subr_witness.c revision 96122
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 96122 2002-05-06 19:31:28Z alfred $
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	{ NULL, NULL },
229	{ NULL, NULL }
230};
231
232/*
233 * Pairs of locks which have been blessed
234 * Don't complain about order problems with blessed locks
235 */
236static struct witness_blessed blessed_list[] = {
237};
238static int blessed_count =
239	sizeof(blessed_list) / sizeof(struct witness_blessed);
240
241/*
242 * List of all locks in the system.
243 */
244STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245
246static struct mtx all_mtx = {
247	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248	  "All locks list",		/* mtx_object.lo_name */
249	  "All locks list",		/* mtx_object.lo_type */
250	  LO_INITIALIZED,		/* mtx_object.lo_flags */
251	  { NULL },			/* mtx_object.lo_list */
252	  NULL },			/* mtx_object.lo_witness */
253	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
254	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
255	{ NULL, NULL }			/* mtx_contested */
256};
257
258/*
259 * This global is set to 0 once it becomes safe to use the witness code.
260 */
261static int witness_cold = 1;
262
263/*
264 * Global variables for book keeping.
265 */
266static int lock_cur_cnt;
267static int lock_max_cnt;
268
269/*
270 * The WITNESS-enabled diagnostic code.
271 */
272static void
273witness_initialize(void *dummy __unused)
274{
275	struct lock_object *lock;
276	struct witness_order_list_entry *order;
277	struct witness *w, *w1;
278	int i;
279
280	/*
281	 * We have to release Giant before initializing its witness
282	 * structure so that WITNESS doesn't get confused.
283	 */
284	mtx_unlock(&Giant);
285	mtx_assert(&Giant, MA_NOTOWNED);
286
287	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
288	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
289	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
290	    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_type,
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_type, 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	if (class->lc_flags & LC_SLEEPLOCK) {
481		/*
482		 * Since spin locks include a critical section, this check
483		 * impliclty enforces a lock order of all sleep locks before
484		 * all spin locks.
485		 */
486		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
487			panic("blockable sleep lock (%s) %s @ %s:%d",
488			    class->lc_name, lock->lo_name, file, line);
489		lock_list = &td->td_sleeplocks;
490	} else
491		lock_list = PCPU_PTR(spinlocks);
492
493	/*
494	 * Try locks do not block if they fail to acquire the lock, thus
495	 * there is no danger of deadlocks or of switching while holding a
496	 * spin lock if we acquire a lock via a try operation.
497	 */
498	if (flags & LOP_TRYLOCK)
499		goto out;
500
501	/*
502	 * Is this the first lock acquired?  If so, then no order checking
503	 * is needed.
504	 */
505	if (*lock_list == NULL)
506		goto out;
507
508	/*
509	 * Check to see if we are recursing on a lock we already own.
510	 */
511	lock1 = find_instance(*lock_list, lock);
512	if (lock1 != NULL) {
513		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
514		    (flags & LOP_EXCLUSIVE) == 0) {
515			printf("shared lock of (%s) %s @ %s:%d\n",
516			    class->lc_name, lock->lo_name, file, line);
517			printf("while exclusively locked from %s:%d\n",
518			    lock1->li_file, lock1->li_line);
519			panic("share->excl");
520		}
521		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
522		    (flags & LOP_EXCLUSIVE) != 0) {
523			printf("exclusive lock of (%s) %s @ %s:%d\n",
524			    class->lc_name, lock->lo_name, file, line);
525			printf("while share locked from %s:%d\n",
526			    lock1->li_file, lock1->li_line);
527			panic("excl->share");
528		}
529		lock1->li_flags++;
530		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
531			printf(
532			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
533			    class->lc_name, lock->lo_name, file, line);
534			printf("first acquired @ %s:%d\n", lock1->li_file,
535			    lock1->li_line);
536			panic("recurse");
537		}
538		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
539		    td->td_proc->p_pid, lock->lo_name,
540		    lock1->li_flags & LI_RECURSEMASK);
541		lock1->li_file = file;
542		lock1->li_line = line;
543		return;
544	}
545
546	/*
547	 * Check for duplicate locks of the same type.  Note that we only
548	 * have to check for this on the last lock we just acquired.  Any
549	 * other cases will be caught as lock order violations.
