xref: /freebsd/sys/kern/subr_witness.c (revision 42c159fe388a3765f69860c84183700af37aca8a)
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  * $FreeBSD$
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 
86 struct witness_child_list_entry;
87 
88 struct 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 
103 struct 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 
109 STAILQ_HEAD(witness_list, witness);
110 
111 struct witness_blessed {
112 	const	char *b_lock1;
113 	const	char *b_lock2;
114 };
115 
116 struct witness_order_list_entry {
117 	const	char *w_name;
118 	struct	lock_class *w_class;
119 };
120 
121 static struct	witness *enroll(const char *description,
122 				struct lock_class *lock_class);
123 static int	itismychild(struct witness *parent, struct witness *child);
124 static void	removechild(struct witness *parent, struct witness *child);
125 static int	isitmychild(struct witness *parent, struct witness *child);
126 static int	isitmydescendant(struct witness *parent, struct witness *child);
127 static int	dup_ok(struct witness *);
128 static int	blessed(struct witness *, struct witness *);
129 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
130 				     struct witness_list *list);
131 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
132 					   struct witness *);
133 static void	witness_leveldescendents(struct witness *parent, int level);
134 static void	witness_levelall(void);
135 static struct	witness *witness_get(void);
136 static void	witness_free(struct witness *m);
137 static struct	witness_child_list_entry *witness_child_get(void);
138 static void	witness_child_free(struct witness_child_list_entry *wcl);
139 static struct	lock_list_entry *witness_lock_list_get(void);
140 static void	witness_lock_list_free(struct lock_list_entry *lle);
141 static void	witness_display(void(*)(const char *fmt, ...));
142 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
143 					     struct lock_object *lock);
144 
145 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146 
147 static int witness_watch = 1;
148 TUNABLE_INT("debug.witness_watch", &witness_watch);
149 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
150 
151 #ifdef DDB
152 /*
153  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
154  * drop into kdebug() when:
155  *	- a lock heirarchy violation occurs
156  *	- locks are held when going to sleep.
157  */
158 #ifdef WITNESS_DDB
159 int	witness_ddb = 1;
160 #else
161 int	witness_ddb = 0;
162 #endif
163 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
164 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
165 #endif /* DDB */
166 
167 #ifdef WITNESS_SKIPSPIN
168 int	witness_skipspin = 1;
169 #else
170 int	witness_skipspin = 0;
171 #endif
172 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
173 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
174     "");
175 
176 static struct mtx w_mtx;
177 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
178 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
179 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
180 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
181 static struct witness_child_list_entry *w_child_free = NULL;
182 static struct lock_list_entry *w_lock_list_free = NULL;
183 static int witness_dead;	/* fatal error, probably no memory */
184 
185 static struct witness w_data[WITNESS_COUNT];
186 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
187 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188 
189 static struct witness_order_list_entry order_lists[] = {
190 	{ "Giant", &lock_class_mtx_sleep },
191 	{ "pgrpsess", &lock_class_sx },
192 	{ "proctree", &lock_class_sx },
193 	{ "allproc", &lock_class_sx },
194 	{ "process group", &lock_class_mtx_sleep },
195 	{ "process lock", &lock_class_mtx_sleep },
196 	{ "session", &lock_class_mtx_sleep },
197 	{ "uidinfo hash", &lock_class_mtx_sleep },
198 	{ "uidinfo struct", &lock_class_mtx_sleep },
199 	{ NULL, NULL },
200 	/*
201 	 * spin locks
202 	 */
203 #ifdef SMP
204 	{ "ap boot", &lock_class_mtx_spin },
205 #ifdef __i386__
206 	{ "com", &lock_class_mtx_spin },
207 #endif
208 #endif
209 	{ "sio", &lock_class_mtx_spin },
210 #ifdef __i386__
211 	{ "cy", &lock_class_mtx_spin },
212 #endif
213 	{ "ng_node", &lock_class_mtx_spin },
214 	{ "ng_worklist", &lock_class_mtx_spin },
215 	{ "ithread table lock", &lock_class_mtx_spin },
216 	{ "sched lock", &lock_class_mtx_spin },
217 	{ "callout", &lock_class_mtx_spin },
218 	/*
219 	 * leaf locks
220 	 */
221 	{ "allpmaps", &lock_class_mtx_spin },
222 	{ "icu", &lock_class_mtx_spin },
223 #ifdef SMP
224 	{ "smp rendezvous", &lock_class_mtx_spin },
225 #endif
226 	{ "clk", &lock_class_mtx_spin },
227 	{ NULL, NULL },
228 	{ NULL, NULL }
229 };
230 
231 static const char *dup_list[] = {
232 	"process lock",
233 	"process group",
234 	NULL
235 };
236 
237 /*
238  * Pairs of locks which have been blessed
239  * Don't complain about order problems with blessed locks
240  */
241 static struct witness_blessed blessed_list[] = {
242 };
243 static int blessed_count =
244 	sizeof(blessed_list) / sizeof(struct witness_blessed);
245 
246 /*
247  * List of all locks in the system.
248  */
249 STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
250 
251 static struct mtx all_mtx = {
252 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
253 	  "All locks list",		/* mtx_object.lo_name */
254 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
255 	  { NULL },			/* mtx_object.lo_list */
256 	  NULL },			/* mtx_object.lo_witness */
257 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
258 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
259 	{ NULL, NULL }			/* mtx_contested */
260 };
261 
262 /*
263  * This global is set to 0 once it becomes safe to use the witness code.
264  */
265 static int witness_cold = 1;
266 
267 /*
268  * Global variables for book keeping.
