subr_witness.c revision 88088
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 88088 2001-12-18 00:27:18Z 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 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	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#ifdef SMP
201	{ "ap boot", &lock_class_mtx_spin },
202#ifdef __i386__
203	{ "com", &lock_class_mtx_spin },
204#endif
205#endif
206	{ "sio", &lock_class_mtx_spin },
207#ifdef __i386__
208	{ "cy", &lock_class_mtx_spin },
209#endif
210	{ "ng_node", &lock_class_mtx_spin },
211	{ "ng_worklist", &lock_class_mtx_spin },
212	{ "ithread table lock", &lock_class_mtx_spin },
213	{ "sched lock", &lock_class_mtx_spin },
214	{ "callout", &lock_class_mtx_spin },
215	/*
216	 * leaf locks
217	 */
218#ifdef SMP
219#ifdef __i386__
220	{ "imen", &lock_class_mtx_spin },
221#endif
222	{ "smp rendezvous", &lock_class_mtx_spin },
223#endif
224	{ "clk", &lock_class_mtx_spin },
225	{ NULL, NULL },
226	{ NULL, NULL }
227};
228
229static const char *dup_list[] = {
230	"process lock",
231	NULL
232};
233
234/*
235 * Pairs of locks which have been blessed
236 * Don't complain about order problems with blessed locks
237 */
238static struct witness_blessed blessed_list[] = {
239};
240static int blessed_count =
241	sizeof(blessed_list) / sizeof(struct witness_blessed);
242
243/*
244 * List of all locks in the system.
245 */
246STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
247
248static struct mtx all_mtx = {
249	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
250	  "All locks list",		/* mtx_object.lo_name */
251	  LO_INITIALIZED,		/* mtx_object.lo_flags */
252	  { 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	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
290	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
291	for (i = 0; i < WITNESS_COUNT; i++)
292		witness_free(&w_data[i]);
293	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
294		witness_child_free(&w_childdata[i]);
295	for (i = 0; i < LOCK_CHILDCOUNT; i++)
296		witness_lock_list_free(&w_locklistdata[i]);
297
298	/* First add in all the specified order lists. */
299	for (order = order_lists; order->w_name != NULL; order++) {
300		w = enroll(order->w_name, order->w_class);
301		if (w == NULL)
302			continue;
303		w->w_file = "order list";
304		for (order++; order->w_name != NULL; order++) {
305			w1 = enroll(order->w_name, order->w_class);
306			if (w1 == NULL)
307				continue;
308			w1->w_file = "order list";
309			itismychild(w, w1);
310			w = w1;
311		}
312	}
313
314	/* Iterate through all locks and add them to witness. */
315	mtx_lock(&all_mtx);
316	STAILQ_FOREACH(lock, &all_locks, lo_list) {
317		if (lock->lo_flags & LO_WITNESS)
318			lock->lo_witness = enroll(lock->lo_name,
319			    lock->lo_class);
320		else
321			lock->lo_witness = NULL;
322	}
323	mtx_unlock(&all_mtx);
324
325	/* Mark the witness code as being ready for use. */
326	atomic_store_rel_int(&witness_cold, 0);
327
328	mtx_lock(&Giant);
329}
330SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
331
332void
333witness_init(struct lock_object *lock)
334{
335	struct lock_class *class;
336
337	class = lock->lo_class;
338	if (lock->lo_flags & LO_INITIALIZED)
339		panic("%s: lock (%s) %s is already initialized", __func__,
340		    class->lc_name, lock->lo_name);
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", __func__,
344		    class->lc_name, lock->lo_name);
345	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
346	    (class->lc_flags & LC_SLEEPABLE) == 0)
347		panic("%s: lock (%s) %s can not be sleepable", __func__,
348		    class->lc_name, lock->lo_name);
349	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
350	    (class->lc_flags & LC_UPGRADABLE) == 0)
351		panic("%s: lock (%s) %s can not be upgradable", __func__,
352		    class->lc_name, lock->lo_name);
353
354	mtx_lock(&all_mtx);
355	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
356	lock->lo_flags |= LO_INITIALIZED;
357	lock_cur_cnt++;
358	if (lock_cur_cnt > lock_max_cnt)
359		lock_max_cnt = lock_cur_cnt;
360	mtx_unlock(&all_mtx);
361	if (!witness_cold && !witness_dead && panicstr == NULL &&
362	    (lock->lo_flags & LO_WITNESS) != 0)
363		lock->lo_witness = enroll(lock->lo_name, class);
364	else
365		lock->lo_witness = NULL;
366}
367
368void
369witness_destroy(struct lock_object *lock)
370{
371	struct witness *w;
372
373	if (witness_cold)
374		panic("lock (%s) %s destroyed while witness_cold",
375		    lock->lo_class->lc_name, lock->lo_name);
376	if ((lock->lo_flags & LO_INITIALIZED) == 0)
377		panic("%s: lock (%s) %s is not initialized", __func__,
378		    lock->lo_class->lc_name, lock->lo_name);
379
380	/* XXX: need to verify that no one holds the lock */
381	w = lock->lo_witness;
382	if (w != NULL) {
383		mtx_lock_spin(&w_mtx);
384		w->w_refcount--;
385		if (w->w_refcount == 0) {
386			CTR2(KTR_WITNESS,
387			    "%s: marking witness %s as dead", __func__, w->w_name);
388			w->w_name = "(dead)";
389			w->w_file = "(dead)";
390			w->w_line = 0;
391		}
392		mtx_unlock_spin(&w_mtx);
393	}
394
395	mtx_lock(&all_mtx);
396	lock_cur_cnt--;
397	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
398	lock->lo_flags &= ~LO_INITIALIZED;
399	mtx_unlock(&all_mtx);
400}
401
402static void
403witness_display_list(void(*prnt)(const char *fmt, ...),
404		     struct witness_list *list)
405{
406	struct witness *w, *w1;
407	int found;
408
409	STAILQ_FOREACH(w, list, w_typelist) {
410		if (w->w_file == NULL)
411			continue;
412		found = 0;
413		STAILQ_FOREACH(w1, list, w_typelist) {
414			if (isitmychild(w1, w)) {
415				found++;
416				break;
417			}
418		}
419		if (found)
420			continue;
421		/*
422		 * This lock has no anscestors, display its descendants.
