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: releng/10.0/sys/kern/kern_rwlock.c 255788 2013-09-22 14:09:07Z davide $");
36
37 #include "opt_ddb.h"
38 #include "opt_hwpmc_hooks.h"
39 #include "opt_kdtrace.h"
40 #include "opt_no_adaptive_rwlocks.h"
41
42 #include <sys/param.h>
43 #include <sys/kdb.h>
44 #include <sys/ktr.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/proc.h>
49 #include <sys/rwlock.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 #include <sys/turnstile.h>
53
54 #include <machine/cpu.h>
55
56 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
57 #define ADAPTIVE_RWLOCKS
58 #endif
59
60 #ifdef HWPMC_HOOKS
61 #include <sys/pmckern.h>
62 PMC_SOFT_DECLARE( , , lock, failed);
63 #endif
64
65 /*
66 * Return the rwlock address when the lock cookie address is provided.
67 * This functionality assumes that struct rwlock* have a member named rw_lock.
68 */
69 #define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock))
70
71 #ifdef ADAPTIVE_RWLOCKS
72 static int rowner_retries = 10;
73 static int rowner_loops = 10000;
74 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
75 "rwlock debugging");
76 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
77 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
78 #endif
79
80 #ifdef DDB
81 #include <ddb/ddb.h>
82
83 static void db_show_rwlock(const struct lock_object *lock);
84 #endif
85 static void assert_rw(const struct lock_object *lock, int what);
86 static void lock_rw(struct lock_object *lock, uintptr_t how);
87 #ifdef KDTRACE_HOOKS
88 static int owner_rw(const struct lock_object *lock, struct thread **owner);
89 #endif
90 static uintptr_t unlock_rw(struct lock_object *lock);
91
92 struct lock_class lock_class_rw = {
93 .lc_name = "rw",
94 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
95 .lc_assert = assert_rw,
96 #ifdef DDB
97 .lc_ddb_show = db_show_rwlock,
98 #endif
99 .lc_lock = lock_rw,
100 .lc_unlock = unlock_rw,
101 #ifdef KDTRACE_HOOKS
102 .lc_owner = owner_rw,
103 #endif
104 };
105
106 /*
107 * Return a pointer to the owning thread if the lock is write-locked or
108 * NULL if the lock is unlocked or read-locked.
109 */
110 #define rw_wowner(rw) \
111 ((rw)->rw_lock & RW_LOCK_READ ? NULL : \
112 (struct thread *)RW_OWNER((rw)->rw_lock))
113
114 /*
115 * Returns if a write owner is recursed. Write ownership is not assured
116 * here and should be previously checked.
117 */
118 #define rw_recursed(rw) ((rw)->rw_recurse != 0)
119
120 /*
121 * Return true if curthread helds the lock.
122 */
123 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread)
124
125 /*
126 * Return a pointer to the owning thread for this lock who should receive
127 * any priority lent by threads that block on this lock. Currently this
128 * is identical to rw_wowner().
129 */
130 #define rw_owner(rw) rw_wowner(rw)
131
132 #ifndef INVARIANTS
133 #define __rw_assert(c, what, file, line)
134 #endif
135
136 void
137 assert_rw(const struct lock_object *lock, int what)
138 {
139
140 rw_assert((const struct rwlock *)lock, what);
141 }
142
143 void
144 lock_rw(struct lock_object *lock, uintptr_t how)
145 {
146 struct rwlock *rw;
147
148 rw = (struct rwlock *)lock;
149 if (how)
150 rw_rlock(rw);
151 else
152 rw_wlock(rw);
153 }
154
155 uintptr_t
156 unlock_rw(struct lock_object *lock)
157 {
158 struct rwlock *rw;
159
160 rw = (struct rwlock *)lock;
161 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
162 if (rw->rw_lock & RW_LOCK_READ) {
163 rw_runlock(rw);
164 return (1);
165 } else {
166 rw_wunlock(rw);
167 return (0);
168 }
169 }
170
171 #ifdef KDTRACE_HOOKS
172 int
173 owner_rw(const struct lock_object *lock, struct thread **owner)
174 {
175 const struct rwlock *rw = (const struct rwlock *)lock;
176 uintptr_t x = rw->rw_lock;
177
178 *owner = rw_wowner(rw);
179 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) :
180 (*owner != NULL));
181 }
182 #endif
183
184 void
185 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
186 {
187 struct rwlock *rw;
188 int flags;
189
190 rw = rwlock2rw(c);
191
192 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
193 RW_RECURSE)) == 0);
194 ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
195 ("%s: rw_lock not aligned for %s: %p", __func__, name,
196 &rw->rw_lock));
197
198 flags = LO_UPGRADABLE;
199 if (opts & RW_DUPOK)
200 flags |= LO_DUPOK;
201 if (opts & RW_NOPROFILE)
202 flags |= LO_NOPROFILE;
203 if (!(opts & RW_NOWITNESS))
204 flags |= LO_WITNESS;
205 if (opts & RW_RECURSE)
206 flags |= LO_RECURSABLE;
207 if (opts & RW_QUIET)
208 flags |= LO_QUIET;
209
210 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
211 rw->rw_lock = RW_UNLOCKED;
212 rw->rw_recurse = 0;
213 }
214
215 void
216 _rw_destroy(volatile uintptr_t *c)
217 {
218 struct rwlock *rw;
219
220 rw = rwlock2rw(c);
221
222 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