550	 */
551	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
552	w1 = lock1->li_lock->lo_witness;
553	if (w1 == w) {
554		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
555			goto out;
556		w->w_same_squawked = 1;
557		printf("acquiring duplicate lock of same type: \"%s\"\n",
558			lock->lo_type);
559		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
560		    lock1->li_file, lock1->li_line);
561		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
562#ifdef DDB
563		go_into_ddb = 1;
564#endif /* DDB */
565		goto out;
566	}
567	MPASS(!mtx_owned(&w_mtx));
568	mtx_lock_spin(&w_mtx);
569	/*
570	 * If we have a known higher number just say ok
571	 */
572	if (witness_watch > 1 && w->w_level > w1->w_level) {
573		mtx_unlock_spin(&w_mtx);
574		goto out;
575	}
576	if (isitmydescendant(w1, w)) {
577		mtx_unlock_spin(&w_mtx);
578		goto out;
579	}
580	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
581		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
582
583			MPASS(j < WITNESS_COUNT);
584			lock1 = &lle->ll_children[i];
585			w1 = lock1->li_lock->lo_witness;
586
587			/*
588			 * If this lock doesn't undergo witness checking,
589			 * then skip it.
590			 */
591			if (w1 == NULL) {
592				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
593				    ("lock missing witness structure"));
594				continue;
595			}
596			/*
597			 * If we are locking Giant and we slept with this
598			 * lock, then skip it.
599			 */
600			if ((lock1->li_flags & LI_SLEPT) != 0 &&
601			    lock == &Giant.mtx_object)
602				continue;
603			/*
604			 * If we are locking a sleepable lock and this lock
605			 * isn't sleepable and isn't Giant, we want to treat
606			 * it as a lock order violation to enfore a general
607			 * lock order of sleepable locks before non-sleepable
608			 * locks.  Thus, we only bother checking the lock
609			 * order hierarchy if we pass the initial test.
610			 */
611			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
612			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
613			    lock1->li_lock != &Giant.mtx_object)) &&
614			    !isitmydescendant(w, w1))
615				continue;
616			/*
617			 * We have a lock order violation, check to see if it
618			 * is allowed or has already been yelled about.
619			 */
620			mtx_unlock_spin(&w_mtx);
621			if (blessed(w, w1))
622				goto out;
623			if (lock1->li_lock == &Giant.mtx_object) {
624				if (w1->w_Giant_squawked)
625					goto out;
626				else
627					w1->w_Giant_squawked = 1;
628			} else {
629				if (w1->w_other_squawked)
630					goto out;
631				else
632					w1->w_other_squawked = 1;
633			}
634			/*
635			 * Ok, yell about it.
636			 */
637			printf("lock order reversal\n");
638			/*
639			 * Try to locate an earlier lock with
640			 * witness w in our list.
641			 */
642			do {
643				lock2 = &lle->ll_children[i];
644				MPASS(lock2->li_lock != NULL);
645				if (lock2->li_lock->lo_witness == w)
646					break;
647				i--;
648				if (i == 0 && lle->ll_next != NULL) {
649					lle = lle->ll_next;
650					i = lle->ll_count - 1;
651					MPASS(i != 0);
652				}
653			} while (i >= 0);
654			if (i < 0) {
655				printf(" 1st %p %s (%s) @ %s:%d\n",
656				    lock1->li_lock, lock1->li_lock->lo_name,
657				    lock1->li_lock->lo_type, lock1->li_file,
658				    lock1->li_line);
659				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
660				    lock->lo_name, lock->lo_type, file, line);
661			} else {
662				printf(" 1st %p %s (%s) @ %s:%d\n",
663				    lock2->li_lock, lock2->li_lock->lo_name,
664				    lock2->li_lock->lo_type, lock2->li_file,
665				    lock2->li_line);
666				printf(" 2nd %p %s (%s) @ %s:%d\n",
667				    lock1->li_lock, lock1->li_lock->lo_name,
668				    lock1->li_lock->lo_type, lock1->li_file,
669				    lock1->li_line);
670				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
671				    lock->lo_name, lock->lo_type, file, line);
672			}
673#ifdef DDB
674			go_into_ddb = 1;
675#endif /* DDB */
676			goto out;
677		}
678	}
679	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
680	/*
681	 * Don't build a new relationship if we are locking Giant just
682	 * after waking up and the previous lock in the list was acquired
683	 * prior to blocking.
