xref: /freebsd/sys/kern/kern_rwlock.c (revision ec4deee4e4f2aef1b97d9424f25d04e91fd7dc10)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
34 
35 #include "opt_ddb.h"
36 #include "opt_hwpmc_hooks.h"
37 #include "opt_no_adaptive_rwlocks.h"
38 
39 #include <sys/param.h>
40 #include <sys/kdb.h>
41 #include <sys/ktr.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/rwlock.h>
47 #include <sys/sched.h>
48 #include <sys/smp.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #include <sys/turnstile.h>
52 
53 #include <machine/cpu.h>
54 
55 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
56 #define	ADAPTIVE_RWLOCKS
57 #endif
58 
59 #ifdef HWPMC_HOOKS
60 #include <sys/pmckern.h>
61 PMC_SOFT_DECLARE( , , lock, failed);
62 #endif
63 
64 /*
65  * Return the rwlock address when the lock cookie address is provided.
66  * This functionality assumes that struct rwlock* have a member named rw_lock.
67  */
68 #define	rwlock2rw(c)	(__containerof(c, struct rwlock, rw_lock))
69 
70 #ifdef DDB
71 #include <ddb/ddb.h>
72 
73 static void	db_show_rwlock(const struct lock_object *lock);
74 #endif
75 static void	assert_rw(const struct lock_object *lock, int what);
76 static void	lock_rw(struct lock_object *lock, uintptr_t how);
77 #ifdef KDTRACE_HOOKS
78 static int	owner_rw(const struct lock_object *lock, struct thread **owner);
79 #endif
80 static uintptr_t unlock_rw(struct lock_object *lock);
81 
82 struct lock_class lock_class_rw = {
83 	.lc_name = "rw",
84 	.lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
85 	.lc_assert = assert_rw,
86 #ifdef DDB
87 	.lc_ddb_show = db_show_rwlock,
88 #endif
89 	.lc_lock = lock_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 void
163 assert_rw(const struct lock_object *lock, int what)
164 {
165 
166 	rw_assert((const struct rwlock *)lock, what);
167 }
168 
169 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 uintptr_t
182 unlock_rw(struct lock_object *lock)
183 {
184 	struct rwlock *rw;
185 
186 	rw = (struct rwlock *)lock;
187 	rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
188 	if (rw->rw_lock & RW_LOCK_READ) {
189 		rw_runlock(rw);
190 		return (1);
191 	} else {
192 		rw_wunlock(rw);
193 		return (0);
194 	}
195 }
196 
197 #ifdef KDTRACE_HOOKS
198 int
199 owner_rw(const struct lock_object *lock, struct thread **owner)
200 {
201 	const struct rwlock *rw = (const struct rwlock *)lock;
202 	uintptr_t x = rw->rw_lock;
203 
204 	*owner = rw_wowner(rw);
205 	return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
206 	    (*owner != NULL));
207 }
208 #endif
209 
210 void
211 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
212 {
213 	struct rwlock *rw;
214 	int flags;
215 
216 	rw = rwlock2rw(c);
217 
218 	MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
219 	    RW_RECURSE | RW_NEW)) == 0);
220 	ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
221 	    ("%s: rw_lock not aligned for %s: %p", __func__, name,
222 	    &rw->rw_lock));
223 
224 	flags = LO_UPGRADABLE;
225 	if (opts & RW_DUPOK)
226 		flags |= LO_DUPOK;
227 	if (opts & RW_NOPROFILE)
228 		flags |= LO_NOPROFILE;
229 	if (!(opts & RW_NOWITNESS))
230 		flags |= LO_WITNESS;
231 	if (opts & RW_RECURSE)
232 		flags |= LO_RECURSABLE;
233 	if (opts & RW_QUIET)
234 		flags |= LO_QUIET;
235 	if (opts & RW_NEW)
236 		flags |= LO_NEW;
237 
238 	lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
239 	rw->rw_lock = RW_UNLOCKED;
240 	rw->rw_recurse = 0;
241 }
242 
243 void
244 _rw_destroy(volatile uintptr_t *c)
245 {
246 	struct rwlock *rw;
247 
248 	rw = rwlock2rw(c);
249 
250 	KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
251 	KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
252 	rw->rw_lock = RW_DESTROYED;
253 	lock_destroy(&rw->lock_object);
254 }
255 
256 void
257 rw_sysinit(void *arg)
258 {
259 	struct rw_args *args;
260 
261 	args = arg;
262 	rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
263 	    args->ra_flags);
264 }
265 
266 int
267 _rw_wowned(const volatile uintptr_t *c)
268 {
269 
270 	return (rw_wowner(rwlock2rw(c)) == curthread);
271 }
272 
273 void
274 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
275 {
276 	struct rwlock *rw;
277 	uintptr_t tid, v;
278 
279 	rw = rwlock2rw(c);
280 
281 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
282 	    !TD_IS_IDLETHREAD(curthread),
283 	    ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
284 	    curthread, rw->lock_object.lo_name, file, line));
285 	KASSERT(rw->rw_lock != RW_DESTROYED,
286 	    ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
287 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
288 	    line, NULL);
289 	tid = (uintptr_t)curthread;
290 	v = RW_UNLOCKED;
291 	if (!_rw_write_lock_fetch(rw, &v, tid))
292 		_rw_wlock_hard(rw, v, file, line);
293 	else
294 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
295 		    0, 0, file, line, LOCKSTAT_WRITER);
296 
297 	LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
298 	WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
299 	TD_LOCKS_INC(curthread);
300 }
301 
302 int
303 __rw_try_wlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
304 {
305 	struct thread *td;
306 	uintptr_t tid, v;
307 	int rval;
308 	bool recursed;
309 
310 	td = curthread;
311 	tid = (uintptr_t)td;
312 	if (SCHEDULER_STOPPED_TD(td))
313 		return (1);
314 
315 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
