xref: /freebsd/sys/kern/kern_sx.c (revision 123af6ec70016f5556da5972d4d63c7d175c06d3)
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 #ifdef HWPMC_HOOKS
75 #include <sys/pmckern.h>
76 PMC_SOFT_DECLARE( , , lock, failed);
77 #endif
78 
79 /* Handy macros for sleep queues. */
80 #define	SQ_EXCLUSIVE_QUEUE	0
81 #define	SQ_SHARED_QUEUE		1
82 
83 /*
84  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
85  * drop Giant anytime we have to sleep or if we adaptively spin.
86  */
87 #define	GIANT_DECLARE							\
88 	int _giantcnt = 0;						\
89 	WITNESS_SAVE_DECL(Giant)					\
90 
91 #define	GIANT_SAVE(work) do {						\
92 	if (__predict_false(mtx_owned(&Giant))) {			\
93 		work++;							\
94 		WITNESS_SAVE(&Giant.lock_object, Giant);		\
95 		while (mtx_owned(&Giant)) {				\
96 			_giantcnt++;					\
97 			mtx_unlock(&Giant);				\
98 		}							\
99 	}								\
100 } while (0)
101 
102 #define GIANT_RESTORE() do {						\
103 	if (_giantcnt > 0) {						\
104 		mtx_assert(&Giant, MA_NOTOWNED);			\
105 		while (_giantcnt--)					\
106 			mtx_lock(&Giant);				\
107 		WITNESS_RESTORE(&Giant.lock_object, Giant);		\
108 	}								\
109 } while (0)
110 
111 /*
112  * Returns true if an exclusive lock is recursed.  It assumes
113  * curthread currently has an exclusive lock.
114  */
115 #define	sx_recursed(sx)		((sx)->sx_recurse != 0)
116 
117 static void	assert_sx(const struct lock_object *lock, int what);
118 #ifdef DDB
119 static void	db_show_sx(const struct lock_object *lock);
120 #endif
121 static void	lock_sx(struct lock_object *lock, uintptr_t how);
122 #ifdef KDTRACE_HOOKS
123 static int	owner_sx(const struct lock_object *lock, struct thread **owner);
124 #endif
125 static uintptr_t unlock_sx(struct lock_object *lock);
126 
127 struct lock_class lock_class_sx = {
128 	.lc_name = "sx",
129 	.lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
130 	.lc_assert = assert_sx,
131 #ifdef DDB
132 	.lc_ddb_show = db_show_sx,
133 #endif
134 	.lc_lock = lock_sx,
135 	.lc_unlock = unlock_sx,
136 #ifdef KDTRACE_HOOKS
137 	.lc_owner = owner_sx,
138 #endif
139 };
140 
141 #ifndef INVARIANTS
142 #define	_sx_assert(sx, what, file, line)
143 #endif
144 
145 #ifdef ADAPTIVE_SX
146 static __read_frequently u_int asx_retries;
147 static __read_frequently u_int asx_loops;
148 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
149 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
150 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
151 
152 static struct lock_delay_config __read_frequently sx_delay;
153 
154 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
155     0, "");
156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
157     0, "");
158 
159 static void
160 sx_lock_delay_init(void *arg __unused)
161 {
162 
163 	lock_delay_default_init(&sx_delay);
164 	asx_retries = 10;
165 	asx_loops = max(10000, sx_delay.max);
166 }
167 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
168 #endif
169 
170 void
171 assert_sx(const struct lock_object *lock, int what)
172 {
173 
174 	sx_assert((const struct sx *)lock, what);
175 }
176 
177 void
178 lock_sx(struct lock_object *lock, uintptr_t how)
179 {
180 	struct sx *sx;
181 
182 	sx = (struct sx *)lock;
183 	if (how)
184 		sx_slock(sx);
185 	else
186 		sx_xlock(sx);
187 }
188 
189 uintptr_t
190 unlock_sx(struct lock_object *lock)
191 {
192 	struct sx *sx;
193 
194 	sx = (struct sx *)lock;
195 	sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
196 	if (sx_xlocked(sx)) {
197 		sx_xunlock(sx);
198 		return (0);
199 	} else {
200 		sx_sunlock(sx);
201 		return (1);
202 	}
203 }
204 
205 #ifdef KDTRACE_HOOKS
206 int
207 owner_sx(const struct lock_object *lock, struct thread **owner)
208 {
209 	const struct sx *sx;
210 	uintptr_t x;
211 
212 	sx = (const struct sx *)lock;
213 	x = sx->sx_lock;
214 	*owner = NULL;
215 	return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
216 	    ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
217 }
218 #endif
219 
220 void
221 sx_sysinit(void *arg)
222 {
223 	struct sx_args *sargs = arg;
224 
225 	sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
226 }
227 
228 void
229 sx_init_flags(struct sx *sx, const char *description, int opts)
230 {
231 	int flags;
232 
233 	MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
234 	    SX_NOPROFILE | SX_NEW)) == 0);
235 	ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
236 	    ("%s: sx_lock not aligned for %s: %p", __func__, description,
237 	    &sx->sx_lock));
238 
239 	flags = LO_SLEEPABLE | LO_UPGRADABLE;
240 	if (opts & SX_DUPOK)
241 		flags |= LO_DUPOK;
242 	if (opts & SX_NOPROFILE)
243 		flags |= LO_NOPROFILE;
244 	if (!(opts & SX_NOWITNESS))
245 		flags |= LO_WITNESS;
246 	if (opts & SX_RECURSE)
247 		flags |= LO_RECURSABLE;
248 	if (opts & SX_QUIET)
249 		flags |= LO_QUIET;
250 	if (opts & SX_NEW)
251 		flags |= LO_NEW;
252 
253 	lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
254 	sx->sx_lock = SX_LOCK_UNLOCKED;
255 	sx->sx_recurse = 0;
256 }
257 
258 void
259 sx_destroy(struct sx *sx)
260 {
261 
262 	KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
263 	KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
264 	sx->sx_lock = SX_LOCK_DESTROYED;
265 	lock_destroy(&sx->lock_object);
266 }
267 
268 int
269 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
270 {
271 	uintptr_t x;
272 
273 	if (SCHEDULER_STOPPED())
274 		return (1);
275 
276 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
277 	    ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
278 	    curthread, sx->lock_object.lo_name, file, line));