269  */
270 static int lock_cur_cnt;
271 static int lock_max_cnt;
272 
273 /*
274  * The WITNESS-enabled diagnostic code.
275  */
276 static void
277 witness_initialize(void *dummy __unused)
278 {
279 	struct lock_object *lock;
280 	struct witness_order_list_entry *order;
281 	struct witness *w, *w1;
282 	int i;
283 
284 	/*
285 	 * We have to release Giant before initializing its witness
286 	 * structure so that WITNESS doesn't get confused.
287 	 */
288 	mtx_unlock(&Giant);
289 	mtx_assert(&Giant, MA_NOTOWNED);
290 
291 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
292 	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
293 	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
294 	for (i = 0; i < WITNESS_COUNT; i++)
295 		witness_free(&w_data[i]);
296 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
297 		witness_child_free(&w_childdata[i]);
298 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
299 		witness_lock_list_free(&w_locklistdata[i]);
300 
301 	/* First add in all the specified order lists. */
302 	for (order = order_lists; order->w_name != NULL; order++) {
303 		w = enroll(order->w_name, order->w_class);
304 		if (w == NULL)
305 			continue;
306 		w->w_file = "order list";
307 		for (order++; order->w_name != NULL; order++) {
308 			w1 = enroll(order->w_name, order->w_class);
309 			if (w1 == NULL)
310 				continue;
311 			w1->w_file = "order list";
312 			itismychild(w, w1);
313 			w = w1;
314 		}
315 	}
316 
317 	/* Iterate through all locks and add them to witness. */
318 	mtx_lock(&all_mtx);
319 	STAILQ_FOREACH(lock, &all_locks, lo_list) {
320 		if (lock->lo_flags & LO_WITNESS)
321 			lock->lo_witness = enroll(lock->lo_name,
322 			    lock->lo_class);
323 		else
324 			lock->lo_witness = NULL;
325 	}
326 	mtx_unlock(&all_mtx);
327 
328 	/* Mark the witness code as being ready for use. */
329 	atomic_store_rel_int(&witness_cold, 0);
330 
331 	mtx_lock(&Giant);
332 }
333 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
334 
335 void
336 witness_init(struct lock_object *lock)
337 {
338 	struct lock_class *class;
339 
340 	class = lock->lo_class;
341 	if (lock->lo_flags & LO_INITIALIZED)
342 		panic("%s: lock (%s) %s is already initialized", __func__,
343 		    class->lc_name, lock->lo_name);
344 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
345 	    (class->lc_flags & LC_RECURSABLE) == 0)
346 		panic("%s: lock (%s) %s can not be recursable", __func__,
347 		    class->lc_name, lock->lo_name);
348 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
349 	    (class->lc_flags & LC_SLEEPABLE) == 0)
350 		panic("%s: lock (%s) %s can not be sleepable", __func__,
351 		    class->lc_name, lock->lo_name);
352 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
353 	    (class->lc_flags & LC_UPGRADABLE) == 0)
354 		panic("%s: lock (%s) %s can not be upgradable", __func__,
355 		    class->lc_name, lock->lo_name);
356 
357 	mtx_lock(&all_mtx);
358 	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
359 	lock->lo_flags |= LO_INITIALIZED;
360 	lock_cur_cnt++;
361 	if (lock_cur_cnt > lock_max_cnt)
362 		lock_max_cnt = lock_cur_cnt;
363 	mtx_unlock(&all_mtx);
364 	if (!witness_cold && !witness_dead && panicstr == NULL &&
365 	    (lock->lo_flags & LO_WITNESS) != 0)
366 		lock->lo_witness = enroll(lock->lo_name, class);
367 	else
368 		lock->lo_witness = NULL;
369 }
370 
371 void
372 witness_destroy(struct lock_object *lock)
373 {
374 	struct witness *w;
375 
376 	if (witness_cold)
377 		panic("lock (%s) %s destroyed while witness_cold",
378 		    lock->lo_class->lc_name, lock->lo_name);
379 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
380 		panic("%s: lock (%s) %s is not initialized", __func__,
381 		    lock->lo_class->lc_name, lock->lo_name);
382 
383 	/* XXX: need to verify that no one holds the lock */
384 	w = lock->lo_witness;
385 	if (w != NULL) {
386 		mtx_lock_spin(&w_mtx);
387 		w->w_refcount--;
388 		if (w->w_refcount == 0) {
389 			CTR2(KTR_WITNESS,
390 			    "%s: marking witness %s as dead", __func__, w->w_name);
391 			w->w_name = "(dead)";
392 			w->w_file = "(dead)";
393 			w->w_line = 0;
394 		}
395 		mtx_unlock_spin(&w_mtx);
396 	}
397 
398 	mtx_lock(&all_mtx);
399 	lock_cur_cnt--;
400 	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
401 	lock->lo_flags &= ~LO_INITIALIZED;
402 	mtx_unlock(&all_mtx);
403 }
404 
405 static void
406 witness_display_list(void(*prnt)(const char *fmt, ...),
407 		     struct witness_list *list)
408 {
409 	struct witness *w, *w1;
410 	int found;
411 
412 	STAILQ_FOREACH(w, list, w_typelist) {
413 		if (w->w_file == NULL)
414 			continue;
415 		found = 0;
416 		STAILQ_FOREACH(w1, list, w_typelist) {
417 			if (isitmychild(w1, w)) {
418 				found++;
419 				break;
420 			}
421 		}
422 		if (found)
423 			continue;
424 		/*
425 		 * This lock has no anscestors, display its descendants.
426 		 */
427 		witness_displaydescendants(prnt, w);
428 	}
429 }
430 
431 static void
432 witness_display(void(*prnt)(const char *fmt, ...))