316 	    ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
317 	    curthread, rw->lock_object.lo_name, file, line));
318 	KASSERT(rw->rw_lock != RW_DESTROYED,
319 	    ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
320 
321 	rval = 1;
322 	recursed = false;
323 	v = RW_UNLOCKED;
324 	for (;;) {
325 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
326 			break;
327 		if (v == RW_UNLOCKED)
328 			continue;
329 		if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
330 			rw->rw_recurse++;
331 			atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
332 			break;
333 		}
334 		rval = 0;
335 		break;
336 	}
337 
338 	LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
339 	if (rval) {
340 		WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
341 		    file, line);
342 		if (!recursed)
343 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
344 			    rw, 0, 0, file, line, LOCKSTAT_WRITER);
345 		TD_LOCKS_INC(curthread);
346 	}
347 	return (rval);
348 }
349 
350 int
351 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
352 {
353 	struct rwlock *rw;
354 
355 	rw = rwlock2rw(c);
356 	return (__rw_try_wlock_int(rw LOCK_FILE_LINE_ARG));
357 }
358 
359 void
360 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
361 {
362 	struct rwlock *rw;
363 
364 	rw = rwlock2rw(c);
365 
366 	KASSERT(rw->rw_lock != RW_DESTROYED,
367 	    ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
368 	__rw_assert(c, RA_WLOCKED, file, line);
369 	WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
370 	LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
371 	    line);
372 
373 #ifdef LOCK_PROFILING
374 	_rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
375 #else
376 	__rw_wunlock(rw, curthread, file, line);
377 #endif
378 
379 	TD_LOCKS_DEC(curthread);
380 }
381 
382 /*
383  * Determines whether a new reader can acquire a lock.  Succeeds if the
384  * reader already owns a read lock and the lock is locked for read to
385  * prevent deadlock from reader recursion.  Also succeeds if the lock
386  * is unlocked and has no writer waiters or spinners.  Failing otherwise
387  * prioritizes writers before readers.
388  */
389 static bool __always_inline
390 __rw_can_read(struct thread *td, uintptr_t v, bool fp)
391 {
392 
393 	if ((v & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER))
394 	    == RW_LOCK_READ)
395 		return (true);
396 	if (!fp && td->td_rw_rlocks && (v & RW_LOCK_READ))
397 		return (true);
398 	return (false);
399 }
400 
401 static bool __always_inline
402 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, bool fp
403     LOCK_FILE_LINE_ARG_DEF)
404 {
405 
406 	/*
407 	 * Handle the easy case.  If no other thread has a write
408 	 * lock, then try to bump up the count of read locks.  Note
409 	 * that we have to preserve the current state of the
410 	 * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
411 	 * read lock, then rw_lock must have changed, so restart
412 	 * the loop.  Note that this handles the case of a
413 	 * completely unlocked rwlock since such a lock is encoded
414 	 * as a read lock with no waiters.
415 	 */
416 	while (__rw_can_read(td, *vp, fp)) {
417 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
418 			*vp + RW_ONE_READER)) {
419 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
420 				CTR4(KTR_LOCK,
421 				    "%s: %p succeed %p -> %p", __func__,
422 				    rw, (void *)*vp,
423 				    (void *)(*vp + RW_ONE_READER));
424 			td->td_rw_rlocks++;
425 			return (true);
426 		}
427 	}
428 	return (false);
429 }
430 
431 static void __noinline
432 __rw_rlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
433     LOCK_FILE_LINE_ARG_DEF)
434 {
435 	struct turnstile *ts;
436 	struct thread *owner;
437 #ifdef ADAPTIVE_RWLOCKS
438 	int spintries = 0;
439 	int i, n;
440 #endif
441 #ifdef LOCK_PROFILING
442 	uint64_t waittime = 0;
443 	int contested = 0;
444 #endif
445 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
446 	struct lock_delay_arg lda;
447 #endif
448 #ifdef KDTRACE_HOOKS
449 	u_int sleep_cnt = 0;
450 	int64_t sleep_time = 0;
451 	int64_t all_time = 0;
452 #endif
453 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
454 	uintptr_t state = 0;
455 	int doing_lockprof = 0;
456 #endif
457 
458 #ifdef KDTRACE_HOOKS
459 	if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
460 		if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
461 			goto out_lockstat;
462 		doing_lockprof = 1;
463 		all_time -= lockstat_nsecs(&rw->lock_object);
464 		state = v;
465 	}
466 #endif
467 #ifdef LOCK_PROFILING
468 	doing_lockprof = 1;
469 	state = v;
470 #endif
471 
472 	if (SCHEDULER_STOPPED())
473 		return;
474 
475 #if defined(ADAPTIVE_RWLOCKS)
476 	lock_delay_arg_init(&lda, &rw_delay);
477 #elif defined(KDTRACE_HOOKS)
478 	lock_delay_arg_init_noadapt(&lda);
479 #endif
480 
481 #ifdef HWPMC_HOOKS
482 	PMC_SOFT_CALL( , , lock, failed);
483 #endif
484 	lock_profile_obtain_lock_failed(&rw->lock_object,
485 	    &contested, &waittime);
486 
487 	for (;;) {
488 		if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
489 			break;
490 #ifdef KDTRACE_HOOKS
491 		lda.spin_cnt++;
492 #endif
493 
494 #ifdef ADAPTIVE_RWLOCKS
495 		/*
496 		 * If the owner is running on another CPU, spin until
497 		 * the owner stops running or the state of the lock
498 		 * changes.
