xref: /netbsd-src/sys/kern/kern_mutex.c (revision 6779c0232f05d52f6057b4bc5f46e515eb8f8989)
1 /*	$NetBSD: kern_mutex.c,v 1.112 2023/10/15 10:28:23 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008, 2019, 2023
5  *     The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Jason R. Thorpe and Andrew Doran.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * Kernel mutex implementation, modeled after those found in Solaris,
35  * a description of which can be found in:
36  *
37  *	Solaris Internals: Core Kernel Architecture, Jim Mauro and
38  *	    Richard McDougall.
39  */
40 
41 #define	__MUTEX_PRIVATE
42 
43 #include <sys/cdefs.h>
44 __KERNEL_RCSID(0, "$NetBSD: kern_mutex.c,v 1.112 2023/10/15 10:28:23 riastradh Exp $");
45 
46 #include <sys/param.h>
47 
48 #include <sys/atomic.h>
49 #include <sys/cpu.h>
50 #include <sys/intr.h>
51 #include <sys/kernel.h>
52 #include <sys/lock.h>
53 #include <sys/lockdebug.h>
54 #include <sys/mutex.h>
55 #include <sys/proc.h>
56 #include <sys/pserialize.h>
57 #include <sys/sched.h>
58 #include <sys/sleepq.h>
59 #include <sys/syncobj.h>
60 #include <sys/systm.h>
61 #include <sys/types.h>
62 
63 #include <dev/lockstat.h>
64 
65 #include <machine/lock.h>
66 
67 /*
68  * When not running a debug kernel, spin mutexes are not much
69  * more than an splraiseipl() and splx() pair.
70  */
71 
72 #if defined(DIAGNOSTIC) || defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
73 #define	FULL
74 #endif
75 
76 /*
77  * Debugging support.
78  */
79 
80 #define	MUTEX_WANTLOCK(mtx)					\
81     LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx),		\
82         (uintptr_t)__builtin_return_address(0), 0)
83 #define	MUTEX_TESTLOCK(mtx)					\
84     LOCKDEBUG_WANTLOCK(MUTEX_DEBUG_P(mtx), (mtx),		\
85         (uintptr_t)__builtin_return_address(0), -1)
86 #define	MUTEX_LOCKED(mtx)					\
87     LOCKDEBUG_LOCKED(MUTEX_DEBUG_P(mtx), (mtx), NULL,		\
88         (uintptr_t)__builtin_return_address(0), 0)
89 #define	MUTEX_UNLOCKED(mtx)					\
90     LOCKDEBUG_UNLOCKED(MUTEX_DEBUG_P(mtx), (mtx),		\
91         (uintptr_t)__builtin_return_address(0), 0)
92 #define	MUTEX_ABORT(mtx, msg)					\
93     mutex_abort(__func__, __LINE__, mtx, msg)
94 
95 #if defined(LOCKDEBUG)
96 
97 #define	MUTEX_DASSERT(mtx, cond)				\
98 do {								\
99 	if (__predict_false(!(cond)))				\
100 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
101 } while (/* CONSTCOND */ 0)
102 
103 #else	/* LOCKDEBUG */
104 
105 #define	MUTEX_DASSERT(mtx, cond)	/* nothing */
106 
107 #endif /* LOCKDEBUG */
108 
109 #if defined(DIAGNOSTIC)
110 
111 #define	MUTEX_ASSERT(mtx, cond)					\
112 do {								\
113 	if (__predict_false(!(cond)))				\
114 		MUTEX_ABORT(mtx, "assertion failed: " #cond);	\
115 } while (/* CONSTCOND */ 0)
116 
117 #else	/* DIAGNOSTIC */
118 
119 #define	MUTEX_ASSERT(mtx, cond)	/* nothing */
120 
121 #endif	/* DIAGNOSTIC */
122 
123 /*
124  * Some architectures can't use __cpu_simple_lock as is so allow a way
125  * for them to use an alternate definition.
