FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_mutex.c
1 /*-
2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 * 3. Berkeley Software Design Inc's name may not be used to endorse or
13 * promote products derived from this software without specific prior
14 * written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 *
28 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30 */
31
32 /*
33 * Machine independent bits of mutex implementation.
34 */
35
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD: src/sys/kern/kern_mutex.c,v 1.147.2.4 2005/03/28 20:17:30 jhb Exp $");
38
39 #include "opt_adaptive_mutexes.h"
40 #include "opt_ddb.h"
41 #include "opt_mprof.h"
42 #include "opt_mutex_wake_all.h"
43 #include "opt_sched.h"
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/bus.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/sysctl.h>
59 #include <sys/turnstile.h>
60 #include <sys/vmmeter.h>
61
62 #include <machine/atomic.h>
63 #include <machine/bus.h>
64 #include <machine/clock.h>
65 #include <machine/cpu.h>
66
67 #include <ddb/ddb.h>
68
69 #include <vm/vm.h>
70 #include <vm/vm_extern.h>
71
72 /*
73 * Force MUTEX_WAKE_ALL for now.
74 * single thread wakeup needs fixes to avoid race conditions with
75 * priority inheritance.
76 */
77 #ifndef MUTEX_WAKE_ALL
78 #define MUTEX_WAKE_ALL
79 #endif
80
81 /*
82 * Internal utility macros.
83 */
84 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
85
86 #define mtx_owner(m) (mtx_unowned((m)) ? NULL \
87 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK))
88
89 /*
90 * Lock classes for sleep and spin mutexes.
91 */
92 struct lock_class lock_class_mtx_sleep = {
93 "sleep mutex",
94 LC_SLEEPLOCK | LC_RECURSABLE
95 };
96 struct lock_class lock_class_mtx_spin = {
97 "spin mutex",
98 LC_SPINLOCK | LC_RECURSABLE
99 };
100
101 /*
102 * System-wide mutexes
103 */
104 struct mtx sched_lock;
105 struct mtx Giant;
106
107 #ifdef MUTEX_PROFILING
108 SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging");
109 SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling");
110 static int mutex_prof_enable = 0;
111 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW,
112 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime");
113
114 struct mutex_prof {
115 const char *name;
116 const char *file;
117 int line;
118 uintmax_t cnt_max;
119 uintmax_t cnt_tot;
120 uintmax_t cnt_cur;
121 uintmax_t cnt_contest_holding;
122 uintmax_t cnt_contest_locking;
123 struct mutex_prof *next;
124 };
125
126 /*
127 * mprof_buf is a static pool of profiling records to avoid possible
128 * reentrance of the memory allocation functions.
129 *
130 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE.
131 */
132 #ifdef MPROF_BUFFERS
133 #define NUM_MPROF_BUFFERS MPROF_BUFFERS
134 #else
135 #define NUM_MPROF_BUFFERS 1000
136 #endif
137 static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS];
138 static int first_free_mprof_buf;
139 #ifndef MPROF_HASH_SIZE
140 #define MPROF_HASH_SIZE 1009
141 #endif
142 #if NUM_MPROF_BUFFERS >= MPROF_HASH_SIZE
143 #error MPROF_BUFFERS must be larger than MPROF_HASH_SIZE
144 #endif
145 static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE];
146 /* SWAG: sbuf size = avg stat. line size * number of locks */
147 #define MPROF_SBUF_SIZE 256 * 400
148
149 static int mutex_prof_acquisitions;
150 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
151 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded");
152 static int mutex_prof_records;
153 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD,
154 &mutex_prof_records, 0, "Number of profiling records");
155 static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS;
156 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
157 &mutex_prof_maxrecords, 0, "Maximum number of profiling records");
158 static int mutex_prof_rejected;
159 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD,
160 &mutex_prof_rejected, 0, "Number of rejected profiling records");
161 static int mutex_prof_hashsize = MPROF_HASH_SIZE;
162 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD,
163 &mutex_prof_hashsize, 0, "Hash size");
164 static int mutex_prof_collisions = 0;
165 SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD,
166 &mutex_prof_collisions, 0, "Number of hash collisions");
167
168 /*
169 * mprof_mtx protects the profiling buffers and the hash.
