xref: /freebsd/sys/kern/kern_rwlock.c (revision 076ad2f8)
1 /*-
2  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * Machine independent bits of reader/writer lock implementation.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_ddb.h"
35 #include "opt_hwpmc_hooks.h"
36 #include "opt_no_adaptive_rwlocks.h"
37 
38 #include <sys/param.h>
39 #include <sys/kdb.h>
40 #include <sys/ktr.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/rwlock.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 #include <sys/turnstile.h>
51 
52 #include <machine/cpu.h>
53 
54 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
55 #define	ADAPTIVE_RWLOCKS
56 #endif
57 
58 #ifdef HWPMC_HOOKS
59 #include <sys/pmckern.h>
60 PMC_SOFT_DECLARE( , , lock, failed);
61 #endif
62 
63 /*
64  * Return the rwlock address when the lock cookie address is provided.
65  * This functionality assumes that struct rwlock* have a member named rw_lock.
66  */
67 #define	rwlock2rw(c)	(__containerof(c, struct rwlock, rw_lock))
68 
69 #ifdef DDB
70 #include <ddb/ddb.h>
71 
72 static void	db_show_rwlock(const struct lock_object *lock);
73 #endif
74 static void	assert_rw(const struct lock_object *lock, int what);
75 static void	lock_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_unlock = unlock_rw,
90 #ifdef KDTRACE_HOOKS
91 	.lc_owner = owner_rw,
92 #endif
93 };
94 
95 #ifdef ADAPTIVE_RWLOCKS
96 static int rowner_retries = 10;
97 static int rowner_loops = 10000;
98 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
99     "rwlock debugging");
100 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
101 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
102 
103 static struct lock_delay_config __read_mostly rw_delay;
104 
105 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base,
106     0, "");
107 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
108     0, "");
109 
110 LOCK_DELAY_SYSINIT_DEFAULT(rw_delay);
111 #endif
112 
113 /*
114  * Return a pointer to the owning thread if the lock is write-locked or
115  * NULL if the lock is unlocked or read-locked.
116  */
117 
118 #define	lv_rw_wowner(v)							\
119 	((v) & RW_LOCK_READ ? NULL :					\
120 	 (struct thread *)RW_OWNER((v)))
121 
122 #define	rw_wowner(rw)	lv_rw_wowner(RW_READ_VALUE(rw))
123 
124 /*
125  * Returns if a write owner is recursed.  Write ownership is not assured
126  * here and should be previously checked.
127  */
128 #define	rw_recursed(rw)		((rw)->rw_recurse != 0)
129 
130 /*
131  * Return true if curthread helds the lock.
132  */
133 #define	rw_wlocked(rw)		(rw_wowner((rw)) == curthread)
134 
135 /*
136  * Return a pointer to the owning thread for this lock who should receive
137  * any priority lent by threads that block on this lock.  Currently this
138  * is identical to rw_wowner().
139  */
140 #define	rw_owner(rw)		rw_wowner(rw)
141 
142 #ifndef INVARIANTS
143 #define	__rw_assert(c, what, file, line)
144 #endif
145 
146 void
147 assert_rw(const struct lock_object *lock, int what)
148 {
149 
150 	rw_assert((const struct rwlock *)lock, what);
151 }
152 
153 void
154 lock_rw(struct lock_object *lock, uintptr_t how)
155 {
156 	struct rwlock *rw;
157 
158 	rw = (struct rwlock *)lock;
159 	if (how)
160 		rw_rlock(rw);
161 	else
162 		rw_wlock(rw);
163 }
164 
165 uintptr_t
166 unlock_rw(struct lock_object *lock)
167 {
168 	struct rwlock *rw;
169 
170 	rw = (struct rwlock *)lock;
171 	rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
172 	if (rw->rw_lock & RW_LOCK_READ) {
173 		rw_runlock(rw);
174 		return (1);
175 	} else {
176 		rw_wunlock(rw);
177 		return (0);
178 	}
179 }
180 
181 #ifdef KDTRACE_HOOKS
182 int
183 owner_rw(const struct lock_object *lock, struct thread **owner)
184 {
185 	const struct rwlock *rw = (const struct rwlock *)lock;
186 	uintptr_t x = rw->rw_lock;
187 
188 	*owner = rw_wowner(rw);
189 	return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
190 	    (*owner != NULL));
191 }
192 #endif
193 
194 void
195 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
196 {
197 	struct rwlock *rw;
198 	int flags;
199 
200 	rw = rwlock2rw(c);
201 
202 	MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
203 	    RW_RECURSE | RW_NEW)) == 0);
204 	ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
205 	    ("%s: rw_lock not aligned for %s: %p", __func__, name,
206 	    &rw->rw_lock));
207 
208 	flags = LO_UPGRADABLE;
209 	if (opts & RW_DUPOK)
210 		flags |= LO_DUPOK;
211 	if (opts & RW_NOPROFILE)
212 		flags |= LO_NOPROFILE;
213 	if (!(opts & RW_NOWITNESS))
214 		flags |= LO_WITNESS;
215 	if (opts & RW_RECURSE)
216 		flags |= LO_RECURSABLE;
217 	if (opts & RW_QUIET)
218 		flags |= LO_QUIET;
219 	if (opts & RW_NEW)
220 		flags |= LO_NEW;
221 
222 	lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
223 	rw->rw_lock = RW_UNLOCKED;
224 	rw->rw_recurse = 0;
225 }
226 
227 void
228 _rw_destroy(volatile uintptr_t *c)
229 {
230 	struct rwlock *rw;
231 
232 	rw = rwlock2rw(c);
233 
234 	KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
235 	KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
236 	rw->rw_lock = RW_DESTROYED;
237 	lock_destroy(&rw->lock_object);
238 }
239 
240 void
241 rw_sysinit(void *arg)
242 {
243 	struct rw_args *args = arg;
244 
245 	rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
246 }
247 
248 void
249 rw_sysinit_flags(void *arg)
250 {
251 	struct rw_args_flags *args = arg;
252 
253 	rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
254 	    args->ra_flags);
255 }
256 
257 int
258 _rw_wowned(const volatile uintptr_t *c)
259 {
260 
261 	return (rw_wowner(rwlock2rw(c)) == curthread);
262 }
263 
264 void
265 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
266 {
267 	struct rwlock *rw;
268 	uintptr_t tid, v;
269 
270 	rw = rwlock2rw(c);
271 
272 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
