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: stable/9/sys/kern/kern_mutex.c 285760 2015-07-21 17:19:03Z markj $");
38 
39 #include "opt_adaptive_mutexes.h"
40 #include "opt_ddb.h"
41 #include "opt_global.h"
42 #include "opt_hwpmc_hooks.h"
43 #include "opt_kdtrace.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/bus.h>
49 #include <sys/conf.h>
50 #include <sys/kdb.h>
51 #include <sys/kernel.h>
52 #include <sys/ktr.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mutex.h>
56 #include <sys/proc.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sched.h>
59 #include <sys/sbuf.h>
60 #include <sys/sysctl.h>
61 #include <sys/turnstile.h>
62 #include <sys/vmmeter.h>
63 #include <sys/lock_profile.h>
64 
65 #include <machine/atomic.h>
66 #include <machine/bus.h>
67 #include <machine/cpu.h>
68 
69 #include <ddb/ddb.h>
70 
71 #include <fs/devfs/devfs_int.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_extern.h>
75 
76 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
77 #define	ADAPTIVE_MUTEXES
78 #endif
79 
80 #ifdef HWPMC_HOOKS
81 #include <sys/pmckern.h>
82 PMC_SOFT_DEFINE( , , lock, failed);
83 #endif
84 
85 /*
86  * Internal utility macros.
87  */
88 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
89 
90 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
91 
92 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
93 
94 static void	assert_mtx(struct lock_object *lock, int what);
95 #ifdef DDB
96 static void	db_show_mtx(struct lock_object *lock);
97 #endif
98 static void	lock_mtx(struct lock_object *lock, int how);
99 static void	lock_spin(struct lock_object *lock, int how);
100 #ifdef KDTRACE_HOOKS
101 static int	owner_mtx(struct lock_object *lock, struct thread **owner);
102 #endif
103 static int	unlock_mtx(struct lock_object *lock);
104 static int	unlock_spin(struct lock_object *lock);
105 
106 /*
107  * Lock classes for sleep and spin mutexes.
108  */
109 struct lock_class lock_class_mtx_sleep = {
110 	.lc_name = "sleep mutex",
111 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
112 	.lc_assert = assert_mtx,
113 #ifdef DDB
114 	.lc_ddb_show = db_show_mtx,
115 #endif
116 	.lc_lock = lock_mtx,
117 	.lc_unlock = unlock_mtx,
118 #ifdef KDTRACE_HOOKS
119 	.lc_owner = owner_mtx,
120 #endif
121 };
122 struct lock_class lock_class_mtx_spin = {
123 	.lc_name = "spin mutex",
124 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
125 	.lc_assert = assert_mtx,
126 #ifdef DDB
127 	.lc_ddb_show = db_show_mtx,
128 #endif
129 	.lc_lock = lock_spin,
130 	.lc_unlock = unlock_spin,
131 #ifdef KDTRACE_HOOKS
132 	.lc_owner = owner_mtx,
133 #endif
134 };
135 
136 /*
137  * System-wide mutexes
138  */
139 struct mtx blocked_lock;
140 struct mtx Giant;
141 
142 void
assert_mtx(struct lock_object * lock,int what)143 assert_mtx(struct lock_object *lock, int what)
144 {
145 
146 	mtx_assert((struct mtx *)lock, what);
147 }
148 
149 void
lock_mtx(struct lock_object * lock,int how)150 lock_mtx(struct lock_object *lock, int how)
151 {
152 
153 	mtx_lock((struct mtx *)lock);
154 }
155 
156 void
lock_spin(struct lock_object * lock,int how)157 lock_spin(struct lock_object *lock, int how)
158 {
159 
160 	panic("spin locks can only use msleep_spin");
161 }
162 
163 int
unlock_mtx(struct lock_object * lock)164 unlock_mtx(struct lock_object *lock)
165 {
166 	struct mtx *m;
167 
168 	m = (struct mtx *)lock;
169 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
170 	mtx_unlock(m);
171 	return (0);
172 }
173 
174 int
unlock_spin(struct lock_object * lock)175 unlock_spin(struct lock_object *lock)
176 {
177 
178 	panic("spin locks can only use msleep_spin");
179 }
180 
181 #ifdef KDTRACE_HOOKS
182 int
owner_mtx(struct lock_object * lock,struct thread ** owner)183 owner_mtx(struct lock_object *lock, struct thread **owner)
184 {
185 	struct mtx *m = (struct mtx *)lock;
186 
187 	*owner = mtx_owner(m);
188 	return (mtx_unowned(m) == 0);
189 }
190 #endif
191 
192 /*
193  * Function versions of the inlined __mtx_* macros.  These are used by
194  * modules and can also be called from assembly language if needed.
