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 * 3. Neither the name of the author nor the names of any co-contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 /*
31 * Machine independent bits of reader/writer lock implementation.
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_ddb.h"
38 #include "opt_no_adaptive_rwlocks.h"
39
40 #include <sys/param.h>
41 #include <sys/ktr.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/rwlock.h>
46 #include <sys/systm.h>
47 #include <sys/turnstile.h>
48
49 #include <machine/cpu.h>
50
51 CTASSERT((RW_RECURSE & LO_CLASSFLAGS) == RW_RECURSE);
52
53 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
54 #define ADAPTIVE_RWLOCKS
55 #endif
56
57 #ifdef DDB
58 #include <ddb/ddb.h>
59
60 static void db_show_rwlock(struct lock_object *lock);
61 #endif
62
63 struct lock_class lock_class_rw = {
64 .lc_name = "rw",
65 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
66 #ifdef DDB
67 .lc_ddb_show = db_show_rwlock,
68 #endif
69 };
70
71 /*
72 * Return a pointer to the owning thread if the lock is write-locked or
73 * NULL if the lock is unlocked or read-locked.
74 */
75 #define rw_wowner(rw) \
76 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \
77 (struct thread *)RW_OWNER((rw)->rw_lock))
78
79 /*
80 * Returns if a write owner is recursed. Write ownership is not assured
81 * here and should be previously checked.
82 */
83 #define rw_recursed(rw) ((rw)->rw_recurse != 0)
84
85 /*
86 * Return true if curthread helds the lock.
87 */
88 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread)
89
90 /*
91 * Return a pointer to the owning thread for this lock who should receive
92 * any priority lent by threads that block on this lock. Currently this
93 * is identical to rw_wowner().
94 */
95 #define rw_owner(rw) rw_wowner(rw)
96
97 #ifndef INVARIANTS
98 #define _rw_assert(rw, what, file, line)
99 #endif
100
101 void
102 rw_init_flags(struct rwlock *rw, const char *name, int opts)
103 {
104 int flags;
105
106 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
107 RW_RECURSE)) == 0);
108
109 flags = LO_UPGRADABLE | LO_RECURSABLE;
110 if (opts & RW_DUPOK)
111 flags |= LO_DUPOK;
112 if (!(opts & RW_NOWITNESS))
113 flags |= LO_WITNESS;
114 if (opts & RW_QUIET)
115 flags |= LO_QUIET;
116 flags |= opts & RW_RECURSE;
117
118 rw->rw_lock = RW_UNLOCKED;
119 rw->rw_recurse = 0;
120 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
121 }
122
123 void
124 rw_destroy(struct rwlock *rw)
125 {
126
127 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock not unlocked"));
128 KASSERT(rw->rw_recurse == 0, ("rw lock still recursed"));
129 rw->rw_lock = RW_DESTROYED;
130 lock_destroy(&rw->lock_object);
131 }
132
133 void
134 rw_sysinit(void *arg)
135 {
136 struct rw_args *args = arg;
137
138 rw_init(args->ra_rw, args->ra_desc);
139 }
140
141 int
142 rw_wowned(struct rwlock *rw)
143 {
144
145 return (rw_wowner(rw) == curthread);
146 }
147
148 void
149 _rw_wlock(struct rwlock *rw, const char *file, int line)
150 {
151
152 MPASS(curthread != NULL);
153 KASSERT(rw->rw_lock != RW_DESTROYED,
154 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
155 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
156 line);
157 __rw_wlock(rw, curthread, file, line);
158 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
159 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
160 curthread->td_locks++;
161 }
162
163 void
164 _rw_wunlock(struct rwlock *rw, const char *file, int line)
165 {
166
167 MPASS(curthread != NULL);
168 KASSERT(rw->rw_lock != RW_DESTROYED,
169 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
170 _rw_assert(rw, RA_WLOCKED, file, line);
171 curthread->td_locks--;
172 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
173 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
174 line);
175 if (!rw_recursed(rw))
176 lock_profile_release_lock(&rw->lock_object);
177 __rw_wunlock(rw, curthread, file, line);
178 }
179
180 void
181 _rw_rlock(struct rwlock *rw, const char *file, int line)
182 {
183 #ifdef ADAPTIVE_RWLOCKS
184 volatile struct thread *owner;
185 #endif
186 //uint64_t waittime = 0; /* XXX: notsup */
187 //int contested = 0; /* XXX: notsup */
188 uintptr_t x;
189
190 KASSERT(rw->rw_lock != RW_DESTROYED,
191 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
192 KASSERT(rw_wowner(rw) != curthread,
193 ("%s (%s): wlock already held @ %s:%d", __func__,
194 rw->lock_object.lo_name, file, line));
195 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line);
196
197 /*
198 * Note that we don't make any attempt to try to block read
199 * locks once a writer has blocked on the lock. The reason is
200 * that we currently allow for read locks to recurse and we
201 * don't keep track of all the holders of read locks. Thus, if
202 * we were to block readers once a writer blocked and a reader
203 * tried to recurse on their reader lock after a writer had
204 * blocked we would end up in a deadlock since the reader would
205 * be blocked on the writer, and the writer would be blocked
206 * waiting for the reader to release its original read lock.
