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 $
34 * Implementation of the `witness' lock verifier. Originally implemented for
35 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
41 * Pronunciation: 'wit-n&s
43 * Etymology: Middle English witnesse, from Old English witnes knowledge,
44 * testimony, witness, from 2wit
45 * Date: before 12th century
46 * 1 : attestation of a fact or event : TESTIMONY
47 * 2 : one that gives evidence; specifically : one who testifies in
48 * a cause or before a judicial tribunal
49 * 3 : one asked to be present at a transaction so as to be able to
50 * testify to its having taken place
51 * 4 : one who has personal knowledge of something
52 * 5 a : something serving as evidence or proof : SIGN
53 * b : public affirmation by word or example of usually
54 * religious faith or conviction <the heroic witness to divine
56 * 6 capitalized : a member of the Jehovah's Witnesses
60 #include "opt_witness.h"
62 #include <sys/param.h>
64 #include <sys/kernel.h>
67 #include <sys/malloc.h>
68 #include <sys/mutex.h>
70 #include <sys/sysctl.h>
71 #include <sys/systm.h>
75 #define WITNESS_COUNT 200
76 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
78 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should
80 * probably be safe for the most part, but it's still a SWAG.
82 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
84 #define WITNESS_NCHILDREN 6
86 struct witness_child_list_entry;
90 struct lock_class *w_class;
91 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
92 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
93 struct witness_child_list_entry *w_children; /* Great evilness... */
98 u_char w_Giant_squawked:1;
99 u_char w_other_squawked:1;
100 u_char w_same_squawked:1;
103 struct witness_child_list_entry {
104 struct witness_child_list_entry *wcl_next;
105 struct witness *wcl_children[WITNESS_NCHILDREN];
109 STAILQ_HEAD(witness_list, witness);
111 struct witness_blessed {
116 struct witness_order_list_entry {
118 struct lock_class *w_class;
121 static struct witness *enroll(const char *description,
122 struct lock_class *lock_class);
123 static int itismychild(struct witness *parent, struct witness *child);
124 static void removechild(struct witness *parent, struct witness *child);
125 static int isitmychild(struct witness *parent, struct witness *child);
126 static int isitmydescendant(struct witness *parent, struct witness *child);
127 static int blessed(struct witness *, struct witness *);
128 static void witness_display_list(void(*prnt)(const char *fmt, ...),
129 struct witness_list *list);
130 static void witness_displaydescendants(void(*)(const char *fmt, ...),
132 static void witness_leveldescendents(struct witness *parent, int level);
133 static void witness_levelall(void);
134 static struct witness *witness_get(void);
135 static void witness_free(struct witness *m);
136 static struct witness_child_list_entry *witness_child_get(void);
137 static void witness_child_free(struct witness_child_list_entry *wcl);
138 static struct lock_list_entry *witness_lock_list_get(void);
139 static void witness_lock_list_free(struct lock_list_entry *lle);
140 static void witness_display(void(*)(const char *fmt, ...));
141 static struct lock_instance *find_instance(struct lock_list_entry *lock_list,
142 struct lock_object *lock);
144 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146 static int witness_watch = 1;
147 TUNABLE_INT("debug.witness_watch", &witness_watch);
148 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
152 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
153 * drop into kdebug() when:
154 * - a lock heirarchy violation occurs
155 * - locks are held when going to sleep.
162 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
163 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
166 #ifdef WITNESS_SKIPSPIN
167 int witness_skipspin = 1;
169 int witness_skipspin = 0;
171 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
172 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
175 static struct mtx w_mtx;
176 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
177 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
178 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
179 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
180 static struct witness_child_list_entry *w_child_free = NULL;
181 static struct lock_list_entry *w_lock_list_free = NULL;
182 static int witness_dead; /* fatal error, probably no memory */
184 static struct witness w_data[WITNESS_COUNT];
185 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
186 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188 static struct witness_order_list_entry order_lists[] = {
189 { "Giant", &lock_class_mtx_sleep },
190 { "proctree", &lock_class_sx },
191 { "allproc", &lock_class_sx },
192 { "process group", &lock_class_mtx_sleep },
193 { "process lock", &lock_class_mtx_sleep },
194 { "session", &lock_class_mtx_sleep },
195 { "uidinfo hash", &lock_class_mtx_sleep },
196 { "uidinfo struct", &lock_class_mtx_sleep },
202 { "ap boot", &lock_class_mtx_spin },
204 { "com", &lock_class_mtx_spin },
207 { "sio", &lock_class_mtx_spin },
209 { "cy", &lock_class_mtx_spin },
211 { "ng_node", &lock_class_mtx_spin },
212 { "ng_worklist", &lock_class_mtx_spin },
213 { "ithread table lock", &lock_class_mtx_spin },
214 { "sched lock", &lock_class_mtx_spin },
215 { "callout", &lock_class_mtx_spin },
219 { "allpmaps", &lock_class_mtx_spin },
220 { "vm page buckets mutex", &lock_class_mtx_spin },
221 { "icu", &lock_class_mtx_spin },
223 { "smp rendezvous", &lock_class_mtx_spin },
225 { "clk", &lock_class_mtx_spin },
226 { "mutex profiling lock", &lock_class_mtx_spin },
232 * Pairs of locks which have been blessed
233 * Don't complain about order problems with blessed locks
235 static struct witness_blessed blessed_list[] = {
237 static int blessed_count =
238 sizeof(blessed_list) / sizeof(struct witness_blessed);
241 * List of all locks in the system.
