xref: /netbsd-src/sys/kern/kern_mutex.c (revision 53b02e147d4ed531c0d2a5ca9b3e8026ba3e99b5)
1 /*	$NetBSD: kern_mutex.c,v 1.98 2021/08/25 04:13:42 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 1.98 2021/08/25 04:13:42 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 #ifndef MUTEX_CAS
218 #define	MUTEX_CAS(p, o, n)		\
219 	(atomic_cas_ulong((volatile unsigned long *)(p), (o), (n)) == (o))
220 #endif /* MUTEX_CAS */
221 
222 #define	MUTEX_DEBUG_P(mtx)	(((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0)
223 #if defined(LOCKDEBUG)
224 #define	MUTEX_OWNED(owner)		(((owner) & ~MUTEX_BIT_NODEBUG) != 0)
225 #define	MUTEX_INHERITDEBUG(n, o)	(n) |= (o) & MUTEX_BIT_NODEBUG
226 #else /* defined(LOCKDEBUG) */
227 #define	MUTEX_OWNED(owner)		((owner) != 0)
228 #define	MUTEX_INHERITDEBUG(n, o)	/* nothing */
229 #endif /* defined(LOCKDEBUG) */
230 
231 static inline int
232 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
233 {
234 	int rv;
235 	uintptr_t oldown = 0;
236 	uintptr_t newown = curthread;
237 
238 	MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner);
239 	MUTEX_INHERITDEBUG(newown, oldown);
240 	rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown);
241 	MUTEX_MEMBAR_ENTER();
242 	return rv;
243 }
244 
245 static inline int
246 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
247 {
248 	int rv;
249 	rv = MUTEX_CAS(&mtx->mtx_owner, owner, owner | MUTEX_BIT_WAITERS);
250 	MUTEX_MEMBAR_ENTER();
251 	return rv;
252 }
253 
254 static inline void
255 MUTEX_RELEASE(kmutex_t *mtx)
256 {
257 	uintptr_t newown;
258 
259 	MUTEX_MEMBAR_EXIT();
260 	newown = 0;
261 	MUTEX_INHERITDEBUG(newown, mtx->mtx_owner);
262 	mtx->mtx_owner = newown;
263 }
264 #endif	/* __HAVE_SIMPLE_MUTEXES */
265 
266 /*
267  * Patch in stubs via strong alias where they are not available.
268  */
269 
270 #if defined(LOCKDEBUG)
271 #undef	__HAVE_MUTEX_STUBS
272 #undef	__HAVE_SPIN_MUTEX_STUBS
273 #endif
274 
275 #ifndef __HAVE_MUTEX_STUBS
276 __strong_alias(mutex_enter,mutex_vector_enter);
277 __strong_alias(mutex_exit,mutex_vector_exit);
278 #endif
279 
280 #ifndef __HAVE_SPIN_MUTEX_STUBS
281 __strong_alias(mutex_spin_enter,mutex_vector_enter);
282 __strong_alias(mutex_spin_exit,mutex_vector_exit);
283 #endif
284 
285 static void	mutex_abort(const char *, size_t, const kmutex_t *,
286     const char *);
287 static void	mutex_dump(const volatile void *, lockop_printer_t);
288 
289 lockops_t mutex_spin_lockops = {
290 	.lo_name = "Mutex",
291 	.lo_type = LOCKOPS_SPIN,
292 	.lo_dump = mutex_dump,
293 };
294 
295 lockops_t mutex_adaptive_lockops = {
296 	.lo_name = "Mutex",
297 	.lo_type = LOCKOPS_SLEEP,
298 	.lo_dump = mutex_dump,
299 };
300 
301 syncobj_t mutex_syncobj = {
302 	.sobj_flag	= SOBJ_SLEEPQ_SORTED,
303 	.sobj_unsleep	= turnstile_unsleep,
304 	.sobj_changepri	= turnstile_changepri,
305 	.sobj_lendpri	= sleepq_lendpri,
306 	.sobj_owner	= (void *)mutex_owner,
307 };
308 
309 /*
310  * mutex_dump:
311  *
312  *	Dump the contents of a mutex structure.