433 {
434 	struct witness *w;
435 
436 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
437 	witness_levelall();
438 
439 	/*
440 	 * First, handle sleep locks which have been acquired at least
441 	 * once.
442 	 */
443 	prnt("Sleep locks:\n");
444 	witness_display_list(prnt, &w_sleep);
445 
446 	/*
447 	 * Now do spin locks which have been acquired at least once.
448 	 */
449 	prnt("\nSpin locks:\n");
450 	witness_display_list(prnt, &w_spin);
451 
452 	/*
453 	 * Finally, any locks which have not been acquired yet.
454 	 */
455 	prnt("\nLocks which were never acquired:\n");
456 	STAILQ_FOREACH(w, &w_all, w_list) {
457 		if (w->w_file != NULL)
458 			continue;
459 		prnt("%s\n", w->w_name);
460 	}
461 }
462 
463 void
464 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
465 {
466 	struct lock_list_entry **lock_list, *lle;
467 	struct lock_instance *lock1, *lock2;
468 	struct lock_class *class;
469 	struct witness *w, *w1;
470 	struct thread *td;
471 	int i, j;
472 #ifdef DDB
473 	int go_into_ddb = 0;
474 #endif /* DDB */
475 
476 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
477 	    panicstr != NULL)
478 		return;
479 	w = lock->lo_witness;
480 	class = lock->lo_class;
481 	td = curthread;
482 
483 	if (class->lc_flags & LC_SLEEPLOCK) {
484 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
485 			panic("blockable sleep lock (%s) %s @ %s:%d",
486 			    class->lc_name, lock->lo_name, file, line);
487 		lock_list = &td->td_sleeplocks;
488 	} else
489 		lock_list = PCPU_PTR(spinlocks);
490 
491 	/*
492 	 * Try locks do not block if they fail to acquire the lock, thus
493 	 * there is no danger of deadlocks or of switching while holding a
494 	 * spin lock if we acquire a lock via a try operation.
495 	 */
496 	if (flags & LOP_TRYLOCK)
497 		goto out;
498 
499 	/*
500 	 * Is this the first lock acquired?  If so, then no order checking
501 	 * is needed.
502 	 */
503 	if (*lock_list == NULL)
504 		goto out;
505 
506 	/*
507 	 * Check to see if we are recursing on a lock we already own.
508 	 */
509 	lock1 = find_instance(*lock_list, lock);
510 	if (lock1 != NULL) {
511 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
512 		    (flags & LOP_EXCLUSIVE) == 0) {
513 			printf("shared lock of (%s) %s @ %s:%d\n",
514 			    class->lc_name, lock->lo_name, file, line);
515 			printf("while exclusively locked from %s:%d\n",
516 			    lock1->li_file, lock1->li_line);
517 			panic("share->excl");
518 		}
519 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
520 		    (flags & LOP_EXCLUSIVE) != 0) {
521 			printf("exclusive lock of (%s) %s @ %s:%d\n",
522 			    class->lc_name, lock->lo_name, file, line);
523 			printf("while share locked from %s:%d\n",
524 			    lock1->li_file, lock1->li_line);
525 			panic("excl->share");
526 		}
527 		lock1->li_flags++;
528 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
529 			printf(
530 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
531 			    class->lc_name, lock->lo_name, file, line);
532 			printf("first acquired @ %s:%d\n", lock1->li_file,
533 			    lock1->li_line);
534 			panic("recurse");
535 		}
536 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
537 		    td->td_proc->p_pid, lock->lo_name,
538 		    lock1->li_flags & LI_RECURSEMASK);
539 		lock1->li_file = file;
540 		lock1->li_line = line;
541 		return;
542 	}
543 
544 	/*
545 	 * Check for duplicate locks of the same type.  Note that we only
546 	 * have to check for this on the last lock we just acquired.  Any
547 	 * other cases will be caught as lock order violations.
548 	 */
549 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
550 	w1 = lock1->li_lock->lo_witness;
551 	if (w1 == w) {
552 		if (w->w_same_squawked || dup_ok(w))
553 			goto out;
554 		w->w_same_squawked = 1;
555 		printf("acquiring duplicate lock of same type: \"%s\"\n",
556 			lock->lo_name);
557 		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
558 		printf(" 2nd @ %s:%d\n", file, line);
559 #ifdef DDB
560 		go_into_ddb = 1;
561 #endif /* DDB */
562 		goto out;
563 	}
564 	MPASS(!mtx_owned(&w_mtx));
565 	mtx_lock_spin(&w_mtx);
566 	/*
567 	 * If we have a known higher number just say ok
568 	 */
569 	if (witness_watch > 1 && w->w_level > w1->w_level) {
570 		mtx_unlock_spin(&w_mtx);
571 		goto out;
572 	}
573 	if (isitmydescendant(w1, w)) {
574 		mtx_unlock_spin(&w_mtx);
575 		goto out;
576 	}
577 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
578 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
579 
580 			MPASS(j < WITNESS_COUNT);
581 			lock1 = &lle->ll_children[i];
582 			w1 = lock1->li_lock->lo_witness;
583 
584 			/*
585 			 * If this lock doesn't undergo witness checking,
586 			 * then skip it.
587 			 */
588 			if (w1 == NULL) {
589 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
590 				    ("lock missing witness structure"));
591 				continue;
592 			}
593 			/*
594 			 * If we are locking Giant and we slept with this
595 			 * lock, then skip it.
