xref: /freebsd/sys/kern/kern_rwlock.c (revision 6463b6b59152fb1695bbe0de78f6e2675c5a765a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * Machine independent bits of reader/writer lock implementation.
30  */
31 
32 #include <sys/cdefs.h>
33 #include "opt_ddb.h"
34 #include "opt_hwpmc_hooks.h"
35 #include "opt_no_adaptive_rwlocks.h"
36 
37 #include <sys/param.h>
38 #include <sys/kdb.h>
39 #include <sys/ktr.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/rwlock.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/systm.h>
49 #include <sys/turnstile.h>
50 
51 #include <machine/cpu.h>
52 
53 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
54 #define	ADAPTIVE_RWLOCKS
55 #endif
56 
57 #ifdef HWPMC_HOOKS
58 #include <sys/pmckern.h>
59 PMC_SOFT_DECLARE( , , lock, failed);
60 #endif
61 
62 /*
63  * Return the rwlock address when the lock cookie address is provided.
64  * This functionality assumes that struct rwlock* have a member named rw_lock.
65  */
66 #define	rwlock2rw(c)	(__containerof(c, struct rwlock, rw_lock))
67 
68 #ifdef DDB
69 #include <ddb/ddb.h>
70 
71 static void	db_show_rwlock(const struct lock_object *lock);
72 #endif
73 static void	assert_rw(const struct lock_object *lock, int what);
74 static void	lock_rw(struct lock_object *lock, uintptr_t how);
75 static int	trylock_rw(struct lock_object *lock, uintptr_t how);
76 #ifdef KDTRACE_HOOKS
77 static int	owner_rw(const struct lock_object *lock, struct thread **owner);
78 #endif
79 static uintptr_t unlock_rw(struct lock_object *lock);
80 
81 struct lock_class lock_class_rw = {
82 	.lc_name = "rw",
83 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
84 	.lc_assert = assert_rw,
85 #ifdef DDB
86 	.lc_ddb_show = db_show_rwlock,
87 #endif
88 	.lc_lock = lock_rw,
89 	.lc_trylock = trylock_rw,
90 	.lc_unlock = unlock_rw,
91 #ifdef KDTRACE_HOOKS
92 	.lc_owner = owner_rw,
93 #endif
94 };
95 
96 #ifdef ADAPTIVE_RWLOCKS
97 #ifdef RWLOCK_CUSTOM_BACKOFF
98 static u_short __read_frequently rowner_retries;
99 static u_short __read_frequently rowner_loops;
100 static SYSCTL_NODE(_debug, OID_AUTO, rwlock,
101     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
102     "rwlock debugging");
103 SYSCTL_U16(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
104 SYSCTL_U16(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
105 
106 static struct lock_delay_config __read_frequently rw_delay;
107 
108 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base,
109     0, "");
110 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
111     0, "");
112 
113 static void
114 rw_lock_delay_init(void *arg __unused)
115 {
116 
117 	lock_delay_default_init(&rw_delay);
118 	rowner_retries = 10;
119 	rowner_loops = max(10000, rw_delay.max);
120 }
121 LOCK_DELAY_SYSINIT(rw_lock_delay_init);
122 #else
123 #define rw_delay	locks_delay
124 #define rowner_retries	locks_delay_retries
125 #define rowner_loops	locks_delay_loops
126 #endif
127 #endif
128 
129 /*
130  * Return a pointer to the owning thread if the lock is write-locked or
131  * NULL if the lock is unlocked or read-locked.
132  */
133 
134 #define	lv_rw_wowner(v)							\
135 	((v) & RW_LOCK_READ ? NULL :					\
136 	 (struct thread *)RW_OWNER((v)))
137 
138 #define	rw_wowner(rw)	lv_rw_wowner(RW_READ_VALUE(rw))
139 
140 /*
141  * Returns if a write owner is recursed.  Write ownership is not assured
142  * here and should be previously checked.
143  */
144 #define	rw_recursed(rw)		((rw)->rw_recurse != 0)
145 
146 /*
147  * Return true if curthread helds the lock.
148  */
149 #define	rw_wlocked(rw)		(rw_wowner((rw)) == curthread)
150 
151 /*
152  * Return a pointer to the owning thread for this lock who should receive
153  * any priority lent by threads that block on this lock.  Currently this
154  * is identical to rw_wowner().
155  */
156 #define	rw_owner(rw)		rw_wowner(rw)
157 
158 #ifndef INVARIANTS
159 #define	__rw_assert(c, what, file, line)
160 #endif
161 
162 static void
163 assert_rw(const struct lock_object *lock, int what)
164 {
165 
166 	rw_assert((const struct rwlock *)lock, what);
167 }
168 
169 static void
170 lock_rw(struct lock_object *lock, uintptr_t how)
171 {
172 	struct rwlock *rw;
173 
174 	rw = (struct rwlock *)lock;
175 	if (how)
176 		rw_rlock(rw);
177 	else
178 		rw_wlock(rw);
179 }
180 
181 static int
182 trylock_rw(struct lock_object *lock, uintptr_t how)
183 {
184 	struct rwlock *rw;
185 
186 	rw = (struct rwlock *)lock;
187 	if (how)
188 		return (rw_try_rlock(rw));
189 	else
190 		return (rw_try_wlock(rw));
191 }
192 
193 static uintptr_t
194 unlock_rw(struct lock_object *lock)
195 {
196 	struct rwlock *rw;
197 
198 	rw = (struct rwlock *)lock;
199 	rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
200 	if (rw->rw_lock & RW_LOCK_READ) {
201 		rw_runlock(rw);
202 		return (1);
203 	} else {
204 		rw_wunlock(rw);
205 		return (0);
206 	}
207 }
208 
209 #ifdef KDTRACE_HOOKS
210 static int
211 owner_rw(const struct lock_object *lock, struct thread **owner)
212 {
213 	const struct rwlock *rw = (const struct rwlock *)lock;
214 	uintptr_t x = rw->rw_lock;
215 
216 	*owner = rw_wowner(rw);
217 	return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
218 	    (*owner != NULL));
219 }
220 #endif
221 
222 void
223 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
224 {
225 	struct rwlock *rw;
226 	int flags;
227 
228 	rw = rwlock2rw(c);
229 
230 	MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
231 	    RW_RECURSE | RW_NEW)) == 0);
232 	ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
233 	    ("%s: rw_lock not aligned for %s: %p", __func__, name,
