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 static 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 static 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 static 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 static 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 %s @ %s:%d",
282 curthread, rw->lock_object.lo_name, file, line));
283 KASSERT(rw->rw_lock != RW_DESTROYED,
284 ("rw_wlock() of destroyed rwlock @ %s:%d", 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())
311 return (1);
312
313 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
314 ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
315 curthread, rw->lock_object.lo_name, file, line));
316 KASSERT(rw->rw_lock != RW_DESTROYED,
317 ("rw_try_wlock() of destroyed rwlock @ %s:%d", 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 @ %s:%d", 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 __always_inline bool
__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 __always_inline bool
__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 THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
486
487 for (;;) {
488 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
489 break;
490 #ifdef KDTRACE_HOOKS
491 lda.spin_cnt++;
492 #endif
493
494 #ifdef ADAPTIVE_RWLOCKS
495 /*
496 * If the owner is running on another CPU, spin until
497 * the owner stops running or the state of the lock
498 * changes.
499 */
500 if ((v & RW_LOCK_READ) == 0) {
501 owner = (struct thread *)RW_OWNER(v);
502 if (TD_IS_RUNNING(owner)) {
503 if (LOCK_LOG_TEST(&rw->lock_object, 0))
504 CTR3(KTR_LOCK,
505 "%s: spinning on %p held by %p",
506 __func__, rw, owner);
507 KTR_STATE1(KTR_SCHED, "thread",
508 sched_tdname(curthread), "spinning",
509 "lockname:\"%s\"", rw->lock_object.lo_name);
510 do {
511 lock_delay(&lda);
512 v = RW_READ_VALUE(rw);
513 owner = lv_rw_wowner(v);
514 } while (owner != NULL && TD_IS_RUNNING(owner));
515 KTR_STATE0(KTR_SCHED, "thread",
516 sched_tdname(curthread), "running");
517 continue;
518 }
519 } else {
520 if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
521 MPASS(!__rw_can_read(td, v, false));
522 lock_delay_spin(2);
523 v = RW_READ_VALUE(rw);
524 continue;
525 }
526 if (spintries < rowner_retries) {
527 spintries++;
528 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
529 "spinning", "lockname:\"%s\"",
530 rw->lock_object.lo_name);
531 n = RW_READERS(v);
532 for (i = 0; i < rowner_loops; i += n) {
533 lock_delay_spin(n);
534 v = RW_READ_VALUE(rw);
535 if (!(v & RW_LOCK_READ))
536 break;
537 n = RW_READERS(v);
538 if (n == 0)
539 break;
540 if (__rw_can_read(td, v, false))
541 break;
542 }
543 #ifdef KDTRACE_HOOKS
544 lda.spin_cnt += rowner_loops - i;
545 #endif
546 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
547 "running");
548 if (i < rowner_loops)
549 continue;
550 }
551 }
552 #endif
553
554 /*
555 * Okay, now it's the hard case. Some other thread already
556 * has a write lock or there are write waiters present,
557 * acquire the turnstile lock so we can begin the process
558 * of blocking.
559 */
560 ts = turnstile_trywait(&rw->lock_object);
561
562 /*
563 * The lock might have been released while we spun, so
564 * recheck its state and restart the loop if needed.
565 */
566 v = RW_READ_VALUE(rw);
567 retry_ts:
568 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
569 __rw_can_read(td, v, false)) {
570 turnstile_cancel(ts);
571 continue;
572 }
573
574 owner = lv_rw_wowner(v);
575
576 #ifdef ADAPTIVE_RWLOCKS
577 /*
578 * The current lock owner might have started executing
579 * on another CPU (or the lock could have changed
580 * owners) while we were waiting on the turnstile
581 * chain lock. If so, drop the turnstile lock and try
582 * again.
583 */
584 if (owner != NULL) {
585 if (TD_IS_RUNNING(owner)) {
586 turnstile_cancel(ts);
587 continue;
588 }
589 }
590 #endif
591
592 /*
593 * The lock is held in write mode or it already has waiters.
594 */
595 MPASS(!__rw_can_read(td, v, false));
596
597 /*
598 * If the RW_LOCK_READ_WAITERS flag is already set, then
599 * we can go ahead and block. If it is not set then try
600 * to set it. If we fail to set it drop the turnstile
601 * lock and restart the loop.