279 
280 	x = sx->sx_lock;
281 	for (;;) {
282 		KASSERT(x != SX_LOCK_DESTROYED,
283 		    ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
284 		if (!(x & SX_LOCK_SHARED))
285 			break;
286 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
287 			LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
288 			WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
289 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
290 			    sx, 0, 0, file, line, LOCKSTAT_READER);
291 			TD_LOCKS_INC(curthread);
292 			curthread->td_sx_slocks++;
293 			return (1);
294 		}
295 	}
296 
297 	LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
298 	return (0);
299 }
300 
301 int
302 sx_try_slock_(struct sx *sx, const char *file, int line)
303 {
304 
305 	return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
306 }
307 
308 int
309 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
310 {
311 	uintptr_t tid, x;
312 	int error = 0;
313 
314 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
315 	    !TD_IS_IDLETHREAD(curthread),
316 	    ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
317 	    curthread, sx->lock_object.lo_name, file, line));
318 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
319 	    ("sx_xlock() of destroyed sx @ %s:%d", file, line));
320 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
321 	    line, NULL);
322 	tid = (uintptr_t)curthread;
323 	x = SX_LOCK_UNLOCKED;
324 	if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
325 		error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
326 	else
327 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
328 		    0, 0, file, line, LOCKSTAT_WRITER);
329 	if (!error) {
330 		LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
331 		    file, line);
332 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
333 		TD_LOCKS_INC(curthread);
334 	}
335 
336 	return (error);
337 }
338 
339 int
340 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
341 {
342 	struct thread *td;
343 	uintptr_t tid, x;
344 	int rval;
345 	bool recursed;
346 
347 	td = curthread;
348 	tid = (uintptr_t)td;
349 	if (SCHEDULER_STOPPED_TD(td))
350 		return (1);
351 
352 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
353 	    ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
354 	    curthread, sx->lock_object.lo_name, file, line));
355 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
356 	    ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
357 
358 	rval = 1;
359 	recursed = false;
360 	x = SX_LOCK_UNLOCKED;
361 	for (;;) {
362 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
363 			break;
364 		if (x == SX_LOCK_UNLOCKED)
365 			continue;
366 		if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
367 			sx->sx_recurse++;
368 			atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
369 			break;
370 		}
371 		rval = 0;
372 		break;
373 	}
374 
375 	LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
376 	if (rval) {
377 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
378 		    file, line);
379 		if (!recursed)
380 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
381 			    sx, 0, 0, file, line, LOCKSTAT_WRITER);
382 		TD_LOCKS_INC(curthread);
383 	}
384 
385 	return (rval);
386 }
387 
388 int
389 sx_try_xlock_(struct sx *sx, const char *file, int line)
390 {
391 
392 	return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
393 }
394 
395 void
396 _sx_xunlock(struct sx *sx, const char *file, int line)
397 {
398 
399 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
400 	    ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
401 	_sx_assert(sx, SA_XLOCKED, file, line);
402 	WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
403 	LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
404 	    line);
405 #if LOCK_DEBUG > 0
406 	_sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
407 #else
408 	__sx_xunlock(sx, curthread, file, line);
409 #endif
410 	TD_LOCKS_DEC(curthread);
411 }
412 
413 /*
414  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
415  * This will only succeed if this thread holds a single shared lock.
416  * Return 1 if if the upgrade succeed, 0 otherwise.
417  */
418 int
419 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
420 {
421 	uintptr_t x;
422 	uintptr_t waiters;
423 	int success;
424 
425 	if (SCHEDULER_STOPPED())
426 		return (1);
427 
428 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
429 	    ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
430 	_sx_assert(sx, SA_SLOCKED, file, line);
431 
432 	/*
433 	 * Try to switch from one shared lock to an exclusive lock.  We need
434 	 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
435 	 * we will wake up the exclusive waiters when we drop the lock.
436 	 */
437 	success = 0;
438 	x = SX_READ_VALUE(sx);
439 	for (;;) {
440 		if (SX_SHARERS(x) > 1)
441 			break;
442 		waiters = (x & SX_LOCK_WAITERS);
443 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
444 		    (uintptr_t)curthread | waiters)) {
445 			success = 1;
446 			break;
447 		}
448 	}
449 	LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
450 	if (success) {
451 		curthread->td_sx_slocks--;
452 		WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
453 		    file, line);
454 		LOCKSTAT_RECORD0(sx__upgrade, sx);
455 	}
456 	return (success);
457 }
458 
459 int
460 sx_try_upgrade_(struct sx *sx, const char *file, int line)
461 {
462 
463 	return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
464 }
465 
466 /*
467  * Downgrade an unrecursed exclusive lock into a single shared lock.
