subr_witness.c revision 93690
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 93690 2002-04-02 19:27:21Z 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	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	{ "pgrpsess", &lock_class_sx },
191	{ "proctree", &lock_class_sx },
192	{ "allproc", &lock_class_sx },
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	{ "icu", &lock_class_mtx_spin },
222#ifdef SMP
223	{ "smp rendezvous", &lock_class_mtx_spin },
224#endif
225	{ "clk", &lock_class_mtx_spin },
226	{ NULL, NULL },
227	{ NULL, NULL }
228};
229
230/*
231 * Pairs of locks which have been blessed
232 * Don't complain about order problems with blessed locks
233 */
234static struct witness_blessed blessed_list[] = {
235};
236static int blessed_count =
237	sizeof(blessed_list) / sizeof(struct witness_blessed);
238
239/*
240 * List of all locks in the system.
241 */
242STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
243
244static struct mtx all_mtx = {
245	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
246	  "All locks list",		/* mtx_object.lo_name */
247	  LO_INITIALIZED,		/* mtx_object.lo_flags */
248	  { NULL },			/* mtx_object.lo_list */
249	  NULL },			/* mtx_object.lo_witness */
250	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
251	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
252	{ NULL, NULL }			/* mtx_contested */
253};
254
255/*
256 * This global is set to 0 once it becomes safe to use the witness code.
257 */
258static int witness_cold = 1;
259
260/*
261 * Global variables for book keeping.
262 */
263static int lock_cur_cnt;
264static int lock_max_cnt;
265
266/*
267 * The WITNESS-enabled diagnostic code.
268 */
269static void
270witness_initialize(void *dummy __unused)
271{
272	struct lock_object *lock;
273	struct witness_order_list_entry *order;
274	struct witness *w, *w1;
275	int i;
276
277	/*
278	 * We have to release Giant before initializing its witness
279	 * structure so that WITNESS doesn't get confused.
280	 */
281	mtx_unlock(&Giant);
282	mtx_assert(&Giant, MA_NOTOWNED);
283
284	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
285	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
286	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
287	for (i = 0; i < WITNESS_COUNT; i++)
288		witness_free(&w_data[i]);
289	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
290		witness_child_free(&w_childdata[i]);
291	for (i = 0; i < LOCK_CHILDCOUNT; i++)
292		witness_lock_list_free(&w_locklistdata[i]);
293
294	/* First add in all the specified order lists. */
295	for (order = order_lists; order->w_name != NULL; order++) {
296		w = enroll(order->w_name, order->w_class);
297		if (w == NULL)
298			continue;
299		w->w_file = "order list";
300		for (order++; order->w_name != NULL; order++) {
301			w1 = enroll(order->w_name, order->w_class);
302			if (w1 == NULL)
303				continue;
304			w1->w_file = "order list";
305			itismychild(w, w1);
306			w = w1;
307		}
308	}
309
310	/* Iterate through all locks and add them to witness. */
311	mtx_lock(&all_mtx);
312	STAILQ_FOREACH(lock, &all_locks, lo_list) {
313		if (lock->lo_flags & LO_WITNESS)
314			lock->lo_witness = enroll(lock->lo_name,
315			    lock->lo_class);
316		else
317			lock->lo_witness = NULL;
318	}
319	mtx_unlock(&all_mtx);
320
321	/* Mark the witness code as being ready for use. */
322	atomic_store_rel_int(&witness_cold, 0);
323
324	mtx_lock(&Giant);
325}
326SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
327
328void
329witness_init(struct lock_object *lock)
330{
331	struct lock_class *class;
332
333	class = lock->lo_class;
334	if (lock->lo_flags & LO_INITIALIZED)
335		panic("%s: lock (%s) %s is already initialized", __func__,
336		    class->lc_name, lock->lo_name);
337	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
338	    (class->lc_flags & LC_RECURSABLE) == 0)
339		panic("%s: lock (%s) %s can not be recursable", __func__,
340		    class->lc_name, lock->lo_name);
341	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
342	    (class->lc_flags & LC_SLEEPABLE) == 0)
343		panic("%s: lock (%s) %s can not be sleepable", __func__,
344		    class->lc_name, lock->lo_name);
345	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
346	    (class->lc_flags & LC_UPGRADABLE) == 0)
