subr_witness.c revision 99862
1280849Scy/*-
2280849Scy * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3280849Scy *
4280849Scy * Redistribution and use in source and binary forms, with or without
554359Sroberto * modification, are permitted provided that the following conditions
654359Sroberto * are met:
7280849Scy * 1. Redistributions of source code must retain the above copyright
8316722Sdelphij *    notice, this list of conditions and the following disclaimer.
9316722Sdelphij * 2. Redistributions in binary form must reproduce the above copyright
10280849Scy *    notice, this list of conditions and the following disclaimer in the
11280849Scy *    documentation and/or other materials provided with the distribution.
12280849Scy * 3. Berkeley Software Design Inc's name may not be used to endorse or
13316722Sdelphij *    promote products derived from this software without specific prior
14280849Scy *    written permission.
15316722Sdelphij *
16316722Sdelphij * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17280849Scy * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18280849Scy * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
1954359Sroberto * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
2082498Sroberto * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
2154359Sroberto * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22132451Sroberto * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23132451Sroberto * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24132451Sroberto * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25132451Sroberto * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26280849Scy * SUCH DAMAGE.
27280849Scy *
28280849Scy *	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 99862 2002-07-12 07:56:11Z peter $
31 */
32
33/*
34 * Implementation of the `witness' lock verifier.  Originally implemented for
35 * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36 * classes in FreeBSD.
37 */
38
39/*
40 *	Main Entry: witness
41 *	Pronunciation: 'wit-n&s
42 *	Function: noun
43 *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44 *	    testimony, witness, from 2wit
45 *	Date: before 12th century
46 *	1 : attestation of a fact or event : TESTIMONY
47 *	2 : one that gives evidence; specifically : one who testifies in
48 *	    a cause or before a judicial tribunal
49 *	3 : one asked to be present at a transaction so as to be able to
50 *	    testify to its having taken place
51 *	4 : one who has personal knowledge of something
52 *	5 a : something serving as evidence or proof : SIGN
53 *	  b : public affirmation by word or example of usually
54 *	      religious faith or conviction <the heroic witness to divine
55 *	      life -- Pilot>
56 *	6 capitalized : a member of the Jehovah's Witnesses
57 */
58
59#include "opt_ddb.h"
60#include "opt_witness.h"
61
62#include <sys/param.h>
63#include <sys/bus.h>
64#include <sys/kernel.h>
65#include <sys/ktr.h>
66#include <sys/lock.h>
67#include <sys/malloc.h>
68#include <sys/mutex.h>
69#include <sys/proc.h>
70#include <sys/sysctl.h>
71#include <sys/systm.h>
72
73#include <ddb/ddb.h>
74
75#define WITNESS_COUNT 200
76#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77/*
78 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79 * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80 * probably be safe for the most part, but it's still a SWAG.
81 */
82#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83
84#define	WITNESS_NCHILDREN 6
85
86struct witness_child_list_entry;
87
88struct witness {
89	const	char *w_name;
90	struct	lock_class *w_class;
91	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94	const	char *w_file;
95	int	w_line;
96	u_int	w_level;
97	u_int	w_refcount;
98	u_char	w_Giant_squawked:1;
99	u_char	w_other_squawked:1;
100	u_char	w_same_squawked:1;
101};
102
103struct witness_child_list_entry {
104	struct	witness_child_list_entry *wcl_next;
105	struct	witness *wcl_children[WITNESS_NCHILDREN];
106	u_int	wcl_count;
107};
108
109STAILQ_HEAD(witness_list, witness);
110
111struct witness_blessed {
112	const	char *b_lock1;
113	const	char *b_lock2;
114};
115
116struct witness_order_list_entry {
117	const	char *w_name;
118	struct	lock_class *w_class;
119};
120
121static struct	witness *enroll(const char *description,
122				struct lock_class *lock_class);
123static int	itismychild(struct witness *parent, struct witness *child);
124static void	removechild(struct witness *parent, struct witness *child);
125static int	isitmychild(struct witness *parent, struct witness *child);
126static int	isitmydescendant(struct witness *parent, struct witness *child);
127static int	blessed(struct witness *, struct witness *);
128static void	witness_display_list(void(*prnt)(const char *fmt, ...),
129				     struct witness_list *list);
130static void	witness_displaydescendants(void(*)(const char *fmt, ...),
131					   struct witness *);
132static void	witness_leveldescendents(struct witness *parent, int level);
133static void	witness_levelall(void);
134static struct	witness *witness_get(void);
135static void	witness_free(struct witness *m);
136static struct	witness_child_list_entry *witness_child_get(void);
137static void	witness_child_free(struct witness_child_list_entry *wcl);
138static struct	lock_list_entry *witness_lock_list_get(void);
139static void	witness_lock_list_free(struct lock_list_entry *lle);
140static void	witness_display(void(*)(const char *fmt, ...));
141static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
142					     struct lock_object *lock);
143
144MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
145
146static int witness_watch = 1;
147TUNABLE_INT("debug.witness_watch", &witness_watch);
148SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
149
150#ifdef DDB
151/*
152 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
153 * drop into kdebug() when:
154 *	- a lock heirarchy violation occurs
155 *	- locks are held when going to sleep.