195  */
196 void
_mtx_lock_flags(struct mtx * m,int opts,const char * file,int line)197 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
198 {
199 
200 	if (SCHEDULER_STOPPED())
201 		return;
202 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
203 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
204 	    curthread, m->lock_object.lo_name, file, line));
205 	KASSERT(m->mtx_lock != MTX_DESTROYED,
206 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
207 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
208 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
209 	    file, line));
210 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
211 	    file, line, NULL);
212 
213 	__mtx_lock(m, curthread, opts, file, line);
214 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
215 	    line);
216 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
217 	curthread->td_locks++;
218 }
219 
220 void
_mtx_unlock_flags(struct mtx * m,int opts,const char * file,int line)221 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
222 {
223 
224 	if (SCHEDULER_STOPPED())
225 		return;
226 	KASSERT(m->mtx_lock != MTX_DESTROYED,
227 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
228 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
229 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
230 	    file, line));
231 	curthread->td_locks--;
232 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
233 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
234 	    line);
235 	mtx_assert(m, MA_OWNED);
236 
237 	if (m->mtx_recurse == 0)
238 		LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m);
239 	__mtx_unlock(m, curthread, opts, file, line);
240 }
241 
242 void
_mtx_lock_spin_flags(struct mtx * m,int opts,const char * file,int line)243 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
244 {
245 
246 	if (SCHEDULER_STOPPED())
247 		return;
248 	KASSERT(m->mtx_lock != MTX_DESTROYED,
249 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
250 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
251 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
252 	    m->lock_object.lo_name, file, line));
253 	if (mtx_owned(m))
254 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
255 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
256 		    m->lock_object.lo_name, file, line));
257 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
258 	    file, line, NULL);
259 	__mtx_lock_spin(m, curthread, opts, file, line);
260 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
261 	    line);
262 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
263 }
264 
265 void
_mtx_unlock_spin_flags(struct mtx * m,int opts,const char * file,int line)266 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
267 {
268 
269 	if (SCHEDULER_STOPPED())
270 		return;
271 	KASSERT(m->mtx_lock != MTX_DESTROYED,
272 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
273 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
274 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
275 	    m->lock_object.lo_name, file, line));
276 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
277 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
278 	    line);
279 	mtx_assert(m, MA_OWNED);
280 
281 	__mtx_unlock_spin(m);
282 }
283 
284 /*
285  * The important part of mtx_trylock{,_flags}()
286  * Tries to acquire lock `m.'  If this function is called on a mutex that
287  * is already owned, it will recursively acquire the lock.
288  */
289 int
_mtx_trylock(struct mtx * m,int opts,const char * file,int line)290 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
291 {
292 #ifdef LOCK_PROFILING
293 	uint64_t waittime = 0;
294 	int contested = 0;
295 #endif
296 	int rval;
297 
298 	if (SCHEDULER_STOPPED())
299 		return (1);
300 
301 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
302 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
303 	    curthread, m->lock_object.lo_name, file, line));
304 	KASSERT(m->mtx_lock != MTX_DESTROYED,
305 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
306 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
307 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
308 	    file, line));
309 
310 	if (mtx_owned(m) && (m->lock_object.lo_flags & LO_RECURSABLE) != 0) {
311 		m->mtx_recurse++;
312 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
313 		rval = 1;
314 	} else
315 		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
316 
317 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
318 	if (rval) {
319 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
320 		    file, line);
321 		curthread->td_locks++;
322 		if (m->mtx_recurse == 0)
323 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
324 			    m, contested, waittime, file, line);
325 
326 	}
327 
328 	return (rval);
329 }
330 
331 /*
332  * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
333  *
334  * We call this if the lock is either contested (i.e. we need to go to
335  * sleep waiting for it), or if we need to recurse on it.