602 */
603 if (!(v & RW_LOCK_READ_WAITERS)) {
604 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
605 v | RW_LOCK_READ_WAITERS))
606 goto retry_ts;
607 if (LOCK_LOG_TEST(&rw->lock_object, 0))
608 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
609 __func__, rw);
610 }
611
612 /*
613 * We were unable to acquire the lock and the read waiters
614 * flag is set, so we must block on the turnstile.
615 */
616 if (LOCK_LOG_TEST(&rw->lock_object, 0))
617 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
618 rw);
619 #ifdef KDTRACE_HOOKS
620 sleep_time -= lockstat_nsecs(&rw->lock_object);
621 #endif
622 MPASS(owner == rw_owner(rw));
623 turnstile_wait(ts, owner, TS_SHARED_QUEUE);
624 #ifdef KDTRACE_HOOKS
625 sleep_time += lockstat_nsecs(&rw->lock_object);
626 sleep_cnt++;
627 #endif
628 if (LOCK_LOG_TEST(&rw->lock_object, 0))
629 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
630 __func__, rw);
631 v = RW_READ_VALUE(rw);
632 }
633 THREAD_CONTENTION_DONE(&rw->lock_object);
634 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
635 if (__predict_true(!doing_lockprof))
636 return;
637 #endif
638 #ifdef KDTRACE_HOOKS
639 all_time += lockstat_nsecs(&rw->lock_object);
640 if (sleep_time)
641 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
642 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
643 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
644
645 /* Record only the loops spinning and not sleeping. */
646 if (lda.spin_cnt > sleep_cnt)
647 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
648 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
649 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
650 out_lockstat:
651 #endif
652 /*
653 * TODO: acquire "owner of record" here. Here be turnstile dragons
654 * however. turnstiles don't like owners changing between calls to
655 * turnstile_wait() currently.
656 */
657 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
658 waittime, file, line, LOCKSTAT_READER);
659 }
660
661 void
__rw_rlock_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)662 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
663 {
664 struct thread *td;
665 uintptr_t v;
666
667 td = curthread;
668
669 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
670 !TD_IS_IDLETHREAD(td),
671 ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
672 td, rw->lock_object.lo_name, file, line));
673 KASSERT(rw->rw_lock != RW_DESTROYED,
674 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
675 KASSERT(rw_wowner(rw) != td,
676 ("rw_rlock: wlock already held for %s @ %s:%d",
677 rw->lock_object.lo_name, file, line));
678 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
679
680 v = RW_READ_VALUE(rw);
681 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
682 !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
683 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
684 else
685 lock_profile_obtain_lock_success(&rw->lock_object, false, 0, 0,
686 file, line);
687
688 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
689 WITNESS_LOCK(&rw->lock_object, 0, file, line);
690 TD_LOCKS_INC(curthread);
691 }
692
693 void
__rw_rlock(volatile uintptr_t * c,const char * file,int line)694 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
695 {
696 struct rwlock *rw;
697
698 rw = rwlock2rw(c);
699 __rw_rlock_int(rw LOCK_FILE_LINE_ARG);
700 }
701
702 int
__rw_try_rlock_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)703 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
704 {
705 uintptr_t x;
706
707 if (SCHEDULER_STOPPED())
708 return (1);
709
710 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
711 ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
712 curthread, rw->lock_object.lo_name, file, line));
713
714 x = rw->rw_lock;
715 for (;;) {
716 KASSERT(rw->rw_lock != RW_DESTROYED,
717 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
718 if (!(x & RW_LOCK_READ))
719 break;
720 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
721 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
722 line);
723 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
724 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
725 rw, 0, 0, file, line, LOCKSTAT_READER);
726 TD_LOCKS_INC(curthread);
727 curthread->td_rw_rlocks++;
728 return (1);
729 }
730 }
731
732 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
733 return (0);
734 }
735
736 int
__rw_try_rlock(volatile uintptr_t * c,const char * file,int line)737 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
738 {
739 struct rwlock *rw;
740
741 rw = rwlock2rw(c);
742 return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
743 }
744
745 static __always_inline bool
__rw_runlock_try(struct rwlock * rw,struct thread * td,uintptr_t * vp)746 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
747 {
748
749 for (;;) {
750 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
751 if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
752 *vp - RW_ONE_READER)) {
753 if (LOCK_LOG_TEST(&rw->lock_object, 0))
754 CTR4(KTR_LOCK,
755 "%s: %p succeeded %p -> %p",
756 __func__, rw, (void *)*vp,
757 (void *)(*vp - RW_ONE_READER));
758 td->td_rw_rlocks--;
759 return (true);
760 }
761 continue;
762 }
763 break;
764 }
765 return (false);
766 }
767
768 static void __noinline
__rw_runlock_hard(struct rwlock * rw,struct thread * td,uintptr_t v LOCK_FILE_LINE_ARG_DEF)769 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
770 LOCK_FILE_LINE_ARG_DEF)
771 {
772 struct turnstile *ts;
773 uintptr_t setv, passedv, queue;
774
775 if (SCHEDULER_STOPPED())
776 return;
777
778 passedv = v;
779 if (__rw_runlock_try(rw, td, &v))
780 goto out_lockstat;
781
782 /*
783 * Ok, we know we have waiters and we think we are the
784 * last reader, so grab the turnstile lock.
