1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
5 * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice(s), this list of conditions and the following disclaimer as
13 * the first lines of this file unmodified other than the possible
14 * addition of one or more copyright notices.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice(s), this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
20 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
23 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
24 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
25 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
26 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
29 * DAMAGE.
30 */
31
32 /*
33 * Shared/exclusive locks. This implementation attempts to ensure
34 * deterministic lock granting behavior, so that slocks and xlocks are
35 * interleaved.
36 *
37 * Priority propagation will not generally raise the priority of lock holders,
38 * so should not be relied upon in combination with sx locks.
39 */
40
41 #include "opt_ddb.h"
42 #include "opt_hwpmc_hooks.h"
43 #include "opt_no_adaptive_sx.h"
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kdb.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/sched.h>
54 #include <sys/sleepqueue.h>
55 #include <sys/sx.h>
56 #include <sys/smp.h>
57 #include <sys/sysctl.h>
58
59 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
60 #include <machine/cpu.h>
61 #endif
62
63 #ifdef DDB
64 #include <ddb/ddb.h>
65 #endif
66
67 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
68 #define ADAPTIVE_SX
69 #endif
70
71 #ifdef HWPMC_HOOKS
72 #include <sys/pmckern.h>
73 PMC_SOFT_DECLARE( , , lock, failed);
74 #endif
75
76 /* Handy macros for sleep queues. */
77 #define SQ_EXCLUSIVE_QUEUE 0
78 #define SQ_SHARED_QUEUE 1
79
80 /*
81 * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We
82 * drop Giant anytime we have to sleep or if we adaptively spin.
83 */
84 #define GIANT_DECLARE \
85 int _giantcnt = 0; \
86 WITNESS_SAVE_DECL(Giant) \
87
88 #define GIANT_SAVE(work) do { \
89 if (__predict_false(mtx_owned(&Giant))) { \
90 work++; \
91 WITNESS_SAVE(&Giant.lock_object, Giant); \
92 while (mtx_owned(&Giant)) { \
93 _giantcnt++; \
94 mtx_unlock(&Giant); \
95 } \
96 } \
97 } while (0)
98
99 #define GIANT_RESTORE() do { \
100 if (_giantcnt > 0) { \
101 mtx_assert(&Giant, MA_NOTOWNED); \
102 while (_giantcnt--) \
103 mtx_lock(&Giant); \
104 WITNESS_RESTORE(&Giant.lock_object, Giant); \
105 } \
106 } while (0)
107
108 /*
109 * Returns true if an exclusive lock is recursed. It assumes
110 * curthread currently has an exclusive lock.
111 */
112 #define sx_recursed(sx) ((sx)->sx_recurse != 0)
113
114 static void assert_sx(const struct lock_object *lock, int what);
115 #ifdef DDB
116 static void db_show_sx(const struct lock_object *lock);
117 #endif
118 static void lock_sx(struct lock_object *lock, uintptr_t how);
119 #ifdef KDTRACE_HOOKS
120 static int owner_sx(const struct lock_object *lock, struct thread **owner);
121 #endif
122 static uintptr_t unlock_sx(struct lock_object *lock);
123
124 struct lock_class lock_class_sx = {
125 .lc_name = "sx",
126 .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
127 .lc_assert = assert_sx,
128 #ifdef DDB
129 .lc_ddb_show = db_show_sx,
130 #endif
131 .lc_lock = lock_sx,
132 .lc_unlock = unlock_sx,
133 #ifdef KDTRACE_HOOKS
134 .lc_owner = owner_sx,
135 #endif
136 };
137
138 #ifndef INVARIANTS
139 #define _sx_assert(sx, what, file, line)
140 #endif
141
142 #ifdef ADAPTIVE_SX
143 #ifdef SX_CUSTOM_BACKOFF
144 static u_short __read_frequently asx_retries;
145 static u_short __read_frequently asx_loops;
146 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
147 "sxlock debugging");
148 SYSCTL_U16(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
149 SYSCTL_U16(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
150
151 static struct lock_delay_config __read_frequently sx_delay;
152
153 SYSCTL_U16(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
154 0, "");
155 SYSCTL_U16(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
156 0, "");
157
158 static void
sx_lock_delay_init(void * arg __unused)159 sx_lock_delay_init(void *arg __unused)
160 {
161
162 lock_delay_default_init(&sx_delay);
163 asx_retries = 10;
164 asx_loops = max(10000, sx_delay.max);
165 }
166 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
167 #else
168 #define sx_delay locks_delay
169 #define asx_retries locks_delay_retries
170 #define asx_loops locks_delay_loops
171 #endif
172 #endif
173
174 void
assert_sx(const struct lock_object * lock,int what)175 assert_sx(const struct lock_object *lock, int what)
176 {
177
178 sx_assert((const struct sx *)lock, what);
179 }
180
181 void
lock_sx(struct lock_object * lock,uintptr_t how)182 lock_sx(struct lock_object *lock, uintptr_t how)
183 {
184 struct sx *sx;
185
186 sx = (struct sx *)lock;
187 if (how)
188 sx_slock(sx);
189 else
190 sx_xlock(sx);
191 }
192
193 uintptr_t
unlock_sx(struct lock_object * lock)194 unlock_sx(struct lock_object *lock)
195 {
196 struct sx *sx;
197
198 sx = (struct sx *)lock;
199 sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
200 if (sx_xlocked(sx)) {
201 sx_xunlock(sx);
202 return (0);
203 } else {
204 sx_sunlock(sx);
205 return (1);
206 }
207 }
208
209 #ifdef KDTRACE_HOOKS
210 int
owner_sx(const struct lock_object * lock,struct thread ** owner)211 owner_sx(const struct lock_object *lock, struct thread **owner)
212 {
213 const struct sx *sx;
214 uintptr_t x;
215
216 sx = (const struct sx *)lock;
217 x = sx->sx_lock;
218 *owner = NULL;
219 return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
220 ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
221 }
222 #endif
223
224 void
sx_sysinit(void * arg)225 sx_sysinit(void *arg)
226 {
227 struct sx_args *sargs = arg;
228
229 sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
230 }
231
232 void
sx_init_flags(struct sx * sx,const char * description,int opts)233 sx_init_flags(struct sx *sx, const char *description, int opts)
234 {
235 int flags;
236
237 MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
238 SX_NOPROFILE | SX_NEW)) == 0);
