xref: /trueos/sys/kern/kern_mutex.c (revision f3fa4bdf8b98edb697d801e65b8b2bd542f15787)
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_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  * Return the mutex address when the lock cookie address is provided.
87  * This functionality assumes that struct mtx* have a member named mtx_lock.
88  */
89 #define	mtxlock2mtx(c)	(__containerof(c, struct mtx, mtx_lock))
90 
91 /*
92  * Internal utility macros.
93  */
94 #define mtx_unowned(m)	((m)->mtx_lock == MTX_UNOWNED)
95 
96 #define	mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
97 
98 #define	mtx_owner(m)	((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
99 
100 static void	assert_mtx(const struct lock_object *lock, int what);
101 #ifdef DDB
102 static void	db_show_mtx(const struct lock_object *lock);
103 #endif
104 static void	lock_mtx(struct lock_object *lock, uintptr_t how);
105 static void	lock_spin(struct lock_object *lock, uintptr_t how);
106 #ifdef KDTRACE_HOOKS
107 static int	owner_mtx(const struct lock_object *lock,
108 		    struct thread **owner);
109 #endif
110 static uintptr_t unlock_mtx(struct lock_object *lock);
111 static uintptr_t unlock_spin(struct lock_object *lock);
112 
113 /*
114  * Lock classes for sleep and spin mutexes.
115  */
116 struct lock_class lock_class_mtx_sleep = {
117 	.lc_name = "sleep mutex",
118 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
119 	.lc_assert = assert_mtx,
120 #ifdef DDB
121 	.lc_ddb_show = db_show_mtx,
122 #endif
123 	.lc_lock = lock_mtx,
124 	.lc_unlock = unlock_mtx,
125 #ifdef KDTRACE_HOOKS
126 	.lc_owner = owner_mtx,
127 #endif
128 };
129 struct lock_class lock_class_mtx_spin = {
130 	.lc_name = "spin mutex",
131 	.lc_flags = LC_SPINLOCK | LC_RECURSABLE,
132 	.lc_assert = assert_mtx,
133 #ifdef DDB
134 	.lc_ddb_show = db_show_mtx,
135 #endif
136 	.lc_lock = lock_spin,
137 	.lc_unlock = unlock_spin,
138 #ifdef KDTRACE_HOOKS
139 	.lc_owner = owner_mtx,
140 #endif
141 };
142 
143 /*
144  * System-wide mutexes
145  */
146 struct mtx blocked_lock;
147 struct mtx Giant;
148 
149 void
assert_mtx(const struct lock_object * lock,int what)150 assert_mtx(const struct lock_object *lock, int what)
151 {
152 
153 	mtx_assert((const struct mtx *)lock, what);
154 }
155 
156 void
lock_mtx(struct lock_object * lock,uintptr_t how)157 lock_mtx(struct lock_object *lock, uintptr_t how)
158 {
159 
160 	mtx_lock((struct mtx *)lock);
161 }
162 
163 void
lock_spin(struct lock_object * lock,uintptr_t how)164 lock_spin(struct lock_object *lock, uintptr_t how)
165 {
166 
167 	panic("spin locks can only use msleep_spin");
168 }
169 
170 uintptr_t
unlock_mtx(struct lock_object * lock)171 unlock_mtx(struct lock_object *lock)
172 {
173 	struct mtx *m;
174 
175 	m = (struct mtx *)lock;
176 	mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
177 	mtx_unlock(m);
178 	return (0);
179 }
180 
181 uintptr_t
unlock_spin(struct lock_object * lock)182 unlock_spin(struct lock_object *lock)
183 {
184 
185 	panic("spin locks can only use msleep_spin");
186 }
187 
188 #ifdef KDTRACE_HOOKS
189 int
owner_mtx(const struct lock_object * lock,struct thread ** owner)190 owner_mtx(const struct lock_object *lock, struct thread **owner)
191 {
192 	const struct mtx *m = (const struct mtx *)lock;
193 
194 	*owner = mtx_owner(m);
195 	return (mtx_unowned(m) == 0);
196 }
197 #endif
198 
199 /*
200  * Function versions of the inlined __mtx_* macros.  These are used by
201  * modules and can also be called from assembly language if needed.
