kern_rwlock.c revision 157846
1154941Sjhb/*-
2154941Sjhb * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
3154941Sjhb * All rights reserved.
4154941Sjhb *
5154941Sjhb * Redistribution and use in source and binary forms, with or without
6154941Sjhb * modification, are permitted provided that the following conditions
7154941Sjhb * are met:
8154941Sjhb * 1. Redistributions of source code must retain the above copyright
9154941Sjhb *    notice, this list of conditions and the following disclaimer.
10154941Sjhb * 2. Redistributions in binary form must reproduce the above copyright
11154941Sjhb *    notice, this list of conditions and the following disclaimer in the
12154941Sjhb *    documentation and/or other materials provided with the distribution.
13154941Sjhb * 3. Neither the name of the author nor the names of any co-contributors
14154941Sjhb *    may be used to endorse or promote products derived from this software
15154941Sjhb *    without specific prior written permission.
16154941Sjhb *
17154941Sjhb * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18154941Sjhb * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19154941Sjhb * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20154941Sjhb * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21154941Sjhb * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22154941Sjhb * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23154941Sjhb * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24154941Sjhb * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25154941Sjhb * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26154941Sjhb * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27154941Sjhb * SUCH DAMAGE.
28154941Sjhb */
29154941Sjhb
30154941Sjhb/*
31154941Sjhb * Machine independent bits of reader/writer lock implementation.
32154941Sjhb */
33154941Sjhb
34154941Sjhb#include <sys/cdefs.h>
35154941Sjhb__FBSDID("$FreeBSD: head/sys/kern/kern_rwlock.c 157846 2006-04-18 18:27:54Z jhb $");
36154941Sjhb
37154941Sjhb#include "opt_ddb.h"
38154941Sjhb
39154941Sjhb#include <sys/param.h>
40154941Sjhb#include <sys/ktr.h>
41154941Sjhb#include <sys/lock.h>
42154941Sjhb#include <sys/mutex.h>
43154941Sjhb#include <sys/proc.h>
44154941Sjhb#include <sys/rwlock.h>
45154941Sjhb#include <sys/systm.h>
46154941Sjhb#include <sys/turnstile.h>
47154941Sjhb
48154941Sjhb#include <machine/cpu.h>
49154941Sjhb
50154941Sjhb#ifdef DDB
51154941Sjhb#include <ddb/ddb.h>
52154941Sjhb
53154941Sjhbstatic void	db_show_rwlock(struct lock_object *lock);
54154941Sjhb#endif
55154941Sjhb
56154941Sjhbstruct lock_class lock_class_rw = {
57154941Sjhb	"rw",
58154941Sjhb	LC_SLEEPLOCK | LC_RECURSABLE /* | LC_UPGRADABLE */,
59154941Sjhb#ifdef DDB
60154941Sjhb	db_show_rwlock
61154941Sjhb#endif
62154941Sjhb};
63154941Sjhb
64157826Sjhb/*
65157826Sjhb * Return a pointer to the owning thread if the lock is write-locked or
66157826Sjhb * NULL if the lock is unlocked or read-locked.
67157826Sjhb */
68157826Sjhb#define	rw_wowner(rw)							\
69154941Sjhb	((rw)->rw_lock & RW_LOCK_READ ? NULL :				\
70154941Sjhb	    (struct thread *)RW_OWNER((rw)->rw_lock))
71154941Sjhb
72157826Sjhb/*
73157826Sjhb * Return a pointer to the owning thread for this lock who should receive
74157826Sjhb * any priority lent by threads that block on this lock.  Currently this
75157826Sjhb * is identical to rw_wowner().
