xref: /freebsd/sys/kern/kern_mutex.c (revision c0b9f4fe659b6839541970eb5675e57f4d814969)
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/conf.h>
49 #include <sys/kdb.h>
50 #include <sys/kernel.h>
51 #include <sys/ktr.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/mutex.h>
55 #include <sys/proc.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sbuf.h>
59 #include <sys/sysctl.h>
60 #include <sys/turnstile.h>
61 #include <sys/vmmeter.h>
62 
63 #include <machine/atomic.h>
64 #include <machine/bus.h>
65 #include <machine/clock.h>
66 #include <machine/cpu.h>
67 
68 #include <ddb/ddb.h>
69 
70 #include <fs/devfs/devfs_int.h>
71 
72 #include <vm/vm.h>
73 #include <vm/vm_extern.h>
74 
75 /*
76  * Force MUTEX_WAKE_ALL for now.
77  * single thread wakeup needs fixes to avoid race conditions with
78  * priority inheritance.
79  */
80 #ifndef MUTEX_WAKE_ALL
81 #define MUTEX_WAKE_ALL
82 #endif
83 
84 /*
85  * Internal utility macros.
86  */
87 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
88 
89 #define mtx_owner(m)	(mtx_unowned((m)) ? NULL \
90 	: (struct thread *)((m)->mtx_lock & MTX_FLAGMASK))
91 
92 #ifdef DDB
93 static void	db_show_mtx(struct lock_object *lock);
94 #endif
95 
96 /*
97  * Lock classes for sleep and spin mutexes.
98  */
99 struct lock_class lock_class_mtx_sleep = {
100 	"sleep mutex",
101 	LC_SLEEPLOCK | LC_RECURSABLE,
102 #ifdef DDB
103 	db_show_mtx
104 #endif
105 };
106 struct lock_class lock_class_mtx_spin = {
107 	"spin mutex",
108 	LC_SPINLOCK | LC_RECURSABLE,
109 #ifdef DDB
110 	db_show_mtx
111 #endif
112 };
113 
114 /*
115  * System-wide mutexes
116  */
117 struct mtx sched_lock;
118 struct mtx Giant;
119 
120 #ifdef MUTEX_PROFILING
121 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging");
122 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling");
123 static int mutex_prof_enable = 0;
124 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW,
125     &mutex_prof_enable, 0, "Enable tracing of mutex holdtime");
126 
127 struct mutex_prof {
128 	const char	*name;
129 	const char	*file;
130 	int		line;
131 	uintmax_t	cnt_max;
132 	uintmax_t	cnt_tot;
133 	uintmax_t	cnt_cur;
134 	uintmax_t	cnt_contest_holding;
135 	uintmax_t	cnt_contest_locking;
136 	struct mutex_prof *next;
137 };
138 
139 /*
140  * mprof_buf is a static pool of profiling records to avoid possible
141  * reentrance of the memory allocation functions.
142  *
143  * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE.
144  */
145 #ifdef MPROF_BUFFERS
146 #define NUM_MPROF_BUFFERS	MPROF_BUFFERS
147 #else
148 #define	NUM_MPROF_BUFFERS	1000
149 #endif
150 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS];
151 static int first_free_mprof_buf;
152 #ifndef MPROF_HASH_SIZE
153 #define	MPROF_HASH_SIZE		1009
154 #endif
155 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE
156 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE
157 #endif
158 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE];
159 /* SWAG: sbuf size = avg stat. line size * number of locks */
160 #define MPROF_SBUF_SIZE		256 * 400
161 
162 static int mutex_prof_acquisitions;
163 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
164     &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded");
165 static int mutex_prof_records;
166 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD,
167     &mutex_prof_records, 0, "Number of profiling records");
168 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS;
169 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
170     &mutex_prof_maxrecords, 0, "Maximum number of profiling records");
171 static int mutex_prof_rejected;
172 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD,
173     &mutex_prof_rejected, 0, "Number of rejected profiling records");
174 static int mutex_prof_hashsize = MPROF_HASH_SIZE;
175 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD,
176     &mutex_prof_hashsize, 0, "Hash size");
177 static int mutex_prof_collisions = 0;
178 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD,
179     &mutex_prof_collisions, 0, "Number of hash collisions");
180 
181 /*
182  * mprof_mtx protects the profiling buffers and the hash.