684	 */
685	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
686		mtx_unlock_spin(&w_mtx);
687	else {
688		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
689		    lock->lo_type, lock1->li_lock->lo_type);
690		if (!itismychild(lock1->li_lock->lo_witness, w))
691			mtx_unlock_spin(&w_mtx);
692	}
693
694out:
695#ifdef DDB
696	if (witness_ddb && go_into_ddb)
697		Debugger(__func__);
698#endif /* DDB */
699	w->w_file = file;
700	w->w_line = line;
701
702	lle = *lock_list;
703	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
704		lle = witness_lock_list_get();
705		if (lle == NULL)
706			return;
707		lle->ll_next = *lock_list;
708		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
709		    td->td_proc->p_pid, lle);
710		*lock_list = lle;
711	}
712	lock1 = &lle->ll_children[lle->ll_count++];
713	lock1->li_lock = lock;
714	lock1->li_line = line;
715	lock1->li_file = file;
716	if ((flags & LOP_EXCLUSIVE) != 0)
717		lock1->li_flags = LI_EXCLUSIVE;
718	else
719		lock1->li_flags = 0;
720	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
721	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
722}
723
724void
725witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
726{
727	struct lock_instance *instance;
728	struct lock_class *class;
729
730	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
731	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
732		return;
733	class = lock->lo_class;
734	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
735		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
736		    class->lc_name, lock->lo_name, file, line);
737	if ((flags & LOP_TRYLOCK) == 0)
738		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
739		    lock->lo_name, file, line);
740	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
741		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
742		    class->lc_name, lock->lo_name, file, line);
743	instance = find_instance(curthread->td_sleeplocks, lock);
744	if (instance == NULL)
745		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
746		    class->lc_name, lock->lo_name, file, line);
747	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
748		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
749		    class->lc_name, lock->lo_name, file, line);
750	if ((instance->li_flags & LI_RECURSEMASK) != 0)
751		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
752		    class->lc_name, lock->lo_name,
753		    instance->li_flags & LI_RECURSEMASK, file, line);
754	instance->li_flags |= LI_EXCLUSIVE;
755}
756
757void
758witness_downgrade(struct lock_object *lock, int flags, const char *file,
759    int line)
760{
761	struct lock_instance *instance;
762	struct lock_class *class;
763
764	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
765	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
766		return;
767	class = lock->lo_class;
768	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
769		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
770		    class->lc_name, lock->lo_name, file, line);
771	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
772		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
773		    class->lc_name, lock->lo_name, file, line);
774	instance = find_instance(curthread->td_sleeplocks, lock);
775	if (instance == NULL)
776		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
777		    class->lc_name, lock->lo_name, file, line);
778	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
779		panic("downgrade of shared lock (%s) %s @ %s:%d",
780		    class->lc_name, lock->lo_name, file, line);
781	if ((instance->li_flags & LI_RECURSEMASK) != 0)
782		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
783		    class->lc_name, lock->lo_name,
784		    instance->li_flags & LI_RECURSEMASK, file, line);
785	instance->li_flags &= ~LI_EXCLUSIVE;
786}
787
788void
789witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
790{
791	struct lock_list_entry **lock_list, *lle;
792	struct lock_instance *instance;
793	struct lock_class *class;
794	struct thread *td;
795	register_t s;
796	int i, j;
797
798	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
799	    panicstr != NULL)
800		return;
801	td = curthread;
802	class = lock->lo_class;
803	if (class->lc_flags & LC_SLEEPLOCK)
804		lock_list = &td->td_sleeplocks;
805	else
806		lock_list = PCPU_PTR(spinlocks);
807	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
808		for (i = 0; i < (*lock_list)->ll_count; i++) {
809			instance = &(*lock_list)->ll_children[i];
810			if (instance->li_lock == lock) {
