xref: /netbsd-src/sys/kern/kern_mutex.c (revision 627f7eb200a4419d89b531d55fccd2ee3ffdcde0)
1 /*	$NetBSD: kern_mutex.c,v 1.97 2021/04/03 14:56:14 thorpej Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008, 2019 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  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Kernel mutex implementation, modeled after those found in Solaris,
34  * a description of which can be found in:
35  *
36  *	Solaris Internals: Core Kernel Architecture, Jim Mauro and
37  *	    Richard McDougall.
38  */
39 
40 #define	__MUTEX_PRIVATE
41 
42 #include <sys/cdefs.h>
43 __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.97 2021/04/03 14:56:14 thorpej Exp $");
44 
45 #include <sys/param.h>
46 #include <sys/atomic.h>
47 #include <sys/proc.h>
48 #include <sys/mutex.h>
49 #include <sys/sched.h>
50 #include <sys/sleepq.h>
51 #include <sys/systm.h>
52 #include <sys/lockdebug.h>
53 #include <sys/kernel.h>
54 #include <sys/intr.h>
55 #include <sys/lock.h>
56 #include <sys/types.h>
57 #include <sys/cpu.h>
58 #include <sys/pserialize.h>
59 
60 #include <dev/lockstat.h>
61 
62 #include <machine/lock.h>
63 
64 /*
65  * When not running a debug kernel, spin mutexes are not much
66  * more than an splraiseipl() and splx() pair.
67  */
68 
69 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
70 #define	FULL
71 #endif
72 
73 /*
74  * Debugging support.
75  */
76 
77 #define	MUTEX_WANTLOCK(mtx)					\
78     LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx),		\
79         (uintptr_t)__builtin_return_address(0), 0)
80 #define	MUTEX_TESTLOCK(mtx)					\
81     LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx),		\
82         (uintptr_t)__builtin_return_address(0), -1)
83 #define	MUTEX_LOCKED(mtx)					\
84     LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), NULL,		\
85         (uintptr_t)__builtin_return_address(0), 0)
86 #define	MUTEX_UNLOCKED(mtx)					\
87     LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx),		\
88         (uintptr_t)__builtin_return_address(0), 0)
89 #define	MUTEX_ABORT(mtx, msg)					\
90     mutex_abort(__func__, __LINE__, mtx, msg)
91 
92 #if defined(LOCKDEBUG)
93 
94 #define	MUTEX_DASSERT(mtx, cond)				\
95 do {								\
96 	if (__predict_false(!(cond)))				\
97 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
98 } while (/* CONSTCOND */ 0)
99 
100 #else	/* LOCKDEBUG */
101 
102 #define	MUTEX_DASSERT(mtx, cond)	/* nothing */
103 
104 #endif /* LOCKDEBUG */
105 
106 #if defined(DIAGNOSTIC)
107 
108 #define	MUTEX_ASSERT(mtx, cond)					\
109 do {								\
110 	if (__predict_false(!(cond)))				\
111 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
112 } while (/* CONSTCOND */ 0)
113 
114 #else	/* DIAGNOSTIC */
115 
116 #define	MUTEX_ASSERT(mtx, cond)	/* nothing */
117 
118 #endif	/* DIAGNOSTIC */
119 
120 /*
121  * Some architectures can't use __cpu_simple_lock as is so allow a way
122  * for them to use an alternate definition.
123  */
124 #ifndef MUTEX_SPINBIT_LOCK_INIT
125 #define MUTEX_SPINBIT_LOCK_INIT(mtx)	__cpu_simple_lock_init(&(mtx)->mtx_lock)
126 #endif
127 #ifndef MUTEX_SPINBIT_LOCKED_P
128 #define MUTEX_SPINBIT_LOCKED_P(mtx)	__SIMPLELOCK_LOCKED_P(&(mtx)->mtx_lock)
129 #endif
130 #ifndef MUTEX_SPINBIT_LOCK_TRY
131 #define MUTEX_SPINBIT_LOCK_TRY(mtx)	__cpu_simple_lock_try(&(mtx)->mtx_lock)
132 #endif
133 #ifndef MUTEX_SPINBIT_LOCK_UNLOCK
134 #define MUTEX_SPINBIT_LOCK_UNLOCK(mtx)	__cpu_simple_unlock(&(mtx)->mtx_lock)
135 #endif
136 
137 #ifndef MUTEX_INITIALIZE_SPIN_IPL
138 #define MUTEX_INITIALIZE_SPIN_IPL(mtx, ipl) \
139 					((mtx)->mtx_ipl = makeiplcookie((ipl)))
140 #endif
141 
142 /*
143  * Spin mutex SPL save / restore.