76157826Sjhb */
77157826Sjhb#define	rw_owner(rw)		rw_wowner(rw)
78157826Sjhb
79154941Sjhb#ifndef INVARIANTS
80154941Sjhb#define	_rw_assert(rw, what, file, line)
81154941Sjhb#endif
82154941Sjhb
83154941Sjhbvoid
84154941Sjhbrw_init(struct rwlock *rw, const char *name)
85154941Sjhb{
86154941Sjhb
87154941Sjhb	rw->rw_lock = RW_UNLOCKED;
88154941Sjhb
89154941Sjhb	lock_init(&rw->rw_object, &lock_class_rw, name, NULL, LO_WITNESS |
90154941Sjhb	    LO_RECURSABLE /* | LO_UPGRADABLE */);
91154941Sjhb}
92154941Sjhb
93154941Sjhbvoid
94154941Sjhbrw_destroy(struct rwlock *rw)
95154941Sjhb{
96154941Sjhb
97154941Sjhb	KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked"));
98154941Sjhb	lock_destroy(&rw->rw_object);
99154941Sjhb}
100154941Sjhb
101154941Sjhbvoid
102154941Sjhbrw_sysinit(void *arg)
103154941Sjhb{
104154941Sjhb	struct rw_args *args = arg;
105154941Sjhb
106154941Sjhb	rw_init(args->ra_rw, args->ra_desc);
107154941Sjhb}
108154941Sjhb
109154941Sjhbvoid
110154941Sjhb_rw_wlock(struct rwlock *rw, const char *file, int line)
111154941Sjhb{
112154941Sjhb
113154941Sjhb	MPASS(curthread != NULL);
114157826Sjhb	KASSERT(rw_wowner(rw) != curthread,
115154941Sjhb	    ("%s (%s): wlock already held @ %s:%d", __func__,
116154941Sjhb	    rw->rw_object.lo_name, file, line));
117154941Sjhb	WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
118154941Sjhb	    line);
119154941Sjhb	__rw_wlock(rw, curthread, file, line);
120154941Sjhb	LOCK_LOG_LOCK("WLOCK", &rw->rw_object, 0, 0, file, line);
121154941Sjhb	WITNESS_LOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line);
122154941Sjhb}
123154941Sjhb
124154941Sjhbvoid
125154941Sjhb_rw_wunlock(struct rwlock *rw, const char *file, int line)
126154941Sjhb{
127154941Sjhb
128154941Sjhb	MPASS(curthread != NULL);
129154941Sjhb	_rw_assert(rw, RA_WLOCKED, file, line);
130154941Sjhb	WITNESS_UNLOCK(&rw->rw_object, LOP_EXCLUSIVE, file, line);
131154941Sjhb	LOCK_LOG_LOCK("WUNLOCK", &rw->rw_object, 0, 0, file, line);
132154941Sjhb	__rw_wunlock(rw, curthread, file, line);
133154941Sjhb}
134154941Sjhb
135154941Sjhbvoid
136154941Sjhb_rw_rlock(struct rwlock *rw, const char *file, int line)
137154941Sjhb{
138157846Sjhb	volatile struct thread *owner;
139154941Sjhb	uintptr_t x;
140154941Sjhb
141157826Sjhb	KASSERT(rw_wowner(rw) != curthread,
142154941Sjhb	    ("%s (%s): wlock already held @ %s:%d", __func__,
143154941Sjhb	    rw->rw_object.lo_name, file, line));
144154941Sjhb	WITNESS_CHECKORDER(&rw->rw_object, LOP_NEWORDER, file, line);
145154941Sjhb
146154941Sjhb	/*
147154941Sjhb	 * Note that we don't make any attempt to try to block read
148154941Sjhb	 * locks once a writer has blocked on the lock.  The reason is
149154941Sjhb	 * that we currently allow for read locks to recurse and we
150154941Sjhb	 * don't keep track of all the holders of read locks.  Thus, if
151154941Sjhb	 * we were to block readers once a writer blocked and a reader
152154941Sjhb	 * tried to recurse on their reader lock after a writer had
153154941Sjhb	 * blocked we would end up in a deadlock since the reader would
154154941Sjhb	 * be blocked on the writer, and the writer would be blocked
155154941Sjhb	 * waiting for the reader to release its original read lock.
156154941Sjhb	 */
157154941Sjhb	for (;;) {
158154941Sjhb		/*
159154941Sjhb		 * Handle the easy case.  If no other thread has a write
160154941Sjhb		 * lock, then try to bump up the count of read locks.  Note
161154941Sjhb		 * that we have to preserve the current state of the
162154941Sjhb		 * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
163154941Sjhb		 * read lock, then rw_lock must have changed, so restart
164154941Sjhb		 * the loop.  Note that this handles the case of a
165154941Sjhb		 * completely unlocked rwlock since such a lock is encoded
166154941Sjhb		 * as a read lock with no waiters.