234 	    &rw->rw_lock));
235 
236 	flags = LO_UPGRADABLE;
237 	if (opts & RW_DUPOK)
238 		flags |= LO_DUPOK;
239 	if (opts & RW_NOPROFILE)
240 		flags |= LO_NOPROFILE;
241 	if (!(opts & RW_NOWITNESS))
242 		flags |= LO_WITNESS;
243 	if (opts & RW_RECURSE)
244 		flags |= LO_RECURSABLE;
245 	if (opts & RW_QUIET)
246 		flags |= LO_QUIET;
247 	if (opts & RW_NEW)
248 		flags |= LO_NEW;
249 
250 	lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
251 	rw->rw_lock = RW_UNLOCKED;
252 	rw->rw_recurse = 0;
253 }
254 
255 void
256 _rw_destroy(volatile uintptr_t *c)
257 {
258 	struct rwlock *rw;
259 
260 	rw = rwlock2rw(c);
261 
262 	KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
263 	KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
264 	rw->rw_lock = RW_DESTROYED;
265 	lock_destroy(&rw->lock_object);
266 }
267 
268 void
269 rw_sysinit(void *arg)
270 {
271 	struct rw_args *args;
272 
273 	args = arg;
274 	rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
275 	    args->ra_flags);
276 }
277 
278 int
279 _rw_wowned(const volatile uintptr_t *c)
280 {
281 
282 	return (rw_wowner(rwlock2rw(c)) == curthread);
283 }
284 
285 void
286 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
287 {
288 	struct rwlock *rw;
289 	uintptr_t tid, v;
290 
291 	rw = rwlock2rw(c);
292 
293 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
294 	    !TD_IS_IDLETHREAD(curthread),
295 	    ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
296 	    curthread, rw->lock_object.lo_name, file, line));
297 	KASSERT(rw->rw_lock != RW_DESTROYED,
298 	    ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
299 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
300 	    line, NULL);
301 	tid = (uintptr_t)curthread;
302 	v = RW_UNLOCKED;
303 	if (!_rw_write_lock_fetch(rw, &v, tid))
304 		_rw_wlock_hard(rw, v, file, line);
305 	else
306 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
307 		    0, 0, file, line, LOCKSTAT_WRITER);
308 
309 	LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
310 	WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
311 	TD_LOCKS_INC(curthread);
312 }
313 
314 int
315 __rw_try_wlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
316 {
317 	struct thread *td;
318 	uintptr_t tid, v;
319 	int rval;
320 	bool recursed;
321 
322 	td = curthread;
323 	tid = (uintptr_t)td;
324 	if (SCHEDULER_STOPPED())
325 		return (1);
326 
327 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
328 	    ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
329 	    curthread, rw->lock_object.lo_name, file, line));
330 	KASSERT(rw->rw_lock != RW_DESTROYED,
331 	    ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
332 
333 	rval = 1;
334 	recursed = false;
335 	v = RW_UNLOCKED;
336 	for (;;) {
337 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
338 			break;
339 		if (v == RW_UNLOCKED)
340 			continue;
341 		if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
342 			rw->rw_recurse++;
343 			atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
344 			break;
345 		}
346 		rval = 0;
347 		break;
348 	}
349 
350 	LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
351 	if (rval) {
352 		WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
353 		    file, line);
354 		if (!recursed)
355 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
356 			    rw, 0, 0, file, line, LOCKSTAT_WRITER);
357 		TD_LOCKS_INC(curthread);
358 	}
359 	return (rval);
360 }
361 
362 int
363 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
364 {
365 	struct rwlock *rw;
366 
367 	rw = rwlock2rw(c);
368 	return (__rw_try_wlock_int(rw LOCK_FILE_LINE_ARG));
369 }
370 
371 void
372 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
373 {
374 	struct rwlock *rw;
375 
376 	rw = rwlock2rw(c);
377 
378 	KASSERT(rw->rw_lock != RW_DESTROYED,
379 	    ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
380 	__rw_assert(c, RA_WLOCKED, file, line);
381 	WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
382 	LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
383 	    line);
384 
385 #ifdef LOCK_PROFILING
386 	_rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
387 #else
388 	__rw_wunlock(rw, curthread, file, line);
389 #endif
390 
391 	TD_LOCKS_DEC(curthread);
392 }
393 
394 /*
395  * Determines whether a new reader can acquire a lock.  Succeeds if the
396  * reader already owns a read lock and the lock is locked for read to
397  * prevent deadlock from reader recursion.  Also succeeds if the lock
398  * is unlocked and has no writer waiters or spinners.  Failing otherwise
399  * prioritizes writers before readers.
400  */
401 static __always_inline bool
402 __rw_can_read(struct thread *td, uintptr_t v, bool fp)
403 {
404 
405 	if ((v & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER))
406 	    == RW_LOCK_READ)
407 		return (true);
408 	if (!fp && td->td_rw_rlocks && (v & RW_LOCK_READ))
409 		return (true);
410 	return (false);
411 }
412 
413 static __always_inline bool
414 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, bool fp
415     LOCK_FILE_LINE_ARG_DEF)
416 {
417 
418 	/*
419 	 * Handle the easy case.  If no other thread has a write
420 	 * lock, then try to bump up the count of read locks.  Note
421 	 * that we have to preserve the current state of the
422 	 * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
423 	 * read lock, then rw_lock must have changed, so restart
424 	 * the loop.  Note that this handles the case of a
425 	 * completely unlocked rwlock since such a lock is encoded
426 	 * as a read lock with no waiters.