336  */
337 void
_mtx_lock_sleep(struct mtx * m,uintptr_t tid,int opts,const char * file,int line)338 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
339     int line)
340 {
341 	struct turnstile *ts;
342 	uintptr_t v;
343 #ifdef ADAPTIVE_MUTEXES
344 	volatile struct thread *owner;
345 #endif
346 #ifdef KTR
347 	int cont_logged = 0;
348 #endif
349 #ifdef LOCK_PROFILING
350 	int contested = 0;
351 	uint64_t waittime = 0;
352 #endif
353 #ifdef KDTRACE_HOOKS
354 	uint64_t spin_cnt = 0;
355 	uint64_t sleep_cnt = 0;
356 	int64_t sleep_time = 0;
357 	int64_t all_time = 0;
358 #endif
359 
360 	if (SCHEDULER_STOPPED())
361 		return;
362 
363 	if (mtx_owned(m)) {
364 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
365 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
366 		    m->lock_object.lo_name, file, line));
367 		m->mtx_recurse++;
368 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
369 		if (LOCK_LOG_TEST(&m->lock_object, opts))
370 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
371 		return;
372 	}
373 
374 #ifdef HWPMC_HOOKS
375 	PMC_SOFT_CALL( , , lock, failed);
376 #endif
377 	lock_profile_obtain_lock_failed(&m->lock_object,
378 		    &contested, &waittime);
379 	if (LOCK_LOG_TEST(&m->lock_object, opts))
380 		CTR4(KTR_LOCK,
381 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
382 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
383 #ifdef KDTRACE_HOOKS
384 	all_time -= lockstat_nsecs(&m->lock_object);
385 #endif
386 
387 	while (!_mtx_obtain_lock(m, tid)) {
388 #ifdef KDTRACE_HOOKS
389 		spin_cnt++;
390 #endif
391 #ifdef ADAPTIVE_MUTEXES
392 		/*
393 		 * If the owner is running on another CPU, spin until the
394 		 * owner stops running or the state of the lock changes.
395 		 */
396 		v = m->mtx_lock;
397 		if (v != MTX_UNOWNED) {
398 			owner = (struct thread *)(v & ~MTX_FLAGMASK);
399 			if (TD_IS_RUNNING(owner)) {
400 				if (LOCK_LOG_TEST(&m->lock_object, 0))
401 					CTR3(KTR_LOCK,
402 					    "%s: spinning on %p held by %p",
403 					    __func__, m, owner);
404 				while (mtx_owner(m) == owner &&
405 				    TD_IS_RUNNING(owner)) {
406 					cpu_spinwait();
407 #ifdef KDTRACE_HOOKS
408 					spin_cnt++;
409 #endif
410 				}
411 				continue;
412 			}
413 		}
414 #endif
415 
416 		ts = turnstile_trywait(&m->lock_object);
417 		v = m->mtx_lock;
418 
419 		/*
420 		 * Check if the lock has been released while spinning for
421 		 * the turnstile chain lock.
422 		 */
423 		if (v == MTX_UNOWNED) {
424 			turnstile_cancel(ts);
425 			continue;
426 		}
427 
428 #ifdef ADAPTIVE_MUTEXES
429 		/*
430 		 * The current lock owner might have started executing
431 		 * on another CPU (or the lock could have changed
432 		 * owners) while we were waiting on the turnstile
433 		 * chain lock.  If so, drop the turnstile lock and try
434 		 * again.
435 		 */
436 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
437 		if (TD_IS_RUNNING(owner)) {
438 			turnstile_cancel(ts);
439 			continue;
440 		}
441 #endif
442 
443 		/*
444 		 * If the mutex isn't already contested and a failure occurs
445 		 * setting the contested bit, the mutex was either released
446 		 * or the state of the MTX_RECURSED bit changed.
447 		 */
448 		if ((v & MTX_CONTESTED) == 0 &&
449 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
450 			turnstile_cancel(ts);
451 			continue;
452 		}
453 
454 		/*
455 		 * We definitely must sleep for this lock.