183  */
184 static struct mtx mprof_mtx;
185 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET);
186 
187 static u_int64_t
188 nanoseconds(void)
189 {
190 	struct timespec tv;
191 
192 	nanotime(&tv);
193 	return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
194 }
195 
196 static int
197 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
198 {
199 	struct sbuf *sb;
200 	int error, i;
201 	static int multiplier = 1;
202 
203 	if (first_free_mprof_buf == 0)
204 		return (SYSCTL_OUT(req, "No locking recorded",
205 		    sizeof("No locking recorded")));
206 
207 retry_sbufops:
208 	sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
209 	sbuf_printf(sb, "\n%6s %12s %11s %5s %12s %12s %s\n",
210 	    "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name");
211 	/*
212 	 * XXX this spinlock seems to be by far the largest perpetrator
213 	 * of spinlock latency (1.6 msec on an Athlon1600 was recorded
214 	 * even before I pessimized it further by moving the average
215 	 * computation here).
216 	 */
217 	mtx_lock_spin(&mprof_mtx);
218 	for (i = 0; i < first_free_mprof_buf; ++i) {
219 		sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n",
220 		    mprof_buf[i].cnt_max / 1000,
221 		    mprof_buf[i].cnt_tot / 1000,
222 		    mprof_buf[i].cnt_cur,
223 		    mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
224 			mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000),
225 		    mprof_buf[i].cnt_contest_holding,
226 		    mprof_buf[i].cnt_contest_locking,
227 		    mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name);
228 		if (sbuf_overflowed(sb)) {
229 			mtx_unlock_spin(&mprof_mtx);
230 			sbuf_delete(sb);
231 			multiplier++;
232 			goto retry_sbufops;
233 		}
234 	}
235 	mtx_unlock_spin(&mprof_mtx);
236 	sbuf_finish(sb);
237 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
238 	sbuf_delete(sb);
239 	return (error);
240 }
241 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
242     NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics");
243 
244 static int
245 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
246 {
247 	int error, v;
248 
249 	if (first_free_mprof_buf == 0)
250 		return (0);
251 
252 	v = 0;
253 	error = sysctl_handle_int(oidp, &v, 0, req);
254 	if (error)
255 		return (error);
256 	if (req->newptr == NULL)
257 		return (error);
258 	if (v == 0)
259 		return (0);
260 
261 	mtx_lock_spin(&mprof_mtx);
262 	bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf);
263 	bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE);
264 	first_free_mprof_buf = 0;
265 	mtx_unlock_spin(&mprof_mtx);
266 	return (0);
267 }
268 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
269     NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics");
270 #endif
271 
272 /*
273  * Function versions of the inlined __mtx_* macros.  These are used by
274  * modules and can also be called from assembly language if needed.
