subr_witness.c revision 112118
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 112118 2003-03-11 22:14:21Z jhb $
31 */
32
33/*
34 * Implementation of the `witness' lock verifier.  Originally implemented for
35 * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36 * classes in FreeBSD.
37 */
38
39/*
40 *	Main Entry: witness
41 *	Pronunciation: 'wit-n&s
42 *	Function: noun
43 *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44 *	    testimony, witness, from 2wit
45 *	Date: before 12th century
46 *	1 : attestation of a fact or event : TESTIMONY
47 *	2 : one that gives evidence; specifically : one who testifies in
48 *	    a cause or before a judicial tribunal
49 *	3 : one asked to be present at a transaction so as to be able to
50 *	    testify to its having taken place
51 *	4 : one who has personal knowledge of something
52 *	5 a : something serving as evidence or proof : SIGN
53 *	  b : public affirmation by word or example of usually
54 *	      religious faith or conviction <the heroic witness to divine
55 *	      life -- Pilot>
56 *	6 capitalized : a member of the Jehovah's Witnesses
57 */
58
59/*
60 * Special rules concerning Giant and lock orders:
61 *
62 * 1) Giant must be acquired before any other mutexes.  Stated another way,
63 *    no other mutex may be held when Giant is acquired.
64 *
65 * 2) Giant must be released when blocking on a sleepable lock.
66 *
67 * This rule is less obvious, but is a result of Giant providing the same
68 * semantics as spl().  Basically, when a thread sleeps, it must release
69 * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
70 * 2).
71 *
72 * 3) Giant may be acquired before or after sleepable locks.
73 *
74 * This rule is also not quite as obvious.  Giant may be acquired after
75 * a sleepable lock because it is a non-sleepable lock and non-sleepable
76 * locks may always be acquired while holding a sleepable lock.  The second
77 * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
78 * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
79 * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
80 * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
81 * execute.  Thus, acquiring Giant both before and after a sleepable lock
82 * will not result in a lock order reversal.
83 */
84
85#include "opt_ddb.h"
86#include "opt_witness.h"
87
88#include <sys/param.h>
89#include <sys/bus.h>
90#include <sys/kernel.h>
91#include <sys/ktr.h>
92#include <sys/lock.h>
93#include <sys/malloc.h>
94#include <sys/mutex.h>
95#include <sys/proc.h>
96#include <sys/sysctl.h>
97#include <sys/systm.h>
98
99#include <ddb/ddb.h>
100
101#include <machine/stdarg.h>
102
103/* Define this to check for blessed mutexes */
104#undef BLESSING
105
106#define WITNESS_COUNT 200
107#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
108/*
109 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
110 * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
111 * probably be safe for the most part, but it's still a SWAG.
112 */
113#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
114
115#define	WITNESS_NCHILDREN 6
116
117struct witness_child_list_entry;
118
119struct witness {
120	const	char *w_name;
121	struct	lock_class *w_class;
122	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
123	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
124	struct	witness_child_list_entry *w_children;	/* Great evilness... */
125	const	char *w_file;
126	int	w_line;
127	u_int	w_level;
128	u_int	w_refcount;
129	u_char	w_Giant_squawked:1;
130	u_char	w_other_squawked:1;
131	u_char	w_same_squawked:1;
132	u_char	w_displayed:1;
133};
134
135struct witness_child_list_entry {
136	struct	witness_child_list_entry *wcl_next;
137	struct	witness *wcl_children[WITNESS_NCHILDREN];
138	u_int	wcl_count;
139};
140
141STAILQ_HEAD(witness_list, witness);
142
143#ifdef BLESSING
144struct witness_blessed {
145	const	char *b_lock1;
146	const	char *b_lock2;
147};
148#endif
149
150struct witness_order_list_entry {
151	const	char *w_name;
152	struct	lock_class *w_class;
153};
154
155#ifdef BLESSING
156static int	blessed(struct witness *, struct witness *);
157#endif
158static int	depart(struct witness *w);
159static struct	witness *enroll(const char *description,
160				struct lock_class *lock_class);
161static int	insertchild(struct witness *parent, struct witness *child);
162static int	isitmychild(struct witness *parent, struct witness *child);
163static int	isitmydescendant(struct witness *parent, struct witness *child);
164static int	itismychild(struct witness *parent, struct witness *child);
165static int	rebalancetree(struct witness_list *list);
166static void	removechild(struct witness *parent, struct witness *child);
167static int	reparentchildren(struct witness *newparent,
168		    struct witness *oldparent);
169static void	witness_displaydescendants(void(*)(const char *fmt, ...),
170					   struct witness *, int indent);
171static const char *fixup_filename(const char *file);
172static void	witness_leveldescendents(struct witness *parent, int level);
173static void	witness_levelall(void);
174static struct	witness *witness_get(void);
175static void	witness_free(struct witness *m);
176static struct	witness_child_list_entry *witness_child_get(void);
177static void	witness_child_free(struct witness_child_list_entry *wcl);
178static struct	lock_list_entry *witness_lock_list_get(void);
179static void	witness_lock_list_free(struct lock_list_entry *lle);
180static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
181					     struct lock_object *lock);
182static void	witness_list_lock(struct lock_instance *instance);
183#ifdef DDB
184static void	witness_list(struct thread *td);
185static void	witness_display_list(void(*prnt)(const char *fmt, ...),
186				     struct witness_list *list);
187static void	witness_display(void(*)(const char *fmt, ...));
188#endif
189
190MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
191
192static int witness_watch = 1;
193TUNABLE_INT("debug.witness_watch", &witness_watch);
194SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
195
196#ifdef DDB
197/*
198 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
199 * drop into kdebug() when:
200 *	- a lock heirarchy violation occurs
201 *	- locks are held when going to sleep.
202 */
203#ifdef WITNESS_DDB
204int	witness_ddb = 1;
205#else
206int	witness_ddb = 0;
207#endif
208TUNABLE_INT("debug.witness_ddb", &witness_ddb);
209SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
210
211/*
212 * When DDB is enabled and witness_trace is set to 1, it will cause the system
213 * to print a stack trace:
214 *	- a lock heirarchy violation occurs
215 *	- locks are held when going to sleep.
