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 %s @ %s:%d",
282 curthread, sx->lock_object.lo_name, file, line));
283
284 x = sx->sx_lock;
285 for (;;) {
286 KASSERT(x != SX_LOCK_DESTROYED,
287 ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
288 if (!(x & SX_LOCK_SHARED))
289 break;
290 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
291 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
292 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
293 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
294 sx, 0, 0, file, line, LOCKSTAT_READER);
295 TD_LOCKS_INC(curthread);
296 curthread->td_sx_slocks++;
297 return (1);
298 }
299 }
300
301 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
302 return (0);
303 }
304
305 int
sx_try_slock_(struct sx * sx,const char * file,int line)306 sx_try_slock_(struct sx *sx, const char *file, int line)
307 {
308
309 return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
310 }
311
312 int
_sx_xlock(struct sx * sx,int opts,const char * file,int line)313 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
314 {
315 uintptr_t tid, x;
316 int error = 0;
317
318 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
319 !TD_IS_IDLETHREAD(curthread),
320 ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
321 curthread, sx->lock_object.lo_name, file, line));
322 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
323 ("sx_xlock() of destroyed sx @ %s:%d", file, line));
324 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
325 line, NULL);
326 tid = (uintptr_t)curthread;
327 x = SX_LOCK_UNLOCKED;
328 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
329 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
330 else
331 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
332 0, 0, file, line, LOCKSTAT_WRITER);
333 if (!error) {
334 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
335 file, line);
336 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
337 TD_LOCKS_INC(curthread);
338 }
339
340 return (error);
341 }
342
343 int
sx_try_xlock_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)344 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
345 {
346 struct thread *td;
347 uintptr_t tid, x;
348 int rval;
349 bool recursed;
350
351 td = curthread;
352 tid = (uintptr_t)td;
353 if (SCHEDULER_STOPPED())
354 return (1);
355
356 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
357 ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
358 curthread, sx->lock_object.lo_name, file, line));
359 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
360 ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
361
362 rval = 1;
363 recursed = false;
364 x = SX_LOCK_UNLOCKED;
365 for (;;) {
366 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
367 break;
368 if (x == SX_LOCK_UNLOCKED)
369 continue;
370 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
371 sx->sx_recurse++;
372 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
373 break;
374 }
375 rval = 0;
376 break;
377 }
378
379 LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
380 if (rval) {
381 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
382 file, line);
383 if (!recursed)
384 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
385 sx, 0, 0, file, line, LOCKSTAT_WRITER);
386 TD_LOCKS_INC(curthread);
387 }
388
389 return (rval);
390 }
391
392 int
sx_try_xlock_(struct sx * sx,const char * file,int line)393 sx_try_xlock_(struct sx *sx, const char *file, int line)
394 {
395
396 return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
397 }
398
399 void
_sx_xunlock(struct sx * sx,const char * file,int line)400 _sx_xunlock(struct sx *sx, const char *file, int line)
401 {
402
403 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
404 ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
405 _sx_assert(sx, SA_XLOCKED, file, line);
406 WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
407 LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
408 line);
409 #if LOCK_DEBUG > 0
410 _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
411 #else
412 __sx_xunlock(sx, curthread, file, line);
413 #endif
414 TD_LOCKS_DEC(curthread);
415 }
416
417 /*
418 * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
419 * This will only succeed if this thread holds a single shared lock.
420 * Return 1 if if the upgrade succeed, 0 otherwise.
421 */
422 int
sx_try_upgrade_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)423 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
424 {
425 uintptr_t x;
426 uintptr_t waiters;
427 int success;
428
429 if (SCHEDULER_STOPPED())
430 return (1);
431
432 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
433 ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
434 _sx_assert(sx, SA_SLOCKED, file, line);
435
436 /*
437 * Try to switch from one shared lock to an exclusive lock. We need
438 * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
439 * we will wake up the exclusive waiters when we drop the lock.
440 */
441 success = 0;
442 x = SX_READ_VALUE(sx);
443 for (;;) {
444 if (SX_SHARERS(x) > 1)
445 break;
446 waiters = (x & SX_LOCK_WAITERS);
447 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
448 (uintptr_t)curthread | waiters)) {
449 success = 1;
450 break;
451 }
452 }
453 LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
454 if (success) {
455 curthread->td_sx_slocks--;
456 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
457 file, line);
458 LOCKSTAT_RECORD0(sx__upgrade, sx);
459 }
460 return (success);
461 }
462
463 int
sx_try_upgrade_(struct sx * sx,const char * file,int line)464 sx_try_upgrade_(struct sx *sx, const char *file, int line)
465 {
466
467 return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
468 }
469
470 /*
471 * Downgrade an unrecursed exclusive lock into a single shared lock.