499 		 */
500 		if ((v & RW_LOCK_READ) == 0) {
501 			owner = (struct thread *)RW_OWNER(v);
502 			if (TD_IS_RUNNING(owner)) {
503 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
504 					CTR3(KTR_LOCK,
505 					    "%s: spinning on %p held by %p",
506 					    __func__, rw, owner);
507 				KTR_STATE1(KTR_SCHED, "thread",
508 				    sched_tdname(curthread), "spinning",
509 				    "lockname:\"%s\"", rw->lock_object.lo_name);
510 				do {
511 					lock_delay(&lda);
512 					v = RW_READ_VALUE(rw);
513 					owner = lv_rw_wowner(v);
514 				} while (owner != NULL && TD_IS_RUNNING(owner));
515 				KTR_STATE0(KTR_SCHED, "thread",
516 				    sched_tdname(curthread), "running");
517 				continue;
518 			}
519 		} else {
520 			if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
521 				MPASS(!__rw_can_read(td, v, false));
522 				lock_delay_spin(2);
523 				v = RW_READ_VALUE(rw);
524 				continue;
525 			}
526 			if (spintries < rowner_retries) {
527 				spintries++;
528 				KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
529 				    "spinning", "lockname:\"%s\"",
530 				    rw->lock_object.lo_name);
531 				n = RW_READERS(v);
532 				for (i = 0; i < rowner_loops; i += n) {
533 					lock_delay_spin(n);
534 					v = RW_READ_VALUE(rw);
535 					if (!(v & RW_LOCK_READ))
536 						break;
537 					n = RW_READERS(v);
538 					if (n == 0)
539 						break;
540 					if (__rw_can_read(td, v, false))
541 						break;
542 				}
543 #ifdef KDTRACE_HOOKS
544 				lda.spin_cnt += rowner_loops - i;
545 #endif
546 				KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
547 				    "running");
548 				if (i < rowner_loops)
549 					continue;
550 			}
551 		}
552 #endif
553 
554 		/*
555 		 * Okay, now it's the hard case.  Some other thread already
556 		 * has a write lock or there are write waiters present,
557 		 * acquire the turnstile lock so we can begin the process
558 		 * of blocking.
559 		 */
560 		ts = turnstile_trywait(&rw->lock_object);
561 
562 		/*
563 		 * The lock might have been released while we spun, so
564 		 * recheck its state and restart the loop if needed.
565 		 */
566 		v = RW_READ_VALUE(rw);
567 retry_ts:
568 		if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
569 		    __rw_can_read(td, v, false)) {
570 			turnstile_cancel(ts);
571 			continue;
572 		}
573 
574 		owner = lv_rw_wowner(v);
575 
576 #ifdef ADAPTIVE_RWLOCKS
577 		/*
578 		 * The current lock owner might have started executing
579 		 * on another CPU (or the lock could have changed
580 		 * owners) while we were waiting on the turnstile
581 		 * chain lock.  If so, drop the turnstile lock and try
582 		 * again.
583 		 */
584 		if (owner != NULL) {
585 			if (TD_IS_RUNNING(owner)) {
586 				turnstile_cancel(ts);
587 				continue;
588 			}
589 		}
590 #endif
591 
592 		/*
593 		 * The lock is held in write mode or it already has waiters.
594 		 */
595 		MPASS(!__rw_can_read(td, v, false));
596 
597 		/*
598 		 * If the RW_LOCK_READ_WAITERS flag is already set, then
599 		 * we can go ahead and block.  If it is not set then try
600 		 * to set it.  If we fail to set it drop the turnstile
601 		 * lock and restart the loop.
602 		 */
603 		if (!(v & RW_LOCK_READ_WAITERS)) {
604 			if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
605 			    v | RW_LOCK_READ_WAITERS))
606 				goto retry_ts;
607 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
608 				CTR2(KTR_LOCK, "%s: %p set read waiters flag",
609 				    __func__, rw);
610 		}
611 
612 		/*
613 		 * We were unable to acquire the lock and the read waiters
614 		 * flag is set, so we must block on the turnstile.
615 		 */
616 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
617 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
618 			    rw);
619 #ifdef KDTRACE_HOOKS
620 		sleep_time -= lockstat_nsecs(&rw->lock_object);
621 #endif
622 		MPASS(owner == rw_owner(rw));
623 		turnstile_wait(ts, owner, TS_SHARED_QUEUE);
624 #ifdef KDTRACE_HOOKS
625 		sleep_time += lockstat_nsecs(&rw->lock_object);
626 		sleep_cnt++;
627 #endif
628 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
629 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
630 			    __func__, rw);
631 		v = RW_READ_VALUE(rw);
632 	}
633 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
634 	if (__predict_true(!doing_lockprof))
635 		return;
636 #endif
637 #ifdef KDTRACE_HOOKS
638 	all_time += lockstat_nsecs(&rw->lock_object);
639 	if (sleep_time)
640 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
641 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
642 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
643 
644 	/* Record only the loops spinning and not sleeping. */
645 	if (lda.spin_cnt > sleep_cnt)
646 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
647 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
648 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
649 out_lockstat:
650 #endif
651 	/*
652 	 * TODO: acquire "owner of record" here.  Here be turnstile dragons
653 	 * however.  turnstiles don't like owners changing between calls to
654 	 * turnstile_wait() currently.
