xref: /netbsd-src/sys/kern/kern_mutex.c (revision 4724848cf0da353df257f730694b7882798e5daf)
1 /*	$NetBSD: kern_mutex.c,v 1.105 2023/04/12 06:35:40 riastradh 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 1.105 2023/04/12 06:35:40 riastradh 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 	const int s = splraiseipl(MUTEX_SPIN_IPL(mtx));			\
149 	struct cpu_info * const x__ci = curcpu();			\
150 	const int x__cnt = x__ci->ci_mtx_count--;			\
151 	__insn_barrier();						\
152 	if (x__cnt == 0)						\
153 		x__ci->ci_mtx_oldspl = s;				\
154 } while (/* CONSTCOND */ 0)
155 
156 #define	MUTEX_SPIN_SPLRESTORE(mtx)					\
157 do {									\
158 	struct cpu_info * const x__ci = curcpu();			\
159 	const int s = x__ci->ci_mtx_oldspl;				\
160 	__insn_barrier();						\
161 	if (++(x__ci->ci_mtx_count) == 0)				\
162 		splx(s);						\
163 } while (/* CONSTCOND */ 0)
164 
165 /*
166  * Memory barriers.
167  */
168 #ifdef __HAVE_ATOMIC_AS_MEMBAR
169 #define	MUTEX_MEMBAR_ENTER()
170 #else
171 #define	MUTEX_MEMBAR_ENTER()		membar_enter()
172 #endif
173 
174 /*
175  * For architectures that provide 'simple' mutexes: they provide a
176  * CAS function that is either MP-safe, or does not need to be MP
177  * safe.  Adaptive mutexes on these architectures do not require an
178  * additional interlock.
179  */
180 
181 #ifdef __HAVE_SIMPLE_MUTEXES
182 
183 #define	MUTEX_OWNER(owner)						\
184 	(owner & MUTEX_THREAD)
185 #define	MUTEX_HAS_WAITERS(mtx)						\
186 	(((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
187 
188 #define	MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug)				\
189 do {									\
190 	if (!dodebug)							\
191 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
192 } while (/* CONSTCOND */ 0)
193 
194 #define	MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl)			\
195 do {									\
196 	(mtx)->mtx_owner = MUTEX_BIT_SPIN;				\
197 	if (!dodebug)							\
198 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
199 	MUTEX_INITIALIZE_SPIN_IPL((mtx), (ipl));			\
200 	MUTEX_SPINBIT_LOCK_INIT((mtx));					\
201 } while (/* CONSTCOND */ 0)
202 
203 #define	MUTEX_DESTROY(mtx)						\
204 do {									\
205 	(mtx)->mtx_owner = MUTEX_THREAD;				\
206 } while (/* CONSTCOND */ 0)
207 
208 #define	MUTEX_SPIN_P(owner)		\
209     (((owner) & MUTEX_BIT_SPIN) != 0)
210 #define	MUTEX_ADAPTIVE_P(owner)		\
211     (((owner) & MUTEX_BIT_SPIN) == 0)
212 
213 #ifndef MUTEX_CAS
214 #define	MUTEX_CAS(p, o, n)		\
215 	(atomic_cas_ulong((volatile unsigned long *)(p), (o), (n)) == (o))
216 #endif /* MUTEX_CAS */
217 
218 #define	MUTEX_DEBUG_P(mtx)	(((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0)
219 #if defined(LOCKDEBUG)
220 #define	MUTEX_OWNED(owner)		(((owner) & ~MUTEX_BIT_NODEBUG) != 0)
221 #define	MUTEX_INHERITDEBUG(n, o)	(n) |= (o) & MUTEX_BIT_NODEBUG
222 #else /* defined(LOCKDEBUG) */
223 #define	MUTEX_OWNED(owner)		((owner) != 0)
224 #define	MUTEX_INHERITDEBUG(n, o)	/* nothing */
225 #endif /* defined(LOCKDEBUG) */
226 
227 static inline int
228 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
229 {
230 	int rv;
231 	uintptr_t oldown = 0;
232 	uintptr_t newown = curthread;
233 
234 	MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner);
235 	MUTEX_INHERITDEBUG(newown, oldown);
236 	rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown);
237 	membar_acquire();
238 	return rv;
239 }
240 
241 static inline int
242 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
243 {
244 	int rv;
245 	rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
246 	MUTEX_MEMBAR_ENTER();
247 	return rv;
248 }
249 
250 static inline void
251 MUTEX_RELEASE(kmutex_t *mtx)
252 {
253 	uintptr_t newown;
254 
255 	newown = 0;
256 	MUTEX_INHERITDEBUG(newown, mtx->mtx_owner);
