subr_witness.c revision 99072
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 99072 2002-06-29 17:26:22Z julian $
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 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.
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	blessed(struct witness *, struct witness *);
128static void	witness_display_list(void(*prnt)(const char *fmt, ...),
129				     struct witness_list *list);
130static void	witness_displaydescendants(void(*)(const char *fmt, ...),
131					   struct witness *);
132static void	witness_leveldescendents(struct witness *parent, int level);
133static void	witness_levelall(void);
134static struct	witness *witness_get(void);
135static void	witness_free(struct witness *m);
136static struct	witness_child_list_entry *witness_child_get(void);
137static void	witness_child_free(struct witness_child_list_entry *wcl);
138static struct	lock_list_entry *witness_lock_list_get(void);
139static void	witness_lock_list_free(struct lock_list_entry *lle);
140static void	witness_display(void(*)(const char *fmt, ...));
141static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
142					     struct lock_object *lock);
143
144MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
145
146static int witness_watch = 1;
147TUNABLE_INT("debug.witness_watch", &witness_watch);
148SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
149
150#ifdef DDB
151/*
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.
156 */
157#ifdef WITNESS_DDB
158int	witness_ddb = 1;
159#else
160int	witness_ddb = 0;
161#endif
162TUNABLE_INT("debug.witness_ddb", &witness_ddb);
163SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
164#endif /* DDB */
165
166#ifdef WITNESS_SKIPSPIN
167int	witness_skipspin = 1;
168#else
169int	witness_skipspin = 0;
170#endif
171TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
172SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
173    "");
174
175static struct mtx w_mtx;
176static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
177static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
178static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
179static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
180static struct witness_child_list_entry *w_child_free = NULL;
181static struct lock_list_entry *w_lock_list_free = NULL;
182static int witness_dead;	/* fatal error, probably no memory */
183
184static struct witness w_data[WITNESS_COUNT];
185static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
186static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
187
188static struct witness_order_list_entry order_lists[] = {
189	{ "Giant", &lock_class_mtx_sleep },
190	{ "proctree", &lock_class_sx },
191	{ "allproc", &lock_class_sx },
192	{ "sigio lock", &lock_class_mtx_sleep },
193	{ "process group", &lock_class_mtx_sleep },
194	{ "process lock", &lock_class_mtx_sleep },
195	{ "session", &lock_class_mtx_sleep },
196	{ "uidinfo hash", &lock_class_mtx_sleep },
197	{ "uidinfo struct", &lock_class_mtx_sleep },
198	{ NULL, NULL },
199	/*
200	 * spin locks
201	 */
202#ifdef SMP
203	{ "ap boot", &lock_class_mtx_spin },
204#ifdef __i386__
205	{ "com", &lock_class_mtx_spin },
206#endif
207#endif
208	{ "sio", &lock_class_mtx_spin },
209#ifdef __i386__
210	{ "cy", &lock_class_mtx_spin },
211#endif
212	{ "ng_node", &lock_class_mtx_spin },
213	{ "ng_worklist", &lock_class_mtx_spin },
214	{ "ithread table lock", &lock_class_mtx_spin },
215	{ "sched lock", &lock_class_mtx_spin },
216	{ "callout", &lock_class_mtx_spin },
217	/*
218	 * leaf locks
219	 */
220	{ "allpmaps", &lock_class_mtx_spin },
221	{ "vm page buckets mutex", &lock_class_mtx_spin },
222	{ "icu", &lock_class_mtx_spin },
223#ifdef SMP
224	{ "smp rendezvous", &lock_class_mtx_spin },
225#endif
226	{ "clk", &lock_class_mtx_spin },
227	{ "mutex profiling lock", &lock_class_mtx_spin },
228	{ "zombie_thread_lock", &lock_class_mtx_spin },
229	{ NULL, NULL },
230	{ NULL, NULL }
231};
232
233/*
234 * Pairs of locks which have been blessed
235 * Don't complain about order problems with blessed locks
236 */
237static struct witness_blessed blessed_list[] = {
238};
239static int blessed_count =
240	sizeof(blessed_list) / sizeof(struct witness_blessed);
241
242/*
243 * List of all locks in the system.
244 */
245TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
246
247static struct mtx all_mtx = {
248	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
249	  "All locks list",		/* mtx_object.lo_name */
250	  "All locks list",		/* mtx_object.lo_type */
251	  LO_INITIALIZED,		/* mtx_object.lo_flags */
252	  { NULL, NULL },		/* mtx_object.lo_list */
253	  NULL },			/* mtx_object.lo_witness */
254	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
255	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
256	{ NULL, NULL }			/* mtx_contested */
257};
258
259/*
260 * This global is set to 0 once it becomes safe to use the witness code.
261 */
262static int witness_cold = 1;
263
264/*
265 * Global variables for book keeping.
266 */
267static int lock_cur_cnt;
268static int lock_max_cnt;
269
270/*
271 * The WITNESS-enabled diagnostic code.
272 */
273static void
274witness_initialize(void *dummy __unused)
275{
276	struct lock_object *lock;
277	struct witness_order_list_entry *order;
278	struct witness *w, *w1;
279	int i;
280
281	/*
282	 * We have to release Giant before initializing its witness
283	 * structure so that WITNESS doesn't get confused.