655 	 */
656 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
657 	    waittime, file, line, LOCKSTAT_READER);
658 }
659 
660 void
661 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
662 {
663 	struct thread *td;
664 	uintptr_t v;
665 
666 	td = curthread;
667 
668 	KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) ||
669 	    !TD_IS_IDLETHREAD(td),
670 	    ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
671 	    td, rw->lock_object.lo_name, file, line));
672 	KASSERT(rw->rw_lock != RW_DESTROYED,
673 	    ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
674 	KASSERT(rw_wowner(rw) != td,
675 	    ("rw_rlock: wlock already held for %s @ %s:%d",
676 	    rw->lock_object.lo_name, file, line));
677 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
678 
679 	v = RW_READ_VALUE(rw);
680 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
681 	    !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
682 		__rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
683 	else
684 		lock_profile_obtain_lock_success(&rw->lock_object, 0, 0,
685 		    file, line);
686 
687 	LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
688 	WITNESS_LOCK(&rw->lock_object, 0, file, line);
689 	TD_LOCKS_INC(curthread);
690 }
691 
692 void
693 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
694 {
695 	struct rwlock *rw;
696 
697 	rw = rwlock2rw(c);
698 	__rw_rlock_int(rw LOCK_FILE_LINE_ARG);
699 }
700 
701 int
702 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
703 {
704 	uintptr_t x;
705 
706 	if (SCHEDULER_STOPPED())
707 		return (1);
708 
709 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
710 	    ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
711 	    curthread, rw->lock_object.lo_name, file, line));
712 
713 	x = rw->rw_lock;
714 	for (;;) {
715 		KASSERT(rw->rw_lock != RW_DESTROYED,
716 		    ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
717 		if (!(x & RW_LOCK_READ))
718 			break;
719 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
720 			LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
721 			    line);
722 			WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
723 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
724 			    rw, 0, 0, file, line, LOCKSTAT_READER);
725 			TD_LOCKS_INC(curthread);
726 			curthread->td_rw_rlocks++;
727 			return (1);
728 		}
729 	}
730 
731 	LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
732 	return (0);
733 }
734 
735 int
736 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
737 {
738 	struct rwlock *rw;
739 
740 	rw = rwlock2rw(c);
741 	return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
742 }
743 
744 static bool __always_inline
745 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
746 {
747 
748 	for (;;) {
749 		if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
750 			if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
751 			    *vp - RW_ONE_READER)) {
752 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
753 					CTR4(KTR_LOCK,
754 					    "%s: %p succeeded %p -> %p",
755 					    __func__, rw, (void *)*vp,
756 					    (void *)(*vp - RW_ONE_READER));
757 				td->td_rw_rlocks--;
758 				return (true);
759 			}
760 			continue;
761 		}
762 		break;
763 	}
764 	return (false);
765 }
766 
767 static void __noinline
768 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
769     LOCK_FILE_LINE_ARG_DEF)
770 {
771 	struct turnstile *ts;
772 	uintptr_t setv, queue;
773 
774 	if (SCHEDULER_STOPPED())
775 		return;
776 
777 	if (__rw_runlock_try(rw, td, &v))
778 		goto out_lockstat;
779 
780 	/*
781 	 * Ok, we know we have waiters and we think we are the
782 	 * last reader, so grab the turnstile lock.
783 	 */
784 	turnstile_chain_lock(&rw->lock_object);
785 	v = RW_READ_VALUE(rw);
786 	for (;;) {
787 		if (__rw_runlock_try(rw, td, &v))
788 			break;
789 
790 		MPASS(v & RW_LOCK_WAITERS);
791 
792 		/*
793 		 * Try to drop our lock leaving the lock in a unlocked
794 		 * state.
795 		 *
796 		 * If you wanted to do explicit lock handoff you'd have to
797 		 * do it here.  You'd also want to use turnstile_signal()
798 		 * and you'd have to handle the race where a higher
799 		 * priority thread blocks on the write lock before the
800 		 * thread you wakeup actually runs and have the new thread
801 		 * "steal" the lock.  For now it's a lot simpler to just
802 		 * wakeup all of the waiters.
803 		 *
804 		 * As above, if we fail, then another thread might have
805 		 * acquired a read lock, so drop the turnstile lock and
806 		 * restart.
807 		 */
808 		setv = RW_UNLOCKED;
809 		queue = TS_SHARED_QUEUE;
810 		if (v & RW_LOCK_WRITE_WAITERS) {
811 			queue = TS_EXCLUSIVE_QUEUE;
812 			setv |= (v & RW_LOCK_READ_WAITERS);
813 		}
814 		setv |= (v & RW_LOCK_WRITE_SPINNER);
815 		if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
816 			continue;
817 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
818 			CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
819 			    __func__, rw);
820 
821 		/*
822 		 * Ok.  The lock is released and all that's left is to
823 		 * wake up the waiters.  Note that the lock might not be
824 		 * free anymore, but in that case the writers will just
825 		 * block again if they run before the new lock holder(s)
826 		 * release the lock.
827 		 */
828 		ts = turnstile_lookup(&rw->lock_object);
829 		MPASS(ts != NULL);
830 		turnstile_broadcast(ts, queue);
831 		turnstile_unpend(ts);
832 		td->td_rw_rlocks--;
833 		break;
834 	}
835 	turnstile_chain_unlock(&rw->lock_object);
836 out_lockstat:
837 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
838 }
839 
840 void
841 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
842 {
843 	struct thread *td;
844 	uintptr_t v;
845 
846 	KASSERT(rw->rw_lock != RW_DESTROYED,
847 	    ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
848 	__rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
849 	WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
850 	LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
851 
852 	td = curthread;
853 	v = RW_READ_VALUE(rw);
854 
855 	if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
856 	    !__rw_runlock_try(rw, td, &v)))
857 		__rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
858 	else
859 		lock_profile_release_lock(&rw->lock_object);
860 
861 	TD_LOCKS_DEC(curthread);
862 }
863 
864 void
865 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
866 {
867 	struct rwlock *rw;
868 
869 	rw = rwlock2rw(c);
870 	_rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
871 }
872 
873 #ifdef ADAPTIVE_RWLOCKS
874 static inline void
875 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
876 {
877 
878 	if (v & RW_LOCK_WRITE_SPINNER)
879 		return;
880 	if (*in_critical) {
881 		critical_exit();
882 		*in_critical = false;
883 		(*extra_work)--;
884 	}
885 }
886 #else
887 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
888 #endif
889 
890 /*
891  * This function is called when we are unable to obtain a write lock on the
892  * first try.  This means that at least one other thread holds either a
893  * read or write lock.