811				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
812				    (flags & LOP_EXCLUSIVE) == 0) {
813					printf(
814					"shared unlock of (%s) %s @ %s:%d\n",
815					    class->lc_name, lock->lo_name,
816					    file, line);
817					printf(
818					"while exclusively locked from %s:%d\n",
819					    instance->li_file,
820					    instance->li_line);
821					panic("excl->ushare");
822				}
823				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
824				    (flags & LOP_EXCLUSIVE) != 0) {
825					printf(
826					"exclusive unlock of (%s) %s @ %s:%d\n",
827					    class->lc_name, lock->lo_name,
828					    file, line);
829					printf(
830					"while share locked from %s:%d\n",
831					    instance->li_file,
832					    instance->li_line);
833					panic("share->uexcl");
834				}
835				/* If we are recursed, unrecurse. */
836				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
837					CTR4(KTR_WITNESS,
838				    "%s: pid %d unrecursed on %s r=%d", __func__,
839					    td->td_proc->p_pid,
840					    instance->li_lock->lo_name,
841					    instance->li_flags);
842					instance->li_flags--;
843					return;
844				}
845				s = intr_disable();
846				CTR4(KTR_WITNESS,
847				    "%s: pid %d removed %s from lle[%d]", __func__,
848				    td->td_proc->p_pid,
849				    instance->li_lock->lo_name,
850				    (*lock_list)->ll_count - 1);
851				(*lock_list)->ll_count--;
852				for (j = i; j < (*lock_list)->ll_count; j++)
853					(*lock_list)->ll_children[j] =
854					    (*lock_list)->ll_children[j + 1];
855				intr_restore(s);
856				if ((*lock_list)->ll_count == 0) {
857					lle = *lock_list;
858					*lock_list = lle->ll_next;
859					CTR3(KTR_WITNESS,
860					    "%s: pid %d removed lle %p", __func__,
861					    td->td_proc->p_pid, lle);
862					witness_lock_list_free(lle);
863				}
864				return;
865			}
866		}
867	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
868	    file, line);
869}
870
871/*
872 * Warn if any held locks are not sleepable.  Note that Giant and the lock
873 * passed in are both special cases since they are both released during the
874 * sleep process and aren't actually held while the thread is asleep.
875 */
876int
877witness_sleep(int check_only, struct lock_object *lock, const char *file,
878	      int line)
879{
880	struct lock_list_entry **lock_list, *lle;
881	struct lock_instance *lock1;
882	struct thread *td;
883	int i, n;
884
885	if (witness_dead || panicstr != NULL)
886		return (0);
887	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
888	n = 0;
889	/*
890	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
891	 */
892	critical_enter();
893	td = curthread;
894	lock_list = &td->td_sleeplocks;
895again:
896	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
897		for (i = lle->ll_count - 1; i >= 0; i--) {
898			lock1 = &lle->ll_children[i];
899			if (lock1->li_lock == lock ||
900			    lock1->li_lock == &Giant.mtx_object)
901				continue;
902			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
903				if (check_only == 0) {
904					CTR3(KTR_WITNESS,
905				    "pid %d: sleeping with lock (%s) %s held",
906					    td->td_proc->p_pid,
907					    lock1->li_lock->lo_class->lc_name,
908					    lock1->li_lock->lo_name);
909					lock1->li_flags |= LI_SLEPT;
910				}
911				continue;
912			}
913			n++;
914			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
915			    file, line, check_only ? "could sleep" : "sleeping",
916			    lock1->li_lock->lo_name, lock1->li_file,
917			    lock1->li_line);
918		}
919	if (lock_list == &td->td_sleeplocks) {
920		lock_list = PCPU_PTR(spinlocks);
921		goto again;
922	}
923#ifdef DDB
924	if (witness_ddb && n)
925		Debugger(__func__);
926#endif /* DDB */
927	critical_exit();
928	return (n);
929}
930
931static struct witness *
932enroll(const char *description, struct lock_class *lock_class)
933{
934	struct witness *w;
935
936	if (!witness_watch || witness_dead || panicstr != NULL)
937		return (NULL);
938	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
939		return (NULL);
940	mtx_lock_spin(&w_mtx);
941	STAILQ_FOREACH(w, &w_all, w_list) {
942		if (strcmp(description, w->w_name) == 0) {
943			w->w_refcount++;
944			mtx_unlock_spin(&w_mtx);
945			if (lock_class != w->w_class)
946				panic(
947				"lock (%s) %s does not match earlier (%s) lock",
948				    description, lock_class->lc_name,
949				    w->w_class->lc_name);
950			return (w);
951		}
952	}
953	/*
954	 * This isn't quite right, as witness_cold is still 0 while we
955	 * enroll all the locks initialized before witness_initialize().