275  */
276 void
277 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
278 {
279 
280 	MPASS(curthread != NULL);
281 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
282 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
283 	    file, line));
284 	WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
285 	    file, line);
286 	_get_sleep_lock(m, curthread, opts, file, line);
287 	LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
288 	    line);
289 	WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
290 #ifdef MUTEX_PROFILING
291 	/* don't reset the timer when/if recursing */
292 	if (m->mtx_acqtime == 0) {
293 		m->mtx_filename = file;
294 		m->mtx_lineno = line;
295 		m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0;
296 		++mutex_prof_acquisitions;
297 	}
298 #endif
299 }
300 
301 void
302 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
303 {
304 
305 	MPASS(curthread != NULL);
306 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
307 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
308 	    file, line));
309 	WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
310 	LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
311 	    line);
312 	mtx_assert(m, MA_OWNED);
313 #ifdef MUTEX_PROFILING
314 	if (m->mtx_acqtime != 0) {
315 		static const char *unknown = "(unknown)";
316 		struct mutex_prof *mpp;
317 		u_int64_t acqtime, now;
318 		const char *p, *q;
319 		volatile u_int hash;
320 
321 		now = nanoseconds();
322 		acqtime = m->mtx_acqtime;
323 		m->mtx_acqtime = 0;
324 		if (now <= acqtime)
325 			goto out;
326 		for (p = m->mtx_filename;
327 		    p != NULL && strncmp(p, "../", 3) == 0; p += 3)
328 			/* nothing */ ;
329 		if (p == NULL || *p == '\0')
330 			p = unknown;
331 		for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q)
332 			hash = (hash * 2 + *q) % MPROF_HASH_SIZE;
333 		mtx_lock_spin(&mprof_mtx);
334 		for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next)
335 			if (mpp->line == m->mtx_lineno &&
336 			    strcmp(mpp->file, p) == 0)
337 				break;
338 		if (mpp == NULL) {
339 			/* Just exit if we cannot get a trace buffer */
340 			if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) {
341 				++mutex_prof_rejected;
342 				goto unlock;
343 			}
344 			mpp = &mprof_buf[first_free_mprof_buf++];
345 			mpp->name = mtx_name(m);
346 			mpp->file = p;
347 			mpp->line = m->mtx_lineno;
348 			mpp->next = mprof_hash[hash];
349 			if (mprof_hash[hash] != NULL)
350 				++mutex_prof_collisions;
351 			mprof_hash[hash] = mpp;
352 			++mutex_prof_records;
353 		}
354 		/*
355 		 * Record if the mutex has been held longer now than ever
356 		 * before.
357 		 */
358 		if (now - acqtime > mpp->cnt_max)
359 			mpp->cnt_max = now - acqtime;
360 		mpp->cnt_tot += now - acqtime;
361 		mpp->cnt_cur++;
362 		/*
363 		 * There's a small race, really we should cmpxchg
364 		 * 0 with the current value, but that would bill
365 		 * the contention to the wrong lock instance if
366 		 * it followed this also.
367 		 */
368 		mpp->cnt_contest_holding += m->mtx_contest_holding;
369 		m->mtx_contest_holding = 0;
370 		mpp->cnt_contest_locking += m->mtx_contest_locking;
371 		m->mtx_contest_locking = 0;
372 unlock:
373 		mtx_unlock_spin(&mprof_mtx);
374 	}
375 out:
376 #endif
377 	_rel_sleep_lock(m, curthread, opts, file, line);
378 }
379 
380 void
381 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
382 {
383 
384 	MPASS(curthread != NULL);
385 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
386 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
387 	    m->mtx_object.lo_name, file, line));
388 	WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
389 	    file, line);
390 	_get_spin_lock(m, curthread, opts, file, line);
391 	LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
392 	    line);
393 	WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
394 }
395 
396 void
397 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
398 {
399 
400 	MPASS(curthread != NULL);
401 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
402 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
403 	    m->mtx_object.lo_name, file, line));
404 	WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
405 	LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
406 	    line);
407 	mtx_assert(m, MA_OWNED);
408 	_rel_spin_lock(m);
409 }
410 
411 /*
412  * The important part of mtx_trylock{,_flags}()
413  * Tries to acquire lock `m.'  If this function is called on a mutex that
414  * is already owned, it will recursively acquire the lock.
415  */
416 int
417 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
418 {
419 	int rval;
420 
421 	MPASS(curthread != NULL);
422 	KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
423 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
424 	    file, line));
425 
426 	if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) {
427 		m->mtx_recurse++;
428 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
429 		rval = 1;
430 	} else
431 		rval = _obtain_lock(m, (uintptr_t)curthread);
432 
433 	LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line);
434 	if (rval)
435 		WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
436 		    file, line);
437 
438 	return (rval);
439 }
440 
441 /*
442  * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
443  *
444  * We call this if the lock is either contested (i.e. we need to go to
445  * sleep waiting for it), or if we need to recurse on it.