468  */
469 void
470 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
471 {
472 	uintptr_t x;
473 	int wakeup_swapper;
474 
475 	if (SCHEDULER_STOPPED())
476 		return;
477 
478 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
479 	    ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
480 	_sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
481 #ifndef INVARIANTS
482 	if (sx_recursed(sx))
483 		panic("downgrade of a recursed lock");
484 #endif
485 
486 	WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
487 
488 	/*
489 	 * Try to switch from an exclusive lock with no shared waiters
490 	 * to one sharer with no shared waiters.  If there are
491 	 * exclusive waiters, we don't need to lock the sleep queue so
492 	 * long as we preserve the flag.  We do one quick try and if
493 	 * that fails we grab the sleepq lock to keep the flags from
494 	 * changing and do it the slow way.
495 	 *
496 	 * We have to lock the sleep queue if there are shared waiters
497 	 * so we can wake them up.
498 	 */
499 	x = sx->sx_lock;
500 	if (!(x & SX_LOCK_SHARED_WAITERS) &&
501 	    atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
502 	    (x & SX_LOCK_EXCLUSIVE_WAITERS)))
503 		goto out;
504 
505 	/*
506 	 * Lock the sleep queue so we can read the waiters bits
507 	 * without any races and wakeup any shared waiters.
508 	 */
509 	sleepq_lock(&sx->lock_object);
510 
511 	/*
512 	 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
513 	 * shared lock.  If there are any shared waiters, wake them up.
514 	 */
515 	wakeup_swapper = 0;
516 	x = sx->sx_lock;
517 	atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
518 	    (x & SX_LOCK_EXCLUSIVE_WAITERS));
519 	if (x & SX_LOCK_SHARED_WAITERS)
520 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
521 		    0, SQ_SHARED_QUEUE);
522 	sleepq_release(&sx->lock_object);
523 
524 	if (wakeup_swapper)
525 		kick_proc0();
526 
527 out:
528 	curthread->td_sx_slocks++;
529 	LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
530 	LOCKSTAT_RECORD0(sx__downgrade, sx);
531 }
532 
533 void
534 sx_downgrade_(struct sx *sx, const char *file, int line)
535 {
536 
537 	sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
538 }
539 
540 #ifdef	ADAPTIVE_SX
541 static inline void
542 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
543 {
544 
545 	if (x & SX_LOCK_WRITE_SPINNER)
546 		return;
547 	if (*in_critical) {
548 		critical_exit();
549 		*in_critical = false;
550 		(*extra_work)--;
551 	}
552 }
553 #else
554 #define sx_drop_critical(x, in_critical, extra_work) do { } while(0)
555 #endif
556 
557 /*
558  * This function represents the so-called 'hard case' for sx_xlock
559  * operation.  All 'easy case' failures are redirected to this.  Note
560  * that ideally this would be a static function, but it needs to be
561  * accessible from at least sx.h.
562  */
563 int
564 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
565 {
566 	GIANT_DECLARE;
567 	uintptr_t tid, setx;
568 #ifdef ADAPTIVE_SX
569 	volatile struct thread *owner;
570 	u_int i, n, spintries = 0;
571 	enum { READERS, WRITER } sleep_reason = READERS;
572 	bool in_critical = false;
573 #endif
574 #ifdef LOCK_PROFILING
575 	uint64_t waittime = 0;
576 	int contested = 0;
577 #endif
578 	int error = 0;
579 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
580 	struct lock_delay_arg lda;
581 #endif
582 #ifdef	KDTRACE_HOOKS
583 	u_int sleep_cnt = 0;
584 	int64_t sleep_time = 0;
585 	int64_t all_time = 0;
586 #endif
587 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
588 	uintptr_t state = 0;
589 	int doing_lockprof = 0;
590 #endif
591 	int extra_work = 0;
592 
593 	tid = (uintptr_t)curthread;
594 
595 #ifdef KDTRACE_HOOKS
596 	if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
597 		while (x == SX_LOCK_UNLOCKED) {
598 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
599 				goto out_lockstat;
600 		}
601 		extra_work = 1;
602 		doing_lockprof = 1;
603 		all_time -= lockstat_nsecs(&sx->lock_object);
604 		state = x;
605 	}
606 #endif
607 #ifdef LOCK_PROFILING
608 	extra_work = 1;
609 	doing_lockprof = 1;
610 	state = x;
611 #endif
612 
613 	if (SCHEDULER_STOPPED())
614 		return (0);
615 
616 #if defined(ADAPTIVE_SX)
617 	lock_delay_arg_init(&lda, &sx_delay);
618 #elif defined(KDTRACE_HOOKS)
619 	lock_delay_arg_init(&lda, NULL);
620 #endif
621 
622 	if (__predict_false(x == SX_LOCK_UNLOCKED))
623 		x = SX_READ_VALUE(sx);
624 
625 	/* If we already hold an exclusive lock, then recurse. */
626 	if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
627 		KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
628 	    ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
629 		    sx->lock_object.lo_name, file, line));
630 		sx->sx_recurse++;
631 		atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
632 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
633 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
634 		return (0);
635 	}
636 
637 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
638 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
639 		    sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
640 
641 #ifdef HWPMC_HOOKS
642 	PMC_SOFT_CALL( , , lock, failed);
643 #endif
644 	lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
645 	    &waittime);
646 
647 #ifndef INVARIANTS
648 	GIANT_SAVE(extra_work);
649 #endif
650 
651 	for (;;) {
652 		if (x == SX_LOCK_UNLOCKED) {
653 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
654 				break;
655 			continue;
656 		}
657 #ifdef INVARIANTS
658 		GIANT_SAVE(extra_work);
659 #endif
660 #ifdef KDTRACE_HOOKS
661 		lda.spin_cnt++;
662 #endif
663 #ifdef ADAPTIVE_SX
664 		/*
665 		 * If the lock is write locked and the owner is
666 		 * running on another CPU, spin until the owner stops
667 		 * running or the state of the lock changes.
