xref: /freebsd/sys/kern/kern_mutex.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
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_mprof.h"
42 #include "opt_mutex_wake_all.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/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resourcevar.h>
56 #include <sys/sched.h>
57 #include <sys/sbuf.h>
58 #include <sys/sysctl.h>
59 #include <sys/turnstile.h>
60 #include <sys/vmmeter.h>
61 
62 #include <machine/atomic.h>
63 #include <machine/bus.h>
64 #include <machine/clock.h>
65 #include <machine/cpu.h>
66 
67 #include <ddb/ddb.h>
68 
69 #include <vm/vm.h>
70 #include <vm/vm_extern.h>
71 
72 /*
73  * Force MUTEX_WAKE_ALL for now.
74  * single thread wakeup needs fixes to avoid race conditions with
75  * priority inheritance.
76  */
77 #ifndef MUTEX_WAKE_ALL
78 #define MUTEX_WAKE_ALL
79 #endif
80 
81 /*
82  * Internal utility macros.
83  */
84 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
85 
86 #define mtx_owner(m)	(mtx_unowned((m)) ? NULL \
87 	: (struct thread *)((m)->mtx_lock & MTX_FLAGMASK))
88 
89 /*
90  * Lock classes for sleep and spin mutexes.
91  */
92 struct lock_class lock_class_mtx_sleep = {
93 	"sleep mutex",
94 	LC_SLEEPLOCK | LC_RECURSABLE
95 };
96 struct lock_class lock_class_mtx_spin = {
97 	"spin mutex",
98 	LC_SPINLOCK | LC_RECURSABLE
99 };
100 
101 /*
102  * System-wide mutexes
103  */
104 struct mtx sched_lock;
105 struct mtx Giant;
106 
107 #ifdef MUTEX_PROFILING
108 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging");
109 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling");
110 static int mutex_prof_enable = 0;
111 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW,
112     &mutex_prof_enable, 0, "Enable tracing of mutex holdtime");
113 
114 struct mutex_prof {
115 	const char	*name;
116 	const char	*file;
117 	int		line;
118 	uintmax_t	cnt_max;
119 	uintmax_t	cnt_tot;
120 	uintmax_t	cnt_cur;
121 	uintmax_t	cnt_contest_holding;
122 	uintmax_t	cnt_contest_locking;
123 	struct mutex_prof *next;
124 };
125 
126 /*
127  * mprof_buf is a static pool of profiling records to avoid possible
128  * reentrance of the memory allocation functions.
129  *
130  * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE.
131  */
132 #ifdef MPROF_BUFFERS
133 #define NUM_MPROF_BUFFERS	MPROF_BUFFERS
134 #else
135 #define	NUM_MPROF_BUFFERS	1000
136 #endif
137 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS];
138 static int first_free_mprof_buf;
139 #ifndef MPROF_HASH_SIZE
140 #define	MPROF_HASH_SIZE		1009
141 #endif
142 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE
143 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE
144 #endif
145 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE];
146 /* SWAG: sbuf size = avg stat. line size * number of locks */
147 #define MPROF_SBUF_SIZE		256 * 400
148 
149 static int mutex_prof_acquisitions;
150 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
151     &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded");
152 static int mutex_prof_records;
153 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD,
154     &mutex_prof_records, 0, "Number of profiling records");
155 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS;
156 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
157     &mutex_prof_maxrecords, 0, "Maximum number of profiling records");
158 static int mutex_prof_rejected;
159 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD,
160     &mutex_prof_rejected, 0, "Number of rejected profiling records");
161 static int mutex_prof_hashsize = MPROF_HASH_SIZE;
162 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD,
163     &mutex_prof_hashsize, 0, "Hash size");
164 static int mutex_prof_collisions = 0;
165 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD,
166     &mutex_prof_collisions, 0, "Number of hash collisions");
167 
168 /*
169  * mprof_mtx protects the profiling buffers and the hash.
