]> CyberLeo.Net >> Repos - FreeBSD/stable/9.git/blob - sys/kern/kern_rwlock.c
MFC r284297: several lockstat improvements
[FreeBSD/stable/9.git] / sys / kern / kern_rwlock.c
1 /*-
2  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26
27 /*
28  * Machine independent bits of reader/writer lock implementation.
29  */
30
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33
34 #include "opt_ddb.h"
35 #include "opt_hwpmc_hooks.h"
36 #include "opt_kdtrace.h"
37 #include "opt_no_adaptive_rwlocks.h"
38
39 #include <sys/param.h>
40 #include <sys/kdb.h>
41 #include <sys/ktr.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/rwlock.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 #ifdef ADAPTIVE_RWLOCKS
63 #define ROWNER_RETRIES  10
64 #define ROWNER_LOOPS    10000
65 #endif
66
67 #ifdef DDB
68 #include <ddb/ddb.h>
69
70 static void     db_show_rwlock(struct lock_object *lock);
71 #endif
72 static void     assert_rw(struct lock_object *lock, int what);
73 static void     lock_rw(struct lock_object *lock, int how);
74 #ifdef KDTRACE_HOOKS
75 static int      owner_rw(struct lock_object *lock, struct thread **owner);
76 #endif
77 static int      unlock_rw(struct lock_object *lock);
78
79 struct lock_class lock_class_rw = {
80         .lc_name = "rw",
81         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
82         .lc_assert = assert_rw,
83 #ifdef DDB
84         .lc_ddb_show = db_show_rwlock,
85 #endif
86         .lc_lock = lock_rw,
87         .lc_unlock = unlock_rw,
88 #ifdef KDTRACE_HOOKS
89         .lc_owner = owner_rw,
90 #endif
91 };
92
93 /*
94  * Return a pointer to the owning thread if the lock is write-locked or
95  * NULL if the lock is unlocked or read-locked.
96  */
97 #define rw_wowner(rw)                                                   \
98         ((rw)->rw_lock & RW_LOCK_READ ? NULL :                          \
99             (struct thread *)RW_OWNER((rw)->rw_lock))
100
101 /*
102  * Returns if a write owner is recursed.  Write ownership is not assured
103  * here and should be previously checked.
104  */
105 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
106
107 /*
108  * Return true if curthread helds the lock.
109  */
110 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
111
112 /*
113  * Return a pointer to the owning thread for this lock who should receive
114  * any priority lent by threads that block on this lock.  Currently this
115  * is identical to rw_wowner().
116  */
117 #define rw_owner(rw)            rw_wowner(rw)
118
119 #ifndef INVARIANTS
120 #define _rw_assert(rw, what, file, line)
121 #endif
122
123 void
124 assert_rw(struct lock_object *lock, int what)
125 {
126
127         rw_assert((struct rwlock *)lock, what);
128 }
129
130 void
131 lock_rw(struct lock_object *lock, int how)
132 {
133         struct rwlock *rw;
134
135         rw = (struct rwlock *)lock;
136         if (how)
137                 rw_wlock(rw);
138         else
139                 rw_rlock(rw);
140 }
141
142 int
143 unlock_rw(struct lock_object *lock)
144 {
145         struct rwlock *rw;
146
147         rw = (struct rwlock *)lock;
148         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
149         if (rw->rw_lock & RW_LOCK_READ) {
150                 rw_runlock(rw);
151                 return (0);
152         } else {
153                 rw_wunlock(rw);
154                 return (1);
155         }
156 }
157
158 #ifdef KDTRACE_HOOKS
159 int
160 owner_rw(struct lock_object *lock, struct thread **owner)
161 {
162         struct rwlock *rw = (struct rwlock *)lock;
163         uintptr_t x = rw->rw_lock;
164
165         *owner = rw_wowner(rw);
166         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
167             (*owner != NULL));
168 }
169 #endif
170
171 void
172 rw_init_flags(struct rwlock *rw, const char *name, int opts)
173 {
174         int flags;
175
176         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
177             RW_RECURSE)) == 0);
178         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
179             ("%s: rw_lock not aligned for %s: %p", __func__, name,
180             &rw->rw_lock));
181
182         flags = LO_UPGRADABLE;
183         if (opts & RW_DUPOK)
184                 flags |= LO_DUPOK;
185         if (opts & RW_NOPROFILE)
186                 flags |= LO_NOPROFILE;
187         if (!(opts & RW_NOWITNESS))
188                 flags |= LO_WITNESS;
189         if (opts & RW_RECURSE)
190                 flags |= LO_RECURSABLE;
191         if (opts & RW_QUIET)
192                 flags |= LO_QUIET;
193
194         rw->rw_lock = RW_UNLOCKED;
195         rw->rw_recurse = 0;
196         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
197 }
198
199 void
200 rw_destroy(struct rwlock *rw)
201 {
202
203         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
204         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
205         rw->rw_lock = RW_DESTROYED;
206         lock_destroy(&rw->lock_object);
207 }
208
209 void
210 rw_sysinit(void *arg)
211 {
212         struct rw_args *args = arg;
213
214         rw_init(args->ra_rw, args->ra_desc);
215 }
216
217 void
218 rw_sysinit_flags(void *arg)
219 {
220         struct rw_args_flags *args = arg;
221
222         rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags);
223 }
224
225 int
226 rw_wowned(struct rwlock *rw)
227 {
228
