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 * $FreeBSD: head/sys/kern/subr_turnstile.c 115568 2003-05-31 21:11:01Z phk $
31 */
32
33/*
34 * Machine independent bits of mutex implementation.
35 */
36
37#include "opt_adaptive_mutexes.h"
38#include "opt_ddb.h"
39
40#include <sys/param.h>
41#include <sys/systm.h>
42#include <sys/bus.h>
43#include <sys/kernel.h>
44#include <sys/ktr.h>
45#include <sys/lock.h>
46#include <sys/malloc.h>
47#include <sys/mutex.h>
48#include <sys/proc.h>
49#include <sys/resourcevar.h>
50#include <sys/sched.h>
51#include <sys/sbuf.h>
52#include <sys/sysctl.h>
53#include <sys/vmmeter.h>
54
55#include <machine/atomic.h>
56#include <machine/bus.h>
57#include <machine/clock.h>
58#include <machine/cpu.h>
59
60#include <ddb/ddb.h>
61
62#include <vm/vm.h>
63#include <vm/vm_extern.h>
64
65/*
66 * Internal utility macros.
67 */
68#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
69
70#define mtx_owner(m) (mtx_unowned((m)) ? NULL \
71 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK))
72
73/*
74 * Lock classes for sleep and spin mutexes.
75 */
76struct lock_class lock_class_mtx_sleep = {
77 "sleep mutex",
78 LC_SLEEPLOCK | LC_RECURSABLE
79};
80struct lock_class lock_class_mtx_spin = {
81 "spin mutex",
82 LC_SPINLOCK | LC_RECURSABLE
83};
84
85/*
86 * System-wide mutexes
87 */
88struct mtx sched_lock;
89struct mtx Giant;
90
91/*
92 * Prototypes for non-exported routines.
93 */
94static void propagate_priority(struct thread *);
95
96static void
97propagate_priority(struct thread *td)
98{
99 int pri = td->td_priority;
100 struct mtx *m = td->td_blocked;
101
102 mtx_assert(&sched_lock, MA_OWNED);
103 for (;;) {
104 struct thread *td1;
105
106 td = mtx_owner(m);
107
108 if (td == NULL) {
109 /*
110 * This really isn't quite right. Really
111 * ought to bump priority of thread that
112 * next acquires the mutex.
113 */
114 MPASS(m->mtx_lock == MTX_CONTESTED);
115 return;
116 }
117
118 MPASS(td->td_proc != NULL);
119 MPASS(td->td_proc->p_magic == P_MAGIC);
120 KASSERT(!TD_IS_SLEEPING(td), ("sleeping thread owns a mutex"));
121 if (td->td_priority <= pri) /* lower is higher priority */
122 return;
123
124
125 /*
126 * If lock holder is actually running, just bump priority.
127 */
128 if (TD_IS_RUNNING(td)) {
129 td->td_priority = pri;
130 return;
131 }
132
133#ifndef SMP
134 /*
135 * For UP, we check to see if td is curthread (this shouldn't
136 * ever happen however as it would mean we are in a deadlock.)
137 */
138 KASSERT(td != curthread, ("Deadlock detected"));
139#endif
140
141 /*
142 * If on run queue move to new run queue, and quit.
143 * XXXKSE this gets a lot more complicated under threads
144 * but try anyhow.
145 */
146 if (TD_ON_RUNQ(td)) {
147 MPASS(td->td_blocked == NULL);
148 sched_prio(td, pri);
149 return;
150 }
151 /*
152 * Adjust for any other cases.
153 */
154 td->td_priority = pri;
155
156 /*
157 * If we aren't blocked on a mutex, we should be.
158 */
159 KASSERT(TD_ON_LOCK(td), (
160 "process %d(%s):%d holds %s but isn't blocked on a mutex\n",
161 td->td_proc->p_pid, td->td_proc->p_comm, td->td_state,
162 m->mtx_object.lo_name));
163
164 /*
165 * Pick up the mutex that td is blocked on.
166 */
167 m = td->td_blocked;
168 MPASS(m != NULL);
169
170 /*
171 * Check if the thread needs to be moved up on
172 * the blocked chain
173 */
174 if (td == TAILQ_FIRST(&m->mtx_blocked)) {
175 continue;
176 }
177
178 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
179 if (td1->td_priority <= pri) {
180 continue;
181 }
182
183 /*
184 * Remove thread from blocked chain and determine where
185 * it should be moved up to. Since we know that td1 has
186 * a lower priority than td, we know that at least one
187 * thread in the chain has a lower priority and that
188 * td1 will thus not be NULL after the loop.
189 */
190 TAILQ_REMOVE(&m->mtx_blocked, td, td_lockq);
191 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) {
192 MPASS(td1->td_proc->p_magic == P_MAGIC);
193 if (td1->td_priority > pri)
194 break;
195 }
196
197 MPASS(td1 != NULL);
198 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
199 CTR4(KTR_LOCK,
200 "propagate_priority: p %p moved before %p on [%p] %s",
201 td, td1, m, m->mtx_object.lo_name);
202 }
203}
204
205#ifdef MUTEX_PROFILING
206SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging");
207SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling");
208static int mutex_prof_enable = 0;
209SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW,
210 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime");
211
212struct mutex_prof {
213 const char *name;
214 const char *file;
215 int line;
216 uintmax_t cnt_max;
217 uintmax_t cnt_tot;
218 uintmax_t cnt_cur;
219 struct mutex_prof *next;
220};
221
222/*
223 * mprof_buf is a static pool of profiling records to avoid possible
224 * reentrance of the memory allocation functions.
225 *
226 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE.
227 */
228#define NUM_MPROF_BUFFERS 1000
229static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS];
230static int first_free_mprof_buf;
231#define MPROF_HASH_SIZE 1009
232static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE];
233/* SWAG: sbuf size = avg stat. line size * number of locks */
234#define MPROF_SBUF_SIZE 256 * 400
235
236static int mutex_prof_acquisitions;
237SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
238 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded");
239static int mutex_prof_records;
240SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD,
241 &mutex_prof_records, 0, "Number of profiling records");
242static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS;
243SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
244 &mutex_prof_maxrecords, 0, "Maximum number of profiling records");
245static int mutex_prof_rejected;
246SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD,
247 &mutex_prof_rejected, 0, "Number of rejected profiling records");
248static int mutex_prof_hashsize = MPROF_HASH_SIZE;
249SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD,
250 &mutex_prof_hashsize, 0, "Hash size");
251static int mutex_prof_collisions = 0;
252SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD,
253 &mutex_prof_collisions, 0, "Number of hash collisions");
254
255/*
256 * mprof_mtx protects the profiling buffers and the hash.
257 */
258static struct mtx mprof_mtx;
259MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET);
260
261static u_int64_t
262nanoseconds(void)
263{
264 struct timespec tv;
265
266 nanotime(&tv);
267 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
268}
269
270static int
271dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
272{
273 struct sbuf *sb;
274 int error, i;
275 static int multiplier = 1;
276
277 if (first_free_mprof_buf == 0)
278 return (SYSCTL_OUT(req, "No locking recorded",
279 sizeof("No locking recorded")));
280
281retry_sbufops:
282 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
283 sbuf_printf(sb, "%6s %12s %11s %5s %s\n",
284 "max", "total", "count", "avg", "name");
285 /*
286 * XXX this spinlock seems to be by far the largest perpetrator
287 * of spinlock latency (1.6 msec on an Athlon1600 was recorded
288 * even before I pessimized it further by moving the average
289 * computation here).