170  */
171 static struct mtx mprof_mtx;
172 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET);
173 
174 static u_int64_t
175 nanoseconds(void)
176 {
177 	struct timespec tv;
178 
179 	nanotime(&tv);
180 	return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
181 }
182 
183 static int
184 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
185 {
186 	struct sbuf *sb;
187 	int error, i;
188 	static int multiplier = 1;
189 
190 	if (first_free_mprof_buf == 0)
191 		return (SYSCTL_OUT(req, "No locking recorded",
192 		    sizeof("No locking recorded")));
193 
194 retry_sbufops:
195 	sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
196 	sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n",
197 	    "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name");
198 	/*
199 	 * XXX this spinlock seems to be by far the largest perpetrator
200 	 * of spinlock latency (1.6 msec on an Athlon1600 was recorded
201 	 * even before I pessimized it further by moving the average
202 	 * computation here).
203 	 */
204 	mtx_lock_spin(&mprof_mtx);
205 	for (i = 0; i < first_free_mprof_buf; ++i) {
206 		sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n",
207 		    mprof_buf[i].cnt_max / 1000,
208 		    mprof_buf[i].cnt_tot / 1000,
209 		    mprof_buf[i].cnt_cur,
210 		    mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
211 			mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000),
212 		    mprof_buf[i].cnt_contest_holding,
213 		    mprof_buf[i].cnt_contest_locking,
214 		    mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name);
215 		if (sbuf_overflowed(sb)) {
216 			mtx_unlock_spin(&mprof_mtx);
217 			sbuf_delete(sb);
218 			multiplier++;
219 			goto retry_sbufops;
220 		}
221 	}
222 	mtx_unlock_spin(&mprof_mtx);
223 	sbuf_finish(sb);
224 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
225 	sbuf_delete(sb);
226 	return (error);
227 }
228 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
229     NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics");
230 
231 static int
232 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
233 {
234 	int error, v;
235 
236 	if (first_free_mprof_buf == 0)
237 		return (0);
238 
239 	v = 0;
240 	error = sysctl_handle_int(oidp, &v, 0, req);
241 	if (error)
242 		return (error);
243 	if (req->newptr == NULL)
244 		return (error);
245 	if (v == 0)
246 		return (0);
247 
248 	mtx_lock_spin(&mprof_mtx);
249 	bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf);
250 	bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE);
251 	first_free_mprof_buf = 0;
252 	mtx_unlock_spin(&mprof_mtx);
253 	return (0);
254 }
255 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
256     NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics");
257 #endif
258 
259 /*
260  * Function versions of the inlined __mtx_* macros.  These are used by
261  * modules and can also be called from assembly language if needed.
262  */
263 void
264 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
265 {
266 
267 	MPASS(curthread != NULL);
268 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
269 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
270 	    file, line));
271 	WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
272 	    file, line);
273 	_get_sleep_lock(m, curthread, opts, file, line);
274 	LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
275 	    line);
276 	WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
277 #ifdef MUTEX_PROFILING
278 	/* don't reset the timer when/if recursing */
279 	if (m->mtx_acqtime == 0) {
280 		m->mtx_filename = file;
281 		m->mtx_lineno = line;
282 		m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0;
283 		++mutex_prof_acquisitions;
284 	}
285 #endif
286 }
287 
288 void
289 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
290 {
291 
292 	MPASS(curthread != NULL);
293 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
294 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
295 	    file, line));
296 	WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
297 	LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
298 	    line);
299 	mtx_assert(m, MA_OWNED);
300 #ifdef MUTEX_PROFILING
301 	if (m->mtx_acqtime != 0) {
302 		static const char *unknown = "(unknown)";
303 		struct mutex_prof *mpp;
304 		u_int64_t acqtime, now;
305 		const char *p, *q;
306 		volatile u_int hash;
307 
308 		now = nanoseconds();
309 		acqtime = m->mtx_acqtime;
310 		m->mtx_acqtime = 0;
311 		if (now <= acqtime)
312 			goto out;
313 		for (p = m->mtx_filename;
314 		    p != NULL && strncmp(p, "../", 3) == 0; p += 3)
315 			/* nothing */ ;
316 		if (p == NULL || *p == '\0')
317 			p = unknown;
318 		for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q)
319 			hash = (hash * 2 + *q) % MPROF_HASH_SIZE;
320 		mtx_lock_spin(&mprof_mtx);
321 		for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next)
322 			if (mpp->line == m->mtx_lineno &&
323 			    strcmp(mpp->file, p) == 0)
324 				break;
325 		if (mpp == NULL) {
326 			/* Just exit if we cannot get a trace buffer */
327 			if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) {
328 				++mutex_prof_rejected;
329 				goto unlock;
330 			}
331 			mpp = &mprof_buf[first_free_mprof_buf++];
332 			mpp->name = mtx_name(m);
333 			mpp->file = p;
334 			mpp->line = m->mtx_lineno;
335 			mpp->next = mprof_hash[hash];
336 			if (mprof_hash[hash] != NULL)
337 				++mutex_prof_collisions;
338 			mprof_hash[hash] = mpp;
339 			++mutex_prof_records;
340 		}
341 		/*
342 		 * Record if the mutex has been held longer now than ever
343 		 * before.