257 	atomic_store_release(&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, volatile const kmutex_t *,
281 		    const char *);
282 static void	mutex_dump(const volatile void *, lockop_printer_t);
283 static lwp_t	*mutex_owner(wchan_t);
284 
285 lockops_t mutex_spin_lockops = {
286 	.lo_name = "Mutex",
287 	.lo_type = LOCKOPS_SPIN,
288 	.lo_dump = mutex_dump,
289 };
290 
291 lockops_t mutex_adaptive_lockops = {
292 	.lo_name = "Mutex",
293 	.lo_type = LOCKOPS_SLEEP,
294 	.lo_dump = mutex_dump,
295 };
296 
297 syncobj_t mutex_syncobj = {
298 	.sobj_flag	= SOBJ_SLEEPQ_SORTED,
299 	.sobj_unsleep	= turnstile_unsleep,
300 	.sobj_changepri	= turnstile_changepri,
301 	.sobj_lendpri	= sleepq_lendpri,
302 	.sobj_owner	= mutex_owner,
303 };
304 
305 /*
306  * mutex_dump:
307  *
308  *	Dump the contents of a mutex structure.
309  */
310 static void
311 mutex_dump(const volatile void *cookie, lockop_printer_t pr)
312 {
313 	const volatile kmutex_t *mtx = cookie;
314 	uintptr_t owner = mtx->mtx_owner;
315 
316 	pr("owner field  : %#018lx wait/spin: %16d/%d\n",
317 	    (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx),
318 	    MUTEX_SPIN_P(owner));
319 }
320 
321 /*
322  * mutex_abort:
323  *
324  *	Dump information about an error and panic the system.  This
325  *	generates a lot of machine code in the DIAGNOSTIC case, so
326  *	we ask the compiler to not inline it.
327  */
328 static void __noinline
329 mutex_abort(const char *func, size_t line, volatile const kmutex_t *mtx,
330     const char *msg)
331 {
332 
333 	LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ?
334 	    &mutex_spin_lockops : &mutex_adaptive_lockops), msg);
335 }
336 
337 /*
338  * mutex_init:
339  *
340  *	Initialize a mutex for use.  Note that adaptive mutexes are in
341  *	essence spin mutexes that can sleep to avoid deadlock and wasting
342  *	CPU time.  We can't easily provide a type of mutex that always
343  *	sleeps - see comments in mutex_vector_enter() about releasing
344  *	mutexes unlocked.
345  */
346 void
347 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl,
348     uintptr_t return_address)
349 {
350 	lockops_t *lockops __unused;
351 	bool dodebug;
352 
353 	memset(mtx, 0, sizeof(*mtx));
354 
355 	if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK ||
356 	    ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET ||
357 	    ipl == IPL_SOFTSERIAL) {
358 		lockops = (type == MUTEX_NODEBUG ?
359 		    NULL : &mutex_adaptive_lockops);
360 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
361 		MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug);
362 	} else {
363 		lockops = (type == MUTEX_NODEBUG ?
364 		    NULL : &mutex_spin_lockops);
365 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
366 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
367 	}
368 }
369 
370 void
371 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
372 {
373 
374 	_mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0));
375 }
376 
377 /*
378  * mutex_destroy:
379  *
380  *	Tear down a mutex.
381  */
382 void
383 mutex_destroy(kmutex_t *mtx)
384 {
385 	uintptr_t owner = mtx->mtx_owner;
386 
387 	if (MUTEX_ADAPTIVE_P(owner)) {
388 		MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner));
389 		MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx));
390 	} else {
391 		MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx));
392 	}
393 
394 	LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx);
395 	MUTEX_DESTROY(mtx);
396 }
397 
398 #ifdef MULTIPROCESSOR
399 /*
400  * mutex_oncpu:
401  *
402  *	Return true if an adaptive mutex owner is running on a CPU in the
403  *	system.  If the target is waiting on the kernel big lock, then we
404  *	must release it.  This is necessary to avoid deadlock.