207 */
208 for (;;) {
209 /*
210 * Handle the easy case. If no other thread has a write
211 * lock, then try to bump up the count of read locks. Note
212 * that we have to preserve the current state of the
213 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
214 * read lock, then rw_lock must have changed, so restart
215 * the loop. Note that this handles the case of a
216 * completely unlocked rwlock since such a lock is encoded
217 * as a read lock with no waiters.
218 */
219 x = rw->rw_lock;
220 if (x & RW_LOCK_READ) {
221
222 /*
223 * The RW_LOCK_READ_WAITERS flag should only be set
224 * if another thread currently holds a write lock,
225 * and in that case RW_LOCK_READ should be clear.
226 */
227 MPASS((x & RW_LOCK_READ_WAITERS) == 0);
228 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x,
229 x + RW_ONE_READER)) {
230 if (LOCK_LOG_TEST(&rw->lock_object, 0))
231 CTR4(KTR_LOCK,
232 "%s: %p succeed %p -> %p", __func__,
233 rw, (void *)x,
234 (void *)(x + RW_ONE_READER));
235 if (RW_READERS(x) == 0)
236 lock_profile_obtain_lock_success(
237 &rw->lock_object, contested, waittime,
238 file, line);
239 break;
240 }
241 cpu_spinwait();
242 continue;
243 }
244 lock_profile_obtain_lock_failed(&rw->lock_object, &contested,
245 &waittime);
246
247 /*
248 * Okay, now it's the hard case. Some other thread already
249 * has a write lock, so acquire the turnstile lock so we can
250 * begin the process of blocking.
251 */
252 turnstile_lock(&rw->lock_object);
253
254 /*
255 * The lock might have been released while we spun, so
256 * recheck its state and restart the loop if there is no
257 * longer a write lock.
258 */
259 x = rw->rw_lock;
260 if (x & RW_LOCK_READ) {
261 turnstile_release(&rw->lock_object);
262 cpu_spinwait();
263 continue;
264 }
265
266 /*
267 * Ok, it's still a write lock. If the RW_LOCK_READ_WAITERS
268 * flag is already set, then we can go ahead and block. If
269 * it is not set then try to set it. If we fail to set it
270 * drop the turnstile lock and restart the loop.
271 */
272 if (!(x & RW_LOCK_READ_WAITERS)) {
273 if (!atomic_cmpset_ptr(&rw->rw_lock, x,
274 x | RW_LOCK_READ_WAITERS)) {
275 turnstile_release(&rw->lock_object);
276 cpu_spinwait();
277 continue;
278 }
279 if (LOCK_LOG_TEST(&rw->lock_object, 0))
280 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
281 __func__, rw);
282 }
283
284 #ifdef ADAPTIVE_RWLOCKS
285 /*
286 * If the owner is running on another CPU, spin until
287 * the owner stops running or the state of the lock
288 * changes.
289 */
290 owner = (struct thread *)RW_OWNER(x);
291 if (TD_IS_RUNNING(owner)) {
292 turnstile_release(&rw->lock_object);
293 if (LOCK_LOG_TEST(&rw->lock_object, 0))
294 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
295 __func__, rw, owner);
296 while ((struct thread*)RW_OWNER(rw->rw_lock)== owner &&
297 TD_IS_RUNNING(owner))
298 cpu_spinwait();
299 continue;
300 }
301 #endif
302
303 /*
304 * We were unable to acquire the lock and the read waiters
305 * flag is set, so we must block on the turnstile.
