xref: /netbsd-src/sys/kern/kern_mutex.c (revision 7c3f385475147b6e1c4753f2bee961630e2dfc40)
1 /*	$NetBSD: kern_mutex.c,v 1.34 2008/04/11 15:28:34 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe and Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the NetBSD
21  *	Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 
39 /*
40  * Kernel mutex implementation, modeled after those found in Solaris,
41  * a description of which can be found in:
42  *
43  *	Solaris Internals: Core Kernel Architecture, Jim Mauro and
44  *	    Richard McDougall.
45  */
46 
47 #define	__MUTEX_PRIVATE
48 
49 #include <sys/cdefs.h>
50 __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.34 2008/04/11 15:28:34 ad Exp $");
51 
52 #include "opt_multiprocessor.h"
53 
54 #include <sys/param.h>
55 #include <sys/proc.h>
56 #include <sys/mutex.h>
57 #include <sys/sched.h>
58 #include <sys/sleepq.h>
59 #include <sys/systm.h>
60 #include <sys/lockdebug.h>
61 #include <sys/kernel.h>
62 #include <sys/atomic.h>
63 #include <sys/intr.h>
64 #include <sys/lock.h>
65 #include <sys/pool.h>
66 
67 #include <dev/lockstat.h>
68 
69 #include <machine/lock.h>
70 
71 /*
72  * When not running a debug kernel, spin mutexes are not much
73  * more than an splraiseipl() and splx() pair.
74  */
75 
76 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
77 #define	FULL
78 #endif
79 
80 /*
81  * Debugging support.
82  */
83 
84 #define	MUTEX_WANTLOCK(mtx)					\
85     LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx),		\
86         (uintptr_t)__builtin_return_address(0), 0)
87 #define	MUTEX_LOCKED(mtx)					\
88     LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx),			\
89         (uintptr_t)__builtin_return_address(0), 0)
90 #define	MUTEX_UNLOCKED(mtx)					\
91     LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx),		\
92         (uintptr_t)__builtin_return_address(0), 0)
93 #define	MUTEX_ABORT(mtx, msg)					\
94     mutex_abort(mtx, __func__, msg)
95 
96 #if defined(LOCKDEBUG)
97 
98 #define	MUTEX_DASSERT(mtx, cond)				\
99 do {								\
100 	if (!(cond))						\
101 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
102 } while (/* CONSTCOND */ 0);
103 
104 #else	/* LOCKDEBUG */
105 
106 #define	MUTEX_DASSERT(mtx, cond)	/* nothing */
107 
108 #endif /* LOCKDEBUG */
109 
110 #if defined(DIAGNOSTIC)
111 
112 #define	MUTEX_ASSERT(mtx, cond)					\
113 do {								\
114 	if (!(cond))						\
115 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
116 } while (/* CONSTCOND */ 0)
117 
118 #else	/* DIAGNOSTIC */
119 
120 #define	MUTEX_ASSERT(mtx, cond)	/* nothing */
121 
122 #endif	/* DIAGNOSTIC */
123 
124 /*
125  * Spin mutex SPL save / restore.
126  */
127 #ifndef MUTEX_COUNT_BIAS
128 #define	MUTEX_COUNT_BIAS	0
129 #endif
130 
131 #define	MUTEX_SPIN_SPLRAISE(mtx)					\
132 do {									\
133 	struct cpu_info *x__ci = curcpu();				\
134 	int x__cnt, s;							\
135 	x__cnt = x__ci->ci_mtx_count--;					\
136 	s = splraiseipl(mtx->mtx_ipl);					\
137 	if (x__cnt == MUTEX_COUNT_BIAS)					\
138 		x__ci->ci_mtx_oldspl = (s);				\
139 } while (/* CONSTCOND */ 0)
140 
141 #define	MUTEX_SPIN_SPLRESTORE(mtx)					\
142 do {									\
143 	struct cpu_info *x__ci = curcpu();				\
144 	int s = x__ci->ci_mtx_oldspl;					\
145 	__insn_barrier();						\
146 	if (++(x__ci->ci_mtx_count) == MUTEX_COUNT_BIAS)		\
147 		splx(s);						\
148 } while (/* CONSTCOND */ 0)
149 
150 /*
151  * For architectures that provide 'simple' mutexes: they provide a
152  * CAS function that is either MP-safe, or does not need to be MP
153  * safe.  Adaptive mutexes on these architectures do not require an
154  * additional interlock.