423		 */
424		witness_displaydescendants(prnt, w);
425	}
426}
427
428static void
429witness_display(void(*prnt)(const char *fmt, ...))
430{
431	struct witness *w;
432
433	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
434	witness_levelall();
435
436	/*
437	 * First, handle sleep locks which have been acquired at least
438	 * once.
439	 */
440	prnt("Sleep locks:\n");
441	witness_display_list(prnt, &w_sleep);
442
443	/*
444	 * Now do spin locks which have been acquired at least once.
445	 */
446	prnt("\nSpin locks:\n");
447	witness_display_list(prnt, &w_spin);
448
449	/*
450	 * Finally, any locks which have not been acquired yet.
451	 */
452	prnt("\nLocks which were never acquired:\n");
453	STAILQ_FOREACH(w, &w_all, w_list) {
454		if (w->w_file != NULL)
455			continue;
456		prnt("%s\n", w->w_name);
457	}
458}
459
460void
461witness_lock(struct lock_object *lock, int flags, const char *file, int line)
462{
463	struct lock_list_entry **lock_list, *lle;
464	struct lock_instance *lock1, *lock2;
465	struct lock_class *class;
466	struct witness *w, *w1;
467	struct thread *td;
468	int i, j;
469#ifdef DDB
470	int go_into_ddb = 0;
471#endif /* DDB */
472
473	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
474	    panicstr != NULL)
475		return;
476	w = lock->lo_witness;
477	class = lock->lo_class;
478	td = curthread;
479
480	/*
481	 * We have to hold a spinlock to keep lock_list valid across the check
482	 * in the LC_SLEEPLOCK case.  In the LC_SPINLOCK case, it is already
483	 * protected by the spinlock we are currently performing the witness
484	 * checks on, so it is ok to release the lock after performing this
485	 * check.  All we have to protect is the LC_SLEEPLOCK case when no
486	 * spinlocks are held as we may get preempted during this check and
487	 * lock_list could end up pointing to some other CPU's spinlock list.
488	 */
489	mtx_lock_spin(&w_mtx);
490	lock_list = PCPU_PTR(spinlocks);
491	if (class->lc_flags & LC_SLEEPLOCK) {
492		if (*lock_list != NULL && (flags & LOP_TRYLOCK) == 0) {
493			mtx_unlock_spin(&w_mtx);
494			panic("blockable sleep lock (%s) %s @ %s:%d",
495			    class->lc_name, lock->lo_name, file, line);
496		}
497		lock_list = &td->td_sleeplocks;
498	}
499	mtx_unlock_spin(&w_mtx);
500
501	/*
502	 * Try locks do not block if they fail to acquire the lock, thus
503	 * there is no danger of deadlocks or of switching while holding a
504	 * spin lock if we acquire a lock via a try operation.
505	 */
506	if (flags & LOP_TRYLOCK)
507		goto out;
508
509	/*
510	 * Is this the first lock acquired?  If so, then no order checking
511	 * is needed.
512	 */
513	if (*lock_list == NULL)
514		goto out;
515
516	/*
517	 * Check to see if we are recursing on a lock we already own.