472 */
473 void
sx_downgrade_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)474 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
475 {
476 uintptr_t x;
477
478 if (SCHEDULER_STOPPED())
479 return;
480
481 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
482 ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
483 _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
484 #ifndef INVARIANTS
485 if (sx_recursed(sx))
486 panic("downgrade of a recursed lock");
487 #endif
488
489 WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
490
491 /*
492 * Try to switch from an exclusive lock with no shared waiters
493 * to one sharer with no shared waiters. If there are
494 * exclusive waiters, we don't need to lock the sleep queue so
495 * long as we preserve the flag. We do one quick try and if
496 * that fails we grab the sleepq lock to keep the flags from
497 * changing and do it the slow way.
498 *
499 * We have to lock the sleep queue if there are shared waiters
500 * so we can wake them up.
501 */
502 x = sx->sx_lock;
503 if (!(x & SX_LOCK_SHARED_WAITERS) &&
504 atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
505 (x & SX_LOCK_EXCLUSIVE_WAITERS)))
506 goto out;
507
508 /*
509 * Lock the sleep queue so we can read the waiters bits
510 * without any races and wakeup any shared waiters.
511 */
512 sleepq_lock(&sx->lock_object);
513
514 /*
515 * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
516 * shared lock. If there are any shared waiters, wake them up.
517 */
518 x = sx->sx_lock;
519 atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
520 (x & SX_LOCK_EXCLUSIVE_WAITERS));
521 if (x & SX_LOCK_SHARED_WAITERS)
522 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
523 SQ_SHARED_QUEUE);
524 sleepq_release(&sx->lock_object);
525
526 out:
527 curthread->td_sx_slocks++;
528 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
529 LOCKSTAT_RECORD0(sx__downgrade, sx);
530 }
531
532 void
sx_downgrade_(struct sx * sx,const char * file,int line)533 sx_downgrade_(struct sx *sx, const char *file, int line)
534 {
535
536 sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
537 }
538
539 #ifdef ADAPTIVE_SX
540 static inline void
sx_drop_critical(uintptr_t x,bool * in_critical,int * extra_work)541 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
542 {
543
544 if (x & SX_LOCK_WRITE_SPINNER)
545 return;
546 if (*in_critical) {
547 critical_exit();
548 *in_critical = false;
549 (*extra_work)--;
550 }
551 }
552 #else
553 #define sx_drop_critical(x, in_critical, extra_work) do { } while (0)
554 #endif
555
556 /*
557 * This function represents the so-called 'hard case' for sx_xlock
558 * operation. All 'easy case' failures are redirected to this. Note
559 * that ideally this would be a static function, but it needs to be
560 * accessible from at least sx.h.
561 */
562 int
_sx_xlock_hard(struct sx * sx,uintptr_t x,int opts LOCK_FILE_LINE_ARG_DEF)563 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
564 {
565 GIANT_DECLARE;
566 uintptr_t tid, setx;
567 #ifdef ADAPTIVE_SX
568 struct thread *owner;
569 u_int i, n, spintries = 0;
570 enum { READERS, WRITER } sleep_reason = READERS;
571 bool in_critical = false;
572 #endif
573 #ifdef LOCK_PROFILING
574 uint64_t waittime = 0;
575 int contested = 0;
576 #endif
577 int error = 0;
578 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
579 struct lock_delay_arg lda;
580 #endif
581 #ifdef KDTRACE_HOOKS
582 u_int sleep_cnt = 0;
583 int64_t sleep_time = 0;
584 int64_t all_time = 0;
585 uintptr_t state = 0;
586 #endif
587 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
588 int doing_lockprof = 0;
589 #endif
590 int extra_work = 0;
591
592 tid = (uintptr_t)curthread;
593
594 #ifdef KDTRACE_HOOKS
595 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
596 while (x == SX_LOCK_UNLOCKED) {
597 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
598 goto out_lockstat;
599 }
600 extra_work = 1;
601 doing_lockprof = 1;
602 all_time -= lockstat_nsecs(&sx->lock_object);
603 }
604 state = x;
605 #endif
606 #ifdef LOCK_PROFILING
607 extra_work = 1;
608 doing_lockprof = 1;
609 #endif
610
611 if (SCHEDULER_STOPPED())
612 return (0);
613
614 if (__predict_false(x == SX_LOCK_UNLOCKED))
615 x = SX_READ_VALUE(sx);
616
617 /* If we already hold an exclusive lock, then recurse. */
618 if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
619 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
620 ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
621 sx->lock_object.lo_name, file, line));
622 sx->sx_recurse++;
623 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
624 if (LOCK_LOG_TEST(&sx->lock_object, 0))
625 CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
626 return (0);
627 }
628
629 if (LOCK_LOG_TEST(&sx->lock_object, 0))
630 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
631 sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
632
633 #if defined(ADAPTIVE_SX)
634 lock_delay_arg_init(&lda, &sx_delay);
635 #elif defined(KDTRACE_HOOKS)
636 lock_delay_arg_init_noadapt(&lda);
637 #endif
638
639 #ifdef HWPMC_HOOKS
640 PMC_SOFT_CALL( , , lock, failed);
641 #endif
642 lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
643 &waittime);
644
645 #ifndef INVARIANTS
646 GIANT_SAVE(extra_work);
647 #endif
648
649 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
650
651 for (;;) {
652 if (x == SX_LOCK_UNLOCKED) {
653 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
654 break;
655 continue;
656 }
657 #ifdef INVARIANTS
658 GIANT_SAVE(extra_work);
659 #endif
660 #ifdef KDTRACE_HOOKS
661 lda.spin_cnt++;
662 #endif
663 #ifdef ADAPTIVE_SX
664 if (x == (SX_LOCK_SHARED | SX_LOCK_WRITE_SPINNER)) {
665 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
666 break;
667 continue;
668 }
669
670 /*
671 * If the lock is write locked and the owner is
672 * running on another CPU, spin until the owner stops
673 * running or the state of the lock changes.
