1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 */
27
28 /*
29 * Machine independent bits of reader/writer lock implementation.
30 */
31
32 #include <sys/cdefs.h>
33 #include "opt_ddb.h"
34 #include "opt_hwpmc_hooks.h"
35 #include "opt_no_adaptive_rwlocks.h"
36
37 #include <sys/param.h>
38 #include <sys/kdb.h>
39 #include <sys/ktr.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/rwlock.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/systm.h>
49 #include <sys/turnstile.h>
50
51 #include <machine/cpu.h>
52
53 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
54 #define ADAPTIVE_RWLOCKS
55 #endif
56
57 #ifdef HWPMC_HOOKS
58 #include <sys/pmckern.h>
59 PMC_SOFT_DECLARE( , , lock, failed);
60 #endif
61
62 /*
63 * Return the rwlock address when the lock cookie address is provided.
64 * This functionality assumes that struct rwlock* have a member named rw_lock.
65 */
66 #define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock))
67
68 #ifdef DDB
69 #include <ddb/ddb.h>
70
71 static void db_show_rwlock(const struct lock_object *lock);
72 #endif
73 static void assert_rw(const struct lock_object *lock, int what);
74 static void lock_rw(struct lock_object *lock, uintptr_t how);
75 #ifdef KDTRACE_HOOKS
76 static int owner_rw(const struct lock_object *lock, struct thread **owner);
77 #endif
78 static uintptr_t unlock_rw(struct lock_object *lock);
79
80 struct lock_class lock_class_rw = {
81 .lc_name = "rw",
82 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
83 .lc_assert = assert_rw,
84 #ifdef DDB
85 .lc_ddb_show = db_show_rwlock,
86 #endif
87 .lc_lock = lock_rw,
88 .lc_unlock = unlock_rw,
89 #ifdef KDTRACE_HOOKS
90 .lc_owner = owner_rw,
91 #endif
92 };
93
94 #ifdef ADAPTIVE_RWLOCKS
95 #ifdef RWLOCK_CUSTOM_BACKOFF
96 static u_short __read_frequently rowner_retries;
97 static u_short __read_frequently rowner_loops;
98 static SYSCTL_NODE(_debug, OID_AUTO, rwlock,
99 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
100 "rwlock debugging");
101 SYSCTL_U16(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
102 SYSCTL_U16(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
103
104 static struct lock_delay_config __read_frequently rw_delay;
105
106 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base,
107 0, "");
108 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
109 0, "");
110
111 static void
rw_lock_delay_init(void * arg __unused)112 rw_lock_delay_init(void *arg __unused)
113 {
114
115 lock_delay_default_init(&rw_delay);
116 rowner_retries = 10;
117 rowner_loops = max(10000, rw_delay.max);
118 }
119 LOCK_DELAY_SYSINIT(rw_lock_delay_init);
120 #else
121 #define rw_delay locks_delay
122 #define rowner_retries locks_delay_retries
123 #define rowner_loops locks_delay_loops
124 #endif
125 #endif
126
127 /*
128 * Return a pointer to the owning thread if the lock is write-locked or
129 * NULL if the lock is unlocked or read-locked.
130 */
131
132 #define lv_rw_wowner(v) \
133 ((v) & RW_LOCK_READ ? NULL : \
134 (struct thread *)RW_OWNER((v)))
135
136 #define rw_wowner(rw) lv_rw_wowner(RW_READ_VALUE(rw))
137
138 /*
139 * Returns if a write owner is recursed. Write ownership is not assured
140 * here and should be previously checked.
141 */
142 #define rw_recursed(rw) ((rw)->rw_recurse != 0)
143
144 /*
145 * Return true if curthread helds the lock.
146 */
147 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread)
148
149 /*
150 * Return a pointer to the owning thread for this lock who should receive
151 * any priority lent by threads that block on this lock. Currently this
152 * is identical to rw_wowner().
153 */
154 #define rw_owner(rw) rw_wowner(rw)
155
156 #ifndef INVARIANTS
157 #define __rw_assert(c, what, file, line)
158 #endif
159
160 void
assert_rw(const struct lock_object * lock,int what)161 assert_rw(const struct lock_object *lock, int what)
162 {
163
164 rw_assert((const struct rwlock *)lock, what);
165 }
166
167 void
lock_rw(struct lock_object * lock,uintptr_t how)168 lock_rw(struct lock_object *lock, uintptr_t how)
169 {
170 struct rwlock *rw;
171
172 rw = (struct rwlock *)lock;
173 if (how)
174 rw_rlock(rw);
175 else
176 rw_wlock(rw);
177 }
178
179 uintptr_t
unlock_rw(struct lock_object * lock)180 unlock_rw(struct lock_object *lock)
181 {
182 struct rwlock *rw;
183
184 rw = (struct rwlock *)lock;
185 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
186 if (rw->rw_lock & RW_LOCK_READ) {
187 rw_runlock(rw);
188 return (1);
189 } else {
190 rw_wunlock(rw);
191 return (0);
192 }
193 }
194
195 #ifdef KDTRACE_HOOKS
196 int
owner_rw(const struct lock_object * lock,struct thread ** owner)197 owner_rw(const struct lock_object *lock, struct thread **owner)
198 {
199 const struct rwlock *rw = (const struct rwlock *)lock;
200 uintptr_t x = rw->rw_lock;
201
202 *owner = rw_wowner(rw);
203 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) :
204 (*owner != NULL));
205 }
206 #endif
207
208 void
_rw_init_flags(volatile uintptr_t * c,const char * name,int opts)209 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
210 {
211 struct rwlock *rw;
212 int flags;
213
214 rw = rwlock2rw(c);
215
216 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
217 RW_RECURSE | RW_NEW)) == 0);
218 ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
219 ("%s: rw_lock not aligned for %s: %p", __func__, name,
220 &rw->rw_lock));
221
222 flags = LO_UPGRADABLE;
223 if (opts & RW_DUPOK)
224 flags |= LO_DUPOK;
225 if (opts & RW_NOPROFILE)
226 flags |= LO_NOPROFILE;
227 if (!(opts & RW_NOWITNESS))
228 flags |= LO_WITNESS;
229 if (opts & RW_RECURSE)