126  */
127 #ifndef MUTEX_SPINBIT_LOCK_INIT
128 #define MUTEX_SPINBIT_LOCK_INIT(mtx)	__cpu_simple_lock_init(&(mtx)->mtx_lock)
129 #endif
130 #ifndef MUTEX_SPINBIT_LOCKED_P
131 #define MUTEX_SPINBIT_LOCKED_P(mtx)	__SIMPLELOCK_LOCKED_P(&(mtx)->mtx_lock)
132 #endif
133 #ifndef MUTEX_SPINBIT_LOCK_TRY
134 #define MUTEX_SPINBIT_LOCK_TRY(mtx)	__cpu_simple_lock_try(&(mtx)->mtx_lock)
135 #endif
136 #ifndef MUTEX_SPINBIT_LOCK_UNLOCK
137 #define MUTEX_SPINBIT_LOCK_UNLOCK(mtx)	__cpu_simple_unlock(&(mtx)->mtx_lock)
138 #endif
139 
140 #ifndef MUTEX_INITIALIZE_SPIN_IPL
141 #define MUTEX_INITIALIZE_SPIN_IPL(mtx, ipl) \
142 					((mtx)->mtx_ipl = makeiplcookie((ipl)))
143 #endif
144 
145 /*
146  * Spin mutex SPL save / restore.
147  */
148 
149 #define	MUTEX_SPIN_SPLRAISE(mtx)					\
150 do {									\
151 	const int s = splraiseipl(MUTEX_SPIN_IPL(mtx));			\
152 	struct cpu_info * const x__ci = curcpu();			\
153 	const int x__cnt = x__ci->ci_mtx_count--;			\
154 	__insn_barrier();						\
155 	if (x__cnt == 0)						\
156 		x__ci->ci_mtx_oldspl = s;				\
157 } while (/* CONSTCOND */ 0)
158 
159 #define	MUTEX_SPIN_SPLRESTORE(mtx)					\
160 do {									\
161 	struct cpu_info * const x__ci = curcpu();			\
162 	const int s = x__ci->ci_mtx_oldspl;				\
163 	__insn_barrier();						\
164 	if (++(x__ci->ci_mtx_count) == 0)				\
165 		splx(s);						\
166 } while (/* CONSTCOND */ 0)
167 
168 /*
169  * Memory barriers.
170  */
171 #ifdef __HAVE_ATOMIC_AS_MEMBAR
172 #define	MUTEX_MEMBAR_ENTER()
173 #else
174 #define	MUTEX_MEMBAR_ENTER()		membar_enter()
175 #endif
176 
177 /*
178  * For architectures that provide 'simple' mutexes: they provide a
179  * CAS function that is either MP-safe, or does not need to be MP
180  * safe.  Adaptive mutexes on these architectures do not require an
181  * additional interlock.
182  */
183 
184 #ifdef __HAVE_SIMPLE_MUTEXES
185 
186 #define	MUTEX_OWNER(owner)						\
187 	(owner & MUTEX_THREAD)
188 #define	MUTEX_HAS_WAITERS(mtx)						\
189 	(((int)(mtx)->mtx_owner & MUTEX_BIT_WAITERS) != 0)
190 
191 #define	MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug)				\
192 do {									\
193 	if (!dodebug)							\
194 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
195 } while (/* CONSTCOND */ 0)
196 
197 #define	MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl)			\
198 do {									\
199 	(mtx)->mtx_owner = MUTEX_BIT_SPIN;				\
200 	if (!dodebug)							\
201 		(mtx)->mtx_owner |= MUTEX_BIT_NODEBUG;			\
202 	MUTEX_INITIALIZE_SPIN_IPL((mtx), (ipl));			\
203 	MUTEX_SPINBIT_LOCK_INIT((mtx));					\
204 } while (/* CONSTCOND */ 0)
205 
206 #define	MUTEX_DESTROY(mtx)						\
207 do {									\
208 	(mtx)->mtx_owner = MUTEX_THREAD;				\
209 } while (/* CONSTCOND */ 0)