144  */
145 
146 #define	MUTEX_SPIN_SPLRAISE(mtx)					\
147 do {									\
148 	struct cpu_info *x__ci;						\
149 	int x__cnt, s;							\
150 	s = splraiseipl(MUTEX_SPIN_IPL(mtx));				\
151 	x__ci = curcpu();						\
152 	x__cnt = x__ci->ci_mtx_count--;					\
153 	__insn_barrier();						\
154 	if (x__cnt == 0)						\
155 		x__ci->ci_mtx_oldspl = (s);				\
156 } while (/* CONSTCOND */ 0)
157 
158 #define	MUTEX_SPIN_SPLRESTORE(mtx)					\
159 do {									\
160 	struct cpu_info *x__ci = curcpu();				\
161 	int s = x__ci->ci_mtx_oldspl;					\
162 	__insn_barrier();						\
163 	if (++(x__ci->ci_mtx_count) == 0)				\
164 		splx(s);						\
165 } while (/* CONSTCOND */ 0)
166 
167 /*
168  * Memory barriers.
169  */
170 #ifdef __HAVE_ATOMIC_AS_MEMBAR
171 #define	MUTEX_MEMBAR_ENTER()
172 #define	MUTEX_MEMBAR_EXIT()
173 #else
174 #define	MUTEX_MEMBAR_ENTER()		membar_enter()
175 #define	MUTEX_MEMBAR_EXIT()		membar_exit()
176 #endif
177 
178 /*
179  * For architectures that provide 'simple' mutexes: they provide a
180  * CAS function that is either MP-safe, or does not need to be MP
181  * safe.  Adaptive mutexes on these architectures do not require an
182  * additional interlock.
183  */
184 
185 #ifdef __HAVE_SIMPLE_MUTEXES
186 
187 #define	MUTEX_OWNER(owner)						\
188 	(owner & MUTEX_THREAD)
189 #define	MUTEX_HAS_WAITERS(mtx)						\
190 	(((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
191 
192 #define	MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug)				\
193 do {									\
194 	if (!dodebug)							\
195 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
196 } while (/* CONSTCOND */ 0)
197 
198 #define	MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl)			\
199 do {									\
200 	(mtx)->mtx_owner = MUTEX_BIT_SPIN;				\
201 	if (!dodebug)							\
202 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
203 	MUTEX_INITIALIZE_SPIN_IPL((mtx), (ipl));			\
204 	MUTEX_SPINBIT_LOCK_INIT((mtx));					\
205 } while (/* CONSTCOND */ 0)
206 
207 #define	MUTEX_DESTROY(mtx)						\
208 do {									\
209 	(mtx)->mtx_owner = MUTEX_THREAD;				\
210 } while (/* CONSTCOND */ 0)
211 
212 #define	MUTEX_SPIN_P(owner)		\
213     (((owner) & MUTEX_BIT_SPIN) != 0)
214 #define	MUTEX_ADAPTIVE_P(owner)		\
215     (((owner) & MUTEX_BIT_SPIN) == 0)
216 
217 #define	MUTEX_DEBUG_P(mtx)	(((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0)
218 #if defined(LOCKDEBUG)
219 #define	MUTEX_OWNED(owner)		(((owner) & ~MUTEX_BIT_NODEBUG) != 0)
220 #define	MUTEX_INHERITDEBUG(n, o)	(n) |= (o) & MUTEX_BIT_NODEBUG
221 #else /* defined(LOCKDEBUG) */
222 #define	MUTEX_OWNED(owner)		((owner) != 0)
223 #define	MUTEX_INHERITDEBUG(n, o)	/* nothing */
224 #endif /* defined(LOCKDEBUG) */
225 
226 static inline int
227 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
228 {
229 	int rv;
230 	uintptr_t oldown = 0;
231 	uintptr_t newown = curthread;
232 
233 	MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner);
234 	MUTEX_INHERITDEBUG(newown, oldown);
235 	rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown);
236 	MUTEX_MEMBAR_ENTER();
237 	return rv;
238 }
239 
240 static inline int
241 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
242 {
243 	int rv;
244 	rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
245 	MUTEX_MEMBAR_ENTER();
246 	return rv;
247 }
248 
249 static inline void
250 MUTEX_RELEASE(kmutex_t *mtx)
251 {
252 	uintptr_t newown;
253 
254 	MUTEX_MEMBAR_EXIT();
255 	newown = 0;
256 	MUTEX_INHERITDEBUG(newown, mtx->mtx_owner);
257 	mtx->mtx_owner = newown;
258 }
259 #endif	/* __HAVE_SIMPLE_MUTEXES */
260 
261 /*
262  * Patch in stubs via strong alias where they are not available.