427 	 */
428 	while (__rw_can_read(td, *vp, fp)) {
429 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
430 			*vp + RW_ONE_READER)) {
431 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
432 				CTR4(KTR_LOCK,
433 				    "%s: %p succeed %p -> %p", __func__,
434 				    rw, (void *)*vp,
435 				    (void *)(*vp + RW_ONE_READER));
436 			td->td_rw_rlocks++;
437 			return (true);
438 		}
439 	}
440 	return (false);
441 }
442 
443 static void __noinline
444 __rw_rlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
445     LOCK_FILE_LINE_ARG_DEF)
446 {
447 	struct turnstile *ts;
448 	struct thread *owner;
449 #ifdef ADAPTIVE_RWLOCKS
450 	int spintries = 0;
451 	int i, n;
452 #endif
453 #ifdef LOCK_PROFILING
454 	uint64_t waittime = 0;
455 	int contested = 0;
456 #endif
457 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
458 	struct lock_delay_arg lda;
459 #endif
460 #ifdef KDTRACE_HOOKS
461 	u_int sleep_cnt = 0;
462 	int64_t sleep_time = 0;
463 	int64_t all_time = 0;
464 #endif
465 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
466 	uintptr_t state = 0;
467 	int doing_lockprof = 0;
468 #endif
469 
470 #ifdef KDTRACE_HOOKS
471 	if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
472 		if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
473 			goto out_lockstat;
474 		doing_lockprof = 1;
475 		all_time -= lockstat_nsecs(&rw->lock_object);
476 		state = v;
477 	}
478 #endif
479 #ifdef LOCK_PROFILING
480 	doing_lockprof = 1;
481 	state = v;
482 #endif
483 
484 	if (SCHEDULER_STOPPED())
485 		return;
486 
487 #if defined(ADAPTIVE_RWLOCKS)
488 	lock_delay_arg_init(&lda, &rw_delay);
489 #elif defined(KDTRACE_HOOKS)
490 	lock_delay_arg_init_noadapt(&lda);
491 #endif
492 
493 #ifdef HWPMC_HOOKS
494 	PMC_SOFT_CALL( , , lock, failed);
495 #endif
496 	lock_profile_obtain_lock_failed(&rw->lock_object, false,
497 	    &contested, &waittime);
498 
499 	THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
500 
501 	for (;;) {
502 		if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
503 			break;
504 #ifdef KDTRACE_HOOKS
505 		lda.spin_cnt++;
506 #endif
507 
508 #ifdef ADAPTIVE_RWLOCKS
509 		/*
510 		 * If the owner is running on another CPU, spin until
511 		 * the owner stops running or the state of the lock
512 		 * changes.
513 		 */
514 		if ((v & RW_LOCK_READ) == 0) {
515 			owner = (struct thread *)RW_OWNER(v);
516 			if (TD_IS_RUNNING(owner)) {
517 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
518 					CTR3(KTR_LOCK,
519 					    "%s: spinning on %p held by %p",
520 					    __func__, rw, owner);
521 				KTR_STATE1(KTR_SCHED, "thread",
522 				    sched_tdname(curthread), "spinning",
523 				    "lockname:\"%s\"", rw->lock_object.lo_name);
524 				do {
525 					lock_delay(&lda);
526 					v = RW_READ_VALUE(rw);
527 					owner = lv_rw_wowner(v);
528 				} while (owner != NULL && TD_IS_RUNNING(owner));
529 				KTR_STATE0(KTR_SCHED, "thread",
530 				    sched_tdname(curthread), "running");
531 				continue;
532 			}
533 		} else {
534 			if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
535 				MPASS(!__rw_can_read(td, v, false));
536 				lock_delay_spin(2);
537 				v = RW_READ_VALUE(rw);
538 				continue;
539 			}
540 			if (spintries < rowner_retries) {
541 				spintries++;
542 				KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
543 				    "spinning", "lockname:\"%s\"",
544 				    rw->lock_object.lo_name);
545 				n = RW_READERS(v);
546 				for (i = 0; i < rowner_loops; i += n) {
547 					lock_delay_spin(n);
548 					v = RW_READ_VALUE(rw);
549 					if (!(v & RW_LOCK_READ))
550 						break;
551 					n = RW_READERS(v);
552 					if (n == 0)
553 						break;
554 					if (__rw_can_read(td, v, false))
555 						break;
556 				}
557 #ifdef KDTRACE_HOOKS
558 				lda.spin_cnt += rowner_loops - i;
559 #endif
560 				KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
561 				    "running");
562 				if (i < rowner_loops)
563 					continue;
564 			}
565 		}
566 #endif
567 
568 		/*
569 		 * Okay, now it's the hard case.  Some other thread already
570 		 * has a write lock or there are write waiters present,
571 		 * acquire the turnstile lock so we can begin the process
572 		 * of blocking.
573 		 */
574 		ts = turnstile_trywait(&rw->lock_object);
575 
576 		/*
577 		 * The lock might have been released while we spun, so
578 		 * recheck its state and restart the loop if needed.
579 		 */
580 		v = RW_READ_VALUE(rw);
581 retry_ts:
582 		if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
583 		    __rw_can_read(td, v, false)) {
584 			turnstile_cancel(ts);
585 			continue;
586 		}
587 
588 		owner = lv_rw_wowner(v);
589 
590 #ifdef ADAPTIVE_RWLOCKS
591 		/*
592 		 * The current lock owner might have started executing
593 		 * on another CPU (or the lock could have changed
594 		 * owners) while we were waiting on the turnstile
595 		 * chain lock.  If so, drop the turnstile lock and try
596 		 * again.
597 		 */
598 		if (owner != NULL) {
599 			if (TD_IS_RUNNING(owner)) {
600 				turnstile_cancel(ts);
601 				continue;
602 			}
603 		}
604 #endif
605 
606 		/*
607 		 * The lock is held in write mode or it already has waiters.
608 		 */
609 		MPASS(!__rw_can_read(td, v, false));
610 
611 		/*
612 		 * If the RW_LOCK_READ_WAITERS flag is already set, then
613 		 * we can go ahead and block.  If it is not set then try
614 		 * to set it.  If we fail to set it drop the turnstile
615 		 * lock and restart the loop.
616 		 */
617 		if (!(v & RW_LOCK_READ_WAITERS)) {
618 			if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
619 			    v | RW_LOCK_READ_WAITERS))
620 				goto retry_ts;
621 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
622 				CTR2(KTR_LOCK, "%s: %p set read waiters flag",
623 				    __func__, rw);
624 		}
625 
626 		/*
627 		 * We were unable to acquire the lock and the read waiters
628 		 * flag is set, so we must block on the turnstile.
