subr_witness.c revision 88899
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 88899 2002-01-05 08:29:54Z jhb $
31 */
32
33/*
34 * Implementation of the `witness' lock verifier.  Originally implemented for
35 * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36 * classes in FreeBSD.
37 */
38
39/*
40 *	Main Entry: witness
41 *	Pronunciation: 'wit-n&s
42 *	Function: noun
43 *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44 *	    testimony, witness, from 2wit
45 *	Date: before 12th century
46 *	1 : attestation of a fact or event : TESTIMONY
47 *	2 : one that gives evidence; specifically : one who testifies in
48 *	    a cause or before a judicial tribunal
49 *	3 : one asked to be present at a transaction so as to be able to
50 *	    testify to its having taken place
51 *	4 : one who has personal knowledge of something
52 *	5 a : something serving as evidence or proof : SIGN
53 *	  b : public affirmation by word or example of usually
54 *	      religious faith or conviction <the heroic witness to divine
55 *	      life -- Pilot>
56 *	6 capitalized : a member of the Jehovah's Witnesses
57 */
58
59#include "opt_ddb.h"
60#include "opt_witness.h"
61
62#include <sys/param.h>
63#include <sys/bus.h>
64#include <sys/kernel.h>
65#include <sys/ktr.h>
66#include <sys/lock.h>
67#include <sys/malloc.h>
68#include <sys/mutex.h>
69#include <sys/proc.h>
70#include <sys/sysctl.h>
71#include <sys/systm.h>
72
73#include <ddb/ddb.h>
74
75#define WITNESS_COUNT 200
76#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77/*
78 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79 * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80 * probably be safe for the most part, but it's still a SWAG.
81 */
82#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83
84#define	WITNESS_NCHILDREN 6
85
86struct witness_child_list_entry;
87
88struct witness {
89	const	char *w_name;
90	struct	lock_class *w_class;
91	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94	const	char *w_file;
95	int	w_line;
96	u_int	w_level;
97	u_int	w_refcount;
98	u_char	w_Giant_squawked:1;
99	u_char	w_other_squawked:1;
100	u_char	w_same_squawked:1;
101};
102
103struct witness_child_list_entry {
104	struct	witness_child_list_entry *wcl_next;
105	struct	witness *wcl_children[WITNESS_NCHILDREN];
106	u_int	wcl_count;
107};
108
109STAILQ_HEAD(witness_list, witness);
110
111struct witness_blessed {
112	const	char *b_lock1;
113	const	char *b_lock2;
114};
115
116struct witness_order_list_entry {
117	const	char *w_name;
118	struct	lock_class *w_class;
119};
120
121static struct	witness *enroll(const char *description,
122				struct lock_class *lock_class);
123static int	itismychild(struct witness *parent, struct witness *child);
124static void	removechild(struct witness *parent, struct witness *child);
125static int	isitmychild(struct witness *parent, struct witness *child);
126static int	isitmydescendant(struct witness *parent, struct witness *child);
127static int	dup_ok(struct witness *);
128static int	blessed(struct witness *, struct witness *);
129static void	witness_display_list(void(*prnt)(const char *fmt, ...),
130				     struct witness_list *list);
131static void	witness_displaydescendants(void(*)(const char *fmt, ...),
132					   struct witness *);
133static void	witness_leveldescendents(struct witness *parent, int level);
134static void	witness_levelall(void);
135static struct	witness *witness_get(void);
136static void	witness_free(struct witness *m);
137static struct	witness_child_list_entry *witness_child_get(void);
138static void	witness_child_free(struct witness_child_list_entry *wcl);
139static struct	lock_list_entry *witness_lock_list_get(void);
140static void	witness_lock_list_free(struct lock_list_entry *lle);
141static void	witness_display(void(*)(const char *fmt, ...));
142static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
143					     struct lock_object *lock);
144
145MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146
147static int witness_watch = 1;
148TUNABLE_INT("debug.witness_watch", &witness_watch);
149SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
150
151#ifdef DDB
152/*
153 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
154 * drop into kdebug() when:
155 *	- a lock heirarchy violation occurs
156 *	- locks are held when going to sleep.
157 */
158#ifdef WITNESS_DDB
159int	witness_ddb = 1;
160#else
161int	witness_ddb = 0;
162#endif
163TUNABLE_INT("debug.witness_ddb", &witness_ddb);
164SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
165#endif /* DDB */
166
167#ifdef WITNESS_SKIPSPIN
168int	witness_skipspin = 1;
169#else
170int	witness_skipspin = 0;
171#endif
172TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
173SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
174    "");
175
176static struct mtx w_mtx;
177static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
178static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
179static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
180static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
181static struct witness_child_list_entry *w_child_free = NULL;
182static struct lock_list_entry *w_lock_list_free = NULL;
183static int witness_dead;	/* fatal error, probably no memory */
184
185static struct witness w_data[WITNESS_COUNT];
186static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
187static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188
189static struct witness_order_list_entry order_lists[] = {
190	{ "Giant", &lock_class_mtx_sleep },
191	{ "proctree", &lock_class_sx },
192	{ "allproc", &lock_class_sx },
193	{ "process lock", &lock_class_mtx_sleep },
194	{ "uidinfo hash", &lock_class_mtx_sleep },
195	{ "uidinfo struct", &lock_class_mtx_sleep },
196	{ NULL, NULL },
197	/*
198	 * spin locks
199	 */
200#ifdef SMP
201	{ "ap boot", &lock_class_mtx_spin },
202#ifdef __i386__
203	{ "com", &lock_class_mtx_spin },
204#endif
205#endif
206	{ "sio", &lock_class_mtx_spin },
207#ifdef __i386__
208	{ "cy", &lock_class_mtx_spin },
209#endif
210	{ "ng_node", &lock_class_mtx_spin },
211	{ "ng_worklist", &lock_class_mtx_spin },
212	{ "ithread table lock", &lock_class_mtx_spin },
213	{ "sched lock", &lock_class_mtx_spin },
214	{ "callout", &lock_class_mtx_spin },
215	/*
216	 * leaf locks
217	 */
218	{ "icu", &lock_class_mtx_spin },
219#ifdef SMP
220	{ "smp rendezvous", &lock_class_mtx_spin },
221#endif
222	{ "clk", &lock_class_mtx_spin },
223	{ NULL, NULL },
224	{ NULL, NULL }
225};
226
227static const char *dup_list[] = {
228	"process lock",
229	NULL
230};
231
232/*
233 * Pairs of locks which have been blessed
234 * Don't complain about order problems with blessed locks
235 */
236static struct witness_blessed blessed_list[] = {
237};
238static int blessed_count =
239	sizeof(blessed_list) / sizeof(struct witness_blessed);
240
241/*
242 * List of all locks in the system.
