xref: /freebsd/sys/kern/subr_witness.c (revision 11f0b352e05306cf6f1f85e9087022c0a92624a3)
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	blessed(struct witness *, struct witness *);
128 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
129 				     struct witness_list *list);
130 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
131 					   struct witness *);
132 static void	witness_leveldescendents(struct witness *parent, int level);
133 static void	witness_levelall(void);
134 static struct	witness *witness_get(void);
135 static void	witness_free(struct witness *m);
136 static struct	witness_child_list_entry *witness_child_get(void);
137 static void	witness_child_free(struct witness_child_list_entry *wcl);
138 static struct	lock_list_entry *witness_lock_list_get(void);
139 static void	witness_lock_list_free(struct lock_list_entry *lle);
140 static void	witness_display(void(*)(const char *fmt, ...));
141 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
142 					     struct lock_object *lock);
143 
144 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
145 
146 static int witness_watch = 1;
147 TUNABLE_INT("debug.witness_watch", &witness_watch);
148 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
149 
150 #ifdef DDB
151 /*
152  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
153  * drop into kdebug() when:
154  *	- a lock heirarchy violation occurs
155  *	- locks are held when going to sleep.
156  */
157 #ifdef WITNESS_DDB
158 int	witness_ddb = 1;
159 #else
160 int	witness_ddb = 0;
161 #endif
162 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
163 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
164 #endif /* DDB */
165 
166 #ifdef WITNESS_SKIPSPIN
167 int	witness_skipspin = 1;
168 #else
169 int	witness_skipspin = 0;
170 #endif
171 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
172 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
173     "");
174 
175 static struct mtx w_mtx;
176 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
177 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
178 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
179 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
180 static struct witness_child_list_entry *w_child_free = NULL;
181 static struct lock_list_entry *w_lock_list_free = NULL;
182 static int witness_dead;	/* fatal error, probably no memory */
183 
184 static struct witness w_data[WITNESS_COUNT];
185 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
186 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
187 
188 static struct witness_order_list_entry order_lists[] = {
189 	{ "Giant", &lock_class_mtx_sleep },
190 	{ "proctree", &lock_class_sx },
191 	{ "allproc", &lock_class_sx },
192 	{ "sigio lock", &lock_class_mtx_sleep },
193 	{ "process group", &lock_class_mtx_sleep },
194 	{ "process lock", &lock_class_mtx_sleep },
195 	{ "session", &lock_class_mtx_sleep },
196 	{ "uidinfo hash", &lock_class_mtx_sleep },
197 	{ "uidinfo struct", &lock_class_mtx_sleep },
198 	{ NULL, NULL },
199 	/*
200 	 * spin locks
201 	 */
202 #ifdef SMP
203 	{ "ap boot", &lock_class_mtx_spin },
204 #ifdef __i386__
205 	{ "com", &lock_class_mtx_spin },
206 #endif
207 #endif
208 	{ "sio", &lock_class_mtx_spin },
209 #ifdef __i386__
210 	{ "cy", &lock_class_mtx_spin },
211 #endif
212 	{ "ng_node", &lock_class_mtx_spin },
213 	{ "ng_worklist", &lock_class_mtx_spin },
214 	{ "ithread table lock", &lock_class_mtx_spin },
215 	{ "sched lock", &lock_class_mtx_spin },
216 	{ "callout", &lock_class_mtx_spin },
217 	/*
218 	 * leaf locks
219 	 */
220 	{ "allpmaps", &lock_class_mtx_spin },
221 	{ "vm page buckets mutex", &lock_class_mtx_spin },
222 	{ "vm page queue free mutex", &lock_class_mtx_spin },
223 	{ "icu", &lock_class_mtx_spin },
224 #ifdef SMP
225 	{ "smp rendezvous", &lock_class_mtx_spin },
226 #endif
227 	{ "clk", &lock_class_mtx_spin },
228 	{ "mutex profiling lock", &lock_class_mtx_spin },
229 	{ "zombie_thread_lock", &lock_class_mtx_spin },
230 	{ NULL, NULL },
231 	{ NULL, NULL }
232 };
233 
234 /*
235  * Pairs of locks which have been blessed
236  * Don't complain about order problems with blessed locks
237  */
238 static struct witness_blessed blessed_list[] = {
239 };
240 static int blessed_count =
241 	sizeof(blessed_list) / sizeof(struct witness_blessed);
242 
243 /*
244  * List of all locks in the system.
