != MTX_DESTROYED,
270 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
271 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
272 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
273 	    m->lock_object.lo_name, file, line));
274 	if (mtx_owned(m))
275 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
276 		    (opts & MTX_RECURSE) != 0,
277 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
278 		    m->lock_object.lo_name, file, line));
279 	opts &= ~MTX_RECURSE;
280 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
281 	    file, line, NULL);
282 	__mtx_lock_spin(m, curthread, opts, file, line);
283 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
284 	    line);
285 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
286 }
287 
288 void
289 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
290     int line)
291 {
292 	struct mtx *m;
293 
294 	if (SCHEDULER_STOPPED())
295 		return;
296 
297 	m = mtxlock2mtx(c);
298 
299 	KASSERT(m->mtx_lock != MTX_DESTROYED,
300 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
301 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
302 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
303 	    m->lock_object.lo_name, file, line));
304 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
305 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
306 	    line);
307 	mtx_assert(m, MA_OWNED);
308 
309 	__mtx_unlock_spin(m);
310 }
311 
312 /*
313  * The important part of mtx_trylock{,_flags}()
314  * Tries to acquire lock `m.'  If this function is called on a mutex that
315  * is already owned, it will recursively acquire the lock.
316  */
317 int
318 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
319 {
320 	struct mtx *m;
321 #ifdef LOCK_PROFILING
322 	uint64_t waittime = 0;
323 	int contested = 0;
324 #endif
325 	int rval;
326 
327 	if (SCHEDULER_STOPPED())
328 		return (1);
329 
330 	m = mtxlock2mtx(c);
331 
332 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
333 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
334 	    curthread, m->lock_object.lo_name, file, line));
335 	KASSERT(m->mtx_lock != MTX_DESTROYED,
336 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
337 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
338 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
339 	    file, line));
340 
341 	if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
342 	    (opts & MTX_RECURSE) != 0)) {
343 		m->mtx_recurse++;
344 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
345 		rval = 1;
346 	} else
347 		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
348 	opts &= ~MTX_RECURSE;
349 
350 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
351 	if (rval) {
352 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
353 		    file, line);
354 		curthread->td_locks++;
355 		if (m->mtx_recurse == 0)
356 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
357 			    m, contested, waittime, file, line);
358 
359 	}
360 
361 	return (rval);
362 }
363 
364 /*
365  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
366  *
367  * We call this if the lock is either contested (i.e. we need to go to
368  * sleep waiting for it), or if we need to recurse on it.
369  */
370 void
371 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
372     const char *file, int line)
373 {
374 	struct mtx *m;
375 	struct turnstile *ts;
376 	uintptr_t v;
377 #ifdef ADAPTIVE_MUTEXES
378 	volatile struct thread *owner;
379 #endif
380 #ifdef KTR
381 	int cont_logged = 0;
382 #endif
383 #ifdef LOCK_PROFILING
384 	int contested = 0;
385 	uint64_t waittime = 0;
386 #endif
387 #ifdef KDTRACE_HOOKS
388 	uint64_t spin_cnt = 0;
389 	uint64_t sleep_cnt = 0;
390 	int64_t sleep_time = 0;
391 #endif
392 
393 	if (SCHEDULER_STOPPED())
394 		return;
395 
396 	m = mtxlock2mtx(c);
397 
398 	if (mtx_owned(m)) {
399 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
400 		    (opts & MTX_RECURSE) != 0,
401 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
402 		    m->lock_object.lo_name, file, line));
403 		opts &= ~MTX_RECURSE;
404 		m->mtx_recurse++;
405 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
406 		if (LOCK_LOG_TEST(&m->lock_object, opts))
407 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
408 		return;
409 	}
410 	opts &= ~MTX_RECURSE;
411 
412 #ifdef HWPMC_HOOKS
413 	PMC_SOFT_CALL( , , lock, failed);
414 #endif
415 	lock_profile_obtain_lock_failed(&m->lock_object,
416 		    &contested, &waittime);
417 	if (LOCK_LOG_TEST(&m->lock_object, opts))
418 		CTR4(KTR_LOCK,
419 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
420 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
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 				while (mtx_owner(m) == owner &&
440 				    TD_IS_RUNNING(owner)) {
441 					cpu_spinwait();
442 #ifdef KDTRACE_HOOKS
443 					spin_cnt++;
444 #endif
445 				}
446 				continue;
447 			}
448 		}
449 #endif
450 
451 		ts = turnstile_trywait(&m->lock_object);
452 		v = m->mtx_lock;
453 
454 		/*
455 		 * Check if the lock has been released while spinning for
456 		 * the turnstile chain lock.