243 */
244STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245
246static struct mtx all_mtx = {
247	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248	  "All locks list",		/* mtx_object.lo_name */
249	  LO_INITIALIZED,		/* mtx_object.lo_flags */
250	  { NULL },			/* mtx_object.lo_list */
251	  NULL },			/* mtx_object.lo_witness */
252	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
253	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
254	{ NULL, NULL }			/* mtx_contested */
255};
256
257/*
258 * This global is set to 0 once it becomes safe to use the witness code.
259 */
260static int witness_cold = 1;
261
262/*
263 * Global variables for book keeping.
264 */
265static int lock_cur_cnt;
266static int lock_max_cnt;
267
268/*
269 * The WITNESS-enabled diagnostic code.
270 */
271static void
272witness_initialize(void *dummy __unused)
273{
274	struct lock_object *lock;
275	struct witness_order_list_entry *order;
276	struct witness *w, *w1;
277	int i;
278
279	/*
280	 * We have to release Giant before initializing its witness
281	 * structure so that WITNESS doesn't get confused.
282	 */
283	mtx_unlock(&Giant);
284	mtx_assert(&Giant, MA_NOTOWNED);
285
286	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
287	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
288	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
289	for (i = 0; i < WITNESS_COUNT; i++)
290		witness_free(&w_data[i]);
291	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
292		witness_child_free(&w_childdata[i]);
293	for (i = 0; i < LOCK_CHILDCOUNT; i++)
294		witness_lock_list_free(&w_locklistdata[i]);
295
296	/* First add in all the specified order lists. */
297	for (order = order_lists; order->w_name != NULL; order++) {
298		w = enroll(order->w_name, order->w_class);
299		if (w == NULL)
300			continue;
301		w->w_file = "order list";
302		for (order++; order->w_name != NULL; order++) {
303			w1 = enroll(order->w_name, order->w_class);
304			if (w1 == NULL)
305				continue;
306			w1->w_file = "order list";
307			itismychild(w, w1);
308			w = w1;
309		}
310	}
311
312	/* Iterate through all locks and add them to witness. */
313	mtx_lock(&all_mtx);
314	STAILQ_FOREACH(lock, &all_locks, lo_list) {
315		if (lock->lo_flags & LO_WITNESS)
316			lock->lo_witness = enroll(lock->lo_name,
317			    lock->lo_class);
318		else
319			lock->lo_witness = NULL;
320	}
321	mtx_unlock(&all_mtx);
322
323	/* Mark the witness code as being ready for use. */
324	atomic_store_rel_int(&witness_cold, 0);
325
326	mtx_lock(&Giant);
327}
328SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
329
330void
331witness_init(struct lock_object *lock)
332{
333	struct lock_class *class;
334
335	class = lock->lo_class;
336	if (lock->lo_flags & LO_INITIALIZED)
337		panic("%s: lock (%s) %s is already initialized", __func__,
338		    class->lc_name, lock->lo_name);
339	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
340	    (class->lc_flags & LC_RECURSABLE) == 0)
341		panic("%s: lock (%s) %s can not be recursable", __func__,
342		    class->lc_name, lock->lo_name);
343	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
344	    (class->lc_flags & LC_SLEEPABLE) == 0)
345		panic("%s: lock (%s) %s can not be sleepable", __func__,
346		    class->lc_name, lock->lo_name);
347	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
348	    (class->lc_flags & LC_UPGRADABLE) == 0)
349		panic("%s: lock (%s) %s can not be upgradable", __func__,
350		    class->lc_name, lock->lo_name);
351
352	mtx_lock(&all_mtx);
353	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
354	lock->lo_flags |= LO_INITIALIZED;
355	lock_cur_cnt++;
356	if (lock_cur_cnt > lock_max_cnt)
357		lock_max_cnt = lock_cur_cnt;
358	mtx_unlock(&all_mtx);
359	if (!witness_cold && !witness_dead && panicstr == NULL &&
360	    (lock->lo_flags & LO_WITNESS) != 0)
361		lock->lo_witness = enroll(lock->lo_name, class);
362	else
363		lock->lo_witness = NULL;
364}
365
366void
367witness_destroy(struct lock_object *lock)
368{
369	struct witness *w;
370
371	if (witness_cold)
372		panic("lock (%s) %s destroyed while witness_cold",
373		    lock->lo_class->lc_name, lock->lo_name);
374	if ((lock->lo_flags & LO_INITIALIZED) == 0)
375		panic("%s: lock (%s) %s is not initialized", __func__,
376		    lock->lo_class->lc_name, lock->lo_name);
377
378	/* XXX: need to verify that no one holds the lock */
379	w = lock->lo_witness;
380	if (w != NULL) {
381		mtx_lock_spin(&w_mtx);
382		w->w_refcount--;
383		if (w->w_refcount == 0) {
384			CTR2(KTR_WITNESS,
385			    "%s: marking witness %s as dead", __func__, w->w_name);
386			w->w_name = "(dead)";
387			w->w_file = "(dead)";
388			w->w_line = 0;
389		}
390		mtx_unlock_spin(&w_mtx);
391	}
392
393	mtx_lock(&all_mtx);
394	lock_cur_cnt--;
395	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
396	lock->lo_flags &= ~LO_INITIALIZED;
397	mtx_unlock(&all_mtx);
398}
399
400static void
401witness_display_list(void(*prnt)(const char *fmt, ...),
402		     struct witness_list *list)
403{
404	struct witness *w, *w1;
405	int found;
406
407	STAILQ_FOREACH(w, list, w_typelist) {
408		if (w->w_file == NULL)
409			continue;
410		found = 0;
411		STAILQ_FOREACH(w1, list, w_typelist) {
412			if (isitmychild(w1, w)) {
413				found++;
414				break;
415			}
416		}
417		if (found)
418			continue;
419		/*
420		 * This lock has no anscestors, display its descendants.