243 STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245 static struct mtx all_mtx = {
246 { &lock_class_mtx_sleep, /* mtx_object.lo_class */
247 "All locks list", /* mtx_object.lo_name */
248 "All locks list", /* mtx_object.lo_type */
249 LO_INITIALIZED, /* mtx_object.lo_flags */
250 { NULL }, /* mtx_object.lo_list */
251 NULL }, /* mtx_object.lo_witness */
252 MTX_UNOWNED, 0, /* mtx_lock, mtx_recurse */
253 TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
254 { NULL, NULL } /* mtx_contested */
258 * This global is set to 0 once it becomes safe to use the witness code.
260 static int witness_cold = 1;
263 * Global variables for book keeping.
265 static int lock_cur_cnt;
266 static int lock_max_cnt;
269 * The WITNESS-enabled diagnostic code.
272 witness_initialize(void *dummy __unused)
274 struct lock_object *lock;
275 struct witness_order_list_entry *order;
276 struct witness *w, *w1;
280 * We have to release Giant before initializing its witness
281 * structure so that WITNESS doesn't get confused.
284 mtx_assert(&Giant, MA_NOTOWNED);
286 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
287 STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
288 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
290 for (i = 0; i < WITNESS_COUNT; i++)
291 witness_free(&w_data[i]);
292 for (i = 0; i < WITNESS_CHILDCOUNT; i++)
293 witness_child_free(&w_childdata[i]);
294 for (i = 0; i < LOCK_CHILDCOUNT; i++)
295 witness_lock_list_free(&w_locklistdata[i]);
297 /* First add in all the specified order lists. */
298 for (order = order_lists; order->w_name != NULL; order++) {
299 w = enroll(order->w_name, order->w_class);
302 w->w_file = "order list";
303 for (order++; order->w_name != NULL; order++) {
304 w1 = enroll(order->w_name, order->w_class);
307 w1->w_file = "order list";
313 /* Iterate through all locks and add them to witness. */
315 STAILQ_FOREACH(lock, &all_locks, lo_list) {
316 if (lock->lo_flags & LO_WITNESS)
317 lock->lo_witness = enroll(lock->lo_type,
320 lock->lo_witness = NULL;
322 mtx_unlock(&all_mtx);
324 /* Mark the witness code as being ready for use. */
325 atomic_store_rel_int(&witness_cold, 0);
329 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
332 witness_init(struct lock_object *lock)
334 struct lock_class *class;
336 class = lock->lo_class;
337 if (lock->lo_flags & LO_INITIALIZED)
338 panic("%s: lock (%s) %s is already initialized", __func__,
339 class->lc_name, lock->lo_name);
340 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
341 (class->lc_flags & LC_RECURSABLE) == 0)
342 panic("%s: lock (%s) %s can not be recursable", __func__,
343 class->lc_name, lock->lo_name);
344 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
345 (class->lc_flags & LC_SLEEPABLE) == 0)
346 panic("%s: lock (%s) %s can not be sleepable", __func__,
347 class->lc_name, lock->lo_name);
348 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
349 (class->lc_flags & LC_UPGRADABLE) == 0)
350 panic("%s: lock (%s) %s can not be upgradable", __func__,
351 class->lc_name, lock->lo_name);
354 STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
355 lock->lo_flags |= LO_INITIALIZED;
357 if (lock_cur_cnt > lock_max_cnt)
358 lock_max_cnt = lock_cur_cnt;
359 mtx_unlock(&all_mtx);
360 if (!witness_cold && !witness_dead && panicstr == NULL &&
361 (lock->lo_flags & LO_WITNESS) != 0)
362 lock->lo_witness = enroll(lock->lo_type, class);
364 lock->lo_witness = NULL;
368 witness_destroy(struct lock_object *lock)
373 panic("lock (%s) %s destroyed while witness_cold",
374 lock->lo_class->lc_name, lock->lo_name);
375 if ((lock->lo_flags & LO_INITIALIZED) == 0)
376 panic("%s: lock (%s) %s is not initialized", __func__,
377 lock->lo_class->lc_name, lock->lo_name);
