xref: /NextBSD/sys/kern/kern_mutex.c (revision eec72aa896cd37a39f1061591568ede6f351c188)
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_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 #include "opt_thrworkq.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 #include <sys/lock_profile.h>
63 
64 #include <machine/atomic.h>
65 #include <machine/bus.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 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
76 #define	ADAPTIVE_MUTEXES
77 #endif
78 
79 #ifdef HWPMC_HOOKS
80 #include <sys/pmckern.h>
81 PMC_SOFT_DEFINE( , , lock, failed);
82 #endif
83 
84 /*
85  * Return the mutex address when the lock cookie address is provided.
86  * This functionality assumes that struct mtx* have a member named mtx_lock.
87  */
88 #define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
89 
90 /*
91  * Internal utility macros.
92  */
93 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
94 
95 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
96 
97 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
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 /*
143  * System-wide mutexes
144  */
145 struct mtx blocked_lock;
146 struct mtx Giant;
147 
148 void
assert_mtx(const struct lock_object * lock,int what)149 assert_mtx(const struct lock_object *lock, int what)
150 {
151 
152 	mtx_assert((const struct mtx *)lock, what);
153 }
154 
155 void
lock_mtx(struct lock_object * lock,uintptr_t how)156 lock_mtx(struct lock_object *lock, uintptr_t how)
157 {
158 
159 	mtx_lock((struct mtx *)lock);
160 }
161 
162 void
lock_spin(struct lock_object * lock,uintptr_t how)163 lock_spin(struct lock_object *lock, uintptr_t how)
164 {
165 
166 	panic("spin locks can only use msleep_spin");
167 }
168 
169 uintptr_t
unlock_mtx(struct lock_object * lock)170 unlock_mtx(struct lock_object *lock)
171 {
172 	struct mtx *m;
173 
174 	m = (struct mtx *)lock;
175 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
176 	mtx_unlock(m);
177 	return (0);
178 }
179 
180 uintptr_t
unlock_spin(struct lock_object * lock)181 unlock_spin(struct lock_object *lock)
182 {
183 
184 	panic("spin locks can only use msleep_spin");
185 }
186 
187 #ifdef KDTRACE_HOOKS
188 int
owner_mtx(const struct lock_object * lock,struct thread ** owner)189 owner_mtx(const struct lock_object *lock, struct thread **owner)
190 {
191 	const struct mtx *m = (const struct mtx *)lock;
192 
193 	*owner = mtx_owner(m);
194 	return (mtx_unowned(m) == 0);
195 }
196 #endif
197 
198 /*
199  * Function versions of the inlined __mtx_* macros.  These are used by
200  * modules and can also be called from assembly language if needed.
201  */
202 void
__mtx_lock_flags(volatile uintptr_t * c,int opts,const char * file,int line)203 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
204 {
205 	struct mtx *m;
206 
207 	if (SCHEDULER_STOPPED())
208 		return;
209 
210 	m = mtxlock2mtx(c);
211 
212 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
213 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
214 	    curthread, m->lock_object.lo_name, file, line));
215 	KASSERT(m->mtx_lock != MTX_DESTROYED,
216 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
217 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
218 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
219 	    file, line));
220 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
221 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
222 
223 	__mtx_lock(m, curthread, opts, file, line);
224 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
225 	    line);
226 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
227 	    file, line);
228 	TD_LOCKS_INC(curthread);
229 }
230 
231 void
__mtx_unlock_flags(volatile uintptr_t * c,int opts,const char * file,int line)232 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
233 {
234 	struct mtx *m;
235 
236 	if (SCHEDULER_STOPPED())
237 		return;
238 
239 	m = mtxlock2mtx(c);
240 
241 	KASSERT(m->mtx_lock != MTX_DESTROYED,
242 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
243 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
244 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
245 	    file, line));
246 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
247 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
248 	    line);
249 	mtx_assert(m, MA_OWNED);
250 
251 	__mtx_unlock(m, curthread, opts, file, line);
252 	TD_LOCKS_DEC(curthread);
253 }
254 
255 void
__mtx_lock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)256 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
257     int line)
258 {
259 	struct mtx *m;
260 
261 	if (SCHEDULER_STOPPED())
262 		return;
263 
264 	m = mtxlock2mtx(c);
265 
266 	KASSERT(m->mtx_lock != MTX_DESTROYED,
267 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
268 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
269 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
270 	    m->lock_object.lo_name, file, line));
271 	if (mtx_owned(m))