284	 */
285	mtx_unlock(&Giant);
286	mtx_assert(&Giant, MA_NOTOWNED);
287
288	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
289	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
290	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
291	    MTX_NOWITNESS);
292	for (i = 0; i < WITNESS_COUNT; i++)
293		witness_free(&w_data[i]);
294	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
295		witness_child_free(&w_childdata[i]);
296	for (i = 0; i < LOCK_CHILDCOUNT; i++)
297		witness_lock_list_free(&w_locklistdata[i]);
298
299	/* First add in all the specified order lists. */
300	for (order = order_lists; order->w_name != NULL; order++) {
301		w = enroll(order->w_name, order->w_class);
302		if (w == NULL)
303			continue;
304		w->w_file = "order list";
305		for (order++; order->w_name != NULL; order++) {
306			w1 = enroll(order->w_name, order->w_class);
307			if (w1 == NULL)
308				continue;
309			w1->w_file = "order list";
310			itismychild(w, w1);
311			w = w1;
312		}
313	}
314
315	/* Iterate through all locks and add them to witness. */
316	mtx_lock(&all_mtx);
317	TAILQ_FOREACH(lock, &all_locks, lo_list) {
318		if (lock->lo_flags & LO_WITNESS)
319			lock->lo_witness = enroll(lock->lo_type,
320			    lock->lo_class);
321		else
322			lock->lo_witness = NULL;
323	}
324	mtx_unlock(&all_mtx);
325
326	/* Mark the witness code as being ready for use. */
327	atomic_store_rel_int(&witness_cold, 0);
328
329	mtx_lock(&Giant);
330}
331SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
332
333void
334witness_init(struct lock_object *lock)
335{
336	struct lock_class *class;
337
338	class = lock->lo_class;
339	if (lock->lo_flags & LO_INITIALIZED)
340		panic("%s: lock (%s) %s is already initialized", __func__,
341		    class->lc_name, lock->lo_name);
342	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
343	    (class->lc_flags & LC_RECURSABLE) == 0)
344		panic("%s: lock (%s) %s can not be recursable", __func__,
345		    class->lc_name, lock->lo_name);
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", __func__,
349		    class->lc_name, lock->lo_name);
350	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
351	    (class->lc_flags & LC_UPGRADABLE) == 0)
352		panic("%s: lock (%s) %s can not be upgradable", __func__,
353		    class->lc_name, lock->lo_name);
354
355	mtx_lock(&all_mtx);
356	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
357	lock->lo_flags |= LO_INITIALIZED;
358	lock_cur_cnt++;
359	if (lock_cur_cnt > lock_max_cnt)
360		lock_max_cnt = lock_cur_cnt;
361	mtx_unlock(&all_mtx);
362	if (!witness_cold && !witness_dead && panicstr == NULL &&
363	    (lock->lo_flags & LO_WITNESS) != 0)
364		lock->lo_witness = enroll(lock->lo_type, class);
365	else
366		lock->lo_witness = NULL;
367}
368
369void
370witness_destroy(struct lock_object *lock)
371{
372	struct witness *w;
373
374	if (witness_cold)
375		panic("lock (%s) %s destroyed while witness_cold",
376		    lock->lo_class->lc_name, lock->lo_name);
377	if ((lock->lo_flags & LO_INITIALIZED) == 0)
378		panic("%s: lock (%s) %s is not initialized", __func__,
379		    lock->lo_class->lc_name, lock->lo_name);
380
381	/* XXX: need to verify that no one holds the lock */
382	w = lock->lo_witness;
383	if (w != NULL) {
384		mtx_lock_spin(&w_mtx);
385		MPASS(w->w_refcount > 0);
386		w->w_refcount--;
387		mtx_unlock_spin(&w_mtx);
388	}
389
390	mtx_lock(&all_mtx);
391	lock_cur_cnt--;
392	TAILQ_REMOVE(&all_locks, lock, lo_list);
393	lock->lo_flags &= ~LO_INITIALIZED;
394	mtx_unlock(&all_mtx);
395}
396
397static void
398witness_display_list(void(*prnt)(const char *fmt, ...),
399		     struct witness_list *list)
400{
401	struct witness *w, *w1;
402	int found;
403
404	STAILQ_FOREACH(w, list, w_typelist) {
405		if (w->w_file == NULL)
406			continue;
407		found = 0;
408		STAILQ_FOREACH(w1, list, w_typelist) {
409			if (isitmychild(w1, w)) {
410				found++;
411				break;
412			}
413		}
414		if (found)
415			continue;
416		/*
417		 * This lock has no anscestors, display its descendants.
418		 */
419		witness_displaydescendants(prnt, w);
420	}
421}
422
423static void
424witness_display(void(*prnt)(const char *fmt, ...))
425{
426	struct witness *w;
427
428	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
429	witness_levelall();
430
431	/*
432	 * First, handle sleep locks which have been acquired at least
433	 * once.
434	 */
435	prnt("Sleep locks:\n");
436	witness_display_list(prnt, &w_sleep);
437
438	/*
439	 * Now do spin locks which have been acquired at least once.
440	 */
441	prnt("\nSpin locks:\n");
442	witness_display_list(prnt, &w_spin);
443
444	/*
445	 * Finally, any locks which have not been acquired yet.