245  */
246 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
247 
248 static struct mtx all_mtx = {
249 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
250 	  "All locks list",		/* mtx_object.lo_name */
251 	  "All locks list",		/* mtx_object.lo_type */
252 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
253 	  { NULL, NULL },		/* mtx_object.lo_list */
254 	  NULL },			/* mtx_object.lo_witness */
255 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
256 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
257 	{ NULL, NULL }			/* mtx_contested */
258 };
259 
260 /*
261  * This global is set to 0 once it becomes safe to use the witness code.
262  */
263 static int witness_cold = 1;
264 
265 /*
266  * Global variables for book keeping.
267  */
268 static int lock_cur_cnt;
269 static int lock_max_cnt;
270 
271 /*
272  * The WITNESS-enabled diagnostic code.
273  */
274 static void
275 witness_initialize(void *dummy __unused)
276 {
277 	struct lock_object *lock;
278 	struct witness_order_list_entry *order;
279 	struct witness *w, *w1;
280 	int i;
281 
282 	/*
283 	 * We have to release Giant before initializing its witness
284 	 * structure so that WITNESS doesn't get confused.
285 	 */
286 	mtx_unlock(&Giant);
287 	mtx_assert(&Giant, MA_NOTOWNED);
288 
289 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
290 	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
291 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
292 	    MTX_NOWITNESS);
293 	for (i = 0; i < WITNESS_COUNT; i++)
294 		witness_free(&w_data[i]);
295 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
296 		witness_child_free(&w_childdata[i]);
297 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
298 		witness_lock_list_free(&w_locklistdata[i]);
299 
300 	/* First add in all the specified order lists. */
301 	for (order = order_lists; order->w_name != NULL; order++) {
302 		w = enroll(order->w_name, order->w_class);
303 		if (w == NULL)
304 			continue;
305 		w->w_file = "order list";
306 		for (order++; order->w_name != NULL; order++) {
307 			w1 = enroll(order->w_name, order->w_class);
308 			if (w1 == NULL)
309 				continue;
310 			w1->w_file = "order list";
311 			itismychild(w, w1);
312 			w = w1;
313 		}
314 	}
315 
316 	/* Iterate through all locks and add them to witness. */
317 	mtx_lock(&all_mtx);
318 	TAILQ_FOREACH(lock, &all_locks, lo_list) {
319 		if (lock->lo_flags & LO_WITNESS)
320 			lock->lo_witness = enroll(lock->lo_type,
321 			    lock->lo_class);
322 		else
323 			lock->lo_witness = NULL;
324 	}
325 	mtx_unlock(&all_mtx);
326 
327 	/* Mark the witness code as being ready for use. */
328 	atomic_store_rel_int(&witness_cold, 0);
329 
330 	mtx_lock(&Giant);
331 }
332 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
333 
334 void
335 witness_init(struct lock_object *lock)
336 {
337 	struct lock_class *class;
338 
339 	class = lock->lo_class;
340 	if (lock->lo_flags & LO_INITIALIZED)
341 		panic("%s: lock (%s) %s is already initialized", __func__,
342 		    class->lc_name, lock->lo_name);
343 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
344 	    (class->lc_flags & LC_RECURSABLE) == 0)
345 		panic("%s: lock (%s) %s can not be recursable", __func__,
346 		    class->lc_name, lock->lo_name);
347 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
348 	    (class->lc_flags & LC_SLEEPABLE) == 0)
349 		panic("%s: lock (%s) %s can not be sleepable", __func__,
350 		    class->lc_name, lock->lo_name);
351 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
352 	    (class->lc_flags & LC_UPGRADABLE) == 0)
353 		panic("%s: lock (%s) %s can not be upgradable", __func__,
354 		    class->lc_name, lock->lo_name);
355 
356 	mtx_lock(&all_mtx);
357 	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
358 	lock->lo_flags |= LO_INITIALIZED;
359 	lock_cur_cnt++;
360 	if (lock_cur_cnt > lock_max_cnt)
361 		lock_max_cnt = lock_cur_cnt;