668 		 */
669 		if ((x & SX_LOCK_SHARED) == 0) {
670 			sx_drop_critical(x, &in_critical, &extra_work);
671 			sleep_reason = WRITER;
672 			owner = lv_sx_owner(x);
673 			if (!TD_IS_RUNNING(owner))
674 				goto sleepq;
675 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
676 				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
677 				    __func__, sx, owner);
678 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
679 			    "spinning", "lockname:\"%s\"",
680 			    sx->lock_object.lo_name);
681 			do {
682 				lock_delay(&lda);
683 				x = SX_READ_VALUE(sx);
684 				owner = lv_sx_owner(x);
685 			} while (owner != NULL && TD_IS_RUNNING(owner));
686 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
687 			    "running");
688 			continue;
689 		} else if (SX_SHARERS(x) > 0) {
690 			sleep_reason = READERS;
691 			if (spintries == asx_retries)
692 				goto sleepq;
693 			if (!(x & SX_LOCK_WRITE_SPINNER)) {
694 				if (!in_critical) {
695 					critical_enter();
696 					in_critical = true;
697 					extra_work++;
698 				}
699 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
700 				    x | SX_LOCK_WRITE_SPINNER)) {
701 					critical_exit();
702 					in_critical = false;
703 					extra_work--;
704 					continue;
705 				}
706 			}
707 			spintries++;
708 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
709 			    "spinning", "lockname:\"%s\"",
710 			    sx->lock_object.lo_name);
711 			n = SX_SHARERS(x);
712 			for (i = 0; i < asx_loops; i += n) {
713 				lock_delay_spin(n);
714 				x = SX_READ_VALUE(sx);
715 				if (!(x & SX_LOCK_WRITE_SPINNER))
716 					break;
717 				if (!(x & SX_LOCK_SHARED))
718 					break;
719 				n = SX_SHARERS(x);
720 				if (n == 0)
721 					break;
722 			}
723 #ifdef KDTRACE_HOOKS
724 			lda.spin_cnt += i;
725 #endif
726 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
727 			    "running");
728 			if (i < asx_loops)
729 				continue;
730 		}
731 sleepq:
732 #endif
733 		sleepq_lock(&sx->lock_object);
734 		x = SX_READ_VALUE(sx);
735 retry_sleepq:
736 
737 		/*
738 		 * If the lock was released while spinning on the
739 		 * sleep queue chain lock, try again.
740 		 */
741 		if (x == SX_LOCK_UNLOCKED) {
742 			sleepq_release(&sx->lock_object);
743 			sx_drop_critical(x, &in_critical, &extra_work);
744 			continue;
745 		}
746 
747 #ifdef ADAPTIVE_SX
748 		/*
749 		 * The current lock owner might have started executing
750 		 * on another CPU (or the lock could have changed
751 		 * owners) while we were waiting on the sleep queue
752 		 * chain lock.  If so, drop the sleep queue lock and try
753 		 * again.
754 		 */
755 		if (!(x & SX_LOCK_SHARED)) {
756 			owner = (struct thread *)SX_OWNER(x);
757 			if (TD_IS_RUNNING(owner)) {
758 				sleepq_release(&sx->lock_object);
759 				sx_drop_critical(x, &in_critical,
760 				    &extra_work);
761 				continue;
762 			}
763 		} else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
764 			sleepq_release(&sx->lock_object);
765 			sx_drop_critical(x, &in_critical, &extra_work);
766 			continue;
767 		}
768 #endif
769 
770 		/*
771 		 * If an exclusive lock was released with both shared
772 		 * and exclusive waiters and a shared waiter hasn't
773 		 * woken up and acquired the lock yet, sx_lock will be
774 		 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
775 		 * If we see that value, try to acquire it once.  Note
776 		 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
777 		 * as there are other exclusive waiters still.  If we
778 		 * fail, restart the loop.
779 		 */
780 		setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
781 		if ((x & ~setx) == SX_LOCK_SHARED) {
782 			setx &= ~SX_LOCK_WRITE_SPINNER;
783 			if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
784 				goto retry_sleepq;
785 			sleepq_release(&sx->lock_object);
786 			CTR2(KTR_LOCK, "%s: %p claimed by new writer",
787 			    __func__, sx);
788 			break;
789 		}
790 
791 #ifdef ADAPTIVE_SX
792 		/*
793 		 * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
794 		 * It is an invariant that when the bit is set, there is
795 		 * a writer ready to grab the lock. Thus clear the bit since
796 		 * we are going to sleep.
797 		 */
798 		if (in_critical) {
799 			if ((x & SX_LOCK_WRITE_SPINNER) ||
800 			    !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
801 				setx = x & ~SX_LOCK_WRITE_SPINNER;
802 				setx |= SX_LOCK_EXCLUSIVE_WAITERS;
803 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
804 				    setx)) {
805 					goto retry_sleepq;
806 				}
807 			}
808 			critical_exit();
809 			in_critical = false;
810 		} else {
811 #endif
812 			/*
813 			 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
814 			 * than loop back and retry.
815 			 */
816 			if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
817 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
818 				    x | SX_LOCK_EXCLUSIVE_WAITERS)) {
819 					goto retry_sleepq;
820 				}
821 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
822 					CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
823 					    __func__, sx);
824 			}
825 #ifdef ADAPTIVE_SX
826 		}
827 #endif
828 
829 		/*
830 		 * Since we have been unable to acquire the exclusive
831 		 * lock and the exclusive waiters flag is set, we have
832 		 * to sleep.