170 */
171 static struct mtx mprof_mtx;
172 MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET);
173
174 static u_int64_t
175 nanoseconds(void)
176 {
177 struct timespec tv;
178
179 nanotime(&tv);
180 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
181 }
182
183 static int
184 dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
185 {
186 struct sbuf *sb;
187 int error, i;
188 static int multiplier = 1;
189
190 if (first_free_mprof_buf == 0)
191 return (SYSCTL_OUT(req, "No locking recorded",
192 sizeof("No locking recorded")));
193
194 retry_sbufops:
195 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
196 sbuf_printf(sb, "%6s %12s %11s %5s %12s %12s %s\n",
197 "max", "total", "count", "avg", "cnt_hold", "cnt_lock", "name");
198 /*
199 * XXX this spinlock seems to be by far the largest perpetrator
200 * of spinlock latency (1.6 msec on an Athlon1600 was recorded
201 * even before I pessimized it further by moving the average
202 * computation here).
203 */
204 mtx_lock_spin(&mprof_mtx);
205 for (i = 0; i < first_free_mprof_buf; ++i) {
206 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %12ju %12ju %s:%d (%s)\n",
207 mprof_buf[i].cnt_max / 1000,
208 mprof_buf[i].cnt_tot / 1000,
209 mprof_buf[i].cnt_cur,
210 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
211 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000),
212 mprof_buf[i].cnt_contest_holding,
213 mprof_buf[i].cnt_contest_locking,
214 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name);
215 if (sbuf_overflowed(sb)) {
216 mtx_unlock_spin(&mprof_mtx);
217 sbuf_delete(sb);
218 multiplier++;
219 goto retry_sbufops;
220 }
221 }
222 mtx_unlock_spin(&mprof_mtx);
223 sbuf_finish(sb);
224 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
225 sbuf_delete(sb);
226 return (error);
227 }
228 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
229 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics");
230
231 static int
232 reset_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
233 {
234 int error, v;
235
236 if (first_free_mprof_buf == 0)
237 return (0);
238
239 v = 0;
240 error = sysctl_handle_int(oidp, &v, 0, req);
241 if (error)
242 return (error);
243 if (req->newptr == NULL)
244 return (error);
245 if (v == 0)
246 return (0);
247
248 mtx_lock_spin(&mprof_mtx);
249 bzero(mprof_buf, sizeof(*mprof_buf) * first_free_mprof_buf);
250 bzero(mprof_hash, sizeof(struct mtx *) * MPROF_HASH_SIZE);
251 first_free_mprof_buf = 0;
252 mtx_unlock_spin(&mprof_mtx);
253 return (0);
254 }
255 SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
256 NULL, 0, reset_mutex_prof_stats, "I", "Reset mutex profiling statistics");
257 #endif
258
259 /*
260 * Function versions of the inlined __mtx_* macros. These are used by
261 * modules and can also be called from assembly language if needed.
262 */
263 void
264 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
265 {
266
267 MPASS(curthread != NULL);
268 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
269 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
270 file, line));
271 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
272 file, line);
273 _get_sleep_lock(m, curthread, opts, file, line);
274 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
275 line);
276 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
277 #ifdef MUTEX_PROFILING
278 /* don't reset the timer when/if recursing */
279 if (m->mtx_acqtime == 0) {
280 m->mtx_filename = file;
281 m->mtx_lineno = line;
282 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0;
283 ++mutex_prof_acquisitions;
284 }
285 #endif
286 }
287
288 void
289 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
290 {
291
292 MPASS(curthread != NULL);
293 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
294 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
295 file, line));
296 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
297 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
298 line);
299 mtx_assert(m, MA_OWNED);
300 #ifdef MUTEX_PROFILING
301 if (m->mtx_acqtime != 0) {
302 static const char *unknown = "(unknown)";
303 struct mutex_prof *mpp;
304 u_int64_t acqtime, now;
305 const char *p, *q;
306 volatile u_int hash;
307
308 now = nanoseconds();
309 acqtime = m->mtx_acqtime;
310 m->mtx_acqtime = 0;
311 if (now <= acqtime)
312 goto out;
313 for (p = m->mtx_filename;
314 p != NULL && strncmp(p, "../", 3) == 0; p += 3)
315 /* nothing */ ;
316 if (p == NULL || *p == '\0')
317 p = unknown;
318 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q)
319 hash = (hash * 2 + *q) % MPROF_HASH_SIZE;
320 mtx_lock_spin(&mprof_mtx);
321 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next)
322 if (mpp->line == m->mtx_lineno &&
323 strcmp(mpp->file, p) == 0)
324 break;
325 if (mpp == NULL) {
326 /* Just exit if we cannot get a trace buffer */
327 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) {
328 ++mutex_prof_rejected;
329 goto unlock;
330 }
331 mpp = &mprof_buf[first_free_mprof_buf++];
332 mpp->name = mtx_name(m);
333 mpp->file = p;
334 mpp->line = m->mtx_lineno;
335 mpp->next = mprof_hash[hash];
336 if (mprof_hash[hash] != NULL)
337 ++mutex_prof_collisions;
338 mprof_hash[hash] = mpp;
339 ++mutex_prof_records;
340 }
341 /*
342 * Record if the mutex has been held longer now than ever
343 * before.
