subr_witness.c revision 82244
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_witness.c 82244 2001-08-23 22:47:05Z jhb $
31 */
32
33/*
34 * Implementation of the `witness' lock verifier.  Originally implemented for
35 * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36 * classes in FreeBSD.
37 */
38
39/*
40 *	Main Entry: witness
41 *	Pronunciation: 'wit-n&s
42 *	Function: noun
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
55 *	      life -- Pilot>
56 *	6 capitalized : a member of the Jehovah's Witnesses
57 */
58
59#include "opt_ddb.h"
60#include "opt_witness.h"
61
62#include <sys/param.h>
63#include <sys/bus.h>
64#include <sys/kernel.h>
65#include <sys/ktr.h>
66#include <sys/lock.h>
67#include <sys/malloc.h>
68#include <sys/mutex.h>
69#include <sys/proc.h>
70#include <sys/sysctl.h>
71#include <sys/systm.h>
72
73#include <ddb/ddb.h>
74
75#define WITNESS_COUNT 200
76#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77/*
78 * XXX: This is somewhat bogus, as we assume here that at most 1024 processes
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.
81 */
82#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83
84#define	WITNESS_NCHILDREN 6
85
86struct witness_child_list_entry;
87
88struct witness {
89	const	char *w_name;
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... */
94	const	char *w_file;
95	int	w_line;
96	u_int	w_level;
97	u_int	w_refcount;
98	u_char	w_Giant_squawked:1;
99	u_char	w_other_squawked:1;
100	u_char	w_same_squawked:1;
101};
102
103struct witness_child_list_entry {
104	struct	witness_child_list_entry *wcl_next;
105	struct	witness *wcl_children[WITNESS_NCHILDREN];
106	u_int	wcl_count;
107};
108
109STAILQ_HEAD(witness_list, witness);
110
111struct witness_blessed {
112	const	char *b_lock1;
113	const	char *b_lock2;
114};
115
116struct witness_order_list_entry {
117	const	char *w_name;
118	struct	lock_class *w_class;
119};
120
121static struct	witness *enroll(const char *description,
122				struct lock_class *lock_class);
123static int	itismychild(struct witness *parent, struct witness *child);
124static void	removechild(struct witness *parent, struct witness *child);
125static int	isitmychild(struct witness *parent, struct witness *child);
126static int	isitmydescendant(struct witness *parent, struct witness *child);
127static int	dup_ok(struct witness *);
128static int	blessed(struct witness *, struct witness *);
129static void	witness_display_list(void(*prnt)(const char *fmt, ...),
130				     struct witness_list *list);
131static void	witness_displaydescendants(void(*)(const char *fmt, ...),
132					   struct witness *);
133static void	witness_leveldescendents(struct witness *parent, int level);
134static void	witness_levelall(void);
135static struct	witness *witness_get(void);
136static void	witness_free(struct witness *m);
137static struct	witness_child_list_entry *witness_child_get(void);
138static void	witness_child_free(struct witness_child_list_entry *wcl);
139static struct	lock_list_entry *witness_lock_list_get(void);
140static void	witness_lock_list_free(struct lock_list_entry *lle);
141static void	witness_display(void(*)(const char *fmt, ...));
142static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
143					     struct lock_object *lock);
144
145MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146
147static int witness_watch = 1;
148TUNABLE_INT("debug.witness_watch", &witness_watch);
149SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
150
151#ifdef DDB
152/*
153 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
154 * drop into kdebug() when:
155 *	- a lock heirarchy violation occurs
156 *	- locks are held when going to sleep.
157 */
158#ifdef WITNESS_DDB
159int	witness_ddb = 1;
160#else
161int	witness_ddb = 0;
162#endif
163TUNABLE_INT("debug.witness_ddb", &witness_ddb);
164SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
165#endif /* DDB */
166
167#ifdef WITNESS_SKIPSPIN
168int	witness_skipspin = 1;
169#else
170int	witness_skipspin = 0;
171#endif
172TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
173SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
174    "");
175
176static struct mtx w_mtx;
177static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
178static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
179static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
180static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
181static struct witness_child_list_entry *w_child_free = NULL;
182static struct lock_list_entry *w_lock_list_free = NULL;
183static int witness_dead;	/* fatal error, probably no memory */
184
185static struct witness w_data[WITNESS_COUNT];
186static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
187static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188
189static struct witness_order_list_entry order_lists[] = {
190	{ "Giant", &lock_class_mtx_sleep },
191	{ "proctree", &lock_class_sx },
192	{ "allproc", &lock_class_sx },
193	{ "process lock", &lock_class_mtx_sleep },
194	{ "uidinfo hash", &lock_class_mtx_sleep },
195	{ "uidinfo struct", &lock_class_mtx_sleep },
196	{ NULL, NULL },
197	/*
198	 * spin locks
199	 */
200#if defined(__i386__) && defined (SMP)
201	{ "com", &lock_class_mtx_spin },
202#endif
203	{ "sio", &lock_class_mtx_spin },
204#ifdef __i386__
205	{ "cy", &lock_class_mtx_spin },
206#endif
207	{ "ng_node", &lock_class_mtx_spin },
208	{ "ng_worklist", &lock_class_mtx_spin },
209	{ "ithread table lock", &lock_class_mtx_spin },
210	{ "sched lock", &lock_class_mtx_spin },
211	{ "callout", &lock_class_mtx_spin },
212	/*
213	 * leaf locks
214	 */
215#ifdef SMP
216	{ "ap boot", &lock_class_mtx_spin },
217#ifdef __i386__
218	{ "imen", &lock_class_mtx_spin },
219#endif
220	{ "smp rendezvous", &lock_class_mtx_spin },
221#endif
222	{ "clk", &lock_class_mtx_spin },
223	{ NULL, NULL },
224	{ NULL, NULL }
225};
226
227static const char *dup_list[] = {
228	"process lock",
229	NULL
230};
231
232/*
233 * Pairs of locks which have been blessed
234 * Don't complain about order problems with blessed locks
235 */
236static struct witness_blessed blessed_list[] = {
237};
238static int blessed_count =
239	sizeof(blessed_list) / sizeof(struct witness_blessed);
240
241/*
242 * List of all locks in the system.