167154941Sjhb		 */
168154941Sjhb		x = rw->rw_lock;
169154941Sjhb		if (x & RW_LOCK_READ) {
170154941Sjhb
171154941Sjhb			/*
172154941Sjhb			 * The RW_LOCK_READ_WAITERS flag should only be set
173154941Sjhb			 * if another thread currently holds a write lock,
174154941Sjhb			 * and in that case RW_LOCK_READ should be clear.
175154941Sjhb			 */
176154941Sjhb			MPASS((x & RW_LOCK_READ_WAITERS) == 0);
177154941Sjhb			if (atomic_cmpset_acq_ptr(&rw->rw_lock, x,
178154941Sjhb			    x + RW_ONE_READER)) {
179154941Sjhb				if (LOCK_LOG_TEST(&rw->rw_object, 0))
180154941Sjhb					CTR4(KTR_LOCK,
181154941Sjhb					    "%s: %p succeed %p -> %p", __func__,
182154941Sjhb					    rw, (void *)x,
183154941Sjhb					    (void *)(x + RW_ONE_READER));
184154941Sjhb				break;
185154941Sjhb			}
186157846Sjhb			cpu_spinwait();
187154941Sjhb			continue;
188154941Sjhb		}
189154941Sjhb
190154941Sjhb		/*
191154941Sjhb		 * Okay, now it's the hard case.  Some other thread already
192154941Sjhb		 * has a write lock, so acquire the turnstile lock so we can
193154941Sjhb		 * begin the process of blocking.
194154941Sjhb		 */
195154941Sjhb		turnstile_lock(&rw->rw_object);
196154941Sjhb
197154941Sjhb		/*
198154941Sjhb		 * The lock might have been released while we spun, so
199154941Sjhb		 * recheck its state and restart the loop if there is no
200154941Sjhb		 * longer a write lock.
201154941Sjhb		 */
202154941Sjhb		x = rw->rw_lock;
203154941Sjhb		if (x & RW_LOCK_READ) {
204154941Sjhb			turnstile_release(&rw->rw_object);
205157846Sjhb			cpu_spinwait();
206154941Sjhb			continue;
207154941Sjhb		}
208154941Sjhb
209154941Sjhb		/*
210154941Sjhb		 * Ok, it's still a write lock.  If the RW_LOCK_READ_WAITERS
211154941Sjhb		 * flag is already set, then we can go ahead and block.  If
212154941Sjhb		 * it is not set then try to set it.  If we fail to set it
213154941Sjhb		 * drop the turnstile lock and restart the loop.
214154941Sjhb		 */
215157826Sjhb		if (!(x & RW_LOCK_READ_WAITERS)) {
216157826Sjhb			if (!atomic_cmpset_ptr(&rw->rw_lock, x,
217157826Sjhb			    x | RW_LOCK_READ_WAITERS)) {
218157826Sjhb				turnstile_release(&rw->rw_object);
219157826Sjhb				cpu_spinwait();
220157826Sjhb				continue;
221157826Sjhb			}
222157826Sjhb			if (LOCK_LOG_TEST(&rw->rw_object, 0))
223157826Sjhb				CTR2(KTR_LOCK, "%s: %p set read waiters flag",
224157826Sjhb				    __func__, rw);
225154941Sjhb		}
226154941Sjhb
227157846Sjhb#ifdef SMP
228154941Sjhb		/*
229157846Sjhb		 * If the owner is running on another CPU, spin until
230157846Sjhb		 * the owner stops running or the state of the lock
231157846Sjhb		 * changes.
232157846Sjhb		 */
233157846Sjhb		owner = (struct thread *)RW_OWNER(x);
234157846Sjhb		if (TD_IS_RUNNING(owner)) {
235157846Sjhb			turnstile_release(&rw->rw_object);
236157846Sjhb			if (LOCK_LOG_TEST(&rw->rw_object, 0))
237157846Sjhb				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
238157846Sjhb				    __func__, rw, owner);
239157846Sjhb			while ((struct thread*)RW_OWNER(rw->rw_lock)== owner &&
240157846Sjhb			    TD_IS_RUNNING(owner))
241157846Sjhb				cpu_spinwait();
242157846Sjhb			continue;
243157846Sjhb		}
244157846Sjhb#endif
245157846Sjhb
246157846Sjhb		/*
247154941Sjhb		 * We were unable to acquire the lock and the read waiters
248154941Sjhb		 * flag is set, so we must block on the turnstile.