785 */
786 turnstile_chain_lock(&rw->lock_object);
787 v = RW_READ_VALUE(rw);
788 for (;;) {
789 if (__rw_runlock_try(rw, td, &v))
790 break;
791
792 MPASS(v & RW_LOCK_WAITERS);
793
794 /*
795 * Try to drop our lock leaving the lock in a unlocked
796 * state.
797 *
798 * If you wanted to do explicit lock handoff you'd have to
799 * do it here. You'd also want to use turnstile_signal()
800 * and you'd have to handle the race where a higher
801 * priority thread blocks on the write lock before the
802 * thread you wakeup actually runs and have the new thread
803 * "steal" the lock. For now it's a lot simpler to just
804 * wakeup all of the waiters.
805 *
806 * As above, if we fail, then another thread might have
807 * acquired a read lock, so drop the turnstile lock and
808 * restart.
809 */
810 setv = RW_UNLOCKED;
811 queue = TS_SHARED_QUEUE;
812 if (v & RW_LOCK_WRITE_WAITERS) {
813 queue = TS_EXCLUSIVE_QUEUE;
814 setv |= (v & RW_LOCK_READ_WAITERS);
815 }
816 setv |= (v & RW_LOCK_WRITE_SPINNER);
817 if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
818 continue;
819 if (LOCK_LOG_TEST(&rw->lock_object, 0))
820 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
821 __func__, rw);
822
823 /*
824 * Ok. The lock is released and all that's left is to
825 * wake up the waiters. Note that the lock might not be
826 * free anymore, but in that case the writers will just
827 * block again if they run before the new lock holder(s)
828 * release the lock.
829 */
830 ts = turnstile_lookup(&rw->lock_object);
831 if (__predict_false(ts == NULL)) {
832 panic("got NULL turnstile on rwlock %p passedv %zx v %zx",
833 rw, passedv, v);
834 }
835 turnstile_broadcast(ts, queue);
836 turnstile_unpend(ts);
837 td->td_rw_rlocks--;
838 break;
839 }
840 turnstile_chain_unlock(&rw->lock_object);
841 out_lockstat:
842 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
843 }
844
845 void
_rw_runlock_cookie_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)846 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
847 {
848 struct thread *td;
849 uintptr_t v;
850
851 KASSERT(rw->rw_lock != RW_DESTROYED,
852 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
853 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
854 WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
855 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
856
857 td = curthread;
858 v = RW_READ_VALUE(rw);
859
860 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
861 !__rw_runlock_try(rw, td, &v)))
862 __rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
863 else
864 lock_profile_release_lock(&rw->lock_object, false);
865
866 TD_LOCKS_DEC(curthread);
867 }
868
869 void
_rw_runlock_cookie(volatile uintptr_t * c,const char * file,int line)870 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
871 {
872 struct rwlock *rw;
873
874 rw = rwlock2rw(c);
875 _rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
876 }
877
878 #ifdef ADAPTIVE_RWLOCKS
879 static inline void
rw_drop_critical(uintptr_t v,bool * in_critical,int * extra_work)880 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
881 {
882
883 if (v & RW_LOCK_WRITE_SPINNER)
884 return;
885 if (*in_critical) {
886 critical_exit();
887 *in_critical = false;
888 (*extra_work)--;
889 }
890 }
891 #else
892 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
893 #endif
894
895 /*
896 * This function is called when we are unable to obtain a write lock on the
897 * first try. This means that at least one other thread holds either a
898 * read or write lock.
