xref: /freebsd/sys/kern/kern_mutex.c (revision 94cba8034ba53725c225c85e35724f0c2b13cea5)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. Berkeley Software Design Inc's name may not be used to endorse or
15  *    promote products derived from this software without specific prior
16  *    written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
31  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
32  */
33 
34 /*
35  * Machine independent bits of mutex implementation.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_adaptive_mutexes.h"
42 #include "opt_ddb.h"
43 #include "opt_hwpmc_hooks.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/smp.h>
61 #include <sys/sysctl.h>
62 #include <sys/turnstile.h>
63 #include <sys/vmmeter.h>
64 #include <sys/lock_profile.h>
65 
66 #include <machine/atomic.h>
67 #include <machine/bus.h>
68 #include <machine/cpu.h>
69 
70 #include <ddb/ddb.h>
71 
72 #include <fs/devfs/devfs_int.h>
73 
74 #include <vm/vm.h>
75 #include <vm/vm_extern.h>
76 
77 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
78 #define	ADAPTIVE_MUTEXES
79 #endif
80 
81 #ifdef HWPMC_HOOKS
82 #include <sys/pmckern.h>
83 PMC_SOFT_DEFINE( , , lock, failed);
84 #endif
85 
86 /*
87  * Return the mutex address when the lock cookie address is provided.
88  * This functionality assumes that struct mtx* have a member named mtx_lock.
89  */
90 #define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
91 
92 /*
93  * Internal utility macros.
94  */
95 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
96 
97 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
98 
99 static void	assert_mtx(const struct lock_object *lock, int what);
100 #ifdef DDB
101 static void	db_show_mtx(const struct lock_object *lock);
102 #endif
103 static void	lock_mtx(struct lock_object *lock, uintptr_t how);
104 static void	lock_spin(struct lock_object *lock, uintptr_t how);
105 #ifdef KDTRACE_HOOKS
106 static int	owner_mtx(const struct lock_object *lock,
107 		    struct thread **owner);
108 #endif
109 static uintptr_t unlock_mtx(struct lock_object *lock);
110 static uintptr_t unlock_spin(struct lock_object *lock);
111 
112 /*
113  * Lock classes for sleep and spin mutexes.
114  */
115 struct lock_class lock_class_mtx_sleep = {
116 	.lc_name = "sleep mutex",
117 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
118 	.lc_assert = assert_mtx,
119 #ifdef DDB
120 	.lc_ddb_show = db_show_mtx,
121 #endif
122 	.lc_lock = lock_mtx,
123 	.lc_unlock = unlock_mtx,
124 #ifdef KDTRACE_HOOKS
125 	.lc_owner = owner_mtx,
126 #endif
127 };
128 struct lock_class lock_class_mtx_spin = {
129 	.lc_name = "spin mutex",
130 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
131 	.lc_assert = assert_mtx,
132 #ifdef DDB
133 	.lc_ddb_show = db_show_mtx,
134 #endif
135 	.lc_lock = lock_spin,
136 	.lc_unlock = unlock_spin,
137 #ifdef KDTRACE_HOOKS
138 	.lc_owner = owner_mtx,
139 #endif
140 };
141 
142 #ifdef ADAPTIVE_MUTEXES
143 #ifdef MUTEX_CUSTOM_BACKOFF
144 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
145     "mtx debugging");
146 
147 static struct lock_delay_config __read_frequently mtx_delay;
148 
149 SYSCTL_U16(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base,
150     0, "");
151 SYSCTL_U16(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max,
152     0, "");
153 
154 LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay);
155 #else
156 #define mtx_delay	locks_delay
157 #endif
158 #endif
159 
160 #ifdef MUTEX_SPIN_CUSTOM_BACKOFF
161 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin,
162     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
163     "mtx spin debugging");
164 
165 static struct lock_delay_config __read_frequently mtx_spin_delay;
166 
167 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW,
168     &mtx_spin_delay.base, 0, "");
169 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW,
170     &mtx_spin_delay.max, 0, "");
171 
172 LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay);
173 #else
174 #define mtx_spin_delay	locks_delay
175 #endif
176 
177 /*
178  * System-wide mutexes
179  */
180 struct mtx blocked_lock;
181 struct mtx __exclusive_cache_line Giant;
182 
183 static void _mtx_lock_indefinite_check(struct mtx *, struct lock_delay_arg *);
184 
185 void
186 assert_mtx(const struct lock_object *lock, int what)
187 {
188 
189 	/*
190 	 * Treat LA_LOCKED as if LA_XLOCKED was asserted.
191 	 *
192 	 * Some callers of lc_assert uses LA_LOCKED to indicate that either
193 	 * a shared lock or write lock was held, while other callers uses
194 	 * the more strict LA_XLOCKED (used as MA_OWNED).
195 	 *
196 	 * Mutex is the only lock class that can not be shared, as a result,
197 	 * we can reasonably consider the caller really intends to assert
198 	 * LA_XLOCKED when they are asserting LA_LOCKED on a mutex object.