674 */
675 if ((x & SX_LOCK_SHARED) == 0) {
676 sx_drop_critical(x, &in_critical, &extra_work);
677 sleep_reason = WRITER;
678 owner = lv_sx_owner(x);
679 if (!TD_IS_RUNNING(owner))
680 goto sleepq;
681 if (LOCK_LOG_TEST(&sx->lock_object, 0))
682 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
683 __func__, sx, owner);
684 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
685 "spinning", "lockname:\"%s\"",
686 sx->lock_object.lo_name);
687 do {
688 lock_delay(&lda);
689 x = SX_READ_VALUE(sx);
690 owner = lv_sx_owner(x);
691 } while (owner != NULL && TD_IS_RUNNING(owner));
692 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
693 "running");
694 continue;
695 } else if (SX_SHARERS(x) > 0) {
696 sleep_reason = READERS;
697 if (spintries == asx_retries)
698 goto sleepq;
699 if (!(x & SX_LOCK_WRITE_SPINNER)) {
700 if (!in_critical) {
701 critical_enter();
702 in_critical = true;
703 extra_work++;
704 }
705 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
706 x | SX_LOCK_WRITE_SPINNER)) {
707 critical_exit();
708 in_critical = false;
709 extra_work--;
710 continue;
711 }
712 }
713 spintries++;
714 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
715 "spinning", "lockname:\"%s\"",
716 sx->lock_object.lo_name);
717 n = SX_SHARERS(x);
718 for (i = 0; i < asx_loops; i += n) {
719 lock_delay_spin(n);
720 x = SX_READ_VALUE(sx);
721 if (!(x & SX_LOCK_WRITE_SPINNER))
722 break;
723 if (!(x & SX_LOCK_SHARED))
724 break;
725 n = SX_SHARERS(x);
726 if (n == 0)
727 break;
728 }
729 #ifdef KDTRACE_HOOKS
730 lda.spin_cnt += i;
731 #endif
732 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
733 "running");
734 if (i < asx_loops)
735 continue;
736 }
737 sleepq:
738 #endif
739 sleepq_lock(&sx->lock_object);
740 x = SX_READ_VALUE(sx);
741 retry_sleepq:
742
743 /*
744 * If the lock was released while spinning on the
745 * sleep queue chain lock, try again.
746 */
747 if (x == SX_LOCK_UNLOCKED) {
748 sleepq_release(&sx->lock_object);
749 sx_drop_critical(x, &in_critical, &extra_work);
750 continue;
751 }
752
753 #ifdef ADAPTIVE_SX
754 /*
755 * The current lock owner might have started executing
756 * on another CPU (or the lock could have changed
757 * owners) while we were waiting on the sleep queue
758 * chain lock. If so, drop the sleep queue lock and try
759 * again.
760 */
761 if (!(x & SX_LOCK_SHARED)) {
762 owner = (struct thread *)SX_OWNER(x);
763 if (TD_IS_RUNNING(owner)) {
764 sleepq_release(&sx->lock_object);
765 sx_drop_critical(x, &in_critical,
766 &extra_work);
767 continue;
768 }
769 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
770 sleepq_release(&sx->lock_object);
771 sx_drop_critical(x, &in_critical, &extra_work);
772 continue;
773 }
774 #endif
775
776 /*
777 * If an exclusive lock was released with both shared
778 * and exclusive waiters and a shared waiter hasn't
779 * woken up and acquired the lock yet, sx_lock will be
780 * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
781 * If we see that value, try to acquire it once. Note
782 * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
783 * as there are other exclusive waiters still. If we
784 * fail, restart the loop.
785 */
786 setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
787 if ((x & ~setx) == SX_LOCK_SHARED) {
788 setx &= ~SX_LOCK_WRITE_SPINNER;
789 if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
790 goto retry_sleepq;
791 sleepq_release(&sx->lock_object);
792 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
793 __func__, sx);
794 break;
795 }
796
797 #ifdef ADAPTIVE_SX
798 /*
799 * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
800 * It is an invariant that when the bit is set, there is
801 * a writer ready to grab the lock. Thus clear the bit since
802 * we are going to sleep.