379 /* XXX: need to verify that no one holds the lock */
380 w = lock->lo_witness;
382 mtx_lock_spin(&w_mtx);
384 if (w->w_refcount == 0) {
386 "%s: marking witness %s as dead", __func__, w->w_name);
387 w->w_name = "(dead)";
388 w->w_file = "(dead)";
391 mtx_unlock_spin(&w_mtx);
396 STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
397 lock->lo_flags &= ~LO_INITIALIZED;
398 mtx_unlock(&all_mtx);
402 witness_display_list(void(*prnt)(const char *fmt, ...),
403 struct witness_list *list)
405 struct witness *w, *w1;
408 STAILQ_FOREACH(w, list, w_typelist) {
409 if (w->w_file == NULL)
412 STAILQ_FOREACH(w1, list, w_typelist) {
413 if (isitmychild(w1, w)) {
421 * This lock has no anscestors, display its descendants.
423 witness_displaydescendants(prnt, w);
428 witness_display(void(*prnt)(const char *fmt, ...))
432 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
436 * First, handle sleep locks which have been acquired at least
439 prnt("Sleep locks:\n");
440 witness_display_list(prnt, &w_sleep);
443 * Now do spin locks which have been acquired at least once.
445 prnt("\nSpin locks:\n");
446 witness_display_list(prnt, &w_spin);
449 * Finally, any locks which have not been acquired yet.
451 prnt("\nLocks which were never acquired:\n");
452 STAILQ_FOREACH(w, &w_all, w_list) {
453 if (w->w_file != NULL)
455 prnt("%s\n", w->w_name);
460 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
462 struct lock_list_entry **lock_list, *lle;
463 struct lock_instance *lock1, *lock2;
464 struct lock_class *class;
465 struct witness *w, *w1;
472 if (witness_cold || witness_dead || lock->lo_witness == NULL ||
475 w = lock->lo_witness;
476 class = lock->lo_class;
479 if (class->lc_flags & LC_SLEEPLOCK) {
481 * Since spin locks include a critical section, this check
482 * impliclty enforces a lock order of all sleep locks before
485 if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
486 panic("blockable sleep lock (%s) %s @ %s:%d",
487 class->lc_name, lock->lo_name, file, line);
488 lock_list = &td->td_sleeplocks;
490 lock_list = PCPU_PTR(spinlocks);
493 * Try locks do not block if they fail to acquire the lock, thus
494 * there is no danger of deadlocks or of switching while holding a
495 * spin lock if we acquire a lock via a try operation.
497 if (flags & LOP_TRYLOCK)
501 * Is this the first lock acquired? If so, then no order checking
504 if (*lock_list == NULL)
508 * Check to see if we are recursing on a lock we already own.
510 lock1 = find_instance(*lock_list, lock);
512 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
513 (flags & LOP_EXCLUSIVE) == 0) {
514 printf("shared lock of (%s) %s @ %s:%d\n",
515 class->lc_name, lock->lo_name, file, line);
516 printf("while exclusively locked from %s:%d\n",
517 lock1->li_file, lock1->li_line);
518 panic("share->excl");
520 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
521 (flags & LOP_EXCLUSIVE) != 0) {
522 printf("exclusive lock of (%s) %s @ %s:%d\n",
523 class->lc_name, lock->lo_name, file, line);
524 printf("while share locked from %s:%d\n",
525 lock1->li_file, lock1->li_line);
526 panic("excl->share");
529 if ((lock->lo_flags & LO_RECURSABLE) == 0) {
531 "recursed on non-recursive lock (%s) %s @ %s:%d\n",
532 class->lc_name, lock->lo_name, file, line);
533 printf("first acquired @ %s:%d\n", lock1->li_file,
537 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
538 td->td_proc->p_pid, lock->lo_name,
539 lock1->li_flags & LI_RECURSEMASK);
540 lock1->li_file = file;
541 lock1->li_line = line;
546 * Check for duplicate locks of the same type. Note that we only
547 * have to check for this on the last lock we just acquired. Any
548 * other cases will be caught as lock order violations.