362 	mtx_unlock(&all_mtx);
363 	if (!witness_cold && !witness_dead && panicstr == NULL &&
364 	    (lock->lo_flags & LO_WITNESS) != 0)
365 		lock->lo_witness = enroll(lock->lo_type, class);
366 	else
367 		lock->lo_witness = NULL;
368 }
369 
370 void
371 witness_destroy(struct lock_object *lock)
372 {
373 	struct witness *w;
374 
375 	if (witness_cold)
376 		panic("lock (%s) %s destroyed while witness_cold",
377 		    lock->lo_class->lc_name, lock->lo_name);
378 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
379 		panic("%s: lock (%s) %s is not initialized", __func__,
380 		    lock->lo_class->lc_name, lock->lo_name);
381 
382 	/* XXX: need to verify that no one holds the lock */
383 	w = lock->lo_witness;
384 	if (w != NULL) {
385 		mtx_lock_spin(&w_mtx);
386 		MPASS(w->w_refcount > 0);
387 		w->w_refcount--;
388 		mtx_unlock_spin(&w_mtx);
389 	}
390 
391 	mtx_lock(&all_mtx);
392 	lock_cur_cnt--;
393 	TAILQ_REMOVE(&all_locks, lock, lo_list);
394 	lock->lo_flags &= ~LO_INITIALIZED;
395 	mtx_unlock(&all_mtx);
396 }
397 
398 static void
399 witness_display_list(void(*prnt)(const char *fmt, ...),
400 		     struct witness_list *list)
401 {
402 	struct witness *w, *w1;
403 	int found;
404 
405 	STAILQ_FOREACH(w, list, w_typelist) {
406 		if (w->w_file == NULL)
407 			continue;
408 		found = 0;
409 		STAILQ_FOREACH(w1, list, w_typelist) {
410 			if (isitmychild(w1, w)) {
411 				found++;
412 				break;
413 			}
414 		}
415 		if (found)
416 			continue;
417 		/*
418 		 * This lock has no anscestors, display its descendants.
419 		 */
420 		witness_displaydescendants(prnt, w);
421 	}
422 }
423 
424 static void
425 witness_display(void(*prnt)(const char *fmt, ...))
426 {
427 	struct witness *w;
428 
429 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
430 	witness_levelall();
431 
432 	/*
433 	 * First, handle sleep locks which have been acquired at least
434 	 * once.
435 	 */
436 	prnt("Sleep locks:\n");
437 	witness_display_list(prnt, &w_sleep);
438 
439 	/*
440 	 * Now do spin locks which have been acquired at least once.
441 	 */
442 	prnt("\nSpin locks:\n");
443 	witness_display_list(prnt, &w_spin);
444 
445 	/*
446 	 * Finally, any locks which have not been acquired yet.
447 	 */
448 	prnt("\nLocks which were never acquired:\n");
449 	STAILQ_FOREACH(w, &w_all, w_list) {
450 		if (w->w_file != NULL || w->w_refcount == 0)
451 			continue;
452 		prnt("%s\n", w->w_name);
453 	}
454 }
455 
456 void
457 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
458 {
459 	struct lock_list_entry **lock_list, *lle;
460 	struct lock_instance *lock1, *lock2;
461 	struct lock_class *class;
462 	struct witness *w, *w1;
463 	struct thread *td;
464 	int i, j;
465 #ifdef DDB
466 	int go_into_ddb = 0;
467 #endif /* DDB */
468 
469 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
470 	    panicstr != NULL)
471 		return;
472 	w = lock->lo_witness;
473 	class = lock->lo_class;
474 	td = curthread;
475 
476 	if (class->lc_flags & LC_SLEEPLOCK) {
477 		/*
478 		 * Since spin locks include a critical section, this check
479 		 * impliclty enforces a lock order of all sleep locks before
480 		 * all spin locks.
481 		 */
482 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
483 			panic("blockable sleep lock (%s) %s @ %s:%d",
484 			    class->lc_name, lock->lo_name, file, line);
485 		lock_list = &td->td_sleeplocks;
486 	} else
487 		lock_list = PCPU_PTR(spinlocks);
488 
489 	/*
490 	 * Try locks do not block if they fail to acquire the lock, thus
491 	 * there is no danger of deadlocks or of switching while holding a
492 	 * spin lock if we acquire a lock via a try operation.
493 	 */
494 	if (flags & LOP_TRYLOCK)
495 		goto out;
496 
497 	/*
498 	 * Is this the first lock acquired?  If so, then no order checking
499 	 * is needed.
500 	 */
501 	if (*lock_list == NULL)
502 		goto out;
503 
504 	/*
505 	 * Check to see if we are recursing on a lock we already own.