446  */
447 void
448 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
449     int line)
450 {
451 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
452 	struct thread *owner;
453 #endif
454 	uintptr_t v;
455 #ifdef KTR
456 	int cont_logged = 0;
457 #endif
458 #ifdef MUTEX_PROFILING
459 	int contested;
460 #endif
461 
462 	if (mtx_owned(m)) {
463 		KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0,
464 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
465 		    m->mtx_object.lo_name, file, line));
466 		m->mtx_recurse++;
467 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
468 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
469 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
470 		return;
471 	}
472 
473 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
474 		CTR4(KTR_LOCK,
475 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
476 		    m->mtx_object.lo_name, (void *)m->mtx_lock, file, line);
477 
478 #ifdef MUTEX_PROFILING
479 	contested = 0;
480 #endif
481 	while (!_obtain_lock(m, tid)) {
482 #ifdef MUTEX_PROFILING
483 		contested = 1;
484 		atomic_add_int(&m->mtx_contest_holding, 1);
485 #endif
486 		turnstile_lock(&m->mtx_object);
487 		v = m->mtx_lock;
488 
489 		/*
490 		 * Check if the lock has been released while spinning for
491 		 * the turnstile chain lock.
492 		 */
493 		if (v == MTX_UNOWNED) {
494 			turnstile_release(&m->mtx_object);
495 			cpu_spinwait();
496 			continue;
497 		}
498 
499 #ifdef MUTEX_WAKE_ALL
500 		MPASS(v != MTX_CONTESTED);
501 #else
502 		/*
503 		 * The mutex was marked contested on release. This means that
504 		 * there are other threads blocked on it.  Grab ownership of
505 		 * it and propagate its priority to the current thread if
506 		 * necessary.
507 		 */
508 		if (v == MTX_CONTESTED) {
509 			m->mtx_lock = tid | MTX_CONTESTED;
510 			turnstile_claim(&m->mtx_object);
511 			break;
512 		}
513 #endif
514 
515 		/*
516 		 * If the mutex isn't already contested and a failure occurs
517 		 * setting the contested bit, the mutex was either released
518 		 * or the state of the MTX_RECURSED bit changed.
519 		 */
520 		if ((v & MTX_CONTESTED) == 0 &&
521 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
522 			turnstile_release(&m->mtx_object);
523 			cpu_spinwait();
524 			continue;
525 		}
526 
527 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
528 		/*
529 		 * If the current owner of the lock is executing on another
530 		 * CPU, spin instead of blocking.
531 		 */
532 		owner = (struct thread *)(v & MTX_FLAGMASK);
533 #ifdef ADAPTIVE_GIANT
534 		if (TD_IS_RUNNING(owner)) {
535 #else
536 		if (m != &Giant && TD_IS_RUNNING(owner)) {
537 #endif
538 			turnstile_release(&m->mtx_object);
539 			while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
540 				cpu_spinwait();
541 			}
542 			continue;
543 		}
544 #endif	/* SMP && !NO_ADAPTIVE_MUTEXES */
545 
546 		/*
547 		 * We definitely must sleep for this lock.
548 		 */
549 		mtx_assert(m, MA_NOTOWNED);
550 
551 #ifdef KTR
552 		if (!cont_logged) {
553 			CTR6(KTR_CONTENTION,
554 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
555 			    (void *)tid, file, line, m->mtx_object.lo_name,
556 			    WITNESS_FILE(&m->mtx_object),
557 			    WITNESS_LINE(&m->mtx_object));
558 			cont_logged = 1;
559 		}
560 #endif
561 
562 		/*
563 		 * Block on the turnstile.
564 		 */
565 		turnstile_wait(&m->mtx_object, mtx_owner(m));
566 	}
567 
568 #ifdef KTR
569 	if (cont_logged) {
570 		CTR4(KTR_CONTENTION,
571 		    "contention end: %s acquired by %p at %s:%d",
572 		    m->mtx_object.lo_name, (void *)tid, file, line);
573 	}
574 #endif
575 #ifdef MUTEX_PROFILING
576 	if (contested)
577 		m->mtx_contest_locking++;
578 	m->mtx_contest_holding = 0;
579 #endif
580 	return;
581 }
582 
583 #ifdef SMP
584 /*
585  * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
586  *
587  * This is only called if we need to actually spin for the lock. Recursion
588  * is handled inline.
