1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Berkeley Software Design Inc's name may not be used to endorse or
15 * promote products derived from this software without specific prior
16 * written permission.
17 *
18 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 *
30 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
31 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
32 */
33
34 /*
35 * Machine independent bits of mutex implementation.
36 */
37
38 #include <sys/cdefs.h>
39 #include "opt_adaptive_mutexes.h"
40 #include "opt_ddb.h"
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resourcevar.h>
56 #include <sys/sched.h>
57 #include <sys/sbuf.h>
58 #include <sys/smp.h>
59 #include <sys/sysctl.h>
60 #include <sys/turnstile.h>
61 #include <sys/vmmeter.h>
62 #include <sys/lock_profile.h>
63
64 #include <machine/atomic.h>
65 #include <machine/bus.h>
66 #include <machine/cpu.h>
67
68 #include <ddb/ddb.h>
69
70 #include <fs/devfs/devfs_int.h>
71
72 #include <vm/vm.h>
73 #include <vm/vm_extern.h>
74
75 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
76 #define ADAPTIVE_MUTEXES
77 #endif
78
79 #ifdef HWPMC_HOOKS
80 #include <sys/pmckern.h>
81 PMC_SOFT_DEFINE( , , lock, failed);
82 #endif
83
84 /*
85 * Return the mutex address when the lock cookie address is provided.
86 * This functionality assumes that struct mtx* have a member named mtx_lock.
87 */
88 #define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock))
89
90 /*
91 * Internal utility macros.
92 */
93 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
94
95 #define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
96
97 static void assert_mtx(const struct lock_object *lock, int what);
98 #ifdef DDB
99 static void db_show_mtx(const struct lock_object *lock);
100 #endif
101 static void lock_mtx(struct lock_object *lock, uintptr_t how);
102 static void lock_spin(struct lock_object *lock, uintptr_t how);
103 static int trylock_mtx(struct lock_object *lock, uintptr_t how);
104 static int trylock_spin(struct lock_object *lock, uintptr_t how);
105 #ifdef KDTRACE_HOOKS
106 static int owner_mtx(const struct lock_object *lock,
107 struct thread **owner);
108 #endif
109 static uintptr_t unlock_mtx(struct lock_object *lock);
110 static uintptr_t unlock_spin(struct lock_object *lock);
111
112 /*
113 * Lock classes for sleep and spin mutexes.
114 */
115 struct lock_class lock_class_mtx_sleep = {
116 .lc_name = "sleep mutex",
117 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
118 .lc_assert = assert_mtx,
119 #ifdef DDB
120 .lc_ddb_show = db_show_mtx,
121 #endif
122 .lc_lock = lock_mtx,
123 .lc_trylock = trylock_mtx,
124 .lc_unlock = unlock_mtx,
125 #ifdef KDTRACE_HOOKS
126 .lc_owner = owner_mtx,
127 #endif
128 };
129 struct lock_class lock_class_mtx_spin = {
130 .lc_name = "spin mutex",
131 .lc_flags = LC_SPINLOCK | LC_RECURSABLE,
132 .lc_assert = assert_mtx,
133 #ifdef DDB
134 .lc_ddb_show = db_show_mtx,
135 #endif
136 .lc_lock = lock_spin,
137 .lc_trylock = trylock_spin,
138 .lc_unlock = unlock_spin,
139 #ifdef KDTRACE_HOOKS
140 .lc_owner = owner_mtx,
141 #endif
142 };
143
144 #ifdef ADAPTIVE_MUTEXES
145 #ifdef MUTEX_CUSTOM_BACKOFF
146 static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
147 "mtx debugging");
148
149 static struct lock_delay_config __read_frequently mtx_delay;
150
151 SYSCTL_U16(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base,
152 0, "");
153 SYSCTL_U16(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max,
154 0, "");
155
156 LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay);
157 #else
158 #define mtx_delay locks_delay
159 #endif
160 #endif
161
162 #ifdef MUTEX_SPIN_CUSTOM_BACKOFF
163 static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin,
164 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
165 "mtx spin debugging");
166
167 static struct lock_delay_config __read_frequently mtx_spin_delay;
168
169 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW,
170 &mtx_spin_delay.base, 0, "");
171 SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW,
172 &mtx_spin_delay.max, 0, "");
173
174 LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay);
175 #else
176 #define mtx_spin_delay locks_delay
177 #endif
178
179 /*
180 * System-wide mutexes
181 */
182 struct mtx blocked_lock;
183 struct mtx __exclusive_cache_line Giant;
184
185 static void _mtx_lock_indefinite_check(struct mtx *, struct lock_delay_arg *);
186
187 static void
assert_mtx(const struct lock_object * lock,int what)188 assert_mtx(const struct lock_object *lock, int what)
189 {
190
191 /*
192 * Treat LA_LOCKED as if LA_XLOCKED was asserted.
193 *
194 * Some callers of lc_assert uses LA_LOCKED to indicate that either
195 * a shared lock or write lock was held, while other callers uses
196 * the more strict LA_XLOCKED (used as MA_OWNED).
197 *
198 * Mutex is the only lock class that can not be shared, as a result,
199 * we can reasonably consider the caller really intends to assert
200 * LA_XLOCKED when they are asserting LA_LOCKED on a mutex object.