456 		 */
457 		mtx_assert(m, MA_NOTOWNED);
458 
459 #ifdef KTR
460 		if (!cont_logged) {
461 			CTR6(KTR_CONTENTION,
462 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
463 			    (void *)tid, file, line, m->lock_object.lo_name,
464 			    WITNESS_FILE(&m->lock_object),
465 			    WITNESS_LINE(&m->lock_object));
466 			cont_logged = 1;
467 		}
468 #endif
469 
470 		/*
471 		 * Block on the turnstile.
472 		 */
473 #ifdef KDTRACE_HOOKS
474 		sleep_time -= lockstat_nsecs(&m->lock_object);
475 #endif
476 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
477 #ifdef KDTRACE_HOOKS
478 		sleep_time += lockstat_nsecs(&m->lock_object);
479 		sleep_cnt++;
480 #endif
481 	}
482 #ifdef KDTRACE_HOOKS
483 	all_time += lockstat_nsecs(&m->lock_object);
484 #endif
485 #ifdef KTR
486 	if (cont_logged) {
487 		CTR4(KTR_CONTENTION,
488 		    "contention end: %s acquired by %p at %s:%d",
489 		    m->lock_object.lo_name, (void *)tid, file, line);
490 	}
491 #endif
492 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
493 	    waittime, file, line);
494 #ifdef KDTRACE_HOOKS
495 	if (sleep_time)
496 		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
497 
498 	/*
499 	 * Only record the loops spinning and not sleeping.
500 	 */
501 	if (spin_cnt > sleep_cnt)
502 		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (all_time - sleep_time));
503 #endif
504 }
505 
506 static void
_mtx_lock_spin_failed(struct mtx * m)507 _mtx_lock_spin_failed(struct mtx *m)
508 {
509 	struct thread *td;
510 
511 	td = mtx_owner(m);
512 
513 	/* If the mutex is unlocked, try again. */
514 	if (td == NULL)
515 		return;
516 
517 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
518 	    m, m->lock_object.lo_name, td, td->td_tid);
519 #ifdef WITNESS
520 	witness_display_spinlock(&m->lock_object, td, printf);
521 #endif
522 	panic("spin lock held too long");
523 }
524 
525 #ifdef SMP
526 /*
527  * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
528  *
529  * This is only called if we need to actually spin for the lock. Recursion
530  * is handled inline.
531  */
532 void
_mtx_lock_spin(struct mtx * m,uintptr_t tid,int opts,const char * file,int line)533 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file,
534     int line)
535 {
536 	int i = 0;
537 #ifdef LOCK_PROFILING
538 	int contested = 0;
539 	uint64_t waittime = 0;
540 #endif
541 #ifdef KDTRACE_HOOKS
542 	int64_t spin_time = 0;
543 #endif
544 
545 	if (SCHEDULER_STOPPED())
546 		return;
547 
548 	if (LOCK_LOG_TEST(&m->lock_object, opts))
549 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
550 
551 #ifdef HWPMC_HOOKS
552 	PMC_SOFT_CALL( , , lock, failed);
553 #endif
554 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
555 #ifdef KDTRACE_HOOKS
556 	spin_time -= lockstat_nsecs(&m->lock_object);
557 #endif
558 	while (!_mtx_obtain_lock(m, tid)) {
559 
560 		/* Give interrupts a chance while we spin. */
561 		spinlock_exit();
562 		while (m->mtx_lock != MTX_UNOWNED) {
563 			if (i++ < 10000000) {
564 				cpu_spinwait();
565 				continue;
566 			}
567 			if (i < 60000000 || kdb_active || panicstr != NULL)
568 				DELAY(1);
569 			else
570 				_mtx_lock_spin_failed(m);
571 			cpu_spinwait();
572 		}
573 		spinlock_enter();
574 	}
575 #ifdef KDTRACE_HOOKS
576 	spin_time += lockstat_nsecs(&m->lock_object);
577 #endif
578 
579 	if (LOCK_LOG_TEST(&m->lock_object, opts))
580 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
581 
582 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
583 	    contested, waittime, (file), (line));
584 #ifdef KDTRACE_HOOKS
585 	if (spin_time != 0)