223 KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
224 rw->rw_lock = RW_DESTROYED;
225 lock_destroy(&rw->lock_object);
226 }
227
228 void
229 rw_sysinit(void *arg)
230 {
231 struct rw_args *args = arg;
232
233 rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
234 }
235
236 void
237 rw_sysinit_flags(void *arg)
238 {
239 struct rw_args_flags *args = arg;
240
241 rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
242 args->ra_flags);
243 }
244
245 int
246 _rw_wowned(const volatile uintptr_t *c)
247 {
248
249 return (rw_wowner(rwlock2rw(c)) == curthread);
250 }
251
252 void
253 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
254 {
255 struct rwlock *rw;
256
257 if (SCHEDULER_STOPPED())
258 return;
259
260 rw = rwlock2rw(c);
261
262 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
263 ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
264 curthread, rw->lock_object.lo_name, file, line));
265 KASSERT(rw->rw_lock != RW_DESTROYED,
266 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
267 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
268 line, NULL);
269 __rw_wlock(rw, curthread, file, line);
270 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
271 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
272 curthread->td_locks++;
273 }
274
275 int
276 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
277 {
278 struct rwlock *rw;
279 int rval;
280
281 if (SCHEDULER_STOPPED())
282 return (1);
283
284 rw = rwlock2rw(c);
285
286 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
287 ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
288 curthread, rw->lock_object.lo_name, file, line));
289 KASSERT(rw->rw_lock != RW_DESTROYED,
290 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
291
292 if (rw_wlocked(rw) &&
293 (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
294 rw->rw_recurse++;
295 rval = 1;
296 } else
297 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
298 (uintptr_t)curthread);
299
300 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
301 if (rval) {
302 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
303 file, line);
304 curthread->td_locks++;
305 }
306 return (rval);
307 }
308
309 void
310 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
311 {
312 struct rwlock *rw;
313
314 if (SCHEDULER_STOPPED())
315 return;
316
317 rw = rwlock2rw(c);
318
319 KASSERT(rw->rw_lock != RW_DESTROYED,
320 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
321 __rw_assert(c, RA_WLOCKED, file, line);
322 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
323 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
324 line);
325 if (!rw_recursed(rw))
326 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
327 __rw_wunlock(rw, curthread, file, line);
328 curthread->td_locks--;
329 }
330 /*
331 * Determines whether a new reader can acquire a lock. Succeeds if the
332 * reader already owns a read lock and the lock is locked for read to
333 * prevent deadlock from reader recursion. Also succeeds if the lock
334 * is unlocked and has no writer waiters or spinners. Failing otherwise
335 * prioritizes writers before readers.
336 */
337 #define RW_CAN_READ(_rw) \
338 ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) & \
339 (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) == \
340 RW_LOCK_READ)
341
342 void
343 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
344 {
345 struct rwlock *rw;
346 struct turnstile *ts;
347 #ifdef ADAPTIVE_RWLOCKS
348 volatile struct thread *owner;
349 int spintries = 0;
350 int i;
351 #endif
352 #ifdef LOCK_PROFILING
353 uint64_t waittime = 0;
354 int contested = 0;
355 #endif
356 uintptr_t v;
357 #ifdef KDTRACE_HOOKS
358 uint64_t spin_cnt = 0;
359 uint64_t sleep_cnt = 0;
360 int64_t sleep_time = 0;
361 #endif
362
363 if (SCHEDULER_STOPPED())
364 return;
365
366 rw = rwlock2rw(c);
367
368 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
369 ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
370 curthread, rw->lock_object.lo_name, file, line));
371 KASSERT(rw->rw_lock != RW_DESTROYED,
372 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
373 KASSERT(rw_wowner(rw) != curthread,
374 ("rw_rlock: wlock already held for %s @ %s:%d",
375 rw->lock_object.lo_name, file, line));
376 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
377
378 for (;;) {
379 #ifdef KDTRACE_HOOKS
380 spin_cnt++;
381 #endif
382 /*
383 * Handle the easy case. If no other thread has a write
384 * lock, then try to bump up the count of read locks. Note
385 * that we have to preserve the current state of the
386 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
387 * read lock, then rw_lock must have changed, so restart
388 * the loop. Note that this handles the case of a
389 * completely unlocked rwlock since such a lock is encoded
390 * as a read lock with no waiters.