313  */
314 static void
315 mutex_dump(const volatile void *cookie, lockop_printer_t pr)
316 {
317 	const volatile kmutex_t *mtx = cookie;
318 	uintptr_t owner = mtx->mtx_owner;
319 
320 	pr("owner field  : %#018lx wait/spin: %16d/%d\n",
321 	    (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx),
322 	    MUTEX_SPIN_P(owner));
323 }
324 
325 /*
326  * mutex_abort:
327  *
328  *	Dump information about an error and panic the system.  This
329  *	generates a lot of machine code in the DIAGNOSTIC case, so
330  *	we ask the compiler to not inline it.
331  */
332 static void __noinline
333 mutex_abort(const char *func, size_t line, const kmutex_t *mtx, const char *msg)
334 {
335 
336 	LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ?
337 	    &mutex_spin_lockops : &mutex_adaptive_lockops), msg);
338 }
339 
340 /*
341  * mutex_init:
342  *
343  *	Initialize a mutex for use.  Note that adaptive mutexes are in
344  *	essence spin mutexes that can sleep to avoid deadlock and wasting
345  *	CPU time.  We can't easily provide a type of mutex that always
346  *	sleeps - see comments in mutex_vector_enter() about releasing
347  *	mutexes unlocked.
348  */
349 void _mutex_init(kmutex_t *, kmutex_type_t, int, uintptr_t);
350 void
351 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl,
352     uintptr_t return_address)
353 {
354 	lockops_t *lockops __unused;
355 	bool dodebug;
356 
357 	memset(mtx, 0, sizeof(*mtx));
358 
359 	if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK ||
360 	    ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET ||
361 	    ipl == IPL_SOFTSERIAL) {
362 		lockops = (type == MUTEX_NODEBUG ?
363 		    NULL : &mutex_adaptive_lockops);
364 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
365 		MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug);
366 	} else {
367 		lockops = (type == MUTEX_NODEBUG ?
368 		    NULL : &mutex_spin_lockops);
369 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
370 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
371 	}
372 }
373 
374 void
375 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
376 {
377 
378 	_mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0));
379 }
380 
381 /*
382  * mutex_destroy:
383  *
384  *	Tear down a mutex.
385  */
386 void
387 mutex_destroy(kmutex_t *mtx)
388 {
389 	uintptr_t owner = mtx->mtx_owner;
390 
391 	if (MUTEX_ADAPTIVE_P(owner)) {
392 		MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner));
393 		MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx));
394 	} else {
395 		MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx));
396 	}
397 
398 	LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx);
399 	MUTEX_DESTROY(mtx);
400 }
401 
402 #ifdef MULTIPROCESSOR
403 /*
404  * mutex_oncpu:
405  *
406  *	Return true if an adaptive mutex owner is running on a CPU in the
407  *	system.  If the target is waiting on the kernel big lock, then we
408  *	must release it.  This is necessary to avoid deadlock.
409  */
410 static bool
411 mutex_oncpu(uintptr_t owner)
412 {
413 	struct cpu_info *ci;
414 	lwp_t *l;
415 
416 	KASSERT(kpreempt_disabled());
417 
418 	if (!MUTEX_OWNED(owner)) {
419 		return false;
420 	}
421 
422 	/*
423 	 * See lwp_dtor() why dereference of the LWP pointer is safe.
424 	 * We must have kernel preemption disabled for that.