202  */
203 void
__mtx_lock_flags(volatile uintptr_t * c,int opts,const char * file,int line)204 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
205 {
206 	struct mtx *m;
207 
208 	if (SCHEDULER_STOPPED())
209 		return;
210 
211 	m = mtxlock2mtx(c);
212 
213 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
214 	    ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d",
215 	    curthread, m->lock_object.lo_name, file, line));
216 	KASSERT(m->mtx_lock != MTX_DESTROYED,
217 	    ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
218 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
219 	    ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
220 	    file, line));
221 	WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
222 	    LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
223 
224 	__mtx_lock(m, curthread, opts, file, line);
225 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
226 	    line);
227 	WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
228 	    file, line);
229 	curthread->td_locks++;
230 }
231 
232 void
__mtx_unlock_flags(volatile uintptr_t * c,int opts,const char * file,int line)233 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
234 {
235 	struct mtx *m;
236 
237 	if (SCHEDULER_STOPPED())
238 		return;
239 
240 	m = mtxlock2mtx(c);
241 
242 	KASSERT(m->mtx_lock != MTX_DESTROYED,
243 	    ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
244 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
245 	    ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
246 	    file, line));
247 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
248 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
249 	    line);
250 	mtx_assert(m, MA_OWNED);
251 
252 	if (m->mtx_recurse == 0)
253 		LOCKSTAT_PROFILE_RELEASE_LOCK(LS_MTX_UNLOCK_RELEASE, m);
254 	__mtx_unlock(m, curthread, opts, file, line);
255 	curthread->td_locks--;
256 }
257 
258 void
__mtx_lock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)259 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
260     int line)
261 {
262 	struct mtx *m;
263 
264 	if (SCHEDULER_STOPPED())
265 		return;
266 
267 	m = mtxlock2mtx(c);
268 
269 	KASSERT(m->mtx_lock != MTX_DESTROYED,
270 	    ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
271 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
272 	    ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
273 	    m->lock_object.lo_name, file, line));
274 	if (mtx_owned(m))
275 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
276 		    (opts & MTX_RECURSE) != 0,
277 	    ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
278 		    m->lock_object.lo_name, file, line));
279 	opts &= ~MTX_RECURSE;
280 	WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
281 	    file, line, NULL);
282 	__mtx_lock_spin(m, curthread, opts, file, line);
283 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
284 	    line);
285 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
286 }
287 
288 void
__mtx_unlock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)289 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
290     int line)
291 {
292 	struct mtx *m;
293 
294 	if (SCHEDULER_STOPPED())
295 		return;
296 
297 	m = mtxlock2mtx(c);
298 
299 	KASSERT(m->mtx_lock != MTX_DESTROYED,
300 	    ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
301 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
302 	    ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
303 	    m->lock_object.lo_name, file, line));
304 	WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
305 	LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
306 	    line);
307 	mtx_assert(m, MA_OWNED);
308 
309 	__mtx_unlock_spin(m);
310 }
311 
312 /*
313  * The important part of mtx_trylock{,_flags}()
314  * Tries to acquire lock `m.'  If this function is called on a mutex that
315  * is already owned, it will recursively acquire the lock.
316  */
317 int
_mtx_trylock_flags_(volatile uintptr_t * c,int opts,const char * file,int line)318 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
319 {
320 	struct mtx *m;
321 #ifdef LOCK_PROFILING
322 	uint64_t waittime = 0;
323 	int contested = 0;
324 #endif
325 	int rval;
326 
327 	if (SCHEDULER_STOPPED())
328 		return (1);
329 
330 	m = mtxlock2mtx(c);
331 
332 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
333 	    ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d",
334 	    curthread, m->lock_object.lo_name, file, line));
335 	KASSERT(m->mtx_lock != MTX_DESTROYED,
336 	    ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
337 	KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
338 	    ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
339 	    file, line));
340 
341 	if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
342 	    (opts & MTX_RECURSE) != 0)) {
343 		m->mtx_recurse++;
344 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
345 		rval = 1;
346 	} else
347 		rval = _mtx_obtain_lock(m, (uintptr_t)curthread);
348 	opts &= ~MTX_RECURSE;
349 
350 	LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
351 	if (rval) {
352 		WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
353 		    file, line);
354 		curthread->td_locks++;
355 		if (m->mtx_recurse == 0)
356 			LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE,
357 			    m, contested, waittime, file, line);
358 
359 	}
360 
361 	return (rval);
362 }
363 
364 /*
365  * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
366  *
367  * We call this if the lock is either contested (i.e. we need to go to
368  * sleep waiting for it), or if we need to recurse on it.