421		 */
422		witness_displaydescendants(prnt, w);
423	}
424}
425
426static void
427witness_display(void(*prnt)(const char *fmt, ...))
428{
429	struct witness *w;
430
431	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
432	witness_levelall();
433
434	/*
435	 * First, handle sleep locks which have been acquired at least
436	 * once.
437	 */
438	prnt("Sleep locks:\n");
439	witness_display_list(prnt, &w_sleep);
440
441	/*
442	 * Now do spin locks which have been acquired at least once.
443	 */
444	prnt("\nSpin locks:\n");
445	witness_display_list(prnt, &w_spin);
446
447	/*
448	 * Finally, any locks which have not been acquired yet.
449	 */
450	prnt("\nLocks which were never acquired:\n");
451	STAILQ_FOREACH(w, &w_all, w_list) {
452		if (w->w_file != NULL)
453			continue;
454		prnt("%s\n", w->w_name);
455	}
456}
457
458void
459witness_lock(struct lock_object *lock, int flags, const char *file, int line)
460{
461	struct lock_list_entry **lock_list, *lle;
462	struct lock_instance *lock1, *lock2;
463	struct lock_class *class;
464	struct witness *w, *w1;
465	struct thread *td;
466	int i, j;
467#ifdef DDB
468	int go_into_ddb = 0;
469#endif /* DDB */
470
471	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
472	    panicstr != NULL)
473		return;
474	w = lock->lo_witness;
475	class = lock->lo_class;
476	td = curthread;
477
478	if (class->lc_flags & LC_SLEEPLOCK) {
479		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
480			panic("blockable sleep lock (%s) %s @ %s:%d",
481			    class->lc_name, lock->lo_name, file, line);
482		lock_list = &td->td_sleeplocks;
483	} else
484		lock_list = PCPU_PTR(spinlocks);
485
486	/*
487	 * Try locks do not block if they fail to acquire the lock, thus
488	 * there is no danger of deadlocks or of switching while holding a
489	 * spin lock if we acquire a lock via a try operation.
490	 */
491	if (flags & LOP_TRYLOCK)
492		goto out;
493
494	/*
495	 * Is this the first lock acquired?  If so, then no order checking
496	 * is needed.
497	 */
498	if (*lock_list == NULL)
499		goto out;
500
501	/*
502	 * Check to see if we are recursing on a lock we already own.
503	 */
504	lock1 = find_instance(*lock_list, lock);
505	if (lock1 != NULL) {
506		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
507		    (flags & LOP_EXCLUSIVE) == 0) {
508			printf("shared lock of (%s) %s @ %s:%d\n",
509			    class->lc_name, lock->lo_name, file, line);
510			printf("while exclusively locked from %s:%d\n",
511			    lock1->li_file, lock1->li_line);
512			panic("share->excl");
513		}
514		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
515		    (flags & LOP_EXCLUSIVE) != 0) {
516			printf("exclusive lock of (%s) %s @ %s:%d\n",
517			    class->lc_name, lock->lo_name, file, line);
518			printf("while share locked from %s:%d\n",
519			    lock1->li_file, lock1->li_line);
520			panic("excl->share");
521		}
522		lock1->li_flags++;
523		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
524			printf(
525			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
526			    class->lc_name, lock->lo_name, file, line);
527			printf("first acquired @ %s:%d\n", lock1->li_file,
528			    lock1->li_line);
529			panic("recurse");
530		}
531		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
532		    td->td_proc->p_pid, lock->lo_name,
533		    lock1->li_flags & LI_RECURSEMASK);
534		lock1->li_file = file;
535		lock1->li_line = line;
536		return;
537	}
538
539	/*
540	 * Check for duplicate locks of the same type.  Note that we only
541	 * have to check for this on the last lock we just acquired.  Any
542	 * other cases will be caught as lock order violations.
543	 */
544	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
545	w1 = lock1->li_lock->lo_witness;
546	if (w1 == w) {
547		if (w->w_same_squawked || dup_ok(w))
548			goto out;
549		w->w_same_squawked = 1;
550		printf("acquiring duplicate lock of same type: \"%s\"\n",
551			lock->lo_name);
552		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
553		printf(" 2nd @ %s:%d\n", file, line);
554#ifdef DDB
555		go_into_ddb = 1;
556#endif /* DDB */
557		goto out;
558	}
559	MPASS(!mtx_owned(&w_mtx));
560	mtx_lock_spin(&w_mtx);
561	/*
562	 * If we have a known higher number just say ok
563	 */
564	if (witness_watch > 1 && w->w_level > w1->w_level) {
565		mtx_unlock_spin(&w_mtx);
566		goto out;
567	}
568	if (isitmydescendant(w1, w)) {
569		mtx_unlock_spin(&w_mtx);
570		goto out;
571	}
572	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
573		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
574
575			MPASS(j < WITNESS_COUNT);
576			lock1 = &lle->ll_children[i];
577			w1 = lock1->li_lock->lo_witness;
578
579			/*
580			 * If this lock doesn't undergo witness checking,
581			 * then skip it.