229         return (rw_wowner(rw) == curthread);
230 }
231
232 void
233 _rw_wlock(struct rwlock *rw, const char *file, int line)
234 {
235
236         if (SCHEDULER_STOPPED())
237                 return;
238         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
239             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
240             curthread, rw->lock_object.lo_name, file, line));
241         KASSERT(rw->rw_lock != RW_DESTROYED,
242             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
243         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
244             line, NULL);
245         __rw_wlock(rw, curthread, file, line);
246         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
247         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
248         curthread->td_locks++;
249 }
250
251 int
252 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
253 {
254         int rval;
255
256         if (SCHEDULER_STOPPED())
257                 return (1);
258
259         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
260             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
261             curthread, rw->lock_object.lo_name, file, line));
262         KASSERT(rw->rw_lock != RW_DESTROYED,
263             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
264
265         if (rw_wlocked(rw) &&
266             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
267                 rw->rw_recurse++;
268                 rval = 1;
269         } else
270                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
271                     (uintptr_t)curthread);
272
273         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
274         if (rval) {
275                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
276                     file, line);
277                 if (!rw_recursed(rw))
278                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE,
279                             rw, 0, 0, file, line);
280                 curthread->td_locks++;
281         }
282         return (rval);
283 }
284
285 void
286 _rw_wunlock(struct rwlock *rw, const char *file, int line)
287 {
288
289         if (SCHEDULER_STOPPED())
290                 return;
291         KASSERT(rw->rw_lock != RW_DESTROYED,
292             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
293         _rw_assert(rw, RA_WLOCKED, file, line);
294         curthread->td_locks--;
295         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
296         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
297             line);
298         if (!rw_recursed(rw))
299                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
300         __rw_wunlock(rw, curthread, file, line);
301 }
302 /*
303  * Determines whether a new reader can acquire a lock.  Succeeds if the
304  * reader already owns a read lock and the lock is locked for read to
305  * prevent deadlock from reader recursion.  Also succeeds if the lock
306  * is unlocked and has no writer waiters or spinners.  Failing otherwise
307  * prioritizes writers before readers.
308  */
309 #define RW_CAN_READ(_rw)                                                \
310     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
311     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
312     RW_LOCK_READ)
313
314 void
315 _rw_rlock(struct rwlock *rw, const char *file, int line)
316 {
317         struct turnstile *ts;
318 #ifdef ADAPTIVE_RWLOCKS
319         volatile struct thread *owner;
320         int spintries = 0;
321         int i;
322 #endif
323 #ifdef LOCK_PROFILING
324         uint64_t waittime = 0;
325         int contested = 0;
326 #endif
327         uintptr_t v;
328 #ifdef KDTRACE_HOOKS
329         uintptr_t state;
330         uint64_t spin_cnt = 0;
331         uint64_t sleep_cnt = 0;
332         int64_t sleep_time = 0;
333         int64_t all_time = 0;
334 #endif
335
336         if (SCHEDULER_STOPPED())
337                 return;
338
339         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
340             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
341             curthread, rw->lock_object.lo_name, file, line));
342         KASSERT(rw->rw_lock != RW_DESTROYED,
343             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
344         KASSERT(rw_wowner(rw) != curthread,
345             ("rw_rlock: wlock already held for %s @ %s:%d",
346             rw->lock_object.lo_name, file, line));
347         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
348
349 #ifdef KDTRACE_HOOKS
350         all_time -= lockstat_nsecs();
351         state = rw->rw_lock;
352 #endif
353         for (;;) {
354 #ifdef KDTRACE_HOOKS
355                 spin_cnt++;
356 #endif
357                 /*
358                  * Handle the easy case.  If no other thread has a write
359                  * lock, then try to bump up the count of read locks.  Note
360                  * that we have to preserve the current state of the
361                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
362                  * read lock, then rw_lock must have changed, so restart
363                  * the loop.  Note that this handles the case of a
364                  * completely unlocked rwlock since such a lock is encoded
365                  * as a read lock with no waiters.