230 flags |= LO_RECURSABLE;
231 if (opts & RW_QUIET)
232 flags |= LO_QUIET;
233 if (opts & RW_NEW)
234 flags |= LO_NEW;
235
236 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
237 rw->rw_lock = RW_UNLOCKED;
238 rw->rw_recurse = 0;
239 }
240
241 void
_rw_destroy(volatile uintptr_t * c)242 _rw_destroy(volatile uintptr_t *c)
243 {
244 struct rwlock *rw;
245
246 rw = rwlock2rw(c);
247
248 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
249 KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
250 rw->rw_lock = RW_DESTROYED;
251 lock_destroy(&rw->lock_object);
252 }
253
254 void
rw_sysinit(void * arg)255 rw_sysinit(void *arg)
256 {
257 struct rw_args *args;
258
259 args = arg;
260 rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
261 args->ra_flags);
262 }
263
264 int
_rw_wowned(const volatile uintptr_t * c)265 _rw_wowned(const volatile uintptr_t *c)
266 {
267
268 return (rw_wowner(rwlock2rw(c)) == curthread);
269 }
270
271 void
_rw_wlock_cookie(volatile uintptr_t * c,const char * file,int line)272 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
273 {
274 struct rwlock *rw;
275 uintptr_t tid, v;
276
277 rw = rwlock2rw(c);
278
279 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
280 !TD_IS_IDLETHREAD(curthread),
281 ("rw_wlock() by idle thread %p on rwlock %p @ %s:%d",
282 curthread, rw, file, line));
283 KASSERT(rw->rw_lock != RW_DESTROYED,
284 ("rw_wlock() of destroyed rwlock %p @ %s:%d", rw, file, line));
285 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
286 line, NULL);
287 tid = (uintptr_t)curthread;
288 v = RW_UNLOCKED;
289 if (!_rw_write_lock_fetch(rw, &v, tid))
290 _rw_wlock_hard(rw, v, file, line);
291 else
292 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
293 0, 0, file, line, LOCKSTAT_WRITER);
294
295 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
296 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
297 TD_LOCKS_INC(curthread);
298 }
299
300 int
__rw_try_wlock_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)301 __rw_try_wlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
302 {
303 struct thread *td;
304 uintptr_t tid, v;
305 int rval;
306 bool recursed;
307
308 td = curthread;
309 tid = (uintptr_t)td;
310 if (SCHEDULER_STOPPED_TD(td))
311 return (1);
312
313 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
314 ("rw_try_wlock() by idle thread %p on rwlock %p @ %s:%d",
315 curthread, rw, file, line));
316 KASSERT(rw->rw_lock != RW_DESTROYED,
317 ("rw_try_wlock() of destroyed rwlock %p @ %s:%d", rw, file, line));
318
319 rval = 1;
320 recursed = false;
321 v = RW_UNLOCKED;
322 for (;;) {
323 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
324 break;
325 if (v == RW_UNLOCKED)
326 continue;
327 if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
328 rw->rw_recurse++;
329 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
330 break;
331 }
332 rval = 0;
333 break;
334 }
335
336 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
337 if (rval) {
338 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
339 file, line);
340 if (!recursed)
341 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
342 rw, 0, 0, file, line, LOCKSTAT_WRITER);
343 TD_LOCKS_INC(curthread);
344 }
345 return (rval);
346 }
347
348 int
__rw_try_wlock(volatile uintptr_t * c,const char * file,int line)349 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
350 {
351 struct rwlock *rw;
352
353 rw = rwlock2rw(c);
354 return (__rw_try_wlock_int(rw LOCK_FILE_LINE_ARG));
355 }
356
357 void
_rw_wunlock_cookie(volatile uintptr_t * c,const char * file,int line)358 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
359 {
360 struct rwlock *rw;
361
362 rw = rwlock2rw(c);
363
364 KASSERT(rw->rw_lock != RW_DESTROYED,
365 ("rw_wunlock() of destroyed rwlock %p @ %s:%d", rw, file, line));
366 __rw_assert(c, RA_WLOCKED, file, line);
367 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
368 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
369 line);
370
371 #ifdef LOCK_PROFILING
372 _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
373 #else
374 __rw_wunlock(rw, curthread, file, line);
375 #endif
376
377 TD_LOCKS_DEC(curthread);
378 }
379
380 /*
381 * Determines whether a new reader can acquire a lock. Succeeds if the
382 * reader already owns a read lock and the lock is locked for read to
383 * prevent deadlock from reader recursion. Also succeeds if the lock
384 * is unlocked and has no writer waiters or spinners. Failing otherwise
385 * prioritizes writers before readers.
386 */
387 static bool __always_inline
__rw_can_read(struct thread * td,uintptr_t v,bool fp)388 __rw_can_read(struct thread *td, uintptr_t v, bool fp)
389 {
390
391 if ((v & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER))
392 == RW_LOCK_READ)
393 return (true);
394 if (!fp && td->td_rw_rlocks && (v & RW_LOCK_READ))
395 return (true);
396 return (false);
397 }
398
399 static bool __always_inline
__rw_rlock_try(struct rwlock * rw,struct thread * td,uintptr_t * vp,bool fp LOCK_FILE_LINE_ARG_DEF)400 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, bool fp
401 LOCK_FILE_LINE_ARG_DEF)
402 {
403
404 /*
405 * Handle the easy case. If no other thread has a write
406 * lock, then try to bump up the count of read locks. Note
407 * that we have to preserve the current state of the
408 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
409 * read lock, then rw_lock must have changed, so restart
410 * the loop. Note that this handles the case of a
411 * completely unlocked rwlock since such a lock is encoded
412 * as a read lock with no waiters.