199 	 */
200 	if (what & LA_LOCKED) {
201 		what &= ~LA_LOCKED;
202 		what |= LA_XLOCKED;
203 	}
204 	mtx_assert((const struct mtx *)lock, what);
205 }
206 
207 void
208 lock_mtx(struct lock_object *lock, uintptr_t how)
209 {
210 
211 	mtx_lock((struct mtx *)lock);
212 }
213 
214 void
215 lock_spin(struct lock_object *lock, uintptr_t how)
216 {
217 
218 	panic("spin locks can only use msleep_spin");
219 }
220 
221 uintptr_t
222 unlock_mtx(struct lock_object *lock)
223 {
224 	struct mtx *m;
225 
226 	m = (struct mtx *)lock;
227 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
228 	mtx_unlock(m);
229 	return (0);
230 }
231 
232 uintptr_t
233 unlock_spin(struct lock_object *lock)
234 {
235 
236 	panic("spin locks can only use msleep_spin");
237 }
238 
239 #ifdef KDTRACE_HOOKS
240 int
241 owner_mtx(const struct lock_object *lock, struct thread **owner)
242 {
243 	const struct mtx *m;
244 	uintptr_t x;
245 
246 	m = (const struct mtx *)lock;
247 	x = m->mtx_lock;
248 	*owner = (struct thread *)(x & ~MTX_FLAGMASK);
249 	return (*owner != NULL);
250 }
251 #endif
252 
253 /*
254  * Function versions of the inlined __mtx_* macros.  These are used by
255  * modules and can also be called from assembly language if needed.
256  */
257 void
258 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
259 {
260 	struct mtx *m;
261 	uintptr_t tid, v;
262 
263 	m = mtxlock2mtx(c);
264 
265 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
266 	    !TD_IS_IDLETHREAD(curthread),
267 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
268 	    curthread, m->lock_object.lo_name, file, line));
269 	KASSERT(m->mtx_lock != MTX_DESTROYED,
270 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
271 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
272 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
273 	    file, line));
274 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
275 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
276 
277 	tid = (uintptr_t)curthread;
278 	v = MTX_UNOWNED;
279 	if (!_mtx_obtain_lock_fetch(m, &v, tid))
280 		_mtx_lock_sleep(m, v, opts, file, line);
281 	else
282 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
283 		    m, 0, 0, file, line);
284 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
285 	    line);
286 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
287 	    file, line);
288 	TD_LOCKS_INC(curthread);
289 }
290 
291 void
292 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
293 {
294 	struct mtx *m;
295 
296 	m = mtxlock2mtx(c);
297 
298 	KASSERT(m->mtx_lock != MTX_DESTROYED,
299 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
300 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
301 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
302 	    file, line));
303 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
304 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
305 	    line);
306 	mtx_assert(m, MA_OWNED);
307 
308 #ifdef LOCK_PROFILING
309 	__mtx_unlock_sleep(c, (uintptr_t)curthread, opts, file, line);
310 #else
311 	__mtx_unlock(m, curthread, opts, file, line);
312 #endif
313 	TD_LOCKS_DEC(curthread);
314 }
315 
316 void
317 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
318     int line)
319 {
320 	struct mtx *m;
321 #ifdef SMP
322 	uintptr_t tid, v;
323 #endif
324 
325 	m = mtxlock2mtx(c);
326 
327 	KASSERT(m->mtx_lock != MTX_DESTROYED,
328 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
329 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
330 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
331 	    m->lock_object.lo_name, file, line));
332 	if (mtx_owned(m))
333 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
334 		    (opts & MTX_RECURSE) != 0,
335 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
336 		    m->lock_object.lo_name, file, line));
337 	opts &= ~MTX_RECURSE;
338 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
339 	    file, line, NULL);
340 #ifdef SMP
341 	spinlock_enter();
342 	tid = (uintptr_t)curthread;
343 	v = MTX_UNOWNED;
344 	if (!_mtx_obtain_lock_fetch(m, &v, tid))
345 		_mtx_lock_spin(m, v, opts, file, line);
346 	else
347 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire,
348 		    m, 0, 0, file, line);
349 #else
350 	__mtx_lock_spin(m, curthread, opts, file, line);
351 #endif
352 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
353 	    line);
354 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
355 }
356 
357 int
358 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
359     int line)
360 {
361 	struct mtx *m;
362 
363 	if (SCHEDULER_STOPPED())
364 		return (1);
365 
366 	m = mtxlock2mtx(c);
367 
368 	KASSERT(m->mtx_lock != MTX_DESTROYED,
369 	    ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line));
370 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
371 	    ("mtx_trylock_spin() of sleep mutex %s @ %s:%d",
372 	    m->lock_object.lo_name, file, line));
373 	KASSERT((opts & MTX_RECURSE) == 0,
374 	    ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n",
375 	    m->lock_object.lo_name, file, line));
376 	if (__mtx_trylock_spin(m, curthread, opts, file, line)) {
377 		LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line);
378 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
379 		return (1);
380 	}
381 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line);
382 	return (0);
383 }
384 
385 void
386 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
387     int line)
388 {
389 	struct mtx *m;
390 
391 	m = mtxlock2mtx(c);
392 
393 	KASSERT(m->mtx_lock != MTX_DESTROYED,
394 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
395 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
396 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
397 	    m->lock_object.lo_name, file, line));
398 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
399 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
400 	    line);
401 	mtx_assert(m, MA_OWNED);
402 
403 	__mtx_unlock_spin(m);
404 }
405 
406 /*
407  * The important part of mtx_trylock{,_flags}()
408  * Tries to acquire lock `m.'  If this function is called on a mutex that
409  * is already owned, it will recursively acquire the lock.