156 */
157#ifdef WITNESS_DDB
158int	witness_ddb = 1;
159#else
160int	witness_ddb = 0;
161#endif
162TUNABLE_INT("debug.witness_ddb", &witness_ddb);
163SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
164#endif /* DDB */
165
166#ifdef WITNESS_SKIPSPIN
167int	witness_skipspin = 1;
168#else
169int	witness_skipspin = 0;
170#endif
171TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
172SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
173    "");
174
175static struct mtx w_mtx;
176static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
177static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
178static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
179static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
180static struct witness_child_list_entry *w_child_free = NULL;
181static struct lock_list_entry *w_lock_list_free = NULL;
182static int witness_dead;	/* fatal error, probably no memory */
183
184static struct witness w_data[WITNESS_COUNT];
185static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
186static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
187
188static struct witness_order_list_entry order_lists[] = {
189	{ "Giant", &lock_class_mtx_sleep },
190	{ "proctree", &lock_class_sx },
191	{ "allproc", &lock_class_sx },
192	{ "sigio lock", &lock_class_mtx_sleep },
193	{ "process group", &lock_class_mtx_sleep },
194	{ "process lock", &lock_class_mtx_sleep },
195	{ "session", &lock_class_mtx_sleep },
196	{ "uidinfo hash", &lock_class_mtx_sleep },
197	{ "uidinfo struct", &lock_class_mtx_sleep },
198	{ NULL, NULL },
199	/*
200	 * spin locks
201	 */
202#ifdef SMP
203	{ "ap boot", &lock_class_mtx_spin },
204#ifdef __i386__
205	{ "com", &lock_class_mtx_spin },
206#endif
207#endif
208	{ "sio", &lock_class_mtx_spin },
209#ifdef __i386__
210	{ "cy", &lock_class_mtx_spin },
211#endif
212	{ "ng_node", &lock_class_mtx_spin },
213	{ "ng_worklist", &lock_class_mtx_spin },
214	{ "ithread table lock", &lock_class_mtx_spin },
215	{ "sched lock", &lock_class_mtx_spin },
216	{ "callout", &lock_class_mtx_spin },
217	/*
218	 * leaf locks
219	 */
220	{ "allpmaps", &lock_class_mtx_spin },
221	{ "vm page buckets mutex", &lock_class_mtx_spin },
222	{ "vm page queue free mutex", &lock_class_mtx_spin },
223	{ "icu", &lock_class_mtx_spin },
224#ifdef SMP
225	{ "smp rendezvous", &lock_class_mtx_spin },
226#if defined(__i386__) && defined(APIC_IO)
227	{ "tlb", &lock_class_mtx_spin },
228#endif
229#endif
230	{ "clk", &lock_class_mtx_spin },
231	{ "mutex profiling lock", &lock_class_mtx_spin },
232	{ "zombie_thread_lock", &lock_class_mtx_spin },
233	{ NULL, NULL },
234	{ NULL, NULL }
235};
236
237/*
238 * Pairs of locks which have been blessed
239 * Don't complain about order problems with blessed locks
240 */
241static struct witness_blessed blessed_list[] = {
242};
243static int blessed_count =
244	sizeof(blessed_list) / sizeof(struct witness_blessed);
245
246/*
247 * List of all locks in the system.
248 */
249TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
250
251static struct mtx all_mtx = {
252	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
253	  "All locks list",		/* mtx_object.lo_name */
254	  "All locks list",		/* mtx_object.lo_type */
255	  LO_INITIALIZED,		/* mtx_object.lo_flags */
256	  { NULL, NULL },		/* mtx_object.lo_list */
257	  NULL },			/* mtx_object.lo_witness */
258	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
259	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
260	{ NULL, NULL }			/* mtx_contested */
261};
262
263/*
264 * This global is set to 0 once it becomes safe to use the witness code.
265 */
266static int witness_cold = 1;
267
268/*
269 * Global variables for book keeping.
270 */
271static int lock_cur_cnt;
272static int lock_max_cnt;
273
274/*
275 * The WITNESS-enabled diagnostic code.
276 */
277static void
278witness_initialize(void *dummy __unused)
279{
280	struct lock_object *lock;
281	struct witness_order_list_entry *order;
282	struct witness *w, *w1;
283	int i;
284
285	/*
286	 * We have to release Giant before initializing its witness
287	 * structure so that WITNESS doesn't get confused.