582			 */
583			if (w1 == NULL) {
584				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
585				    ("lock missing witness structure"));
586				continue;
587			}
588			/*
589			 * If we are locking Giant and we slept with this
590			 * lock, then skip it.
591			 */
592			if ((lock1->li_flags & LI_SLEPT) != 0 &&
593			    lock == &Giant.mtx_object)
594				continue;
595			if (!isitmydescendant(w, w1))
596				continue;
597			/*
598			 * We have a lock order violation, check to see if it
599			 * is allowed or has already been yelled about.
600			 */
601			mtx_unlock_spin(&w_mtx);
602			if (blessed(w, w1))
603				goto out;
604			if (lock1->li_lock == &Giant.mtx_object) {
605				if (w1->w_Giant_squawked)
606					goto out;
607				else
608					w1->w_Giant_squawked = 1;
609			} else {
610				if (w1->w_other_squawked)
611					goto out;
612				else
613					w1->w_other_squawked = 1;
614			}
615			/*
616			 * Ok, yell about it.
617			 */
618			printf("lock order reversal\n");
619			/*
620			 * Try to locate an earlier lock with
621			 * witness w in our list.
622			 */
623			do {
624				lock2 = &lle->ll_children[i];
625				MPASS(lock2->li_lock != NULL);
626				if (lock2->li_lock->lo_witness == w)
627					break;
628				i--;
629				if (i == 0 && lle->ll_next != NULL) {
630					lle = lle->ll_next;
631					i = lle->ll_count - 1;
632					MPASS(i != 0);
633				}
634			} while (i >= 0);
635			if (i < 0) {
636				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
637				    lock1->li_lock->lo_name, lock1->li_file,
638				    lock1->li_line);
639				printf(" 2nd %p %s @ %s:%d\n", lock,
640				    lock->lo_name, file, line);
641			} else {
642				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
643				    lock2->li_lock->lo_name, lock2->li_file,
644				    lock2->li_line);
645				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
646				    lock1->li_lock->lo_name, lock1->li_file,
647				    lock1->li_line);
648				printf(" 3rd %p %s @ %s:%d\n", lock,
649				    lock->lo_name, file, line);
650			}
651#ifdef DDB
652			go_into_ddb = 1;
653#endif /* DDB */
654			goto out;
655		}
656	}
657	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
658	/*
659	 * Don't build a new relationship if we are locking Giant just
660	 * after waking up and the previous lock in the list was acquired
661	 * prior to blocking.
662	 */
663	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
664		mtx_unlock_spin(&w_mtx);
665	else {
666		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
667		    lock->lo_name, lock1->li_lock->lo_name);
668		if (!itismychild(lock1->li_lock->lo_witness, w))
669			mtx_unlock_spin(&w_mtx);
670	}
671
672out:
673#ifdef DDB
674	if (witness_ddb && go_into_ddb)
675		Debugger(__func__);
676#endif /* DDB */
677	w->w_file = file;
678	w->w_line = line;
679
680	lle = *lock_list;
681	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
682		lle = witness_lock_list_get();
683		if (lle == NULL)
684			return;
685		lle->ll_next = *lock_list;
686		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
687		    td->td_proc->p_pid, lle);
688		*lock_list = lle;
689	}
690	lock1 = &lle->ll_children[lle->ll_count++];
691	lock1->li_lock = lock;
692	lock1->li_line = line;
693	lock1->li_file = file;
694	if ((flags & LOP_EXCLUSIVE) != 0)
695		lock1->li_flags = LI_EXCLUSIVE;
696	else
697		lock1->li_flags = 0;
698	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
699	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
700}
701
702void
703witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
704{
705	struct lock_instance *instance;
706	struct lock_class *class;
707
708	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
709	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
710		return;
711	class = lock->lo_class;
712	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
713		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
714		    class->lc_name, lock->lo_name, file, line);
715	if ((flags & LOP_TRYLOCK) == 0)
716		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
717		    lock->lo_name, file, line);
718	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
719		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
720		    class->lc_name, lock->lo_name, file, line);
721	instance = find_instance(curthread->td_sleeplocks, lock);
722	if (instance == NULL)
723		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
724		    class->lc_name, lock->lo_name, file, line);
725	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
726		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
727		    class->lc_name, lock->lo_name, file, line);
728	if ((instance->li_flags & LI_RECURSEMASK) != 0)
729		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
730		    class->lc_name, lock->lo_name,
731		    instance->li_flags & LI_RECURSEMASK, file, line);
732	instance->li_flags |= LI_EXCLUSIVE;
733}
734
735void
736witness_downgrade(struct lock_object *lock, int flags, const char *file,
737    int line)
738{
739	struct lock_instance *instance;
740	struct lock_class *class;
741
742	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
743	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
744		return;
745	class = lock->lo_class;
746	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
747		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
748		    class->lc_name, lock->lo_name, file, line);
749	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
750		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
751		    class->lc_name, lock->lo_name, file, line);
752	instance = find_instance(curthread->td_sleeplocks, lock);
753	if (instance == NULL)
754		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
755		    class->lc_name, lock->lo_name, file, line);
756	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
757		panic("downgrade of shared lock (%s) %s @ %s:%d",
758		    class->lc_name, lock->lo_name, file, line);
759	if ((instance->li_flags & LI_RECURSEMASK) != 0)
760		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