803 */
804 if (in_critical) {
805 if ((x & SX_LOCK_WRITE_SPINNER) ||
806 !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
807 setx = x & ~SX_LOCK_WRITE_SPINNER;
808 setx |= SX_LOCK_EXCLUSIVE_WAITERS;
809 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
810 setx)) {
811 goto retry_sleepq;
812 }
813 }
814 critical_exit();
815 in_critical = false;
816 } else {
817 #endif
818 /*
819 * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
820 * than loop back and retry.
821 */
822 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
823 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
824 x | SX_LOCK_EXCLUSIVE_WAITERS)) {
825 goto retry_sleepq;
826 }
827 if (LOCK_LOG_TEST(&sx->lock_object, 0))
828 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
829 __func__, sx);
830 }
831 #ifdef ADAPTIVE_SX
832 }
833 #endif
834
835 /*
836 * Since we have been unable to acquire the exclusive
837 * lock and the exclusive waiters flag is set, we have
838 * to sleep.
839 */
840 if (LOCK_LOG_TEST(&sx->lock_object, 0))
841 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
842 __func__, sx);
843
844 #ifdef KDTRACE_HOOKS
845 sleep_time -= lockstat_nsecs(&sx->lock_object);
846 #endif
847 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
848 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
849 SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
850 /*
851 * Hack: this can land in thread_suspend_check which will
852 * conditionally take a mutex, tripping over an assert if a
853 * lock we are waiting for is set.
854 */
855 THREAD_CONTENTION_DONE(&sx->lock_object);
856 if (!(opts & SX_INTERRUPTIBLE))
857 sleepq_wait(&sx->lock_object, 0);
858 else
859 error = sleepq_wait_sig(&sx->lock_object, 0);
860 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
861 #ifdef KDTRACE_HOOKS
862 sleep_time += lockstat_nsecs(&sx->lock_object);
863 sleep_cnt++;
864 #endif
865 if (error) {
866 if (LOCK_LOG_TEST(&sx->lock_object, 0))
867 CTR2(KTR_LOCK,
868 "%s: interruptible sleep by %p suspended by signal",
869 __func__, sx);
870 break;
871 }
872 if (LOCK_LOG_TEST(&sx->lock_object, 0))
873 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
874 __func__, sx);
875 x = SX_READ_VALUE(sx);
876 }
877 THREAD_CONTENTION_DONE(&sx->lock_object);
878 if (__predict_true(!extra_work))
879 return (error);
880 #ifdef ADAPTIVE_SX
881 if (in_critical)
882 critical_exit();
883 #endif
884 GIANT_RESTORE();
885 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
886 if (__predict_true(!doing_lockprof))
887 return (error);
888 #endif
889 #ifdef KDTRACE_HOOKS
890 all_time += lockstat_nsecs(&sx->lock_object);
891 if (sleep_time)
892 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
893 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
894 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
895 if (lda.spin_cnt > sleep_cnt)
896 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
897 LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
898 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
899 out_lockstat:
900 #endif
901 if (!error)
902 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
903 contested, waittime, file, line, LOCKSTAT_WRITER);
904 return (error);
905 }
906
907 /*
908 * This function represents the so-called 'hard case' for sx_xunlock
909 * operation. All 'easy case' failures are redirected to this. Note
910 * that ideally this would be a static function, but it needs to be
911 * accessible from at least sx.h.
912 */
913 void
_sx_xunlock_hard(struct sx * sx,uintptr_t x LOCK_FILE_LINE_ARG_DEF)914 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
915 {
916 uintptr_t tid, setx;
917 int queue;
918
919 if (SCHEDULER_STOPPED())
920 return;
921
922 tid = (uintptr_t)curthread;
923
924 if (__predict_false(x == tid))
925 x = SX_READ_VALUE(sx);
926
927 MPASS(!(x & SX_LOCK_SHARED));
928
929 if (__predict_false(x & SX_LOCK_RECURSED)) {
930 /* The lock is recursed, unrecurse one level. */
931 if ((--sx->sx_recurse) == 0)
932 atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
933 if (LOCK_LOG_TEST(&sx->lock_object, 0))
934 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
935 return;
936 }
937
938 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
939 if (x == tid &&
940 atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
941 return;
942
943 if (LOCK_LOG_TEST(&sx->lock_object, 0))
944 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
945
946 sleepq_lock(&sx->lock_object);
947 x = SX_READ_VALUE(sx);
948 MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
949
950 /*
951 * The wake up algorithm here is quite simple and probably not
952 * ideal. It gives precedence to shared waiters if they are
953 * present. For this condition, we have to preserve the
954 * state of the exclusive waiters flag.
955 * If interruptible sleeps left the shared queue empty avoid a
956 * starvation for the threads sleeping on the exclusive queue by giving
957 * them precedence and cleaning up the shared waiters bit anyway.