273 	    ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
274 	    curthread, rw->lock_object.lo_name, file, line));
275 	KASSERT(rw->rw_lock != RW_DESTROYED,
276 	    ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
277 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
278 	    line, NULL);
279 	tid = (uintptr_t)curthread;
280 	v = RW_UNLOCKED;
281 	if (!_rw_write_lock_fetch(rw, &v, tid))
282 		_rw_wlock_hard(rw, v, tid, file, line);
283 	else
284 		LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
285 		    0, 0, file, line, LOCKSTAT_WRITER);
286 
287 	LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
288 	WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
289 	TD_LOCKS_INC(curthread);
290 }
291 
292 int
293 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
294 {
295 	struct rwlock *rw;
296 	struct thread *td;
297 	uintptr_t tid, v;
298 	int rval;
299 	bool recursed;
300 
301 	td = curthread;
302 	tid = (uintptr_t)td;
303 	if (SCHEDULER_STOPPED_TD(td))
304 		return (1);
305 
306 	rw = rwlock2rw(c);
307 
308 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
309 	    ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
310 	    curthread, rw->lock_object.lo_name, file, line));
311 	KASSERT(rw->rw_lock != RW_DESTROYED,
312 	    ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
313 
314 	rval = 1;
315 	recursed = false;
316 	v = RW_UNLOCKED;
317 	for (;;) {
318 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
319 			break;
320 		if (v == RW_UNLOCKED)
321 			continue;
322 		if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
323 			rw->rw_recurse++;
324 			atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
325 			break;
326 		}
327 		rval = 0;
328 		break;
329 	}
330 
331 	LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
332 	if (rval) {
333 		WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
334 		    file, line);
335 		if (!recursed)
336 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
337 			    rw, 0, 0, file, line, LOCKSTAT_WRITER);
338 		TD_LOCKS_INC(curthread);
339 	}
340 	return (rval);
341 }
342 
343 void
344 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
345 {
346 	struct rwlock *rw;
347 
348 	rw = rwlock2rw(c);
349 
350 	KASSERT(rw->rw_lock != RW_DESTROYED,
351 	    ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
352 	__rw_assert(c, RA_WLOCKED, file, line);
353 	WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
354 	LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
355 	    line);
356 
357 #ifdef LOCK_PROFILING
358 	_rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
359 #else
360 	__rw_wunlock(rw, curthread, file, line);
361 #endif
362 
363 	TD_LOCKS_DEC(curthread);
364 }
365 
366 /*
367  * Determines whether a new reader can acquire a lock.  Succeeds if the
368  * reader already owns a read lock and the lock is locked for read to
369  * prevent deadlock from reader recursion.  Also succeeds if the lock
370  * is unlocked and has no writer waiters or spinners.  Failing otherwise
371  * prioritizes writers before readers.
372  */
373 #define	RW_CAN_READ(td, _rw)						\
374     (((td)->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &	\
375     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==	\
376     RW_LOCK_READ)
377 
378 static bool __always_inline
379 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp,
380     const char *file, int line)
381 {
382 
383 	/*
384 	 * Handle the easy case.  If no other thread has a write
385 	 * lock, then try to bump up the count of read locks.  Note
386 	 * that we have to preserve the current state of the
387 	 * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
388 	 * read lock, then rw_lock must have changed, so restart
389 	 * the loop.  Note that this handles the case of a
390 	 * completely unlocked rwlock since such a lock is encoded
391 	 * as a read lock with no waiters.
392 	 */
393 	while (RW_CAN_READ(td, *vp)) {
394 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
395 			*vp + RW_ONE_READER)) {
396 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
397 				CTR4(KTR_LOCK,
398 				    "%s: %p succeed %p -> %p", __func__,
399 				    rw, (void *)*vp,
400 				    (void *)(*vp + RW_ONE_READER));
401 			td->td_rw_rlocks++;
402 			return (true);
403 		}
404 	}
405 	return (false);
406 }
407 
408 static void __noinline
409 __rw_rlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
410     const char *file, int line)
411 {
412 	struct rwlock *rw;
413 	struct turnstile *ts;
414 #ifdef ADAPTIVE_RWLOCKS
415 	volatile struct thread *owner;
416 	int spintries = 0;
417 	int i;
418 #endif
419 #ifdef LOCK_PROFILING
420 	uint64_t waittime = 0;
421 	int contested = 0;
422 #endif
423 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
424 	struct lock_delay_arg lda;
425 #endif
426 #ifdef KDTRACE_HOOKS
427 	uintptr_t state;
428 	u_int sleep_cnt = 0;
429 	int64_t sleep_time = 0;
430 	int64_t all_time = 0;
431 #endif
432 
433 	if (SCHEDULER_STOPPED())
434 		return;
435 
436 #if defined(ADAPTIVE_RWLOCKS)
437 	lock_delay_arg_init(&lda, &rw_delay);
438 #elif defined(KDTRACE_HOOKS)
439 	lock_delay_arg_init(&lda, NULL);
440 #endif
441 	rw = rwlock2rw(c);
442 
443 #ifdef KDTRACE_HOOKS
444 	all_time -= lockstat_nsecs(&rw->lock_object);
445 #endif
446 #ifdef KDTRACE_HOOKS
447 	state = v;
448 #endif
449 	for (;;) {
450 		if (__rw_rlock_try(rw, td, &v, file, line))
451 			break;
452 #ifdef KDTRACE_HOOKS
453 		lda.spin_cnt++;
454 #endif
455 #ifdef HWPMC_HOOKS
456 		PMC_SOFT_CALL( , , lock, failed);
457 #endif
458 		lock_profile_obtain_lock_failed(&rw->lock_object,
459 		    &contested, &waittime);
460 
461 #ifdef ADAPTIVE_RWLOCKS
462 		/*
463 		 * If the owner is running on another CPU, spin until
464 		 * the owner stops running or the state of the lock
465 		 * changes.