956	 */
957	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
958		mtx_unlock_spin(&w_mtx);
959		panic("spin lock %s not in order list", description);
960	}
961	if ((w = witness_get()) == NULL)
962		return (NULL);
963	w->w_name = description;
964	w->w_class = lock_class;
965	w->w_refcount = 1;
966	STAILQ_INSERT_HEAD(&w_all, w, w_list);
967	if (lock_class->lc_flags & LC_SPINLOCK)
968		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
969	else if (lock_class->lc_flags & LC_SLEEPLOCK)
970		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
971	else {
972		mtx_unlock_spin(&w_mtx);
973		panic("lock class %s is not sleep or spin",
974		    lock_class->lc_name);
975	}
976	mtx_unlock_spin(&w_mtx);
977	return (w);
978}
979
980static int
981itismychild(struct witness *parent, struct witness *child)
982{
983	static int recursed;
984	struct witness_child_list_entry **wcl;
985	struct witness_list *list;
986
987	MPASS(child != NULL && parent != NULL);
988	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
989	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
990		panic(
991		"%s: parent (%s) and child (%s) are not the same lock type",
992		    __func__, parent->w_class->lc_name,
993		    child->w_class->lc_name);
994
995	/*
996	 * Insert "child" after "parent"
997	 */
998	wcl = &parent->w_children;
999	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1000		wcl = &(*wcl)->wcl_next;
1001	if (*wcl == NULL) {
1002		*wcl = witness_child_get();
1003		if (*wcl == NULL)
1004			return (1);
1005	}
1006	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1007
1008	/*
1009	 * Now prune whole tree.  We look for cases where a lock is now
1010	 * both a descendant and a direct child of a given lock.  In that
1011	 * case, we want to remove the direct child link from the tree.
1012	 */
1013	if (recursed)
1014		return (0);
1015	recursed = 1;
1016	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1017		list = &w_sleep;
1018	else
1019		list = &w_spin;
1020	STAILQ_FOREACH(child, list, w_typelist) {
1021		STAILQ_FOREACH(parent, list, w_typelist) {
1022			if (!isitmychild(parent, child))
1023				continue;
1024			removechild(parent, child);
1025			if (isitmydescendant(parent, child))
1026				continue;
1027			itismychild(parent, child);
1028		}
1029	}
1030	recursed = 0;
1031	witness_levelall();
1032	return (0);
1033}
1034
1035static void
1036removechild(struct witness *parent, struct witness *child)
1037{
1038	struct witness_child_list_entry **wcl, *wcl1;
1039	int i;
1040
1041	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1042		for (i = 0; i < (*wcl)->wcl_count; i++)
1043			if ((*wcl)->wcl_children[i] == child)
1044				goto found;
1045	return;
1046found:
1047	(*wcl)->wcl_count--;
1048	if ((*wcl)->wcl_count > i)
1049		(*wcl)->wcl_children[i] =
1050		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1051	MPASS((*wcl)->wcl_children[i] != NULL);
1052	if ((*wcl)->wcl_count != 0)
1053		return;
1054	wcl1 = *wcl;
1055	*wcl = wcl1->wcl_next;
1056	witness_child_free(wcl1);
1057}
1058
1059static int
1060isitmychild(struct witness *parent, struct witness *child)
1061{
1062	struct witness_child_list_entry *wcl;
1063	int i;
1064
1065	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1066		for (i = 0; i < wcl->wcl_count; i++) {
1067			if (wcl->wcl_children[i] == child)
1068				return (1);
1069		}
1070	}
1071	return (0);
1072}
1073
1074static int
1075isitmydescendant(struct witness *parent, struct witness *child)
1076{
1077	struct witness_child_list_entry *wcl;
1078	int i, j;
1079
1080	if (isitmychild(parent, child))
1081		return (1);
1082	j = 0;
1083	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1084		MPASS(j < 1000);
1085		for (i = 0; i < wcl->wcl_count; i++) {
1086			if (isitmydescendant(wcl->wcl_children[i], child))
1087				return (1);
1088		}
1089		j++;
1090	}
1091	return (0);
1092}
1093
1094void
1095witness_levelall (void)
1096{
1097	struct witness_list *list;
1098	struct witness *w, *w1;
1099
1100	/*
1101	 * First clear all levels.