629 		 */
630 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
631 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
632 			    rw);
633 #ifdef KDTRACE_HOOKS
634 		sleep_time -= lockstat_nsecs(&rw->lock_object);
635 #endif
636 		MPASS(owner == rw_owner(rw));
637 		turnstile_wait(ts, owner, TS_SHARED_QUEUE);
638 #ifdef KDTRACE_HOOKS
639 		sleep_time += lockstat_nsecs(&rw->lock_object);
640 		sleep_cnt++;
641 #endif
642 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
643 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
644 			    __func__, rw);
645 		v = RW_READ_VALUE(rw);
646 	}
647 	THREAD_CONTENTION_DONE(&rw->lock_object);
648 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
649 	if (__predict_true(!doing_lockprof))
650 		return;
651 #endif
652 #ifdef KDTRACE_HOOKS
653 	all_time += lockstat_nsecs(&rw->lock_object);
654 	if (sleep_time)
655 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
656 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
657 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
658 
659 	/* Record only the loops spinning and not sleeping. */
660 	if (lda.spin_cnt > sleep_cnt)
661 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
662 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
663 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
664 out_lockstat:
665 #endif
666 	/*
667 	 * TODO: acquire "owner of record" here.  Here be turnstile dragons
668 	 * however.  turnstiles don't like owners changing between calls to
669 	 * turnstile_wait() currently.
670 	 */
671 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
672 	    waittime, file, line, LOCKSTAT_READER);
673 }
674 
675 void
676 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
677 {
678 	struct thread *td;
679 	uintptr_t v;
680 
681 	td = curthread;
682 
683 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
684 	    !TD_IS_IDLETHREAD(td),
685 	    ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
686 	    td, rw->lock_object.lo_name, file, line));
687 	KASSERT(rw->rw_lock != RW_DESTROYED,
688 	    ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
689 	KASSERT(rw_wowner(rw) != td,
690 	    ("rw_rlock: wlock already held for %s @ %s:%d",
691 	    rw->lock_object.lo_name, file, line));
692 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
693 
694 	v = RW_READ_VALUE(rw);
695 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
696 	    !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
697 		__rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
698 	else
699 		lock_profile_obtain_lock_success(&rw->lock_object, false, 0, 0,
700 		    file, line);
701 
702 	LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
703 	WITNESS_LOCK(&rw->lock_object, 0, file, line);
704 	TD_LOCKS_INC(curthread);
705 }
706 
707 void
708 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
709 {
710 	struct rwlock *rw;
711 
712 	rw = rwlock2rw(c);
713 	__rw_rlock_int(rw LOCK_FILE_LINE_ARG);
714 }
715 
716 int
717 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
718 {
719 	uintptr_t x;
720 
721 	if (SCHEDULER_STOPPED())
722 		return (1);
723 
724 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
725 	    ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
726 	    curthread, rw->lock_object.lo_name, file, line));
727 
728 	x = rw->rw_lock;
729 	for (;;) {
730 		KASSERT(rw->rw_lock != RW_DESTROYED,
731 		    ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
732 		if (!(x & RW_LOCK_READ))
733 			break;
734 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
735 			LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
736 			    line);
737 			WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
738 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
739 			    rw, 0, 0, file, line, LOCKSTAT_READER);
740 			TD_LOCKS_INC(curthread);
741 			curthread->td_rw_rlocks++;
742 			return (1);
743 		}
744 	}
745 
746 	LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
747 	return (0);
748 }
749 
750 int
751 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
752 {
753 	struct rwlock *rw;
754 
755 	rw = rwlock2rw(c);
756 	return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
757 }
758 
759 static __always_inline bool
760 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
761 {
762 
763 	for (;;) {
764 		if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
765 			if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
766 			    *vp - RW_ONE_READER)) {
767 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
768 					CTR4(KTR_LOCK,
769 					    "%s: %p succeeded %p -> %p",
770 					    __func__, rw, (void *)*vp,
771 					    (void *)(*vp - RW_ONE_READER));
772 				td->td_rw_rlocks--;
773 				return (true);
774 			}
775 			continue;
776 		}
777 		break;
778 	}
779 	return (false);
780 }
781 
782 static void __noinline
783 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
784     LOCK_FILE_LINE_ARG_DEF)
785 {
786 	struct turnstile *ts;
787 	uintptr_t setv, passedv, queue;
788 
789 	if (SCHEDULER_STOPPED())
790 		return;
791 
792 	passedv = v;
793 	if (__rw_runlock_try(rw, td, &v))
794 		goto out_lockstat;
795 
796 	/*
797 	 * Ok, we know we have waiters and we think we are the
798 	 * last reader, so grab the turnstile lock.
799 	 */
800 	turnstile_chain_lock(&rw->lock_object);
801 	v = RW_READ_VALUE(rw);
802 	for (;;) {
803 		if (__rw_runlock_try(rw, td, &v))
804 			break;
805 
806 		MPASS(v & RW_LOCK_WAITERS);
807 
808 		/*
809 		 * Try to drop our lock leaving the lock in a unlocked
810 		 * state.
811 		 *
812 		 * If you wanted to do explicit lock handoff you'd have to
813 		 * do it here.  You'd also want to use turnstile_signal()
814 		 * and you'd have to handle the race where a higher
815 		 * priority thread blocks on the write lock before the
816 		 * thread you wakeup actually runs and have the new thread
817 		 * "steal" the lock.  For now it's a lot simpler to just
818 		 * wakeup all of the waiters.
819 		 *
820 		 * As above, if we fail, then another thread might have
821 		 * acquired a read lock, so drop the turnstile lock and
822 		 * restart.
823 		 */
824 		setv = RW_UNLOCKED;
825 		queue = TS_SHARED_QUEUE;
826 		if (v & RW_LOCK_WRITE_WAITERS) {
827 			queue = TS_EXCLUSIVE_QUEUE;
828 			setv |= (v & RW_LOCK_READ_WAITERS);
829 		}
830 		setv |= (v & RW_LOCK_WRITE_SPINNER);
831 		if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
832 			continue;
833 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
834 			CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
835 			    __func__, rw);
836 
837 		/*
838 		 * Ok.  The lock is released and all that's left is to
839 		 * wake up the waiters.  Note that the lock might not be
840 		 * free anymore, but in that case the writers will just
841 		 * block again if they run before the new lock holder(s)
842 		 * release the lock.