506 	 */
507 	lock1 = find_instance(*lock_list, lock);
508 	if (lock1 != NULL) {
509 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
510 		    (flags & LOP_EXCLUSIVE) == 0) {
511 			printf("shared lock of (%s) %s @ %s:%d\n",
512 			    class->lc_name, lock->lo_name, file, line);
513 			printf("while exclusively locked from %s:%d\n",
514 			    lock1->li_file, lock1->li_line);
515 			panic("share->excl");
516 		}
517 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
518 		    (flags & LOP_EXCLUSIVE) != 0) {
519 			printf("exclusive lock of (%s) %s @ %s:%d\n",
520 			    class->lc_name, lock->lo_name, file, line);
521 			printf("while share locked from %s:%d\n",
522 			    lock1->li_file, lock1->li_line);
523 			panic("excl->share");
524 		}
525 		lock1->li_flags++;
526 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
527 			printf(
528 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
529 			    class->lc_name, lock->lo_name, file, line);
530 			printf("first acquired @ %s:%d\n", lock1->li_file,
531 			    lock1->li_line);
532 			panic("recurse");
533 		}
534 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
535 		    td->td_proc->p_pid, lock->lo_name,
536 		    lock1->li_flags & LI_RECURSEMASK);
537 		lock1->li_file = file;
538 		lock1->li_line = line;
539 		return;
540 	}
541 
542 	/*
543 	 * Check for duplicate locks of the same type.  Note that we only
544 	 * have to check for this on the last lock we just acquired.  Any
545 	 * other cases will be caught as lock order violations.
546 	 */
547 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
548 	w1 = lock1->li_lock->lo_witness;
549 	if (w1 == w) {
550 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
551 			goto out;
552 		w->w_same_squawked = 1;
553 		printf("acquiring duplicate lock of same type: \"%s\"\n",
554 			lock->lo_type);
555 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
556 		    lock1->li_file, lock1->li_line);
557 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
558 #ifdef DDB
559 		go_into_ddb = 1;
560 #endif /* DDB */
561 		goto out;
562 	}
563 	MPASS(!mtx_owned(&w_mtx));
564 	mtx_lock_spin(&w_mtx);
565 	/*
566 	 * If we have a known higher number just say ok
567 	 */
568 	if (witness_watch > 1 && w->w_level > w1->w_level) {
569 		mtx_unlock_spin(&w_mtx);
570 		goto out;
571 	}
572 	if (isitmydescendant(w1, w)) {
573 		mtx_unlock_spin(&w_mtx);
574 		goto out;
575 	}
576 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
577 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
578 
579 			MPASS(j < WITNESS_COUNT);
580 			lock1 = &lle->ll_children[i];
581 			w1 = lock1->li_lock->lo_witness;
582 
583 			/*
584 			 * If this lock doesn't undergo witness checking,
585 			 * then skip it.
586 			 */
587 			if (w1 == NULL) {
588 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
589 				    ("lock missing witness structure"));
590 				continue;
591 			}
592 			/*
593 			 * If we are locking Giant and we slept with this
594 			 * lock, then skip it.
595 			 */
596 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
597 			    lock == &Giant.mtx_object)
598 				continue;
599 			/*
600 			 * If we are locking a sleepable lock and this lock
601 			 * isn't sleepable and isn't Giant, we want to treat
602 			 * it as a lock order violation to enfore a general
603 			 * lock order of sleepable locks before non-sleepable
604 			 * locks.  Thus, we only bother checking the lock
605 			 * order hierarchy if we pass the initial test.
606 			 */
607 			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
608 			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
609 			    lock1->li_lock != &Giant.mtx_object)) &&
610 			    !isitmydescendant(w, w1))
611 				continue;
612 			/*
613 			 * We have a lock order violation, check to see if it
614 			 * is allowed or has already been yelled about.
615 			 */
616 			mtx_unlock_spin(&w_mtx);
617 			if (blessed(w, w1))
618 				goto out;
619 			if (lock1->li_lock == &Giant.mtx_object) {
620 				if (w1->w_Giant_squawked)
621 					goto out;
622 				else
623 					w1->w_Giant_squawked = 1;
624 			} else {
625 				if (w1->w_other_squawked)
626 					goto out;
627 				else
628 					w1->w_other_squawked = 1;
629 			}
630 			/*
631 			 * Ok, yell about it.
632 			 */
633 			printf("lock order reversal\n");
634 			/*
635 			 * Try to locate an earlier lock with
636 			 * witness w in our list.