550 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
551 w1 = lock1->li_lock->lo_witness;
553 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
555 w->w_same_squawked = 1;
556 printf("acquiring duplicate lock of same type: \"%s\"\n",
558 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
559 lock1->li_file, lock1->li_line);
560 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
566 MPASS(!mtx_owned(&w_mtx));
567 mtx_lock_spin(&w_mtx);
569 * If we have a known higher number just say ok
571 if (witness_watch > 1 && w->w_level > w1->w_level) {
572 mtx_unlock_spin(&w_mtx);
575 if (isitmydescendant(w1, w)) {
576 mtx_unlock_spin(&w_mtx);
579 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
580 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
582 MPASS(j < WITNESS_COUNT);
583 lock1 = &lle->ll_children[i];
584 w1 = lock1->li_lock->lo_witness;
587 * If this lock doesn't undergo witness checking,
591 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
592 ("lock missing witness structure"));
596 * If we are locking Giant and we slept with this
597 * lock, then skip it.
599 if ((lock1->li_flags & LI_SLEPT) != 0 &&
600 lock == &Giant.mtx_object)
603 * If we are locking a sleepable lock and this lock
604 * isn't sleepable and isn't Giant, we want to treat
605 * it as a lock order violation to enfore a general
606 * lock order of sleepable locks before non-sleepable
607 * locks. Thus, we only bother checking the lock
608 * order hierarchy if we pass the initial test.
610 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
611 ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
612 lock1->li_lock != &Giant.mtx_object)) &&
613 !isitmydescendant(w, w1))
616 * We have a lock order violation, check to see if it
617 * is allowed or has already been yelled about.
619 mtx_unlock_spin(&w_mtx);
622 if (lock1->li_lock == &Giant.mtx_object) {
623 if (w1->w_Giant_squawked)
626 w1->w_Giant_squawked = 1;
628 if (w1->w_other_squawked)
631 w1->w_other_squawked = 1;
636 printf("lock order reversal\n");
638 * Try to locate an earlier lock with
639 * witness w in our list.
642 lock2 = &lle->ll_children[i];
643 MPASS(lock2->li_lock != NULL);
644 if (lock2->li_lock->lo_witness == w)
647 if (i == 0 && lle->ll_next != NULL) {
649 i = lle->ll_count - 1;
654 printf(" 1st %p %s (%s) @ %s:%d\n",
655 lock1->li_lock, lock1->li_lock->lo_name,
656 lock1->li_lock->lo_type, lock1->li_file,
658 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
659 lock->lo_name, lock->lo_type, file, line);
661 printf(" 1st %p %s (%s) @ %s:%d\n",
662 lock2->li_lock, lock2->li_lock->lo_name,
663 lock2->li_lock->lo_type, lock2->li_file,
665 printf(" 2nd %p %s (%s) @ %s:%d\n",
666 lock1->li_lock, lock1->li_lock->lo_name,
667 lock1->li_lock->lo_type, lock1->li_file,
669 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
670 lock->lo_name, lock->lo_type, file, line);
678 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
680 * Don't build a new relationship if we are locking Giant just
681 * after waking up and the previous lock in the list was acquired
684 if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
685 mtx_unlock_spin(&w_mtx);
687 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
688 lock->lo_type, lock1->li_lock->lo_type);
689 if (!itismychild(lock1->li_lock->lo_witness, w))
690 mtx_unlock_spin(&w_mtx);
695 if (witness_ddb && go_into_ddb)
702 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
703 lle = witness_lock_list_get();
706 lle->ll_next = *lock_list;
707 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
708 td->td_proc->p_pid, lle);
711 lock1 = &lle->ll_children[lle->ll_count++];
712 lock1->li_lock = lock;
713 lock1->li_line = line;
714 lock1->li_file = file;
715 if ((flags & LOP_EXCLUSIVE) != 0)
716 lock1->li_flags = LI_EXCLUSIVE;
719 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
720 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
724 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
726 struct lock_instance *instance;
727 struct lock_class *class;
729 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
730 if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
732 class = lock->lo_class;
733 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
734 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
735 class->lc_name, lock->lo_name, file, line);
736 if ((flags & LOP_TRYLOCK) == 0)
737 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
738 lock->lo_name, file, line);
739 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
740 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
741 class->lc_name, lock->lo_name, file, line);
742 instance = find_instance(curthread->td_sleeplocks, lock);
743 if (instance == NULL)
744 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
745 class->lc_name, lock->lo_name, file, line);
746 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
747 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
748 class->lc_name, lock->lo_name, file, line);
749 if ((instance->li_flags & LI_RECURSEMASK) != 0)