366                  */
367                 v = rw->rw_lock;
368                 if (RW_CAN_READ(v)) {
369                         /*
370                          * The RW_LOCK_READ_WAITERS flag should only be set
371                          * if the lock has been unlocked and write waiters
372                          * were present.
373                          */
374                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
375                             v + RW_ONE_READER)) {
376                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
377                                         CTR4(KTR_LOCK,
378                                             "%s: %p succeed %p -> %p", __func__,
379                                             rw, (void *)v,
380                                             (void *)(v + RW_ONE_READER));
381                                 break;
382                         }
383                         continue;
384                 }
385 #ifdef HWPMC_HOOKS
386                 PMC_SOFT_CALL( , , lock, failed);
387 #endif
388                 lock_profile_obtain_lock_failed(&rw->lock_object,
389                     &contested, &waittime);
390
391 #ifdef ADAPTIVE_RWLOCKS
392                 /*
393                  * If the owner is running on another CPU, spin until
394                  * the owner stops running or the state of the lock
395                  * changes.
396                  */
397                 if ((v & RW_LOCK_READ) == 0) {
398                         owner = (struct thread *)RW_OWNER(v);
399                         if (TD_IS_RUNNING(owner)) {
400                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
401                                         CTR3(KTR_LOCK,
402                                             "%s: spinning on %p held by %p",
403                                             __func__, rw, owner);
404                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
405                                     owner && TD_IS_RUNNING(owner)) {
406                                         cpu_spinwait();
407 #ifdef KDTRACE_HOOKS
408                                         spin_cnt++;
409 #endif
410                                 }
411                                 continue;
412                         }
413                 } else if (spintries < ROWNER_RETRIES) {
414                         spintries++;
415                         for (i = 0; i < ROWNER_LOOPS; i++) {
416                                 v = rw->rw_lock;
417                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
418                                         break;
419                                 cpu_spinwait();
420                         }
421                         if (i != ROWNER_LOOPS)
422                                 continue;
423                 }
424 #endif
425
426                 /*
427                  * Okay, now it's the hard case.  Some other thread already
428                  * has a write lock or there are write waiters present,
429                  * acquire the turnstile lock so we can begin the process
430                  * of blocking.
431                  */
432                 ts = turnstile_trywait(&rw->lock_object);
433
434                 /*
435                  * The lock might have been released while we spun, so
436                  * recheck its state and restart the loop if needed.
437                  */
438                 v = rw->rw_lock;
439                 if (RW_CAN_READ(v)) {
440                         turnstile_cancel(ts);
441                         continue;
442                 }
443
444 #ifdef ADAPTIVE_RWLOCKS
445                 /*
446                  * The current lock owner might have started executing
447                  * on another CPU (or the lock could have changed
448                  * owners) while we were waiting on the turnstile
449                  * chain lock.  If so, drop the turnstile lock and try
450                  * again.
451                  */
452                 if ((v & RW_LOCK_READ) == 0) {
453                         owner = (struct thread *)RW_OWNER(v);
454                         if (TD_IS_RUNNING(owner)) {
455                                 turnstile_cancel(ts);
456                                 continue;
457                         }
458                 }
459 #endif
460
461                 /*
462                  * The lock is held in write mode or it already has waiters.
463                  */
464                 MPASS(!RW_CAN_READ(v));
465
466                 /*
467                  * If the RW_LOCK_READ_WAITERS flag is already set, then
468                  * we can go ahead and block.  If it is not set then try
469                  * to set it.  If we fail to set it drop the turnstile
470                  * lock and restart the loop.
471                  */
472                 if (!(v & RW_LOCK_READ_WAITERS)) {
473                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
474                             v | RW_LOCK_READ_WAITERS)) {
475                                 turnstile_cancel(ts);
476                                 continue;
477                         }
478                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
479                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
480                                     __func__, rw);
481                 }
482
483                 /*
484                  * We were unable to acquire the lock and the read waiters
485                  * flag is set, so we must block on the turnstile.