596 			 */
597 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
598 			    lock == &Giant.mtx_object)
599 				continue;
600 			if (!isitmydescendant(w, w1))
601 				continue;
602 			/*
603 			 * We have a lock order violation, check to see if it
604 			 * is allowed or has already been yelled about.
605 			 */
606 			mtx_unlock_spin(&w_mtx);
607 			if (blessed(w, w1))
608 				goto out;
609 			if (lock1->li_lock == &Giant.mtx_object) {
610 				if (w1->w_Giant_squawked)
611 					goto out;
612 				else
613 					w1->w_Giant_squawked = 1;
614 			} else {
615 				if (w1->w_other_squawked)
616 					goto out;
617 				else
618 					w1->w_other_squawked = 1;
619 			}
620 			/*
621 			 * Ok, yell about it.
622 			 */
623 			printf("lock order reversal\n");
624 			/*
625 			 * Try to locate an earlier lock with
626 			 * witness w in our list.
627 			 */
628 			do {
629 				lock2 = &lle->ll_children[i];
630 				MPASS(lock2->li_lock != NULL);
631 				if (lock2->li_lock->lo_witness == w)
632 					break;
633 				i--;
634 				if (i == 0 && lle->ll_next != NULL) {
635 					lle = lle->ll_next;
636 					i = lle->ll_count - 1;
637 					MPASS(i != 0);
638 				}
639 			} while (i >= 0);
640 			if (i < 0) {
641 				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
642 				    lock1->li_lock->lo_name, lock1->li_file,
643 				    lock1->li_line);
644 				printf(" 2nd %p %s @ %s:%d\n", lock,
645 				    lock->lo_name, file, line);
646 			} else {
647 				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
648 				    lock2->li_lock->lo_name, lock2->li_file,
649 				    lock2->li_line);
650 				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
651 				    lock1->li_lock->lo_name, lock1->li_file,
652 				    lock1->li_line);
653 				printf(" 3rd %p %s @ %s:%d\n", lock,
654 				    lock->lo_name, file, line);
655 			}
656 #ifdef DDB
657 			go_into_ddb = 1;
658 #endif /* DDB */
659 			goto out;
660 		}
661 	}
662 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
663 	/*
664 	 * Don't build a new relationship if we are locking Giant just
665 	 * after waking up and the previous lock in the list was acquired
666 	 * prior to blocking.
667 	 */
668 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
669 		mtx_unlock_spin(&w_mtx);
670 	else {
671 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
672 		    lock->lo_name, lock1->li_lock->lo_name);
673 		if (!itismychild(lock1->li_lock->lo_witness, w))
674 			mtx_unlock_spin(&w_mtx);
675 	}
676 
677 out:
678 #ifdef DDB
679 	if (witness_ddb && go_into_ddb)
680 		Debugger(__func__);
681 #endif /* DDB */
682 	w->w_file = file;
683 	w->w_line = line;
684 
685 	lle = *lock_list;
686 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
687 		lle = witness_lock_list_get();
688 		if (lle == NULL)
689 			return;
690 		lle->ll_next = *lock_list;
691 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
692 		    td->td_proc->p_pid, lle);
693 		*lock_list = lle;
694 	}
695 	lock1 = &lle->ll_children[lle->ll_count++];
696 	lock1->li_lock = lock;
697 	lock1->li_line = line;
698 	lock1->li_file = file;
699 	if ((flags & LOP_EXCLUSIVE) != 0)
700 		lock1->li_flags = LI_EXCLUSIVE;
701 	else
702 		lock1->li_flags = 0;
703 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
704 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
705 }
706 
707 void
708 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
709 {
710 	struct lock_instance *instance;
711 	struct lock_class *class;
712 
713 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
714 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
715 		return;
716 	class = lock->lo_class;
717 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
718 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
719 		    class->lc_name, lock->lo_name, file, line);
720 	if ((flags & LOP_TRYLOCK) == 0)
721 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
722 		    lock->lo_name, file, line);
723 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
724 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
725 		    class->lc_name, lock->lo_name, file, line);
726 	instance = find_instance(curthread->td_sleeplocks, lock);
727 	if (instance == NULL)
728 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
729 		    class->lc_name, lock->lo_name, file, line);
730 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
731 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
732 		    class->lc_name, lock->lo_name, file, line);
733 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
734 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
735 		    class->lc_name, lock->lo_name,
736 		    instance->li_flags & LI_RECURSEMASK, file, line);
737 	instance->li_flags |= LI_EXCLUSIVE;
738 }
739 
740 void
741 witness_downgrade(struct lock_object *lock, int flags, const char *file,
742     int line)
743 {
744 	struct lock_instance *instance;
745 	struct lock_class *class;
746 
747 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
748 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
749 		return;
750 	class = lock->lo_class;
751 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
752 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
753 		    class->lc_name, lock->lo_name, file, line);
754 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
755 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
756 		    class->lc_name, lock->lo_name, file, line);
757 	instance = find_instance(curthread->td_sleeplocks, lock);
758 	if (instance == NULL)
759 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
760 		    class->lc_name, lock->lo_name, file, line);
761 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
762 		panic("downgrade of shared lock (%s) %s @ %s:%d",