410  */
411 int
412 _mtx_trylock_flags_int(struct mtx *m, int opts LOCK_FILE_LINE_ARG_DEF)
413 {
414 	struct thread *td;
415 	uintptr_t tid, v;
416 #ifdef LOCK_PROFILING
417 	uint64_t waittime = 0;
418 	int contested = 0;
419 #endif
420 	int rval;
421 	bool recursed;
422 
423 	td = curthread;
424 	tid = (uintptr_t)td;
425 	if (SCHEDULER_STOPPED_TD(td))
426 		return (1);
427 
428 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
429 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
430 	    curthread, m->lock_object.lo_name, file, line));
431 	KASSERT(m->mtx_lock != MTX_DESTROYED,
432 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
433 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
434 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
435 	    file, line));
436 
437 	rval = 1;
438 	recursed = false;
439 	v = MTX_UNOWNED;
440 	for (;;) {
441 		if (_mtx_obtain_lock_fetch(m, &v, tid))
442 			break;
443 		if (v == MTX_UNOWNED)
444 			continue;
445 		if (v == tid &&
446 		    ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
447 		    (opts & MTX_RECURSE) != 0)) {
448 			m->mtx_recurse++;
449 			atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
450 			recursed = true;
451 			break;
452 		}
453 		rval = 0;
454 		break;
455 	}
456 
457 	opts &= ~MTX_RECURSE;
458 
459 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
460 	if (rval) {
461 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
462 		    file, line);
463 		TD_LOCKS_INC(curthread);
464 		if (!recursed)
465 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
466 			    m, contested, waittime, file, line);
467 	}
468 
469 	return (rval);
470 }
471 
472 int
473 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
474 {
475 	struct mtx *m;
476 
477 	m = mtxlock2mtx(c);
478 	return (_mtx_trylock_flags_int(m, opts LOCK_FILE_LINE_ARG));
479 }
480 
481 /*
482  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
483  *
484  * We call this if the lock is either contested (i.e. we need to go to
485  * sleep waiting for it), or if we need to recurse on it.
486  */
487 #if LOCK_DEBUG > 0
488 void
489 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, int opts, const char *file,
490     int line)
491 #else
492 void
493 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v)
494 #endif
495 {
496 	struct thread *td;
497 	struct mtx *m;
498 	struct turnstile *ts;
499 	uintptr_t tid;
500 	struct thread *owner;
501 #ifdef LOCK_PROFILING
502 	int contested = 0;
503 	uint64_t waittime = 0;
504 #endif
505 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
506 	struct lock_delay_arg lda;
507 #endif
508 #ifdef KDTRACE_HOOKS
509 	u_int sleep_cnt = 0;
510 	int64_t sleep_time = 0;
511 	int64_t all_time = 0;
512 #endif
513 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
514 	int doing_lockprof = 0;
515 #endif
516 
517 	td = curthread;
518 	tid = (uintptr_t)td;
519 	m = mtxlock2mtx(c);
520 
521 #ifdef KDTRACE_HOOKS
522 	if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
523 		while (v == MTX_UNOWNED) {
524 			if (_mtx_obtain_lock_fetch(m, &v, tid))
525 				goto out_lockstat;
526 		}
527 		doing_lockprof = 1;
528 		all_time -= lockstat_nsecs(&m->lock_object);
529 	}
530 #endif
531 #ifdef LOCK_PROFILING
532 	doing_lockprof = 1;
533 #endif
534 
535 	if (SCHEDULER_STOPPED_TD(td))
536 		return;
537 
538 #if defined(ADAPTIVE_MUTEXES)
539 	lock_delay_arg_init(&lda, &mtx_delay);
540 #elif defined(KDTRACE_HOOKS)
541 	lock_delay_arg_init_noadapt(&lda);
542 #endif
543 
544 	if (__predict_false(v == MTX_UNOWNED))
545 		v = MTX_READ_VALUE(m);
546 
547 	if (__predict_false(lv_mtx_owner(v) == td)) {
548 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
549 		    (opts & MTX_RECURSE) != 0,
550 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
551 		    m->lock_object.lo_name, file, line));
552 #if LOCK_DEBUG > 0
553 		opts &= ~MTX_RECURSE;
554 #endif
555 		m->mtx_recurse++;
556 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
557 		if (LOCK_LOG_TEST(&m->lock_object, opts))
558 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
559 		return;
560 	}
561 #if LOCK_DEBUG > 0
562 	opts &= ~MTX_RECURSE;
563 #endif
564 
565 #ifdef HWPMC_HOOKS
566 	PMC_SOFT_CALL( , , lock, failed);
567 #endif
568 	lock_profile_obtain_lock_failed(&m->lock_object,
569 		    &contested, &waittime);
570 	if (LOCK_LOG_TEST(&m->lock_object, opts))
571 		CTR4(KTR_LOCK,
572 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
573 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
574 
575 	for (;;) {
576 		if (v == MTX_UNOWNED) {
577 			if (_mtx_obtain_lock_fetch(m, &v, tid))
578 				break;
579 			continue;
580 		}
581 #ifdef KDTRACE_HOOKS
582 		lda.spin_cnt++;
583 #endif
584 #ifdef ADAPTIVE_MUTEXES
585 		/*
586 		 * If the owner is running on another CPU, spin until the
587 		 * owner stops running or the state of the lock changes.