272 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
273 		    (opts & MTX_RECURSE) != 0,
274 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
275 		    m->lock_object.lo_name, file, line));
276 	opts &= ~MTX_RECURSE;
277 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
278 	    file, line, NULL);
279 	__mtx_lock_spin(m, curthread, opts, file, line);
280 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
281 	    line);
282 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
283 }
284 
285 void
__mtx_unlock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)286 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
287     int line)
288 {
289 	struct mtx *m;
290 
291 	if (SCHEDULER_STOPPED())
292 		return;
293 
294 	m = mtxlock2mtx(c);
295 
296 	KASSERT(m->mtx_lock != MTX_DESTROYED,
297 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
298 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
299 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
300 	    m->lock_object.lo_name, file, line));
301 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
302 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
303 	    line);
304 	mtx_assert(m, MA_OWNED);
305 
306 	__mtx_unlock_spin(m);
307 }
308 
309 /*
310  * The important part of mtx_trylock{,_flags}()
311  * Tries to acquire lock `m.'  If this function is called on a mutex that
312  * is already owned, it will recursively acquire the lock.
313  */
314 int
_mtx_trylock_flags_(volatile uintptr_t * c,int opts,const char * file,int line)315 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
316 {
317 	struct mtx *m;
318 #ifdef LOCK_PROFILING
319 	uint64_t waittime = 0;
320 	int contested = 0;
321 #endif
322 	int rval;
323 
324 	if (SCHEDULER_STOPPED())
325 		return (1);
326 
327 	m = mtxlock2mtx(c);
328 
329 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
330 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
331 	    curthread, m->lock_object.lo_name, file, line));
332 	KASSERT(m->mtx_lock != MTX_DESTROYED,
333 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
334 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
335 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
336 	    file, line));
337 
338 	if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
339 	    (opts & MTX_RECURSE) != 0)) {
340 		m->mtx_recurse++;
341 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
342 		rval = 1;
343 	} else
344 		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
345 	opts &= ~MTX_RECURSE;
346 
347 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
348 	if (rval) {
349 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
350 		    file, line);
351 		TD_LOCKS_INC(curthread);
352 		if (m->mtx_recurse == 0)
353 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
354 			    m, contested, waittime, file, line);
355 
356 	}
357 
358 	return (rval);
359 }
360 
361 /*
362  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
363  *
364  * We call this if the lock is either contested (i.e. we need to go to
365  * sleep waiting for it), or if we need to recurse on it.
366  */
367 void
__mtx_lock_sleep(volatile uintptr_t * c,uintptr_t tid,int opts,const char * file,int line)368 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
369     const char *file, int line)
370 {
371 	struct mtx *m;
372 	struct turnstile *ts;
373 	uintptr_t v;
374 #ifdef ADAPTIVE_MUTEXES
375 	volatile struct thread *owner;
376 #endif
377 #ifdef KTR
378 	int cont_logged = 0;
379 #endif
380 #ifdef LOCK_PROFILING
381 	int contested = 0;
382 	uint64_t waittime = 0;
383 #endif
384 #ifdef KDTRACE_HOOKS
385 	uint64_t spin_cnt = 0;
386 	uint64_t sleep_cnt = 0;
387 	int64_t sleep_time = 0;
388 	int64_t all_time = 0;
389 #endif
390 
391 	if (SCHEDULER_STOPPED())
392 		return;
393 
394 	m = mtxlock2mtx(c);
395 
396 	if (mtx_owned(m)) {
397 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
398 		    (opts & MTX_RECURSE) != 0,
399 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
400 		    m->lock_object.lo_name, file, line));
401 		opts &= ~MTX_RECURSE;
402 		m->mtx_recurse++;
403 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
404 		if (LOCK_LOG_TEST(&m->lock_object, opts))
405 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
406 		return;
407 	}
408 	opts &= ~MTX_RECURSE;
409 
410 #ifdef HWPMC_HOOKS
411 	PMC_SOFT_CALL( , , lock, failed);
412 #endif
413 	lock_profile_obtain_lock_failed(&m->lock_object,
414 		    &contested, &waittime);
415 	if (LOCK_LOG_TEST(&m->lock_object, opts))
416 		CTR4(KTR_LOCK,
417 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
418 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
419 #ifdef KDTRACE_HOOKS
420 	all_time -= lockstat_nsecs(&m->lock_object);
421 #endif
422 
423 	while (!_mtx_obtain_lock(m, tid)) {
424 #ifdef KDTRACE_HOOKS
425 		spin_cnt++;
426 #endif
427 #ifdef ADAPTIVE_MUTEXES
428 		/*
429 		 * If the owner is running on another CPU, spin until the
430 		 * owner stops running or the state of the lock changes.