899 */
900 void
__rw_wlock_hard(volatile uintptr_t * c,uintptr_t v LOCK_FILE_LINE_ARG_DEF)901 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
902 {
903 uintptr_t tid;
904 struct rwlock *rw;
905 struct turnstile *ts;
906 struct thread *owner;
907 #ifdef ADAPTIVE_RWLOCKS
908 int spintries = 0;
909 int i, n;
910 enum { READERS, WRITER } sleep_reason = READERS;
911 bool in_critical = false;
912 #endif
913 uintptr_t setv;
914 #ifdef LOCK_PROFILING
915 uint64_t waittime = 0;
916 int contested = 0;
917 #endif
918 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
919 struct lock_delay_arg lda;
920 #endif
921 #ifdef KDTRACE_HOOKS
922 u_int sleep_cnt = 0;
923 int64_t sleep_time = 0;
924 int64_t all_time = 0;
925 #endif
926 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
927 uintptr_t state = 0;
928 int doing_lockprof = 0;
929 #endif
930 int extra_work = 0;
931
932 tid = (uintptr_t)curthread;
933 rw = rwlock2rw(c);
934
935 #ifdef KDTRACE_HOOKS
936 if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
937 while (v == RW_UNLOCKED) {
938 if (_rw_write_lock_fetch(rw, &v, tid))
939 goto out_lockstat;
940 }
941 extra_work = 1;
942 doing_lockprof = 1;
943 all_time -= lockstat_nsecs(&rw->lock_object);
944 state = v;
945 }
946 #endif
947 #ifdef LOCK_PROFILING
948 extra_work = 1;
949 doing_lockprof = 1;
950 state = v;
951 #endif
952
953 if (SCHEDULER_STOPPED())
954 return;
955
956 if (__predict_false(v == RW_UNLOCKED))
957 v = RW_READ_VALUE(rw);
958
959 if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
960 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
961 ("%s: recursing but non-recursive rw %s @ %s:%d\n",
962 __func__, rw->lock_object.lo_name, file, line));
963 rw->rw_recurse++;
964 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
965 if (LOCK_LOG_TEST(&rw->lock_object, 0))
966 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
967 return;
968 }
969
970 if (LOCK_LOG_TEST(&rw->lock_object, 0))
971 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
972 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
973
974 #if defined(ADAPTIVE_RWLOCKS)
975 lock_delay_arg_init(&lda, &rw_delay);
976 #elif defined(KDTRACE_HOOKS)
977 lock_delay_arg_init_noadapt(&lda);
978 #endif
979
980 #ifdef HWPMC_HOOKS
981 PMC_SOFT_CALL( , , lock, failed);
982 #endif
983 lock_profile_obtain_lock_failed(&rw->lock_object, false,
984 &contested, &waittime);
985
986 THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
987
988 for (;;) {
989 if (v == RW_UNLOCKED) {
990 if (_rw_write_lock_fetch(rw, &v, tid))
991 break;
992 continue;
993 }
994 #ifdef KDTRACE_HOOKS
995 lda.spin_cnt++;
996 #endif
997
998 #ifdef ADAPTIVE_RWLOCKS
999 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
1000 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
1001 break;
1002 continue;
1003 }
1004
1005 /*
1006 * If the lock is write locked and the owner is
1007 * running on another CPU, spin until the owner stops
1008 * running or the state of the lock changes.
1009 */
1010 if (!(v & RW_LOCK_READ)) {
1011 rw_drop_critical(v, &in_critical, &extra_work);
1012 sleep_reason = WRITER;
1013 owner = lv_rw_wowner(v);
1014 if (!TD_IS_RUNNING(owner))
1015 goto ts;
1016 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1017 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1018 __func__, rw, owner);
1019 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1020 "spinning", "lockname:\"%s\"",
1021 rw->lock_object.lo_name);
1022 do {
1023 lock_delay(&lda);
1024 v = RW_READ_VALUE(rw);
1025 owner = lv_rw_wowner(v);
1026 } while (owner != NULL && TD_IS_RUNNING(owner));
1027 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1028 "running");
1029 continue;
1030 } else if (RW_READERS(v) > 0) {
1031 sleep_reason = READERS;
1032 if (spintries == rowner_retries)
1033 goto ts;
1034 if (!(v & RW_LOCK_WRITE_SPINNER)) {
1035 if (!in_critical) {
1036 critical_enter();
1037 in_critical = true;
1038 extra_work++;
1039 }
1040 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1041 v | RW_LOCK_WRITE_SPINNER)) {
1042 critical_exit();
1043 in_critical = false;
1044 extra_work--;
1045 continue;
1046 }
1047 }
1048 spintries++;
1049 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1050 "spinning", "lockname:\"%s\"",
1051 rw->lock_object.lo_name);
1052 n = RW_READERS(v);
1053 for (i = 0; i < rowner_loops; i += n) {
1054 lock_delay_spin(n);
1055 v = RW_READ_VALUE(rw);
1056 if (!(v & RW_LOCK_WRITE_SPINNER))
1057 break;
1058 if (!(v & RW_LOCK_READ))
1059 break;
1060 n = RW_READERS(v);
1061 if (n == 0)
1062 break;
1063 }
1064 #ifdef KDTRACE_HOOKS
1065 lda.spin_cnt += i;
1066 #endif
1067 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1068 "running");
1069 if (i < rowner_loops)
1070 continue;
1071 }
1072 ts:
1073 #endif
1074 ts = turnstile_trywait(&rw->lock_object);
1075 v = RW_READ_VALUE(rw);
1076 retry_ts:
1077 owner = lv_rw_wowner(v);
1078
1079 #ifdef ADAPTIVE_RWLOCKS
1080 /*
1081 * The current lock owner might have started executing
1082 * on another CPU (or the lock could have changed
1083 * owners) while we were waiting on the turnstile
1084 * chain lock. If so, drop the turnstile lock and try
1085 * again.