239 ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
240 ("%s: sx_lock not aligned for %s: %p", __func__, description,
241 &sx->sx_lock));
242
243 flags = LO_SLEEPABLE | LO_UPGRADABLE;
244 if (opts & SX_DUPOK)
245 flags |= LO_DUPOK;
246 if (opts & SX_NOPROFILE)
247 flags |= LO_NOPROFILE;
248 if (!(opts & SX_NOWITNESS))
249 flags |= LO_WITNESS;
250 if (opts & SX_RECURSE)
251 flags |= LO_RECURSABLE;
252 if (opts & SX_QUIET)
253 flags |= LO_QUIET;
254 if (opts & SX_NEW)
255 flags |= LO_NEW;
256
257 lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
258 sx->sx_lock = SX_LOCK_UNLOCKED;
259 sx->sx_recurse = 0;
260 }
261
262 void
sx_destroy(struct sx * sx)263 sx_destroy(struct sx *sx)
264 {
265
266 KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
267 KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
268 sx->sx_lock = SX_LOCK_DESTROYED;
269 lock_destroy(&sx->lock_object);
270 }
271
272 int
sx_try_slock_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)273 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
274 {
275 uintptr_t x;
276
277 if (SCHEDULER_STOPPED())
278 return (1);
279
280 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
281 ("sx_try_slock() by idle thread %p on sx %p @ %s:%d",
282 curthread, sx, file, line));
283
284 x = sx->sx_lock;
285 for (;;) {
286 KASSERT(x != SX_LOCK_DESTROYED,
287 ("sx_try_slock() of destroyed sx %p @ %s:%d", sx, file,
288 line));
289 if (!(x & SX_LOCK_SHARED))
290 break;
291 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
292 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
293 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
294 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
295 sx, 0, 0, file, line, LOCKSTAT_READER);
296 TD_LOCKS_INC(curthread);
297 curthread->td_sx_slocks++;
298 return (1);
299 }
300 }
301
302 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
303 return (0);
304 }
305
306 int
sx_try_slock_(struct sx * sx,const char * file,int line)307 sx_try_slock_(struct sx *sx, const char *file, int line)
308 {
309
310 return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
311 }
312
313 int
_sx_xlock(struct sx * sx,int opts,const char * file,int line)314 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
315 {
316 uintptr_t tid, x;
317 int error = 0;
318
319 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
320 !TD_IS_IDLETHREAD(curthread),
321 ("sx_xlock() by idle thread %p on sx %p @ %s:%d",
322 curthread, sx, file, line));
323 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
324 ("sx_xlock() of destroyed sx %p @ %s:%d", sx, file, line));
325 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
326 line, NULL);
327 tid = (uintptr_t)curthread;
328 x = SX_LOCK_UNLOCKED;
329 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
330 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
331 else
332 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
333 0, 0, file, line, LOCKSTAT_WRITER);
334 if (!error) {
335 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
336 file, line);
337 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
338 TD_LOCKS_INC(curthread);
339 }
340
341 return (error);
342 }
343
344 int
sx_try_xlock_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)345 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
346 {
347 struct thread *td;
348 uintptr_t tid, x;
349 int rval;
350 bool recursed;
351
352 td = curthread;
353 tid = (uintptr_t)td;
354 if (SCHEDULER_STOPPED())
355 return (1);
356
357 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
358 ("sx_try_xlock() by idle thread %p on sx %p @ %s:%d",
359 curthread, sx, file, line));
360 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
361 ("sx_try_xlock() of destroyed sx %p @ %s:%d", sx, file, line));
362
363 rval = 1;
364 recursed = false;
365 x = SX_LOCK_UNLOCKED;
366 for (;;) {
367 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
368 break;
369 if (x == SX_LOCK_UNLOCKED)
370 continue;
371 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
372 sx->sx_recurse++;
373 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
374 break;
375 }
376 rval = 0;
377 break;
378 }
379
380 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
381 if (rval) {
382 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
383 file, line);
384 if (!recursed)
385 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
386 sx, 0, 0, file, line, LOCKSTAT_WRITER);
387 TD_LOCKS_INC(curthread);
388 }
389
390 return (rval);
391 }
392
393 int
sx_try_xlock_(struct sx * sx,const char * file,int line)394 sx_try_xlock_(struct sx *sx, const char *file, int line)
395 {
396
397 return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
398 }
399
400 void
_sx_xunlock(struct sx * sx,const char * file,int line)401 _sx_xunlock(struct sx *sx, const char *file, int line)
402 {
403
404 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
405 ("sx_xunlock() of destroyed sx %p @ %s:%d", sx, file, line));
406 _sx_assert(sx, SA_XLOCKED, file, line);
407 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
408 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
409 line);
410 #if LOCK_DEBUG > 0
411 _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
412 #else
413 __sx_xunlock(sx, curthread, file, line);
414 #endif
415 TD_LOCKS_DEC(curthread);
416 }
417
418 /*
419 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
420 * This will only succeed if this thread holds a single shared lock.
421 * Return 1 if if the upgrade succeed, 0 otherwise.
422 */
423 int
sx_try_upgrade_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)424 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
425 {
426 uintptr_t x;
427 uintptr_t waiters;
428 int success;
429
430 if (SCHEDULER_STOPPED())
431 return (1);
432
433 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
434 ("sx_try_upgrade() of destroyed sx %p @ %s:%d", sx, file, line));
435 _sx_assert(sx, SA_SLOCKED, file, line);
436
437 /*
438 * Try to switch from one shared lock to an exclusive lock. We need
439 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
440 * we will wake up the exclusive waiters when we drop the lock.