405  */
406 static bool
407 mutex_oncpu(uintptr_t owner)
408 {
409 	struct cpu_info *ci;
410 	lwp_t *l;
411 
412 	KASSERT(kpreempt_disabled());
413 
414 	if (!MUTEX_OWNED(owner)) {
415 		return false;
416 	}
417 
418 	/*
419 	 * See lwp_dtor() why dereference of the LWP pointer is safe.
420 	 * We must have kernel preemption disabled for that.
421 	 */
422 	l = (lwp_t *)MUTEX_OWNER(owner);
423 	ci = l->l_cpu;
424 
425 	if (ci && ci->ci_curlwp == l) {
426 		/* Target is running; do we need to block? */
427 		return (atomic_load_relaxed(&ci->ci_biglock_wanted) != l);
428 	}
429 
430 	/* Not running.  It may be safe to block now. */
431 	return false;
432 }
433 #endif	/* MULTIPROCESSOR */
434 
435 /*
436  * mutex_vector_enter:
437  *
438  *	Support routine for mutex_enter() that must handle all cases.  In
439  *	the LOCKDEBUG case, mutex_enter() is always aliased here, even if
440  *	fast-path stubs are available.  If a mutex_spin_enter() stub is
441  *	not available, then it is also aliased directly here.
442  */
443 void
444 mutex_vector_enter(kmutex_t *mtx)
445 {
446 	uintptr_t owner, curthread;
447 	turnstile_t *ts;
448 #ifdef MULTIPROCESSOR
449 	u_int count;
450 #endif
451 	LOCKSTAT_COUNTER(spincnt);
452 	LOCKSTAT_COUNTER(slpcnt);
453 	LOCKSTAT_TIMER(spintime);
454 	LOCKSTAT_TIMER(slptime);
455 	LOCKSTAT_FLAG(lsflag);
456 
457 	/*
458 	 * Handle spin mutexes.
459 	 */
460 	KPREEMPT_DISABLE(curlwp);
461 	owner = mtx->mtx_owner;
462 	if (MUTEX_SPIN_P(owner)) {
463 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
464 		u_int spins = 0;
465 #endif
466 		KPREEMPT_ENABLE(curlwp);
467 		MUTEX_SPIN_SPLRAISE(mtx);
468 		MUTEX_WANTLOCK(mtx);
469 #ifdef FULL
470 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
471 			MUTEX_LOCKED(mtx);
472 			return;
473 		}
474 #if !defined(MULTIPROCESSOR)
475 		MUTEX_ABORT(mtx, "locking against myself");
476 #else /* !MULTIPROCESSOR */
477 
478 		LOCKSTAT_ENTER(lsflag);
479 		LOCKSTAT_START_TIMER(lsflag, spintime);
480 		count = SPINLOCK_BACKOFF_MIN;
481 
482 		/*
483 		 * Spin testing the lock word and do exponential backoff
484 		 * to reduce cache line ping-ponging between CPUs.
485 		 */
486 		do {
487 			while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
488 				SPINLOCK_SPIN_HOOK;
489 				SPINLOCK_BACKOFF(count);
490 #ifdef LOCKDEBUG
491 				if (SPINLOCK_SPINOUT(spins))
492 					MUTEX_ABORT(mtx, "spinout");
493 #endif	/* LOCKDEBUG */
494 			}
495 		} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
496 
497 		if (count != SPINLOCK_BACKOFF_MIN) {
498 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
499 			LOCKSTAT_EVENT(lsflag, mtx,
500 			    LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
501 		}
502 		LOCKSTAT_EXIT(lsflag);
503 #endif	/* !MULTIPROCESSOR */
504 #endif	/* FULL */
505 		MUTEX_LOCKED(mtx);
506 		return;
507 	}
508 
509 	curthread = (uintptr_t)curlwp;
510 
511 	MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner));
512 	MUTEX_ASSERT(mtx, curthread != 0);
513 	MUTEX_ASSERT(mtx, !cpu_intr_p());
514 	MUTEX_WANTLOCK(mtx);
515 
516 	if (__predict_true(panicstr == NULL)) {
517 		KDASSERT(pserialize_not_in_read_section());
518 		LOCKDEBUG_BARRIER(&kernel_lock, 1);
519 	}
520 
521 	LOCKSTAT_ENTER(lsflag);
522 
523 	/*
524 	 * Adaptive mutex; spin trying to acquire the mutex.  If we
525 	 * determine that the owner is not running on a processor,
526 	 * then we stop spinning, and sleep instead.