155  */
156 
157 #ifdef __HAVE_SIMPLE_MUTEXES
158 
159 #define	MUTEX_OWNER(owner)						\
160 	(owner & MUTEX_THREAD)
161 #define	MUTEX_HAS_WAITERS(mtx)						\
162 	(((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
163 
164 #define	MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug)				\
165 do {									\
166 	if (dodebug)							\
167 		(mtx)->mtx_owner |= MUTEX_BIT_DEBUG;			\
168 } while (/* CONSTCOND */ 0);
169 
170 #define	MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl)			\
171 do {									\
172 	(mtx)->mtx_owner = MUTEX_BIT_SPIN;				\
173 	if (dodebug)							\
174 		(mtx)->mtx_owner |= MUTEX_BIT_DEBUG;			\
175 	(mtx)->mtx_ipl = makeiplcookie((ipl));				\
176 	__cpu_simple_lock_init(&(mtx)->mtx_lock);			\
177 } while (/* CONSTCOND */ 0)
178 
179 #define	MUTEX_DESTROY(mtx)						\
180 do {									\
181 	(mtx)->mtx_owner = MUTEX_THREAD;				\
182 } while (/* CONSTCOND */ 0);
183 
184 #define	MUTEX_SPIN_P(mtx)		\
185     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) != 0)
186 #define	MUTEX_ADAPTIVE_P(mtx)		\
187     (((mtx)->mtx_owner & MUTEX_BIT_SPIN) == 0)
188 
189 #define	MUTEX_DEBUG_P(mtx)	(((mtx)->mtx_owner & MUTEX_BIT_DEBUG) != 0)
190 #if defined(LOCKDEBUG)
191 #define	MUTEX_OWNED(owner)		(((owner) & ~MUTEX_BIT_DEBUG) != 0)
192 #define	MUTEX_INHERITDEBUG(new, old)	(new) |= (old) & MUTEX_BIT_DEBUG
193 #else /* defined(LOCKDEBUG) */
194 #define	MUTEX_OWNED(owner)		((owner) != 0)
195 #define	MUTEX_INHERITDEBUG(new, old)	/* nothing */
196 #endif /* defined(LOCKDEBUG) */
197 
198 static inline int
199 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
200 {
201 	int rv;
202 	uintptr_t old = 0;
203 	uintptr_t new = curthread;
204 
205 	MUTEX_INHERITDEBUG(old, mtx->mtx_owner);
206 	MUTEX_INHERITDEBUG(new, old);
207 	rv = MUTEX_CAS(&mtx->mtx_owner, old, new);
208 	MUTEX_RECEIVE(mtx);
209 	return rv;
210 }
211 
212 static inline int
213 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
214 {
215 	int rv;
216 	rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
217 	MUTEX_RECEIVE(mtx);
218 	return rv;
219 }
220 
221 static inline void
222 MUTEX_RELEASE(kmutex_t *mtx)
223 {
224 	uintptr_t new;
225 
226 	MUTEX_GIVE(mtx);
227 	new = 0;
228 	MUTEX_INHERITDEBUG(new, mtx->mtx_owner);
229 	mtx->mtx_owner = new;
230 }
231 
232 static inline void
233 MUTEX_CLEAR_WAITERS(kmutex_t *mtx)
234 {
235 	/* nothing */
236 }
237 #endif	/* __HAVE_SIMPLE_MUTEXES */
238 
239 /*
240  * Patch in stubs via strong alias where they are not available.
241  */
242 
243 #if defined(LOCKDEBUG)
244 #undef	__HAVE_MUTEX_STUBS
245 #undef	__HAVE_SPIN_MUTEX_STUBS
246 #endif
247 
248 #ifndef __HAVE_MUTEX_STUBS
249 __strong_alias(mutex_enter,mutex_vector_enter);
250 __strong_alias(mutex_exit,mutex_vector_exit);
251 #endif
252 
253 #ifndef __HAVE_SPIN_MUTEX_STUBS
254 __strong_alias(mutex_spin_enter,mutex_vector_enter);
255 __strong_alias(mutex_spin_exit,mutex_vector_exit);
256 #endif
257 
258 void	mutex_abort(kmutex_t *, const char *, const char *);
259 void	mutex_dump(volatile void *);
260 int	mutex_onproc(uintptr_t, struct cpu_info **);
261 
262 lockops_t mutex_spin_lockops = {
263 	"Mutex",
264 	0,
265 	mutex_dump
266 };
267 
268 lockops_t mutex_adaptive_lockops = {
269 	"Mutex",
270 	1,
271 	mutex_dump
272 };
273 
274 syncobj_t mutex_syncobj = {
275 	SOBJ_SLEEPQ_SORTED,
276 	turnstile_unsleep,
277 	turnstile_changepri,
278 	sleepq_lendpri,
279 	(void *)mutex_owner,
280 };
281 
282 /* Mutex cache */
283 #define	MUTEX_OBJ_MAGIC	0x5aa3c85d
284 struct kmutexobj {
285 	kmutex_t	mo_lock;
286 	u_int		mo_magic;
287 	u_int		mo_refcnt;
288 };
289 
290 static int	mutex_obj_ctor(void *, void *, int);
291 
292 static pool_cache_t	mutex_obj_cache;
293 
294 /*
295  * mutex_dump:
296  *
297  *	Dump the contents of a mutex structure.