431 		 */
432 		v = m->mtx_lock;
433 		if (v != MTX_UNOWNED) {
434 			owner = (struct thread *)(v & ~MTX_FLAGMASK);
435 			if (TD_IS_RUNNING(owner)) {
436 				if (LOCK_LOG_TEST(&m->lock_object, 0))
437 					CTR3(KTR_LOCK,
438 					    "%s: spinning on %p held by %p",
439 					    __func__, m, owner);
440 				KTR_STATE1(KTR_SCHED, "thread",
441 				    sched_tdname((struct thread *)tid),
442 				    "spinning", "lockname:\"%s\"",
443 				    m->lock_object.lo_name);
444 				while (mtx_owner(m) == owner &&
445 				    TD_IS_RUNNING(owner)) {
446 					cpu_spinwait();
447 #ifdef KDTRACE_HOOKS
448 					spin_cnt++;
449 #endif
450 				}
451 				KTR_STATE0(KTR_SCHED, "thread",
452 				    sched_tdname((struct thread *)tid),
453 				    "running");
454 				continue;
455 			}
456 		}
457 #endif
458 
459 		ts = turnstile_trywait(&m->lock_object);
460 		v = m->mtx_lock;
461 
462 		/*
463 		 * Check if the lock has been released while spinning for
464 		 * the turnstile chain lock.
465 		 */
466 		if (v == MTX_UNOWNED) {
467 			turnstile_cancel(ts);
468 			continue;
469 		}
470 
471 #ifdef ADAPTIVE_MUTEXES
472 		/*
473 		 * The current lock owner might have started executing
474 		 * on another CPU (or the lock could have changed
475 		 * owners) while we were waiting on the turnstile
476 		 * chain lock.  If so, drop the turnstile lock and try
477 		 * again.
478 		 */
479 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
480 		if (TD_IS_RUNNING(owner)) {
481 			turnstile_cancel(ts);
482 			continue;
483 		}
484 #endif
485 
486 		/*
487 		 * If the mutex isn't already contested and a failure occurs
488 		 * setting the contested bit, the mutex was either released
489 		 * or the state of the MTX_RECURSED bit changed.
490 		 */
491 		if ((v & MTX_CONTESTED) == 0 &&
492 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
493 			turnstile_cancel(ts);
494 			continue;
495 		}
496 
497 		/*
498 		 * We definitely must sleep for this lock.
499 		 */
500 		mtx_assert(m, MA_NOTOWNED);
501 
502 #ifdef KTR
503 		if (!cont_logged) {
504 			CTR6(KTR_CONTENTION,
505 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
506 			    (void *)tid, file, line, m->lock_object.lo_name,
507 			    WITNESS_FILE(&m->lock_object),
508 			    WITNESS_LINE(&m->lock_object));
509 			cont_logged = 1;
510 		}
511 #endif
512 
513 		/*
514 		 * Block on the turnstile.