369  */
370 void
__mtx_lock_sleep(volatile uintptr_t * c,uintptr_t tid,int opts,const char * file,int line)371 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
372     const char *file, int line)
373 {
374 	struct mtx *m;
375 	struct turnstile *ts;
376 	uintptr_t v;
377 #ifdef ADAPTIVE_MUTEXES
378 	volatile struct thread *owner;
379 #endif
380 #ifdef KTR
381 	int cont_logged = 0;
382 #endif
383 #ifdef LOCK_PROFILING
384 	int contested = 0;
385 	uint64_t waittime = 0;
386 #endif
387 #ifdef KDTRACE_HOOKS
388 	uint64_t spin_cnt = 0;
389 	uint64_t sleep_cnt = 0;
390 	int64_t sleep_time = 0;
391 #endif
392 
393 	if (SCHEDULER_STOPPED())
394 		return;
395 
396 	m = mtxlock2mtx(c);
397 
398 	if (mtx_owned(m)) {
399 		KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
400 		    (opts & MTX_RECURSE) != 0,
401 	    ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
402 		    m->lock_object.lo_name, file, line));
403 		opts &= ~MTX_RECURSE;
404 		m->mtx_recurse++;
405 		atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
406 		if (LOCK_LOG_TEST(&m->lock_object, opts))
407 			CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
408 		return;
409 	}
410 	opts &= ~MTX_RECURSE;
411 
412 #ifdef HWPMC_HOOKS
413 	PMC_SOFT_CALL( , , lock, failed);
414 #endif
415 	lock_profile_obtain_lock_failed(&m->lock_object,
416 		    &contested, &waittime);
417 	if (LOCK_LOG_TEST(&m->lock_object, opts))
418 		CTR4(KTR_LOCK,
419 		    "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
420 		    m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
421 
422 	while (!_mtx_obtain_lock(m, tid)) {
423 #ifdef KDTRACE_HOOKS
424 		spin_cnt++;
425 #endif
426 #ifdef ADAPTIVE_MUTEXES
427 		/*
428 		 * If the owner is running on another CPU, spin until the
429 		 * owner stops running or the state of the lock changes.
430 		 */
431 		v = m->mtx_lock;
432 		if (v != MTX_UNOWNED) {
433 			owner = (struct thread *)(v & ~MTX_FLAGMASK);
434 			if (TD_IS_RUNNING(owner)) {
435 				if (LOCK_LOG_TEST(&m->lock_object, 0))
436 					CTR3(KTR_LOCK,
437 					    "%s: spinning on %p held by %p",
438 					    __func__, m, owner);
439 				KTR_STATE1(KTR_SCHED, "thread",
440 				    sched_tdname((struct thread *)tid),
441 				    "spinning", "lockname:\"%s\"",
442 				    m->lock_object.lo_name);
443 				while (mtx_owner(m) == owner &&
444 				    TD_IS_RUNNING(owner)) {
445 					cpu_spinwait();
446 #ifdef KDTRACE_HOOKS
447 					spin_cnt++;
448 #endif
449 				}
450 				KTR_STATE0(KTR_SCHED, "thread",
451 				    sched_tdname((struct thread *)tid),
452 				    "running");
453 				continue;
454 			}
455 		}
456 #endif
457 
458 		ts = turnstile_trywait(&m->lock_object);
459 		v = m->mtx_lock;
460 
461 		/*
462 		 * Check if the lock has been released while spinning for
463 		 * the turnstile chain lock.
464 		 */
465 		if (v == MTX_UNOWNED) {
466 			turnstile_cancel(ts);
467 			continue;
468 		}
469 
470 #ifdef ADAPTIVE_MUTEXES
471 		/*
472 		 * The current lock owner might have started executing
473 		 * on another CPU (or the lock could have changed
474 		 * owners) while we were waiting on the turnstile
475 		 * chain lock.  If so, drop the turnstile lock and try
476 		 * again.
477 		 */
478 		owner = (struct thread *)(v & ~MTX_FLAGMASK);
479 		if (TD_IS_RUNNING(owner)) {
480 			turnstile_cancel(ts);
481 			continue;
482 		}
483 #endif
484 
485 		/*
486 		 * If the mutex isn't already contested and a failure occurs
487 		 * setting the contested bit, the mutex was either released
488 		 * or the state of the MTX_RECURSED bit changed.