441 */
442 success = 0;
443 x = SX_READ_VALUE(sx);
444 for (;;) {
445 if (SX_SHARERS(x) > 1)
446 break;
447 waiters = (x & SX_LOCK_WAITERS);
448 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
449 (uintptr_t)curthread | waiters)) {
450 success = 1;
451 break;
452 }
453 }
454 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
455 if (success) {
456 curthread->td_sx_slocks--;
457 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
458 file, line);
459 LOCKSTAT_RECORD0(sx__upgrade, sx);
460 }
461 return (success);
462 }
463
464 int
sx_try_upgrade_(struct sx * sx,const char * file,int line)465 sx_try_upgrade_(struct sx *sx, const char *file, int line)
466 {
467
468 return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
469 }
470
471 /*
472 * Downgrade an unrecursed exclusive lock into a single shared lock.
473 */
474 void
sx_downgrade_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)475 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
476 {
477 uintptr_t x;
478
479 if (SCHEDULER_STOPPED())
480 return;
481
482 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
483 ("sx_downgrade() of destroyed sx %p @ %s:%d", sx, file, line));
484 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
485 #ifndef INVARIANTS
486 if (sx_recursed(sx))
487 panic("downgrade of a recursed lock");
488 #endif
489
490 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
491
492 /*
493 * Try to switch from an exclusive lock with no shared waiters
494 * to one sharer with no shared waiters. If there are
495 * exclusive waiters, we don't need to lock the sleep queue so
496 * long as we preserve the flag. We do one quick try and if
497 * that fails we grab the sleepq lock to keep the flags from
498 * changing and do it the slow way.
499 *
500 * We have to lock the sleep queue if there are shared waiters
501 * so we can wake them up.
502 */
503 x = sx->sx_lock;
504 if (!(x & SX_LOCK_SHARED_WAITERS) &&
505 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
506 (x & SX_LOCK_EXCLUSIVE_WAITERS)))
507 goto out;
508
509 /*
510 * Lock the sleep queue so we can read the waiters bits
511 * without any races and wakeup any shared waiters.
512 */
513 sleepq_lock(&sx->lock_object);
514
515 /*
516 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
517 * shared lock. If there are any shared waiters, wake them up.
518 */
519 x = sx->sx_lock;
520 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
521 (x & SX_LOCK_EXCLUSIVE_WAITERS));
522 if (x & SX_LOCK_SHARED_WAITERS)
523 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
524 SQ_SHARED_QUEUE);
525 sleepq_release(&sx->lock_object);
526
527 out:
528 curthread->td_sx_slocks++;
529 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
530 LOCKSTAT_RECORD0(sx__downgrade, sx);
531 }
532
533 void
sx_downgrade_(struct sx * sx,const char * file,int line)534 sx_downgrade_(struct sx *sx, const char *file, int line)
535 {
536
537 sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
538 }
539
540 #ifdef ADAPTIVE_SX
541 static inline void
sx_drop_critical(uintptr_t x,bool * in_critical,int * extra_work)542 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
543 {
544
545 if (x & SX_LOCK_WRITE_SPINNER)
546 return;
547 if (*in_critical) {
548 critical_exit();
549 *in_critical = false;
550 (*extra_work)--;
551 }
552 }
553 #else
554 #define sx_drop_critical(x, in_critical, extra_work) do { } while (0)
555 #endif
556
557 /*
558 * This function represents the so-called 'hard case' for sx_xlock
559 * operation. All 'easy case' failures are redirected to this. Note
560 * that ideally this would be a static function, but it needs to be
561 * accessible from at least sx.h.
562 */
563 int
_sx_xlock_hard(struct sx * sx,uintptr_t x,int opts LOCK_FILE_LINE_ARG_DEF)564 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
565 {
566 GIANT_DECLARE;
567 uintptr_t tid, setx;
568 #ifdef ADAPTIVE_SX
569 struct thread *owner;
570 u_int i, n, spintries = 0;
571 enum { READERS, WRITER } sleep_reason = READERS;
572 bool in_critical = false;
573 #endif
574 #ifdef LOCK_PROFILING
575 uint64_t waittime = 0;
576 int contested = 0;
577 #endif
578 int error = 0;
579 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
580 struct lock_delay_arg lda;
581 #endif
582 #ifdef KDTRACE_HOOKS
583 u_int sleep_cnt = 0;
584 int64_t sleep_time = 0;
585 int64_t all_time = 0;
586 uintptr_t state = 0;
587 #endif
588 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
589 int doing_lockprof = 0;
590 #endif
591 int extra_work = 0;
592
593 tid = (uintptr_t)curthread;
594
595 #ifdef KDTRACE_HOOKS
596 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
597 while (x == SX_LOCK_UNLOCKED) {
598 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
599 goto out_lockstat;
600 }
601 extra_work = 1;
602 doing_lockprof = 1;
603 all_time -= lockstat_nsecs(&sx->lock_object);
604 }
605 state = x;
606 #endif
607 #ifdef LOCK_PROFILING
608 extra_work = 1;
609 doing_lockprof = 1;
610 #endif
611
612 if (SCHEDULER_STOPPED())
613 return (0);
614
615 if (__predict_false(x == SX_LOCK_UNLOCKED))
616 x = SX_READ_VALUE(sx);
617
618 /* If we already hold an exclusive lock, then recurse. */
619 if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
620 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
621 ("_sx_xlock_hard: recursed on non-recursive sx %p @ %s:%d\n",
622 sx, file, line));
623 sx->sx_recurse++;
624 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
625 if (LOCK_LOG_TEST(&sx->lock_object, 0))
626 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
627 return (0);
628 }
629
630 if (LOCK_LOG_TEST(&sx->lock_object, 0))
631 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
632 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
633
634 #if defined(ADAPTIVE_SX)
635 lock_delay_arg_init(&lda, &sx_delay);
636 #elif defined(KDTRACE_HOOKS)
637 lock_delay_arg_init_noadapt(&lda);
638 #endif
639
640 #ifdef HWPMC_HOOKS
641 PMC_SOFT_CALL( , , lock, failed);
642 #endif
643 lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
644 &waittime);
645
646 #ifndef INVARIANTS
647 GIANT_SAVE(extra_work);
648 #endif
649
650 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
651
652 for (;;) {
653 if (x == SX_LOCK_UNLOCKED) {
654 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
655 break;
656 continue;
657 }
658 #ifdef INVARIANTS
659 GIANT_SAVE(extra_work);
660 #endif
661 #ifdef KDTRACE_HOOKS
662 lda.spin_cnt++;
663 #endif
664 #ifdef ADAPTIVE_SX
665 if (x == (SX_LOCK_SHARED | SX_LOCK_WRITE_SPINNER)) {
666 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
667 break;
668 continue;
669 }
670
671 /*
672 * If the lock is write locked and the owner is
673 * running on another CPU, spin until the owner stops
674 * running or the state of the lock changes.