1102	 */
1103	STAILQ_FOREACH(w, &w_all, w_list) {
1104		w->w_level = 0;
1105	}
1106
1107	/*
1108	 * Look for locks with no parent and level all their descendants.
1109	 */
1110	STAILQ_FOREACH(w, &w_all, w_list) {
1111		/*
1112		 * This is just an optimization, technically we could get
1113		 * away just walking the all list each time.
1114		 */
1115		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1116			list = &w_sleep;
1117		else
1118			list = &w_spin;
1119		STAILQ_FOREACH(w1, list, w_typelist) {
1120			if (isitmychild(w1, w))
1121				goto skip;
1122		}
1123		witness_leveldescendents(w, 0);
1124	skip:
1125		;	/* silence GCC 3.x */
1126	}
1127}
1128
1129static void
1130witness_leveldescendents(struct witness *parent, int level)
1131{
1132	struct witness_child_list_entry *wcl;
1133	int i;
1134
1135	if (parent->w_level < level)
1136		parent->w_level = level;
1137	level++;
1138	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1139		for (i = 0; i < wcl->wcl_count; i++)
1140			witness_leveldescendents(wcl->wcl_children[i], level);
1141}
1142
1143static void
1144witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1145			   struct witness *parent)
1146{
1147	struct witness_child_list_entry *wcl;
1148	int i, level;
1149
1150	level = parent->w_level;
1151	prnt("%-2d", level);
1152	for (i = 0; i < level; i++)
1153		prnt(" ");
1154	prnt("%s", parent->w_name);
1155	if (parent->w_file != NULL)
1156		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1157		    parent->w_line);
1158	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1159		for (i = 0; i < wcl->wcl_count; i++)
1160			    witness_displaydescendants(prnt,
1161				wcl->wcl_children[i]);
1162}
1163
1164static int
1165blessed(struct witness *w1, struct witness *w2)
1166{
1167	int i;
1168	struct witness_blessed *b;
1169
1170	for (i = 0; i < blessed_count; i++) {
1171		b = &blessed_list[i];
1172		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1173			if (strcmp(w2->w_name, b->b_lock2) == 0)
1174				return (1);
1175			continue;
1176		}
1177		if (strcmp(w1->w_name, b->b_lock2) == 0)
1178			if (strcmp(w2->w_name, b->b_lock1) == 0)
1179				return (1);
1180	}
1181	return (0);
1182}
1183
1184static struct witness *
1185witness_get(void)
1186{
1187	struct witness *w;
1188
1189	if (witness_dead) {
1190		mtx_unlock_spin(&w_mtx);
1191		return (NULL);
1192	}
1193	if (STAILQ_EMPTY(&w_free)) {
1194		witness_dead = 1;
1195		mtx_unlock_spin(&w_mtx);
1196		printf("%s: witness exhausted\n", __func__);
1197		return (NULL);
1198	}
1199	w = STAILQ_FIRST(&w_free);
1200	STAILQ_REMOVE_HEAD(&w_free, w_list);
1201	bzero(w, sizeof(*w));
1202	return (w);
1203}
1204
1205static void
1206witness_free(struct witness *w)
1207{
1208
1209	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1210}
1211
1212static struct witness_child_list_entry *
1213witness_child_get(void)
1214{
1215	struct witness_child_list_entry *wcl;
1216
1217	if (witness_dead) {
1218		mtx_unlock_spin(&w_mtx);
1219		return (NULL);
1220	}
1221	wcl = w_child_free;
1222	if (wcl == NULL) {
1223		witness_dead = 1;
1224		mtx_unlock_spin(&w_mtx);
1225		printf("%s: witness exhausted\n", __func__);
1226		return (NULL);
1227	}
1228	w_child_free = wcl->wcl_next;
1229	bzero(wcl, sizeof(*wcl));
1230	return (wcl);
1231}
1232
1233static void
1234witness_child_free(struct witness_child_list_entry *wcl)
1235{
1236
1237	wcl->wcl_next = w_child_free;