290 */
291 mtx_lock_spin(&mprof_mtx);
292 for (i = 0; i < first_free_mprof_buf; ++i) {
293 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n",
294 mprof_buf[i].cnt_max / 1000,
295 mprof_buf[i].cnt_tot / 1000,
296 mprof_buf[i].cnt_cur,
297 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
298 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000),
299 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name);
300 if (sbuf_overflowed(sb)) {
301 mtx_unlock_spin(&mprof_mtx);
302 sbuf_delete(sb);
303 multiplier++;
304 goto retry_sbufops;
305 }
306 }
307 mtx_unlock_spin(&mprof_mtx);
308 sbuf_finish(sb);
309 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
310 sbuf_delete(sb);
311 return (error);
312}
313SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
314 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics");
315#endif
316
317/*
318 * Function versions of the inlined __mtx_* macros. These are used by
319 * modules and can also be called from assembly language if needed.
320 */
321void
322_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
323{
324
325 MPASS(curthread != NULL);
326 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
327 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
328 file, line));
329 _get_sleep_lock(m, curthread, opts, file, line);
330 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
331 line);
332 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
333#ifdef MUTEX_PROFILING
334 /* don't reset the timer when/if recursing */
335 if (m->mtx_acqtime == 0) {
336 m->mtx_filename = file;
337 m->mtx_lineno = line;
338 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0;
339 ++mutex_prof_acquisitions;
340 }
341#endif
342}
343
344void
345_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
346{
347
348 MPASS(curthread != NULL);
349 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
350 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
351 file, line));
352 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
353 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
354 line);
355 mtx_assert(m, MA_OWNED);
356#ifdef MUTEX_PROFILING
357 if (m->mtx_acqtime != 0) {
358 static const char *unknown = "(unknown)";
359 struct mutex_prof *mpp;
360 u_int64_t acqtime, now;
361 const char *p, *q;
362 volatile u_int hash;
363
364 now = nanoseconds();
365 acqtime = m->mtx_acqtime;
366 m->mtx_acqtime = 0;
367 if (now <= acqtime)
368 goto out;
369 for (p = m->mtx_filename;
370 p != NULL && strncmp(p, "../", 3) == 0; p += 3)
371 /* nothing */ ;
372 if (p == NULL || *p == '\0')
373 p = unknown;
374 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q)
375 hash = (hash * 2 + *q) % MPROF_HASH_SIZE;
376 mtx_lock_spin(&mprof_mtx);
377 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next)
378 if (mpp->line == m->mtx_lineno &&
379 strcmp(mpp->file, p) == 0)
380 break;
381 if (mpp == NULL) {
382 /* Just exit if we cannot get a trace buffer */
383 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) {
384 ++mutex_prof_rejected;
385 goto unlock;
386 }
387 mpp = &mprof_buf[first_free_mprof_buf++];
388 mpp->name = mtx_name(m);
389 mpp->file = p;
390 mpp->line = m->mtx_lineno;
391 mpp->next = mprof_hash[hash];
392 if (mprof_hash[hash] != NULL)
393 ++mutex_prof_collisions;
394 mprof_hash[hash] = mpp;
395 ++mutex_prof_records;
396 }
397 /*
398 * Record if the mutex has been held longer now than ever
399 * before.
400 */
401 if (now - acqtime > mpp->cnt_max)
402 mpp->cnt_max = now - acqtime;
403 mpp->cnt_tot += now - acqtime;
404 mpp->cnt_cur++;
405unlock:
406 mtx_unlock_spin(&mprof_mtx);
407 }
408out:
409#endif
410 _rel_sleep_lock(m, curthread, opts, file, line);
411}
412
413void
414_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
415{
416
417 MPASS(curthread != NULL);
418 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
419 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
420 m->mtx_object.lo_name, file, line));
421#if defined(SMP) || LOCK_DEBUG > 0 || 1
422 _get_spin_lock(m, curthread, opts, file, line);
423#else
424 critical_enter();
425#endif
426 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
427 line);
428 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
429}
430
431void
432_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
433{
434
435 MPASS(curthread != NULL);
436 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
437 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
438 m->mtx_object.lo_name, file, line));
439 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
440 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
441 line);
442 mtx_assert(m, MA_OWNED);
443#if defined(SMP) || LOCK_DEBUG > 0 || 1
444 _rel_spin_lock(m);
445#else
446 critical_exit();
447#endif
448}
449
450/*
451 * The important part of mtx_trylock{,_flags}()
452 * Tries to acquire lock `m.' We do NOT handle recursion here. If this
453 * function is called on a recursed mutex, it will return failure and
454 * will not recursively acquire the lock. You are expected to know what
455 * you are doing.
456 */
457int
458_mtx_trylock(struct mtx *m, int opts, const char *file, int line)
459{
460 int rval;
461
462 MPASS(curthread != NULL);
463
464 rval = _obtain_lock(m, curthread);
465
466 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line);
467 if (rval)
468 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
469 file, line);
470
471 return (rval);
472}
473
474/*
475 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
476 *
477 * We call this if the lock is either contested (i.e. we need to go to
478 * sleep waiting for it), or if we need to recurse on it.
479 */
480void
481_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line)
482{
483 struct thread *td = curthread;
484 struct thread *td1;
485#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
486 struct thread *owner;
487#endif
488 uintptr_t v;
489#ifdef KTR
490 int cont_logged = 0;
491#endif
492
493 if (mtx_owned(m)) {
494 m->mtx_recurse++;
495 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
496 if (LOCK_LOG_TEST(&m->mtx_object, opts))
497 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
498 return;
499 }
500
501 if (LOCK_LOG_TEST(&m->mtx_object, opts))
502 CTR4(KTR_LOCK,
503 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
504 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line);
505
506 while (!_obtain_lock(m, td)) {
507
508 mtx_lock_spin(&sched_lock);
509 v = m->mtx_lock;
510
511 /*
512 * Check if the lock has been released while spinning for
513 * the sched_lock.
514 */
515 if (v == MTX_UNOWNED) {
516 mtx_unlock_spin(&sched_lock);
517#ifdef __i386__
518 ia32_pause();
519#endif
520 continue;
521 }
522
523 /*
524 * The mutex was marked contested on release. This means that
525 * there are threads blocked on it.
526 */
527 if (v == MTX_CONTESTED) {
528 td1 = TAILQ_FIRST(&m->mtx_blocked);
529 MPASS(td1 != NULL);
530 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED;
531
532 if (td1->td_priority < td->td_priority)
533 td->td_priority = td1->td_priority;
534 mtx_unlock_spin(&sched_lock);
535 return;
536 }
537
538 /*
539 * If the mutex isn't already contested and a failure occurs
540 * setting the contested bit, the mutex was either released
541 * or the state of the MTX_RECURSED bit changed.
542 */
543 if ((v & MTX_CONTESTED) == 0 &&
544 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v,
545 (void *)(v | MTX_CONTESTED))) {
546 mtx_unlock_spin(&sched_lock);
547#ifdef __i386__
548 ia32_pause();
549#endif
550 continue;
551 }
552
553#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
554 /*
555 * If the current owner of the lock is executing on another
556 * CPU, spin instead of blocking.