243 */
244STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245
246static struct mtx all_mtx = {
247	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248	  "All locks list",		/* mtx_object.lo_name */
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	0,				/* mtx_savecrit */
254	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
255	{ NULL, NULL }			/* mtx_contested */
256};
257
258/*
259 * This global is set to 0 once it becomes safe to use the witness code.
260 */
261static int witness_cold = 1;
262
263/*
264 * Global variables for book keeping.
265 */
266static int lock_cur_cnt;
267static int lock_max_cnt;
268
269/*
270 * The WITNESS-enabled diagnostic code.
271 */
272static void
273witness_initialize(void *dummy __unused)
274{
275	struct lock_object *lock;
276	struct witness_order_list_entry *order;
277	struct witness *w, *w1;
278	int i;
279
280	/*
281	 * We have to release Giant before initializing its witness
282	 * structure so that WITNESS doesn't get confused.
283	 */
284	mtx_unlock(&Giant);
285	mtx_assert(&Giant, MA_NOTOWNED);
286
287	CTR0(KTR_WITNESS, __func__ ": initializing witness");
288	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
289	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
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]);
296
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);
300		if (w == NULL)
301			continue;
302		w->w_file = "order list";
303		for (order++; order->w_name != NULL; order++) {
304			w1 = enroll(order->w_name, order->w_class);
305			if (w1 == NULL)
306				continue;
307			w1->w_file = "order list";
308			itismychild(w, w1);
309			w = w1;
310		}
311	}
312
313	/* Iterate through all locks and add them to witness. */
314	mtx_lock(&all_mtx);
315	STAILQ_FOREACH(lock, &all_locks, lo_list) {
316		if (lock->lo_flags & LO_WITNESS)
317			lock->lo_witness = enroll(lock->lo_name,
318			    lock->lo_class);
319		else
320			lock->lo_witness = NULL;
321	}
322	mtx_unlock(&all_mtx);
323
324	/* Mark the witness code as being ready for use. */
325	atomic_store_rel_int(&witness_cold, 0);
326
327	mtx_lock(&Giant);
328}
329SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
330
331void
332witness_init(struct lock_object *lock)
333{
334	struct lock_class *class;
335
336	class = lock->lo_class;
337	if (lock->lo_flags & LO_INITIALIZED)
338		panic("%s: lock (%s) %s is already initialized!\n", __func__,
339		    class->lc_name, lock->lo_name);
340
341	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
342	    (class->lc_flags & LC_RECURSABLE) == 0)
343		panic("%s: lock (%s) %s can not be recursable!\n", __func__,
344		    class->lc_name, lock->lo_name);
345
346	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
347	    (class->lc_flags & LC_SLEEPABLE) == 0)
348		panic("%s: lock (%s) %s can not be sleepable!\n", __func__,
349		    class->lc_name, lock->lo_name);
350
351	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
352	    (class->lc_flags & LC_UPGRADABLE) == 0)
353		panic("%s: lock (%s) %s can not be upgradable!\n", __func__,
354		    class->lc_name, lock->lo_name);
355
356	mtx_lock(&all_mtx);
357	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
358	lock->lo_flags |= LO_INITIALIZED;
359	lock_cur_cnt++;
360	if (lock_cur_cnt > lock_max_cnt)
361		lock_max_cnt = lock_cur_cnt;
362	mtx_unlock(&all_mtx);
363	if (!witness_cold && !witness_dead && panicstr == NULL &&
364	    (lock->lo_flags & LO_WITNESS) != 0)
365		lock->lo_witness = enroll(lock->lo_name, class);
366	else
367		lock->lo_witness = NULL;
368}
369
370void
371witness_destroy(struct lock_object *lock)
372{
373	struct witness *w;
374
375	if (witness_cold)
376		panic("lock (%s) %s destroyed while witness_cold",
377		    lock->lo_class->lc_name, lock->lo_name);
378
379	if ((lock->lo_flags & LO_INITIALIZED) == 0)
380		panic("%s: lock (%s) %s is not initialized!\n", __func__,
381		    lock->lo_class->lc_name, lock->lo_name);
382
383	/* XXX: need to verify that no one holds the lock */
384	w = lock->lo_witness;
385	if (w != NULL) {
386		mtx_lock_spin(&w_mtx);
387		w->w_refcount--;
388		if (w->w_refcount == 0) {
389			CTR1(KTR_WITNESS,
390			    __func__ ": marking witness %s as dead", w->w_name);
391			w->w_name = "(dead)";
392			w->w_file = "(dead)";
393			w->w_line = 0;
394		}
395		mtx_unlock_spin(&w_mtx);
396	}
397
398	mtx_lock(&all_mtx);
399	lock_cur_cnt--;
400	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
401	lock->lo_flags &= ~LO_INITIALIZED;
402	mtx_unlock(&all_mtx);
403}
404
405static void
406witness_display_list(void(*prnt)(const char *fmt, ...),
407		     struct witness_list *list)
408{
409	struct witness *w, *w1;
410	int found;
411
412	STAILQ_FOREACH(w, list, w_typelist) {
413		if (w->w_file == NULL)
414			continue;
415		found = 0;
416		STAILQ_FOREACH(w1, list, w_typelist) {
417			if (isitmychild(w1, w)) {
418				found++;
419				break;
420			}
421		}
422		if (found)
423			continue;
424		/*
425		 * This lock has no anscestors, display its descendants.