843 		 */
844 		ts = turnstile_lookup(&rw->lock_object);
845 		if (__predict_false(ts == NULL)) {
846 			panic("got NULL turnstile on rwlock %p passedv %p v %p",
847 			    rw, (void *)passedv, (void *)v);
848 		}
849 		turnstile_broadcast(ts, queue);
850 		turnstile_unpend(ts);
851 		td->td_rw_rlocks--;
852 		break;
853 	}
854 	turnstile_chain_unlock(&rw->lock_object);
855 out_lockstat:
856 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
857 }
858 
859 void
860 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
861 {
862 	struct thread *td;
863 	uintptr_t v;
864 
865 	KASSERT(rw->rw_lock != RW_DESTROYED,
866 	    ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
867 	__rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
868 	WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
869 	LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
870 
871 	td = curthread;
872 	v = RW_READ_VALUE(rw);
873 
874 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
875 	    !__rw_runlock_try(rw, td, &v)))
876 		__rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
877 	else
878 		lock_profile_release_lock(&rw->lock_object, false);
879 
880 	TD_LOCKS_DEC(curthread);
881 }
882 
883 void
884 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
885 {
886 	struct rwlock *rw;
887 
888 	rw = rwlock2rw(c);
889 	_rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
890 }
891 
892 #ifdef ADAPTIVE_RWLOCKS
893 static inline void
894 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
895 {
896 
897 	if (v & RW_LOCK_WRITE_SPINNER)
898 		return;
899 	if (*in_critical) {
900 		critical_exit();
901 		*in_critical = false;
902 		(*extra_work)--;
903 	}
904 }
905 #else
906 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
907 #endif
908 
909 /*
910  * This function is called when we are unable to obtain a write lock on the
911  * first try.  This means that at least one other thread holds either a
912  * read or write lock.
913  */
914 void
915 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
916 {
917 	uintptr_t tid;
918 	struct rwlock *rw;
919 	struct turnstile *ts;
920 	struct thread *owner;
921 #ifdef ADAPTIVE_RWLOCKS
922 	int spintries = 0;
923 	int i, n;
924 	enum { READERS, WRITER } sleep_reason = READERS;
925 	bool in_critical = false;
926 #endif
927 	uintptr_t setv;
928 #ifdef LOCK_PROFILING
929 	uint64_t waittime = 0;
930 	int contested = 0;
931 #endif
932 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
933 	struct lock_delay_arg lda;
934 #endif
935 #ifdef KDTRACE_HOOKS
936 	u_int sleep_cnt = 0;
937 	int64_t sleep_time = 0;
938 	int64_t all_time = 0;
939 #endif
940 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
941 	uintptr_t state = 0;
942 	int doing_lockprof = 0;
943 #endif
944 	int extra_work = 0;
945 
946 	tid = (uintptr_t)curthread;
947 	rw = rwlock2rw(c);
948 
949 #ifdef KDTRACE_HOOKS
950 	if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
951 		while (v == RW_UNLOCKED) {
952 			if (_rw_write_lock_fetch(rw, &v, tid))
953 				goto out_lockstat;
954 		}
955 		extra_work = 1;
956 		doing_lockprof = 1;
957 		all_time -= lockstat_nsecs(&rw->lock_object);
958 		state = v;
959 	}
960 #endif
961 #ifdef LOCK_PROFILING
962 	extra_work = 1;
963 	doing_lockprof = 1;
964 	state = v;
965 #endif
966 
967 	if (SCHEDULER_STOPPED())
968 		return;
969 
970 	if (__predict_false(v == RW_UNLOCKED))
971 		v = RW_READ_VALUE(rw);
972 
973 	if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
974 		KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
975 		    ("%s: recursing but non-recursive rw %s @ %s:%d\n",
976 		    __func__, rw->lock_object.lo_name, file, line));
977 		rw->rw_recurse++;
978 		atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
979 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
980 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
981 		return;
982 	}
983 
984 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
985 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
986 		    rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
987 
988 #if defined(ADAPTIVE_RWLOCKS)
989 	lock_delay_arg_init(&lda, &rw_delay);
990 #elif defined(KDTRACE_HOOKS)
991 	lock_delay_arg_init_noadapt(&lda);
992 #endif
993 
994 #ifdef HWPMC_HOOKS
995 	PMC_SOFT_CALL( , , lock, failed);
996 #endif
997 	lock_profile_obtain_lock_failed(&rw->lock_object, false,
998 	    &contested, &waittime);
999 
1000 	THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
1001 
1002 	for (;;) {
1003 		if (v == RW_UNLOCKED) {
1004 			if (_rw_write_lock_fetch(rw, &v, tid))
1005 				break;
1006 			continue;
1007 		}
1008 #ifdef KDTRACE_HOOKS
1009 		lda.spin_cnt++;
1010 #endif
1011 
1012 #ifdef ADAPTIVE_RWLOCKS
1013 		if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
1014 			if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
1015 				break;
1016 			continue;
1017 		}
1018 
1019 		/*
1020 		 * If the lock is write locked and the owner is
1021 		 * running on another CPU, spin until the owner stops
1022 		 * running or the state of the lock changes.