515 		 */
516 #ifdef KDTRACE_HOOKS
517 		sleep_time -= lockstat_nsecs(&m->lock_object);
518 #endif
519 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
520 #ifdef KDTRACE_HOOKS
521 		sleep_time += lockstat_nsecs(&m->lock_object);
522 		sleep_cnt++;
523 #endif
524 	}
525 #ifdef KDTRACE_HOOKS
526 	all_time += lockstat_nsecs(&m->lock_object);
527 #endif
528 #ifdef KTR
529 	if (cont_logged) {
530 		CTR4(KTR_CONTENTION,
531 		    "contention end: %s acquired by %p at %s:%d",
532 		    m->lock_object.lo_name, (void *)tid, file, line);
533 	}
534 #endif
535 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
536 	    waittime, file, line);
537 #ifdef KDTRACE_HOOKS
538 	if (sleep_time)
539 		LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
540 
541 	/*
542 	 * Only record the loops spinning and not sleeping.
543 	 */
544 	if (spin_cnt > sleep_cnt)
545 		LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
546 #endif
547 }
548 
549 static void
_mtx_lock_spin_failed(struct mtx * m)550 _mtx_lock_spin_failed(struct mtx *m)
551 {
552 	struct thread *td;
553 
554 	td = mtx_owner(m);
555 
556 	/* If the mutex is unlocked, try again. */
557 	if (td == NULL)
558 		return;
559 
560 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
561 	    m, m->lock_object.lo_name, td, td->td_tid);
562 #ifdef WITNESS
563 	witness_display_spinlock(&m->lock_object, td, printf);
564 #endif
565 	panic("spin lock held too long");
566 }
567 
568 #ifdef SMP
569 /*
570  * _mtx_lock_spin_cookie: 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
_mtx_lock_spin_cookie(volatile uintptr_t * c,uintptr_t tid,int opts,const char * file,int line)576 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
577     const char *file, int line)
578 {
579 	struct mtx *m;
580 	int i = 0;
581 #ifdef LOCK_PROFILING
582 	int contested = 0;
583 	uint64_t waittime = 0;
584 #endif
585 #ifdef KDTRACE_HOOKS
586 	int64_t spin_time = 0;
587 #endif
588 
589 	if (SCHEDULER_STOPPED())
590 		return;
591 
592 	m = mtxlock2mtx(c);
593 
594 	if (LOCK_LOG_TEST(&m->lock_object, opts))
595 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
596 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
597 	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
598 
599 #ifdef HWPMC_HOOKS
600 	PMC_SOFT_CALL( , , lock, failed);
601 #endif
602 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
603 #ifdef KDTRACE_HOOKS
604 	spin_time -= lockstat_nsecs(&m->lock_object);
605 #endif
606 	while (!_mtx_obtain_lock(m, tid)) {
607 
608 		/* Give interrupts a chance while we spin. */
609 		spinlock_exit();
610 		while (m->mtx_lock != MTX_UNOWNED) {
611 			if (i++ < 10000000) {
612 				cpu_spinwait();
613 				continue;
614 			}
615 			if (i < 60000000 || kdb_active || panicstr != NULL)
616 				DELAY(1);
617 			else
618 				_mtx_lock_spin_failed(m);
619 			cpu_spinwait();
620 		}
621 		spinlock_enter();
622 	}
623 #ifdef KDTRACE_HOOKS
624 	spin_time += lockstat_nsecs(&m->lock_object);
625 #endif
626 
627 	if (LOCK_LOG_TEST(&m->lock_object, opts))
628 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
629 	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
630 	    "running");
631 
632 #ifdef KDTRACE_HOOKS
633 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
634 	    contested, waittime, file, line);
635 	if (spin_time != 0)
636 		LOCKSTAT_RECORD1(spin__spin, m, spin_time);
637 #endif
638 }
639 #endif /* SMP */
640 
641 void
thread_lock_flags_(struct thread * td,int opts,const char * file,int line)642 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
643 {
644 	struct mtx *m;
645 	uintptr_t tid;
646 	int i;
647 #ifdef LOCK_PROFILING
648 	int contested = 0;
649 	uint64_t waittime = 0;
650 #endif
651 #ifdef KDTRACE_HOOKS
652 	int64_t spin_time = 0;
653 #endif
654 
655 	i = 0;
656 	tid = (uintptr_t)curthread;
657 
658 	if (SCHEDULER_STOPPED())
659 		return;
660 
661 #ifdef KDTRACE_HOOKS
662 	spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
663 #endif
664 	for (;;) {
665 retry:
666 		spinlock_enter();