527 	 */
528 	for (;;) {
529 		if (!MUTEX_OWNED(owner)) {
530 			/*
531 			 * Mutex owner clear could mean two things:
532 			 *
533 			 *	* The mutex has been released.
534 			 *	* The owner field hasn't been set yet.
535 			 *
536 			 * Try to acquire it again.  If that fails,
537 			 * we'll just loop again.
538 			 */
539 			if (MUTEX_ACQUIRE(mtx, curthread))
540 				break;
541 			owner = mtx->mtx_owner;
542 			continue;
543 		}
544 		if (__predict_false(MUTEX_OWNER(owner) == curthread)) {
545 			MUTEX_ABORT(mtx, "locking against myself");
546 		}
547 #ifdef MULTIPROCESSOR
548 		/*
549 		 * Check to see if the owner is running on a processor.
550 		 * If so, then we should just spin, as the owner will
551 		 * likely release the lock very soon.
552 		 */
553 		if (mutex_oncpu(owner)) {
554 			LOCKSTAT_START_TIMER(lsflag, spintime);
555 			count = SPINLOCK_BACKOFF_MIN;
556 			do {
557 				KPREEMPT_ENABLE(curlwp);
558 				SPINLOCK_BACKOFF(count);
559 				KPREEMPT_DISABLE(curlwp);
560 				owner = mtx->mtx_owner;
561 			} while (mutex_oncpu(owner));
562 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
563 			LOCKSTAT_COUNT(spincnt, 1);
564 			if (!MUTEX_OWNED(owner))
565 				continue;
566 		}
567 #endif
568 
569 		ts = turnstile_lookup(mtx);
570 
571 		/*
572 		 * Once we have the turnstile chain interlock, mark the
573 		 * mutex as having waiters.  If that fails, spin again:
574 		 * chances are that the mutex has been released.
575 		 */
576 		if (!MUTEX_SET_WAITERS(mtx, owner)) {
577 			turnstile_exit(mtx);
578 			owner = mtx->mtx_owner;
579 			continue;
580 		}
581 
582 #ifdef MULTIPROCESSOR
583 		/*
584 		 * mutex_exit() is permitted to release the mutex without
585 		 * any interlocking instructions, and the following can
586 		 * occur as a result:
587 		 *
588 		 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
589 		 * ---------------------------- ----------------------------
590 		 *		..		    acquire cache line
591 		 *		..                   test for waiters
592 		 *	acquire cache line    <-      lose cache line
593 		 *	 lock cache line	           ..
594 		 *     verify mutex is held                ..
595 		 *	    set waiters  	           ..
596 		 *	 unlock cache line		   ..
597 		 *	  lose cache line     ->    acquire cache line
598 		 *		..	          clear lock word, waiters
599 		 *	  return success
600 		 *
601 		 * There is another race that can occur: a third CPU could
602 		 * acquire the mutex as soon as it is released.  Since
603 		 * adaptive mutexes are primarily spin mutexes, this is not
604 		 * something that we need to worry about too much.  What we
605 		 * do need to ensure is that the waiters bit gets set.
606 		 *
607 		 * To allow the unlocked release, we need to make some
608 		 * assumptions here:
609 		 *
610 		 * o Release is the only non-atomic/unlocked operation
611 		 *   that can be performed on the mutex.  (It must still
612 		 *   be atomic on the local CPU, e.g. in case interrupted
613 		 *   or preempted).
614 		 *
615 		 * o At any given time, MUTEX_SET_WAITERS() can only ever
616 		 *   be in progress on one CPU in the system - guaranteed
617 		 *   by the turnstile chain lock.