675 */
676 if ((x & SX_LOCK_SHARED) == 0) {
677 sx_drop_critical(x, &in_critical, &extra_work);
678 sleep_reason = WRITER;
679 owner = lv_sx_owner(x);
680 if (!TD_IS_RUNNING(owner))
681 goto sleepq;
682 if (LOCK_LOG_TEST(&sx->lock_object, 0))
683 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
684 __func__, sx, owner);
685 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
686 "spinning", "lockname:\"%s\"",
687 sx->lock_object.lo_name);
688 do {
689 lock_delay(&lda);
690 x = SX_READ_VALUE(sx);
691 owner = lv_sx_owner(x);
692 } while (owner != NULL && TD_IS_RUNNING(owner));
693 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
694 "running");
695 continue;
696 } else if (SX_SHARERS(x) > 0) {
697 sleep_reason = READERS;
698 if (spintries == asx_retries)
699 goto sleepq;
700 if (!(x & SX_LOCK_WRITE_SPINNER)) {
701 if (!in_critical) {
702 critical_enter();
703 in_critical = true;
704 extra_work++;
705 }
706 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
707 x | SX_LOCK_WRITE_SPINNER)) {
708 critical_exit();
709 in_critical = false;
710 extra_work--;
711 continue;
712 }
713 }
714 spintries++;
715 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
716 "spinning", "lockname:\"%s\"",
717 sx->lock_object.lo_name);
718 n = SX_SHARERS(x);
719 for (i = 0; i < asx_loops; i += n) {
720 lock_delay_spin(n);
721 x = SX_READ_VALUE(sx);
722 if (!(x & SX_LOCK_WRITE_SPINNER))
723 break;
724 if (!(x & SX_LOCK_SHARED))
725 break;
726 n = SX_SHARERS(x);
727 if (n == 0)
728 break;
729 }
730 #ifdef KDTRACE_HOOKS
731 lda.spin_cnt += i;
732 #endif
733 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
734 "running");
735 if (i < asx_loops)
736 continue;
737 }
738 sleepq:
739 #endif
740 sleepq_lock(&sx->lock_object);
741 x = SX_READ_VALUE(sx);
742 retry_sleepq:
743
744 /*
745 * If the lock was released while spinning on the
746 * sleep queue chain lock, try again.
747 */
748 if (x == SX_LOCK_UNLOCKED) {
749 sleepq_release(&sx->lock_object);
750 sx_drop_critical(x, &in_critical, &extra_work);
751 continue;
752 }
753
754 #ifdef ADAPTIVE_SX
755 /*
756 * The current lock owner might have started executing
757 * on another CPU (or the lock could have changed
758 * owners) while we were waiting on the sleep queue
759 * chain lock. If so, drop the sleep queue lock and try
760 * again.
761 */
762 if (!(x & SX_LOCK_SHARED)) {
763 owner = (struct thread *)SX_OWNER(x);
764 if (TD_IS_RUNNING(owner)) {
765 sleepq_release(&sx->lock_object);
766 sx_drop_critical(x, &in_critical,
767 &extra_work);
768 continue;
769 }
770 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
771 sleepq_release(&sx->lock_object);
772 sx_drop_critical(x, &in_critical, &extra_work);
773 continue;
774 }
775 #endif
776
777 /*
778 * If an exclusive lock was released with both shared
779 * and exclusive waiters and a shared waiter hasn't
780 * woken up and acquired the lock yet, sx_lock will be
781 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
782 * If we see that value, try to acquire it once. Note
783 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
784 * as there are other exclusive waiters still. If we
785 * fail, restart the loop.
786 */
787 setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
788 if ((x & ~setx) == SX_LOCK_SHARED) {
789 setx &= ~SX_LOCK_WRITE_SPINNER;
790 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
791 goto retry_sleepq;
792 sleepq_release(&sx->lock_object);
793 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
794 __func__, sx);
795 break;
796 }
797
798 #ifdef ADAPTIVE_SX
799 /*
800 * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
801 * It is an invariant that when the bit is set, there is
802 * a writer ready to grab the lock. Thus clear the bit since
803 * we are going to sleep.
804 */
805 if (in_critical) {
806 if ((x & SX_LOCK_WRITE_SPINNER) ||
807 !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
808 setx = x & ~SX_LOCK_WRITE_SPINNER;
809 setx |= SX_LOCK_EXCLUSIVE_WAITERS;
810 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
811 setx)) {
812 goto retry_sleepq;
813 }
814 }
815 critical_exit();
816 in_critical = false;
817 } else {
818 #endif
819 /*
820 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
821 * than loop back and retry.