391 */
392 v = rw->rw_lock;
393 if (RW_CAN_READ(v)) {
394 /*
395 * The RW_LOCK_READ_WAITERS flag should only be set
396 * if the lock has been unlocked and write waiters
397 * were present.
398 */
399 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
400 v + RW_ONE_READER)) {
401 if (LOCK_LOG_TEST(&rw->lock_object, 0))
402 CTR4(KTR_LOCK,
403 "%s: %p succeed %p -> %p", __func__,
404 rw, (void *)v,
405 (void *)(v + RW_ONE_READER));
406 break;
407 }
408 continue;
409 }
410 #ifdef HWPMC_HOOKS
411 PMC_SOFT_CALL( , , lock, failed);
412 #endif
413 lock_profile_obtain_lock_failed(&rw->lock_object,
414 &contested, &waittime);
415
416 #ifdef ADAPTIVE_RWLOCKS
417 /*
418 * If the owner is running on another CPU, spin until
419 * the owner stops running or the state of the lock
420 * changes.
421 */
422 if ((v & RW_LOCK_READ) == 0) {
423 owner = (struct thread *)RW_OWNER(v);
424 if (TD_IS_RUNNING(owner)) {
425 if (LOCK_LOG_TEST(&rw->lock_object, 0))
426 CTR3(KTR_LOCK,
427 "%s: spinning on %p held by %p",
428 __func__, rw, owner);
429 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
430 owner && TD_IS_RUNNING(owner)) {
431 cpu_spinwait();
432 #ifdef KDTRACE_HOOKS
433 spin_cnt++;
434 #endif
435 }
436 continue;
437 }
438 } else if (spintries < rowner_retries) {
439 spintries++;
440 for (i = 0; i < rowner_loops; i++) {
441 v = rw->rw_lock;
442 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
443 break;
444 cpu_spinwait();
445 }
446 if (i != rowner_loops)
447 continue;
448 }
449 #endif
450
451 /*
452 * Okay, now it's the hard case. Some other thread already
453 * has a write lock or there are write waiters present,
454 * acquire the turnstile lock so we can begin the process
455 * of blocking.
456 */
457 ts = turnstile_trywait(&rw->lock_object);
458
459 /*
460 * The lock might have been released while we spun, so
461 * recheck its state and restart the loop if needed.
462 */
463 v = rw->rw_lock;
464 if (RW_CAN_READ(v)) {
465 turnstile_cancel(ts);
466 continue;
467 }
468
469 #ifdef ADAPTIVE_RWLOCKS
470 /*
471 * The current lock owner might have started executing
472 * on another CPU (or the lock could have changed
473 * owners) while we were waiting on the turnstile
474 * chain lock. If so, drop the turnstile lock and try
475 * again.
476 */
477 if ((v & RW_LOCK_READ) == 0) {
478 owner = (struct thread *)RW_OWNER(v);
479 if (TD_IS_RUNNING(owner)) {
480 turnstile_cancel(ts);
481 continue;
482 }
483 }
484 #endif
485
486 /*
487 * The lock is held in write mode or it already has waiters.
488 */
489 MPASS(!RW_CAN_READ(v));
490
491 /*
492 * If the RW_LOCK_READ_WAITERS flag is already set, then
493 * we can go ahead and block. If it is not set then try
494 * to set it. If we fail to set it drop the turnstile
495 * lock and restart the loop.
496 */
497 if (!(v & RW_LOCK_READ_WAITERS)) {
498 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
499 v | RW_LOCK_READ_WAITERS)) {
500 turnstile_cancel(ts);
501 continue;
502 }
503 if (LOCK_LOG_TEST(&rw->lock_object, 0))
504 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
505 __func__, rw);
506 }
507
508 /*
509 * We were unable to acquire the lock and the read waiters
510 * flag is set, so we must block on the turnstile.