518	 */
519	lock1 = find_instance(*lock_list, lock);
520	if (lock1 != NULL) {
521		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
522		    (flags & LOP_EXCLUSIVE) == 0) {
523			printf("shared lock of (%s) %s @ %s:%d\n",
524			    class->lc_name, lock->lo_name, file, line);
525			printf("while exclusively locked from %s:%d\n",
526			    lock1->li_file, lock1->li_line);
527			panic("share->excl");
528		}
529		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
530		    (flags & LOP_EXCLUSIVE) != 0) {
531			printf("exclusive lock of (%s) %s @ %s:%d\n",
532			    class->lc_name, lock->lo_name, file, line);
533			printf("while share locked from %s:%d\n",
534			    lock1->li_file, lock1->li_line);
535			panic("excl->share");
536		}
537		lock1->li_flags++;
538		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
539			printf(
540			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
541			    class->lc_name, lock->lo_name, file, line);
542			printf("first acquired @ %s:%d\n", lock1->li_file,
543			    lock1->li_line);
544			panic("recurse");
545		}
546		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
547		    td->td_proc->p_pid, lock->lo_name,
548		    lock1->li_flags & LI_RECURSEMASK);
549		lock1->li_file = file;
550		lock1->li_line = line;
551		return;
552	}
553
554	/*
555	 * Check for duplicate locks of the same type.  Note that we only
556	 * have to check for this on the last lock we just acquired.  Any
557	 * other cases will be caught as lock order violations.
558	 */
559	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
560	w1 = lock1->li_lock->lo_witness;
561	if (w1 == w) {
562		if (w->w_same_squawked || dup_ok(w))
563			goto out;
564		w->w_same_squawked = 1;
565		printf("acquiring duplicate lock of same type: \"%s\"\n",
566			lock->lo_name);
567		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
568		printf(" 2nd @ %s:%d\n", file, line);
569#ifdef DDB
570		go_into_ddb = 1;
571#endif /* DDB */
572		goto out;
573	}
574	MPASS(!mtx_owned(&w_mtx));
575	mtx_lock_spin(&w_mtx);
576	/*
577	 * If we have a known higher number just say ok
578	 */
579	if (witness_watch > 1 && w->w_level > w1->w_level) {
580		mtx_unlock_spin(&w_mtx);
581		goto out;
582	}
583	if (isitmydescendant(w1, w)) {
584		mtx_unlock_spin(&w_mtx);
585		goto out;
586	}
587	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
588		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
589
590			MPASS(j < WITNESS_COUNT);
591			lock1 = &lle->ll_children[i];
592			w1 = lock1->li_lock->lo_witness;
593
594			/*
595			 * If this lock doesn't undergo witness checking,
596			 * then skip it.
597			 */
598			if (w1 == NULL) {
599				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
600				    ("lock missing witness structure"));
601				continue;
602			}
603			/*
604			 * If we are locking Giant and we slept with this
605			 * lock, then skip it.
606			 */
607			if ((lock1->li_flags & LI_SLEPT) != 0 &&
608			    lock == &Giant.mtx_object)
609				continue;
610			if (!isitmydescendant(w, w1))
611				continue;
612			/*
613			 * We have a lock order violation, check to see if it
614			 * is allowed or has already been yelled about.
615			 */
616			mtx_unlock_spin(&w_mtx);
617			if (blessed(w, w1))
618				goto out;
619			if (lock1->li_lock == &Giant.mtx_object) {
620				if (w1->w_Giant_squawked)
621					goto out;
622				else
623					w1->w_Giant_squawked = 1;
624			} else {
625				if (w1->w_other_squawked)
626					goto out;
627				else
628					w1->w_other_squawked = 1;
629			}
630			/*
631			 * Ok, yell about it.
632			 */
633			printf("lock order reversal\n");
634			/*
635			 * Try to locate an earlier lock with
636			 * witness w in our list.
637			 */
638			do {
639				lock2 = &lle->ll_children[i];
640				MPASS(lock2->li_lock != NULL);
641				if (lock2->li_lock->lo_witness == w)
642					break;
643				i--;
644				if (i == 0 && lle->ll_next != NULL) {
645					lle = lle->ll_next;
646					i = lle->ll_count - 1;
647					MPASS(i != 0);
648				}
649			} while (i >= 0);
650			if (i < 0) {
651				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
652				    lock1->li_lock->lo_name, lock1->li_file,
653				    lock1->li_line);
654				printf(" 2nd %p %s @ %s:%d\n", lock,
655				    lock->lo_name, file, line);
656			} else {
657				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
658				    lock2->li_lock->lo_name, lock2->li_file,
659				    lock2->li_line);
660				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
661				    lock1->li_lock->lo_name, lock1->li_file,
662				    lock1->li_line);
663				printf(" 3rd %p %s @ %s:%d\n", lock,
664				    lock->lo_name, file, line);
665			}
666#ifdef DDB
667			go_into_ddb = 1;
668#endif /* DDB */
669			goto out;
670		}
671	}
672	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
673	/*
674	 * Don't build a new relationship if we are locking Giant just
675	 * after waking up and the previous lock in the list was acquired
676	 * prior to blocking.