763 		    class->lc_name, lock->lo_name, file, line);
764 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
765 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
766 		    class->lc_name, lock->lo_name,
767 		    instance->li_flags & LI_RECURSEMASK, file, line);
768 	instance->li_flags &= ~LI_EXCLUSIVE;
769 }
770 
771 void
772 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
773 {
774 	struct lock_list_entry **lock_list, *lle;
775 	struct lock_instance *instance;
776 	struct lock_class *class;
777 	struct thread *td;
778 	register_t s;
779 	int i, j;
780 
781 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
782 	    panicstr != NULL)
783 		return;
784 	td = curthread;
785 	class = lock->lo_class;
786 	if (class->lc_flags & LC_SLEEPLOCK)
787 		lock_list = &td->td_sleeplocks;
788 	else
789 		lock_list = PCPU_PTR(spinlocks);
790 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
791 		for (i = 0; i < (*lock_list)->ll_count; i++) {
792 			instance = &(*lock_list)->ll_children[i];
793 			if (instance->li_lock == lock) {
794 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
795 				    (flags & LOP_EXCLUSIVE) == 0) {
796 					printf(
797 					"shared unlock of (%s) %s @ %s:%d\n",
798 					    class->lc_name, lock->lo_name,
799 					    file, line);
800 					printf(
801 					"while exclusively locked from %s:%d\n",
802 					    instance->li_file,
803 					    instance->li_line);
804 					panic("excl->ushare");
805 				}
806 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
807 				    (flags & LOP_EXCLUSIVE) != 0) {
808 					printf(
809 					"exclusive unlock of (%s) %s @ %s:%d\n",
810 					    class->lc_name, lock->lo_name,
811 					    file, line);
812 					printf(
813 					"while share locked from %s:%d\n",
814 					    instance->li_file,
815 					    instance->li_line);
816 					panic("share->uexcl");
817 				}
818 				/* If we are recursed, unrecurse. */
819 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
820 					CTR4(KTR_WITNESS,
821 				    "%s: pid %d unrecursed on %s r=%d", __func__,
822 					    td->td_proc->p_pid,
823 					    instance->li_lock->lo_name,
824 					    instance->li_flags);
825 					instance->li_flags--;
826 					return;
827 				}
828 				s = intr_disable();
829 				CTR4(KTR_WITNESS,
830 				    "%s: pid %d removed %s from lle[%d]", __func__,
831 				    td->td_proc->p_pid,
832 				    instance->li_lock->lo_name,
833 				    (*lock_list)->ll_count - 1);
834 				(*lock_list)->ll_count--;
835 				for (j = i; j < (*lock_list)->ll_count; j++)
836 					(*lock_list)->ll_children[j] =
837 					    (*lock_list)->ll_children[j + 1];
838 				intr_restore(s);
839 				if ((*lock_list)->ll_count == 0) {
840 					lle = *lock_list;
841 					*lock_list = lle->ll_next;
842 					CTR3(KTR_WITNESS,
843 					    "%s: pid %d removed lle %p", __func__,
844 					    td->td_proc->p_pid, lle);
845 					witness_lock_list_free(lle);
846 				}
847 				return;
848 			}
849 		}
850 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
851 	    file, line);
852 }
853 
854 /*
855  * Warn if any held locks are not sleepable.  Note that Giant and the lock
856  * passed in are both special cases since they are both released during the
857  * sleep process and aren't actually held while the thread is asleep.
858  */
859 int
860 witness_sleep(int check_only, struct lock_object *lock, const char *file,
861 	      int line)
862 {
863 	struct lock_list_entry **lock_list, *lle;
864 	struct lock_instance *lock1;
865 	struct thread *td;
866 	int i, n;
867 
868 	if (witness_dead || panicstr != NULL)
869 		return (0);
870 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
871 	n = 0;
872 	/*
873 	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
874 	 */
875 	critical_enter();
876 	td = curthread;
877 	lock_list = &td->td_sleeplocks;
878 again:
879 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
880 		for (i = lle->ll_count - 1; i >= 0; i--) {
881 			lock1 = &lle->ll_children[i];
882 			if (lock1->li_lock == lock ||
883 			    lock1->li_lock == &Giant.mtx_object)
884 				continue;
885 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
886 				if (check_only == 0) {
887 					CTR3(KTR_WITNESS,
888 				    "pid %d: sleeping with lock (%s) %s held",
889 					    td->td_proc->p_pid,
890 					    lock1->li_lock->lo_class->lc_name,
891 					    lock1->li_lock->lo_name);
892 					lock1->li_flags |= LI_SLEPT;
893 				}
894 				continue;
895 			}
896 			n++;
897 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
898 			    file, line, check_only ? "could sleep" : "sleeping",
899 			    lock1->li_lock->lo_name, lock1->li_file,
900 			    lock1->li_line);
901 		}
902 	if (lock_list == &td->td_sleeplocks) {
903 		lock_list = PCPU_PTR(spinlocks);
904 		goto again;
905 	}
906 #ifdef DDB
907 	if (witness_ddb && n)
908 		Debugger(__func__);
909 #endif /* DDB */
910 	critical_exit();
911 	return (n);
912 }
913 
914 static struct witness *
915 enroll(const char *description, struct lock_class *lock_class)
916 {
917 	struct witness *w;
918 
919 	if (!witness_watch || witness_dead || panicstr != NULL)
920 		return (NULL);
921 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
922 		return (NULL);
923 	mtx_lock_spin(&w_mtx);
924 	STAILQ_FOREACH(w, &w_all, w_list) {
925 		if (strcmp(description, w->w_name) == 0) {
926 			w->w_refcount++;
927 			mtx_unlock_spin(&w_mtx);
928 			if (lock_class != w->w_class)
929 				panic(
930 				"lock (%s) %s does not match earlier (%s) lock",
931 				    description, lock_class->lc_name,
932 				    w->w_class->lc_name);
933 			return (w);
934 		}
935 	}
936 	/*
937 	 * This isn't quite right, as witness_cold is still 0 while we
938 	 * enroll all the locks initialized before witness_initialize().