426		 */
427		witness_displaydescendants(prnt, w);
428	}
429}
430
431static void
432witness_display(void(*prnt)(const char *fmt, ...))
433{
434	struct witness *w;
435
436	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
437	witness_levelall();
438
439	/*
440	 * First, handle sleep locks which have been acquired at least
441	 * once.
442	 */
443	prnt("Sleep locks:\n");
444	witness_display_list(prnt, &w_sleep);
445
446	/*
447	 * Now do spin locks which have been acquired at least once.
448	 */
449	prnt("\nSpin locks:\n");
450	witness_display_list(prnt, &w_spin);
451
452	/*
453	 * Finally, any locks which have not been acquired yet.
454	 */
455	prnt("\nLocks which were never acquired:\n");
456	STAILQ_FOREACH(w, &w_all, w_list) {
457		if (w->w_file != NULL)
458			continue;
459		prnt("%s\n", w->w_name);
460	}
461}
462
463void
464witness_lock(struct lock_object *lock, int flags, const char *file, int line)
465{
466	struct lock_list_entry **lock_list, *lle;
467	struct lock_instance *lock1, *lock2;
468	struct lock_class *class;
469	struct witness *w, *w1;
470	struct proc *p;
471	int i, j;
472#ifdef DDB
473	int go_into_ddb = 0;
474#endif /* DDB */
475
476	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
477	    panicstr != NULL)
478		return;
479	w = lock->lo_witness;
480	class = lock->lo_class;
481	p = curproc;
482
483	/*
484	 * We have to hold a spinlock to keep lock_list valid across the check
485	 * in the LC_SLEEPLOCK case.  In the LC_SPINLOCK case, it is already
486	 * protected by the spinlock we are currently performing the witness
487	 * checks on, so it is ok to release the lock after performing this
488	 * check.  All we have to protect is the LC_SLEEPLOCK case when no
489	 * spinlocks are held as we may get preempted during this check and
490	 * lock_list could end up pointing to some other CPU's spinlock list.
491	 */
492	mtx_lock_spin(&w_mtx);
493	lock_list = PCPU_PTR(spinlocks);
494	if (class->lc_flags & LC_SLEEPLOCK) {
495		if (*lock_list != NULL && (flags & LOP_TRYLOCK) == 0) {
496			mtx_unlock_spin(&w_mtx);
497			panic("blockable sleep lock (%s) %s @ %s:%d",
498			    class->lc_name, lock->lo_name, file, line);
499		}
500		lock_list = &p->p_sleeplocks;
501	}
502	mtx_unlock_spin(&w_mtx);
503
504	/*
505	 * Try locks do not block if they fail to acquire the lock, thus
506	 * there is no danger of deadlocks or of switching while holding a
507	 * spin lock if we acquire a lock via a try operation.
508	 */
509	if (flags & LOP_TRYLOCK)
510		goto out;
511
512	/*
513	 * Is this the first lock acquired?  If so, then no order checking
514	 * is needed.
515	 */
516	if (*lock_list == NULL)
517		goto out;
518
519	/*
520	 * Check to see if we are recursing on a lock we already own.
521	 */
522	lock1 = find_instance(*lock_list, lock);
523	if (lock1 != NULL) {
524		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
525		    (flags & LOP_EXCLUSIVE) == 0) {
526			printf("shared lock of (%s) %s @ %s:%d\n",
527			    class->lc_name, lock->lo_name, file, line);
528			printf("while exclusively locked from %s:%d\n",
529			    lock1->li_file, lock1->li_line);
530			panic("share->excl");
531		}
532		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
533		    (flags & LOP_EXCLUSIVE) != 0) {
534			printf("exclusive lock of (%s) %s @ %s:%d\n",
535			    class->lc_name, lock->lo_name, file, line);
536			printf("while share locked from %s:%d\n",
537			    lock1->li_file, lock1->li_line);
538			panic("excl->share");
539		}
540		lock1->li_flags++;
541		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
542			printf(
543			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
544			    class->lc_name, lock->lo_name, file, line);
545			printf("first acquired @ %s:%d\n", lock1->li_file,
546			    lock1->li_line);
547			panic("recurse");
548		}
549		CTR3(KTR_WITNESS, __func__ ": pid %d recursed on %s r=%d",
550		    curproc->p_pid, lock->lo_name,
551		    lock1->li_flags & LI_RECURSEMASK);
552		lock1->li_file = file;
553		lock1->li_line = line;
554		return;
555	}
556
557	/*
558	 * Check for duplicate locks of the same type.  Note that we only
559	 * have to check for this on the last lock we just acquired.  Any
560	 * other cases will be caught as lock order violations.
561	 */
562	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
563	w1 = lock1->li_lock->lo_witness;
564	if (w1 == w) {
565		if (w->w_same_squawked || dup_ok(w))
566			goto out;
567		w->w_same_squawked = 1;
568		printf("acquiring duplicate lock of same type: \"%s\"\n",
569			lock->lo_name);
570		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
571		printf(" 2nd @ %s:%d\n", file, line);
572#ifdef DDB
573		go_into_ddb = 1;
574#endif /* DDB */
575		goto out;
576	}
577	MPASS(!mtx_owned(&w_mtx));
578	mtx_lock_spin(&w_mtx);
579	/*
580	 * If we have a known higher number just say ok
581	 */
582	if (witness_watch > 1 && w->w_level > w1->w_level) {
583		mtx_unlock_spin(&w_mtx);
584		goto out;
585	}
586	if (isitmydescendant(w1, w)) {
587		mtx_unlock_spin(&w_mtx);
588		goto out;
589	}
590	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
591		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
592
593			MPASS(j < WITNESS_COUNT);
594			lock1 = &lle->ll_children[i];
595			w1 = lock1->li_lock->lo_witness;
596
597			/*
598			 * If this lock doesn't undergo witness checking,
599			 * then skip it.