588 		 */
589 		owner = lv_mtx_owner(v);
590 		if (TD_IS_RUNNING(owner)) {
591 			if (LOCK_LOG_TEST(&m->lock_object, 0))
592 				CTR3(KTR_LOCK,
593 				    "%s: spinning on %p held by %p",
594 				    __func__, m, owner);
595 			KTR_STATE1(KTR_SCHED, "thread",
596 			    sched_tdname((struct thread *)tid),
597 			    "spinning", "lockname:\"%s\"",
598 			    m->lock_object.lo_name);
599 			do {
600 				lock_delay(&lda);
601 				v = MTX_READ_VALUE(m);
602 				owner = lv_mtx_owner(v);
603 			} while (v != MTX_UNOWNED && TD_IS_RUNNING(owner));
604 			KTR_STATE0(KTR_SCHED, "thread",
605 			    sched_tdname((struct thread *)tid),
606 			    "running");
607 			continue;
608 		}
609 #endif
610 
611 		ts = turnstile_trywait(&m->lock_object);
612 		v = MTX_READ_VALUE(m);
613 retry_turnstile:
614 
615 		/*
616 		 * Check if the lock has been released while spinning for
617 		 * the turnstile chain lock.
618 		 */
619 		if (v == MTX_UNOWNED) {
620 			turnstile_cancel(ts);
621 			continue;
622 		}
623 
624 #ifdef ADAPTIVE_MUTEXES
625 		/*
626 		 * The current lock owner might have started executing
627 		 * on another CPU (or the lock could have changed
628 		 * owners) while we were waiting on the turnstile
629 		 * chain lock.  If so, drop the turnstile lock and try
630 		 * again.
631 		 */
632 		owner = lv_mtx_owner(v);
633 		if (TD_IS_RUNNING(owner)) {
634 			turnstile_cancel(ts);
635 			continue;
636 		}
637 #endif
638 
639 		/*
640 		 * If the mutex isn't already contested and a failure occurs
641 		 * setting the contested bit, the mutex was either released
642 		 * or the state of the MTX_RECURSED bit changed.
643 		 */
644 		if ((v & MTX_CONTESTED) == 0 &&
645 		    !atomic_fcmpset_ptr(&m->mtx_lock, &v, v | MTX_CONTESTED)) {
646 			goto retry_turnstile;
647 		}
648 
649 		/*
650 		 * We definitely must sleep for this lock.
651 		 */
652 		mtx_assert(m, MA_NOTOWNED);
653 
654 		/*
655 		 * Block on the turnstile.
656 		 */
657 #ifdef KDTRACE_HOOKS
658 		sleep_time -= lockstat_nsecs(&m->lock_object);
659 #endif
660 #ifndef ADAPTIVE_MUTEXES
661 		owner = mtx_owner(m);
662 #endif
663 		MPASS(owner == mtx_owner(m));
664 		turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
665 #ifdef KDTRACE_HOOKS
666 		sleep_time += lockstat_nsecs(&m->lock_object);
667 		sleep_cnt++;
668 #endif
669 		v = MTX_READ_VALUE(m);
670 	}
671 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
672 	if (__predict_true(!doing_lockprof))
673 		return;
674 #endif
675 #ifdef KDTRACE_HOOKS
676 	all_time += lockstat_nsecs(&m->lock_object);
677 	if (sleep_time)
678 		LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
679 
680 	/*
681 	 * Only record the loops spinning and not sleeping.
682 	 */
683 	if (lda.spin_cnt > sleep_cnt)
684 		LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
685 out_lockstat:
686 #endif
687 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
688 	    waittime, file, line);
689 }
690 
691 #ifdef SMP
692 /*
693  * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
694  *
695  * This is only called if we need to actually spin for the lock. Recursion
696  * is handled inline.