344 		 */
345 		if (now - acqtime > mpp->cnt_max)
346 			mpp->cnt_max = now - acqtime;
347 		mpp->cnt_tot += now - acqtime;
348 		mpp->cnt_cur++;
349 		/*
350 		 * There's a small race, really we should cmpxchg
351 		 * 0 with the current value, but that would bill
352 		 * the contention to the wrong lock instance if
353 		 * it followed this also.
354 		 */
355 		mpp->cnt_contest_holding += m->mtx_contest_holding;
356 		m->mtx_contest_holding = 0;
357 		mpp->cnt_contest_locking += m->mtx_contest_locking;
358 		m->mtx_contest_locking = 0;
359 unlock:
360 		mtx_unlock_spin(&mprof_mtx);
361 	}
362 out:
363 #endif
364 	_rel_sleep_lock(m, curthread, opts, file, line);
365 }
366 
367 void
368 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
369 {
370 
371 	MPASS(curthread != NULL);
372 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
373 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
374 	    m->mtx_object.lo_name, file, line));
375 	WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
376 	    file, line);
377 	_get_spin_lock(m, curthread, opts, file, line);
378 	LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
379 	    line);
380 	WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
381 }
382 
383 void
384 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
385 {
386 
387 	MPASS(curthread != NULL);
388 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
389 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
390 	    m->mtx_object.lo_name, file, line));
391 	WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
392 	LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
393 	    line);
394 	mtx_assert(m, MA_OWNED);
395 	_rel_spin_lock(m);
396 }
397 
398 /*
399  * The important part of mtx_trylock{,_flags}()
400  * Tries to acquire lock `m.'  If this function is called on a mutex that
401  * is already owned, it will recursively acquire the lock.
402  */
403 int
404 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
405 {
406 	int rval;
407 
408 	MPASS(curthread != NULL);
409 
410 	if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) {
411 		m->mtx_recurse++;
412 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
413 		rval = 1;
414 	} else
415 		rval = _obtain_lock(m, curthread);
416 
417 	LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line);
418 	if (rval)
419 		WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
420 		    file, line);
421 
422 	return (rval);
423 }
424 
425 /*
426  * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
427  *
428  * We call this if the lock is either contested (i.e. we need to go to
429  * sleep waiting for it), or if we need to recurse on it.
430  */
431 void
432 _mtx_lock_sleep(struct mtx *m, struct thread *td, int opts, const char *file,
433     int line)
434 {
435 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
436 	struct thread *owner;
437 #endif
438 	uintptr_t v;
439 #ifdef KTR
440 	int cont_logged = 0;
441 #endif
442 #ifdef MUTEX_PROFILING
443 	int contested;
444 #endif
445 
446 	if (mtx_owned(m)) {
447 		KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0,
448 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
449 		    m->mtx_object.lo_name, file, line));
450 		m->mtx_recurse++;
451 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
452 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
453 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
454 		return;
455 	}
456 
457 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
458 		CTR4(KTR_LOCK,
459 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
460 		    m->mtx_object.lo_name, (void *)m->mtx_lock, file, line);
461 
462 #ifdef MUTEX_PROFILING
463 	contested = 0;
464 #endif
465 	while (!_obtain_lock(m, td)) {
466 #ifdef MUTEX_PROFILING
467 		contested = 1;
468 		atomic_add_int(&m->mtx_contest_holding, 1);
469 #endif
470 		turnstile_lock(&m->mtx_object);
471 		v = m->mtx_lock;
472 
473 		/*
474 		 * Check if the lock has been released while spinning for
475 		 * the turnstile chain lock.
476 		 */
477 		if (v == MTX_UNOWNED) {
478 			turnstile_release(&m->mtx_object);
479 			cpu_spinwait();
480 			continue;
481 		}
482 
483 #ifdef MUTEX_WAKE_ALL
484 		MPASS(v != MTX_CONTESTED);
485 #else
486 		/*
487 		 * The mutex was marked contested on release. This means that
488 		 * there are other threads blocked on it.  Grab ownership of
489 		 * it and propagate its priority to the current thread if
490 		 * necessary.