489 		 */
490 		if ((v & MTX_CONTESTED) == 0 &&
491 		    !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
492 			turnstile_cancel(ts);
493 			continue;
494 		}
495 
496 		/*
497 		 * We definitely must sleep for this lock.
498 		 */
499 		mtx_assert(m, MA_NOTOWNED);
500 
501 #ifdef KTR
502 		if (!cont_logged) {
503 			CTR6(KTR_CONTENTION,
504 			    "contention: %p at %s:%d wants %s, taken by %s:%d",
505 			    (void *)tid, file, line, m->lock_object.lo_name,
506 			    WITNESS_FILE(&m->lock_object),
507 			    WITNESS_LINE(&m->lock_object));
508 			cont_logged = 1;
509 		}
510 #endif
511 
512 		/*
513 		 * Block on the turnstile.
514 		 */
515 #ifdef KDTRACE_HOOKS
516 		sleep_time -= lockstat_nsecs();
517 #endif
518 		turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
519 #ifdef KDTRACE_HOOKS
520 		sleep_time += lockstat_nsecs();
521 		sleep_cnt++;
522 #endif
523 	}
524 #ifdef KTR
525 	if (cont_logged) {
526 		CTR4(KTR_CONTENTION,
527 		    "contention end: %s acquired by %p at %s:%d",
528 		    m->lock_object.lo_name, (void *)tid, file, line);
529 	}
530 #endif
531 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
532 	    waittime, file, line);
533 #ifdef KDTRACE_HOOKS
534 	if (sleep_time)
535 		LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
536 
537 	/*
538 	 * Only record the loops spinning and not sleeping.
539 	 */
540 	if (spin_cnt > sleep_cnt)
541 		LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt));
542 #endif
543 }
544 
545 static void
_mtx_lock_spin_failed(struct mtx * m)546 _mtx_lock_spin_failed(struct mtx *m)
547 {
548 	struct thread *td;
549 
550 	td = mtx_owner(m);
551 
552 	/* If the mutex is unlocked, try again. */
553 	if (td == NULL)
554 		return;
555 
556 	printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
557 	    m, m->lock_object.lo_name, td, td->td_tid);
558 #ifdef WITNESS
559 	witness_display_spinlock(&m->lock_object, td, printf);
560 #endif
561 	panic("spin lock held too long");
562 }
563 
564 #ifdef SMP
565 /*
566  * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
567  *
568  * This is only called if we need to actually spin for the lock. Recursion
569  * is handled inline.
570  */
571 void
_mtx_lock_spin_cookie(volatile uintptr_t * c,uintptr_t tid,int opts,const char * file,int line)572 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
573     const char *file, int line)
574 {
575 	struct mtx *m;
576 	int i = 0;
577 #ifdef LOCK_PROFILING
578 	int contested = 0;
579 	uint64_t waittime = 0;
580 #endif
581 
582 	if (SCHEDULER_STOPPED())
583 		return;
584 
585 	m = mtxlock2mtx(c);
586 
587 	if (LOCK_LOG_TEST(&m->lock_object, opts))
588 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
589 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
590 	    "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
591 
592 #ifdef HWPMC_HOOKS
593 	PMC_SOFT_CALL( , , lock, failed);
594 #endif
595 	lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
596 	while (!_mtx_obtain_lock(m, tid)) {
597 
598 		/* Give interrupts a chance while we spin. */
599 		spinlock_exit();
600 		while (m->mtx_lock != MTX_UNOWNED) {
601 			if (i++ < 10000000) {
602 				cpu_spinwait();
603 				continue;
604 			}
605 			if (i < 60000000 || kdb_active || panicstr != NULL)
606 				DELAY(1);
607 			else
608 				_mtx_lock_spin_failed(m);
609 			cpu_spinwait();
610 		}
611 		spinlock_enter();
612 	}
613 
614 	if (LOCK_LOG_TEST(&m->lock_object, opts))
615 		CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
616 	KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
617 	    "running");
618 
619 	LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
620 	    contested, waittime, (file), (line));
621 	LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, i);
622 }
623 #endif /* SMP */
624 
625 void
thread_lock_flags_(struct thread * td,int opts,const char * file,int line)626 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
627 {
628 	struct mtx *m;
629 	uintptr_t tid;