833 		 */
834 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
835 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
836 			    __func__, sx);
837 
838 #ifdef KDTRACE_HOOKS
839 		sleep_time -= lockstat_nsecs(&sx->lock_object);
840 #endif
841 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
842 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
843 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
844 		if (!(opts & SX_INTERRUPTIBLE))
845 			sleepq_wait(&sx->lock_object, 0);
846 		else
847 			error = sleepq_wait_sig(&sx->lock_object, 0);
848 #ifdef KDTRACE_HOOKS
849 		sleep_time += lockstat_nsecs(&sx->lock_object);
850 		sleep_cnt++;
851 #endif
852 		if (error) {
853 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
854 				CTR2(KTR_LOCK,
855 			"%s: interruptible sleep by %p suspended by signal",
856 				    __func__, sx);
857 			break;
858 		}
859 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
860 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
861 			    __func__, sx);
862 		x = SX_READ_VALUE(sx);
863 	}
864 	if (__predict_true(!extra_work))
865 		return (error);
866 #ifdef ADAPTIVE_SX
867 	if (in_critical)
868 		critical_exit();
869 #endif
870 	GIANT_RESTORE();
871 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
872 	if (__predict_true(!doing_lockprof))
873 		return (error);
874 #endif
875 #ifdef KDTRACE_HOOKS
876 	all_time += lockstat_nsecs(&sx->lock_object);
877 	if (sleep_time)
878 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
879 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
880 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
881 	if (lda.spin_cnt > sleep_cnt)
882 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
883 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
884 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
885 out_lockstat:
886 #endif
887 	if (!error)
888 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
889 		    contested, waittime, file, line, LOCKSTAT_WRITER);
890 	return (error);
891 }
892 
893 /*
894  * This function represents the so-called 'hard case' for sx_xunlock
895  * operation.  All 'easy case' failures are redirected to this.  Note
896  * that ideally this would be a static function, but it needs to be
897  * accessible from at least sx.h.
898  */
899 void
900 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
901 {
902 	uintptr_t tid, setx;
903 	int queue, wakeup_swapper;
904 
905 	if (SCHEDULER_STOPPED())
906 		return;
907 
908 	tid = (uintptr_t)curthread;
909 
910 	if (__predict_false(x == tid))
911 		x = SX_READ_VALUE(sx);
912 
913 	MPASS(!(x & SX_LOCK_SHARED));
914 
915 	if (__predict_false(x & SX_LOCK_RECURSED)) {
916 		/* The lock is recursed, unrecurse one level. */
917 		if ((--sx->sx_recurse) == 0)
918 			atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
919 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
920 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
921 		return;
922 	}
923 
924 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
925 	if (x == tid &&
926 	    atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
927 		return;
928 
929 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
930 		CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
931 
932 	sleepq_lock(&sx->lock_object);
933 	x = SX_READ_VALUE(sx);
934 	MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
935 
936 	/*
937 	 * The wake up algorithm here is quite simple and probably not
938 	 * ideal.  It gives precedence to shared waiters if they are
939 	 * present.  For this condition, we have to preserve the
940 	 * state of the exclusive waiters flag.
941 	 * If interruptible sleeps left the shared queue empty avoid a
942 	 * starvation for the threads sleeping on the exclusive queue by giving
943 	 * them precedence and cleaning up the shared waiters bit anyway.
944 	 */
945 	setx = SX_LOCK_UNLOCKED;
946 	queue = SQ_SHARED_QUEUE;
947 	if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
948 	    sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
949 		queue = SQ_EXCLUSIVE_QUEUE;
950 		setx |= (x & SX_LOCK_SHARED_WAITERS);
951 	}
952 	atomic_store_rel_ptr(&sx->sx_lock, setx);
953 
954 	/* Wake up all the waiters for the specific queue. */
955 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
956 		CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
957 		    __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
958 		    "exclusive");
959 
960 	wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
961 	    queue);
962 	sleepq_release(&sx->lock_object);
963 	if (wakeup_swapper)
964 		kick_proc0();
965 }
966 
967 static bool __always_inline
968 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
969 {
970 
971 	if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
972 			== SX_LOCK_SHARED)
973 		return (true);
974 	if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
975 		return (true);
976 	return (false);
977 }
978 
979 static bool __always_inline
980 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
981     LOCK_FILE_LINE_ARG_DEF)
982 {
983 
984 	/*
985 	 * If no other thread has an exclusive lock then try to bump up
986 	 * the count of sharers.  Since we have to preserve the state
987 	 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
988 	 * shared lock loop back and retry.
989 	 */
990 	while (__sx_can_read(td, *xp, fp)) {
991 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
992 		    *xp + SX_ONE_SHARER)) {
993 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
994 				CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
995 				    __func__, sx, (void *)*xp,
996 				    (void *)(*xp + SX_ONE_SHARER));
997 			td->td_sx_slocks++;
998 			return (true);
999 		}
1000 	}
1001 	return (false);
1002 }
1003 
1004 static int __noinline
1005 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1006 {
1007 	GIANT_DECLARE;
1008 	struct thread *td;
1009 #ifdef ADAPTIVE_SX
1010 	volatile struct thread *owner;
1011 	u_int i, n, spintries = 0;
1012 #endif
1013 #ifdef LOCK_PROFILING
1014 	uint64_t waittime = 0;
1015 	int contested = 0;
1016 #endif
1017 	int error = 0;
1018 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1019 	struct lock_delay_arg lda;
1020 #endif
1021 #ifdef KDTRACE_HOOKS
1022 	u_int sleep_cnt = 0;
1023 	int64_t sleep_time = 0;
1024 	int64_t all_time = 0;
1025 #endif
1026 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1027 	uintptr_t state = 0;
1028 #endif
1029 	int extra_work = 0;
1030 
1031 	td = curthread;
1032 
1033 #ifdef KDTRACE_HOOKS
1034 	if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1035 		if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1036 			goto out_lockstat;
1037 		extra_work = 1;
1038 		all_time -= lockstat_nsecs(&sx->lock_object);
1039 		state = x;
1040 	}
1041 #endif
1042 #ifdef LOCK_PROFILING
1043 	extra_work = 1;
1044 	state = x;
1045 #endif
1046 
1047 	if (SCHEDULER_STOPPED())
1048 		return (0);
1049 
1050 #if defined(ADAPTIVE_SX)
1051 	lock_delay_arg_init(&lda, &sx_delay);
1052 #elif defined(KDTRACE_HOOKS)
1053 	lock_delay_arg_init(&lda, NULL);
1054 #endif
1055 
1056 #ifdef HWPMC_HOOKS
1057 	PMC_SOFT_CALL( , , lock, failed);
1058 #endif
1059 	lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
1060 	    &waittime);
1061 
1062 #ifndef INVARIANTS
1063 	GIANT_SAVE(extra_work);
1064 #endif
1065 
1066 	/*
1067 	 * As with rwlocks, we don't make any attempt to try to block
1068 	 * shared locks once there is an exclusive waiter.