466 		 */
467 		if ((v & RW_LOCK_READ) == 0) {
468 			owner = (struct thread *)RW_OWNER(v);
469 			if (TD_IS_RUNNING(owner)) {
470 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
471 					CTR3(KTR_LOCK,
472 					    "%s: spinning on %p held by %p",
473 					    __func__, rw, owner);
474 				KTR_STATE1(KTR_SCHED, "thread",
475 				    sched_tdname(curthread), "spinning",
476 				    "lockname:\"%s\"", rw->lock_object.lo_name);
477 				do {
478 					lock_delay(&lda);
479 					v = RW_READ_VALUE(rw);
480 					owner = lv_rw_wowner(v);
481 				} while (owner != NULL && TD_IS_RUNNING(owner));
482 				KTR_STATE0(KTR_SCHED, "thread",
483 				    sched_tdname(curthread), "running");
484 				continue;
485 			}
486 		} else if (spintries < rowner_retries) {
487 			spintries++;
488 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
489 			    "spinning", "lockname:\"%s\"",
490 			    rw->lock_object.lo_name);
491 			for (i = 0; i < rowner_loops; i++) {
492 				v = RW_READ_VALUE(rw);
493 				if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(td, v))
494 					break;
495 				cpu_spinwait();
496 			}
497 			v = RW_READ_VALUE(rw);
498 #ifdef KDTRACE_HOOKS
499 			lda.spin_cnt += rowner_loops - i;
500 #endif
501 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
502 			    "running");
503 			if (i != rowner_loops)
504 				continue;
505 		}
506 #endif
507 
508 		/*
509 		 * Okay, now it's the hard case.  Some other thread already
510 		 * has a write lock or there are write waiters present,
511 		 * acquire the turnstile lock so we can begin the process
512 		 * of blocking.
513 		 */
514 		ts = turnstile_trywait(&rw->lock_object);
515 
516 		/*
517 		 * The lock might have been released while we spun, so
518 		 * recheck its state and restart the loop if needed.
519 		 */
520 		v = RW_READ_VALUE(rw);
521 		if (RW_CAN_READ(td, v)) {
522 			turnstile_cancel(ts);
523 			continue;
524 		}
525 
526 #ifdef ADAPTIVE_RWLOCKS
527 		/*
528 		 * The current lock owner might have started executing
529 		 * on another CPU (or the lock could have changed
530 		 * owners) while we were waiting on the turnstile
531 		 * chain lock.  If so, drop the turnstile lock and try
532 		 * again.
533 		 */
534 		if ((v & RW_LOCK_READ) == 0) {
535 			owner = (struct thread *)RW_OWNER(v);
536 			if (TD_IS_RUNNING(owner)) {
537 				turnstile_cancel(ts);
538 				continue;
539 			}
540 		}
541 #endif
542 
543 		/*
544 		 * The lock is held in write mode or it already has waiters.
545 		 */
546 		MPASS(!RW_CAN_READ(td, v));
547 
548 		/*
549 		 * If the RW_LOCK_READ_WAITERS flag is already set, then
550 		 * we can go ahead and block.  If it is not set then try
551 		 * to set it.  If we fail to set it drop the turnstile
552 		 * lock and restart the loop.
553 		 */
554 		if (!(v & RW_LOCK_READ_WAITERS)) {
555 			if (!atomic_cmpset_ptr(&rw->rw_lock, v,
556 			    v | RW_LOCK_READ_WAITERS)) {
557 				turnstile_cancel(ts);
558 				v = RW_READ_VALUE(rw);
559 				continue;
560 			}
561 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
562 				CTR2(KTR_LOCK, "%s: %p set read waiters flag",
563 				    __func__, rw);
564 		}
565 
566 		/*
567 		 * We were unable to acquire the lock and the read waiters
568 		 * flag is set, so we must block on the turnstile.