486                  */
487                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
488                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
489                             rw);
490 #ifdef KDTRACE_HOOKS
491                 sleep_time -= lockstat_nsecs();
492 #endif
493                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
494 #ifdef KDTRACE_HOOKS
495                 sleep_time += lockstat_nsecs();
496                 sleep_cnt++;
497 #endif
498                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
499                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
500                             __func__, rw);
501         }
502 #ifdef KDTRACE_HOOKS
503         all_time += lockstat_nsecs();
504         if (sleep_time)
505                 LOCKSTAT_RECORD4(LS_RW_RLOCK_BLOCK, rw, sleep_time,
506                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
507                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
508
509         /* Record only the loops spinning and not sleeping. */
510         if (spin_cnt > sleep_cnt)
511                 LOCKSTAT_RECORD4(LS_RW_RLOCK_SPIN, rw, all_time - sleep_time,
512                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
513                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
514 #endif
515         /*
516          * TODO: acquire "owner of record" here.  Here be turnstile dragons
517          * however.  turnstiles don't like owners changing between calls to
518          * turnstile_wait() currently.
519          */
520         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
521             waittime, file, line);
522         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
523         WITNESS_LOCK(&rw->lock_object, 0, file, line);
524         curthread->td_locks++;
525         curthread->td_rw_rlocks++;
526 }
527
528 int
529 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
530 {
531         uintptr_t x;
532
533         if (SCHEDULER_STOPPED())
534                 return (1);
535
536         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
537             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
538             curthread, rw->lock_object.lo_name, file, line));
539
540         for (;;) {
541                 x = rw->rw_lock;
542                 KASSERT(rw->rw_lock != RW_DESTROYED,
543                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
544                 if (!(x & RW_LOCK_READ))
545                         break;
546                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
547                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
548                             line);
549                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
550                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE,
551                             rw, 0, 0, file, line);
552                         curthread->td_locks++;
553                         curthread->td_rw_rlocks++;
554                         return (1);
555                 }
556         }
557
558         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
559         return (0);
560 }
561
562 void
563 _rw_runlock(struct rwlock *rw, const char *file, int line)
564 {
565         struct turnstile *ts;
566         uintptr_t x, v, queue;
567
568         if (SCHEDULER_STOPPED())
569                 return;
570
571         KASSERT(rw->rw_lock != RW_DESTROYED,
572             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
573         _rw_assert(rw, RA_RLOCKED, file, line);
574         curthread->td_locks--;
575         curthread->td_rw_rlocks--;
576         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
577         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
578
579         /* TODO: drop "owner of record" here. */
580
581         for (;;) {
582                 /*
583                  * See if there is more than one read lock held.  If so,
584                  * just drop one and return.
585                  */
586                 x = rw->rw_lock;
587                 if (RW_READERS(x) > 1) {
588                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
589                             x - RW_ONE_READER)) {
590                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
591                                         CTR4(KTR_LOCK,
592                                             "%s: %p succeeded %p -> %p",
593                                             __func__, rw, (void *)x,
594                                             (void *)(x - RW_ONE_READER));
595                                 break;
596                         }
597                         continue;
598                 }
599                 /*
600                  * If there aren't any waiters for a write lock, then try
601                  * to drop it quickly.
602                  */
603                 if (!(x & RW_LOCK_WAITERS)) {
604                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
605                             RW_READERS_LOCK(1));
606                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
607                             RW_UNLOCKED)) {
608                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
609                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
610                                             __func__, rw);
611                                 break;
612                         }
613                         continue;
614                 }
615                 /*
616                  * Ok, we know we have waiters and we think we are the
617                  * last reader, so grab the turnstile lock.
618                  */
619                 turnstile_chain_lock(&rw->lock_object);
620                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
621                 MPASS(v & RW_LOCK_WAITERS);
622
623                 /*
624                  * Try to drop our lock leaving the lock in a unlocked
625                  * state.
626                  *
627                  * If you wanted to do explicit lock handoff you'd have to
628                  * do it here.  You'd also want to use turnstile_signal()
629                  * and you'd have to handle the race where a higher
630                  * priority thread blocks on the write lock before the
631                  * thread you wakeup actually runs and have the new thread
632                  * "steal" the lock.  For now it's a lot simpler to just
633                  * wakeup all of the waiters.
634                  *
635                  * As above, if we fail, then another thread might have
636                  * acquired a read lock, so drop the turnstile lock and
637                  * restart.