637 			 */
638 			do {
639 				lock2 = &lle->ll_children[i];
640 				MPASS(lock2->li_lock != NULL);
641 				if (lock2->li_lock->lo_witness == w)
642 					break;
643 				i--;
644 				if (i == 0 && lle->ll_next != NULL) {
645 					lle = lle->ll_next;
646 					i = lle->ll_count - 1;
647 					MPASS(i != 0);
648 				}
649 			} while (i >= 0);
650 			if (i < 0) {
651 				printf(" 1st %p %s (%s) @ %s:%d\n",
652 				    lock1->li_lock, lock1->li_lock->lo_name,
653 				    lock1->li_lock->lo_type, lock1->li_file,
654 				    lock1->li_line);
655 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
656 				    lock->lo_name, lock->lo_type, file, line);
657 			} else {
658 				printf(" 1st %p %s (%s) @ %s:%d\n",
659 				    lock2->li_lock, lock2->li_lock->lo_name,
660 				    lock2->li_lock->lo_type, lock2->li_file,
661 				    lock2->li_line);
662 				printf(" 2nd %p %s (%s) @ %s:%d\n",
663 				    lock1->li_lock, lock1->li_lock->lo_name,
664 				    lock1->li_lock->lo_type, lock1->li_file,
665 				    lock1->li_line);
666 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
667 				    lock->lo_name, lock->lo_type, file, line);
668 			}
669 #ifdef DDB
670 			go_into_ddb = 1;
671 #endif /* DDB */
672 			goto out;
673 		}
674 	}
675 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
676 	/*
677 	 * Don't build a new relationship if we are locking Giant just
678 	 * after waking up and the previous lock in the list was acquired
679 	 * prior to blocking.
680 	 */
681 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
682 		mtx_unlock_spin(&w_mtx);
683 	else {
684 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
685 		    lock->lo_type, lock1->li_lock->lo_type);
686 		if (!itismychild(lock1->li_lock->lo_witness, w))
687 			mtx_unlock_spin(&w_mtx);
688 	}
689 
690 out:
691 #ifdef DDB
692 	if (witness_ddb && go_into_ddb)
693 		Debugger(__func__);
694 #endif /* DDB */
695 	w->w_file = file;
696 	w->w_line = line;
697 
698 	lle = *lock_list;
699 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
700 		lle = witness_lock_list_get();
701 		if (lle == NULL)
702 			return;
703 		lle->ll_next = *lock_list;
704 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
705 		    td->td_proc->p_pid, lle);
706 		*lock_list = lle;
707 	}
708 	lock1 = &lle->ll_children[lle->ll_count++];
709 	lock1->li_lock = lock;
710 	lock1->li_line = line;
711 	lock1->li_file = file;
712 	if ((flags & LOP_EXCLUSIVE) != 0)
713 		lock1->li_flags = LI_EXCLUSIVE;
714 	else
715 		lock1->li_flags = 0;
716 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
717 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
718 }
719 
720 void
721 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
722 {
723 	struct lock_instance *instance;
724 	struct lock_class *class;
725 
726 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
727 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
728 		return;
729 	class = lock->lo_class;
730 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
731 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
732 		    class->lc_name, lock->lo_name, file, line);
733 	if ((flags & LOP_TRYLOCK) == 0)
734 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
735 		    lock->lo_name, file, line);
736 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
737 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
738 		    class->lc_name, lock->lo_name, file, line);
739 	instance = find_instance(curthread->td_sleeplocks, lock);
740 	if (instance == NULL)
741 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
742 		    class->lc_name, lock->lo_name, file, line);
743 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
744 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
745 		    class->lc_name, lock->lo_name, file, line);
746 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
747 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
748 		    class->lc_name, lock->lo_name,
749 		    instance->li_flags & LI_RECURSEMASK, file, line);
750 	instance->li_flags |= LI_EXCLUSIVE;
751 }
752 
753 void
754 witness_downgrade(struct lock_object *lock, int flags, const char *file,
755     int line)
756 {
757 	struct lock_instance *instance;
758 	struct lock_class *class;
759 
760 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
761 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
762 		return;
763 	class = lock->lo_class;
764 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
765 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
766 		    class->lc_name, lock->lo_name, file, line);
767 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
768 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
769 		    class->lc_name, lock->lo_name, file, line);
770 	instance = find_instance(curthread->td_sleeplocks, lock);
771 	if (instance == NULL)
772 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