347		panic("%s: lock (%s) %s can not be upgradable", __func__,
348		    class->lc_name, lock->lo_name);
349
350	mtx_lock(&all_mtx);
351	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
352	lock->lo_flags |= LO_INITIALIZED;
353	lock_cur_cnt++;
354	if (lock_cur_cnt > lock_max_cnt)
355		lock_max_cnt = lock_cur_cnt;
356	mtx_unlock(&all_mtx);
357	if (!witness_cold && !witness_dead && panicstr == NULL &&
358	    (lock->lo_flags & LO_WITNESS) != 0)
359		lock->lo_witness = enroll(lock->lo_name, class);
360	else
361		lock->lo_witness = NULL;
362}
363
364void
365witness_destroy(struct lock_object *lock)
366{
367	struct witness *w;
368
369	if (witness_cold)
370		panic("lock (%s) %s destroyed while witness_cold",
371		    lock->lo_class->lc_name, lock->lo_name);
372	if ((lock->lo_flags & LO_INITIALIZED) == 0)
373		panic("%s: lock (%s) %s is not initialized", __func__,
374		    lock->lo_class->lc_name, lock->lo_name);
375
376	/* XXX: need to verify that no one holds the lock */
377	w = lock->lo_witness;
378	if (w != NULL) {
379		mtx_lock_spin(&w_mtx);
380		w->w_refcount--;
381		if (w->w_refcount == 0) {
382			CTR2(KTR_WITNESS,
383			    "%s: marking witness %s as dead", __func__, w->w_name);
384			w->w_name = "(dead)";
385			w->w_file = "(dead)";
386			w->w_line = 0;
387		}
388		mtx_unlock_spin(&w_mtx);
389	}
390
391	mtx_lock(&all_mtx);
392	lock_cur_cnt--;
393	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
394	lock->lo_flags &= ~LO_INITIALIZED;
395	mtx_unlock(&all_mtx);
396}
397
398static void
399witness_display_list(void(*prnt)(const char *fmt, ...),
400		     struct witness_list *list)
401{
402	struct witness *w, *w1;
403	int found;
404
405	STAILQ_FOREACH(w, list, w_typelist) {
406		if (w->w_file == NULL)
407			continue;
408		found = 0;
409		STAILQ_FOREACH(w1, list, w_typelist) {
410			if (isitmychild(w1, w)) {
411				found++;
412				break;
413			}
414		}
415		if (found)
416			continue;
417		/*
418		 * This lock has no anscestors, display its descendants.
419		 */
420		witness_displaydescendants(prnt, w);
421	}
422}
423
424static void
425witness_display(void(*prnt)(const char *fmt, ...))
426{
427	struct witness *w;
428
429	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
430	witness_levelall();
431
432	/*
433	 * First, handle sleep locks which have been acquired at least
434	 * once.
435	 */
436	prnt("Sleep locks:\n");
437	witness_display_list(prnt, &w_sleep);
438
439	/*
440	 * Now do spin locks which have been acquired at least once.
441	 */
442	prnt("\nSpin locks:\n");
443	witness_display_list(prnt, &w_spin);
444
445	/*
446	 * Finally, any locks which have not been acquired yet.
447	 */
448	prnt("\nLocks which were never acquired:\n");
449	STAILQ_FOREACH(w, &w_all, w_list) {
450		if (w->w_file != NULL)
451			continue;
452		prnt("%s\n", w->w_name);
453	}
454}
455
456void
457witness_lock(struct lock_object *lock, int flags, const char *file, int line)
458{
459	struct lock_list_entry **lock_list, *lle;
460	struct lock_instance *lock1, *lock2;
461	struct lock_class *class;
462	struct witness *w, *w1;
463	struct thread *td;
464	int i, j;
465#ifdef DDB
466	int go_into_ddb = 0;
467#endif /* DDB */
468
469	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
470	    panicstr != NULL)
471		return;
472	w = lock->lo_witness;
473	class = lock->lo_class;
474	td = curthread;
475
476	if (class->lc_flags & LC_SLEEPLOCK) {
477		/*
478		 * Since spin locks include a critical section, this check
479		 * impliclty enforces a lock order of all sleep locks before
480		 * all spin locks.
481		 */
482		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
483			panic("blockable sleep lock (%s) %s @ %s:%d",
484			    class->lc_name, lock->lo_name, file, line);
485		lock_list = &td->td_sleeplocks;
486	} else
487		lock_list = PCPU_PTR(spinlocks);
488
489	/*
490	 * Try locks do not block if they fail to acquire the lock, thus
491	 * there is no danger of deadlocks or of switching while holding a
492	 * spin lock if we acquire a lock via a try operation.
493	 */
494	if (flags & LOP_TRYLOCK)
495		goto out;
496
497	/*
498	 * Is this the first lock acquired?  If so, then no order checking
499	 * is needed.
500	 */
501	if (*lock_list == NULL)
502		goto out;
503
504	/*
505	 * Check to see if we are recursing on a lock we already own.