822 */
823 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
824 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
825 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
826 goto retry_sleepq;
827 }
828 if (LOCK_LOG_TEST(&sx->lock_object, 0))
829 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
830 __func__, sx);
831 }
832 #ifdef ADAPTIVE_SX
833 }
834 #endif
835
836 /*
837 * Since we have been unable to acquire the exclusive
838 * lock and the exclusive waiters flag is set, we have
839 * to sleep.
840 */
841 if (LOCK_LOG_TEST(&sx->lock_object, 0))
842 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
843 __func__, sx);
844
845 #ifdef KDTRACE_HOOKS
846 sleep_time -= lockstat_nsecs(&sx->lock_object);
847 #endif
848 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
849 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
850 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
851 /*
852 * Hack: this can land in thread_suspend_check which will
853 * conditionally take a mutex, tripping over an assert if a
854 * lock we are waiting for is set.
855 */
856 THREAD_CONTENTION_DONE(&sx->lock_object);
857 if (!(opts & SX_INTERRUPTIBLE))
858 sleepq_wait(&sx->lock_object, 0);
859 else
860 error = sleepq_wait_sig(&sx->lock_object, 0);
861 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
862 #ifdef KDTRACE_HOOKS
863 sleep_time += lockstat_nsecs(&sx->lock_object);
864 sleep_cnt++;
865 #endif
866 if (error) {
867 if (LOCK_LOG_TEST(&sx->lock_object, 0))
868 CTR2(KTR_LOCK,
869 "%s: interruptible sleep by %p suspended by signal",
870 __func__, sx);
871 break;
872 }
873 if (LOCK_LOG_TEST(&sx->lock_object, 0))
874 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
875 __func__, sx);
876 x = SX_READ_VALUE(sx);
877 }
878 THREAD_CONTENTION_DONE(&sx->lock_object);
879 if (__predict_true(!extra_work))
880 return (error);
881 #ifdef ADAPTIVE_SX
882 if (in_critical)
883 critical_exit();
884 #endif
885 GIANT_RESTORE();
886 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
887 if (__predict_true(!doing_lockprof))
888 return (error);
889 #endif
890 #ifdef KDTRACE_HOOKS
891 all_time += lockstat_nsecs(&sx->lock_object);
892 if (sleep_time)
893 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
894 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
895 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
896 if (lda.spin_cnt > sleep_cnt)
897 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
898 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
899 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
900 out_lockstat:
901 #endif
902 if (!error)
903 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
904 contested, waittime, file, line, LOCKSTAT_WRITER);
905 return (error);
906 }
907
908 /*
909 * This function represents the so-called 'hard case' for sx_xunlock
910 * operation. All 'easy case' failures are redirected to this. Note
911 * that ideally this would be a static function, but it needs to be
912 * accessible from at least sx.h.
913 */
914 void
_sx_xunlock_hard(struct sx * sx,uintptr_t x LOCK_FILE_LINE_ARG_DEF)915 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
916 {
917 uintptr_t tid, setx;
918 int queue;
919
920 if (SCHEDULER_STOPPED())
921 return;
922
923 tid = (uintptr_t)curthread;
924
925 if (__predict_false(x == tid))
926 x = SX_READ_VALUE(sx);
927
928 MPASS(!(x & SX_LOCK_SHARED));
929
930 if (__predict_false(x & SX_LOCK_RECURSED)) {
931 /* The lock is recursed, unrecurse one level. */
932 if ((--sx->sx_recurse) == 0)
933 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
934 if (LOCK_LOG_TEST(&sx->lock_object, 0))
935 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
936 return;
937 }
938
939 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
940 if (x == tid &&
941 atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
942 return;
943
944 if (LOCK_LOG_TEST(&sx->lock_object, 0))
945 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
946
947 sleepq_lock(&sx->lock_object);
948 x = SX_READ_VALUE(sx);
949 MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
950
951 /*
952 * The wake up algorithm here is quite simple and probably not
953 * ideal. It gives precedence to shared waiters if they are
954 * present. For this condition, we have to preserve the
955 * state of the exclusive waiters flag.
956 * If interruptible sleeps left the shared queue empty avoid a
957 * starvation for the threads sleeping on the exclusive queue by giving
958 * them precedence and cleaning up the shared waiters bit anyway.
959 */
960 setx = SX_LOCK_UNLOCKED;
961 queue = SQ_SHARED_QUEUE;
962 if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
963 sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
964 queue = SQ_EXCLUSIVE_QUEUE;
965 setx |= (x & SX_LOCK_SHARED_WAITERS);
966 }
967 atomic_store_rel_ptr(&sx->sx_lock, setx);
968
969 /* Wake up all the waiters for the specific queue. */
970 if (LOCK_LOG_TEST(&sx->lock_object, 0))
971 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
972 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
973 "exclusive");
974
975 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, queue);
976 sleepq_release(&sx->lock_object);
977 }
978
979 static __always_inline bool
__sx_can_read(struct thread * td,uintptr_t x,bool fp)980 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
981 {
982
983 if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
984 == SX_LOCK_SHARED)
985 return (true);
986 if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
987 return (true);
988 return (false);
989 }
990
991 static __always_inline bool
__sx_slock_try(struct sx * sx,struct thread * td,uintptr_t * xp,bool fp LOCK_FILE_LINE_ARG_DEF)992 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
993 LOCK_FILE_LINE_ARG_DEF)
994 {
995
996 /*
997 * If no other thread has an exclusive lock then try to bump up
998 * the count of sharers. Since we have to preserve the state
999 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
1000 * shared lock loop back and retry.
