xref: /freebsd/sys/kern/kern_sx.c (revision 137a344c6341d1469432e9deb3a25593f96672ad)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
5  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice(s), this list of conditions and the following disclaimer as
13  *    the first lines of this file unmodified other than the possible
14  *    addition of one or more copyright notices.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice(s), 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 COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
20  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
23  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
24  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
25  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
26  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
29  * DAMAGE.
30  */
31 
32 /*
33  * Shared/exclusive locks.  This implementation attempts to ensure
34  * deterministic lock granting behavior, so that slocks and xlocks are
35  * interleaved.
36  *
37  * Priority propagation will not generally raise the priority of lock holders,
38  * so should not be relied upon in combination with sx locks.
39  */
40 
41 #include "opt_ddb.h"
42 #include "opt_hwpmc_hooks.h"
43 #include "opt_no_adaptive_sx.h"
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/kdb.h>
51 #include <sys/kernel.h>
52 #include <sys/ktr.h>
53 #include <sys/lock.h>
54 #include <sys/mutex.h>
55 #include <sys/proc.h>
56 #include <sys/sched.h>
57 #include <sys/sleepqueue.h>
58 #include <sys/sx.h>
59 #include <sys/smp.h>
60 #include <sys/sysctl.h>
61 
62 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
63 #include <machine/cpu.h>
64 #endif
65 
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #endif
69 
70 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
71 #define	ADAPTIVE_SX
72 #endif
73 
74 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
75 
76 #ifdef HWPMC_HOOKS
77 #include <sys/pmckern.h>
78 PMC_SOFT_DECLARE( , , lock, failed);
79 #endif
80 
81 /* Handy macros for sleep queues. */
82 #define	SQ_EXCLUSIVE_QUEUE	0
83 #define	SQ_SHARED_QUEUE		1
84 
85 /*
86  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
87  * drop Giant anytime we have to sleep or if we adaptively spin.
88  */
89 #define	GIANT_DECLARE							\
90 	int _giantcnt = 0;						\
91 	WITNESS_SAVE_DECL(Giant)					\
92 
93 #define	GIANT_SAVE(work) do {						\
94 	if (__predict_false(mtx_owned(&Giant))) {			\
95 		work++;							\
96 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
97 		while (mtx_owned(&Giant)) {				\
98 			_giantcnt++;					\
99 			mtx_unlock(&Giant);				\
100 		}							\
101 	}								\
102 } while (0)
103 
104 #define GIANT_RESTORE() do {						\
105 	if (_giantcnt > 0) {						\
106 		mtx_assert(&Giant, MA_NOTOWNED);			\
107 		while (_giantcnt--)					\
108 			mtx_lock(&Giant);				\
109 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
110 	}								\
111 } while (0)
112 
113 /*
114  * Returns true if an exclusive lock is recursed.  It assumes
115  * curthread currently has an exclusive lock.
116  */
117 #define	sx_recursed(sx)		((sx)->sx_recurse != 0)
118 
119 static void	assert_sx(const struct lock_object *lock, int what);
120 #ifdef DDB
121 static void	db_show_sx(const struct lock_object *lock);
122 #endif
123 static void	lock_sx(struct lock_object *lock, uintptr_t how);
124 #ifdef KDTRACE_HOOKS
125 static int	owner_sx(const struct lock_object *lock, struct thread **owner);
126 #endif
127 static uintptr_t unlock_sx(struct lock_object *lock);
128 
129 struct lock_class lock_class_sx = {
130 	.lc_name = "sx",
131 	.lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
132 	.lc_assert = assert_sx,
133 #ifdef DDB
134 	.lc_ddb_show = db_show_sx,
135 #endif
136 	.lc_lock = lock_sx,
137 	.lc_unlock = unlock_sx,
138 #ifdef KDTRACE_HOOKS
139 	.lc_owner = owner_sx,
140 #endif
141 };
142 
143 #ifndef INVARIANTS
144 #define	_sx_assert(sx, what, file, line)
145 #endif
146 
147 #ifdef ADAPTIVE_SX
148 static __read_frequently u_int asx_retries = 10;
149 static __read_frequently u_int asx_loops = 10000;
150 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
151 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
152 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
153 
154 static struct lock_delay_config __read_frequently sx_delay;
155 
156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
157     0, "");
158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
159     0, "");
160 
161 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
162 #endif
163 
164 void
165 assert_sx(const struct lock_object *lock, int what)
166 {
167 
168 	sx_assert((const struct sx *)lock, what);
169 }
170 
171 void
172 lock_sx(struct lock_object *lock, uintptr_t how)
173 {
174 	struct sx *sx;
175 
176 	sx = (struct sx *)lock;
177 	if (how)
178 		sx_slock(sx);
179 	else
180 		sx_xlock(sx);
181 }
182 
183 uintptr_t
184 unlock_sx(struct lock_object *lock)
185 {
186 	struct sx *sx;
187 
188 	sx = (struct sx *)lock;
189 	sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
190 	if (sx_xlocked(sx)) {
191 		sx_xunlock(sx);
192 		return (0);
193 	} else {
194 		sx_sunlock(sx);
195 		return (1);
196 	}
197 }
198 
199 #ifdef KDTRACE_HOOKS
200 int
201 owner_sx(const struct lock_object *lock, struct thread **owner)
202 {
203 	const struct sx *sx;
204 	uintptr_t x;
205 
206 	sx = (const struct sx *)lock;
207 	x = sx->sx_lock;
208 	*owner = NULL;
209 	return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
210 	    ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
211 }
212 #endif
213 
214 void
215 sx_sysinit(void *arg)
216 {
217 	struct sx_args *sargs = arg;
218 