298  */
299 void
300 mutex_dump(volatile void *cookie)
301 {
302 	volatile kmutex_t *mtx = cookie;
303 
304 	printf_nolog("owner field  : %#018lx wait/spin: %16d/%d\n",
305 	    (long)MUTEX_OWNER(mtx->mtx_owner), MUTEX_HAS_WAITERS(mtx),
306 	    MUTEX_SPIN_P(mtx));
307 }
308 
309 /*
310  * mutex_abort:
311  *
312  *	Dump information about an error and panic the system.  This
313  *	generates a lot of machine code in the DIAGNOSTIC case, so
314  *	we ask the compiler to not inline it.
315  */
316 
317 #if __GNUC_PREREQ__(3, 0)
318 __attribute ((noinline)) __attribute ((noreturn))
319 #endif
320 void
321 mutex_abort(kmutex_t *mtx, const char *func, const char *msg)
322 {
323 
324 	LOCKDEBUG_ABORT(mtx, (MUTEX_SPIN_P(mtx) ?
325 	    &mutex_spin_lockops : &mutex_adaptive_lockops), func, msg);
326 	/* NOTREACHED */
327 }
328 
329 /*
330  * mutex_init:
331  *
332  *	Initialize a mutex for use.  Note that adaptive mutexes are in
333  *	essence spin mutexes that can sleep to avoid deadlock and wasting
334  *	CPU time.  We can't easily provide a type of mutex that always
335  *	sleeps - see comments in mutex_vector_enter() about releasing
336  *	mutexes unlocked.
337  */
338 void
339 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
340 {
341 	bool dodebug;
342 
343 	memset(mtx, 0, sizeof(*mtx));
344 
345 	switch (type) {
346 	case MUTEX_ADAPTIVE:
347 		KASSERT(ipl == IPL_NONE);
348 		break;
349 	case MUTEX_DEFAULT:
350 	case MUTEX_DRIVER:
351 		if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK ||
352 		    ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET ||
353 		    ipl == IPL_SOFTSERIAL) {
354 			type = MUTEX_ADAPTIVE;
355 		} else {
356 			type = MUTEX_SPIN;
357 		}
358 		break;
359 	default:
360 		break;
361 	}
362 
363 	switch (type) {
364 	case MUTEX_NODEBUG:
365 		dodebug = LOCKDEBUG_ALLOC(mtx, NULL,
366 		    (uintptr_t)__builtin_return_address(0));
367 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
368 		break;
369 	case MUTEX_ADAPTIVE:
370 		dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_adaptive_lockops,
371 		    (uintptr_t)__builtin_return_address(0));
372 		MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug);
373 		break;
374 	case MUTEX_SPIN:
375 		dodebug = LOCKDEBUG_ALLOC(mtx, &mutex_spin_lockops,
376 		    (uintptr_t)__builtin_return_address(0));
377 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
378 		break;
379 	default:
380 		panic("mutex_init: impossible type");
381 		break;
382 	}
383 }
384 
385 /*
386  * mutex_destroy:
387  *
388  *	Tear down a mutex.
389  */
390 void
391 mutex_destroy(kmutex_t *mtx)
392 {
393 
394 	if (MUTEX_ADAPTIVE_P(mtx)) {
395 		MUTEX_ASSERT(mtx, !MUTEX_OWNED(mtx->mtx_owner) &&
396 		    !MUTEX_HAS_WAITERS(mtx));
397 	} else {
398 		MUTEX_ASSERT(mtx, !__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock));
399 	}
400 
401 	LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx);
402 	MUTEX_DESTROY(mtx);
403 }
404 
405 /*
406  * mutex_onproc:
407  *
408  *	Return true if an adaptive mutex owner is running on a CPU in the
409  *	system.  If the target is waiting on the kernel big lock, then we
410  *	must release it.  This is necessary to avoid deadlock.