413 */
414 while (__rw_can_read(td, *vp, fp)) {
415 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
416 *vp + RW_ONE_READER)) {
417 if (LOCK_LOG_TEST(&rw->lock_object, 0))
418 CTR4(KTR_LOCK,
419 "%s: %p succeed %p -> %p", __func__,
420 rw, (void *)*vp,
421 (void *)(*vp + RW_ONE_READER));
422 td->td_rw_rlocks++;
423 return (true);
424 }
425 }
426 return (false);
427 }
428
429 static void __noinline
__rw_rlock_hard(struct rwlock * rw,struct thread * td,uintptr_t v LOCK_FILE_LINE_ARG_DEF)430 __rw_rlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
431 LOCK_FILE_LINE_ARG_DEF)
432 {
433 struct turnstile *ts;
434 struct thread *owner;
435 #ifdef ADAPTIVE_RWLOCKS
436 int spintries = 0;
437 int i, n;
438 #endif
439 #ifdef LOCK_PROFILING
440 uint64_t waittime = 0;
441 int contested = 0;
442 #endif
443 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
444 struct lock_delay_arg lda;
445 #endif
446 #ifdef KDTRACE_HOOKS
447 u_int sleep_cnt = 0;
448 int64_t sleep_time = 0;
449 int64_t all_time = 0;
450 #endif
451 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
452 uintptr_t state = 0;
453 int doing_lockprof = 0;
454 #endif
455
456 #ifdef KDTRACE_HOOKS
457 if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
458 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
459 goto out_lockstat;
460 doing_lockprof = 1;
461 all_time -= lockstat_nsecs(&rw->lock_object);
462 state = v;
463 }
464 #endif
465 #ifdef LOCK_PROFILING
466 doing_lockprof = 1;
467 state = v;
468 #endif
469
470 if (SCHEDULER_STOPPED())
471 return;
472
473 #if defined(ADAPTIVE_RWLOCKS)
474 lock_delay_arg_init(&lda, &rw_delay);
475 #elif defined(KDTRACE_HOOKS)
476 lock_delay_arg_init_noadapt(&lda);
477 #endif
478
479 #ifdef HWPMC_HOOKS
480 PMC_SOFT_CALL( , , lock, failed);
481 #endif
482 lock_profile_obtain_lock_failed(&rw->lock_object, false,
483 &contested, &waittime);
484
485 for (;;) {
486 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
487 break;
488 #ifdef KDTRACE_HOOKS
489 lda.spin_cnt++;
490 #endif
491
492 #ifdef ADAPTIVE_RWLOCKS
493 /*
494 * If the owner is running on another CPU, spin until
495 * the owner stops running or the state of the lock
496 * changes.
497 */
498 if ((v & RW_LOCK_READ) == 0) {
499 owner = (struct thread *)RW_OWNER(v);
500 if (TD_IS_RUNNING(owner)) {
501 if (LOCK_LOG_TEST(&rw->lock_object, 0))
502 CTR3(KTR_LOCK,
503 "%s: spinning on %p held by %p",
504 __func__, rw, owner);
505 KTR_STATE1(KTR_SCHED, "thread",
506 sched_tdname(curthread), "spinning",
507 "lockname:\"%s\"", rw->lock_object.lo_name);
508 do {
509 lock_delay(&lda);
510 v = RW_READ_VALUE(rw);
511 owner = lv_rw_wowner(v);
512 } while (owner != NULL && TD_IS_RUNNING(owner));
513 KTR_STATE0(KTR_SCHED, "thread",
514 sched_tdname(curthread), "running");
515 continue;
516 }
517 } else {
518 if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
519 MPASS(!__rw_can_read(td, v, false));
520 lock_delay_spin(2);
521 v = RW_READ_VALUE(rw);
522 continue;
523 }
524 if (spintries < rowner_retries) {
525 spintries++;
526 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
527 "spinning", "lockname:\"%s\"",
528 rw->lock_object.lo_name);
529 n = RW_READERS(v);
530 for (i = 0; i < rowner_loops; i += n) {
531 lock_delay_spin(n);
532 v = RW_READ_VALUE(rw);
533 if (!(v & RW_LOCK_READ))
534 break;
535 n = RW_READERS(v);
536 if (n == 0)
537 break;
538 if (__rw_can_read(td, v, false))
539 break;
540 }
541 #ifdef KDTRACE_HOOKS
542 lda.spin_cnt += rowner_loops - i;
543 #endif
544 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
545 "running");
546 if (i < rowner_loops)
547 continue;
548 }
549 }
550 #endif
551
552 /*
553 * Okay, now it's the hard case. Some other thread already
554 * has a write lock or there are write waiters present,
555 * acquire the turnstile lock so we can begin the process
556 * of blocking.
557 */
558 ts = turnstile_trywait(&rw->lock_object);
559
560 /*
561 * The lock might have been released while we spun, so
562 * recheck its state and restart the loop if needed.
563 */
564 v = RW_READ_VALUE(rw);
565 retry_ts:
566 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
567 __rw_can_read(td, v, false)) {
568 turnstile_cancel(ts);
569 continue;
570 }
571
572 owner = lv_rw_wowner(v);
573
574 #ifdef ADAPTIVE_RWLOCKS
575 /*
576 * The current lock owner might have started executing
577 * on another CPU (or the lock could have changed
578 * owners) while we were waiting on the turnstile
579 * chain lock. If so, drop the turnstile lock and try
580 * again.
581 */
582 if (owner != NULL) {
583 if (TD_IS_RUNNING(owner)) {
584 turnstile_cancel(ts);
585 continue;
586 }
587 }
588 #endif
589
590 /*
591 * The lock is held in write mode or it already has waiters.
592 */
593 MPASS(!__rw_can_read(td, v, false));
594
595 /*
596 * If the RW_LOCK_READ_WAITERS flag is already set, then
597 * we can go ahead and block. If it is not set then try
598 * to set it. If we fail to set it drop the turnstile
599 * lock and restart the loop.
600 */
601 if (!(v & RW_LOCK_READ_WAITERS)) {
602 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
603 v | RW_LOCK_READ_WAITERS))
604 goto retry_ts;
605 if (LOCK_LOG_TEST(&rw->lock_object, 0))
606 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
607 __func__, rw);
608 }
609
610 /*
611 * We were unable to acquire the lock and the read waiters
612 * flag is set, so we must block on the turnstile.
