xref: /freebsd/sys/kern/kern_sx.c (revision ce9f95bd835ec7d3ac6def5a910282509b1413f0)
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 #ifdef SX_CUSTOM_BACKOFF
147 static u_short __read_frequently asx_retries;
148 static u_short __read_frequently asx_loops;
149 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
150     "sxlock debugging");
151 SYSCTL_U16(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
152 SYSCTL_U16(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
153 
154 static struct lock_delay_config __read_frequently sx_delay;
155 
156 SYSCTL_U16(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
157     0, "");
158 SYSCTL_U16(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
159     0, "");
160 
161 static void
162 sx_lock_delay_init(void *arg __unused)
163 {
164 
165 	lock_delay_default_init(&sx_delay);
166 	asx_retries = 10;
167 	asx_loops = max(10000, sx_delay.max);
168 }
169 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
170 #else
171 #define sx_delay	locks_delay
172 #define asx_retries	locks_delay_retries
173 #define asx_loops	locks_delay_loops
174 #endif
175 #endif
176 
177 void
178 assert_sx(const struct lock_object *lock, int what)
179 {
180 
181 	sx_assert((const struct sx *)lock, what);
182 }
183 
184 void
185 lock_sx(struct lock_object *lock, uintptr_t how)
186 {
187 	struct sx *sx;
188 
189 	sx = (struct sx *)lock;
190 	if (how)
191 		sx_slock(sx);
192 	else
193 		sx_xlock(sx);
194 }
195 
196 uintptr_t
197 unlock_sx(struct lock_object *lock)
198 {
199 	struct sx *sx;
200 
201 	sx = (struct sx *)lock;
202 	sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
203 	if (sx_xlocked(sx)) {
204 		sx_xunlock(sx);
205 		return (0);
206 	} else {
207 		sx_sunlock(sx);
208 		return (1);
209 	}
210 }
211 
212 #ifdef KDTRACE_HOOKS
213 int
214 owner_sx(const struct lock_object *lock, struct thread **owner)
215 {
216 	const struct sx *sx;
217 	uintptr_t x;
218 
219 	sx = (const struct sx *)lock;
220 	x = sx->sx_lock;
221 	*owner = NULL;
222 	return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
223 	    ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
224 }
225 #endif
226 
227 void
228 sx_sysinit(void *arg)
229 {
230 	struct sx_args *sargs = arg;
231 
232 	sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
233 }
234 
235 void
236 sx_init_flags(struct sx *sx, const char *description, int opts)
237 {
238 	int flags;
239 
240 	MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
241 	    SX_NOPROFILE | SX_NEW)) == 0);
242 	ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
243 	    ("%s: sx_lock not aligned for %s: %p", __func__, description,
244 	    &sx->sx_lock));
245 
246 	flags = LO_SLEEPABLE | LO_UPGRADABLE;
247 	if (opts & SX_DUPOK)
248 		flags |= LO_DUPOK;
249 	if (opts & SX_NOPROFILE)
250 		flags |= LO_NOPROFILE;
251 	if (!(opts & SX_NOWITNESS))
252 		flags |= LO_WITNESS;
253 	if (opts & SX_RECURSE)
254 		flags |= LO_RECURSABLE;
255 	if (opts & SX_QUIET)
256 		flags |= LO_QUIET;
257 	if (opts & SX_NEW)
258 		flags |= LO_NEW;
259 
260 	lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
261 	sx->sx_lock = SX_LOCK_UNLOCKED;
262 	sx->sx_recurse = 0;
263 }
264 
265 void
266 sx_destroy(struct sx *sx)
267 {
268 
269 	KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
270 	KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
271 	sx->sx_lock = SX_LOCK_DESTROYED;
272 	lock_destroy(&sx->lock_object);
273 }
274 
275 int
276 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
277 {
278 	uintptr_t x;
279 
280 	if (SCHEDULER_STOPPED())
281 		return (1);
282 
283 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
284 	    ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
285 	    curthread, sx->lock_object.lo_name, file, line));
286 
287 	x = sx->sx_lock;
288 	for (;;) {
289 		KASSERT(x != SX_LOCK_DESTROYED,
290 		    ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
291 		if (!(x & SX_LOCK_SHARED))
292 			break;
293 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
294 			LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
295 			WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
296 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
297 			    sx, 0, 0, file, line, LOCKSTAT_READER);
298 			TD_LOCKS_INC(curthread);
299 			curthread->td_sx_slocks++;
300 			return (1);
301 		}
302 	}
303 
304 	LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
305 	return (0);
306 }
307 
308 int
309 sx_try_slock_(struct sx *sx, const char *file, int line)
310 {
311 
312 	return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
313 }
314 
315 int
316 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
317 {
318 	uintptr_t tid, x;
319 	int error = 0;
320 
321 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
322 	    !TD_IS_IDLETHREAD(curthread),
323 	    ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
324 	    curthread, sx->lock_object.lo_name, file, line));
325 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
326 	    ("sx_xlock() of destroyed sx @ %s:%d", file, line));
327 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
328 	    line, NULL);
329 	tid = (uintptr_t)curthread;
330 	x = SX_LOCK_UNLOCKED;
331 	if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
332 		error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
333 	else
334 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
335 		    0, 0, file, line, LOCKSTAT_WRITER);
336 	if (!error) {
337 		LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
338 		    file, line);
339 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
340 		TD_LOCKS_INC(curthread);
341 	}
342 
343 	return (error);
344 }
345 
346 int
347 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
348 {
349 	struct thread *td;
350 	uintptr_t tid, x;
351 	int rval;
352 	bool recursed;
353 
354 	td = curthread;
355 	tid = (uintptr_t)td;
356 	if (SCHEDULER_STOPPED_TD(td))
357 		return (1);
358 
359 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
360 	    ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
361 	    curthread, sx->lock_object.lo_name, file, line));