446	 */
447	prnt("\nLocks which were never acquired:\n");
448	STAILQ_FOREACH(w, &w_all, w_list) {
449		if (w->w_file != NULL || w->w_refcount == 0)
450			continue;
451		prnt("%s\n", w->w_name);
452	}
453}
454
455void
456witness_lock(struct lock_object *lock, int flags, const char *file, int line)
457{
458	struct lock_list_entry **lock_list, *lle;
459	struct lock_instance *lock1, *lock2;
460	struct lock_class *class;
461	struct witness *w, *w1;
462	struct thread *td;
463	int i, j;
464#ifdef DDB
465	int go_into_ddb = 0;
466#endif /* DDB */
467
468	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
469	    panicstr != NULL)
470		return;
471	w = lock->lo_witness;
472	class = lock->lo_class;
473	td = curthread;
474
475	if (class->lc_flags & LC_SLEEPLOCK) {
476		/*
477		 * Since spin locks include a critical section, this check
478		 * impliclty enforces a lock order of all sleep locks before
479		 * all spin locks.
480		 */
481		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
482			panic("blockable sleep lock (%s) %s @ %s:%d",
483			    class->lc_name, lock->lo_name, file, line);
484		lock_list = &td->td_sleeplocks;
485	} else
486		lock_list = PCPU_PTR(spinlocks);
487
488	/*
489	 * Try locks do not block if they fail to acquire the lock, thus
490	 * there is no danger of deadlocks or of switching while holding a
491	 * spin lock if we acquire a lock via a try operation.
492	 */
493	if (flags & LOP_TRYLOCK)
494		goto out;
495
496	/*
497	 * Is this the first lock acquired?  If so, then no order checking
498	 * is needed.
499	 */
500	if (*lock_list == NULL)
501		goto out;
502
503	/*
504	 * Check to see if we are recursing on a lock we already own.
505	 */
506	lock1 = find_instance(*lock_list, lock);
507	if (lock1 != NULL) {
508		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
509		    (flags & LOP_EXCLUSIVE) == 0) {
510			printf("shared lock of (%s) %s @ %s:%d\n",
511			    class->lc_name, lock->lo_name, file, line);
512			printf("while exclusively locked from %s:%d\n",
513			    lock1->li_file, lock1->li_line);
514			panic("share->excl");
515		}
516		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
517		    (flags & LOP_EXCLUSIVE) != 0) {
518			printf("exclusive lock of (%s) %s @ %s:%d\n",
519			    class->lc_name, lock->lo_name, file, line);
520			printf("while share locked from %s:%d\n",
521			    lock1->li_file, lock1->li_line);
522			panic("excl->share");
523		}
524		lock1->li_flags++;
525		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
526			printf(
527			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
528			    class->lc_name, lock->lo_name, file, line);
529			printf("first acquired @ %s:%d\n", lock1->li_file,
530			    lock1->li_line);
531			panic("recurse");
532		}
533		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
534		    td->td_proc->p_pid, lock->lo_name,
535		    lock1->li_flags & LI_RECURSEMASK);
536		lock1->li_file = file;
537		lock1->li_line = line;
538		return;
539	}
540
541	/*
542	 * Check for duplicate locks of the same type.  Note that we only
543	 * have to check for this on the last lock we just acquired.  Any
544	 * other cases will be caught as lock order violations.
545	 */
546	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
547	w1 = lock1->li_lock->lo_witness;
548	if (w1 == w) {
549		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
550			goto out;
551		w->w_same_squawked = 1;
552		printf("acquiring duplicate lock of same type: \"%s\"\n",
553			lock->lo_type);
554		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
555		    lock1->li_file, lock1->li_line);
556		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
557#ifdef DDB
558		go_into_ddb = 1;
559#endif /* DDB */
560		goto out;
561	}
562	MPASS(!mtx_owned(&w_mtx));
563	mtx_lock_spin(&w_mtx);
564	/*
565	 * If we have a known higher number just say ok
566	 */
567	if (witness_watch > 1 && w->w_level > w1->w_level) {
568		mtx_unlock_spin(&w_mtx);
569		goto out;
570	}
571	if (isitmydescendant(w1, w)) {
572		mtx_unlock_spin(&w_mtx);
573		goto out;
574	}
575	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
576		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
577
578			MPASS(j < WITNESS_COUNT);
579			lock1 = &lle->ll_children[i];
580			w1 = lock1->li_lock->lo_witness;
581
582			/*
583			 * If this lock doesn't undergo witness checking,
584			 * then skip it.
585			 */
586			if (w1 == NULL) {
587				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
588				    ("lock missing witness structure"));
589				continue;
590			}
591			/*
592			 * If we are locking Giant and we slept with this
593			 * lock, then skip it.
594			 */
595			if ((lock1->li_flags & LI_SLEPT) != 0 &&
596			    lock == &Giant.mtx_object)
597				continue;
598			/*
599			 * If we are locking a sleepable lock and this lock
600			 * isn't sleepable and isn't Giant, we want to treat
601			 * it as a lock order violation to enfore a general
602			 * lock order of sleepable locks before non-sleepable
603			 * locks.  Thus, we only bother checking the lock
604			 * order hierarchy if we pass the initial test.
605			 */
606			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
607			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
608			    lock1->li_lock != &Giant.mtx_object)) &&
609			    !isitmydescendant(w, w1))
610				continue;
611			/*
612			 * We have a lock order violation, check to see if it
613			 * is allowed or has already been yelled about.
614			 */
615			mtx_unlock_spin(&w_mtx);
616			if (blessed(w, w1))
617				goto out;
618			if (lock1->li_lock == &Giant.mtx_object) {
619				if (w1->w_Giant_squawked)
620					goto out;
621				else
622					w1->w_Giant_squawked = 1;
623			} else {
624				if (w1->w_other_squawked)
625					goto out;
626				else
627					w1->w_other_squawked = 1;
628			}
629			/*
630			 * Ok, yell about it.