219 	sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
220 }
221 
222 void
223 sx_init_flags(struct sx *sx, const char *description, int opts)
224 {
225 	int flags;
226 
227 	MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
228 	    SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
229 	ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
230 	    ("%s: sx_lock not aligned for %s: %p", __func__, description,
231 	    &sx->sx_lock));
232 
233 	flags = LO_SLEEPABLE | LO_UPGRADABLE;
234 	if (opts & SX_DUPOK)
235 		flags |= LO_DUPOK;
236 	if (opts & SX_NOPROFILE)
237 		flags |= LO_NOPROFILE;
238 	if (!(opts & SX_NOWITNESS))
239 		flags |= LO_WITNESS;
240 	if (opts & SX_RECURSE)
241 		flags |= LO_RECURSABLE;
242 	if (opts & SX_QUIET)
243 		flags |= LO_QUIET;
244 	if (opts & SX_NEW)
245 		flags |= LO_NEW;
246 
247 	flags |= opts & SX_NOADAPTIVE;
248 	lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
249 	sx->sx_lock = SX_LOCK_UNLOCKED;
250 	sx->sx_recurse = 0;
251 }
252 
253 void
254 sx_destroy(struct sx *sx)
255 {
256 
257 	KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
258 	KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
259 	sx->sx_lock = SX_LOCK_DESTROYED;
260 	lock_destroy(&sx->lock_object);
261 }
262 
263 int
264 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
265 {
266 	uintptr_t x;
267 
268 	if (SCHEDULER_STOPPED())
269 		return (1);
270 
271 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
272 	    ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
273 	    curthread, sx->lock_object.lo_name, file, line));
274 
275 	x = sx->sx_lock;
276 	for (;;) {
277 		KASSERT(x != SX_LOCK_DESTROYED,
278 		    ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
279 		if (!(x & SX_LOCK_SHARED))
280 			break;
281 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
282 			LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
283 			WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
284 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
285 			    sx, 0, 0, file, line, LOCKSTAT_READER);
286 			TD_LOCKS_INC(curthread);
287 			return (1);
288 		}
289 	}
290 
291 	LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
292 	return (0);
293 }
294 
295 int
296 sx_try_slock_(struct sx *sx, const char *file, int line)
297 {
298 
299 	return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
300 }
301 
302 int
303 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
304 {
305 	uintptr_t tid, x;
306 	int error = 0;
307 
308 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
309 	    !TD_IS_IDLETHREAD(curthread),
310 	    ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
311 	    curthread, sx->lock_object.lo_name, file, line));
312 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
313 	    ("sx_xlock() of destroyed sx @ %s:%d", file, line));
314 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
315 	    line, NULL);
316 	tid = (uintptr_t)curthread;
317 	x = SX_LOCK_UNLOCKED;
318 	if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
319 		error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
320 	else
321 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
322 		    0, 0, file, line, LOCKSTAT_WRITER);
323 	if (!error) {
324 		LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
325 		    file, line);
326 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
327 		TD_LOCKS_INC(curthread);
328 	}
329 
330 	return (error);
331 }
332 
333 int
334 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
335 {
336 	struct thread *td;
337 	uintptr_t tid, x;
338 	int rval;
339 	bool recursed;
340 
341 	td = curthread;
342 	tid = (uintptr_t)td;
343 	if (SCHEDULER_STOPPED_TD(td))
344 		return (1);
345 
346 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
347 	    ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
348 	    curthread, sx->lock_object.lo_name, file, line));
349 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
350 	    ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
351 
352 	rval = 1;
353 	recursed = false;
354 	x = SX_LOCK_UNLOCKED;
355 	for (;;) {
356 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
357 			break;
358 		if (x == SX_LOCK_UNLOCKED)
359 			continue;
360 		if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
361 			sx->sx_recurse++;
362 			atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
363 			break;
364 		}
365 		rval = 0;
366 		break;
367 	}
368 
369 	LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
370 	if (rval) {
371 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
372 		    file, line);
373 		if (!recursed)
374 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
375 			    sx, 0, 0, file, line, LOCKSTAT_WRITER);
376 		TD_LOCKS_INC(curthread);
377 	}
378 
379 	return (rval);
380 }
381 
382 int
383 sx_try_xlock_(struct sx *sx, const char *file, int line)
384 {
385 
386 	return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
387 }
388 
389 void
390 _sx_xunlock(struct sx *sx, const char *file, int line)
391 {
392 
393 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
394 	    ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
395 	_sx_assert(sx, SA_XLOCKED, file, line);
396 	WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
397 	LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
398 	    line);
399 #if LOCK_DEBUG > 0
400 	_sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
401 #else
402 	__sx_xunlock(sx, curthread, file, line);
403 #endif
404 	TD_LOCKS_DEC(curthread);
405 }
406 
407 /*
408  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
409  * This will only succeed if this thread holds a single shared lock.
410  * Return 1 if if the upgrade succeed, 0 otherwise.