411  *
412  *	Note that we can't use the mutex owner field as an LWP pointer.  We
413  *	don't have full control over the timing of our execution, and so the
414  *	pointer could be completely invalid by the time we dereference it.
415  */
416 #ifdef MULTIPROCESSOR
417 int
418 mutex_onproc(uintptr_t owner, struct cpu_info **cip)
419 {
420 	CPU_INFO_ITERATOR cii;
421 	struct cpu_info *ci;
422 	struct lwp *l;
423 
424 	if (!MUTEX_OWNED(owner))
425 		return 0;
426 	l = (struct lwp *)MUTEX_OWNER(owner);
427 
428 	/* See if the target is running on a CPU somewhere. */
429 	if ((ci = *cip) != NULL && ci->ci_curlwp == l)
430 		goto run;
431 	for (CPU_INFO_FOREACH(cii, ci))
432 		if (ci->ci_curlwp == l)
433 			goto run;
434 
435 	/* No: it may be safe to block now. */
436 	*cip = NULL;
437 	return 0;
438 
439  run:
440  	/* Target is running; do we need to block? */
441  	*cip = ci;
442 	return ci->ci_biglock_wanted != l;
443 }
444 #endif	/* MULTIPROCESSOR */
445 
446 /*
447  * mutex_vector_enter:
448  *
449  *	Support routine for mutex_enter() that must handles all cases.  In
450  *	the LOCKDEBUG case, mutex_enter() is always aliased here, even if
451  *	fast-path stubs are available.  If an mutex_spin_enter() stub is
452  *	not available, then it is also aliased directly here.
453  */
454 void
455 mutex_vector_enter(kmutex_t *mtx)
456 {
457 	uintptr_t owner, curthread;
458 	turnstile_t *ts;
459 #ifdef MULTIPROCESSOR
460 	struct cpu_info *ci = NULL;
461 	u_int count;
462 #endif
463 	LOCKSTAT_COUNTER(spincnt);
464 	LOCKSTAT_COUNTER(slpcnt);
465 	LOCKSTAT_TIMER(spintime);
466 	LOCKSTAT_TIMER(slptime);
467 	LOCKSTAT_FLAG(lsflag);
468 
469 	/*
470 	 * Handle spin mutexes.
471 	 */
472 	if (MUTEX_SPIN_P(mtx)) {
473 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
474 		u_int spins = 0;
475 #endif
476 		MUTEX_SPIN_SPLRAISE(mtx);
477 		MUTEX_WANTLOCK(mtx);
478 #ifdef FULL
479 		if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
480 			MUTEX_LOCKED(mtx);
481 			return;
482 		}
483 #if !defined(MULTIPROCESSOR)
484 		MUTEX_ABORT(mtx, "locking against myself");
485 #else /* !MULTIPROCESSOR */
486 
487 		LOCKSTAT_ENTER(lsflag);
488 		LOCKSTAT_START_TIMER(lsflag, spintime);
489 		count = SPINLOCK_BACKOFF_MIN;
490 
491 		/*
492 		 * Spin testing the lock word and do exponential backoff
493 		 * to reduce cache line ping-ponging between CPUs.
494 		 */
495 		do {
496 			if (panicstr != NULL)
497 				break;
498 			while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) {
499 				SPINLOCK_BACKOFF(count);
500 #ifdef LOCKDEBUG
501 				if (SPINLOCK_SPINOUT(spins))
502 					MUTEX_ABORT(mtx, "spinout");
503 #endif	/* LOCKDEBUG */
504 			}
505 		} while (!__cpu_simple_lock_try(&mtx->mtx_lock));
506 
507 		if (count != SPINLOCK_BACKOFF_MIN) {
508 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
509 			LOCKSTAT_EVENT(lsflag, mtx,
510 			    LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
511 		}
512 		LOCKSTAT_EXIT(lsflag);
513 #endif	/* !MULTIPROCESSOR */
514 #endif	/* FULL */
515 		MUTEX_LOCKED(mtx);
516 		return;
517 	}
518 
519 	curthread = (uintptr_t)curlwp;
520 
521 	MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
522 	MUTEX_ASSERT(mtx, curthread != 0);
523 	MUTEX_WANTLOCK(mtx);
524 
525 	if (panicstr == NULL) {
526 		LOCKDEBUG_BARRIER(&kernel_lock, 1);
527 	}
528 
529 	LOCKSTAT_ENTER(lsflag);
530 
531 	/*
532 	 * Adaptive mutex; spin trying to acquire the mutex.  If we
533 	 * determine that the owner is not running on a processor,
534 	 * then we stop spinning, and sleep instead.