1023 		 */
1024 		if (!(v & RW_LOCK_READ)) {
1025 			rw_drop_critical(v, &in_critical, &extra_work);
1026 			sleep_reason = WRITER;
1027 			owner = lv_rw_wowner(v);
1028 			if (!TD_IS_RUNNING(owner))
1029 				goto ts;
1030 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
1031 				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1032 				    __func__, rw, owner);
1033 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1034 			    "spinning", "lockname:\"%s\"",
1035 			    rw->lock_object.lo_name);
1036 			do {
1037 				lock_delay(&lda);
1038 				v = RW_READ_VALUE(rw);
1039 				owner = lv_rw_wowner(v);
1040 			} while (owner != NULL && TD_IS_RUNNING(owner));
1041 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1042 			    "running");
1043 			continue;
1044 		} else if (RW_READERS(v) > 0) {
1045 			sleep_reason = READERS;
1046 			if (spintries == rowner_retries)
1047 				goto ts;
1048 			if (!(v & RW_LOCK_WRITE_SPINNER)) {
1049 				if (!in_critical) {
1050 					critical_enter();
1051 					in_critical = true;
1052 					extra_work++;
1053 				}
1054 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1055 				    v | RW_LOCK_WRITE_SPINNER)) {
1056 					critical_exit();
1057 					in_critical = false;
1058 					extra_work--;
1059 					continue;
1060 				}
1061 			}
1062 			spintries++;
1063 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1064 			    "spinning", "lockname:\"%s\"",
1065 			    rw->lock_object.lo_name);
1066 			n = RW_READERS(v);
1067 			for (i = 0; i < rowner_loops; i += n) {
1068 				lock_delay_spin(n);
1069 				v = RW_READ_VALUE(rw);
1070 				if (!(v & RW_LOCK_WRITE_SPINNER))
1071 					break;
1072 				if (!(v & RW_LOCK_READ))
1073 					break;
1074 				n = RW_READERS(v);
1075 				if (n == 0)
1076 					break;
1077 			}
1078 #ifdef KDTRACE_HOOKS
1079 			lda.spin_cnt += i;
1080 #endif
1081 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1082 			    "running");
1083 			if (i < rowner_loops)
1084 				continue;
1085 		}
1086 ts:
1087 #endif
1088 		ts = turnstile_trywait(&rw->lock_object);
1089 		v = RW_READ_VALUE(rw);
1090 retry_ts:
1091 		owner = lv_rw_wowner(v);
1092 
1093 #ifdef ADAPTIVE_RWLOCKS
1094 		/*
1095 		 * The current lock owner might have started executing
1096 		 * on another CPU (or the lock could have changed
1097 		 * owners) while we were waiting on the turnstile
1098 		 * chain lock.  If so, drop the turnstile lock and try
1099 		 * again.
1100 		 */
1101 		if (owner != NULL) {
1102 			if (TD_IS_RUNNING(owner)) {
1103 				turnstile_cancel(ts);
1104 				rw_drop_critical(v, &in_critical, &extra_work);
1105 				continue;
1106 			}
1107 		} else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1108 			turnstile_cancel(ts);
1109 			rw_drop_critical(v, &in_critical, &extra_work);
1110 			continue;
1111 		}
1112 #endif
1113 		/*
1114 		 * Check for the waiters flags about this rwlock.
1115 		 * If the lock was released, without maintain any pending
1116 		 * waiters queue, simply try to acquire it.
1117 		 * If a pending waiters queue is present, claim the lock
1118 		 * ownership and maintain the pending queue.
1119 		 */
1120 		setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1121 		if ((v & ~setv) == RW_UNLOCKED) {
1122 			setv &= ~RW_LOCK_WRITE_SPINNER;
1123 			if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1124 				if (setv)
1125 					turnstile_claim(ts);
1126 				else
1127 					turnstile_cancel(ts);
1128 				break;
1129 			}
1130 			goto retry_ts;
1131 		}
1132 
1133 #ifdef ADAPTIVE_RWLOCKS
1134 		if (in_critical) {
1135 			if ((v & RW_LOCK_WRITE_SPINNER) ||
1136 			    !((v & RW_LOCK_WRITE_WAITERS))) {
1137 				setv = v & ~RW_LOCK_WRITE_SPINNER;
1138 				setv |= RW_LOCK_WRITE_WAITERS;
1139 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1140 					goto retry_ts;
1141 			}
1142 			critical_exit();
1143 			in_critical = false;
1144 			extra_work--;
1145 		} else {
1146 #endif
1147 			/*
1148 			 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1149 			 * set it.  If we fail to set it, then loop back and try
1150 			 * again.
1151 			 */
1152 			if (!(v & RW_LOCK_WRITE_WAITERS)) {
1153 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1154 				    v | RW_LOCK_WRITE_WAITERS))
1155 					goto retry_ts;
1156 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
1157 					CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1158 					    __func__, rw);
1159 			}
1160 #ifdef ADAPTIVE_RWLOCKS
1161 		}
1162 #endif
1163 		/*
1164 		 * We were unable to acquire the lock and the write waiters
1165 		 * flag is set, so we must block on the turnstile.
1166 		 */
1167 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1168 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1169 			    rw);
1170 #ifdef KDTRACE_HOOKS
1171 		sleep_time -= lockstat_nsecs(&rw->lock_object);
1172 #endif
1173 		MPASS(owner == rw_owner(rw));
1174 		turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1175 #ifdef KDTRACE_HOOKS
1176 		sleep_time += lockstat_nsecs(&rw->lock_object);
1177 		sleep_cnt++;
1178 #endif
1179 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1180 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1181 			    __func__, rw);
1182 #ifdef ADAPTIVE_RWLOCKS
1183 		spintries = 0;
1184 #endif
1185 		v = RW_READ_VALUE(rw);
1186 	}
1187 	THREAD_CONTENTION_DONE(&rw->lock_object);
1188 	if (__predict_true(!extra_work))
1189 		return;
1190 #ifdef ADAPTIVE_RWLOCKS
1191 	if (in_critical)
1192 		critical_exit();
1193 #endif
1194 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1195 	if (__predict_true(!doing_lockprof))
1196 		return;
1197 #endif
1198 #ifdef KDTRACE_HOOKS
1199 	all_time += lockstat_nsecs(&rw->lock_object);
1200 	if (sleep_time)
1201 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1202 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1203 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1204 
1205 	/* Record only the loops spinning and not sleeping. */
1206 	if (lda.spin_cnt > sleep_cnt)
1207 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1208 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1209 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1210 out_lockstat:
1211 #endif
1212 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1213 	    waittime, file, line, LOCKSTAT_WRITER);
1214 }
1215 
1216 /*
1217  * This function is called if lockstat is active or the first try at releasing
1218  * a write lock failed.  The latter means that the lock is recursed or one of
1219  * the 2 waiter bits must be set indicating that at least one thread is waiting
1220  * on this lock.
1221  */
1222 void
1223 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1224 {
1225 	struct rwlock *rw;
1226 	struct turnstile *ts;
1227 	uintptr_t tid, setv, passedv;
1228 	int queue;
1229 
1230 	tid = (uintptr_t)curthread;
1231 	if (SCHEDULER_STOPPED())
1232 		return;
1233 
1234 	rw = rwlock2rw(c);
1235 	if (__predict_false(v == tid))
1236 		v = RW_READ_VALUE(rw);
1237 
1238 	if (v & RW_LOCK_WRITER_RECURSED) {
1239 		if (--(rw->rw_recurse) == 0)
1240 			atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1241 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1242 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1243 		return;
1244 	}
1245 
1246 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1247 	if (v == tid && _rw_write_unlock(rw, tid))
1248 		return;
1249 
1250 	KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1251 	    ("%s: neither of the waiter flags are set", __func__));
1252 
1253 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1254 		CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1255 
1256 	turnstile_chain_lock(&rw->lock_object);
1257 
1258 	/*
1259 	 * Use the same algo as sx locks for now.  Prefer waking up shared
1260 	 * waiters if we have any over writers.  This is probably not ideal.