491 		 */
492 		if (v == MTX_CONTESTED) {
493 			m->mtx_lock = (uintptr_t)td | MTX_CONTESTED;
494 			turnstile_claim(&m->mtx_object);
495 			break;
496 		}
497 #endif
498 
499 		/*
500 		 * If the mutex isn't already contested and a failure occurs
501 		 * setting the contested bit, the mutex was either released
502 		 * or the state of the MTX_RECURSED bit changed.
503 		 */
504 		if ((v & MTX_CONTESTED) == 0 &&
505 		    !atomic_cmpset_ptr(&m->mtx_lock, (void *)v,
506 			(void *)(v | MTX_CONTESTED))) {
507 			turnstile_release(&m->mtx_object);
508 			cpu_spinwait();
509 			continue;
510 		}
511 
512 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
513 		/*
514 		 * If the current owner of the lock is executing on another
515 		 * CPU, spin instead of blocking.
516 		 */
517 		owner = (struct thread *)(v & MTX_FLAGMASK);
518 #ifdef ADAPTIVE_GIANT
519 		if (TD_IS_RUNNING(owner)) {
520 #else
521 		if (m != &Giant && TD_IS_RUNNING(owner)) {
522 #endif
523 			turnstile_release(&m->mtx_object);
524 			while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
525 				cpu_spinwait();
526 			}
527 			continue;
528 		}
529 #endif	/* SMP && !NO_ADAPTIVE_MUTEXES */
530 
531 		/*
532 		 * We definitely must sleep for this lock.
533 		 */
534 		mtx_assert(m, MA_NOTOWNED);
535 
536 #ifdef KTR
537 		if (!cont_logged) {
538 			CTR6(KTR_CONTENTION,
539 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
540 			    td, file, line, m->mtx_object.lo_name,
541 			    WITNESS_FILE(&m->mtx_object),
542 			    WITNESS_LINE(&m->mtx_object));
543 			cont_logged = 1;
544 		}
545 #endif
546 
547 		/*
548 		 * Block on the turnstile.
549 		 */
550 		turnstile_wait(&m->mtx_object, mtx_owner(m));
551 	}
552 
553 #ifdef KTR
554 	if (cont_logged) {
555 		CTR4(KTR_CONTENTION,
556 		    "contention end: %s acquired by %p at %s:%d",
557 		    m->mtx_object.lo_name, td, file, line);
558 	}
559 #endif
560 #ifdef MUTEX_PROFILING
561 	if (contested)
562 		m->mtx_contest_locking++;
563 	m->mtx_contest_holding = 0;
564 #endif
565 	return;
566 }
567 
568 #ifdef SMP
569 /*
570  * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
571  *
572  * This is only called if we need to actually spin for the lock. Recursion
573  * is handled inline.
574  */
575 void
576 _mtx_lock_spin(struct mtx *m, struct thread *td, int opts, const char *file,
577     int line)
578 {
579 	int i = 0;
580 
581 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
582 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
583 
584 	for (;;) {
585 		if (_obtain_lock(m, td))
586 			break;
587 
588 		/* Give interrupts a chance while we spin. */
589 		critical_exit();
590 		while (m->mtx_lock != MTX_UNOWNED) {
591 			if (i++ < 10000000) {
592 				cpu_spinwait();
593 				continue;
594 			}
595 			if (i < 60000000)
596 				DELAY(1);
597 			else if (!kdb_active) {
598 				printf("spin lock %s held by %p for > 5 seconds\n",
599 				    m->mtx_object.lo_name, (void *)m->mtx_lock);
600 #ifdef WITNESS
601 				witness_display_spinlock(&m->mtx_object,
602 				    mtx_owner(m));
603 #endif
604 				panic("spin lock held too long");
605 			}
606 			cpu_spinwait();
607 		}
608 		critical_enter();
609 	}
610 
611 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
612 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
613 
614 	return;
615 }
616 #endif /* SMP */
617 
618 /*
619  * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
620  *
621  * We are only called here if the lock is recursed or contested (i.e. we
622  * need to wake up a blocked thread).