761		    class->lc_name, lock->lo_name,
762		    instance->li_flags & LI_RECURSEMASK, file, line);
763	instance->li_flags &= ~LI_EXCLUSIVE;
764}
765
766void
767witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
768{
769	struct lock_list_entry **lock_list, *lle;
770	struct lock_instance *instance;
771	struct lock_class *class;
772	struct thread *td;
773	critical_t s;
774	int i, j;
775
776	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
777	    panicstr != NULL)
778		return;
779	td = curthread;
780	class = lock->lo_class;
781	if (class->lc_flags & LC_SLEEPLOCK)
782		lock_list = &td->td_sleeplocks;
783	else
784		lock_list = PCPU_PTR(spinlocks);
785	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
786		for (i = 0; i < (*lock_list)->ll_count; i++) {
787			instance = &(*lock_list)->ll_children[i];
788			if (instance->li_lock == lock) {
789				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
790				    (flags & LOP_EXCLUSIVE) == 0) {
791					printf(
792					"shared unlock of (%s) %s @ %s:%d\n",
793					    class->lc_name, lock->lo_name,
794					    file, line);
795					printf(
796					"while exclusively locked from %s:%d\n",
797					    instance->li_file,
798					    instance->li_line);
799					panic("excl->ushare");
800				}
801				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
802				    (flags & LOP_EXCLUSIVE) != 0) {
803					printf(
804					"exclusive unlock of (%s) %s @ %s:%d\n",
805					    class->lc_name, lock->lo_name,
806					    file, line);
807					printf(
808					"while share locked from %s:%d\n",
809					    instance->li_file,
810					    instance->li_line);
811					panic("share->uexcl");
812				}
813				/* If we are recursed, unrecurse. */
814				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
815					CTR4(KTR_WITNESS,
816				    "%s: pid %d unrecursed on %s r=%d", __func__,
817					    td->td_proc->p_pid,
818					    instance->li_lock->lo_name,
819					    instance->li_flags);
820					instance->li_flags--;
821					goto out;
822				}
823				s = cpu_critical_enter();
824				CTR4(KTR_WITNESS,
825				    "%s: pid %d removed %s from lle[%d]", __func__,
826				    td->td_proc->p_pid,
827				    instance->li_lock->lo_name,
828				    (*lock_list)->ll_count - 1);
829				(*lock_list)->ll_count--;
830				for (j = i; j < (*lock_list)->ll_count; j++)
831					(*lock_list)->ll_children[j] =
832					    (*lock_list)->ll_children[j + 1];
833				cpu_critical_exit(s);
834				if ((*lock_list)->ll_count == 0) {
835					lle = *lock_list;
836					*lock_list = lle->ll_next;
837					CTR3(KTR_WITNESS,
838					    "%s: pid %d removed lle %p", __func__,
839					    td->td_proc->p_pid, lle);
840					witness_lock_list_free(lle);
841				}
842				goto out;
843			}
844		}
845	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
846	    file, line);
847out:
848	/*
849	 * We don't need to protect this PCPU_GET() here against preemption
850	 * because if we hold any spinlocks then we are already protected,
851	 * and if we don't we will get NULL if we hold no spinlocks even if
852	 * we switch CPU's while reading it.
853	 */
854	if (class->lc_flags & LC_SLEEPLOCK) {
855		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
856			panic("switchable sleep unlock (%s) %s @ %s:%d",
857			    class->lc_name, lock->lo_name, file, line);
858	}
859}
860
861/*
862 * Warn if any held locks are not sleepable.  Note that Giant and the lock
863 * passed in are both special cases since they are both released during the
864 * sleep process and aren't actually held while the thread is asleep.
865 */
866int
867witness_sleep(int check_only, struct lock_object *lock, const char *file,
868	      int line)
869{
870	struct lock_list_entry **lock_list, *lle;
871	struct lock_instance *lock1;
872	struct thread *td;
873	critical_t savecrit;
874	int i, n;
875
876	if (witness_dead || panicstr != NULL)
877		return (0);
878	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
879	n = 0;
880	/*
881	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
882	 */
883	savecrit = cpu_critical_enter();
884	td = curthread;
885	lock_list = &td->td_sleeplocks;
886again:
887	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
888		for (i = lle->ll_count - 1; i >= 0; i--) {
889			lock1 = &lle->ll_children[i];
890			if (lock1->li_lock == lock ||
891			    lock1->li_lock == &Giant.mtx_object)
892				continue;
893			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
894				if (check_only == 0) {
895					CTR3(KTR_WITNESS,
896				    "pid %d: sleeping with lock (%s) %s held",
897					    td->td_proc->p_pid,
898					    lock1->li_lock->lo_class->lc_name,
899					    lock1->li_lock->lo_name);
900					lock1->li_flags |= LI_SLEPT;
901				}
902				continue;
903			}
904			n++;
905			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
906			    file, line, check_only ? "could sleep" : "sleeping",
907			    lock1->li_lock->lo_name, lock1->li_file,
908			    lock1->li_line);
909		}
910	if (lock_list == &td->td_sleeplocks) {
911		lock_list = PCPU_PTR(spinlocks);
912		goto again;
913	}
914#ifdef DDB
915	if (witness_ddb && n)
916		Debugger(__func__);
917#endif /* DDB */
918	cpu_critical_exit(savecrit);
919	return (n);
920}
921
922static struct witness *
923enroll(const char *description, struct lock_class *lock_class)
924{
925	struct witness *w;
926
927	if (!witness_watch || witness_dead || panicstr != NULL)
928		return (NULL);
929	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
930		return (NULL);
931	mtx_lock_spin(&w_mtx);
932	STAILQ_FOREACH(w, &w_all, w_list) {
933		if (strcmp(description, w->w_name) == 0) {
934			w->w_refcount++;
935			mtx_unlock_spin(&w_mtx);
936			if (lock_class != w->w_class)
937				panic(
938				"lock (%s) %s does not match earlier (%s) lock",
939				    description, lock_class->lc_name,
940				    w->w_class->lc_name);
941			return (w);
942		}
943	}
944	/*
945	 * This isn't quite right, as witness_cold is still 0 while we
946	 * enroll all the locks initialized before witness_initialize().