344 */
345 if (now - acqtime > mpp->cnt_max)
346 mpp->cnt_max = now - acqtime;
347 mpp->cnt_tot += now - acqtime;
348 mpp->cnt_cur++;
349 /*
350 * There's a small race, really we should cmpxchg
351 * 0 with the current value, but that would bill
352 * the contention to the wrong lock instance if
353 * it followed this also.
354 */
355 mpp->cnt_contest_holding += m->mtx_contest_holding;
356 m->mtx_contest_holding = 0;
357 mpp->cnt_contest_locking += m->mtx_contest_locking;
358 m->mtx_contest_locking = 0;
359 unlock:
360 mtx_unlock_spin(&mprof_mtx);
361 }
362 out:
363 #endif
364 _rel_sleep_lock(m, curthread, opts, file, line);
365 }
366
367 void
368 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
369 {
370
371 MPASS(curthread != NULL);
372 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
373 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
374 m->mtx_object.lo_name, file, line));
375 WITNESS_CHECKORDER(&m->mtx_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
376 file, line);
377 _get_spin_lock(m, curthread, opts, file, line);
378 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
379 line);
380 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
381 }
382
383 void
384 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
385 {
386
387 MPASS(curthread != NULL);
388 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
389 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
390 m->mtx_object.lo_name, file, line));
391 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
392 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
393 line);
394 mtx_assert(m, MA_OWNED);
395 _rel_spin_lock(m);
396 }
397
398 /*
399 * The important part of mtx_trylock{,_flags}()
400 * Tries to acquire lock `m.' If this function is called on a mutex that
401 * is already owned, it will recursively acquire the lock.
402 */
403 int
404 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
405 {
406 int rval;
407
408 MPASS(curthread != NULL);
409
410 if (mtx_owned(m) && (m->mtx_object.lo_flags & LO_RECURSABLE) != 0) {
411 m->mtx_recurse++;
412 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
413 rval = 1;
414 } else
415 rval = _obtain_lock(m, curthread);
416
417 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line);
418 if (rval)
419 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
420 file, line);
421
422 return (rval);
423 }
424
425 /*
426 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
427 *
428 * We call this if the lock is either contested (i.e. we need to go to
429 * sleep waiting for it), or if we need to recurse on it.
430 */
431 void
432 _mtx_lock_sleep(struct mtx *m, struct thread *td, int opts, const char *file,
433 int line)
434 {
435 struct turnstile *ts;
436 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
437 struct thread *owner;
438 #endif
439 uintptr_t v;
440 #ifdef KTR
441 int cont_logged = 0;
442 #endif
443 #ifdef MUTEX_PROFILING
444 int contested;
445 #endif
446
447 if (mtx_owned(m)) {
448 KASSERT((m->mtx_object.lo_flags & LO_RECURSABLE) != 0,
449 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
450 m->mtx_object.lo_name, file, line));
451 m->mtx_recurse++;
452 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
453 if (LOCK_LOG_TEST(&m->mtx_object, opts))
454 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
455 return;
456 }
457
458 if (LOCK_LOG_TEST(&m->mtx_object, opts))
459 CTR4(KTR_LOCK,
460 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
461 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line);
462
463 #ifdef MUTEX_PROFILING
464 contested = 0;
465 #endif
466 while (!_obtain_lock(m, td)) {
467 #ifdef MUTEX_PROFILING
468 contested = 1;
469 atomic_add_int(&m->mtx_contest_holding, 1);
470 #endif
471 ts = turnstile_lookup(&m->mtx_object);
472 v = m->mtx_lock;
473
474 /*
475 * Check if the lock has been released while spinning for
476 * the turnstile chain lock.