201 */
202 if (what & LA_LOCKED) {
203 what &= ~LA_LOCKED;
204 what |= LA_XLOCKED;
205 }
206 mtx_assert((const struct mtx *)lock, what);
207 }
208
209 static void
lock_mtx(struct lock_object * lock,uintptr_t how)210 lock_mtx(struct lock_object *lock, uintptr_t how)
211 {
212
213 mtx_lock((struct mtx *)lock);
214 }
215
216 static void
lock_spin(struct lock_object * lock,uintptr_t how)217 lock_spin(struct lock_object *lock, uintptr_t how)
218 {
219
220 mtx_lock_spin((struct mtx *)lock);
221 }
222
223 static int
trylock_mtx(struct lock_object * lock,uintptr_t how)224 trylock_mtx(struct lock_object *lock, uintptr_t how)
225 {
226
227 return (mtx_trylock((struct mtx *)lock));
228 }
229
230 static int
trylock_spin(struct lock_object * lock,uintptr_t how)231 trylock_spin(struct lock_object *lock, uintptr_t how)
232 {
233
234 return (mtx_trylock_spin((struct mtx *)lock));
235 }
236
237 static uintptr_t
unlock_mtx(struct lock_object * lock)238 unlock_mtx(struct lock_object *lock)
239 {
240 struct mtx *m;
241
242 m = (struct mtx *)lock;
243 mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
244 mtx_unlock(m);
245 return (0);
246 }
247
248 static uintptr_t
unlock_spin(struct lock_object * lock)249 unlock_spin(struct lock_object *lock)
250 {
251 struct mtx *m;
252
253 m = (struct mtx *)lock;
254 mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
255 mtx_unlock_spin(m);
256 return (0);
257 }
258
259 #ifdef KDTRACE_HOOKS
260 static int
owner_mtx(const struct lock_object * lock,struct thread ** owner)261 owner_mtx(const struct lock_object *lock, struct thread **owner)
262 {
263 const struct mtx *m;
264 uintptr_t x;
265
266 m = (const struct mtx *)lock;
267 x = m->mtx_lock;
268 *owner = (struct thread *)(x & ~MTX_FLAGMASK);
269 return (*owner != NULL);
270 }
271 #endif
272
273 /*
274 * Function versions of the inlined __mtx_* macros. These are used by
275 * modules and can also be called from assembly language if needed.
276 */
277 void
__mtx_lock_flags(volatile uintptr_t * c,int opts,const char * file,int line)278 __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
279 {
280 struct mtx *m;
281 uintptr_t tid, v;
282
283 m = mtxlock2mtx(c);
284
285 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
286 !TD_IS_IDLETHREAD(curthread),
287 ("mtx_lock() by idle thread %p on mutex %p @ %s:%d",
288 curthread, m, file, line));
289 KASSERT(m->mtx_lock != MTX_DESTROYED,
290 ("mtx_lock() of destroyed mutex %p @ %s:%d", m, file, line));
291 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_spin,
292 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
293 file, line));
294 WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) |
295 LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
296
297 tid = (uintptr_t)curthread;
298 v = MTX_UNOWNED;
299 if (!_mtx_obtain_lock_fetch(m, &v, tid))
300 _mtx_lock_sleep(m, v, opts, file, line);
301 else
302 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
303 m, 0, 0, file, line);
304 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
305 line);
306 WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE,
307 file, line);
308 TD_LOCKS_INC(curthread);
309 }
310
311 void
__mtx_unlock_flags(volatile uintptr_t * c,int opts,const char * file,int line)312 __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line)
313 {
314 struct mtx *m;
315
316 m = mtxlock2mtx(c);
317
318 KASSERT(m->mtx_lock != MTX_DESTROYED,
319 ("mtx_unlock() of destroyed mutex %p @ %s:%d", m, file, line));
320 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_spin,
321 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
322 file, line));
323 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
324 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
325 line);
326 mtx_assert(m, MA_OWNED);
327
328 #ifdef LOCK_PROFILING
329 __mtx_unlock_sleep(c, (uintptr_t)curthread, opts, file, line);
330 #else
331 __mtx_unlock(m, curthread, opts, file, line);
332 #endif
333 TD_LOCKS_DEC(curthread);
334 }
335
336 void
__mtx_lock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)337 __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
338 int line)
339 {
340 struct mtx *m;
341 #ifdef SMP
342 uintptr_t tid, v;
343 #endif
344
345 m = mtxlock2mtx(c);
346
347 KASSERT(m->mtx_lock != MTX_DESTROYED,
348 ("mtx_lock_spin() of destroyed mutex %p @ %s:%d", m, file, line));
349 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_sleep,
350 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
351 m->lock_object.lo_name, file, line));
352 if (mtx_owned(m))
353 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
354 (opts & MTX_RECURSE) != 0,
355 ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n",
356 m->lock_object.lo_name, file, line));
357 opts &= ~MTX_RECURSE;
358 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
359 file, line, NULL);
360 #ifdef SMP
361 spinlock_enter();
362 tid = (uintptr_t)curthread;
363 v = MTX_UNOWNED;
364 if (!_mtx_obtain_lock_fetch(m, &v, tid))
365 _mtx_lock_spin(m, v, opts, file, line);
366 else
367 LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire,
368 m, 0, 0, file, line);
369 #else
370 __mtx_lock_spin(m, curthread, opts, file, line);
371 #endif
372 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
373 line);
374 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
375 }
376
377 int
__mtx_trylock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)378 __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
379 int line)
380 {
381 struct mtx *m;
382
383 if (SCHEDULER_STOPPED())
384 return (1);
385
386 m = mtxlock2mtx(c);
387
388 KASSERT(m->mtx_lock != MTX_DESTROYED,
389 ("mtx_trylock_spin() of destroyed mutex %p @ %s:%d", m, file,
390 line));
391 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_sleep,
392 ("mtx_trylock_spin() of sleep mutex %s @ %s:%d",
393 m->lock_object.lo_name, file, line));
394 KASSERT((opts & MTX_RECURSE) == 0,
395 ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n",
396 m->lock_object.lo_name, file, line));
397 if (__mtx_trylock_spin(m, curthread, opts, file, line)) {
398 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line);
399 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
400 return (1);
401 }
402 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line);
403 return (0);
404 }
405
406 void
__mtx_unlock_spin_flags(volatile uintptr_t * c,int opts,const char * file,int line)407 __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file,
408 int line)
409 {
410 struct mtx *m;
411
412 m = mtxlock2mtx(c);
413
414 KASSERT(m->mtx_lock != MTX_DESTROYED,
415 ("mtx_unlock_spin() of destroyed mutex %p @ %s:%d", m, file,
416 line));
417 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_sleep,
418 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
419 m->lock_object.lo_name, file, line));
420 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
421 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
422 line);
423 mtx_assert(m, MA_OWNED);
424
425 __mtx_unlock_spin(m);
426 }
427
428 /*
429 * The important part of mtx_trylock{,_flags}()
430 * Tries to acquire lock `m.' If this function is called on a mutex that
431 * is already owned, it will recursively acquire the lock.