623  */
624 void
625 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
626 {
627 	struct turnstile *ts;
628 #ifndef PREEMPTION
629 	struct thread *td, *td1;
630 #endif
631 
632 	if (mtx_recursed(m)) {
633 		if (--(m->mtx_recurse) == 0)
634 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
635 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
636 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
637 		return;
638 	}
639 
640 	turnstile_lock(&m->mtx_object);
641 	ts = turnstile_lookup(&m->mtx_object);
642 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
643 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
644 
645 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
646 	if (ts == NULL) {
647 		_release_lock_quick(m);
648 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
649 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
650 		turnstile_release(&m->mtx_object);
651 		return;
652 	}
653 #else
654 	MPASS(ts != NULL);
655 #endif
656 #ifndef PREEMPTION
657 	/* XXX */
658 	td1 = turnstile_head(ts);
659 #endif
660 #ifdef MUTEX_WAKE_ALL
661 	turnstile_broadcast(ts);
662 	_release_lock_quick(m);
663 #else
664 	if (turnstile_signal(ts)) {
665 		_release_lock_quick(m);
666 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
667 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m);
668 	} else {
669 		m->mtx_lock = MTX_CONTESTED;
670 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
671 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested",
672 			    m);
673 	}
674 #endif
675 	turnstile_unpend(ts);
676 
677 #ifndef PREEMPTION
678 	/*
679 	 * XXX: This is just a hack until preemption is done.  However,
680 	 * once preemption is done we need to either wrap the
681 	 * turnstile_signal() and release of the actual lock in an
682 	 * extra critical section or change the preemption code to
683 	 * always just set a flag and never do instant-preempts.
684 	 */
685 	td = curthread;
686 	if (td->td_critnest > 0 || td1->td_priority >= td->td_priority)
687 		return;
688 	mtx_lock_spin(&sched_lock);
689 	if (!TD_IS_RUNNING(td1)) {
690 #ifdef notyet
691 		if (td->td_ithd != NULL) {
692 			struct ithd *it = td->td_ithd;
693 
694 			if (it->it_interrupted) {
695 				if (LOCK_LOG_TEST(&m->mtx_object, opts))
696 					CTR2(KTR_LOCK,
697 				    "_mtx_unlock_sleep: %p interrupted %p",
698 					    it, it->it_interrupted);
699 				intr_thd_fixup(it);
700 			}
701 		}
702 #endif
703 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
704 			CTR2(KTR_LOCK,
705 			    "_mtx_unlock_sleep: %p switching out lock=%p", m,
706 			    (void *)m->mtx_lock);
707 
708 		mi_switch(SW_INVOL, NULL);
709 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
710 			CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p",
711 			    m, (void *)m->mtx_lock);
712 	}
713 	mtx_unlock_spin(&sched_lock);
714 #endif
715 
716 	return;
717 }
718 
719 /*
720  * All the unlocking of MTX_SPIN locks is done inline.
721  * See the _rel_spin_lock() macro for the details.
722  */
723 
724 /*
725  * The backing function for the INVARIANTS-enabled mtx_assert()
726  */
727 #ifdef INVARIANT_SUPPORT
728 void
729 _mtx_assert(struct mtx *m, int what, const char *file, int line)
730 {
731 
732 	if (panicstr != NULL)
733 		return;
734 	switch (what) {
735 	case MA_OWNED:
736 	case MA_OWNED | MA_RECURSED:
737 	case MA_OWNED | MA_NOTRECURSED:
738 		if (!mtx_owned(m))
739 			panic("mutex %s not owned at %s:%d",
740 			    m->mtx_object.lo_name, file, line);
741 		if (mtx_recursed(m)) {
742 			if ((what & MA_NOTRECURSED) != 0)
743 				panic("mutex %s recursed at %s:%d",
744 				    m->mtx_object.lo_name, file, line);
745 		} else if ((what & MA_RECURSED) != 0) {
746 			panic("mutex %s unrecursed at %s:%d",
747 			    m->mtx_object.lo_name, file, line);
748 		}
749 		break;
750 	case MA_NOTOWNED:
751 		if (mtx_owned(m))
752 			panic("mutex %s owned at %s:%d",
753 			    m->mtx_object.lo_name, file, line);
754 		break;
755 	default:
756 		panic("unknown mtx_assert at %s:%d", file, line);
757 	}
758 }
759 #endif
760 
761 /*
762  * The MUTEX_DEBUG-enabled mtx_validate()
763  *
764  * Most of these checks have been moved off into the LO_INITIALIZED flag
765  * maintained by the witness code.