631			 */
632			printf("lock order reversal\n");
633			/*
634			 * Try to locate an earlier lock with
635			 * witness w in our list.
636			 */
637			do {
638				lock2 = &lle->ll_children[i];
639				MPASS(lock2->li_lock != NULL);
640				if (lock2->li_lock->lo_witness == w)
641					break;
642				i--;
643				if (i == 0 && lle->ll_next != NULL) {
644					lle = lle->ll_next;
645					i = lle->ll_count - 1;
646					MPASS(i != 0);
647				}
648			} while (i >= 0);
649			if (i < 0) {
650				printf(" 1st %p %s (%s) @ %s:%d\n",
651				    lock1->li_lock, lock1->li_lock->lo_name,
652				    lock1->li_lock->lo_type, lock1->li_file,
653				    lock1->li_line);
654				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
655				    lock->lo_name, lock->lo_type, file, line);
656			} else {
657				printf(" 1st %p %s (%s) @ %s:%d\n",
658				    lock2->li_lock, lock2->li_lock->lo_name,
659				    lock2->li_lock->lo_type, lock2->li_file,
660				    lock2->li_line);
661				printf(" 2nd %p %s (%s) @ %s:%d\n",
662				    lock1->li_lock, lock1->li_lock->lo_name,
663				    lock1->li_lock->lo_type, lock1->li_file,
664				    lock1->li_line);
665				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
666				    lock->lo_name, lock->lo_type, file, line);
667			}
668#ifdef DDB
669			go_into_ddb = 1;
670#endif /* DDB */
671			goto out;
672		}
673	}
674	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
675	/*
676	 * Don't build a new relationship if we are locking Giant just
677	 * after waking up and the previous lock in the list was acquired
678	 * prior to blocking.
679	 */
680	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
681		mtx_unlock_spin(&w_mtx);
682	else {
683		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
684		    lock->lo_type, lock1->li_lock->lo_type);
685		if (!itismychild(lock1->li_lock->lo_witness, w))
686			mtx_unlock_spin(&w_mtx);
687	}
688
689out:
690#ifdef DDB
691	if (witness_ddb && go_into_ddb)
692		Debugger(__func__);
693#endif /* DDB */
694	w->w_file = file;
695	w->w_line = line;
696
697	lle = *lock_list;
698	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
699		lle = witness_lock_list_get();
700		if (lle == NULL)
701			return;
702		lle->ll_next = *lock_list;
703		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
704		    td->td_proc->p_pid, lle);
705		*lock_list = lle;
706	}
707	lock1 = &lle->ll_children[lle->ll_count++];
708	lock1->li_lock = lock;
709	lock1->li_line = line;
710	lock1->li_file = file;
711	if ((flags & LOP_EXCLUSIVE) != 0)
712		lock1->li_flags = LI_EXCLUSIVE;
713	else
714		lock1->li_flags = 0;
715	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
716	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
717}
718
719void
720witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
721{
722	struct lock_instance *instance;
723	struct lock_class *class;
724
725	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
726	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
727		return;
728	class = lock->lo_class;
729	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
730		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
731		    class->lc_name, lock->lo_name, file, line);
732	if ((flags & LOP_TRYLOCK) == 0)
733		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
734		    lock->lo_name, file, line);
735	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
736		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
737		    class->lc_name, lock->lo_name, file, line);
738	instance = find_instance(curthread->td_sleeplocks, lock);
739	if (instance == NULL)
740		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
741		    class->lc_name, lock->lo_name, file, line);
742	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
743		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
744		    class->lc_name, lock->lo_name, file, line);
745	if ((instance->li_flags & LI_RECURSEMASK) != 0)
746		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
747		    class->lc_name, lock->lo_name,
748		    instance->li_flags & LI_RECURSEMASK, file, line);
749	instance->li_flags |= LI_EXCLUSIVE;
750}
751
752void
753witness_downgrade(struct lock_object *lock, int flags, const char *file,
754    int line)
755{
756	struct lock_instance *instance;
757	struct lock_class *class;
758
759	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
760	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
761		return;
762	class = lock->lo_class;
763	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
764		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
765		    class->lc_name, lock->lo_name, file, line);
766	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
767		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
768		    class->lc_name, lock->lo_name, file, line);
769	instance = find_instance(curthread->td_sleeplocks, lock);
770	if (instance == NULL)
771		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
772		    class->lc_name, lock->lo_name, file, line);
773	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
774		panic("downgrade of shared lock (%s) %s @ %s:%d",
775		    class->lc_name, lock->lo_name, file, line);
776	if ((instance->li_flags & LI_RECURSEMASK) != 0)
777		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
778		    class->lc_name, lock->lo_name,
779		    instance->li_flags & LI_RECURSEMASK, file, line);