249154941Sjhb		 */
250154941Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
251154941Sjhb			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
252154941Sjhb			    rw);
253154941Sjhb		turnstile_wait(&rw->rw_object, rw_owner(rw), TS_SHARED_QUEUE);
254154941Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
255154941Sjhb			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
256154941Sjhb			    __func__, rw);
257154941Sjhb	}
258154941Sjhb
259154941Sjhb	/*
260154941Sjhb	 * TODO: acquire "owner of record" here.  Here be turnstile dragons
261154941Sjhb	 * however.  turnstiles don't like owners changing between calls to
262154941Sjhb	 * turnstile_wait() currently.
263154941Sjhb	 */
264154941Sjhb
265154941Sjhb	LOCK_LOG_LOCK("RLOCK", &rw->rw_object, 0, 0, file, line);
266154941Sjhb	WITNESS_LOCK(&rw->rw_object, 0, file, line);
267154941Sjhb}
268154941Sjhb
269154941Sjhbvoid
270154941Sjhb_rw_runlock(struct rwlock *rw, const char *file, int line)
271154941Sjhb{
272154941Sjhb	struct turnstile *ts;
273154941Sjhb	uintptr_t x;
274154941Sjhb
275154941Sjhb	_rw_assert(rw, RA_RLOCKED, file, line);
276154941Sjhb	WITNESS_UNLOCK(&rw->rw_object, 0, file, line);
277154941Sjhb	LOCK_LOG_LOCK("RUNLOCK", &rw->rw_object, 0, 0, file, line);
278154941Sjhb
279154941Sjhb	/* TODO: drop "owner of record" here. */
280154941Sjhb
281154941Sjhb	for (;;) {
282154941Sjhb		/*
283154941Sjhb		 * See if there is more than one read lock held.  If so,
284154941Sjhb		 * just drop one and return.
285154941Sjhb		 */
286154941Sjhb		x = rw->rw_lock;
287154941Sjhb		if (RW_READERS(x) > 1) {
288154941Sjhb			if (atomic_cmpset_ptr(&rw->rw_lock, x,
289154941Sjhb			    x - RW_ONE_READER)) {
290154941Sjhb				if (LOCK_LOG_TEST(&rw->rw_object, 0))
291154941Sjhb					CTR4(KTR_LOCK,
292154941Sjhb					    "%s: %p succeeded %p -> %p",
293154941Sjhb					    __func__, rw, (void *)x,
294154941Sjhb					    (void *)(x - RW_ONE_READER));
295154941Sjhb				break;
296154941Sjhb			}
297154941Sjhb			continue;
298154941Sjhb		}
299154941Sjhb
300154941Sjhb		/*
301154941Sjhb		 * We should never have read waiters while at least one
302154941Sjhb		 * thread holds a read lock.  (See note above)
303154941Sjhb		 */
304154941Sjhb		KASSERT(!(x & RW_LOCK_READ_WAITERS),
305154941Sjhb		    ("%s: waiting readers", __func__));
306154941Sjhb
307154941Sjhb		/*
308154941Sjhb		 * If there aren't any waiters for a write lock, then try
309154941Sjhb		 * to drop it quickly.
310154941Sjhb		 */
311154941Sjhb		if (!(x & RW_LOCK_WRITE_WAITERS)) {
312154941Sjhb
313154941Sjhb			/*
314154941Sjhb			 * There shouldn't be any flags set and we should
315154941Sjhb			 * be the only read lock.  If we fail to release
316154941Sjhb			 * the single read lock, then another thread might
317154941Sjhb			 * have just acquired a read lock, so go back up
318154941Sjhb			 * to the multiple read locks case.
319154941Sjhb			 */
320154941Sjhb			MPASS(x == RW_READERS_LOCK(1));
321154941Sjhb			if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1),
322154941Sjhb			    RW_UNLOCKED)) {
323154941Sjhb				if (LOCK_LOG_TEST(&rw->rw_object, 0))
324154941Sjhb					CTR2(KTR_LOCK, "%s: %p last succeeded",
325154941Sjhb					    __func__, rw);
326154941Sjhb				break;
327154941Sjhb			}
328154941Sjhb			continue;
329154941Sjhb		}
330154941Sjhb
331154941Sjhb		/*
332154941Sjhb		 * There should just be one reader with one or more
333154941Sjhb		 * writers waiting.