210 
211 #define	MUTEX_SPIN_P(owner)		\
212     (((owner) & MUTEX_BIT_SPIN) != 0)
213 #define	MUTEX_ADAPTIVE_P(owner)		\
214     (((owner) & MUTEX_BIT_SPIN) == 0)
215 
216 #ifndef MUTEX_CAS
217 #define	MUTEX_CAS(p, o, n)		\
218 	(atomic_cas_ulong((volatile unsigned long *)(p), (o), (n)) == (o))
219 #endif /* MUTEX_CAS */
220 
221 #define	MUTEX_DEBUG_P(mtx)	(((mtx)->mtx_owner & MUTEX_BIT_NODEBUG) == 0)
222 #if defined(LOCKDEBUG)
223 #define	MUTEX_OWNED(owner)		(((owner) & ~MUTEX_BIT_NODEBUG) != 0)
224 #define	MUTEX_INHERITDEBUG(n, o)	(n) |= (o) & MUTEX_BIT_NODEBUG
225 #else /* defined(LOCKDEBUG) */
226 #define	MUTEX_OWNED(owner)		((owner) != 0)
227 #define	MUTEX_INHERITDEBUG(n, o)	/* nothing */
228 #endif /* defined(LOCKDEBUG) */
229 
230 static inline int
MUTEX_ACQUIRE(kmutex_t * mtx,uintptr_t curthread)231 MUTEX_ACQUIRE(kmutex_t *mtx, uintptr_t curthread)
232 {
233 	int rv;
234 	uintptr_t oldown = 0;
235 	uintptr_t newown = curthread;
236 
237 	MUTEX_INHERITDEBUG(oldown, mtx->mtx_owner);
238 	MUTEX_INHERITDEBUG(newown, oldown);
239 	rv = MUTEX_CAS(&mtx->mtx_owner, oldown, newown);
240 	membar_acquire();
241 	return rv;
242 }
243 
244 static inline int
MUTEX_SET_WAITERS(kmutex_t * mtx,uintptr_t owner)245 MUTEX_SET_WAITERS(kmutex_t *mtx, uintptr_t owner)
246 {
247 	int rv;
248 
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
MUTEX_RELEASE(kmutex_t * mtx)255 MUTEX_RELEASE(kmutex_t *mtx)
256 {
257 	uintptr_t newown;
258 
259 	newown = 0;
260 	MUTEX_INHERITDEBUG(newown, mtx->mtx_owner);
261 	atomic_store_release(&mtx->mtx_owner, newown);
262 }
263 #endif	/* __HAVE_SIMPLE_MUTEXES */
264 
265 /*
266  * Patch in stubs via strong alias where they are not available.
267  */
268 
269 #if defined(LOCKDEBUG)
270 #undef	__HAVE_MUTEX_STUBS
271 #undef	__HAVE_SPIN_MUTEX_STUBS
272 #endif
273 
274 #ifndef __HAVE_MUTEX_STUBS
275 __strong_alias(mutex_enter,mutex_vector_enter);
276 __strong_alias(mutex_exit,mutex_vector_exit);
277 #endif
278 
279 #ifndef __HAVE_SPIN_MUTEX_STUBS
280 __strong_alias(mutex_spin_enter,mutex_vector_enter);
281 __strong_alias(mutex_spin_exit,mutex_vector_exit);
282 #endif
283 
284 static void	mutex_abort(const char *, size_t, volatile const kmutex_t *,
285 		    const char *);
286 static void	mutex_dump(const volatile void *, lockop_printer_t);
287 static lwp_t	*mutex_owner(wchan_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_name	= "mutex",
303 	.sobj_flag	= SOBJ_SLEEPQ_SORTED,
304 	.sobj_boostpri  = PRI_KERNEL,
305 	.sobj_unsleep	= turnstile_unsleep,
306 	.sobj_changepri	= turnstile_changepri,
307 	.sobj_lendpri	= sleepq_lendpri,
308 	.sobj_owner	= mutex_owner,
309 };
310 
311 /*
312  * mutex_dump:
313  *
314  *	Dump the contents of a mutex structure.