697  */
698 #if LOCK_DEBUG > 0
699 void
700 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, int opts,
701     const char *file, int line)
702 #else
703 void
704 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v)
705 #endif
706 {
707 	struct mtx *m;
708 	struct lock_delay_arg lda;
709 	uintptr_t tid;
710 #ifdef LOCK_PROFILING
711 	int contested = 0;
712 	uint64_t waittime = 0;
713 #endif
714 #ifdef KDTRACE_HOOKS
715 	int64_t spin_time = 0;
716 #endif
717 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
718 	int doing_lockprof = 0;
719 #endif
720 
721 	tid = (uintptr_t)curthread;
722 	m = mtxlock2mtx(c);
723 
724 #ifdef KDTRACE_HOOKS
725 	if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
726 		while (v == MTX_UNOWNED) {
727 			if (_mtx_obtain_lock_fetch(m, &v, tid))
728 				goto out_lockstat;
729 		}
730 		doing_lockprof = 1;
731 		spin_time -= lockstat_nsecs(&m->lock_object);
732 	}
733 #endif
734 #ifdef LOCK_PROFILING
735 	doing_lockprof = 1;
736 #endif
737 
738 	if (__predict_false(v == MTX_UNOWNED))
739 		v = MTX_READ_VALUE(m);
740 
741 	if (__predict_false(v == tid)) {
742 		m->mtx_recurse++;
743 		return;
744 	}
745 
746 	if (SCHEDULER_STOPPED())
747 		return;
748 
749 	lock_delay_arg_init(&lda, &mtx_spin_delay);
750 
751 	if (LOCK_LOG_TEST(&m->lock_object, opts))
752 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
753 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
754 	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
755 
756 #ifdef HWPMC_HOOKS
757 	PMC_SOFT_CALL( , , lock, failed);
758 #endif
759 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
760 
761 	for (;;) {
762 		if (v == MTX_UNOWNED) {
763 			if (_mtx_obtain_lock_fetch(m, &v, tid))
764 				break;
765 			continue;
766 		}
767 		/* Give interrupts a chance while we spin. */
768 		spinlock_exit();
769 		do {
770 			if (__predict_true(lda.spin_cnt < 10000000)) {
771 				lock_delay(&lda);
772 			} else {
773 				_mtx_lock_indefinite_check(m, &lda);
774 			}
775 			v = MTX_READ_VALUE(m);
776 		} while (v != MTX_UNOWNED);
777 		spinlock_enter();
778 	}
779 
780 	if (LOCK_LOG_TEST(&m->lock_object, opts))
781 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
782 	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
783 	    "running");
784 
785 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
786 	if (__predict_true(!doing_lockprof))
787 		return;
788 #endif
789 #ifdef KDTRACE_HOOKS
790 	spin_time += lockstat_nsecs(&m->lock_object);
791 	if (lda.spin_cnt != 0)
792 		LOCKSTAT_RECORD1(spin__spin, m, spin_time);
793 out_lockstat:
794 #endif
795 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
796 	    contested, waittime, file, line);
797 }
798 #endif /* SMP */
799 
800 #ifdef INVARIANTS
801 static void
802 thread_lock_validate(struct mtx *m, int opts, const char *file, int line)
803 {
804 
805 	KASSERT(m->mtx_lock != MTX_DESTROYED,
806 	    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
807 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
808 	    ("thread_lock() of sleep mutex %s @ %s:%d",
809 	    m->lock_object.lo_name, file, line));
810 	KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) == 0,
811 	    ("thread_lock: got a recursive mutex %s @ %s:%d\n",
812 	    m->lock_object.lo_name, file, line));
813 	WITNESS_CHECKORDER(&m->lock_object,
814 	    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
815 }
816 #else
817 #define thread_lock_validate(m, opts, file, line) do { } while (0)
818 #endif
819 
820 #ifndef LOCK_PROFILING
821 #if LOCK_DEBUG > 0
822 void
823 _thread_lock(struct thread *td, int opts, const char *file, int line)
824 #else
825 void
826 _thread_lock(struct thread *td)
827 #endif
828 {
829 	struct mtx *m;
830 	uintptr_t tid;
831 
832 	tid = (uintptr_t)curthread;
833 
834 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(spin__acquire)))
835 		goto slowpath_noirq;
836 	spinlock_enter();
837 	m = td->td_lock;
838 	thread_lock_validate(m, 0, file, line);
839 	if (__predict_false(m == &blocked_lock))
840 		goto slowpath_unlocked;
841 	if (__predict_false(!_mtx_obtain_lock(m, tid)))
842 		goto slowpath_unlocked;
843 	if (__predict_true(m == td->td_lock)) {
844 		WITNESS_LOCK(&m->lock_object, LOP_EXCLUSIVE, file, line);
845 		return;
846 	}
847 	_mtx_release_lock_quick(m);
848 slowpath_unlocked:
849 	spinlock_exit();
850 slowpath_noirq:
851 #if LOCK_DEBUG > 0
852 	thread_lock_flags_(td, opts, file, line);
853 #else
854 	thread_lock_flags_(td, 0, 0, 0);
855 #endif
856 }
857 #endif
858 
859 void
860 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
861 {
862 	struct mtx *m;
863 	uintptr_t tid, v;
864 	struct lock_delay_arg lda;
865 #ifdef LOCK_PROFILING
866 	int contested = 0;
867 	uint64_t waittime = 0;
868 #endif
869 #ifdef KDTRACE_HOOKS
870 	int64_t spin_time = 0;
871 #endif
872 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
873 	int doing_lockprof = 1;
874 #endif
875 
876 	tid = (uintptr_t)curthread;
877 
878 	if (SCHEDULER_STOPPED()) {
879 		/*
880 		 * Ensure that spinlock sections are balanced even when the
881 		 * scheduler is stopped, since we may otherwise inadvertently
882 		 * re-enable interrupts while dumping core.