362 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
363 	    ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
364 
365 	rval = 1;
366 	recursed = false;
367 	x = SX_LOCK_UNLOCKED;
368 	for (;;) {
369 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
370 			break;
371 		if (x == SX_LOCK_UNLOCKED)
372 			continue;
373 		if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
374 			sx->sx_recurse++;
375 			atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
376 			break;
377 		}
378 		rval = 0;
379 		break;
380 	}
381 
382 	LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
383 	if (rval) {
384 		WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
385 		    file, line);
386 		if (!recursed)
387 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
388 			    sx, 0, 0, file, line, LOCKSTAT_WRITER);
389 		TD_LOCKS_INC(curthread);
390 	}
391 
392 	return (rval);
393 }
394 
395 int
396 sx_try_xlock_(struct sx *sx, const char *file, int line)
397 {
398 
399 	return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
400 }
401 
402 void
403 _sx_xunlock(struct sx *sx, const char *file, int line)
404 {
405 
406 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
407 	    ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
408 	_sx_assert(sx, SA_XLOCKED, file, line);
409 	WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
410 	LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
411 	    line);
412 #if LOCK_DEBUG > 0
413 	_sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
414 #else
415 	__sx_xunlock(sx, curthread, file, line);
416 #endif
417 	TD_LOCKS_DEC(curthread);
418 }
419 
420 /*
421  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
422  * This will only succeed if this thread holds a single shared lock.
423  * Return 1 if if the upgrade succeed, 0 otherwise.
424  */
425 int
426 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
427 {
428 	uintptr_t x;
429 	uintptr_t waiters;
430 	int success;
431 
432 	if (SCHEDULER_STOPPED())
433 		return (1);
434 
435 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
436 	    ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
437 	_sx_assert(sx, SA_SLOCKED, file, line);
438 
439 	/*
440 	 * Try to switch from one shared lock to an exclusive lock.  We need
441 	 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
442 	 * we will wake up the exclusive waiters when we drop the lock.
443 	 */
444 	success = 0;
445 	x = SX_READ_VALUE(sx);
446 	for (;;) {
447 		if (SX_SHARERS(x) > 1)
448 			break;
449 		waiters = (x & SX_LOCK_WAITERS);
450 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
451 		    (uintptr_t)curthread | waiters)) {
452 			success = 1;
453 			break;
454 		}
455 	}
456 	LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
457 	if (success) {
458 		curthread->td_sx_slocks--;
459 		WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
460 		    file, line);
461 		LOCKSTAT_RECORD0(sx__upgrade, sx);
462 	}
463 	return (success);
464 }
465 
466 int
467 sx_try_upgrade_(struct sx *sx, const char *file, int line)
468 {
469 
470 	return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
471 }
472 
473 /*
474  * Downgrade an unrecursed exclusive lock into a single shared lock.
475  */
476 void
477 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
478 {
479 	uintptr_t x;
480 	int wakeup_swapper;
481 
482 	if (SCHEDULER_STOPPED())
483 		return;
484 
485 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
486 	    ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
487 	_sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
488 #ifndef INVARIANTS
489 	if (sx_recursed(sx))
490 		panic("downgrade of a recursed lock");
491 #endif
492 
493 	WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
494 
495 	/*
496 	 * Try to switch from an exclusive lock with no shared waiters
497 	 * to one sharer with no shared waiters.  If there are
498 	 * exclusive waiters, we don't need to lock the sleep queue so
499 	 * long as we preserve the flag.  We do one quick try and if
500 	 * that fails we grab the sleepq lock to keep the flags from
501 	 * changing and do it the slow way.
502 	 *
503 	 * We have to lock the sleep queue if there are shared waiters
504 	 * so we can wake them up.
505 	 */
506 	x = sx->sx_lock;
507 	if (!(x & SX_LOCK_SHARED_WAITERS) &&
508 	    atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
509 	    (x & SX_LOCK_EXCLUSIVE_WAITERS)))
510 		goto out;
511 
512 	/*
513 	 * Lock the sleep queue so we can read the waiters bits
514 	 * without any races and wakeup any shared waiters.
515 	 */
516 	sleepq_lock(&sx->lock_object);
517 
518 	/*
519 	 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
520 	 * shared lock.  If there are any shared waiters, wake them up.
521 	 */
522 	wakeup_swapper = 0;
523 	x = sx->sx_lock;
524 	atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
525 	    (x & SX_LOCK_EXCLUSIVE_WAITERS));
526 	if (x & SX_LOCK_SHARED_WAITERS)
527 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
528 		    0, SQ_SHARED_QUEUE);
529 	sleepq_release(&sx->lock_object);
530 
531 	if (wakeup_swapper)
532 		kick_proc0();
533 
534 out:
535 	curthread->td_sx_slocks++;
536 	LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
537 	LOCKSTAT_RECORD0(sx__downgrade, sx);
538 }
539 
540 void
541 sx_downgrade_(struct sx *sx, const char *file, int line)
542 {
543 
544 	sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
545 }
546 
547 #ifdef	ADAPTIVE_SX
548 static inline void
549 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
550 {
551 
552 	if (x & SX_LOCK_WRITE_SPINNER)
553 		return;
554 	if (*in_critical) {
555 		critical_exit();
556 		*in_critical = false;
557 		(*extra_work)--;
558 	}
559 }
560 #else
561 #define sx_drop_critical(x, in_critical, extra_work) do { } while (0)
562 #endif
563 
564 /*
565  * This function represents the so-called 'hard case' for sx_xlock
566  * operation.  All 'easy case' failures are redirected to this.  Note
567  * that ideally this would be a static function, but it needs to be
568  * accessible from at least sx.h.