334154941Sjhb		 */
335154941Sjhb		MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS));
336154941Sjhb
337154941Sjhb		/*
338154941Sjhb		 * Ok, we know we have a waiting writer and we think we
339154941Sjhb		 * are the last reader, so grab the turnstile lock.
340154941Sjhb		 */
341154941Sjhb		turnstile_lock(&rw->rw_object);
342154941Sjhb
343154941Sjhb		/*
344154941Sjhb		 * Try to drop our lock leaving the lock in a unlocked
345154941Sjhb		 * state.
346154941Sjhb		 *
347154941Sjhb		 * If you wanted to do explicit lock handoff you'd have to
348154941Sjhb		 * do it here.  You'd also want to use turnstile_signal()
349154941Sjhb		 * and you'd have to handle the race where a higher
350154941Sjhb		 * priority thread blocks on the write lock before the
351154941Sjhb		 * thread you wakeup actually runs and have the new thread
352154941Sjhb		 * "steal" the lock.  For now it's a lot simpler to just
353154941Sjhb		 * wakeup all of the waiters.
354154941Sjhb		 *
355154941Sjhb		 * As above, if we fail, then another thread might have
356154941Sjhb		 * acquired a read lock, so drop the turnstile lock and
357154941Sjhb		 * restart.
358154941Sjhb		 */
359154941Sjhb		if (!atomic_cmpset_ptr(&rw->rw_lock,
360154941Sjhb		    RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) {
361154941Sjhb			turnstile_release(&rw->rw_object);
362154941Sjhb			continue;
363154941Sjhb		}
364154941Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
365154941Sjhb			CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
366154941Sjhb			    __func__, rw);
367154941Sjhb
368154941Sjhb		/*
369154941Sjhb		 * Ok.  The lock is released and all that's left is to
370154941Sjhb		 * wake up the waiters.  Note that the lock might not be
371154941Sjhb		 * free anymore, but in that case the writers will just
372154941Sjhb		 * block again if they run before the new lock holder(s)
373154941Sjhb		 * release the lock.
374154941Sjhb		 */
375154941Sjhb		ts = turnstile_lookup(&rw->rw_object);
376157846Sjhb		MPASS(ts != NULL);
377154941Sjhb		turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
378154941Sjhb		turnstile_unpend(ts, TS_SHARED_LOCK);
379154941Sjhb		break;
380154941Sjhb	}
381154941Sjhb}
382154941Sjhb
383154941Sjhb/*
384154941Sjhb * This function is called when we are unable to obtain a write lock on the
385154941Sjhb * first try.  This means that at least one other thread holds either a
386154941Sjhb * read or write lock.
387154941Sjhb */
388154941Sjhbvoid
389154941Sjhb_rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
390154941Sjhb{
391157846Sjhb	volatile struct thread *owner;
392154941Sjhb	uintptr_t v;
393154941Sjhb
394154941Sjhb	if (LOCK_LOG_TEST(&rw->rw_object, 0))
395154941Sjhb		CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
396154941Sjhb		    rw->rw_object.lo_name, (void *)rw->rw_lock, file, line);
397154941Sjhb
398154941Sjhb	while (!_rw_write_lock(rw, tid)) {
399154941Sjhb		turnstile_lock(&rw->rw_object);
400154941Sjhb		v = rw->rw_lock;
401154941Sjhb
402154941Sjhb		/*
403154941Sjhb		 * If the lock was released while spinning on the
404154941Sjhb		 * turnstile chain lock, try again.
405154941Sjhb		 */
406154941Sjhb		if (v == RW_UNLOCKED) {
407154941Sjhb			turnstile_release(&rw->rw_object);
408154941Sjhb			cpu_spinwait();
409154941Sjhb			continue;
410154941Sjhb		}
411154941Sjhb
412154941Sjhb		/*
413154941Sjhb		 * If the lock was released by a writer with both readers
414154941Sjhb		 * and writers waiting and a reader hasn't woken up and
415154941Sjhb		 * acquired the lock yet, rw_lock will be set to the
416154941Sjhb		 * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS.  If we see
417154941Sjhb		 * that value, try to acquire it once.  Note that we have
418154941Sjhb		 * to preserve the RW_LOCK_WRITE_WAITERS flag as there are
419154941Sjhb		 * other writers waiting still. If we fail, restart the
420154941Sjhb		 * loop.