315  */
316 static void
mutex_dump(const volatile void * cookie,lockop_printer_t pr)317 mutex_dump(const volatile void *cookie, lockop_printer_t pr)
318 {
319 	const volatile kmutex_t *mtx = cookie;
320 	uintptr_t owner = mtx->mtx_owner;
321 
322 	pr("owner field  : %#018lx wait/spin: %16d/%d\n",
323 	    (long)MUTEX_OWNER(owner), MUTEX_HAS_WAITERS(mtx),
324 	    MUTEX_SPIN_P(owner));
325 }
326 
327 /*
328  * mutex_abort:
329  *
330  *	Dump information about an error and panic the system.  This
331  *	generates a lot of machine code in the DIAGNOSTIC case, so
332  *	we ask the compiler to not inline it.
333  */
334 static void __noinline
mutex_abort(const char * func,size_t line,volatile const kmutex_t * mtx,const char * msg)335 mutex_abort(const char *func, size_t line, volatile const kmutex_t *mtx,
336     const char *msg)
337 {
338 
339 	LOCKDEBUG_ABORT(func, line, mtx, (MUTEX_SPIN_P(mtx->mtx_owner) ?
340 	    &mutex_spin_lockops : &mutex_adaptive_lockops), msg);
341 }
342 
343 /*
344  * mutex_init:
345  *
346  *	Initialize a mutex for use.  Note that adaptive mutexes are in
347  *	essence spin mutexes that can sleep to avoid deadlock and wasting
348  *	CPU time.  We can't easily provide a type of mutex that always
349  *	sleeps - see comments in mutex_vector_enter() about releasing
350  *	mutexes unlocked.
351  */
352 void
_mutex_init(kmutex_t * mtx,kmutex_type_t type,int ipl,uintptr_t return_address)353 _mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl,
354     uintptr_t return_address)
355 {
356 	lockops_t *lockops __unused;
357 	bool dodebug;
358 
359 	memset(mtx, 0, sizeof(*mtx));
360 
361 	if (ipl == IPL_NONE || ipl == IPL_SOFTCLOCK ||
362 	    ipl == IPL_SOFTBIO || ipl == IPL_SOFTNET ||
363 	    ipl == IPL_SOFTSERIAL) {
364 		lockops = (type == MUTEX_NODEBUG ?
365 		    NULL : &mutex_adaptive_lockops);
366 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
367 		MUTEX_INITIALIZE_ADAPTIVE(mtx, dodebug);
368 	} else {
369 		lockops = (type == MUTEX_NODEBUG ?
370 		    NULL : &mutex_spin_lockops);
371 		dodebug = LOCKDEBUG_ALLOC(mtx, lockops, return_address);
372 		MUTEX_INITIALIZE_SPIN(mtx, dodebug, ipl);
373 	}
374 }
375 
376 void
mutex_init(kmutex_t * mtx,kmutex_type_t type,int ipl)377 mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
378 {
379 
380 	_mutex_init(mtx, type, ipl, (uintptr_t)__builtin_return_address(0));
381 }
382 
383 /*
384  * mutex_destroy:
385  *
386  *	Tear down a mutex.
387  */
388 void
mutex_destroy(kmutex_t * mtx)389 mutex_destroy(kmutex_t *mtx)
390 {
391 	uintptr_t owner = mtx->mtx_owner;
392 
393 	if (MUTEX_ADAPTIVE_P(owner)) {
394 		MUTEX_ASSERT(mtx, !MUTEX_OWNED(owner));
395 		MUTEX_ASSERT(mtx, !MUTEX_HAS_WAITERS(mtx));
396 	} else {
397 		MUTEX_ASSERT(mtx, !MUTEX_SPINBIT_LOCKED_P(mtx));
398 	}
399 
400 	LOCKDEBUG_FREE(MUTEX_DEBUG_P(mtx), mtx);
401 	MUTEX_DESTROY(mtx);
402 }
403 
404 #ifdef MULTIPROCESSOR
405 /*
406  * mutex_oncpu:
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 static bool
mutex_oncpu(uintptr_t owner)413 mutex_oncpu(uintptr_t owner)
414 {
415 	struct cpu_info *ci;
416 	lwp_t *l;
417 
418 	KASSERT(kpreempt_disabled());
419 
420 	if (!MUTEX_OWNED(owner)) {
421 		return false;
422 	}
423 
424 	/*
425 	 * See lwp_dtor() why dereference of the LWP pointer is safe.
426 	 * We must have kernel preemption disabled for that.
