]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/kern/kern_sx.c
Merge ACPICA 20120320.
[FreeBSD/FreeBSD.git] / sys / kern / kern_sx.c
1 /*-
2  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
3  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
4  * All rights reserved.
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(s), this list of conditions and the following disclaimer as
11  *    the first lines of this file unmodified other than the possible
12  *    addition of one or more copyright notices.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice(s), this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
18  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
19  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
20  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
21  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
24  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
27  * DAMAGE.
28  */
29
30 /*
31  * Shared/exclusive locks.  This implementation attempts to ensure
32  * deterministic lock granting behavior, so that slocks and xlocks are
33  * interleaved.
34  *
35  * Priority propagation will not generally raise the priority of lock holders,
36  * so should not be relied upon in combination with sx locks.
37  */
38
39 #include "opt_ddb.h"
40 #include "opt_kdtrace.h"
41 #include "opt_no_adaptive_sx.h"
42
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/ktr.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/sleepqueue.h>
53 #include <sys/sx.h>
54 #include <sys/sysctl.h>
55
56 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
57 #include <machine/cpu.h>
58 #endif
59
60 #ifdef DDB
61 #include <ddb/ddb.h>
62 #endif
63
64 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
65 #define ADAPTIVE_SX
66 #endif
67
68 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
69
70 /* Handy macros for sleep queues. */
71 #define SQ_EXCLUSIVE_QUEUE      0
72 #define SQ_SHARED_QUEUE         1
73
74 /*
75  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
76  * drop Giant anytime we have to sleep or if we adaptively spin.
77  */
78 #define GIANT_DECLARE                                                   \
79         int _giantcnt = 0;                                              \
80         WITNESS_SAVE_DECL(Giant)                                        \
81
82 #define GIANT_SAVE() do {                                               \
83         if (mtx_owned(&Giant)) {                                        \
84                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
85                 while (mtx_owned(&Giant)) {                             \
86                         _giantcnt++;                                    \
87                         mtx_unlock(&Giant);                             \
88                 }                                                       \
89         }                                                               \
90 } while (0)
91
92 #define GIANT_RESTORE() do {                                            \
93         if (_giantcnt > 0) {                                            \
94                 mtx_assert(&Giant, MA_NOTOWNED);                        \
95                 while (_giantcnt--)                                     \
96                         mtx_lock(&Giant);                               \
97                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
98         }                                                               \
99 } while (0)
100
101 /*
102  * Returns true if an exclusive lock is recursed.  It assumes
103  * curthread currently has an exclusive lock.
104  */
105 #define sx_recurse              lock_object.lo_data
106 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
107
108 static void     assert_sx(const struct lock_object *lock, int what);
109 #ifdef DDB
110 static void     db_show_sx(const struct lock_object *lock);
111 #endif
112 static void     lock_sx(struct lock_object *lock, int how);
113 #ifdef KDTRACE_HOOKS
114 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
115 #endif
116 static int      unlock_sx(struct lock_object *lock);
117
118 struct lock_class lock_class_sx = {
119         .lc_name = "sx",
120         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
121         .lc_assert = assert_sx,
122 #ifdef DDB
123         .lc_ddb_show = db_show_sx,
124 #endif
125         .lc_lock = lock_sx,
126         .lc_unlock = unlock_sx,
127 #ifdef KDTRACE_HOOKS
128         .lc_owner = owner_sx,
129 #endif
130 };
131
132 #ifndef INVARIANTS
133 #define _sx_assert(sx, what, file, line)
134 #endif
135
136 #ifdef ADAPTIVE_SX
137 static u_int asx_retries = 10;
138 static u_int asx_loops = 10000;
139 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
140 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
141 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
142 #endif
143
144 void
145 assert_sx(const struct lock_object *lock, int what)
146 {
147
148         sx_assert((const struct sx *)lock, what);
149 }
150
151 void
152 lock_sx(struct lock_object *lock, int how)
153 {
154         struct sx *sx;
155
156         sx = (struct sx *)lock;
157         if (how)
158                 sx_xlock(sx);
159         else
160                 sx_slock(sx);
161 }
162
163 int
164 unlock_sx(struct lock_object *lock)
165 {
166         struct sx *sx;
167
168         sx = (struct sx *)lock;
169         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
170         if (sx_xlocked(sx)) {
171                 sx_xunlock(sx);
172                 return (1);
173         } else {
174                 sx_sunlock(sx);
175                 return (0);
176         }
177 }
178
179 #ifdef KDTRACE_HOOKS
180 int
181 owner_sx(const struct lock_object *lock, struct thread **owner)
182 {
183         const struct sx *sx = (const struct sx *)lock;
184         uintptr_t x = sx->sx_lock;
185
186         *owner = (struct thread *)SX_OWNER(x);
187         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
188             (*owner != NULL));
189 }
190 #endif
191
192 void
193 sx_sysinit(void *arg)
194 {
195         struct sx_args *sargs = arg;
196
197         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
198 }
199
200 void
201 sx_init_flags(struct sx *sx, const char *description, int opts)