1001 */
1002 while (__sx_can_read(td, *xp, fp)) {
1003 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1004 *xp + SX_ONE_SHARER)) {
1005 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1006 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1007 __func__, sx, (void *)*xp,
1008 (void *)(*xp + SX_ONE_SHARER));
1009 td->td_sx_slocks++;
1010 return (true);
1011 }
1012 }
1013 return (false);
1014 }
1015
1016 static int __noinline
_sx_slock_hard(struct sx * sx,int opts,uintptr_t x LOCK_FILE_LINE_ARG_DEF)1017 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1018 {
1019 GIANT_DECLARE;
1020 struct thread *td;
1021 #ifdef ADAPTIVE_SX
1022 struct thread *owner;
1023 u_int i, n, spintries = 0;
1024 #endif
1025 #ifdef LOCK_PROFILING
1026 uint64_t waittime = 0;
1027 int contested = 0;
1028 #endif
1029 int error = 0;
1030 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1031 struct lock_delay_arg lda;
1032 #endif
1033 #ifdef KDTRACE_HOOKS
1034 u_int sleep_cnt = 0;
1035 int64_t sleep_time = 0;
1036 int64_t all_time = 0;
1037 uintptr_t state = 0;
1038 #endif
1039 int extra_work __sdt_used = 0;
1040
1041 td = curthread;
1042
1043 #ifdef KDTRACE_HOOKS
1044 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1045 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1046 goto out_lockstat;
1047 extra_work = 1;
1048 all_time -= lockstat_nsecs(&sx->lock_object);
1049 }
1050 state = x;
1051 #endif
1052 #ifdef LOCK_PROFILING
1053 extra_work = 1;
1054 #endif
1055
1056 if (SCHEDULER_STOPPED())
1057 return (0);
1058
1059 #if defined(ADAPTIVE_SX)
1060 lock_delay_arg_init(&lda, &sx_delay);
1061 #elif defined(KDTRACE_HOOKS)
1062 lock_delay_arg_init_noadapt(&lda);
1063 #endif
1064
1065 #ifdef HWPMC_HOOKS
1066 PMC_SOFT_CALL( , , lock, failed);
1067 #endif
1068 lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
1069 &waittime);
1070
1071 #ifndef INVARIANTS
1072 GIANT_SAVE(extra_work);
1073 #endif
1074
1075 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1076
1077 /*
1078 * As with rwlocks, we don't make any attempt to try to block
1079 * shared locks once there is an exclusive waiter.
1080 */
1081 for (;;) {
1082 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1083 break;
1084 #ifdef INVARIANTS
1085 GIANT_SAVE(extra_work);
1086 #endif
1087 #ifdef KDTRACE_HOOKS
1088 lda.spin_cnt++;
1089 #endif
1090
1091 #ifdef ADAPTIVE_SX
1092 /*
1093 * If the owner is running on another CPU, spin until
1094 * the owner stops running or the state of the lock
1095 * changes.
1096 */
1097 if ((x & SX_LOCK_SHARED) == 0) {
1098 owner = lv_sx_owner(x);
1099 if (TD_IS_RUNNING(owner)) {
1100 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1101 CTR3(KTR_LOCK,
1102 "%s: spinning on %p held by %p",
1103 __func__, sx, owner);
1104 KTR_STATE1(KTR_SCHED, "thread",
1105 sched_tdname(curthread), "spinning",
1106 "lockname:\"%s\"", sx->lock_object.lo_name);
1107 do {
1108 lock_delay(&lda);
1109 x = SX_READ_VALUE(sx);
1110 owner = lv_sx_owner(x);
1111 } while (owner != NULL && TD_IS_RUNNING(owner));
1112 KTR_STATE0(KTR_SCHED, "thread",
1113 sched_tdname(curthread), "running");
1114 continue;
1115 }
1116 } else {
1117 if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1118 MPASS(!__sx_can_read(td, x, false));
1119 lock_delay_spin(2);
1120 x = SX_READ_VALUE(sx);
1121 continue;
1122 }
1123 if (spintries < asx_retries) {
1124 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1125 "spinning", "lockname:\"%s\"",
1126 sx->lock_object.lo_name);
1127 n = SX_SHARERS(x);
1128 for (i = 0; i < asx_loops; i += n) {
1129 lock_delay_spin(n);
1130 x = SX_READ_VALUE(sx);
1131 if (!(x & SX_LOCK_SHARED))
1132 break;
1133 n = SX_SHARERS(x);
1134 if (n == 0)
1135 break;
1136 if (__sx_can_read(td, x, false))
1137 break;
1138 }
1139 #ifdef KDTRACE_HOOKS
1140 lda.spin_cnt += i;
1141 #endif
1142 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1143 "running");
1144 if (i < asx_loops)
1145 continue;
1146 }
1147 }
1148 #endif
1149
1150 /*
1151 * Some other thread already has an exclusive lock, so
1152 * start the process of blocking.
1153 */
1154 sleepq_lock(&sx->lock_object);
1155 x = SX_READ_VALUE(sx);
1156 retry_sleepq:
1157 if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1158 __sx_can_read(td, x, false)) {
1159 sleepq_release(&sx->lock_object);
1160 continue;
1161 }
1162
1163 #ifdef ADAPTIVE_SX
1164 /*
1165 * If the owner is running on another CPU, spin until
1166 * the owner stops running or the state of the lock
1167 * changes.
1168 */
1169 if (!(x & SX_LOCK_SHARED)) {
1170 owner = (struct thread *)SX_OWNER(x);
1171 if (TD_IS_RUNNING(owner)) {
1172 sleepq_release(&sx->lock_object);
1173 x = SX_READ_VALUE(sx);
1174 continue;
1175 }
1176 }
1177 #endif
1178
1179 /*
1180 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
1181 * fail to set it drop the sleep queue lock and loop
1182 * back.
1183 */
1184 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1185 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1186 x | SX_LOCK_SHARED_WAITERS))
1187 goto retry_sleepq;
1188 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1189 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1190 __func__, sx);
1191 }
1192
1193 /*
1194 * Since we have been unable to acquire the shared lock,
1195 * we have to sleep.