432 */
433 int
_mtx_trylock_flags_int(struct mtx * m,int opts LOCK_FILE_LINE_ARG_DEF)434 _mtx_trylock_flags_int(struct mtx *m, int opts LOCK_FILE_LINE_ARG_DEF)
435 {
436 struct thread *td;
437 uintptr_t tid, v;
438 #ifdef LOCK_PROFILING
439 uint64_t waittime = 0;
440 int contested = 0;
441 #endif
442 int rval;
443 bool recursed;
444
445 td = curthread;
446 tid = (uintptr_t)td;
447 if (SCHEDULER_STOPPED())
448 return (1);
449
450 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
451 ("mtx_trylock() by idle thread %p on mutex %p @ %s:%d",
452 curthread, m, file, line));
453 KASSERT(m->mtx_lock != MTX_DESTROYED,
454 ("mtx_trylock() of destroyed mutex %p @ %s:%d", m, file, line));
455 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_spin,
456 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
457 file, line));
458
459 rval = 1;
460 recursed = false;
461 v = MTX_UNOWNED;
462 for (;;) {
463 if (_mtx_obtain_lock_fetch(m, &v, tid))
464 break;
465 if (v == MTX_UNOWNED)
466 continue;
467 if (v == tid &&
468 ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
469 (opts & MTX_RECURSE) != 0)) {
470 m->mtx_recurse++;
471 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
472 recursed = true;
473 break;
474 }
475 rval = 0;
476 break;
477 }
478
479 opts &= ~MTX_RECURSE;
480
481 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
482 if (rval) {
483 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
484 file, line);
485 TD_LOCKS_INC(curthread);
486 if (!recursed)
487 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire,
488 m, contested, waittime, file, line);
489 }
490
491 return (rval);
492 }
493
494 int
_mtx_trylock_flags_(volatile uintptr_t * c,int opts,const char * file,int line)495 _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line)
496 {
497 struct mtx *m;
498
499 m = mtxlock2mtx(c);
500 return (_mtx_trylock_flags_int(m, opts LOCK_FILE_LINE_ARG));
501 }
502
503 /*
504 * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
505 *
506 * We get here if lock profiling is enabled, the lock is already held by
507 * someone else or we are recursing on it.
508 */
509 #if LOCK_DEBUG > 0
510 void
__mtx_lock_sleep(volatile uintptr_t * c,uintptr_t v,int opts,const char * file,int line)511 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, int opts, const char *file,
512 int line)
513 #else
514 void
515 __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v)
516 #endif
517 {
518 struct thread *td;
519 struct mtx *m;
520 struct turnstile *ts;
521 uintptr_t tid;
522 struct thread *owner;
523 #ifdef LOCK_PROFILING
524 int contested = 0;
525 uint64_t waittime = 0;
526 #endif
527 #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
528 struct lock_delay_arg lda;
529 #endif
530 #ifdef KDTRACE_HOOKS
531 u_int sleep_cnt = 0;
532 int64_t sleep_time = 0;
533 int64_t all_time = 0;
534 #endif
535 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
536 int doing_lockprof = 0;
537 #endif
538
539 td = curthread;
540 tid = (uintptr_t)td;
541 m = mtxlock2mtx(c);
542
543 #ifdef KDTRACE_HOOKS
544 if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
545 while (v == MTX_UNOWNED) {
546 if (_mtx_obtain_lock_fetch(m, &v, tid))
547 goto out_lockstat;
548 }
549 doing_lockprof = 1;
550 all_time -= lockstat_nsecs(&m->lock_object);
551 }
552 #endif
553 #ifdef LOCK_PROFILING
554 doing_lockprof = 1;
555 #endif
556
557 if (SCHEDULER_STOPPED())
558 return;
559
560 if (__predict_false(v == MTX_UNOWNED))
561 v = MTX_READ_VALUE(m);
562
563 if (__predict_false(lv_mtx_owner(v) == td)) {
564 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
565 (opts & MTX_RECURSE) != 0,
566 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
567 m->lock_object.lo_name, file, line));
568 #if LOCK_DEBUG > 0
569 opts &= ~MTX_RECURSE;
570 #endif
571 m->mtx_recurse++;
572 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
573 if (LOCK_LOG_TEST(&m->lock_object, opts))
574 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
575 return;
576 }
577 #if LOCK_DEBUG > 0
578 opts &= ~MTX_RECURSE;
579 #endif
580
581 #if defined(ADAPTIVE_MUTEXES)
582 lock_delay_arg_init(&lda, &mtx_delay);
583 #elif defined(KDTRACE_HOOKS)
584 lock_delay_arg_init_noadapt(&lda);
585 #endif
586
587 #ifdef HWPMC_HOOKS
588 PMC_SOFT_CALL( , , lock, failed);
589 #endif
590 lock_profile_obtain_lock_failed(&m->lock_object, false,
591 &contested, &waittime);
592 if (LOCK_LOG_TEST(&m->lock_object, opts))
593 CTR4(KTR_LOCK,
594 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
595 m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
596
597 THREAD_CONTENDS_ON_LOCK(&m->lock_object);
598
599 for (;;) {
600 if (v == MTX_UNOWNED) {
601 if (_mtx_obtain_lock_fetch(m, &v, tid))
602 break;
603 continue;
604 }
605 #ifdef KDTRACE_HOOKS
606 lda.spin_cnt++;
607 #endif
608 #ifdef ADAPTIVE_MUTEXES
609 /*
610 * If the owner is running on another CPU, spin until the
611 * owner stops running or the state of the lock changes.