427 	 */
428 	l = (lwp_t *)MUTEX_OWNER(owner);
429 	ci = l->l_cpu;
430 
431 	if (ci && ci->ci_curlwp == l) {
432 		/* Target is running; do we need to block? */
433 		return (atomic_load_relaxed(&ci->ci_biglock_wanted) != l);
434 	}
435 
436 	/* Not running.  It may be safe to block now. */
437 	return false;
438 }
439 #endif	/* MULTIPROCESSOR */
440 
441 /*
442  * mutex_vector_enter:
443  *
444  *	Support routine for mutex_enter() that must handle all cases.  In
445  *	the LOCKDEBUG case, mutex_enter() is always aliased here, even if
446  *	fast-path stubs are available.  If a mutex_spin_enter() stub is
447  *	not available, then it is also aliased directly here.
448  */
449 void
mutex_vector_enter(kmutex_t * mtx)450 mutex_vector_enter(kmutex_t *mtx)
451 {
452 	uintptr_t owner, curthread;
453 	turnstile_t *ts;
454 #ifdef MULTIPROCESSOR
455 	u_int count;
456 #endif
457 	LOCKSTAT_COUNTER(spincnt);
458 	LOCKSTAT_COUNTER(slpcnt);
459 	LOCKSTAT_TIMER(spintime);
460 	LOCKSTAT_TIMER(slptime);
461 	LOCKSTAT_FLAG(lsflag);
462 
463 	/*
464 	 * Handle spin mutexes.
465 	 */
466 	KPREEMPT_DISABLE(curlwp);
467 	owner = mtx->mtx_owner;
468 	if (MUTEX_SPIN_P(owner)) {
469 #if defined(LOCKDEBUG) && defined(MULTIPROCESSOR)
470 		u_int spins = 0;
471 #endif
472 		KPREEMPT_ENABLE(curlwp);
473 		MUTEX_SPIN_SPLRAISE(mtx);
474 		MUTEX_WANTLOCK(mtx);
475 #ifdef FULL
476 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
477 			MUTEX_LOCKED(mtx);
478 			return;
479 		}
480 #if !defined(MULTIPROCESSOR)
481 		MUTEX_ABORT(mtx, "locking against myself");
482 #else /* !MULTIPROCESSOR */
483 
484 		LOCKSTAT_ENTER(lsflag);
485 		LOCKSTAT_START_TIMER(lsflag, spintime);
486 		count = SPINLOCK_BACKOFF_MIN;
487 
488 		/*
489 		 * Spin testing the lock word and do exponential backoff
490 		 * to reduce cache line ping-ponging between CPUs.
491 		 */
492 		do {
493 			while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
494 				SPINLOCK_SPIN_HOOK;
495 				SPINLOCK_BACKOFF(count);
496 #ifdef LOCKDEBUG
497 				if (SPINLOCK_SPINOUT(spins))
498 					MUTEX_ABORT(mtx, "spinout");
499 #endif	/* LOCKDEBUG */
500 			}
501 		} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
502 
503 		if (count != SPINLOCK_BACKOFF_MIN) {
504 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
505 			LOCKSTAT_EVENT(lsflag, mtx,
506 			    LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
507 		}
508 		LOCKSTAT_EXIT(lsflag);
509 #endif	/* !MULTIPROCESSOR */
510 #endif	/* FULL */
511 		MUTEX_LOCKED(mtx);
512 		return;
513 	}
514 
515 	curthread = (uintptr_t)curlwp;
516 
517 	MUTEX_DASSERT(mtx, MUTEX_ADAPTIVE_P(owner));
518 	MUTEX_ASSERT(mtx, curthread != 0);
519 	MUTEX_ASSERT(mtx, !cpu_intr_p());
520 	MUTEX_WANTLOCK(mtx);
521 
522 	if (__predict_true(panicstr == NULL)) {
523 		KDASSERT(pserialize_not_in_read_section());
524 		LOCKDEBUG_BARRIER(&kernel_lock, 1);
525 	}
526 
527 	LOCKSTAT_ENTER(lsflag);
528 
529 	/*
530 	 * Adaptive mutex; spin trying to acquire the mutex.  If we
531 	 * determine that the owner is not running on a processor,
532 	 * then we stop spinning, and sleep instead.