750 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
751 class->lc_name, lock->lo_name,
752 instance->li_flags & LI_RECURSEMASK, file, line);
753 instance->li_flags |= LI_EXCLUSIVE;
757 witness_downgrade(struct lock_object *lock, int flags, const char *file,
760 struct lock_instance *instance;
761 struct lock_class *class;
763 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
764 if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
766 class = lock->lo_class;
767 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
768 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
769 class->lc_name, lock->lo_name, file, line);
770 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
771 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
772 class->lc_name, lock->lo_name, file, line);
773 instance = find_instance(curthread->td_sleeplocks, lock);
774 if (instance == NULL)
775 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
776 class->lc_name, lock->lo_name, file, line);
777 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
778 panic("downgrade of shared lock (%s) %s @ %s:%d",
779 class->lc_name, lock->lo_name, file, line);
780 if ((instance->li_flags & LI_RECURSEMASK) != 0)
781 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
782 class->lc_name, lock->lo_name,
783 instance->li_flags & LI_RECURSEMASK, file, line);
784 instance->li_flags &= ~LI_EXCLUSIVE;
788 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
790 struct lock_list_entry **lock_list, *lle;
791 struct lock_instance *instance;
792 struct lock_class *class;
797 if (witness_cold || witness_dead || lock->lo_witness == NULL ||
801 class = lock->lo_class;
802 if (class->lc_flags & LC_SLEEPLOCK)
803 lock_list = &td->td_sleeplocks;
805 lock_list = PCPU_PTR(spinlocks);
806 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
807 for (i = 0; i < (*lock_list)->ll_count; i++) {
808 instance = &(*lock_list)->ll_children[i];
809 if (instance->li_lock == lock) {
810 if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
811 (flags & LOP_EXCLUSIVE) == 0) {
813 "shared unlock of (%s) %s @ %s:%d\n",
814 class->lc_name, lock->lo_name,
817 "while exclusively locked from %s:%d\n",
820 panic("excl->ushare");
822 if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
823 (flags & LOP_EXCLUSIVE) != 0) {
825 "exclusive unlock of (%s) %s @ %s:%d\n",
826 class->lc_name, lock->lo_name,
829 "while share locked from %s:%d\n",
832 panic("share->uexcl");
834 /* If we are recursed, unrecurse. */
835 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
837 "%s: pid %d unrecursed on %s r=%d", __func__,
839 instance->li_lock->lo_name,
841 instance->li_flags--;
846 "%s: pid %d removed %s from lle[%d]", __func__,
848 instance->li_lock->lo_name,
849 (*lock_list)->ll_count - 1);
850 (*lock_list)->ll_count--;
851 for (j = i; j < (*lock_list)->ll_count; j++)
852 (*lock_list)->ll_children[j] =
853 (*lock_list)->ll_children[j + 1];
855 if ((*lock_list)->ll_count == 0) {
857 *lock_list = lle->ll_next;
859 "%s: pid %d removed lle %p", __func__,
860 td->td_proc->p_pid, lle);
861 witness_lock_list_free(lle);
866 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
871 * Warn if any held locks are not sleepable. Note that Giant and the lock
872 * passed in are both special cases since they are both released during the
873 * sleep process and aren't actually held while the thread is asleep.
876 witness_sleep(int check_only, struct lock_object *lock, const char *file,
879 struct lock_list_entry **lock_list, *lle;
880 struct lock_instance *lock1;
884 if (witness_dead || panicstr != NULL)
886 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
889 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
893 lock_list = &td->td_sleeplocks;
895 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
896 for (i = lle->ll_count - 1; i >= 0; i--) {
897 lock1 = &lle->ll_children[i];
898 if (lock1->li_lock == lock ||
899 lock1->li_lock == &Giant.mtx_object)
901 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
902 if (check_only == 0) {
904 "pid %d: sleeping with lock (%s) %s held",
906 lock1->li_lock->lo_class->lc_name,
907 lock1->li_lock->lo_name);
908 lock1->li_flags |= LI_SLEPT;
913 printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
914 file, line, check_only ? "could sleep" : "sleeping",
915 lock1->li_lock->lo_name, lock1->li_file,
918 if (lock_list == &td->td_sleeplocks) {
919 lock_list = PCPU_PTR(spinlocks);
923 if (witness_ddb && n)
930 static struct witness *
931 enroll(const char *description, struct lock_class *lock_class)
935 if (!witness_watch || witness_dead || panicstr != NULL)
937 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
939 mtx_lock_spin(&w_mtx);
940 STAILQ_FOREACH(w, &w_all, w_list) {
941 if (strcmp(description, w->w_name) == 0) {
943 mtx_unlock_spin(&w_mtx);
944 if (lock_class != w->w_class)
946 "lock (%s) %s does not match earlier (%s) lock",
947 description, lock_class->lc_name,
948 w->w_class->lc_name);
953 * This isn't quite right, as witness_cold is still 0 while we
954 * enroll all the locks initialized before witness_initialize().