1086 */
1087 if (owner != NULL) {
1088 if (TD_IS_RUNNING(owner)) {
1089 turnstile_cancel(ts);
1090 rw_drop_critical(v, &in_critical, &extra_work);
1091 continue;
1092 }
1093 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1094 turnstile_cancel(ts);
1095 rw_drop_critical(v, &in_critical, &extra_work);
1096 continue;
1097 }
1098 #endif
1099 /*
1100 * Check for the waiters flags about this rwlock.
1101 * If the lock was released, without maintain any pending
1102 * waiters queue, simply try to acquire it.
1103 * If a pending waiters queue is present, claim the lock
1104 * ownership and maintain the pending queue.
1105 */
1106 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1107 if ((v & ~setv) == RW_UNLOCKED) {
1108 setv &= ~RW_LOCK_WRITE_SPINNER;
1109 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1110 if (setv)
1111 turnstile_claim(ts);
1112 else
1113 turnstile_cancel(ts);
1114 break;
1115 }
1116 goto retry_ts;
1117 }
1118
1119 #ifdef ADAPTIVE_RWLOCKS
1120 if (in_critical) {
1121 if ((v & RW_LOCK_WRITE_SPINNER) ||
1122 !((v & RW_LOCK_WRITE_WAITERS))) {
1123 setv = v & ~RW_LOCK_WRITE_SPINNER;
1124 setv |= RW_LOCK_WRITE_WAITERS;
1125 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1126 goto retry_ts;
1127 }
1128 critical_exit();
1129 in_critical = false;
1130 extra_work--;
1131 } else {
1132 #endif
1133 /*
1134 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1135 * set it. If we fail to set it, then loop back and try
1136 * again.
1137 */
1138 if (!(v & RW_LOCK_WRITE_WAITERS)) {
1139 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1140 v | RW_LOCK_WRITE_WAITERS))
1141 goto retry_ts;
1142 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1143 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1144 __func__, rw);
1145 }
1146 #ifdef ADAPTIVE_RWLOCKS
1147 }
1148 #endif
1149 /*
1150 * We were unable to acquire the lock and the write waiters
1151 * flag is set, so we must block on the turnstile.
1152 */
1153 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1154 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1155 rw);
1156 #ifdef KDTRACE_HOOKS
1157 sleep_time -= lockstat_nsecs(&rw->lock_object);
1158 #endif
1159 MPASS(owner == rw_owner(rw));
1160 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1161 #ifdef KDTRACE_HOOKS
1162 sleep_time += lockstat_nsecs(&rw->lock_object);
1163 sleep_cnt++;
1164 #endif
1165 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1166 CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1167 __func__, rw);
1168 #ifdef ADAPTIVE_RWLOCKS
1169 spintries = 0;
1170 #endif
1171 v = RW_READ_VALUE(rw);
1172 }
1173 THREAD_CONTENTION_DONE(&rw->lock_object);
1174 if (__predict_true(!extra_work))
1175 return;
1176 #ifdef ADAPTIVE_RWLOCKS
1177 if (in_critical)
1178 critical_exit();
1179 #endif
1180 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1181 if (__predict_true(!doing_lockprof))
1182 return;
1183 #endif
1184 #ifdef KDTRACE_HOOKS
1185 all_time += lockstat_nsecs(&rw->lock_object);
1186 if (sleep_time)
1187 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1188 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1189 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1190
1191 /* Record only the loops spinning and not sleeping. */
1192 if (lda.spin_cnt > sleep_cnt)
1193 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1194 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1195 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1196 out_lockstat:
1197 #endif
1198 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1199 waittime, file, line, LOCKSTAT_WRITER);
1200 }
1201
1202 /*
1203 * This function is called if lockstat is active or the first try at releasing
1204 * a write lock failed. The latter means that the lock is recursed or one of
1205 * the 2 waiter bits must be set indicating that at least one thread is waiting
1206 * on this lock.