1261 	 *
1262 	 * 'v' is the value we are going to write back to rw_lock.  If we
1263 	 * have waiters on both queues, we need to preserve the state of
1264 	 * the waiter flag for the queue we don't wake up.  For now this is
1265 	 * hardcoded for the algorithm mentioned above.
1266 	 *
1267 	 * In the case of both readers and writers waiting we wakeup the
1268 	 * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1269 	 * new writer comes in before a reader it will claim the lock up
1270 	 * above.  There is probably a potential priority inversion in
1271 	 * there that could be worked around either by waking both queues
1272 	 * of waiters or doing some complicated lock handoff gymnastics.
1273 	 */
1274 	setv = RW_UNLOCKED;
1275 	passedv = v;
1276 	v = RW_READ_VALUE(rw);
1277 	queue = TS_SHARED_QUEUE;
1278 	if (v & RW_LOCK_WRITE_WAITERS) {
1279 		queue = TS_EXCLUSIVE_QUEUE;
1280 		setv |= (v & RW_LOCK_READ_WAITERS);
1281 	}
1282 	atomic_store_rel_ptr(&rw->rw_lock, setv);
1283 
1284 	/* Wake up all waiters for the specific queue. */
1285 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1286 		CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1287 		    queue == TS_SHARED_QUEUE ? "read" : "write");
1288 
1289 	ts = turnstile_lookup(&rw->lock_object);
1290 	if (__predict_false(ts == NULL)) {
1291 		panic("got NULL turnstile on rwlock %p passedv %p v %p", rw,
1292 		    (void *)passedv, (void *)v);
1293 	}
1294 	turnstile_broadcast(ts, queue);
1295 	turnstile_unpend(ts);
1296 	turnstile_chain_unlock(&rw->lock_object);
1297 }
1298 
1299 /*
1300  * Attempt to do a non-blocking upgrade from a read lock to a write
1301  * lock.  This will only succeed if this thread holds a single read
1302  * lock.  Returns true if the upgrade succeeded and false otherwise.
1303  */
1304 int
1305 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1306 {
1307 	uintptr_t v, setv, tid;
1308 	struct turnstile *ts;
1309 	int success;
1310 
1311 	if (SCHEDULER_STOPPED())
1312 		return (1);
1313 
1314 	KASSERT(rw->rw_lock != RW_DESTROYED,
1315 	    ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1316 	__rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1317 
1318 	/*
1319 	 * Attempt to switch from one reader to a writer.  If there
1320 	 * are any write waiters, then we will have to lock the
1321 	 * turnstile first to prevent races with another writer
1322 	 * calling turnstile_wait() before we have claimed this
1323 	 * turnstile.  So, do the simple case of no waiters first.
1324 	 */
1325 	tid = (uintptr_t)curthread;
1326 	success = 0;
1327 	v = RW_READ_VALUE(rw);
1328 	for (;;) {
1329 		if (RW_READERS(v) > 1)
1330 			break;
1331 		if (!(v & RW_LOCK_WAITERS)) {
1332 			success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1333 			if (!success)
1334 				continue;
1335 			break;
1336 		}
1337 
1338 		/*
1339 		 * Ok, we think we have waiters, so lock the turnstile.
1340 		 */
1341 		ts = turnstile_trywait(&rw->lock_object);
1342 		v = RW_READ_VALUE(rw);
1343 retry_ts:
1344 		if (RW_READERS(v) > 1) {
1345 			turnstile_cancel(ts);
1346 			break;
1347 		}
1348 		/*
1349 		 * Try to switch from one reader to a writer again.  This time
1350 		 * we honor the current state of the waiters flags.
1351 		 * If we obtain the lock with the flags set, then claim
1352 		 * ownership of the turnstile.
1353 		 */
1354 		setv = tid | (v & RW_LOCK_WAITERS);
1355 		success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1356 		if (success) {
1357 			if (v & RW_LOCK_WAITERS)
1358 				turnstile_claim(ts);
1359 			else
1360 				turnstile_cancel(ts);
1361 			break;
1362 		}
1363 		goto retry_ts;
1364 	}
1365 	LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1366 	if (success) {
1367 		curthread->td_rw_rlocks--;
1368 		WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1369 		    file, line);
1370 		LOCKSTAT_RECORD0(rw__upgrade, rw);
1371 	}
1372 	return (success);
1373 }
1374 
1375 int
1376 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1377 {
1378 	struct rwlock *rw;
1379 
1380 	rw = rwlock2rw(c);
1381 	return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1382 }
1383 
1384 /*
1385  * Downgrade a write lock into a single read lock.
1386  */
1387 void
1388 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1389 {
1390 	struct turnstile *ts;
1391 	uintptr_t tid, v;
1392 	int rwait, wwait;
1393 
1394 	if (SCHEDULER_STOPPED())
1395 		return;
1396 
1397 	KASSERT(rw->rw_lock != RW_DESTROYED,
1398 	    ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1399 	__rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1400 #ifndef INVARIANTS
1401 	if (rw_recursed(rw))
1402 		panic("downgrade of a recursed lock");
1403 #endif
1404 
1405 	WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1406 
1407 	/*
1408 	 * Convert from a writer to a single reader.  First we handle
1409 	 * the easy case with no waiters.  If there are any waiters, we
1410 	 * lock the turnstile and "disown" the lock.
1411 	 */
1412 	tid = (uintptr_t)curthread;
1413 	if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1414 		goto out;
1415 
1416 	/*
1417 	 * Ok, we think we have waiters, so lock the turnstile so we can
1418 	 * read the waiter flags without any races.