780	instance->li_flags &= ~LI_EXCLUSIVE;
781}
782
783void
784witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
785{
786	struct lock_list_entry **lock_list, *lle;
787	struct lock_instance *instance;
788	struct lock_class *class;
789	struct thread *td;
790	register_t s;
791	int i, j;
792
793	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
794	    panicstr != NULL)
795		return;
796	td = curthread;
797	class = lock->lo_class;
798	if (class->lc_flags & LC_SLEEPLOCK)
799		lock_list = &td->td_sleeplocks;
800	else
801		lock_list = PCPU_PTR(spinlocks);
802	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
803		for (i = 0; i < (*lock_list)->ll_count; i++) {
804			instance = &(*lock_list)->ll_children[i];
805			if (instance->li_lock == lock) {
806				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
807				    (flags & LOP_EXCLUSIVE) == 0) {
808					printf(
809					"shared unlock of (%s) %s @ %s:%d\n",
810					    class->lc_name, lock->lo_name,
811					    file, line);
812					printf(
813					"while exclusively locked from %s:%d\n",
814					    instance->li_file,
815					    instance->li_line);
816					panic("excl->ushare");
817				}
818				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
819				    (flags & LOP_EXCLUSIVE) != 0) {
820					printf(
821					"exclusive unlock of (%s) %s @ %s:%d\n",
822					    class->lc_name, lock->lo_name,
823					    file, line);
824					printf(
825					"while share locked from %s:%d\n",
826					    instance->li_file,
827					    instance->li_line);
828					panic("share->uexcl");
829				}
830				/* If we are recursed, unrecurse. */
831				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
832					CTR4(KTR_WITNESS,
833				    "%s: pid %d unrecursed on %s r=%d", __func__,
834					    td->td_proc->p_pid,
835					    instance->li_lock->lo_name,
836					    instance->li_flags);
837					instance->li_flags--;
838					return;
839				}
840				s = intr_disable();
841				CTR4(KTR_WITNESS,
842				    "%s: pid %d removed %s from lle[%d]", __func__,
843				    td->td_proc->p_pid,
844				    instance->li_lock->lo_name,
845				    (*lock_list)->ll_count - 1);
846				for (j = i; j < (*lock_list)->ll_count - 1; j++)
847					(*lock_list)->ll_children[j] =
848					    (*lock_list)->ll_children[j + 1];
849				(*lock_list)->ll_count--;
850				intr_restore(s);
851				if ((*lock_list)->ll_count == 0) {
852					lle = *lock_list;
853					*lock_list = lle->ll_next;
854					CTR3(KTR_WITNESS,
855					    "%s: pid %d removed lle %p", __func__,
856					    td->td_proc->p_pid, lle);
857					witness_lock_list_free(lle);
858				}
859				return;
860			}
861		}
862	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
863	    file, line);
864}
865
866/*
867 * Warn if any held locks are not sleepable.  Note that Giant and the lock
868 * passed in are both special cases since they are both released during the
869 * sleep process and aren't actually held while the thread is asleep.
870 */
871int
872witness_sleep(int check_only, struct lock_object *lock, const char *file,
873	      int line)
874{
875	struct lock_list_entry **lock_list, *lle;
876	struct lock_instance *lock1;
877	struct thread *td;
878	int i, n;
879
880	if (witness_cold || witness_dead || panicstr != NULL)
881		return (0);
882	n = 0;
883	td = curthread;
884	lock_list = &td->td_sleeplocks;
885again:
886	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
887		for (i = lle->ll_count - 1; i >= 0; i--) {
888			lock1 = &lle->ll_children[i];
889			if (lock1->li_lock == lock ||
890			    lock1->li_lock == &Giant.mtx_object)
891				continue;
892			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
893				if (check_only == 0) {
894					CTR3(KTR_WITNESS,
895				    "pid %d: sleeping with lock (%s) %s held",
896					    td->td_proc->p_pid,
897					    lock1->li_lock->lo_class->lc_name,
898					    lock1->li_lock->lo_name);
899					lock1->li_flags |= LI_SLEPT;
900				}
901				continue;
902			}
903			n++;
904			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
905			    file, line, check_only ? "could sleep" : "sleeping",
906			    lock1->li_lock->lo_name, lock1->li_file,
907			    lock1->li_line);
908		}
909	if (lock_list == &td->td_sleeplocks && PCPU_GET(spinlocks) != NULL) {
910		/*
911		 * Since we already hold a spinlock preemption is
912		 * already blocked.
913		 */
914		lock_list = PCPU_PTR(spinlocks);
915		goto again;
916	}
917#ifdef DDB
918	if (witness_ddb && n)
919		Debugger(__func__);
920#endif /* DDB */
921	return (n);
922}
923
924static struct witness *
925enroll(const char *description, struct lock_class *lock_class)
926{
927	struct witness *w;
928
929	if (!witness_watch || witness_dead || panicstr != NULL)
930		return (NULL);
931	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
932		return (NULL);
933	mtx_lock_spin(&w_mtx);
934	STAILQ_FOREACH(w, &w_all, w_list) {
935		if (w->w_name == description || (w->w_refcount > 0 &&
936		    strcmp(description, w->w_name) == 0)) {
937			w->w_refcount++;
938			mtx_unlock_spin(&w_mtx);
939			if (lock_class != w->w_class)
940				panic(
941				"lock (%s) %s does not match earlier (%s) lock",
942				    description, lock_class->lc_name,
943				    w->w_class->lc_name);
944			return (w);
945		}
946	}
947	/*
948	 * This isn't quite right, as witness_cold is still 0 while we
949	 * enroll all the locks initialized before witness_initialize().