511 */
512 if (LOCK_LOG_TEST(&rw->lock_object, 0))
513 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
514 rw);
515 #ifdef KDTRACE_HOOKS
516 sleep_time -= lockstat_nsecs();
517 #endif
518 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
519 #ifdef KDTRACE_HOOKS
520 sleep_time += lockstat_nsecs();
521 sleep_cnt++;
522 #endif
523 if (LOCK_LOG_TEST(&rw->lock_object, 0))
524 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
525 __func__, rw);
526 }
527
528 /*
529 * TODO: acquire "owner of record" here. Here be turnstile dragons
530 * however. turnstiles don't like owners changing between calls to
531 * turnstile_wait() currently.
532 */
533 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
534 waittime, file, line);
535 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
536 WITNESS_LOCK(&rw->lock_object, 0, file, line);
537 curthread->td_locks++;
538 curthread->td_rw_rlocks++;
539 #ifdef KDTRACE_HOOKS
540 if (sleep_time)
541 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
542
543 /*
544 * Record only the loops spinning and not sleeping.
545 */
546 if (spin_cnt > sleep_cnt)
547 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
548 #endif
549 }
550
551 int
552 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
553 {
554 struct rwlock *rw;
555 uintptr_t x;
556
557 if (SCHEDULER_STOPPED())
558 return (1);
559
560 rw = rwlock2rw(c);
561
562 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
563 ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
564 curthread, rw->lock_object.lo_name, file, line));
565
566 for (;;) {
567 x = rw->rw_lock;
568 KASSERT(rw->rw_lock != RW_DESTROYED,
569 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
570 if (!(x & RW_LOCK_READ))
571 break;
572 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
573 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
574 line);
575 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
576 curthread->td_locks++;
577 curthread->td_rw_rlocks++;
578 return (1);
579 }
580 }
581
582 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
583 return (0);
584 }
585
586 void
587 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
588 {
589 struct rwlock *rw;
590 struct turnstile *ts;
591 uintptr_t x, v, queue;
592
593 if (SCHEDULER_STOPPED())
594 return;
595
596 rw = rwlock2rw(c);
597
598 KASSERT(rw->rw_lock != RW_DESTROYED,
599 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
600 __rw_assert(c, RA_RLOCKED, file, line);
601 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
602 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
603
604 /* TODO: drop "owner of record" here. */
605
606 for (;;) {
607 /*
608 * See if there is more than one read lock held. If so,
609 * just drop one and return.
610 */
611 x = rw->rw_lock;
612 if (RW_READERS(x) > 1) {
613 if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
614 x - RW_ONE_READER)) {
615 if (LOCK_LOG_TEST(&rw->lock_object, 0))
616 CTR4(KTR_LOCK,
617 "%s: %p succeeded %p -> %p",
618 __func__, rw, (void *)x,
619 (void *)(x - RW_ONE_READER));
620 break;
621 }
622 continue;
623 }
624 /*
625 * If there aren't any waiters for a write lock, then try
626 * to drop it quickly.
627 */
628 if (!(x & RW_LOCK_WAITERS)) {
629 MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
630 RW_READERS_LOCK(1));
631 if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
632 RW_UNLOCKED)) {
633 if (LOCK_LOG_TEST(&rw->lock_object, 0))
634 CTR2(KTR_LOCK, "%s: %p last succeeded",
635 __func__, rw);
636 break;
637 }
638 continue;
639 }
640 /*
641 * Ok, we know we have waiters and we think we are the
642 * last reader, so grab the turnstile lock.
643 */
644 turnstile_chain_lock(&rw->lock_object);
645 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
646 MPASS(v & RW_LOCK_WAITERS);
647
648 /*
649 * Try to drop our lock leaving the lock in a unlocked
650 * state.
651 *
652 * If you wanted to do explicit lock handoff you'd have to
653 * do it here. You'd also want to use turnstile_signal()
654 * and you'd have to handle the race where a higher
655 * priority thread blocks on the write lock before the
656 * thread you wakeup actually runs and have the new thread
657 * "steal" the lock. For now it's a lot simpler to just
658 * wakeup all of the waiters.
659 *
660 * As above, if we fail, then another thread might have
661 * acquired a read lock, so drop the turnstile lock and
662 * restart.