677	 */
678	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
679		mtx_unlock_spin(&w_mtx);
680	else {
681		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
682		    lock->lo_name, lock1->li_lock->lo_name);
683		if (!itismychild(lock1->li_lock->lo_witness, w))
684			mtx_unlock_spin(&w_mtx);
685	}
686
687out:
688#ifdef DDB
689	if (witness_ddb && go_into_ddb)
690		Debugger(__func__);
691#endif /* DDB */
692	w->w_file = file;
693	w->w_line = line;
694
695	lle = *lock_list;
696	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
697		lle = witness_lock_list_get();
698		if (lle == NULL)
699			return;
700		lle->ll_next = *lock_list;
701		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
702		    td->td_proc->p_pid, lle);
703		*lock_list = lle;
704	}
705	lock1 = &lle->ll_children[lle->ll_count++];
706	lock1->li_lock = lock;
707	lock1->li_line = line;
708	lock1->li_file = file;
709	if ((flags & LOP_EXCLUSIVE) != 0)
710		lock1->li_flags = LI_EXCLUSIVE;
711	else
712		lock1->li_flags = 0;
713	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
714	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
715}
716
717void
718witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
719{
720	struct lock_instance *instance;
721	struct lock_class *class;
722
723	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
724	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
725		return;
726	class = lock->lo_class;
727	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
728		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
729		    class->lc_name, lock->lo_name, file, line);
730	if ((flags & LOP_TRYLOCK) == 0)
731		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
732		    lock->lo_name, file, line);
733	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
734		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
735		    class->lc_name, lock->lo_name, file, line);
736	instance = find_instance(curthread->td_sleeplocks, lock);
737	if (instance == NULL)
738		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
739		    class->lc_name, lock->lo_name, file, line);
740	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
741		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
742		    class->lc_name, lock->lo_name, file, line);
743	if ((instance->li_flags & LI_RECURSEMASK) != 0)
744		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
745		    class->lc_name, lock->lo_name,
746		    instance->li_flags & LI_RECURSEMASK, file, line);
747	instance->li_flags |= LI_EXCLUSIVE;
748}
749
750void
751witness_downgrade(struct lock_object *lock, int flags, const char *file,
752    int line)
753{
754	struct lock_instance *instance;
755	struct lock_class *class;
756
757	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
758	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
759		return;
760	class = lock->lo_class;
761	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
762		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
763		    class->lc_name, lock->lo_name, file, line);
764	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
765		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
766		    class->lc_name, lock->lo_name, file, line);
767	instance = find_instance(curthread->td_sleeplocks, lock);
768	if (instance == NULL)
769		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
770		    class->lc_name, lock->lo_name, file, line);
771	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
772		panic("downgrade of shared lock (%s) %s @ %s:%d",
773		    class->lc_name, lock->lo_name, file, line);
774	if ((instance->li_flags & LI_RECURSEMASK) != 0)
775		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
776		    class->lc_name, lock->lo_name,
777		    instance->li_flags & LI_RECURSEMASK, file, line);
778	instance->li_flags &= ~LI_EXCLUSIVE;
779}
780
781void
782witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
783{
784	struct lock_list_entry **lock_list, *lle;
785	struct lock_instance *instance;
786	struct lock_class *class;
787	struct thread *td;
788	critical_t s;
789	int i, j;
790
791	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
792	    panicstr != NULL)
793		return;
794	td = curthread;
795	class = lock->lo_class;
796	if (class->lc_flags & LC_SLEEPLOCK)
797		lock_list = &td->td_sleeplocks;
798	else
799		lock_list = PCPU_PTR(spinlocks);
800	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
801		for (i = 0; i < (*lock_list)->ll_count; i++) {
802			instance = &(*lock_list)->ll_children[i];
803			if (instance->li_lock == lock) {
804				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
805				    (flags & LOP_EXCLUSIVE) == 0) {
806					printf(
807					"shared unlock of (%s) %s @ %s:%d\n",
808					    class->lc_name, lock->lo_name,
809					    file, line);
810					printf(
811					"while exclusively locked from %s:%d\n",
812					    instance->li_file,
813					    instance->li_line);
814					panic("excl->ushare");
815				}
816				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
817				    (flags & LOP_EXCLUSIVE) != 0) {
818					printf(
819					"exclusive unlock of (%s) %s @ %s:%d\n",
820					    class->lc_name, lock->lo_name,
821					    file, line);
822					printf(
823					"while share locked from %s:%d\n",
824					    instance->li_file,
825					    instance->li_line);
826					panic("share->uexcl");
827				}
828				/* If we are recursed, unrecurse. */
829				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
830					CTR4(KTR_WITNESS,
831				    "%s: pid %d unrecursed on %s r=%d", __func__,
832					    td->td_proc->p_pid,
833					    instance->li_lock->lo_name,
834					    instance->li_flags);
835					instance->li_flags--;
836					goto out;
837				}
838				s = cpu_critical_enter();
839				CTR4(KTR_WITNESS,
840				    "%s: pid %d removed %s from lle[%d]", __func__,
841				    td->td_proc->p_pid,
842				    instance->li_lock->lo_name,
843				    (*lock_list)->ll_count - 1);
844				(*lock_list)->ll_count--;
845				for (j = i; j < (*lock_list)->ll_count; j++)
846					(*lock_list)->ll_children[j] =
847					    (*lock_list)->ll_children[j + 1];
848				cpu_critical_exit(s);
849				if ((*lock_list)->ll_count == 0) {
850					lle = *lock_list;
851					*lock_list = lle->ll_next;
852					CTR3(KTR_WITNESS,
853					    "%s: pid %d removed lle %p", __func__,
854					    td->td_proc->p_pid, lle);
855					witness_lock_list_free(lle);
856				}
857				goto out;
858			}
859		}
860	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
861	    file, line);
862out:
863	/*
864	 * We don't need to protect this PCPU_GET() here against preemption
865	 * because if we hold any spinlocks then we are already protected,
866	 * and if we don't we will get NULL if we hold no spinlocks even if
867	 * we switch CPU's while reading it.