557 */
558 owner = (struct thread *)(v & MTX_FLAGMASK);
559 if (m != &Giant && TD_IS_RUNNING(owner)) {
560 mtx_unlock_spin(&sched_lock);
561 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
562#ifdef __i386__
563 ia32_pause();
564#endif
565 }
566 continue;
567 }
568#endif /* SMP && ADAPTIVE_MUTEXES */
569
570 /*
571 * We definitely must sleep for this lock.
572 */
573 mtx_assert(m, MA_NOTOWNED);
574
575#ifdef notyet
576 /*
577 * If we're borrowing an interrupted thread's VM context, we
578 * must clean up before going to sleep.
579 */
580 if (td->td_ithd != NULL) {
581 struct ithd *it = td->td_ithd;
582
583 if (it->it_interrupted) {
584 if (LOCK_LOG_TEST(&m->mtx_object, opts))
585 CTR2(KTR_LOCK,
586 "_mtx_lock_sleep: %p interrupted %p",
587 it, it->it_interrupted);
588 intr_thd_fixup(it);
589 }
590 }
591#endif
592
593 /*
594 * Put us on the list of threads blocked on this mutex.
595 */
596 if (TAILQ_EMPTY(&m->mtx_blocked)) {
597 td1 = mtx_owner(m);
598 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested);
599 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq);
600 } else {
601 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq)
602 if (td1->td_priority > td->td_priority)
603 break;
604 if (td1)
605 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
606 else
607 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq);
608 }
609#ifdef KTR
610 if (!cont_logged) {
611 CTR6(KTR_CONTENTION,
612 "contention: %p at %s:%d wants %s, taken by %s:%d",
613 td, file, line, m->mtx_object.lo_name,
614 WITNESS_FILE(&m->mtx_object),
615 WITNESS_LINE(&m->mtx_object));
616 cont_logged = 1;
617 }
618#endif
619
620 /*
621 * Save who we're blocked on.
622 */
623 td->td_blocked = m;
624 td->td_lockname = m->mtx_object.lo_name;
625 TD_SET_LOCK(td);
626 propagate_priority(td);
627
628 if (LOCK_LOG_TEST(&m->mtx_object, opts))
629 CTR3(KTR_LOCK,
630 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m,
631 m->mtx_object.lo_name);
632
633 td->td_proc->p_stats->p_ru.ru_nvcsw++;
634 mi_switch();
635
636 if (LOCK_LOG_TEST(&m->mtx_object, opts))
637 CTR3(KTR_LOCK,
638 "_mtx_lock_sleep: p %p free from blocked on [%p] %s",
639 td, m, m->mtx_object.lo_name);
640
641 mtx_unlock_spin(&sched_lock);
642 }
643
644#ifdef KTR
645 if (cont_logged) {
646 CTR4(KTR_CONTENTION,
647 "contention end: %s acquired by %p at %s:%d",
648 m->mtx_object.lo_name, td, file, line);
649 }
650#endif
651 return;
652}
653
654/*
655 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
656 *
657 * This is only called if we need to actually spin for the lock. Recursion
658 * is handled inline.
659 */
660void
661_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line)
662{
663 int i = 0;
664
665 if (LOCK_LOG_TEST(&m->mtx_object, opts))
666 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
667
668 for (;;) {
669 if (_obtain_lock(m, curthread))
670 break;
671
672 /* Give interrupts a chance while we spin. */
673 critical_exit();
674 while (m->mtx_lock != MTX_UNOWNED) {
675 if (i++ < 10000000) {
676#ifdef __i386__
677 ia32_pause();
678#endif
679 continue;
680 }
681 if (i < 60000000)
682 DELAY(1);
683#ifdef DDB
684 else if (!db_active)
685#else
686 else
687#endif
688 panic("spin lock %s held by %p for > 5 seconds",
689 m->mtx_object.lo_name, (void *)m->mtx_lock);
690#ifdef __i386__
691 ia32_pause();
692#endif
693 }
694 critical_enter();
695 }
696
697 if (LOCK_LOG_TEST(&m->mtx_object, opts))
698 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
699
700 return;
701}
702
703/*
704 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
705 *
706 * We are only called here if the lock is recursed or contested (i.e. we
707 * need to wake up a blocked thread).
708 */
709void
710_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
711{
712 struct thread *td, *td1;
713 struct mtx *m1;
714 int pri;
715
716 td = curthread;
717
718 if (mtx_recursed(m)) {
719 if (--(m->mtx_recurse) == 0)
720 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
721 if (LOCK_LOG_TEST(&m->mtx_object, opts))
722 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
723 return;
724 }
725
726 mtx_lock_spin(&sched_lock);
727 if (LOCK_LOG_TEST(&m->mtx_object, opts))
728 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
729
730 td1 = TAILQ_FIRST(&m->mtx_blocked);
731#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
732 if (td1 == NULL) {
733 _release_lock_quick(m);
734 if (LOCK_LOG_TEST(&m->mtx_object, opts))
735 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
736 mtx_unlock_spin(&sched_lock);
737 return;
738 }
739#endif
740 MPASS(td->td_proc->p_magic == P_MAGIC);
741 MPASS(td1->td_proc->p_magic == P_MAGIC);
742
743 TAILQ_REMOVE(&m->mtx_blocked, td1, td_lockq);
744
745 if (TAILQ_EMPTY(&m->mtx_blocked)) {
746 LIST_REMOVE(m, mtx_contested);
747 _release_lock_quick(m);
748 if (LOCK_LOG_TEST(&m->mtx_object, opts))
749 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m);
750 } else
751 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED);
752
753 pri = PRI_MAX;
754 LIST_FOREACH(m1, &td->td_contested, mtx_contested) {
755 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority;
756 if (cp < pri)
757 pri = cp;
758 }
759
760 if (pri > td->td_base_pri)
761 pri = td->td_base_pri;
762 td->td_priority = pri;
763
764 if (LOCK_LOG_TEST(&m->mtx_object, opts))
765 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p",
766 m, td1);
767
768 td1->td_blocked = NULL;
769 TD_CLR_LOCK(td1);
770 if (!TD_CAN_RUN(td1)) {
771 mtx_unlock_spin(&sched_lock);
772 return;
773 }
774 setrunqueue(td1);
775
776 if (td->td_critnest == 1 && td1->td_priority < pri) {
777#ifdef notyet
778 if (td->td_ithd != NULL) {
779 struct ithd *it = td->td_ithd;
780
781 if (it->it_interrupted) {
782 if (LOCK_LOG_TEST(&m->mtx_object, opts))
783 CTR2(KTR_LOCK,
784 "_mtx_unlock_sleep: %p interrupted %p",
785 it, it->it_interrupted);
786 intr_thd_fixup(it);
787 }
788 }
789#endif
790 if (LOCK_LOG_TEST(&m->mtx_object, opts))
791 CTR2(KTR_LOCK,
792 "_mtx_unlock_sleep: %p switching out lock=%p", m,
793 (void *)m->mtx_lock);
794
795 td->td_proc->p_stats->p_ru.ru_nivcsw++;
796 mi_switch();
797 if (LOCK_LOG_TEST(&m->mtx_object, opts))
798 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p",
799 m, (void *)m->mtx_lock);
800 }
801
802 mtx_unlock_spin(&sched_lock);
803
804 return;
805}
806
807/*
808 * All the unlocking of MTX_SPIN locks is done inline.