457 		 */
458 		if (v == MTX_UNOWNED) {
459 			turnstile_cancel(ts);
460 			continue;
461 		}
462 
463 #ifdef ADAPTIVE_MUTEXES
464 		/*
465 		 * The current lock owner might have started executing
466 		 * on another CPU (or the lock could have changed
467 		 * owners) while we were waiting on the turnstile
468 		 * chain lock.  If so, drop the turnstile lock and try
469 		 * again.
470 		 */
471 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
472 		if (TD_IS_RUNNING(owner)) {
473 			turnstile_cancel(ts);
474 			continue;
475 		}
476 #endif
477 
478 		/*
479 		 * If the mutex isn't already contested and a failure occurs
480 		 * setting the contested bit, the mutex was either released
481 		 * or the state of the MTX_RECURSED bit changed.
482 		 */
483 		if ((v & MTX_CONTESTED) == 0 &&
484 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
485 			turnstile_cancel(ts);
486 			continue;
487 		}
488 
489 		/*
490 		 * We definitely must sleep for this lock.
491 		 */
492 		mtx_assert(m, MA_NOTOWNED);
493 
494 #ifdef KTR
495 		if (!cont_logged) {
496 			CTR6(KTR_CONTENTION,
497 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
498 			    (void *)tid, file, line, m->lock_object.lo_name,
499 			    WITNESS_FILE(&m->lock_object),
500 			    WITNESS_LINE(&m->lock_object));
501 			cont_logged = 1;
502 		}
503 #endif
504 
505 		/*
506 		 * Block on the turnstile.
507 		 */
508 #ifdef KDTRACE_HOOKS
509 		sleep_time -= lockstat_nsecs();
510 #endif
511 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
512 #ifdef KDTRACE_HOOKS
513 		sleep_time += lockstat_nsecs();
514 		sleep_cnt++;
515 #endif
516 	}
517 #ifdef KTR
518 	if (cont_logged) {
519 		CTR4(KTR_CONTENTION,
520 		    "contention end: %s acquired by %p at %s:%d",
521 		    m->lock_object.lo_name, (void *)tid, file, line);
522 	}
523 #endif
524 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
525 	    waittime, file, line);
526 #ifdef KDTRACE_HOOKS
527 	if (sleep_time)
528 		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
529 
530 	/*
531 	 * Only record the loops spinning and not sleeping.
532 	 */
533 	if (spin_cnt > sleep_cnt)
534 		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt));
535 #endif
536 }
537 
538 static void
539 _mtx_lock_spin_failed(struct mtx *m)
540 {
541 	struct thread *td;
542 
543 	td = mtx_owner(m);
544 
545 	/* If the mutex is unlocked, try again. */
546 	if (td == NULL)
547 		return;
548 
549 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
550 	    m, m->lock_object.lo_name, td, td->td_tid);
551 #ifdef WITNESS
552 	witness_display_spinlock(&m->lock_object, td, printf);
553 #endif
554 	panic("spin lock held too long");
555 }
556 
557 #ifdef SMP
558 /*
559  * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
560  *
561  * This is only called if we need to actually spin for the lock. Recursion
562  * is handled inline.