773 		    class->lc_name, lock->lo_name, file, line);
774 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
775 		panic("downgrade of shared lock (%s) %s @ %s:%d",
776 		    class->lc_name, lock->lo_name, file, line);
777 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
778 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
779 		    class->lc_name, lock->lo_name,
780 		    instance->li_flags & LI_RECURSEMASK, file, line);
781 	instance->li_flags &= ~LI_EXCLUSIVE;
782 }
783 
784 void
785 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
786 {
787 	struct lock_list_entry **lock_list, *lle;
788 	struct lock_instance *instance;
789 	struct lock_class *class;
790 	struct thread *td;
791 	register_t s;
792 	int i, j;
793 
794 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
795 	    panicstr != NULL)
796 		return;
797 	td = curthread;
798 	class = lock->lo_class;
799 	if (class->lc_flags & LC_SLEEPLOCK)
800 		lock_list = &td->td_sleeplocks;
801 	else
802 		lock_list = PCPU_PTR(spinlocks);
803 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
804 		for (i = 0; i < (*lock_list)->ll_count; i++) {
805 			instance = &(*lock_list)->ll_children[i];
806 			if (instance->li_lock == lock) {
807 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
808 				    (flags & LOP_EXCLUSIVE) == 0) {
809 					printf(
810 					"shared unlock of (%s) %s @ %s:%d\n",
811 					    class->lc_name, lock->lo_name,
812 					    file, line);
813 					printf(
814 					"while exclusively locked from %s:%d\n",
815 					    instance->li_file,
816 					    instance->li_line);
817 					panic("excl->ushare");
818 				}
819 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
820 				    (flags & LOP_EXCLUSIVE) != 0) {
821 					printf(
822 					"exclusive unlock of (%s) %s @ %s:%d\n",
823 					    class->lc_name, lock->lo_name,
824 					    file, line);
825 					printf(
826 					"while share locked from %s:%d\n",
827 					    instance->li_file,
828 					    instance->li_line);
829 					panic("share->uexcl");
830 				}
831 				/* If we are recursed, unrecurse. */
832 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
833 					CTR4(KTR_WITNESS,
834 				    "%s: pid %d unrecursed on %s r=%d", __func__,
835 					    td->td_proc->p_pid,
836 					    instance->li_lock->lo_name,
837 					    instance->li_flags);
838 					instance->li_flags--;
839 					return;
840 				}
841 				s = intr_disable();
842 				CTR4(KTR_WITNESS,
843 				    "%s: pid %d removed %s from lle[%d]", __func__,
844 				    td->td_proc->p_pid,
845 				    instance->li_lock->lo_name,
846 				    (*lock_list)->ll_count - 1);
847 				for (j = i; j < (*lock_list)->ll_count - 1; j++)
848 					(*lock_list)->ll_children[j] =
849 					    (*lock_list)->ll_children[j + 1];
850 				(*lock_list)->ll_count--;
851 				intr_restore(s);
852 				if ((*lock_list)->ll_count == 0) {
853 					lle = *lock_list;
854 					*lock_list = lle->ll_next;
855 					CTR3(KTR_WITNESS,
856 					    "%s: pid %d removed lle %p", __func__,
857 					    td->td_proc->p_pid, lle);
858 					witness_lock_list_free(lle);
859 				}
860 				return;
861 			}
862 		}
863 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
864 	    file, line);
865 }
866 
867 /*
868  * Warn if any held locks are not sleepable.  Note that Giant and the lock
869  * passed in are both special cases since they are both released during the
870  * sleep process and aren't actually held while the thread is asleep.
871  */
872 int
873 witness_sleep(int check_only, struct lock_object *lock, const char *file,
874 	      int line)
875 {
876 	struct lock_list_entry **lock_list, *lle;
877 	struct lock_instance *lock1;
878 	struct thread *td;
879 	int i, n;
880 
881 	if (witness_cold || witness_dead || panicstr != NULL)
882 		return (0);
883 	n = 0;
884 	td = curthread;
885 	lock_list = &td->td_sleeplocks;
886 again:
887 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
888 		for (i = lle->ll_count - 1; i >= 0; i--) {
889 			lock1 = &lle->ll_children[i];
890 			if (lock1->li_lock == lock ||
891 			    lock1->li_lock == &Giant.mtx_object)
892 				continue;
893 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
894 				if (check_only == 0) {
895 					CTR3(KTR_WITNESS,
896 				    "pid %d: sleeping with lock (%s) %s held",
897 					    td->td_proc->p_pid,
898 					    lock1->li_lock->lo_class->lc_name,
899 					    lock1->li_lock->lo_name);
900 					lock1->li_flags |= LI_SLEPT;
901 				}
902 				continue;
903 			}
904 			n++;
905 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
906 			    file, line, check_only ? "could sleep" : "sleeping",
907 			    lock1->li_lock->lo_name, lock1->li_file,
908 			    lock1->li_line);
909 		}
910 	if (lock_list == &td->td_sleeplocks && PCPU_GET(spinlocks) != NULL) {
911 		/*
912 		 * Since we already hold a spinlock preemption is
913 		 * already blocked.