569  */
570 int
571 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
572 {
573 	GIANT_DECLARE;
574 	uintptr_t tid, setx;
575 #ifdef ADAPTIVE_SX
576 	struct thread *owner;
577 	u_int i, n, spintries = 0;
578 	enum { READERS, WRITER } sleep_reason = READERS;
579 	bool in_critical = false;
580 #endif
581 #ifdef LOCK_PROFILING
582 	uint64_t waittime = 0;
583 	int contested = 0;
584 #endif
585 	int error = 0;
586 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
587 	struct lock_delay_arg lda;
588 #endif
589 #ifdef	KDTRACE_HOOKS
590 	u_int sleep_cnt = 0;
591 	int64_t sleep_time = 0;
592 	int64_t all_time = 0;
593 #endif
594 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
595 	uintptr_t state = 0;
596 	int doing_lockprof = 0;
597 #endif
598 	int extra_work = 0;
599 
600 	tid = (uintptr_t)curthread;
601 
602 #ifdef KDTRACE_HOOKS
603 	if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
604 		while (x == SX_LOCK_UNLOCKED) {
605 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
606 				goto out_lockstat;
607 		}
608 		extra_work = 1;
609 		doing_lockprof = 1;
610 		all_time -= lockstat_nsecs(&sx->lock_object);
611 		state = x;
612 	}
613 #endif
614 #ifdef LOCK_PROFILING
615 	extra_work = 1;
616 	doing_lockprof = 1;
617 	state = x;
618 #endif
619 
620 	if (SCHEDULER_STOPPED())
621 		return (0);
622 
623 	if (__predict_false(x == SX_LOCK_UNLOCKED))
624 		x = SX_READ_VALUE(sx);
625 
626 	/* If we already hold an exclusive lock, then recurse. */
627 	if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
628 		KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
629 	    ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
630 		    sx->lock_object.lo_name, file, line));
631 		sx->sx_recurse++;
632 		atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
633 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
634 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
635 		return (0);
636 	}
637 
638 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
639 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
640 		    sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
641 
642 #if defined(ADAPTIVE_SX)
643 	lock_delay_arg_init(&lda, &sx_delay);
644 #elif defined(KDTRACE_HOOKS)
645 	lock_delay_arg_init_noadapt(&lda);
646 #endif
647 
648 #ifdef HWPMC_HOOKS
649 	PMC_SOFT_CALL( , , lock, failed);
650 #endif
651 	lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
652 	    &waittime);
653 
654 #ifndef INVARIANTS
655 	GIANT_SAVE(extra_work);
656 #endif
657 
658 	for (;;) {
659 		if (x == SX_LOCK_UNLOCKED) {
660 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
661 				break;
662 			continue;
663 		}
664 #ifdef INVARIANTS
665 		GIANT_SAVE(extra_work);
666 #endif
667 #ifdef KDTRACE_HOOKS
668 		lda.spin_cnt++;
669 #endif
670 #ifdef ADAPTIVE_SX
671 		if (x == (SX_LOCK_SHARED | SX_LOCK_WRITE_SPINNER)) {
672 			if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
673 				break;
674 			continue;
675 		}
676 
677 		/*
678 		 * If the lock is write locked and the owner is
679 		 * running on another CPU, spin until the owner stops
680 		 * running or the state of the lock changes.
681 		 */
682 		if ((x & SX_LOCK_SHARED) == 0) {
683 			sx_drop_critical(x, &in_critical, &extra_work);
684 			sleep_reason = WRITER;
685 			owner = lv_sx_owner(x);
686 			if (!TD_IS_RUNNING(owner))
687 				goto sleepq;
688 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
689 				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
690 				    __func__, sx, owner);
691 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
692 			    "spinning", "lockname:\"%s\"",
693 			    sx->lock_object.lo_name);
694 			do {
695 				lock_delay(&lda);
696 				x = SX_READ_VALUE(sx);
697 				owner = lv_sx_owner(x);
698 			} while (owner != NULL && TD_IS_RUNNING(owner));
699 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
700 			    "running");
701 			continue;
702 		} else if (SX_SHARERS(x) > 0) {
703 			sleep_reason = READERS;
704 			if (spintries == asx_retries)
705 				goto sleepq;
706 			if (!(x & SX_LOCK_WRITE_SPINNER)) {
707 				if (!in_critical) {
708 					critical_enter();
709 					in_critical = true;
710 					extra_work++;
711 				}
712 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
713 				    x | SX_LOCK_WRITE_SPINNER)) {
714 					critical_exit();
715 					in_critical = false;
716 					extra_work--;
717 					continue;
718 				}
719 			}
720 			spintries++;
721 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
722 			    "spinning", "lockname:\"%s\"",
723 			    sx->lock_object.lo_name);
724 			n = SX_SHARERS(x);
725 			for (i = 0; i < asx_loops; i += n) {
726 				lock_delay_spin(n);
727 				x = SX_READ_VALUE(sx);
728 				if (!(x & SX_LOCK_WRITE_SPINNER))
729 					break;
730 				if (!(x & SX_LOCK_SHARED))
731 					break;
732 				n = SX_SHARERS(x);
733 				if (n == 0)
734 					break;
735 			}
736 #ifdef KDTRACE_HOOKS
737 			lda.spin_cnt += i;
738 #endif
739 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
740 			    "running");
741 			if (i < asx_loops)
742 				continue;
743 		}
744 sleepq:
745 #endif
746 		sleepq_lock(&sx->lock_object);
747 		x = SX_READ_VALUE(sx);
748 retry_sleepq:
749 
750 		/*
751 		 * If the lock was released while spinning on the
752 		 * sleep queue chain lock, try again.