288	 */
289	mtx_unlock(&Giant);
290	mtx_assert(&Giant, MA_NOTOWNED);
291
292	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
293	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
294	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
295	    MTX_NOWITNESS);
296	for (i = 0; i < WITNESS_COUNT; i++)
297		witness_free(&w_data[i]);
298	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
299		witness_child_free(&w_childdata[i]);
300	for (i = 0; i < LOCK_CHILDCOUNT; i++)
301		witness_lock_list_free(&w_locklistdata[i]);
302
303	/* First add in all the specified order lists. */
304	for (order = order_lists; order->w_name != NULL; order++) {
305		w = enroll(order->w_name, order->w_class);
306		if (w == NULL)
307			continue;
308		w->w_file = "order list";
309		for (order++; order->w_name != NULL; order++) {
310			w1 = enroll(order->w_name, order->w_class);
311			if (w1 == NULL)
312				continue;
313			w1->w_file = "order list";
314			itismychild(w, w1);
315			w = w1;
316		}
317	}
318
319	/* Iterate through all locks and add them to witness. */
320	mtx_lock(&all_mtx);
321	TAILQ_FOREACH(lock, &all_locks, lo_list) {
322		if (lock->lo_flags & LO_WITNESS)
323			lock->lo_witness = enroll(lock->lo_type,
324			    lock->lo_class);
325		else
326			lock->lo_witness = NULL;
327	}
328	mtx_unlock(&all_mtx);
329
330	/* Mark the witness code as being ready for use. */
331	atomic_store_rel_int(&witness_cold, 0);
332
333	mtx_lock(&Giant);
334}
335SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
336
337void
338witness_init(struct lock_object *lock)
339{
340	struct lock_class *class;
341
342	class = lock->lo_class;
343	if (lock->lo_flags & LO_INITIALIZED)
344		panic("%s: lock (%s) %s is already initialized", __func__,
345		    class->lc_name, lock->lo_name);
346	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
347	    (class->lc_flags & LC_RECURSABLE) == 0)
348		panic("%s: lock (%s) %s can not be recursable", __func__,
349		    class->lc_name, lock->lo_name);
350	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
351	    (class->lc_flags & LC_SLEEPABLE) == 0)
352		panic("%s: lock (%s) %s can not be sleepable", __func__,
353		    class->lc_name, lock->lo_name);
354	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
355	    (class->lc_flags & LC_UPGRADABLE) == 0)
356		panic("%s: lock (%s) %s can not be upgradable", __func__,
357		    class->lc_name, lock->lo_name);
358
359	mtx_lock(&all_mtx);
360	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
361	lock->lo_flags |= LO_INITIALIZED;
362	lock_cur_cnt++;
363	if (lock_cur_cnt > lock_max_cnt)
364		lock_max_cnt = lock_cur_cnt;
365	mtx_unlock(&all_mtx);
366	if (!witness_cold && !witness_dead && panicstr == NULL &&
367	    (lock->lo_flags & LO_WITNESS) != 0)
368		lock->lo_witness = enroll(lock->lo_type, class);
369	else
370		lock->lo_witness = NULL;
371}
372
373void
374witness_destroy(struct lock_object *lock)
375{
376	struct witness *w;
377
378	if (witness_cold)
379		panic("lock (%s) %s destroyed while witness_cold",
380		    lock->lo_class->lc_name, lock->lo_name);
381	if ((lock->lo_flags & LO_INITIALIZED) == 0)
382		panic("%s: lock (%s) %s is not initialized", __func__,
383		    lock->lo_class->lc_name, lock->lo_name);
384
385	/* XXX: need to verify that no one holds the lock */
386	w = lock->lo_witness;
387	if (w != NULL) {
388		mtx_lock_spin(&w_mtx);
389		MPASS(w->w_refcount > 0);
390		w->w_refcount--;
391		mtx_unlock_spin(&w_mtx);
392	}
393
394	mtx_lock(&all_mtx);
395	lock_cur_cnt--;
396	TAILQ_REMOVE(&all_locks, lock, lo_list);
397	lock->lo_flags &= ~LO_INITIALIZED;
398	mtx_unlock(&all_mtx);
399}
400
401static void
402witness_display_list(void(*prnt)(const char *fmt, ...),
403		     struct witness_list *list)
404{
405	struct witness *w, *w1;
406	int found;
407
408	STAILQ_FOREACH(w, list, w_typelist) {
409		if (w->w_file == NULL)
410			continue;
411		found = 0;
412		STAILQ_FOREACH(w1, list, w_typelist) {
413			if (isitmychild(w1, w)) {
414				found++;
415				break;
416			}
417		}
418		if (found)
419			continue;
420		/*
421		 * This lock has no anscestors, display its descendants.
422		 */
423		witness_displaydescendants(prnt, w);
424	}
425}
426
427static void
428witness_display(void(*prnt)(const char *fmt, ...))
429{
430	struct witness *w;
431
432	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
433	witness_levelall();
434
435	/*
436	 * First, handle sleep locks which have been acquired at least
437	 * once.
438	 */
439	prnt("Sleep locks:\n");
440	witness_display_list(prnt, &w_sleep);
441
442	/*
443	 * Now do spin locks which have been acquired at least once.
444	 */
445	prnt("\nSpin locks:\n");
446	witness_display_list(prnt, &w_spin);
447
448	/*
449	 * Finally, any locks which have not been acquired yet.
450	 */
451	prnt("\nLocks which were never acquired:\n");
452	STAILQ_FOREACH(w, &w_all, w_list) {
453		if (w->w_file != NULL || w->w_refcount == 0)
454			continue;
455		prnt("%s\n", w->w_name);
456	}
457}
458
459void
460witness_lock(struct lock_object *lock, int flags, const char *file, int line)
461{
462	struct lock_list_entry **lock_list, *lle;
463	struct lock_instance *lock1, *lock2;
464	struct lock_class *class;
465	struct witness *w, *w1;
466	struct thread *td;
467	int i, j;
468#ifdef DDB
469	int go_into_ddb = 0;
470#endif /* DDB */
471
472	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
473	    panicstr != NULL)
474		return;
475	w = lock->lo_witness;
476	class = lock->lo_class;
477	td = curthread;
478
479	if (class->lc_flags & LC_SLEEPLOCK) {
480		/*
481		 * Since spin locks include a critical section, this check
482		 * impliclty enforces a lock order of all sleep locks before
483		 * all spin locks.
484		 */
485		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
486			panic("blockable sleep lock (%s) %s @ %s:%d",
487			    class->lc_name, lock->lo_name, file, line);
488		lock_list = &td->td_sleeplocks;
489	} else
490		lock_list = PCPU_PTR(spinlocks);
491
492	/*
493	 * Try locks do not block if they fail to acquire the lock, thus
494	 * there is no danger of deadlocks or of switching while holding a
495	 * spin lock if we acquire a lock via a try operation.