638                  */
639                 x = RW_UNLOCKED;
640                 if (v & RW_LOCK_WRITE_WAITERS) {
641                         queue = TS_EXCLUSIVE_QUEUE;
642                         x |= (v & RW_LOCK_READ_WAITERS);
643                 } else
644                         queue = TS_SHARED_QUEUE;
645                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
646                     x)) {
647                         turnstile_chain_unlock(&rw->lock_object);
648                         continue;
649                 }
650                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
651                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
652                             __func__, rw);
653
654                 /*
655                  * Ok.  The lock is released and all that's left is to
656                  * wake up the waiters.  Note that the lock might not be
657                  * free anymore, but in that case the writers will just
658                  * block again if they run before the new lock holder(s)
659                  * release the lock.
660                  */
661                 ts = turnstile_lookup(&rw->lock_object);
662                 MPASS(ts != NULL);
663                 turnstile_broadcast(ts, queue);
664                 turnstile_unpend(ts, TS_SHARED_LOCK);
665                 turnstile_chain_unlock(&rw->lock_object);
666                 break;
667         }
668         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
669 }
670
671 /*
672  * This function is called when we are unable to obtain a write lock on the
673  * first try.  This means that at least one other thread holds either a
674  * read or write lock.
675  */
676 void
677 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
678 {
679         struct turnstile *ts;
680 #ifdef ADAPTIVE_RWLOCKS
681         volatile struct thread *owner;
682         int spintries = 0;
683         int i;
684 #endif
685         uintptr_t v, x;
686 #ifdef LOCK_PROFILING
687         uint64_t waittime = 0;
688         int contested = 0;
689 #endif
690 #ifdef KDTRACE_HOOKS
691         uintptr_t state;
692         uint64_t spin_cnt = 0;
693         uint64_t sleep_cnt = 0;
694         int64_t sleep_time = 0;
695         int64_t all_time = 0;
696 #endif
697
698         if (SCHEDULER_STOPPED())
699                 return;
700
701         if (rw_wlocked(rw)) {
702                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
703                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
704                     __func__, rw->lock_object.lo_name, file, line));
705                 rw->rw_recurse++;
706                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
707                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
708                 return;
709         }
710
711         if (LOCK_LOG_TEST(&rw->lock_object, 0))
712                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
713                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
714
715 #ifdef KDTRACE_HOOKS
716         all_time -= lockstat_nsecs();
717         state = rw->rw_lock;
718 #endif
719         while (!_rw_write_lock(rw, tid)) {
720 #ifdef KDTRACE_HOOKS
721                 spin_cnt++;
722 #endif
723 #ifdef HWPMC_HOOKS
724                 PMC_SOFT_CALL( , , lock, failed);
725 #endif
726                 lock_profile_obtain_lock_failed(&rw->lock_object,
727                     &contested, &waittime);
728 #ifdef ADAPTIVE_RWLOCKS
729                 /*
730                  * If the lock is write locked and the owner is
731                  * running on another CPU, spin until the owner stops
732                  * running or the state of the lock changes.
733                  */
734                 v = rw->rw_lock;
735                 owner = (struct thread *)RW_OWNER(v);
736                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
737                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
738                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
739                                     __func__, rw, owner);
740                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
741                             TD_IS_RUNNING(owner)) {
742                                 cpu_spinwait();
743 #ifdef KDTRACE_HOOKS
744                                 spin_cnt++;
745 #endif
746                         }
747                         continue;
748                 }
749                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
750                     spintries < ROWNER_RETRIES) {
751                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
752                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
753                                     v | RW_LOCK_WRITE_SPINNER)) {
754                                         continue;
755                                 }
756                         }
757                         spintries++;
758                         for (i = 0; i < ROWNER_LOOPS; i++) {
759                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
760                                         break;
761                                 cpu_spinwait();
762                         }
763 #ifdef KDTRACE_HOOKS
764                         spin_cnt += ROWNER_LOOPS - i;
765 #endif
766                         if (i != ROWNER_LOOPS)
767                                 continue;
768                 }
769 #endif
770                 ts = turnstile_trywait(&rw->lock_object);
771                 v = rw->rw_lock;
772
773 #ifdef ADAPTIVE_RWLOCKS
774                 /*
775                  * The current lock owner might have started executing
776                  * on another CPU (or the lock could have changed
777                  * owners) while we were waiting on the turnstile
778                  * chain lock.  If so, drop the turnstile lock and try
779                  * again.
780                  */
781                 if (!(v & RW_LOCK_READ)) {
782                         owner = (struct thread *)RW_OWNER(v);
783                         if (TD_IS_RUNNING(owner)) {
784                                 turnstile_cancel(ts);
785                                 continue;
786                         }
787                 }
788 #endif
789                 /*
790                  * Check for the waiters flags about this rwlock.