216 */
217int	witness_trace = 1;
218TUNABLE_INT("debug.witness_trace", &witness_trace);
219SYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, "");
220#endif /* DDB */
221
222#ifdef WITNESS_SKIPSPIN
223int	witness_skipspin = 1;
224#else
225int	witness_skipspin = 0;
226#endif
227TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
228SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
229    "");
230
231static struct mtx w_mtx;
232static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
233static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
234static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
235static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
236static struct witness_child_list_entry *w_child_free = NULL;
237static struct lock_list_entry *w_lock_list_free = NULL;
238static int witness_dead;	/* fatal error, probably no memory */
239
240static struct witness w_data[WITNESS_COUNT];
241static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
242static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
243
244static struct witness_order_list_entry order_lists[] = {
245	{ "proctree", &lock_class_sx },
246	{ "allproc", &lock_class_sx },
247	{ "Giant", &lock_class_mtx_sleep },
248	{ "filedesc structure", &lock_class_mtx_sleep },
249	{ "pipe mutex", &lock_class_mtx_sleep },
250	{ "sigio lock", &lock_class_mtx_sleep },
251	{ "process group", &lock_class_mtx_sleep },
252	{ "process lock", &lock_class_mtx_sleep },
253	{ "session", &lock_class_mtx_sleep },
254	{ "uidinfo hash", &lock_class_mtx_sleep },
255	{ "uidinfo struct", &lock_class_mtx_sleep },
256	{ NULL, NULL },
257	/*
258	 * spin locks
259	 */
260#ifdef SMP
261	{ "ap boot", &lock_class_mtx_spin },
262#ifdef __i386__
263	{ "com", &lock_class_mtx_spin },
264#endif
265#endif
266	{ "sio", &lock_class_mtx_spin },
267#ifdef __i386__
268	{ "cy", &lock_class_mtx_spin },
269#endif
270	{ "sabtty", &lock_class_mtx_spin },
271	{ "zstty", &lock_class_mtx_spin },
272	{ "ng_node", &lock_class_mtx_spin },
273	{ "ng_worklist", &lock_class_mtx_spin },
274	{ "ithread table lock", &lock_class_mtx_spin },
275	{ "sched lock", &lock_class_mtx_spin },
276	{ "callout", &lock_class_mtx_spin },
277	/*
278	 * leaf locks
279	 */
280	{ "allpmaps", &lock_class_mtx_spin },
281	{ "vm page queue free mutex", &lock_class_mtx_spin },
282	{ "icu", &lock_class_mtx_spin },
283#ifdef SMP
284	{ "smp rendezvous", &lock_class_mtx_spin },
285#if defined(__i386__) && defined(APIC_IO)
286	{ "tlb", &lock_class_mtx_spin },
287#endif
288#ifdef __sparc64__
289	{ "ipi", &lock_class_mtx_spin },
290#endif
291#endif
292	{ "clk", &lock_class_mtx_spin },
293	{ "mutex profiling lock", &lock_class_mtx_spin },
294	{ "kse zombie lock", &lock_class_mtx_spin },
295	{ "ALD Queue", &lock_class_mtx_spin },
296#ifdef __ia64__
297	{ "MCA spin lock", &lock_class_mtx_spin },
298#endif
299#ifdef __i386__
300	{ "pcicfg", &lock_class_mtx_spin },
301#endif
302	{ NULL, NULL },
303	{ NULL, NULL }
304};
305
306#ifdef BLESSING
307/*
308 * Pairs of locks which have been blessed
309 * Don't complain about order problems with blessed locks
310 */
311static struct witness_blessed blessed_list[] = {
312};
313static int blessed_count =
314	sizeof(blessed_list) / sizeof(struct witness_blessed);
315#endif
316
317/*
318 * List of all locks in the system.
319 */
320TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
321
322static struct mtx all_mtx = {
323	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
324	  "All locks list",		/* mtx_object.lo_name */
325	  "All locks list",		/* mtx_object.lo_type */
326	  LO_INITIALIZED,		/* mtx_object.lo_flags */
327	  { NULL, NULL },		/* mtx_object.lo_list */
328	  NULL },			/* mtx_object.lo_witness */
329	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
330	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
331	{ NULL, NULL }			/* mtx_contested */
332};
333
334/*
335 * This global is set to 0 once it becomes safe to use the witness code.
336 */
337static int witness_cold = 1;
338
339/*
340 * Global variables for book keeping.
341 */
342static int lock_cur_cnt;
343static int lock_max_cnt;
344
345/*
346 * The WITNESS-enabled diagnostic code.
347 */
348static void
349witness_initialize(void *dummy __unused)
350{
351	struct lock_object *lock;
352	struct witness_order_list_entry *order;
353	struct witness *w, *w1;
354	int i;
355
356	/*
357	 * We have to release Giant before initializing its witness
358	 * structure so that WITNESS doesn't get confused.
359	 */
360	mtx_unlock(&Giant);
361	mtx_assert(&Giant, MA_NOTOWNED);
362
363	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
364	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
365	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
366	    MTX_NOWITNESS);
367	for (i = 0; i < WITNESS_COUNT; i++)
368		witness_free(&w_data[i]);
369	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
370		witness_child_free(&w_childdata[i]);
371	for (i = 0; i < LOCK_CHILDCOUNT; i++)
372		witness_lock_list_free(&w_locklistdata[i]);
373
374	/* First add in all the specified order lists. */
375	for (order = order_lists; order->w_name != NULL; order++) {
376		w = enroll(order->w_name, order->w_class);
377		if (w == NULL)
378			continue;
379		w->w_file = "order list";
380		for (order++; order->w_name != NULL; order++) {
381			w1 = enroll(order->w_name, order->w_class);
382			if (w1 == NULL)
383				continue;
384			w1->w_file = "order list";
385			if (!itismychild(w, w1))
386				panic("Not enough memory for static orders!");
387			w = w1;
388		}
389	}
390
391	/* Iterate through all locks and add them to witness. */
392	mtx_lock(&all_mtx);
393	TAILQ_FOREACH(lock, &all_locks, lo_list) {
394		if (lock->lo_flags & LO_WITNESS)
395			lock->lo_witness = enroll(lock->lo_type,
396			    lock->lo_class);
397		else
398			lock->lo_witness = NULL;
399	}
400	mtx_unlock(&all_mtx);
401
402	/* Mark the witness code as being ready for use. */
403	atomic_store_rel_int(&witness_cold, 0);
404
405	mtx_lock(&Giant);
406}
407SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
408
409void
410witness_init(struct lock_object *lock)
411{
412	struct lock_class *class;
413
414	class = lock->lo_class;
415	if (lock->lo_flags & LO_INITIALIZED)
416		panic("%s: lock (%s) %s is already initialized", __func__,
417		    class->lc_name, lock->lo_name);
418	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
419	    (class->lc_flags & LC_RECURSABLE) == 0)
420		panic("%s: lock (%s) %s can not be recursable", __func__,
421		    class->lc_name, lock->lo_name);
422	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
423	    (class->lc_flags & LC_SLEEPABLE) == 0)
424		panic("%s: lock (%s) %s can not be sleepable", __func__,
425		    class->lc_name, lock->lo_name);
426	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
427	    (class->lc_flags & LC_UPGRADABLE) == 0)
428		panic("%s: lock (%s) %s can not be upgradable", __func__,
429		    class->lc_name, lock->lo_name);
430
431	mtx_lock(&all_mtx);
432	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
433	lock->lo_flags |= LO_INITIALIZED;
434	lock_cur_cnt++;
435	if (lock_cur_cnt > lock_max_cnt)
436		lock_max_cnt = lock_cur_cnt;
437	mtx_unlock(&all_mtx);
438	if (!witness_cold && !witness_dead && panicstr == NULL &&
439	    (lock->lo_flags & LO_WITNESS) != 0)
440		lock->lo_witness = enroll(lock->lo_type, class);
441	else
442		lock->lo_witness = NULL;
443}
444
445void
446witness_destroy(struct lock_object *lock)
447{
448	struct witness *w;
449
450	if (witness_cold)
451		panic("lock (%s) %s destroyed while witness_cold",
452		    lock->lo_class->lc_name, lock->lo_name);
453	if ((lock->lo_flags & LO_INITIALIZED) == 0)
454		panic("%s: lock (%s) %s is not initialized", __func__,
455		    lock->lo_class->lc_name, lock->lo_name);
456
457	/* XXX: need to verify that no one holds the lock */
458	w = lock->lo_witness;
459	if (w != NULL) {
460		mtx_lock_spin(&w_mtx);
461		MPASS(w->w_refcount > 0);
462		w->w_refcount--;
463
464		/*
465		 * Lock is already released if we have an allocation failure
466		 * and depart() fails.