496	 */
497	if (flags & LOP_TRYLOCK)
498		goto out;
499
500	/*
501	 * Is this the first lock acquired?  If so, then no order checking
502	 * is needed.
503	 */
504	if (*lock_list == NULL)
505		goto out;
506
507	/*
508	 * Check to see if we are recursing on a lock we already own.
509	 */
510	lock1 = find_instance(*lock_list, lock);
511	if (lock1 != NULL) {
512		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
513		    (flags & LOP_EXCLUSIVE) == 0) {
514			printf("shared lock of (%s) %s @ %s:%d\n",
515			    class->lc_name, lock->lo_name, file, line);
516			printf("while exclusively locked from %s:%d\n",
517			    lock1->li_file, lock1->li_line);
518			panic("share->excl");
519		}
520		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
521		    (flags & LOP_EXCLUSIVE) != 0) {
522			printf("exclusive lock of (%s) %s @ %s:%d\n",
523			    class->lc_name, lock->lo_name, file, line);
524			printf("while share locked from %s:%d\n",
525			    lock1->li_file, lock1->li_line);
526			panic("excl->share");
527		}
528		lock1->li_flags++;
529		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
530			printf(
531			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
532			    class->lc_name, lock->lo_name, file, line);
533			printf("first acquired @ %s:%d\n", lock1->li_file,
534			    lock1->li_line);
535			panic("recurse");
536		}
537		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
538		    td->td_proc->p_pid, lock->lo_name,
539		    lock1->li_flags & LI_RECURSEMASK);
540		lock1->li_file = file;
541		lock1->li_line = line;
542		return;
543	}
544
545	/*
546	 * Check for duplicate locks of the same type.  Note that we only
547	 * have to check for this on the last lock we just acquired.  Any
548	 * other cases will be caught as lock order violations.
549	 */
550	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
551	w1 = lock1->li_lock->lo_witness;
552	if (w1 == w) {
553		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
554			goto out;
555		w->w_same_squawked = 1;
556		printf("acquiring duplicate lock of same type: \"%s\"\n",
557			lock->lo_type);
558		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
559		    lock1->li_file, lock1->li_line);
560		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
561#ifdef DDB
562		go_into_ddb = 1;
563#endif /* DDB */
564		goto out;
565	}
566	MPASS(!mtx_owned(&w_mtx));
567	mtx_lock_spin(&w_mtx);
568	/*
569	 * If we have a known higher number just say ok
570	 */
571	if (witness_watch > 1 && w->w_level > w1->w_level) {
572		mtx_unlock_spin(&w_mtx);
573		goto out;
574	}
575	if (isitmydescendant(w1, w)) {
576		mtx_unlock_spin(&w_mtx);
577		goto out;
578	}
579	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
580		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
581
582			MPASS(j < WITNESS_COUNT);
583			lock1 = &lle->ll_children[i];
584			w1 = lock1->li_lock->lo_witness;
585
586			/*
587			 * If this lock doesn't undergo witness checking,
588			 * then skip it.
589			 */
590			if (w1 == NULL) {
591				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
592				    ("lock missing witness structure"));
593				continue;
594			}
595			/*
596			 * If we are locking Giant and we slept with this
597			 * lock, then skip it.
598			 */
599			if ((lock1->li_flags & LI_SLEPT) != 0 &&
600			    lock == &Giant.mtx_object)
601				continue;
602			/*
603			 * If we are locking a sleepable lock and this lock
604			 * isn't sleepable and isn't Giant, we want to treat
605			 * it as a lock order violation to enfore a general
606			 * lock order of sleepable locks before non-sleepable
607			 * locks.  Thus, we only bother checking the lock
608			 * order hierarchy if we pass the initial test.
609			 */
610			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
611			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
612			    lock1->li_lock != &Giant.mtx_object)) &&
613			    !isitmydescendant(w, w1))
614				continue;
615			/*
616			 * We have a lock order violation, check to see if it
617			 * is allowed or has already been yelled about.
618			 */
619			mtx_unlock_spin(&w_mtx);
620			if (blessed(w, w1))
621				goto out;
622			if (lock1->li_lock == &Giant.mtx_object) {
623				if (w1->w_Giant_squawked)
624					goto out;
625				else
626					w1->w_Giant_squawked = 1;
627			} else {
628				if (w1->w_other_squawked)
629					goto out;
630				else
631					w1->w_other_squawked = 1;
632			}
633			/*
634			 * Ok, yell about it.
635			 */
636			printf("lock order reversal\n");
637			/*
638			 * Try to locate an earlier lock with
639			 * witness w in our list.
640			 */
641			do {
642				lock2 = &lle->ll_children[i];
643				MPASS(lock2->li_lock != NULL);
644				if (lock2->li_lock->lo_witness == w)
645					break;
646				i--;
647				if (i == 0 && lle->ll_next != NULL) {
648					lle = lle->ll_next;
649					i = lle->ll_count - 1;
650					MPASS(i != 0);
651				}
652			} while (i >= 0);
653			if (i < 0) {
654				printf(" 1st %p %s (%s) @ %s:%d\n",
655				    lock1->li_lock, lock1->li_lock->lo_name,
656				    lock1->li_lock->lo_type, lock1->li_file,
657				    lock1->li_line);
658				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
659				    lock->lo_name, lock->lo_type, file, line);
660			} else {
661				printf(" 1st %p %s (%s) @ %s:%d\n",
662				    lock2->li_lock, lock2->li_lock->lo_name,
663				    lock2->li_lock->lo_type, lock2->li_file,
664				    lock2->li_line);
665				printf(" 2nd %p %s (%s) @ %s:%d\n",
666				    lock1->li_lock, lock1->li_lock->lo_name,
667				    lock1->li_lock->lo_type, lock1->li_file,
668				    lock1->li_line);
669				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
670				    lock->lo_name, lock->lo_type, file, line);
671			}
672#ifdef DDB
673			go_into_ddb = 1;
674#endif /* DDB */
675			goto out;
676		}
677	}
678	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
679	/*
680	 * Don't build a new relationship if we are locking Giant just
681	 * after waking up and the previous lock in the list was acquired
682	 * prior to blocking.