753 		 */
754 		if (x == SX_LOCK_UNLOCKED) {
755 			sleepq_release(&sx->lock_object);
756 			sx_drop_critical(x, &in_critical, &extra_work);
757 			continue;
758 		}
759 
760 #ifdef ADAPTIVE_SX
761 		/*
762 		 * The current lock owner might have started executing
763 		 * on another CPU (or the lock could have changed
764 		 * owners) while we were waiting on the sleep queue
765 		 * chain lock.  If so, drop the sleep queue lock and try
766 		 * again.
767 		 */
768 		if (!(x & SX_LOCK_SHARED)) {
769 			owner = (struct thread *)SX_OWNER(x);
770 			if (TD_IS_RUNNING(owner)) {
771 				sleepq_release(&sx->lock_object);
772 				sx_drop_critical(x, &in_critical,
773 				    &extra_work);
774 				continue;
775 			}
776 		} else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
777 			sleepq_release(&sx->lock_object);
778 			sx_drop_critical(x, &in_critical, &extra_work);
779 			continue;
780 		}
781 #endif
782 
783 		/*
784 		 * If an exclusive lock was released with both shared
785 		 * and exclusive waiters and a shared waiter hasn't
786 		 * woken up and acquired the lock yet, sx_lock will be
787 		 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
788 		 * If we see that value, try to acquire it once.  Note
789 		 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
790 		 * as there are other exclusive waiters still.  If we
791 		 * fail, restart the loop.
792 		 */
793 		setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
794 		if ((x & ~setx) == SX_LOCK_SHARED) {
795 			setx &= ~SX_LOCK_WRITE_SPINNER;
796 			if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
797 				goto retry_sleepq;
798 			sleepq_release(&sx->lock_object);
799 			CTR2(KTR_LOCK, "%s: %p claimed by new writer",
800 			    __func__, sx);
801 			break;
802 		}
803 
804 #ifdef ADAPTIVE_SX
805 		/*
806 		 * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
807 		 * It is an invariant that when the bit is set, there is
808 		 * a writer ready to grab the lock. Thus clear the bit since
809 		 * we are going to sleep.
810 		 */
811 		if (in_critical) {
812 			if ((x & SX_LOCK_WRITE_SPINNER) ||
813 			    !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
814 				setx = x & ~SX_LOCK_WRITE_SPINNER;
815 				setx |= SX_LOCK_EXCLUSIVE_WAITERS;
816 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
817 				    setx)) {
818 					goto retry_sleepq;
819 				}
820 			}
821 			critical_exit();
822 			in_critical = false;
823 		} else {
824 #endif
825 			/*
826 			 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
827 			 * than loop back and retry.
828 			 */
829 			if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
830 				if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
831 				    x | SX_LOCK_EXCLUSIVE_WAITERS)) {
832 					goto retry_sleepq;
833 				}
834 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
835 					CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
836 					    __func__, sx);
837 			}
838 #ifdef ADAPTIVE_SX
839 		}
840 #endif
841 
842 		/*
843 		 * Since we have been unable to acquire the exclusive
844 		 * lock and the exclusive waiters flag is set, we have
845 		 * to sleep.
846 		 */
847 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
848 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
849 			    __func__, sx);
850 
851 #ifdef KDTRACE_HOOKS
852 		sleep_time -= lockstat_nsecs(&sx->lock_object);
853 #endif
854 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
855 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
856 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
857 		if (!(opts & SX_INTERRUPTIBLE))
858 			sleepq_wait(&sx->lock_object, 0);
859 		else
860 			error = sleepq_wait_sig(&sx->lock_object, 0);
861 #ifdef KDTRACE_HOOKS
862 		sleep_time += lockstat_nsecs(&sx->lock_object);
863 		sleep_cnt++;
864 #endif
865 		if (error) {
866 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
867 				CTR2(KTR_LOCK,
868 			"%s: interruptible sleep by %p suspended by signal",
869 				    __func__, sx);
870 			break;
871 		}
872 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
873 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
874 			    __func__, sx);
875 		x = SX_READ_VALUE(sx);
876 	}
877 	if (__predict_true(!extra_work))
878 		return (error);
879 #ifdef ADAPTIVE_SX
880 	if (in_critical)
881 		critical_exit();
882 #endif
883 	GIANT_RESTORE();
884 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
885 	if (__predict_true(!doing_lockprof))
886 		return (error);
887 #endif
888 #ifdef KDTRACE_HOOKS
889 	all_time += lockstat_nsecs(&sx->lock_object);
890 	if (sleep_time)
891 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
892 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
893 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
894 	if (lda.spin_cnt > sleep_cnt)
895 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
896 		    LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
897 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
898 out_lockstat:
899 #endif
900 	if (!error)
901 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
902 		    contested, waittime, file, line, LOCKSTAT_WRITER);
903 	return (error);
904 }
905 
906 /*
907  * This function represents the so-called 'hard case' for sx_xunlock
908  * operation.  All 'easy case' failures are redirected to this.  Note
909  * that ideally this would be a static function, but it needs to be
910  * accessible from at least sx.h.