202 {
203         int flags;
204
205         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
206             SX_NOPROFILE | SX_NOADAPTIVE)) == 0);
207         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
208             ("%s: sx_lock not aligned for %s: %p", __func__, description,
209             &sx->sx_lock));
210
211         flags = LO_SLEEPABLE | LO_UPGRADABLE;
212         if (opts & SX_DUPOK)
213                 flags |= LO_DUPOK;
214         if (opts & SX_NOPROFILE)
215                 flags |= LO_NOPROFILE;
216         if (!(opts & SX_NOWITNESS))
217                 flags |= LO_WITNESS;
218         if (opts & SX_RECURSE)
219                 flags |= LO_RECURSABLE;
220         if (opts & SX_QUIET)
221                 flags |= LO_QUIET;
222
223         flags |= opts & SX_NOADAPTIVE;
224         sx->sx_lock = SX_LOCK_UNLOCKED;
225         sx->sx_recurse = 0;
226         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
227 }
228
229 void
230 sx_destroy(struct sx *sx)
231 {
232
233         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
234         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
235         sx->sx_lock = SX_LOCK_DESTROYED;
236         lock_destroy(&sx->lock_object);
237 }
238
239 int
240 _sx_slock(struct sx *sx, int opts, const char *file, int line)
241 {
242         int error = 0;
243
244         if (SCHEDULER_STOPPED())
245                 return (0);
246         MPASS(curthread != NULL);
247         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
248             ("sx_slock() of destroyed sx @ %s:%d", file, line));
249         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
250         error = __sx_slock(sx, opts, file, line);
251         if (!error) {
252                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
253                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
254                 curthread->td_locks++;
255         }
256
257         return (error);
258 }
259
260 int
261 sx_try_slock_(struct sx *sx, const char *file, int line)
262 {
263         uintptr_t x;
264
265         if (SCHEDULER_STOPPED())
266                 return (1);
267
268         for (;;) {
269                 x = sx->sx_lock;
270                 KASSERT(x != SX_LOCK_DESTROYED,
271                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
272                 if (!(x & SX_LOCK_SHARED))
273                         break;
274                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
275                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
276                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
277                         curthread->td_locks++;
278                         return (1);
279                 }
280         }
281
282         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
283         return (0);
284 }
285
286 int
287 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
288 {
289         int error = 0;
290
291         if (SCHEDULER_STOPPED())
292                 return (0);
293         MPASS(curthread != NULL);
294         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
295             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
296         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
297             line, NULL);
298         error = __sx_xlock(sx, curthread, opts, file, line);
299         if (!error) {
300                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
301                     file, line);
302                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
303                 curthread->td_locks++;
304         }
305
306         return (error);
307 }
308
309 int
310 sx_try_xlock_(struct sx *sx, const char *file, int line)
311 {
312         int rval;
313
314         if (SCHEDULER_STOPPED())
315                 return (1);
316
317         MPASS(curthread != NULL);
318         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
319             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
320
321         if (sx_xlocked(sx) &&
322             (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
323                 sx->sx_recurse++;
324                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
325                 rval = 1;
326         } else
327                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
328                     (uintptr_t)curthread);
329         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
330         if (rval) {
331                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
332                     file, line);
333                 curthread->td_locks++;
334         }
335
336         return (rval);
337 }
338
339 void
340 _sx_sunlock(struct sx *sx, const char *file, int line)
341 {
342
343         if (SCHEDULER_STOPPED())
344                 return;
345         MPASS(curthread != NULL);
346         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
347             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
348         _sx_assert(sx, SA_SLOCKED, file, line);
349         curthread->td_locks--;
350         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
351         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
352         __sx_sunlock(sx, file, line);
353         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_SUNLOCK_RELEASE, sx);
354 }
355
356 void
357 _sx_xunlock(struct sx *sx, const char *file, int line)
358 {
359
360         if (SCHEDULER_STOPPED())
361                 return;
362         MPASS(curthread != NULL);
363         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
364             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
365         _sx_assert(sx, SA_XLOCKED, file, line);
366         curthread->td_locks--;
367         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
368         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
369             line);
370         if (!sx_recursed(sx))
371                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_XUNLOCK_RELEASE, sx);
372         __sx_xunlock(sx, curthread, file, line);
373 }
374
375 /*
376  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
377  * This will only succeed if this thread holds a single shared lock.