613 */
614 if (LOCK_LOG_TEST(&rw->lock_object, 0))
615 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
616 rw);
617 #ifdef KDTRACE_HOOKS
618 sleep_time -= lockstat_nsecs(&rw->lock_object);
619 #endif
620 MPASS(owner == rw_owner(rw));
621 turnstile_wait(ts, owner, TS_SHARED_QUEUE);
622 #ifdef KDTRACE_HOOKS
623 sleep_time += lockstat_nsecs(&rw->lock_object);
624 sleep_cnt++;
625 #endif
626 if (LOCK_LOG_TEST(&rw->lock_object, 0))
627 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
628 __func__, rw);
629 v = RW_READ_VALUE(rw);
630 }
631 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
632 if (__predict_true(!doing_lockprof))
633 return;
634 #endif
635 #ifdef KDTRACE_HOOKS
636 all_time += lockstat_nsecs(&rw->lock_object);
637 if (sleep_time)
638 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
639 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
640 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
641
642 /* Record only the loops spinning and not sleeping. */
643 if (lda.spin_cnt > sleep_cnt)
644 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
645 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
646 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
647 out_lockstat:
648 #endif
649 /*
650 * TODO: acquire "owner of record" here. Here be turnstile dragons
651 * however. turnstiles don't like owners changing between calls to
652 * turnstile_wait() currently.
653 */
654 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
655 waittime, file, line, LOCKSTAT_READER);
656 }
657
658 void
__rw_rlock_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)659 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
660 {
661 struct thread *td;
662 uintptr_t v;
663
664 td = curthread;
665
666 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) ||
667 !TD_IS_IDLETHREAD(td),
668 ("rw_rlock() by idle thread %p on rwlock %p @ %s:%d",
669 td, rw, file, line));
670 KASSERT(rw->rw_lock != RW_DESTROYED,
671 ("rw_rlock() of destroyed rwlock %p @ %s:%d", rw, file, line));
672 KASSERT(rw_wowner(rw) != td,
673 ("rw_rlock: wlock already held for %p @ %s:%d",
674 rw, file, line));
675 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
676
677 v = RW_READ_VALUE(rw);
678 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
679 !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
680 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
681 else
682 lock_profile_obtain_lock_success(&rw->lock_object, false, 0, 0,
683 file, line);
684
685 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
686 WITNESS_LOCK(&rw->lock_object, 0, file, line);
687 TD_LOCKS_INC(curthread);
688 }
689
690 void
__rw_rlock(volatile uintptr_t * c,const char * file,int line)691 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
692 {
693 struct rwlock *rw;
694
695 rw = rwlock2rw(c);
696 __rw_rlock_int(rw LOCK_FILE_LINE_ARG);
697 }
698
699 int
__rw_try_rlock_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)700 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
701 {
702 uintptr_t x;
703
704 if (SCHEDULER_STOPPED())
705 return (1);
706
707 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
708 ("rw_try_rlock() by idle thread %p on rwlock %p @ %s:%d",
709 curthread, rw, file, line));
710
711 x = rw->rw_lock;
712 for (;;) {
713 KASSERT(rw->rw_lock != RW_DESTROYED,
714 ("rw_try_rlock() of destroyed rwlock %p @ %s:%d", rw, file,
715 line));
716 if (!(x & RW_LOCK_READ))
717 break;
718 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
719 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
720 line);
721 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
722 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
723 rw, 0, 0, file, line, LOCKSTAT_READER);
724 TD_LOCKS_INC(curthread);
725 curthread->td_rw_rlocks++;
726 return (1);
727 }
728 }
729
730 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
731 return (0);
732 }
733
734 int
__rw_try_rlock(volatile uintptr_t * c,const char * file,int line)735 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
736 {
737 struct rwlock *rw;
738
739 rw = rwlock2rw(c);
740 return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
741 }
742
743 static bool __always_inline
__rw_runlock_try(struct rwlock * rw,struct thread * td,uintptr_t * vp)744 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
745 {
746
747 for (;;) {
748 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
749 if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
750 *vp - RW_ONE_READER)) {
751 if (LOCK_LOG_TEST(&rw->lock_object, 0))
752 CTR4(KTR_LOCK,
753 "%s: %p succeeded %p -> %p",
754 __func__, rw, (void *)*vp,
755 (void *)(*vp - RW_ONE_READER));
756 td->td_rw_rlocks--;
757 return (true);
758 }
759 continue;
760 }
761 break;
762 }
763 return (false);
764 }
765
766 static void __noinline
__rw_runlock_hard(struct rwlock * rw,struct thread * td,uintptr_t v LOCK_FILE_LINE_ARG_DEF)767 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
768 LOCK_FILE_LINE_ARG_DEF)
769 {
770 struct turnstile *ts;
771 uintptr_t setv, queue;
772
773 if (SCHEDULER_STOPPED())
774 return;
775
776 if (__rw_runlock_try(rw, td, &v))
777 goto out_lockstat;
778
779 /*
780 * Ok, we know we have waiters and we think we are the
781 * last reader, so grab the turnstile lock.
782 */
783 turnstile_chain_lock(&rw->lock_object);
784 v = RW_READ_VALUE(rw);
785 for (;;) {
786 if (__rw_runlock_try(rw, td, &v))
787 break;
788
789 MPASS(v & RW_LOCK_WAITERS);
790
791 /*
792 * Try to drop our lock leaving the lock in a unlocked
793 * state.
794 *
795 * If you wanted to do explicit lock handoff you'd have to
796 * do it here. You'd also want to use turnstile_signal()
797 * and you'd have to handle the race where a higher
798 * priority thread blocks on the write lock before the
799 * thread you wakeup actually runs and have the new thread
800 * "steal" the lock. For now it's a lot simpler to just
801 * wakeup all of the waiters.
802 *
803 * As above, if we fail, then another thread might have
804 * acquired a read lock, so drop the turnstile lock and
805 * restart.