956 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
957 mtx_unlock_spin(&w_mtx);
958 panic("spin lock %s not in order list", description);
960 if ((w = witness_get()) == NULL)
962 w->w_name = description;
963 w->w_class = lock_class;
965 STAILQ_INSERT_HEAD(&w_all, w, w_list);
966 if (lock_class->lc_flags & LC_SPINLOCK)
967 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
968 else if (lock_class->lc_flags & LC_SLEEPLOCK)
969 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
971 mtx_unlock_spin(&w_mtx);
972 panic("lock class %s is not sleep or spin",
973 lock_class->lc_name);
975 mtx_unlock_spin(&w_mtx);
980 itismychild(struct witness *parent, struct witness *child)
983 struct witness_child_list_entry **wcl;
984 struct witness_list *list;
986 MPASS(child != NULL && parent != NULL);
987 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
988 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
990 "%s: parent (%s) and child (%s) are not the same lock type",
991 __func__, parent->w_class->lc_name,
992 child->w_class->lc_name);
995 * Insert "child" after "parent"
997 wcl = &parent->w_children;
998 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
999 wcl = &(*wcl)->wcl_next;
1001 *wcl = witness_child_get();
1005 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1008 * Now prune whole tree. We look for cases where a lock is now
1009 * both a descendant and a direct child of a given lock. In that
1010 * case, we want to remove the direct child link from the tree.
1015 if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1019 STAILQ_FOREACH(child, list, w_typelist) {
1020 STAILQ_FOREACH(parent, list, w_typelist) {
1021 if (!isitmychild(parent, child))
1023 removechild(parent, child);
1024 if (isitmydescendant(parent, child))
1026 itismychild(parent, child);
1035 removechild(struct witness *parent, struct witness *child)
1037 struct witness_child_list_entry **wcl, *wcl1;
1040 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1041 for (i = 0; i < (*wcl)->wcl_count; i++)
1042 if ((*wcl)->wcl_children[i] == child)
1046 (*wcl)->wcl_count--;
1047 if ((*wcl)->wcl_count > i)
1048 (*wcl)->wcl_children[i] =
1049 (*wcl)->wcl_children[(*wcl)->wcl_count];
1050 MPASS((*wcl)->wcl_children[i] != NULL);
1051 if ((*wcl)->wcl_count != 0)
1054 *wcl = wcl1->wcl_next;
1055 witness_child_free(wcl1);
1059 isitmychild(struct witness *parent, struct witness *child)
1061 struct witness_child_list_entry *wcl;
1064 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1065 for (i = 0; i < wcl->wcl_count; i++) {
1066 if (wcl->wcl_children[i] == child)
1074 isitmydescendant(struct witness *parent, struct witness *child)
1076 struct witness_child_list_entry *wcl;
1079 if (isitmychild(parent, child))
1082 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1084 for (i = 0; i < wcl->wcl_count; i++) {
1085 if (isitmydescendant(wcl->wcl_children[i], child))
1094 witness_levelall (void)
1096 struct witness_list *list;
1097 struct witness *w, *w1;
1100 * First clear all levels.
1102 STAILQ_FOREACH(w, &w_all, w_list) {
1107 * Look for locks with no parent and level all their descendants.
1109 STAILQ_FOREACH(w, &w_all, w_list) {
1111 * This is just an optimization, technically we could get
1112 * away just walking the all list each time.