630 	int i;
631 #ifdef LOCK_PROFILING
632 	int contested = 0;
633 	uint64_t waittime = 0;
634 #endif
635 #ifdef KDTRACE_HOOKS
636 	uint64_t spin_cnt = 0;
637 #endif
638 
639 	i = 0;
640 	tid = (uintptr_t)curthread;
641 
642 	if (SCHEDULER_STOPPED())
643 		return;
644 
645 	for (;;) {
646 retry:
647 		spinlock_enter();
648 		m = td->td_lock;
649 		KASSERT(m->mtx_lock != MTX_DESTROYED,
650 		    ("thread_lock() of destroyed mutex @ %s:%d", file, line));
651 		KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
652 		    ("thread_lock() of sleep mutex %s @ %s:%d",
653 		    m->lock_object.lo_name, file, line));
654 		if (mtx_owned(m))
655 			KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
656 	    ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n",
657 			    m->lock_object.lo_name, file, line));
658 		WITNESS_CHECKORDER(&m->lock_object,
659 		    opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
660 		while (!_mtx_obtain_lock(m, tid)) {
661 #ifdef KDTRACE_HOOKS
662 			spin_cnt++;
663 #endif
664 			if (m->mtx_lock == tid) {
665 				m->mtx_recurse++;
666 				break;
667 			}
668 #ifdef HWPMC_HOOKS
669 			PMC_SOFT_CALL( , , lock, failed);
670 #endif
671 			lock_profile_obtain_lock_failed(&m->lock_object,
672 			    &contested, &waittime);
673 			/* Give interrupts a chance while we spin. */
674 			spinlock_exit();
675 			while (m->mtx_lock != MTX_UNOWNED) {
676 				if (i++ < 10000000)
677 					cpu_spinwait();
678 				else if (i < 60000000 ||
679 				    kdb_active || panicstr != NULL)
680 					DELAY(1);
681 				else
682 					_mtx_lock_spin_failed(m);
683 				cpu_spinwait();
684 				if (m != td->td_lock)
685 					goto retry;
686 			}
687 			spinlock_enter();
688 		}
689 		if (m == td->td_lock)
690 			break;
691 		__mtx_unlock_spin(m);	/* does spinlock_exit() */
692 #ifdef KDTRACE_HOOKS
693 		spin_cnt++;
694 #endif
695 	}
696 	if (m->mtx_recurse == 0)
697 		LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE,
698 		    m, contested, waittime, (file), (line));
699 	LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
700 	    line);
701 	WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
702 	LOCKSTAT_RECORD1(LS_THREAD_LOCK_SPIN, m, spin_cnt);
703 }
704 
705 struct mtx *
thread_lock_block(struct thread * td)706 thread_lock_block(struct thread *td)
707 {
708 	struct mtx *lock;
709 
710 	THREAD_LOCK_ASSERT(td, MA_OWNED);
711 	lock = td->td_lock;
712 	td->td_lock = &blocked_lock;
713 	mtx_unlock_spin(lock);
714 
715 	return (lock);
716 }
717 
718 void
thread_lock_unblock(struct thread * td,struct mtx * new)719 thread_lock_unblock(struct thread *td, struct mtx *new)
720 {
721 	mtx_assert(new, MA_OWNED);
722 	MPASS(td->td_lock == &blocked_lock);
723 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
724 }
725 
726 void
thread_lock_set(struct thread * td,struct mtx * new)727 thread_lock_set(struct thread *td, struct mtx *new)
728 {
729 	struct mtx *lock;
730 
731 	mtx_assert(new, MA_OWNED);
732 	THREAD_LOCK_ASSERT(td, MA_OWNED);
733 	lock = td->td_lock;
734 	td->td_lock = new;
735 	mtx_unlock_spin(lock);
736 }
737 
738 /*
739  * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
740  *
741  * We are only called here if the lock is recursed or contested (i.e. we
742  * need to wake up a blocked thread).
743  */
744 void
__mtx_unlock_sleep(volatile uintptr_t * c,int opts,const char * file,int line)745 __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line)
746 {
747 	struct mtx *m;
748 	struct turnstile *ts;
749 
750 	if (SCHEDULER_STOPPED())
751 		return;
752 
753 	m = mtxlock2mtx(c);
754 
755 	if (mtx_recursed(m)) {
756 		if (--(m->mtx_recurse) == 0)
757 			atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
758 		if (LOCK_LOG_TEST(&m->lock_object, opts))
759 			CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
760 		return;
761 	}
762 
763 	/*
764 	 * We have to lock the chain before the turnstile so this turnstile
765 	 * can be removed from the hash list if it is empty.