883 		 */
884 		spinlock_enter();
885 		return;
886 	}
887 
888 	lock_delay_arg_init(&lda, &mtx_spin_delay);
889 
890 #ifdef HWPMC_HOOKS
891 	PMC_SOFT_CALL( , , lock, failed);
892 #endif
893 
894 #ifdef LOCK_PROFILING
895 	doing_lockprof = 1;
896 #elif defined(KDTRACE_HOOKS)
897 	doing_lockprof = lockstat_enabled;
898 	if (__predict_false(doing_lockprof))
899 		spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
900 #endif
901 	spinlock_enter();
902 
903 	for (;;) {
904 retry:
905 		m = td->td_lock;
906 		thread_lock_validate(m, opts, file, line);
907 		v = MTX_READ_VALUE(m);
908 		for (;;) {
909 			if (v == MTX_UNOWNED) {
910 				if (_mtx_obtain_lock_fetch(m, &v, tid))
911 					break;
912 				continue;
913 			}
914 			MPASS(v != tid);
915 			lock_profile_obtain_lock_failed(&m->lock_object,
916 			    &contested, &waittime);
917 			/* Give interrupts a chance while we spin. */
918 			spinlock_exit();
919 			do {
920 				if (__predict_true(lda.spin_cnt < 10000000)) {
921 					lock_delay(&lda);
922 				} else {
923 					_mtx_lock_indefinite_check(m, &lda);
924 				}
925 				if (m != td->td_lock) {
926 					spinlock_enter();
927 					goto retry;
928 				}
929 				v = MTX_READ_VALUE(m);
930 			} while (v != MTX_UNOWNED);
931 			spinlock_enter();
932 		}
933 		if (m == td->td_lock)
934 			break;
935 		_mtx_release_lock_quick(m);
936 	}
937 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
938 	    line);
939 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
940 
941 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
942 	if (__predict_true(!doing_lockprof))
943 		return;
944 #endif
945 #ifdef KDTRACE_HOOKS
946 	spin_time += lockstat_nsecs(&m->lock_object);
947 #endif
948 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, contested,
949 	    waittime, file, line);
950 #ifdef KDTRACE_HOOKS
951 	if (lda.spin_cnt != 0)
952 		LOCKSTAT_RECORD1(thread__spin, m, spin_time);
953 #endif
954 }
955 
956 struct mtx *
957 thread_lock_block(struct thread *td)
958 {
959 	struct mtx *lock;
960 
961 	lock = td->td_lock;
962 	mtx_assert(lock, MA_OWNED);
963 	td->td_lock = &blocked_lock;
964 
965 	return (lock);
966 }
967 
968 void
969 thread_lock_unblock(struct thread *td, struct mtx *new)
970 {
971 
972 	mtx_assert(new, MA_OWNED);
973 	KASSERT(td->td_lock == &blocked_lock,
974 	    ("thread %p lock %p not blocked_lock %p",
975 	    td, td->td_lock, &blocked_lock));
976 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
977 }
978 
979 void
980 thread_lock_block_wait(struct thread *td)
981 {
982 
983 	while (td->td_lock == &blocked_lock)
984 		cpu_spinwait();
985 
986 	/* Acquire fence to be certain that all thread state is visible. */
987 	atomic_thread_fence_acq();
988 }
989 
990 void
991 thread_lock_set(struct thread *td, struct mtx *new)
992 {
993 	struct mtx *lock;
994 
995 	mtx_assert(new, MA_OWNED);
996 	lock = td->td_lock;
997 	mtx_assert(lock, MA_OWNED);
998 	td->td_lock = new;
999 	mtx_unlock_spin(lock);
1000 }
1001 
1002 /*
1003  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
1004  *
1005  * We are only called here if the lock is recursed, contested (i.e. we
1006  * need to wake up a blocked thread) or lockstat probe is active.
1007  */
1008 #if LOCK_DEBUG > 0
1009 void
1010 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v, int opts,
1011     const char *file, int line)
1012 #else
1013 void
1014 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v)
1015 #endif
1016 {
1017 	struct mtx *m;
1018 	struct turnstile *ts;
1019 	uintptr_t tid;
1020 
1021 	if (SCHEDULER_STOPPED())
1022 		return;
1023 
1024 	tid = (uintptr_t)curthread;
1025 	m = mtxlock2mtx(c);
1026 
1027 	if (__predict_false(v == tid))
1028 		v = MTX_READ_VALUE(m);
1029 
1030 	if (__predict_false(v & MTX_RECURSED)) {
1031 		if (--(m->mtx_recurse) == 0)
1032 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
1033 		if (LOCK_LOG_TEST(&m->lock_object, opts))
1034 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
1035 		return;
1036 	}
1037 
1038 	LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m);
1039 	if (v == tid && _mtx_release_lock(m, tid))
1040 		return;
1041 
1042 	/*
1043 	 * We have to lock the chain before the turnstile so this turnstile
1044 	 * can be removed from the hash list if it is empty.