586 		LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, spin_time);
587 #endif
588 }
589 #endif /* SMP */
590 
591 void
_thread_lock_flags(struct thread * td,int opts,const char * file,int line)592 _thread_lock_flags(struct thread *td, int opts, const char *file, int line)
593 {
594 	struct mtx *m;
595 	uintptr_t tid;
596 	int i;
597 #ifdef LOCK_PROFILING
598 	int contested = 0;
599 	uint64_t waittime = 0;
600 #endif
601 #ifdef KDTRACE_HOOKS
602 	int64_t spin_time = 0;
603 #endif
604 
605 	i = 0;
606 	tid = (uintptr_t)curthread;
607 
608 	if (SCHEDULER_STOPPED())
609 		return;
610 
611 #ifdef KDTRACE_HOOKS
612 	spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
613 #endif
614 	for (;;) {
615 retry:
616 		spinlock_enter();
617 		m = td->td_lock;
618 		KASSERT(m->mtx_lock != MTX_DESTROYED,
619 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
620 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
621 		    ("thread_lock() of sleep mutex %s @ %s:%d",
622 		    m->lock_object.lo_name, file, line));
623 		if (mtx_owned(m))
624 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
625 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
626 			    m->lock_object.lo_name, file, line));
627 		WITNESS_CHECKORDER(&m->lock_object,
628 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
629 		while (!_mtx_obtain_lock(m, tid)) {
630 			if (m->mtx_lock == tid) {
631 				m->mtx_recurse++;
632 				break;
633 			}
634 #ifdef HWPMC_HOOKS
635 			PMC_SOFT_CALL( , , lock, failed);
636 #endif
637 			lock_profile_obtain_lock_failed(&m->lock_object,
638 			    &contested, &waittime);
639 			/* Give interrupts a chance while we spin. */
640 			spinlock_exit();
641 			while (m->mtx_lock != MTX_UNOWNED) {
642 				if (i++ < 10000000)
643 					cpu_spinwait();
644 				else if (i < 60000000 ||
645 				    kdb_active || panicstr != NULL)
646 					DELAY(1);
647 				else
648 					_mtx_lock_spin_failed(m);
649 				cpu_spinwait();
650 				if (m != td->td_lock)
651 					goto retry;
652 			}
653 			spinlock_enter();
654 		}
655 		if (m == td->td_lock)
656 			break;
657 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
658 	}
659 #ifdef KDTRACE_HOOKS
660 	spin_time += lockstat_nsecs(&m->lock_object);
661 #endif
662 	if (m->mtx_recurse == 0)
663 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
664 		    m, contested, waittime, (file), (line));
665 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
666 	    line);
667 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
668 	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_time);
669 }
670 
671 struct mtx *
thread_lock_block(struct thread * td)672 thread_lock_block(struct thread *td)
673 {
674 	struct mtx *lock;
675 
676 	THREAD_LOCK_ASSERT(td, MA_OWNED);
677 	lock = td->td_lock;
678 	td->td_lock = &blocked_lock;
679 	mtx_unlock_spin(lock);
680 
681 	return (lock);
682 }
683 
684 void
thread_lock_unblock(struct thread * td,struct mtx * new)685 thread_lock_unblock(struct thread *td, struct mtx *new)
686 {
687 	mtx_assert(new, MA_OWNED);
688 	MPASS(td->td_lock == &blocked_lock);
689 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
690 }
691 
692 void
thread_lock_set(struct thread * td,struct mtx * new)693 thread_lock_set(struct thread *td, struct mtx *new)
694 {
695 	struct mtx *lock;
696 
697 	mtx_assert(new, MA_OWNED);
698 	THREAD_LOCK_ASSERT(td, MA_OWNED);
699 	lock = td->td_lock;
700 	td->td_lock = new;
701 	mtx_unlock_spin(lock);
702 }
703 
704 /*
705  * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
706  *
707  * We are only called here if the lock is recursed or contested (i.e. we
708  * need to wake up a blocked thread).