425 	 */
426 	l = (lwp_t *)MUTEX_OWNER(owner);
427 	ci = l->l_cpu;
428 
429 	if (ci && ci->ci_curlwp == l) {
430 		/* Target is running; do we need to block? */
431 		return (ci->ci_biglock_wanted != l);
432 	}
433 
434 	/* Not running.  It may be safe to block now. */
435 	return false;
436 }
437 #endif	/* MULTIPROCESSOR */
438 
439 /*
440  * mutex_vector_enter:
441  *
442  *	Support routine for mutex_enter() that must handle all cases.  In
443  *	the LOCKDEBUG case, mutex_enter() is always aliased here, even if
444  *	fast-path stubs are available.  If a mutex_spin_enter() stub is
445  *	not available, then it is also aliased directly here.
446  */
447 void
448 mutex_vector_enter(kmutex_t *mtx)
449 {
450 	uintptr_t owner, curthread;
451 	turnstile_t *ts;
452 #ifdef MULTIPROCESSOR
453 	u_int count;
454 #endif
455 	LOCKSTAT_COUNTER(spincnt);
456 	LOCKSTAT_COUNTER(slpcnt);
457 	LOCKSTAT_TIMER(spintime);
458 	LOCKSTAT_TIMER(slptime);
459 	LOCKSTAT_FLAG(lsflag);
460 
461 	/*
462 	 * Handle spin mutexes.
463 	 */
464 	KPREEMPT_DISABLE(curlwp);
465 	owner = mtx->mtx_owner;
466 	if (MUTEX_SPIN_P(owner)) {
467 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
468 		u_int spins = 0;
469 #endif
470 		KPREEMPT_ENABLE(curlwp);
471 		MUTEX_SPIN_SPLRAISE(mtx);
472 		MUTEX_WANTLOCK(mtx);
473 #ifdef FULL
474 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
475 			MUTEX_LOCKED(mtx);
476 			return;
477 		}
478 #if !defined(MULTIPROCESSOR)
479 		MUTEX_ABORT(mtx, "locking against myself");
480 #else /* !MULTIPROCESSOR */
481 
482 		LOCKSTAT_ENTER(lsflag);
483 		LOCKSTAT_START_TIMER(lsflag, spintime);
484 		count = SPINLOCK_BACKOFF_MIN;
485 
486 		/*
487 		 * Spin testing the lock word and do exponential backoff
488 		 * to reduce cache line ping-ponging between CPUs.
489 		 */
490 		do {
491 			while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
492 				SPINLOCK_SPIN_HOOK;
493 				SPINLOCK_BACKOFF(count);
494 #ifdef LOCKDEBUG
495 				if (SPINLOCK_SPINOUT(spins))
496 					MUTEX_ABORT(mtx, "spinout");
497 #endif	/* LOCKDEBUG */
498 			}
499 		} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
500 
501 		if (count != SPINLOCK_BACKOFF_MIN) {
502 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
503 			LOCKSTAT_EVENT(lsflag, mtx,
504 			    LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
505 		}
506 		LOCKSTAT_EXIT(lsflag);
507 #endif	/* !MULTIPROCESSOR */
508 #endif	/* FULL */
509 		MUTEX_LOCKED(mtx);
510 		return;
511 	}
512 
513 	curthread = (uintptr_t)curlwp;
514 
515 	MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner));
516 	MUTEX_ASSERT(mtx, curthread != 0);
517 	MUTEX_ASSERT(mtx, !cpu_intr_p());
518 	MUTEX_WANTLOCK(mtx);
519 
520 	if (panicstr == NULL) {
521 		KDASSERT(pserialize_not_in_read_section());
522 		LOCKDEBUG_BARRIER(&kernel_lock, 1);
523 	}
524 
525 	LOCKSTAT_ENTER(lsflag);
526 
527 	/*
528 	 * Adaptive mutex; spin trying to acquire the mutex.  If we
529 	 * determine that the owner is not running on a processor,
530 	 * then we stop spinning, and sleep instead.
531 	 */
532 	for (;;) {
533 		if (!MUTEX_OWNED(owner)) {
534 			/*
535 			 * Mutex owner clear could mean two things:
536 			 *
537 			 *	* The mutex has been released.