378  * Return 1 if if the upgrade succeed, 0 otherwise.
379  */
380 int
381 sx_try_upgrade_(struct sx *sx, const char *file, int line)
382 {
383         uintptr_t x;
384         int success;
385
386         if (SCHEDULER_STOPPED())
387                 return (1);
388
389         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
390             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
391         _sx_assert(sx, SA_SLOCKED, file, line);
392
393         /*
394          * Try to switch from one shared lock to an exclusive lock.  We need
395          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
396          * we will wake up the exclusive waiters when we drop the lock.
397          */
398         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
399         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
400             (uintptr_t)curthread | x);
401         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
402         if (success) {
403                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
404                     file, line);
405                 LOCKSTAT_RECORD0(LS_SX_TRYUPGRADE_UPGRADE, sx);
406         }
407         return (success);
408 }
409
410 /*
411  * Downgrade an unrecursed exclusive lock into a single shared lock.
412  */
413 void
414 sx_downgrade_(struct sx *sx, const char *file, int line)
415 {
416         uintptr_t x;
417         int wakeup_swapper;
418
419         if (SCHEDULER_STOPPED())
420                 return;
421
422         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
423             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
424         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
425 #ifndef INVARIANTS
426         if (sx_recursed(sx))
427                 panic("downgrade of a recursed lock");
428 #endif
429
430         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
431
432         /*
433          * Try to switch from an exclusive lock with no shared waiters
434          * to one sharer with no shared waiters.  If there are
435          * exclusive waiters, we don't need to lock the sleep queue so
436          * long as we preserve the flag.  We do one quick try and if
437          * that fails we grab the sleepq lock to keep the flags from
438          * changing and do it the slow way.
439          *
440          * We have to lock the sleep queue if there are shared waiters
441          * so we can wake them up.
442          */
443         x = sx->sx_lock;
444         if (!(x & SX_LOCK_SHARED_WAITERS) &&
445             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
446             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
447                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
448                 return;
449         }
450
451         /*
452          * Lock the sleep queue so we can read the waiters bits
453          * without any races and wakeup any shared waiters.
454          */
455         sleepq_lock(&sx->lock_object);
456
457         /*
458          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
459          * shared lock.  If there are any shared waiters, wake them up.
460          */
461         wakeup_swapper = 0;
462         x = sx->sx_lock;
463         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
464             (x & SX_LOCK_EXCLUSIVE_WAITERS));
465         if (x & SX_LOCK_SHARED_WAITERS)
466                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
467                     0, SQ_SHARED_QUEUE);
468         sleepq_release(&sx->lock_object);
469
470         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
471         LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx);
472
473         if (wakeup_swapper)
474                 kick_proc0();
475 }
476
477 /*
478  * This function represents the so-called 'hard case' for sx_xlock
479  * operation.  All 'easy case' failures are redirected to this.  Note
480  * that ideally this would be a static function, but it needs to be
481  * accessible from at least sx.h.
