2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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
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
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 $
33 * Machine independent bits of mutex implementation.
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
39 #include "opt_adaptive_mutexes.h"
41 #include "opt_global.h"
42 #include "opt_sched.h"
44 #include <sys/param.h>
45 #include <sys/systm.h>
49 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
55 #include <sys/resourcevar.h>
56 #include <sys/sched.h>
58 #include <sys/sysctl.h>
59 #include <sys/turnstile.h>
60 #include <sys/vmmeter.h>
61 #include <sys/lock_profile.h>
63 #include <machine/atomic.h>
64 #include <machine/bus.h>
65 #include <machine/cpu.h>
69 #include <fs/devfs/devfs_int.h>
72 #include <vm/vm_extern.h>
74 #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES)
75 #define ADAPTIVE_MUTEXES
79 * Internal utility macros.
81 #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED)
83 #define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED)
85 #define mtx_owner(m) ((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK))
88 static void db_show_mtx(struct lock_object *lock);
90 static void lock_mtx(struct lock_object *lock, int how);
91 static void lock_spin(struct lock_object *lock, int how);
92 static int unlock_mtx(struct lock_object *lock);
93 static int unlock_spin(struct lock_object *lock);
96 * Lock classes for sleep and spin mutexes.
98 struct lock_class lock_class_mtx_sleep = {
99 .lc_name = "sleep mutex",
100 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE,
102 .lc_ddb_show = db_show_mtx,
105 .lc_unlock = unlock_mtx,
107 struct lock_class lock_class_mtx_spin = {
108 .lc_name = "spin mutex",
109 .lc_flags = LC_SPINLOCK | LC_RECURSABLE,
111 .lc_ddb_show = db_show_mtx,
113 .lc_lock = lock_spin,
114 .lc_unlock = unlock_spin,
118 * System-wide mutexes
120 struct mtx blocked_lock;
123 #ifdef LOCK_PROFILING
124 static inline void lock_profile_init(void)
127 /* Initialize the mutex profiling locks */
128 for (i = 0; i < LPROF_LOCK_SIZE; i++) {
129 mtx_init(&lprof_locks[i], "mprof lock",
130 NULL, MTX_SPIN|MTX_QUIET|MTX_NOPROFILE);
134 static inline void lock_profile_init(void) {;}
138 lock_mtx(struct lock_object *lock, int how)
141 mtx_lock((struct mtx *)lock);
145 lock_spin(struct lock_object *lock, int how)
148 panic("spin locks can only use msleep_spin");
152 unlock_mtx(struct lock_object *lock)
156 m = (struct mtx *)lock;
157 mtx_assert(m, MA_OWNED | MA_NOTRECURSED);
163 unlock_spin(struct lock_object *lock)
166 panic("spin locks can only use msleep_spin");
170 * Function versions of the inlined __mtx_* macros. These are used by
171 * modules and can also be called from assembly language if needed.
174 _mtx_lock_flags(struct mtx *m, int opts, const char *file, int line)
177 MPASS(curthread != NULL);
178 KASSERT(m->mtx_lock != MTX_DESTROYED,
179 ("mtx_lock() of destroyed mutex @ %s:%d", file, line));
180 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
181 ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
183 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
186 _get_sleep_lock(m, curthread, opts, file, line);
187 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
189 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
190 curthread->td_locks++;
194 _mtx_unlock_flags(struct mtx *m, int opts, const char *file, int line)
196 MPASS(curthread != NULL);
197 KASSERT(m->mtx_lock != MTX_DESTROYED,
198 ("mtx_unlock() of destroyed mutex @ %s:%d", file, line));
199 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
200 ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
202 curthread->td_locks--;
203 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
204 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
206 mtx_assert(m, MA_OWNED);
208 if (m->mtx_recurse == 0)
209 lock_profile_release_lock(&m->lock_object);
210 _rel_sleep_lock(m, curthread, opts, file, line);
214 _mtx_lock_spin_flags(struct mtx *m, int opts, const char *file, int line)
217 MPASS(curthread != NULL);
218 KASSERT(m->mtx_lock != MTX_DESTROYED,
219 ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line));
220 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
221 ("mtx_lock_spin() of sleep mutex %s @ %s:%d",
222 m->lock_object.lo_name, file, line));
223 WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE,
225 _get_spin_lock(m, curthread, opts, file, line);
226 LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file,
228 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
232 _mtx_unlock_spin_flags(struct mtx *m, int opts, const char *file, int line)
235 MPASS(curthread != NULL);
236 KASSERT(m->mtx_lock != MTX_DESTROYED,
237 ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line));
238 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin,
239 ("mtx_unlock_spin() of sleep mutex %s @ %s:%d",
240 m->lock_object.lo_name, file, line));
241 WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
242 LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file,
244 mtx_assert(m, MA_OWNED);
250 * The important part of mtx_trylock{,_flags}()
251 * Tries to acquire lock `m.' If this function is called on a mutex that
252 * is already owned, it will recursively acquire the lock.