600			 */
601			if (w1 == NULL) {
602				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
603				    ("lock missing witness structure"));
604				continue;
605			}
606			/*
607			 * If we are locking Giant and we slept with this
608			 * lock, then skip it.
609			 */
610			if ((lock1->li_flags & LI_SLEPT) != 0 &&
611			    lock == &Giant.mtx_object)
612				continue;
613			if (!isitmydescendant(w, w1))
614				continue;
615			/*
616			 * We have a lock order violation, check to see if it
617			 * is allowed or has already been yelled about.
618			 */
619			mtx_unlock_spin(&w_mtx);
620			if (blessed(w, w1))
621				goto out;
622			if (lock1->li_lock == &Giant.mtx_object) {
623				if (w1->w_Giant_squawked)
624					goto out;
625				else
626					w1->w_Giant_squawked = 1;
627			} else {
628				if (w1->w_other_squawked)
629					goto out;
630				else
631					w1->w_other_squawked = 1;
632			}
633			/*
634			 * Ok, yell about it.
635			 */
636			printf("lock order reversal\n");
637			/*
638			 * Try to locate an earlier lock with
639			 * witness w in our list.
640			 */
641			do {
642				lock2 = &lle->ll_children[i];
643				MPASS(lock2->li_lock != NULL);
644				if (lock2->li_lock->lo_witness == w)
645					break;
646				i--;
647				if (i == 0 && lle->ll_next != NULL) {
648					lle = lle->ll_next;
649					i = lle->ll_count - 1;
650					MPASS(i != 0);
651				}
652			} while (i >= 0);
653			if (i < 0) {
654				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
655				    lock1->li_lock->lo_name, lock1->li_file,
656				    lock1->li_line);
657				printf(" 2nd %p %s @ %s:%d\n", lock,
658				    lock->lo_name, file, line);
659			} else {
660				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
661				    lock2->li_lock->lo_name, lock2->li_file,
662				    lock2->li_line);
663				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
664				    lock1->li_lock->lo_name, lock1->li_file,
665				    lock1->li_line);
666				printf(" 3rd %p %s @ %s:%d\n", lock,
667				    lock->lo_name, file, line);
668			}
669#ifdef DDB
670			go_into_ddb = 1;
671#endif /* DDB */
672			goto out;
673		}
674	}
675	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
676	/*
677	 * Don't build a new relationship if we are locking Giant just
678	 * after waking up and the previous lock in the list was acquired
679	 * prior to blocking.
680	 */
681	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
682		mtx_unlock_spin(&w_mtx);
683	else {
684		CTR2(KTR_WITNESS, __func__ ": adding %s as a child of %s",
685		    lock->lo_name, lock1->li_lock->lo_name);
686		if (!itismychild(lock1->li_lock->lo_witness, w))
687			mtx_unlock_spin(&w_mtx);
688	}
689
690out:
691#ifdef DDB
692	if (witness_ddb && go_into_ddb)
693		Debugger(__func__);
694#endif /* DDB */
695	w->w_file = file;
696	w->w_line = line;
697
698	lle = *lock_list;
699	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
700		lle = witness_lock_list_get();
701		if (lle == NULL)
702			return;
703		lle->ll_next = *lock_list;
704		CTR2(KTR_WITNESS, __func__ ": pid %d added lle %p",
705		    curproc->p_pid, lle);
706		*lock_list = lle;
707	}
708	lock1 = &lle->ll_children[lle->ll_count++];
709	lock1->li_lock = lock;
710	lock1->li_line = line;
711	lock1->li_file = file;
712	if ((flags & LOP_EXCLUSIVE) != 0)
713		lock1->li_flags = LI_EXCLUSIVE;
714	else
715		lock1->li_flags = 0;
716	CTR3(KTR_WITNESS, __func__ ": pid %d added %s as lle[%d]",
717	    curproc->p_pid, lock->lo_name, lle->ll_count - 1);
718}
719
720void
721witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
722{
723	struct lock_instance *instance;
724	struct lock_class *class;
725
726	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
727	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
728		return;
729
730	class = lock->lo_class;
731	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
732		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
733		    class->lc_name, lock->lo_name, file, line);
734	if ((flags & LOP_TRYLOCK) == 0)
735		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
736		    lock->lo_name, file, line);
737	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
738		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
739		    class->lc_name, lock->lo_name, file, line);
740	instance = find_instance(curproc->p_sleeplocks, lock);
741	if (instance == NULL)
742		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
743		    class->lc_name, lock->lo_name, file, line);
744	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
745		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
746		    class->lc_name, lock->lo_name, file, line);
747	if ((instance->li_flags & LI_RECURSEMASK) != 0)
748		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
749		    class->lc_name, lock->lo_name,
750		    instance->li_flags & LI_RECURSEMASK, file, line);
751	instance->li_flags |= LI_EXCLUSIVE;
752}
753
754void
755witness_downgrade(struct lock_object *lock, int flags, const char *file,
756    int line)
757{
758	struct lock_instance *instance;
759	struct lock_class *class;
760
761	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
762	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
763		return;
764
765	class = lock->lo_class;
766	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
767		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
768		    class->lc_name, lock->lo_name, file, line);
769	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
770		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
771		    class->lc_name, lock->lo_name, file, line);
772	instance = find_instance(curproc->p_sleeplocks, lock);
773	if (instance == NULL)
774		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
775		    class->lc_name, lock->lo_name, file, line);
776	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
777		panic("downgrade of shared lock (%s) %s @ %s:%d",
778		    class->lc_name, lock->lo_name, file, line);
779	if ((instance->li_flags & LI_RECURSEMASK) != 0)