667 		m = td->td_lock;
668 		KASSERT(m->mtx_lock != MTX_DESTROYED,
669 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
670 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
671 		    ("thread_lock() of sleep mutex %s @ %s:%d",
672 		    m->lock_object.lo_name, file, line));
673 		if (mtx_owned(m))
674 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
675 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
676 			    m->lock_object.lo_name, file, line));
677 		WITNESS_CHECKORDER(&m->lock_object,
678 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
679 		while (!_mtx_obtain_lock(m, tid)) {
680 			if (m->mtx_lock == tid) {
681 				m->mtx_recurse++;
682 				break;
683 			}
684 #ifdef HWPMC_HOOKS
685 			PMC_SOFT_CALL( , , lock, failed);
686 #endif
687 			lock_profile_obtain_lock_failed(&m->lock_object,
688 			    &contested, &waittime);
689 			/* Give interrupts a chance while we spin. */
690 			spinlock_exit();
691 			while (m->mtx_lock != MTX_UNOWNED) {
692 				if (i++ < 10000000)
693 					cpu_spinwait();
694 				else if (i < 60000000 ||
695 				    kdb_active || panicstr != NULL)
696 					DELAY(1);
697 				else
698 					_mtx_lock_spin_failed(m);
699 				cpu_spinwait();
700 				if (m != td->td_lock)
701 					goto retry;
702 			}
703 			spinlock_enter();
704 		}
705 		if (m == td->td_lock)
706 			break;
707 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
708 	}
709 #ifdef KDTRACE_HOOKS
710 	spin_time += lockstat_nsecs(&m->lock_object);
711 #endif
712 	if (m->mtx_recurse == 0)
713 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
714 		    contested, waittime, file, line);
715 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
716 	    line);
717 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
718 	LOCKSTAT_RECORD1(thread__spin, m, spin_time);
719 }
720 
721 struct mtx *
thread_lock_block(struct thread * td)722 thread_lock_block(struct thread *td)
723 {
724 	struct mtx *lock;
725 
726 	THREAD_LOCK_ASSERT(td, MA_OWNED);
727 	lock = td->td_lock;
728 	td->td_lock = &blocked_lock;
729 	mtx_unlock_spin(lock);
730 
731 	return (lock);
732 }
733 
734 void
thread_lock_unblock(struct thread * td,struct mtx * new)735 thread_lock_unblock(struct thread *td, struct mtx *new)
736 {
737 	mtx_assert(new, MA_OWNED);
738 	MPASS(td->td_lock == &blocked_lock);
739 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
740 }
741 
742 void
thread_lock_set(struct thread * td,struct mtx * new)743 thread_lock_set(struct thread *td, struct mtx *new)
744 {
745 	struct mtx *lock;
746 
747 	mtx_assert(new, MA_OWNED);
748 	THREAD_LOCK_ASSERT(td, MA_OWNED);
749 	lock = td->td_lock;
750 	td->td_lock = new;
751 	mtx_unlock_spin(lock);
752 }
753 
754 /*
755  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
756  *
757  * We are only called here if the lock is recursed or contested (i.e. we
758  * need to wake up a blocked thread).
759  */
760 void
__mtx_unlock_sleep(volatile uintptr_t * c,int opts,const char * file,int line)761 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
762 {
763 	struct mtx *m;
764 	struct turnstile *ts;
765 
766 	if (SCHEDULER_STOPPED())
767 		return;
768 
769 	m = mtxlock2mtx(c);
770 
771 	if (mtx_recursed(m)) {
772 		if (--(m->mtx_recurse) == 0)
773 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
774 		if (LOCK_LOG_TEST(&m->lock_object, opts))
775 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
776 		return;
777 	}
778 
779 	/*
780 	 * We have to lock the chain before the turnstile so this turnstile
781 	 * can be removed from the hash list if it is empty.
782 	 */
783 	turnstile_chain_lock(&m->lock_object);
784 	ts = turnstile_lookup(&m->lock_object);
785 	if (LOCK_LOG_TEST(&m->lock_object, opts))
786 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
787 	MPASS(ts != NULL);
788 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
789 	_mtx_release_lock_quick(m);
790 
791 	/*
792 	 * This turnstile is now no longer associated with the mutex.  We can
793 	 * unlock the chain lock so a new turnstile may take it's place.