663 */
664 x = RW_UNLOCKED;
665 if (v & RW_LOCK_WRITE_WAITERS) {
666 queue = TS_EXCLUSIVE_QUEUE;
667 x |= (v & RW_LOCK_READ_WAITERS);
668 } else
669 queue = TS_SHARED_QUEUE;
670 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
671 x)) {
672 turnstile_chain_unlock(&rw->lock_object);
673 continue;
674 }
675 if (LOCK_LOG_TEST(&rw->lock_object, 0))
676 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
677 __func__, rw);
678
679 /*
680 * Ok. The lock is released and all that's left is to
681 * wake up the waiters. Note that the lock might not be
682 * free anymore, but in that case the writers will just
683 * block again if they run before the new lock holder(s)
684 * release the lock.
685 */
686 ts = turnstile_lookup(&rw->lock_object);
687 MPASS(ts != NULL);
688 turnstile_broadcast(ts, queue);
689 turnstile_unpend(ts, TS_SHARED_LOCK);
690 turnstile_chain_unlock(&rw->lock_object);
691 break;
692 }
693 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
694 curthread->td_locks--;
695 curthread->td_rw_rlocks--;
696 }
697
698 /*
699 * This function is called when we are unable to obtain a write lock on the
700 * first try. This means that at least one other thread holds either a
701 * read or write lock.
702 */
703 void
704 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
705 int line)
706 {
707 struct rwlock *rw;
708 struct turnstile *ts;
709 #ifdef ADAPTIVE_RWLOCKS
710 volatile struct thread *owner;
711 int spintries = 0;
712 int i;
713 #endif
714 uintptr_t v, x;
715 #ifdef LOCK_PROFILING
716 uint64_t waittime = 0;
717 int contested = 0;
718 #endif
719 #ifdef KDTRACE_HOOKS
720 uint64_t spin_cnt = 0;
721 uint64_t sleep_cnt = 0;
722 int64_t sleep_time = 0;
723 #endif
724
725 if (SCHEDULER_STOPPED())
726 return;
727
728 rw = rwlock2rw(c);
729
730 if (rw_wlocked(rw)) {
731 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
732 ("%s: recursing but non-recursive rw %s @ %s:%d\n",
733 __func__, rw->lock_object.lo_name, file, line));
734 rw->rw_recurse++;
735 if (LOCK_LOG_TEST(&rw->lock_object, 0))
736 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
737 return;
738 }
739
740 if (LOCK_LOG_TEST(&rw->lock_object, 0))
741 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
742 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
743
744 while (!_rw_write_lock(rw, tid)) {
745 #ifdef KDTRACE_HOOKS
746 spin_cnt++;
747 #endif
748 #ifdef HWPMC_HOOKS
749 PMC_SOFT_CALL( , , lock, failed);
750 #endif
751 lock_profile_obtain_lock_failed(&rw->lock_object,
752 &contested, &waittime);
753 #ifdef ADAPTIVE_RWLOCKS
754 /*
755 * If the lock is write locked and the owner is
756 * running on another CPU, spin until the owner stops
757 * running or the state of the lock changes.
758 */
759 v = rw->rw_lock;
760 owner = (struct thread *)RW_OWNER(v);
761 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
762 if (LOCK_LOG_TEST(&rw->lock_object, 0))
763 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
764 __func__, rw, owner);
765 while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
766 TD_IS_RUNNING(owner)) {
767 cpu_spinwait();
768 #ifdef KDTRACE_HOOKS
769 spin_cnt++;
770 #endif
771 }
772 continue;
773 }
774 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
775 spintries < rowner_retries) {
776 if (!(v & RW_LOCK_WRITE_SPINNER)) {
777 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
778 v | RW_LOCK_WRITE_SPINNER)) {
779 continue;
780 }
781 }
782 spintries++;
783 for (i = 0; i < rowner_loops; i++) {
784 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
785 break;
786 cpu_spinwait();
787 }
788 #ifdef KDTRACE_HOOKS
789 spin_cnt += rowner_loops - i;
790 #endif
791 if (i != rowner_loops)
792 continue;
793 }
794 #endif
795 ts = turnstile_trywait(&rw->lock_object);
796 v = rw->rw_lock;
797
798 #ifdef ADAPTIVE_RWLOCKS
799 /*
800 * The current lock owner might have started executing
801 * on another CPU (or the lock could have changed
802 * owners) while we were waiting on the turnstile
803 * chain lock. If so, drop the turnstile lock and try
804 * again.
805 */
806 if (!(v & RW_LOCK_READ)) {
807 owner = (struct thread *)RW_OWNER(v);
808 if (TD_IS_RUNNING(owner)) {
809 turnstile_cancel(ts);
810 continue;
811 }
812 }
813 #endif
814 /*
815 * Check for the waiters flags about this rwlock.