958 */
959 setx = SX_LOCK_UNLOCKED;
960 queue = SQ_SHARED_QUEUE;
961 if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
962 sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
963 queue = SQ_EXCLUSIVE_QUEUE;
964 setx |= (x & SX_LOCK_SHARED_WAITERS);
965 }
966 atomic_store_rel_ptr(&sx->sx_lock, setx);
967
968 /* Wake up all the waiters for the specific queue. */
969 if (LOCK_LOG_TEST(&sx->lock_object, 0))
970 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
971 __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
972 "exclusive");
973
974 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, queue);
975 sleepq_release(&sx->lock_object);
976 }
977
978 static __always_inline bool
__sx_can_read(struct thread * td,uintptr_t x,bool fp)979 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
980 {
981
982 if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
983 == SX_LOCK_SHARED)
984 return (true);
985 if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
986 return (true);
987 return (false);
988 }
989
990 static __always_inline bool
__sx_slock_try(struct sx * sx,struct thread * td,uintptr_t * xp,bool fp LOCK_FILE_LINE_ARG_DEF)991 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
992 LOCK_FILE_LINE_ARG_DEF)
993 {
994
995 /*
996 * If no other thread has an exclusive lock then try to bump up
997 * the count of sharers. Since we have to preserve the state
998 * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
999 * shared lock loop back and retry.
1000 */
1001 while (__sx_can_read(td, *xp, fp)) {
1002 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1003 *xp + SX_ONE_SHARER)) {
1004 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1005 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1006 __func__, sx, (void *)*xp,
1007 (void *)(*xp + SX_ONE_SHARER));
1008 td->td_sx_slocks++;
1009 return (true);
1010 }
1011 }
1012 return (false);
1013 }
1014
1015 static int __noinline
_sx_slock_hard(struct sx * sx,int opts,uintptr_t x LOCK_FILE_LINE_ARG_DEF)1016 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1017 {
1018 GIANT_DECLARE;
1019 struct thread *td;
1020 #ifdef ADAPTIVE_SX
1021 struct thread *owner;
1022 u_int i, n, spintries = 0;
1023 #endif
1024 #ifdef LOCK_PROFILING
1025 uint64_t waittime = 0;
1026 int contested = 0;
1027 #endif
1028 int error = 0;
1029 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1030 struct lock_delay_arg lda;
1031 #endif
1032 #ifdef KDTRACE_HOOKS
1033 u_int sleep_cnt = 0;
1034 int64_t sleep_time = 0;
1035 int64_t all_time = 0;
1036 uintptr_t state = 0;
1037 #endif
1038 int extra_work __sdt_used = 0;
1039
1040 td = curthread;
1041
1042 #ifdef KDTRACE_HOOKS
1043 if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1044 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1045 goto out_lockstat;
1046 extra_work = 1;
1047 all_time -= lockstat_nsecs(&sx->lock_object);
1048 }
1049 state = x;
1050 #endif
1051 #ifdef LOCK_PROFILING
1052 extra_work = 1;
1053 #endif
1054
1055 if (SCHEDULER_STOPPED())
1056 return (0);
1057
1058 #if defined(ADAPTIVE_SX)
1059 lock_delay_arg_init(&lda, &sx_delay);
1060 #elif defined(KDTRACE_HOOKS)
1061 lock_delay_arg_init_noadapt(&lda);
1062 #endif
1063
1064 #ifdef HWPMC_HOOKS
1065 PMC_SOFT_CALL( , , lock, failed);
1066 #endif
1067 lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
1068 &waittime);
1069
1070 #ifndef INVARIANTS
1071 GIANT_SAVE(extra_work);
1072 #endif
1073
1074 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1075
1076 /*
1077 * As with rwlocks, we don't make any attempt to try to block
1078 * shared locks once there is an exclusive waiter.
1079 */
1080 for (;;) {
1081 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1082 break;
1083 #ifdef INVARIANTS
1084 GIANT_SAVE(extra_work);
1085 #endif
1086 #ifdef KDTRACE_HOOKS
1087 lda.spin_cnt++;
1088 #endif
1089
1090 #ifdef ADAPTIVE_SX
1091 /*
1092 * If the owner is running on another CPU, spin until
1093 * the owner stops running or the state of the lock
1094 * changes.