612 */
613 owner = lv_mtx_owner(v);
614 if (TD_IS_RUNNING(owner)) {
615 if (LOCK_LOG_TEST(&m->lock_object, 0))
616 CTR3(KTR_LOCK,
617 "%s: spinning on %p held by %p",
618 __func__, m, owner);
619 KTR_STATE1(KTR_SCHED, "thread",
620 sched_tdname((struct thread *)tid),
621 "spinning", "lockname:\"%s\"",
622 m->lock_object.lo_name);
623 do {
624 lock_delay(&lda);
625 v = MTX_READ_VALUE(m);
626 owner = lv_mtx_owner(v);
627 } while (v != MTX_UNOWNED && TD_IS_RUNNING(owner));
628 KTR_STATE0(KTR_SCHED, "thread",
629 sched_tdname((struct thread *)tid),
630 "running");
631 continue;
632 }
633 #endif
634
635 ts = turnstile_trywait(&m->lock_object);
636 v = MTX_READ_VALUE(m);
637 retry_turnstile:
638
639 /*
640 * Check if the lock has been released while spinning for
641 * the turnstile chain lock.
642 */
643 if (v == MTX_UNOWNED) {
644 turnstile_cancel(ts);
645 continue;
646 }
647
648 #ifdef ADAPTIVE_MUTEXES
649 /*
650 * The current lock owner might have started executing
651 * on another CPU (or the lock could have changed
652 * owners) while we were waiting on the turnstile
653 * chain lock. If so, drop the turnstile lock and try
654 * again.
655 */
656 owner = lv_mtx_owner(v);
657 if (TD_IS_RUNNING(owner)) {
658 turnstile_cancel(ts);
659 continue;
660 }
661 #endif
662
663 if ((v & MTX_WAITERS) == 0 &&
664 !atomic_fcmpset_ptr(&m->mtx_lock, &v, v | MTX_WAITERS)) {
665 goto retry_turnstile;
666 }
667
668 /*
669 * We definitely must sleep for this lock.
670 */
671 mtx_assert(m, MA_NOTOWNED);
672
673 /*
674 * Block on the turnstile.
675 */
676 #ifdef KDTRACE_HOOKS
677 sleep_time -= lockstat_nsecs(&m->lock_object);
678 #endif
679 #ifndef ADAPTIVE_MUTEXES
680 owner = mtx_owner(m);
681 #endif
682 MPASS(owner == mtx_owner(m));
683 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
684 #ifdef KDTRACE_HOOKS
685 sleep_time += lockstat_nsecs(&m->lock_object);
686 sleep_cnt++;
687 #endif
688 v = MTX_READ_VALUE(m);
689 }
690 THREAD_CONTENTION_DONE(&m->lock_object);
691 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
692 if (__predict_true(!doing_lockprof))
693 return;
694 #endif
695 #ifdef KDTRACE_HOOKS
696 all_time += lockstat_nsecs(&m->lock_object);
697 if (sleep_time)
698 LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
699
700 /*
701 * Only record the loops spinning and not sleeping.
702 */
703 if (lda.spin_cnt > sleep_cnt)
704 LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
705 out_lockstat:
706 #endif
707 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
708 waittime, file, line);
709 }
710
711 #ifdef SMP
712 /*
713 * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
714 *
715 * This is only called if we need to actually spin for the lock. Recursion
716 * is handled inline.