618 		 *
619 		 * o No other operations other than MUTEX_SET_WAITERS()
620 		 *   and release can modify a mutex with a non-zero
621 		 *   owner field.
622 		 *
623 		 * o The result of a successful MUTEX_SET_WAITERS() call
624 		 *   is an unbuffered write that is immediately visible
625 		 *   to all other processors in the system.
626 		 *
627 		 * o If the holding LWP switches away, it posts a store
628 		 *   fence before changing curlwp, ensuring that any
629 		 *   overwrite of the mutex waiters flag by mutex_exit()
630 		 *   completes before the modification of curlwp becomes
631 		 *   visible to this CPU.
632 		 *
633 		 * o cpu_switchto() posts a store fence after setting curlwp
634 		 *   and before resuming execution of an LWP.
635 		 *
636 		 * o _kernel_lock() posts a store fence before setting
637 		 *   curcpu()->ci_biglock_wanted, and after clearing it.
638 		 *   This ensures that any overwrite of the mutex waiters
639 		 *   flag by mutex_exit() completes before the modification
640 		 *   of ci_biglock_wanted becomes visible.
641 		 *
642 		 * We now post a read memory barrier (after setting the
643 		 * waiters field) and check the lock holder's status again.
644 		 * Some of the possible outcomes (not an exhaustive list):
645 		 *
646 		 * 1. The on-CPU check returns true: the holding LWP is
647 		 *    running again.  The lock may be released soon and
648 		 *    we should spin.  Importantly, we can't trust the
649 		 *    value of the waiters flag.
650 		 *
651 		 * 2. The on-CPU check returns false: the holding LWP is
652 		 *    not running.  We now have the opportunity to check
653 		 *    if mutex_exit() has blatted the modifications made
654 		 *    by MUTEX_SET_WAITERS().
655 		 *
656 		 * 3. The on-CPU check returns false: the holding LWP may
657 		 *    or may not be running.  It has context switched at
658 		 *    some point during our check.  Again, we have the
659 		 *    chance to see if the waiters bit is still set or
660 		 *    has been overwritten.
661 		 *
662 		 * 4. The on-CPU check returns false: the holding LWP is
663 		 *    running on a CPU, but wants the big lock.  It's OK
664 		 *    to check the waiters field in this case.
665 		 *
666 		 * 5. The has-waiters check fails: the mutex has been
667 		 *    released, the waiters flag cleared and another LWP
668 		 *    now owns the mutex.
669 		 *
670 		 * 6. The has-waiters check fails: the mutex has been
671 		 *    released.
672 		 *
673 		 * If the waiters bit is not set it's unsafe to go asleep,
674 		 * as we might never be awoken.
675 		 */
676 		membar_consumer();
677 		if (mutex_oncpu(owner)) {
678 			turnstile_exit(mtx);
679 			owner = mtx->mtx_owner;
680 			continue;
681 		}
682 		membar_consumer();
683 		if (!MUTEX_HAS_WAITERS(mtx)) {
684 			turnstile_exit(mtx);
685 			owner = mtx->mtx_owner;
686 			continue;
687 		}
688 #endif	/* MULTIPROCESSOR */
689 
690 		LOCKSTAT_START_TIMER(lsflag, slptime);
691 
692 		turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj);
693 
694 		LOCKSTAT_STOP_TIMER(lsflag, slptime);
695 		LOCKSTAT_COUNT(slpcnt, 1);
696 
697 		owner = mtx->mtx_owner;
698 	}
699 	KPREEMPT_ENABLE(curlwp);
700 
701 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SLEEP1,
702 	    slpcnt, slptime);
703 	LOCKSTAT_EVENT(lsflag, mtx, LB_ADAPTIVE_MUTEX | LB_SPIN,
704 	    spincnt, spintime);
705 	LOCKSTAT_EXIT(lsflag);
706 
707 	MUTEX_DASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
708 	MUTEX_LOCKED(mtx);
709 }
710 
711 /*
712  * mutex_vector_exit:
713  *
714  *	Support routine for mutex_exit() that handles all cases.