1095 */
1096 if ((x & SX_LOCK_SHARED) == 0) {
1097 owner = lv_sx_owner(x);
1098 if (TD_IS_RUNNING(owner)) {
1099 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1100 CTR3(KTR_LOCK,
1101 "%s: spinning on %p held by %p",
1102 __func__, sx, owner);
1103 KTR_STATE1(KTR_SCHED, "thread",
1104 sched_tdname(curthread), "spinning",
1105 "lockname:\"%s\"", sx->lock_object.lo_name);
1106 do {
1107 lock_delay(&lda);
1108 x = SX_READ_VALUE(sx);
1109 owner = lv_sx_owner(x);
1110 } while (owner != NULL && TD_IS_RUNNING(owner));
1111 KTR_STATE0(KTR_SCHED, "thread",
1112 sched_tdname(curthread), "running");
1113 continue;
1114 }
1115 } else {
1116 if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1117 MPASS(!__sx_can_read(td, x, false));
1118 lock_delay_spin(2);
1119 x = SX_READ_VALUE(sx);
1120 continue;
1121 }
1122 if (spintries < asx_retries) {
1123 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1124 "spinning", "lockname:\"%s\"",
1125 sx->lock_object.lo_name);
1126 n = SX_SHARERS(x);
1127 for (i = 0; i < asx_loops; i += n) {
1128 lock_delay_spin(n);
1129 x = SX_READ_VALUE(sx);
1130 if (!(x & SX_LOCK_SHARED))
1131 break;
1132 n = SX_SHARERS(x);
1133 if (n == 0)
1134 break;
1135 if (__sx_can_read(td, x, false))
1136 break;
1137 }
1138 #ifdef KDTRACE_HOOKS
1139 lda.spin_cnt += i;
1140 #endif
1141 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1142 "running");
1143 if (i < asx_loops)
1144 continue;
1145 }
1146 }
1147 #endif
1148
1149 /*
1150 * Some other thread already has an exclusive lock, so
1151 * start the process of blocking.
1152 */
1153 sleepq_lock(&sx->lock_object);
1154 x = SX_READ_VALUE(sx);
1155 retry_sleepq:
1156 if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1157 __sx_can_read(td, x, false)) {
1158 sleepq_release(&sx->lock_object);
1159 continue;
1160 }
1161
1162 #ifdef ADAPTIVE_SX
1163 /*
1164 * If the owner is running on another CPU, spin until
1165 * the owner stops running or the state of the lock
1166 * changes.
1167 */
1168 if (!(x & SX_LOCK_SHARED)) {
1169 owner = (struct thread *)SX_OWNER(x);
1170 if (TD_IS_RUNNING(owner)) {
1171 sleepq_release(&sx->lock_object);
1172 x = SX_READ_VALUE(sx);
1173 continue;
1174 }
1175 }
1176 #endif
1177
1178 /*
1179 * Try to set the SX_LOCK_SHARED_WAITERS flag. If we
1180 * fail to set it drop the sleep queue lock and loop
1181 * back.
1182 */
1183 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1184 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1185 x | SX_LOCK_SHARED_WAITERS))
1186 goto retry_sleepq;
1187 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1188 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1189 __func__, sx);
1190 }
1191
1192 /*
1193 * Since we have been unable to acquire the shared lock,
1194 * we have to sleep.
1195 */
1196 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1197 CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1198 __func__, sx);
1199
1200 #ifdef KDTRACE_HOOKS
1201 sleep_time -= lockstat_nsecs(&sx->lock_object);
1202 #endif
1203 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1204 SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1205 SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1206 /*
1207 * Hack: this can land in thread_suspend_check which will
1208 * conditionally take a mutex, tripping over an assert if a
1209 * lock we are waiting for is set.
1210 */
1211 THREAD_CONTENTION_DONE(&sx->lock_object);
1212 if (!(opts & SX_INTERRUPTIBLE))
1213 sleepq_wait(&sx->lock_object, 0);
1214 else
1215 error = sleepq_wait_sig(&sx->lock_object, 0);
1216 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1217 #ifdef KDTRACE_HOOKS
1218 sleep_time += lockstat_nsecs(&sx->lock_object);
1219 sleep_cnt++;
1220 #endif
1221 if (error) {
1222 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1223 CTR2(KTR_LOCK,
1224 "%s: interruptible sleep by %p suspended by signal",
1225 __func__, sx);
1226 break;
1227 }
1228 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1229 CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1230 __func__, sx);
1231 x = SX_READ_VALUE(sx);
1232 }
1233 THREAD_CONTENTION_DONE(&sx->lock_object);
1234 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1235 if (__predict_true(!extra_work))
1236 return (error);
1237 #endif
1238 #ifdef KDTRACE_HOOKS
1239 all_time += lockstat_nsecs(&sx->lock_object);
1240 if (sleep_time)
1241 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1242 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1243 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1244 if (lda.spin_cnt > sleep_cnt)
1245 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1246 LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1247 (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1248 out_lockstat:
1249 #endif
1250 if (error == 0) {
1251 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1252 contested, waittime, file, line, LOCKSTAT_READER);
1253 }
1254 GIANT_RESTORE();
1255 return (error);
1256 }
1257
1258 int
_sx_slock_int(struct sx * sx,int opts LOCK_FILE_LINE_ARG_DEF)1259 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1260 {
1261 struct thread *td;
1262 uintptr_t x;
1263 int error;
1264