506	 */
507	lock1 = find_instance(*lock_list, lock);
508	if (lock1 != NULL) {
509		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
510		    (flags & LOP_EXCLUSIVE) == 0) {
511			printf("shared lock of (%s) %s @ %s:%d\n",
512			    class->lc_name, lock->lo_name, file, line);
513			printf("while exclusively locked from %s:%d\n",
514			    lock1->li_file, lock1->li_line);
515			panic("share->excl");
516		}
517		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
518		    (flags & LOP_EXCLUSIVE) != 0) {
519			printf("exclusive lock of (%s) %s @ %s:%d\n",
520			    class->lc_name, lock->lo_name, file, line);
521			printf("while share locked from %s:%d\n",
522			    lock1->li_file, lock1->li_line);
523			panic("excl->share");
524		}
525		lock1->li_flags++;
526		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
527			printf(
528			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
529			    class->lc_name, lock->lo_name, file, line);
530			printf("first acquired @ %s:%d\n", lock1->li_file,
531			    lock1->li_line);
532			panic("recurse");
533		}
534		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
535		    td->td_proc->p_pid, lock->lo_name,
536		    lock1->li_flags & LI_RECURSEMASK);
537		lock1->li_file = file;
538		lock1->li_line = line;
539		return;
540	}
541
542	/*
543	 * Check for duplicate locks of the same type.  Note that we only
544	 * have to check for this on the last lock we just acquired.  Any
545	 * other cases will be caught as lock order violations.
546	 */
547	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
548	w1 = lock1->li_lock->lo_witness;
549	if (w1 == w) {
550		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
551			goto out;
552		w->w_same_squawked = 1;
553		printf("acquiring duplicate lock of same type: \"%s\"\n",
554			lock->lo_name);
555		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
556		printf(" 2nd @ %s:%d\n", file, line);
557#ifdef DDB
558		go_into_ddb = 1;
559#endif /* DDB */
560		goto out;
561	}
562	MPASS(!mtx_owned(&w_mtx));
563	mtx_lock_spin(&w_mtx);
564	/*
565	 * If we have a known higher number just say ok
566	 */
567	if (witness_watch > 1 && w->w_level > w1->w_level) {
568		mtx_unlock_spin(&w_mtx);
569		goto out;
570	}
571	if (isitmydescendant(w1, w)) {
572		mtx_unlock_spin(&w_mtx);
573		goto out;
574	}
575	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
576		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
577
578			MPASS(j < WITNESS_COUNT);
579			lock1 = &lle->ll_children[i];
580			w1 = lock1->li_lock->lo_witness;
581
582			/*
583			 * If this lock doesn't undergo witness checking,
584			 * then skip it.
585			 */
586			if (w1 == NULL) {
587				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
588				    ("lock missing witness structure"));
589				continue;
590			}
591			/*
592			 * If we are locking Giant and we slept with this
593			 * lock, then skip it.
594			 */
595			if ((lock1->li_flags & LI_SLEPT) != 0 &&
596			    lock == &Giant.mtx_object)
597				continue;
598			/*
599			 * If we are locking a sleepable lock and this lock
600			 * isn't sleepable and isn't Giant, we want to treat
601			 * it as a lock order violation to enfore a general
602			 * lock order of sleepable locks before non-sleepable
603			 * locks.  Thus, we only bother checking the lock
604			 * order hierarchy if we pass the initial test.
605			 */
606			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
607			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
608			    lock1->li_lock != &Giant.mtx_object)) &&
609			    !isitmydescendant(w, w1))
610				continue;
611			/*
612			 * We have a lock order violation, check to see if it
613			 * is allowed or has already been yelled about.
614			 */
615			mtx_unlock_spin(&w_mtx);
616			if (blessed(w, w1))
617				goto out;
618			if (lock1->li_lock == &Giant.mtx_object) {
619				if (w1->w_Giant_squawked)
620					goto out;
621				else
622					w1->w_Giant_squawked = 1;
623			} else {
624				if (w1->w_other_squawked)
625					goto out;
626				else
627					w1->w_other_squawked = 1;
628			}
629			/*
630			 * Ok, yell about it.
631			 */
632			printf("lock order reversal\n");
633			/*
634			 * Try to locate an earlier lock with
635			 * witness w in our list.
636			 */
637			do {
638				lock2 = &lle->ll_children[i];
639				MPASS(lock2->li_lock != NULL);
640				if (lock2->li_lock->lo_witness == w)
641					break;
642				i--;
643				if (i == 0 && lle->ll_next != NULL) {
644					lle = lle->ll_next;
645					i = lle->ll_count - 1;
646					MPASS(i != 0);
647				}
648			} while (i >= 0);
649			if (i < 0) {
650				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
651				    lock1->li_lock->lo_name, lock1->li_file,
652				    lock1->li_line);
653				printf(" 2nd %p %s @ %s:%d\n", lock,
654				    lock->lo_name, file, line);
655			} else {
656				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
657				    lock2->li_lock->lo_name, lock2->li_file,
658				    lock2->li_line);
659				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
660				    lock1->li_lock->lo_name, lock1->li_file,
661				    lock1->li_line);
662				printf(" 3rd %p %s @ %s:%d\n", lock,
663				    lock->lo_name, file, line);
664			}
665#ifdef DDB
666			go_into_ddb = 1;
667#endif /* DDB */
668			goto out;
669		}
670	}
671	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
672	/*
673	 * Don't build a new relationship if we are locking Giant just
674	 * after waking up and the previous lock in the list was acquired
675	 * prior to blocking.