306 */
307 if (LOCK_LOG_TEST(&rw->lock_object, 0))
308 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
309 rw);
310 turnstile_wait(&rw->lock_object, rw_owner(rw),
311 TS_SHARED_QUEUE);
312 if (LOCK_LOG_TEST(&rw->lock_object, 0))
313 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
314 __func__, rw);
315 }
316
317 /*
318 * TODO: acquire "owner of record" here. Here be turnstile dragons
319 * however. turnstiles don't like owners changing between calls to
320 * turnstile_wait() currently.
321 */
322
323 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
324 WITNESS_LOCK(&rw->lock_object, 0, file, line);
325 curthread->td_locks++;
326 }
327
328 void
329 _rw_runlock(struct rwlock *rw, const char *file, int line)
330 {
331 struct turnstile *ts;
332 uintptr_t x;
333
334 KASSERT(rw->rw_lock != RW_DESTROYED,
335 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
336 _rw_assert(rw, RA_RLOCKED, file, line);
337 curthread->td_locks--;
338 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
339 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
340
341 /* TODO: drop "owner of record" here. */
342
343 for (;;) {
344 /*
345 * See if there is more than one read lock held. If so,
346 * just drop one and return.
347 */
348 x = rw->rw_lock;
349 if (RW_READERS(x) > 1) {
350 if (atomic_cmpset_ptr(&rw->rw_lock, x,
351 x - RW_ONE_READER)) {
352 if (LOCK_LOG_TEST(&rw->lock_object, 0))
353 CTR4(KTR_LOCK,
354 "%s: %p succeeded %p -> %p",
355 __func__, rw, (void *)x,
356 (void *)(x - RW_ONE_READER));
357 break;
358 }
359 continue;
360 }
361
362
363 /*
364 * We should never have read waiters while at least one
365 * thread holds a read lock. (See note above)
366 */
367 KASSERT(!(x & RW_LOCK_READ_WAITERS),
368 ("%s: waiting readers", __func__));
369
370 /*
371 * If there aren't any waiters for a write lock, then try
372 * to drop it quickly.
373 */
374 if (!(x & RW_LOCK_WRITE_WAITERS)) {
375
376 /*
377 * There shouldn't be any flags set and we should
378 * be the only read lock. If we fail to release
379 * the single read lock, then another thread might
380 * have just acquired a read lock, so go back up
381 * to the multiple read locks case.
382 */
383 MPASS(x == RW_READERS_LOCK(1));
384 if (atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1),
385 RW_UNLOCKED)) {
386 if (LOCK_LOG_TEST(&rw->lock_object, 0))
387 CTR2(KTR_LOCK, "%s: %p last succeeded",
388 __func__, rw);
389 break;
390 }
391 continue;
392 }
393
394 /*
395 * There should just be one reader with one or more
396 * writers waiting.
397 */
398 MPASS(x == (RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS));
399
400 /*
401 * Ok, we know we have a waiting writer and we think we
402 * are the last reader, so grab the turnstile lock.
403 */
404 turnstile_lock(&rw->lock_object);
405
406 /*
407 * Try to drop our lock leaving the lock in a unlocked
408 * state.
409 *
410 * If you wanted to do explicit lock handoff you'd have to
411 * do it here. You'd also want to use turnstile_signal()
412 * and you'd have to handle the race where a higher
413 * priority thread blocks on the write lock before the
414 * thread you wakeup actually runs and have the new thread
415 * "steal" the lock. For now it's a lot simpler to just
416 * wakeup all of the waiters.
417 *
418 * As above, if we fail, then another thread might have
419 * acquired a read lock, so drop the turnstile lock and
420 * restart.
421 */
422 if (!atomic_cmpset_ptr(&rw->rw_lock,
423 RW_READERS_LOCK(1) | RW_LOCK_WRITE_WAITERS, RW_UNLOCKED)) {
424 turnstile_release(&rw->lock_object);
425 continue;
426 }
427 if (LOCK_LOG_TEST(&rw->lock_object, 0))
428 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
429 __func__, rw);
430
431 /*
432 * Ok. The lock is released and all that's left is to
433 * wake up the waiters. Note that the lock might not be
434 * free anymore, but in that case the writers will just
435 * block again if they run before the new lock holder(s)
436 * release the lock.