809 * See the _rel_spin_lock() macro for the details.
810 */
811
812/*
813 * The backing function for the INVARIANTS-enabled mtx_assert()
814 */
815#ifdef INVARIANT_SUPPORT
816void
817_mtx_assert(struct mtx *m, int what, const char *file, int line)
818{
819
820 if (panicstr != NULL)
821 return;
822 switch (what) {
823 case MA_OWNED:
824 case MA_OWNED | MA_RECURSED:
825 case MA_OWNED | MA_NOTRECURSED:
826 if (!mtx_owned(m))
827 panic("mutex %s not owned at %s:%d",
828 m->mtx_object.lo_name, file, line);
829 if (mtx_recursed(m)) {
830 if ((what & MA_NOTRECURSED) != 0)
831 panic("mutex %s recursed at %s:%d",
832 m->mtx_object.lo_name, file, line);
833 } else if ((what & MA_RECURSED) != 0) {
834 panic("mutex %s unrecursed at %s:%d",
835 m->mtx_object.lo_name, file, line);
836 }
837 break;
838 case MA_NOTOWNED:
839 if (mtx_owned(m))
840 panic("mutex %s owned at %s:%d",
841 m->mtx_object.lo_name, file, line);
842 break;
843 default:
844 panic("unknown mtx_assert at %s:%d", file, line);
845 }
846}
847#endif
848
849/*
850 * The MUTEX_DEBUG-enabled mtx_validate()
851 *
852 * Most of these checks have been moved off into the LO_INITIALIZED flag
853 * maintained by the witness code.
854 */
855#ifdef MUTEX_DEBUG
856
857void mtx_validate(struct mtx *);
858
859void
860mtx_validate(struct mtx *m)
861{
862
863/*
864 * XXX: When kernacc() does not require Giant we can reenable this check
865 */
866#ifdef notyet
867/*
868 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly
869 * we can re-enable the kernacc() checks.
870 */
871#ifndef __alpha__
872 /*
873 * Can't call kernacc() from early init386(), especially when
874 * initializing Giant mutex, because some stuff in kernacc()
875 * requires Giant itself.
876 */
877 if (!cold)
878 if (!kernacc((caddr_t)m, sizeof(m),
879 VM_PROT_READ | VM_PROT_WRITE))
880 panic("Can't read and write to mutex %p", m);
881#endif
882#endif
883}
884#endif
885
886/*
887 * General init routine used by the MTX_SYSINIT() macro.
888 */
889void
890mtx_sysinit(void *arg)
891{
892 struct mtx_args *margs = arg;
893
894 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
895}
896
897/*
898 * Mutex initialization routine; initialize lock `m' of type contained in
899 * `opts' with options contained in `opts' and name `name.' The optional
900 * lock type `type' is used as a general lock category name for use with
901 * witness.
902 */
903void
904mtx_init(struct mtx *m, const char *name, const char *type, int opts)
905{
906 struct lock_object *lock;
907
908 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
909 MTX_NOWITNESS | MTX_DUPOK)) == 0);
910
911#ifdef MUTEX_DEBUG
912 /* Diagnostic and error correction */
913 mtx_validate(m);
914#endif
915
916 lock = &m->mtx_object;
917 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0,
918 ("mutex \"%s\" %p already initialized", name, m));
919 bzero(m, sizeof(*m));
920 if (opts & MTX_SPIN)
921 lock->lo_class = &lock_class_mtx_spin;
922 else
923 lock->lo_class = &lock_class_mtx_sleep;
924 lock->lo_name = name;
925 lock->lo_type = type != NULL ? type : name;
926 if (opts & MTX_QUIET)
927 lock->lo_flags = LO_QUIET;
928 if (opts & MTX_RECURSE)
929 lock->lo_flags |= LO_RECURSABLE;
930 if ((opts & MTX_NOWITNESS) == 0)
931 lock->lo_flags |= LO_WITNESS;
932 if (opts & MTX_DUPOK)
933 lock->lo_flags |= LO_DUPOK;
934
935 m->mtx_lock = MTX_UNOWNED;
936 TAILQ_INIT(&m->mtx_blocked);
937
938 LOCK_LOG_INIT(lock, opts);
939
940 WITNESS_INIT(lock);
941}
942
943/*
944 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
945 * passed in as a flag here because if the corresponding mtx_init() was
946 * called with MTX_QUIET set, then it will already be set in the mutex's
947 * flags.
948 */
949void
950mtx_destroy(struct mtx *m)
951{
952
953 LOCK_LOG_DESTROY(&m->mtx_object, 0);
954
955 if (!mtx_owned(m))
956 MPASS(mtx_unowned(m));
957 else {
958 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
959
960 /* Tell witness this isn't locked to make it happy. */
961 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__,
962 __LINE__);
963 }
964
965 WITNESS_DESTROY(&m->mtx_object);
966}
967
968/*
969 * Intialize the mutex code and system mutexes. This is called from the MD
970 * startup code prior to mi_startup(). The per-CPU data space needs to be
971 * setup before this is called.
972 */
973void
974mutex_init(void)
975{
976
977 /* Setup thread0 so that mutexes work. */
978 LIST_INIT(&thread0.td_contested);
979
980 /*
981 * Initialize mutexes.
982 */
983 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
984 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
985 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
986 mtx_lock(&Giant);
987}
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 * $FreeBSD: head/sys/kern/subr_turnstile.c 115568 2003-05-31 21:11:01Z phk $
31 */
32
33/*
34 * Machine independent bits of mutex implementation.
35 */
36
37#include "opt_adaptive_mutexes.h"
38#include "opt_ddb.h"
39
40#include <sys/param.h>
41#include <sys/systm.h>
42#include <sys/bus.h>
43#include <sys/kernel.h>
44#include <sys/ktr.h>
45#include <sys/lock.h>
46#include <sys/malloc.h>
47#include <sys/mutex.h>
48#include <sys/proc.h>
49#include <sys/resourcevar.h>
50#include <sys/sched.h>
51#include <sys/sbuf.h>
52#include <sys/sysctl.h>
53#include <sys/vmmeter.h>
54
55#include <machine/atomic.h>
56#include <machine/bus.h>
57#include <machine/clock.h>
58#include <machine/cpu.h>
59
60#include <ddb/ddb.h>
61
62#include <vm/vm.h>
63#include <vm/vm_extern.h>
64
65/*
66 * Internal utility macros.
67 */
68#define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
69
70#define mtx_owner(m) (mtx_unowned((m)) ? NULL \
71 : (struct thread *)((m)->mtx_lock & MTX_FLAGMASK))
72
73/*
74 * Lock classes for sleep and spin mutexes.
75 */
76struct lock_class lock_class_mtx_sleep = {
77 "sleep mutex",
78 LC_SLEEPLOCK | LC_RECURSABLE
79};
80struct lock_class lock_class_mtx_spin = {
81 "spin mutex",
82 LC_SPINLOCK | LC_RECURSABLE
83};
84
85/*
86 * System-wide mutexes
87 */
88struct mtx sched_lock;
89struct mtx Giant;
90
91/*
92 * Prototypes for non-exported routines.
93 */
94static void propagate_priority(struct thread *);
95
96static void
97propagate_priority(struct thread *td)
98{
99 int pri = td->td_priority;
100 struct mtx *m = td->td_blocked;
101
102 mtx_assert(&sched_lock, MA_OWNED);
103 for (;;) {
104 struct thread *td1;
105
106 td = mtx_owner(m);
107
108 if (td == NULL) {
109 /*
110 * This really isn't quite right. Really
111 * ought to bump priority of thread that
112 * next acquires the mutex.