715  */
716 void
717 mutex_vector_exit(kmutex_t *mtx)
718 {
719 	turnstile_t *ts;
720 	uintptr_t curthread;
721 
722 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
723 #ifdef FULL
724 		if (__predict_false(!MUTEX_SPINBIT_LOCKED_P(mtx))) {
725 			MUTEX_ABORT(mtx, "exiting unheld spin mutex");
726 		}
727 		MUTEX_UNLOCKED(mtx);
728 		MUTEX_SPINBIT_LOCK_UNLOCK(mtx);
729 #endif
730 		MUTEX_SPIN_SPLRESTORE(mtx);
731 		return;
732 	}
733 
734 #ifndef __HAVE_MUTEX_STUBS
735 	/*
736 	 * On some architectures without mutex stubs, we can enter here to
737 	 * release mutexes before interrupts and whatnot are up and running.
738 	 * We need this hack to keep them sweet.
739 	 */
740 	if (__predict_false(cold)) {
741 		MUTEX_UNLOCKED(mtx);
742 		MUTEX_RELEASE(mtx);
743 		return;
744 	}
745 #endif
746 
747 	curthread = (uintptr_t)curlwp;
748 	MUTEX_DASSERT(mtx, curthread != 0);
749 	MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
750 	MUTEX_UNLOCKED(mtx);
751 #if !defined(LOCKDEBUG)
752 	__USE(curthread);
753 #endif
754 
755 #ifdef LOCKDEBUG
756 	/*
757 	 * Avoid having to take the turnstile chain lock every time
758 	 * around.  Raise the priority level to splhigh() in order
759 	 * to disable preemption and so make the following atomic.
760 	 */
761 	{
762 		int s = splhigh();
763 		if (!MUTEX_HAS_WAITERS(mtx)) {
764 			MUTEX_RELEASE(mtx);
765 			splx(s);
766 			return;
767 		}
768 		splx(s);
769 	}
770 #endif
771 
772 	/*
773 	 * Get this lock's turnstile.  This gets the interlock on
774 	 * the sleep queue.  Once we have that, we can clear the
775 	 * lock.  If there was no turnstile for the lock, there
776 	 * were no waiters remaining.
777 	 */
778 	ts = turnstile_lookup(mtx);
779 
780 	if (ts == NULL) {
781 		MUTEX_RELEASE(mtx);
782 		turnstile_exit(mtx);
783 	} else {
784 		MUTEX_RELEASE(mtx);
785 		turnstile_wakeup(ts, TS_WRITER_Q,
786 		    TS_WAITERS(ts, TS_WRITER_Q), NULL);
787 	}
788 }
789 
790 #ifndef __HAVE_SIMPLE_MUTEXES
791 /*
792  * mutex_wakeup:
793  *
794  *	Support routine for mutex_exit() that wakes up all waiters.
795  *	We assume that the mutex has been released, but it need not
796  *	be.
797  */
798 void
799 mutex_wakeup(kmutex_t *mtx)
800 {
801 	turnstile_t *ts;
802 
803 	ts = turnstile_lookup(mtx);
804 	if (ts == NULL) {
805 		turnstile_exit(mtx);
806 		return;
807 	}
808 	MUTEX_CLEAR_WAITERS(mtx);
809 	turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
810 }
811 #endif	/* !__HAVE_SIMPLE_MUTEXES */
812 
813 /*
814  * mutex_owned:
815  *
816  *	Return true if the current LWP (adaptive) or CPU (spin)
817  *	holds the mutex.
818  */
819 int
820 mutex_owned(const kmutex_t *mtx)
821 {
822 
823 	if (mtx == NULL)
824 		return 0;
825 	if (MUTEX_ADAPTIVE_P(mtx->mtx_owner))
826 		return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
827 #ifdef FULL
828 	return MUTEX_SPINBIT_LOCKED_P(mtx);
829 #else
830 	return 1;
831 #endif
832 }
833 
834 /*
835  * mutex_owner:
836  *
837  *	Return the current owner of an adaptive mutex.  Used for
838  *	priority inheritance.
839  */
840 static lwp_t *
841 mutex_owner(wchan_t wchan)
842 {
843 	volatile const kmutex_t *mtx = wchan;
844 
845 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
846 	return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
847 }
848 
849 /*
850  * mutex_owner_running:
851  *
852  *	Return true if an adaptive mutex is unheld, or held and the owner is
853  *	running on a CPU.  For the pagedaemon only - do not document or use
854  *	in other code.