794 	 */
795 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
796 	turnstile_chain_unlock(&m->lock_object);
797 }
798 
799 /*
800  * All the unlocking of MTX_SPIN locks is done inline.
801  * See the __mtx_unlock_spin() macro for the details.
802  */
803 
804 /*
805  * The backing function for the INVARIANTS-enabled mtx_assert()
806  */
807 #ifdef INVARIANT_SUPPORT
808 void
__mtx_assert(const volatile uintptr_t * c,int what,const char * file,int line)809 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
810 {
811 	const struct mtx *m;
812 
813 	if (panicstr != NULL || dumping)
814 		return;
815 
816 	m = mtxlock2mtx(c);
817 
818 	switch (what) {
819 	case MA_OWNED:
820 	case MA_OWNED | MA_RECURSED:
821 	case MA_OWNED | MA_NOTRECURSED:
822 		if (!mtx_owned(m))
823 			panic("mutex %s not owned at %s:%d",
824 			    m->lock_object.lo_name, file, line);
825 		if (mtx_recursed(m)) {
826 			if ((what & MA_NOTRECURSED) != 0)
827 				panic("mutex %s recursed at %s:%d",
828 				    m->lock_object.lo_name, file, line);
829 		} else if ((what & MA_RECURSED) != 0) {
830 			panic("mutex %s unrecursed at %s:%d",
831 			    m->lock_object.lo_name, file, line);
832 		}
833 		break;
834 	case MA_NOTOWNED:
835 		if (mtx_owned(m))
836 			panic("mutex %s owned at %s:%d",
837 			    m->lock_object.lo_name, file, line);
838 		break;
839 	default:
840 		panic("unknown mtx_assert at %s:%d", file, line);
841 	}
842 }
843 #endif
844 
845 /*
846  * The MUTEX_DEBUG-enabled mtx_validate()
847  *
848  * Most of these checks have been moved off into the LO_INITIALIZED flag
849  * maintained by the witness code.
850  */
851 #ifdef MUTEX_DEBUG
852 
853 void	mtx_validate(struct mtx *);
854 
855 void
mtx_validate(struct mtx * m)856 mtx_validate(struct mtx *m)
857 {
858 
859 /*
860  * XXX: When kernacc() does not require Giant we can reenable this check
861  */
862 #ifdef notyet
863 	/*
864 	 * Can't call kernacc() from early init386(), especially when
865 	 * initializing Giant mutex, because some stuff in kernacc()
866 	 * requires Giant itself.
867 	 */
868 	if (!cold)
869 		if (!kernacc((caddr_t)m, sizeof(m),
870 		    VM_PROT_READ | VM_PROT_WRITE))
871 			panic("Can't read and write to mutex %p", m);
872 #endif
873 }
874 #endif
875 
876 /*
877  * General init routine used by the MTX_SYSINIT() macro.
878  */
879 void
mtx_sysinit(void * arg)880 mtx_sysinit(void *arg)
881 {
882 	struct mtx_args *margs = arg;
883 
884 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
885 	    margs->ma_opts);
886 }
887 
888 /*
889  * Mutex initialization routine; initialize lock `m' of type contained in
890  * `opts' with options contained in `opts' and name `name.'  The optional
891  * lock type `type' is used as a general lock category name for use with
892  * witness.