1419 	 */
1420 	turnstile_chain_lock(&rw->lock_object);
1421 	v = rw->rw_lock & RW_LOCK_WAITERS;
1422 	rwait = v & RW_LOCK_READ_WAITERS;
1423 	wwait = v & RW_LOCK_WRITE_WAITERS;
1424 	MPASS(rwait | wwait);
1425 
1426 	/*
1427 	 * Downgrade from a write lock while preserving waiters flag
1428 	 * and give up ownership of the turnstile.
1429 	 */
1430 	ts = turnstile_lookup(&rw->lock_object);
1431 	MPASS(ts != NULL);
1432 	if (!wwait)
1433 		v &= ~RW_LOCK_READ_WAITERS;
1434 	atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1435 	/*
1436 	 * Wake other readers if there are no writers pending.  Otherwise they
1437 	 * won't be able to acquire the lock anyway.
1438 	 */
1439 	if (rwait && !wwait) {
1440 		turnstile_broadcast(ts, TS_SHARED_QUEUE);
1441 		turnstile_unpend(ts);
1442 	} else
1443 		turnstile_disown(ts);
1444 	turnstile_chain_unlock(&rw->lock_object);
1445 out:
1446 	curthread->td_rw_rlocks++;
1447 	LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1448 	LOCKSTAT_RECORD0(rw__downgrade, rw);
1449 }
1450 
1451 void
1452 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1453 {
1454 	struct rwlock *rw;
1455 
1456 	rw = rwlock2rw(c);
1457 	__rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1458 }
1459 
1460 #ifdef INVARIANT_SUPPORT
1461 #ifndef INVARIANTS
1462 #undef __rw_assert
1463 #endif
1464 
1465 /*
1466  * In the non-WITNESS case, rw_assert() can only detect that at least
1467  * *some* thread owns an rlock, but it cannot guarantee that *this*
1468  * thread owns an rlock.
1469  */
1470 void
1471 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1472 {
1473 	const struct rwlock *rw;
1474 
1475 	if (SCHEDULER_STOPPED())
1476 		return;
1477 
1478 	rw = rwlock2rw(c);
1479 
1480 	switch (what) {
1481 	case RA_LOCKED:
1482 	case RA_LOCKED | RA_RECURSED:
1483 	case RA_LOCKED | RA_NOTRECURSED:
1484 	case RA_RLOCKED:
1485 	case RA_RLOCKED | RA_RECURSED:
1486 	case RA_RLOCKED | RA_NOTRECURSED:
1487 #ifdef WITNESS
1488 		witness_assert(&rw->lock_object, what, file, line);
1489 #else
1490 		/*
1491 		 * If some other thread has a write lock or we have one
1492 		 * and are asserting a read lock, fail.  Also, if no one
1493 		 * has a lock at all, fail.
1494 		 */
1495 		if (rw->rw_lock == RW_UNLOCKED ||
1496 		    (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1497 		    rw_wowner(rw) != curthread)))
1498 			panic("Lock %s not %slocked @ %s:%d\n",
1499 			    rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1500 			    "read " : "", file, line);
1501 
1502 		if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1503 			if (rw_recursed(rw)) {
1504 				if (what & RA_NOTRECURSED)
1505 					panic("Lock %s recursed @ %s:%d\n",
1506 					    rw->lock_object.lo_name, file,
1507 					    line);
1508 			} else if (what & RA_RECURSED)
1509 				panic("Lock %s not recursed @ %s:%d\n",
1510 				    rw->lock_object.lo_name, file, line);
1511 		}
1512 #endif
1513 		break;
1514 	case RA_WLOCKED:
1515 	case RA_WLOCKED | RA_RECURSED:
1516 	case RA_WLOCKED | RA_NOTRECURSED:
1517 		if (rw_wowner(rw) != curthread)
1518 			panic("Lock %s not exclusively locked @ %s:%d\n",
1519 			    rw->lock_object.lo_name, file, line);
1520 		if (rw_recursed(rw)) {
1521 			if (what & RA_NOTRECURSED)
1522 				panic("Lock %s recursed @ %s:%d\n",
1523 				    rw->lock_object.lo_name, file, line);
1524 		} else if (what & RA_RECURSED)
1525 			panic("Lock %s not recursed @ %s:%d\n",
1526 			    rw->lock_object.lo_name, file, line);
1527 		break;
1528 	case RA_UNLOCKED:
1529 #ifdef WITNESS
1530 		witness_assert(&rw->lock_object, what, file, line);
1531 #else
1532 		/*
1533 		 * If we hold a write lock fail.  We can't reliably check
1534 		 * to see if we hold a read lock or not.
1535 		 */
1536 		if (rw_wowner(rw) == curthread)
1537 			panic("Lock %s exclusively locked @ %s:%d\n",
1538 			    rw->lock_object.lo_name, file, line);
1539 #endif
1540 		break;
1541 	default:
1542 		panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1543 		    line);
1544 	}
1545 }
1546 #endif /* INVARIANT_SUPPORT */
1547 
1548 #ifdef DDB
1549 static void
1550 db_show_rwlock(const struct lock_object *lock)
1551 {
1552 	const struct rwlock *rw;
1553 	struct thread *td;
1554 
1555 	rw = (const struct rwlock *)lock;
1556 
1557 	db_printf(" state: ");
1558 	if (rw->rw_lock == RW_UNLOCKED)
1559 		db_printf("UNLOCKED\n");
1560 	else if (rw->rw_lock == RW_DESTROYED) {
1561 		db_printf("DESTROYED\n");
1562 		return;
1563 	} else if (rw->rw_lock & RW_LOCK_READ)
1564 		db_printf("RLOCK: %ju locks\n",
1565 		    (uintmax_t)(RW_READERS(rw->rw_lock)));
1566 	else {
1567 		td = rw_wowner(rw);
1568 		db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1569 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1570 		if (rw_recursed(rw))
1571 			db_printf(" recursed: %u\n", rw->rw_recurse);
1572 	}
1573 	db_printf(" waiters: ");
1574 	switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1575 	case RW_LOCK_READ_WAITERS:
1576 		db_printf("readers\n");
1577 		break;
1578 	case RW_LOCK_WRITE_WAITERS:
1579 		db_printf("writers\n");
1580 		break;
1581 	case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1582 		db_printf("readers and writers\n");
1583 		break;
1584 	default:
1585 		db_printf("none\n");
1586 		break;
1587 	}
1588 }
1589 
1590 #endif
1591