911  */
912 void
913 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
914 {
915 	uintptr_t tid, setx;
916 	int queue, wakeup_swapper;
917 
918 	if (SCHEDULER_STOPPED())
919 		return;
920 
921 	tid = (uintptr_t)curthread;
922 
923 	if (__predict_false(x == tid))
924 		x = SX_READ_VALUE(sx);
925 
926 	MPASS(!(x & SX_LOCK_SHARED));
927 
928 	if (__predict_false(x & SX_LOCK_RECURSED)) {
929 		/* The lock is recursed, unrecurse one level. */
930 		if ((--sx->sx_recurse) == 0)
931 			atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
932 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
933 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
934 		return;
935 	}
936 
937 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
938 	if (x == tid &&
939 	    atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
940 		return;
941 
942 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
943 		CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
944 
945 	sleepq_lock(&sx->lock_object);
946 	x = SX_READ_VALUE(sx);
947 	MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
948 
949 	/*
950 	 * The wake up algorithm here is quite simple and probably not
951 	 * ideal.  It gives precedence to shared waiters if they are
952 	 * present.  For this condition, we have to preserve the
953 	 * state of the exclusive waiters flag.
954 	 * If interruptible sleeps left the shared queue empty avoid a
955 	 * starvation for the threads sleeping on the exclusive queue by giving
956 	 * them precedence and cleaning up the shared waiters bit anyway.
957 	 */
958 	setx = SX_LOCK_UNLOCKED;
959 	queue = SQ_SHARED_QUEUE;
960 	if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
961 	    sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
962 		queue = SQ_EXCLUSIVE_QUEUE;
963 		setx |= (x & SX_LOCK_SHARED_WAITERS);
964 	}
965 	atomic_store_rel_ptr(&sx->sx_lock, setx);
966 
967 	/* Wake up all the waiters for the specific queue. */
968 	if (LOCK_LOG_TEST(&sx->lock_object, 0))
969 		CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
970 		    __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
971 		    "exclusive");
972 
973 	wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
974 	    queue);
975 	sleepq_release(&sx->lock_object);
976 	if (wakeup_swapper)
977 		kick_proc0();
978 }
979 
980 static bool __always_inline
981 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
982 {
983 
984 	if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
985 			== SX_LOCK_SHARED)
986 		return (true);
987 	if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
988 		return (true);
989 	return (false);
990 }
991 
992 static bool __always_inline
993 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
994     LOCK_FILE_LINE_ARG_DEF)
995 {
996 
997 	/*
998 	 * If no other thread has an exclusive lock then try to bump up
999 	 * the count of sharers.  Since we have to preserve the state
1000 	 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
1001 	 * shared lock loop back and retry.
1002 	 */
1003 	while (__sx_can_read(td, *xp, fp)) {
1004 		if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1005 		    *xp + SX_ONE_SHARER)) {
1006 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1007 				CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1008 				    __func__, sx, (void *)*xp,
1009 				    (void *)(*xp + SX_ONE_SHARER));
1010 			td->td_sx_slocks++;
1011 			return (true);
1012 		}
1013 	}
1014 	return (false);
1015 }
1016 
1017 static int __noinline
1018 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1019 {
1020 	GIANT_DECLARE;
1021 	struct thread *td;
1022 #ifdef ADAPTIVE_SX
1023 	struct thread *owner;
1024 	u_int i, n, spintries = 0;
1025 #endif
1026 #ifdef LOCK_PROFILING
1027 	uint64_t waittime = 0;
1028 	int contested = 0;
1029 #endif
1030 	int error = 0;
1031 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1032 	struct lock_delay_arg lda;
1033 #endif
1034 #ifdef KDTRACE_HOOKS
1035 	u_int sleep_cnt = 0;
1036 	int64_t sleep_time = 0;
1037 	int64_t all_time = 0;
1038 #endif
1039 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1040 	uintptr_t state = 0;
1041 #endif
1042 	int extra_work __sdt_used = 0;
1043 
1044 	td = curthread;
1045 
1046 #ifdef KDTRACE_HOOKS
1047 	if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1048 		if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1049 			goto out_lockstat;
1050 		extra_work = 1;
1051 		all_time -= lockstat_nsecs(&sx->lock_object);
1052 		state = x;
1053 	}
1054 #endif
1055 #ifdef LOCK_PROFILING
1056 	extra_work = 1;
1057 	state = x;
1058 #endif
1059 
1060 	if (SCHEDULER_STOPPED())
1061 		return (0);
1062 
1063 #if defined(ADAPTIVE_SX)
1064 	lock_delay_arg_init(&lda, &sx_delay);
1065 #elif defined(KDTRACE_HOOKS)
1066 	lock_delay_arg_init_noadapt(&lda);
1067 #endif
1068 
1069 #ifdef HWPMC_HOOKS
1070 	PMC_SOFT_CALL( , , lock, failed);
1071 #endif
1072 	lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
1073 	    &waittime);
1074 
1075 #ifndef INVARIANTS
1076 	GIANT_SAVE(extra_work);
1077 #endif
1078 
1079 	/*
1080 	 * As with rwlocks, we don't make any attempt to try to block
1081 	 * shared locks once there is an exclusive waiter.