1196 */
1197 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1198 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1199 __func__, sx);
1200
1201 #ifdef KDTRACE_HOOKS
1202 sleep_time -= lockstat_nsecs(&sx->lock_object);
1203 #endif
1204 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1205 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1206 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1207 /*
1208 * Hack: this can land in thread_suspend_check which will
1209 * conditionally take a mutex, tripping over an assert if a
1210 * lock we are waiting for is set.
1211 */
1212 THREAD_CONTENTION_DONE(&sx->lock_object);
1213 if (!(opts & SX_INTERRUPTIBLE))
1214 sleepq_wait(&sx->lock_object, 0);
1215 else
1216 error = sleepq_wait_sig(&sx->lock_object, 0);
1217 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1218 #ifdef KDTRACE_HOOKS
1219 sleep_time += lockstat_nsecs(&sx->lock_object);
1220 sleep_cnt++;
1221 #endif
1222 if (error) {
1223 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1224 CTR2(KTR_LOCK,
1225 "%s: interruptible sleep by %p suspended by signal",
1226 __func__, sx);
1227 break;
1228 }
1229 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1230 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1231 __func__, sx);
1232 x = SX_READ_VALUE(sx);
1233 }
1234 THREAD_CONTENTION_DONE(&sx->lock_object);
1235 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1236 if (__predict_true(!extra_work))
1237 return (error);
1238 #endif
1239 #ifdef KDTRACE_HOOKS
1240 all_time += lockstat_nsecs(&sx->lock_object);
1241 if (sleep_time)
1242 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1243 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1244 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1245 if (lda.spin_cnt > sleep_cnt)
1246 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1247 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1248 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1249 out_lockstat:
1250 #endif
1251 if (error == 0) {
1252 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1253 contested, waittime, file, line, LOCKSTAT_READER);
1254 }
1255 GIANT_RESTORE();
1256 return (error);
1257 }
1258
1259 int
_sx_slock_int(struct sx * sx,int opts LOCK_FILE_LINE_ARG_DEF)1260 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1261 {
1262 struct thread *td;
1263 uintptr_t x;
1264 int error;
1265
1266 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1267 !TD_IS_IDLETHREAD(curthread),
1268 ("sx_slock() by idle thread %p on sx %p @ %s:%d",
1269 curthread, sx, file, line));
1270 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1271 ("sx_slock() of destroyed sx %p @ %s:%d", sx, file, line));
1272 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1273
1274 error = 0;
1275 td = curthread;
1276 x = SX_READ_VALUE(sx);
1277 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1278 !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1279 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1280 else
1281 lock_profile_obtain_lock_success(&sx->lock_object, false, 0, 0,
1282 file, line);
1283 if (error == 0) {
1284 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1285 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1286 TD_LOCKS_INC(curthread);
1287 }
1288 return (error);
1289 }
1290
1291 int
_sx_slock(struct sx * sx,int opts,const char * file,int line)1292 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1293 {
1294
1295 return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1296 }
1297
1298 static __always_inline bool
_sx_sunlock_try(struct sx * sx,struct thread * td,uintptr_t * xp)1299 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1300 {
1301
1302 for (;;) {
1303 if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1304 if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1305 *xp - SX_ONE_SHARER)) {
1306 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1307 CTR4(KTR_LOCK,
1308 "%s: %p succeeded %p -> %p",
1309 __func__, sx, (void *)*xp,
1310 (void *)(*xp - SX_ONE_SHARER));
1311 td->td_sx_slocks--;
1312 return (true);
1313 }
1314 continue;
1315 }
1316 break;
1317 }
1318 return (false);
1319 }
1320
1321 static void __noinline
_sx_sunlock_hard(struct sx * sx,struct thread * td,uintptr_t x LOCK_FILE_LINE_ARG_DEF)1322 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1323 LOCK_FILE_LINE_ARG_DEF)
1324 {
1325 uintptr_t setx, queue;
1326
1327 if (SCHEDULER_STOPPED())
1328 return;
1329
1330 if (_sx_sunlock_try(sx, td, &x))
1331 goto out_lockstat;
1332
1333 sleepq_lock(&sx->lock_object);
1334 x = SX_READ_VALUE(sx);
1335 for (;;) {
1336 if (_sx_sunlock_try(sx, td, &x))
1337 break;
1338
1339 /*
1340 * Wake up semantic here is quite simple:
1341 * Just wake up all the exclusive waiters.
1342 * Note that the state of the lock could have changed,
1343 * so if it fails loop back and retry.
1344 */
1345 setx = SX_LOCK_UNLOCKED;
1346 queue = SQ_SHARED_QUEUE;
1347 if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1348 setx |= (x & SX_LOCK_SHARED_WAITERS);
1349 queue = SQ_EXCLUSIVE_QUEUE;
1350 }
1351 setx |= (x & SX_LOCK_WRITE_SPINNER);
1352 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1353 continue;
1354 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1355 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1356 "exclusive queue", __func__, sx);
1357 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, queue);
1358 td->td_sx_slocks--;
1359 break;
1360 }
1361 sleepq_release(&sx->lock_object);
1362 out_lockstat:
1363 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1364 }
1365
1366 void
_sx_sunlock_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)1367 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1368 {
1369 struct thread *td;
1370 uintptr_t x;
1371
1372 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1373 ("sx_sunlock() of destroyed sx %p @ %s:%d", sx, file, line));
1374 _sx_assert(sx, SA_SLOCKED, file, line);
1375 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1376 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1377
1378 td = curthread;
1379 x = SX_READ_VALUE(sx);
1380 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1381 !_sx_sunlock_try(sx, td, &x)))
1382 _sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1383 else
1384 lock_profile_release_lock(&sx->lock_object, false);
1385
1386 TD_LOCKS_DEC(curthread);
1387 }
1388
1389 void
_sx_sunlock(struct sx * sx,const char * file,int line)1390 _sx_sunlock(struct sx *sx, const char *file, int line)
1391 {
1392
1393 _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1394 }
1395
1396 #ifdef INVARIANT_SUPPORT
1397 #ifndef INVARIANTS
1398 #undef _sx_assert
1399 #endif
1400
1401 /*
1402 * In the non-WITNESS case, sx_assert() can only detect that at least
1403 * *some* thread owns an slock, but it cannot guarantee that *this*
1404 * thread owns an slock.