766 	 */
767 	turnstile_chain_lock(&m->lock_object);
768 	ts = turnstile_lookup(&m->lock_object);
769 	if (LOCK_LOG_TEST(&m->lock_object, opts))
770 		CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
771 	MPASS(ts != NULL);
772 	turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
773 	_mtx_release_lock_quick(m);
774 
775 	/*
776 	 * This turnstile is now no longer associated with the mutex.  We can
777 	 * unlock the chain lock so a new turnstile may take it's place.
778 	 */
779 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
780 	turnstile_chain_unlock(&m->lock_object);
781 }
782 
783 /*
784  * All the unlocking of MTX_SPIN locks is done inline.
785  * See the __mtx_unlock_spin() macro for the details.
786  */
787 
788 /*
789  * The backing function for the INVARIANTS-enabled mtx_assert()
790  */
791 #ifdef INVARIANT_SUPPORT
792 void
__mtx_assert(const volatile uintptr_t * c,int what,const char * file,int line)793 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
794 {
795 	const struct mtx *m;
796 
797 	if (panicstr != NULL || dumping)
798 		return;
799 
800 	m = mtxlock2mtx(c);
801 
802 	switch (what) {
803 	case MA_OWNED:
804 	case MA_OWNED | MA_RECURSED:
805 	case MA_OWNED | MA_NOTRECURSED:
806 		if (!mtx_owned(m))
807 			panic("mutex %s not owned at %s:%d",
808 			    m->lock_object.lo_name, file, line);
809 		if (mtx_recursed(m)) {
810 			if ((what & MA_NOTRECURSED) != 0)
811 				panic("mutex %s recursed at %s:%d",
812 				    m->lock_object.lo_name, file, line);
813 		} else if ((what & MA_RECURSED) != 0) {
814 			panic("mutex %s unrecursed at %s:%d",
815 			    m->lock_object.lo_name, file, line);
816 		}
817 		break;
818 	case MA_NOTOWNED:
819 		if (mtx_owned(m))
820 			panic("mutex %s owned at %s:%d",
821 			    m->lock_object.lo_name, file, line);
822 		break;
823 	default:
824 		panic("unknown mtx_assert at %s:%d", file, line);
825 	}
826 }
827 #endif
828 
829 /*
830  * The MUTEX_DEBUG-enabled mtx_validate()
831  *
832  * Most of these checks have been moved off into the LO_INITIALIZED flag
833  * maintained by the witness code.
834  */
835 #ifdef MUTEX_DEBUG
836 
837 void	mtx_validate(struct mtx *);
838 
839 void
mtx_validate(struct mtx * m)840 mtx_validate(struct mtx *m)
841 {
842 
843 /*
844  * XXX: When kernacc() does not require Giant we can reenable this check
845  */
846 #ifdef notyet
847 	/*
848 	 * Can't call kernacc() from early init386(), especially when
849 	 * initializing Giant mutex, because some stuff in kernacc()
850 	 * requires Giant itself.
851 	 */
852 	if (!cold)
853 		if (!kernacc((caddr_t)m, sizeof(m),
854 		    VM_PROT_READ | VM_PROT_WRITE))
855 			panic("Can't read and write to mutex %p", m);
856 #endif
857 }
858 #endif
859 
860 /*
861  * General init routine used by the MTX_SYSINIT() macro.
862  */
863 void
mtx_sysinit(void * arg)864 mtx_sysinit(void *arg)
865 {
866 	struct mtx_args *margs = arg;
867 
868 	mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
869 	    margs->ma_opts);
870 }
871 
872 /*
873  * Mutex initialization routine; initialize lock `m' of type contained in
874  * `opts' with options contained in `opts' and name `name.'  The optional
875  * lock type `type' is used as a general lock category name for use with
876  * witness.