535 	 */
536 	for (owner = mtx->mtx_owner;;) {
537 		if (!MUTEX_OWNED(owner)) {
538 			/*
539 			 * Mutex owner clear could mean two things:
540 			 *
541 			 *	* The mutex has been released.
542 			 *	* The owner field hasn't been set yet.
543 			 *
544 			 * Try to acquire it again.  If that fails,
545 			 * we'll just loop again.
546 			 */
547 			if (MUTEX_ACQUIRE(mtx, curthread))
548 				break;
549 			owner = mtx->mtx_owner;
550 			continue;
551 		}
552 
553 		if (panicstr != NULL)
554 			return;
555 		if (MUTEX_OWNER(owner) == curthread)
556 			MUTEX_ABORT(mtx, "locking against myself");
557 
558 #ifdef MULTIPROCESSOR
559 		/*
560 		 * Check to see if the owner is running on a processor.
561 		 * If so, then we should just spin, as the owner will
562 		 * likely release the lock very soon.
563 		 */
564 		if (mutex_onproc(owner, &ci)) {
565 			LOCKSTAT_START_TIMER(lsflag, spintime);
566 			count = SPINLOCK_BACKOFF_MIN;
567 			for (;;) {
568 				SPINLOCK_BACKOFF(count);
569 				owner = mtx->mtx_owner;
570 				if (!mutex_onproc(owner, &ci))
571 					break;
572 			}
573 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
574 			LOCKSTAT_COUNT(spincnt, 1);
575 			if (!MUTEX_OWNED(owner))
576 				continue;
577 		}
578 #endif
579 
580 		ts = turnstile_lookup(mtx);
581 
582 		/*
583 		 * Once we have the turnstile chain interlock, mark the
584 		 * mutex has having waiters.  If that fails, spin again:
585 		 * chances are that the mutex has been released.
586 		 */
587 		if (!MUTEX_SET_WAITERS(mtx, owner)) {
588 			turnstile_exit(mtx);
589 			owner = mtx->mtx_owner;
590 			continue;
591 		}
592 
593 #ifdef MULTIPROCESSOR
594 		/*
595 		 * mutex_exit() is permitted to release the mutex without
596 		 * any interlocking instructions, and the following can
597 		 * occur as a result:
598 		 *
599 		 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
600 		 * ---------------------------- ----------------------------
601 		 *		..		    acquire cache line
602 		 *		..                   test for waiters
603 		 *	acquire cache line    <-      lose cache line
604 		 *	 lock cache line	           ..
605 		 *     verify mutex is held                ..
606 		 *	    set waiters  	           ..
607 		 *	 unlock cache line		   ..
608 		 *	  lose cache line     ->    acquire cache line
609 		 *		..	          clear lock word, waiters
610 		 *	  return success
611 		 *
612 		 * There is a another race that can occur: a third CPU could
613 		 * acquire the mutex as soon as it is released.  Since
614 		 * adaptive mutexes are primarily spin mutexes, this is not
615 		 * something that we need to worry about too much.  What we
616 		 * do need to ensure is that the waiters bit gets set.
617 		 *
618 		 * To allow the unlocked release, we need to make some
619 		 * assumptions here:
620 		 *
621 		 * o Release is the only non-atomic/unlocked operation
622 		 *   that can be performed on the mutex.  (It must still
623 		 *   be atomic on the local CPU, e.g. in case interrupted
624 		 *   or preempted).
625 		 *
626 		 * o At any given time, MUTEX_SET_WAITERS() can only ever
627 		 *   be in progress on one CPU in the system - guaranteed
628 		 *   by the turnstile chain lock.
629 		 *
630 		 * o No other operations other than MUTEX_SET_WAITERS()
631 		 *   and release can modify a mutex with a non-zero
632 		 *   owner field.
633 		 *
634 		 * o The result of a successful MUTEX_SET_WAITERS() call
635 		 *   is an unbuffered write that is immediately visible
636 		 *   to all other processors in the system.