816 * If the lock was released, without maintain any pending
817 * waiters queue, simply try to acquire it.
818 * If a pending waiters queue is present, claim the lock
819 * ownership and maintain the pending queue.
820 */
821 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
822 if ((v & ~x) == RW_UNLOCKED) {
823 x &= ~RW_LOCK_WRITE_SPINNER;
824 if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
825 if (x)
826 turnstile_claim(ts);
827 else
828 turnstile_cancel(ts);
829 break;
830 }
831 turnstile_cancel(ts);
832 continue;
833 }
834 /*
835 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
836 * set it. If we fail to set it, then loop back and try
837 * again.
838 */
839 if (!(v & RW_LOCK_WRITE_WAITERS)) {
840 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
841 v | RW_LOCK_WRITE_WAITERS)) {
842 turnstile_cancel(ts);
843 continue;
844 }
845 if (LOCK_LOG_TEST(&rw->lock_object, 0))
846 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
847 __func__, rw);
848 }
849 /*
850 * We were unable to acquire the lock and the write waiters
851 * flag is set, so we must block on the turnstile.
852 */
853 if (LOCK_LOG_TEST(&rw->lock_object, 0))
854 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
855 rw);
856 #ifdef KDTRACE_HOOKS
857 sleep_time -= lockstat_nsecs();
858 #endif
859 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
860 #ifdef KDTRACE_HOOKS
861 sleep_time += lockstat_nsecs();
862 sleep_cnt++;
863 #endif
864 if (LOCK_LOG_TEST(&rw->lock_object, 0))
865 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
866 __func__, rw);
867 #ifdef ADAPTIVE_RWLOCKS
868 spintries = 0;
869 #endif
870 }
871 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
872 waittime, file, line);
873 #ifdef KDTRACE_HOOKS
874 if (sleep_time)
875 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
876
877 /*
878 * Record only the loops spinning and not sleeping.
879 */
880 if (spin_cnt > sleep_cnt)
881 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
882 #endif
883 }
884
885 /*
886 * This function is called if the first try at releasing a write lock failed.
887 * This means that one of the 2 waiter bits must be set indicating that at
888 * least one thread is waiting on this lock.
889 */
890 void
891 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
892 int line)
893 {
894 struct rwlock *rw;
895 struct turnstile *ts;
896 uintptr_t v;
897 int queue;
898
899 if (SCHEDULER_STOPPED())
900 return;
901
902 rw = rwlock2rw(c);
903
904 if (rw_wlocked(rw) && rw_recursed(rw)) {
905 rw->rw_recurse--;
906 if (LOCK_LOG_TEST(&rw->lock_object, 0))
907 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
908 return;
909 }
910
911 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
912 ("%s: neither of the waiter flags are set", __func__));
913
914 if (LOCK_LOG_TEST(&rw->lock_object, 0))
915 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
916
917 turnstile_chain_lock(&rw->lock_object);
918 ts = turnstile_lookup(&rw->lock_object);
919 MPASS(ts != NULL);
920
921 /*
922 * Use the same algo as sx locks for now. Prefer waking up shared
923 * waiters if we have any over writers. This is probably not ideal.
924 *
925 * 'v' is the value we are going to write back to rw_lock. If we
926 * have waiters on both queues, we need to preserve the state of
927 * the waiter flag for the queue we don't wake up. For now this is
928 * hardcoded for the algorithm mentioned above.
929 *
930 * In the case of both readers and writers waiting we wakeup the
931 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
932 * new writer comes in before a reader it will claim the lock up
933 * above. There is probably a potential priority inversion in
934 * there that could be worked around either by waking both queues
935 * of waiters or doing some complicated lock handoff gymnastics.
936 */
937 v = RW_UNLOCKED;
938 if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
939 queue = TS_EXCLUSIVE_QUEUE;
940 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
941 } else
942 queue = TS_SHARED_QUEUE;
943
944 /* Wake up all waiters for the specific queue. */
945 if (LOCK_LOG_TEST(&rw->lock_object, 0))
946 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
947 queue == TS_SHARED_QUEUE ? "read" : "write");
948 turnstile_broadcast(ts, queue);
949 atomic_store_rel_ptr(&rw->rw_lock, v);
950 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
951 turnstile_chain_unlock(&rw->lock_object);
952 }
953
954 /*
955 * Attempt to do a non-blocking upgrade from a read lock to a write
956 * lock. This will only succeed if this thread holds a single read
957 * lock. Returns true if the upgrade succeeded and false otherwise.