1114 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1118 STAILQ_FOREACH(w1, list, w_typelist) {
1119 if (isitmychild(w1, w))
1122 witness_leveldescendents(w, 0);
1124 ; /* silence GCC 3.x */
1129 witness_leveldescendents(struct witness *parent, int level)
1131 struct witness_child_list_entry *wcl;
1134 if (parent->w_level < level)
1135 parent->w_level = level;
1137 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1138 for (i = 0; i < wcl->wcl_count; i++)
1139 witness_leveldescendents(wcl->wcl_children[i], level);
1143 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1144 struct witness *parent)
1146 struct witness_child_list_entry *wcl;
1149 level = parent->w_level;
1150 prnt("%-2d", level);
1151 for (i = 0; i < level; i++)
1153 prnt("%s", parent->w_name);
1154 if (parent->w_file != NULL)
1155 prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1157 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1158 for (i = 0; i < wcl->wcl_count; i++)
1159 witness_displaydescendants(prnt,
1160 wcl->wcl_children[i]);
1164 blessed(struct witness *w1, struct witness *w2)
1167 struct witness_blessed *b;
1169 for (i = 0; i < blessed_count; i++) {
1170 b = &blessed_list[i];
1171 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1172 if (strcmp(w2->w_name, b->b_lock2) == 0)
1176 if (strcmp(w1->w_name, b->b_lock2) == 0)
1177 if (strcmp(w2->w_name, b->b_lock1) == 0)
1183 static struct witness *
1189 mtx_unlock_spin(&w_mtx);
1192 if (STAILQ_EMPTY(&w_free)) {
1194 mtx_unlock_spin(&w_mtx);
1195 printf("%s: witness exhausted\n", __func__);
1198 w = STAILQ_FIRST(&w_free);
1199 STAILQ_REMOVE_HEAD(&w_free, w_list);
1200 bzero(w, sizeof(*w));
1205 witness_free(struct witness *w)
1208 STAILQ_INSERT_HEAD(&w_free, w, w_list);
1211 static struct witness_child_list_entry *
1212 witness_child_get(void)
1214 struct witness_child_list_entry *wcl;
1217 mtx_unlock_spin(&w_mtx);
1223 mtx_unlock_spin(&w_mtx);
1224 printf("%s: witness exhausted\n", __func__);
1227 w_child_free = wcl->wcl_next;
1228 bzero(wcl, sizeof(*wcl));
1233 witness_child_free(struct witness_child_list_entry *wcl)
1236 wcl->wcl_next = w_child_free;
1240 static struct lock_list_entry *
1241 witness_lock_list_get(void)
1243 struct lock_list_entry *lle;
1247 mtx_lock_spin(&w_mtx);
1248 lle = w_lock_list_free;
1251 mtx_unlock_spin(&w_mtx);
1252 printf("%s: witness exhausted\n", __func__);
1255 w_lock_list_free = lle->ll_next;
1256 mtx_unlock_spin(&w_mtx);
1257 bzero(lle, sizeof(*lle));
1262 witness_lock_list_free(struct lock_list_entry *lle)
1265 mtx_lock_spin(&w_mtx);
1266 lle->ll_next = w_lock_list_free;
1267 w_lock_list_free = lle;
1268 mtx_unlock_spin(&w_mtx);
1271 static struct lock_instance *
1272 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1274 struct lock_list_entry *lle;
1275 struct lock_instance *instance;
1278 for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1279 for (i = lle->ll_count - 1; i >= 0; i--) {
1280 instance = &lle->ll_children[i];
1281 if (instance->li_lock == lock)
1288 witness_list_locks(struct lock_list_entry **lock_list)
1290 struct lock_list_entry *lle;
1291 struct lock_instance *instance;
1292 struct lock_object *lock;
1296 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1297 for (i = lle->ll_count - 1; i >= 0; i--) {
1298 instance = &lle->ll_children[i];
1299 lock = instance->li_lock;
1301 (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1302 "exclusive" : "shared",
1303 lock->lo_class->lc_name, lock->lo_name);
1304 if (lock->lo_type != lock->lo_name)
1305 printf(" (%s)", lock->lo_type);
1306 printf(" r = %d (%p) locked @ %s:%d\n",
1307 instance->li_flags & LI_RECURSEMASK, lock,
1308 instance->li_file, instance->li_line);
1315 * Calling this on td != curthread is bad unless we are in ddb.
1318 witness_list(struct thread *td)
1322 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1324 KASSERT(td == curthread || db_active,
1325 ("%s: td != curthread and we aren't in the debugger", __func__));
1326 if (!db_active && witness_dead)
1329 KASSERT(td == curthread, ("%s: p != curthread", __func__));
1333 nheld = witness_list_locks(&td->td_sleeplocks);
1336 * We only handle spinlocks if td == curthread. This is somewhat broken
1337 * if td is currently executing on some other CPU and holds spin locks
1338 * as we won't display those locks. If we had a MI way of getting
1339 * the per-cpu data for a given cpu then we could use
1340 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1343 * That still wouldn't really fix this unless we locked sched_lock
1344 * or stopped the other CPU to make sure it wasn't changing the list
1345 * out from under us. It is probably best to just not try to handle
1346 * threads on other CPU's for now.