1069 	 */
1070 	for (;;) {
1071 		if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1072 			break;
1073 #ifdef INVARIANTS
1074 		GIANT_SAVE(extra_work);
1075 #endif
1076 #ifdef KDTRACE_HOOKS
1077 		lda.spin_cnt++;
1078 #endif
1079 
1080 #ifdef ADAPTIVE_SX
1081 		/*
1082 		 * If the owner is running on another CPU, spin until
1083 		 * the owner stops running or the state of the lock
1084 		 * changes.
1085 		 */
1086 		if ((x & SX_LOCK_SHARED) == 0) {
1087 			owner = lv_sx_owner(x);
1088 			if (TD_IS_RUNNING(owner)) {
1089 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1090 					CTR3(KTR_LOCK,
1091 					    "%s: spinning on %p held by %p",
1092 					    __func__, sx, owner);
1093 				KTR_STATE1(KTR_SCHED, "thread",
1094 				    sched_tdname(curthread), "spinning",
1095 				    "lockname:\"%s\"", sx->lock_object.lo_name);
1096 				do {
1097 					lock_delay(&lda);
1098 					x = SX_READ_VALUE(sx);
1099 					owner = lv_sx_owner(x);
1100 				} while (owner != NULL && TD_IS_RUNNING(owner));
1101 				KTR_STATE0(KTR_SCHED, "thread",
1102 				    sched_tdname(curthread), "running");
1103 				continue;
1104 			}
1105 		} else {
1106 			if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1107 				MPASS(!__sx_can_read(td, x, false));
1108 				lock_delay_spin(2);
1109 				x = SX_READ_VALUE(sx);
1110 				continue;
1111 			}
1112 			if (spintries < asx_retries) {
1113 				KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1114 				    "spinning", "lockname:\"%s\"",
1115 				    sx->lock_object.lo_name);
1116 				n = SX_SHARERS(x);
1117 				for (i = 0; i < asx_loops; i += n) {
1118 					lock_delay_spin(n);
1119 					x = SX_READ_VALUE(sx);
1120 					if (!(x & SX_LOCK_SHARED))
1121 						break;
1122 					n = SX_SHARERS(x);
1123 					if (n == 0)
1124 						break;
1125 					if (__sx_can_read(td, x, false))
1126 						break;
1127 				}
1128 #ifdef KDTRACE_HOOKS
1129 				lda.spin_cnt += i;
1130 #endif
1131 				KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1132 				    "running");
1133 				if (i < asx_loops)
1134 					continue;
1135 			}
1136 		}
1137 #endif
1138 
1139 		/*
1140 		 * Some other thread already has an exclusive lock, so
1141 		 * start the process of blocking.
1142 		 */
1143 		sleepq_lock(&sx->lock_object);
1144 		x = SX_READ_VALUE(sx);
1145 retry_sleepq:
1146 		if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1147 		    __sx_can_read(td, x, false)) {
1148 			sleepq_release(&sx->lock_object);
1149 			continue;
1150 		}
1151 
1152 #ifdef ADAPTIVE_SX
1153 		/*
1154 		 * If the owner is running on another CPU, spin until
1155 		 * the owner stops running or the state of the lock
1156 		 * changes.
1157 		 */
1158 		if (!(x & SX_LOCK_SHARED)) {
1159 			owner = (struct thread *)SX_OWNER(x);
1160 			if (TD_IS_RUNNING(owner)) {
1161 				sleepq_release(&sx->lock_object);
1162 				x = SX_READ_VALUE(sx);
1163 				continue;
1164 			}
1165 		}
1166 #endif
1167 
1168 		/*
1169 		 * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1170 		 * fail to set it drop the sleep queue lock and loop
1171 		 * back.
1172 		 */
1173 		if (!(x & SX_LOCK_SHARED_WAITERS)) {
1174 			if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1175 			    x | SX_LOCK_SHARED_WAITERS))
1176 				goto retry_sleepq;
1177 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1178 				CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1179 				    __func__, sx);
1180 		}
1181 
1182 		/*
1183 		 * Since we have been unable to acquire the shared lock,
1184 		 * we have to sleep.