806 */
807 setv = RW_UNLOCKED;
808 queue = TS_SHARED_QUEUE;
809 if (v & RW_LOCK_WRITE_WAITERS) {
810 queue = TS_EXCLUSIVE_QUEUE;
811 setv |= (v & RW_LOCK_READ_WAITERS);
812 }
813 setv |= (v & RW_LOCK_WRITE_SPINNER);
814 if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
815 continue;
816 if (LOCK_LOG_TEST(&rw->lock_object, 0))
817 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
818 __func__, rw);
819
820 /*
821 * Ok. The lock is released and all that's left is to
822 * wake up the waiters. Note that the lock might not be
823 * free anymore, but in that case the writers will just
824 * block again if they run before the new lock holder(s)
825 * release the lock.
826 */
827 ts = turnstile_lookup(&rw->lock_object);
828 MPASS(ts != NULL);
829 turnstile_broadcast(ts, queue);
830 turnstile_unpend(ts);
831 td->td_rw_rlocks--;
832 break;
833 }
834 turnstile_chain_unlock(&rw->lock_object);
835 out_lockstat:
836 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
837 }
838
839 void
_rw_runlock_cookie_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)840 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
841 {
842 struct thread *td;
843 uintptr_t v;
844
845 KASSERT(rw->rw_lock != RW_DESTROYED,
846 ("rw_runlock() of destroyed rwlock %p @ %s:%d", rw, file, line));
847 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
848 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
849 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
850
851 td = curthread;
852 v = RW_READ_VALUE(rw);
853
854 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
855 !__rw_runlock_try(rw, td, &v)))
856 __rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
857 else
858 lock_profile_release_lock(&rw->lock_object, false);
859
860 TD_LOCKS_DEC(curthread);
861 }
862
863 void
_rw_runlock_cookie(volatile uintptr_t * c,const char * file,int line)864 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
865 {
866 struct rwlock *rw;
867
868 rw = rwlock2rw(c);
869 _rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
870 }
871
872 #ifdef ADAPTIVE_RWLOCKS
873 static inline void
rw_drop_critical(uintptr_t v,bool * in_critical,int * extra_work)874 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
875 {
876
877 if (v & RW_LOCK_WRITE_SPINNER)
878 return;
879 if (*in_critical) {
880 critical_exit();
881 *in_critical = false;
882 (*extra_work)--;
883 }
884 }
885 #else
886 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
887 #endif
888
889 /*
890 * This function is called when we are unable to obtain a write lock on the
891 * first try. This means that at least one other thread holds either a
892 * read or write lock.
893 */
894 void
__rw_wlock_hard(volatile uintptr_t * c,uintptr_t v LOCK_FILE_LINE_ARG_DEF)895 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
896 {
897 uintptr_t tid;
898 struct rwlock *rw;
899 struct turnstile *ts;
900 struct thread *owner;
901 #ifdef ADAPTIVE_RWLOCKS
902 int spintries = 0;
903 int i, n;
904 enum { READERS, WRITER } sleep_reason = READERS;
905 bool in_critical = false;
906 #endif
907 uintptr_t setv;
908 #ifdef LOCK_PROFILING
909 uint64_t waittime = 0;
910 int contested = 0;
911 #endif
912 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
913 struct lock_delay_arg lda;
914 #endif
915 #ifdef KDTRACE_HOOKS
916 u_int sleep_cnt = 0;
917 int64_t sleep_time = 0;
918 int64_t all_time = 0;
919 #endif
920 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
921 uintptr_t state = 0;
922 int doing_lockprof = 0;
923 #endif
924 int extra_work = 0;
925
926 tid = (uintptr_t)curthread;
927 rw = rwlock2rw(c);
928
929 #ifdef KDTRACE_HOOKS
930 if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
931 while (v == RW_UNLOCKED) {
932 if (_rw_write_lock_fetch(rw, &v, tid))
933 goto out_lockstat;
934 }
935 extra_work = 1;
936 doing_lockprof = 1;
937 all_time -= lockstat_nsecs(&rw->lock_object);
938 state = v;
939 }
940 #endif
941 #ifdef LOCK_PROFILING
942 extra_work = 1;
943 doing_lockprof = 1;
944 state = v;
945 #endif
946
947 if (SCHEDULER_STOPPED())
948 return;
949
950 if (__predict_false(v == RW_UNLOCKED))
951 v = RW_READ_VALUE(rw);
952
953 if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
954 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
955 ("%s: recursing but non-recursive rw %p @ %s:%d\n",
956 __func__, rw, file, line));
957 rw->rw_recurse++;
958 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
959 if (LOCK_LOG_TEST(&rw->lock_object, 0))
960 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
961 return;
962 }
963
964 if (LOCK_LOG_TEST(&rw->lock_object, 0))
965 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
966 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
967
968 #if defined(ADAPTIVE_RWLOCKS)
969 lock_delay_arg_init(&lda, &rw_delay);
970 #elif defined(KDTRACE_HOOKS)
971 lock_delay_arg_init_noadapt(&lda);
972 #endif
973
974 #ifdef HWPMC_HOOKS
975 PMC_SOFT_CALL( , , lock, failed);
976 #endif
977 lock_profile_obtain_lock_failed(&rw->lock_object, false,
978 &contested, &waittime);
979
980 for (;;) {
981 if (v == RW_UNLOCKED) {
982 if (_rw_write_lock_fetch(rw, &v, tid))
983 break;
984 continue;
985 }
986 #ifdef KDTRACE_HOOKS
987 lda.spin_cnt++;
988 #endif
989
990 #ifdef ADAPTIVE_RWLOCKS
991 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
992 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
993 break;
994 continue;
995 }
996
997 /*
998 * If the lock is write locked and the owner is
999 * running on another CPU, spin until the owner stops
1000 * running or the state of the lock changes.