563  */
564 void
565 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
566     const char *file, int line)
567 {
568 	struct mtx *m;
569 	int i = 0;
570 #ifdef LOCK_PROFILING
571 	int contested = 0;
572 	uint64_t waittime = 0;
573 #endif
574 
575 	if (SCHEDULER_STOPPED())
576 		return;
577 
578 	m = mtxlock2mtx(c);
579 
580 	if (LOCK_LOG_TEST(&m->lock_object, opts))
581 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
582 
583 #ifdef HWPMC_HOOKS
584 	PMC_SOFT_CALL( , , lock, failed);
585 #endif
586 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
587 	while (!_mtx_obtain_lock(m, tid)) {
588 
589 		/* Give interrupts a chance while we spin. */
590 		spinlock_exit();
591 		while (m->mtx_lock != MTX_UNOWNED) {
592 			if (i++ < 10000000) {
593 				cpu_spinwait();
594 				continue;
595 			}
596 			if (i < 60000000 || kdb_active || panicstr != NULL)
597 				DELAY(1);
598 			else
599 				_mtx_lock_spin_failed(m);
600 			cpu_spinwait();
601 		}
602 		spinlock_enter();
603 	}
604 
605 	if (LOCK_LOG_TEST(&m->lock_object, opts))
606 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
607 
608 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
609 	    contested, waittime, (file), (line));
610 	LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, i);
611 }
612 #endif /* SMP */
613 
614 void
615 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
616 {
617 	struct mtx *m;
618 	uintptr_t tid;
619 	int i;
620 #ifdef LOCK_PROFILING
621 	int contested = 0;
622 	uint64_t waittime = 0;
623 #endif
624 #ifdef KDTRACE_HOOKS
625 	uint64_t spin_cnt = 0;
626 #endif
627 
628 	i = 0;
629 	tid = (uintptr_t)curthread;
630 
631 	if (SCHEDULER_STOPPED())
632 		return;
633 
634 	for (;;) {
635 retry:
636 		spinlock_enter();
637 		m = td->td_lock;
638 		KASSERT(m->mtx_lock != MTX_DESTROYED,
639 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
640 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
641 		    ("thread_lock() of sleep mutex %s @ %s:%d",
642 		    m->lock_object.lo_name, file, line));
643 		if (mtx_owned(m))
644 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
645 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
646 			    m->lock_object.lo_name, file, line));
647 		WITNESS_CHECKORDER(&m->lock_object,
648 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
649 		while (!_mtx_obtain_lock(m, tid)) {
650 #ifdef KDTRACE_HOOKS
651 			spin_cnt++;
652 #endif
653 			if (m->mtx_lock == tid) {
654 				m->mtx_recurse++;
655 				break;
656 			}
657 #ifdef HWPMC_HOOKS
658 			PMC_SOFT_CALL( , , lock, failed);
659 #endif
660 			lock_profile_obtain_lock_failed(&m->lock_object,
661 			    &contested, &waittime);
662 			/* Give interrupts a chance while we spin. */
663 			spinlock_exit();
664 			while (m->mtx_lock != MTX_UNOWNED) {
665 				if (i++ < 10000000)
666 					cpu_spinwait();
667 				else if (i < 60000000 ||
668 				    kdb_active || panicstr != NULL)
669 					DELAY(1);
670 				else
671 					_mtx_lock_spin_failed(m);
672 				cpu_spinwait();
673 				if (m != td->td_lock)
674 					goto retry;
675 			}
676 			spinlock_enter();
677 		}
678 		if (m == td->td_lock)
679 			break;
680 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
681 #ifdef KDTRACE_HOOKS
682 		spin_cnt++;
683 #endif
684 	}
685 	if (m->mtx_recurse == 0)
686 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
687 		    m, contested, waittime, (file), (line));
688 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
689 	    line);
690 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
691 	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_cnt);
692 }
693 
694 struct mtx *
695 thread_lock_block(struct thread *td)
696 {
697 	struct mtx *lock;
698 
699 	THREAD_LOCK_ASSERT(td, MA_OWNED);
700 	lock = td->td_lock;
701 	td->td_lock = &blocked_lock;
702 	mtx_unlock_spin(lock);
703 
704 	return (lock);
705 }
706 
707 void
708 thread_lock_unblock(struct thread *td, struct mtx *new)
709 {
710 	mtx_assert(new, MA_OWNED);
711 	MPASS(td->td_lock == &blocked_lock);
712 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
713 }
714 
715 void
716 thread_lock_set(struct thread *td, struct mtx *new)
717 {
718 	struct mtx *lock;
719 
720 	mtx_assert(new, MA_OWNED);
721 	THREAD_LOCK_ASSERT(td, MA_OWNED);
722 	lock = td->td_lock;
723 	td->td_lock = new;
724 	mtx_unlock_spin(lock);
725 }
726 
727 /*
728  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
729  *
730  * We are only called here if the lock is recursed or contested (i.e. we
731  * need to wake up a blocked thread).