1082 	 */
1083 	for (;;) {
1084 		if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1085 			break;
1086 #ifdef INVARIANTS
1087 		GIANT_SAVE(extra_work);
1088 #endif
1089 #ifdef KDTRACE_HOOKS
1090 		lda.spin_cnt++;
1091 #endif
1092 
1093 #ifdef ADAPTIVE_SX
1094 		/*
1095 		 * If the owner is running on another CPU, spin until
1096 		 * the owner stops running or the state of the lock
1097 		 * changes.
1098 		 */
1099 		if ((x & SX_LOCK_SHARED) == 0) {
1100 			owner = lv_sx_owner(x);
1101 			if (TD_IS_RUNNING(owner)) {
1102 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1103 					CTR3(KTR_LOCK,
1104 					    "%s: spinning on %p held by %p",
1105 					    __func__, sx, owner);
1106 				KTR_STATE1(KTR_SCHED, "thread",
1107 				    sched_tdname(curthread), "spinning",
1108 				    "lockname:\"%s\"", sx->lock_object.lo_name);
1109 				do {
1110 					lock_delay(&lda);
1111 					x = SX_READ_VALUE(sx);
1112 					owner = lv_sx_owner(x);
1113 				} while (owner != NULL && TD_IS_RUNNING(owner));
1114 				KTR_STATE0(KTR_SCHED, "thread",
1115 				    sched_tdname(curthread), "running");
1116 				continue;
1117 			}
1118 		} else {
1119 			if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1120 				MPASS(!__sx_can_read(td, x, false));
1121 				lock_delay_spin(2);
1122 				x = SX_READ_VALUE(sx);
1123 				continue;
1124 			}
1125 			if (spintries < asx_retries) {
1126 				KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1127 				    "spinning", "lockname:\"%s\"",
1128 				    sx->lock_object.lo_name);
1129 				n = SX_SHARERS(x);
1130 				for (i = 0; i < asx_loops; i += n) {
1131 					lock_delay_spin(n);
1132 					x = SX_READ_VALUE(sx);
1133 					if (!(x & SX_LOCK_SHARED))
1134 						break;
1135 					n = SX_SHARERS(x);
1136 					if (n == 0)
1137 						break;
1138 					if (__sx_can_read(td, x, false))
1139 						break;
1140 				}
1141 #ifdef KDTRACE_HOOKS
1142 				lda.spin_cnt += i;
1143 #endif
1144 				KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1145 				    "running");
1146 				if (i < asx_loops)
1147 					continue;
1148 			}
1149 		}
1150 #endif
1151 
1152 		/*
1153 		 * Some other thread already has an exclusive lock, so
1154 		 * start the process of blocking.
1155 		 */
1156 		sleepq_lock(&sx->lock_object);
1157 		x = SX_READ_VALUE(sx);
1158 retry_sleepq:
1159 		if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1160 		    __sx_can_read(td, x, false)) {
1161 			sleepq_release(&sx->lock_object);
1162 			continue;
1163 		}
1164 
1165 #ifdef ADAPTIVE_SX
1166 		/*
1167 		 * If the owner is running on another CPU, spin until
1168 		 * the owner stops running or the state of the lock
1169 		 * changes.
1170 		 */
1171 		if (!(x & SX_LOCK_SHARED)) {
1172 			owner = (struct thread *)SX_OWNER(x);
1173 			if (TD_IS_RUNNING(owner)) {
1174 				sleepq_release(&sx->lock_object);
1175 				x = SX_READ_VALUE(sx);
1176 				continue;
1177 			}
1178 		}
1179 #endif
1180 
1181 		/*
1182 		 * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1183 		 * fail to set it drop the sleep queue lock and loop
1184 		 * back.
1185 		 */
1186 		if (!(x & SX_LOCK_SHARED_WAITERS)) {
1187 			if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1188 			    x | SX_LOCK_SHARED_WAITERS))
1189 				goto retry_sleepq;
1190 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1191 				CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1192 				    __func__, sx);
1193 		}
1194 
1195 		/*
1196 		 * Since we have been unable to acquire the shared lock,
1197 		 * we have to sleep.