950	 */
951	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
952		mtx_unlock_spin(&w_mtx);
953		panic("spin lock %s not in order list", description);
954	}
955	if ((w = witness_get()) == NULL)
956		return (NULL);
957	w->w_name = description;
958	w->w_class = lock_class;
959	w->w_refcount = 1;
960	STAILQ_INSERT_HEAD(&w_all, w, w_list);
961	if (lock_class->lc_flags & LC_SPINLOCK)
962		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
963	else if (lock_class->lc_flags & LC_SLEEPLOCK)
964		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
965	else {
966		mtx_unlock_spin(&w_mtx);
967		panic("lock class %s is not sleep or spin",
968		    lock_class->lc_name);
969	}
970	mtx_unlock_spin(&w_mtx);
971	return (w);
972}
973
974static int
975itismychild(struct witness *parent, struct witness *child)
976{
977	static int recursed;
978	struct witness_child_list_entry **wcl;
979	struct witness_list *list;
980
981	MPASS(child != NULL && parent != NULL);
982	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
983	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
984		panic(
985		"%s: parent (%s) and child (%s) are not the same lock type",
986		    __func__, parent->w_class->lc_name,
987		    child->w_class->lc_name);
988
989	/*
990	 * Insert "child" after "parent"
991	 */
992	wcl = &parent->w_children;
993	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
994		wcl = &(*wcl)->wcl_next;
995	if (*wcl == NULL) {
996		*wcl = witness_child_get();
997		if (*wcl == NULL)
998			return (1);
999	}
1000	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1001
1002	/*
1003	 * Now prune whole tree.  We look for cases where a lock is now
1004	 * both a descendant and a direct child of a given lock.  In that
1005	 * case, we want to remove the direct child link from the tree.
1006	 */
1007	if (recursed)
1008		return (0);
1009	recursed = 1;
1010	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1011		list = &w_sleep;
1012	else
1013		list = &w_spin;
1014	STAILQ_FOREACH(child, list, w_typelist) {
1015		STAILQ_FOREACH(parent, list, w_typelist) {
1016			if (!isitmychild(parent, child))
1017				continue;
1018			removechild(parent, child);
1019			if (isitmydescendant(parent, child))
1020				continue;
1021			itismychild(parent, child);
1022		}
1023	}
1024	recursed = 0;
1025	witness_levelall();
1026	return (0);
1027}
1028
1029static void
1030removechild(struct witness *parent, struct witness *child)
1031{
1032	struct witness_child_list_entry **wcl, *wcl1;
1033	int i;
1034
1035	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1036		for (i = 0; i < (*wcl)->wcl_count; i++)
1037			if ((*wcl)->wcl_children[i] == child)
1038				goto found;
1039	return;
1040found:
1041	(*wcl)->wcl_count--;
1042	if ((*wcl)->wcl_count > i)
1043		(*wcl)->wcl_children[i] =
1044		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1045	MPASS((*wcl)->wcl_children[i] != NULL);
1046	if ((*wcl)->wcl_count != 0)
1047		return;
1048	wcl1 = *wcl;
1049	*wcl = wcl1->wcl_next;
1050	witness_child_free(wcl1);
1051}
1052
1053static int
1054isitmychild(struct witness *parent, struct witness *child)
1055{
1056	struct witness_child_list_entry *wcl;
1057	int i;
1058
1059	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1060		for (i = 0; i < wcl->wcl_count; i++) {
1061			if (wcl->wcl_children[i] == child)
1062				return (1);
1063		}
1064	}
1065	return (0);
1066}
1067
1068static int
1069isitmydescendant(struct witness *parent, struct witness *child)
1070{
1071	struct witness_child_list_entry *wcl;
1072	int i, j;
1073
1074	if (isitmychild(parent, child))
1075		return (1);
1076	j = 0;
1077	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1078		MPASS(j < 1000);
1079		for (i = 0; i < wcl->wcl_count; i++) {
1080			if (isitmydescendant(wcl->wcl_children[i], child))
1081				return (1);
1082		}
1083		j++;
1084	}
1085	return (0);
1086}
1087
1088void
1089witness_levelall (void)
1090{
1091	struct witness_list *list;
1092	struct witness *w, *w1;
1093
1094	/*
1095	 * First clear all levels.
1096	 */
1097	STAILQ_FOREACH(w, &w_all, w_list) {
1098		w->w_level = 0;
1099	}
1100
1101	/*
1102	 * Look for locks with no parent and level all their descendants.
1103	 */
1104	STAILQ_FOREACH(w, &w_all, w_list) {
1105		/*
1106		 * This is just an optimization, technically we could get
1107		 * away just walking the all list each time.