780		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
781		    class->lc_name, lock->lo_name,
782		    instance->li_flags & LI_RECURSEMASK, file, line);
783	instance->li_flags &= ~LI_EXCLUSIVE;
784}
785
786void
787witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
788{
789	struct lock_list_entry **lock_list, *lle;
790	struct lock_instance *instance;
791	struct lock_class *class;
792	struct proc *p;
793	critical_t s;
794	int i, j;
795
796	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
797	    panicstr != NULL)
798		return;
799	p = curproc;
800	class = lock->lo_class;
801	if (class->lc_flags & LC_SLEEPLOCK)
802		lock_list = &p->p_sleeplocks;
803	else
804		lock_list = PCPU_PTR(spinlocks);
805	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
806		for (i = 0; i < (*lock_list)->ll_count; i++) {
807			instance = &(*lock_list)->ll_children[i];
808			if (instance->li_lock == lock) {
809				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
810				    (flags & LOP_EXCLUSIVE) == 0) {
811					printf(
812					"shared unlock of (%s) %s @ %s:%d\n",
813					    class->lc_name, lock->lo_name,
814					    file, line);
815					printf(
816					"while exclusively locked from %s:%d\n",
817					    instance->li_file,
818					    instance->li_line);
819					panic("excl->ushare");
820				}
821				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
822				    (flags & LOP_EXCLUSIVE) != 0) {
823					printf(
824					"exclusive unlock of (%s) %s @ %s:%d\n",
825					    class->lc_name, lock->lo_name,
826					    file, line);
827					printf(
828					"while share locked from %s:%d\n",
829					    instance->li_file,
830					    instance->li_line);
831					panic("share->uexcl");
832				}
833				/* If we are recursed, unrecurse. */
834				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
835					CTR3(KTR_WITNESS,
836				    __func__ ": pid %d unrecursed on %s r=%d",
837					    curproc->p_pid,
838					    instance->li_lock->lo_name,
839					    instance->li_flags);
840					instance->li_flags--;
841					goto out;
842				}
843				s = critical_enter();
844				CTR3(KTR_WITNESS,
845				    __func__ ": pid %d removed %s from lle[%d]",
846				    curproc->p_pid, instance->li_lock->lo_name,
847				    (*lock_list)->ll_count - 1);
848				(*lock_list)->ll_count--;
849				for (j = i; j < (*lock_list)->ll_count; j++)
850					(*lock_list)->ll_children[j] =
851					    (*lock_list)->ll_children[j + 1];
852				critical_exit(s);
853				if ((*lock_list)->ll_count == 0) {
854					lle = *lock_list;
855					*lock_list = lle->ll_next;
856					CTR2(KTR_WITNESS,
857					    __func__ ": pid %d removed lle %p",
858					    curproc->p_pid, lle);
859					witness_lock_list_free(lle);
860				}
861				goto out;
862			}
863		}
864	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
865	    file, line);
866out:
867	/*
868	 * We don't need to protect this PCPU_GET() here against preemption
869	 * because if we hold any spinlocks then we are already protected,
870	 * and if we don't we will get NULL if we hold no spinlocks even if
871	 * we switch CPU's while reading it.
872	 */
873	if (class->lc_flags & LC_SLEEPLOCK) {
874		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
875			panic("switchable sleep unlock (%s) %s @ %s:%d",
876			    class->lc_name, lock->lo_name, file, line);
877	}
878}
879
880/*
881 * Warn if any held locks are not sleepable.  Note that Giant and the lock
882 * passed in are both special cases since they are both released during the
883 * sleep process and aren't actually held while the process is asleep.
884 */
885int
886witness_sleep(int check_only, struct lock_object *lock, const char *file,
887	      int line)
888{
889	struct lock_list_entry **lock_list, *lle;
890	struct lock_instance *lock1;
891	struct proc *p;
892	critical_t savecrit;
893	int i, n;
894
895	if (witness_dead || panicstr != NULL)
896		return (0);
897	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
898	n = 0;
899	/*
900	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
901	 */
902	savecrit = critical_enter();
903	p = curproc;
904	lock_list = &p->p_sleeplocks;
905again:
906	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
907		for (i = lle->ll_count - 1; i >= 0; i--) {
908			lock1 = &lle->ll_children[i];
909			if (lock1->li_lock == lock ||
910			    lock1->li_lock == &Giant.mtx_object)
911				continue;
912			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
913				if (check_only == 0) {
914					CTR3(KTR_WITNESS,
915				    "pid %d: sleeping with lock (%s) %s held",
916					    curproc->p_pid,
917					    lock1->li_lock->lo_class->lc_name,
918					    lock1->li_lock->lo_name);
919					lock1->li_flags |= LI_SLEPT;
920				}
921				continue;
922			}
923			n++;
924			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
925			    file, line, check_only ? "could sleep" : "sleeping",
926			    lock1->li_lock->lo_name, lock1->li_file,
927			    lock1->li_line);
928		}
929	if (lock_list == &p->p_sleeplocks) {
930		lock_list = PCPU_PTR(spinlocks);
931		goto again;
932	}
933#ifdef DDB
934	if (witness_ddb && n)
935		Debugger(__func__);
936#endif /* DDB */
937	critical_exit(savecrit);
938	return (n);
939}
940
941static struct witness *
942enroll(const char *description, struct lock_class *lock_class)
943{
944	struct witness *w;
945
946	if (!witness_watch || witness_dead || panicstr != NULL)
947		return (NULL);
948
949	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
950		return (NULL);
951	mtx_lock_spin(&w_mtx);
952	STAILQ_FOREACH(w, &w_all, w_list) {
953		if (strcmp(description, w->w_name) == 0) {
954			w->w_refcount++;
955			mtx_unlock_spin(&w_mtx);
956			if (lock_class != w->w_class)
957				panic(
958				"lock (%s) %s does not match earlier (%s) lock",
959				    description, lock_class->lc_name,
960				    w->w_class->lc_name);
961			return (w);
962		}
963	}
964	/*
965	 * This isn't quite right, as witness_cold is still 0 while we
966	 * enroll all the locks initialized before witness_initialize().