589  */
590 void
591 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file,
592     int line)
593 {
594 	int i = 0;
595 
596 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
597 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
598 
599 	for (;;) {
600 		if (_obtain_lock(m, tid))
601 			break;
602 
603 		/* Give interrupts a chance while we spin. */
604 		spinlock_exit();
605 		while (m->mtx_lock != MTX_UNOWNED) {
606 			if (i++ < 10000000) {
607 				cpu_spinwait();
608 				continue;
609 			}
610 			if (i < 60000000)
611 				DELAY(1);
612 			else if (!kdb_active && !panicstr) {
613 				printf("spin lock %s held by %p for > 5 seconds\n",
614 				    m->mtx_object.lo_name, (void *)m->mtx_lock);
615 #ifdef WITNESS
616 				witness_display_spinlock(&m->mtx_object,
617 				    mtx_owner(m));
618 #endif
619 				panic("spin lock held too long");
620 			}
621 			cpu_spinwait();
622 		}
623 		spinlock_enter();
624 	}
625 
626 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
627 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
628 
629 	return;
630 }
631 #endif /* SMP */
632 
633 /*
634  * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
635  *
636  * We are only called here if the lock is recursed or contested (i.e. we
637  * need to wake up a blocked thread).
638  */
639 void
640 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
641 {
642 	struct turnstile *ts;
643 #ifndef PREEMPTION
644 	struct thread *td, *td1;
645 #endif
646 
647 	if (mtx_recursed(m)) {
648 		if (--(m->mtx_recurse) == 0)
649 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
650 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
651 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
652 		return;
653 	}
654 
655 	turnstile_lock(&m->mtx_object);
656 	ts = turnstile_lookup(&m->mtx_object);
657 	if (LOCK_LOG_TEST(&m->mtx_object, opts))
658 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
659 
660 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
661 	if (ts == NULL) {
662 		_release_lock_quick(m);
663 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
664 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
665 		turnstile_release(&m->mtx_object);
666 		return;
667 	}
668 #else
669 	MPASS(ts != NULL);
670 #endif
671 #ifndef PREEMPTION
672 	/* XXX */
673 	td1 = turnstile_head(ts);
674 #endif
675 #ifdef MUTEX_WAKE_ALL
676 	turnstile_broadcast(ts);
677 	_release_lock_quick(m);
678 #else
679 	if (turnstile_signal(ts)) {
680 		_release_lock_quick(m);
681 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
682 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m);
683 	} else {
684 		m->mtx_lock = MTX_CONTESTED;
685 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
686 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested",
687 			    m);
688 	}
689 #endif
690 	turnstile_unpend(ts);
691 
692 #ifndef PREEMPTION
693 	/*
694 	 * XXX: This is just a hack until preemption is done.  However,
695 	 * once preemption is done we need to either wrap the
696 	 * turnstile_signal() and release of the actual lock in an
697 	 * extra critical section or change the preemption code to
698 	 * always just set a flag and never do instant-preempts.
699 	 */
700 	td = curthread;
701 	if (td->td_critnest > 0 || td1->td_priority >= td->td_priority)
702 		return;
703 	mtx_lock_spin(&sched_lock);
704 	if (!TD_IS_RUNNING(td1)) {
705 #ifdef notyet
706 		if (td->td_ithd != NULL) {
707 			struct ithd *it = td->td_ithd;
708 
709 			if (it->it_interrupted) {
710 				if (LOCK_LOG_TEST(&m->mtx_object, opts))
711 					CTR2(KTR_LOCK,
712 				    "_mtx_unlock_sleep: %p interrupted %p",
713 					    it, it->it_interrupted);
714 				intr_thd_fixup(it);
715 			}
716 		}
717 #endif
718 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
719 			CTR2(KTR_LOCK,
720 			    "_mtx_unlock_sleep: %p switching out lock=%p", m,
721 			    (void *)m->mtx_lock);
722 
723 		mi_switch(SW_INVOL, NULL);
724 		if (LOCK_LOG_TEST(&m->mtx_object, opts))
725 			CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p",
726 			    m, (void *)m->mtx_lock);
727 	}
728 	mtx_unlock_spin(&sched_lock);
729 #endif
730 
731 	return;
732 }
733 
734 /*
735  * All the unlocking of MTX_SPIN locks is done inline.