477 */
478 if (v == MTX_UNOWNED) {
479 turnstile_release(&m->mtx_object);
480 cpu_spinwait();
481 continue;
482 }
483
484 #ifdef MUTEX_WAKE_ALL
485 MPASS(v != MTX_CONTESTED);
486 #else
487 /*
488 * The mutex was marked contested on release. This means that
489 * there are other threads blocked on it. Grab ownership of
490 * it and propagate its priority to the current thread if
491 * necessary.
492 */
493 if (v == MTX_CONTESTED) {
494 MPASS(ts != NULL);
495 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED;
496 turnstile_claim(ts);
497 break;
498 }
499 #endif
500
501 /*
502 * If the mutex isn't already contested and a failure occurs
503 * setting the contested bit, the mutex was either released
504 * or the state of the MTX_RECURSED bit changed.
505 */
506 if ((v & MTX_CONTESTED) == 0 &&
507 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v,
508 (void *)(v | MTX_CONTESTED))) {
509 turnstile_release(&m->mtx_object);
510 cpu_spinwait();
511 continue;
512 }
513
514 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
515 /*
516 * If the current owner of the lock is executing on another
517 * CPU, spin instead of blocking.
518 */
519 owner = (struct thread *)(v & MTX_FLAGMASK);
520 #ifdef ADAPTIVE_GIANT
521 if (TD_IS_RUNNING(owner)) {
522 #else
523 if (m != &Giant && TD_IS_RUNNING(owner)) {
524 #endif
525 turnstile_release(&m->mtx_object);
526 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
527 cpu_spinwait();
528 }
529 continue;
530 }
531 #endif /* SMP && !NO_ADAPTIVE_MUTEXES */
532
533 /*
534 * We definitely must sleep for this lock.
535 */
536 mtx_assert(m, MA_NOTOWNED);
537
538 #ifdef KTR
539 if (!cont_logged) {
540 CTR6(KTR_CONTENTION,
541 "contention: %p at %s:%d wants %s, taken by %s:%d",
542 td, file, line, m->mtx_object.lo_name,
543 WITNESS_FILE(&m->mtx_object),
544 WITNESS_LINE(&m->mtx_object));
545 cont_logged = 1;
546 }
547 #endif
548
549 /*
550 * Block on the turnstile.
551 */
552 turnstile_wait(ts, &m->mtx_object, mtx_owner(m));
553 }
554
555 #ifdef KTR
556 if (cont_logged) {
557 CTR4(KTR_CONTENTION,
558 "contention end: %s acquired by %p at %s:%d",
559 m->mtx_object.lo_name, td, file, line);
560 }
561 #endif
562 #ifdef MUTEX_PROFILING
563 if (contested)
564 m->mtx_contest_locking++;
565 m->mtx_contest_holding = 0;
566 #endif
567 return;
568 }
569
570 #ifdef SMP
571 /*
572 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
573 *
574 * This is only called if we need to actually spin for the lock. Recursion
575 * is handled inline.
576 */
577 void
578 _mtx_lock_spin(struct mtx *m, struct thread *td, int opts, const char *file,
579 int line)
580 {
581 int i = 0;
582
583 if (LOCK_LOG_TEST(&m->mtx_object, opts))
584 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
585
586 for (;;) {
587 if (_obtain_lock(m, td))
588 break;
589
590 /* Give interrupts a chance while we spin. */
591 critical_exit();
592 while (m->mtx_lock != MTX_UNOWNED) {
593 if (i++ < 10000000) {
594 cpu_spinwait();
595 continue;
596 }
597 if (i < 60000000)
598 DELAY(1);
599 else if (!kdb_active) {
600 printf("spin lock %s held by %p for > 5 seconds\n",
601 m->mtx_object.lo_name, (void *)m->mtx_lock);
602 #ifdef WITNESS
603 witness_display_spinlock(&m->mtx_object,
604 mtx_owner(m));
605 #endif
606 panic("spin lock held too long");
607 }
608 cpu_spinwait();
609 }
610 critical_enter();
611 }
612
613 if (LOCK_LOG_TEST(&m->mtx_object, opts))
614 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
615
616 return;
617 }
618 #endif /* SMP */
619
620 /*
621 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
622 *
623 * We are only called here if the lock is recursed or contested (i.e. we
624 * need to wake up a blocked thread).