676	 */
677	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
678		mtx_unlock_spin(&w_mtx);
679	else {
680		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
681		    lock->lo_name, lock1->li_lock->lo_name);
682		if (!itismychild(lock1->li_lock->lo_witness, w))
683			mtx_unlock_spin(&w_mtx);
684	}
685
686out:
687#ifdef DDB
688	if (witness_ddb && go_into_ddb)
689		Debugger(__func__);
690#endif /* DDB */
691	w->w_file = file;
692	w->w_line = line;
693
694	lle = *lock_list;
695	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
696		lle = witness_lock_list_get();
697		if (lle == NULL)
698			return;
699		lle->ll_next = *lock_list;
700		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
701		    td->td_proc->p_pid, lle);
702		*lock_list = lle;
703	}
704	lock1 = &lle->ll_children[lle->ll_count++];
705	lock1->li_lock = lock;
706	lock1->li_line = line;
707	lock1->li_file = file;
708	if ((flags & LOP_EXCLUSIVE) != 0)
709		lock1->li_flags = LI_EXCLUSIVE;
710	else
711		lock1->li_flags = 0;
712	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
713	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
714}
715
716void
717witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
718{
719	struct lock_instance *instance;
720	struct lock_class *class;
721
722	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
723	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
724		return;
725	class = lock->lo_class;
726	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
727		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
728		    class->lc_name, lock->lo_name, file, line);
729	if ((flags & LOP_TRYLOCK) == 0)
730		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
731		    lock->lo_name, file, line);
732	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
733		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
734		    class->lc_name, lock->lo_name, file, line);
735	instance = find_instance(curthread->td_sleeplocks, lock);
736	if (instance == NULL)
737		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
738		    class->lc_name, lock->lo_name, file, line);
739	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
740		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
741		    class->lc_name, lock->lo_name, file, line);
742	if ((instance->li_flags & LI_RECURSEMASK) != 0)
743		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
744		    class->lc_name, lock->lo_name,
745		    instance->li_flags & LI_RECURSEMASK, file, line);
746	instance->li_flags |= LI_EXCLUSIVE;
747}
748
749void
750witness_downgrade(struct lock_object *lock, int flags, const char *file,
751    int line)
752{
753	struct lock_instance *instance;
754	struct lock_class *class;
755
756	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
757	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
758		return;
759	class = lock->lo_class;
760	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
761		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
762		    class->lc_name, lock->lo_name, file, line);
763	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
764		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
765		    class->lc_name, lock->lo_name, file, line);
766	instance = find_instance(curthread->td_sleeplocks, lock);
767	if (instance == NULL)
768		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
769		    class->lc_name, lock->lo_name, file, line);
770	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
771		panic("downgrade of shared lock (%s) %s @ %s:%d",
772		    class->lc_name, lock->lo_name, file, line);
773	if ((instance->li_flags & LI_RECURSEMASK) != 0)
774		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
775		    class->lc_name, lock->lo_name,
776		    instance->li_flags & LI_RECURSEMASK, file, line);
777	instance->li_flags &= ~LI_EXCLUSIVE;
778}
779
780void
781witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
782{
783	struct lock_list_entry **lock_list, *lle;
784	struct lock_instance *instance;
785	struct lock_class *class;
786	struct thread *td;
787	register_t s;
788	int i, j;
789
790	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
791	    panicstr != NULL)
792		return;
793	td = curthread;
794	class = lock->lo_class;
795	if (class->lc_flags & LC_SLEEPLOCK)
796		lock_list = &td->td_sleeplocks;
797	else
798		lock_list = PCPU_PTR(spinlocks);
799	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
800		for (i = 0; i < (*lock_list)->ll_count; i++) {
801			instance = &(*lock_list)->ll_children[i];
802			if (instance->li_lock == lock) {
803				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
804				    (flags & LOP_EXCLUSIVE) == 0) {
805					printf(
806					"shared unlock of (%s) %s @ %s:%d\n",
807					    class->lc_name, lock->lo_name,
808					    file, line);
809					printf(
810					"while exclusively locked from %s:%d\n",
811					    instance->li_file,
812					    instance->li_line);
813					panic("excl->ushare");
814				}
815				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
816				    (flags & LOP_EXCLUSIVE) != 0) {
817					printf(
818					"exclusive unlock of (%s) %s @ %s:%d\n",
819					    class->lc_name, lock->lo_name,
820					    file, line);
821					printf(
822					"while share locked from %s:%d\n",
823					    instance->li_file,
824					    instance->li_line);
825					panic("share->uexcl");
826				}
827				/* If we are recursed, unrecurse. */
828				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
829					CTR4(KTR_WITNESS,
830				    "%s: pid %d unrecursed on %s r=%d", __func__,
831					    td->td_proc->p_pid,
832					    instance->li_lock->lo_name,
833					    instance->li_flags);
834					instance->li_flags--;
835					return;
836				}
837				s = intr_disable();
838				CTR4(KTR_WITNESS,
839				    "%s: pid %d removed %s from lle[%d]", __func__,
840				    td->td_proc->p_pid,
841				    instance->li_lock->lo_name,
842				    (*lock_list)->ll_count - 1);
843				(*lock_list)->ll_count--;
844				for (j = i; j < (*lock_list)->ll_count; j++)
845					(*lock_list)->ll_children[j] =
846					    (*lock_list)->ll_children[j + 1];
847				intr_restore(s);
848				if ((*lock_list)->ll_count == 0) {
849					lle = *lock_list;
850					*lock_list = lle->ll_next;
851					CTR3(KTR_WITNESS,
852					    "%s: pid %d removed lle %p", __func__,
853					    td->td_proc->p_pid, lle);
854					witness_lock_list_free(lle);
855				}
856				return;
857			}
858		}
859	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
860	    file, line);
861}
862
863/*
864 * Warn if any held locks are not sleepable.  Note that Giant and the lock
865 * passed in are both special cases since they are both released during the
866 * sleep process and aren't actually held while the thread is asleep.