1207 */
1208 void
__rw_wunlock_hard(volatile uintptr_t * c,uintptr_t v LOCK_FILE_LINE_ARG_DEF)1209 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1210 {
1211 struct rwlock *rw;
1212 struct turnstile *ts;
1213 uintptr_t tid, setv, passedv;
1214 int queue;
1215
1216 tid = (uintptr_t)curthread;
1217 if (SCHEDULER_STOPPED())
1218 return;
1219
1220 rw = rwlock2rw(c);
1221 if (__predict_false(v == tid))
1222 v = RW_READ_VALUE(rw);
1223
1224 if (v & RW_LOCK_WRITER_RECURSED) {
1225 if (--(rw->rw_recurse) == 0)
1226 atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1227 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1228 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1229 return;
1230 }
1231
1232 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1233 if (v == tid && _rw_write_unlock(rw, tid))
1234 return;
1235
1236 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1237 ("%s: neither of the waiter flags are set", __func__));
1238
1239 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1240 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1241
1242 turnstile_chain_lock(&rw->lock_object);
1243
1244 /*
1245 * Use the same algo as sx locks for now. Prefer waking up shared
1246 * waiters if we have any over writers. This is probably not ideal.
1247 *
1248 * 'v' is the value we are going to write back to rw_lock. If we
1249 * have waiters on both queues, we need to preserve the state of
1250 * the waiter flag for the queue we don't wake up. For now this is
1251 * hardcoded for the algorithm mentioned above.
1252 *
1253 * In the case of both readers and writers waiting we wakeup the
1254 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
1255 * new writer comes in before a reader it will claim the lock up
1256 * above. There is probably a potential priority inversion in
1257 * there that could be worked around either by waking both queues
1258 * of waiters or doing some complicated lock handoff gymnastics.
1259 */
1260 setv = RW_UNLOCKED;
1261 passedv = v;
1262 v = RW_READ_VALUE(rw);
1263 queue = TS_SHARED_QUEUE;
1264 if (v & RW_LOCK_WRITE_WAITERS) {
1265 queue = TS_EXCLUSIVE_QUEUE;
1266 setv |= (v & RW_LOCK_READ_WAITERS);
1267 }
1268 atomic_store_rel_ptr(&rw->rw_lock, setv);
1269
1270 /* Wake up all waiters for the specific queue. */
1271 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1272 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1273 queue == TS_SHARED_QUEUE ? "read" : "write");
1274
1275 ts = turnstile_lookup(&rw->lock_object);
1276 if (__predict_false(ts == NULL)) {
1277 panic("got NULL turnstile on rwlock %p passedv %zx v %zx", rw,
1278 passedv, v);
1279 }
1280 turnstile_broadcast(ts, queue);
1281 turnstile_unpend(ts);
1282 turnstile_chain_unlock(&rw->lock_object);
1283 }
1284
1285 /*
1286 * Attempt to do a non-blocking upgrade from a read lock to a write
1287 * lock. This will only succeed if this thread holds a single read
1288 * lock. Returns true if the upgrade succeeded and false otherwise.
1289 */
1290 int
__rw_try_upgrade_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)1291 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1292 {
1293 uintptr_t v, setv, tid;
1294 struct turnstile *ts;
1295 int success;
1296
1297 if (SCHEDULER_STOPPED())
1298 return (1);
1299
1300 KASSERT(rw->rw_lock != RW_DESTROYED,
1301 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1302 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1303
1304 /*
1305 * Attempt to switch from one reader to a writer. If there
1306 * are any write waiters, then we will have to lock the
1307 * turnstile first to prevent races with another writer
1308 * calling turnstile_wait() before we have claimed this
1309 * turnstile. So, do the simple case of no waiters first.