437 */
438 ts = turnstile_lookup(&rw->lock_object);
439 MPASS(ts != NULL);
440 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
441 turnstile_unpend(ts, TS_SHARED_LOCK);
442 break;
443 }
444 lock_profile_release_lock(&rw->lock_object);
445 }
446
447 /*
448 * This function is called when we are unable to obtain a write lock on the
449 * first try. This means that at least one other thread holds either a
450 * read or write lock.
451 */
452 void
453 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
454 {
455 #ifdef ADAPTIVE_RWLOCKS
456 volatile struct thread *owner;
457 #endif
458 uintptr_t v;
459
460 if (rw_wlocked(rw)) {
461 KASSERT(rw->lock_object.lo_flags & RW_RECURSE,
462 ("%s: recursing but non-recursive rw %s @ %s:%d\n",
463 __func__, rw->lock_object.lo_name, file, line));
464 rw->rw_recurse++;
465 atomic_set_ptr(&rw->rw_lock, RW_LOCK_RECURSED);
466 if (LOCK_LOG_TEST(&rw->lock_object, 0))
467 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
468 return;
469 }
470
471 if (LOCK_LOG_TEST(&rw->lock_object, 0))
472 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
473 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
474
475 while (!_rw_write_lock(rw, tid)) {
476 turnstile_lock(&rw->lock_object);
477 v = rw->rw_lock;
478
479 /*
480 * If the lock was released while spinning on the
481 * turnstile chain lock, try again.
482 */
483 if (v == RW_UNLOCKED) {
484 turnstile_release(&rw->lock_object);
485 cpu_spinwait();
486 continue;
487 }
488
489 /*
490 * If the lock was released by a writer with both readers
491 * and writers waiting and a reader hasn't woken up and
492 * acquired the lock yet, rw_lock will be set to the
493 * value RW_UNLOCKED | RW_LOCK_WRITE_WAITERS. If we see
494 * that value, try to acquire it once. Note that we have
495 * to preserve the RW_LOCK_WRITE_WAITERS flag as there are
496 * other writers waiting still. If we fail, restart the
497 * loop.
498 */
499 if (v == (RW_UNLOCKED | RW_LOCK_WRITE_WAITERS)) {
500 if (atomic_cmpset_acq_ptr(&rw->rw_lock,
501 RW_UNLOCKED | RW_LOCK_WRITE_WAITERS,
502 tid | RW_LOCK_WRITE_WAITERS)) {
503 turnstile_claim(&rw->lock_object);
504 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
505 __func__, rw);
506 break;
507 }
508 turnstile_release(&rw->lock_object);
509 cpu_spinwait();
510 continue;
511 }
512
513 /*
514 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
515 * set it. If we fail to set it, then loop back and try
516 * again.
517 */
518 if (!(v & RW_LOCK_WRITE_WAITERS)) {
519 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
520 v | RW_LOCK_WRITE_WAITERS)) {
521 turnstile_release(&rw->lock_object);
522 cpu_spinwait();
523 continue;
524 }
525 if (LOCK_LOG_TEST(&rw->lock_object, 0))
526 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
527 __func__, rw);
528 }
529
530 #ifdef ADAPTIVE_RWLOCKS
531 /*
532 * If the lock is write locked and the owner is
533 * running on another CPU, spin until the owner stops
534 * running or the state of the lock changes.
535 */
536 owner = (struct thread *)RW_OWNER(v);
537 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
538 turnstile_release(&rw->lock_object);
539 if (LOCK_LOG_TEST(&rw->lock_object, 0))
540 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
541 __func__, rw, owner);
542 while ((struct thread*)RW_OWNER(rw->rw_lock)== owner &&
543 TD_IS_RUNNING(owner))
544 cpu_spinwait();
545 continue;
546 }
547 #endif
548
549 /*
550 * We were unable to acquire the lock and the write waiters
551 * flag is set, so we must block on the turnstile.
552 */
553 if (LOCK_LOG_TEST(&rw->lock_object, 0))
554 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
555 rw);
556 turnstile_wait(&rw->lock_object, rw_owner(rw),
557 TS_EXCLUSIVE_QUEUE);
558 if (LOCK_LOG_TEST(&rw->lock_object, 0))
559 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
560 __func__, rw);
561 }
562 }
563
564 /*
565 * This function is called if the first try at releasing a write lock failed.
566 * This means that one of the 2 waiter bits must be set indicating that at
567 * least one thread is waiting on this lock.