538 			 *	* The owner field hasn't been set yet.
539 			 *
540 			 * Try to acquire it again.  If that fails,
541 			 * we'll just loop again.
542 			 */
543 			if (MUTEX_ACQUIRE(mtx, curthread))
544 				break;
545 			owner = mtx->mtx_owner;
546 			continue;
547 		}
548 		if (__predict_false(MUTEX_OWNER(owner) == curthread)) {
549 			MUTEX_ABORT(mtx, "locking against myself");
550 		}
551 #ifdef MULTIPROCESSOR
552 		/*
553 		 * Check to see if the owner is running on a processor.
554 		 * If so, then we should just spin, as the owner will
555 		 * likely release the lock very soon.
556 		 */
557 		if (mutex_oncpu(owner)) {
558 			LOCKSTAT_START_TIMER(lsflag, spintime);
559 			count = SPINLOCK_BACKOFF_MIN;
560 			do {
561 				KPREEMPT_ENABLE(curlwp);
562 				SPINLOCK_BACKOFF(count);
563 				KPREEMPT_DISABLE(curlwp);
564 				owner = mtx->mtx_owner;
565 			} while (mutex_oncpu(owner));
566 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
567 			LOCKSTAT_COUNT(spincnt, 1);
568 			if (!MUTEX_OWNED(owner))
569 				continue;
570 		}
571 #endif
572 
573 		ts = turnstile_lookup(mtx);
574 
575 		/*
576 		 * Once we have the turnstile chain interlock, mark the
577 		 * mutex as having waiters.  If that fails, spin again:
578 		 * chances are that the mutex has been released.
579 		 */
580 		if (!MUTEX_SET_WAITERS(mtx, owner)) {
581 			turnstile_exit(mtx);
582 			owner = mtx->mtx_owner;
583 			continue;
584 		}
585 
586 #ifdef MULTIPROCESSOR
587 		/*
588 		 * mutex_exit() is permitted to release the mutex without
589 		 * any interlocking instructions, and the following can
590 		 * occur as a result:
591 		 *
592 		 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
593 		 * ---------------------------- ----------------------------
594 		 *		..		    acquire cache line
595 		 *		..                   test for waiters
596 		 *	acquire cache line    <-      lose cache line
597 		 *	 lock cache line	           ..
598 		 *     verify mutex is held                ..
599 		 *	    set waiters  	           ..
600 		 *	 unlock cache line		   ..
601 		 *	  lose cache line     ->    acquire cache line
602 		 *		..	          clear lock word, waiters
603 		 *	  return success
604 		 *
605 		 * There is another race that can occur: a third CPU could
606 		 * acquire the mutex as soon as it is released.  Since
607 		 * adaptive mutexes are primarily spin mutexes, this is not
608 		 * something that we need to worry about too much.  What we
609 		 * do need to ensure is that the waiters bit gets set.
610 		 *
611 		 * To allow the unlocked release, we need to make some
612 		 * assumptions here:
613 		 *
614 		 * o Release is the only non-atomic/unlocked operation
615 		 *   that can be performed on the mutex.  (It must still
616 		 *   be atomic on the local CPU, e.g. in case interrupted
617 		 *   or preempted).
618 		 *
619 		 * o At any given time, MUTEX_SET_WAITERS() can only ever
620 		 *   be in progress on one CPU in the system - guaranteed
621 		 *   by the turnstile chain lock.
622 		 *
623 		 * o No other operations other than MUTEX_SET_WAITERS()
624 		 *   and release can modify a mutex with a non-zero
625 		 *   owner field.
626 		 *
627 		 * o The result of a successful MUTEX_SET_WAITERS() call
628 		 *   is an unbuffered write that is immediately visible
629 		 *   to all other processors in the system.