421154941Sjhb		 */
422154941Sjhb		if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) {
423154941Sjhb			if (atomic_cmpset_acq_ptr(&rw->rw_lock,
424154941Sjhb			    RW_UNLOCKED | RW_LOCK_WRITE_WAITERS,
425154941Sjhb			    tid | RW_LOCK_WRITE_WAITERS)) {
426154941Sjhb				turnstile_claim(&rw->rw_object);
427154941Sjhb				CTR2(KTR_LOCK, "%s: %p claimed by new writer",
428154941Sjhb				    __func__, rw);
429154941Sjhb				break;
430154941Sjhb			}
431154941Sjhb			turnstile_release(&rw->rw_object);
432154941Sjhb			cpu_spinwait();
433154941Sjhb			continue;
434154941Sjhb		}
435154941Sjhb
436154941Sjhb		/*
437154941Sjhb		 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
438154941Sjhb		 * set it.  If we fail to set it, then loop back and try
439154941Sjhb		 * again.
440154941Sjhb		 */
441157826Sjhb		if (!(v & RW_LOCK_WRITE_WAITERS)) {
442157826Sjhb			if (!atomic_cmpset_ptr(&rw->rw_lock, v,
443157826Sjhb			    v | RW_LOCK_WRITE_WAITERS)) {
444157826Sjhb				turnstile_release(&rw->rw_object);
445157826Sjhb				cpu_spinwait();
446157826Sjhb				continue;
447157826Sjhb			}
448157826Sjhb			if (LOCK_LOG_TEST(&rw->rw_object, 0))
449157826Sjhb				CTR2(KTR_LOCK, "%s: %p set write waiters flag",
450157826Sjhb				    __func__, rw);
451154941Sjhb		}
452154941Sjhb
453157846Sjhb#ifdef SMP
454157846Sjhb		/*
455157846Sjhb		 * If the lock is write locked and the owner is
456157846Sjhb		 * running on another CPU, spin until the owner stops
457157846Sjhb		 * running or the state of the lock changes.
458157846Sjhb		 */
459157846Sjhb		owner = (struct thread *)RW_OWNER(v);
460157846Sjhb		if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
461157846Sjhb			turnstile_release(&rw->rw_object);
462157846Sjhb			if (LOCK_LOG_TEST(&rw->rw_object, 0))
463157846Sjhb				CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
464157846Sjhb				    __func__, rw, owner);
465157846Sjhb			while ((struct thread*)RW_OWNER(rw->rw_lock)== owner &&
466157846Sjhb			    TD_IS_RUNNING(owner))
467157846Sjhb				cpu_spinwait();
468157846Sjhb			continue;
469157846Sjhb		}
470157846Sjhb#endif
471154941Sjhb
472154941Sjhb		/*
473154941Sjhb		 * We were unable to acquire the lock and the write waiters
474154941Sjhb		 * flag is set, so we must block on the turnstile.
475154941Sjhb		 */
476154941Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
477154941Sjhb			CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
478154941Sjhb			    rw);
479154941Sjhb		turnstile_wait(&rw->rw_object, rw_owner(rw),
480154941Sjhb		    TS_EXCLUSIVE_QUEUE);
481154941Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
482154941Sjhb			CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
483154941Sjhb			    __func__, rw);
484154941Sjhb	}
485154941Sjhb}
486154941Sjhb
487154941Sjhb/*
488154941Sjhb * This function is called if the first try at releasing a write lock failed.
489154941Sjhb * This means that one of the 2 waiter bits must be set indicating that at
490154941Sjhb * least one thread is waiting on this lock.
491154941Sjhb */
492154941Sjhbvoid
493154941Sjhb_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
494154941Sjhb{
495154941Sjhb	struct turnstile *ts;
496154941Sjhb	uintptr_t v;
497154941Sjhb	int queue;
498154941Sjhb
499154941Sjhb	KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
500154941Sjhb	    ("%s: neither of the waiter flags are set", __func__));
501154941Sjhb
502154941Sjhb	if (LOCK_LOG_TEST(&rw->rw_object, 0))
503154941Sjhb		CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
504154941Sjhb
505154941Sjhb	turnstile_lock(&rw->rw_object);
506154941Sjhb	ts = turnstile_lookup(&rw->rw_object);
507154941Sjhb
508157846Sjhb#ifdef SMP
509157846Sjhb	/*
510157846Sjhb	 * There might not be a turnstile for this lock if all of
511157846Sjhb	 * the waiters are adaptively spinning.  In that case, just
512157846Sjhb	 * reset the lock to the unlocked state and return.