411  */
412 int
413 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
414 {
415 	uintptr_t x;
416 	int success;
417 
418 	if (SCHEDULER_STOPPED())
419 		return (1);
420 
421 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
422 	    ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
423 	_sx_assert(sx, SA_SLOCKED, file, line);
424 
425 	/*
426 	 * Try to switch from one shared lock to an exclusive lock.  We need
427 	 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
428 	 * we will wake up the exclusive waiters when we drop the lock.
429 	 */
430 	x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
431 	success = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
432 	    (uintptr_t)curthread | x);
433 	LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
434 	if (success) {
435 		WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
436 		    file, line);
437 		LOCKSTAT_RECORD0(sx__upgrade, sx);
438 	}
439 	return (success);
440 }
441 
442 int
443 sx_try_upgrade_(struct sx *sx, const char *file, int line)
444 {
445 
446 	return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
447 }
448 
449 /*
450  * Downgrade an unrecursed exclusive lock into a single shared lock.
451  */
452 void
453 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
454 {
455 	uintptr_t x;
456 	int wakeup_swapper;
457 
458 	if (SCHEDULER_STOPPED())
459 		return;
460 
461 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
462 	    ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
463 	_sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
464 #ifndef INVARIANTS
465 	if (sx_recursed(sx))
466 		panic("downgrade of a recursed lock");
467 #endif
468 
469 	WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
470 
471 	/*
472 	 * Try to switch from an exclusive lock with no shared waiters
473 	 * to one sharer with no shared waiters.  If there are
474 	 * exclusive waiters, we don't need to lock the sleep queue so
475 	 * long as we preserve the flag.  We do one quick try and if
476 	 * that fails we grab the sleepq lock to keep the flags from
477 	 * changing and do it the slow way.
478 	 *
479 	 * We have to lock the sleep queue if there are shared waiters
480 	 * so we can wake them up.
481 	 */
482 	x = sx->sx_lock;
483 	if (!(x & SX_LOCK_SHARED_WAITERS) &&
484 	    atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
485 	    (x & SX_LOCK_EXCLUSIVE_WAITERS)))
486 		goto out;
487 
488 	/*
489 	 * Lock the sleep queue so we can read the waiters bits
490 	 * without any races and wakeup any shared waiters.
491 	 */
492 	sleepq_lock(&sx->lock_object);
493 
494 	/*
495 	 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
496 	 * shared lock.  If there are any shared waiters, wake them up.
497 	 */
498 	wakeup_swapper = 0;
499 	x = sx->sx_lock;
500 	atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
501 	    (x & SX_LOCK_EXCLUSIVE_WAITERS));
502 	if (x & SX_LOCK_SHARED_WAITERS)
503 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
504 		    0, SQ_SHARED_QUEUE);
505 	sleepq_release(&sx->lock_object);
506 
507 	if (wakeup_swapper)
508 		kick_proc0();
509 
510 out:
511 	LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
512 	LOCKSTAT_RECORD0(sx__downgrade, sx);
513 }
514 
515 void
516 sx_downgrade_(struct sx *sx, const char *file, int line)
517 {
518 
519 	sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
520 }
521 
522 /*
523  * This function represents the so-called 'hard case' for sx_xlock
524  * operation.  All 'easy case' failures are redirected to this.  Note
525  * that ideally this would be a static function, but it needs to be
526  * accessible from at least sx.h.
527  */
528 int
529 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
530 {
531 	GIANT_DECLARE;
532 	uintptr_t tid;
533 #ifdef ADAPTIVE_SX
534 	volatile struct thread *owner;
535 	u_int i, n, spintries = 0;
536 	bool adaptive;
537 	int sleep_reason = 0;
538 #endif
539 #ifdef LOCK_PROFILING
540 	uint64_t waittime = 0;
541 	int contested = 0;
542 #endif
543 	int error = 0;
544 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
545 	struct lock_delay_arg lda;
546 #endif
547 #ifdef	KDTRACE_HOOKS
548 	u_int sleep_cnt = 0;
549 	int64_t sleep_time = 0;
550 	int64_t all_time = 0;
551 #endif
552 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
553 	uintptr_t state;
554 #endif
555 	int extra_work = 0;
556 
557 	tid = (uintptr_t)curthread;
558 	if (SCHEDULER_STOPPED())
559 		return (0);
560 
561 #if defined(ADAPTIVE_SX)
562 	lock_delay_arg_init(&lda, &sx_delay);
563 #elif defined(KDTRACE_HOOKS)
564 	lock_delay_arg_init(&lda, NULL);
565 #endif
566 
567 	if (__predict_false(x == SX_LOCK_UNLOCKED))
568 		x = SX_READ_VALUE(sx);
569 
570 	/* If we already hold an exclusive lock, then recurse. */
571 	if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
572 		KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
573 	    ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
574 		    sx->lock_object.lo_name, file, line));
575 		sx->sx_recurse++;
576 		atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
577 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
578 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
579 		return (0);
580 	}
581 
582 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
583 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
584 		    sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
585 
586 #ifdef ADAPTIVE_SX
587 	adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) != 0);
588 #endif
589 
590 #ifdef HWPMC_HOOKS
591 	PMC_SOFT_CALL( , , lock, failed);
592 #endif
593 	lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
594 	    &waittime);
595 
596 #ifdef LOCK_PROFILING
597 	extra_work = 1;
598 	state = x;
599 #elif defined(KDTRACE_HOOKS)
600 	extra_work = lockstat_enabled;
601 	if (__predict_false(extra_work)) {
602 		all_time -= lockstat_nsecs(&sx->lock_object);
603 		state = x;
604 	}
605 #endif
606 #ifndef INVARIANTS
607 	GIANT_SAVE(extra_work);
608 #endif
609 
610 	for (;;) {
611 		if (x == SX_LOCK_UNLOCKED) {
612 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
613 				break;
614 			continue;
615 		}
616 #ifdef INVARIANTS
617 		GIANT_SAVE(extra_work);
618 #endif
619 #ifdef KDTRACE_HOOKS
620 		lda.spin_cnt++;
621 #endif
622 #ifdef ADAPTIVE_SX
623 		if (__predict_false(!adaptive))
624 			goto sleepq;
625 		/*
626 		 * If the lock is write locked and the owner is
627 		 * running on another CPU, spin until the owner stops
628 		 * running or the state of the lock changes.