625 */
626 void
627 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
628 {
629 struct turnstile *ts;
630 #ifndef PREEMPTION
631 struct thread *td, *td1;
632 #endif
633
634 if (mtx_recursed(m)) {
635 if (--(m->mtx_recurse) == 0)
636 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
637 if (LOCK_LOG_TEST(&m->mtx_object, opts))
638 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
639 return;
640 }
641
642 ts = turnstile_lookup(&m->mtx_object);
643 if (LOCK_LOG_TEST(&m->mtx_object, opts))
644 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
645
646 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
647 if (ts == NULL) {
648 _release_lock_quick(m);
649 if (LOCK_LOG_TEST(&m->mtx_object, opts))
650 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
651 turnstile_release(&m->mtx_object);
652 return;
653 }
654 #else
655 MPASS(ts != NULL);
656 #endif
657 #ifndef PREEMPTION
658 /* XXX */
659 td1 = turnstile_head(ts);
660 #endif
661 #ifdef MUTEX_WAKE_ALL
662 turnstile_broadcast(ts);
663 _release_lock_quick(m);
664 #else
665 if (turnstile_signal(ts)) {
666 _release_lock_quick(m);
667 if (LOCK_LOG_TEST(&m->mtx_object, opts))
668 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m);
669 } else {
670 m->mtx_lock = MTX_CONTESTED;
671 if (LOCK_LOG_TEST(&m->mtx_object, opts))
672 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p still contested",
673 m);
674 }
675 #endif
676 turnstile_unpend(ts);
677
678 #ifndef PREEMPTION
679 /*
680 * XXX: This is just a hack until preemption is done. However,
681 * once preemption is done we need to either wrap the
682 * turnstile_signal() and release of the actual lock in an
683 * extra critical section or change the preemption code to
684 * always just set a flag and never do instant-preempts.
685 */
686 td = curthread;
687 if (td->td_critnest > 0 || td1->td_priority >= td->td_priority)
688 return;
689 mtx_lock_spin(&sched_lock);
690 if (!TD_IS_RUNNING(td1)) {
691 #ifdef notyet
692 if (td->td_ithd != NULL) {
693 struct ithd *it = td->td_ithd;
694
695 if (it->it_interrupted) {
696 if (LOCK_LOG_TEST(&m->mtx_object, opts))
697 CTR2(KTR_LOCK,
698 "_mtx_unlock_sleep: %p interrupted %p",
699 it, it->it_interrupted);
700 intr_thd_fixup(it);
701 }
702 }
703 #endif
704 if (LOCK_LOG_TEST(&m->mtx_object, opts))
705 CTR2(KTR_LOCK,
706 "_mtx_unlock_sleep: %p switching out lock=%p", m,
707 (void *)m->mtx_lock);
708
709 mi_switch(SW_INVOL, NULL);
710 if (LOCK_LOG_TEST(&m->mtx_object, opts))
711 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p",
712 m, (void *)m->mtx_lock);
713 }
714 mtx_unlock_spin(&sched_lock);
715 #endif
716
717 return;
718 }
719
720 /*
721 * All the unlocking of MTX_SPIN locks is done inline.
722 * See the _rel_spin_lock() macro for the details.
723 */
724
725 /*
726 * The backing function for the INVARIANTS-enabled mtx_assert()
727 */
728 #ifdef INVARIANT_SUPPORT
729 void
730 _mtx_assert(struct mtx *m, int what, const char *file, int line)
731 {
732
733 if (panicstr != NULL)
734 return;
735 switch (what) {
736 case MA_OWNED:
737 case MA_OWNED | MA_RECURSED:
738 case MA_OWNED | MA_NOTRECURSED:
739 if (!mtx_owned(m))
740 panic("mutex %s not owned at %s:%d",
741 m->mtx_object.lo_name, file, line);
742 if (mtx_recursed(m)) {
743 if ((what & MA_NOTRECURSED) != 0)
744 panic("mutex %s recursed at %s:%d",
745 m->mtx_object.lo_name, file, line);
746 } else if ((what & MA_RECURSED) != 0) {
747 panic("mutex %s unrecursed at %s:%d",
748 m->mtx_object.lo_name, file, line);
749 }
750 break;
751 case MA_NOTOWNED:
752 if (mtx_owned(m))
753 panic("mutex %s owned at %s:%d",
754 m->mtx_object.lo_name, file, line);
755 break;
756 default:
757 panic("unknown mtx_assert at %s:%d", file, line);
758 }
759 }
760 #endif
761
762 /*
763 * The MUTEX_DEBUG-enabled mtx_validate()
764 *
765 * Most of these checks have been moved off into the LO_INITIALIZED flag
766 * maintained by the witness code.