717 */
718 #if LOCK_DEBUG > 0
719 void
_mtx_lock_spin_cookie(volatile uintptr_t * c,uintptr_t v,int opts,const char * file,int line)720 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, int opts,
721 const char *file, int line)
722 #else
723 void
724 _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v)
725 #endif
726 {
727 struct mtx *m;
728 struct lock_delay_arg lda;
729 uintptr_t tid;
730 #ifdef LOCK_PROFILING
731 int contested = 0;
732 uint64_t waittime = 0;
733 #endif
734 #ifdef KDTRACE_HOOKS
735 int64_t spin_time = 0;
736 #endif
737 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
738 int doing_lockprof = 0;
739 #endif
740
741 tid = (uintptr_t)curthread;
742 m = mtxlock2mtx(c);
743
744 #ifdef KDTRACE_HOOKS
745 if (LOCKSTAT_PROFILE_ENABLED(adaptive__acquire)) {
746 while (v == MTX_UNOWNED) {
747 if (_mtx_obtain_lock_fetch(m, &v, tid))
748 goto out_lockstat;
749 }
750 doing_lockprof = 1;
751 spin_time -= lockstat_nsecs(&m->lock_object);
752 }
753 #endif
754 #ifdef LOCK_PROFILING
755 doing_lockprof = 1;
756 #endif
757
758 if (__predict_false(v == MTX_UNOWNED))
759 v = MTX_READ_VALUE(m);
760
761 if (__predict_false(v == tid)) {
762 m->mtx_recurse++;
763 return;
764 }
765
766 if (SCHEDULER_STOPPED())
767 return;
768
769 if (LOCK_LOG_TEST(&m->lock_object, opts))
770 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
771 KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
772 "spinning", "lockname:\"%s\"", m->lock_object.lo_name);
773
774 lock_delay_arg_init(&lda, &mtx_spin_delay);
775
776 #ifdef HWPMC_HOOKS
777 PMC_SOFT_CALL( , , lock, failed);
778 #endif
779 lock_profile_obtain_lock_failed(&m->lock_object, true, &contested, &waittime);
780
781 for (;;) {
782 if (v == MTX_UNOWNED) {
783 if (_mtx_obtain_lock_fetch(m, &v, tid))
784 break;
785 continue;
786 }
787 /* Give interrupts a chance while we spin. */
788 spinlock_exit();
789 do {
790 if (__predict_true(lda.spin_cnt < 10000000)) {
791 lock_delay(&lda);
792 } else {
793 _mtx_lock_indefinite_check(m, &lda);
794 }
795 v = MTX_READ_VALUE(m);
796 } while (v != MTX_UNOWNED);
797 spinlock_enter();
798 }
799
800 if (LOCK_LOG_TEST(&m->lock_object, opts))
801 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
802 KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
803 "running");
804
805 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
806 if (__predict_true(!doing_lockprof))
807 return;
808 #endif
809 #ifdef KDTRACE_HOOKS
810 spin_time += lockstat_nsecs(&m->lock_object);
811 if (lda.spin_cnt != 0)
812 LOCKSTAT_RECORD1(spin__spin, m, spin_time);
813 out_lockstat:
814 #endif
815 LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire, m,
816 contested, waittime, file, line);
817 }
818 #endif /* SMP */
819
820 #ifdef INVARIANTS
821 static void
thread_lock_validate(struct mtx * m,int opts,const char * file,int line)822 thread_lock_validate(struct mtx *m, int opts, const char *file, int line)
823 {
824
825 KASSERT(m->mtx_lock != MTX_DESTROYED,
826 ("thread_lock() of destroyed mutex %p @ %s:%d", m, file, line));
827 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_sleep,
828 ("thread_lock() of sleep mutex %s @ %s:%d",
829 m->lock_object.lo_name, file, line));
830 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) == 0,
831 ("thread_lock: got a recursive mutex %s @ %s:%d\n",
832 m->lock_object.lo_name, file, line));
833 WITNESS_CHECKORDER(&m->lock_object,
834 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
835 }
836 #else
837 #define thread_lock_validate(m, opts, file, line) do { } while (0)
838 #endif
839
840 #ifndef LOCK_PROFILING
841 #if LOCK_DEBUG > 0
842 void
_thread_lock(struct thread * td,int opts,const char * file,int line)843 _thread_lock(struct thread *td, int opts, const char *file, int line)
844 #else
845 void
846 _thread_lock(struct thread *td)
847 #endif
848 {
849 struct mtx *m;
850 uintptr_t tid;
851
852 tid = (uintptr_t)curthread;
853
854 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(spin__acquire)))
855 goto slowpath_noirq;
856 spinlock_enter();
857 m = td->td_lock;
858 thread_lock_validate(m, 0, file, line);
859 if (__predict_false(m == &blocked_lock))
860 goto slowpath_unlocked;
861 if (__predict_false(!_mtx_obtain_lock(m, tid)))
862 goto slowpath_unlocked;
863 if (__predict_true(m == td->td_lock)) {
864 WITNESS_LOCK(&m->lock_object, LOP_EXCLUSIVE, file, line);
865 return;
866 }
867 atomic_store_rel_ptr(&m->mtx_lock, MTX_UNOWNED);
868 slowpath_unlocked:
869 spinlock_exit();
870 slowpath_noirq:
871 #if LOCK_DEBUG > 0
872 thread_lock_flags_(td, opts, file, line);
873 #else
874 thread_lock_flags_(td, 0, 0, 0);
875 #endif
876 }
877 #endif
878
879 void
thread_lock_flags_(struct thread * td,int opts,const char * file,int line)880 thread_lock_flags_(struct thread *td, int opts, const char *file, int line)
881 {
882 struct mtx *m;
883 uintptr_t tid, v;
884 struct lock_delay_arg lda;
885 #ifdef LOCK_PROFILING
886 int contested = 0;
887 uint64_t waittime = 0;
888 #endif
889 #ifdef KDTRACE_HOOKS
890 int64_t spin_time = 0;
891 #endif
892 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
893 int doing_lockprof = 1;
894 #endif
895
896 tid = (uintptr_t)curthread;
897
898 if (SCHEDULER_STOPPED()) {
899 /*
900 * Ensure that spinlock sections are balanced even when the
901 * scheduler is stopped, since we may otherwise inadvertently
902 * re-enable interrupts while dumping core.