263  */
264 
265 #if defined(LOCKDEBUG)
266 #undef	__HAVE_MUTEX_STUBS
267 #undef	__HAVE_SPIN_MUTEX_STUBS
268 #endif
269 
270 #ifndef __HAVE_MUTEX_STUBS
271 __strong_alias(mutex_enter,mutex_vector_enter);
272 __strong_alias(mutex_exit,mutex_vector_exit);
273 #endif
274 
275 #ifndef __HAVE_SPIN_MUTEX_STUBS
276 __strong_alias(mutex_spin_enter,mutex_vector_enter);
277 __strong_alias(mutex_spin_exit,mutex_vector_exit);
278 #endif
279 
280 static void	mutex_abort(const char *, size_t, const kmutex_t *,
281     const char *);
282 static void	mutex_dump(const volatile void *, lockop_printer_t);
283 
284 lockops_t mutex_spin_lockops = {
285 	.lo_name = "Mutex",
286 	.lo_type = LOCKOPS_SPIN,
287 	.lo_dump = mutex_dump,
288 };
289 
290 lockops_t mutex_adaptive_lockops = {
291 	.lo_name = "Mutex",
292 	.lo_type = LOCKOPS_SLEEP,
293 	.lo_dump = mutex_dump,
294 };
295 
296 syncobj_t mutex_syncobj = {
297 	.sobj_flag	= SOBJ_SLEEPQ_SORTED,
298 	.sobj_unsleep	= turnstile_unsleep,
299 	.sobj_changepri	= turnstile_changepri,
300 	.sobj_lendpri	= sleepq_lendpri,
301 	.sobj_owner	= (void *)mutex_owner,
302 };
303 
304 /*
305  * mutex_dump:
306  *
307  *	Dump the contents of a mutex structure.
308  */
309 static void
310 mutex_dump(const volatile void *cookie, lockop_printer_t pr)
311 {
312 	const volatile kmutex_t *mtx = cookie;
313 	uintptr_t owner = mtx->mtx_owner;
314 
315 	pr("owner field  : %#018lx wait/spin: %16d/%d\n",
316 	    (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx),
317 	    MUTEX_SPIN_P(owner));
318 }
319 
320 /*
321  * mutex_abort:
322  *
323  *	Dump information about an error and panic the system.  This
324  *	generates a lot of machine code in the DIAGNOSTIC case, so
325  *	we ask the compiler to not inline it.
326  */
327 static void __noinline
328 mutex_abort(const char *func, size_t line, const kmutex_t *mtx, const char *msg)
329 {
330 
331 	LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ?
332 	    &mutex_spin_lockops : &mutex_adaptive_lockops), msg);
333 }
334 
335 /*
336  * mutex_init:
337  *
338  *	Initialize a mutex for use.  Note that adaptive mutexes are in
339  *	essence spin mutexes that can sleep to avoid deadlock and wasting
340  *	CPU time.  We can't easily provide a type of mutex that always
341  *	sleeps - see comments in mutex_vector_enter() about releasing
342  *	mutexes unlocked.
343  */
344 void _mutex_init(kmutex_t *, kmutex_type_t, int, uintptr_t);
345 void
346 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl,
347     uintptr_t return_address)
348 {
349 	lockops_t *lockops __unused;
350 	bool dodebug;
351 
352 	memset(mtx, 0, sizeof(*mtx));
353 
354 	if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK ||
355 	    ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET ||
356 	    ipl == IPL_SOFTSERIAL) {
357 		lockops = (type == MUTEX_NODEBUG ?