967	 */
968	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
969		mtx_unlock_spin(&w_mtx);
970		panic("spin lock %s not in order list", description);
971	}
972	if ((w = witness_get()) == NULL)
973		return (NULL);
974	w->w_name = description;
975	w->w_class = lock_class;
976	w->w_refcount = 1;
977	STAILQ_INSERT_HEAD(&w_all, w, w_list);
978	if (lock_class->lc_flags & LC_SPINLOCK)
979		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
980	else if (lock_class->lc_flags & LC_SLEEPLOCK)
981		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
982	else {
983		mtx_unlock_spin(&w_mtx);
984		panic("lock class %s is not sleep or spin",
985		    lock_class->lc_name);
986	}
987	mtx_unlock_spin(&w_mtx);
988
989	return (w);
990}
991
992static int
993itismychild(struct witness *parent, struct witness *child)
994{
995	static int recursed;
996	struct witness_child_list_entry **wcl;
997	struct witness_list *list;
998
999	MPASS(child != NULL && parent != NULL);
1000	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1001	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1002		panic(
1003		"%s: parent (%s) and child (%s) are not the same lock type",
1004		    __func__, parent->w_class->lc_name,
1005		    child->w_class->lc_name);
1006
1007	/*
1008	 * Insert "child" after "parent"
1009	 */
1010	wcl = &parent->w_children;
1011	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1012		wcl = &(*wcl)->wcl_next;
1013
1014	if (*wcl == NULL) {
1015		*wcl = witness_child_get();
1016		if (*wcl == NULL)
1017			return (1);
1018	}
1019
1020	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1021
1022	/*
1023	 * Now prune whole tree.  We look for cases where a lock is now
1024	 * both a descendant and a direct child of a given lock.  In that
1025	 * case, we want to remove the direct child link from the tree.
1026	 */
1027	if (recursed)
1028		return (0);
1029	recursed = 1;
1030	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1031		list = &w_sleep;
1032	else
1033		list = &w_spin;
1034	STAILQ_FOREACH(child, list, w_typelist) {
1035		STAILQ_FOREACH(parent, list, w_typelist) {
1036			if (!isitmychild(parent, child))
1037				continue;
1038			removechild(parent, child);
1039			if (isitmydescendant(parent, child))
1040				continue;
1041			itismychild(parent, child);
1042		}
1043	}
1044	recursed = 0;
1045	witness_levelall();
1046	return (0);
1047}
1048
1049static void
1050removechild(struct witness *parent, struct witness *child)
1051{
1052	struct witness_child_list_entry **wcl, *wcl1;
1053	int i;
1054
1055	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1056		for (i = 0; i < (*wcl)->wcl_count; i++)
1057			if ((*wcl)->wcl_children[i] == child)
1058				goto found;
1059	return;
1060found:
1061	(*wcl)->wcl_count--;
1062	if ((*wcl)->wcl_count > i)
1063		(*wcl)->wcl_children[i] =
1064		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1065	MPASS((*wcl)->wcl_children[i] != NULL);
1066
1067	if ((*wcl)->wcl_count != 0)
1068		return;
1069
1070	wcl1 = *wcl;
1071	*wcl = wcl1->wcl_next;
1072	witness_child_free(wcl1);
1073}
1074
1075static int
1076isitmychild(struct witness *parent, struct witness *child)
1077{
1078	struct witness_child_list_entry *wcl;
1079	int i;
1080
1081	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1082		for (i = 0; i < wcl->wcl_count; i++) {
1083			if (wcl->wcl_children[i] == child)
1084				return (1);
1085		}
1086	}
1087	return (0);
1088}
1089
1090static int
1091isitmydescendant(struct witness *parent, struct witness *child)
1092{
1093	struct witness_child_list_entry *wcl;
1094	int i, j;
1095
1096	if (isitmychild(parent, child))
1097		return (1);
1098	j = 0;
1099	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1100		MPASS(j < 1000);
1101		for (i = 0; i < wcl->wcl_count; i++) {
1102			if (isitmydescendant(wcl->wcl_children[i], child))
1103				return (1);
1104		}
1105		j++;
1106	}
1107	return (0);
1108}
1109
1110void
1111witness_levelall (void)
1112{
1113	struct witness_list *list;
1114	struct witness *w, *w1;
1115
1116	/*
1117	 * First clear all levels.
1118	 */
1119	STAILQ_FOREACH(w, &w_all, w_list) {
1120		w->w_level = 0;
1121	}
1122
1123	/*
1124	 * Look for locks with no parent and level all their descendants.
1125	 */
1126	STAILQ_FOREACH(w, &w_all, w_list) {
1127		/*
1128		 * This is just an optimization, technically we could get
1129		 * away just walking the all list each time.