569 		 */
570 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
571 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
572 			    rw);
573 #ifdef KDTRACE_HOOKS
574 		sleep_time -= lockstat_nsecs(&rw->lock_object);
575 #endif
576 		turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
577 #ifdef KDTRACE_HOOKS
578 		sleep_time += lockstat_nsecs(&rw->lock_object);
579 		sleep_cnt++;
580 #endif
581 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
582 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
583 			    __func__, rw);
584 		v = RW_READ_VALUE(rw);
585 	}
586 #ifdef KDTRACE_HOOKS
587 	all_time += lockstat_nsecs(&rw->lock_object);
588 	if (sleep_time)
589 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
590 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
591 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
592 
593 	/* Record only the loops spinning and not sleeping. */
594 	if (lda.spin_cnt > sleep_cnt)
595 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
596 		    LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
597 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
598 #endif
599 	/*
600 	 * TODO: acquire "owner of record" here.  Here be turnstile dragons
601 	 * however.  turnstiles don't like owners changing between calls to
602 	 * turnstile_wait() currently.
603 	 */
604 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
605 	    waittime, file, line, LOCKSTAT_READER);
606 }
607 
608 void
609 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
610 {
611 	struct rwlock *rw;
612 	struct thread *td;
613 	uintptr_t v;
614 
615 	td = curthread;
616 	rw = rwlock2rw(c);
617 
618 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
619 	    ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
620 	    td, rw->lock_object.lo_name, file, line));
621 	KASSERT(rw->rw_lock != RW_DESTROYED,
622 	    ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
623 	KASSERT(rw_wowner(rw) != td,
624 	    ("rw_rlock: wlock already held for %s @ %s:%d",
625 	    rw->lock_object.lo_name, file, line));
626 	WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
627 
628 	v = RW_READ_VALUE(rw);
629 	if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__acquire) ||
630 	    !__rw_rlock_try(rw, td, &v, file, line)))
631 		__rw_rlock_hard(c, td, v, file, line);
632 
633 	LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
634 	WITNESS_LOCK(&rw->lock_object, 0, file, line);
635 	TD_LOCKS_INC(curthread);
636 }
637 
638 int
639 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
640 {
641 	struct rwlock *rw;
642 	uintptr_t x;
643 
644 	if (SCHEDULER_STOPPED())
645 		return (1);
646 
647 	rw = rwlock2rw(c);
648 
649 	KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
650 	    ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
651 	    curthread, rw->lock_object.lo_name, file, line));
652 
653 	x = rw->rw_lock;
654 	for (;;) {
655 		KASSERT(rw->rw_lock != RW_DESTROYED,
656 		    ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
657 		if (!(x & RW_LOCK_READ))
658 			break;
659 		if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
660 			LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
661 			    line);
662 			WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
663 			LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
664 			    rw, 0, 0, file, line, LOCKSTAT_READER);
665 			TD_LOCKS_INC(curthread);
666 			curthread->td_rw_rlocks++;
667 			return (1);
668 		}
669 	}
670 
671 	LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
672 	return (0);
673 }
674 
675 static bool __always_inline
676 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
677 {
678 
679 	for (;;) {
680 		/*
681 		 * See if there is more than one read lock held.  If so,
682 		 * just drop one and return.
683 		 */
684 		if (RW_READERS(*vp) > 1) {
685 			if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
686 			    *vp - RW_ONE_READER)) {
687 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
688 					CTR4(KTR_LOCK,
689 					    "%s: %p succeeded %p -> %p",
690 					    __func__, rw, (void *)*vp,
691 					    (void *)(*vp - RW_ONE_READER));
692 				td->td_rw_rlocks--;
693 				return (true);
694 			}
695 			continue;
696 		}
697 		/*
698 		 * If there aren't any waiters for a write lock, then try
699 		 * to drop it quickly.
700 		 */
701 		if (!(*vp & RW_LOCK_WAITERS)) {
702 			MPASS((*vp & ~RW_LOCK_WRITE_SPINNER) ==
703 			    RW_READERS_LOCK(1));
704 			if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
705 			    RW_UNLOCKED)) {
706 				if (LOCK_LOG_TEST(&rw->lock_object, 0))
707 					CTR2(KTR_LOCK, "%s: %p last succeeded",
708 					    __func__, rw);
709 				td->td_rw_rlocks--;
710 				return (true);
711 			}
712 			continue;
713 		}
714 		break;
715 	}
716 	return (false);
717 }
718 
719 static void __noinline
720 __rw_runlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
721     const char *file, int line)
722 {
723 	struct rwlock *rw;
724 	struct turnstile *ts;
725 	uintptr_t x, queue;
726 
727 	if (SCHEDULER_STOPPED())
728 		return;
729 
730 	rw = rwlock2rw(c);
731 
732 	for (;;) {
733 		if (__rw_runlock_try(rw, td, &v))
734 			break;
735 
736 		/*
737 		 * Ok, we know we have waiters and we think we are the
738 		 * last reader, so grab the turnstile lock.
739 		 */
740 		turnstile_chain_lock(&rw->lock_object);
741 		v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
742 		MPASS(v & RW_LOCK_WAITERS);
743 
744 		/*
745 		 * Try to drop our lock leaving the lock in a unlocked
746 		 * state.
747 		 *
748 		 * If you wanted to do explicit lock handoff you'd have to
749 		 * do it here.  You'd also want to use turnstile_signal()
750 		 * and you'd have to handle the race where a higher
751 		 * priority thread blocks on the write lock before the
752 		 * thread you wakeup actually runs and have the new thread
753 		 * "steal" the lock.  For now it's a lot simpler to just
754 		 * wakeup all of the waiters.
755 		 *
756 		 * As above, if we fail, then another thread might have
757 		 * acquired a read lock, so drop the turnstile lock and
758 		 * restart.