513157846Sjhb	 */
514157846Sjhb	if (ts == NULL) {
515157846Sjhb		atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED);
516157846Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
517157846Sjhb			CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw);
518157846Sjhb		turnstile_release(&rw->rw_object);
519157846Sjhb		return;
520157846Sjhb	}
521157846Sjhb#else
522154941Sjhb	MPASS(ts != NULL);
523157846Sjhb#endif
524154941Sjhb
525154941Sjhb	/*
526154941Sjhb	 * Use the same algo as sx locks for now.  Prefer waking up shared
527154941Sjhb	 * waiters if we have any over writers.  This is probably not ideal.
528154941Sjhb	 *
529154941Sjhb	 * 'v' is the value we are going to write back to rw_lock.  If we
530154941Sjhb	 * have waiters on both queues, we need to preserve the state of
531154941Sjhb	 * the waiter flag for the queue we don't wake up.  For now this is
532154941Sjhb	 * hardcoded for the algorithm mentioned above.
533154941Sjhb	 *
534154941Sjhb	 * In the case of both readers and writers waiting we wakeup the
535154941Sjhb	 * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
536154941Sjhb	 * new writer comes in before a reader it will claim the lock up
537154941Sjhb	 * above.  There is probably a potential priority inversion in
538154941Sjhb	 * there that could be worked around either by waking both queues
539154941Sjhb	 * of waiters or doing some complicated lock handoff gymnastics.
540157846Sjhb	 *
541157846Sjhb	 * Note that in the SMP case, if both flags are set, there might
542157846Sjhb	 * not be any actual writers on the turnstile as they might all
543157846Sjhb	 * be spinning.  In that case, we don't want to preserve the
544157846Sjhb	 * RW_LOCK_WRITE_WAITERS flag as the turnstile is going to go
545157846Sjhb	 * away once we wakeup all the readers.
546154941Sjhb	 */
547157846Sjhb	v = RW_UNLOCKED;
548154941Sjhb	if (rw->rw_lock & RW_LOCK_READ_WAITERS) {
549154941Sjhb		queue = TS_SHARED_QUEUE;
550157846Sjhb#ifdef SMP
551157846Sjhb		if (rw->rw_lock & RW_LOCK_WRITE_WAITERS &&
552157846Sjhb		    !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE))
553157846Sjhb			v |= RW_LOCK_WRITE_WAITERS;
554157846Sjhb#else
555157846Sjhb		v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS);
556157846Sjhb#endif
557157846Sjhb	} else
558154941Sjhb		queue = TS_EXCLUSIVE_QUEUE;
559157846Sjhb
560157846Sjhb#ifdef SMP
561157846Sjhb	/*
562157846Sjhb	 * We have to make sure that we actually have waiters to
563157846Sjhb	 * wakeup.  If they are all spinning, then we just need to
564157846Sjhb	 * disown the turnstile and return.
565157846Sjhb	 */
566157846Sjhb	if (turnstile_empty(ts, queue)) {
567157846Sjhb		if (LOCK_LOG_TEST(&rw->rw_object, 0))
568157846Sjhb			CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw);
569157846Sjhb		atomic_store_rel_ptr(&rw->rw_lock, v);
570157846Sjhb		turnstile_disown(ts);
571157846Sjhb		return;
572154941Sjhb	}
573157846Sjhb#endif
574157846Sjhb
575157846Sjhb	/* Wake up all waiters for the specific queue. */
576154941Sjhb	if (LOCK_LOG_TEST(&rw->rw_object, 0))
577154941Sjhb		CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
578154941Sjhb		    queue == TS_SHARED_QUEUE ? "read" : "write");
579154941Sjhb	turnstile_broadcast(ts, queue);
580154941Sjhb	atomic_store_rel_ptr(&rw->rw_lock, v);
581154941Sjhb	turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
582154941Sjhb}
583154941Sjhb
584154941Sjhb#ifdef INVARIANT_SUPPORT
585155162Sscottl#ifndef INVARIANTS
586154941Sjhb#undef _rw_assert
587154941Sjhb#endif
588154941Sjhb
589154941Sjhb/*
590154941Sjhb * In the non-WITNESS case, rw_assert() can only detect that at least
591154941Sjhb * *some* thread owns an rlock, but it cannot guarantee that *this*
592154941Sjhb * thread owns an rlock.