1130		 */
1131		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1132			list = &w_sleep;
1133		else
1134			list = &w_spin;
1135		STAILQ_FOREACH(w1, list, w_typelist) {
1136			if (isitmychild(w1, w))
1137				goto skip;
1138		}
1139		witness_leveldescendents(w, 0);
1140	skip:
1141	}
1142}
1143
1144static void
1145witness_leveldescendents(struct witness *parent, int level)
1146{
1147	struct witness_child_list_entry *wcl;
1148	int i;
1149
1150	if (parent->w_level < level)
1151		parent->w_level = level;
1152	level++;
1153	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1154		for (i = 0; i < wcl->wcl_count; i++)
1155			witness_leveldescendents(wcl->wcl_children[i], level);
1156}
1157
1158static void
1159witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1160			   struct witness *parent)
1161{
1162	struct witness_child_list_entry *wcl;
1163	int i, level;
1164
1165	level =  parent->w_level;
1166
1167	prnt("%-2d", level);
1168	for (i = 0; i < level; i++)
1169		prnt(" ");
1170	prnt("%s", parent->w_name);
1171	if (parent->w_file != NULL)
1172		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1173		    parent->w_line);
1174
1175	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1176		for (i = 0; i < wcl->wcl_count; i++)
1177			    witness_displaydescendants(prnt,
1178				wcl->wcl_children[i]);
1179}
1180
1181static int
1182dup_ok(struct witness *w)
1183{
1184	const char **dup;
1185
1186	for (dup = dup_list; *dup != NULL; dup++)
1187		if (strcmp(w->w_name, *dup) == 0)
1188			return (1);
1189	return (0);
1190}
1191
1192static int
1193blessed(struct witness *w1, struct witness *w2)
1194{
1195	int i;
1196	struct witness_blessed *b;
1197
1198	for (i = 0; i < blessed_count; i++) {
1199		b = &blessed_list[i];
1200		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1201			if (strcmp(w2->w_name, b->b_lock2) == 0)
1202				return (1);
1203			continue;
1204		}
1205		if (strcmp(w1->w_name, b->b_lock2) == 0)
1206			if (strcmp(w2->w_name, b->b_lock1) == 0)
1207				return (1);
1208	}
1209	return (0);
1210}
1211
1212static struct witness *
1213witness_get(void)
1214{
1215	struct witness *w;
1216
1217	if (witness_dead) {
1218		mtx_unlock_spin(&w_mtx);
1219		return (NULL);
1220	}
1221	if (STAILQ_EMPTY(&w_free)) {
1222		witness_dead = 1;
1223		mtx_unlock_spin(&w_mtx);
1224		printf("%s: witness exhausted\n", __func__);
1225		return (NULL);
1226	}
1227	w = STAILQ_FIRST(&w_free);
1228	STAILQ_REMOVE_HEAD(&w_free, w_list);
1229	bzero(w, sizeof(*w));
1230	return (w);
1231}
1232
1233static void
1234witness_free(struct witness *w)
1235{
1236
1237	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1238}
1239
1240static struct witness_child_list_entry *
1241witness_child_get(void)
1242{
1243	struct witness_child_list_entry *wcl;
1244
1245	if (witness_dead) {
1246		mtx_unlock_spin(&w_mtx);
1247		return (NULL);
1248	}
1249	wcl = w_child_free;
1250	if (wcl == NULL) {
1251		witness_dead = 1;
1252		mtx_unlock_spin(&w_mtx);
1253		printf("%s: witness exhausted\n", __func__);
1254		return (NULL);
1255	}
1256	w_child_free = wcl->wcl_next;
1257	bzero(wcl, sizeof(*wcl));
1258	return (wcl);
1259}
1260
1261static void
1262witness_child_free(struct witness_child_list_entry *wcl)
1263{
1264
1265	wcl->wcl_next = w_child_free;
1266	w_child_free = wcl;
1267}
1268
1269static struct lock_list_entry *
1270witness_lock_list_get(void)
1271{
1272	struct lock_list_entry *lle;
1273
1274	if (witness_dead)
1275		return (NULL);
1276	mtx_lock_spin(&w_mtx);
1277	lle = w_lock_list_free;
1278	if (lle == NULL) {
1279		witness_dead = 1;
1280		mtx_unlock_spin(&w_mtx);
1281		printf("%s: witness exhausted\n", __func__);
1282		return (NULL);
1283	}
1284	w_lock_list_free = lle->ll_next;
1285	mtx_unlock_spin(&w_mtx);
1286	bzero(lle, sizeof(*lle));
1287	return (lle);
1288}
1289
1290static void
1291witness_lock_list_free(struct lock_list_entry *lle)
1292{
1293
1294	mtx_lock_spin(&w_mtx);
1295	lle->ll_next = w_lock_list_free;
1296	w_lock_list_free = lle;
1297	mtx_unlock_spin(&w_mtx);
1298}
1299
1300static struct lock_instance *
1301find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1302{
1303	struct lock_list_entry *lle;
1304	struct lock_instance *instance;
1305	int i;
1306
1307	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1308		for (i = lle->ll_count - 1; i >= 0; i--) {
1309			instance = &lle->ll_children[i];
1310			if (instance->li_lock == lock)
1311				return (instance);
1312		}
1313	return (NULL);
1314}
1315
1316int
1317witness_list_locks(struct lock_list_entry **lock_list)
1318{
1319	struct lock_list_entry *lle;
1320	struct lock_instance *instance;
1321	struct lock_object *lock;
1322	int i, nheld;
1323
1324	nheld = 0;
1325	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1326		for (i = lle->ll_count - 1; i >= 0; i--) {
1327			instance = &lle->ll_children[i];
1328			lock = instance->li_lock;
1329			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1330			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1331			    "exclusive" : "shared",
1332			    lock->lo_class->lc_name, lock->lo_name, lock,
1333			    instance->li_file, instance->li_line);
1334			nheld++;
1335		}
1336	return (nheld);
1337}
1338
1339/*
1340 * Calling this on p != curproc is bad unless we are in ddb.