358 		    NULL : &mutex_adaptive_lockops);
359 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
360 		MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug);
361 	} else {
362 		lockops = (type == MUTEX_NODEBUG ?
363 		    NULL : &mutex_spin_lockops);
364 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
365 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
366 	}
367 }
368 
369 void
370 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
371 {
372 
373 	_mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0));
374 }
375 
376 /*
377  * mutex_destroy:
378  *
379  *	Tear down a mutex.
380  */
381 void
382 mutex_destroy(kmutex_t *mtx)
383 {
384 	uintptr_t owner = mtx->mtx_owner;
385 
386 	if (MUTEX_ADAPTIVE_P(owner)) {
387 		MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner));
388 		MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx));
389 	} else {
390 		MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx));
391 	}
392 
393 	LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx);
394 	MUTEX_DESTROY(mtx);
395 }
396 
397 #ifdef MULTIPROCESSOR
398 /*
399  * mutex_oncpu:
400  *
401  *	Return true if an adaptive mutex owner is running on a CPU in the
402  *	system.  If the target is waiting on the kernel big lock, then we
403  *	must release it.  This is necessary to avoid deadlock.
404  */
405 static bool
406 mutex_oncpu(uintptr_t owner)
407 {
408 	struct cpu_info *ci;
409 	lwp_t *l;
410 
411 	KASSERT(kpreempt_disabled());
412 
413 	if (!MUTEX_OWNED(owner)) {
414 		return false;
415 	}
416 
417 	/*
418 	 * See lwp_dtor() why dereference of the LWP pointer is safe.
419 	 * We must have kernel preemption disabled for that.
420 	 */
421 	l = (lwp_t *)MUTEX_OWNER(owner);
422 	ci = l->l_cpu;
423 
424 	if (ci && ci->ci_curlwp == l) {
425 		/* Target is running; do we need to block? */
426 		return (ci->ci_biglock_wanted != l);
427 	}
428 
429 	/* Not running.  It may be safe to block now. */
430 	return false;
431 }
432 #endif	/* MULTIPROCESSOR */
433 
434 /*
435  * mutex_vector_enter:
436  *
437  *	Support routine for mutex_enter() that must handle all cases.  In
438  *	the LOCKDEBUG case, mutex_enter() is always aliased here, even if
439  *	fast-path stubs are available.  If a mutex_spin_enter() stub is
440  *	not available, then it is also aliased directly here.
441  */
442 void
443 mutex_vector_enter(kmutex_t *mtx)
444 {
445 	uintptr_t owner, curthread;
446 	turnstile_t *ts;
447 #ifdef MULTIPROCESSOR
448 	u_int count;
449 #endif
450 	LOCKSTAT_COUNTER(spincnt);
451 	LOCKSTAT_COUNTER(slpcnt);
452 	LOCKSTAT_TIMER(spintime);
453 	LOCKSTAT_TIMER(slptime);
454 	LOCKSTAT_FLAG(lsflag);
455 
456 	/*
457 	 * Handle spin mutexes.
458 	 */
459 	KPREEMPT_DISABLE(curlwp);
460 	owner = mtx->mtx_owner;
461 	if (MUTEX_SPIN_P(owner)) {
462 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
463 		u_int spins = 0;
464 #endif
465 		KPREEMPT_ENABLE(curlwp);
466 		MUTEX_SPIN_SPLRAISE(mtx);
467 		MUTEX_WANTLOCK(mtx);
468 #ifdef FULL
469 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
470 			MUTEX_LOCKED(mtx);
471 			return;
472 		}
473 #if !defined(MULTIPROCESSOR)
474 		MUTEX_ABORT(mtx, "locking against myself");
475 #else /* !MULTIPROCESSOR */
476 
477 		LOCKSTAT_ENTER(lsflag);
478 		LOCKSTAT_START_TIMER(lsflag, spintime);
479 		count = SPINLOCK_BACKOFF_MIN;
480 
481 		/*
482 		 * Spin testing the lock word and do exponential backoff
483 		 * to reduce cache line ping-ponging between CPUs.