255 _mtx_trylock(struct mtx *m, int opts, const char *file, int line)
257 int rval, contested = 0;
258 uint64_t waittime = 0;
260 MPASS(curthread != NULL);
261 KASSERT(m->mtx_lock != MTX_DESTROYED,
262 ("mtx_trylock() of destroyed mutex @ %s:%d", file, line));
263 KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep,
264 ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name,
267 if (mtx_owned(m) && (m->lock_object.lo_flags & LO_RECURSABLE) != 0) {
269 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
272 rval = _obtain_lock(m, (uintptr_t)curthread);
274 LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line);
276 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK,
278 curthread->td_locks++;
279 if (m->mtx_recurse == 0)
280 lock_profile_obtain_lock_success(&m->lock_object, contested,
281 waittime, file, line);
289 * _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
291 * We call this if the lock is either contested (i.e. we need to go to
292 * sleep waiting for it), or if we need to recurse on it.
295 _mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
298 struct turnstile *ts;
299 #ifdef ADAPTIVE_MUTEXES
300 volatile struct thread *owner;
306 uint64_t waittime = 0;
310 KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
311 ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
312 m->lock_object.lo_name, file, line));
314 atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
315 if (LOCK_LOG_TEST(&m->lock_object, opts))
316 CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
320 lock_profile_obtain_lock_failed(&m->lock_object,
321 &contested, &waittime);
322 if (LOCK_LOG_TEST(&m->lock_object, opts))
324 "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
325 m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
327 while (!_obtain_lock(m, tid)) {
328 ts = turnstile_trywait(&m->lock_object);
332 * Check if the lock has been released while spinning for
333 * the turnstile chain lock.
335 if (v == MTX_UNOWNED) {
336 turnstile_cancel(ts);
341 MPASS(v != MTX_CONTESTED);
344 * If the mutex isn't already contested and a failure occurs
345 * setting the contested bit, the mutex was either released
346 * or the state of the MTX_RECURSED bit changed.
348 if ((v & MTX_CONTESTED) == 0 &&
349 !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
350 turnstile_cancel(ts);
355 #ifdef ADAPTIVE_MUTEXES
357 * If the current owner of the lock is executing on another
358 * CPU, spin instead of blocking.
360 owner = (struct thread *)(v & ~MTX_FLAGMASK);
361 #ifdef ADAPTIVE_GIANT
362 if (TD_IS_RUNNING(owner))
364 if (m != &Giant && TD_IS_RUNNING(owner))
367 turnstile_cancel(ts);
368 while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) {
373 #endif /* ADAPTIVE_MUTEXES */
376 * We definitely must sleep for this lock.
378 mtx_assert(m, MA_NOTOWNED);
383 "contention: %p at %s:%d wants %s, taken by %s:%d",
384 (void *)tid, file, line, m->lock_object.lo_name,
385 WITNESS_FILE(&m->lock_object),
386 WITNESS_LINE(&m->lock_object));
392 * Block on the turnstile.