791                  * If the lock was released, without maintain any pending
792                  * waiters queue, simply try to acquire it.
793                  * If a pending waiters queue is present, claim the lock
794                  * ownership and maintain the pending queue.
795                  */
796                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
797                 if ((v & ~x) == RW_UNLOCKED) {
798                         x &= ~RW_LOCK_WRITE_SPINNER;
799                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
800                                 if (x)
801                                         turnstile_claim(ts);
802                                 else
803                                         turnstile_cancel(ts);
804                                 break;
805                         }
806                         turnstile_cancel(ts);
807                         continue;
808                 }
809                 /*
810                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
811                  * set it.  If we fail to set it, then loop back and try
812                  * again.
813                  */
814                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
815                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
816                             v | RW_LOCK_WRITE_WAITERS)) {
817                                 turnstile_cancel(ts);
818                                 continue;
819                         }
820                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
821                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
822                                     __func__, rw);
823                 }
824                 /*
825                  * We were unable to acquire the lock and the write waiters
826                  * flag is set, so we must block on the turnstile.
827                  */
828                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
829                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
830                             rw);
831 #ifdef KDTRACE_HOOKS
832                 sleep_time -= lockstat_nsecs();
833 #endif
834                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
835 #ifdef KDTRACE_HOOKS
836                 sleep_time += lockstat_nsecs();
837                 sleep_cnt++;
838 #endif
839                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
840                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
841                             __func__, rw);
842 #ifdef ADAPTIVE_RWLOCKS
843                 spintries = 0;
844 #endif
845         }
846 #ifdef KDTRACE_HOOKS
847         all_time += lockstat_nsecs();
848         if (sleep_time)
849                 LOCKSTAT_RECORD4(LS_RW_WLOCK_BLOCK, rw, sleep_time,
850                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
851                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
852
853         /* Record only the loops spinning and not sleeping. */
854         if (spin_cnt > sleep_cnt)
855                 LOCKSTAT_RECORD4(LS_RW_WLOCK_SPIN, rw, all_time - sleep_time,
856                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
857                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
858 #endif
859         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
860             waittime, file, line);
861 }
862
863 /*
864  * This function is called if the first try at releasing a write lock failed.
865  * This means that one of the 2 waiter bits must be set indicating that at
866  * least one thread is waiting on this lock.
867  */
868 void
869 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
870 {
871         struct turnstile *ts;
872         uintptr_t v;
873         int queue;
874
875         if (SCHEDULER_STOPPED())
876                 return;
877
878         if (rw_wlocked(rw) && rw_recursed(rw)) {
879                 rw->rw_recurse--;
880                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
881                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
882                 return;
883         }
884
885         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
886             ("%s: neither of the waiter flags are set", __func__));
887
888         if (LOCK_LOG_TEST(&rw->lock_object, 0))
889                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
890
891         turnstile_chain_lock(&rw->lock_object);
892         ts = turnstile_lookup(&rw->lock_object);
893         MPASS(ts != NULL);
894
895         /*
896          * Use the same algo as sx locks for now.  Prefer waking up shared
897          * waiters if we have any over writers.  This is probably not ideal.
898          *
899          * 'v' is the value we are going to write back to rw_lock.  If we
900          * have waiters on both queues, we need to preserve the state of
901          * the waiter flag for the queue we don't wake up.  For now this is
902          * hardcoded for the algorithm mentioned above.
903          *
904          * In the case of both readers and writers waiting we wakeup the
905          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
906          * new writer comes in before a reader it will claim the lock up
907          * above.  There is probably a potential priority inversion in
908          * there that could be worked around either by waking both queues
909          * of waiters or doing some complicated lock handoff gymnastics.
910          */
911         v = RW_UNLOCKED;
912         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
913                 queue = TS_EXCLUSIVE_QUEUE;
914                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
915         } else
916                 queue = TS_SHARED_QUEUE;
917
918         /* Wake up all waiters for the specific queue. */
919         if (LOCK_LOG_TEST(&rw->lock_object, 0))
920                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
921                     queue == TS_SHARED_QUEUE ? "read" : "write");
922         turnstile_broadcast(ts, queue);
923         atomic_store_rel_ptr(&rw->rw_lock, v);
924         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
925         turnstile_chain_unlock(&rw->lock_object);
926 }
927
928 /*
929  * Attempt to do a non-blocking upgrade from a read lock to a write
930  * lock.  This will only succeed if this thread holds a single read
931  * lock.  Returns true if the upgrade succeeded and false otherwise.