709  */
710 void
_mtx_unlock_sleep(struct mtx * m,int opts,const char * file,int line)711 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
712 {
713 	struct turnstile *ts;
714 
715 	if (SCHEDULER_STOPPED())
716 		return;
717 
718 	if (mtx_recursed(m)) {
719 		if (--(m->mtx_recurse) == 0)
720 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
721 		if (LOCK_LOG_TEST(&m->lock_object, opts))
722 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
723 		return;
724 	}
725 
726 	/*
727 	 * We have to lock the chain before the turnstile so this turnstile
728 	 * can be removed from the hash list if it is empty.
729 	 */
730 	turnstile_chain_lock(&m->lock_object);
731 	ts = turnstile_lookup(&m->lock_object);
732 	if (LOCK_LOG_TEST(&m->lock_object, opts))
733 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
734 	MPASS(ts != NULL);
735 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
736 	_mtx_release_lock_quick(m);
737 
738 	/*
739 	 * This turnstile is now no longer associated with the mutex.  We can
740 	 * unlock the chain lock so a new turnstile may take it's place.
741 	 */
742 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
743 	turnstile_chain_unlock(&m->lock_object);
744 }
745 
746 /*
747  * All the unlocking of MTX_SPIN locks is done inline.
748  * See the __mtx_unlock_spin() macro for the details.
749  */
750 
751 /*
752  * The backing function for the INVARIANTS-enabled mtx_assert()
753  */
754 #ifdef INVARIANT_SUPPORT
755 void
_mtx_assert(struct mtx * m,int what,const char * file,int line)756 _mtx_assert(struct mtx *m, int what, const char *file, int line)
757 {
758 
759 	if (panicstr != NULL || dumping)
760 		return;
761 	switch (what) {
762 	case MA_OWNED:
763 	case MA_OWNED | MA_RECURSED:
764 	case MA_OWNED | MA_NOTRECURSED:
765 		if (!mtx_owned(m))
766 			panic("mutex %s not owned at %s:%d",
767 			    m->lock_object.lo_name, file, line);
768 		if (mtx_recursed(m)) {
769 			if ((what & MA_NOTRECURSED) != 0)
770 				panic("mutex %s recursed at %s:%d",
771 				    m->lock_object.lo_name, file, line);
772 		} else if ((what & MA_RECURSED) != 0) {
773 			panic("mutex %s unrecursed at %s:%d",
774 			    m->lock_object.lo_name, file, line);
775 		}
776 		break;
777 	case MA_NOTOWNED:
778 		if (mtx_owned(m))
779 			panic("mutex %s owned at %s:%d",
780 			    m->lock_object.lo_name, file, line);
781 		break;
782 	default:
783 		panic("unknown mtx_assert at %s:%d", file, line);
784 	}
785 }
786 #endif
787 
788 /*
789  * The MUTEX_DEBUG-enabled mtx_validate()
790  *
791  * Most of these checks have been moved off into the LO_INITIALIZED flag
792  * maintained by the witness code.
793  */
794 #ifdef MUTEX_DEBUG
795 
796 void	mtx_validate(struct mtx *);
797 
798 void
mtx_validate(struct mtx * m)799 mtx_validate(struct mtx *m)
800 {
801 
802 /*
803  * XXX: When kernacc() does not require Giant we can reenable this check
804  */
805 #ifdef notyet
806 	/*
807 	 * Can't call kernacc() from early init386(), especially when
808 	 * initializing Giant mutex, because some stuff in kernacc()
809 	 * requires Giant itself.
810 	 */
811 	if (!cold)
812 		if (!kernacc((caddr_t)m, sizeof(m),
813 		    VM_PROT_READ | VM_PROT_WRITE))
814 			panic("Can't read and write to mutex %p", m);
815 #endif
816 }
817 #endif
818 
819 /*
820  * General init routine used by the MTX_SYSINIT() macro.
821  */
822 void
mtx_sysinit(void * arg)823 mtx_sysinit(void *arg)
824 {
825 	struct mtx_args *margs = arg;
826 
827 	mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
828 }
829 
830 /*
831  * Mutex initialization routine; initialize lock `m' of type contained in
832  * `opts' with options contained in `opts' and name `name.'  The optional
833  * lock type `type' is used as a general lock category name for use with
834  * witness.