113 */
114 MPASS(m->mtx_lock == MTX_CONTESTED);
115 return;
116 }
117
118 MPASS(td->td_proc != NULL);
119 MPASS(td->td_proc->p_magic == P_MAGIC);
120 KASSERT(!TD_IS_SLEEPING(td), ("sleeping thread owns a mutex"));
121 if (td->td_priority <= pri) /* lower is higher priority */
122 return;
123
124
125 /*
126 * If lock holder is actually running, just bump priority.
127 */
128 if (TD_IS_RUNNING(td)) {
129 td->td_priority = pri;
130 return;
131 }
132
133#ifndef SMP
134 /*
135 * For UP, we check to see if td is curthread (this shouldn't
136 * ever happen however as it would mean we are in a deadlock.)
137 */
138 KASSERT(td != curthread, ("Deadlock detected"));
139#endif
140
141 /*
142 * If on run queue move to new run queue, and quit.
143 * XXXKSE this gets a lot more complicated under threads
144 * but try anyhow.
145 */
146 if (TD_ON_RUNQ(td)) {
147 MPASS(td->td_blocked == NULL);
148 sched_prio(td, pri);
149 return;
150 }
151 /*
152 * Adjust for any other cases.
153 */
154 td->td_priority = pri;
155
156 /*
157 * If we aren't blocked on a mutex, we should be.
158 */
159 KASSERT(TD_ON_LOCK(td), (
160 "process %d(%s):%d holds %s but isn't blocked on a mutex\n",
161 td->td_proc->p_pid, td->td_proc->p_comm, td->td_state,
162 m->mtx_object.lo_name));
163
164 /*
165 * Pick up the mutex that td is blocked on.
166 */
167 m = td->td_blocked;
168 MPASS(m != NULL);
169
170 /*
171 * Check if the thread needs to be moved up on
172 * the blocked chain
173 */
174 if (td == TAILQ_FIRST(&m->mtx_blocked)) {
175 continue;
176 }
177
178 td1 = TAILQ_PREV(td, threadqueue, td_lockq);
179 if (td1->td_priority <= pri) {
180 continue;
181 }
182
183 /*
184 * Remove thread from blocked chain and determine where
185 * it should be moved up to. Since we know that td1 has
186 * a lower priority than td, we know that at least one
187 * thread in the chain has a lower priority and that
188 * td1 will thus not be NULL after the loop.
189 */
190 TAILQ_REMOVE(&m->mtx_blocked, td, td_lockq);
191 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq) {
192 MPASS(td1->td_proc->p_magic == P_MAGIC);
193 if (td1->td_priority > pri)
194 break;
195 }
196
197 MPASS(td1 != NULL);
198 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
199 CTR4(KTR_LOCK,
200 "propagate_priority: p %p moved before %p on [%p] %s",
201 td, td1, m, m->mtx_object.lo_name);
202 }
203}
204
205#ifdef MUTEX_PROFILING
206SYSCTL_NODE(_debug, OID_AUTO, mutex, CTLFLAG_RD, NULL, "mutex debugging");
207SYSCTL_NODE(_debug_mutex, OID_AUTO, prof, CTLFLAG_RD, NULL, "mutex profiling");
208static int mutex_prof_enable = 0;
209SYSCTL_INT(_debug_mutex_prof, OID_AUTO, enable, CTLFLAG_RW,
210 &mutex_prof_enable, 0, "Enable tracing of mutex holdtime");
211
212struct mutex_prof {
213 const char *name;
214 const char *file;
215 int line;
216 uintmax_t cnt_max;
217 uintmax_t cnt_tot;
218 uintmax_t cnt_cur;
219 struct mutex_prof *next;
220};
221
222/*
223 * mprof_buf is a static pool of profiling records to avoid possible
224 * reentrance of the memory allocation functions.
225 *
226 * Note: NUM_MPROF_BUFFERS must be smaller than MPROF_HASH_SIZE.
227 */
228#define NUM_MPROF_BUFFERS 1000
229static struct mutex_prof mprof_buf[NUM_MPROF_BUFFERS];
230static int first_free_mprof_buf;
231#define MPROF_HASH_SIZE 1009
232static struct mutex_prof *mprof_hash[MPROF_HASH_SIZE];
233/* SWAG: sbuf size = avg stat. line size * number of locks */
234#define MPROF_SBUF_SIZE 256 * 400
235
236static int mutex_prof_acquisitions;
237SYSCTL_INT(_debug_mutex_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
238 &mutex_prof_acquisitions, 0, "Number of mutex acquistions recorded");
239static int mutex_prof_records;
240SYSCTL_INT(_debug_mutex_prof, OID_AUTO, records, CTLFLAG_RD,
241 &mutex_prof_records, 0, "Number of profiling records");
242static int mutex_prof_maxrecords = NUM_MPROF_BUFFERS;
243SYSCTL_INT(_debug_mutex_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
244 &mutex_prof_maxrecords, 0, "Maximum number of profiling records");
245static int mutex_prof_rejected;
246SYSCTL_INT(_debug_mutex_prof, OID_AUTO, rejected, CTLFLAG_RD,
247 &mutex_prof_rejected, 0, "Number of rejected profiling records");
248static int mutex_prof_hashsize = MPROF_HASH_SIZE;
249SYSCTL_INT(_debug_mutex_prof, OID_AUTO, hashsize, CTLFLAG_RD,
250 &mutex_prof_hashsize, 0, "Hash size");
251static int mutex_prof_collisions = 0;
252SYSCTL_INT(_debug_mutex_prof, OID_AUTO, collisions, CTLFLAG_RD,
253 &mutex_prof_collisions, 0, "Number of hash collisions");
254
255/*
256 * mprof_mtx protects the profiling buffers and the hash.
257 */
258static struct mtx mprof_mtx;
259MTX_SYSINIT(mprof, &mprof_mtx, "mutex profiling lock", MTX_SPIN | MTX_QUIET);
260
261static u_int64_t
262nanoseconds(void)
263{
264 struct timespec tv;
265
266 nanotime(&tv);
267 return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
268}
269
270static int
271dump_mutex_prof_stats(SYSCTL_HANDLER_ARGS)
272{
273 struct sbuf *sb;
274 int error, i;
275 static int multiplier = 1;
276
277 if (first_free_mprof_buf == 0)
278 return (SYSCTL_OUT(req, "No locking recorded",
279 sizeof("No locking recorded")));
280
281retry_sbufops:
282 sb = sbuf_new(NULL, NULL, MPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
283 sbuf_printf(sb, "%6s %12s %11s %5s %s\n",
284 "max", "total", "count", "avg", "name");
285 /*
286 * XXX this spinlock seems to be by far the largest perpetrator
287 * of spinlock latency (1.6 msec on an Athlon1600 was recorded
288 * even before I pessimized it further by moving the average
289 * computation here).