629 		 */
630 		if ((x & SX_LOCK_SHARED) == 0) {
631 			owner = lv_sx_owner(x);
632 			if (TD_IS_RUNNING(owner)) {
633 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
634 					CTR3(KTR_LOCK,
635 				    "%s: spinning on %p held by %p",
636 					    __func__, sx, owner);
637 				KTR_STATE1(KTR_SCHED, "thread",
638 				    sched_tdname(curthread), "spinning",
639 				    "lockname:\"%s\"",
640 				    sx->lock_object.lo_name);
641 				do {
642 					lock_delay(&lda);
643 					x = SX_READ_VALUE(sx);
644 					owner = lv_sx_owner(x);
645 				} while (owner != NULL &&
646 					    TD_IS_RUNNING(owner));
647 				KTR_STATE0(KTR_SCHED, "thread",
648 				    sched_tdname(curthread), "running");
649 				continue;
650 			}
651 			sleep_reason = 1;
652 		} else if (SX_SHARERS(x) && spintries < asx_retries) {
653 			KTR_STATE1(KTR_SCHED, "thread",
654 			    sched_tdname(curthread), "spinning",
655 			    "lockname:\"%s\"", sx->lock_object.lo_name);
656 			spintries++;
657 			for (i = 0; i < asx_loops; i += n) {
658 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
659 					CTR4(KTR_LOCK,
660 			    "%s: shared spinning on %p with %u and %u",
661 					    __func__, sx, spintries, i);
662 				n = SX_SHARERS(x);
663 				lock_delay_spin(n);
664 				x = SX_READ_VALUE(sx);
665 				if ((x & SX_LOCK_SHARED) == 0 ||
666 				    SX_SHARERS(x) == 0)
667 					break;
668 			}
669 #ifdef KDTRACE_HOOKS
670 			lda.spin_cnt += i;
671 #endif
672 			KTR_STATE0(KTR_SCHED, "thread",
673 			    sched_tdname(curthread), "running");
674 			if (i < asx_loops)
675 				continue;
676 			sleep_reason = 2;
677 		}
678 sleepq:
679 #endif
680 		sleepq_lock(&sx->lock_object);
681 		x = SX_READ_VALUE(sx);
682 retry_sleepq:
683 
684 		/*
685 		 * If the lock was released while spinning on the
686 		 * sleep queue chain lock, try again.
687 		 */
688 		if (x == SX_LOCK_UNLOCKED) {
689 			sleepq_release(&sx->lock_object);
690 			continue;
691 		}
692 
693 #ifdef ADAPTIVE_SX
694 		/*
695 		 * The current lock owner might have started executing
696 		 * on another CPU (or the lock could have changed
697 		 * owners) while we were waiting on the sleep queue
698 		 * chain lock.  If so, drop the sleep queue lock and try
699 		 * again.
700 		 */
701 		if (adaptive) {
702 			if (!(x & SX_LOCK_SHARED)) {
703 				owner = (struct thread *)SX_OWNER(x);
704 				if (TD_IS_RUNNING(owner)) {
705 					sleepq_release(&sx->lock_object);
706 					continue;
707 				}
708 			} else if (SX_SHARERS(x) > 0 && sleep_reason == 1) {
709 				sleepq_release(&sx->lock_object);
710 				continue;
711 			}
712 		}
713 #endif
714 
715 		/*
716 		 * If an exclusive lock was released with both shared
717 		 * and exclusive waiters and a shared waiter hasn't
718 		 * woken up and acquired the lock yet, sx_lock will be
719 		 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
720 		 * If we see that value, try to acquire it once.  Note
721 		 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
722 		 * as there are other exclusive waiters still.  If we
723 		 * fail, restart the loop.
724 		 */
725 		if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
726 			if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
727 			    tid | SX_LOCK_EXCLUSIVE_WAITERS))
728 				goto retry_sleepq;
729 			sleepq_release(&sx->lock_object);
730 			CTR2(KTR_LOCK, "%s: %p claimed by new writer",
731 			    __func__, sx);
732 			break;
733 		}
734 
735 		/*
736 		 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
737 		 * than loop back and retry.
738 		 */
739 		if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
740 			if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
741 			    x | SX_LOCK_EXCLUSIVE_WAITERS)) {
742 				goto retry_sleepq;
743 			}
744 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
745 				CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
746 				    __func__, sx);
747 		}
748 
749 		/*
750 		 * Since we have been unable to acquire the exclusive
751 		 * lock and the exclusive waiters flag is set, we have
752 		 * to sleep.