868	 */
869	if (class->lc_flags & LC_SLEEPLOCK) {
870		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
871			panic("switchable sleep unlock (%s) %s @ %s:%d",
872			    class->lc_name, lock->lo_name, file, line);
873	}
874}
875
876/*
877 * Warn if any held locks are not sleepable.  Note that Giant and the lock
878 * passed in are both special cases since they are both released during the
879 * sleep process and aren't actually held while the thread is asleep.
880 */
881int
882witness_sleep(int check_only, struct lock_object *lock, const char *file,
883	      int line)
884{
885	struct lock_list_entry **lock_list, *lle;
886	struct lock_instance *lock1;
887	struct thread *td;
888	critical_t savecrit;
889	int i, n;
890
891	if (witness_dead || panicstr != NULL)
892		return (0);
893	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
894	n = 0;
895	/*
896	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
897	 */
898	savecrit = cpu_critical_enter();
899	td = curthread;
900	lock_list = &td->td_sleeplocks;
901again:
902	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
903		for (i = lle->ll_count - 1; i >= 0; i--) {
904			lock1 = &lle->ll_children[i];
905			if (lock1->li_lock == lock ||
906			    lock1->li_lock == &Giant.mtx_object)
907				continue;
908			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
909				if (check_only == 0) {
910					CTR3(KTR_WITNESS,
911				    "pid %d: sleeping with lock (%s) %s held",
912					    td->td_proc->p_pid,
913					    lock1->li_lock->lo_class->lc_name,
914					    lock1->li_lock->lo_name);
915					lock1->li_flags |= LI_SLEPT;
916				}
917				continue;
918			}
919			n++;
920			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
921			    file, line, check_only ? "could sleep" : "sleeping",
922			    lock1->li_lock->lo_name, lock1->li_file,
923			    lock1->li_line);
924		}
925	if (lock_list == &td->td_sleeplocks) {
926		lock_list = PCPU_PTR(spinlocks);
927		goto again;
928	}
929#ifdef DDB
930	if (witness_ddb && n)
931		Debugger(__func__);
932#endif /* DDB */
933	cpu_critical_exit(savecrit);
934	return (n);
935}
936
937static struct witness *
938enroll(const char *description, struct lock_class *lock_class)
939{
940	struct witness *w;
941
942	if (!witness_watch || witness_dead || panicstr != NULL)
943		return (NULL);
944	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
945		return (NULL);
946	mtx_lock_spin(&w_mtx);
947	STAILQ_FOREACH(w, &w_all, w_list) {
948		if (strcmp(description, w->w_name) == 0) {
949			w->w_refcount++;
950			mtx_unlock_spin(&w_mtx);
951			if (lock_class != w->w_class)
952				panic(
953				"lock (%s) %s does not match earlier (%s) lock",
954				    description, lock_class->lc_name,
955				    w->w_class->lc_name);
956			return (w);
957		}
958	}
959	/*
960	 * This isn't quite right, as witness_cold is still 0 while we
961	 * enroll all the locks initialized before witness_initialize().
962	 */
963	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
964		mtx_unlock_spin(&w_mtx);
965		panic("spin lock %s not in order list", description);
966	}
967	if ((w = witness_get()) == NULL)
968		return (NULL);
969	w->w_name = description;
970	w->w_class = lock_class;
971	w->w_refcount = 1;
972	STAILQ_INSERT_HEAD(&w_all, w, w_list);
973	if (lock_class->lc_flags & LC_SPINLOCK)
974		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
975	else if (lock_class->lc_flags & LC_SLEEPLOCK)
976		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
977	else {
978		mtx_unlock_spin(&w_mtx);
979		panic("lock class %s is not sleep or spin",
980		    lock_class->lc_name);
981	}
982	mtx_unlock_spin(&w_mtx);
983	return (w);
984}
985
986static int
987itismychild(struct witness *parent, struct witness *child)
988{
989	static int recursed;
990	struct witness_child_list_entry **wcl;
991	struct witness_list *list;
992
993	MPASS(child != NULL && parent != NULL);
994	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
995	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
996		panic(
997		"%s: parent (%s) and child (%s) are not the same lock type",
998		    __func__, parent->w_class->lc_name,
999		    child->w_class->lc_name);
1000
1001	/*
1002	 * Insert "child" after "parent"
1003	 */
1004	wcl = &parent->w_children;
1005	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1006		wcl = &(*wcl)->wcl_next;
1007	if (*wcl == NULL) {
1008		*wcl = witness_child_get();
1009		if (*wcl == NULL)
1010			return (1);
1011	}
1012	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1013
1014	/*
1015	 * Now prune whole tree.  We look for cases where a lock is now
1016	 * both a descendant and a direct child of a given lock.  In that
1017	 * case, we want to remove the direct child link from the tree.