1310 */
1311 tid = (uintptr_t)curthread;
1312 success = 0;
1313 v = RW_READ_VALUE(rw);
1314 for (;;) {
1315 if (RW_READERS(v) > 1)
1316 break;
1317 if (!(v & RW_LOCK_WAITERS)) {
1318 success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1319 if (!success)
1320 continue;
1321 break;
1322 }
1323
1324 /*
1325 * Ok, we think we have waiters, so lock the turnstile.
1326 */
1327 ts = turnstile_trywait(&rw->lock_object);
1328 v = RW_READ_VALUE(rw);
1329 retry_ts:
1330 if (RW_READERS(v) > 1) {
1331 turnstile_cancel(ts);
1332 break;
1333 }
1334 /*
1335 * Try to switch from one reader to a writer again. This time
1336 * we honor the current state of the waiters flags.
1337 * If we obtain the lock with the flags set, then claim
1338 * ownership of the turnstile.
1339 */
1340 setv = tid | (v & RW_LOCK_WAITERS);
1341 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1342 if (success) {
1343 if (v & RW_LOCK_WAITERS)
1344 turnstile_claim(ts);
1345 else
1346 turnstile_cancel(ts);
1347 break;
1348 }
1349 goto retry_ts;
1350 }
1351 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1352 if (success) {
1353 curthread->td_rw_rlocks--;
1354 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1355 file, line);
1356 LOCKSTAT_RECORD0(rw__upgrade, rw);
1357 }
1358 return (success);
1359 }
1360
1361 int
__rw_try_upgrade(volatile uintptr_t * c,const char * file,int line)1362 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1363 {
1364 struct rwlock *rw;
1365
1366 rw = rwlock2rw(c);
1367 return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1368 }
1369
1370 /*
1371 * Downgrade a write lock into a single read lock.
1372 */
1373 void
__rw_downgrade_int(struct rwlock * rw LOCK_FILE_LINE_ARG_DEF)1374 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1375 {
1376 struct turnstile *ts;
1377 uintptr_t tid, v;
1378 int rwait, wwait;
1379
1380 if (SCHEDULER_STOPPED())
1381 return;
1382
1383 KASSERT(rw->rw_lock != RW_DESTROYED,
1384 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1385 __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1386 #ifndef INVARIANTS
1387 if (rw_recursed(rw))
1388 panic("downgrade of a recursed lock");
1389 #endif
1390
1391 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1392
1393 /*
1394 * Convert from a writer to a single reader. First we handle
1395 * the easy case with no waiters. If there are any waiters, we
1396 * lock the turnstile and "disown" the lock.
1397 */
1398 tid = (uintptr_t)curthread;
1399 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1400 goto out;
1401
1402 /*
1403 * Ok, we think we have waiters, so lock the turnstile so we can
1404 * read the waiter flags without any races.
1405 */
1406 turnstile_chain_lock(&rw->lock_object);
1407 v = rw->rw_lock & RW_LOCK_WAITERS;
1408 rwait = v & RW_LOCK_READ_WAITERS;
1409 wwait = v & RW_LOCK_WRITE_WAITERS;
1410 MPASS(rwait | wwait);
1411
1412 /*
1413 * Downgrade from a write lock while preserving waiters flag
1414 * and give up ownership of the turnstile.
1415 */
1416 ts = turnstile_lookup(&rw->lock_object);
1417 MPASS(ts != NULL);
1418 if (!wwait)
1419 v &= ~RW_LOCK_READ_WAITERS;
1420 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1421 /*
1422 * Wake other readers if there are no writers pending. Otherwise they
1423 * won't be able to acquire the lock anyway.
1424 */
1425 if (rwait && !wwait) {
1426 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1427 turnstile_unpend(ts);
1428 } else
1429 turnstile_disown(ts);
1430 turnstile_chain_unlock(&rw->lock_object);
1431 out:
1432 curthread->td_rw_rlocks++;
1433 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1434 LOCKSTAT_RECORD0(rw__downgrade, rw);
1435 }
1436
1437 void
__rw_downgrade(volatile uintptr_t * c,const char * file,int line)1438 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1439 {
1440 struct rwlock *rw;
1441
1442 rw = rwlock2rw(c);
1443 __rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1444 }
1445
1446 #ifdef INVARIANT_SUPPORT
1447 #ifndef INVARIANTS
1448 #undef __rw_assert
1449 #endif
1450
1451 /*
1452 * In the non-WITNESS case, rw_assert() can only detect that at least
1453 * *some* thread owns an rlock, but it cannot guarantee that *this*
1454 * thread owns an rlock.