893  */
894 void
_mtx_init(volatile uintptr_t * c,const char * name,const char * type,int opts)895 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
896 {
897 	struct mtx *m;
898 	struct lock_class *class;
899 	int flags;
900 
901 	m = mtxlock2mtx(c);
902 
903 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
904 	    MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
905 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
906 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
907 	    &m->mtx_lock));
908 
909 #ifdef MUTEX_DEBUG
910 	/* Diagnostic and error correction */
911 	mtx_validate(m);
912 #endif
913 
914 	/* Determine lock class and lock flags. */
915 	if (opts & MTX_SPIN)
916 		class = &lock_class_mtx_spin;
917 	else
918 		class = &lock_class_mtx_sleep;
919 	flags = 0;
920 	if (opts & MTX_QUIET)
921 		flags |= LO_QUIET;
922 	if (opts & MTX_RECURSE)
923 		flags |= LO_RECURSABLE;
924 	if ((opts & MTX_NOWITNESS) == 0)
925 		flags |= LO_WITNESS;
926 	if (opts & MTX_DUPOK)
927 		flags |= LO_DUPOK;
928 	if (opts & MTX_NOPROFILE)
929 		flags |= LO_NOPROFILE;
930 	if (opts & MTX_NEW)
931 		flags |= LO_NEW;
932 
933 	/* Initialize mutex. */
934 	lock_init(&m->lock_object, class, name, type, flags);
935 
936 	m->mtx_lock = MTX_UNOWNED;
937 	m->mtx_recurse = 0;
938 }
939 
940 /*
941  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
942  * passed in as a flag here because if the corresponding mtx_init() was
943  * called with MTX_QUIET set, then it will already be set in the mutex's
944  * flags.
945  */
946 void
_mtx_destroy(volatile uintptr_t * c)947 _mtx_destroy(volatile uintptr_t *c)
948 {
949 	struct mtx *m;
950 
951 	m = mtxlock2mtx(c);
952 
953 	if (!mtx_owned(m))
954 		MPASS(mtx_unowned(m));
955 	else {
956 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
957 
958 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
959 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
960 			spinlock_exit();
961 		else
962 			TD_LOCKS_DEC(curthread);
963 
964 		lock_profile_release_lock(&m->lock_object);
965 		/* Tell witness this isn't locked to make it happy. */
966 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
967 		    __LINE__);
968 	}
969 
970 	m->mtx_lock = MTX_DESTROYED;
971 	lock_destroy(&m->lock_object);
972 }
973 
974 /*
975  * Intialize the mutex code and system mutexes.  This is called from the MD
976  * startup code prior to mi_startup().  The per-CPU data space needs to be
977  * setup before this is called.
978  */
979 void
mutex_init(void)980 mutex_init(void)
981 {
982 
983 	/* Setup turnstiles so that sleep mutexes work. */
984 	init_turnstiles();
985 
986 	/*
987 	 * Initialize mutexes.
988 	 */
989 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
990 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
991 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
992 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
993 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
994 #ifdef THRWORKQ
995 	mtx_init(&proc0.p_twqlock, "thr workq lock", NULL, MTX_DEF | MTX_DUPOK);
996 	proc0.p_twq = NULL;
997 #endif /* THRWORKQ */
998 	mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
999 	mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1000 	mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1001 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1002 	mtx_lock(&Giant);
1003 }
1004 
1005 #ifdef DDB
1006 void
db_show_mtx(const struct lock_object * lock)1007 db_show_mtx(const struct lock_object *lock)
1008 {
1009 	struct thread *td;
1010 	const struct mtx *m;
1011 
1012 	m = (const struct mtx *)lock;
1013 
1014 	db_printf(" flags: {");
1015 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1016 		db_printf("SPIN");
1017 	else
1018 		db_printf("DEF");
1019 	if (m->lock_object.lo_flags & LO_RECURSABLE)
1020 		db_printf(", RECURSE");
1021 	if (m->lock_object.lo_flags & LO_DUPOK)
1022 		db_printf(", DUPOK");
1023 	db_printf("}\n");
1024 	db_printf(" state: {");
1025 	if (mtx_unowned(m))
1026 		db_printf("UNOWNED");
1027 	else if (mtx_destroyed(m))
1028 		db_printf("DESTROYED");
1029 	else {
1030 		db_printf("OWNED");
1031 		if (m->mtx_lock & MTX_CONTESTED)
1032 			db_printf(", CONTESTED");
1033 		if (m->mtx_lock & MTX_RECURSED)
1034 			db_printf(", RECURSED");
1035 	}
1036 	db_printf("}\n");
1037 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1038 		td = mtx_owner(m);
1039 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1040 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1041 		if (mtx_recursed(m))
1042 			db_printf(" recursed: %d\n", m->mtx_recurse);
1043 	}
1044 }
1045 #endif
1046