683	 */
684	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
685		mtx_unlock_spin(&w_mtx);
686	else {
687		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
688		    lock->lo_type, lock1->li_lock->lo_type);
689		if (!itismychild(lock1->li_lock->lo_witness, w))
690			mtx_unlock_spin(&w_mtx);
691	}
692
693out:
694#ifdef DDB
695	if (witness_ddb && go_into_ddb)
696		Debugger(__func__);
697#endif /* DDB */
698	w->w_file = file;
699	w->w_line = line;
700
701	lle = *lock_list;
702	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
703		lle = witness_lock_list_get();
704		if (lle == NULL)
705			return;
706		lle->ll_next = *lock_list;
707		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
708		    td->td_proc->p_pid, lle);
709		*lock_list = lle;
710	}
711	lock1 = &lle->ll_children[lle->ll_count++];
712	lock1->li_lock = lock;
713	lock1->li_line = line;
714	lock1->li_file = file;
715	if ((flags & LOP_EXCLUSIVE) != 0)
716		lock1->li_flags = LI_EXCLUSIVE;
717	else
718		lock1->li_flags = 0;
719	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
720	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
721}
722
723void
724witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
725{
726	struct lock_instance *instance;
727	struct lock_class *class;
728
729	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
730	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
731		return;
732	class = lock->lo_class;
733	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
734		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
735		    class->lc_name, lock->lo_name, file, line);
736	if ((flags & LOP_TRYLOCK) == 0)
737		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
738		    lock->lo_name, file, line);
739	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
740		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
741		    class->lc_name, lock->lo_name, file, line);
742	instance = find_instance(curthread->td_sleeplocks, lock);
743	if (instance == NULL)
744		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
745		    class->lc_name, lock->lo_name, file, line);
746	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
747		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
748		    class->lc_name, lock->lo_name, file, line);
749	if ((instance->li_flags & LI_RECURSEMASK) != 0)
750		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
751		    class->lc_name, lock->lo_name,
752		    instance->li_flags & LI_RECURSEMASK, file, line);
753	instance->li_flags |= LI_EXCLUSIVE;
754}
755
756void
757witness_downgrade(struct lock_object *lock, int flags, const char *file,
758    int line)
759{
760	struct lock_instance *instance;
761	struct lock_class *class;
762
763	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
764	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
765		return;
766	class = lock->lo_class;
767	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
768		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
769		    class->lc_name, lock->lo_name, file, line);
770	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
771		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
772		    class->lc_name, lock->lo_name, file, line);
773	instance = find_instance(curthread->td_sleeplocks, lock);
774	if (instance == NULL)
775		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
776		    class->lc_name, lock->lo_name, file, line);
777	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
778		panic("downgrade of shared lock (%s) %s @ %s:%d",
779		    class->lc_name, lock->lo_name, file, line);
780	if ((instance->li_flags & LI_RECURSEMASK) != 0)
781		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
782		    class->lc_name, lock->lo_name,
783		    instance->li_flags & LI_RECURSEMASK, file, line);
784	instance->li_flags &= ~LI_EXCLUSIVE;
785}
786
787void
788witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
789{
790	struct lock_list_entry **lock_list, *lle;
791	struct lock_instance *instance;
792	struct lock_class *class;
793	struct thread *td;
794	register_t s;
795	int i, j;
796
797	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
798	    panicstr != NULL)
799		return;
800	td = curthread;
801	class = lock->lo_class;
802	if (class->lc_flags & LC_SLEEPLOCK)
803		lock_list = &td->td_sleeplocks;
804	else
805		lock_list = PCPU_PTR(spinlocks);
806	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
807		for (i = 0; i < (*lock_list)->ll_count; i++) {
808			instance = &(*lock_list)->ll_children[i];
809			if (instance->li_lock == lock) {
810				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
811				    (flags & LOP_EXCLUSIVE) == 0) {
812					printf(
813					"shared unlock of (%s) %s @ %s:%d\n",
814					    class->lc_name, lock->lo_name,
815					    file, line);
816					printf(
817					"while exclusively locked from %s:%d\n",
818					    instance->li_file,
819					    instance->li_line);
820					panic("excl->ushare");
821				}
822				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
823				    (flags & LOP_EXCLUSIVE) != 0) {
824					printf(
825					"exclusive unlock of (%s) %s @ %s:%d\n",
826					    class->lc_name, lock->lo_name,
827					    file, line);
828					printf(
829					"while share locked from %s:%d\n",
830					    instance->li_file,
831					    instance->li_line);
832					panic("share->uexcl");
833				}
834				/* If we are recursed, unrecurse. */
835				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
836					CTR4(KTR_WITNESS,
837				    "%s: pid %d unrecursed on %s r=%d", __func__,
838					    td->td_proc->p_pid,
839					    instance->li_lock->lo_name,
840					    instance->li_flags);
841					instance->li_flags--;
842					return;
843				}
844				s = intr_disable();
845				CTR4(KTR_WITNESS,
846				    "%s: pid %d removed %s from lle[%d]", __func__,
847				    td->td_proc->p_pid,
848				    instance->li_lock->lo_name,
849				    (*lock_list)->ll_count - 1);
850				for (j = i; j < (*lock_list)->ll_count - 1; j++)
851					(*lock_list)->ll_children[j] =
852					    (*lock_list)->ll_children[j + 1];
853				(*lock_list)->ll_count--;
854				intr_restore(s);
855				if ((*lock_list)->ll_count == 0) {
856					lle = *lock_list;
857					*lock_list = lle->ll_next;
858					CTR3(KTR_WITNESS,
859					    "%s: pid %d removed lle %p", __func__,
860					    td->td_proc->p_pid, lle);
861					witness_lock_list_free(lle);
862				}
863				return;
864			}
865		}
866	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
867	    file, line);
868}
869
870/*
871 * Warn if any held locks are not sleepable.  Note that Giant and the lock
872 * passed in are both special cases since they are both released during the
873 * sleep process and aren't actually held while the thread is asleep.