759 		 */
760 		x = RW_UNLOCKED;
761 		if (v & RW_LOCK_WRITE_WAITERS) {
762 			queue = TS_EXCLUSIVE_QUEUE;
763 			x |= (v & RW_LOCK_READ_WAITERS);
764 		} else
765 			queue = TS_SHARED_QUEUE;
766 		if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
767 		    x)) {
768 			turnstile_chain_unlock(&rw->lock_object);
769 			v = RW_READ_VALUE(rw);
770 			continue;
771 		}
772 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
773 			CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
774 			    __func__, rw);
775 
776 		/*
777 		 * Ok.  The lock is released and all that's left is to
778 		 * wake up the waiters.  Note that the lock might not be
779 		 * free anymore, but in that case the writers will just
780 		 * block again if they run before the new lock holder(s)
781 		 * release the lock.
782 		 */
783 		ts = turnstile_lookup(&rw->lock_object);
784 		MPASS(ts != NULL);
785 		turnstile_broadcast(ts, queue);
786 		turnstile_unpend(ts, TS_SHARED_LOCK);
787 		turnstile_chain_unlock(&rw->lock_object);
788 		td->td_rw_rlocks--;
789 		break;
790 	}
791 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
792 }
793 
794 void
795 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
796 {
797 	struct rwlock *rw;
798 	struct thread *td;
799 	uintptr_t v;
800 
801 	rw = rwlock2rw(c);
802 
803 	KASSERT(rw->rw_lock != RW_DESTROYED,
804 	    ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
805 	__rw_assert(c, RA_RLOCKED, file, line);
806 	WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
807 	LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
808 
809 	td = curthread;
810 	v = RW_READ_VALUE(rw);
811 
812 	if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__release) ||
813 	    !__rw_runlock_try(rw, td, &v)))
814 		__rw_runlock_hard(c, td, v, file, line);
815 
816 	TD_LOCKS_DEC(curthread);
817 }
818 
819 
820 /*
821  * This function is called when we are unable to obtain a write lock on the
822  * first try.  This means that at least one other thread holds either a
823  * read or write lock.
824  */
825 void
826 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v, uintptr_t tid,
827     const char *file, int line)
828 {
829 	struct rwlock *rw;
830 	struct turnstile *ts;
831 #ifdef ADAPTIVE_RWLOCKS
832 	volatile struct thread *owner;
833 	int spintries = 0;
834 	int i;
835 #endif
836 	uintptr_t x;
837 #ifdef LOCK_PROFILING
838 	uint64_t waittime = 0;
839 	int contested = 0;
840 #endif
841 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
842 	struct lock_delay_arg lda;
843 #endif
844 #ifdef KDTRACE_HOOKS
845 	uintptr_t state;
846 	u_int sleep_cnt = 0;
847 	int64_t sleep_time = 0;
848 	int64_t all_time = 0;
849 #endif
850 
851 	if (SCHEDULER_STOPPED())
852 		return;
853 
854 #if defined(ADAPTIVE_RWLOCKS)
855 	lock_delay_arg_init(&lda, &rw_delay);
856 #elif defined(KDTRACE_HOOKS)
857 	lock_delay_arg_init(&lda, NULL);
858 #endif
859 	rw = rwlock2rw(c);
860 	if (__predict_false(v == RW_UNLOCKED))
861 		v = RW_READ_VALUE(rw);
862 
863 	if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
864 		KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
865 		    ("%s: recursing but non-recursive rw %s @ %s:%d\n",
866 		    __func__, rw->lock_object.lo_name, file, line));
867 		rw->rw_recurse++;
868 		atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
869 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
870 			CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
871 		return;
872 	}
873 
874 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
875 		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
876 		    rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
877 
878 #ifdef KDTRACE_HOOKS
879 	all_time -= lockstat_nsecs(&rw->lock_object);
880 	state = v;
881 #endif
882 	for (;;) {
883 		if (v == RW_UNLOCKED) {
884 			if (_rw_write_lock_fetch(rw, &v, tid))
885 				break;
886 			continue;
887 		}
888 #ifdef KDTRACE_HOOKS
889 		lda.spin_cnt++;
890 #endif
891 #ifdef HWPMC_HOOKS
892 		PMC_SOFT_CALL( , , lock, failed);
893 #endif
894 		lock_profile_obtain_lock_failed(&rw->lock_object,
895 		    &contested, &waittime);
896 #ifdef ADAPTIVE_RWLOCKS
897 		/*
898 		 * If the lock is write locked and the owner is
899 		 * running on another CPU, spin until the owner stops
900 		 * running or the state of the lock changes.
901 		 */
902 		owner = lv_rw_wowner(v);
903 		if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
904 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
905 				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
906 				    __func__, rw, owner);
907 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
908 			    "spinning", "lockname:\"%s\"",
909 			    rw->lock_object.lo_name);
910 			do {
911 				lock_delay(&lda);
912 				v = RW_READ_VALUE(rw);
913 				owner = lv_rw_wowner(v);
914 			} while (owner != NULL && TD_IS_RUNNING(owner));
915 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
916 			    "running");
917 			continue;
918 		}
919 		if ((v & RW_LOCK_READ) && RW_READERS(v) &&
920 		    spintries < rowner_retries) {
921 			if (!(v & RW_LOCK_WRITE_SPINNER)) {
922 				if (!atomic_cmpset_ptr(&rw->rw_lock, v,
923 				    v | RW_LOCK_WRITE_SPINNER)) {
924 					v = RW_READ_VALUE(rw);
925 					continue;
926 				}
927 			}
928 			spintries++;
929 			KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
930 			    "spinning", "lockname:\"%s\"",
931 			    rw->lock_object.lo_name);
932 			for (i = 0; i < rowner_loops; i++) {
933 				if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
934 					break;
935 				cpu_spinwait();
936 			}
937 			KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
938 			    "running");
939 			v = RW_READ_VALUE(rw);
940 #ifdef KDTRACE_HOOKS
941 			lda.spin_cnt += rowner_loops - i;
942 #endif
943 			if (i != rowner_loops)
944 				continue;
945 		}
946 #endif
947 		ts = turnstile_trywait(&rw->lock_object);
948 		v = RW_READ_VALUE(rw);
949 
950 #ifdef ADAPTIVE_RWLOCKS
951 		/*
952 		 * The current lock owner might have started executing
953 		 * on another CPU (or the lock could have changed
954 		 * owners) while we were waiting on the turnstile
955 		 * chain lock.  If so, drop the turnstile lock and try
956 		 * again.