1185 		 */
1186 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1187 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1188 			    __func__, sx);
1189 
1190 #ifdef KDTRACE_HOOKS
1191 		sleep_time -= lockstat_nsecs(&sx->lock_object);
1192 #endif
1193 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1194 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1195 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1196 		if (!(opts & SX_INTERRUPTIBLE))
1197 			sleepq_wait(&sx->lock_object, 0);
1198 		else
1199 			error = sleepq_wait_sig(&sx->lock_object, 0);
1200 #ifdef KDTRACE_HOOKS
1201 		sleep_time += lockstat_nsecs(&sx->lock_object);
1202 		sleep_cnt++;
1203 #endif
1204 		if (error) {
1205 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1206 				CTR2(KTR_LOCK,
1207 			"%s: interruptible sleep by %p suspended by signal",
1208 				    __func__, sx);
1209 			break;
1210 		}
1211 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1212 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1213 			    __func__, sx);
1214 		x = SX_READ_VALUE(sx);
1215 	}
1216 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1217 	if (__predict_true(!extra_work))
1218 		return (error);
1219 #endif
1220 #ifdef KDTRACE_HOOKS
1221 	all_time += lockstat_nsecs(&sx->lock_object);
1222 	if (sleep_time)
1223 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1224 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1225 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1226 	if (lda.spin_cnt > sleep_cnt)
1227 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1228 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1229 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1230 out_lockstat:
1231 #endif
1232 	if (error == 0) {
1233 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1234 		    contested, waittime, file, line, LOCKSTAT_READER);
1235 	}
1236 	GIANT_RESTORE();
1237 	return (error);
1238 }
1239 
1240 int
1241 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1242 {
1243 	struct thread *td;
1244 	uintptr_t x;
1245 	int error;
1246 
1247 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1248 	    !TD_IS_IDLETHREAD(curthread),
1249 	    ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1250 	    curthread, sx->lock_object.lo_name, file, line));
1251 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1252 	    ("sx_slock() of destroyed sx @ %s:%d", file, line));
1253 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1254 
1255 	error = 0;
1256 	td = curthread;
1257 	x = SX_READ_VALUE(sx);
1258 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1259 	    !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1260 		error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1261 	else
1262 		lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1263 		    file, line);
1264 	if (error == 0) {
1265 		LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1266 		WITNESS_LOCK(&sx->lock_object, 0, file, line);
1267 		TD_LOCKS_INC(curthread);
1268 	}
1269 	return (error);
1270 }
1271 
1272 int
1273 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1274 {
1275 
1276 	return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1277 }
1278 
1279 static bool __always_inline
1280 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1281 {
1282 
1283 	for (;;) {
1284 		if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1285 			if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1286 			    *xp - SX_ONE_SHARER)) {
1287 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1288 					CTR4(KTR_LOCK,
1289 					    "%s: %p succeeded %p -> %p",
1290 					    __func__, sx, (void *)*xp,
1291 					    (void *)(*xp - SX_ONE_SHARER));
1292 				td->td_sx_slocks--;
1293 				return (true);
1294 			}
1295 			continue;
1296 		}
1297 		break;
1298 	}
1299 	return (false);
1300 }
1301 
1302 static void __noinline
1303 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1304     LOCK_FILE_LINE_ARG_DEF)
1305 {
1306 	int wakeup_swapper = 0;
1307 	uintptr_t setx, queue;
1308 
1309 	if (SCHEDULER_STOPPED())
1310 		return;
1311 
1312 	if (_sx_sunlock_try(sx, td, &x))
1313 		goto out_lockstat;
1314 
1315 	sleepq_lock(&sx->lock_object);
1316 	x = SX_READ_VALUE(sx);
1317 	for (;;) {
1318 		if (_sx_sunlock_try(sx, td, &x))
1319 			break;
1320 
1321 		/*
1322 		 * Wake up semantic here is quite simple:
1323 		 * Just wake up all the exclusive waiters.
1324 		 * Note that the state of the lock could have changed,
1325 		 * so if it fails loop back and retry.
1326 		 */
1327 		setx = SX_LOCK_UNLOCKED;
1328 		queue = SQ_SHARED_QUEUE;
1329 		if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1330 			setx |= (x & SX_LOCK_SHARED_WAITERS);
1331 			queue = SQ_EXCLUSIVE_QUEUE;
1332 		}
1333 		setx |= (x & SX_LOCK_WRITE_SPINNER);
1334 		if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1335 			continue;
1336 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1337 			CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1338 			    "exclusive queue", __func__, sx);
1339 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1340 		    0, queue);
1341 		td->td_sx_slocks--;
1342 		break;
1343 	}
1344 	sleepq_release(&sx->lock_object);
1345 	if (wakeup_swapper)
1346 		kick_proc0();
1347 out_lockstat:
1348 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1349 }
1350 
1351 void
1352 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1353 {
1354 	struct thread *td;
1355 	uintptr_t x;
1356 
1357 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1358 	    ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1359 	_sx_assert(sx, SA_SLOCKED, file, line);
1360 	WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1361 	LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1362 
1363 	td = curthread;
1364 	x = SX_READ_VALUE(sx);
1365 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1366 	    !_sx_sunlock_try(sx, td, &x)))
1367 		_sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1368 	else
1369 		lock_profile_release_lock(&sx->lock_object);
1370 
1371 	TD_LOCKS_DEC(curthread);
1372 }
1373 
1374 void
1375 _sx_sunlock(struct sx *sx, const char *file, int line)
1376 {
1377 
1378 	_sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1379 }
1380 
1381 #ifdef INVARIANT_SUPPORT
1382 #ifndef INVARIANTS
1383 #undef	_sx_assert
1384 #endif
1385 
1386 /*
1387  * In the non-WITNESS case, sx_assert() can only detect that at least
1388  * *some* thread owns an slock, but it cannot guarantee that *this*
1389  * thread owns an slock.