630 		 *
631 		 * o If the holding LWP switches away, it posts a store
632 		 *   fence before changing curlwp, ensuring that any
633 		 *   overwrite of the mutex waiters flag by mutex_exit()
634 		 *   completes before the modification of curlwp becomes
635 		 *   visible to this CPU.
636 		 *
637 		 * o cpu_switchto() posts a store fence after setting curlwp
638 		 *   and before resuming execution of an LWP.
639 		 *
640 		 * o _kernel_lock() posts a store fence before setting
641 		 *   curcpu()->ci_biglock_wanted, and after clearing it.
642 		 *   This ensures that any overwrite of the mutex waiters
643 		 *   flag by mutex_exit() completes before the modification
644 		 *   of ci_biglock_wanted becomes visible.
645 		 *
646 		 * We now post a read memory barrier (after setting the
647 		 * waiters field) and check the lock holder's status again.
648 		 * Some of the possible outcomes (not an exhaustive list):
649 		 *
650 		 * 1. The on-CPU check returns true: the holding LWP is
651 		 *    running again.  The lock may be released soon and
652 		 *    we should spin.  Importantly, we can't trust the
653 		 *    value of the waiters flag.
654 		 *
655 		 * 2. The on-CPU check returns false: the holding LWP is
656 		 *    not running.  We now have the opportunity to check
657 		 *    if mutex_exit() has blatted the modifications made
658 		 *    by MUTEX_SET_WAITERS().
659 		 *
660 		 * 3. The on-CPU check returns false: the holding LWP may
661 		 *    or may not be running.  It has context switched at
662 		 *    some point during our check.  Again, we have the
663 		 *    chance to see if the waiters bit is still set or
664 		 *    has been overwritten.
665 		 *
666 		 * 4. The on-CPU check returns false: the holding LWP is
667 		 *    running on a CPU, but wants the big lock.  It's OK
668 		 *    to check the waiters field in this case.
669 		 *
670 		 * 5. The has-waiters check fails: the mutex has been
671 		 *    released, the waiters flag cleared and another LWP
672 		 *    now owns the mutex.
673 		 *
674 		 * 6. The has-waiters check fails: the mutex has been
675 		 *    released.
676 		 *
677 		 * If the waiters bit is not set it's unsafe to go asleep,
678 		 * as we might never be awoken.
679 		 */
680 		membar_consumer();
681 		if (mutex_oncpu(owner)) {
682 			turnstile_exit(mtx);
683 			owner = mtx->mtx_owner;
684 			continue;
685 		}
686 		membar_consumer();
687 		if (!MUTEX_HAS_WAITERS(mtx)) {
688 			turnstile_exit(mtx);
689 			owner = mtx->mtx_owner;
690 			continue;
691 		}
692 #endif	/* MULTIPROCESSOR */
693 
694 		LOCKSTAT_START_TIMER(lsflag, slptime);
695 
696 		turnstile_block(ts, TS_WRITER_Q, mtx, &mutex_syncobj);
697 
698 		LOCKSTAT_STOP_TIMER(lsflag, slptime);
699 		LOCKSTAT_COUNT(slpcnt, 1);
700 
701 		owner = mtx->mtx_owner;
702 	}
703 	KPREEMPT_ENABLE(curlwp);
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->mtx_owner)) {
727 #ifdef FULL
728 		if (__predict_false(!MUTEX_SPINBIT_LOCKED_P(mtx))) {
729 			MUTEX_ABORT(mtx, "exiting unheld spin mutex");
730 		}
731 		MUTEX_UNLOCKED(mtx);
732 		MUTEX_SPINBIT_LOCK_UNLOCK(mtx);
733 #endif
734 		MUTEX_SPIN_SPLRESTORE(mtx);
735 		return;
736 	}
737 
738 #ifndef __HAVE_MUTEX_STUBS
739 	/*
740 	 * On some architectures without mutex stubs, we can enter here to
741 	 * release mutexes before interrupts and whatnot are up and running.