957 		 */
958 		if (!(v & RW_LOCK_READ)) {
959 			owner = (struct thread *)RW_OWNER(v);
960 			if (TD_IS_RUNNING(owner)) {
961 				turnstile_cancel(ts);
962 				continue;
963 			}
964 		}
965 #endif
966 		/*
967 		 * Check for the waiters flags about this rwlock.
968 		 * If the lock was released, without maintain any pending
969 		 * waiters queue, simply try to acquire it.
970 		 * If a pending waiters queue is present, claim the lock
971 		 * ownership and maintain the pending queue.
972 		 */
973 		x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
974 		if ((v & ~x) == RW_UNLOCKED) {
975 			x &= ~RW_LOCK_WRITE_SPINNER;
976 			if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
977 				if (x)
978 					turnstile_claim(ts);
979 				else
980 					turnstile_cancel(ts);
981 				break;
982 			}
983 			turnstile_cancel(ts);
984 			v = RW_READ_VALUE(rw);
985 			continue;
986 		}
987 		/*
988 		 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
989 		 * set it.  If we fail to set it, then loop back and try
990 		 * again.
991 		 */
992 		if (!(v & RW_LOCK_WRITE_WAITERS)) {
993 			if (!atomic_cmpset_ptr(&rw->rw_lock, v,
994 			    v | RW_LOCK_WRITE_WAITERS)) {
995 				turnstile_cancel(ts);
996 				v = RW_READ_VALUE(rw);
997 				continue;
998 			}
999 			if (LOCK_LOG_TEST(&rw->lock_object, 0))
1000 				CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1001 				    __func__, rw);
1002 		}
1003 		/*
1004 		 * We were unable to acquire the lock and the write waiters
1005 		 * flag is set, so we must block on the turnstile.
1006 		 */
1007 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1008 			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1009 			    rw);
1010 #ifdef KDTRACE_HOOKS
1011 		sleep_time -= lockstat_nsecs(&rw->lock_object);
1012 #endif
1013 		turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
1014 #ifdef KDTRACE_HOOKS
1015 		sleep_time += lockstat_nsecs(&rw->lock_object);
1016 		sleep_cnt++;
1017 #endif
1018 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1019 			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1020 			    __func__, rw);
1021 #ifdef ADAPTIVE_RWLOCKS
1022 		spintries = 0;
1023 #endif
1024 		v = RW_READ_VALUE(rw);
1025 	}
1026 #ifdef KDTRACE_HOOKS
1027 	all_time += lockstat_nsecs(&rw->lock_object);
1028 	if (sleep_time)
1029 		LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1030 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1031 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1032 
1033 	/* Record only the loops spinning and not sleeping. */
1034 	if (lda.spin_cnt > sleep_cnt)
1035 		LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1036 		    LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1037 		    (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1038 #endif
1039 	LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1040 	    waittime, file, line, LOCKSTAT_WRITER);
1041 }
1042 
1043 /*
1044  * This function is called if lockstat is active or the first try at releasing
1045  * a write lock failed.  The latter means that the lock is recursed or one of
1046  * the 2 waiter bits must be set indicating that at least one thread is waiting
1047  * on this lock.
1048  */
1049 void
1050 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
1051     int line)
1052 {
1053 	struct rwlock *rw;
1054 	struct turnstile *ts;
1055 	uintptr_t v;
1056 	int queue;
1057 
1058 	if (SCHEDULER_STOPPED())
1059 		return;
1060 
1061 	rw = rwlock2rw(c);
1062 	v = RW_READ_VALUE(rw);
1063 	if (v & RW_LOCK_WRITER_RECURSED) {
1064 		if (--(rw->rw_recurse) == 0)
1065 			atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1066 		if (LOCK_LOG_TEST(&rw->lock_object, 0))
1067 			CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1068 		return;
1069 	}
1070 
1071 	LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1072 	if (v == tid && _rw_write_unlock(rw, tid))
1073 		return;
1074 
1075 	KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1076 	    ("%s: neither of the waiter flags are set", __func__));
1077 
1078 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1079 		CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1080 
1081 	turnstile_chain_lock(&rw->lock_object);
1082 	ts = turnstile_lookup(&rw->lock_object);
1083 	MPASS(ts != NULL);
1084 
1085 	/*
1086 	 * Use the same algo as sx locks for now.  Prefer waking up shared
1087 	 * waiters if we have any over writers.  This is probably not ideal.
1088 	 *
1089 	 * 'v' is the value we are going to write back to rw_lock.  If we
1090 	 * have waiters on both queues, we need to preserve the state of
1091 	 * the waiter flag for the queue we don't wake up.  For now this is
1092 	 * hardcoded for the algorithm mentioned above.