467		 */
468		if (w->w_refcount != 0 || depart(w))
469			mtx_unlock_spin(&w_mtx);
470	}
471
472	mtx_lock(&all_mtx);
473	lock_cur_cnt--;
474	TAILQ_REMOVE(&all_locks, lock, lo_list);
475	lock->lo_flags &= ~LO_INITIALIZED;
476	mtx_unlock(&all_mtx);
477}
478
479#ifdef DDB
480static void
481witness_display_list(void(*prnt)(const char *fmt, ...),
482		     struct witness_list *list)
483{
484	struct witness *w;
485
486	STAILQ_FOREACH(w, list, w_typelist) {
487		if (w->w_file == NULL || w->w_level > 0)
488			continue;
489		/*
490		 * This lock has no anscestors, display its descendants.
491		 */
492		witness_displaydescendants(prnt, w, 0);
493	}
494}
495
496static void
497witness_display(void(*prnt)(const char *fmt, ...))
498{
499	struct witness *w;
500
501	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
502	witness_levelall();
503
504	/* Clear all the displayed flags. */
505	STAILQ_FOREACH(w, &w_all, w_list) {
506		w->w_displayed = 0;
507	}
508
509	/*
510	 * First, handle sleep locks which have been acquired at least
511	 * once.
512	 */
513	prnt("Sleep locks:\n");
514	witness_display_list(prnt, &w_sleep);
515
516	/*
517	 * Now do spin locks which have been acquired at least once.
518	 */
519	prnt("\nSpin locks:\n");
520	witness_display_list(prnt, &w_spin);
521
522	/*
523	 * Finally, any locks which have not been acquired yet.
524	 */
525	prnt("\nLocks which were never acquired:\n");
526	STAILQ_FOREACH(w, &w_all, w_list) {
527		if (w->w_file != NULL || w->w_refcount == 0)
528			continue;
529		prnt("%s\n", w->w_name);
530	}
531}
532#endif /* DDB */
533
534/* Trim useless garbage from filenames. */
535static const char *
536fixup_filename(const char *file)
537{
538
539	if (file == NULL)
540		return (NULL);
541	while (strncmp(file, "../", 3) == 0)
542		file += 3;
543	return (file);
544}
545
546void
547witness_lock(struct lock_object *lock, int flags, const char *file, int line)
548{
549	struct lock_list_entry **lock_list, *lle;
550	struct lock_instance *lock1, *lock2;
551	struct lock_class *class;
552	struct witness *w, *w1;
553	struct thread *td;
554	int i, j;
555#ifdef DDB
556	int go_into_ddb = 0;
557#endif
558
559	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
560	    panicstr != NULL)
561		return;
562	w = lock->lo_witness;
563	class = lock->lo_class;
564	td = curthread;
565	file = fixup_filename(file);
566
567	if (class->lc_flags & LC_SLEEPLOCK) {
568		/*
569		 * Since spin locks include a critical section, this check
570		 * impliclty enforces a lock order of all sleep locks before
571		 * all spin locks.
572		 */
573		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
574			panic("blockable sleep lock (%s) %s @ %s:%d",
575			    class->lc_name, lock->lo_name, file, line);
576		lock_list = &td->td_sleeplocks;
577	} else
578		lock_list = PCPU_PTR(spinlocks);
579
580	/*
581	 * Is this the first lock acquired?  If so, then no order checking
582	 * is needed.
583	 */
584	if (*lock_list == NULL)
585		goto out;
586
587	/*
588	 * Check to see if we are recursing on a lock we already own.
589	 */
590	lock1 = find_instance(*lock_list, lock);
591	if (lock1 != NULL) {
592		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
593		    (flags & LOP_EXCLUSIVE) == 0) {
594			printf("shared lock of (%s) %s @ %s:%d\n",
595			    class->lc_name, lock->lo_name, file, line);
596			printf("while exclusively locked from %s:%d\n",
597			    lock1->li_file, lock1->li_line);
598			panic("share->excl");
599		}
600		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
601		    (flags & LOP_EXCLUSIVE) != 0) {
602			printf("exclusive lock of (%s) %s @ %s:%d\n",
603			    class->lc_name, lock->lo_name, file, line);
604			printf("while share locked from %s:%d\n",
605			    lock1->li_file, lock1->li_line);
606			panic("excl->share");
607		}
608		lock1->li_flags++;
609		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
610			printf(
611			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
612			    class->lc_name, lock->lo_name, file, line);
613			printf("first acquired @ %s:%d\n", lock1->li_file,
614			    lock1->li_line);
615			panic("recurse");
616		}
617		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
618		    td->td_proc->p_pid, lock->lo_name,
619		    lock1->li_flags & LI_RECURSEMASK);
620		lock1->li_file = file;
621		lock1->li_line = line;
622		return;
623	}
624
625	/*
626	 * Try locks do not block if they fail to acquire the lock, thus
627	 * there is no danger of deadlocks or of switching while holding a
628	 * spin lock if we acquire a lock via a try operation.
629	 */
630	if (flags & LOP_TRYLOCK)
631		goto out;
632
633	/*
634	 * Check for duplicate locks of the same type.  Note that we only
635	 * have to check for this on the last lock we just acquired.  Any
636	 * other cases will be caught as lock order violations.