732  */
733 void
734 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
735 {
736 	struct mtx *m;
737 	struct turnstile *ts;
738 
739 	if (SCHEDULER_STOPPED())
740 		return;
741 
742 	m = mtxlock2mtx(c);
743 
744 	if (mtx_recursed(m)) {
745 		if (--(m->mtx_recurse) == 0)
746 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
747 		if (LOCK_LOG_TEST(&m->lock_object, opts))
748 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
749 		return;
750 	}
751 
752 	/*
753 	 * We have to lock the chain before the turnstile so this turnstile
754 	 * can be removed from the hash list if it is empty.
755 	 */
756 	turnstile_chain_lock(&m->lock_object);
757 	ts = turnstile_lookup(&m->lock_object);
758 	if (LOCK_LOG_TEST(&m->lock_object, opts))
759 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
760 	MPASS(ts != NULL);
761 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
762 	_mtx_release_lock_quick(m);
763 
764 	/*
765 	 * This turnstile is now no longer associated with the mutex.  We can
766 	 * unlock the chain lock so a new turnstile may take it's place.
767 	 */
768 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
769 	turnstile_chain_unlock(&m->lock_object);
770 }
771 
772 /*
773  * All the unlocking of MTX_SPIN locks is done inline.
774  * See the __mtx_unlock_spin() macro for the details.
775  */
776 
777 /*
778  * The backing function for the INVARIANTS-enabled mtx_assert()
779  */
780 #ifdef INVARIANT_SUPPORT
781 void
782 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
783 {
784 	const struct mtx *m;
785 
786 	if (panicstr != NULL || dumping)
787 		return;
788 
789 	m = mtxlock2mtx(c);
790 
791 	switch (what) {
792 	case MA_OWNED:
793 	case MA_OWNED | MA_RECURSED:
794 	case MA_OWNED | MA_NOTRECURSED:
795 		if (!mtx_owned(m))
796 			panic("mutex %s not owned at %s:%d",
797 			    m->lock_object.lo_name, file, line);
798 		if (mtx_recursed(m)) {
799 			if ((what & MA_NOTRECURSED) != 0)
800 				panic("mutex %s recursed at %s:%d",
801 				    m->lock_object.lo_name, file, line);
802 		} else if ((what & MA_RECURSED) != 0) {
803 			panic("mutex %s unrecursed at %s:%d",
804 			    m->lock_object.lo_name, file, line);
805 		}
806 		break;
807 	case MA_NOTOWNED:
808 		if (mtx_owned(m))
809 			panic("mutex %s owned at %s:%d",
810 			    m->lock_object.lo_name, file, line);
811 		break;
812 	default:
813 		panic("unknown mtx_assert at %s:%d", file, line);
814 	}
815 }
816 #endif
817 
818 /*
819  * The MUTEX_DEBUG-enabled mtx_validate()
820  *
821  * Most of these checks have been moved off into the LO_INITIALIZED flag
822  * maintained by the witness code.
823  */
824 #ifdef MUTEX_DEBUG
825 
826 void	mtx_validate(struct mtx *);
827 
828 void
829 mtx_validate(struct mtx *m)
830 {
831 
832 /*
833  * XXX: When kernacc() does not require Giant we can reenable this check
834  */
835 #ifdef notyet
836 	/*
837 	 * Can't call kernacc() from early init386(), especially when
838 	 * initializing Giant mutex, because some stuff in kernacc()
839 	 * requires Giant itself.
840 	 */
841 	if (!cold)
842 		if (!kernacc((caddr_t)m, sizeof(m),
843 		    VM_PROT_READ | VM_PROT_WRITE))
844 			panic("Can't read and write to mutex %p", m);
845 #endif
846 }
847 #endif
848 
849 /*
850  * General init routine used by the MTX_SYSINIT() macro.
851  */
852 void
853 mtx_sysinit(void *arg)
854 {
855 	struct mtx_args *margs = arg;
856 
857 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
858 	    margs->ma_opts);
859 }
860 
861 /*
862  * Mutex initialization routine; initialize lock `m' of type contained in
863  * `opts' with options contained in `opts' and name `name.'  The optional
864  * lock type `type' is used as a general lock category name for use with
865  * witness.