958 */
959 int
960 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
961 {
962 struct rwlock *rw;
963 uintptr_t v, x, tid;
964 struct turnstile *ts;
965 int success;
966
967 if (SCHEDULER_STOPPED())
968 return (1);
969
970 rw = rwlock2rw(c);
971
972 KASSERT(rw->rw_lock != RW_DESTROYED,
973 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
974 __rw_assert(c, RA_RLOCKED, file, line);
975
976 /*
977 * Attempt to switch from one reader to a writer. If there
978 * are any write waiters, then we will have to lock the
979 * turnstile first to prevent races with another writer
980 * calling turnstile_wait() before we have claimed this
981 * turnstile. So, do the simple case of no waiters first.
982 */
983 tid = (uintptr_t)curthread;
984 success = 0;
985 for (;;) {
986 v = rw->rw_lock;
987 if (RW_READERS(v) > 1)
988 break;
989 if (!(v & RW_LOCK_WAITERS)) {
990 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
991 if (!success)
992 continue;
993 break;
994 }
995
996 /*
997 * Ok, we think we have waiters, so lock the turnstile.
998 */
999 ts = turnstile_trywait(&rw->lock_object);
1000 v = rw->rw_lock;
1001 if (RW_READERS(v) > 1) {
1002 turnstile_cancel(ts);
1003 break;
1004 }
1005 /*
1006 * Try to switch from one reader to a writer again. This time
1007 * we honor the current state of the waiters flags.
1008 * If we obtain the lock with the flags set, then claim
1009 * ownership of the turnstile.
1010 */
1011 x = rw->rw_lock & RW_LOCK_WAITERS;
1012 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1013 if (success) {
1014 if (x)
1015 turnstile_claim(ts);
1016 else
1017 turnstile_cancel(ts);
1018 break;
1019 }
1020 turnstile_cancel(ts);
1021 }
1022 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1023 if (success) {
1024 curthread->td_rw_rlocks--;
1025 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1026 file, line);
1027 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
1028 }
1029 return (success);
1030 }
1031
1032 /*
1033 * Downgrade a write lock into a single read lock.
1034 */
1035 void
1036 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1037 {
1038 struct rwlock *rw;
1039 struct turnstile *ts;
1040 uintptr_t tid, v;
1041 int rwait, wwait;
1042
1043 if (SCHEDULER_STOPPED())
1044 return;
1045
1046 rw = rwlock2rw(c);
1047
1048 KASSERT(rw->rw_lock != RW_DESTROYED,
1049 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1050 __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1051 #ifndef INVARIANTS
1052 if (rw_recursed(rw))
1053 panic("downgrade of a recursed lock");
1054 #endif
1055
1056 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1057
1058 /*
1059 * Convert from a writer to a single reader. First we handle
1060 * the easy case with no waiters. If there are any waiters, we
1061 * lock the turnstile and "disown" the lock.
1062 */
1063 tid = (uintptr_t)curthread;
1064 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1065 goto out;
1066
1067 /*
1068 * Ok, we think we have waiters, so lock the turnstile so we can
1069 * read the waiter flags without any races.
1070 */
1071 turnstile_chain_lock(&rw->lock_object);
1072 v = rw->rw_lock & RW_LOCK_WAITERS;
1073 rwait = v & RW_LOCK_READ_WAITERS;
1074 wwait = v & RW_LOCK_WRITE_WAITERS;
1075 MPASS(rwait | wwait);
1076
1077 /*
1078 * Downgrade from a write lock while preserving waiters flag
1079 * and give up ownership of the turnstile.
1080 */
1081 ts = turnstile_lookup(&rw->lock_object);
1082 MPASS(ts != NULL);
1083 if (!wwait)
1084 v &= ~RW_LOCK_READ_WAITERS;
1085 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1086 /*
1087 * Wake other readers if there are no writers pending. Otherwise they
1088 * won't be able to acquire the lock anyway.
1089 */
1090 if (rwait && !wwait) {
1091 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1092 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1093 } else
1094 turnstile_disown(ts);
1095 turnstile_chain_unlock(&rw->lock_object);
1096 out:
1097 curthread->td_rw_rlocks++;
1098 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1099 LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1100 }
1101
1102 #ifdef INVARIANT_SUPPORT
1103 #ifndef INVARIANTS
1104 #undef __rw_assert
1105 #endif
1106
1107 /*
1108 * In the non-WITNESS case, rw_assert() can only detect that at least
1109 * *some* thread owns an rlock, but it cannot guarantee that *this*
1110 * thread owns an rlock.