1045 	 */
1046 	turnstile_chain_lock(&m->lock_object);
1047 	_mtx_release_lock_quick(m);
1048 	ts = turnstile_lookup(&m->lock_object);
1049 	MPASS(ts != NULL);
1050 	if (LOCK_LOG_TEST(&m->lock_object, opts))
1051 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
1052 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
1053 
1054 	/*
1055 	 * This turnstile is now no longer associated with the mutex.  We can
1056 	 * unlock the chain lock so a new turnstile may take it's place.
1057 	 */
1058 	turnstile_unpend(ts);
1059 	turnstile_chain_unlock(&m->lock_object);
1060 }
1061 
1062 /*
1063  * All the unlocking of MTX_SPIN locks is done inline.
1064  * See the __mtx_unlock_spin() macro for the details.
1065  */
1066 
1067 /*
1068  * The backing function for the INVARIANTS-enabled mtx_assert()
1069  */
1070 #ifdef INVARIANT_SUPPORT
1071 void
1072 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1073 {
1074 	const struct mtx *m;
1075 
1076 	if (KERNEL_PANICKED() || dumping || SCHEDULER_STOPPED())
1077 		return;
1078 
1079 	m = mtxlock2mtx(c);
1080 
1081 	switch (what) {
1082 	case MA_OWNED:
1083 	case MA_OWNED | MA_RECURSED:
1084 	case MA_OWNED | MA_NOTRECURSED:
1085 		if (!mtx_owned(m))
1086 			panic("mutex %s not owned at %s:%d",
1087 			    m->lock_object.lo_name, file, line);
1088 		if (mtx_recursed(m)) {
1089 			if ((what & MA_NOTRECURSED) != 0)
1090 				panic("mutex %s recursed at %s:%d",
1091 				    m->lock_object.lo_name, file, line);
1092 		} else if ((what & MA_RECURSED) != 0) {
1093 			panic("mutex %s unrecursed at %s:%d",
1094 			    m->lock_object.lo_name, file, line);
1095 		}
1096 		break;
1097 	case MA_NOTOWNED:
1098 		if (mtx_owned(m))
1099 			panic("mutex %s owned at %s:%d",
1100 			    m->lock_object.lo_name, file, line);
1101 		break;
1102 	default:
1103 		panic("unknown mtx_assert at %s:%d", file, line);
1104 	}
1105 }
1106 #endif
1107 
1108 /*
1109  * General init routine used by the MTX_SYSINIT() macro.
1110  */
1111 void
1112 mtx_sysinit(void *arg)
1113 {
1114 	struct mtx_args *margs = arg;
1115 
1116 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
1117 	    margs->ma_opts);
1118 }
1119 
1120 /*
1121  * Mutex initialization routine; initialize lock `m' of type contained in
1122  * `opts' with options contained in `opts' and name `name.'  The optional
1123  * lock type `type' is used as a general lock category name for use with
1124  * witness.
1125  */
1126 void
1127 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
1128 {
1129 	struct mtx *m;
1130 	struct lock_class *class;
1131 	int flags;
1132 
1133 	m = mtxlock2mtx(c);
1134 
1135 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
1136 	    MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
1137 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
1138 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
1139 	    &m->mtx_lock));
1140 
1141 	/* Determine lock class and lock flags. */
1142 	if (opts & MTX_SPIN)
1143 		class = &lock_class_mtx_spin;
1144 	else
1145 		class = &lock_class_mtx_sleep;
1146 	flags = 0;
1147 	if (opts & MTX_QUIET)
1148 		flags |= LO_QUIET;
1149 	if (opts & MTX_RECURSE)
1150 		flags |= LO_RECURSABLE;
1151 	if ((opts & MTX_NOWITNESS) == 0)
1152 		flags |= LO_WITNESS;
1153 	if (opts & MTX_DUPOK)
1154 		flags |= LO_DUPOK;
1155 	if (opts & MTX_NOPROFILE)
1156 		flags |= LO_NOPROFILE;
1157 	if (opts & MTX_NEW)
1158 		flags |= LO_NEW;
1159 
1160 	/* Initialize mutex. */
1161 	lock_init(&m->lock_object, class, name, type, flags);
1162 
1163 	m->mtx_lock = MTX_UNOWNED;
1164 	m->mtx_recurse = 0;
1165 }
1166 
1167 /*
1168  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
1169  * passed in as a flag here because if the corresponding mtx_init() was
1170  * called with MTX_QUIET set, then it will already be set in the mutex's
1171  * flags.
1172  */
1173 void
1174 _mtx_destroy(volatile uintptr_t *c)
1175 {
1176 	struct mtx *m;
1177 
1178 	m = mtxlock2mtx(c);
1179 
1180 	if (!mtx_owned(m))
1181 		MPASS(mtx_unowned(m));
1182 	else {
1183 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
1184 
1185 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
1186 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
1187 			spinlock_exit();
1188 		else
1189 			TD_LOCKS_DEC(curthread);
1190 
1191 		lock_profile_release_lock(&m->lock_object);
1192 		/* Tell witness this isn't locked to make it happy. */
1193 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
1194 		    __LINE__);
1195 	}
1196 
1197 	m->mtx_lock = MTX_DESTROYED;
1198 	lock_destroy(&m->lock_object);
1199 }
1200 
1201 /*
1202  * Intialize the mutex code and system mutexes.  This is called from the MD
1203  * startup code prior to mi_startup().  The per-CPU data space needs to be
1204  * setup before this is called.