290 */
291 mtx_lock_spin(&mprof_mtx);
292 for (i = 0; i < first_free_mprof_buf; ++i) {
293 sbuf_printf(sb, "%6ju %12ju %11ju %5ju %s:%d (%s)\n",
294 mprof_buf[i].cnt_max / 1000,
295 mprof_buf[i].cnt_tot / 1000,
296 mprof_buf[i].cnt_cur,
297 mprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
298 mprof_buf[i].cnt_tot / (mprof_buf[i].cnt_cur * 1000),
299 mprof_buf[i].file, mprof_buf[i].line, mprof_buf[i].name);
300 if (sbuf_overflowed(sb)) {
301 mtx_unlock_spin(&mprof_mtx);
302 sbuf_delete(sb);
303 multiplier++;
304 goto retry_sbufops;
305 }
306 }
307 mtx_unlock_spin(&mprof_mtx);
308 sbuf_finish(sb);
309 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
310 sbuf_delete(sb);
311 return (error);
312}
313SYSCTL_PROC(_debug_mutex_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
314 NULL, 0, dump_mutex_prof_stats, "A", "Mutex profiling statistics");
315#endif
316
317/*
318 * Function versions of the inlined __mtx_* macros. These are used by
319 * modules and can also be called from assembly language if needed.
320 */
321void
322_mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
323{
324
325 MPASS(curthread != NULL);
326 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
327 ("mtx_lock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
328 file, line));
329 _get_sleep_lock(m, curthread, opts, file, line);
330 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
331 line);
332 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
333#ifdef MUTEX_PROFILING
334 /* don't reset the timer when/if recursing */
335 if (m->mtx_acqtime == 0) {
336 m->mtx_filename = file;
337 m->mtx_lineno = line;
338 m->mtx_acqtime = mutex_prof_enable ? nanoseconds() : 0;
339 ++mutex_prof_acquisitions;
340 }
341#endif
342}
343
344void
345_mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
346{
347
348 MPASS(curthread != NULL);
349 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_sleep,
350 ("mtx_unlock() of spin mutex %s @ %s:%d", m->mtx_object.lo_name,
351 file, line));
352 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
353 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
354 line);
355 mtx_assert(m, MA_OWNED);
356#ifdef MUTEX_PROFILING
357 if (m->mtx_acqtime != 0) {
358 static const char *unknown = "(unknown)";
359 struct mutex_prof *mpp;
360 u_int64_t acqtime, now;
361 const char *p, *q;
362 volatile u_int hash;
363
364 now = nanoseconds();
365 acqtime = m->mtx_acqtime;
366 m->mtx_acqtime = 0;
367 if (now <= acqtime)
368 goto out;
369 for (p = m->mtx_filename;
370 p != NULL && strncmp(p, "../", 3) == 0; p += 3)
371 /* nothing */ ;
372 if (p == NULL || *p == '\0')
373 p = unknown;
374 for (hash = m->mtx_lineno, q = p; *q != '\0'; ++q)
375 hash = (hash * 2 + *q) % MPROF_HASH_SIZE;
376 mtx_lock_spin(&mprof_mtx);
377 for (mpp = mprof_hash[hash]; mpp != NULL; mpp = mpp->next)
378 if (mpp->line == m->mtx_lineno &&
379 strcmp(mpp->file, p) == 0)
380 break;
381 if (mpp == NULL) {
382 /* Just exit if we cannot get a trace buffer */
383 if (first_free_mprof_buf >= NUM_MPROF_BUFFERS) {
384 ++mutex_prof_rejected;
385 goto unlock;
386 }
387 mpp = &mprof_buf[first_free_mprof_buf++];
388 mpp->name = mtx_name(m);
389 mpp->file = p;
390 mpp->line = m->mtx_lineno;
391 mpp->next = mprof_hash[hash];
392 if (mprof_hash[hash] != NULL)
393 ++mutex_prof_collisions;
394 mprof_hash[hash] = mpp;
395 ++mutex_prof_records;
396 }
397 /*
398 * Record if the mutex has been held longer now than ever
399 * before.
400 */
401 if (now - acqtime > mpp->cnt_max)
402 mpp->cnt_max = now - acqtime;
403 mpp->cnt_tot += now - acqtime;
404 mpp->cnt_cur++;
405unlock:
406 mtx_unlock_spin(&mprof_mtx);
407 }
408out:
409#endif
410 _rel_sleep_lock(m, curthread, opts, file, line);
411}
412
413void
414_mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
415{
416
417 MPASS(curthread != NULL);
418 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
419 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
420 m->mtx_object.lo_name, file, line));
421#if defined(SMP) || LOCK_DEBUG > 0 || 1
422 _get_spin_lock(m, curthread, opts, file, line);
423#else
424 critical_enter();
425#endif
426 LOCK_LOG_LOCK("LOCK", &m->mtx_object, opts, m->mtx_recurse, file,
427 line);
428 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
429}
430
431void
432_mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
433{
434
435 MPASS(curthread != NULL);
436 KASSERT(m->mtx_object.lo_class == &lock_class_mtx_spin,
437 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
438 m->mtx_object.lo_name, file, line));
439 WITNESS_UNLOCK(&m->mtx_object, opts | LOP_EXCLUSIVE, file, line);
440 LOCK_LOG_LOCK("UNLOCK", &m->mtx_object, opts, m->mtx_recurse, file,
441 line);
442 mtx_assert(m, MA_OWNED);
443#if defined(SMP) || LOCK_DEBUG > 0 || 1
444 _rel_spin_lock(m);
445#else
446 critical_exit();
447#endif
448}
449
450/*
451 * The important part of mtx_trylock{,_flags}()
452 * Tries to acquire lock `m.' We do NOT handle recursion here. If this
453 * function is called on a recursed mutex, it will return failure and
454 * will not recursively acquire the lock. You are expected to know what
455 * you are doing.
456 */
457int
458_mtx_trylock(struct mtx *m, int opts, const char *file, int line)
459{
460 int rval;
461
462 MPASS(curthread != NULL);
463
464 rval = _obtain_lock(m, curthread);
465
466 LOCK_LOG_TRY("LOCK", &m->mtx_object, opts, rval, file, line);
467 if (rval)
468 WITNESS_LOCK(&m->mtx_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
469 file, line);
470
471 return (rval);
472}
473
474/*
475 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
476 *
477 * We call this if the lock is either contested (i.e. we need to go to
478 * sleep waiting for it), or if we need to recurse on it.
479 */
480void
481_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line)
482{
483 struct thread *td = curthread;
484 struct thread *td1;
485#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
486 struct thread *owner;
487#endif
488 uintptr_t v;
489#ifdef KTR
490 int cont_logged = 0;
491#endif
492
493 if (mtx_owned(m)) {
494 m->mtx_recurse++;
495 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
496 if (LOCK_LOG_TEST(&m->mtx_object, opts))
497 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
498 return;
499 }
500
501 if (LOCK_LOG_TEST(&m->mtx_object, opts))
502 CTR4(KTR_LOCK,
503 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
504 m->mtx_object.lo_name, (void *)m->mtx_lock, file, line);
505
506 while (!_obtain_lock(m, td)) {
507
508 mtx_lock_spin(&sched_lock);
509 v = m->mtx_lock;
510
511 /*
512 * Check if the lock has been released while spinning for
513 * the sched_lock.
514 */
515 if (v == MTX_UNOWNED) {
516 mtx_unlock_spin(&sched_lock);
517#ifdef __i386__
518 ia32_pause();
519#endif
520 continue;
521 }
522
523 /*
524 * The mutex was marked contested on release. This means that
525 * there are threads blocked on it.