939 	 */
940 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
941 		mtx_unlock_spin(&w_mtx);
942 		panic("spin lock %s not in order list", description);
943 	}
944 	if ((w = witness_get()) == NULL)
945 		return (NULL);
946 	w->w_name = description;
947 	w->w_class = lock_class;
948 	w->w_refcount = 1;
949 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
950 	if (lock_class->lc_flags & LC_SPINLOCK)
951 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
952 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
953 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
954 	else {
955 		mtx_unlock_spin(&w_mtx);
956 		panic("lock class %s is not sleep or spin",
957 		    lock_class->lc_name);
958 	}
959 	mtx_unlock_spin(&w_mtx);
960 	return (w);
961 }
962 
963 static int
964 itismychild(struct witness *parent, struct witness *child)
965 {
966 	static int recursed;
967 	struct witness_child_list_entry **wcl;
968 	struct witness_list *list;
969 
970 	MPASS(child != NULL && parent != NULL);
971 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
972 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
973 		panic(
974 		"%s: parent (%s) and child (%s) are not the same lock type",
975 		    __func__, parent->w_class->lc_name,
976 		    child->w_class->lc_name);
977 
978 	/*
979 	 * Insert "child" after "parent"
980 	 */
981 	wcl = &parent->w_children;
982 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
983 		wcl = &(*wcl)->wcl_next;
984 	if (*wcl == NULL) {
985 		*wcl = witness_child_get();
986 		if (*wcl == NULL)
987 			return (1);
988 	}
989 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
990 
991 	/*
992 	 * Now prune whole tree.  We look for cases where a lock is now
993 	 * both a descendant and a direct child of a given lock.  In that
994 	 * case, we want to remove the direct child link from the tree.
995 	 */
996 	if (recursed)
997 		return (0);
998 	recursed = 1;
999 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1000 		list = &w_sleep;
1001 	else
1002 		list = &w_spin;
1003 	STAILQ_FOREACH(child, list, w_typelist) {
1004 		STAILQ_FOREACH(parent, list, w_typelist) {
1005 			if (!isitmychild(parent, child))
1006 				continue;
1007 			removechild(parent, child);
1008 			if (isitmydescendant(parent, child))
1009 				continue;
1010 			itismychild(parent, child);
1011 		}
1012 	}
1013 	recursed = 0;
1014 	witness_levelall();
1015 	return (0);
1016 }
1017 
1018 static void
1019 removechild(struct witness *parent, struct witness *child)
1020 {
1021 	struct witness_child_list_entry **wcl, *wcl1;
1022 	int i;
1023 
1024 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1025 		for (i = 0; i < (*wcl)->wcl_count; i++)
1026 			if ((*wcl)->wcl_children[i] == child)
1027 				goto found;
1028 	return;
1029 found:
1030 	(*wcl)->wcl_count--;
1031 	if ((*wcl)->wcl_count > i)
1032 		(*wcl)->wcl_children[i] =
1033 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1034 	MPASS((*wcl)->wcl_children[i] != NULL);
1035 	if ((*wcl)->wcl_count != 0)
1036 		return;
1037 	wcl1 = *wcl;
1038 	*wcl = wcl1->wcl_next;
1039 	witness_child_free(wcl1);
1040 }
1041 
1042 static int
1043 isitmychild(struct witness *parent, struct witness *child)
1044 {
1045 	struct witness_child_list_entry *wcl;
1046 	int i;
1047 
1048 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1049 		for (i = 0; i < wcl->wcl_count; i++) {
1050 			if (wcl->wcl_children[i] == child)
1051 				return (1);
1052 		}
1053 	}
1054 	return (0);
1055 }
1056 
1057 static int
1058 isitmydescendant(struct witness *parent, struct witness *child)
1059 {
1060 	struct witness_child_list_entry *wcl;
1061 	int i, j;
1062 
1063 	if (isitmychild(parent, child))
1064 		return (1);
1065 	j = 0;
1066 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1067 		MPASS(j < 1000);
1068 		for (i = 0; i < wcl->wcl_count; i++) {
1069 			if (isitmydescendant(wcl->wcl_children[i], child))
1070 				return (1);
1071 		}
1072 		j++;
1073 	}
1074 	return (0);
1075 }
1076 
1077 void
1078 witness_levelall (void)
1079 {
1080 	struct witness_list *list;
1081 	struct witness *w, *w1;
1082 
1083 	/*
1084 	 * First clear all levels.
1085 	 */
1086 	STAILQ_FOREACH(w, &w_all, w_list) {
1087 		w->w_level = 0;
1088 	}
1089 
1090 	/*
1091 	 * Look for locks with no parent and level all their descendants.
1092 	 */
1093 	STAILQ_FOREACH(w, &w_all, w_list) {
1094 		/*
1095 		 * This is just an optimization, technically we could get
1096 		 * away just walking the all list each time.