903 */
904 spinlock_enter();
905 return;
906 }
907
908 lock_delay_arg_init(&lda, &mtx_spin_delay);
909
910 #ifdef HWPMC_HOOKS
911 PMC_SOFT_CALL( , , lock, failed);
912 #endif
913
914 #ifdef LOCK_PROFILING
915 doing_lockprof = 1;
916 #elif defined(KDTRACE_HOOKS)
917 doing_lockprof = lockstat_enabled;
918 #endif
919 #ifdef KDTRACE_HOOKS
920 if (__predict_false(doing_lockprof))
921 spin_time -= lockstat_nsecs(&td->td_lock->lock_object);
922 #endif
923 spinlock_enter();
924
925 for (;;) {
926 retry:
927 m = td->td_lock;
928 thread_lock_validate(m, opts, file, line);
929 v = MTX_READ_VALUE(m);
930 for (;;) {
931 if (v == MTX_UNOWNED) {
932 if (_mtx_obtain_lock_fetch(m, &v, tid))
933 break;
934 continue;
935 }
936 MPASS(v != tid);
937 lock_profile_obtain_lock_failed(&m->lock_object, true,
938 &contested, &waittime);
939 /* Give interrupts a chance while we spin. */
940 spinlock_exit();
941 do {
942 if (__predict_true(lda.spin_cnt < 10000000)) {
943 lock_delay(&lda);
944 } else {
945 _mtx_lock_indefinite_check(m, &lda);
946 }
947 if (m != td->td_lock) {
948 spinlock_enter();
949 goto retry;
950 }
951 v = MTX_READ_VALUE(m);
952 } while (v != MTX_UNOWNED);
953 spinlock_enter();
954 }
955 if (m == td->td_lock)
956 break;
957 atomic_store_rel_ptr(&m->mtx_lock, MTX_UNOWNED);
958 }
959 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
960 line);
961 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
962
963 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
964 if (__predict_true(!doing_lockprof))
965 return;
966 #endif
967 #ifdef KDTRACE_HOOKS
968 spin_time += lockstat_nsecs(&m->lock_object);
969 #endif
970 LOCKSTAT_PROFILE_OBTAIN_SPIN_LOCK_SUCCESS(spin__acquire, m, contested,
971 waittime, file, line);
972 #ifdef KDTRACE_HOOKS
973 if (lda.spin_cnt != 0)
974 LOCKSTAT_RECORD1(thread__spin, m, spin_time);
975 #endif
976 }
977
978 struct mtx *
thread_lock_block(struct thread * td)979 thread_lock_block(struct thread *td)
980 {
981 struct mtx *lock;
982
983 lock = td->td_lock;
984 mtx_assert(lock, MA_OWNED);
985 td->td_lock = &blocked_lock;
986
987 return (lock);
988 }
989
990 void
thread_lock_unblock(struct thread * td,struct mtx * new)991 thread_lock_unblock(struct thread *td, struct mtx *new)
992 {
993
994 mtx_assert(new, MA_OWNED);
995 KASSERT(td->td_lock == &blocked_lock,
996 ("thread %p lock %p not blocked_lock %p",
997 td, td->td_lock, &blocked_lock));
998 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
999 }
1000
1001 void
thread_lock_block_wait(struct thread * td)1002 thread_lock_block_wait(struct thread *td)
1003 {
1004
1005 while (td->td_lock == &blocked_lock)
1006 cpu_spinwait();
1007
1008 /* Acquire fence to be certain that all thread state is visible. */
1009 atomic_thread_fence_acq();
1010 }
1011
1012 void
thread_lock_set(struct thread * td,struct mtx * new)1013 thread_lock_set(struct thread *td, struct mtx *new)
1014 {
1015 struct mtx *lock;
1016
1017 mtx_assert(new, MA_OWNED);
1018 lock = td->td_lock;
1019 mtx_assert(lock, MA_OWNED);
1020 td->td_lock = new;
1021 mtx_unlock_spin(lock);
1022 }
1023
1024 /*
1025 * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
1026 *
1027 * We get here if lock profiling is enabled, the lock is already held by
1028 * someone else or we are recursing on it.
1029 */
1030 #if LOCK_DEBUG > 0
1031 void
__mtx_unlock_sleep(volatile uintptr_t * c,uintptr_t v,int opts,const char * file,int line)1032 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v, int opts,
1033 const char *file, int line)
1034 #else
1035 void
1036 __mtx_unlock_sleep(volatile uintptr_t *c, uintptr_t v)
1037 #endif
1038 {
1039 struct mtx *m;
1040 struct turnstile *ts;
1041 uintptr_t tid;
1042
1043 if (SCHEDULER_STOPPED())
1044 return;
1045
1046 tid = (uintptr_t)curthread;
1047 m = mtxlock2mtx(c);
1048
1049 if (__predict_false(v == tid))
1050 v = MTX_READ_VALUE(m);
1051
1052 if (__predict_false(v & MTX_RECURSED)) {
1053 if (--(m->mtx_recurse) == 0)
1054 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
1055 if (LOCK_LOG_TEST(&m->lock_object, opts))
1056 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
1057 return;
1058 }
1059
1060 LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m);
1061 if (v == tid && _mtx_release_lock(m, tid))
1062 return;
1063
1064 /*
1065 * We have to lock the chain before the turnstile so this turnstile
1066 * can be removed from the hash list if it is empty.