874 */
875int
876witness_sleep(int check_only, struct lock_object *lock, const char *file,
877	      int line)
878{
879	struct lock_list_entry **lock_list, *lle;
880	struct lock_instance *lock1;
881	struct thread *td;
882	int i, n;
883
884	if (witness_cold || witness_dead || panicstr != NULL)
885		return (0);
886	n = 0;
887	td = curthread;
888	lock_list = &td->td_sleeplocks;
889again:
890	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
891		for (i = lle->ll_count - 1; i >= 0; i--) {
892			lock1 = &lle->ll_children[i];
893			if (lock1->li_lock == lock ||
894			    lock1->li_lock == &Giant.mtx_object)
895				continue;
896			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
897				if (check_only == 0) {
898					CTR3(KTR_WITNESS,
899				    "pid %d: sleeping with lock (%s) %s held",
900					    td->td_proc->p_pid,
901					    lock1->li_lock->lo_class->lc_name,
902					    lock1->li_lock->lo_name);
903					lock1->li_flags |= LI_SLEPT;
904				}
905				continue;
906			}
907			n++;
908			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
909			    file, line, check_only ? "could sleep" : "sleeping",
910			    lock1->li_lock->lo_name, lock1->li_file,
911			    lock1->li_line);
912		}
913	if (lock_list == &td->td_sleeplocks && PCPU_GET(spinlocks) != NULL) {
914		/*
915		 * Since we already hold a spinlock preemption is
916		 * already blocked.
917		 */
918		lock_list = PCPU_PTR(spinlocks);
919		goto again;
920	}
921#ifdef DDB
922	if (witness_ddb && n)
923		Debugger(__func__);
924#endif /* DDB */
925	return (n);
926}
927
928static struct witness *
929enroll(const char *description, struct lock_class *lock_class)
930{
931	struct witness *w;
932
933	if (!witness_watch || witness_dead || panicstr != NULL)
934		return (NULL);
935	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
936		return (NULL);
937	mtx_lock_spin(&w_mtx);
938	STAILQ_FOREACH(w, &w_all, w_list) {
939		if (w->w_name == description || (w->w_refcount > 0 &&
940		    strcmp(description, w->w_name) == 0)) {
941			w->w_refcount++;
942			mtx_unlock_spin(&w_mtx);
943			if (lock_class != w->w_class)
944				panic(
945				"lock (%s) %s does not match earlier (%s) lock",
946				    description, lock_class->lc_name,
947				    w->w_class->lc_name);
948			return (w);
949		}
950	}
951	/*
952	 * This isn't quite right, as witness_cold is still 0 while we
953	 * enroll all the locks initialized before witness_initialize().
954	 */
955	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
956		mtx_unlock_spin(&w_mtx);
957		panic("spin lock %s not in order list", description);
958	}
959	if ((w = witness_get()) == NULL)
960		return (NULL);
961	w->w_name = description;
962	w->w_class = lock_class;
963	w->w_refcount = 1;
964	STAILQ_INSERT_HEAD(&w_all, w, w_list);
965	if (lock_class->lc_flags & LC_SPINLOCK)
966		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
967	else if (lock_class->lc_flags & LC_SLEEPLOCK)
968		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
969	else {
970		mtx_unlock_spin(&w_mtx);
971		panic("lock class %s is not sleep or spin",
972		    lock_class->lc_name);
973	}
974	mtx_unlock_spin(&w_mtx);
975	return (w);
976}
977
978static int
979itismychild(struct witness *parent, struct witness *child)
980{
981	static int recursed;
982	struct witness_child_list_entry **wcl;
983	struct witness_list *list;
984
985	MPASS(child != NULL && parent != NULL);
986	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
987	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
988		panic(
989		"%s: parent (%s) and child (%s) are not the same lock type",
990		    __func__, parent->w_class->lc_name,
991		    child->w_class->lc_name);
992
993	/*
994	 * Insert "child" after "parent"
995	 */
996	wcl = &parent->w_children;
997	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
998		wcl = &(*wcl)->wcl_next;
999	if (*wcl == NULL) {
1000		*wcl = witness_child_get();
1001		if (*wcl == NULL)
1002			return (1);
1003	}
1004	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1005
1006	/*
1007	 * Now prune whole tree.  We look for cases where a lock is now
1008	 * both a descendant and a direct child of a given lock.  In that
1009	 * case, we want to remove the direct child link from the tree.