484 		 */
485 		do {
486 			while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
487 				SPINLOCK_SPIN_HOOK;
488 				SPINLOCK_BACKOFF(count);
489 #ifdef LOCKDEBUG
490 				if (SPINLOCK_SPINOUT(spins))
491 					MUTEX_ABORT(mtx, "spinout");
492 #endif	/* LOCKDEBUG */
493 			}
494 		} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
495 
496 		if (count != SPINLOCK_BACKOFF_MIN) {
497 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
498 			LOCKSTAT_EVENT(lsflag, mtx,
499 			    LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
500 		}
501 		LOCKSTAT_EXIT(lsflag);
502 #endif	/* !MULTIPROCESSOR */
503 #endif	/* FULL */
504 		MUTEX_LOCKED(mtx);
505 		return;
506 	}
507 
508 	curthread = (uintptr_t)curlwp;
509 
510 	MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner));
511 	MUTEX_ASSERT(mtx, curthread != 0);
512 	MUTEX_ASSERT(mtx, !cpu_intr_p());
513 	MUTEX_WANTLOCK(mtx);
514 
515 	if (panicstr == NULL) {
516 		KDASSERT(pserialize_not_in_read_section());
517 		LOCKDEBUG_BARRIER(&kernel_lock, 1);
518 	}
519 
520 	LOCKSTAT_ENTER(lsflag);
521 
522 	/*
523 	 * Adaptive mutex; spin trying to acquire the mutex.  If we
524 	 * determine that the owner is not running on a processor,
525 	 * then we stop spinning, and sleep instead.
526 	 */
527 	for (;;) {
528 		if (!MUTEX_OWNED(owner)) {
529 			/*
530 			 * Mutex owner clear could mean two things:
531 			 *
532 			 *	* The mutex has been released.
533 			 *	* The owner field hasn't been set yet.
534 			 *
535 			 * Try to acquire it again.  If that fails,
536 			 * we'll just loop again.
537 			 */
538 			if (MUTEX_ACQUIRE(mtx, curthread))
539 				break;
540 			owner = mtx->mtx_owner;
541 			continue;
542 		}
543 		if (__predict_false(MUTEX_OWNER(owner) == curthread)) {
544 			MUTEX_ABORT(mtx, "locking against myself");
545 		}
546 #ifdef MULTIPROCESSOR
547 		/*
548 		 * Check to see if the owner is running on a processor.
549 		 * If so, then we should just spin, as the owner will
550 		 * likely release the lock very soon.
551 		 */
552 		if (mutex_oncpu(owner)) {
553 			LOCKSTAT_START_TIMER(lsflag, spintime);
554 			count = SPINLOCK_BACKOFF_MIN;
555 			do {
556 				KPREEMPT_ENABLE(curlwp);
557 				SPINLOCK_BACKOFF(count);
558 				KPREEMPT_DISABLE(curlwp);
559 				owner = mtx->mtx_owner;
560 			} while (mutex_oncpu(owner));
561 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
562 			LOCKSTAT_COUNT(spincnt, 1);
563 			if (!MUTEX_OWNED(owner))
564 				continue;
565 		}
566 #endif
567 
568 		ts = turnstile_lookup(mtx);
569 
570 		/*
571 		 * Once we have the turnstile chain interlock, mark the
572 		 * mutex as having waiters.  If that fails, spin again:
573 		 * chances are that the mutex has been released.
574 		 */
575 		if (!MUTEX_SET_WAITERS(mtx, owner)) {
576 			turnstile_exit(mtx);
577 			owner = mtx->mtx_owner;
578 			continue;
579 		}
580 
581 #ifdef MULTIPROCESSOR
582 		/*
583 		 * mutex_exit() is permitted to release the mutex without
584 		 * any interlocking instructions, and the following can
585 		 * occur as a result:
586 		 *
587 		 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
588 		 * ---------------------------- ----------------------------
589 		 *		..		    acquire cache line
590 		 *		..                   test for waiters
591 		 *	acquire cache line    <-      lose cache line
592 		 *	 lock cache line	           ..
593 		 *     verify mutex is held                ..
594 		 *	    set waiters  	           ..
595 		 *	 unlock cache line		   ..
596 		 *	  lose cache line     ->    acquire cache line
597 		 *		..	          clear lock word, waiters
598 		 *	  return success
599 		 *
600 		 * There is another race that can occur: a third CPU could
601 		 * acquire the mutex as soon as it is released.  Since
602 		 * adaptive mutexes are primarily spin mutexes, this is not
603 		 * something that we need to worry about too much.  What we
604 		 * do need to ensure is that the waiters bit gets set.