894  */
895 void
896 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
897 {
898 	uintptr_t tid;
899 	struct rwlock *rw;
900 	struct turnstile *ts;
901 	struct thread *owner;
902 #ifdef ADAPTIVE_RWLOCKS
903 	int spintries = 0;
904 	int i, n;
905 	enum { READERS, WRITER } sleep_reason = READERS;
906 	bool in_critical = false;
907 #endif
908 	uintptr_t setv;
909 #ifdef LOCK_PROFILING
910 	uint64_t waittime = 0;
911 	int contested = 0;
912 #endif
913 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
914 	struct lock_delay_arg lda;
915 #endif
916 #ifdef KDTRACE_HOOKS
917 	u_int sleep_cnt = 0;
918 	int64_t sleep_time = 0;
919 	int64_t all_time = 0;
920 #endif
921 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
922 	uintptr_t state = 0;
923 	int doing_lockprof = 0;
924 #endif
925 	int extra_work = 0;
926 
927 	tid = (uintptr_t)curthread;
928 	rw = rwlock2rw(c);
929 
930 #ifdef KDTRACE_HOOKS
931 	if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
932 		while (v == RW_UNLOCKED) {
933 			if (_rw_write_lock_fetch(rw, &v, tid))
934 				goto out_lockstat;
935 		}
936 		extra_work = 1;
937 		doing_lockprof = 1;
938 		all_time -= lockstat_nsecs(&rw->lock_object);
939 		state = v;
940 	}
941 #endif
942 #ifdef LOCK_PROFILING
943 	extra_work = 1;
944 	doing_lockprof = 1;
945 	state = v;
946 #endif
947 
948 	if (SCHEDULER_STOPPED())
949 		return;
950 
951 #if defined(ADAPTIVE_RWLOCKS)
952 	lock_delay_arg_init(&lda, &rw_delay);
953 #elif defined(KDTRACE_HOOKS)
954 	lock_delay_arg_init_noadapt(&lda);
955 #endif
956 	if (__predict_false(v == RW_UNLOCKED))
957 		v = RW_READ_VALUE(rw);
958 
959 	if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
960 		KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
961 		    ("%s: recursing but non-recursive rw %s @ %s:%d\n",
962 		    __func__, rw->lock_object.lo_name, file, line));
963 		rw->rw_recurse++;
964 		atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
965 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
966 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
967 		return;
968 	}
969 
970 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
971 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
972 		    rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
973 
974 #ifdef HWPMC_HOOKS
975 	PMC_SOFT_CALL( , , lock, failed);
976 #endif
977 	lock_profile_obtain_lock_failed(&rw->lock_object,
978 	    &contested, &waittime);
979 
980 	for (;;) {
981 		if (v == RW_UNLOCKED) {
982 			if (_rw_write_lock_fetch(rw, &v, tid))
983 				break;
984 			continue;
985 		}
986 #ifdef KDTRACE_HOOKS
987 		lda.spin_cnt++;
988 #endif
989 
990 #ifdef ADAPTIVE_RWLOCKS
991 		if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
992 			if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
993 				break;
994 			continue;
995 		}
996 
997 		/*
998 		 * If the lock is write locked and the owner is
999 		 * running on another CPU, spin until the owner stops
1000 		 * running or the state of the lock changes.
1001 		 */
1002 		if (!(v & RW_LOCK_READ)) {
1003 			rw_drop_critical(v, &in_critical, &extra_work);
1004 			sleep_reason = WRITER;
1005 			owner = lv_rw_wowner(v);
1006 			if (!TD_IS_RUNNING(owner))
1007 				goto ts;
1008 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
1009 				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1010 				    __func__, rw, owner);
1011 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1012 			    "spinning", "lockname:\"%s\"",
1013 			    rw->lock_object.lo_name);
1014 			do {
1015 				lock_delay(&lda);
1016 				v = RW_READ_VALUE(rw);
1017 				owner = lv_rw_wowner(v);
1018 			} while (owner != NULL && TD_IS_RUNNING(owner));
1019 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1020 			    "running");
1021 			continue;
1022 		} else if (RW_READERS(v) > 0) {
1023 			sleep_reason = READERS;
1024 			if (spintries == rowner_retries)
1025 				goto ts;
1026 			if (!(v & RW_LOCK_WRITE_SPINNER)) {
1027 				if (!in_critical) {
1028 					critical_enter();
1029 					in_critical = true;
1030 					extra_work++;
1031 				}
1032 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1033 				    v | RW_LOCK_WRITE_SPINNER)) {
1034 					critical_exit();
1035 					in_critical = false;
1036 					extra_work--;
1037 					continue;
1038 				}
1039 			}
1040 			spintries++;
1041 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1042 			    "spinning", "lockname:\"%s\"",
1043 			    rw->lock_object.lo_name);
1044 			n = RW_READERS(v);
1045 			for (i = 0; i < rowner_loops; i += n) {
1046 				lock_delay_spin(n);
1047 				v = RW_READ_VALUE(rw);
1048 				if (!(v & RW_LOCK_WRITE_SPINNER))
1049 					break;
1050 				if (!(v & RW_LOCK_READ))
1051 					break;
1052 				n = RW_READERS(v);
1053 				if (n == 0)
1054 					break;
1055 			}
1056 #ifdef KDTRACE_HOOKS
1057 			lda.spin_cnt += i;
1058 #endif
1059 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1060 			    "running");
1061 			if (i < rowner_loops)
1062 				continue;
1063 		}
1064 ts:
1065 #endif
1066 		ts = turnstile_trywait(&rw->lock_object);
1067 		v = RW_READ_VALUE(rw);
1068 retry_ts:
1069 		owner = lv_rw_wowner(v);
1070 
1071 #ifdef ADAPTIVE_RWLOCKS
1072 		/*
1073 		 * The current lock owner might have started executing
1074 		 * on another CPU (or the lock could have changed
1075 		 * owners) while we were waiting on the turnstile
1076 		 * chain lock.  If so, drop the turnstile lock and try
1077 		 * again.