533 	 */
534 	for (;;) {
535 		if (!MUTEX_OWNED(owner)) {
536 			/*
537 			 * Mutex owner clear could mean two things:
538 			 *
539 			 *	* The mutex has been released.
540 			 *	* The owner field hasn't been set yet.
541 			 *
542 			 * Try to acquire it again.  If that fails,
543 			 * we'll just loop again.
544 			 */
545 			if (MUTEX_ACQUIRE(mtx, curthread))
546 				break;
547 			owner = mtx->mtx_owner;
548 			continue;
549 		}
550 		if (__predict_false(MUTEX_OWNER(owner) == curthread)) {
551 			MUTEX_ABORT(mtx, "locking against myself");
552 		}
553 #ifdef MULTIPROCESSOR
554 		/*
555 		 * Check to see if the owner is running on a processor.
556 		 * If so, then we should just spin, as the owner will
557 		 * likely release the lock very soon.
558 		 */
559 		if (mutex_oncpu(owner)) {
560 			LOCKSTAT_START_TIMER(lsflag, spintime);
561 			count = SPINLOCK_BACKOFF_MIN;
562 			do {
563 				KPREEMPT_ENABLE(curlwp);
564 				SPINLOCK_BACKOFF(count);
565 				KPREEMPT_DISABLE(curlwp);
566 				owner = mtx->mtx_owner;
567 			} while (mutex_oncpu(owner));
568 			LOCKSTAT_STOP_TIMER(lsflag, spintime);
569 			LOCKSTAT_COUNT(spincnt, 1);
570 			if (!MUTEX_OWNED(owner))
571 				continue;
572 		}
573 #endif
574 
575 		ts = turnstile_lookup(mtx);
576 
577 		/*
578 		 * Once we have the turnstile chain interlock, mark the
579 		 * mutex as having waiters.  If that fails, spin again:
580 		 * chances are that the mutex has been released.
581 		 */
582 		if (!MUTEX_SET_WAITERS(mtx, owner)) {
583 			turnstile_exit(mtx);
584 			owner = mtx->mtx_owner;
585 			continue;
586 		}
587 
588 #ifdef MULTIPROCESSOR
589 		/*
590 		 * mutex_exit() is permitted to release the mutex without
591 		 * any interlocking instructions, and the following can
592 		 * occur as a result:
593 		 *
594 		 *  CPU 1: MUTEX_SET_WAITERS()      CPU2: mutex_exit()
595 		 * ---------------------------- ----------------------------
596 		 *		..		load mtx->mtx_owner
597 		 *		..		see has-waiters bit clear
598 		 *	set has-waiters bit  	           ..
599 		 *		..		store mtx->mtx_owner := 0
600 		 *	  return success
601 		 *
602 		 * There is another race that can occur: a third CPU could
603 		 * acquire the mutex as soon as it is released.  Since
604 		 * adaptive mutexes are primarily spin mutexes, this is not
605 		 * something that we need to worry about too much.  What we
606 		 * do need to ensure is that the waiters bit gets set.
607 		 *
608 		 * To allow the unlocked release, we need to make some
609 		 * assumptions here:
610 		 *
611 		 * o Release is the only non-atomic/unlocked operation
612 		 *   that can be performed on the mutex.  (It must still
613 		 *   be atomic on the local CPU, e.g. in case interrupted
614 		 *   or preempted).
615 		 *
616 		 * o At any given time on each mutex, MUTEX_SET_WAITERS()
617 		 *   can only ever be in progress on one CPU in the
618 		 *   system - guaranteed by the turnstile chain lock.
619 		 *
620 		 * o No other operations other than MUTEX_SET_WAITERS()
621 		 *   and release can modify a mutex with a non-zero
622 		 *   owner field.
623 		 *
624 		 * o If the holding LWP switches away, it posts a store
625 		 *   fence before changing curlwp, ensuring that any
626 		 *   overwrite of the mutex waiters flag by mutex_exit()
627 		 *   completes before the modification of curlwp becomes
628 		 *   visible to this CPU.