605 		 *
606 		 * To allow the unlocked release, we need to make some
607 		 * assumptions here:
608 		 *
609 		 * o Release is the only non-atomic/unlocked operation
610 		 *   that can be performed on the mutex.  (It must still
611 		 *   be atomic on the local CPU, e.g. in case interrupted
612 		 *   or preempted).
613 		 *
614 		 * o At any given time, MUTEX_SET_WAITERS() can only ever
615 		 *   be in progress on one CPU in the system - guaranteed
616 		 *   by the turnstile chain lock.
617 		 *
618 		 * o No other operations other than MUTEX_SET_WAITERS()
619 		 *   and release can modify a mutex with a non-zero
620 		 *   owner field.
621 		 *
622 		 * o The result of a successful MUTEX_SET_WAITERS() call
623 		 *   is an unbuffered write that is immediately visible
624 		 *   to all other processors in the system.
625 		 *
626 		 * o If the holding LWP switches away, it posts a store
627 		 *   fence before changing curlwp, ensuring that any
628 		 *   overwrite of the mutex waiters flag by mutex_exit()
629 		 *   completes before the modification of curlwp becomes
630 		 *   visible to this CPU.
631 		 *
632 		 * o cpu_switchto() posts a store fence after setting curlwp
633 		 *   and before resuming execution of an LWP.
634 		 *
635 		 * o _kernel_lock() posts a store fence before setting
636 		 *   curcpu()->ci_biglock_wanted, and after clearing it.
637 		 *   This ensures that any overwrite of the mutex waiters
638 		 *   flag by mutex_exit() completes before the modification
639 		 *   of ci_biglock_wanted becomes visible.
640 		 *
641 		 * We now post a read memory barrier (after setting the
642 		 * waiters field) and check the lock holder's status again.
643 		 * Some of the possible outcomes (not an exhaustive list):
644 		 *
645 		 * 1. The on-CPU check returns true: the holding LWP is
646 		 *    running again.  The lock may be released soon and
647 		 *    we should spin.  Importantly, we can't trust the
648 		 *    value of the waiters flag.
649 		 *
650 		 * 2. The on-CPU check returns false: the holding LWP is
651 		 *    not running.  We now have the opportunity to check
652 		 *    if mutex_exit() has blatted the modifications made
653 		 *    by MUTEX_SET_WAITERS().
654 		 *
655 		 * 3. The on-CPU check returns false: the holding LWP may
656 		 *    or may not be running.  It has context switched at
657 		 *    some point during our check.  Again, we have the
658 		 *    chance to see if the waiters bit is still set or
659 		 *    has been overwritten.
660 		 *
661 		 * 4. The on-CPU check returns false: the holding LWP is
662 		 *    running on a CPU, but wants the big lock.  It's OK
663 		 *    to check the waiters field in this case.
664 		 *
665 		 * 5. The has-waiters check fails: the mutex has been
666 		 *    released, the waiters flag cleared and another LWP
667 		 *    now owns the mutex.
668 		 *
669 		 * 6. The has-waiters check fails: the mutex has been
670 		 *    released.
671 		 *
672 		 * If the waiters bit is not set it's unsafe to go asleep,
673 		 * as we might never be awoken.
674 		 */
675 		membar_consumer();
676 		if (mutex_oncpu(owner)) {
677 			turnstile_exit(mtx);
678 			owner = mtx->mtx_owner;
679 			continue;
680 		}
681 		membar_consumer();
682 		if (!MUTEX_HAS_WAITERS(mtx)) {
683 			turnstile_exit(mtx);
684 			owner = mtx->mtx_owner;
685 			continue;
686 		}
687 #endif	/* MULTIPROCESSOR */
688 
689 		LOCKSTAT_START_TIMER(lsflag, slptime);
690 
691 		turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj);
692 
693 		LOCKSTAT_STOP_TIMER(lsflag, slptime);
694 		LOCKSTAT_COUNT(slpcnt, 1);
695 
696 		owner = mtx->mtx_owner;
697 	}
698 	KPREEMPT_ENABLE(curlwp);
699 
700 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1,
701 	    slpcnt, slptime);
702 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN,
703 	    spincnt, spintime);
704 	LOCKSTAT_EXIT(lsflag);
705 
706 	MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
707 	MUTEX_LOCKED(mtx);
708 }
709 
710 /*
711  * mutex_vector_exit:
712  *
713  *	Support routine for mutex_exit() that handles all cases.