867 */
868int
869witness_sleep(int check_only, struct lock_object *lock, const char *file,
870	      int line)
871{
872	struct lock_list_entry **lock_list, *lle;
873	struct lock_instance *lock1;
874	struct thread *td;
875	int i, n;
876
877	if (witness_dead || panicstr != NULL)
878		return (0);
879	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
880	n = 0;
881	/*
882	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
883	 */
884	critical_enter();
885	td = curthread;
886	lock_list = &td->td_sleeplocks;
887again:
888	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
889		for (i = lle->ll_count - 1; i >= 0; i--) {
890			lock1 = &lle->ll_children[i];
891			if (lock1->li_lock == lock ||
892			    lock1->li_lock == &Giant.mtx_object)
893				continue;
894			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
895				if (check_only == 0) {
896					CTR3(KTR_WITNESS,
897				    "pid %d: sleeping with lock (%s) %s held",
898					    td->td_proc->p_pid,
899					    lock1->li_lock->lo_class->lc_name,
900					    lock1->li_lock->lo_name);
901					lock1->li_flags |= LI_SLEPT;
902				}
903				continue;
904			}
905			n++;
906			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
907			    file, line, check_only ? "could sleep" : "sleeping",
908			    lock1->li_lock->lo_name, lock1->li_file,
909			    lock1->li_line);
910		}
911	if (lock_list == &td->td_sleeplocks) {
912		lock_list = PCPU_PTR(spinlocks);
913		goto again;
914	}
915#ifdef DDB
916	if (witness_ddb && n)
917		Debugger(__func__);
918#endif /* DDB */
919	critical_exit();
920	return (n);
921}
922
923static struct witness *
924enroll(const char *description, struct lock_class *lock_class)
925{
926	struct witness *w;
927
928	if (!witness_watch || witness_dead || panicstr != NULL)
929		return (NULL);
930	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
931		return (NULL);
932	mtx_lock_spin(&w_mtx);
933	STAILQ_FOREACH(w, &w_all, w_list) {
934		if (strcmp(description, w->w_name) == 0) {
935			w->w_refcount++;
936			mtx_unlock_spin(&w_mtx);
937			if (lock_class != w->w_class)
938				panic(
939				"lock (%s) %s does not match earlier (%s) lock",
940				    description, lock_class->lc_name,
941				    w->w_class->lc_name);
942			return (w);
943		}
944	}
945	/*
946	 * This isn't quite right, as witness_cold is still 0 while we
947	 * enroll all the locks initialized before witness_initialize().
948	 */
949	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
950		mtx_unlock_spin(&w_mtx);
951		panic("spin lock %s not in order list", description);
952	}
953	if ((w = witness_get()) == NULL)
954		return (NULL);
955	w->w_name = description;
956	w->w_class = lock_class;
957	w->w_refcount = 1;
958	STAILQ_INSERT_HEAD(&w_all, w, w_list);
959	if (lock_class->lc_flags & LC_SPINLOCK)
960		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
961	else if (lock_class->lc_flags & LC_SLEEPLOCK)
962		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
963	else {
964		mtx_unlock_spin(&w_mtx);
965		panic("lock class %s is not sleep or spin",
966		    lock_class->lc_name);
967	}
968	mtx_unlock_spin(&w_mtx);
969	return (w);
970}
971
972static int
973itismychild(struct witness *parent, struct witness *child)
974{
975	static int recursed;
976	struct witness_child_list_entry **wcl;
977	struct witness_list *list;
978
979	MPASS(child != NULL && parent != NULL);
980	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
981	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
982		panic(
983		"%s: parent (%s) and child (%s) are not the same lock type",
984		    __func__, parent->w_class->lc_name,
985		    child->w_class->lc_name);
986
987	/*
988	 * Insert "child" after "parent"
989	 */
990	wcl = &parent->w_children;
991	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
992		wcl = &(*wcl)->wcl_next;
993	if (*wcl == NULL) {
994		*wcl = witness_child_get();
995		if (*wcl == NULL)
996			return (1);
997	}
998	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
999
1000	/*
1001	 * Now prune whole tree.  We look for cases where a lock is now
1002	 * both a descendant and a direct child of a given lock.  In that
1003	 * case, we want to remove the direct child link from the tree.