932  */
933 int
934 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
935 {
936         uintptr_t v, x, tid;
937         struct turnstile *ts;
938         int success;
939
940         if (SCHEDULER_STOPPED())
941                 return (1);
942
943         KASSERT(rw->rw_lock != RW_DESTROYED,
944             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
945         _rw_assert(rw, RA_RLOCKED, file, line);
946
947         /*
948          * Attempt to switch from one reader to a writer.  If there
949          * are any write waiters, then we will have to lock the
950          * turnstile first to prevent races with another writer
951          * calling turnstile_wait() before we have claimed this
952          * turnstile.  So, do the simple case of no waiters first.
953          */
954         tid = (uintptr_t)curthread;
955         success = 0;
956         for (;;) {
957                 v = rw->rw_lock;
958                 if (RW_READERS(v) > 1)
959                         break;
960                 if (!(v & RW_LOCK_WAITERS)) {
961                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
962                         if (!success)
963                                 continue;
964                         break;
965                 }
966
967                 /*
968                  * Ok, we think we have waiters, so lock the turnstile.
969                  */
970                 ts = turnstile_trywait(&rw->lock_object);
971                 v = rw->rw_lock;
972                 if (RW_READERS(v) > 1) {
973                         turnstile_cancel(ts);
974                         break;
975                 }
976                 /*
977                  * Try to switch from one reader to a writer again.  This time
978                  * we honor the current state of the waiters flags.
979                  * If we obtain the lock with the flags set, then claim
980                  * ownership of the turnstile.
981                  */
982                 x = rw->rw_lock & RW_LOCK_WAITERS;
983                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
984                 if (success) {
985                         if (x)
986                                 turnstile_claim(ts);
987                         else
988                                 turnstile_cancel(ts);
989                         break;
990                 }
991                 turnstile_cancel(ts);
992         }
993         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
994         if (success) {
995                 curthread->td_rw_rlocks--;
996                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
997                     file, line);
998                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
999         }
1000         return (success);
1001 }
1002
1003 /*
1004  * Downgrade a write lock into a single read lock.
1005  */
1006 void
1007 _rw_downgrade(struct rwlock *rw, const char *file, int line)
1008 {
1009         struct turnstile *ts;
1010         uintptr_t tid, v;
1011         int rwait, wwait;
1012
1013         if (SCHEDULER_STOPPED())
1014                 return;
1015
1016         KASSERT(rw->rw_lock != RW_DESTROYED,
1017             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1018         _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
1019 #ifndef INVARIANTS
1020         if (rw_recursed(rw))
1021                 panic("downgrade of a recursed lock");
1022 #endif
1023
1024         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1025
1026         /*
1027          * Convert from a writer to a single reader.  First we handle
1028          * the easy case with no waiters.  If there are any waiters, we
1029          * lock the turnstile and "disown" the lock.
1030          */
1031         tid = (uintptr_t)curthread;
1032         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1033                 goto out;
1034
1035         /*
1036          * Ok, we think we have waiters, so lock the turnstile so we can
1037          * read the waiter flags without any races.
1038          */
1039         turnstile_chain_lock(&rw->lock_object);
1040         v = rw->rw_lock & RW_LOCK_WAITERS;
1041         rwait = v & RW_LOCK_READ_WAITERS;
1042         wwait = v & RW_LOCK_WRITE_WAITERS;
1043         MPASS(rwait | wwait);
1044
1045         /*
1046          * Downgrade from a write lock while preserving waiters flag
1047          * and give up ownership of the turnstile.
1048          */
1049         ts = turnstile_lookup(&rw->lock_object);
1050         MPASS(ts != NULL);
1051         if (!wwait)
1052                 v &= ~RW_LOCK_READ_WAITERS;
1053         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1054         /*
1055          * Wake other readers if there are no writers pending.  Otherwise they
1056          * won't be able to acquire the lock anyway.
1057          */
1058         if (rwait && !wwait) {
1059                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1060                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1061         } else
1062                 turnstile_disown(ts);
1063         turnstile_chain_unlock(&rw->lock_object);
1064 out:
1065         curthread->td_rw_rlocks++;
1066         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1067         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1068 }
1069
1070 #ifdef INVARIANT_SUPPORT
1071 #ifndef INVARIANTS
1072 #undef _rw_assert
1073 #endif
1074
1075 /*
1076  * In the non-WITNESS case, rw_assert() can only detect that at least
1077  * *some* thread owns an rlock, but it cannot guarantee that *this*
1078  * thread owns an rlock.