1078 		 */
1079 		if (owner != NULL) {
1080 			if (TD_IS_RUNNING(owner)) {
1081 				turnstile_cancel(ts);
1082 				rw_drop_critical(v, &in_critical, &extra_work);
1083 				continue;
1084 			}
1085 		} else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1086 			turnstile_cancel(ts);
1087 			rw_drop_critical(v, &in_critical, &extra_work);
1088 			continue;
1089 		}
1090 #endif
1091 		/*
1092 		 * Check for the waiters flags about this rwlock.
1093 		 * If the lock was released, without maintain any pending
1094 		 * waiters queue, simply try to acquire it.
1095 		 * If a pending waiters queue is present, claim the lock
1096 		 * ownership and maintain the pending queue.
1097 		 */
1098 		setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1099 		if ((v & ~setv) == RW_UNLOCKED) {
1100 			setv &= ~RW_LOCK_WRITE_SPINNER;
1101 			if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1102 				if (setv)
1103 					turnstile_claim(ts);
1104 				else
1105 					turnstile_cancel(ts);
1106 				break;
1107 			}
1108 			goto retry_ts;
1109 		}
1110 
1111 #ifdef ADAPTIVE_RWLOCKS
1112 		if (in_critical) {
1113 			if ((v & RW_LOCK_WRITE_SPINNER) ||
1114 			    !((v & RW_LOCK_WRITE_WAITERS))) {
1115 				setv = v & ~RW_LOCK_WRITE_SPINNER;
1116 				setv |= RW_LOCK_WRITE_WAITERS;
1117 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1118 					goto retry_ts;
1119 			}
1120 			critical_exit();
1121 			in_critical = false;
1122 			extra_work--;
1123 		} else {
1124 #endif
1125 			/*
1126 			 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1127 			 * set it.  If we fail to set it, then loop back and try
1128 			 * again.
1129 			 */
1130 			if (!(v & RW_LOCK_WRITE_WAITERS)) {
1131 				if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1132 				    v | RW_LOCK_WRITE_WAITERS))
1133 					goto retry_ts;
1134 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
1135 					CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1136 					    __func__, rw);
1137 			}
1138 #ifdef ADAPTIVE_RWLOCKS
1139 		}
1140 #endif
1141 		/*
1142 		 * We were unable to acquire the lock and the write waiters
1143 		 * flag is set, so we must block on the turnstile.
1144 		 */
1145 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1146 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1147 			    rw);
1148 #ifdef KDTRACE_HOOKS
1149 		sleep_time -= lockstat_nsecs(&rw->lock_object);
1150 #endif
1151 		MPASS(owner == rw_owner(rw));
1152 		turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1153 #ifdef KDTRACE_HOOKS
1154 		sleep_time += lockstat_nsecs(&rw->lock_object);
1155 		sleep_cnt++;
1156 #endif
1157 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1158 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1159 			    __func__, rw);
1160 #ifdef ADAPTIVE_RWLOCKS
1161 		spintries = 0;
1162 #endif
1163 		v = RW_READ_VALUE(rw);
1164 	}
1165 	if (__predict_true(!extra_work))
1166 		return;
1167 #ifdef ADAPTIVE_RWLOCKS
1168 	if (in_critical)
1169 		critical_exit();
1170 #endif
1171 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1172 	if (__predict_true(!doing_lockprof))
1173 		return;
1174 #endif
1175 #ifdef KDTRACE_HOOKS
1176 	all_time += lockstat_nsecs(&rw->lock_object);
1177 	if (sleep_time)
1178 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1179 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1180 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1181 
1182 	/* Record only the loops spinning and not sleeping. */
1183 	if (lda.spin_cnt > sleep_cnt)
1184 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1185 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1186 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1187 out_lockstat:
1188 #endif
1189 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1190 	    waittime, file, line, LOCKSTAT_WRITER);
1191 }
1192 
1193 /*
1194  * This function is called if lockstat is active or the first try at releasing
1195  * a write lock failed.  The latter means that the lock is recursed or one of
1196  * the 2 waiter bits must be set indicating that at least one thread is waiting
1197  * on this lock.
1198  */
1199 void
1200 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1201 {
1202 	struct rwlock *rw;
1203 	struct turnstile *ts;
1204 	uintptr_t tid, setv;
1205 	int queue;
1206 
1207 	tid = (uintptr_t)curthread;
1208 	if (SCHEDULER_STOPPED())
1209 		return;
1210 
1211 	rw = rwlock2rw(c);
1212 	if (__predict_false(v == tid))
1213 		v = RW_READ_VALUE(rw);
1214 
1215 	if (v & RW_LOCK_WRITER_RECURSED) {
1216 		if (--(rw->rw_recurse) == 0)
1217 			atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1218 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1219 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1220 		return;
1221 	}
1222 
1223 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1224 	if (v == tid && _rw_write_unlock(rw, tid))
1225 		return;
1226 
1227 	KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1228 	    ("%s: neither of the waiter flags are set", __func__));
1229 
1230 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1231 		CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1232 
1233 	turnstile_chain_lock(&rw->lock_object);
1234 
1235 	/*
1236 	 * Use the same algo as sx locks for now.  Prefer waking up shared
1237 	 * waiters if we have any over writers.  This is probably not ideal.
1238 	 *
1239 	 * 'v' is the value we are going to write back to rw_lock.  If we
1240 	 * have waiters on both queues, we need to preserve the state of
1241 	 * the waiter flag for the queue we don't wake up.  For now this is
1242 	 * hardcoded for the algorithm mentioned above.