753 		 */
754 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
755 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
756 			    __func__, sx);
757 
758 #ifdef KDTRACE_HOOKS
759 		sleep_time -= lockstat_nsecs(&sx->lock_object);
760 #endif
761 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
762 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
763 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
764 		if (!(opts & SX_INTERRUPTIBLE))
765 			sleepq_wait(&sx->lock_object, 0);
766 		else
767 			error = sleepq_wait_sig(&sx->lock_object, 0);
768 #ifdef KDTRACE_HOOKS
769 		sleep_time += lockstat_nsecs(&sx->lock_object);
770 		sleep_cnt++;
771 #endif
772 		if (error) {
773 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
774 				CTR2(KTR_LOCK,
775 			"%s: interruptible sleep by %p suspended by signal",
776 				    __func__, sx);
777 			break;
778 		}
779 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
780 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
781 			    __func__, sx);
782 		x = SX_READ_VALUE(sx);
783 	}
784 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
785 	if (__predict_true(!extra_work))
786 		return (error);
787 #endif
788 #ifdef KDTRACE_HOOKS
789 	all_time += lockstat_nsecs(&sx->lock_object);
790 	if (sleep_time)
791 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
792 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
793 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
794 	if (lda.spin_cnt > sleep_cnt)
795 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
796 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
797 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
798 #endif
799 	if (!error)
800 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
801 		    contested, waittime, file, line, LOCKSTAT_WRITER);
802 	GIANT_RESTORE();
803 	return (error);
804 }
805 
806 /*
807  * This function represents the so-called 'hard case' for sx_xunlock
808  * operation.  All 'easy case' failures are redirected to this.  Note
809  * that ideally this would be a static function, but it needs to be
810  * accessible from at least sx.h.
811  */
812 void
813 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
814 {
815 	uintptr_t tid, setx;
816 	int queue, wakeup_swapper;
817 
818 	if (SCHEDULER_STOPPED())
819 		return;
820 
821 	tid = (uintptr_t)curthread;
822 
823 	if (__predict_false(x == tid))
824 		x = SX_READ_VALUE(sx);
825 
826 	MPASS(!(x & SX_LOCK_SHARED));
827 
828 	if (__predict_false(x & SX_LOCK_RECURSED)) {
829 		/* The lock is recursed, unrecurse one level. */
830 		if ((--sx->sx_recurse) == 0)
831 			atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
832 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
833 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
834 		return;
835 	}
836 
837 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
838 	if (x == tid &&
839 	    atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
840 		return;
841 
842 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
843 		CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
844 
845 	sleepq_lock(&sx->lock_object);
846 	x = SX_READ_VALUE(sx);
847 	MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
848 
849 	/*
850 	 * The wake up algorithm here is quite simple and probably not
851 	 * ideal.  It gives precedence to shared waiters if they are
852 	 * present.  For this condition, we have to preserve the
853 	 * state of the exclusive waiters flag.
854 	 * If interruptible sleeps left the shared queue empty avoid a
855 	 * starvation for the threads sleeping on the exclusive queue by giving
856 	 * them precedence and cleaning up the shared waiters bit anyway.
857 	 */
858 	setx = SX_LOCK_UNLOCKED;
859 	queue = SQ_EXCLUSIVE_QUEUE;
860 	if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
861 	    sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
862 		queue = SQ_SHARED_QUEUE;
863 		setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
864 	}
865 	atomic_store_rel_ptr(&sx->sx_lock, setx);
866 
867 	/* Wake up all the waiters for the specific queue. */
868 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
869 		CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
870 		    __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
871 		    "exclusive");
872 
873 	wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
874 	    queue);
875 	sleepq_release(&sx->lock_object);
876 	if (wakeup_swapper)
877 		kick_proc0();
878 }
879 
880 static bool __always_inline
881 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
882 {
883 
884 	/*
885 	 * If no other thread has an exclusive lock then try to bump up
886 	 * the count of sharers.  Since we have to preserve the state
887 	 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
888 	 * shared lock loop back and retry.
889 	 */
890 	while (*xp & SX_LOCK_SHARED) {
891 		MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
892 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
893 		    *xp + SX_ONE_SHARER)) {
894 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
895 				CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
896 				    __func__, sx, (void *)*xp,
897 				    (void *)(*xp + SX_ONE_SHARER));
898 			return (true);
899 		}
900 	}
901 	return (false);
902 }
903 
904 static int __noinline
905 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
906 {
907 	GIANT_DECLARE;
908 #ifdef ADAPTIVE_SX
909 	volatile struct thread *owner;
910 	bool adaptive;
911 #endif
912 #ifdef LOCK_PROFILING
913 	uint64_t waittime = 0;
914 	int contested = 0;
915 #endif
916 	int error = 0;
917 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
918 	struct lock_delay_arg lda;
919 #endif
920 #ifdef KDTRACE_HOOKS
921 	u_int sleep_cnt = 0;
922 	int64_t sleep_time = 0;
923 	int64_t all_time = 0;
924 #endif
925 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
926 	uintptr_t state;
927 #endif
928 	int extra_work = 0;
929 
930 	if (SCHEDULER_STOPPED())
931 		return (0);
932 
933 #if defined(ADAPTIVE_SX)
934 	lock_delay_arg_init(&lda, &sx_delay);
935 #elif defined(KDTRACE_HOOKS)
936 	lock_delay_arg_init(&lda, NULL);
937 #endif
938 
939 #ifdef ADAPTIVE_SX
940 	adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) != 0);
941 #endif
942 
943 #ifdef HWPMC_HOOKS
944 	PMC_SOFT_CALL( , , lock, failed);
945 #endif
946 	lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
947 	    &waittime);
948 
949 #ifdef LOCK_PROFILING
950 	extra_work = 1;
951 	state = x;
952 #elif defined(KDTRACE_HOOKS)
953 	extra_work = lockstat_enabled;
954 	if (__predict_false(extra_work)) {
955 		all_time -= lockstat_nsecs(&sx->lock_object);
956 		state = x;
957 	}
958 #endif
959 #ifndef INVARIANTS
960 	GIANT_SAVE(extra_work);
961 #endif
962 
963 	/*
964 	 * As with rwlocks, we don't make any attempt to try to block
965 	 * shared locks once there is an exclusive waiter.