1198 		 */
1199 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1200 			CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1201 			    __func__, sx);
1202 
1203 #ifdef KDTRACE_HOOKS
1204 		sleep_time -= lockstat_nsecs(&sx->lock_object);
1205 #endif
1206 		sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1207 		    SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1208 		    SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1209 		if (!(opts & SX_INTERRUPTIBLE))
1210 			sleepq_wait(&sx->lock_object, 0);
1211 		else
1212 			error = sleepq_wait_sig(&sx->lock_object, 0);
1213 #ifdef KDTRACE_HOOKS
1214 		sleep_time += lockstat_nsecs(&sx->lock_object);
1215 		sleep_cnt++;
1216 #endif
1217 		if (error) {
1218 			if (LOCK_LOG_TEST(&sx->lock_object, 0))
1219 				CTR2(KTR_LOCK,
1220 			"%s: interruptible sleep by %p suspended by signal",
1221 				    __func__, sx);
1222 			break;
1223 		}
1224 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1225 			CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1226 			    __func__, sx);
1227 		x = SX_READ_VALUE(sx);
1228 	}
1229 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1230 	if (__predict_true(!extra_work))
1231 		return (error);
1232 #endif
1233 #ifdef KDTRACE_HOOKS
1234 	all_time += lockstat_nsecs(&sx->lock_object);
1235 	if (sleep_time)
1236 		LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1237 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1238 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1239 	if (lda.spin_cnt > sleep_cnt)
1240 		LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1241 		    LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1242 		    (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1243 out_lockstat:
1244 #endif
1245 	if (error == 0) {
1246 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1247 		    contested, waittime, file, line, LOCKSTAT_READER);
1248 	}
1249 	GIANT_RESTORE();
1250 	return (error);
1251 }
1252 
1253 int
1254 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1255 {
1256 	struct thread *td;
1257 	uintptr_t x;
1258 	int error;
1259 
1260 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1261 	    !TD_IS_IDLETHREAD(curthread),
1262 	    ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1263 	    curthread, sx->lock_object.lo_name, file, line));
1264 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1265 	    ("sx_slock() of destroyed sx @ %s:%d", file, line));
1266 	WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1267 
1268 	error = 0;
1269 	td = curthread;
1270 	x = SX_READ_VALUE(sx);
1271 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1272 	    !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1273 		error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1274 	else
1275 		lock_profile_obtain_lock_success(&sx->lock_object, false, 0, 0,
1276 		    file, line);
1277 	if (error == 0) {
1278 		LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1279 		WITNESS_LOCK(&sx->lock_object, 0, file, line);
1280 		TD_LOCKS_INC(curthread);
1281 	}
1282 	return (error);
1283 }
1284 
1285 int
1286 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1287 {
1288 
1289 	return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1290 }
1291 
1292 static bool __always_inline
1293 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1294 {
1295 
1296 	for (;;) {
1297 		if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1298 			if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1299 			    *xp - SX_ONE_SHARER)) {
1300 				if (LOCK_LOG_TEST(&sx->lock_object, 0))
1301 					CTR4(KTR_LOCK,
1302 					    "%s: %p succeeded %p -> %p",
1303 					    __func__, sx, (void *)*xp,
1304 					    (void *)(*xp - SX_ONE_SHARER));
1305 				td->td_sx_slocks--;
1306 				return (true);
1307 			}
1308 			continue;
1309 		}
1310 		break;
1311 	}
1312 	return (false);
1313 }
1314 
1315 static void __noinline
1316 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1317     LOCK_FILE_LINE_ARG_DEF)
1318 {
1319 	int wakeup_swapper = 0;
1320 	uintptr_t setx, queue;
1321 
1322 	if (SCHEDULER_STOPPED())
1323 		return;
1324 
1325 	if (_sx_sunlock_try(sx, td, &x))
1326 		goto out_lockstat;
1327 
1328 	sleepq_lock(&sx->lock_object);
1329 	x = SX_READ_VALUE(sx);
1330 	for (;;) {
1331 		if (_sx_sunlock_try(sx, td, &x))
1332 			break;
1333 
1334 		/*
1335 		 * Wake up semantic here is quite simple:
1336 		 * Just wake up all the exclusive waiters.
1337 		 * Note that the state of the lock could have changed,
1338 		 * so if it fails loop back and retry.
1339 		 */
1340 		setx = SX_LOCK_UNLOCKED;
1341 		queue = SQ_SHARED_QUEUE;
1342 		if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1343 			setx |= (x & SX_LOCK_SHARED_WAITERS);
1344 			queue = SQ_EXCLUSIVE_QUEUE;
1345 		}
1346 		setx |= (x & SX_LOCK_WRITE_SPINNER);
1347 		if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1348 			continue;
1349 		if (LOCK_LOG_TEST(&sx->lock_object, 0))
1350 			CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1351 			    "exclusive queue", __func__, sx);
1352 		wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1353 		    0, queue);
1354 		td->td_sx_slocks--;
1355 		break;
1356 	}
1357 	sleepq_release(&sx->lock_object);
1358 	if (wakeup_swapper)
1359 		kick_proc0();
1360 out_lockstat:
1361 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1362 }
1363 
1364 void
1365 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1366 {
1367 	struct thread *td;
1368 	uintptr_t x;
1369 
1370 	KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1371 	    ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1372 	_sx_assert(sx, SA_SLOCKED, file, line);
1373 	WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1374 	LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1375 
1376 	td = curthread;
1377 	x = SX_READ_VALUE(sx);
1378 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1379 	    !_sx_sunlock_try(sx, td, &x)))
1380 		_sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1381 	else
1382 		lock_profile_release_lock(&sx->lock_object, false);
1383 
1384 	TD_LOCKS_DEC(curthread);
1385 }
1386 
1387 void
1388 _sx_sunlock(struct sx *sx, const char *file, int line)
1389 {
1390 
1391 	_sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1392 }
1393 
1394 #ifdef INVARIANT_SUPPORT
1395 #ifndef INVARIANTS
1396 #undef	_sx_assert
1397 #endif
1398 
1399 /*
1400  * In the non-WITNESS case, sx_assert() can only detect that at least
1401  * *some* thread owns an slock, but it cannot guarantee that *this*
1402  * thread owns an slock.