394 turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
399 "contention end: %s acquired by %p at %s:%d",
400 m->lock_object.lo_name, (void *)tid, file, line);
403 lock_profile_obtain_lock_success(&m->lock_object, contested,
404 waittime, (file), (line));
408 _mtx_lock_spin_failed(struct mtx *m)
414 /* If the mutex is unlocked, try again. */
418 printf( "spin lock %p (%s) held by %p (tid %d) too long\n",
419 m, m->lock_object.lo_name, td, td->td_tid);
421 witness_display_spinlock(&m->lock_object, td);
423 panic("spin lock held too long");
428 * _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
430 * This is only called if we need to actually spin for the lock. Recursion
434 _mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file,
437 int i = 0, contested = 0;
438 uint64_t waittime = 0;
440 if (LOCK_LOG_TEST(&m->lock_object, opts))
441 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
443 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
444 while (!_obtain_lock(m, tid)) {
446 /* Give interrupts a chance while we spin. */
448 while (m->mtx_lock != MTX_UNOWNED) {
449 if (i++ < 10000000) {
453 if (i < 60000000 || kdb_active || panicstr != NULL)
456 _mtx_lock_spin_failed(m);
462 if (LOCK_LOG_TEST(&m->lock_object, opts))
463 CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
465 lock_profile_obtain_lock_success(&m->lock_object, contested,
466 waittime, (file), (line));
471 _thread_lock_flags(struct thread *td, int opts, const char *file, int line)
481 tid = (uintptr_t)curthread;
486 WITNESS_CHECKORDER(&m->lock_object,
487 opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line);
488 while (!_obtain_lock(m, tid)) {
489 if (m->mtx_lock == tid) {
493 lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
494 /* Give interrupts a chance while we spin. */
496 while (m->mtx_lock != MTX_UNOWNED) {
499 else if (i < 60000000 ||
500 kdb_active || panicstr != NULL)
503 _mtx_lock_spin_failed(m);
505 if (m != td->td_lock)
510 if (m == td->td_lock)
512 _rel_spin_lock(m); /* does spinlock_exit() */
514 lock_profile_obtain_lock_success(&m->lock_object, contested,
515 waittime, (file), (line));
516 WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line);
520 thread_lock_block(struct thread *td)
525 THREAD_LOCK_ASSERT(td, MA_OWNED);
527 td->td_lock = &blocked_lock;
528 mtx_unlock_spin(lock);
534 thread_lock_unblock(struct thread *td, struct mtx *new)
536 mtx_assert(new, MA_OWNED);
537 MPASS(td->td_lock == &blocked_lock);
538 atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
543 thread_lock_set(struct thread *td, struct mtx *new)
547 mtx_assert(new, MA_OWNED);
548 THREAD_LOCK_ASSERT(td, MA_OWNED);
551 mtx_unlock_spin(lock);
555 * _mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock.
557 * We are only called here if the lock is recursed or contested (i.e. we
558 * need to wake up a blocked thread).
561 _mtx_unlock_sleep(struct mtx *m, int opts, const char *file, int line)
563 struct turnstile *ts;
565 if (mtx_recursed(m)) {
566 if (--(m->mtx_recurse) == 0)
567 atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED);
568 if (LOCK_LOG_TEST(&m->lock_object, opts))
569 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m);
574 * We have to lock the chain before the turnstile so this turnstile
575 * can be removed from the hash list if it is empty.
577 turnstile_chain_lock(&m->lock_object);
578 ts = turnstile_lookup(&m->lock_object);
579 if (LOCK_LOG_TEST(&m->lock_object, opts))
580 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m);
582 #ifdef ADAPTIVE_MUTEXES
584 _release_lock_quick(m);
585 if (LOCK_LOG_TEST(&m->lock_object, opts))
586 CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p no sleepers", m);
587 turnstile_chain_unlock(&m->lock_object);
593 turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE);
594 _release_lock_quick(m);
596 * This turnstile is now no longer associated with the mutex. We can
597 * unlock the chain lock so a new turnstile may take it's place.
599 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
600 turnstile_chain_unlock(&m->lock_object);
604 * All the unlocking of MTX_SPIN locks is done inline.
605 * See the _rel_spin_lock() macro for the details.