855  */
856 bool
857 mutex_owner_running(const kmutex_t *mtx)
858 {
859 #ifdef MULTIPROCESSOR
860 	uintptr_t owner;
861 	bool rv;
862 
863 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
864 	kpreempt_disable();
865 	owner = mtx->mtx_owner;
866 	rv = !MUTEX_OWNED(owner) || mutex_oncpu(MUTEX_OWNER(owner));
867 	kpreempt_enable();
868 	return rv;
869 #else
870 	return mutex_owner(mtx) == curlwp;
871 #endif
872 }
873 
874 /*
875  * mutex_ownable:
876  *
877  *	When compiled with DEBUG and LOCKDEBUG defined, ensure that
878  *	the mutex is available.  We cannot use !mutex_owned() since
879  *	that won't work correctly for spin mutexes.
880  */
881 int
882 mutex_ownable(const kmutex_t *mtx)
883 {
884 
885 #ifdef LOCKDEBUG
886 	MUTEX_TESTLOCK(mtx);
887 #endif
888 	return 1;
889 }
890 
891 /*
892  * mutex_tryenter:
893  *
894  *	Try to acquire the mutex; return non-zero if we did.
895  */
896 int
897 mutex_tryenter(kmutex_t *mtx)
898 {
899 	uintptr_t curthread;
900 
901 	/*
902 	 * Handle spin mutexes.
903 	 */
904 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
905 		MUTEX_SPIN_SPLRAISE(mtx);
906 #ifdef FULL
907 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
908 			MUTEX_WANTLOCK(mtx);
909 			MUTEX_LOCKED(mtx);
910 			return 1;
911 		}
912 		MUTEX_SPIN_SPLRESTORE(mtx);
913 #else
914 		MUTEX_WANTLOCK(mtx);
915 		MUTEX_LOCKED(mtx);
916 		return 1;
917 #endif
918 	} else {
919 		curthread = (uintptr_t)curlwp;
920 		MUTEX_ASSERT(mtx, curthread != 0);
921 		if (MUTEX_ACQUIRE(mtx, curthread)) {
922 			MUTEX_WANTLOCK(mtx);
923 			MUTEX_LOCKED(mtx);
924 			MUTEX_DASSERT(mtx,
925 			    MUTEX_OWNER(mtx->mtx_owner) == curthread);
926 			return 1;
927 		}
928 	}
929 
930 	return 0;
931 }
932 
933 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
934 /*
935  * mutex_spin_retry:
936  *
937  *	Support routine for mutex_spin_enter().  Assumes that the caller
938  *	has already raised the SPL, and adjusted counters.
939  */
940 void
941 mutex_spin_retry(kmutex_t *mtx)
942 {
943 #ifdef MULTIPROCESSOR
944 	u_int count;
945 	LOCKSTAT_TIMER(spintime);
946 	LOCKSTAT_FLAG(lsflag);
947 #ifdef LOCKDEBUG
948 	u_int spins = 0;
949 #endif	/* LOCKDEBUG */
950 
951 	MUTEX_WANTLOCK(mtx);
952 
953 	LOCKSTAT_ENTER(lsflag);
954 	LOCKSTAT_START_TIMER(lsflag, spintime);
955 	count = SPINLOCK_BACKOFF_MIN;
956 
957 	/*
958 	 * Spin testing the lock word and do exponential backoff
959 	 * to reduce cache line ping-ponging between CPUs.
960 	 */
961 	do {
962 		while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
963 			SPINLOCK_BACKOFF(count);
964 #ifdef LOCKDEBUG
965 			if (SPINLOCK_SPINOUT(spins))
966 				MUTEX_ABORT(mtx, "spinout");
967 #endif	/* LOCKDEBUG */
968 		}
969 	} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
970 
971 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
972 	LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
973 	LOCKSTAT_EXIT(lsflag);
974 
975 	MUTEX_LOCKED(mtx);
976 #else	/* MULTIPROCESSOR */
977 	MUTEX_ABORT(mtx, "locking against myself");
978 #endif	/* MULTIPROCESSOR */
979 }
980 #endif	/* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
981