637	 */
638	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
639	w1 = lock1->li_lock->lo_witness;
640	if (w1 == w) {
641		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
642			goto out;
643		w->w_same_squawked = 1;
644		printf("acquiring duplicate lock of same type: \"%s\"\n",
645			lock->lo_type);
646		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
647		    lock1->li_file, lock1->li_line);
648		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
649#ifdef DDB
650		go_into_ddb = 1;
651#endif
652		goto out;
653	}
654	MPASS(!mtx_owned(&w_mtx));
655	mtx_lock_spin(&w_mtx);
656	/*
657	 * If we have a known higher number just say ok
658	 */
659	if (witness_watch > 1 && w->w_level > w1->w_level) {
660		mtx_unlock_spin(&w_mtx);
661		goto out;
662	}
663	/*
664	 * If we know that the the lock we are acquiring comes after
665	 * the lock we most recently acquired in the lock order tree,
666	 * then there is no need for any further checks.
667	 */
668	if (isitmydescendant(w1, w)) {
669		mtx_unlock_spin(&w_mtx);
670		goto out;
671	}
672	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
673		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
674
675			MPASS(j < WITNESS_COUNT);
676			lock1 = &lle->ll_children[i];
677			w1 = lock1->li_lock->lo_witness;
678
679			/*
680			 * If this lock doesn't undergo witness checking,
681			 * then skip it.
682			 */
683			if (w1 == NULL) {
684				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
685				    ("lock missing witness structure"));
686				continue;
687			}
688			/*
689			 * If we are locking Giant and this is a sleepable
690			 * lock, then skip it.
691			 */
692			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
693			    lock == &Giant.mtx_object)
694				continue;
695			/*
696			 * If we are locking a sleepable lock and this lock
697			 * is Giant, then skip it.
698			 */
699			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
700			    lock1->li_lock == &Giant.mtx_object)
701				continue;
702			/*
703			 * If we are locking a sleepable lock and this lock
704			 * isn't sleepable, we want to treat it as a lock
705			 * order violation to enfore a general lock order of
706			 * sleepable locks before non-sleepable locks.
707			 */
708			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
709			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
710			    /*
711			     * Check the lock order hierarchy for a reveresal.
712			     */
713			    if (!isitmydescendant(w, w1))
714				continue;
715			/*
716			 * We have a lock order violation, check to see if it
717			 * is allowed or has already been yelled about.
718			 */
719			mtx_unlock_spin(&w_mtx);
720#ifdef BLESSING
721			if (blessed(w, w1))
722				goto out;
723#endif
724			if (lock1->li_lock == &Giant.mtx_object) {
725				if (w1->w_Giant_squawked)
726					goto out;
727				else
728					w1->w_Giant_squawked = 1;
729			} else {
730				if (w1->w_other_squawked)
731					goto out;
732				else
733					w1->w_other_squawked = 1;
734			}
735			/*
736			 * Ok, yell about it.
737			 */
738			printf("lock order reversal\n");
739			/*
740			 * Try to locate an earlier lock with
741			 * witness w in our list.
742			 */
743			do {
744				lock2 = &lle->ll_children[i];
745				MPASS(lock2->li_lock != NULL);
746				if (lock2->li_lock->lo_witness == w)
747					break;
748				i--;
749				if (i == 0 && lle->ll_next != NULL) {
750					lle = lle->ll_next;
751					i = lle->ll_count - 1;
752					MPASS(i >= 0 && i < LOCK_NCHILDREN);
753				}
754			} while (i >= 0);
755			if (i < 0) {
756				printf(" 1st %p %s (%s) @ %s:%d\n",
757				    lock1->li_lock, lock1->li_lock->lo_name,
758				    lock1->li_lock->lo_type, lock1->li_file,
759				    lock1->li_line);
760				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
761				    lock->lo_name, lock->lo_type, file, line);
762			} else {
763				printf(" 1st %p %s (%s) @ %s:%d\n",
764				    lock2->li_lock, lock2->li_lock->lo_name,
765				    lock2->li_lock->lo_type, lock2->li_file,
766				    lock2->li_line);
767				printf(" 2nd %p %s (%s) @ %s:%d\n",
768				    lock1->li_lock, lock1->li_lock->lo_name,
769				    lock1->li_lock->lo_type, lock1->li_file,
770				    lock1->li_line);
771				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
772				    lock->lo_name, lock->lo_type, file, line);
773			}
774#ifdef DDB
775			go_into_ddb = 1;
776#endif
777			goto out;
778		}
779	}
780	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
781	/*
782	 * Don't build a new relationship between a sleepable lock and
783	 * Giant if it is the wrong direction.  The real lock order is that
784	 * sleepable locks come before Giant.
785	 */
786	if (!(lock1->li_lock == &Giant.mtx_object &&
787	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
788		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
789		    lock->lo_type, lock1->li_lock->lo_type);
790		if (!itismychild(lock1->li_lock->lo_witness, w))
791			/* Witness is dead. */
792			return;
793	}
794	mtx_unlock_spin(&w_mtx);
795
796out:
797#ifdef DDB
798	if (go_into_ddb) {
799		if (witness_trace)
800			backtrace();
801		if (witness_ddb)
802			Debugger(__func__);
803	}
804#endif
805	w->w_file = file;
806	w->w_line = line;
807
808	lle = *lock_list;
809	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
810		lle = witness_lock_list_get();
811		if (lle == NULL)
812			return;
813		lle->ll_next = *lock_list;
814		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
815		    td->td_proc->p_pid, lle);
816		*lock_list = lle;
817	}
818	lock1 = &lle->ll_children[lle->ll_count++];
819	lock1->li_lock = lock;
820	lock1->li_line = line;
821	lock1->li_file = file;
822	if ((flags & LOP_EXCLUSIVE) != 0)
823		lock1->li_flags = LI_EXCLUSIVE;
824	else
825		lock1->li_flags = 0;
826	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
827	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
828}
829
830void
831witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
832{
833	struct lock_instance *instance;
834	struct lock_class *class;
835
836	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
837	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
838		return;
839	class = lock->lo_class;
840	file = fixup_filename(file);
841	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
842		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
843		    class->lc_name, lock->lo_name, file, line);
844	if ((flags & LOP_TRYLOCK) == 0)
845		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
846		    lock->lo_name, file, line);
847	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
848		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
849		    class->lc_name, lock->lo_name, file, line);
850	instance = find_instance(curthread->td_sleeplocks, lock);
851	if (instance == NULL)
852		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
853		    class->lc_name, lock->lo_name, file, line);
854	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
855		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
856		    class->lc_name, lock->lo_name, file, line);
857	if ((instance->li_flags & LI_RECURSEMASK) != 0)
858		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
859		    class->lc_name, lock->lo_name,
860		    instance->li_flags & LI_RECURSEMASK, file, line);
861	instance->li_flags |= LI_EXCLUSIVE;
862}
863
864void