714  */
715 void
716 mutex_vector_exit(kmutex_t *mtx)
717 {
718 	turnstile_t *ts;
719 	uintptr_t curthread;
720 
721 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
722 #ifdef FULL
723 		if (__predict_false(!MUTEX_SPINBIT_LOCKED_P(mtx))) {
724 			MUTEX_ABORT(mtx, "exiting unheld spin mutex");
725 		}
726 		MUTEX_UNLOCKED(mtx);
727 		MUTEX_SPINBIT_LOCK_UNLOCK(mtx);
728 #endif
729 		MUTEX_SPIN_SPLRESTORE(mtx);
730 		return;
731 	}
732 
733 #ifndef __HAVE_MUTEX_STUBS
734 	/*
735 	 * On some architectures without mutex stubs, we can enter here to
736 	 * release mutexes before interrupts and whatnot are up and running.
737 	 * We need this hack to keep them sweet.
738 	 */
739 	if (__predict_false(cold)) {
740 		MUTEX_UNLOCKED(mtx);
741 		MUTEX_RELEASE(mtx);
742 		return;
743 	}
744 #endif
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 #if !defined(LOCKDEBUG)
751 	__USE(curthread);
752 #endif
753 
754 #ifdef LOCKDEBUG
755 	/*
756 	 * Avoid having to take the turnstile chain lock every time
757 	 * around.  Raise the priority level to splhigh() in order
758 	 * to disable preemption and so make the following atomic.
759 	 */
760 	{
761 		int s = splhigh();
762 		if (!MUTEX_HAS_WAITERS(mtx)) {
763 			MUTEX_RELEASE(mtx);
764 			splx(s);
765 			return;
766 		}
767 		splx(s);
768 	}
769 #endif
770 
771 	/*
772 	 * Get this lock's turnstile.  This gets the interlock on
773 	 * the sleep queue.  Once we have that, we can clear the
774 	 * lock.  If there was no turnstile for the lock, there
775 	 * were no waiters remaining.
776 	 */
777 	ts = turnstile_lookup(mtx);
778 
779 	if (ts == NULL) {
780 		MUTEX_RELEASE(mtx);
781 		turnstile_exit(mtx);
782 	} else {
783 		MUTEX_RELEASE(mtx);
784 		turnstile_wakeup(ts, TS_WRITER_Q,
785 		    TS_WAITERS(ts, TS_WRITER_Q), NULL);
786 	}
787 }
788 
789 #ifndef __HAVE_SIMPLE_MUTEXES
790 /*
791  * mutex_wakeup:
792  *
793  *	Support routine for mutex_exit() that wakes up all waiters.
794  *	We assume that the mutex has been released, but it need not
795  *	be.
796  */
797 void
798 mutex_wakeup(kmutex_t *mtx)
799 {
800 	turnstile_t *ts;
801 
802 	ts = turnstile_lookup(mtx);
803 	if (ts == NULL) {
804 		turnstile_exit(mtx);
805 		return;
806 	}
807 	MUTEX_CLEAR_WAITERS(mtx);
808 	turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
809 }
810 #endif	/* !__HAVE_SIMPLE_MUTEXES */
811 
812 /*
813  * mutex_owned:
814  *
815  *	Return true if the current LWP (adaptive) or CPU (spin)
816  *	holds the mutex.
817  */
818 int
819 mutex_owned(const kmutex_t *mtx)
820 {
821 
822 	if (mtx == NULL)
823 		return 0;
824 	if (MUTEX_ADAPTIVE_P(mtx->mtx_owner))
825 		return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
826 #ifdef FULL
827 	return MUTEX_SPINBIT_LOCKED_P(mtx);
828 #else
829 	return 1;
830 #endif
831 }
832 
833 /*
834  * mutex_owner:
835  *
836  *	Return the current owner of an adaptive mutex.  Used for
837  *	priority inheritance.