947	 */
948	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
949		mtx_unlock_spin(&w_mtx);
950		panic("spin lock %s not in order list", description);
951	}
952	if ((w = witness_get()) == NULL)
953		return (NULL);
954	w->w_name = description;
955	w->w_class = lock_class;
956	w->w_refcount = 1;
957	STAILQ_INSERT_HEAD(&w_all, w, w_list);
958	if (lock_class->lc_flags & LC_SPINLOCK)
959		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
960	else if (lock_class->lc_flags & LC_SLEEPLOCK)
961		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
962	else {
963		mtx_unlock_spin(&w_mtx);
964		panic("lock class %s is not sleep or spin",
965		    lock_class->lc_name);
966	}
967	mtx_unlock_spin(&w_mtx);
968	return (w);
969}
970
971static int
972itismychild(struct witness *parent, struct witness *child)
973{
974	static int recursed;
975	struct witness_child_list_entry **wcl;
976	struct witness_list *list;
977
978	MPASS(child != NULL && parent != NULL);
979	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
980	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
981		panic(
982		"%s: parent (%s) and child (%s) are not the same lock type",
983		    __func__, parent->w_class->lc_name,
984		    child->w_class->lc_name);
985
986	/*
987	 * Insert "child" after "parent"
988	 */
989	wcl = &parent->w_children;
990	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
991		wcl = &(*wcl)->wcl_next;
992	if (*wcl == NULL) {
993		*wcl = witness_child_get();
994		if (*wcl == NULL)
995			return (1);
996	}
997	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
998
999	/*
1000	 * Now prune whole tree.  We look for cases where a lock is now
1001	 * both a descendant and a direct child of a given lock.  In that
1002	 * case, we want to remove the direct child link from the tree.
1003	 */
1004	if (recursed)
1005		return (0);
1006	recursed = 1;
1007	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1008		list = &w_sleep;
1009	else
1010		list = &w_spin;
1011	STAILQ_FOREACH(child, list, w_typelist) {
1012		STAILQ_FOREACH(parent, list, w_typelist) {
1013			if (!isitmychild(parent, child))
1014				continue;
1015			removechild(parent, child);
1016			if (isitmydescendant(parent, child))
1017				continue;
1018			itismychild(parent, child);
1019		}
1020	}
1021	recursed = 0;
1022	witness_levelall();
1023	return (0);
1024}
1025
1026static void
1027removechild(struct witness *parent, struct witness *child)
1028{
1029	struct witness_child_list_entry **wcl, *wcl1;
1030	int i;
1031
1032	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1033		for (i = 0; i < (*wcl)->wcl_count; i++)
1034			if ((*wcl)->wcl_children[i] == child)
1035				goto found;
1036	return;
1037found:
1038	(*wcl)->wcl_count--;
1039	if ((*wcl)->wcl_count > i)
1040		(*wcl)->wcl_children[i] =
1041		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1042	MPASS((*wcl)->wcl_children[i] != NULL);
1043	if ((*wcl)->wcl_count != 0)
1044		return;
1045	wcl1 = *wcl;
1046	*wcl = wcl1->wcl_next;
1047	witness_child_free(wcl1);
1048}
1049
1050static int
1051isitmychild(struct witness *parent, struct witness *child)
1052{
1053	struct witness_child_list_entry *wcl;
1054	int i;
1055
1056	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1057		for (i = 0; i < wcl->wcl_count; i++) {
1058			if (wcl->wcl_children[i] == child)
1059				return (1);
1060		}
1061	}
1062	return (0);
1063}
1064
1065static int
1066isitmydescendant(struct witness *parent, struct witness *child)
1067{
1068	struct witness_child_list_entry *wcl;
1069	int i, j;
1070
1071	if (isitmychild(parent, child))
1072		return (1);
1073	j = 0;
1074	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1075		MPASS(j < 1000);
1076		for (i = 0; i < wcl->wcl_count; i++) {
1077			if (isitmydescendant(wcl->wcl_children[i], child))
1078				return (1);
1079		}
1080		j++;
1081	}
1082	return (0);
1083}
1084
1085void
1086witness_levelall (void)
1087{
1088	struct witness_list *list;
1089	struct witness *w, *w1;
1090
1091	/*
1092	 * First clear all levels.
1093	 */
1094	STAILQ_FOREACH(w, &w_all, w_list) {
1095		w->w_level = 0;
1096	}
1097
1098	/*
1099	 * Look for locks with no parent and level all their descendants.
1100	 */
1101	STAILQ_FOREACH(w, &w_all, w_list) {
1102		/*
1103		 * This is just an optimization, technically we could get
1104		 * away just walking the all list each time.