568 */
569 void
570 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
571 {
572 struct turnstile *ts;
573 uintptr_t v;
574 int queue;
575
576 if (rw_wlocked(rw) && rw_recursed(rw)) {
577 if ((--rw->rw_recurse) == 0)
578 atomic_clear_ptr(&rw->rw_lock, RW_LOCK_RECURSED);
579 if (LOCK_LOG_TEST(&rw->lock_object, 0))
580 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
581 return;
582 }
583
584 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
585 ("%s: neither of the waiter flags are set", __func__));
586
587 if (LOCK_LOG_TEST(&rw->lock_object, 0))
588 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
589
590 turnstile_lock(&rw->lock_object);
591 ts = turnstile_lookup(&rw->lock_object);
592
593 #ifdef ADAPTIVE_RWLOCKS
594 /*
595 * There might not be a turnstile for this lock if all of
596 * the waiters are adaptively spinning. In that case, just
597 * reset the lock to the unlocked state and return.
598 */
599 if (ts == NULL) {
600 atomic_store_rel_ptr(&rw->rw_lock, RW_UNLOCKED);
601 if (LOCK_LOG_TEST(&rw->lock_object, 0))
602 CTR2(KTR_LOCK, "%s: %p no sleepers", __func__, rw);
603 turnstile_release(&rw->lock_object);
604 return;
605 }
606 #else
607 MPASS(ts != NULL);
608 #endif
609
610 /*
611 * Use the same algo as sx locks for now. Prefer waking up shared
612 * waiters if we have any over writers. This is probably not ideal.
613 *
614 * 'v' is the value we are going to write back to rw_lock. If we
615 * have waiters on both queues, we need to preserve the state of
616 * the waiter flag for the queue we don't wake up. For now this is
617 * hardcoded for the algorithm mentioned above.
618 *
619 * In the case of both readers and writers waiting we wakeup the
620 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
621 * new writer comes in before a reader it will claim the lock up
622 * above. There is probably a potential priority inversion in
623 * there that could be worked around either by waking both queues
624 * of waiters or doing some complicated lock handoff gymnastics.
625 *
626 * Note that in the ADAPTIVE_RWLOCKS case, if both flags are
627 * set, there might not be any actual writers on the turnstile
628 * as they might all be spinning. In that case, we don't want
629 * to preserve the RW_LOCK_WRITE_WAITERS flag as the turnstile
630 * is going to go away once we wakeup all the readers.
631 */
632 v = RW_UNLOCKED;
633 if (rw->rw_lock & RW_LOCK_READ_WAITERS) {
634 queue = TS_SHARED_QUEUE;
635 #ifdef ADAPTIVE_RWLOCKS
636 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS &&
637 !turnstile_empty(ts, TS_EXCLUSIVE_QUEUE))
638 v |= RW_LOCK_WRITE_WAITERS;
639 #else
640 v |= (rw->rw_lock & RW_LOCK_WRITE_WAITERS);
641 #endif
642 } else
643 queue = TS_EXCLUSIVE_QUEUE;
644
645 #ifdef ADAPTIVE_RWLOCKS
646 /*
647 * We have to make sure that we actually have waiters to
648 * wakeup. If they are all spinning, then we just need to
649 * disown the turnstile and return.
650 */
651 if (turnstile_empty(ts, queue)) {
652 if (LOCK_LOG_TEST(&rw->lock_object, 0))
653 CTR2(KTR_LOCK, "%s: %p no sleepers 2", __func__, rw);
654 atomic_store_rel_ptr(&rw->rw_lock, v);
655 turnstile_disown(ts);
656 turnstile_release(&rw->lock_object);
657 return;
658 }
659 #endif
660
661 /* Wake up all waiters for the specific queue. */
662 if (LOCK_LOG_TEST(&rw->lock_object, 0))
663 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
664 queue == TS_SHARED_QUEUE ? "read" : "write");
665 turnstile_broadcast(ts, queue);
666 atomic_store_rel_ptr(&rw->rw_lock, v);
667 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
668 }
669
670 /*
671 * Attempt to do a non-blocking upgrade from a read lock to a write
672 * lock. This will only succeed if this thread holds a single read
673 * lock. Returns true if the upgrade succeeded and false otherwise.