1097 		 */
1098 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1099 			list = &w_sleep;
1100 		else
1101 			list = &w_spin;
1102 		STAILQ_FOREACH(w1, list, w_typelist) {
1103 			if (isitmychild(w1, w))
1104 				goto skip;
1105 		}
1106 		witness_leveldescendents(w, 0);
1107 	skip:
1108 	}
1109 }
1110 
1111 static void
1112 witness_leveldescendents(struct witness *parent, int level)
1113 {
1114 	struct witness_child_list_entry *wcl;
1115 	int i;
1116 
1117 	if (parent->w_level < level)
1118 		parent->w_level = level;
1119 	level++;
1120 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1121 		for (i = 0; i < wcl->wcl_count; i++)
1122 			witness_leveldescendents(wcl->wcl_children[i], level);
1123 }
1124 
1125 static void
1126 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1127 			   struct witness *parent)
1128 {
1129 	struct witness_child_list_entry *wcl;
1130 	int i, level;
1131 
1132 	level =  parent->w_level;
1133 	prnt("%-2d", level);
1134 	for (i = 0; i < level; i++)
1135 		prnt(" ");
1136 	prnt("%s", parent->w_name);
1137 	if (parent->w_file != NULL)
1138 		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1139 		    parent->w_line);
1140 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1141 		for (i = 0; i < wcl->wcl_count; i++)
1142 			    witness_displaydescendants(prnt,
1143 				wcl->wcl_children[i]);
1144 }
1145 
1146 static int
1147 dup_ok(struct witness *w)
1148 {
1149 	const char **dup;
1150 
1151 	for (dup = dup_list; *dup != NULL; dup++)
1152 		if (strcmp(w->w_name, *dup) == 0)
1153 			return (1);
1154 	return (0);
1155 }
1156 
1157 static int
1158 blessed(struct witness *w1, struct witness *w2)
1159 {
1160 	int i;
1161 	struct witness_blessed *b;
1162 
1163 	for (i = 0; i < blessed_count; i++) {
1164 		b = &blessed_list[i];
1165 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1166 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1167 				return (1);
1168 			continue;
1169 		}
1170 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1171 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1172 				return (1);
1173 	}
1174 	return (0);
1175 }
1176 
1177 static struct witness *
1178 witness_get(void)
1179 {
1180 	struct witness *w;
1181 
1182 	if (witness_dead) {
1183 		mtx_unlock_spin(&w_mtx);
1184 		return (NULL);
1185 	}
1186 	if (STAILQ_EMPTY(&w_free)) {
1187 		witness_dead = 1;
1188 		mtx_unlock_spin(&w_mtx);
1189 		printf("%s: witness exhausted\n", __func__);
1190 		return (NULL);
1191 	}
1192 	w = STAILQ_FIRST(&w_free);
1193 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1194 	bzero(w, sizeof(*w));
1195 	return (w);
1196 }
1197 
1198 static void
1199 witness_free(struct witness *w)
1200 {
1201 
1202 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1203 }
1204 
1205 static struct witness_child_list_entry *
1206 witness_child_get(void)
1207 {
1208 	struct witness_child_list_entry *wcl;
1209 
1210 	if (witness_dead) {
1211 		mtx_unlock_spin(&w_mtx);
1212 		return (NULL);
1213 	}
1214 	wcl = w_child_free;
1215 	if (wcl == NULL) {
1216 		witness_dead = 1;
1217 		mtx_unlock_spin(&w_mtx);
1218 		printf("%s: witness exhausted\n", __func__);
1219 		return (NULL);
1220 	}
1221 	w_child_free = wcl->wcl_next;
1222 	bzero(wcl, sizeof(*wcl));
1223 	return (wcl);
1224 }
1225 
1226 static void
1227 witness_child_free(struct witness_child_list_entry *wcl)
1228 {
1229 
1230 	wcl->wcl_next = w_child_free;
1231 	w_child_free = wcl;
1232 }
1233 
1234 static struct lock_list_entry *
1235 witness_lock_list_get(void)
1236 {
1237 	struct lock_list_entry *lle;
1238 
1239 	if (witness_dead)
1240 		return (NULL);
1241 	mtx_lock_spin(&w_mtx);
1242 	lle = w_lock_list_free;
1243 	if (lle == NULL) {
1244 		witness_dead = 1;
1245 		mtx_unlock_spin(&w_mtx);
1246 		printf("%s: witness exhausted\n", __func__);
1247 		return (NULL);
1248 	}
1249 	w_lock_list_free = lle->ll_next;
1250 	mtx_unlock_spin(&w_mtx);
1251 	bzero(lle, sizeof(*lle));
1252 	return (lle);
1253 }
1254 
1255 static void
1256 witness_lock_list_free(struct lock_list_entry *lle)
1257 {
1258 
1259 	mtx_lock_spin(&w_mtx);
1260 	lle->ll_next = w_lock_list_free;
1261 	w_lock_list_free = lle;
1262 	mtx_unlock_spin(&w_mtx);
1263 }
1264 
1265 static struct lock_instance *
1266 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1267 {
1268 	struct lock_list_entry *lle;
1269 	struct lock_instance *instance;
1270 	int i;
1271 
1272 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1273 		for (i = lle->ll_count - 1; i >= 0; i--) {
1274 			instance = &lle->ll_children[i];
1275 			if (instance->li_lock == lock)
1276 				return (instance);
1277 		}
1278 	return (NULL);
1279 }
1280 
1281 int
1282 witness_list_locks(struct lock_list_entry **lock_list)
1283 {
1284 	struct lock_list_entry *lle;
1285 	struct lock_instance *instance;
1286 	struct lock_object *lock;
1287 	int i, nheld;
1288 
1289 	nheld = 0;
1290 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1291 		for (i = lle->ll_count - 1; i >= 0; i--) {
1292 			instance = &lle->ll_children[i];
1293 			lock = instance->li_lock;
1294 			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1295 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1296 			    "exclusive" : "shared",
1297 			    lock->lo_class->lc_name, lock->lo_name, lock,
1298 			    instance->li_file, instance->li_line);
1299 			nheld++;
1300 		}
1301 	return (nheld);
1302 }
1303 
1304 /*
1305  * Calling this on td != curthread is bad unless we are in ddb.