1105		 */
1106		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1107			list = &w_sleep;
1108		else
1109			list = &w_spin;
1110		STAILQ_FOREACH(w1, list, w_typelist) {
1111			if (isitmychild(w1, w))
1112				goto skip;
1113		}
1114		witness_leveldescendents(w, 0);
1115	skip:
1116	}
1117}
1118
1119static void
1120witness_leveldescendents(struct witness *parent, int level)
1121{
1122	struct witness_child_list_entry *wcl;
1123	int i;
1124
1125	if (parent->w_level < level)
1126		parent->w_level = level;
1127	level++;
1128	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1129		for (i = 0; i < wcl->wcl_count; i++)
1130			witness_leveldescendents(wcl->wcl_children[i], level);
1131}
1132
1133static void
1134witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1135			   struct witness *parent)
1136{
1137	struct witness_child_list_entry *wcl;
1138	int i, level;
1139
1140	level =  parent->w_level;
1141	prnt("%-2d", level);
1142	for (i = 0; i < level; i++)
1143		prnt(" ");
1144	prnt("%s", parent->w_name);
1145	if (parent->w_file != NULL)
1146		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1147		    parent->w_line);
1148	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1149		for (i = 0; i < wcl->wcl_count; i++)
1150			    witness_displaydescendants(prnt,
1151				wcl->wcl_children[i]);
1152}
1153
1154static int
1155dup_ok(struct witness *w)
1156{
1157	const char **dup;
1158
1159	for (dup = dup_list; *dup != NULL; dup++)
1160		if (strcmp(w->w_name, *dup) == 0)
1161			return (1);
1162	return (0);
1163}
1164
1165static int
1166blessed(struct witness *w1, struct witness *w2)
1167{
1168	int i;
1169	struct witness_blessed *b;
1170
1171	for (i = 0; i < blessed_count; i++) {
1172		b = &blessed_list[i];
1173		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1174			if (strcmp(w2->w_name, b->b_lock2) == 0)
1175				return (1);
1176			continue;
1177		}
1178		if (strcmp(w1->w_name, b->b_lock2) == 0)
1179			if (strcmp(w2->w_name, b->b_lock1) == 0)
1180				return (1);
1181	}
1182	return (0);
1183}
1184
1185static struct witness *
1186witness_get(void)
1187{
1188	struct witness *w;
1189
1190	if (witness_dead) {
1191		mtx_unlock_spin(&w_mtx);
1192		return (NULL);
1193	}
1194	if (STAILQ_EMPTY(&w_free)) {
1195		witness_dead = 1;
1196		mtx_unlock_spin(&w_mtx);
1197		printf("%s: witness exhausted\n", __func__);
1198		return (NULL);
1199	}
1200	w = STAILQ_FIRST(&w_free);
1201	STAILQ_REMOVE_HEAD(&w_free, w_list);
1202	bzero(w, sizeof(*w));
1203	return (w);
1204}
1205
1206static void
1207witness_free(struct witness *w)
1208{
1209
1210	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1211}
1212
1213static struct witness_child_list_entry *
1214witness_child_get(void)
1215{
1216	struct witness_child_list_entry *wcl;
1217
1218	if (witness_dead) {
1219		mtx_unlock_spin(&w_mtx);
1220		return (NULL);
1221	}
1222	wcl = w_child_free;
1223	if (wcl == NULL) {
1224		witness_dead = 1;
1225		mtx_unlock_spin(&w_mtx);
1226		printf("%s: witness exhausted\n", __func__);
1227		return (NULL);
1228	}
1229	w_child_free = wcl->wcl_next;
1230	bzero(wcl, sizeof(*wcl));
1231	return (wcl);
1232}
1233
1234static void
1235witness_child_free(struct witness_child_list_entry *wcl)
1236{
1237
1238	wcl->wcl_next = w_child_free;
1239	w_child_free = wcl;
1240}
1241
1242static struct lock_list_entry *
1243witness_lock_list_get(void)
1244{
1245	struct lock_list_entry *lle;
1246
1247	if (witness_dead)
1248		return (NULL);
1249	mtx_lock_spin(&w_mtx);
1250	lle = w_lock_list_free;
1251	if (lle == NULL) {
1252		witness_dead = 1;
1253		mtx_unlock_spin(&w_mtx);
1254		printf("%s: witness exhausted\n", __func__);
1255		return (NULL);
1256	}
1257	w_lock_list_free = lle->ll_next;
1258	mtx_unlock_spin(&w_mtx);
1259	bzero(lle, sizeof(*lle));
1260	return (lle);
1261}
1262
1263static void
1264witness_lock_list_free(struct lock_list_entry *lle)
1265{
1266
1267	mtx_lock_spin(&w_mtx);
1268	lle->ll_next = w_lock_list_free;
1269	w_lock_list_free = lle;
1270	mtx_unlock_spin(&w_mtx);
1271}
1272
1273static struct lock_instance *
1274find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1275{
1276	struct lock_list_entry *lle;
1277	struct lock_instance *instance;
1278	int i;
1279
1280	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1281		for (i = lle->ll_count - 1; i >= 0; i--) {
1282			instance = &lle->ll_children[i];
1283			if (instance->li_lock == lock)
1284				return (instance);
1285		}
1286	return (NULL);
1287}
1288
1289int
1290witness_list_locks(struct lock_list_entry **lock_list)
1291{
1292	struct lock_list_entry *lle;
1293	struct lock_instance *instance;
1294	struct lock_object *lock;
1295	int i, nheld;
1296
1297	nheld = 0;
1298	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1299		for (i = lle->ll_count - 1; i >= 0; i--) {
1300			instance = &lle->ll_children[i];
1301			lock = instance->li_lock;
1302			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1303			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1304			    "exclusive" : "shared",
1305			    lock->lo_class->lc_name, lock->lo_name, lock,
1306			    instance->li_file, instance->li_line);
1307			nheld++;
1308		}
1309	return (nheld);
1310}
1311
1312/*
1313 * Calling this on td != curthread is bad unless we are in ddb.