1265 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1266 !TD_IS_IDLETHREAD(curthread),
1267 ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1268 curthread, sx->lock_object.lo_name, file, line));
1269 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1270 ("sx_slock() of destroyed sx @ %s:%d", file, line));
1271 WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1272
1273 error = 0;
1274 td = curthread;
1275 x = SX_READ_VALUE(sx);
1276 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1277 !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1278 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1279 else
1280 lock_profile_obtain_lock_success(&sx->lock_object, false, 0, 0,
1281 file, line);
1282 if (error == 0) {
1283 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1284 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1285 TD_LOCKS_INC(curthread);
1286 }
1287 return (error);
1288 }
1289
1290 int
_sx_slock(struct sx * sx,int opts,const char * file,int line)1291 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1292 {
1293
1294 return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1295 }
1296
1297 static __always_inline bool
_sx_sunlock_try(struct sx * sx,struct thread * td,uintptr_t * xp)1298 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1299 {
1300
1301 for (;;) {
1302 if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1303 if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1304 *xp - SX_ONE_SHARER)) {
1305 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1306 CTR4(KTR_LOCK,
1307 "%s: %p succeeded %p -> %p",
1308 __func__, sx, (void *)*xp,
1309 (void *)(*xp - SX_ONE_SHARER));
1310 td->td_sx_slocks--;
1311 return (true);
1312 }
1313 continue;
1314 }
1315 break;
1316 }
1317 return (false);
1318 }
1319
1320 static void __noinline
_sx_sunlock_hard(struct sx * sx,struct thread * td,uintptr_t x LOCK_FILE_LINE_ARG_DEF)1321 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1322 LOCK_FILE_LINE_ARG_DEF)
1323 {
1324 uintptr_t setx, queue;
1325
1326 if (SCHEDULER_STOPPED())
1327 return;
1328
1329 if (_sx_sunlock_try(sx, td, &x))
1330 goto out_lockstat;
1331
1332 sleepq_lock(&sx->lock_object);
1333 x = SX_READ_VALUE(sx);
1334 for (;;) {
1335 if (_sx_sunlock_try(sx, td, &x))
1336 break;
1337
1338 /*
1339 * Wake up semantic here is quite simple:
1340 * Just wake up all the exclusive waiters.
1341 * Note that the state of the lock could have changed,
1342 * so if it fails loop back and retry.
1343 */
1344 setx = SX_LOCK_UNLOCKED;
1345 queue = SQ_SHARED_QUEUE;
1346 if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1347 setx |= (x & SX_LOCK_SHARED_WAITERS);
1348 queue = SQ_EXCLUSIVE_QUEUE;
1349 }
1350 setx |= (x & SX_LOCK_WRITE_SPINNER);
1351 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1352 continue;
1353 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1354 CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1355 "exclusive queue", __func__, sx);
1356 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, queue);
1357 td->td_sx_slocks--;
1358 break;
1359 }
1360 sleepq_release(&sx->lock_object);
1361 out_lockstat:
1362 LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1363 }
1364
1365 void
_sx_sunlock_int(struct sx * sx LOCK_FILE_LINE_ARG_DEF)1366 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1367 {
1368 struct thread *td;
1369 uintptr_t x;
1370
1371 KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1372 ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1373 _sx_assert(sx, SA_SLOCKED, file, line);
1374 WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1375 LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1376
1377 td = curthread;
1378 x = SX_READ_VALUE(sx);
1379 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1380 !_sx_sunlock_try(sx, td, &x)))
1381 _sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1382 else
1383 lock_profile_release_lock(&sx->lock_object, false);
1384
1385 TD_LOCKS_DEC(curthread);
1386 }
1387
1388 void
_sx_sunlock(struct sx * sx,const char * file,int line)1389 _sx_sunlock(struct sx *sx, const char *file, int line)
1390 {
1391
1392 _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1393 }
1394
1395 #ifdef INVARIANT_SUPPORT
1396 #ifndef INVARIANTS
1397 #undef _sx_assert
1398 #endif
1399
1400 /*
1401 * In the non-WITNESS case, sx_assert() can only detect that at least
1402 * *some* thread owns an slock, but it cannot guarantee that *this*
1403 * thread owns an slock.
1404 */
1405 void
_sx_assert(const struct sx * sx,int what,const char * file,int line)1406 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1407 {
1408 #ifndef WITNESS
1409 int slocked = 0;
1410 #endif
1411
1412 if (SCHEDULER_STOPPED())
1413 return;
1414 switch (what) {
1415 case SA_SLOCKED:
1416 case SA_SLOCKED | SA_NOTRECURSED:
1417 case SA_SLOCKED | SA_RECURSED:
1418 #ifndef WITNESS
1419 slocked = 1;
1420 /* FALLTHROUGH */
1421 #endif
1422 case SA_LOCKED:
1423 case SA_LOCKED | SA_NOTRECURSED:
1424 case SA_LOCKED | SA_RECURSED:
1425 #ifdef WITNESS
1426 witness_assert(&sx->lock_object, what, file, line);
1427 #else
1428 /*
1429 * If some other thread has an exclusive lock or we
1430 * have one and are asserting a shared lock, fail.