526 */
527 if (v == MTX_CONTESTED) {
528 td1 = TAILQ_FIRST(&m->mtx_blocked);
529 MPASS(td1 != NULL);
530 m->mtx_lock = (uintptr_t)td | MTX_CONTESTED;
531
532 if (td1->td_priority < td->td_priority)
533 td->td_priority = td1->td_priority;
534 mtx_unlock_spin(&sched_lock);
535 return;
536 }
537
538 /*
539 * If the mutex isn't already contested and a failure occurs
540 * setting the contested bit, the mutex was either released
541 * or the state of the MTX_RECURSED bit changed.
542 */
543 if ((v & MTX_CONTESTED) == 0 &&
544 !atomic_cmpset_ptr(&m->mtx_lock, (void *)v,
545 (void *)(v | MTX_CONTESTED))) {
546 mtx_unlock_spin(&sched_lock);
547#ifdef __i386__
548 ia32_pause();
549#endif
550 continue;
551 }
552
553#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
554 /*
555 * If the current owner of the lock is executing on another
556 * CPU, spin instead of blocking.
557 */
558 owner = (struct thread *)(v & MTX_FLAGMASK);
559 if (m != &Giant && TD_IS_RUNNING(owner)) {
560 mtx_unlock_spin(&sched_lock);
561 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
562#ifdef __i386__
563 ia32_pause();
564#endif
565 }
566 continue;
567 }
568#endif /* SMP && ADAPTIVE_MUTEXES */
569
570 /*
571 * We definitely must sleep for this lock.
572 */
573 mtx_assert(m, MA_NOTOWNED);
574
575#ifdef notyet
576 /*
577 * If we're borrowing an interrupted thread's VM context, we
578 * must clean up before going to sleep.
579 */
580 if (td->td_ithd != NULL) {
581 struct ithd *it = td->td_ithd;
582
583 if (it->it_interrupted) {
584 if (LOCK_LOG_TEST(&m->mtx_object, opts))
585 CTR2(KTR_LOCK,
586 "_mtx_lock_sleep: %p interrupted %p",
587 it, it->it_interrupted);
588 intr_thd_fixup(it);
589 }
590 }
591#endif
592
593 /*
594 * Put us on the list of threads blocked on this mutex.
595 */
596 if (TAILQ_EMPTY(&m->mtx_blocked)) {
597 td1 = mtx_owner(m);
598 LIST_INSERT_HEAD(&td1->td_contested, m, mtx_contested);
599 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq);
600 } else {
601 TAILQ_FOREACH(td1, &m->mtx_blocked, td_lockq)
602 if (td1->td_priority > td->td_priority)
603 break;
604 if (td1)
605 TAILQ_INSERT_BEFORE(td1, td, td_lockq);
606 else
607 TAILQ_INSERT_TAIL(&m->mtx_blocked, td, td_lockq);
608 }
609#ifdef KTR
610 if (!cont_logged) {
611 CTR6(KTR_CONTENTION,
612 "contention: %p at %s:%d wants %s, taken by %s:%d",
613 td, file, line, m->mtx_object.lo_name,
614 WITNESS_FILE(&m->mtx_object),
615 WITNESS_LINE(&m->mtx_object));
616 cont_logged = 1;
617 }
618#endif
619
620 /*
621 * Save who we're blocked on.
622 */
623 td->td_blocked = m;
624 td->td_lockname = m->mtx_object.lo_name;
625 TD_SET_LOCK(td);
626 propagate_priority(td);
627
628 if (LOCK_LOG_TEST(&m->mtx_object, opts))
629 CTR3(KTR_LOCK,
630 "_mtx_lock_sleep: p %p blocked on [%p] %s", td, m,
631 m->mtx_object.lo_name);
632
633 td->td_proc->p_stats->p_ru.ru_nvcsw++;
634 mi_switch();
635
636 if (LOCK_LOG_TEST(&m->mtx_object, opts))
637 CTR3(KTR_LOCK,
638 "_mtx_lock_sleep: p %p free from blocked on [%p] %s",
639 td, m, m->mtx_object.lo_name);
640
641 mtx_unlock_spin(&sched_lock);
642 }
643
644#ifdef KTR
645 if (cont_logged) {
646 CTR4(KTR_CONTENTION,
647 "contention end: %s acquired by %p at %s:%d",
648 m->mtx_object.lo_name, td, file, line);
649 }
650#endif
651 return;
652}
653
654/*
655 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
656 *
657 * This is only called if we need to actually spin for the lock. Recursion
658 * is handled inline.
659 */
660void
661_mtx_lock_spin(struct mtx *m, int opts, const char *file, int line)
662{
663 int i = 0;
664
665 if (LOCK_LOG_TEST(&m->mtx_object, opts))
666 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
667
668 for (;;) {
669 if (_obtain_lock(m, curthread))
670 break;
671
672 /* Give interrupts a chance while we spin. */
673 critical_exit();
674 while (m->mtx_lock != MTX_UNOWNED) {
675 if (i++ < 10000000) {
676#ifdef __i386__
677 ia32_pause();
678#endif
679 continue;
680 }
681 if (i < 60000000)
682 DELAY(1);
683#ifdef DDB
684 else if (!db_active)
685#else
686 else
687#endif
688 panic("spin lock %s held by %p for > 5 seconds",
689 m->mtx_object.lo_name, (void *)m->mtx_lock);
690#ifdef __i386__
691 ia32_pause();
692#endif
693 }
694 critical_enter();
695 }
696
697 if (LOCK_LOG_TEST(&m->mtx_object, opts))
698 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
699
700 return;
701}
702
703/*
704 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
705 *
706 * We are only called here if the lock is recursed or contested (i.e. we
707 * need to wake up a blocked thread).