865witness_downgrade(struct lock_object *lock, int flags, const char *file,
866    int line)
867{
868	struct lock_instance *instance;
869	struct lock_class *class;
870
871	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
872	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
873		return;
874	class = lock->lo_class;
875	file = fixup_filename(file);
876	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
877		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
878		    class->lc_name, lock->lo_name, file, line);
879	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
880		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
881		    class->lc_name, lock->lo_name, file, line);
882	instance = find_instance(curthread->td_sleeplocks, lock);
883	if (instance == NULL)
884		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
885		    class->lc_name, lock->lo_name, file, line);
886	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
887		panic("downgrade of shared lock (%s) %s @ %s:%d",
888		    class->lc_name, lock->lo_name, file, line);
889	if ((instance->li_flags & LI_RECURSEMASK) != 0)
890		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
891		    class->lc_name, lock->lo_name,
892		    instance->li_flags & LI_RECURSEMASK, file, line);
893	instance->li_flags &= ~LI_EXCLUSIVE;
894}
895
896void
897witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
898{
899	struct lock_list_entry **lock_list, *lle;
900	struct lock_instance *instance;
901	struct lock_class *class;
902	struct thread *td;
903	register_t s;
904	int i, j;
905
906	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
907	    panicstr != NULL)
908		return;
909	td = curthread;
910	class = lock->lo_class;
911	file = fixup_filename(file);
912	if (class->lc_flags & LC_SLEEPLOCK)
913		lock_list = &td->td_sleeplocks;
914	else
915		lock_list = PCPU_PTR(spinlocks);
916	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
917		for (i = 0; i < (*lock_list)->ll_count; i++) {
918			instance = &(*lock_list)->ll_children[i];
919			if (instance->li_lock == lock) {
920				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
921				    (flags & LOP_EXCLUSIVE) == 0) {
922					printf(
923					"shared unlock of (%s) %s @ %s:%d\n",
924					    class->lc_name, lock->lo_name,
925					    file, line);
926					printf(
927					"while exclusively locked from %s:%d\n",
928					    instance->li_file,
929					    instance->li_line);
930					panic("excl->ushare");
931				}
932				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
933				    (flags & LOP_EXCLUSIVE) != 0) {
934					printf(
935					"exclusive unlock of (%s) %s @ %s:%d\n",
936					    class->lc_name, lock->lo_name,
937					    file, line);
938					printf(
939					"while share locked from %s:%d\n",
940					    instance->li_file,
941					    instance->li_line);
942					panic("share->uexcl");
943				}
944				/* If we are recursed, unrecurse. */
945				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
946					CTR4(KTR_WITNESS,
947				    "%s: pid %d unrecursed on %s r=%d", __func__,
948					    td->td_proc->p_pid,
949					    instance->li_lock->lo_name,
950					    instance->li_flags);
951					instance->li_flags--;
952					return;
953				}
954				s = intr_disable();
955				CTR4(KTR_WITNESS,
956				    "%s: pid %d removed %s from lle[%d]", __func__,
957				    td->td_proc->p_pid,
958				    instance->li_lock->lo_name,
959				    (*lock_list)->ll_count - 1);
960				for (j = i; j < (*lock_list)->ll_count - 1; j++)
961					(*lock_list)->ll_children[j] =
962					    (*lock_list)->ll_children[j + 1];
963				(*lock_list)->ll_count--;
964				intr_restore(s);
965				if ((*lock_list)->ll_count == 0) {
966					lle = *lock_list;
967					*lock_list = lle->ll_next;
968					CTR3(KTR_WITNESS,
969					    "%s: pid %d removed lle %p", __func__,
970					    td->td_proc->p_pid, lle);
971					witness_lock_list_free(lle);
972				}
973				return;
974			}
975		}
976	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
977	    file, line);
978}
979
980/*
981 * Warn if any locks other than 'lock' are held.  Flags can be passed in to
982 * exempt Giant and sleepable locks from the checks as well.  If any
983 * non-exempt locks are held, then a supplied message is printed to the
984 * console along with a list of the offending locks.  If indicated in the
985 * flags then a failure results in a panic as well.
986 */
987int
988witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
989{
990	struct lock_list_entry *lle;
991	struct lock_instance *lock1;
992	struct thread *td;
993	va_list ap;
994	int i, n;
995
996	if (witness_cold || witness_dead || panicstr != NULL)
997		return (0);
998	n = 0;
999	td = curthread;
1000	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1001		for (i = lle->ll_count - 1; i >= 0; i--) {
1002			lock1 = &lle->ll_children[i];
1003			if (lock1->li_lock == lock)
1004				continue;
1005			if (flags & WARN_GIANTOK &&
1006			    lock1->li_lock == &Giant.mtx_object)
1007				continue;
1008			if (flags & WARN_SLEEPOK &&
1009			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1010				continue;
1011			if (n == 0) {
1012				va_start(ap, fmt);
1013				vprintf(fmt, ap);
1014				va_end(ap);
1015				printf(" with the following");
1016				if (flags & WARN_SLEEPOK)
1017					printf(" non-sleepable");
1018				printf("locks held:\n");
1019			}
1020			n++;
1021			witness_list_lock(lock1);
1022		}
1023	if (PCPU_GET(spinlocks) != NULL) {
1024		/*
1025		 * Since we already hold a spinlock preemption is
1026		 * already blocked.
1027		 */
1028		if (n == 0) {
1029			va_start(ap, fmt);
1030			vprintf(fmt, ap);
1031			va_end(ap);
1032			printf(" with the following");
1033			if (flags & WARN_SLEEPOK)
1034				printf(" non-sleepable");
1035			printf("locks held:\n");
1036		}
1037		n += witness_list_locks(PCPU_PTR(spinlocks));
1038	}
1039	if (flags & WARN_PANIC && n)
1040		panic("witness_warn");
1041#ifdef DDB
1042	else if (witness_ddb && n)
1043		Debugger(__func__);
1044#endif
1045	return (n);
1046}
1047
1048const char *
1049witness_file(struct lock_object *lock)
1050{
1051	struct witness *w;
1052
1053	if (witness_cold || witness_dead || lock->lo_witness == NULL)
1054		return ("?");
1055	w = lock->lo_witness;
1056	return (w->w_file);
1057}
1058
1059int
1060witness_line(struct lock_object *lock)
1061{
1062	struct witness *w;
1063
1064	if (witness_cold || witness_dead || lock->lo_witness == NULL)
1065		return (0);
1066	w = lock->lo_witness;
1067	return (w->w_line);
1068}
1069
1070static struct witness *
1071enroll(const char *description, struct lock_class *lock_class)
1072{
1073	struct witness *w;
1074
1075	if (!witness_watch || witness_dead || panicstr != NULL)
1076		return (NULL);
1077	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1078		return (NULL);
1079	mtx_lock_spin(&w_mtx);
1080	STAILQ_FOREACH(w, &w_all, w_list) {
1081		if (w->w_name == description || (w->w_refcount > 0 &&
1082		    strcmp(description, w->w_name) == 0)) {
1083			w->w_refcount++;
1084			mtx_unlock_spin(&w_mtx);
1085			if (lock_class != w->w_class)
1086				panic(
1087				"lock (%s) %s does not match earlier (%s) lock",
1088				    description, lock_class->lc_name,
1089				    w->w_class->lc_name);
1090			return (w);
1091		}
1092	}
1093	/*
1094	 * This isn't quite right, as witness_cold is still 0 while we
1095	 * enroll all the locks initialized before witness_initialize().