742 	 * We need this hack to keep them sweet.
743 	 */
744 	if (__predict_false(cold)) {
745 		MUTEX_UNLOCKED(mtx);
746 		MUTEX_RELEASE(mtx);
747 		return;
748 	}
749 #endif
750 
751 	curthread = (uintptr_t)curlwp;
752 	MUTEX_DASSERT(mtx, curthread != 0);
753 	MUTEX_ASSERT(mtx, MUTEX_OWNER(mtx->mtx_owner) == curthread);
754 	MUTEX_UNLOCKED(mtx);
755 #if !defined(LOCKDEBUG)
756 	__USE(curthread);
757 #endif
758 
759 #ifdef LOCKDEBUG
760 	/*
761 	 * Avoid having to take the turnstile chain lock every time
762 	 * around.  Raise the priority level to splhigh() in order
763 	 * to disable preemption and so make the following atomic.
764 	 */
765 	{
766 		int s = splhigh();
767 		if (!MUTEX_HAS_WAITERS(mtx)) {
768 			MUTEX_RELEASE(mtx);
769 			splx(s);
770 			return;
771 		}
772 		splx(s);
773 	}
774 #endif
775 
776 	/*
777 	 * Get this lock's turnstile.  This gets the interlock on
778 	 * the sleep queue.  Once we have that, we can clear the
779 	 * lock.  If there was no turnstile for the lock, there
780 	 * were no waiters remaining.
781 	 */
782 	ts = turnstile_lookup(mtx);
783 
784 	if (ts == NULL) {
785 		MUTEX_RELEASE(mtx);
786 		turnstile_exit(mtx);
787 	} else {
788 		MUTEX_RELEASE(mtx);
789 		turnstile_wakeup(ts, TS_WRITER_Q,
790 		    TS_WAITERS(ts, TS_WRITER_Q), NULL);
791 	}
792 }
793 
794 #ifndef __HAVE_SIMPLE_MUTEXES
795 /*
796  * mutex_wakeup:
797  *
798  *	Support routine for mutex_exit() that wakes up all waiters.
799  *	We assume that the mutex has been released, but it need not
800  *	be.
801  */
802 void
803 mutex_wakeup(kmutex_t *mtx)
804 {
805 	turnstile_t *ts;
806 
807 	ts = turnstile_lookup(mtx);
808 	if (ts == NULL) {
809 		turnstile_exit(mtx);
810 		return;
811 	}
812 	MUTEX_CLEAR_WAITERS(mtx);
813 	turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
814 }
815 #endif	/* !__HAVE_SIMPLE_MUTEXES */
816 
817 /*
818  * mutex_owned:
819  *
820  *	Return true if the current LWP (adaptive) or CPU (spin)
821  *	holds the mutex.
822  */
823 int
824 mutex_owned(const kmutex_t *mtx)
825 {
826 
827 	if (mtx == NULL)
828 		return 0;
829 	if (MUTEX_ADAPTIVE_P(mtx->mtx_owner))
830 		return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
831 #ifdef FULL
832 	return MUTEX_SPINBIT_LOCKED_P(mtx);
833 #else
834 	return 1;
835 #endif
836 }
837 
838 /*
839  * mutex_owner:
840  *
841  *	Return the current owner of an adaptive mutex.  Used for
842  *	priority inheritance.
843  */
844 lwp_t *
845 mutex_owner(const kmutex_t *mtx)
846 {
847 
848 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
849 	return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
850 }
851 
852 /*
853  * mutex_owner_running:
854  *
855  *	Return true if an adaptive mutex is unheld, or held and the owner is
856  *	running on a CPU.  For the pagedaemon only - do not document or use
857  *	in other code.