1205  */
1206 void
1207 mutex_init(void)
1208 {
1209 
1210 	/* Setup turnstiles so that sleep mutexes work. */
1211 	init_turnstiles();
1212 
1213 	/*
1214 	 * Initialize mutexes.
1215 	 */
1216 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
1217 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
1218 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
1219 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
1220 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
1221 	mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
1222 	mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1223 	mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1224 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1225 	mtx_lock(&Giant);
1226 }
1227 
1228 static void __noinline
1229 _mtx_lock_indefinite_check(struct mtx *m, struct lock_delay_arg *ldap)
1230 {
1231 	struct thread *td;
1232 
1233 	ldap->spin_cnt++;
1234 	if (ldap->spin_cnt < 60000000 || kdb_active || KERNEL_PANICKED())
1235 		cpu_lock_delay();
1236 	else {
1237 		td = mtx_owner(m);
1238 
1239 		/* If the mutex is unlocked, try again. */
1240 		if (td == NULL)
1241 			return;
1242 
1243 		printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
1244 		    m, m->lock_object.lo_name, td, td->td_tid);
1245 #ifdef WITNESS
1246 		witness_display_spinlock(&m->lock_object, td, printf);
1247 #endif
1248 		panic("spin lock held too long");
1249 	}
1250 	cpu_spinwait();
1251 }
1252 
1253 void
1254 mtx_spin_wait_unlocked(struct mtx *m)
1255 {
1256 	struct lock_delay_arg lda;
1257 
1258 	KASSERT(m->mtx_lock != MTX_DESTROYED,
1259 	    ("%s() of destroyed mutex %p", __func__, m));
1260 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
1261 	    ("%s() of sleep mutex %p (%s)", __func__, m,
1262 	    m->lock_object.lo_name));
1263 	KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1264 	    m->lock_object.lo_name));
1265 
1266 	lda.spin_cnt = 0;
1267 
1268 	while (atomic_load_acq_ptr(&m->mtx_lock) != MTX_UNOWNED) {
1269 		if (__predict_true(lda.spin_cnt < 10000000)) {
1270 			cpu_spinwait();
1271 			lda.spin_cnt++;
1272 		} else {
1273 			_mtx_lock_indefinite_check(m, &lda);
1274 		}
1275 	}
1276 }
1277 
1278 void
1279 mtx_wait_unlocked(struct mtx *m)
1280 {
1281 	struct thread *owner;
1282 	uintptr_t v;
1283 
1284 	KASSERT(m->mtx_lock != MTX_DESTROYED,
1285 	    ("%s() of destroyed mutex %p", __func__, m));
1286 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
1287 	    ("%s() not a sleep mutex %p (%s)", __func__, m,
1288 	    m->lock_object.lo_name));
1289 	KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1290 	    m->lock_object.lo_name));
1291 
1292 	for (;;) {
1293 		v = atomic_load_acq_ptr(&m->mtx_lock);
1294 		if (v == MTX_UNOWNED) {
1295 			break;
1296 		}
1297 		owner = lv_mtx_owner(v);
1298 		if (!TD_IS_RUNNING(owner)) {
1299 			mtx_lock(m);
1300 			mtx_unlock(m);
1301 			break;
1302 		}
1303 		cpu_spinwait();
1304 	}
1305 }
1306 
1307 #ifdef DDB
1308 void
1309 db_show_mtx(const struct lock_object *lock)
1310 {
1311 	struct thread *td;
1312 	const struct mtx *m;
1313 
1314 	m = (const struct mtx *)lock;
1315 
1316 	db_printf(" flags: {");
1317 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1318 		db_printf("SPIN");
1319 	else
1320 		db_printf("DEF");
1321 	if (m->lock_object.lo_flags & LO_RECURSABLE)
1322 		db_printf(", RECURSE");
1323 	if (m->lock_object.lo_flags & LO_DUPOK)
1324 		db_printf(", DUPOK");
1325 	db_printf("}\n");
1326 	db_printf(" state: {");
1327 	if (mtx_unowned(m))
1328 		db_printf("UNOWNED");
1329 	else if (mtx_destroyed(m))
1330 		db_printf("DESTROYED");
1331 	else {
1332 		db_printf("OWNED");
1333 		if (m->mtx_lock & MTX_CONTESTED)
1334 			db_printf(", CONTESTED");
1335 		if (m->mtx_lock & MTX_RECURSED)
1336 			db_printf(", RECURSED");
1337 	}
1338 	db_printf("}\n");
1339 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1340 		td = mtx_owner(m);
1341 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1342 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1343 		if (mtx_recursed(m))
1344 			db_printf(" recursed: %d\n", m->mtx_recurse);
1345 	}
1346 }
1347 #endif
1348