1341 */
1342int
1343witness_list(struct proc *p)
1344{
1345	critical_t savecrit;
1346	int nheld;
1347
1348	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1349#ifdef DDB
1350	KASSERT(p == curproc || db_active,
1351	    ("%s: p != curproc and we aren't in the debugger", __func__));
1352
1353	if (!db_active && witness_dead)
1354		return (0);
1355#else
1356	KASSERT(p == curproc, ("%s: p != curproc", __func__));
1357
1358	if (witness_dead)
1359		return (0);
1360#endif
1361	nheld = witness_list_locks(&p->p_sleeplocks);
1362
1363	/*
1364	 * We only handle spinlocks if p == curproc.  This is somewhat broken
1365	 * if p is currently executing on some other CPU and holds spin locks
1366	 * as we won't display those locks.  If we had a MI way of getting
1367	 * the per-cpu data for a given cpu then we could use p->p_oncpu to
1368	 * get the list of spinlocks for this process and "fix" this.
1369	 */
1370	if (p == curproc) {
1371		/*
1372		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1373		 * change.
1374		 */
1375		savecrit = critical_enter();
1376		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1377		critical_exit(savecrit);
1378	}
1379
1380	return (nheld);
1381}
1382
1383void
1384witness_save(struct lock_object *lock, const char **filep, int *linep)
1385{
1386	struct lock_instance *instance;
1387
1388	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
1389	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1390		return;
1391
1392	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1393		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1394		    lock->lo_class->lc_name, lock->lo_name);
1395	instance = find_instance(curproc->p_sleeplocks, lock);
1396	if (instance == NULL)
1397		panic("%s: lock (%s) %s not locked", __func__,
1398		    lock->lo_class->lc_name, lock->lo_name);
1399
1400	*filep = instance->li_file;
1401	*linep = instance->li_line;
1402}
1403
1404void
1405witness_restore(struct lock_object *lock, const char *file, int line)
1406{
1407	struct lock_instance *instance;
1408
1409	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
1410	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1411		return;
1412
1413	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1414		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1415		    lock->lo_class->lc_name, lock->lo_name);
1416	instance = find_instance(curproc->p_sleeplocks, lock);
1417	if (instance == NULL)
1418		panic("%s: lock (%s) %s not locked", __func__,
1419		    lock->lo_class->lc_name, lock->lo_name);
1420
1421	lock->lo_witness->w_file = file;
1422	lock->lo_witness->w_line = line;
1423	instance->li_file = file;
1424	instance->li_line = line;
1425}
1426
1427void
1428witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1429{
1430#ifdef INVARIANT_SUPPORT
1431	struct lock_instance *instance;
1432
1433	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1434		return;
1435
1436	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1437		instance = find_instance(curproc->p_sleeplocks, lock);
1438	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1439		instance = find_instance(PCPU_GET(spinlocks), lock);
1440	else
1441		panic("Lock (%s) %s is not sleep or spin!",
1442		    lock->lo_class->lc_name, lock->lo_name);
1443	switch (flags) {
1444	case LA_UNLOCKED:
1445		if (instance != NULL)
1446			panic("Lock (%s) %s locked @ %s:%d.",
1447			    lock->lo_class->lc_name, lock->lo_name, file, line);
1448		break;
1449	case LA_LOCKED:
1450	case LA_LOCKED | LA_RECURSED:
1451	case LA_LOCKED | LA_NOTRECURSED:
1452	case LA_SLOCKED:
1453	case LA_SLOCKED | LA_RECURSED:
1454	case LA_SLOCKED | LA_NOTRECURSED:
1455	case LA_XLOCKED:
1456	case LA_XLOCKED | LA_RECURSED:
1457	case LA_XLOCKED | LA_NOTRECURSED:
1458		if (instance == NULL)
1459			panic("Lock (%s) %s not locked @ %s:%d.",
1460			    lock->lo_class->lc_name, lock->lo_name, file, line);
1461		if ((flags & LA_XLOCKED) != 0 &&
1462		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1463			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1464			    lock->lo_class->lc_name, lock->lo_name, file, line);
1465		if ((flags & LA_SLOCKED) != 0 &&
1466		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1467			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1468			    lock->lo_class->lc_name, lock->lo_name, file, line);
1469		if ((flags & LA_RECURSED) != 0 &&
1470		    (instance->li_flags & LI_RECURSEMASK) == 0)
1471			panic("Lock (%s) %s not recursed @ %s:%d.",
1472			    lock->lo_class->lc_name, lock->lo_name, file, line);
1473		if ((flags & LA_NOTRECURSED) != 0 &&
1474		    (instance->li_flags & LI_RECURSEMASK) != 0)
1475			panic("Lock (%s) %s recursed @ %s:%d.",
1476			    lock->lo_class->lc_name, lock->lo_name, file, line);
1477		break;
1478	default:
1479		panic("Invalid lock assertion at %s:%d.", file, line);
1480
1481	}
1482#endif	/* INVARIANT_SUPPORT */
1483}
1484
1485#ifdef DDB
1486
1487DB_SHOW_COMMAND(locks, db_witness_list)
1488{
1489	struct proc *p;
1490	pid_t pid;
1491
1492	if (have_addr) {
1493		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1494		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1495		    ((addr >> 16) % 16) * 10000;
1496
1497		/* sx_slock(&allproc_lock); */
1498		LIST_FOREACH(p, &allproc, p_list) {
1499			if (p->p_pid == pid)
1500				break;
1501		}
1502		/* sx_sunlock(&allproc_lock); */
1503		if (p == NULL) {
1504			db_printf("pid %d not found\n", pid);
1505			return;
1506		}
1507	} else
1508		p = curproc;
1509
1510	witness_list(p);
1511}
1512
1513DB_SHOW_COMMAND(witness, db_witness_display)
1514{
1515
1516	witness_display(db_printf);
1517}
1518#endif
1519