1010	 */
1011	if (recursed)
1012		return (0);
1013	recursed = 1;
1014	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1015		list = &w_sleep;
1016	else
1017		list = &w_spin;
1018	STAILQ_FOREACH(child, list, w_typelist) {
1019		STAILQ_FOREACH(parent, list, w_typelist) {
1020			if (!isitmychild(parent, child))
1021				continue;
1022			removechild(parent, child);
1023			if (isitmydescendant(parent, child))
1024				continue;
1025			itismychild(parent, child);
1026		}
1027	}
1028	recursed = 0;
1029	witness_levelall();
1030	return (0);
1031}
1032
1033static void
1034removechild(struct witness *parent, struct witness *child)
1035{
1036	struct witness_child_list_entry **wcl, *wcl1;
1037	int i;
1038
1039	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1040		for (i = 0; i < (*wcl)->wcl_count; i++)
1041			if ((*wcl)->wcl_children[i] == child)
1042				goto found;
1043	return;
1044found:
1045	(*wcl)->wcl_count--;
1046	if ((*wcl)->wcl_count > i)
1047		(*wcl)->wcl_children[i] =
1048		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1049	MPASS((*wcl)->wcl_children[i] != NULL);
1050	if ((*wcl)->wcl_count != 0)
1051		return;
1052	wcl1 = *wcl;
1053	*wcl = wcl1->wcl_next;
1054	witness_child_free(wcl1);
1055}
1056
1057static int
1058isitmychild(struct witness *parent, struct witness *child)
1059{
1060	struct witness_child_list_entry *wcl;
1061	int i;
1062
1063	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1064		for (i = 0; i < wcl->wcl_count; i++) {
1065			if (wcl->wcl_children[i] == child)
1066				return (1);
1067		}
1068	}
1069	return (0);
1070}
1071
1072static int
1073isitmydescendant(struct witness *parent, struct witness *child)
1074{
1075	struct witness_child_list_entry *wcl;
1076	int i, j;
1077
1078	if (isitmychild(parent, child))
1079		return (1);
1080	j = 0;
1081	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1082		MPASS(j < 1000);
1083		for (i = 0; i < wcl->wcl_count; i++) {
1084			if (isitmydescendant(wcl->wcl_children[i], child))
1085				return (1);
1086		}
1087		j++;
1088	}
1089	return (0);
1090}
1091
1092void
1093witness_levelall (void)
1094{
1095	struct witness_list *list;
1096	struct witness *w, *w1;
1097
1098	/*
1099	 * First clear all levels.
1100	 */
1101	STAILQ_FOREACH(w, &w_all, w_list) {
1102		w->w_level = 0;
1103	}
1104
1105	/*
1106	 * Look for locks with no parent and level all their descendants.
1107	 */
1108	STAILQ_FOREACH(w, &w_all, w_list) {
1109		/*
1110		 * This is just an optimization, technically we could get
1111		 * away just walking the all list each time.
1112		 */
1113		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1114			list = &w_sleep;
1115		else
1116			list = &w_spin;
1117		STAILQ_FOREACH(w1, list, w_typelist) {
1118			if (isitmychild(w1, w))
1119				goto skip;
1120		}
1121		witness_leveldescendents(w, 0);
1122	skip:
1123		;	/* silence GCC 3.x */
1124	}
1125}
1126
1127static void
1128witness_leveldescendents(struct witness *parent, int level)
1129{
1130	struct witness_child_list_entry *wcl;
1131	int i;
1132
1133	if (parent->w_level < level)
1134		parent->w_level = level;
1135	level++;
1136	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1137		for (i = 0; i < wcl->wcl_count; i++)
1138			witness_leveldescendents(wcl->wcl_children[i], level);
1139}
1140
1141static void
1142witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1143			   struct witness *parent)
1144{
1145	struct witness_child_list_entry *wcl;
1146	int i, level;
1147
1148	level = parent->w_level;
1149	prnt("%-2d", level);
1150	for (i = 0; i < level; i++)
1151		prnt(" ");
1152	if (parent->w_refcount > 0) {
1153		prnt("%s", parent->w_name);
1154		if (parent->w_file != NULL)
1155			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1156			    parent->w_line);
1157	} else
1158		prnt("(dead)\n");
1159	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1160		for (i = 0; i < wcl->wcl_count; i++)
1161			    witness_displaydescendants(prnt,
1162				wcl->wcl_children[i]);
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",
1303			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1304			    "exclusive" : "shared",
1305			    lock->lo_class->lc_name, lock->lo_name);
1306			if (lock->lo_type != lock->lo_name)
1307				printf(" (%s)", lock->lo_type);
1308			printf(" r = %d (%p) locked @ %s:%d\n",
1309			    instance->li_flags & LI_RECURSEMASK, lock,
1310			    instance->li_file, instance->li_line);
1311			nheld++;
1312		}
1313	return (nheld);
1314}
1315
1316/*
1317 * Calling this on td != curthread is bad unless we are in ddb.
1318 */
1319int
1320witness_list(struct thread *td)
1321{
1322	int nheld;
1323
1324	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1325#ifdef DDB
1326	KASSERT(td == curthread || db_active,
1327	    ("%s: td != curthread and we aren't in the debugger", __func__));
1328	if (!db_active && witness_dead)
1329		return (0);
1330#else
1331	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1332	if (witness_dead)
1333		return (0);
1334#endif
1335	nheld = witness_list_locks(&td->td_sleeplocks);
1336
1337	/*
1338	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1339	 * if td is currently executing on some other CPU and holds spin locks
1340	 * as we won't display those locks.  If we had a MI way of getting
1341	 * the per-cpu data for a given cpu then we could use
1342	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1343	 * and "fix" this.