835  */
836 void
mtx_init(struct mtx * m,const char * name,const char * type,int opts)837 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
838 {
839 	struct lock_class *class;
840 	int flags;
841 
842 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
843 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
844 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
845 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
846 	    &m->mtx_lock));
847 
848 #ifdef MUTEX_DEBUG
849 	/* Diagnostic and error correction */
850 	mtx_validate(m);
851 #endif
852 
853 	/* Determine lock class and lock flags. */
854 	if (opts & MTX_SPIN)
855 		class = &lock_class_mtx_spin;
856 	else
857 		class = &lock_class_mtx_sleep;
858 	flags = 0;
859 	if (opts & MTX_QUIET)
860 		flags |= LO_QUIET;
861 	if (opts & MTX_RECURSE)
862 		flags |= LO_RECURSABLE;
863 	if ((opts & MTX_NOWITNESS) == 0)
864 		flags |= LO_WITNESS;
865 	if (opts & MTX_DUPOK)
866 		flags |= LO_DUPOK;
867 	if (opts & MTX_NOPROFILE)
868 		flags |= LO_NOPROFILE;
869 
870 	/* Initialize mutex. */
871 	m->mtx_lock = MTX_UNOWNED;
872 	m->mtx_recurse = 0;
873 
874 	lock_init(&m->lock_object, class, name, type, flags);
875 }
876 
877 /*
878  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
879  * passed in as a flag here because if the corresponding mtx_init() was
880  * called with MTX_QUIET set, then it will already be set in the mutex's
881  * flags.
882  */
883 void
mtx_destroy(struct mtx * m)884 mtx_destroy(struct mtx *m)
885 {
886 
887 	if (!mtx_owned(m))
888 		MPASS(mtx_unowned(m));
889 	else {
890 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
891 
892 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
893 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
894 			spinlock_exit();
895 		else
896 			curthread->td_locks--;
897 
898 		lock_profile_release_lock(&m->lock_object);
899 		/* Tell witness this isn't locked to make it happy. */
900 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
901 		    __LINE__);
902 	}
903 
904 	m->mtx_lock = MTX_DESTROYED;
905 	lock_destroy(&m->lock_object);
906 }
907 
908 /*
909  * Intialize the mutex code and system mutexes.  This is called from the MD
910  * startup code prior to mi_startup().  The per-CPU data space needs to be
911  * setup before this is called.
912  */
913 void
mutex_init(void)914 mutex_init(void)
915 {
916 
917 	/* Setup turnstiles so that sleep mutexes work. */
918 	init_turnstiles();
919 
920 	/*
921 	 * Initialize mutexes.
922 	 */
923 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
924 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
925 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
926 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
927 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
928 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
929 	mtx_lock(&Giant);
930 }
931 
932 #ifdef DDB
933 void
db_show_mtx(struct lock_object * lock)934 db_show_mtx(struct lock_object *lock)
935 {
936 	struct thread *td;
937 	struct mtx *m;
938 
939 	m = (struct mtx *)lock;
940 
941 	db_printf(" flags: {");
942 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
943 		db_printf("SPIN");
944 	else
945 		db_printf("DEF");
946 	if (m->lock_object.lo_flags & LO_RECURSABLE)
947 		db_printf(", RECURSE");
948 	if (m->lock_object.lo_flags & LO_DUPOK)
949 		db_printf(", DUPOK");
950 	db_printf("}\n");
951 	db_printf(" state: {");
952 	if (mtx_unowned(m))
953 		db_printf("UNOWNED");
954 	else if (mtx_destroyed(m))
955 		db_printf("DESTROYED");
956 	else {
957 		db_printf("OWNED");
958 		if (m->mtx_lock & MTX_CONTESTED)
959 			db_printf(", CONTESTED");
960 		if (m->mtx_lock & MTX_RECURSED)
961 			db_printf(", RECURSED");
962 	}
963 	db_printf("}\n");
964 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
965 		td = mtx_owner(m);
966 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
967 		    td->td_tid, td->td_proc->p_pid, td->td_name);
968 		if (mtx_recursed(m))
969 			db_printf(" recursed: %d\n", m->mtx_recurse);
970 	}
971 }
972 #endif
973