877  */
878 void
_mtx_init(volatile uintptr_t * c,const char * name,const char * type,int opts)879 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
880 {
881 	struct mtx *m;
882 	struct lock_class *class;
883 	int flags;
884 
885 	m = mtxlock2mtx(c);
886 
887 	MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
888 		MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
889 	ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
890 	    ("%s: mtx_lock not aligned for %s: %p", __func__, name,
891 	    &m->mtx_lock));
892 
893 #ifdef MUTEX_DEBUG
894 	/* Diagnostic and error correction */
895 	mtx_validate(m);
896 #endif
897 
898 	/* Determine lock class and lock flags. */
899 	if (opts & MTX_SPIN)
900 		class = &lock_class_mtx_spin;
901 	else
902 		class = &lock_class_mtx_sleep;
903 	flags = 0;
904 	if (opts & MTX_QUIET)
905 		flags |= LO_QUIET;
906 	if (opts & MTX_RECURSE)
907 		flags |= LO_RECURSABLE;
908 	if ((opts & MTX_NOWITNESS) == 0)
909 		flags |= LO_WITNESS;
910 	if (opts & MTX_DUPOK)
911 		flags |= LO_DUPOK;
912 	if (opts & MTX_NOPROFILE)
913 		flags |= LO_NOPROFILE;
914 
915 	/* Initialize mutex. */
916 	lock_init(&m->lock_object, class, name, type, flags);
917 
918 	m->mtx_lock = MTX_UNOWNED;
919 	m->mtx_recurse = 0;
920 }
921 
922 /*
923  * Remove lock `m' from all_mtx queue.  We don't allow MTX_QUIET to be
924  * passed in as a flag here because if the corresponding mtx_init() was
925  * called with MTX_QUIET set, then it will already be set in the mutex's
926  * flags.
927  */
928 void
_mtx_destroy(volatile uintptr_t * c)929 _mtx_destroy(volatile uintptr_t *c)
930 {
931 	struct mtx *m;
932 
933 	m = mtxlock2mtx(c);
934 
935 	if (!mtx_owned(m))
936 		MPASS(mtx_unowned(m));
937 	else {
938 		MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
939 
940 		/* Perform the non-mtx related part of mtx_unlock_spin(). */
941 		if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
942 			spinlock_exit();
943 		else
944 			curthread->td_locks--;
945 
946 		lock_profile_release_lock(&m->lock_object);
947 		/* Tell witness this isn't locked to make it happy. */
948 		WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
949 		    __LINE__);
950 	}
951 
952 	m->mtx_lock = MTX_DESTROYED;
953 	lock_destroy(&m->lock_object);
954 }
955 
956 /*
957  * Intialize the mutex code and system mutexes.  This is called from the MD
958  * startup code prior to mi_startup().  The per-CPU data space needs to be
959  * setup before this is called.
960  */
961 void
mutex_init(void)962 mutex_init(void)
963 {
964 
965 	/* Setup turnstiles so that sleep mutexes work. */
966 	init_turnstiles();
967 
968 	/*
969 	 * Initialize mutexes.
970 	 */
971 	mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
972 	mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
973 	blocked_lock.mtx_lock = 0xdeadc0de;	/* Always blocked. */
974 	mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
975 	mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
976 	mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
977 	mtx_lock(&Giant);
978 }
979 
980 #ifdef DDB
981 void
db_show_mtx(const struct lock_object * lock)982 db_show_mtx(const struct lock_object *lock)
983 {
984 	struct thread *td;
985 	const struct mtx *m;
986 
987 	m = (const struct mtx *)lock;
988 
989 	db_printf(" flags: {");
990 	if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
991 		db_printf("SPIN");
992 	else
993 		db_printf("DEF");
994 	if (m->lock_object.lo_flags & LO_RECURSABLE)
995 		db_printf(", RECURSE");
996 	if (m->lock_object.lo_flags & LO_DUPOK)
997 		db_printf(", DUPOK");
998 	db_printf("}\n");
999 	db_printf(" state: {");
1000 	if (mtx_unowned(m))
1001 		db_printf("UNOWNED");
1002 	else if (mtx_destroyed(m))
1003 		db_printf("DESTROYED");
1004 	else {
1005 		db_printf("OWNED");
1006 		if (m->mtx_lock & MTX_CONTESTED)
1007 			db_printf(", CONTESTED");
1008 		if (m->mtx_lock & MTX_RECURSED)
1009 			db_printf(", RECURSED");
1010 	}
1011 	db_printf("}\n");
1012 	if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1013 		td = mtx_owner(m);
1014 		db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1015 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1016 		if (mtx_recursed(m))
1017 			db_printf(" recursed: %d\n", m->mtx_recurse);
1018 	}
1019 }
1020 #endif
1021