593154941Sjhb */
594154941Sjhbvoid
595154941Sjhb_rw_assert(struct rwlock *rw, int what, const char *file, int line)
596154941Sjhb{
597154941Sjhb
598154941Sjhb	if (panicstr != NULL)
599154941Sjhb		return;
600154941Sjhb	switch (what) {
601154941Sjhb	case RA_LOCKED:
602154941Sjhb	case RA_RLOCKED:
603154941Sjhb#ifdef WITNESS
604154941Sjhb		witness_assert(&rw->rw_object, what, file, line);
605154941Sjhb#else
606154941Sjhb		/*
607154941Sjhb		 * If some other thread has a write lock or we have one
608154941Sjhb		 * and are asserting a read lock, fail.  Also, if no one
609154941Sjhb		 * has a lock at all, fail.
610154941Sjhb		 */
611155061Sscottl		if (rw->rw_lock == RW_UNLOCKED ||
612155061Sscottl		    (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED ||
613157826Sjhb		    rw_wowner(rw) != curthread)))
614154941Sjhb			panic("Lock %s not %slocked @ %s:%d\n",
615155012Sscottl			    rw->rw_object.lo_name, (what == RA_RLOCKED) ?
616154941Sjhb			    "read " : "", file, line);
617154941Sjhb#endif
618154941Sjhb		break;
619154941Sjhb	case RA_WLOCKED:
620157826Sjhb		if (rw_wowner(rw) != curthread)
621154941Sjhb			panic("Lock %s not exclusively locked @ %s:%d\n",
622154941Sjhb			    rw->rw_object.lo_name, file, line);
623154941Sjhb		break;
624154941Sjhb	case RA_UNLOCKED:
625154941Sjhb#ifdef WITNESS
626154941Sjhb		witness_assert(&rw->rw_object, what, file, line);
627154941Sjhb#else
628154941Sjhb		/*
629154941Sjhb		 * If we hold a write lock fail.  We can't reliably check
630154941Sjhb		 * to see if we hold a read lock or not.
631154941Sjhb		 */
632157826Sjhb		if (rw_wowner(rw) == curthread)
633154941Sjhb			panic("Lock %s exclusively locked @ %s:%d\n",
634154941Sjhb			    rw->rw_object.lo_name, file, line);
635154941Sjhb#endif
636154941Sjhb		break;
637154941Sjhb	default:
638154941Sjhb		panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
639154941Sjhb		    line);
640154941Sjhb	}
641154941Sjhb}
642154941Sjhb#endif /* INVARIANT_SUPPORT */
643154941Sjhb
644154941Sjhb#ifdef DDB
645154941Sjhbvoid
646154941Sjhbdb_show_rwlock(struct lock_object *lock)
647154941Sjhb{
648154941Sjhb	struct rwlock *rw;
649154941Sjhb	struct thread *td;
650154941Sjhb
651154941Sjhb	rw = (struct rwlock *)lock;
652154941Sjhb
653154941Sjhb	db_printf(" state: ");
654154941Sjhb	if (rw->rw_lock == RW_UNLOCKED)
655154941Sjhb		db_printf("UNLOCKED\n");
656154941Sjhb	else if (rw->rw_lock & RW_LOCK_READ)
657154973Smlaier		db_printf("RLOCK: %jd locks\n",
658154973Smlaier		    (intmax_t)(RW_READERS(rw->rw_lock)));
659154941Sjhb	else {
660157826Sjhb		td = rw_wowner(rw);
661154941Sjhb		db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
662154941Sjhb		    td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
663154941Sjhb	}
664154941Sjhb	db_printf(" waiters: ");
665154941Sjhb	switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
666154941Sjhb	case RW_LOCK_READ_WAITERS:
667154941Sjhb		db_printf("readers\n");
668154941Sjhb		break;
669154941Sjhb	case RW_LOCK_WRITE_WAITERS:
670154941Sjhb		db_printf("writers\n");
671154941Sjhb		break;
672154941Sjhb	case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
673154941Sjhb		db_printf("readers and waiters\n");
674154941Sjhb		break;
675154941Sjhb	default:
676154941Sjhb		db_printf("none\n");
677154941Sjhb		break;
678154941Sjhb	}
679154941Sjhb}
680154941Sjhb
681154941Sjhb#endif
682