966 	 */
967 	for (;;) {
968 		if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
969 			break;
970 #ifdef INVARIANTS
971 		GIANT_SAVE(extra_work);
972 #endif
973 #ifdef KDTRACE_HOOKS
974 		lda.spin_cnt++;
975 #endif
976 
977 #ifdef ADAPTIVE_SX
978 		if (__predict_false(!adaptive))
979 			goto sleepq;
980 		/*
981 		 * If the owner is running on another CPU, spin until
982 		 * the owner stops running or the state of the lock
983 		 * changes.
984 		 */
985 		owner = lv_sx_owner(x);
986 		if (TD_IS_RUNNING(owner)) {
987 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
988 				CTR3(KTR_LOCK,
989 				    "%s: spinning on %p held by %p",
990 				    __func__, sx, owner);
991 			KTR_STATE1(KTR_SCHED, "thread",
992 			    sched_tdname(curthread), "spinning",
993 			    "lockname:\"%s\"", sx->lock_object.lo_name);
994 			do {
995 				lock_delay(&lda);
996 				x = SX_READ_VALUE(sx);
997 				owner = lv_sx_owner(x);
998 			} while (owner != NULL && TD_IS_RUNNING(owner));
999 			KTR_STATE0(KTR_SCHED, "thread",
1000 			    sched_tdname(curthread), "running");
1001 			continue;
1002 		}
1003 sleepq:
1004 #endif
1005 
1006 		/*
1007 		 * Some other thread already has an exclusive lock, so
1008 		 * start the process of blocking.
1009 		 */
1010 		sleepq_lock(&sx->lock_object);
1011 		x = SX_READ_VALUE(sx);
1012 retry_sleepq:
1013 		/*
1014 		 * The lock could have been released while we spun.
1015 		 * In this case loop back and retry.
1016 		 */
1017 		if (x & SX_LOCK_SHARED) {
1018 			sleepq_release(&sx->lock_object);
1019 			continue;
1020 		}
1021 
1022 #ifdef ADAPTIVE_SX
1023 		/*
1024 		 * If the owner is running on another CPU, spin until
1025 		 * the owner stops running or the state of the lock
1026 		 * changes.
1027 		 */
1028 		if (!(x & SX_LOCK_SHARED) && adaptive) {
1029 			owner = (struct thread *)SX_OWNER(x);
1030 			if (TD_IS_RUNNING(owner)) {
1031 				sleepq_release(&sx->lock_object);
1032 				x = SX_READ_VALUE(sx);
1033 				continue;
1034 			}
1035 		}
1036 #endif
1037 
1038 		/*
1039 		 * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1040 		 * fail to set it drop the sleep queue lock and loop
1041 		 * back.
1042 		 */
1043 		if (!(x & SX_LOCK_SHARED_WAITERS)) {
1044 			if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1045 			    x | SX_LOCK_SHARED_WAITERS))
1046 				goto retry_sleepq;
1047 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1048 				CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1049 				    __func__, sx);
1050 		}
1051 
1052 		/*
1053 		 * Since we have been unable to acquire the shared lock,
1054 		 * we have to sleep.
1055 		 */
1056 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1057 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1058 			    __func__, sx);
1059 
1060 #ifdef KDTRACE_HOOKS
1061 		sleep_time -= lockstat_nsecs(&sx->lock_object);
1062 #endif
1063 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1064 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1065 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1066 		if (!(opts & SX_INTERRUPTIBLE))
1067 			sleepq_wait(&sx->lock_object, 0);
1068 		else
1069 			error = sleepq_wait_sig(&sx->lock_object, 0);
1070 #ifdef KDTRACE_HOOKS
1071 		sleep_time += lockstat_nsecs(&sx->lock_object);
1072 		sleep_cnt++;
1073 #endif
1074 		if (error) {
1075 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1076 				CTR2(KTR_LOCK,
1077 			"%s: interruptible sleep by %p suspended by signal",
1078 				    __func__, sx);
1079 			break;
1080 		}
1081 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1082 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1083 			    __func__, sx);
1084 		x = SX_READ_VALUE(sx);
1085 	}
1086 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1087 	if (__predict_true(!extra_work))
1088 		return (error);
1089 #endif
1090 #ifdef KDTRACE_HOOKS
1091 	all_time += lockstat_nsecs(&sx->lock_object);
1092 	if (sleep_time)
1093 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1094 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1095 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1096 	if (lda.spin_cnt > sleep_cnt)
1097 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1098 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1099 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1100 #endif
1101 	if (error == 0) {
1102 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1103 		    contested, waittime, file, line, LOCKSTAT_READER);
1104 	}
1105 	GIANT_RESTORE();
1106 	return (error);
1107 }
1108 
1109 int
1110 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1111 {
1112 	uintptr_t x;
1113 	int error;
1114 
1115 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1116 	    !TD_IS_IDLETHREAD(curthread),
1117 	    ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1118 	    curthread, sx->lock_object.lo_name, file, line));
1119 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1120 	    ("sx_slock() of destroyed sx @ %s:%d", file, line));
1121 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1122 
1123 	error = 0;
1124 	x = SX_READ_VALUE(sx);
1125 	if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1126 	    !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1127 		error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1128 	if (error == 0) {
1129 		LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1130 		WITNESS_LOCK(&sx->lock_object, 0, file, line);
1131 		TD_LOCKS_INC(curthread);
1132 	}
1133 	return (error);
1134 }
1135 
1136 int
1137 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1138 {
1139 
1140 	return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1141 }
1142 
1143 static bool __always_inline
1144 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1145 {
1146 
1147 	for (;;) {
1148 		/*
1149 		 * We should never have sharers while at least one thread
1150 		 * holds a shared lock.