1093 	 *
1094 	 * In the case of both readers and writers waiting we wakeup the
1095 	 * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1096 	 * new writer comes in before a reader it will claim the lock up
1097 	 * above.  There is probably a potential priority inversion in
1098 	 * there that could be worked around either by waking both queues
1099 	 * of waiters or doing some complicated lock handoff gymnastics.
1100 	 */
1101 	v = RW_UNLOCKED;
1102 	if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1103 		queue = TS_EXCLUSIVE_QUEUE;
1104 		v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1105 	} else
1106 		queue = TS_SHARED_QUEUE;
1107 
1108 	/* Wake up all waiters for the specific queue. */
1109 	if (LOCK_LOG_TEST(&rw->lock_object, 0))
1110 		CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1111 		    queue == TS_SHARED_QUEUE ? "read" : "write");
1112 	turnstile_broadcast(ts, queue);
1113 	atomic_store_rel_ptr(&rw->rw_lock, v);
1114 	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1115 	turnstile_chain_unlock(&rw->lock_object);
1116 }
1117 
1118 /*
1119  * Attempt to do a non-blocking upgrade from a read lock to a write
1120  * lock.  This will only succeed if this thread holds a single read
1121  * lock.  Returns true if the upgrade succeeded and false otherwise.
1122  */
1123 int
1124 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1125 {
1126 	struct rwlock *rw;
1127 	uintptr_t v, x, tid;
1128 	struct turnstile *ts;
1129 	int success;
1130 
1131 	if (SCHEDULER_STOPPED())
1132 		return (1);
1133 
1134 	rw = rwlock2rw(c);
1135 
1136 	KASSERT(rw->rw_lock != RW_DESTROYED,
1137 	    ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1138 	__rw_assert(c, RA_RLOCKED, file, line);
1139 
1140 	/*
1141 	 * Attempt to switch from one reader to a writer.  If there
1142 	 * are any write waiters, then we will have to lock the
1143 	 * turnstile first to prevent races with another writer
1144 	 * calling turnstile_wait() before we have claimed this
1145 	 * turnstile.  So, do the simple case of no waiters first.
1146 	 */
1147 	tid = (uintptr_t)curthread;
1148 	success = 0;
1149 	for (;;) {
1150 		v = rw->rw_lock;
1151 		if (RW_READERS(v) > 1)
1152 			break;
1153 		if (!(v & RW_LOCK_WAITERS)) {
1154 			success = atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid);
1155 			if (!success)
1156 				continue;
1157 			break;
1158 		}
1159 
1160 		/*
1161 		 * Ok, we think we have waiters, so lock the turnstile.
1162 		 */
1163 		ts = turnstile_trywait(&rw->lock_object);
1164 		v = rw->rw_lock;
1165 		if (RW_READERS(v) > 1) {
1166 			turnstile_cancel(ts);
1167 			break;
1168 		}
1169 		/*
1170 		 * Try to switch from one reader to a writer again.  This time
1171 		 * we honor the current state of the waiters flags.
1172 		 * If we obtain the lock with the flags set, then claim
1173 		 * ownership of the turnstile.
1174 		 */
1175 		x = rw->rw_lock & RW_LOCK_WAITERS;
1176 		success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1177 		if (success) {
1178 			if (x)
1179 				turnstile_claim(ts);
1180 			else
1181 				turnstile_cancel(ts);
1182 			break;
1183 		}
1184 		turnstile_cancel(ts);
1185 	}
1186 	LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1187 	if (success) {
1188 		curthread->td_rw_rlocks--;
1189 		WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1190 		    file, line);
1191 		LOCKSTAT_RECORD0(rw__upgrade, rw);
1192 	}
1193 	return (success);
1194 }
1195 
1196 /*
1197  * Downgrade a write lock into a single read lock.
1198  */
1199 void
1200 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1201 {
1202 	struct rwlock *rw;
1203 	struct turnstile *ts;
1204 	uintptr_t tid, v;
1205 	int rwait, wwait;
1206 
1207 	if (SCHEDULER_STOPPED())
1208 		return;
1209 
1210 	rw = rwlock2rw(c);
1211 
1212 	KASSERT(rw->rw_lock != RW_DESTROYED,
1213 	    ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1214 	__rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1215 #ifndef INVARIANTS
1216 	if (rw_recursed(rw))
1217 		panic("downgrade of a recursed lock");
1218 #endif
1219 
1220 	WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1221 
1222 	/*
1223 	 * Convert from a writer to a single reader.  First we handle
1224 	 * the easy case with no waiters.  If there are any waiters, we
1225 	 * lock the turnstile and "disown" the lock.
1226 	 */
1227 	tid = (uintptr_t)curthread;
1228 	if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1229 		goto out;
1230 
1231 	/*
1232 	 * Ok, we think we have waiters, so lock the turnstile so we can
1233 	 * read the waiter flags without any races.
1234 	 */
1235 	turnstile_chain_lock(&rw->lock_object);
1236 	v = rw->rw_lock & RW_LOCK_WAITERS;
1237 	rwait = v & RW_LOCK_READ_WAITERS;
1238 	wwait = v & RW_LOCK_WRITE_WAITERS;
1239 	MPASS(rwait | wwait);
1240 
1241 	/*
1242 	 * Downgrade from a write lock while preserving waiters flag
1243 	 * and give up ownership of the turnstile.
1244 	 */
1245 	ts = turnstile_lookup(&rw->lock_object);
1246 	MPASS(ts != NULL);
1247 	if (!wwait)
1248 		v &= ~RW_LOCK_READ_WAITERS;
1249 	atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1250 	/*
1251 	 * Wake other readers if there are no writers pending.  Otherwise they
1252 	 * won't be able to acquire the lock anyway.