1018	 */
1019	if (recursed)
1020		return (0);
1021	recursed = 1;
1022	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1023		list = &w_sleep;
1024	else
1025		list = &w_spin;
1026	STAILQ_FOREACH(child, list, w_typelist) {
1027		STAILQ_FOREACH(parent, list, w_typelist) {
1028			if (!isitmychild(parent, child))
1029				continue;
1030			removechild(parent, child);
1031			if (isitmydescendant(parent, child))
1032				continue;
1033			itismychild(parent, child);
1034		}
1035	}
1036	recursed = 0;
1037	witness_levelall();
1038	return (0);
1039}
1040
1041static void
1042removechild(struct witness *parent, struct witness *child)
1043{
1044	struct witness_child_list_entry **wcl, *wcl1;
1045	int i;
1046
1047	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1048		for (i = 0; i < (*wcl)->wcl_count; i++)
1049			if ((*wcl)->wcl_children[i] == child)
1050				goto found;
1051	return;
1052found:
1053	(*wcl)->wcl_count--;
1054	if ((*wcl)->wcl_count > i)
1055		(*wcl)->wcl_children[i] =
1056		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1057	MPASS((*wcl)->wcl_children[i] != NULL);
1058	if ((*wcl)->wcl_count != 0)
1059		return;
1060	wcl1 = *wcl;
1061	*wcl = wcl1->wcl_next;
1062	witness_child_free(wcl1);
1063}
1064
1065static int
1066isitmychild(struct witness *parent, struct witness *child)
1067{
1068	struct witness_child_list_entry *wcl;
1069	int i;
1070
1071	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1072		for (i = 0; i < wcl->wcl_count; i++) {
1073			if (wcl->wcl_children[i] == child)
1074				return (1);
1075		}
1076	}
1077	return (0);
1078}
1079
1080static int
1081isitmydescendant(struct witness *parent, struct witness *child)
1082{
1083	struct witness_child_list_entry *wcl;
1084	int i, j;
1085
1086	if (isitmychild(parent, child))
1087		return (1);
1088	j = 0;
1089	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1090		MPASS(j < 1000);
1091		for (i = 0; i < wcl->wcl_count; i++) {
1092			if (isitmydescendant(wcl->wcl_children[i], child))
1093				return (1);
1094		}
1095		j++;
1096	}
1097	return (0);
1098}
1099
1100void
1101witness_levelall (void)
1102{
1103	struct witness_list *list;
1104	struct witness *w, *w1;
1105
1106	/*
1107	 * First clear all levels.
1108	 */
1109	STAILQ_FOREACH(w, &w_all, w_list) {
1110		w->w_level = 0;
1111	}
1112
1113	/*
1114	 * Look for locks with no parent and level all their descendants.
1115	 */
1116	STAILQ_FOREACH(w, &w_all, w_list) {
1117		/*
1118		 * This is just an optimization, technically we could get
1119		 * away just walking the all list each time.
1120		 */
1121		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1122			list = &w_sleep;
1123		else
1124			list = &w_spin;
1125		STAILQ_FOREACH(w1, list, w_typelist) {
1126			if (isitmychild(w1, w))
1127				goto skip;
1128		}
1129		witness_leveldescendents(w, 0);
1130	skip:
1131	}
1132}
1133
1134static void
1135witness_leveldescendents(struct witness *parent, int level)
1136{
1137	struct witness_child_list_entry *wcl;
1138	int i;
1139
1140	if (parent->w_level < level)
1141		parent->w_level = level;
1142	level++;
1143	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1144		for (i = 0; i < wcl->wcl_count; i++)
1145			witness_leveldescendents(wcl->wcl_children[i], level);
1146}
1147
1148static void
1149witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1150			   struct witness *parent)
1151{
1152	struct witness_child_list_entry *wcl;
1153	int i, level;
1154
1155	level =  parent->w_level;
1156	prnt("%-2d", level);
1157	for (i = 0; i < level; i++)
1158		prnt(" ");
1159	prnt("%s", parent->w_name);
1160	if (parent->w_file != NULL)
1161		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1162		    parent->w_line);
1163	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1164		for (i = 0; i < wcl->wcl_count; i++)
1165			    witness_displaydescendants(prnt,
1166				wcl->wcl_children[i]);
1167}
1168
1169static int
1170dup_ok(struct witness *w)
1171{
1172	const char **dup;
1173
1174	for (dup = dup_list; *dup != NULL; dup++)
1175		if (strcmp(w->w_name, *dup) == 0)
1176			return (1);
1177	return (0);
1178}
1179
1180static int
1181blessed(struct witness *w1, struct witness *w2)
1182{
1183	int i;
1184	struct witness_blessed *b;
1185
1186	for (i = 0; i < blessed_count; i++) {
1187		b = &blessed_list[i];
1188		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1189			if (strcmp(w2->w_name, b->b_lock2) == 0)