674 */
675 int
676 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
677 {
678 uintptr_t v, tid;
679 int success;
680
681 KASSERT(rw->rw_lock != RW_DESTROYED,
682 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
683 _rw_assert(rw, RA_RLOCKED, file, line);
684
685 /*
686 * Attempt to switch from one reader to a writer. If there
687 * are any write waiters, then we will have to lock the
688 * turnstile first to prevent races with another writer
689 * calling turnstile_wait() before we have claimed this
690 * turnstile. So, do the simple case of no waiters first.
691 */
692 tid = (uintptr_t)curthread;
693 if (!(rw->rw_lock & RW_LOCK_WRITE_WAITERS)) {
694 success = atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1),
695 tid);
696 goto out;
697 }
698
699 /*
700 * Ok, we think we have write waiters, so lock the
701 * turnstile.
702 */
703 turnstile_lock(&rw->lock_object);
704
705 /*
706 * Try to switch from one reader to a writer again. This time
707 * we honor the current state of the RW_LOCK_WRITE_WAITERS
708 * flag. If we obtain the lock with the flag set, then claim
709 * ownership of the turnstile. In the ADAPTIVE_RWLOCKS case
710 * it is possible for there to not be an associated turnstile
711 * even though there are waiters if all of the waiters are
712 * spinning.
713 */
714 v = rw->rw_lock & RW_LOCK_WRITE_WAITERS;
715 success = atomic_cmpset_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
716 tid | v);
717 #ifdef ADAPTIVE_RWLOCKS
718 if (success && v && turnstile_lookup(&rw->lock_object) != NULL)
719 #else
720 if (success && v)
721 #endif
722 turnstile_claim(&rw->lock_object);
723 else
724 turnstile_release(&rw->lock_object);
725 out:
726 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
727 if (success)
728 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
729 file, line);
730 return (success);
731 }
732
733 /*
734 * Downgrade a write lock into a single read lock.
735 */
736 void
737 _rw_downgrade(struct rwlock *rw, const char *file, int line)
738 {
739 struct turnstile *ts;
740 uintptr_t tid, v;
741
742 KASSERT(rw->rw_lock != RW_DESTROYED,
743 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
744 _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
745 #ifndef INVARIANTS
746 if (rw_recursed(rw))
747 panic("downgrade of a recursed lock");
748 #endif
749
750 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
751
752 /*
753 * Convert from a writer to a single reader. First we handle
754 * the easy case with no waiters. If there are any waiters, we
755 * lock the turnstile, "disown" the lock, and awaken any read
756 * waiters.
757 */
758 tid = (uintptr_t)curthread;
759 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
760 goto out;
761
762 /*
763 * Ok, we think we have waiters, so lock the turnstile so we can
764 * read the waiter flags without any races.
765 */
766 turnstile_lock(&rw->lock_object);
767 v = rw->rw_lock;
768 MPASS(v & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS));
769
770 /*
771 * Downgrade from a write lock while preserving
772 * RW_LOCK_WRITE_WAITERS and give up ownership of the
773 * turnstile. If there are any read waiters, wake them up.
774 *
775 * For ADAPTIVE_RWLOCKS, we have to allow for the fact that
776 * all of the read waiters might be spinning. In that case,
777 * act as if RW_LOCK_READ_WAITERS is not set. Also, only
778 * preserve the RW_LOCK_WRITE_WAITERS flag if at least one
779 * writer is blocked on the turnstile.
780 */
781 ts = turnstile_lookup(&rw->lock_object);
782 #ifdef ADAPTIVE_RWLOCKS
783 if (ts == NULL)
784 v &= ~(RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS);
785 else if (v & RW_LOCK_READ_WAITERS &&
786 turnstile_empty(ts, TS_SHARED_QUEUE))
787 v &= ~RW_LOCK_READ_WAITERS;
788 else if (v & RW_LOCK_WRITE_WAITERS &&
789 turnstile_empty(ts, TS_EXCLUSIVE_QUEUE))
790 v &= ~RW_LOCK_WRITE_WAITERS;
791 #else
792 MPASS(ts != NULL);
793 #endif
794 if (v & RW_LOCK_READ_WAITERS)
795 turnstile_broadcast(ts, TS_SHARED_QUEUE);
796 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) |
797 (v & RW_LOCK_WRITE_WAITERS));
798 if (v & RW_LOCK_READ_WAITERS) {
799 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
800 } else if (ts) {
801 turnstile_disown(ts);
802 turnstile_release(&rw->lock_object);
803 }
804 out:
805 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
806 }
807
808 #ifdef INVARIANT_SUPPORT
809 #ifndef INVARIANTS
810 #undef _rw_assert
811 #endif
812
813 /*
814 * In the non-WITNESS case, rw_assert() can only detect that at least
815 * *some* thread owns an rlock, but it cannot guarantee that *this*
816 * thread owns an rlock.