482  */
483 int
484 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
485     int line)
486 {
487         GIANT_DECLARE;
488 #ifdef ADAPTIVE_SX
489         volatile struct thread *owner;
490         u_int i, spintries = 0;
491 #endif
492         uintptr_t x;
493 #ifdef LOCK_PROFILING
494         uint64_t waittime = 0;
495         int contested = 0;
496 #endif
497         int error = 0;
498 #ifdef  KDTRACE_HOOKS
499         uint64_t spin_cnt = 0;
500         uint64_t sleep_cnt = 0;
501         int64_t sleep_time = 0;
502 #endif
503
504         if (SCHEDULER_STOPPED())
505                 return (0);
506
507         /* If we already hold an exclusive lock, then recurse. */
508         if (sx_xlocked(sx)) {
509                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
510             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
511                     sx->lock_object.lo_name, file, line));
512                 sx->sx_recurse++;
513                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
514                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
515                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
516                 return (0);
517         }
518
519         if (LOCK_LOG_TEST(&sx->lock_object, 0))
520                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
521                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
522
523         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
524 #ifdef KDTRACE_HOOKS
525                 spin_cnt++;
526 #endif
527                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
528                     &waittime);
529 #ifdef ADAPTIVE_SX
530                 /*
531                  * If the lock is write locked and the owner is
532                  * running on another CPU, spin until the owner stops
533                  * running or the state of the lock changes.
534                  */
535                 x = sx->sx_lock;
536                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
537                         if ((x & SX_LOCK_SHARED) == 0) {
538                                 x = SX_OWNER(x);
539                                 owner = (struct thread *)x;
540                                 if (TD_IS_RUNNING(owner)) {
541                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
542                                                 CTR3(KTR_LOCK,
543                                             "%s: spinning on %p held by %p",
544                                                     __func__, sx, owner);
545                                         GIANT_SAVE();
546                                         while (SX_OWNER(sx->sx_lock) == x &&
547                                             TD_IS_RUNNING(owner)) {
548                                                 cpu_spinwait();
549 #ifdef KDTRACE_HOOKS
550                                                 spin_cnt++;
551 #endif
552                                         }
553                                         continue;
554                                 }
555                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
556                                 GIANT_SAVE();
557                                 spintries++;
558                                 for (i = 0; i < asx_loops; i++) {
559                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
560                                                 CTR4(KTR_LOCK,
561                                     "%s: shared spinning on %p with %u and %u",
562                                                     __func__, sx, spintries, i);
563                                         x = sx->sx_lock;
564                                         if ((x & SX_LOCK_SHARED) == 0 ||
565                                             SX_SHARERS(x) == 0)
566                                                 break;
567                                         cpu_spinwait();
568 #ifdef KDTRACE_HOOKS
569                                         spin_cnt++;
570 #endif
571                                 }
572                                 if (i != asx_loops)
573                                         continue;
574                         }
575                 }
576 #endif
577
578                 sleepq_lock(&sx->lock_object);
579                 x = sx->sx_lock;
580
581                 /*
582                  * If the lock was released while spinning on the
583                  * sleep queue chain lock, try again.
584                  */
585                 if (x == SX_LOCK_UNLOCKED) {
586                         sleepq_release(&sx->lock_object);
587                         continue;
588                 }
589
590 #ifdef ADAPTIVE_SX
591                 /*
592                  * The current lock owner might have started executing
593                  * on another CPU (or the lock could have changed
594                  * owners) while we were waiting on the sleep queue
595                  * chain lock.  If so, drop the sleep queue lock and try
596                  * again.
597                  */
598                 if (!(x & SX_LOCK_SHARED) &&
599                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
600                         owner = (struct thread *)SX_OWNER(x);
601                         if (TD_IS_RUNNING(owner)) {
602                                 sleepq_release(&sx->lock_object);
603                                 continue;
604                         }
605                 }
606 #endif
607
608                 /*
609                  * If an exclusive lock was released with both shared
610                  * and exclusive waiters and a shared waiter hasn't
611                  * woken up and acquired the lock yet, sx_lock will be
612                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
613                  * If we see that value, try to acquire it once.  Note
614                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
615                  * as there are other exclusive waiters still.  If we
616                  * fail, restart the loop.
617                  */
618                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
619                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
620                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
621                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
622                                 sleepq_release(&sx->lock_object);
623                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
624                                     __func__, sx);
625                                 break;
626                         }
627                         sleepq_release(&sx->lock_object);
628                         continue;
629                 }
630
631                 /*
632                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
633                  * than loop back and retry.
634                  */
635                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
636                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
637                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
638                                 sleepq_release(&sx->lock_object);
639                                 continue;
640                         }
641                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
642                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
643                                     __func__, sx);
644                 }
645
646                 /*
647                  * Since we have been unable to acquire the exclusive
648                  * lock and the exclusive waiters flag is set, we have
649                  * to sleep.