1190				return (1);
1191			continue;
1192		}
1193		if (strcmp(w1->w_name, b->b_lock2) == 0)
1194			if (strcmp(w2->w_name, b->b_lock1) == 0)
1195				return (1);
1196	}
1197	return (0);
1198}
1199
1200static struct witness *
1201witness_get(void)
1202{
1203	struct witness *w;
1204
1205	if (witness_dead) {
1206		mtx_unlock_spin(&w_mtx);
1207		return (NULL);
1208	}
1209	if (STAILQ_EMPTY(&w_free)) {
1210		witness_dead = 1;
1211		mtx_unlock_spin(&w_mtx);
1212		printf("%s: witness exhausted\n", __func__);
1213		return (NULL);
1214	}
1215	w = STAILQ_FIRST(&w_free);
1216	STAILQ_REMOVE_HEAD(&w_free, w_list);
1217	bzero(w, sizeof(*w));
1218	return (w);
1219}
1220
1221static void
1222witness_free(struct witness *w)
1223{
1224
1225	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1226}
1227
1228static struct witness_child_list_entry *
1229witness_child_get(void)
1230{
1231	struct witness_child_list_entry *wcl;
1232
1233	if (witness_dead) {
1234		mtx_unlock_spin(&w_mtx);
1235		return (NULL);
1236	}
1237	wcl = w_child_free;
1238	if (wcl == NULL) {
1239		witness_dead = 1;
1240		mtx_unlock_spin(&w_mtx);
1241		printf("%s: witness exhausted\n", __func__);
1242		return (NULL);
1243	}
1244	w_child_free = wcl->wcl_next;
1245	bzero(wcl, sizeof(*wcl));
1246	return (wcl);
1247}
1248
1249static void
1250witness_child_free(struct witness_child_list_entry *wcl)
1251{
1252
1253	wcl->wcl_next = w_child_free;
1254	w_child_free = wcl;
1255}
1256
1257static struct lock_list_entry *
1258witness_lock_list_get(void)
1259{
1260	struct lock_list_entry *lle;
1261
1262	if (witness_dead)
1263		return (NULL);
1264	mtx_lock_spin(&w_mtx);
1265	lle = w_lock_list_free;
1266	if (lle == NULL) {
1267		witness_dead = 1;
1268		mtx_unlock_spin(&w_mtx);
1269		printf("%s: witness exhausted\n", __func__);
1270		return (NULL);
1271	}
1272	w_lock_list_free = lle->ll_next;
1273	mtx_unlock_spin(&w_mtx);
1274	bzero(lle, sizeof(*lle));
1275	return (lle);
1276}
1277
1278static void
1279witness_lock_list_free(struct lock_list_entry *lle)
1280{
1281
1282	mtx_lock_spin(&w_mtx);
1283	lle->ll_next = w_lock_list_free;
1284	w_lock_list_free = lle;
1285	mtx_unlock_spin(&w_mtx);
1286}
1287
1288static struct lock_instance *
1289find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1290{
1291	struct lock_list_entry *lle;
1292	struct lock_instance *instance;
1293	int i;
1294
1295	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1296		for (i = lle->ll_count - 1; i >= 0; i--) {
1297			instance = &lle->ll_children[i];
1298			if (instance->li_lock == lock)
1299				return (instance);
1300		}
1301	return (NULL);
1302}
1303
1304int
1305witness_list_locks(struct lock_list_entry **lock_list)
1306{
1307	struct lock_list_entry *lle;
1308	struct lock_instance *instance;
1309	struct lock_object *lock;
1310	int i, nheld;
1311
1312	nheld = 0;
1313	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1314		for (i = lle->ll_count - 1; i >= 0; i--) {
1315			instance = &lle->ll_children[i];
1316			lock = instance->li_lock;
1317			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1318			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1319			    "exclusive" : "shared",
1320			    lock->lo_class->lc_name, lock->lo_name, lock,
1321			    instance->li_file, instance->li_line);
1322			nheld++;
1323		}
1324	return (nheld);
1325}
1326
1327/*
1328 * Calling this on td != curthread is bad unless we are in ddb.
1329 */
1330int
1331witness_list(struct thread *td)
1332{
1333	critical_t savecrit;
1334	int nheld;
1335
1336	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1337#ifdef DDB
1338	KASSERT(td == curthread || db_active,
1339	    ("%s: td != curthread and we aren't in the debugger", __func__));
1340	if (!db_active && witness_dead)
1341		return (0);
1342#else
1343	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1344	if (witness_dead)
1345		return (0);
1346#endif
1347	nheld = witness_list_locks(&td->td_sleeplocks);
1348
1349	/*
1350	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1351	 * if td is currently executing on some other CPU and holds spin locks
1352	 * as we won't display those locks.  If we had a MI way of getting
1353	 * the per-cpu data for a given cpu then we could use
1354	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1355	 * and "fix" this.