1096	 */
1097	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1098		mtx_unlock_spin(&w_mtx);
1099		panic("spin lock %s not in order list", description);
1100	}
1101	if ((w = witness_get()) == NULL)
1102		return (NULL);
1103	w->w_name = description;
1104	w->w_class = lock_class;
1105	w->w_refcount = 1;
1106	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1107	if (lock_class->lc_flags & LC_SPINLOCK)
1108		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1109	else if (lock_class->lc_flags & LC_SLEEPLOCK)
1110		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1111	else {
1112		mtx_unlock_spin(&w_mtx);
1113		panic("lock class %s is not sleep or spin",
1114		    lock_class->lc_name);
1115	}
1116	mtx_unlock_spin(&w_mtx);
1117	return (w);
1118}
1119
1120/* Don't let the door bang you on the way out... */
1121static int
1122depart(struct witness *w)
1123{
1124	struct witness_child_list_entry *wcl, *nwcl;
1125	struct witness_list *list;
1126	struct witness *parent;
1127
1128	MPASS(w->w_refcount == 0);
1129	if (w->w_class->lc_flags & LC_SLEEPLOCK)
1130		list = &w_sleep;
1131	else
1132		list = &w_spin;
1133	/*
1134	 * First, we run through the entire tree looking for any
1135	 * witnesses that the outgoing witness is a child of.  For
1136	 * each parent that we find, we reparent all the direct
1137	 * children of the outgoing witness to its parent.
1138	 */
1139	STAILQ_FOREACH(parent, list, w_typelist) {
1140		if (!isitmychild(parent, w))
1141			continue;
1142		removechild(parent, w);
1143		if (!reparentchildren(parent, w))
1144			return (0);
1145	}
1146
1147	/*
1148	 * Now we go through and free up the child list of the
1149	 * outgoing witness.
1150	 */
1151	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1152		nwcl = wcl->wcl_next;
1153		witness_child_free(wcl);
1154	}
1155
1156	/*
1157	 * Detach from various lists and free.
1158	 */
1159	STAILQ_REMOVE(list, w, witness, w_typelist);
1160	STAILQ_REMOVE(&w_all, w, witness, w_list);
1161	witness_free(w);
1162
1163	/* Finally, fixup the tree. */
1164	return (rebalancetree(list));
1165}
1166
1167
1168/*
1169 * Prune an entire lock order tree.  We look for cases where a lock
1170 * is now both a descendant and a direct child of a given lock.  In
1171 * that case, we want to remove the direct child link from the tree.
1172 *
1173 * Returns false if insertchild() fails.
1174 */
1175static int
1176rebalancetree(struct witness_list *list)
1177{
1178	struct witness *child, *parent;
1179
1180	STAILQ_FOREACH(child, list, w_typelist) {
1181		STAILQ_FOREACH(parent, list, w_typelist) {
1182			if (!isitmychild(parent, child))
1183				continue;
1184			removechild(parent, child);
1185			if (isitmydescendant(parent, child))
1186				continue;
1187			if (!insertchild(parent, child))
1188				return (0);
1189		}
1190	}
1191	witness_levelall();
1192	return (1);
1193}
1194
1195/*
1196 * Add "child" as a direct child of "parent".  Returns false if
1197 * we fail due to out of memory.
1198 */
1199static int
1200insertchild(struct witness *parent, struct witness *child)
1201{
1202	struct witness_child_list_entry **wcl;
1203
1204	MPASS(child != NULL && parent != NULL);
1205
1206	/*
1207	 * Insert "child" after "parent"
1208	 */
1209	wcl = &parent->w_children;
1210	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1211		wcl = &(*wcl)->wcl_next;
1212	if (*wcl == NULL) {
1213		*wcl = witness_child_get();
1214		if (*wcl == NULL)
1215			return (0);
1216	}
1217	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1218
1219	return (1);
1220}
1221
1222/*
1223 * Make all the direct descendants of oldparent be direct descendants
1224 * of newparent.
1225 */
1226static int
1227reparentchildren(struct witness *newparent, struct witness *oldparent)
1228{
1229	struct witness_child_list_entry *wcl;
1230	int i;
1231
1232	/* Avoid making a witness a child of itself. */
1233	MPASS(!isitmychild(oldparent, newparent));
1234
1235	for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1236		for (i = 0; i < wcl->wcl_count; i++)
1237			if (!insertchild(newparent, wcl->wcl_children[i]))
1238				return (0);
1239	return (1);
1240}
1241
1242static int
1243itismychild(struct witness *parent, struct witness *child)
1244{
1245	struct witness_list *list;
1246
1247	MPASS(child != NULL && parent != NULL);
1248	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1249	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1250		panic(
1251		"%s: parent (%s) and child (%s) are not the same lock type",
1252		    __func__, parent->w_class->lc_name,
1253		    child->w_class->lc_name);
1254
1255	if (!insertchild(parent, child))
1256		return (0);
1257
1258	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1259		list = &w_sleep;
1260	else
1261		list = &w_spin;
1262	return (rebalancetree(list));
1263}
1264
1265static void
1266removechild(struct witness *parent, struct witness *child)
1267{
1268	struct witness_child_list_entry **wcl, *wcl1;
1269	int i;
1270
1271	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1272		for (i = 0; i < (*wcl)->wcl_count; i++)
1273			if ((*wcl)->wcl_children[i] == child)
1274				goto found;
1275	return;
1276found:
1277	(*wcl)->wcl_count--;
1278	if ((*wcl)->wcl_count > i)
1279		(*wcl)->wcl_children[i] =
1280		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1281	MPASS((*wcl)->wcl_children[i] != NULL);
1282	if ((*wcl)->wcl_count != 0)
1283		return;
1284	wcl1 = *wcl;
1285	*wcl = wcl1->wcl_next;
1286	witness_child_free(wcl1);
1287}
1288
1289static int
1290isitmychild(struct witness *parent, struct witness *child)
1291{
1292	struct witness_child_list_entry *wcl;
1293	int i;
1294
1295	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1296		for (i = 0; i < wcl->wcl_count; i++) {
1297			if (wcl->wcl_children[i] == child)
1298				return (1);
1299		}
1300	}
1301	return (0);
1302}
1303
1304static int
1305isitmydescendant(struct witness *parent, struct witness *child)
1306{
1307	struct witness_child_list_entry *wcl;
1308	int i, j;
1309
1310	if (isitmychild(parent, child))
1311		return (1);
1312	j = 0;
1313	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1314		MPASS(j < 1000);
1315		for (i = 0; i < wcl->wcl_count; i++) {
1316			if (isitmydescendant(wcl->wcl_children[i], child))
1317				return (1);
1318		}
1319		j++;
1320	}
1321	return (0);
1322}
1323
1324static void
1325witness_levelall (void)
1326{
1327	struct witness_list *list;
1328	struct witness *w, *w1;
1329
1330	/*
1331	 * First clear all levels.