1431 * Also, if no one has a lock at all, fail.
1432 */
1433 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1434 (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1435 sx_xholder(sx) != curthread)))
1436 panic("Lock %s not %slocked @ %s:%d\n",
1437 sx->lock_object.lo_name, slocked ? "share " : "",
1438 file, line);
1439
1440 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1441 if (sx_recursed(sx)) {
1442 if (what & SA_NOTRECURSED)
1443 panic("Lock %s recursed @ %s:%d\n",
1444 sx->lock_object.lo_name, file,
1445 line);
1446 } else if (what & SA_RECURSED)
1447 panic("Lock %s not recursed @ %s:%d\n",
1448 sx->lock_object.lo_name, file, line);
1449 }
1450 #endif
1451 break;
1452 case SA_XLOCKED:
1453 case SA_XLOCKED | SA_NOTRECURSED:
1454 case SA_XLOCKED | SA_RECURSED:
1455 if (sx_xholder(sx) != curthread)
1456 panic("Lock %s not exclusively locked @ %s:%d\n",
1457 sx->lock_object.lo_name, file, line);
1458 if (sx_recursed(sx)) {
1459 if (what & SA_NOTRECURSED)
1460 panic("Lock %s recursed @ %s:%d\n",
1461 sx->lock_object.lo_name, file, line);
1462 } else if (what & SA_RECURSED)
1463 panic("Lock %s not recursed @ %s:%d\n",
1464 sx->lock_object.lo_name, file, line);
1465 break;
1466 case SA_UNLOCKED:
1467 #ifdef WITNESS
1468 witness_assert(&sx->lock_object, what, file, line);
1469 #else
1470 /*
1471 * If we hold an exclusve lock fail. We can't
1472 * reliably check to see if we hold a shared lock or
1473 * not.
1474 */
1475 if (sx_xholder(sx) == curthread)
1476 panic("Lock %s exclusively locked @ %s:%d\n",
1477 sx->lock_object.lo_name, file, line);
1478 #endif
1479 break;
1480 default:
1481 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1482 line);
1483 }
1484 }
1485 #endif /* INVARIANT_SUPPORT */
1486
1487 #ifdef DDB
1488 static void
db_show_sx(const struct lock_object * lock)1489 db_show_sx(const struct lock_object *lock)
1490 {
1491 struct thread *td;
1492 const struct sx *sx;
1493
1494 sx = (const struct sx *)lock;
1495
1496 db_printf(" state: ");
1497 if (sx->sx_lock == SX_LOCK_UNLOCKED)
1498 db_printf("UNLOCKED\n");
1499 else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1500 db_printf("DESTROYED\n");
1501 return;
1502 } else if (sx->sx_lock & SX_LOCK_SHARED)
1503 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1504 else {
1505 td = sx_xholder(sx);
1506 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1507 td->td_tid, td->td_proc->p_pid, td->td_name);
1508 if (sx_recursed(sx))
1509 db_printf(" recursed: %d\n", sx->sx_recurse);
1510 }
1511
1512 db_printf(" waiters: ");
1513 switch(sx->sx_lock &
1514 (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1515 case SX_LOCK_SHARED_WAITERS:
1516 db_printf("shared\n");
1517 break;
1518 case SX_LOCK_EXCLUSIVE_WAITERS:
1519 db_printf("exclusive\n");
1520 break;
1521 case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1522 db_printf("exclusive and shared\n");
1523 break;
1524 default:
1525 db_printf("none\n");
1526 }
1527 }
1528
1529 /*
1530 * Check to see if a thread that is blocked on a sleep queue is actually
1531 * blocked on an sx lock. If so, output some details and return true.
1532 * If the lock has an exclusive owner, return that in *ownerp.
1533 */
1534 int
sx_chain(struct thread * td,struct thread ** ownerp)1535 sx_chain(struct thread *td, struct thread **ownerp)
1536 {
1537 const struct sx *sx;
1538
1539 /*
1540 * Check to see if this thread is blocked on an sx lock.
1541 * First, we check the lock class. If that is ok, then we
1542 * compare the lock name against the wait message.
1543 */
1544 sx = td->td_wchan;
1545 if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1546 sx->lock_object.lo_name != td->td_wmesg)
1547 return (0);
1548
1549 /* We think we have an sx lock, so output some details. */
1550 db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1551 *ownerp = sx_xholder(sx);
1552 if (sx->sx_lock & SX_LOCK_SHARED)
1553 db_printf("SLOCK (count %ju)\n",
1554 (uintmax_t)SX_SHARERS(sx->sx_lock));
1555 else
1556 db_printf("XLOCK\n");
1557 return (1);
1558 }
1559 #endif
1560