1403  */
1404 void
1405 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1406 {
1407 #ifndef WITNESS
1408 	int slocked = 0;
1409 #endif
1410 
1411 	if (SCHEDULER_STOPPED())
1412 		return;
1413 	switch (what) {
1414 	case SA_SLOCKED:
1415 	case SA_SLOCKED | SA_NOTRECURSED:
1416 	case SA_SLOCKED | SA_RECURSED:
1417 #ifndef WITNESS
1418 		slocked = 1;
1419 		/* FALLTHROUGH */
1420 #endif
1421 	case SA_LOCKED:
1422 	case SA_LOCKED | SA_NOTRECURSED:
1423 	case SA_LOCKED | SA_RECURSED:
1424 #ifdef WITNESS
1425 		witness_assert(&sx->lock_object, what, file, line);
1426 #else
1427 		/*
1428 		 * If some other thread has an exclusive lock or we
1429 		 * have one and are asserting a shared lock, fail.
1430 		 * Also, if no one has a lock at all, fail.
1431 		 */
1432 		if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1433 		    (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1434 		    sx_xholder(sx) != curthread)))
1435 			panic("Lock %s not %slocked @ %s:%d\n",
1436 			    sx->lock_object.lo_name, slocked ? "share " : "",
1437 			    file, line);
1438 
1439 		if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1440 			if (sx_recursed(sx)) {
1441 				if (what & SA_NOTRECURSED)
1442 					panic("Lock %s recursed @ %s:%d\n",
1443 					    sx->lock_object.lo_name, file,
1444 					    line);
1445 			} else if (what & SA_RECURSED)
1446 				panic("Lock %s not recursed @ %s:%d\n",
1447 				    sx->lock_object.lo_name, file, line);
1448 		}
1449 #endif
1450 		break;
1451 	case SA_XLOCKED:
1452 	case SA_XLOCKED | SA_NOTRECURSED:
1453 	case SA_XLOCKED | SA_RECURSED:
1454 		if (sx_xholder(sx) != curthread)
1455 			panic("Lock %s not exclusively locked @ %s:%d\n",
1456 			    sx->lock_object.lo_name, file, line);
1457 		if (sx_recursed(sx)) {
1458 			if (what & SA_NOTRECURSED)
1459 				panic("Lock %s recursed @ %s:%d\n",
1460 				    sx->lock_object.lo_name, file, line);
1461 		} else if (what & SA_RECURSED)
1462 			panic("Lock %s not recursed @ %s:%d\n",
1463 			    sx->lock_object.lo_name, file, line);
1464 		break;
1465 	case SA_UNLOCKED:
1466 #ifdef WITNESS
1467 		witness_assert(&sx->lock_object, what, file, line);
1468 #else
1469 		/*
1470 		 * If we hold an exclusve lock fail.  We can't
1471 		 * reliably check to see if we hold a shared lock or
1472 		 * not.
1473 		 */
1474 		if (sx_xholder(sx) == curthread)
1475 			panic("Lock %s exclusively locked @ %s:%d\n",
1476 			    sx->lock_object.lo_name, file, line);
1477 #endif
1478 		break;
1479 	default:
1480 		panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1481 		    line);
1482 	}
1483 }
1484 #endif	/* INVARIANT_SUPPORT */
1485 
1486 #ifdef DDB
1487 static void
1488 db_show_sx(const struct lock_object *lock)
1489 {
1490 	struct thread *td;
1491 	const struct sx *sx;
1492 
1493 	sx = (const struct sx *)lock;
1494 
1495 	db_printf(" state: ");
1496 	if (sx->sx_lock == SX_LOCK_UNLOCKED)
1497 		db_printf("UNLOCKED\n");
1498 	else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1499 		db_printf("DESTROYED\n");
1500 		return;
1501 	} else if (sx->sx_lock & SX_LOCK_SHARED)
1502 		db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1503 	else {
1504 		td = sx_xholder(sx);
1505 		db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1506 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1507 		if (sx_recursed(sx))
1508 			db_printf(" recursed: %d\n", sx->sx_recurse);
1509 	}
1510 
1511 	db_printf(" waiters: ");
1512 	switch(sx->sx_lock &
1513 	    (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1514 	case SX_LOCK_SHARED_WAITERS:
1515 		db_printf("shared\n");
1516 		break;
1517 	case SX_LOCK_EXCLUSIVE_WAITERS:
1518 		db_printf("exclusive\n");
1519 		break;
1520 	case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1521 		db_printf("exclusive and shared\n");
1522 		break;
1523 	default:
1524 		db_printf("none\n");
1525 	}
1526 }
1527 
1528 /*
1529  * Check to see if a thread that is blocked on a sleep queue is actually
1530  * blocked on an sx lock.  If so, output some details and return true.
1531  * If the lock has an exclusive owner, return that in *ownerp.
1532  */
1533 int
1534 sx_chain(struct thread *td, struct thread **ownerp)
1535 {
1536 	const struct sx *sx;
1537 
1538 	/*
1539 	 * Check to see if this thread is blocked on an sx lock.
1540 	 * First, we check the lock class.  If that is ok, then we
1541 	 * compare the lock name against the wait message.
1542 	 */
1543 	sx = td->td_wchan;
1544 	if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1545 	    sx->lock_object.lo_name != td->td_wmesg)
1546 		return (0);
1547 
1548 	/* We think we have an sx lock, so output some details. */
1549 	db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1550 	*ownerp = sx_xholder(sx);
1551 	if (sx->sx_lock & SX_LOCK_SHARED)
1552 		db_printf("SLOCK (count %ju)\n",
1553 		    (uintmax_t)SX_SHARERS(sx->sx_lock));
1554 	else
1555 		db_printf("XLOCK\n");
1556 	return (1);
1557 }
1558 #endif
1559