629 		 *
630 		 * o cpu_switchto() posts a store fence after setting curlwp
631 		 *   and before resuming execution of an LWP.
632 		 *
633 		 * o _kernel_lock() posts a store fence before setting
634 		 *   curcpu()->ci_biglock_wanted, and after clearing it.
635 		 *   This ensures that any overwrite of the mutex waiters
636 		 *   flag by mutex_exit() completes before the modification
637 		 *   of ci_biglock_wanted becomes visible.
638 		 *
639 		 * After MUTEX_SET_WAITERS() succeeds, simultaneously
640 		 * confirming that the same LWP still holds the mutex
641 		 * since we took the turnstile lock and notifying it that
642 		 * we're waiting, we check the lock holder's status again.
643 		 * Some of the possible outcomes (not an exhaustive list;
644 		 * XXX this should be made exhaustive):
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 		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
mutex_vector_exit(kmutex_t * mtx)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 	 * This also blocks out soft interrupts that could set the
760 	 * waiters bit.
761 	 */
762 	{
763 		int s = splhigh();
764 		if (!MUTEX_HAS_WAITERS(mtx)) {
765 			MUTEX_RELEASE(mtx);
766 			splx(s);
767 			return;
768 		}
769 		splx(s);
770 	}
771 #endif
772 
773 	/*
774 	 * Get this lock's turnstile.  This gets the interlock on
775 	 * the sleep queue.  Once we have that, we can clear the
776 	 * lock.  If there was no turnstile for the lock, there
777 	 * were no waiters remaining.
778 	 */
779 	ts = turnstile_lookup(mtx);
780 
781 	if (ts == NULL) {
782 		MUTEX_RELEASE(mtx);
783 		turnstile_exit(mtx);
784 	} else {
785 		MUTEX_RELEASE(mtx);
786 		turnstile_wakeup(ts, TS_WRITER_Q,
787 		    TS_WAITERS(ts, TS_WRITER_Q), NULL);
788 	}
789 }
790 
791 #ifndef __HAVE_SIMPLE_MUTEXES
792 /*
793  * mutex_wakeup:
794  *
795  *	Support routine for mutex_exit() that wakes up all waiters.
796  *	We assume that the mutex has been released, but it need not
797  *	be.
798  */
799 void
mutex_wakeup(kmutex_t * mtx)800 mutex_wakeup(kmutex_t *mtx)
801 {
802 	turnstile_t *ts;
803 
804 	ts = turnstile_lookup(mtx);
805 	if (ts == NULL) {
806 		turnstile_exit(mtx);
807 		return;
808 	}
809 	MUTEX_CLEAR_WAITERS(mtx);
810 	turnstile_wakeup(ts, TS_WRITER_Q, TS_WAITERS(ts, TS_WRITER_Q), NULL);
811 }
812 #endif	/* !__HAVE_SIMPLE_MUTEXES */
813 
814 /*
815  * mutex_owned:
816  *
817  *	Return true if the current LWP (adaptive) or CPU (spin)
818  *	holds the mutex.
819  */
820 int
mutex_owned(const kmutex_t * mtx)821 mutex_owned(const kmutex_t *mtx)
822 {
823 
824 	if (mtx == NULL)
825 		return 0;
826 	if (MUTEX_ADAPTIVE_P(mtx->mtx_owner))
827 		return MUTEX_OWNER(mtx->mtx_owner) == (uintptr_t)curlwp;
828 #ifdef FULL
829 	return MUTEX_SPINBIT_LOCKED_P(mtx);
830 #else
831 	return 1;
832 #endif
833 }
834 
835 /*
836  * mutex_owner:
837  *
838  *	Return the current owner of an adaptive mutex.  Used for
839  *	priority inheritance.
840  */
841 static lwp_t *
mutex_owner(wchan_t wchan)842 mutex_owner(wchan_t wchan)
843 {
844 	volatile const kmutex_t *mtx = wchan;
845 
846 	MUTEX_ASSERT(mtx, MUTEX_ADAPTIVE_P(mtx->mtx_owner));
847 	return (struct lwp *)MUTEX_OWNER(mtx->mtx_owner);
848 }
849 
850 /*
851  * mutex_ownable:
852  *
853  *	When compiled with DEBUG and LOCKDEBUG defined, ensure that
854  *	the mutex is available.  We cannot use !mutex_owned() since
855  *	that won't work correctly for spin mutexes.