708 */
709void
710_mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
711{
712 struct thread *td, *td1;
713 struct mtx *m1;
714 int pri;
715
716 td = curthread;
717
718 if (mtx_recursed(m)) {
719 if (--(m->mtx_recurse) == 0)
720 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
721 if (LOCK_LOG_TEST(&m->mtx_object, opts))
722 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
723 return;
724 }
725
726 mtx_lock_spin(&sched_lock);
727 if (LOCK_LOG_TEST(&m->mtx_object, opts))
728 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
729
730 td1 = TAILQ_FIRST(&m->mtx_blocked);
731#if defined(SMP) && defined(ADAPTIVE_MUTEXES)
732 if (td1 == NULL) {
733 _release_lock_quick(m);
734 if (LOCK_LOG_TEST(&m->mtx_object, opts))
735 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
736 mtx_unlock_spin(&sched_lock);
737 return;
738 }
739#endif
740 MPASS(td->td_proc->p_magic == P_MAGIC);
741 MPASS(td1->td_proc->p_magic == P_MAGIC);
742
743 TAILQ_REMOVE(&m->mtx_blocked, td1, td_lockq);
744
745 if (TAILQ_EMPTY(&m->mtx_blocked)) {
746 LIST_REMOVE(m, mtx_contested);
747 _release_lock_quick(m);
748 if (LOCK_LOG_TEST(&m->mtx_object, opts))
749 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p not held", m);
750 } else
751 atomic_store_rel_ptr(&m->mtx_lock, (void *)MTX_CONTESTED);
752
753 pri = PRI_MAX;
754 LIST_FOREACH(m1, &td->td_contested, mtx_contested) {
755 int cp = TAILQ_FIRST(&m1->mtx_blocked)->td_priority;
756 if (cp < pri)
757 pri = cp;
758 }
759
760 if (pri > td->td_base_pri)
761 pri = td->td_base_pri;
762 td->td_priority = pri;
763
764 if (LOCK_LOG_TEST(&m->mtx_object, opts))
765 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p contested setrunqueue %p",
766 m, td1);
767
768 td1->td_blocked = NULL;
769 TD_CLR_LOCK(td1);
770 if (!TD_CAN_RUN(td1)) {
771 mtx_unlock_spin(&sched_lock);
772 return;
773 }
774 setrunqueue(td1);
775
776 if (td->td_critnest == 1 && td1->td_priority < pri) {
777#ifdef notyet
778 if (td->td_ithd != NULL) {
779 struct ithd *it = td->td_ithd;
780
781 if (it->it_interrupted) {
782 if (LOCK_LOG_TEST(&m->mtx_object, opts))
783 CTR2(KTR_LOCK,
784 "_mtx_unlock_sleep: %p interrupted %p",
785 it, it->it_interrupted);
786 intr_thd_fixup(it);
787 }
788 }
789#endif
790 if (LOCK_LOG_TEST(&m->mtx_object, opts))
791 CTR2(KTR_LOCK,
792 "_mtx_unlock_sleep: %p switching out lock=%p", m,
793 (void *)m->mtx_lock);
794
795 td->td_proc->p_stats->p_ru.ru_nivcsw++;
796 mi_switch();
797 if (LOCK_LOG_TEST(&m->mtx_object, opts))
798 CTR2(KTR_LOCK, "_mtx_unlock_sleep: %p resuming lock=%p",
799 m, (void *)m->mtx_lock);
800 }
801
802 mtx_unlock_spin(&sched_lock);
803
804 return;
805}
806
807/*
808 * All the unlocking of MTX_SPIN locks is done inline.
809 * See the _rel_spin_lock() macro for the details.
810 */
811
812/*
813 * The backing function for the INVARIANTS-enabled mtx_assert()
814 */
815#ifdef INVARIANT_SUPPORT
816void
817_mtx_assert(struct mtx *m, int what, const char *file, int line)
818{
819
820 if (panicstr != NULL)
821 return;
822 switch (what) {
823 case MA_OWNED:
824 case MA_OWNED | MA_RECURSED:
825 case MA_OWNED | MA_NOTRECURSED:
826 if (!mtx_owned(m))
827 panic("mutex %s not owned at %s:%d",
828 m->mtx_object.lo_name, file, line);
829 if (mtx_recursed(m)) {
830 if ((what & MA_NOTRECURSED) != 0)
831 panic("mutex %s recursed at %s:%d",
832 m->mtx_object.lo_name, file, line);
833 } else if ((what & MA_RECURSED) != 0) {
834 panic("mutex %s unrecursed at %s:%d",
835 m->mtx_object.lo_name, file, line);
836 }
837 break;
838 case MA_NOTOWNED:
839 if (mtx_owned(m))
840 panic("mutex %s owned at %s:%d",
841 m->mtx_object.lo_name, file, line);
842 break;
843 default:
844 panic("unknown mtx_assert at %s:%d", file, line);
845 }
846}
847#endif
848
849/*
850 * The MUTEX_DEBUG-enabled mtx_validate()
851 *
852 * Most of these checks have been moved off into the LO_INITIALIZED flag
853 * maintained by the witness code.
854 */
855#ifdef MUTEX_DEBUG
856
857void mtx_validate(struct mtx *);
858
859void
860mtx_validate(struct mtx *m)
861{
862
863/*
864 * XXX: When kernacc() does not require Giant we can reenable this check
865 */
866#ifdef notyet
867/*
868 * XXX - When kernacc() is fixed on the alpha to handle K0_SEG memory properly
869 * we can re-enable the kernacc() checks.
870 */
871#ifndef __alpha__
872 /*
873 * Can't call kernacc() from early init386(), especially when
874 * initializing Giant mutex, because some stuff in kernacc()
875 * requires Giant itself.
876 */
877 if (!cold)
878 if (!kernacc((caddr_t)m, sizeof(m),
879 VM_PROT_READ | VM_PROT_WRITE))
880 panic("Can't read and write to mutex %p", m);
881#endif
882#endif
883}
884#endif
885
886/*
887 * General init routine used by the MTX_SYSINIT() macro.
888 */
889void
890mtx_sysinit(void *arg)
891{
892 struct mtx_args *margs = arg;
893
894 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
895}
896
897/*
898 * Mutex initialization routine; initialize lock `m' of type contained in
899 * `opts' with options contained in `opts' and name `name.' The optional
900 * lock type `type' is used as a general lock category name for use with
901 * witness.
902 */
903void
904mtx_init(struct mtx *m, const char *name, const char *type, int opts)
905{
906 struct lock_object *lock;
907
908 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
909 MTX_NOWITNESS | MTX_DUPOK)) == 0);
910
911#ifdef MUTEX_DEBUG
912 /* Diagnostic and error correction */
913 mtx_validate(m);
914#endif
915
916 lock = &m->mtx_object;
917 KASSERT((lock->lo_flags & LO_INITIALIZED) == 0,
918 ("mutex \"%s\" %p already initialized", name, m));
919 bzero(m, sizeof(*m));
920 if (opts & MTX_SPIN)
921 lock->lo_class = &lock_class_mtx_spin;
922 else
923 lock->lo_class = &lock_class_mtx_sleep;
924 lock->lo_name = name;
925 lock->lo_type = type != NULL ? type : name;
926 if (opts & MTX_QUIET)
927 lock->lo_flags = LO_QUIET;
928 if (opts & MTX_RECURSE)
929 lock->lo_flags |= LO_RECURSABLE;
930 if ((opts & MTX_NOWITNESS) == 0)
931 lock->lo_flags |= LO_WITNESS;
932 if (opts & MTX_DUPOK)
933 lock->lo_flags |= LO_DUPOK;
934
935 m->mtx_lock = MTX_UNOWNED;
936 TAILQ_INIT(&m->mtx_blocked);
937
938 LOCK_LOG_INIT(lock, opts);
939
940 WITNESS_INIT(lock);
941}
942
943/*
944 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
945 * passed in as a flag here because if the corresponding mtx_init() was
946 * called with MTX_QUIET set, then it will already be set in the mutex's
947 * flags.
948 */
949void
950mtx_destroy(struct mtx *m)
951{
952
953 LOCK_LOG_DESTROY(&m->mtx_object, 0);
954
955 if (!mtx_owned(m))
956 MPASS(mtx_unowned(m));
957 else {
958 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
959
960 /* Tell witness this isn't locked to make it happy. */
961 WITNESS_UNLOCK(&m->mtx_object, LOP_EXCLUSIVE, __FILE__,
962 __LINE__);
963 }
964
965 WITNESS_DESTROY(&m->mtx_object);
966}
967
968/*
969 * Intialize the mutex code and system mutexes. This is called from the MD
970 * startup code prior to mi_startup(). The per-CPU data space needs to be
971 * setup before this is called.
972 */
973void
974mutex_init(void)
975{
976
977 /* Setup thread0 so that mutexes work. */
978 LIST_INIT(&thread0.td_contested);
979
980 /*
981 * Initialize mutexes.
982 */
983 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
984 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
985 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
986 mtx_lock(&Giant);
987}