1238	w_child_free = wcl;
1239}
1240
1241static struct lock_list_entry *
1242witness_lock_list_get(void)
1243{
1244	struct lock_list_entry *lle;
1245
1246	if (witness_dead)
1247		return (NULL);
1248	mtx_lock_spin(&w_mtx);
1249	lle = w_lock_list_free;
1250	if (lle == NULL) {
1251		witness_dead = 1;
1252		mtx_unlock_spin(&w_mtx);
1253		printf("%s: witness exhausted\n", __func__);
1254		return (NULL);
1255	}
1256	w_lock_list_free = lle->ll_next;
1257	mtx_unlock_spin(&w_mtx);
1258	bzero(lle, sizeof(*lle));
1259	return (lle);
1260}
1261
1262static void
1263witness_lock_list_free(struct lock_list_entry *lle)
1264{
1265
1266	mtx_lock_spin(&w_mtx);
1267	lle->ll_next = w_lock_list_free;
1268	w_lock_list_free = lle;
1269	mtx_unlock_spin(&w_mtx);
1270}
1271
1272static struct lock_instance *
1273find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1274{
1275	struct lock_list_entry *lle;
1276	struct lock_instance *instance;
1277	int i;
1278
1279	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1280		for (i = lle->ll_count - 1; i >= 0; i--) {
1281			instance = &lle->ll_children[i];
1282			if (instance->li_lock == lock)
1283				return (instance);
1284		}
1285	return (NULL);
1286}
1287
1288int
1289witness_list_locks(struct lock_list_entry **lock_list)
1290{
1291	struct lock_list_entry *lle;
1292	struct lock_instance *instance;
1293	struct lock_object *lock;
1294	int i, nheld;
1295
1296	nheld = 0;
1297	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1298		for (i = lle->ll_count - 1; i >= 0; i--) {
1299			instance = &lle->ll_children[i];
1300			lock = instance->li_lock;
1301			printf("%s %s %s",
1302			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1303			    "exclusive" : "shared",
1304			    lock->lo_class->lc_name, lock->lo_name);
1305			if (lock->lo_type != lock->lo_name)
1306				printf(" (%s)", lock->lo_type);
1307			printf(" r = %d (%p) locked @ %s:%d\n",
1308			    instance->li_flags & LI_RECURSEMASK, lock,
1309			    instance->li_file, instance->li_line);
1310			nheld++;
1311		}
1312	return (nheld);
1313}
1314
1315/*
1316 * Calling this on td != curthread is bad unless we are in ddb.
1317 */
1318int
1319witness_list(struct thread *td)
1320{
1321	int nheld;
1322
1323	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1324#ifdef DDB
1325	KASSERT(td == curthread || db_active,
1326	    ("%s: td != curthread and we aren't in the debugger", __func__));
1327	if (!db_active && witness_dead)
1328		return (0);
1329#else
1330	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1331	if (witness_dead)
1332		return (0);
1333#endif
1334	nheld = witness_list_locks(&td->td_sleeplocks);
1335
1336	/*
1337	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1338	 * if td is currently executing on some other CPU and holds spin locks
1339	 * as we won't display those locks.  If we had a MI way of getting
1340	 * the per-cpu data for a given cpu then we could use
1341	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1342	 * and "fix" this.
1343	 *
1344	 * That still wouldn't really fix this unless we locked sched_lock
1345	 * or stopped the other CPU to make sure it wasn't changing the list
1346	 * out from under us.  It is probably best to just not try to handle
1347	 * threads on other CPU's for now.
1348	 */
1349	if (td == curthread) {
1350		/*
1351		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1352		 * change.