1067 */
1068 turnstile_chain_lock(&m->lock_object);
1069 atomic_store_rel_ptr(&m->mtx_lock, MTX_UNOWNED);
1070 ts = turnstile_lookup(&m->lock_object);
1071 MPASS(ts != NULL);
1072 if (LOCK_LOG_TEST(&m->lock_object, opts))
1073 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
1074 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
1075
1076 /*
1077 * This turnstile is now no longer associated with the mutex. We can
1078 * unlock the chain lock so a new turnstile may take it's place.
1079 */
1080 turnstile_unpend(ts);
1081 turnstile_chain_unlock(&m->lock_object);
1082 }
1083
1084 /*
1085 * All the unlocking of MTX_SPIN locks is done inline.
1086 * See the __mtx_unlock_spin() macro for the details.
1087 */
1088
1089 /*
1090 * The backing function for the INVARIANTS-enabled mtx_assert()
1091 */
1092 #ifdef INVARIANT_SUPPORT
1093 void
__mtx_assert(const volatile uintptr_t * c,int what,const char * file,int line)1094 __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1095 {
1096 const struct mtx *m;
1097
1098 if (KERNEL_PANICKED() || dumping || SCHEDULER_STOPPED())
1099 return;
1100
1101 m = mtxlock2mtx(c);
1102
1103 switch (what) {
1104 case MA_OWNED:
1105 case MA_OWNED | MA_RECURSED:
1106 case MA_OWNED | MA_NOTRECURSED:
1107 if (!mtx_owned(m))
1108 panic("mutex %s not owned at %s:%d",
1109 m->lock_object.lo_name, file, line);
1110 if (mtx_recursed(m)) {
1111 if ((what & MA_NOTRECURSED) != 0)
1112 panic("mutex %s recursed at %s:%d",
1113 m->lock_object.lo_name, file, line);
1114 } else if ((what & MA_RECURSED) != 0) {
1115 panic("mutex %s unrecursed at %s:%d",
1116 m->lock_object.lo_name, file, line);
1117 }
1118 break;
1119 case MA_NOTOWNED:
1120 if (mtx_owned(m))
1121 panic("mutex %s owned at %s:%d",
1122 m->lock_object.lo_name, file, line);
1123 break;
1124 default:
1125 panic("unknown mtx_assert at %s:%d", file, line);
1126 }
1127 }
1128 #endif
1129
1130 /*
1131 * General init routine used by the MTX_SYSINIT() macro.
1132 */
1133 void
mtx_sysinit(const void * arg)1134 mtx_sysinit(const void *arg)
1135 {
1136 const struct mtx_args *margs = arg;
1137
1138 mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL,
1139 margs->ma_opts);
1140 }
1141
1142 /*
1143 * Mutex initialization routine; initialize lock `m' of type contained in
1144 * `opts' with options contained in `opts' and name `name.' The optional
1145 * lock type `type' is used as a general lock category name for use with
1146 * witness.
1147 */
1148 void
_mtx_init(volatile uintptr_t * c,const char * name,const char * type,int opts)1149 _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts)
1150 {
1151 struct mtx *m;
1152 struct lock_class *class;
1153 int flags;
1154
1155 m = mtxlock2mtx(c);
1156
1157 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
1158 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0);
1159 ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock,
1160 ("%s: mtx_lock not aligned for %s: %p", __func__, name,
1161 &m->mtx_lock));
1162
1163 /* Determine lock class and lock flags. */
1164 if (opts & MTX_SPIN)
1165 class = &lock_class_mtx_spin;
1166 else
1167 class = &lock_class_mtx_sleep;
1168 flags = 0;
1169 if (opts & MTX_QUIET)
1170 flags |= LO_QUIET;
1171 if (opts & MTX_RECURSE)
1172 flags |= LO_RECURSABLE;
1173 if ((opts & MTX_NOWITNESS) == 0)
1174 flags |= LO_WITNESS;
1175 if (opts & MTX_DUPOK)
1176 flags |= LO_DUPOK;
1177 if (opts & MTX_NOPROFILE)
1178 flags |= LO_NOPROFILE;
1179 if (opts & MTX_NEW)
1180 flags |= LO_NEW;
1181
1182 /* Initialize mutex. */
1183 lock_init(&m->lock_object, class, name, type, flags);
1184
1185 m->mtx_lock = MTX_UNOWNED;
1186 m->mtx_recurse = 0;
1187 }
1188
1189 /*
1190 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
1191 * passed in as a flag here because if the corresponding mtx_init() was
1192 * called with MTX_QUIET set, then it will already be set in the mutex's
1193 * flags.
1194 */
1195 void
_mtx_destroy(volatile uintptr_t * c)1196 _mtx_destroy(volatile uintptr_t *c)
1197 {
1198 struct mtx *m;
1199
1200 m = mtxlock2mtx(c);
1201
1202 if (!mtx_owned(m))
1203 MPASS(mtx_unowned(m));
1204 else {
1205 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_WAITERS)) == 0);
1206
1207 /* Perform the non-mtx related part of mtx_unlock_spin(). */
1208 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) {
1209 lock_profile_release_lock(&m->lock_object, true);
1210 spinlock_exit();
1211 } else {
1212 TD_LOCKS_DEC(curthread);
1213 lock_profile_release_lock(&m->lock_object, false);
1214 }
1215
1216 /* Tell witness this isn't locked to make it happy. */
1217 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
1218 __LINE__);
1219 }
1220
1221 m->mtx_lock = MTX_DESTROYED;
1222 lock_destroy(&m->lock_object);
1223 }
1224
1225 /*
1226 * Intialize the mutex code and system mutexes. This is called from the MD
1227 * startup code prior to mi_startup(). The per-CPU data space needs to be
1228 * setup before this is called.