866  */
867 void
868 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
869 {
870 	struct mtx *m;
871 	struct lock_class *class;
872 	int flags;
873 
874 	m = mtxlock2mtx(c);
875 
876 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
877 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
878 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
879 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
880 	    &m->mtx_lock));
881 
882 #ifdef MUTEX_DEBUG
883 	/* Diagnostic and error correction */
884 	mtx_validate(m);
885 #endif
886 
887 	/* Determine lock class and lock flags. */
888 	if (opts & MTX_SPIN)
889 		class = &lock_class_mtx_spin;
890 	else
891 		class = &lock_class_mtx_sleep;
892 	flags = 0;
893 	if (opts & MTX_QUIET)
894 		flags |= LO_QUIET;
895 	if (opts & MTX_RECURSE)
896 		flags |= LO_RECURSABLE;
897 	if ((opts & MTX_NOWITNESS) == 0)
898 		flags |= LO_WITNESS;
899 	if (opts & MTX_DUPOK)
900 		flags |= LO_DUPOK;
901 	if (opts & MTX_NOPROFILE)
902 		flags |= LO_NOPROFILE;
903 
904 	/* Initialize mutex. */
905 	lock_init(&m->lock_object, class, name, type, flags);
906 
907 	m->mtx_lock = MTX_UNOWNED;
908 	m->mtx_recurse = 0;
909 }
910 
911 /*
912  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
913  * passed in as a flag here because if the corresponding mtx_init() was
914  * called with MTX_QUIET set, then it will already be set in the mutex's
915  * flags.
916  */
917 void
918 _mtx_destroy(volatile uintptr_t *c)
919 {
920 	struct mtx *m;
921 
922 	m = mtxlock2mtx(c);
923 
924 	if (!mtx_owned(m))
925 		MPASS(mtx_unowned(m));
926 	else {
927 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
928 
929 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
930 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
931 			spinlock_exit();
932 		else
933 			curthread->td_locks--;
934 
935 		lock_profile_release_lock(&m->lock_object);
936 		/* Tell witness this isn't locked to make it happy. */
937 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
938 		    __LINE__);
939 	}
940 
941 	m->mtx_lock = MTX_DESTROYED;
942 	lock_destroy(&m->lock_object);
943 }
944 
945 /*
946  * Intialize the mutex code and system mutexes.  This is called from the MD
947  * startup code prior to mi_startup().  The per-CPU data space needs to be
948  * setup before this is called.
949  */
950 void
951 mutex_init(void)
952 {
953 
954 	/* Setup turnstiles so that sleep mutexes work. */
955 	init_turnstiles();
956 
957 	/*
958 	 * Initialize mutexes.
959 	 */
960 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
961 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
962 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
963 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
964 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
965 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
966 	mtx_lock(&Giant);
967 }
968 
969 #ifdef DDB
970 void
971 db_show_mtx(const struct lock_object *lock)
972 {
973 	struct thread *td;
974 	const struct mtx *m;
975 
976 	m = (const struct mtx *)lock;
977 
978 	db_printf(" flags: {");
979 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
980 		db_printf("SPIN");
981 	else
982 		db_printf("DEF");
983 	if (m->lock_object.lo_flags & LO_RECURSABLE)
984 		db_printf(", RECURSE");
985 	if (m->lock_object.lo_flags & LO_DUPOK)
986 		db_printf(", DUPOK");
987 	db_printf("}\n");
988 	db_printf(" state: {");
989 	if (mtx_unowned(m))
990 		db_printf("UNOWNED");
991 	else if (mtx_destroyed(m))
992 		db_printf("DESTROYED");
993 	else {
994 		db_printf("OWNED");
995 		if (m->mtx_lock & MTX_CONTESTED)
996 			db_printf(", CONTESTED");
997 		if (m->mtx_lock & MTX_RECURSED)
998 			db_printf(", RECURSED");
999 	}
1000 	db_printf("}\n");
1001 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1002 		td = mtx_owner(m);
1003 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1004 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1005 		if (mtx_recursed(m))
1006 			db_printf(" recursed: %d\n", m->mtx_recurse);
1007 	}
1008 }
1009 #endif
1010