650                  */
651                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
652                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
653                             __func__, sx);
654
655 #ifdef KDTRACE_HOOKS
656                 sleep_time -= lockstat_nsecs();
657 #endif
658                 GIANT_SAVE();
659                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
660                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
661                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
662                 if (!(opts & SX_INTERRUPTIBLE))
663                         sleepq_wait(&sx->lock_object, 0);
664                 else
665                         error = sleepq_wait_sig(&sx->lock_object, 0);
666 #ifdef KDTRACE_HOOKS
667                 sleep_time += lockstat_nsecs();
668                 sleep_cnt++;
669 #endif
670                 if (error) {
671                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
672                                 CTR2(KTR_LOCK,
673                         "%s: interruptible sleep by %p suspended by signal",
674                                     __func__, sx);
675                         break;
676                 }
677                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
678                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
679                             __func__, sx);
680         }
681
682         GIANT_RESTORE();
683         if (!error)
684                 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx,
685                     contested, waittime, file, line);
686 #ifdef KDTRACE_HOOKS
687         if (sleep_time)
688                 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time);
689         if (spin_cnt > sleep_cnt)
690                 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt));
691 #endif
692         return (error);
693 }
694
695 /*
696  * This function represents the so-called 'hard case' for sx_xunlock
697  * operation.  All 'easy case' failures are redirected to this.  Note
698  * that ideally this would be a static function, but it needs to be
699  * accessible from at least sx.h.
700  */
701 void
702 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
703 {
704         uintptr_t x;
705         int queue, wakeup_swapper;
706
707         if (SCHEDULER_STOPPED())
708                 return;
709
710         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
711
712         /* If the lock is recursed, then unrecurse one level. */
713         if (sx_xlocked(sx) && sx_recursed(sx)) {
714                 if ((--sx->sx_recurse) == 0)
715                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
716                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
717                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
718                 return;
719         }
720         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
721             SX_LOCK_EXCLUSIVE_WAITERS));
722         if (LOCK_LOG_TEST(&sx->lock_object, 0))
723                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
724
725         sleepq_lock(&sx->lock_object);
726         x = SX_LOCK_UNLOCKED;
727
728         /*
729          * The wake up algorithm here is quite simple and probably not
730          * ideal.  It gives precedence to shared waiters if they are
731          * present.  For this condition, we have to preserve the
732          * state of the exclusive waiters flag.
733          * If interruptible sleeps left the shared queue empty avoid a
734          * starvation for the threads sleeping on the exclusive queue by giving
735          * them precedence and cleaning up the shared waiters bit anyway.
736          */
737         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
738             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
739                 queue = SQ_SHARED_QUEUE;
740                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
741         } else
742                 queue = SQ_EXCLUSIVE_QUEUE;
743
744         /* Wake up all the waiters for the specific queue. */
745         if (LOCK_LOG_TEST(&sx->lock_object, 0))
746                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
747                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
748                     "exclusive");
749         atomic_store_rel_ptr(&sx->sx_lock, x);
750         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
751             queue);
752         sleepq_release(&sx->lock_object);
753         if (wakeup_swapper)
754                 kick_proc0();
755 }
756
757 /*
758  * This function represents the so-called 'hard case' for sx_slock
759  * operation.  All 'easy case' failures are redirected to this.  Note
760  * that ideally this would be a static function, but it needs to be
761  * accessible from at least sx.h.
762  */
763 int
764 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
765 {
766         GIANT_DECLARE;
767 #ifdef ADAPTIVE_SX
768         volatile struct thread *owner;
769 #endif
770 #ifdef LOCK_PROFILING
771         uint64_t waittime = 0;
772         int contested = 0;
773 #endif
774         uintptr_t x;
775         int error = 0;
776 #ifdef KDTRACE_HOOKS
777         uint64_t spin_cnt = 0;
778         uint64_t sleep_cnt = 0;
779         int64_t sleep_time = 0;
780 #endif
781
782         if (SCHEDULER_STOPPED())
783                 return (0);
784
785         /*
786          * As with rwlocks, we don't make any attempt to try to block
787          * shared locks once there is an exclusive waiter.