1306  */
1307 int
1308 witness_list(struct thread *td)
1309 {
1310 	int nheld;
1311 
1312 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1313 #ifdef DDB
1314 	KASSERT(td == curthread || db_active,
1315 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1316 	if (!db_active && witness_dead)
1317 		return (0);
1318 #else
1319 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1320 	if (witness_dead)
1321 		return (0);
1322 #endif
1323 	nheld = witness_list_locks(&td->td_sleeplocks);
1324 
1325 	/*
1326 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1327 	 * if td is currently executing on some other CPU and holds spin locks
1328 	 * as we won't display those locks.  If we had a MI way of getting
1329 	 * the per-cpu data for a given cpu then we could use
1330 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1331 	 * and "fix" this.
1332 	 *
1333 	 * That still wouldn't really fix this unless we locked sched_lock
1334 	 * or stopped the other CPU to make sure it wasn't changing the list
1335 	 * out from under us.  It is probably best to just not try to handle
1336 	 * threads on other CPU's for now.
1337 	 */
1338 	if (td == curthread) {
1339 		/*
1340 		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1341 		 * change.
1342 		 */
1343 		critical_enter();
1344 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1345 		critical_exit();
1346 	}
1347 	return (nheld);
1348 }
1349 
1350 void
1351 witness_save(struct lock_object *lock, const char **filep, int *linep)
1352 {
1353 	struct lock_instance *instance;
1354 
1355 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1356 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1357 		return;
1358 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1359 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1360 		    lock->lo_class->lc_name, lock->lo_name);
1361 	instance = find_instance(curthread->td_sleeplocks, lock);
1362 	if (instance == NULL)
1363 		panic("%s: lock (%s) %s not locked", __func__,
1364 		    lock->lo_class->lc_name, lock->lo_name);
1365 	*filep = instance->li_file;
1366 	*linep = instance->li_line;
1367 }
1368 
1369 void
1370 witness_restore(struct lock_object *lock, const char *file, int line)
1371 {
1372 	struct lock_instance *instance;
1373 
1374 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1375 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1376 		return;
1377 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1378 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1379 		    lock->lo_class->lc_name, lock->lo_name);
1380 	instance = find_instance(curthread->td_sleeplocks, lock);
1381 	if (instance == NULL)
1382 		panic("%s: lock (%s) %s not locked", __func__,
1383 		    lock->lo_class->lc_name, lock->lo_name);
1384 	lock->lo_witness->w_file = file;
1385 	lock->lo_witness->w_line = line;
1386 	instance->li_file = file;
1387 	instance->li_line = line;
1388 }
1389 
1390 void
1391 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1392 {
1393 #ifdef INVARIANT_SUPPORT
1394 	struct lock_instance *instance;
1395 
1396 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1397 		return;
1398 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1399 		instance = find_instance(curthread->td_sleeplocks, lock);
1400 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1401 		instance = find_instance(PCPU_GET(spinlocks), lock);
1402 	else {
1403 		panic("Lock (%s) %s is not sleep or spin!",
1404 		    lock->lo_class->lc_name, lock->lo_name);
1405 		return;
1406 	}
1407 	switch (flags) {
1408 	case LA_UNLOCKED:
1409 		if (instance != NULL)
1410 			panic("Lock (%s) %s locked @ %s:%d.",
1411 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1412 		break;
1413 	case LA_LOCKED:
1414 	case LA_LOCKED | LA_RECURSED:
1415 	case LA_LOCKED | LA_NOTRECURSED:
1416 	case LA_SLOCKED:
1417 	case LA_SLOCKED | LA_RECURSED:
1418 	case LA_SLOCKED | LA_NOTRECURSED:
1419 	case LA_XLOCKED:
1420 	case LA_XLOCKED | LA_RECURSED:
1421 	case LA_XLOCKED | LA_NOTRECURSED:
1422 		if (instance == NULL) {
1423 			panic("Lock (%s) %s not locked @ %s:%d.",
1424 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1425 			break;
1426 		}
1427 		if ((flags & LA_XLOCKED) != 0 &&
1428 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1429 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1430 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1431 		if ((flags & LA_SLOCKED) != 0 &&
1432 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1433 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1434 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1435 		if ((flags & LA_RECURSED) != 0 &&
1436 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1437 			panic("Lock (%s) %s not recursed @ %s:%d.",
1438 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1439 		if ((flags & LA_NOTRECURSED) != 0 &&
1440 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1441 			panic("Lock (%s) %s recursed @ %s:%d.",
1442 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1443 		break;
1444 	default:
1445 		panic("Invalid lock assertion at %s:%d.", file, line);
1446 
1447 	}
1448 #endif	/* INVARIANT_SUPPORT */
1449 }
1450 
1451 #ifdef DDB
1452 
1453 DB_SHOW_COMMAND(locks, db_witness_list)
1454 {
1455 	struct thread *td;
1456 	pid_t pid;
1457 	struct proc *p;
1458 
1459 	if (have_addr) {
1460 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1461 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1462 		    ((addr >> 16) % 16) * 10000;
1463 		/* sx_slock(&allproc_lock); */
1464 		FOREACH_PROC_IN_SYSTEM(p) {
1465 			if (p->p_pid == pid)
1466 				break;
1467 		}
1468 		/* sx_sunlock(&allproc_lock); */
1469 		if (p == NULL) {
1470 			db_printf("pid %d not found\n", pid);
1471 			return;
1472 		}
1473 		FOREACH_THREAD_IN_PROC(p, td) {
1474 			witness_list(td);
1475 		}
1476 	} else {
1477 		td = curthread;
1478 		witness_list(td);
1479 	}
1480 }
1481 
1482 DB_SHOW_COMMAND(witness, db_witness_display)
1483 {
1484 
1485 	witness_display(db_printf);
1486 }
1487 #endif
1488