1332	 */
1333	STAILQ_FOREACH(w, &w_all, w_list) {
1334		w->w_level = 0;
1335	}
1336
1337	/*
1338	 * Look for locks with no parent and level all their descendants.
1339	 */
1340	STAILQ_FOREACH(w, &w_all, w_list) {
1341		/*
1342		 * This is just an optimization, technically we could get
1343		 * away just walking the all list each time.
1344		 */
1345		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1346			list = &w_sleep;
1347		else
1348			list = &w_spin;
1349		STAILQ_FOREACH(w1, list, w_typelist) {
1350			if (isitmychild(w1, w))
1351				goto skip;
1352		}
1353		witness_leveldescendents(w, 0);
1354	skip:
1355		;	/* silence GCC 3.x */
1356	}
1357}
1358
1359static void
1360witness_leveldescendents(struct witness *parent, int level)
1361{
1362	struct witness_child_list_entry *wcl;
1363	int i;
1364
1365	if (parent->w_level < level)
1366		parent->w_level = level;
1367	level++;
1368	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1369		for (i = 0; i < wcl->wcl_count; i++)
1370			witness_leveldescendents(wcl->wcl_children[i], level);
1371}
1372
1373static void
1374witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1375			   struct witness *parent, int indent)
1376{
1377	struct witness_child_list_entry *wcl;
1378	int i, level;
1379
1380	level = parent->w_level;
1381	prnt("%-2d", level);
1382	for (i = 0; i < indent; i++)
1383		prnt(" ");
1384	if (parent->w_refcount > 0)
1385		prnt("%s", parent->w_name);
1386	else
1387		prnt("(dead)");
1388	if (parent->w_displayed) {
1389		prnt(" -- (already displayed)\n");
1390		return;
1391	}
1392	parent->w_displayed = 1;
1393	if (parent->w_refcount > 0) {
1394		if (parent->w_file != NULL)
1395			prnt(" -- last acquired @ %s:%d", parent->w_file,
1396			    parent->w_line);
1397	}
1398	prnt("\n");
1399	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1400		for (i = 0; i < wcl->wcl_count; i++)
1401			    witness_displaydescendants(prnt,
1402				wcl->wcl_children[i], indent + 1);
1403}
1404
1405#ifdef BLESSING
1406static int
1407blessed(struct witness *w1, struct witness *w2)
1408{
1409	int i;
1410	struct witness_blessed *b;
1411
1412	for (i = 0; i < blessed_count; i++) {
1413		b = &blessed_list[i];
1414		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1415			if (strcmp(w2->w_name, b->b_lock2) == 0)
1416				return (1);
1417			continue;
1418		}
1419		if (strcmp(w1->w_name, b->b_lock2) == 0)
1420			if (strcmp(w2->w_name, b->b_lock1) == 0)
1421				return (1);
1422	}
1423	return (0);
1424}
1425#endif
1426
1427static struct witness *
1428witness_get(void)
1429{
1430	struct witness *w;
1431
1432	if (witness_dead) {
1433		mtx_unlock_spin(&w_mtx);
1434		return (NULL);
1435	}
1436	if (STAILQ_EMPTY(&w_free)) {
1437		witness_dead = 1;
1438		mtx_unlock_spin(&w_mtx);
1439		printf("%s: witness exhausted\n", __func__);
1440		return (NULL);
1441	}
1442	w = STAILQ_FIRST(&w_free);
1443	STAILQ_REMOVE_HEAD(&w_free, w_list);
1444	bzero(w, sizeof(*w));
1445	return (w);
1446}
1447
1448static void
1449witness_free(struct witness *w)
1450{
1451
1452	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1453}
1454
1455static struct witness_child_list_entry *
1456witness_child_get(void)
1457{
1458	struct witness_child_list_entry *wcl;
1459
1460	if (witness_dead) {
1461		mtx_unlock_spin(&w_mtx);
1462		return (NULL);
1463	}
1464	wcl = w_child_free;
1465	if (wcl == NULL) {
1466		witness_dead = 1;
1467		mtx_unlock_spin(&w_mtx);
1468		printf("%s: witness exhausted\n", __func__);
1469		return (NULL);
1470	}
1471	w_child_free = wcl->wcl_next;
1472	bzero(wcl, sizeof(*wcl));
1473	return (wcl);
1474}
1475
1476static void
1477witness_child_free(struct witness_child_list_entry *wcl)
1478{
1479
1480	wcl->wcl_next = w_child_free;
1481	w_child_free = wcl;
1482}
1483
1484static struct lock_list_entry *
1485witness_lock_list_get(void)
1486{
1487	struct lock_list_entry *lle;
1488
1489	if (witness_dead)
1490		return (NULL);
1491	mtx_lock_spin(&w_mtx);
1492	lle = w_lock_list_free;
1493	if (lle == NULL) {
1494		witness_dead = 1;
1495		mtx_unlock_spin(&w_mtx);
1496		printf("%s: witness exhausted\n", __func__);
1497		return (NULL);
1498	}
1499	w_lock_list_free = lle->ll_next;
1500	mtx_unlock_spin(&w_mtx);
1501	bzero(lle, sizeof(*lle));
1502	return (lle);
1503}
1504
1505static void
1506witness_lock_list_free(struct lock_list_entry *lle)
1507{
1508
1509	mtx_lock_spin(&w_mtx);
1510	lle->ll_next = w_lock_list_free;
1511	w_lock_list_free = lle;
1512	mtx_unlock_spin(&w_mtx);
1513}
1514
1515static struct lock_instance *
1516find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1517{
1518	struct lock_list_entry *lle;
1519	struct lock_instance *instance;
1520	int i;
1521
1522	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1523		for (i = lle->ll_count - 1; i >= 0; i--) {
1524			instance = &lle->ll_children[i];
1525			if (instance->li_lock == lock)
1526				return (instance);
1527		}
1528	return (NULL);
1529}
1530
1531static void
1532witness_list_lock(struct lock_instance *instance)
1533{
1534	struct lock_object *lock;
1535
1536	lock = instance->li_lock;
1537	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1538	    "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1539	if (lock->lo_type != lock->lo_name)
1540		printf(" (%s)", lock->lo_type);
1541	printf(" r = %d (%p) locked @ %s:%d\n",
1542	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1543	    instance->li_line);
1544}
1545
1546int
1547witness_list_locks(struct lock_list_entry **lock_list)
1548{
1549	struct lock_list_entry *lle;
1550	int i, nheld;
1551
1552	nheld = 0;
1553	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1554		for (i = lle->ll_count - 1; i >= 0; i--) {
1555			witness_list_lock(&lle->ll_children[i]);