788          */
789         for (;;) {
790 #ifdef KDTRACE_HOOKS
791                 spin_cnt++;
792 #endif
793                 x = sx->sx_lock;
794
795                 /*
796                  * If no other thread has an exclusive lock then try to bump up
797                  * the count of sharers.  Since we have to preserve the state
798                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
799                  * shared lock loop back and retry.
800                  */
801                 if (x & SX_LOCK_SHARED) {
802                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
803                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
804                             x + SX_ONE_SHARER)) {
805                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
806                                         CTR4(KTR_LOCK,
807                                             "%s: %p succeed %p -> %p", __func__,
808                                             sx, (void *)x,
809                                             (void *)(x + SX_ONE_SHARER));
810                                 break;
811                         }
812                         continue;
813                 }
814                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
815                     &waittime);
816
817 #ifdef ADAPTIVE_SX
818                 /*
819                  * If the owner is running on another CPU, spin until
820                  * the owner stops running or the state of the lock
821                  * changes.
822                  */
823                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
824                         x = SX_OWNER(x);
825                         owner = (struct thread *)x;
826                         if (TD_IS_RUNNING(owner)) {
827                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
828                                         CTR3(KTR_LOCK,
829                                             "%s: spinning on %p held by %p",
830                                             __func__, sx, owner);
831                                 GIANT_SAVE();
832                                 while (SX_OWNER(sx->sx_lock) == x &&
833                                     TD_IS_RUNNING(owner)) {
834 #ifdef KDTRACE_HOOKS
835                                         spin_cnt++;
836 #endif
837                                         cpu_spinwait();
838                                 }
839                                 continue;
840                         }
841                 }
842 #endif
843
844                 /*
845                  * Some other thread already has an exclusive lock, so
846                  * start the process of blocking.
847                  */
848                 sleepq_lock(&sx->lock_object);
849                 x = sx->sx_lock;
850
851                 /*
852                  * The lock could have been released while we spun.
853                  * In this case loop back and retry.
854                  */
855                 if (x & SX_LOCK_SHARED) {
856                         sleepq_release(&sx->lock_object);
857                         continue;
858                 }
859
860 #ifdef ADAPTIVE_SX
861                 /*
862                  * If the owner is running on another CPU, spin until
863                  * the owner stops running or the state of the lock
864                  * changes.
865                  */
866                 if (!(x & SX_LOCK_SHARED) &&
867                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
868                         owner = (struct thread *)SX_OWNER(x);
869                         if (TD_IS_RUNNING(owner)) {
870                                 sleepq_release(&sx->lock_object);
871                                 continue;
872                         }
873                 }
874 #endif
875
876                 /*
877                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
878                  * fail to set it drop the sleep queue lock and loop
879                  * back.
880                  */
881                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
882                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
883                             x | SX_LOCK_SHARED_WAITERS)) {
884                                 sleepq_release(&sx->lock_object);
885                                 continue;
886                         }
887                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
888                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
889                                     __func__, sx);
890                 }
891
892                 /*
893                  * Since we have been unable to acquire the shared lock,
894                  * we have to sleep.
895                  */
896                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
897                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
898                             __func__, sx);
899
900 #ifdef KDTRACE_HOOKS
901                 sleep_time -= lockstat_nsecs();
902 #endif
903                 GIANT_SAVE();
904                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
905                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
906                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
907                 if (!(opts & SX_INTERRUPTIBLE))
908                         sleepq_wait(&sx->lock_object, 0);
909                 else
910                         error = sleepq_wait_sig(&sx->lock_object, 0);
911 #ifdef KDTRACE_HOOKS
912                 sleep_time += lockstat_nsecs();
913                 sleep_cnt++;
914 #endif
915                 if (error) {
916                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
917                                 CTR2(KTR_LOCK,
918                         "%s: interruptible sleep by %p suspended by signal",
919                                     __func__, sx);
920                         break;
921                 }
922                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
923                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
924                             __func__, sx);
925         }
926         if (error == 0)
927                 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx,
928                     contested, waittime, file, line);
929 #ifdef KDTRACE_HOOKS
930         if (sleep_time)
931                 LOCKSTAT_RECORD1(LS_SX_XLOCK_BLOCK, sx, sleep_time);
932         if (spin_cnt > sleep_cnt)
933                 LOCKSTAT_RECORD1(LS_SX_XLOCK_SPIN, sx, (spin_cnt - sleep_cnt));
934 #endif
935         GIANT_RESTORE();
936         return (error);
937 }
938
939 /*
940  * This function represents the so-called 'hard case' for sx_sunlock
941  * operation.  All 'easy case' failures are redirected to this.  Note
942  * that ideally this would be a static function, but it needs to be
943  * accessible from at least sx.h.