1356	 */
1357	if (td == curthread) {
1358		/*
1359		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1360		 * change.
1361		 */
1362		savecrit = cpu_critical_enter();
1363		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1364		cpu_critical_exit(savecrit);
1365	}
1366	return (nheld);
1367}
1368
1369void
1370witness_save(struct lock_object *lock, const char **filep, int *linep)
1371{
1372	struct lock_instance *instance;
1373
1374	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1375	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1376		return;
1377	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1378		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1379		    lock->lo_class->lc_name, lock->lo_name);
1380	instance = find_instance(curthread->td_sleeplocks, lock);
1381	if (instance == NULL)
1382		panic("%s: lock (%s) %s not locked", __func__,
1383		    lock->lo_class->lc_name, lock->lo_name);
1384	*filep = instance->li_file;
1385	*linep = instance->li_line;
1386}
1387
1388void
1389witness_restore(struct lock_object *lock, const char *file, int line)
1390{
1391	struct lock_instance *instance;
1392
1393	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1394	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1395		return;
1396	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1397		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1398		    lock->lo_class->lc_name, lock->lo_name);
1399	instance = find_instance(curthread->td_sleeplocks, lock);
1400	if (instance == NULL)
1401		panic("%s: lock (%s) %s not locked", __func__,
1402		    lock->lo_class->lc_name, lock->lo_name);
1403	lock->lo_witness->w_file = file;
1404	lock->lo_witness->w_line = line;
1405	instance->li_file = file;
1406	instance->li_line = line;
1407}
1408
1409void
1410witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1411{
1412#ifdef INVARIANT_SUPPORT
1413	struct lock_instance *instance;
1414
1415	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1416		return;
1417	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1418		instance = find_instance(curthread->td_sleeplocks, lock);
1419	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1420		instance = find_instance(PCPU_GET(spinlocks), lock);
1421	else {
1422		panic("Lock (%s) %s is not sleep or spin!",
1423		    lock->lo_class->lc_name, lock->lo_name);
1424		return;
1425	}
1426	switch (flags) {
1427	case LA_UNLOCKED:
1428		if (instance != NULL)
1429			panic("Lock (%s) %s locked @ %s:%d.",
1430			    lock->lo_class->lc_name, lock->lo_name, file, line);
1431		break;
1432	case LA_LOCKED:
1433	case LA_LOCKED | LA_RECURSED:
1434	case LA_LOCKED | LA_NOTRECURSED:
1435	case LA_SLOCKED:
1436	case LA_SLOCKED | LA_RECURSED:
1437	case LA_SLOCKED | LA_NOTRECURSED:
1438	case LA_XLOCKED:
1439	case LA_XLOCKED | LA_RECURSED:
1440	case LA_XLOCKED | LA_NOTRECURSED:
1441		if (instance == NULL) {
1442			panic("Lock (%s) %s not locked @ %s:%d.",
1443			    lock->lo_class->lc_name, lock->lo_name, file, line);
1444			break;
1445		}
1446		if ((flags & LA_XLOCKED) != 0 &&
1447		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1448			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1449			    lock->lo_class->lc_name, lock->lo_name, file, line);
1450		if ((flags & LA_SLOCKED) != 0 &&
1451		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1452			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1453			    lock->lo_class->lc_name, lock->lo_name, file, line);
1454		if ((flags & LA_RECURSED) != 0 &&
1455		    (instance->li_flags & LI_RECURSEMASK) == 0)
1456			panic("Lock (%s) %s not recursed @ %s:%d.",
1457			    lock->lo_class->lc_name, lock->lo_name, file, line);
1458		if ((flags & LA_NOTRECURSED) != 0 &&
1459		    (instance->li_flags & LI_RECURSEMASK) != 0)
1460			panic("Lock (%s) %s recursed @ %s:%d.",
1461			    lock->lo_class->lc_name, lock->lo_name, file, line);
1462		break;
1463	default:
1464		panic("Invalid lock assertion at %s:%d.", file, line);
1465
1466	}
1467#endif	/* INVARIANT_SUPPORT */
1468}
1469
1470#ifdef DDB
1471
1472DB_SHOW_COMMAND(locks, db_witness_list)
1473{
1474	struct thread *td;
1475	pid_t pid;
1476	struct proc *p;
1477
1478	if (have_addr) {
1479		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1480		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1481		    ((addr >> 16) % 16) * 10000;
1482		/* sx_slock(&allproc_lock); */
1483		FOREACH_PROC_IN_SYSTEM(p) {
1484			if (p->p_pid == pid)
1485				break;
1486		}
1487		/* sx_sunlock(&allproc_lock); */
1488		if (p == NULL) {
1489			db_printf("pid %d not found\n", pid);
1490			return;
1491		}
1492		td = &p->p_thread; /* XXXKSE */
1493	} else {
1494		td = curthread;
1495	}
1496	witness_list(td);
1497}
1498
1499DB_SHOW_COMMAND(witness, db_witness_display)
1500{
1501
1502	witness_display(db_printf);
1503}
1504#endif
1505