1111 */
1112 void
1113 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1114 {
1115 const struct rwlock *rw;
1116
1117 if (panicstr != NULL)
1118 return;
1119
1120 rw = rwlock2rw(c);
1121
1122 switch (what) {
1123 case RA_LOCKED:
1124 case RA_LOCKED | RA_RECURSED:
1125 case RA_LOCKED | RA_NOTRECURSED:
1126 case RA_RLOCKED:
1127 case RA_RLOCKED | RA_RECURSED:
1128 case RA_RLOCKED | RA_NOTRECURSED:
1129 #ifdef WITNESS
1130 witness_assert(&rw->lock_object, what, file, line);
1131 #else
1132 /*
1133 * If some other thread has a write lock or we have one
1134 * and are asserting a read lock, fail. Also, if no one
1135 * has a lock at all, fail.
1136 */
1137 if (rw->rw_lock == RW_UNLOCKED ||
1138 (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1139 rw_wowner(rw) != curthread)))
1140 panic("Lock %s not %slocked @ %s:%d\n",
1141 rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1142 "read " : "", file, line);
1143
1144 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1145 if (rw_recursed(rw)) {
1146 if (what & RA_NOTRECURSED)
1147 panic("Lock %s recursed @ %s:%d\n",
1148 rw->lock_object.lo_name, file,
1149 line);
1150 } else if (what & RA_RECURSED)
1151 panic("Lock %s not recursed @ %s:%d\n",
1152 rw->lock_object.lo_name, file, line);
1153 }
1154 #endif
1155 break;
1156 case RA_WLOCKED:
1157 case RA_WLOCKED | RA_RECURSED:
1158 case RA_WLOCKED | RA_NOTRECURSED:
1159 if (rw_wowner(rw) != curthread)
1160 panic("Lock %s not exclusively locked @ %s:%d\n",
1161 rw->lock_object.lo_name, file, line);
1162 if (rw_recursed(rw)) {
1163 if (what & RA_NOTRECURSED)
1164 panic("Lock %s recursed @ %s:%d\n",
1165 rw->lock_object.lo_name, file, line);
1166 } else if (what & RA_RECURSED)
1167 panic("Lock %s not recursed @ %s:%d\n",
1168 rw->lock_object.lo_name, file, line);
1169 break;
1170 case RA_UNLOCKED:
1171 #ifdef WITNESS
1172 witness_assert(&rw->lock_object, what, file, line);
1173 #else
1174 /*
1175 * If we hold a write lock fail. We can't reliably check
1176 * to see if we hold a read lock or not.
1177 */
1178 if (rw_wowner(rw) == curthread)
1179 panic("Lock %s exclusively locked @ %s:%d\n",
1180 rw->lock_object.lo_name, file, line);
1181 #endif
1182 break;
1183 default:
1184 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1185 line);
1186 }
1187 }
1188 #endif /* INVARIANT_SUPPORT */
1189
1190 #ifdef DDB
1191 void
1192 db_show_rwlock(const struct lock_object *lock)
1193 {
1194 const struct rwlock *rw;
1195 struct thread *td;
1196
1197 rw = (const struct rwlock *)lock;
1198
1199 db_printf(" state: ");
1200 if (rw->rw_lock == RW_UNLOCKED)
1201 db_printf("UNLOCKED\n");
1202 else if (rw->rw_lock == RW_DESTROYED) {
1203 db_printf("DESTROYED\n");
1204 return;
1205 } else if (rw->rw_lock & RW_LOCK_READ)
1206 db_printf("RLOCK: %ju locks\n",
1207 (uintmax_t)(RW_READERS(rw->rw_lock)));
1208 else {
1209 td = rw_wowner(rw);
1210 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1211 td->td_tid, td->td_proc->p_pid, td->td_name);
1212 if (rw_recursed(rw))
1213 db_printf(" recursed: %u\n", rw->rw_recurse);
1214 }
1215 db_printf(" waiters: ");
1216 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1217 case RW_LOCK_READ_WAITERS:
1218 db_printf("readers\n");
1219 break;
1220 case RW_LOCK_WRITE_WAITERS:
1221 db_printf("writers\n");
1222 break;
1223 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1224 db_printf("readers and writers\n");
1225 break;
1226 default:
1227 db_printf("none\n");
1228 break;
1229 }
1230 }
1231
1232 #endif
Cache object: 9b0d53e5fe610855beaf3a35e8f8f536
|