914 		 */
915 		lock_list = PCPU_PTR(spinlocks);
916 		goto again;
917 	}
918 #ifdef DDB
919 	if (witness_ddb && n)
920 		Debugger(__func__);
921 #endif /* DDB */
922 	return (n);
923 }
924 
925 static struct witness *
926 enroll(const char *description, struct lock_class *lock_class)
927 {
928 	struct witness *w;
929 
930 	if (!witness_watch || witness_dead || panicstr != NULL)
931 		return (NULL);
932 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
933 		return (NULL);
934 	mtx_lock_spin(&w_mtx);
935 	STAILQ_FOREACH(w, &w_all, w_list) {
936 		if (w->w_name == description || (w->w_refcount > 0 &&
937 		    strcmp(description, w->w_name) == 0)) {
938 			w->w_refcount++;
939 			mtx_unlock_spin(&w_mtx);
940 			if (lock_class != w->w_class)
941 				panic(
942 				"lock (%s) %s does not match earlier (%s) lock",
943 				    description, lock_class->lc_name,
944 				    w->w_class->lc_name);
945 			return (w);
946 		}
947 	}
948 	/*
949 	 * This isn't quite right, as witness_cold is still 0 while we
950 	 * enroll all the locks initialized before witness_initialize().
951 	 */
952 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
953 		mtx_unlock_spin(&w_mtx);
954 		panic("spin lock %s not in order list", description);
955 	}
956 	if ((w = witness_get()) == NULL)
957 		return (NULL);
958 	w->w_name = description;
959 	w->w_class = lock_class;
960 	w->w_refcount = 1;
961 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
962 	if (lock_class->lc_flags & LC_SPINLOCK)
963 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
964 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
965 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
966 	else {
967 		mtx_unlock_spin(&w_mtx);
968 		panic("lock class %s is not sleep or spin",
969 		    lock_class->lc_name);
970 	}
971 	mtx_unlock_spin(&w_mtx);
972 	return (w);
973 }
974 
975 static int
976 itismychild(struct witness *parent, struct witness *child)
977 {
978 	static int recursed;
979 	struct witness_child_list_entry **wcl;
980 	struct witness_list *list;
981 
982 	MPASS(child != NULL && parent != NULL);
983 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
984 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
985 		panic(
986 		"%s: parent (%s) and child (%s) are not the same lock type",
987 		    __func__, parent->w_class->lc_name,
988 		    child->w_class->lc_name);
989 
990 	/*
991 	 * Insert "child" after "parent"
992 	 */
993 	wcl = &parent->w_children;
994 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
995 		wcl = &(*wcl)->wcl_next;
996 	if (*wcl == NULL) {
997 		*wcl = witness_child_get();
998 		if (*wcl == NULL)
999 			return (1);
1000 	}
1001 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1002 
1003 	/*
1004 	 * Now prune whole tree.  We look for cases where a lock is now
1005 	 * both a descendant and a direct child of a given lock.  In that
1006 	 * case, we want to remove the direct child link from the tree.
1007 	 */
1008 	if (recursed)
1009 		return (0);
1010 	recursed = 1;
1011 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1012 		list = &w_sleep;
1013 	else
1014 		list = &w_spin;
1015 	STAILQ_FOREACH(child, list, w_typelist) {
1016 		STAILQ_FOREACH(parent, list, w_typelist) {
1017 			if (!isitmychild(parent, child))
1018 				continue;
1019 			removechild(parent, child);
1020 			if (isitmydescendant(parent, child))
1021 				continue;
1022 			itismychild(parent, child);
1023 		}
1024 	}
1025 	recursed = 0;
1026 	witness_levelall();
1027 	return (0);
1028 }
1029 
1030 static void
1031 removechild(struct witness *parent, struct witness *child)
1032 {
1033 	struct witness_child_list_entry **wcl, *wcl1;
1034 	int i;
1035 
1036 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1037 		for (i = 0; i < (*wcl)->wcl_count; i++)
1038 			if ((*wcl)->wcl_children[i] == child)
1039 				goto found;
1040 	return;
1041 found:
1042 	(*wcl)->wcl_count--;
1043 	if ((*wcl)->wcl_count > i)
1044 		(*wcl)->wcl_children[i] =
1045 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1046 	MPASS((*wcl)->wcl_children[i] != NULL);
1047 	if ((*wcl)->wcl_count != 0)
1048 		return;
1049 	wcl1 = *wcl;
1050 	*wcl = wcl1->wcl_next;
1051 	witness_child_free(wcl1);
1052 }
1053 
1054 static int
1055 isitmychild(struct witness *parent, struct witness *child)
1056 {
1057 	struct witness_child_list_entry *wcl;
1058 	int i;
1059 
1060 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1061 		for (i = 0; i < wcl->wcl_count; i++) {
1062 			if (wcl->wcl_children[i] == child)
1063 				return (1);
1064 		}
1065 	}
1066 	return (0);
1067 }
1068 
1069 static int
1070 isitmydescendant(struct witness *parent, struct witness *child)
1071 {
1072 	struct witness_child_list_entry *wcl;
1073 	int i, j;
1074 
1075 	if (isitmychild(parent, child))
1076 		return (1);
1077 	j = 0;
1078 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1079 		MPASS(j < 1000);
1080 		for (i = 0; i < wcl->wcl_count; i++) {
1081 			if (isitmydescendant(wcl->wcl_children[i], child))
1082 				return (1);
1083 		}
1084 		j++;
1085 	}
1086 	return (0);
1087 }
1088 
1089 void
1090 witness_levelall (void)
1091 {
1092 	struct witness_list *list;
1093 	struct witness *w, *w1;
1094 
1095 	/*
1096 	 * First clear all levels.