838  */
839 lwp_t *
840 mutex_owner(const kmutex_t *mtx)
841 {
842 
843 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
844 	return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
845 }
846 
847 /*
848  * mutex_owner_running:
849  *
850  *	Return true if an adaptive mutex is unheld, or held and the owner is
851  *	running on a CPU.  For the pagedaemon only - do not document or use
852  *	in other code.
853  */
854 bool
855 mutex_owner_running(const kmutex_t *mtx)
856 {
857 #ifdef MULTIPROCESSOR
858 	uintptr_t owner;
859 	bool rv;
860 
861 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
862 	kpreempt_disable();
863 	owner = mtx->mtx_owner;
864 	rv = !MUTEX_OWNED(owner) || mutex_oncpu(MUTEX_OWNER(owner));
865 	kpreempt_enable();
866 	return rv;
867 #else
868 	return mutex_owner(mtx) == curlwp;
869 #endif
870 }
871 
872 /*
873  * mutex_ownable:
874  *
875  *	When compiled with DEBUG and LOCKDEBUG defined, ensure that
876  *	the mutex is available.  We cannot use !mutex_owned() since
877  *	that won't work correctly for spin mutexes.
878  */
879 int
880 mutex_ownable(const kmutex_t *mtx)
881 {
882 
883 #ifdef LOCKDEBUG
884 	MUTEX_TESTLOCK(mtx);
885 #endif
886 	return 1;
887 }
888 
889 /*
890  * mutex_tryenter:
891  *
892  *	Try to acquire the mutex; return non-zero if we did.
893  */
894 int
895 mutex_tryenter(kmutex_t *mtx)
896 {
897 	uintptr_t curthread;
898 
899 	/*
900 	 * Handle spin mutexes.
901 	 */
902 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
903 		MUTEX_SPIN_SPLRAISE(mtx);
904 #ifdef FULL
905 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
906 			MUTEX_WANTLOCK(mtx);
907 			MUTEX_LOCKED(mtx);
908 			return 1;
909 		}
910 		MUTEX_SPIN_SPLRESTORE(mtx);
911 #else
912 		MUTEX_WANTLOCK(mtx);
913 		MUTEX_LOCKED(mtx);
914 		return 1;
915 #endif
916 	} else {
917 		curthread = (uintptr_t)curlwp;
918 		MUTEX_ASSERT(mtx, curthread != 0);
919 		if (MUTEX_ACQUIRE(mtx, curthread)) {
920 			MUTEX_WANTLOCK(mtx);
921 			MUTEX_LOCKED(mtx);
922 			MUTEX_DASSERT(mtx,
923 			    MUTEX_OWNER(mtx->mtx_owner) == curthread);
924 			return 1;
925 		}
926 	}
927 
928 	return 0;
929 }
930 
931 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
932 /*
933  * mutex_spin_retry:
934  *
935  *	Support routine for mutex_spin_enter().  Assumes that the caller
936  *	has already raised the SPL, and adjusted counters.
937  */
938 void
939 mutex_spin_retry(kmutex_t *mtx)
940 {
941 #ifdef MULTIPROCESSOR
942 	u_int count;
943 	LOCKSTAT_TIMER(spintime);
944 	LOCKSTAT_FLAG(lsflag);
945 #ifdef LOCKDEBUG
946 	u_int spins = 0;
947 #endif	/* LOCKDEBUG */
948 
949 	MUTEX_WANTLOCK(mtx);
950 
951 	LOCKSTAT_ENTER(lsflag);
952 	LOCKSTAT_START_TIMER(lsflag, spintime);
953 	count = SPINLOCK_BACKOFF_MIN;
954 
955 	/*
956 	 * Spin testing the lock word and do exponential backoff
957 	 * to reduce cache line ping-ponging between CPUs.
958 	 */
959 	do {
960 		while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
961 			SPINLOCK_BACKOFF(count);
962 #ifdef LOCKDEBUG
963 			if (SPINLOCK_SPINOUT(spins))
964 				MUTEX_ABORT(mtx, "spinout");
965 #endif	/* LOCKDEBUG */
966 		}
967 	} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
968 
969 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
970 	LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
971 	LOCKSTAT_EXIT(lsflag);
972 
973 	MUTEX_LOCKED(mtx);
974 #else	/* MULTIPROCESSOR */
975 	MUTEX_ABORT(mtx, "locking against myself");
976 #endif	/* MULTIPROCESSOR */
977 }
978 #endif	/* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
979