766  */
767 #ifdef MUTEX_DEBUG
768 
769 void	mtx_validate(struct mtx *);
770 
771 void
772 mtx_validate(struct mtx *m)
773 {
774 
775 /*
776  * XXX: When kernacc() does not require Giant we can reenable this check
777  */
778 #ifdef notyet
779 /*
780  * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly
781  * we can re-enable the kernacc() checks.
782  */
783 #ifndef __alpha__
784 	/*
785 	 * Can't call kernacc() from early init386(), especially when
786 	 * initializing Giant mutex, because some stuff in kernacc()
787 	 * requires Giant itself.
788 	 */
789 	if (!cold)
790 		if (!kernacc((caddr_t)m, sizeof(m),
791 		    VM_PROT_READ | VM_PROT_WRITE))
792 			panic("Can't read and write to mutex %p", m);
793 #endif
794 #endif
795 }
796 #endif
797 
798 /*
799  * General init routine used by the MTX_SYSINIT() macro.
800  */
801 void
802 mtx_sysinit(void *arg)
803 {
804 	struct mtx_args *margs = arg;
805 
806 	mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
807 }
808 
809 /*
810  * Mutex initialization routine; initialize lock `m' of type contained in
811  * `opts' with options contained in `opts' and name `name.'  The optional
812  * lock type `type' is used as a general lock category name for use with
813  * witness.
814  */
815 void
816 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
817 {
818 	struct lock_object *lock;
819 
820 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
821 	    MTX_NOWITNESS | MTX_DUPOK)) == 0);
822 
823 #ifdef MUTEX_DEBUG
824 	/* Diagnostic and error correction */
825 	mtx_validate(m);
826 #endif
827 
828 	lock = &m->mtx_object;
829 	KASSERT((lock->lo_flags & LO_INITIALIZED) == 0,
830 	    ("mutex \"%s\" %p already initialized", name, m));
831 	bzero(m, sizeof(*m));
832 	if (opts & MTX_SPIN)
833 		lock->lo_class = &lock_class_mtx_spin;
834 	else
835 		lock->lo_class = &lock_class_mtx_sleep;
836 	lock->lo_name = name;
837 	lock->lo_type = type != NULL ? type : name;
838 	if (opts & MTX_QUIET)
839 		lock->lo_flags = LO_QUIET;
840 	if (opts & MTX_RECURSE)
841 		lock->lo_flags |= LO_RECURSABLE;
842 	if ((opts & MTX_NOWITNESS) == 0)
843 		lock->lo_flags |= LO_WITNESS;
844 	if (opts & MTX_DUPOK)
845 		lock->lo_flags |= LO_DUPOK;
846 
847 	m->mtx_lock = MTX_UNOWNED;
848 
849 	LOCK_LOG_INIT(lock, opts);
850 
851 	WITNESS_INIT(lock);
852 }
853 
854 /*
855  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
856  * passed in as a flag here because if the corresponding mtx_init() was
857  * called with MTX_QUIET set, then it will already be set in the mutex's
858  * flags.
859  */
860 void
861 mtx_destroy(struct mtx *m)
862 {
863 
864 	LOCK_LOG_DESTROY(&m->mtx_object, 0);
865 
866 	if (!mtx_owned(m))
867 		MPASS(mtx_unowned(m));
868 	else {
869 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
870 
871 		/* Tell witness this isn't locked to make it happy. */
872 		WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__,
873 		    __LINE__);
874 	}
875 
876 	WITNESS_DESTROY(&m->mtx_object);
877 }
878 
879 /*
880  * Intialize the mutex code and system mutexes.  This is called from the MD
881  * startup code prior to mi_startup().  The per-CPU data space needs to be
882  * setup before this is called.
883  */
884 void
885 mutex_init(void)
886 {
887 
888 	/* Setup thread0 so that mutexes work. */
889 	LIST_INIT(&thread0.td_contested);
890 
891 	/* Setup turnstiles so that sleep mutexes work. */
892 	init_turnstiles();
893 
894 	/*
895 	 * Initialize mutexes.
896 	 */
897 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
898 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
899 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
900 	mtx_lock(&Giant);
901 }
902