1229 */
1230 void
mutex_init(void)1231 mutex_init(void)
1232 {
1233
1234 /* Setup turnstiles so that sleep mutexes work. */
1235 init_turnstiles();
1236
1237 /*
1238 * Initialize mutexes.
1239 */
1240 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
1241 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
1242 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */
1243 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
1244 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN);
1245 mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN);
1246 mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN);
1247 mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN);
1248 mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
1249 mtx_lock(&Giant);
1250 }
1251
1252 static void __noinline
_mtx_lock_indefinite_check(struct mtx * m,struct lock_delay_arg * ldap)1253 _mtx_lock_indefinite_check(struct mtx *m, struct lock_delay_arg *ldap)
1254 {
1255 struct thread *td;
1256
1257 ldap->spin_cnt++;
1258 if (ldap->spin_cnt < 60000000 || kdb_active || KERNEL_PANICKED())
1259 cpu_lock_delay();
1260 else {
1261 td = mtx_owner(m);
1262
1263 /* If the mutex is unlocked, try again. */
1264 if (td == NULL)
1265 return;
1266
1267 printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
1268 m, m->lock_object.lo_name, td, td->td_tid);
1269 #ifdef WITNESS
1270 witness_display_spinlock(&m->lock_object, td, printf);
1271 #endif
1272 panic("spin lock held too long");
1273 }
1274 cpu_spinwait();
1275 }
1276
1277 void
mtx_spin_wait_unlocked(struct mtx * m)1278 mtx_spin_wait_unlocked(struct mtx *m)
1279 {
1280 struct lock_delay_arg lda;
1281
1282 KASSERT(m->mtx_lock != MTX_DESTROYED,
1283 ("%s() of destroyed mutex %p", __func__, m));
1284 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_sleep,
1285 ("%s() of sleep mutex %p (%s)", __func__, m,
1286 m->lock_object.lo_name));
1287 KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1288 m->lock_object.lo_name));
1289
1290 lda.spin_cnt = 0;
1291
1292 while (atomic_load_acq_ptr(&m->mtx_lock) != MTX_UNOWNED) {
1293 if (__predict_true(lda.spin_cnt < 10000000)) {
1294 cpu_spinwait();
1295 lda.spin_cnt++;
1296 } else {
1297 _mtx_lock_indefinite_check(m, &lda);
1298 }
1299 }
1300 }
1301
1302 void
mtx_wait_unlocked(struct mtx * m)1303 mtx_wait_unlocked(struct mtx *m)
1304 {
1305 struct thread *owner;
1306 uintptr_t v;
1307
1308 KASSERT(m->mtx_lock != MTX_DESTROYED,
1309 ("%s() of destroyed mutex %p", __func__, m));
1310 KASSERT(LOCK_CLASS(&m->lock_object) != &lock_class_mtx_spin,
1311 ("%s() of spin mutex %p (%s)", __func__, m,
1312 m->lock_object.lo_name));
1313 KASSERT(!mtx_owned(m), ("%s() waiting on myself on lock %p (%s)", __func__, m,
1314 m->lock_object.lo_name));
1315
1316 for (;;) {
1317 v = atomic_load_acq_ptr(&m->mtx_lock);
1318 if (v == MTX_UNOWNED) {
1319 break;
1320 }
1321 owner = lv_mtx_owner(v);
1322 if (!TD_IS_RUNNING(owner)) {
1323 mtx_lock(m);
1324 mtx_unlock(m);
1325 break;
1326 }
1327 cpu_spinwait();
1328 }
1329 }
1330
1331 #ifdef DDB
1332 static void
db_show_mtx(const struct lock_object * lock)1333 db_show_mtx(const struct lock_object *lock)
1334 {
1335 struct thread *td;
1336 const struct mtx *m;
1337
1338 m = (const struct mtx *)lock;
1339
1340 db_printf(" flags: {");
1341 if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
1342 db_printf("SPIN");
1343 else
1344 db_printf("DEF");
1345 if (m->lock_object.lo_flags & LO_RECURSABLE)
1346 db_printf(", RECURSE");
1347 if (m->lock_object.lo_flags & LO_DUPOK)
1348 db_printf(", DUPOK");
1349 db_printf("}\n");
1350 db_printf(" state: {");
1351 if (mtx_unowned(m))
1352 db_printf("UNOWNED");
1353 else if (mtx_destroyed(m))
1354 db_printf("DESTROYED");
1355 else {
1356 db_printf("OWNED");
1357 if (m->mtx_lock & MTX_WAITERS)
1358 db_printf(", WAITERS");
1359 if (m->mtx_lock & MTX_RECURSED)
1360 db_printf(", RECURSED");
1361 }
1362 db_printf("}\n");
1363 if (!mtx_unowned(m) && !mtx_destroyed(m)) {
1364 td = mtx_owner(m);
1365 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
1366 td->td_tid, td->td_proc->p_pid, td->td_name);
1367 if (mtx_recursed(m))
1368 db_printf(" recursed: %d\n", m->mtx_recurse);
1369 }
1370 }
1371 #endif
1372