1079  */
1080 void
1081 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
1082 {
1083
1084         if (panicstr != NULL)
1085                 return;
1086         switch (what) {
1087         case RA_LOCKED:
1088         case RA_LOCKED | RA_RECURSED:
1089         case RA_LOCKED | RA_NOTRECURSED:
1090         case RA_RLOCKED:
1091         case RA_RLOCKED | RA_RECURSED:
1092         case RA_RLOCKED | RA_NOTRECURSED:
1093 #ifdef WITNESS
1094                 witness_assert(&rw->lock_object, what, file, line);
1095 #else
1096                 /*
1097                  * If some other thread has a write lock or we have one
1098                  * and are asserting a read lock, fail.  Also, if no one
1099                  * has a lock at all, fail.
1100                  */
1101                 if (rw->rw_lock == RW_UNLOCKED ||
1102                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1103                     rw_wowner(rw) != curthread)))
1104                         panic("Lock %s not %slocked @ %s:%d\n",
1105                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1106                             "read " : "", file, line);
1107
1108                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1109                         if (rw_recursed(rw)) {
1110                                 if (what & RA_NOTRECURSED)
1111                                         panic("Lock %s recursed @ %s:%d\n",
1112                                             rw->lock_object.lo_name, file,
1113                                             line);
1114                         } else if (what & RA_RECURSED)
1115                                 panic("Lock %s not recursed @ %s:%d\n",
1116                                     rw->lock_object.lo_name, file, line);
1117                 }
1118 #endif
1119                 break;
1120         case RA_WLOCKED:
1121         case RA_WLOCKED | RA_RECURSED:
1122         case RA_WLOCKED | RA_NOTRECURSED:
1123                 if (rw_wowner(rw) != curthread)
1124                         panic("Lock %s not exclusively locked @ %s:%d\n",
1125                             rw->lock_object.lo_name, file, line);
1126                 if (rw_recursed(rw)) {
1127                         if (what & RA_NOTRECURSED)
1128                                 panic("Lock %s recursed @ %s:%d\n",
1129                                     rw->lock_object.lo_name, file, line);
1130                 } else if (what & RA_RECURSED)
1131                         panic("Lock %s not recursed @ %s:%d\n",
1132                             rw->lock_object.lo_name, file, line);
1133                 break;
1134         case RA_UNLOCKED:
1135 #ifdef WITNESS
1136                 witness_assert(&rw->lock_object, what, file, line);
1137 #else
1138                 /*
1139                  * If we hold a write lock fail.  We can't reliably check
1140                  * to see if we hold a read lock or not.
1141                  */
1142                 if (rw_wowner(rw) == curthread)
1143                         panic("Lock %s exclusively locked @ %s:%d\n",
1144                             rw->lock_object.lo_name, file, line);
1145 #endif
1146                 break;
1147         default:
1148                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1149                     line);
1150         }
1151 }
1152 #endif /* INVARIANT_SUPPORT */
1153
1154 #ifdef DDB
1155 void
1156 db_show_rwlock(struct lock_object *lock)
1157 {
1158         struct rwlock *rw;
1159         struct thread *td;
1160
1161         rw = (struct rwlock *)lock;
1162
1163         db_printf(" state: ");
1164         if (rw->rw_lock == RW_UNLOCKED)
1165                 db_printf("UNLOCKED\n");
1166         else if (rw->rw_lock == RW_DESTROYED) {
1167                 db_printf("DESTROYED\n");
1168                 return;
1169         } else if (rw->rw_lock & RW_LOCK_READ)
1170                 db_printf("RLOCK: %ju locks\n",
1171                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1172         else {
1173                 td = rw_wowner(rw);
1174                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1175                     td->td_tid, td->td_proc->p_pid, td->td_name);
1176                 if (rw_recursed(rw))
1177                         db_printf(" recursed: %u\n", rw->rw_recurse);
1178         }
1179         db_printf(" waiters: ");
1180         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1181         case RW_LOCK_READ_WAITERS:
1182                 db_printf("readers\n");
1183                 break;
1184         case RW_LOCK_WRITE_WAITERS:
1185                 db_printf("writers\n");
1186                 break;
1187         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1188                 db_printf("readers and writers\n");
1189                 break;
1190         default:
1191                 db_printf("none\n");
1192                 break;
1193         }
1194 }
1195
1196 #endif