1353		 */
1354		critical_enter();
1355		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1356		critical_exit();
1357	}
1358	return (nheld);
1359}
1360
1361void
1362witness_save(struct lock_object *lock, const char **filep, int *linep)
1363{
1364	struct lock_instance *instance;
1365
1366	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1367	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1368		return;
1369	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1370		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1371		    lock->lo_class->lc_name, lock->lo_name);
1372	instance = find_instance(curthread->td_sleeplocks, lock);
1373	if (instance == NULL)
1374		panic("%s: lock (%s) %s not locked", __func__,
1375		    lock->lo_class->lc_name, lock->lo_name);
1376	*filep = instance->li_file;
1377	*linep = instance->li_line;
1378}
1379
1380void
1381witness_restore(struct lock_object *lock, const char *file, int line)
1382{
1383	struct lock_instance *instance;
1384
1385	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1386	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1387		return;
1388	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1389		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1390		    lock->lo_class->lc_name, lock->lo_name);
1391	instance = find_instance(curthread->td_sleeplocks, lock);
1392	if (instance == NULL)
1393		panic("%s: lock (%s) %s not locked", __func__,
1394		    lock->lo_class->lc_name, lock->lo_name);
1395	lock->lo_witness->w_file = file;
1396	lock->lo_witness->w_line = line;
1397	instance->li_file = file;
1398	instance->li_line = line;
1399}
1400
1401void
1402witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1403{
1404#ifdef INVARIANT_SUPPORT
1405	struct lock_instance *instance;
1406
1407	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1408		return;
1409	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1410		instance = find_instance(curthread->td_sleeplocks, lock);
1411	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1412		instance = find_instance(PCPU_GET(spinlocks), lock);
1413	else {
1414		panic("Lock (%s) %s is not sleep or spin!",
1415		    lock->lo_class->lc_name, lock->lo_name);
1416		return;
1417	}
1418	switch (flags) {
1419	case LA_UNLOCKED:
1420		if (instance != NULL)
1421			panic("Lock (%s) %s locked @ %s:%d.",
1422			    lock->lo_class->lc_name, lock->lo_name, file, line);
1423		break;
1424	case LA_LOCKED:
1425	case LA_LOCKED | LA_RECURSED:
1426	case LA_LOCKED | LA_NOTRECURSED:
1427	case LA_SLOCKED:
1428	case LA_SLOCKED | LA_RECURSED:
1429	case LA_SLOCKED | LA_NOTRECURSED:
1430	case LA_XLOCKED:
1431	case LA_XLOCKED | LA_RECURSED:
1432	case LA_XLOCKED | LA_NOTRECURSED:
1433		if (instance == NULL) {
1434			panic("Lock (%s) %s not locked @ %s:%d.",
1435			    lock->lo_class->lc_name, lock->lo_name, file, line);
1436			break;
1437		}
1438		if ((flags & LA_XLOCKED) != 0 &&
1439		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1440			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1441			    lock->lo_class->lc_name, lock->lo_name, file, line);
1442		if ((flags & LA_SLOCKED) != 0 &&
1443		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1444			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1445			    lock->lo_class->lc_name, lock->lo_name, file, line);
1446		if ((flags & LA_RECURSED) != 0 &&
1447		    (instance->li_flags & LI_RECURSEMASK) == 0)
1448			panic("Lock (%s) %s not recursed @ %s:%d.",
1449			    lock->lo_class->lc_name, lock->lo_name, file, line);
1450		if ((flags & LA_NOTRECURSED) != 0 &&
1451		    (instance->li_flags & LI_RECURSEMASK) != 0)
1452			panic("Lock (%s) %s recursed @ %s:%d.",
1453			    lock->lo_class->lc_name, lock->lo_name, file, line);
1454		break;
1455	default:
1456		panic("Invalid lock assertion at %s:%d.", file, line);
1457
1458	}
1459#endif	/* INVARIANT_SUPPORT */
1460}
1461
1462#ifdef DDB
1463
1464DB_SHOW_COMMAND(locks, db_witness_list)
1465{
1466	struct thread *td;
1467	pid_t pid;
1468	struct proc *p;
1469
1470	if (have_addr) {
1471		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1472		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1473		    ((addr >> 16) % 16) * 10000;
1474		/* sx_slock(&allproc_lock); */
1475		FOREACH_PROC_IN_SYSTEM(p) {
1476			if (p->p_pid == pid)
1477				break;
1478		}
1479		/* sx_sunlock(&allproc_lock); */
1480		if (p == NULL) {
1481			db_printf("pid %d not found\n", pid);
1482			return;
1483		}
1484		FOREACH_THREAD_IN_PROC(p, td) {
1485			witness_list(td);
1486		}
1487	} else {
1488		td = curthread;
1489		witness_list(td);
1490	}
1491}
1492
1493DB_SHOW_COMMAND(witness, db_witness_display)
1494{
1495
1496	witness_display(db_printf);
1497}
1498#endif
1499