1001 */
1002 if (!(v & RW_LOCK_READ)) {
1003 rw_drop_critical(v, &in_critical, &extra_work);
1004 sleep_reason = WRITER;
1005 owner = lv_rw_wowner(v);
1006 if (!TD_IS_RUNNING(owner))
1007 goto ts;
1008 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1009 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1010 __func__, rw, owner);
1011 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1012 "spinning", "lockname:\"%s\"",
1013 rw->lock_object.lo_name);
1014 do {
1015 lock_delay(&lda);
1016 v = RW_READ_VALUE(rw);
1017 owner = lv_rw_wowner(v);
1018 } while (owner != NULL && TD_IS_RUNNING(owner));
1019 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1020 "running");
1021 continue;
1022 } else if (RW_READERS(v) > 0) {
1023 sleep_reason = READERS;
1024 if (spintries == rowner_retries)
1025 goto ts;
1026 if (!(v & RW_LOCK_WRITE_SPINNER)) {
1027 if (!in_critical) {
1028 critical_enter();
1029 in_critical = true;
1030 extra_work++;
1031 }
1032 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1033 v | RW_LOCK_WRITE_SPINNER)) {
1034 critical_exit();
1035 in_critical = false;
1036 extra_work--;
1037 continue;
1038 }
1039 }
1040 spintries++;
1041 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1042 "spinning", "lockname:\"%s\"",
1043 rw->lock_object.lo_name);
1044 n = RW_READERS(v);
1045 for (i = 0; i < rowner_loops; i += n) {
1046 lock_delay_spin(n);
1047 v = RW_READ_VALUE(rw);
1048 if (!(v & RW_LOCK_WRITE_SPINNER))
1049 break;
1050 if (!(v & RW_LOCK_READ))
1051 break;
1052 n = RW_READERS(v);
1053 if (n == 0)
1054 break;
1055 }
1056 #ifdef KDTRACE_HOOKS
1057 lda.spin_cnt += i;
1058 #endif
1059 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1060 "running");
1061 if (i < rowner_loops)
1062 continue;
1063 }
1064 ts:
1065 #endif
1066 ts = turnstile_trywait(&rw->lock_object);
1067 v = RW_READ_VALUE(rw);
1068 retry_ts:
1069 owner = lv_rw_wowner(v);
1070
1071 #ifdef ADAPTIVE_RWLOCKS
1072 /*
1073 * The current lock owner might have started executing
1074 * on another CPU (or the lock could have changed
1075 * owners) while we were waiting on the turnstile
1076 * chain lock. If so, drop the turnstile lock and try
1077 * again.
1078 */
1079 if (owner != NULL) {
1080 if (TD_IS_RUNNING(owner)) {
1081 turnstile_cancel(ts);
1082 rw_drop_critical(v, &in_critical, &extra_work);
1083 continue;
1084 }
1085 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1086 turnstile_cancel(ts);
1087 rw_drop_critical(v, &in_critical, &extra_work);
1088 continue;
1089 }
1090 #endif
1091 /*
1092 * Check for the waiters flags about this rwlock.
1093 * If the lock was released, without maintain any pending
1094 * waiters queue, simply try to acquire it.
1095 * If a pending waiters queue is present, claim the lock
1096 * ownership and maintain the pending queue.
1097 */
1098 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1099 if ((v & ~setv) == RW_UNLOCKED) {
1100 setv &= ~RW_LOCK_WRITE_SPINNER;
1101 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1102 if (setv)
1103 turnstile_claim(ts);
1104 else
1105 turnstile_cancel(ts);
1106 break;
1107 }
1108 goto retry_ts;
1109 }
1110
1111 #ifdef ADAPTIVE_RWLOCKS
1112 if (in_critical) {
1113 if ((v & RW_LOCK_WRITE_SPINNER) ||
1114 !((v & RW_LOCK_WRITE_WAITERS))) {
1115 setv = v & ~RW_LOCK_WRITE_SPINNER;
1116 setv |= RW_LOCK_WRITE_WAITERS;
1117 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1118 goto retry_ts;
1119 }
1120 critical_exit();
1121 in_critical = false;
1122 extra_work--;
1123 } else {
1124 #endif
1125 /*
1126 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1127 * set it. If we fail to set it, then loop back and try
1128 * again.
1129 */
1130 if (!(v & RW_LOCK_WRITE_WAITERS)) {
1131 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1132 v | RW_LOCK_WRITE_WAITERS))
1133 goto retry_ts;
1134 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1135 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1136 __func__, rw);
1137 }
1138 #ifdef ADAPTIVE_RWLOCKS
1139 }
1140 #endif
1141 /*
1142 * We were unable to acquire the lock and the write waiters
1143 * flag is set, so we must block on the turnstile.
1144 */
1145 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1146 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1147 rw);
1148 #ifdef KDTRACE_HOOKS
1149 sleep_time -= lockstat_nsecs(&rw->lock_object);
1150 #endif
1151 MPASS(owner == rw_owner(rw));
1152 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1153 #ifdef KDTRACE_HOOKS
1154 sleep_time += lockstat_nsecs(&rw->lock_object);
1155 sleep_cnt++;
1156 #endif
1157 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1158 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1159 __func__, rw);
1160 #ifdef ADAPTIVE_RWLOCKS
1161 spintries = 0;
1162 #endif
1163 v = RW_READ_VALUE(rw);
1164 }
1165 if (__predict_true(!extra_work))
1166 return;
1167 #ifdef ADAPTIVE_RWLOCKS
1168 if (in_critical)
1169 critical_exit();
1170 #endif
1171 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1172 if (__predict_true(!doing_lockprof))
1173 return;
1174 #endif
1175 #ifdef KDTRACE_HOOKS
1176 all_time += lockstat_nsecs(&rw->lock_object);
1177 if (sleep_time)
1178 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1179 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1180 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1181
1182 /* Record only the loops spinning and not sleeping. */
1183 if (lda.spin_cnt > sleep_cnt)
1184 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1185 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1186 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1187 out_lockstat:
1188 #endif
1189 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1190 waittime, file, line, LOCKSTAT_WRITER);
1191 }
1192
1193 /*
1194 * This function is called if lockstat is active or the first try at releasing
1195 * a write lock failed. The latter means that the lock is recursed or one of
1196 * the 2 waiter bits must be set indicating that at least one thread is waiting
1197 * on this lock.