856  */
857 int
mutex_ownable(const kmutex_t * mtx)858 mutex_ownable(const kmutex_t *mtx)
859 {
860 
861 #ifdef LOCKDEBUG
862 	MUTEX_TESTLOCK(mtx);
863 #endif
864 	return 1;
865 }
866 
867 /*
868  * mutex_tryenter:
869  *
870  *	Try to acquire the mutex; return non-zero if we did.
871  */
872 int
mutex_tryenter(kmutex_t * mtx)873 mutex_tryenter(kmutex_t *mtx)
874 {
875 	uintptr_t curthread;
876 
877 	/*
878 	 * Handle spin mutexes.
879 	 */
880 	if (MUTEX_SPIN_P(mtx->mtx_owner)) {
881 		MUTEX_SPIN_SPLRAISE(mtx);
882 #ifdef FULL
883 		if (MUTEX_SPINBIT_LOCK_TRY(mtx)) {
884 			MUTEX_WANTLOCK(mtx);
885 			MUTEX_LOCKED(mtx);
886 			return 1;
887 		}
888 		MUTEX_SPIN_SPLRESTORE(mtx);
889 #else
890 		MUTEX_WANTLOCK(mtx);
891 		MUTEX_LOCKED(mtx);
892 		return 1;
893 #endif
894 	} else {
895 		curthread = (uintptr_t)curlwp;
896 		MUTEX_ASSERT(mtx, curthread != 0);
897 		if (MUTEX_ACQUIRE(mtx, curthread)) {
898 			MUTEX_WANTLOCK(mtx);
899 			MUTEX_LOCKED(mtx);
900 			MUTEX_DASSERT(mtx,
901 			    MUTEX_OWNER(mtx->mtx_owner) == curthread);
902 			return 1;
903 		}
904 	}
905 
906 	return 0;
907 }
908 
909 #if defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL)
910 /*
911  * mutex_spin_retry:
912  *
913  *	Support routine for mutex_spin_enter().  Assumes that the caller
914  *	has already raised the SPL, and adjusted counters.
915  */
916 void
mutex_spin_retry(kmutex_t * mtx)917 mutex_spin_retry(kmutex_t *mtx)
918 {
919 #ifdef MULTIPROCESSOR
920 	u_int count;
921 	LOCKSTAT_TIMER(spintime);
922 	LOCKSTAT_FLAG(lsflag);
923 #ifdef LOCKDEBUG
924 	u_int spins = 0;
925 #endif	/* LOCKDEBUG */
926 
927 	MUTEX_WANTLOCK(mtx);
928 
929 	LOCKSTAT_ENTER(lsflag);
930 	LOCKSTAT_START_TIMER(lsflag, spintime);
931 	count = SPINLOCK_BACKOFF_MIN;
932 
933 	/*
934 	 * Spin testing the lock word and do exponential backoff
935 	 * to reduce cache line ping-ponging between CPUs.
936 	 */
937 	do {
938 		while (MUTEX_SPINBIT_LOCKED_P(mtx)) {
939 			SPINLOCK_BACKOFF(count);
940 #ifdef LOCKDEBUG
941 			if (SPINLOCK_SPINOUT(spins))
942 				MUTEX_ABORT(mtx, "spinout");
943 #endif	/* LOCKDEBUG */
944 		}
945 	} while (!MUTEX_SPINBIT_LOCK_TRY(mtx));
946 
947 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
948 	LOCKSTAT_EVENT(lsflag, mtx, LB_SPIN_MUTEX | LB_SPIN, 1, spintime);
949 	LOCKSTAT_EXIT(lsflag);
950 
951 	MUTEX_LOCKED(mtx);
952 #else	/* MULTIPROCESSOR */
953 	MUTEX_ABORT(mtx, "locking against myself");
954 #endif	/* MULTIPROCESSOR */
955 }
956 #endif	/* defined(__HAVE_SPIN_MUTEX_STUBS) || defined(FULL) */
957