1004	 */
1005	if (recursed)
1006		return (0);
1007	recursed = 1;
1008	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1009		list = &w_sleep;
1010	else
1011		list = &w_spin;
1012	STAILQ_FOREACH(child, list, w_typelist) {
1013		STAILQ_FOREACH(parent, list, w_typelist) {
1014			if (!isitmychild(parent, child))
1015				continue;
1016			removechild(parent, child);
1017			if (isitmydescendant(parent, child))
1018				continue;
1019			itismychild(parent, child);
1020		}
1021	}
1022	recursed = 0;
1023	witness_levelall();
1024	return (0);
1025}
1026
1027static void
1028removechild(struct witness *parent, struct witness *child)
1029{
1030	struct witness_child_list_entry **wcl, *wcl1;
1031	int i;
1032
1033	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1034		for (i = 0; i < (*wcl)->wcl_count; i++)
1035			if ((*wcl)->wcl_children[i] == child)
1036				goto found;
1037	return;
1038found:
1039	(*wcl)->wcl_count--;
1040	if ((*wcl)->wcl_count > i)
1041		(*wcl)->wcl_children[i] =
1042		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1043	MPASS((*wcl)->wcl_children[i] != NULL);
1044	if ((*wcl)->wcl_count != 0)
1045		return;
1046	wcl1 = *wcl;
1047	*wcl = wcl1->wcl_next;
1048	witness_child_free(wcl1);
1049}
1050
1051static int
1052isitmychild(struct witness *parent, struct witness *child)
1053{
1054	struct witness_child_list_entry *wcl;
1055	int i;
1056
1057	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1058		for (i = 0; i < wcl->wcl_count; i++) {
1059			if (wcl->wcl_children[i] == child)
1060				return (1);
1061		}
1062	}
1063	return (0);
1064}
1065
1066static int
1067isitmydescendant(struct witness *parent, struct witness *child)
1068{
1069	struct witness_child_list_entry *wcl;
1070	int i, j;
1071
1072	if (isitmychild(parent, child))
1073		return (1);
1074	j = 0;
1075	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1076		MPASS(j < 1000);
1077		for (i = 0; i < wcl->wcl_count; i++) {
1078			if (isitmydescendant(wcl->wcl_children[i], child))
1079				return (1);
1080		}
1081		j++;
1082	}
1083	return (0);
1084}
1085
1086void
1087witness_levelall (void)
1088{
1089	struct witness_list *list;
1090	struct witness *w, *w1;
1091
1092	/*
1093	 * First clear all levels.
1094	 */
1095	STAILQ_FOREACH(w, &w_all, w_list) {
1096		w->w_level = 0;
1097	}
1098
1099	/*
1100	 * Look for locks with no parent and level all their descendants.
1101	 */
1102	STAILQ_FOREACH(w, &w_all, w_list) {
1103		/*
1104		 * This is just an optimization, technically we could get
1105		 * away just walking the all list each time.
1106		 */
1107		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1108			list = &w_sleep;
1109		else
1110			list = &w_spin;
1111		STAILQ_FOREACH(w1, list, w_typelist) {
1112			if (isitmychild(w1, w))
1113				goto skip;
1114		}
1115		witness_leveldescendents(w, 0);
1116	skip:
1117	}
1118}
1119
1120static void
1121witness_leveldescendents(struct witness *parent, int level)
1122{
1123	struct witness_child_list_entry *wcl;
1124	int i;
1125
1126	if (parent->w_level < level)
1127		parent->w_level = level;
1128	level++;
1129	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1130		for (i = 0; i < wcl->wcl_count; i++)
1131			witness_leveldescendents(wcl->wcl_children[i], level);
1132}
1133
1134static void
1135witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1136			   struct witness *parent)
1137{
1138	struct witness_child_list_entry *wcl;
1139	int i, level;
1140
1141	level =  parent->w_level;
1142	prnt("%-2d", level);
1143	for (i = 0; i < level; i++)
1144		prnt(" ");
1145	prnt("%s", parent->w_name);
1146	if (parent->w_file != NULL)
1147		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1148		    parent->w_line);
1149	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1150		for (i = 0; i < wcl->wcl_count; i++)
1151			    witness_displaydescendants(prnt,
1152				wcl->wcl_children[i]);
1153}
1154
1155static int
1156blessed(struct witness *w1, struct witness *w2)
1157{
1158	int i;
1159	struct witness_blessed *b;
1160
1161	for (i = 0; i < blessed_count; i++) {
1162		b = &blessed_list[i];
1163		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1164			if (strcmp(w2->w_name, b->b_lock2) == 0)
1165				return (1);
1166			continue;
1167		}
1168		if (strcmp(w1->w_name, b->b_lock2) == 0)
1169			if (strcmp(w2->w_name, b->b_lock1) == 0)
1170				return (1);
1171	}
1172	return (0);
1173}
1174
1175static struct witness *
1176witness_get(void)