637 		 *
638 		 * o If the holding LWP switches away, it posts a store
639 		 *   fence before changing curlwp, ensuring that any
640 		 *   overwrite of the mutex waiters flag by mutex_exit()
641 		 *   completes before the modification of curlwp becomes
642 		 *   visible to this CPU.
643 		 *
644 		 * o mi_switch() posts a store fence before setting curlwp
645 		 *   and before resuming execution of an LWP.
646 		 *
647 		 * o _kernel_lock() posts a store fence before setting
648 		 *   curcpu()->ci_biglock_wanted, and after clearing it.
649 		 *   This ensures that any overwrite of the mutex waiters
650 		 *   flag by mutex_exit() completes before the modification
651 		 *   of ci_biglock_wanted becomes visible.
652 		 *
653 		 * We now post a read memory barrier (after setting the
654 		 * waiters field) and check the lock holder's status again.
655 		 * Some of the possible outcomes (not an exhaustive list):
656 		 *
657 		 * 1. The onproc check returns true: the holding LWP is
658 		 *    running again.  The lock may be released soon and
659 		 *    we should spin.  Importantly, we can't trust the
660 		 *    value of the waiters flag.
661 		 *
662 		 * 2. The onproc check returns false: the holding LWP is
663 		 *    not running.  We now have the oppertunity to check
664 		 *    if mutex_exit() has blatted the modifications made
665 		 *    by MUTEX_SET_WAITERS().
666 		 *
667 		 * 3. The onproc check returns false: the holding LWP may
668 		 *    or may not be running.  It has context switched at
669 		 *    some point during our check.  Again, we have the
670 		 *    chance to see if the waiters bit is still set or
671 		 *    has been overwritten.
672 		 *
673 		 * 4. The onproc check returns false: the holding LWP is
674 		 *    running on a CPU, but wants the big lock.  It's OK
675 		 *    to check the waiters field in this case.
676 		 *
677 		 * 5. The has-waiters check fails: the mutex has been
678 		 *    released, the waiters flag cleared and another LWP
679 		 *    now owns the mutex.
680 		 *
681 		 * 6. The has-waiters check fails: the mutex has been
682 		 *    released.
683 		 *
684 		 * If the waiters bit is not set it's unsafe to go asleep,
685 		 * as we might never be awoken.
686 		 */
687 		if ((membar_consumer(), mutex_onproc(owner, &ci)) ||
688 		    (membar_consumer(), !MUTEX_HAS_WAITERS(mtx))) {
689 			turnstile_exit(mtx);
690 			owner = mtx->mtx_owner;
691 			continue;
692 		}
693 #endif	/* MULTIPROCESSOR */
694 
695 		LOCKSTAT_START_TIMER(lsflag, slptime);
696 
697 		turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj);
698 
699 		LOCKSTAT_STOP_TIMER(lsflag, slptime);
700 		LOCKSTAT_COUNT(slpcnt, 1);
701 
702 		owner = mtx->mtx_owner;
703 	}
704 
705 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1,
706 	    slpcnt, slptime);
707 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN,
708 	    spincnt, spintime);
709 	LOCKSTAT_EXIT(lsflag);
710 
711 	MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
712 	MUTEX_LOCKED(mtx);
713 }
714 
715 /*
716  * mutex_vector_exit:
717  *
718  *	Support routine for mutex_exit() that handles all cases.
719  */
720 void
721 mutex_vector_exit(kmutex_t *mtx)
722 {
723 	turnstile_t *ts;
724 	uintptr_t curthread;
725 
726 	if (MUTEX_SPIN_P(mtx)) {
727 #ifdef FULL
728 		if (__predict_false(!__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock))) {
729 			if (panicstr != NULL)
730 				return;
731 			MUTEX_ABORT(mtx, "exiting unheld spin mutex");
732 		}
733 		MUTEX_UNLOCKED(mtx);
734 		__cpu_simple_unlock(&mtx->mtx_lock);
735 #endif
736 		MUTEX_SPIN_SPLRESTORE(mtx);
737 		return;
738 	}
739 
740 	if (__predict_false((uintptr_t)panicstr | cold)) {
741 		MUTEX_UNLOCKED(mtx);
742 		MUTEX_RELEASE(mtx);
743 		return;
744 	}
745 
746 	curthread = (uintptr_t)curlwp;
747 	MUTEX_DASSERT(mtx, curthread != 0);
748 	MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
749 	MUTEX_UNLOCKED(mtx);
750 
751 #ifdef LOCKDEBUG
752 	/*
753 	 * Avoid having to take the turnstile chain lock every time
754 	 * around.  Raise the priority level to splhigh() in order
755 	 * to disable preemption and so make the following atomic.