736  * See the _rel_spin_lock() macro for the details.
737  */
738 
739 /*
740  * The backing function for the INVARIANTS-enabled mtx_assert()
741  */
742 #ifdef INVARIANT_SUPPORT
743 void
744 _mtx_assert(struct mtx *m, int what, const char *file, int line)
745 {
746 
747 	if (panicstr != NULL || dumping)
748 		return;
749 	switch (what) {
750 	case MA_OWNED:
751 	case MA_OWNED | MA_RECURSED:
752 	case MA_OWNED | MA_NOTRECURSED:
753 		if (!mtx_owned(m))
754 			panic("mutex %s not owned at %s:%d",
755 			    m->mtx_object.lo_name, file, line);
756 		if (mtx_recursed(m)) {
757 			if ((what & MA_NOTRECURSED) != 0)
758 				panic("mutex %s recursed at %s:%d",
759 				    m->mtx_object.lo_name, file, line);
760 		} else if ((what & MA_RECURSED) != 0) {
761 			panic("mutex %s unrecursed at %s:%d",
762 			    m->mtx_object.lo_name, file, line);
763 		}
764 		break;
765 	case MA_NOTOWNED:
766 		if (mtx_owned(m))
767 			panic("mutex %s owned at %s:%d",
768 			    m->mtx_object.lo_name, file, line);
769 		break;
770 	default:
771 		panic("unknown mtx_assert at %s:%d", file, line);
772 	}
773 }
774 #endif
775 
776 /*
777  * The MUTEX_DEBUG-enabled mtx_validate()
778  *
779  * Most of these checks have been moved off into the LO_INITIALIZED flag
780  * maintained by the witness code.
781  */
782 #ifdef MUTEX_DEBUG
783 
784 void	mtx_validate(struct mtx *);
785 
786 void
787 mtx_validate(struct mtx *m)
788 {
789 
790 /*
791  * XXX: When kernacc() does not require Giant we can reenable this check
792  */
793 #ifdef notyet
794 /*
795  * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly
796  * we can re-enable the kernacc() checks.
797  */
798 #ifndef __alpha__
799 	/*
800 	 * Can't call kernacc() from early init386(), especially when
801 	 * initializing Giant mutex, because some stuff in kernacc()
802 	 * requires Giant itself.
803 	 */
804 	if (!cold)
805 		if (!kernacc((caddr_t)m, sizeof(m),
806 		    VM_PROT_READ | VM_PROT_WRITE))
807 			panic("Can't read and write to mutex %p", m);
808 #endif
809 #endif
810 }
811 #endif
812 
813 /*
814  * General init routine used by the MTX_SYSINIT() macro.
815  */
816 void
817 mtx_sysinit(void *arg)
818 {
819 	struct mtx_args *margs = arg;
820 
821 	mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
822 }
823 
824 /*
825  * Mutex initialization routine; initialize lock `m' of type contained in
826  * `opts' with options contained in `opts' and name `name.'  The optional
827  * lock type `type' is used as a general lock category name for use with
828  * witness.
829  */
830 void
831 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
832 {
833 	struct lock_object *lock;
834 
835 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
836 	    MTX_NOWITNESS | MTX_DUPOK)) == 0);
837 
838 #ifdef MUTEX_DEBUG
839 	/* Diagnostic and error correction */
840 	mtx_validate(m);
841 #endif
842 
843 	lock = &m->mtx_object;
844 	KASSERT((lock->lo_flags & LO_INITIALIZED) == 0,
845 	    ("mutex \"%s\" %p already initialized", name, m));
846 	bzero(m, sizeof(*m));
847 	if (opts & MTX_SPIN)
848 		lock->lo_class = &lock_class_mtx_spin;
849 	else
850 		lock->lo_class = &lock_class_mtx_sleep;
851 	lock->lo_name = name;
852 	lock->lo_type = type != NULL ? type : name;
853 	if (opts & MTX_QUIET)
854 		lock->lo_flags = LO_QUIET;
855 	if (opts & MTX_RECURSE)
856 		lock->lo_flags |= LO_RECURSABLE;
857 	if ((opts & MTX_NOWITNESS) == 0)
858 		lock->lo_flags |= LO_WITNESS;
859 	if (opts & MTX_DUPOK)
860 		lock->lo_flags |= LO_DUPOK;
861 
862 	m->mtx_lock = MTX_UNOWNED;
863 
864 	LOCK_LOG_INIT(lock, opts);
865 
866 	WITNESS_INIT(lock);
867 }
868 
869 /*
870  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
871  * passed in as a flag here because if the corresponding mtx_init() was
872  * called with MTX_QUIET set, then it will already be set in the mutex's
873  * flags.