944  */
945 void
946 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
947 {
948         uintptr_t x;
949         int wakeup_swapper;
950
951         if (SCHEDULER_STOPPED())
952                 return;
953
954         for (;;) {
955                 x = sx->sx_lock;
956
957                 /*
958                  * We should never have sharers while at least one thread
959                  * holds a shared lock.
960                  */
961                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
962                     ("%s: waiting sharers", __func__));
963
964                 /*
965                  * See if there is more than one shared lock held.  If
966                  * so, just drop one and return.
967                  */
968                 if (SX_SHARERS(x) > 1) {
969                         if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
970                             x - SX_ONE_SHARER)) {
971                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
972                                         CTR4(KTR_LOCK,
973                                             "%s: %p succeeded %p -> %p",
974                                             __func__, sx, (void *)x,
975                                             (void *)(x - SX_ONE_SHARER));
976                                 break;
977                         }
978                         continue;
979                 }
980
981                 /*
982                  * If there aren't any waiters for an exclusive lock,
983                  * then try to drop it quickly.
984                  */
985                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
986                         MPASS(x == SX_SHARERS_LOCK(1));
987                         if (atomic_cmpset_rel_ptr(&sx->sx_lock,
988                             SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
989                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
990                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
991                                             __func__, sx);
992                                 break;
993                         }
994                         continue;
995                 }
996
997                 /*
998                  * At this point, there should just be one sharer with
999                  * exclusive waiters.
1000                  */
1001                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1002
1003                 sleepq_lock(&sx->lock_object);
1004
1005                 /*
1006                  * Wake up semantic here is quite simple:
1007                  * Just wake up all the exclusive waiters.
1008                  * Note that the state of the lock could have changed,
1009                  * so if it fails loop back and retry.
1010                  */
1011                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1012                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1013                     SX_LOCK_UNLOCKED)) {
1014                         sleepq_release(&sx->lock_object);
1015                         continue;
1016                 }
1017                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1018                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1019                             "exclusive queue", __func__, sx);
1020                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1021                     0, SQ_EXCLUSIVE_QUEUE);
1022                 sleepq_release(&sx->lock_object);
1023                 if (wakeup_swapper)
1024                         kick_proc0();
1025                 break;
1026         }
1027 }
1028
1029 #ifdef INVARIANT_SUPPORT
1030 #ifndef INVARIANTS
1031 #undef  _sx_assert
1032 #endif
1033
1034 /*
1035  * In the non-WITNESS case, sx_assert() can only detect that at least
1036  * *some* thread owns an slock, but it cannot guarantee that *this*
1037  * thread owns an slock.
1038  */
1039 void
1040 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1041 {
1042 #ifndef WITNESS
1043         int slocked = 0;
1044 #endif
1045
1046         if (panicstr != NULL)
1047                 return;
1048         switch (what) {
1049         case SA_SLOCKED:
1050         case SA_SLOCKED | SA_NOTRECURSED:
1051         case SA_SLOCKED | SA_RECURSED:
1052 #ifndef WITNESS
1053                 slocked = 1;
1054                 /* FALLTHROUGH */
1055 #endif
1056         case SA_LOCKED:
1057         case SA_LOCKED | SA_NOTRECURSED:
1058         case SA_LOCKED | SA_RECURSED:
1059 #ifdef WITNESS
1060                 witness_assert(&sx->lock_object, what, file, line);
1061 #else
1062                 /*
1063                  * If some other thread has an exclusive lock or we
1064                  * have one and are asserting a shared lock, fail.
1065                  * Also, if no one has a lock at all, fail.