1177{
1178	struct witness *w;
1179
1180	if (witness_dead) {
1181		mtx_unlock_spin(&w_mtx);
1182		return (NULL);
1183	}
1184	if (STAILQ_EMPTY(&w_free)) {
1185		witness_dead = 1;
1186		mtx_unlock_spin(&w_mtx);
1187		printf("%s: witness exhausted\n", __func__);
1188		return (NULL);
1189	}
1190	w = STAILQ_FIRST(&w_free);
1191	STAILQ_REMOVE_HEAD(&w_free, w_list);
1192	bzero(w, sizeof(*w));
1193	return (w);
1194}
1195
1196static void
1197witness_free(struct witness *w)
1198{
1199
1200	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1201}
1202
1203static struct witness_child_list_entry *
1204witness_child_get(void)
1205{
1206	struct witness_child_list_entry *wcl;
1207
1208	if (witness_dead) {
1209		mtx_unlock_spin(&w_mtx);
1210		return (NULL);
1211	}
1212	wcl = w_child_free;
1213	if (wcl == NULL) {
1214		witness_dead = 1;
1215		mtx_unlock_spin(&w_mtx);
1216		printf("%s: witness exhausted\n", __func__);
1217		return (NULL);
1218	}
1219	w_child_free = wcl->wcl_next;
1220	bzero(wcl, sizeof(*wcl));
1221	return (wcl);
1222}
1223
1224static void
1225witness_child_free(struct witness_child_list_entry *wcl)
1226{
1227
1228	wcl->wcl_next = w_child_free;
1229	w_child_free = wcl;
1230}
1231
1232static struct lock_list_entry *
1233witness_lock_list_get(void)
1234{
1235	struct lock_list_entry *lle;
1236
1237	if (witness_dead)
1238		return (NULL);
1239	mtx_lock_spin(&w_mtx);
1240	lle = w_lock_list_free;
1241	if (lle == NULL) {
1242		witness_dead = 1;
1243		mtx_unlock_spin(&w_mtx);
1244		printf("%s: witness exhausted\n", __func__);
1245		return (NULL);
1246	}
1247	w_lock_list_free = lle->ll_next;
1248	mtx_unlock_spin(&w_mtx);
1249	bzero(lle, sizeof(*lle));
1250	return (lle);
1251}
1252
1253static void
1254witness_lock_list_free(struct lock_list_entry *lle)
1255{
1256
1257	mtx_lock_spin(&w_mtx);
1258	lle->ll_next = w_lock_list_free;
1259	w_lock_list_free = lle;
1260	mtx_unlock_spin(&w_mtx);
1261}
1262
1263static struct lock_instance *
1264find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1265{
1266	struct lock_list_entry *lle;
1267	struct lock_instance *instance;
1268	int i;
1269
1270	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1271		for (i = lle->ll_count - 1; i >= 0; i--) {
1272			instance = &lle->ll_children[i];
1273			if (instance->li_lock == lock)
1274				return (instance);
1275		}
1276	return (NULL);
1277}
1278
1279int
1280witness_list_locks(struct lock_list_entry **lock_list)
1281{
1282	struct lock_list_entry *lle;
1283	struct lock_instance *instance;
1284	struct lock_object *lock;
1285	int i, nheld;
1286
1287	nheld = 0;
1288	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1289		for (i = lle->ll_count - 1; i >= 0; i--) {
1290			instance = &lle->ll_children[i];
1291			lock = instance->li_lock;
1292			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1293			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1294			    "exclusive" : "shared",
1295			    lock->lo_class->lc_name, lock->lo_name, lock,
1296			    instance->li_file, instance->li_line);
1297			nheld++;
1298		}
1299	return (nheld);
1300}
1301
1302/*
1303 * Calling this on td != curthread is bad unless we are in ddb.
1304 */
1305int
1306witness_list(struct thread *td)
1307{
1308	int nheld;
1309
1310	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1311#ifdef DDB
1312	KASSERT(td == curthread || db_active,
1313	    ("%s: td != curthread and we aren't in the debugger", __func__));
1314	if (!db_active && witness_dead)
1315		return (0);
1316#else
1317	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1318	if (witness_dead)
1319		return (0);
1320#endif
1321	nheld = witness_list_locks(&td->td_sleeplocks);
1322
1323	/*
1324	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1325	 * if td is currently executing on some other CPU and holds spin locks
1326	 * as we won't display those locks.  If we had a MI way of getting
1327	 * the per-cpu data for a given cpu then we could use
1328	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1329	 * and "fix" this.
1330	 *
1331	 * That still wouldn't really fix this unless we locked sched_lock
1332	 * or stopped the other CPU to make sure it wasn't changing the list
1333	 * out from under us.  It is probably best to just not try to handle
1334	 * threads on other CPU's for now.
1335	 */
1336	if (td == curthread) {
1337		/*
1338		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1339		 * change.