1390  */
1391 void
1392 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1393 {
1394 #ifndef WITNESS
1395 	int slocked = 0;
1396 #endif
1397 
1398 	if (SCHEDULER_STOPPED())
1399 		return;
1400 	switch (what) {
1401 	case SA_SLOCKED:
1402 	case SA_SLOCKED | SA_NOTRECURSED:
1403 	case SA_SLOCKED | SA_RECURSED:
1404 #ifndef WITNESS
1405 		slocked = 1;
1406 		/* FALLTHROUGH */
1407 #endif
1408 	case SA_LOCKED:
1409 	case SA_LOCKED | SA_NOTRECURSED:
1410 	case SA_LOCKED | SA_RECURSED:
1411 #ifdef WITNESS
1412 		witness_assert(&sx->lock_object, what, file, line);
1413 #else
1414 		/*
1415 		 * If some other thread has an exclusive lock or we
1416 		 * have one and are asserting a shared lock, fail.
1417 		 * Also, if no one has a lock at all, fail.
1418 		 */
1419 		if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1420 		    (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1421 		    sx_xholder(sx) != curthread)))
1422 			panic("Lock %s not %slocked @ %s:%d\n",
1423 			    sx->lock_object.lo_name, slocked ? "share " : "",
1424 			    file, line);
1425 
1426 		if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1427 			if (sx_recursed(sx)) {
1428 				if (what & SA_NOTRECURSED)
1429 					panic("Lock %s recursed @ %s:%d\n",
1430 					    sx->lock_object.lo_name, file,
1431 					    line);
1432 			} else if (what & SA_RECURSED)
1433 				panic("Lock %s not recursed @ %s:%d\n",
1434 				    sx->lock_object.lo_name, file, line);
1435 		}
1436 #endif
1437 		break;
1438 	case SA_XLOCKED:
1439 	case SA_XLOCKED | SA_NOTRECURSED:
1440 	case SA_XLOCKED | SA_RECURSED:
1441 		if (sx_xholder(sx) != curthread)
1442 			panic("Lock %s not exclusively locked @ %s:%d\n",
1443 			    sx->lock_object.lo_name, file, line);
1444 		if (sx_recursed(sx)) {
1445 			if (what & SA_NOTRECURSED)
1446 				panic("Lock %s recursed @ %s:%d\n",
1447 				    sx->lock_object.lo_name, file, line);
1448 		} else if (what & SA_RECURSED)
1449 			panic("Lock %s not recursed @ %s:%d\n",
1450 			    sx->lock_object.lo_name, file, line);
1451 		break;
1452 	case SA_UNLOCKED:
1453 #ifdef WITNESS
1454 		witness_assert(&sx->lock_object, what, file, line);
1455 #else
1456 		/*
1457 		 * If we hold an exclusve lock fail.  We can't
1458 		 * reliably check to see if we hold a shared lock or
1459 		 * not.
1460 		 */
1461 		if (sx_xholder(sx) == curthread)
1462 			panic("Lock %s exclusively locked @ %s:%d\n",
1463 			    sx->lock_object.lo_name, file, line);
1464 #endif
1465 		break;
1466 	default:
1467 		panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1468 		    line);
1469 	}
1470 }
1471 #endif	/* INVARIANT_SUPPORT */
1472 
1473 #ifdef DDB
1474 static void
1475 db_show_sx(const struct lock_object *lock)
1476 {
1477 	struct thread *td;
1478 	const struct sx *sx;
1479 
1480 	sx = (const struct sx *)lock;
1481 
1482 	db_printf(" state: ");
1483 	if (sx->sx_lock == SX_LOCK_UNLOCKED)
1484 		db_printf("UNLOCKED\n");
1485 	else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1486 		db_printf("DESTROYED\n");
1487 		return;
1488 	} else if (sx->sx_lock & SX_LOCK_SHARED)
1489 		db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1490 	else {
1491 		td = sx_xholder(sx);
1492 		db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1493 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1494 		if (sx_recursed(sx))
1495 			db_printf(" recursed: %d\n", sx->sx_recurse);
1496 	}
1497 
1498 	db_printf(" waiters: ");
1499 	switch(sx->sx_lock &
1500 	    (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1501 	case SX_LOCK_SHARED_WAITERS:
1502 		db_printf("shared\n");
1503 		break;
1504 	case SX_LOCK_EXCLUSIVE_WAITERS:
1505 		db_printf("exclusive\n");
1506 		break;
1507 	case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1508 		db_printf("exclusive and shared\n");
1509 		break;
1510 	default:
1511 		db_printf("none\n");
1512 	}
1513 }
1514 
1515 /*
1516  * Check to see if a thread that is blocked on a sleep queue is actually
1517  * blocked on an sx lock.  If so, output some details and return true.
1518  * If the lock has an exclusive owner, return that in *ownerp.
1519  */
1520 int
1521 sx_chain(struct thread *td, struct thread **ownerp)
1522 {
1523 	struct sx *sx;
1524 
1525 	/*
1526 	 * Check to see if this thread is blocked on an sx lock.
1527 	 * First, we check the lock class.  If that is ok, then we
1528 	 * compare the lock name against the wait message.
1529 	 */
1530 	sx = td->td_wchan;
1531 	if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1532 	    sx->lock_object.lo_name != td->td_wmesg)
1533 		return (0);
1534 
1535 	/* We think we have an sx lock, so output some details. */
1536 	db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1537 	*ownerp = sx_xholder(sx);
1538 	if (sx->sx_lock & SX_LOCK_SHARED)
1539 		db_printf("SLOCK (count %ju)\n",
1540 		    (uintmax_t)SX_SHARERS(sx->sx_lock));
1541 	else
1542 		db_printf("XLOCK\n");
1543 	return (1);
1544 }
1545 #endif
1546