1253 	 */
1254 	if (rwait && !wwait) {
1255 		turnstile_broadcast(ts, TS_SHARED_QUEUE);
1256 		turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1257 	} else
1258 		turnstile_disown(ts);
1259 	turnstile_chain_unlock(&rw->lock_object);
1260 out:
1261 	curthread->td_rw_rlocks++;
1262 	LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1263 	LOCKSTAT_RECORD0(rw__downgrade, rw);
1264 }
1265 
1266 #ifdef INVARIANT_SUPPORT
1267 #ifndef INVARIANTS
1268 #undef __rw_assert
1269 #endif
1270 
1271 /*
1272  * In the non-WITNESS case, rw_assert() can only detect that at least
1273  * *some* thread owns an rlock, but it cannot guarantee that *this*
1274  * thread owns an rlock.
1275  */
1276 void
1277 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1278 {
1279 	const struct rwlock *rw;
1280 
1281 	if (panicstr != NULL)
1282 		return;
1283 
1284 	rw = rwlock2rw(c);
1285 
1286 	switch (what) {
1287 	case RA_LOCKED:
1288 	case RA_LOCKED | RA_RECURSED:
1289 	case RA_LOCKED | RA_NOTRECURSED:
1290 	case RA_RLOCKED:
1291 	case RA_RLOCKED | RA_RECURSED:
1292 	case RA_RLOCKED | RA_NOTRECURSED:
1293 #ifdef WITNESS
1294 		witness_assert(&rw->lock_object, what, file, line);
1295 #else
1296 		/*
1297 		 * If some other thread has a write lock or we have one
1298 		 * and are asserting a read lock, fail.  Also, if no one
1299 		 * has a lock at all, fail.
1300 		 */
1301 		if (rw->rw_lock == RW_UNLOCKED ||
1302 		    (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1303 		    rw_wowner(rw) != curthread)))
1304 			panic("Lock %s not %slocked @ %s:%d\n",
1305 			    rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1306 			    "read " : "", file, line);
1307 
1308 		if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1309 			if (rw_recursed(rw)) {
1310 				if (what & RA_NOTRECURSED)
1311 					panic("Lock %s recursed @ %s:%d\n",
1312 					    rw->lock_object.lo_name, file,
1313 					    line);
1314 			} else if (what & RA_RECURSED)
1315 				panic("Lock %s not recursed @ %s:%d\n",
1316 				    rw->lock_object.lo_name, file, line);
1317 		}
1318 #endif
1319 		break;
1320 	case RA_WLOCKED:
1321 	case RA_WLOCKED | RA_RECURSED:
1322 	case RA_WLOCKED | RA_NOTRECURSED:
1323 		if (rw_wowner(rw) != curthread)
1324 			panic("Lock %s not exclusively locked @ %s:%d\n",
1325 			    rw->lock_object.lo_name, file, line);
1326 		if (rw_recursed(rw)) {
1327 			if (what & RA_NOTRECURSED)
1328 				panic("Lock %s recursed @ %s:%d\n",
1329 				    rw->lock_object.lo_name, file, line);
1330 		} else if (what & RA_RECURSED)
1331 			panic("Lock %s not recursed @ %s:%d\n",
1332 			    rw->lock_object.lo_name, file, line);
1333 		break;
1334 	case RA_UNLOCKED:
1335 #ifdef WITNESS
1336 		witness_assert(&rw->lock_object, what, file, line);
1337 #else
1338 		/*
1339 		 * If we hold a write lock fail.  We can't reliably check
1340 		 * to see if we hold a read lock or not.
1341 		 */
1342 		if (rw_wowner(rw) == curthread)
1343 			panic("Lock %s exclusively locked @ %s:%d\n",
1344 			    rw->lock_object.lo_name, file, line);
1345 #endif
1346 		break;
1347 	default:
1348 		panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1349 		    line);
1350 	}
1351 }
1352 #endif /* INVARIANT_SUPPORT */
1353 
1354 #ifdef DDB
1355 void
1356 db_show_rwlock(const struct lock_object *lock)
1357 {
1358 	const struct rwlock *rw;
1359 	struct thread *td;
1360 
1361 	rw = (const struct rwlock *)lock;
1362 
1363 	db_printf(" state: ");
1364 	if (rw->rw_lock == RW_UNLOCKED)
1365 		db_printf("UNLOCKED\n");
1366 	else if (rw->rw_lock == RW_DESTROYED) {
1367 		db_printf("DESTROYED\n");
1368 		return;
1369 	} else if (rw->rw_lock & RW_LOCK_READ)
1370 		db_printf("RLOCK: %ju locks\n",
1371 		    (uintmax_t)(RW_READERS(rw->rw_lock)));
1372 	else {
1373 		td = rw_wowner(rw);
1374 		db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1375 		    td->td_tid, td->td_proc->p_pid, td->td_name);
1376 		if (rw_recursed(rw))
1377 			db_printf(" recursed: %u\n", rw->rw_recurse);
1378 	}
1379 	db_printf(" waiters: ");
1380 	switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1381 	case RW_LOCK_READ_WAITERS:
1382 		db_printf("readers\n");
1383 		break;
1384 	case RW_LOCK_WRITE_WAITERS:
1385 		db_printf("writers\n");
1386 		break;
1387 	case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1388 		db_printf("readers and writers\n");
1389 		break;
1390 	default:
1391 		db_printf("none\n");
1392 		break;
1393 	}
1394 }
1395 
1396 #endif
1397