1243 	 *
1244 	 * In the case of both readers and writers waiting we wakeup the
1245 	 * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1246 	 * new writer comes in before a reader it will claim the lock up
1247 	 * above.  There is probably a potential priority inversion in
1248 	 * there that could be worked around either by waking both queues
1249 	 * of waiters or doing some complicated lock handoff gymnastics.
1250 	 */
1251 	setv = RW_UNLOCKED;
1252 	v = RW_READ_VALUE(rw);
1253 	queue = TS_SHARED_QUEUE;
1254 	if (v & RW_LOCK_WRITE_WAITERS) {
1255 		queue = TS_EXCLUSIVE_QUEUE;
1256 		setv |= (v & RW_LOCK_READ_WAITERS);
1257 	}
1258 	atomic_store_rel_ptr(&rw->rw_lock, setv);
1259 
1260 	/* Wake up all waiters for the specific queue. */
1261 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1262 		CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1263 		    queue == TS_SHARED_QUEUE ? "read" : "write");
1264 
1265 	ts = turnstile_lookup(&rw->lock_object);
1266 	MPASS(ts != NULL);
1267 	turnstile_broadcast(ts, queue);
1268 	turnstile_unpend(ts);
1269 	turnstile_chain_unlock(&rw->lock_object);
1270 }
1271 
1272 /*
1273  * Attempt to do a non-blocking upgrade from a read lock to a write
1274  * lock.  This will only succeed if this thread holds a single read
1275  * lock.  Returns true if the upgrade succeeded and false otherwise.
1276  */
1277 int
1278 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1279 {
1280 	uintptr_t v, setv, tid;
1281 	struct turnstile *ts;
1282 	int success;
1283 
1284 	if (SCHEDULER_STOPPED())
1285 		return (1);
1286 
1287 	KASSERT(rw->rw_lock != RW_DESTROYED,
1288 	    ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1289 	__rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1290 
1291 	/*
1292 	 * Attempt to switch from one reader to a writer.  If there
1293 	 * are any write waiters, then we will have to lock the
1294 	 * turnstile first to prevent races with another writer
1295 	 * calling turnstile_wait() before we have claimed this
1296 	 * turnstile.  So, do the simple case of no waiters first.
1297 	 */
1298 	tid = (uintptr_t)curthread;
1299 	success = 0;
1300 	v = RW_READ_VALUE(rw);
1301 	for (;;) {
1302 		if (RW_READERS(v) > 1)
1303 			break;
1304 		if (!(v & RW_LOCK_WAITERS)) {
1305 			success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1306 			if (!success)
1307 				continue;
1308 			break;
1309 		}
1310 
1311 		/*
1312 		 * Ok, we think we have waiters, so lock the turnstile.
1313 		 */
1314 		ts = turnstile_trywait(&rw->lock_object);
1315 		v = RW_READ_VALUE(rw);
1316 retry_ts:
1317 		if (RW_READERS(v) > 1) {
1318 			turnstile_cancel(ts);
1319 			break;
1320 		}
1321 		/*
1322 		 * Try to switch from one reader to a writer again.  This time
1323 		 * we honor the current state of the waiters flags.
1324 		 * If we obtain the lock with the flags set, then claim
1325 		 * ownership of the turnstile.
1326 		 */
1327 		setv = tid | (v & RW_LOCK_WAITERS);
1328 		success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1329 		if (success) {
1330 			if (v & RW_LOCK_WAITERS)
1331 				turnstile_claim(ts);
1332 			else
1333 				turnstile_cancel(ts);
1334 			break;
1335 		}
1336 		goto retry_ts;
1337 	}
1338 	LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1339 	if (success) {
1340 		curthread->td_rw_rlocks--;
1341 		WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1342 		    file, line);
1343 		LOCKSTAT_RECORD0(rw__upgrade, rw);
1344 	}
1345 	return (success);
1346 }
1347 
1348 int
1349 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1350 {
1351 	struct rwlock *rw;
1352 
1353 	rw = rwlock2rw(c);
1354 	return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1355 }
1356 
1357 /*
1358  * Downgrade a write lock into a single read lock.
1359  */
1360 void
1361 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1362 {
1363 	struct turnstile *ts;
1364 	uintptr_t tid, v;
1365 	int rwait, wwait;
1366 
1367 	if (SCHEDULER_STOPPED())
1368 		return;
1369 
1370 	KASSERT(rw->rw_lock != RW_DESTROYED,
1371 	    ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1372 	__rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1373 #ifndef INVARIANTS
1374 	if (rw_recursed(rw))
1375 		panic("downgrade of a recursed lock");
1376 #endif
1377 
1378 	WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1379 
1380 	/*
1381 	 * Convert from a writer to a single reader.  First we handle
1382 	 * the easy case with no waiters.  If there are any waiters, we
1383 	 * lock the turnstile and "disown" the lock.
1384 	 */
1385 	tid = (uintptr_t)curthread;
1386 	if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1387 		goto out;
1388 
1389 	/*
1390 	 * Ok, we think we have waiters, so lock the turnstile so we can
1391 	 * read the waiter flags without any races.
1392 	 */
1393 	turnstile_chain_lock(&rw->lock_object);
1394 	v = rw->rw_lock & RW_LOCK_WAITERS;
1395 	rwait = v & RW_LOCK_READ_WAITERS;
1396 	wwait = v & RW_LOCK_WRITE_WAITERS;
1397 	MPASS(rwait | wwait);
1398 
1399 	/*
1400 	 * Downgrade from a write lock while preserving waiters flag
1401 	 * and give up ownership of the turnstile.