1556			nheld++;
1557		}
1558	return (nheld);
1559}
1560
1561void
1562witness_save(struct lock_object *lock, const char **filep, int *linep)
1563{
1564	struct lock_instance *instance;
1565
1566	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1567	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1568		return;
1569	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1570		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1571		    lock->lo_class->lc_name, lock->lo_name);
1572	instance = find_instance(curthread->td_sleeplocks, lock);
1573	if (instance == NULL)
1574		panic("%s: lock (%s) %s not locked", __func__,
1575		    lock->lo_class->lc_name, lock->lo_name);
1576	*filep = instance->li_file;
1577	*linep = instance->li_line;
1578}
1579
1580void
1581witness_restore(struct lock_object *lock, const char *file, int line)
1582{
1583	struct lock_instance *instance;
1584
1585	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1586	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1587		return;
1588	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1589		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1590		    lock->lo_class->lc_name, lock->lo_name);
1591	instance = find_instance(curthread->td_sleeplocks, lock);
1592	if (instance == NULL)
1593		panic("%s: lock (%s) %s not locked", __func__,
1594		    lock->lo_class->lc_name, lock->lo_name);
1595	lock->lo_witness->w_file = file;
1596	lock->lo_witness->w_line = line;
1597	instance->li_file = file;
1598	instance->li_line = line;
1599}
1600
1601void
1602witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1603{
1604#ifdef INVARIANT_SUPPORT
1605	struct lock_instance *instance;
1606
1607	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1608		return;
1609	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1610		instance = find_instance(curthread->td_sleeplocks, lock);
1611	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1612		instance = find_instance(PCPU_GET(spinlocks), lock);
1613	else {
1614		panic("Lock (%s) %s is not sleep or spin!",
1615		    lock->lo_class->lc_name, lock->lo_name);
1616		return;
1617	}
1618	file = fixup_filename(file);
1619	switch (flags) {
1620	case LA_UNLOCKED:
1621		if (instance != NULL)
1622			panic("Lock (%s) %s locked @ %s:%d.",
1623			    lock->lo_class->lc_name, lock->lo_name, file, line);
1624		break;
1625	case LA_LOCKED:
1626	case LA_LOCKED | LA_RECURSED:
1627	case LA_LOCKED | LA_NOTRECURSED:
1628	case LA_SLOCKED:
1629	case LA_SLOCKED | LA_RECURSED:
1630	case LA_SLOCKED | LA_NOTRECURSED:
1631	case LA_XLOCKED:
1632	case LA_XLOCKED | LA_RECURSED:
1633	case LA_XLOCKED | LA_NOTRECURSED:
1634		if (instance == NULL) {
1635			panic("Lock (%s) %s not locked @ %s:%d.",
1636			    lock->lo_class->lc_name, lock->lo_name, file, line);
1637			break;
1638		}
1639		if ((flags & LA_XLOCKED) != 0 &&
1640		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1641			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1642			    lock->lo_class->lc_name, lock->lo_name, file, line);
1643		if ((flags & LA_SLOCKED) != 0 &&
1644		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1645			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1646			    lock->lo_class->lc_name, lock->lo_name, file, line);
1647		if ((flags & LA_RECURSED) != 0 &&
1648		    (instance->li_flags & LI_RECURSEMASK) == 0)
1649			panic("Lock (%s) %s not recursed @ %s:%d.",
1650			    lock->lo_class->lc_name, lock->lo_name, file, line);
1651		if ((flags & LA_NOTRECURSED) != 0 &&
1652		    (instance->li_flags & LI_RECURSEMASK) != 0)
1653			panic("Lock (%s) %s recursed @ %s:%d.",
1654			    lock->lo_class->lc_name, lock->lo_name, file, line);
1655		break;
1656	default:
1657		panic("Invalid lock assertion at %s:%d.", file, line);
1658
1659	}
1660#endif	/* INVARIANT_SUPPORT */
1661}
1662
1663#ifdef DDB
1664static void
1665witness_list(struct thread *td)
1666{
1667
1668	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1669	KASSERT(db_active, ("%s: not in the debugger", __func__));
1670
1671	if (witness_dead)
1672		return;
1673
1674	witness_list_locks(&td->td_sleeplocks);
1675
1676	/*
1677	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1678	 * if td is currently executing on some other CPU and holds spin locks
1679	 * as we won't display those locks.  If we had a MI way of getting
1680	 * the per-cpu data for a given cpu then we could use
1681	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1682	 * and "fix" this.
1683	 *
1684	 * That still wouldn't really fix this unless we locked sched_lock
1685	 * or stopped the other CPU to make sure it wasn't changing the list
1686	 * out from under us.  It is probably best to just not try to handle
1687	 * threads on other CPU's for now.
1688	 */
1689	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1690		witness_list_locks(PCPU_PTR(spinlocks));
1691}
1692
1693DB_SHOW_COMMAND(locks, db_witness_list)
1694{
1695	struct thread *td;
1696	pid_t pid;
1697	struct proc *p;
1698
1699	if (have_addr) {
1700		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1701		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1702		    ((addr >> 16) % 16) * 10000;
1703		/* sx_slock(&allproc_lock); */
1704		FOREACH_PROC_IN_SYSTEM(p) {
1705			if (p->p_pid == pid)
1706				break;
1707		}
1708		/* sx_sunlock(&allproc_lock); */
1709		if (p == NULL) {
1710			db_printf("pid %d not found\n", pid);
1711			return;
1712		}
1713		FOREACH_THREAD_IN_PROC(p, td) {
1714			witness_list(td);
1715		}
1716	} else {
1717		td = curthread;
1718		witness_list(td);
1719	}
1720}
1721
1722DB_SHOW_COMMAND(witness, db_witness_display)
1723{
1724
1725	witness_display(db_printf);
1726}
1727#endif
1728