767 */
768 #ifdef MUTEX_DEBUG
769
770 void mtx_validate(struct mtx *);
771
772 void
773 mtx_validate(struct mtx *m)
774 {
775
776 /*
777 * XXX: When kernacc() does not require Giant we can reenable this check
778 */
779 #ifdef notyet
780 /*
781 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly
782 * we can re-enable the kernacc() checks.
783 */
784 #ifndef __alpha__
785 /*
786 * Can't call kernacc() from early init386(), especially when
787 * initializing Giant mutex, because some stuff in kernacc()
788 * requires Giant itself.
789 */
790 if (!cold)
791 if (!kernacc((caddr_t)m, sizeof(m),
792 VM_PROT_READ | VM_PROT_WRITE))
793 panic("Can't read and write to mutex %p", m);
794 #endif
795 #endif
796 }
797 #endif
798
799 /*
800 * General init routine used by the MTX_SYSINIT() macro.
801 */
802 void
803 mtx_sysinit(void *arg)
804 {
805 struct mtx_args *margs = arg;
806
807 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
808 }
809
810 /*
811 * Mutex initialization routine; initialize lock `m' of type contained in
812 * `opts' with options contained in `opts' and name `name.' The optional
813 * lock type `type' is used as a general lock category name for use with
814 * witness.
815 */
816 void
817 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
818 {
819 struct lock_object *lock;
820
821 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
822 MTX_NOWITNESS | MTX_DUPOK)) == 0);
823
824 #ifdef MUTEX_DEBUG
825 /* Diagnostic and error correction */
826 mtx_validate(m);
827 #endif
828
829 lock = &m->mtx_object;
830 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0,
831 ("mutex \"%s\" %p already initialized", name, m));
832 bzero(m, sizeof(*m));
833 if (opts & MTX_SPIN)
834 lock->lo_class = &lock_class_mtx_spin;
835 else
836 lock->lo_class = &lock_class_mtx_sleep;
837 lock->lo_name = name;
838 lock->lo_type = type != NULL ? type : name;
839 if (opts & MTX_QUIET)
840 lock->lo_flags = LO_QUIET;
841 if (opts & MTX_RECURSE)
842 lock->lo_flags |= LO_RECURSABLE;
843 if ((opts & MTX_NOWITNESS) == 0)
844 lock->lo_flags |= LO_WITNESS;
845 if (opts & MTX_DUPOK)
846 lock->lo_flags |= LO_DUPOK;
847
848 m->mtx_lock = MTX_UNOWNED;
849
850 LOCK_LOG_INIT(lock, opts);
851
852 WITNESS_INIT(lock);
853 }
854
855 /*
856 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
857 * passed in as a flag here because if the corresponding mtx_init() was
858 * called with MTX_QUIET set, then it will already be set in the mutex's
859 * flags.
860 */
861 void
862 mtx_destroy(struct mtx *m)
863 {
864
865 LOCK_LOG_DESTROY(&m->mtx_object, 0);
866
867 if (!mtx_owned(m))
868 MPASS(mtx_unowned(m));
869 else {
870 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
871
872 /* Tell witness this isn't locked to make it happy. */
873 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__,
874 __LINE__);
875 }
876
877 WITNESS_DESTROY(&m->mtx_object);
878 }
879
880 /*
881 * Intialize the mutex code and system mutexes. This is called from the MD
882 * startup code prior to mi_startup(). The per-CPU data space needs to be
883 * setup before this is called.
884 */
885 void
886 mutex_init(void)
887 {
888
889 /* Setup thread0 so that mutexes work. */
890 LIST_INIT(&thread0.td_contested);
891
892 /* Setup turnstiles so that sleep mutexes work. */
893 init_turnstiles();
894
895 /*
896 * Initialize mutexes.
897 */
898 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
899 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
900 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
901 mtx_lock(&Giant);
902 }
Cache object: 71ad9619830ed447588e500e9efc77e7
|