609 * The backing function for the INVARIANTS-enabled mtx_assert()
611 #ifdef INVARIANT_SUPPORT
613 _mtx_assert(struct mtx *m, int what, const char *file, int line)
616 if (panicstr != NULL || dumping)
620 case MA_OWNED | MA_RECURSED:
621 case MA_OWNED | MA_NOTRECURSED:
623 panic("mutex %s not owned at %s:%d",
624 m->lock_object.lo_name, file, line);
625 if (mtx_recursed(m)) {
626 if ((what & MA_NOTRECURSED) != 0)
627 panic("mutex %s recursed at %s:%d",
628 m->lock_object.lo_name, file, line);
629 } else if ((what & MA_RECURSED) != 0) {
630 panic("mutex %s unrecursed at %s:%d",
631 m->lock_object.lo_name, file, line);
636 panic("mutex %s owned at %s:%d",
637 m->lock_object.lo_name, file, line);
640 panic("unknown mtx_assert at %s:%d", file, line);
646 * The MUTEX_DEBUG-enabled mtx_validate()
648 * Most of these checks have been moved off into the LO_INITIALIZED flag
649 * maintained by the witness code.
653 void mtx_validate(struct mtx *);
656 mtx_validate(struct mtx *m)
660 * XXX: When kernacc() does not require Giant we can reenable this check
664 * Can't call kernacc() from early init386(), especially when
665 * initializing Giant mutex, because some stuff in kernacc()
666 * requires Giant itself.
669 if (!kernacc((caddr_t)m, sizeof(m),
670 VM_PROT_READ | VM_PROT_WRITE))
671 panic("Can't read and write to mutex %p", m);
677 * General init routine used by the MTX_SYSINIT() macro.
680 mtx_sysinit(void *arg)
682 struct mtx_args *margs = arg;
684 mtx_init(margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts);
688 * Mutex initialization routine; initialize lock `m' of type contained in
689 * `opts' with options contained in `opts' and name `name.' The optional
690 * lock type `type' is used as a general lock category name for use with
694 mtx_init(struct mtx *m, const char *name, const char *type, int opts)
696 struct lock_class *class;
699 MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE |
700 MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE)) == 0);
703 /* Diagnostic and error correction */
707 /* Determine lock class and lock flags. */
709 class = &lock_class_mtx_spin;
711 class = &lock_class_mtx_sleep;
713 if (opts & MTX_QUIET)
715 if (opts & MTX_RECURSE)
716 flags |= LO_RECURSABLE;
717 if ((opts & MTX_NOWITNESS) == 0)
719 if (opts & MTX_DUPOK)
721 if (opts & MTX_NOPROFILE)
722 flags |= LO_NOPROFILE;
724 /* Initialize mutex. */
725 m->mtx_lock = MTX_UNOWNED;
728 lock_init(&m->lock_object, class, name, type, flags);
732 * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be
733 * passed in as a flag here because if the corresponding mtx_init() was
734 * called with MTX_QUIET set, then it will already be set in the mutex's
738 mtx_destroy(struct mtx *m)
742 MPASS(mtx_unowned(m));
744 MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0);
746 /* Perform the non-mtx related part of mtx_unlock_spin(). */
747 if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin)
750 curthread->td_locks--;
752 /* Tell witness this isn't locked to make it happy. */
753 WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__,
757 m->mtx_lock = MTX_DESTROYED;
758 lock_destroy(&m->lock_object);
762 * Intialize the mutex code and system mutexes. This is called from the MD
763 * startup code prior to mi_startup(). The per-CPU data space needs to be
764 * setup before this is called.
770 /* Setup turnstiles so that sleep mutexes work. */
774 * Initialize mutexes.
776 mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE);
777 mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN);
778 blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */
779 mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK);
780 mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN | MTX_RECURSE);
781 mtx_init(&devmtx, "cdev", NULL, MTX_DEF);
789 db_show_mtx(struct lock_object *lock)
794 m = (struct mtx *)lock;
796 db_printf(" flags: {");
797 if (LOCK_CLASS(lock) == &lock_class_mtx_spin)
801 if (m->lock_object.lo_flags & LO_RECURSABLE)
802 db_printf(", RECURSE");
803 if (m->lock_object.lo_flags & LO_DUPOK)
804 db_printf(", DUPOK");
806 db_printf(" state: {");
808 db_printf("UNOWNED");
809 else if (mtx_destroyed(m))
810 db_printf("DESTROYED");
813 if (m->mtx_lock & MTX_CONTESTED)
814 db_printf(", CONTESTED");
815 if (m->mtx_lock & MTX_RECURSED)
816 db_printf(", RECURSED");
819 if (!mtx_unowned(m) && !mtx_destroyed(m)) {
821 db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td,
822 td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
824 db_printf(" recursed: %d\n", m->mtx_recurse);