1108		 */
1109		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1110			list = &w_sleep;
1111		else
1112			list = &w_spin;
1113		STAILQ_FOREACH(w1, list, w_typelist) {
1114			if (isitmychild(w1, w))
1115				goto skip;
1116		}
1117		witness_leveldescendents(w, 0);
1118	skip:
1119		;	/* silence GCC 3.x */
1120	}
1121}
1122
1123static void
1124witness_leveldescendents(struct witness *parent, int level)
1125{
1126	struct witness_child_list_entry *wcl;
1127	int i;
1128
1129	if (parent->w_level < level)
1130		parent->w_level = level;
1131	level++;
1132	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1133		for (i = 0; i < wcl->wcl_count; i++)
1134			witness_leveldescendents(wcl->wcl_children[i], level);
1135}
1136
1137static void
1138witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1139			   struct witness *parent)
1140{
1141	struct witness_child_list_entry *wcl;
1142	int i, level;
1143
1144	level = parent->w_level;
1145	prnt("%-2d", level);
1146	for (i = 0; i < level; i++)
1147		prnt(" ");
1148	if (parent->w_refcount > 0) {
1149		prnt("%s", parent->w_name);
1150		if (parent->w_file != NULL)
1151			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1152			    parent->w_line);
1153	} else
1154		prnt("(dead)\n");
1155	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1156		for (i = 0; i < wcl->wcl_count; i++)
1157			    witness_displaydescendants(prnt,
1158				wcl->wcl_children[i]);
1159}
1160
1161static int
1162blessed(struct witness *w1, struct witness *w2)
1163{
1164	int i;
1165	struct witness_blessed *b;
1166
1167	for (i = 0; i < blessed_count; i++) {
1168		b = &blessed_list[i];
1169		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1170			if (strcmp(w2->w_name, b->b_lock2) == 0)
1171				return (1);
1172			continue;
1173		}
1174		if (strcmp(w1->w_name, b->b_lock2) == 0)
1175			if (strcmp(w2->w_name, b->b_lock1) == 0)
1176				return (1);
1177	}
1178	return (0);
1179}
1180
1181static struct witness *
1182witness_get(void)
1183{
1184	struct witness *w;
1185
1186	if (witness_dead) {
1187		mtx_unlock_spin(&w_mtx);
1188		return (NULL);
1189	}
1190	if (STAILQ_EMPTY(&w_free)) {
1191		witness_dead = 1;
1192		mtx_unlock_spin(&w_mtx);
1193		printf("%s: witness exhausted\n", __func__);
1194		return (NULL);
1195	}
1196	w = STAILQ_FIRST(&w_free);
1197	STAILQ_REMOVE_HEAD(&w_free, w_list);
1198	bzero(w, sizeof(*w));
1199	return (w);
1200}
1201
1202static void
1203witness_free(struct witness *w)
1204{
1205
1206	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1207}
1208
1209static struct witness_child_list_entry *
1210witness_child_get(void)
1211{
1212	struct witness_child_list_entry *wcl;
1213
1214	if (witness_dead) {
1215		mtx_unlock_spin(&w_mtx);
1216		return (NULL);
1217	}
1218	wcl = w_child_free;
1219	if (wcl == NULL) {
1220		witness_dead = 1;
1221		mtx_unlock_spin(&w_mtx);
1222		printf("%s: witness exhausted\n", __func__);
1223		return (NULL);
1224	}
1225	w_child_free = wcl->wcl_next;
1226	bzero(wcl, sizeof(*wcl));
1227	return (wcl);
1228}
1229
1230static void
1231witness_child_free(struct witness_child_list_entry *wcl)
1232{
1233
1234	wcl->wcl_next = w_child_free;
1235	w_child_free = wcl;
1236}
1237
1238static struct lock_list_entry *
1239witness_lock_list_get(void)
1240{
1241	struct lock_list_entry *lle;
1242
1243	if (witness_dead)
1244		return (NULL);
1245	mtx_lock_spin(&w_mtx);
1246	lle = w_lock_list_free;
1247	if (lle == NULL) {
1248		witness_dead = 1;
1249		mtx_unlock_spin(&w_mtx);
1250		printf("%s: witness exhausted\n", __func__);
1251		return (NULL);
1252	}
1253	w_lock_list_free = lle->ll_next;
1254	mtx_unlock_spin(&w_mtx);
1255	bzero(lle, sizeof(*lle));
1256	return (lle);
1257}
1258
1259static void
1260witness_lock_list_free(struct lock_list_entry *lle)
1261{
1262
1263	mtx_lock_spin(&w_mtx);
1264	lle->ll_next = w_lock_list_free;
1265	w_lock_list_free = lle;
1266	mtx_unlock_spin(&w_mtx);
1267}
1268
1269static struct lock_instance *
1270find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1271{
1272	struct lock_list_entry *lle;
1273	struct lock_instance *instance;
1274	int i;
1275
1276	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1277		for (i = lle->ll_count - 1; i >= 0; i--) {
1278			instance = &lle->ll_children[i];
1279			if (instance->li_lock == lock)
1280				return (instance);
1281		}
1282	return (NULL);
1283}
1284
1285int
1286witness_list_locks(struct lock_list_entry **lock_list)
1287{
1288	struct lock_list_entry *lle;
1289	struct lock_instance *instance;
1290	struct lock_object *lock;
1291	int i, nheld;
1292
1293	nheld = 0;
1294	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1295		for (i = lle->ll_count - 1; i >= 0; i--) {
1296			instance = &lle->ll_children[i];
1297			lock = instance->li_lock;
1298			printf("%s %s %s",
1299			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1300			    "exclusive" : "shared",
1301			    lock->lo_class->lc_name, lock->lo_name);
1302			if (lock->lo_type != lock->lo_name)
1303				printf(" (%s)", lock->lo_type);
1304			printf(" r = %d (%p) locked @ %s:%d\n",
1305			    instance->li_flags & LI_RECURSEMASK, lock,
1306			    instance->li_file, instance->li_line);
1307			nheld++;
1308		}
1309	return (nheld);
1310}
1311
1312/*
1313 * Calling this on td != curthread is bad unless we are in ddb.
1314 */
1315int
1316witness_list(struct thread *td)
1317{
1318	int nheld;
1319
1320	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1321#ifdef DDB
1322	KASSERT(td == curthread || db_active,
1323	    ("%s: td != curthread and we aren't in the debugger", __func__));
1324	if (!db_active && witness_dead)
1325		return (0);
1326#else
1327	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1328	if (witness_dead)
1329		return (0);
1330#endif
1331	nheld = witness_list_locks(&td->td_sleeplocks);
1332
1333	/*
1334	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1335	 * if td is currently executing on some other CPU and holds spin locks
1336	 * as we won't display those locks.  If we had a MI way of getting
1337	 * the per-cpu data for a given cpu then we could use
1338	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1339	 * and "fix" this.