858  */
859 bool
860 mutex_owner_running(const kmutex_t *mtx)
861 {
862 #ifdef MULTIPROCESSOR
863 	uintptr_t owner;
864 	bool rv;
865 
866 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
867 	kpreempt_disable();
868 	owner = mtx->mtx_owner;
869 	rv = !MUTEX_OWNED(owner) || mutex_oncpu(MUTEX_OWNER(owner));
870 	kpreempt_enable();
871 	return rv;
872 #else
873 	return mutex_owner(mtx) == curlwp;
874 #endif
875 }
876 
877 /*
878  * mutex_ownable:
879  *
880  *	When compiled with DEBUG and LOCKDEBUG defined, ensure that
881  *	the mutex is available.  We cannot use !mutex_owned() since
882  *	that won't work correctly for spin mutexes.
883  */
884 int
885 mutex_ownable(const kmutex_t *mtx)
886 {
887 
888 #ifdef LOCKDEBUG
889 	MUTEX_TESTLOCK(mtx);
890 #endif
891 	return 1;
892 }
893 
894 /*
895  * mutex_tryenter:
896  *
897  *	Try to acquire the mutex; return non-zero if we did.
898  */
899 int
900 mutex_tryenter(kmutex_t *mtx)
901 {
902 	uintptr_t curthread;
903 
904 	/*
905 	 * Handle spin mutexes.
906 	 */
907 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
908 		MUTEX_SPIN_SPLRAISE(mtx);
909 #ifdef FULL
910 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
911 			MUTEX_WANTLOCK(mtx);
912 			MUTEX_LOCKED(mtx);
913 			return 1;
914 		}
915 		MUTEX_SPIN_SPLRESTORE(mtx);
916 #else
917 		MUTEX_WANTLOCK(mtx);
918 		MUTEX_LOCKED(mtx);
919 		return 1;
920 #endif
921 	} else {
922 		curthread = (uintptr_t)curlwp;
923 		MUTEX_ASSERT(mtx, curthread != 0);
924 		if (MUTEX_ACQUIRE(mtx, curthread)) {
925 			MUTEX_WANTLOCK(mtx);
926 			MUTEX_LOCKED(mtx);
927 			MUTEX_DASSERT(mtx,
928 			    MUTEX_OWNER(mtx->mtx_owner) == curthread);
929 			return 1;
930 		}
931 	}
932 
933 	return 0;
934 }
935 
936 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
937 /*
938  * mutex_spin_retry:
939  *
940  *	Support routine for mutex_spin_enter().  Assumes that the caller
941  *	has already raised the SPL, and adjusted counters.
942  */
943 void
944 mutex_spin_retry(kmutex_t *mtx)
945 {
946 #ifdef MULTIPROCESSOR
947 	u_int count;
948 	LOCKSTAT_TIMER(spintime);
949 	LOCKSTAT_FLAG(lsflag);
950 #ifdef LOCKDEBUG
951 	u_int spins = 0;
952 #endif	/* LOCKDEBUG */
953 
954 	MUTEX_WANTLOCK(mtx);
955 
956 	LOCKSTAT_ENTER(lsflag);
957 	LOCKSTAT_START_TIMER(lsflag, spintime);
958 	count = SPINLOCK_BACKOFF_MIN;
959 
960 	/*
961 	 * Spin testing the lock word and do exponential backoff
962 	 * to reduce cache line ping-ponging between CPUs.
963 	 */
964 	do {
965 		while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
966 			SPINLOCK_BACKOFF(count);
967 #ifdef LOCKDEBUG
968 			if (SPINLOCK_SPINOUT(spins))
969 				MUTEX_ABORT(mtx, "spinout");
970 #endif	/* LOCKDEBUG */
971 		}
972 	} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
973 
974 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
975 	LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
976 	LOCKSTAT_EXIT(lsflag);
977 
978 	MUTEX_LOCKED(mtx);
979 #else	/* MULTIPROCESSOR */
980 	MUTEX_ABORT(mtx, "locking against myself");
981 #endif	/* MULTIPROCESSOR */
982 }
983 #endif	/* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
984