1455 */
1456 void
__rw_assert(const volatile uintptr_t * c,int what,const char * file,int line)1457 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1458 {
1459 const struct rwlock *rw;
1460
1461 if (SCHEDULER_STOPPED())
1462 return;
1463
1464 rw = rwlock2rw(c);
1465
1466 switch (what) {
1467 case RA_LOCKED:
1468 case RA_LOCKED | RA_RECURSED:
1469 case RA_LOCKED | RA_NOTRECURSED:
1470 case RA_RLOCKED:
1471 case RA_RLOCKED | RA_RECURSED:
1472 case RA_RLOCKED | RA_NOTRECURSED:
1473 #ifdef WITNESS
1474 witness_assert(&rw->lock_object, what, file, line);
1475 #else
1476 /*
1477 * If some other thread has a write lock or we have one
1478 * and are asserting a read lock, fail. Also, if no one
1479 * has a lock at all, fail.
1480 */
1481 if (rw->rw_lock == RW_UNLOCKED ||
1482 (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1483 rw_wowner(rw) != curthread)))
1484 panic("Lock %s not %slocked @ %s:%d\n",
1485 rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1486 "read " : "", file, line);
1487
1488 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1489 if (rw_recursed(rw)) {
1490 if (what & RA_NOTRECURSED)
1491 panic("Lock %s recursed @ %s:%d\n",
1492 rw->lock_object.lo_name, file,
1493 line);
1494 } else if (what & RA_RECURSED)
1495 panic("Lock %s not recursed @ %s:%d\n",
1496 rw->lock_object.lo_name, file, line);
1497 }
1498 #endif
1499 break;
1500 case RA_WLOCKED:
1501 case RA_WLOCKED | RA_RECURSED:
1502 case RA_WLOCKED | RA_NOTRECURSED:
1503 if (rw_wowner(rw) != curthread)
1504 panic("Lock %s not exclusively locked @ %s:%d\n",
1505 rw->lock_object.lo_name, file, line);
1506 if (rw_recursed(rw)) {
1507 if (what & RA_NOTRECURSED)
1508 panic("Lock %s recursed @ %s:%d\n",
1509 rw->lock_object.lo_name, file, line);
1510 } else if (what & RA_RECURSED)
1511 panic("Lock %s not recursed @ %s:%d\n",
1512 rw->lock_object.lo_name, file, line);
1513 break;
1514 case RA_UNLOCKED:
1515 #ifdef WITNESS
1516 witness_assert(&rw->lock_object, what, file, line);
1517 #else
1518 /*
1519 * If we hold a write lock fail. We can't reliably check
1520 * to see if we hold a read lock or not.
1521 */
1522 if (rw_wowner(rw) == curthread)
1523 panic("Lock %s exclusively locked @ %s:%d\n",
1524 rw->lock_object.lo_name, file, line);
1525 #endif
1526 break;
1527 default:
1528 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1529 line);
1530 }
1531 }
1532 #endif /* INVARIANT_SUPPORT */
1533
1534 #ifdef DDB
1535 static void
db_show_rwlock(const struct lock_object * lock)1536 db_show_rwlock(const struct lock_object *lock)
1537 {
1538 const struct rwlock *rw;
1539 struct thread *td;
1540
1541 rw = (const struct rwlock *)lock;
1542
1543 db_printf(" state: ");
1544 if (rw->rw_lock == RW_UNLOCKED)
1545 db_printf("UNLOCKED\n");
1546 else if (rw->rw_lock == RW_DESTROYED) {
1547 db_printf("DESTROYED\n");
1548 return;
1549 } else if (rw->rw_lock & RW_LOCK_READ)
1550 db_printf("RLOCK: %ju locks\n",
1551 (uintmax_t)(RW_READERS(rw->rw_lock)));
1552 else {
1553 td = rw_wowner(rw);
1554 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1555 td->td_tid, td->td_proc->p_pid, td->td_name);
1556 if (rw_recursed(rw))
1557 db_printf(" recursed: %u\n", rw->rw_recurse);
1558 }
1559 db_printf(" waiters: ");
1560 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1561 case RW_LOCK_READ_WAITERS:
1562 db_printf("readers\n");
1563 break;
1564 case RW_LOCK_WRITE_WAITERS:
1565 db_printf("writers\n");
1566 break;
1567 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1568 db_printf("readers and writers\n");
1569 break;
1570 default:
1571 db_printf("none\n");
1572 break;
1573 }
1574 }
1575
1576 #endif
1577