756 	 */
757 	{
758 		int s = splhigh();
759 		if (!MUTEX_HAS_WAITERS(mtx)) {
760 			MUTEX_RELEASE(mtx);
761 			splx(s);
762 			return;
763 		}
764 		splx(s);
765 	}
766 #endif
767 
768 	/*
769 	 * Get this lock's turnstile.  This gets the interlock on
770 	 * the sleep queue.  Once we have that, we can clear the
771 	 * lock.  If there was no turnstile for the lock, there
772 	 * were no waiters remaining.
773 	 */
774 	ts = turnstile_lookup(mtx);
775 
776 	if (ts == NULL) {
777 		MUTEX_RELEASE(mtx);
778 		turnstile_exit(mtx);
779 	} else {
780 		MUTEX_RELEASE(mtx);
781 		turnstile_wakeup(ts, TS_WRITER_Q,
782 		    TS_WAITERS(ts, TS_WRITER_Q), NULL);
783 	}
784 }
785 
786 #ifndef __HAVE_SIMPLE_MUTEXES
787 /*
788  * mutex_wakeup:
789  *
790  *	Support routine for mutex_exit() that wakes up all waiters.
791  *	We assume that the mutex has been released, but it need not
792  *	be.
793  */
794 void
795 mutex_wakeup(kmutex_t *mtx)
796 {
797 	turnstile_t *ts;
798 
799 	ts = turnstile_lookup(mtx);
800 	if (ts == NULL) {
801 		turnstile_exit(mtx);
802 		return;
803 	}
804 	MUTEX_CLEAR_WAITERS(mtx);
805 	turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
806 }
807 #endif	/* !__HAVE_SIMPLE_MUTEXES */
808 
809 /*
810  * mutex_owned:
811  *
812  *	Return true if the current LWP (adaptive) or CPU (spin)
813  *	holds the mutex.
814  */
815 int
816 mutex_owned(kmutex_t *mtx)
817 {
818 
819 	if (MUTEX_ADAPTIVE_P(mtx))
820 		return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
821 #ifdef FULL
822 	return __SIMPLELOCK_LOCKED_P(&mtx->mtx_lock);
823 #else
824 	return 1;
825 #endif
826 }
827 
828 /*
829  * mutex_owner:
830  *
831  *	Return the current owner of an adaptive mutex.  Used for
832  *	priority inheritance.
833  */
834 lwp_t *
835 mutex_owner(kmutex_t *mtx)
836 {
837 
838 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx));
839 	return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
840 }
841 
842 /*
843  * mutex_tryenter:
844  *
845  *	Try to acquire the mutex; return non-zero if we did.
846  */
847 int
848 mutex_tryenter(kmutex_t *mtx)
849 {
850 	uintptr_t curthread;
851 
852 	/*
853 	 * Handle spin mutexes.
854 	 */
855 	if (MUTEX_SPIN_P(mtx)) {
856 		MUTEX_SPIN_SPLRAISE(mtx);
857 #ifdef FULL
858 		if (__cpu_simple_lock_try(&mtx->mtx_lock)) {
859 			MUTEX_WANTLOCK(mtx);
860 			MUTEX_LOCKED(mtx);
861 			return 1;
862 		}
863 		MUTEX_SPIN_SPLRESTORE(mtx);
864 #else
865 		MUTEX_WANTLOCK(mtx);
866 		MUTEX_LOCKED(mtx);
867 		return 1;
868 #endif
869 	} else {
870 		curthread = (uintptr_t)curlwp;
871 		MUTEX_ASSERT(mtx, curthread != 0);
872 		if (MUTEX_ACQUIRE(mtx, curthread)) {
873 			MUTEX_WANTLOCK(mtx);
874 			MUTEX_LOCKED(mtx);
875 			MUTEX_DASSERT(mtx,
876 			    MUTEX_OWNER(mtx->mtx_owner) == curthread);
877 			return 1;
878 		}
879 	}
880 
881 	return 0;
882 }
883 
884 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
885 /*
886  * mutex_spin_retry:
887  *
888  *	Support routine for mutex_spin_enter().  Assumes that the caller
889  *	has already raised the SPL, and adjusted counters.