1348 if (td == curthread) {
1350 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1354 nheld += witness_list_locks(PCPU_PTR(spinlocks));
1361 witness_save(struct lock_object *lock, const char **filep, int *linep)
1363 struct lock_instance *instance;
1365 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1366 if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1368 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1369 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1370 lock->lo_class->lc_name, lock->lo_name);
1371 instance = find_instance(curthread->td_sleeplocks, lock);
1372 if (instance == NULL)
1373 panic("%s: lock (%s) %s not locked", __func__,
1374 lock->lo_class->lc_name, lock->lo_name);
1375 *filep = instance->li_file;
1376 *linep = instance->li_line;
1380 witness_restore(struct lock_object *lock, const char *file, int line)
1382 struct lock_instance *instance;
1384 KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1385 if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1387 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1388 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1389 lock->lo_class->lc_name, lock->lo_name);
1390 instance = find_instance(curthread->td_sleeplocks, lock);
1391 if (instance == NULL)
1392 panic("%s: lock (%s) %s not locked", __func__,
1393 lock->lo_class->lc_name, lock->lo_name);
1394 lock->lo_witness->w_file = file;
1395 lock->lo_witness->w_line = line;
1396 instance->li_file = file;
1397 instance->li_line = line;
1401 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1403 #ifdef INVARIANT_SUPPORT
1404 struct lock_instance *instance;
1406 if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1408 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1409 instance = find_instance(curthread->td_sleeplocks, lock);
1410 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1411 instance = find_instance(PCPU_GET(spinlocks), lock);
1413 panic("Lock (%s) %s is not sleep or spin!",
1414 lock->lo_class->lc_name, lock->lo_name);
1419 if (instance != NULL)
1420 panic("Lock (%s) %s locked @ %s:%d.",
1421 lock->lo_class->lc_name, lock->lo_name, file, line);
1424 case LA_LOCKED | LA_RECURSED:
1425 case LA_LOCKED | LA_NOTRECURSED:
1427 case LA_SLOCKED | LA_RECURSED:
1428 case LA_SLOCKED | LA_NOTRECURSED:
1430 case LA_XLOCKED | LA_RECURSED:
1431 case LA_XLOCKED | LA_NOTRECURSED:
1432 if (instance == NULL) {
1433 panic("Lock (%s) %s not locked @ %s:%d.",
1434 lock->lo_class->lc_name, lock->lo_name, file, line);
1437 if ((flags & LA_XLOCKED) != 0 &&
1438 (instance->li_flags & LI_EXCLUSIVE) == 0)
1439 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1440 lock->lo_class->lc_name, lock->lo_name, file, line);
1441 if ((flags & LA_SLOCKED) != 0 &&
1442 (instance->li_flags & LI_EXCLUSIVE) != 0)
1443 panic("Lock (%s) %s exclusively locked @ %s:%d.",
1444 lock->lo_class->lc_name, lock->lo_name, file, line);
1445 if ((flags & LA_RECURSED) != 0 &&
1446 (instance->li_flags & LI_RECURSEMASK) == 0)
1447 panic("Lock (%s) %s not recursed @ %s:%d.",
1448 lock->lo_class->lc_name, lock->lo_name, file, line);
1449 if ((flags & LA_NOTRECURSED) != 0 &&
1450 (instance->li_flags & LI_RECURSEMASK) != 0)
1451 panic("Lock (%s) %s recursed @ %s:%d.",
1452 lock->lo_class->lc_name, lock->lo_name, file, line);
1455 panic("Invalid lock assertion at %s:%d.", file, line);
1458 #endif /* INVARIANT_SUPPORT */
1463 DB_SHOW_COMMAND(locks, db_witness_list)
1470 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1471 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1472 ((addr >> 16) % 16) * 10000;
1473 /* sx_slock(&allproc_lock); */
1474 FOREACH_PROC_IN_SYSTEM(p) {
1475 if (p->p_pid == pid)
1478 /* sx_sunlock(&allproc_lock); */
1480 db_printf("pid %d not found\n", pid);
1483 FOREACH_THREAD_IN_PROC(p, td) {
1492 DB_SHOW_COMMAND(witness, db_witness_display)
1495 witness_display(db_printf);