1198 */
1199 void
__rw_wunlock_hard(volatile uintptr_t * c,uintptr_t v LOCK_FILE_LINE_ARG_DEF)1200 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1201 {
1202 struct rwlock *rw;
1203 struct turnstile *ts;
1204 uintptr_t tid, setv;
1205 int queue;
1206
1207 tid = (uintptr_t)curthread;
1208 if (SCHEDULER_STOPPED())
1209 return;
1210
1211 rw = rwlock2rw(c);
1212 if (__predict_false(v == tid))
1213 v = RW_READ_VALUE(rw);
1214
1215 if (v & RW_LOCK_WRITER_RECURSED) {
1216 if (--(rw->rw_recurse) == 0)
1217 atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1218 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1219 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1220 return;
1221 }
1222
1223 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1224 if (v == tid && _rw_write_unlock(rw, tid))
1225 return;
1226
1227 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1228 ("%s: neither of the waiter flags are set", __func__));
1229
1230 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1231 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1232
1233 turnstile_chain_lock(&rw->lock_object);
1234
1235 /*
1236 * Use the same algo as sx locks for now. Prefer waking up shared
1237 * waiters if we have any over writers. This is probably not ideal.
1238 *
1239 * 'v' is the value we are going to write back to rw_lock. If we
1240 * have waiters on both queues, we need to preserve the state of
1241 * the waiter flag for the queue we don't wake up. For now this is
1242 * hardcoded for the algorithm mentioned above.
1243 *
1244 * In the case of both readers and writers waiting we wakeup the
1245 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
1246 * new writer comes in before a reader it will claim the lock up
1247 * above. There is probably a potential priority inversion in
1248 * there that could be worked around either by waking both queues
1249 * of waiters or doing some complicated lock handoff gymnastics.
1250 */
1251 setv = RW_UNLOCKED;
1252 v = RW_READ_VALUE(rw);
1253 queue = TS_SHARED_QUEUE;
1254 if (v & RW_LOCK_WRITE_WAITERS) {
1255 queue = TS_EXCLUSIVE_QUEUE;
1256 setv |= (v & RW_LOCK_READ_WAITERS);
1257 }
1258 atomic_store_rel_ptr(&rw->rw_lock, setv);
1259
1260 /* Wake up all waiters for the specific queue. */
1261 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1262 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1263 queue == TS_SHARED_QUEUE ? "read" : "write");
1264
1265 ts = turnstile_lookup(&rw->lock_object);
1266 MPASS(ts != NULL);
1267 turnstile_broadcast(ts, queue);
1268 turnstile_unpend(ts);
1269 turnstile_chain_unlock(&rw->lock_object);
1270 }
1271
1272 /*
1273 * Attempt to do a non-blocking upgrade from a read lock to a write
1274 * lock. This will only succeed if this thread holds a single read
1275 * lock. Returns true if the upgrade succeeded and false otherwise.
1276 */
1277 int
__rw_try_upgrade_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)1278 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1279 {
1280 uintptr_t v, setv, tid;
1281 struct turnstile *ts;
1282 int success;
1283
1284 if (SCHEDULER_STOPPED())
1285 return (1);
1286
1287 KASSERT(rw->rw_lock != RW_DESTROYED,
1288 ("rw_try_upgrade() of destroyed rwlock %p @ %s:%d", rw, file,
1289 line));
1290 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1291
1292 /*
1293 * Attempt to switch from one reader to a writer. If there
1294 * are any write waiters, then we will have to lock the
1295 * turnstile first to prevent races with another writer
1296 * calling turnstile_wait() before we have claimed this
1297 * turnstile. So, do the simple case of no waiters first.
1298 */
1299 tid = (uintptr_t)curthread;
1300 success = 0;
1301 v = RW_READ_VALUE(rw);
1302 for (;;) {
1303 if (RW_READERS(v) > 1)
1304 break;
1305 if (!(v & RW_LOCK_WAITERS)) {
1306 success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1307 if (!success)
1308 continue;
1309 break;
1310 }
1311
1312 /*
1313 * Ok, we think we have waiters, so lock the turnstile.
1314 */
1315 ts = turnstile_trywait(&rw->lock_object);
1316 v = RW_READ_VALUE(rw);
1317 retry_ts:
1318 if (RW_READERS(v) > 1) {
1319 turnstile_cancel(ts);
1320 break;
1321 }
1322 /*
1323 * Try to switch from one reader to a writer again. This time
1324 * we honor the current state of the waiters flags.
1325 * If we obtain the lock with the flags set, then claim
1326 * ownership of the turnstile.
1327 */
1328 setv = tid | (v & RW_LOCK_WAITERS);
1329 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1330 if (success) {
1331 if (v & RW_LOCK_WAITERS)
1332 turnstile_claim(ts);
1333 else
1334 turnstile_cancel(ts);
1335 break;
1336 }
1337 goto retry_ts;
1338 }
1339 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1340 if (success) {
1341 curthread->td_rw_rlocks--;
1342 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1343 file, line);
1344 LOCKSTAT_RECORD0(rw__upgrade, rw);
1345 }
1346 return (success);
1347 }
1348
1349 int
__rw_try_upgrade(volatile uintptr_t * c,const char * file,int line)1350 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1351 {
1352 struct rwlock *rw;
1353
1354 rw = rwlock2rw(c);
1355 return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1356 }
1357
1358 /*
1359 * Downgrade a write lock into a single read lock.
1360 */
1361 void
__rw_downgrade_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)1362 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1363 {
1364 struct turnstile *ts;
1365 uintptr_t tid, v;
1366 int rwait, wwait;
1367
1368 if (SCHEDULER_STOPPED())
1369 return;
1370
1371 KASSERT(rw->rw_lock != RW_DESTROYED,
1372 ("rw_downgrade() of destroyed rwlock %p @ %s:%d", rw, file, line));
1373 __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1374 #ifndef INVARIANTS
1375 if (rw_recursed(rw))
1376 panic("downgrade of a recursed lock");
1377 #endif
1378
1379 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1380
1381 /*
1382 * Convert from a writer to a single reader. First we handle
1383 * the easy case with no waiters. If there are any waiters, we
1384 * lock the turnstile and "disown" the lock.
1385 */
1386 tid = (uintptr_t)curthread;
1387 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1388 goto out;
1389
1390 /*
1391 * Ok, we think we have waiters, so lock the turnstile so we can
1392 * read the waiter flags without any races.