1402 	 */
1403 	ts = turnstile_lookup(&rw->lock_object);
1404 	MPASS(ts != NULL);
1405 	if (!wwait)
1406 		v &= ~RW_LOCK_READ_WAITERS;
1407 	atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1408 	/*
1409 	 * Wake other readers if there are no writers pending.  Otherwise they
1410 	 * won't be able to acquire the lock anyway.
1411 	 */
1412 	if (rwait && !wwait) {
1413 		turnstile_broadcast(ts, TS_SHARED_QUEUE);
1414 		turnstile_unpend(ts);
1415 	} else
1416 		turnstile_disown(ts);
1417 	turnstile_chain_unlock(&rw->lock_object);
1418 out:
1419 	curthread->td_rw_rlocks++;
1420 	LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1421 	LOCKSTAT_RECORD0(rw__downgrade, rw);
1422 }
1423 
1424 void
1425 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1426 {
1427 	struct rwlock *rw;
1428 
1429 	rw = rwlock2rw(c);
1430 	__rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1431 }
1432 
1433 #ifdef INVARIANT_SUPPORT
1434 #ifndef INVARIANTS
1435 #undef __rw_assert
1436 #endif
1437 
1438 /*
1439  * In the non-WITNESS case, rw_assert() can only detect that at least
1440  * *some* thread owns an rlock, but it cannot guarantee that *this*
1441  * thread owns an rlock.
1442  */
1443 void
1444 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1445 {
1446 	const struct rwlock *rw;
1447 
1448 	if (SCHEDULER_STOPPED())
1449 		return;
1450 
1451 	rw = rwlock2rw(c);
1452 
1453 	switch (what) {
1454 	case RA_LOCKED:
1455 	case RA_LOCKED | RA_RECURSED:
1456 	case RA_LOCKED | RA_NOTRECURSED:
1457 	case RA_RLOCKED:
1458 	case RA_RLOCKED | RA_RECURSED:
1459 	case RA_RLOCKED | RA_NOTRECURSED:
1460 #ifdef WITNESS
1461 		witness_assert(&rw->lock_object, what, file, line);
1462 #else
1463 		/*
1464 		 * If some other thread has a write lock or we have one
1465 		 * and are asserting a read lock, fail.  Also, if no one
1466 		 * has a lock at all, fail.
1467 		 */
1468 		if (rw->rw_lock == RW_UNLOCKED ||
1469 		    (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1470 		    rw_wowner(rw) != curthread)))
1471 			panic("Lock %s not %slocked @ %s:%d\n",
1472 			    rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1473 			    "read " : "", file, line);
1474 
1475 		if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1476 			if (rw_recursed(rw)) {
1477 				if (what & RA_NOTRECURSED)
1478 					panic("Lock %s recursed @ %s:%d\n",
1479 					    rw->lock_object.lo_name, file,
1480 					    line);
1481 			} else if (what & RA_RECURSED)
1482 				panic("Lock %s not recursed @ %s:%d\n",
1483 				    rw->lock_object.lo_name, file, line);
1484 		}
1485 #endif
1486 		break;
1487 	case RA_WLOCKED:
1488 	case RA_WLOCKED | RA_RECURSED:
1489 	case RA_WLOCKED | RA_NOTRECURSED:
1490 		if (rw_wowner(rw) != curthread)
1491 			panic("Lock %s not exclusively locked @ %s:%d\n",
1492 			    rw->lock_object.lo_name, file, line);
1493 		if (rw_recursed(rw)) {
1494 			if (what & RA_NOTRECURSED)
1495 				panic("Lock %s recursed @ %s:%d\n",
1496 				    rw->lock_object.lo_name, file, line);
1497 		} else if (what & RA_RECURSED)
1498 			panic("Lock %s not recursed @ %s:%d\n",
1499 			    rw->lock_object.lo_name, file, line);
1500 		break;
1501 	case RA_UNLOCKED:
1502 #ifdef WITNESS
1503 		witness_assert(&rw->lock_object, what, file, line);
1504 #else
1505 		/*
1506 		 * If we hold a write lock fail.  We can't reliably check
1507 		 * to see if we hold a read lock or not.
1508 		 */
1509 		if (rw_wowner(rw) == curthread)
1510 			panic("Lock %s exclusively locked @ %s:%d\n",
1511 			    rw->lock_object.lo_name, file, line);
1512 #endif
1513 		break;
1514 	default:
1515 		panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1516 		    line);
1517 	}
1518 }
1519 #endif /* INVARIANT_SUPPORT */
1520 
1521 #ifdef DDB
1522 void
1523 db_show_rwlock(const struct lock_object *lock)
1524 {
1525 	const struct rwlock *rw;
1526 	struct thread *td;
1527 
1528 	rw = (const struct rwlock *)lock;
1529 
1530 	db_printf(" state: ");
1531 	if (rw->rw_lock == RW_UNLOCKED)
1532 		db_printf("UNLOCKED\n");
1533 	else if (rw->rw_lock == RW_DESTROYED) {
1534 		db_printf("DESTROYED\n");
1535 		return;
1536 	} else if (rw->rw_lock & RW_LOCK_READ)
1537 		db_printf("RLOCK: %ju locks\n",
1538 		    (uintmax_t)(RW_READERS(rw->rw_lock)));
1539 	else {
1540 		td = rw_wowner(rw);
1541 		db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1542 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1543 		if (rw_recursed(rw))
1544 			db_printf(" recursed: %u\n", rw->rw_recurse);
1545 	}
1546 	db_printf(" waiters: ");
1547 	switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1548 	case RW_LOCK_READ_WAITERS:
1549 		db_printf("readers\n");
1550 		break;
1551 	case RW_LOCK_WRITE_WAITERS:
1552 		db_printf("writers\n");
1553 		break;
1554 	case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1555 		db_printf("readers and writers\n");
1556 		break;
1557 	default:
1558 		db_printf("none\n");
1559 		break;
1560 	}
1561 }
1562 
1563 #endif
1564