1340		 */
1341		critical_enter();
1342		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1343		critical_exit();
1344	}
1345	return (nheld);
1346}
1347
1348void
1349witness_save(struct lock_object *lock, const char **filep, int *linep)
1350{
1351	struct lock_instance *instance;
1352
1353	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1354	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1355		return;
1356	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1357		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1358		    lock->lo_class->lc_name, lock->lo_name);
1359	instance = find_instance(curthread->td_sleeplocks, lock);
1360	if (instance == NULL)
1361		panic("%s: lock (%s) %s not locked", __func__,
1362		    lock->lo_class->lc_name, lock->lo_name);
1363	*filep = instance->li_file;
1364	*linep = instance->li_line;
1365}
1366
1367void
1368witness_restore(struct lock_object *lock, const char *file, int line)
1369{
1370	struct lock_instance *instance;
1371
1372	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1373	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1374		return;
1375	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1376		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1377		    lock->lo_class->lc_name, lock->lo_name);
1378	instance = find_instance(curthread->td_sleeplocks, lock);
1379	if (instance == NULL)
1380		panic("%s: lock (%s) %s not locked", __func__,
1381		    lock->lo_class->lc_name, lock->lo_name);
1382	lock->lo_witness->w_file = file;
1383	lock->lo_witness->w_line = line;
1384	instance->li_file = file;
1385	instance->li_line = line;
1386}
1387
1388void
1389witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1390{
1391#ifdef INVARIANT_SUPPORT
1392	struct lock_instance *instance;
1393
1394	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1395		return;
1396	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1397		instance = find_instance(curthread->td_sleeplocks, lock);
1398	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1399		instance = find_instance(PCPU_GET(spinlocks), lock);
1400	else {
1401		panic("Lock (%s) %s is not sleep or spin!",
1402		    lock->lo_class->lc_name, lock->lo_name);
1403		return;
1404	}
1405	switch (flags) {
1406	case LA_UNLOCKED:
1407		if (instance != NULL)
1408			panic("Lock (%s) %s locked @ %s:%d.",
1409			    lock->lo_class->lc_name, lock->lo_name, file, line);
1410		break;
1411	case LA_LOCKED:
1412	case LA_LOCKED | LA_RECURSED:
1413	case LA_LOCKED | LA_NOTRECURSED:
1414	case LA_SLOCKED:
1415	case LA_SLOCKED | LA_RECURSED:
1416	case LA_SLOCKED | LA_NOTRECURSED:
1417	case LA_XLOCKED:
1418	case LA_XLOCKED | LA_RECURSED:
1419	case LA_XLOCKED | LA_NOTRECURSED:
1420		if (instance == NULL) {
1421			panic("Lock (%s) %s not locked @ %s:%d.",
1422			    lock->lo_class->lc_name, lock->lo_name, file, line);
1423			break;
1424		}
1425		if ((flags & LA_XLOCKED) != 0 &&
1426		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1427			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1428			    lock->lo_class->lc_name, lock->lo_name, file, line);
1429		if ((flags & LA_SLOCKED) != 0 &&
1430		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1431			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1432			    lock->lo_class->lc_name, lock->lo_name, file, line);
1433		if ((flags & LA_RECURSED) != 0 &&
1434		    (instance->li_flags & LI_RECURSEMASK) == 0)
1435			panic("Lock (%s) %s not recursed @ %s:%d.",
1436			    lock->lo_class->lc_name, lock->lo_name, file, line);
1437		if ((flags & LA_NOTRECURSED) != 0 &&
1438		    (instance->li_flags & LI_RECURSEMASK) != 0)
1439			panic("Lock (%s) %s recursed @ %s:%d.",
1440			    lock->lo_class->lc_name, lock->lo_name, file, line);
1441		break;
1442	default:
1443		panic("Invalid lock assertion at %s:%d.", file, line);
1444
1445	}
1446#endif	/* INVARIANT_SUPPORT */
1447}
1448
1449#ifdef DDB
1450
1451DB_SHOW_COMMAND(locks, db_witness_list)
1452{
1453	struct thread *td;
1454	pid_t pid;
1455	struct proc *p;
1456
1457	if (have_addr) {
1458		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1459		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1460		    ((addr >> 16) % 16) * 10000;
1461		/* sx_slock(&allproc_lock); */
1462		FOREACH_PROC_IN_SYSTEM(p) {
1463			if (p->p_pid == pid)
1464				break;
1465		}
1466		/* sx_sunlock(&allproc_lock); */
1467		if (p == NULL) {
1468			db_printf("pid %d not found\n", pid);
1469			return;
1470		}
1471		FOREACH_THREAD_IN_PROC(p, td) {
1472			witness_list(td);
1473		}
1474	} else {
1475		td = curthread;
1476		witness_list(td);
1477	}
1478}
1479
1480DB_SHOW_COMMAND(witness, db_witness_display)
1481{
1482
1483	witness_display(db_printf);
1484}
1485#endif
1486