1340	 *
1341	 * That still wouldn't really fix this unless we locked sched_lock
1342	 * or stopped the other CPU to make sure it wasn't changing the list
1343	 * out from under us.  It is probably best to just not try to handle
1344	 * threads on other CPU's for now.
1345	 */
1346	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1347		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1348
1349	return (nheld);
1350}
1351
1352void
1353witness_save(struct lock_object *lock, const char **filep, int *linep)
1354{
1355	struct lock_instance *instance;
1356
1357	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1358	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1359		return;
1360	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1361		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1362		    lock->lo_class->lc_name, lock->lo_name);
1363	instance = find_instance(curthread->td_sleeplocks, lock);
1364	if (instance == NULL)
1365		panic("%s: lock (%s) %s not locked", __func__,
1366		    lock->lo_class->lc_name, lock->lo_name);
1367	*filep = instance->li_file;
1368	*linep = instance->li_line;
1369}
1370
1371void
1372witness_restore(struct lock_object *lock, const char *file, int line)
1373{
1374	struct lock_instance *instance;
1375
1376	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1377	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1378		return;
1379	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1380		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1381		    lock->lo_class->lc_name, lock->lo_name);
1382	instance = find_instance(curthread->td_sleeplocks, lock);
1383	if (instance == NULL)
1384		panic("%s: lock (%s) %s not locked", __func__,
1385		    lock->lo_class->lc_name, lock->lo_name);
1386	lock->lo_witness->w_file = file;
1387	lock->lo_witness->w_line = line;
1388	instance->li_file = file;
1389	instance->li_line = line;
1390}
1391
1392void
1393witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1394{
1395#ifdef INVARIANT_SUPPORT
1396	struct lock_instance *instance;
1397
1398	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1399		return;
1400	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1401		instance = find_instance(curthread->td_sleeplocks, lock);
1402	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1403		instance = find_instance(PCPU_GET(spinlocks), lock);
1404	else {
1405		panic("Lock (%s) %s is not sleep or spin!",
1406		    lock->lo_class->lc_name, lock->lo_name);
1407		return;
1408	}
1409	switch (flags) {
1410	case LA_UNLOCKED:
1411		if (instance != NULL)
1412			panic("Lock (%s) %s locked @ %s:%d.",
1413			    lock->lo_class->lc_name, lock->lo_name, file, line);
1414		break;
1415	case LA_LOCKED:
1416	case LA_LOCKED | LA_RECURSED:
1417	case LA_LOCKED | LA_NOTRECURSED:
1418	case LA_SLOCKED:
1419	case LA_SLOCKED | LA_RECURSED:
1420	case LA_SLOCKED | LA_NOTRECURSED:
1421	case LA_XLOCKED:
1422	case LA_XLOCKED | LA_RECURSED:
1423	case LA_XLOCKED | LA_NOTRECURSED:
1424		if (instance == NULL) {
1425			panic("Lock (%s) %s not locked @ %s:%d.",
1426			    lock->lo_class->lc_name, lock->lo_name, file, line);
1427			break;
1428		}
1429		if ((flags & LA_XLOCKED) != 0 &&
1430		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1431			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1432			    lock->lo_class->lc_name, lock->lo_name, file, line);
1433		if ((flags & LA_SLOCKED) != 0 &&
1434		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1435			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1436			    lock->lo_class->lc_name, lock->lo_name, file, line);
1437		if ((flags & LA_RECURSED) != 0 &&
1438		    (instance->li_flags & LI_RECURSEMASK) == 0)
1439			panic("Lock (%s) %s not recursed @ %s:%d.",
1440			    lock->lo_class->lc_name, lock->lo_name, file, line);
1441		if ((flags & LA_NOTRECURSED) != 0 &&
1442		    (instance->li_flags & LI_RECURSEMASK) != 0)
1443			panic("Lock (%s) %s recursed @ %s:%d.",
1444			    lock->lo_class->lc_name, lock->lo_name, file, line);
1445		break;
1446	default:
1447		panic("Invalid lock assertion at %s:%d.", file, line);
1448
1449	}
1450#endif	/* INVARIANT_SUPPORT */
1451}
1452
1453#ifdef DDB
1454
1455DB_SHOW_COMMAND(locks, db_witness_list)
1456{
1457	struct thread *td;
1458	pid_t pid;
1459	struct proc *p;
1460
1461	if (have_addr) {
1462		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1463		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1464		    ((addr >> 16) % 16) * 10000;
1465		/* sx_slock(&allproc_lock); */
1466		FOREACH_PROC_IN_SYSTEM(p) {
1467			if (p->p_pid == pid)
1468				break;
1469		}
1470		/* sx_sunlock(&allproc_lock); */
1471		if (p == NULL) {
1472			db_printf("pid %d not found\n", pid);
1473			return;
1474		}
1475		FOREACH_THREAD_IN_PROC(p, td) {
1476			witness_list(td);
1477		}
1478	} else {
1479		td = curthread;
1480		witness_list(td);
1481	}
1482}
1483
1484DB_SHOW_COMMAND(witness, db_witness_display)
1485{
1486
1487	witness_display(db_printf);
1488}
1489#endif
1490