1066                  */
1067                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1068                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1069                     sx_xholder(sx) != curthread)))
1070                         panic("Lock %s not %slocked @ %s:%d\n",
1071                             sx->lock_object.lo_name, slocked ? "share " : "",
1072                             file, line);
1073
1074                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1075                         if (sx_recursed(sx)) {
1076                                 if (what & SA_NOTRECURSED)
1077                                         panic("Lock %s recursed @ %s:%d\n",
1078                                             sx->lock_object.lo_name, file,
1079                                             line);
1080                         } else if (what & SA_RECURSED)
1081                                 panic("Lock %s not recursed @ %s:%d\n",
1082                                     sx->lock_object.lo_name, file, line);
1083                 }
1084 #endif
1085                 break;
1086         case SA_XLOCKED:
1087         case SA_XLOCKED | SA_NOTRECURSED:
1088         case SA_XLOCKED | SA_RECURSED:
1089                 if (sx_xholder(sx) != curthread)
1090                         panic("Lock %s not exclusively locked @ %s:%d\n",
1091                             sx->lock_object.lo_name, file, line);
1092                 if (sx_recursed(sx)) {
1093                         if (what & SA_NOTRECURSED)
1094                                 panic("Lock %s recursed @ %s:%d\n",
1095                                     sx->lock_object.lo_name, file, line);
1096                 } else if (what & SA_RECURSED)
1097                         panic("Lock %s not recursed @ %s:%d\n",
1098                             sx->lock_object.lo_name, file, line);
1099                 break;
1100         case SA_UNLOCKED:
1101 #ifdef WITNESS
1102                 witness_assert(&sx->lock_object, what, file, line);
1103 #else
1104                 /*
1105                  * If we hold an exclusve lock fail.  We can't
1106                  * reliably check to see if we hold a shared lock or
1107                  * not.
1108                  */
1109                 if (sx_xholder(sx) == curthread)
1110                         panic("Lock %s exclusively locked @ %s:%d\n",
1111                             sx->lock_object.lo_name, file, line);
1112 #endif
1113                 break;
1114         default:
1115                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1116                     line);
1117         }
1118 }
1119 #endif  /* INVARIANT_SUPPORT */
1120
1121 #ifdef DDB
1122 static void
1123 db_show_sx(const struct lock_object *lock)
1124 {
1125         struct thread *td;
1126         const struct sx *sx;
1127
1128         sx = (const struct sx *)lock;
1129
1130         db_printf(" state: ");
1131         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1132                 db_printf("UNLOCKED\n");
1133         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1134                 db_printf("DESTROYED\n");
1135                 return;
1136         } else if (sx->sx_lock & SX_LOCK_SHARED)
1137                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1138         else {
1139                 td = sx_xholder(sx);
1140                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1141                     td->td_tid, td->td_proc->p_pid, td->td_name);
1142                 if (sx_recursed(sx))
1143                         db_printf(" recursed: %d\n", sx->sx_recurse);
1144         }
1145
1146         db_printf(" waiters: ");
1147         switch(sx->sx_lock &
1148             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1149         case SX_LOCK_SHARED_WAITERS:
1150                 db_printf("shared\n");
1151                 break;
1152         case SX_LOCK_EXCLUSIVE_WAITERS:
1153                 db_printf("exclusive\n");
1154                 break;
1155         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1156                 db_printf("exclusive and shared\n");
1157                 break;
1158         default:
1159                 db_printf("none\n");
1160         }
1161 }
1162
1163 /*
1164  * Check to see if a thread that is blocked on a sleep queue is actually
1165  * blocked on an sx lock.  If so, output some details and return true.
1166  * If the lock has an exclusive owner, return that in *ownerp.
1167  */
1168 int
1169 sx_chain(struct thread *td, struct thread **ownerp)
1170 {
1171         struct sx *sx;
1172
1173         /*
1174          * Check to see if this thread is blocked on an sx lock.
1175          * First, we check the lock class.  If that is ok, then we
1176          * compare the lock name against the wait message.
1177          */
1178         sx = td->td_wchan;
1179         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1180             sx->lock_object.lo_name != td->td_wmesg)
1181                 return (0);
1182
1183         /* We think we have an sx lock, so output some details. */
1184         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1185         *ownerp = sx_xholder(sx);
1186         if (sx->sx_lock & SX_LOCK_SHARED)
1187                 db_printf("SLOCK (count %ju)\n",
1188                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1189         else
1190                 db_printf("XLOCK\n");
1191         return (1);
1192 }
1193 #endif