817 */
818 void
819 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
820 {
821
822 if (panicstr != NULL)
823 return;
824 switch (what) {
825 case RA_LOCKED:
826 case RA_LOCKED | RA_RECURSED:
827 case RA_LOCKED | RA_NOTRECURSED:
828 case RA_RLOCKED:
829 #ifdef WITNESS
830 witness_assert(&rw->lock_object, what, file, line);
831 #else
832 /*
833 * If some other thread has a write lock or we have one
834 * and are asserting a read lock, fail. Also, if no one
835 * has a lock at all, fail.
836 */
837 if (rw->rw_lock == RW_UNLOCKED ||
838 (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED ||
839 rw_wowner(rw) != curthread)))
840 panic("Lock %s not %slocked @ %s:%d\n",
841 rw->lock_object.lo_name, (what == RA_RLOCKED) ?
842 "read " : "", file, line);
843
844 if (!(rw->rw_lock & RW_LOCK_READ)) {
845 if (rw_recursed(rw)) {
846 if (what & RA_NOTRECURSED)
847 panic("Lock %s recursed @ %s:%d\n",
848 rw->lock_object.lo_name, file,
849 line);
850 } else if (what & RA_RECURSED)
851 panic("Lock %s not recursed @ %s:%d\n",
852 rw->lock_object.lo_name, file, line);
853 }
854 #endif
855 break;
856 case RA_WLOCKED:
857 case RA_WLOCKED | RA_RECURSED:
858 case RA_WLOCKED | RA_NOTRECURSED:
859 if (rw_wowner(rw) != curthread)
860 panic("Lock %s not exclusively locked @ %s:%d\n",
861 rw->lock_object.lo_name, file, line);
862 if (rw_recursed(rw)) {
863 if (what & RA_NOTRECURSED)
864 panic("Lock %s recursed @ %s:%d\n",
865 rw->lock_object.lo_name, file, line);
866 } else if (what & RA_RECURSED)
867 panic("Lock %s not recursed @ %s:%d\n",
868 rw->lock_object.lo_name, file, line);
869 break;
870 case RA_UNLOCKED:
871 #ifdef WITNESS
872 witness_assert(&rw->lock_object, what, file, line);
873 #else
874 /*
875 * If we hold a write lock fail. We can't reliably check
876 * to see if we hold a read lock or not.
877 */
878 if (rw_wowner(rw) == curthread)
879 panic("Lock %s exclusively locked @ %s:%d\n",
880 rw->lock_object.lo_name, file, line);
881 #endif
882 break;
883 default:
884 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
885 line);
886 }
887 }
888 #endif /* INVARIANT_SUPPORT */
889
890 #ifdef DDB
891 void
892 db_show_rwlock(struct lock_object *lock)
893 {
894 struct rwlock *rw;
895 struct thread *td;
896
897 rw = (struct rwlock *)lock;
898
899 db_printf(" state: ");
900 if (rw->rw_lock == RW_UNLOCKED)
901 db_printf("UNLOCKED\n");
902 else if (rw->rw_lock == RW_DESTROYED) {
903 db_printf("DESTROYED\n");
904 return;
905 } else if (rw->rw_lock & RW_LOCK_READ)
906 db_printf("RLOCK: %ju locks\n",
907 (uintmax_t)(RW_READERS(rw->rw_lock)));
908 else {
909 td = rw_wowner(rw);
910 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
911 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
912 if (rw_recursed(rw))
913 db_printf(" recursed: %u\n", rw->rw_recurse);
914 }
915 db_printf(" waiters: ");
916 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
917 case RW_LOCK_READ_WAITERS:
918 db_printf("readers\n");
919 break;
920 case RW_LOCK_WRITE_WAITERS:
921 db_printf("writers\n");
922 break;
923 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
924 db_printf("readers and writers\n");
925 break;
926 default:
927 db_printf("none\n");
928 break;
929 }
930 }
931
932 #endif
Cache object: e6370c0e7914236e31b901239fc03502
|