1314 */
1315int
1316witness_list(struct thread *td)
1317{
1318	critical_t savecrit;
1319	int nheld;
1320
1321	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1322#ifdef DDB
1323	KASSERT(td == curthread || db_active,
1324	    ("%s: td != curthread and we aren't in the debugger", __func__));
1325	if (!db_active && witness_dead)
1326		return (0);
1327#else
1328	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1329	if (witness_dead)
1330		return (0);
1331#endif
1332	nheld = witness_list_locks(&td->td_sleeplocks);
1333
1334	/*
1335	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1336	 * if td is currently executing on some other CPU and holds spin locks
1337	 * as we won't display those locks.  If we had a MI way of getting
1338	 * the per-cpu data for a given cpu then we could use
1339	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1340	 * and "fix" this.
1341	 */
1342	if (td == curthread) {
1343		/*
1344		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1345		 * change.
1346		 */
1347		savecrit = cpu_critical_enter();
1348		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1349		cpu_critical_exit(savecrit);
1350	}
1351	return (nheld);
1352}
1353
1354void
1355witness_save(struct lock_object *lock, const char **filep, int *linep)
1356{
1357	struct lock_instance *instance;
1358
1359	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1360	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1361		return;
1362	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1363		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1364		    lock->lo_class->lc_name, lock->lo_name);
1365	instance = find_instance(curthread->td_sleeplocks, lock);
1366	if (instance == NULL)
1367		panic("%s: lock (%s) %s not locked", __func__,
1368		    lock->lo_class->lc_name, lock->lo_name);
1369	*filep = instance->li_file;
1370	*linep = instance->li_line;
1371}
1372
1373void
1374witness_restore(struct lock_object *lock, const char *file, int line)
1375{
1376	struct lock_instance *instance;
1377
1378	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1379	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1380		return;
1381	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1382		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1383		    lock->lo_class->lc_name, lock->lo_name);
1384	instance = find_instance(curthread->td_sleeplocks, lock);
1385	if (instance == NULL)
1386		panic("%s: lock (%s) %s not locked", __func__,
1387		    lock->lo_class->lc_name, lock->lo_name);
1388	lock->lo_witness->w_file = file;
1389	lock->lo_witness->w_line = line;
1390	instance->li_file = file;
1391	instance->li_line = line;
1392}
1393
1394void
1395witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1396{
1397#ifdef INVARIANT_SUPPORT
1398	struct lock_instance *instance;
1399
1400	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1401		return;
1402	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1403		instance = find_instance(curthread->td_sleeplocks, lock);
1404	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1405		instance = find_instance(PCPU_GET(spinlocks), lock);
1406	else {
1407		panic("Lock (%s) %s is not sleep or spin!",
1408		    lock->lo_class->lc_name, lock->lo_name);
1409		return;
1410	}
1411	switch (flags) {
1412	case LA_UNLOCKED:
1413		if (instance != NULL)
1414			panic("Lock (%s) %s locked @ %s:%d.",
1415			    lock->lo_class->lc_name, lock->lo_name, file, line);
1416		break;
1417	case LA_LOCKED:
1418	case LA_LOCKED | LA_RECURSED:
1419	case LA_LOCKED | LA_NOTRECURSED:
1420	case LA_SLOCKED:
1421	case LA_SLOCKED | LA_RECURSED:
1422	case LA_SLOCKED | LA_NOTRECURSED:
1423	case LA_XLOCKED:
1424	case LA_XLOCKED | LA_RECURSED:
1425	case LA_XLOCKED | LA_NOTRECURSED:
1426		if (instance == NULL) {
1427			panic("Lock (%s) %s not locked @ %s:%d.",
1428			    lock->lo_class->lc_name, lock->lo_name, file, line);
1429			break;
1430		}
1431		if ((flags & LA_XLOCKED) != 0 &&
1432		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1433			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1434			    lock->lo_class->lc_name, lock->lo_name, file, line);
1435		if ((flags & LA_SLOCKED) != 0 &&
1436		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1437			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1438			    lock->lo_class->lc_name, lock->lo_name, file, line);
1439		if ((flags & LA_RECURSED) != 0 &&
1440		    (instance->li_flags & LI_RECURSEMASK) == 0)
1441			panic("Lock (%s) %s not recursed @ %s:%d.",
1442			    lock->lo_class->lc_name, lock->lo_name, file, line);
1443		if ((flags & LA_NOTRECURSED) != 0 &&
1444		    (instance->li_flags & LI_RECURSEMASK) != 0)
1445			panic("Lock (%s) %s recursed @ %s:%d.",
1446			    lock->lo_class->lc_name, lock->lo_name, file, line);
1447		break;
1448	default:
1449		panic("Invalid lock assertion at %s:%d.", file, line);
1450
1451	}
1452#endif	/* INVARIANT_SUPPORT */
1453}
1454
1455#ifdef DDB
1456
1457DB_SHOW_COMMAND(locks, db_witness_list)
1458{
1459	struct thread *td;
1460	pid_t pid;
1461	struct proc *p;
1462
1463	if (have_addr) {
1464		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1465		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1466		    ((addr >> 16) % 16) * 10000;
1467		/* sx_slock(&allproc_lock); */
1468		FOREACH_PROC_IN_SYSTEM(p) {
1469			if (p->p_pid == pid)
1470				break;
1471		}
1472		/* sx_sunlock(&allproc_lock); */
1473		if (p == NULL) {
1474			db_printf("pid %d not found\n", pid);
1475			return;
1476		}
1477		td = &p->p_thread; /* XXXKSE */
1478	} else {
1479		td = curthread;
1480	}
1481	witness_list(td);
1482}
1483
1484DB_SHOW_COMMAND(witness, db_witness_display)
1485{
1486
1487	witness_display(db_printf);
1488}
1489#endif
1490