874  */
875 void
876 mtx_destroy(struct mtx *m)
877 {
878 
879 	LOCK_LOG_DESTROY(&m->mtx_object, 0);
880 
881 	if (!mtx_owned(m))
882 		MPASS(mtx_unowned(m));
883 	else {
884 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
885 
886 		/* Tell witness this isn't locked to make it happy. */
887 		WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__,
888 		    __LINE__);
889 	}
890 
891 	WITNESS_DESTROY(&m->mtx_object);
892 }
893 
894 /*
895  * Intialize the mutex code and system mutexes.  This is called from the MD
896  * startup code prior to mi_startup().  The per-CPU data space needs to be
897  * setup before this is called.
898  */
899 void
900 mutex_init(void)
901 {
902 
903 	/* Setup thread0 so that mutexes work. */
904 	LIST_INIT(&thread0.td_contested);
905 
906 	/* Setup turnstiles so that sleep mutexes work. */
907 	init_turnstiles();
908 
909 	/*
910 	 * Initialize mutexes.
911 	 */
912 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
913 	mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
914 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
915 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
916 	mtx_lock(&Giant);
917 }
918 
919 #ifdef DDB
920 /* XXX: This function is not mutex-specific. */
921 DB_SHOW_COMMAND(lock, db_show_lock)
922 {
923 	struct lock_object *lock;
924 
925 	if (!have_addr)
926 		return;
927 	lock = (struct lock_object *)addr;
928 	if (lock->lo_class != &lock_class_mtx_sleep &&
929 	    lock->lo_class != &lock_class_mtx_spin &&
930 	    lock->lo_class != &lock_class_sx) {
931 		db_printf("Unknown lock class\n");
932 		return;
933 	}
934 	db_printf(" class: %s\n", lock->lo_class->lc_name);
935 	db_printf(" name: %s\n", lock->lo_name);
936 	if (lock->lo_type && lock->lo_type != lock->lo_name)
937 		db_printf(" type: %s\n", lock->lo_type);
938 	lock->lo_class->lc_ddb_show(lock);
939 }
940 
941 void
942 db_show_mtx(struct lock_object *lock)
943 {
944 	struct thread *td;
945 	struct mtx *m;
946 
947 	m = (struct mtx *)lock;
948 
949 	db_printf(" flags: {");
950 	if (m->mtx_object.lo_class == &lock_class_mtx_spin)
951 		db_printf("SPIN");
952 	else
953 		db_printf("DEF");
954 	if (m->mtx_object.lo_flags & LO_RECURSABLE)
955 		db_printf(", RECURSE");
956 	if (m->mtx_object.lo_flags & LO_DUPOK)
957 		db_printf(", DUPOK");
958 	db_printf("}\n");
959 	db_printf(" state: {");
960 	if (mtx_unowned(m))
961 		db_printf("UNOWNED");
962 	else {
963 		db_printf("OWNED");
964 		if (m->mtx_lock & MTX_CONTESTED)
965 			db_printf(", CONTESTED");
966 		if (m->mtx_lock & MTX_RECURSED)
967 			db_printf(", RECURSED");
968 	}
969 	db_printf("}\n");
970 	if (!mtx_unowned(m)) {
971 		td = mtx_owner(m);
972 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
973 		    td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
974 		if (mtx_recursed(m))
975 			db_printf(" recursed: %d\n", m->mtx_recurse);
976 	}
977 }
978 #endif
979