1151 		 */
1152 		KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1153 		    ("%s: waiting sharers", __func__));
1154 
1155 		/*
1156 		 * See if there is more than one shared lock held.  If
1157 		 * so, just drop one and return.
1158 		 */
1159 		if (SX_SHARERS(*xp) > 1) {
1160 			if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1161 			    *xp - SX_ONE_SHARER)) {
1162 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1163 					CTR4(KTR_LOCK,
1164 					    "%s: %p succeeded %p -> %p",
1165 					    __func__, sx, (void *)*xp,
1166 					    (void *)(*xp - SX_ONE_SHARER));
1167 				return (true);
1168 			}
1169 			continue;
1170 		}
1171 
1172 		/*
1173 		 * If there aren't any waiters for an exclusive lock,
1174 		 * then try to drop it quickly.
1175 		 */
1176 		if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1177 			MPASS(*xp == SX_SHARERS_LOCK(1));
1178 			*xp = SX_SHARERS_LOCK(1);
1179 			if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1180 			    xp, SX_LOCK_UNLOCKED)) {
1181 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1182 					CTR2(KTR_LOCK, "%s: %p last succeeded",
1183 					    __func__, sx);
1184 				return (true);
1185 			}
1186 			continue;
1187 		}
1188 		break;
1189 	}
1190 	return (false);
1191 }
1192 
1193 static void __noinline
1194 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1195 {
1196 	int wakeup_swapper = 0;
1197 	uintptr_t setx;
1198 
1199 	if (SCHEDULER_STOPPED())
1200 		return;
1201 
1202 	if (_sx_sunlock_try(sx, &x))
1203 		goto out_lockstat;
1204 
1205 	/*
1206 	 * At this point, there should just be one sharer with
1207 	 * exclusive waiters.
1208 	 */
1209 	MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1210 
1211 	sleepq_lock(&sx->lock_object);
1212 	x = SX_READ_VALUE(sx);
1213 	for (;;) {
1214 		MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1215 		MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1216 		if (_sx_sunlock_try(sx, &x))
1217 			break;
1218 
1219 		/*
1220 		 * Wake up semantic here is quite simple:
1221 		 * Just wake up all the exclusive waiters.
1222 		 * Note that the state of the lock could have changed,
1223 		 * so if it fails loop back and retry.
1224 		 */
1225 		setx = x - SX_ONE_SHARER;
1226 		setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1227 		if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1228 			continue;
1229 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1230 			CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1231 			    "exclusive queue", __func__, sx);
1232 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1233 		    0, SQ_EXCLUSIVE_QUEUE);
1234 		break;
1235 	}
1236 	sleepq_release(&sx->lock_object);
1237 	if (wakeup_swapper)
1238 		kick_proc0();
1239 out_lockstat:
1240 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1241 }
1242 
1243 void
1244 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1245 {
1246 	uintptr_t x;
1247 
1248 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1249 	    ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1250 	_sx_assert(sx, SA_SLOCKED, file, line);
1251 	WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1252 	LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1253 
1254 	x = SX_READ_VALUE(sx);
1255 	if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1256 	    !_sx_sunlock_try(sx, &x)))
1257 		_sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1258 
1259 	TD_LOCKS_DEC(curthread);
1260 }
1261 
1262 void
1263 _sx_sunlock(struct sx *sx, const char *file, int line)
1264 {
1265 
1266 	_sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1267 }
1268 
1269 #ifdef INVARIANT_SUPPORT
1270 #ifndef INVARIANTS
1271 #undef	_sx_assert
1272 #endif
1273 
1274 /*
1275  * In the non-WITNESS case, sx_assert() can only detect that at least
1276  * *some* thread owns an slock, but it cannot guarantee that *this*
1277  * thread owns an slock.