890  */
891 void
892 mutex_spin_retry(kmutex_t *mtx)
893 {
894 #ifdef MULTIPROCESSOR
895 	u_int count;
896 	LOCKSTAT_TIMER(spintime);
897 	LOCKSTAT_FLAG(lsflag);
898 #ifdef LOCKDEBUG
899 	u_int spins = 0;
900 #endif	/* LOCKDEBUG */
901 
902 	MUTEX_WANTLOCK(mtx);
903 
904 	LOCKSTAT_ENTER(lsflag);
905 	LOCKSTAT_START_TIMER(lsflag, spintime);
906 	count = SPINLOCK_BACKOFF_MIN;
907 
908 	/*
909 	 * Spin testing the lock word and do exponential backoff
910 	 * to reduce cache line ping-ponging between CPUs.
911 	 */
912 	do {
913 		if (panicstr != NULL)
914 			break;
915 		while (__SIMPLELOCK_LOCKED_P(&mtx->mtx_lock)) {
916 			SPINLOCK_BACKOFF(count);
917 #ifdef LOCKDEBUG
918 			if (SPINLOCK_SPINOUT(spins))
919 				MUTEX_ABORT(mtx, "spinout");
920 #endif	/* LOCKDEBUG */
921 		}
922 	} while (!__cpu_simple_lock_try(&mtx->mtx_lock));
923 
924 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
925 	LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
926 	LOCKSTAT_EXIT(lsflag);
927 
928 	MUTEX_LOCKED(mtx);
929 #else	/* MULTIPROCESSOR */
930 	MUTEX_ABORT(mtx, "locking against myself");
931 #endif	/* MULTIPROCESSOR */
932 }
933 #endif	/* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
934 
935 /*
936  * mutex_obj_init:
937  *
938  *	Initialize the mutex object store.
939  */
940 void
941 mutex_obj_init(void)
942 {
943 
944 	mutex_obj_cache = pool_cache_init(sizeof(struct kmutexobj),
945 	    coherency_unit, 0, 0, "mutex", NULL, IPL_NONE, mutex_obj_ctor,
946 	    NULL, NULL);
947 }
948 
949 /*
950  * mutex_obj_ctor:
951  *
952  *	Initialize a new lock for the cache.
953  */
954 static int
955 mutex_obj_ctor(void *arg, void *obj, int flags)
956 {
957 	struct kmutexobj * mo = obj;
958 
959 	mo->mo_magic = MUTEX_OBJ_MAGIC;
960 
961 	return 0;
962 }
963 
964 /*
965  * mutex_obj_alloc:
966  *
967  *	Allocate a single lock object.
968  */
969 kmutex_t *
970 mutex_obj_alloc(kmutex_type_t type, int ipl)
971 {
972 	struct kmutexobj *mo;
973 
974 	mo = pool_cache_get(mutex_obj_cache, PR_WAITOK);
975 	mutex_init(&mo->mo_lock, type, ipl);
976 	mo->mo_refcnt = 1;
977 
978 	return (kmutex_t *)mo;
979 }
980 
981 /*
982  * mutex_obj_hold:
983  *
984  *	Add a single reference to a lock object.  A reference to the object
985  *	must already be held, and must be held across this call.
986  */
987 void
988 mutex_obj_hold(kmutex_t *lock)
989 {
990 	struct kmutexobj *mo = (struct kmutexobj *)lock;
991 
992 	KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC);
993 	KASSERT(mo->mo_refcnt > 0);
994 
995 	atomic_inc_uint(&mo->mo_refcnt);
996 }
997 
998 /*
999  * mutex_obj_free:
1000  *
1001  *	Drop a reference from a lock object.  If the last reference is being
1002  *	dropped, free the object and return true.  Otherwise, return false.
1003  */
1004 bool
1005 mutex_obj_free(kmutex_t *lock)
1006 {
1007 	struct kmutexobj *mo = (struct kmutexobj *)lock;
1008 
1009 	KASSERT(mo->mo_magic == MUTEX_OBJ_MAGIC);
1010 	KASSERT(mo->mo_refcnt > 0);
1011 
1012 	if (atomic_dec_uint_nv(&mo->mo_refcnt) > 0) {
1013 		return false;
1014 	}
1015 	mutex_destroy(&mo->mo_lock);
1016 	pool_cache_put(mutex_obj_cache, mo);
1017 	return true;
1018 }
1019