1393 */
1394 turnstile_chain_lock(&rw->lock_object);
1395 v = rw->rw_lock & RW_LOCK_WAITERS;
1396 rwait = v & RW_LOCK_READ_WAITERS;
1397 wwait = v & RW_LOCK_WRITE_WAITERS;
1398 MPASS(rwait | wwait);
1399
1400 /*
1401 * Downgrade from a write lock while preserving waiters flag
1402 * and give up ownership of the turnstile.
1403 */
1404 ts = turnstile_lookup(&rw->lock_object);
1405 MPASS(ts != NULL);
1406 if (!wwait)
1407 v &= ~RW_LOCK_READ_WAITERS;
1408 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1409 /*
1410 * Wake other readers if there are no writers pending. Otherwise they
1411 * won't be able to acquire the lock anyway.
1412 */
1413 if (rwait && !wwait) {
1414 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1415 turnstile_unpend(ts);
1416 } else
1417 turnstile_disown(ts);
1418 turnstile_chain_unlock(&rw->lock_object);
1419 out:
1420 curthread->td_rw_rlocks++;
1421 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1422 LOCKSTAT_RECORD0(rw__downgrade, rw);
1423 }
1424
1425 void
__rw_downgrade(volatile uintptr_t * c,const char * file,int line)1426 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1427 {
1428 struct rwlock *rw;
1429
1430 rw = rwlock2rw(c);
1431 __rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1432 }
1433
1434 #ifdef INVARIANT_SUPPORT
1435 #ifndef INVARIANTS
1436 #undef __rw_assert
1437 #endif
1438
1439 /*
1440 * In the non-WITNESS case, rw_assert() can only detect that at least
1441 * *some* thread owns an rlock, but it cannot guarantee that *this*
1442 * thread owns an rlock.
1443 */
1444 void
__rw_assert(const volatile uintptr_t * c,int what,const char * file,int line)1445 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1446 {
1447 const struct rwlock *rw;
1448
1449 if (SCHEDULER_STOPPED())
1450 return;
1451
1452 rw = rwlock2rw(c);
1453
1454 switch (what) {
1455 case RA_LOCKED:
1456 case RA_LOCKED | RA_RECURSED:
1457 case RA_LOCKED | RA_NOTRECURSED:
1458 case RA_RLOCKED:
1459 case RA_RLOCKED | RA_RECURSED:
1460 case RA_RLOCKED | RA_NOTRECURSED:
1461 #ifdef WITNESS
1462 witness_assert(&rw->lock_object, what, file, line);
1463 #else
1464 /*
1465 * If some other thread has a write lock or we have one
1466 * and are asserting a read lock, fail. Also, if no one
1467 * has a lock at all, fail.
1468 */
1469 if (rw->rw_lock == RW_UNLOCKED ||
1470 (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1471 rw_wowner(rw) != curthread)))
1472 panic("Lock %s not %slocked @ %s:%d\n",
1473 rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1474 "read " : "", file, line);
1475
1476 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1477 if (rw_recursed(rw)) {
1478 if (what & RA_NOTRECURSED)
1479 panic("Lock %s recursed @ %s:%d\n",
1480 rw->lock_object.lo_name, file,
1481 line);
1482 } else if (what & RA_RECURSED)
1483 panic("Lock %s not recursed @ %s:%d\n",
1484 rw->lock_object.lo_name, file, line);
1485 }
1486 #endif
1487 break;
1488 case RA_WLOCKED:
1489 case RA_WLOCKED | RA_RECURSED:
1490 case RA_WLOCKED | RA_NOTRECURSED:
1491 if (rw_wowner(rw) != curthread)
1492 panic("Lock %s not exclusively locked @ %s:%d\n",
1493 rw->lock_object.lo_name, file, line);
1494 if (rw_recursed(rw)) {
1495 if (what & RA_NOTRECURSED)
1496 panic("Lock %s recursed @ %s:%d\n",
1497 rw->lock_object.lo_name, file, line);
1498 } else if (what & RA_RECURSED)
1499 panic("Lock %s not recursed @ %s:%d\n",
1500 rw->lock_object.lo_name, file, line);
1501 break;
1502 case RA_UNLOCKED:
1503 #ifdef WITNESS
1504 witness_assert(&rw->lock_object, what, file, line);
1505 #else
1506 /*
1507 * If we hold a write lock fail. We can't reliably check
1508 * to see if we hold a read lock or not.
1509 */
1510 if (rw_wowner(rw) == curthread)
1511 panic("Lock %s exclusively locked @ %s:%d\n",
1512 rw->lock_object.lo_name, file, line);
1513 #endif
1514 break;
1515 default:
1516 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1517 line);
1518 }
1519 }
1520 #endif /* INVARIANT_SUPPORT */
1521
1522 #ifdef DDB
1523 void
db_show_rwlock(const struct lock_object * lock)1524 db_show_rwlock(const struct lock_object *lock)
1525 {
1526 const struct rwlock *rw;
1527 struct thread *td;
1528
1529 rw = (const struct rwlock *)lock;
1530
1531 db_printf(" state: ");
1532 if (rw->rw_lock == RW_UNLOCKED)
1533 db_printf("UNLOCKED\n");
1534 else if (rw->rw_lock == RW_DESTROYED) {
1535 db_printf("DESTROYED\n");
1536 return;
1537 } else if (rw->rw_lock & RW_LOCK_READ)
1538 db_printf("RLOCK: %ju locks\n",
1539 (uintmax_t)(RW_READERS(rw->rw_lock)));
1540 else {
1541 td = rw_wowner(rw);
1542 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1543 td->td_tid, td->td_proc->p_pid, td->td_name);
1544 if (rw_recursed(rw))
1545 db_printf(" recursed: %u\n", rw->rw_recurse);
1546 }
1547 db_printf(" waiters: ");
1548 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1549 case RW_LOCK_READ_WAITERS:
1550 db_printf("readers\n");
1551 break;
1552 case RW_LOCK_WRITE_WAITERS:
1553 db_printf("writers\n");
1554 break;
1555 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1556 db_printf("readers and writers\n");
1557 break;
1558 default:
1559 db_printf("none\n");
1560 break;
1561 }
1562 }
1563
1564 #endif
1565