1278  */
1279 void
1280 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1281 {
1282 #ifndef WITNESS
1283 	int slocked = 0;
1284 #endif
1285 
1286 	if (panicstr != NULL)
1287 		return;
1288 	switch (what) {
1289 	case SA_SLOCKED:
1290 	case SA_SLOCKED | SA_NOTRECURSED:
1291 	case SA_SLOCKED | SA_RECURSED:
1292 #ifndef WITNESS
1293 		slocked = 1;
1294 		/* FALLTHROUGH */
1295 #endif
1296 	case SA_LOCKED:
1297 	case SA_LOCKED | SA_NOTRECURSED:
1298 	case SA_LOCKED | SA_RECURSED:
1299 #ifdef WITNESS
1300 		witness_assert(&sx->lock_object, what, file, line);
1301 #else
1302 		/*
1303 		 * If some other thread has an exclusive lock or we
1304 		 * have one and are asserting a shared lock, fail.
1305 		 * Also, if no one has a lock at all, fail.
1306 		 */
1307 		if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1308 		    (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1309 		    sx_xholder(sx) != curthread)))
1310 			panic("Lock %s not %slocked @ %s:%d\n",
1311 			    sx->lock_object.lo_name, slocked ? "share " : "",
1312 			    file, line);
1313 
1314 		if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1315 			if (sx_recursed(sx)) {
1316 				if (what & SA_NOTRECURSED)
1317 					panic("Lock %s recursed @ %s:%d\n",
1318 					    sx->lock_object.lo_name, file,
1319 					    line);
1320 			} else if (what & SA_RECURSED)
1321 				panic("Lock %s not recursed @ %s:%d\n",
1322 				    sx->lock_object.lo_name, file, line);
1323 		}
1324 #endif
1325 		break;
1326 	case SA_XLOCKED:
1327 	case SA_XLOCKED | SA_NOTRECURSED:
1328 	case SA_XLOCKED | SA_RECURSED:
1329 		if (sx_xholder(sx) != curthread)
1330 			panic("Lock %s not exclusively locked @ %s:%d\n",
1331 			    sx->lock_object.lo_name, file, line);
1332 		if (sx_recursed(sx)) {
1333 			if (what & SA_NOTRECURSED)
1334 				panic("Lock %s recursed @ %s:%d\n",
1335 				    sx->lock_object.lo_name, file, line);
1336 		} else if (what & SA_RECURSED)
1337 			panic("Lock %s not recursed @ %s:%d\n",
1338 			    sx->lock_object.lo_name, file, line);
1339 		break;
1340 	case SA_UNLOCKED:
1341 #ifdef WITNESS
1342 		witness_assert(&sx->lock_object, what, file, line);
1343 #else
1344 		/*
1345 		 * If we hold an exclusve lock fail.  We can't
1346 		 * reliably check to see if we hold a shared lock or
1347 		 * not.
1348 		 */
1349 		if (sx_xholder(sx) == curthread)
1350 			panic("Lock %s exclusively locked @ %s:%d\n",
1351 			    sx->lock_object.lo_name, file, line);
1352 #endif
1353 		break;
1354 	default:
1355 		panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1356 		    line);
1357 	}
1358 }
1359 #endif	/* INVARIANT_SUPPORT */
1360 
1361 #ifdef DDB
1362 static void
1363 db_show_sx(const struct lock_object *lock)
1364 {
1365 	struct thread *td;
1366 	const struct sx *sx;
1367 
1368 	sx = (const struct sx *)lock;
1369 
1370 	db_printf(" state: ");
1371 	if (sx->sx_lock == SX_LOCK_UNLOCKED)
1372 		db_printf("UNLOCKED\n");
1373 	else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1374 		db_printf("DESTROYED\n");
1375 		return;
1376 	} else if (sx->sx_lock & SX_LOCK_SHARED)
1377 		db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1378 	else {
1379 		td = sx_xholder(sx);
1380 		db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1381 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1382 		if (sx_recursed(sx))
1383 			db_printf(" recursed: %d\n", sx->sx_recurse);
1384 	}
1385 
1386 	db_printf(" waiters: ");
1387 	switch(sx->sx_lock &
1388 	    (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1389 	case SX_LOCK_SHARED_WAITERS:
1390 		db_printf("shared\n");
1391 		break;
1392 	case SX_LOCK_EXCLUSIVE_WAITERS:
1393 		db_printf("exclusive\n");
1394 		break;
1395 	case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1396 		db_printf("exclusive and shared\n");
1397 		break;
1398 	default:
1399 		db_printf("none\n");
1400 	}
1401 }
1402 
1403 /*
1404  * Check to see if a thread that is blocked on a sleep queue is actually
1405  * blocked on an sx lock.  If so, output some details and return true.
1406  * If the lock has an exclusive owner, return that in *ownerp.
1407  */
1408 int
1409 sx_chain(struct thread *td, struct thread **ownerp)
1410 {
1411 	struct sx *sx;
1412 
1413 	/*
1414 	 * Check to see if this thread is blocked on an sx lock.
1415 	 * First, we check the lock class.  If that is ok, then we
1416 	 * compare the lock name against the wait message.
1417 	 */
1418 	sx = td->td_wchan;
1419 	if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1420 	    sx->lock_object.lo_name != td->td_wmesg)
1421 		return (0);
1422 
1423 	/* We think we have an sx lock, so output some details. */
1424 	db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1425 	*ownerp = sx_xholder(sx);
1426 	if (sx->sx_lock & SX_LOCK_SHARED)
1427 		db_printf("SLOCK (count %ju)\n",
1428 		    (uintmax_t)SX_SHARERS(sx->sx_lock));
1429 	else
1430 		db_printf("XLOCK\n");
1431 	return (1);
1432 }
1433 #endif
1434