1405 */
1406 void
_sx_assert(const struct sx * sx,int what,const char * file,int line)1407 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1408 {
1409 #ifndef WITNESS
1410 int slocked = 0;
1411 #endif
1412
1413 if (SCHEDULER_STOPPED())
1414 return;
1415 switch (what) {
1416 case SA_SLOCKED:
1417 case SA_SLOCKED | SA_NOTRECURSED:
1418 case SA_SLOCKED | SA_RECURSED:
1419 #ifndef WITNESS
1420 slocked = 1;
1421 /* FALLTHROUGH */
1422 #endif
1423 case SA_LOCKED:
1424 case SA_LOCKED | SA_NOTRECURSED:
1425 case SA_LOCKED | SA_RECURSED:
1426 #ifdef WITNESS
1427 witness_assert(&sx->lock_object, what, file, line);
1428 #else
1429 /*
1430 * If some other thread has an exclusive lock or we
1431 * have one and are asserting a shared lock, fail.
1432 * Also, if no one has a lock at all, fail.
1433 */
1434 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1435 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1436 sx_xholder(sx) != curthread)))
1437 panic("Lock %s not %slocked @ %s:%d\n",
1438 sx->lock_object.lo_name, slocked ? "share " : "",
1439 file, line);
1440
1441 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1442 if (sx_recursed(sx)) {
1443 if (what & SA_NOTRECURSED)
1444 panic("Lock %s recursed @ %s:%d\n",
1445 sx->lock_object.lo_name, file,
1446 line);
1447 } else if (what & SA_RECURSED)
1448 panic("Lock %s not recursed @ %s:%d\n",
1449 sx->lock_object.lo_name, file, line);
1450 }
1451 #endif
1452 break;
1453 case SA_XLOCKED:
1454 case SA_XLOCKED | SA_NOTRECURSED:
1455 case SA_XLOCKED | SA_RECURSED:
1456 if (sx_xholder(sx) != curthread)
1457 panic("Lock %s not exclusively locked @ %s:%d\n",
1458 sx->lock_object.lo_name, file, line);
1459 if (sx_recursed(sx)) {
1460 if (what & SA_NOTRECURSED)
1461 panic("Lock %s recursed @ %s:%d\n",
1462 sx->lock_object.lo_name, file, line);
1463 } else if (what & SA_RECURSED)
1464 panic("Lock %s not recursed @ %s:%d\n",
1465 sx->lock_object.lo_name, file, line);
1466 break;
1467 case SA_UNLOCKED:
1468 #ifdef WITNESS
1469 witness_assert(&sx->lock_object, what, file, line);
1470 #else
1471 /*
1472 * If we hold an exclusve lock fail. We can't
1473 * reliably check to see if we hold a shared lock or
1474 * not.
1475 */
1476 if (sx_xholder(sx) == curthread)
1477 panic("Lock %s exclusively locked @ %s:%d\n",
1478 sx->lock_object.lo_name, file, line);
1479 #endif
1480 break;
1481 default:
1482 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1483 line);
1484 }
1485 }
1486 #endif /* INVARIANT_SUPPORT */
1487
1488 #ifdef DDB
1489 static void
db_show_sx(const struct lock_object * lock)1490 db_show_sx(const struct lock_object *lock)
1491 {
1492 struct thread *td;
1493 const struct sx *sx;
1494
1495 sx = (const struct sx *)lock;
1496
1497 db_printf(" state: ");
1498 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1499 db_printf("UNLOCKED\n");
1500 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1501 db_printf("DESTROYED\n");
1502 return;
1503 } else if (sx->sx_lock & SX_LOCK_SHARED)
1504 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1505 else {
1506 td = sx_xholder(sx);
1507 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1508 td->td_tid, td->td_proc->p_pid, td->td_name);
1509 if (sx_recursed(sx))
1510 db_printf(" recursed: %d\n", sx->sx_recurse);
1511 }
1512
1513 db_printf(" waiters: ");
1514 switch(sx->sx_lock &
1515 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1516 case SX_LOCK_SHARED_WAITERS:
1517 db_printf("shared\n");
1518 break;
1519 case SX_LOCK_EXCLUSIVE_WAITERS:
1520 db_printf("exclusive\n");
1521 break;
1522 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1523 db_printf("exclusive and shared\n");
1524 break;
1525 default:
1526 db_printf("none\n");
1527 }
1528 }
1529
1530 /*
1531 * Check to see if a thread that is blocked on a sleep queue is actually
1532 * blocked on an sx lock. If so, output some details and return true.
1533 * If the lock has an exclusive owner, return that in *ownerp.
1534 */
1535 int
sx_chain(struct thread * td,struct thread ** ownerp)1536 sx_chain(struct thread *td, struct thread **ownerp)
1537 {
1538 const struct sx *sx;
1539
1540 /*
1541 * Check to see if this thread is blocked on an sx lock.
1542 * First, we check the lock class. If that is ok, then we
1543 * compare the lock name against the wait message.
1544 */
1545 sx = td->td_wchan;
1546 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1547 sx->lock_object.lo_name != td->td_wmesg)
1548 return (0);
1549
1550 /* We think we have an sx lock, so output some details. */
1551 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1552 *ownerp = sx_xholder(sx);
1553 if (sx->sx_lock & SX_LOCK_SHARED)
1554 db_printf("SLOCK (count %ju)\n",
1555 (uintmax_t)SX_SHARERS(sx->sx_lock));
1556 else
1557 db_printf("XLOCK\n");
1558 return (1);
1559 }
1560 #endif
1561