1344	 *
1345	 * That still wouldn't really fix this unless we locked sched_lock
1346	 * or stopped the other CPU to make sure it wasn't changing the list
1347	 * out from under us.  It is probably best to just not try to handle
1348	 * threads on other CPU's for now.
1349	 */
1350	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1351		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1352
1353	return (nheld);
1354}
1355
1356void
1357witness_save(struct lock_object *lock, const char **filep, int *linep)
1358{
1359	struct lock_instance *instance;
1360
1361	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1362	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1363		return;
1364	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1365		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1366		    lock->lo_class->lc_name, lock->lo_name);
1367	instance = find_instance(curthread->td_sleeplocks, lock);
1368	if (instance == NULL)
1369		panic("%s: lock (%s) %s not locked", __func__,
1370		    lock->lo_class->lc_name, lock->lo_name);
1371	*filep = instance->li_file;
1372	*linep = instance->li_line;
1373}
1374
1375void
1376witness_restore(struct lock_object *lock, const char *file, int line)
1377{
1378	struct lock_instance *instance;
1379
1380	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1381	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1382		return;
1383	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1384		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1385		    lock->lo_class->lc_name, lock->lo_name);
1386	instance = find_instance(curthread->td_sleeplocks, lock);
1387	if (instance == NULL)
1388		panic("%s: lock (%s) %s not locked", __func__,
1389		    lock->lo_class->lc_name, lock->lo_name);
1390	lock->lo_witness->w_file = file;
1391	lock->lo_witness->w_line = line;
1392	instance->li_file = file;
1393	instance->li_line = line;
1394}
1395
1396void
1397witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1398{
1399#ifdef INVARIANT_SUPPORT
1400	struct lock_instance *instance;
1401
1402	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1403		return;
1404	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1405		instance = find_instance(curthread->td_sleeplocks, lock);
1406	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1407		instance = find_instance(PCPU_GET(spinlocks), lock);
1408	else {
1409		panic("Lock (%s) %s is not sleep or spin!",
1410		    lock->lo_class->lc_name, lock->lo_name);
1411		return;
1412	}
1413	switch (flags) {
1414	case LA_UNLOCKED:
1415		if (instance != NULL)
1416			panic("Lock (%s) %s locked @ %s:%d.",
1417			    lock->lo_class->lc_name, lock->lo_name, file, line);
1418		break;
1419	case LA_LOCKED:
1420	case LA_LOCKED | LA_RECURSED:
1421	case LA_LOCKED | LA_NOTRECURSED:
1422	case LA_SLOCKED:
1423	case LA_SLOCKED | LA_RECURSED:
1424	case LA_SLOCKED | LA_NOTRECURSED:
1425	case LA_XLOCKED:
1426	case LA_XLOCKED | LA_RECURSED:
1427	case LA_XLOCKED | LA_NOTRECURSED:
1428		if (instance == NULL) {
1429			panic("Lock (%s) %s not locked @ %s:%d.",
1430			    lock->lo_class->lc_name, lock->lo_name, file, line);
1431			break;
1432		}
1433		if ((flags & LA_XLOCKED) != 0 &&
1434		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1435			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1436			    lock->lo_class->lc_name, lock->lo_name, file, line);
1437		if ((flags & LA_SLOCKED) != 0 &&
1438		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1439			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1440			    lock->lo_class->lc_name, lock->lo_name, file, line);
1441		if ((flags & LA_RECURSED) != 0 &&
1442		    (instance->li_flags & LI_RECURSEMASK) == 0)
1443			panic("Lock (%s) %s not recursed @ %s:%d.",
1444			    lock->lo_class->lc_name, lock->lo_name, file, line);
1445		if ((flags & LA_NOTRECURSED) != 0 &&
1446		    (instance->li_flags & LI_RECURSEMASK) != 0)
1447			panic("Lock (%s) %s recursed @ %s:%d.",
1448			    lock->lo_class->lc_name, lock->lo_name, file, line);
1449		break;
1450	default:
1451		panic("Invalid lock assertion at %s:%d.", file, line);
1452
1453	}
1454#endif	/* INVARIANT_SUPPORT */
1455}
1456
1457#ifdef DDB
1458
1459DB_SHOW_COMMAND(locks, db_witness_list)
1460{
1461	struct thread *td;
1462	pid_t pid;
1463	struct proc *p;
1464
1465	if (have_addr) {
1466		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1467		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1468		    ((addr >> 16) % 16) * 10000;
1469		/* sx_slock(&allproc_lock); */
1470		FOREACH_PROC_IN_SYSTEM(p) {
1471			if (p->p_pid == pid)
1472				break;
1473		}
1474		/* sx_sunlock(&allproc_lock); */
1475		if (p == NULL) {
1476			db_printf("pid %d not found\n", pid);
1477			return;
1478		}
1479		FOREACH_THREAD_IN_PROC(p, td) {
1480			witness_list(td);
1481		}
1482	} else {
1483		td = curthread;
1484		witness_list(td);
1485	}
1486}
1487
1488DB_SHOW_COMMAND(witness, db_witness_display)
1489{
1490
1491	witness_display(db_printf);
1492}
1493#endif
1494