1097 	 */
1098 	STAILQ_FOREACH(w, &w_all, w_list) {
1099 		w->w_level = 0;
1100 	}
1101 
1102 	/*
1103 	 * Look for locks with no parent and level all their descendants.
1104 	 */
1105 	STAILQ_FOREACH(w, &w_all, w_list) {
1106 		/*
1107 		 * This is just an optimization, technically we could get
1108 		 * away just walking the all list each time.
1109 		 */
1110 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1111 			list = &w_sleep;
1112 		else
1113 			list = &w_spin;
1114 		STAILQ_FOREACH(w1, list, w_typelist) {
1115 			if (isitmychild(w1, w))
1116 				goto skip;
1117 		}
1118 		witness_leveldescendents(w, 0);
1119 	skip:
1120 		;	/* silence GCC 3.x */
1121 	}
1122 }
1123 
1124 static void
1125 witness_leveldescendents(struct witness *parent, int level)
1126 {
1127 	struct witness_child_list_entry *wcl;
1128 	int i;
1129 
1130 	if (parent->w_level < level)
1131 		parent->w_level = level;
1132 	level++;
1133 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1134 		for (i = 0; i < wcl->wcl_count; i++)
1135 			witness_leveldescendents(wcl->wcl_children[i], level);
1136 }
1137 
1138 static void
1139 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1140 			   struct witness *parent)
1141 {
1142 	struct witness_child_list_entry *wcl;
1143 	int i, level;
1144 
1145 	level = parent->w_level;
1146 	prnt("%-2d", level);
1147 	for (i = 0; i < level; i++)
1148 		prnt(" ");
1149 	if (parent->w_refcount > 0) {
1150 		prnt("%s", parent->w_name);
1151 		if (parent->w_file != NULL)
1152 			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1153 			    parent->w_line);
1154 	} else
1155 		prnt("(dead)\n");
1156 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1157 		for (i = 0; i < wcl->wcl_count; i++)
1158 			    witness_displaydescendants(prnt,
1159 				wcl->wcl_children[i]);
1160 }
1161 
1162 static int
1163 blessed(struct witness *w1, struct witness *w2)
1164 {
1165 	int i;
1166 	struct witness_blessed *b;
1167 
1168 	for (i = 0; i < blessed_count; i++) {
1169 		b = &blessed_list[i];
1170 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1171 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1172 				return (1);
1173 			continue;
1174 		}
1175 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1176 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1177 				return (1);
1178 	}
1179 	return (0);
1180 }
1181 
1182 static struct witness *
1183 witness_get(void)
1184 {
1185 	struct witness *w;
1186 
1187 	if (witness_dead) {
1188 		mtx_unlock_spin(&w_mtx);
1189 		return (NULL);
1190 	}
1191 	if (STAILQ_EMPTY(&w_free)) {
1192 		witness_dead = 1;
1193 		mtx_unlock_spin(&w_mtx);
1194 		printf("%s: witness exhausted\n", __func__);
1195 		return (NULL);
1196 	}
1197 	w = STAILQ_FIRST(&w_free);
1198 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1199 	bzero(w, sizeof(*w));
1200 	return (w);
1201 }
1202 
1203 static void
1204 witness_free(struct witness *w)
1205 {
1206 
1207 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1208 }
1209 
1210 static struct witness_child_list_entry *
1211 witness_child_get(void)
1212 {
1213 	struct witness_child_list_entry *wcl;
1214 
1215 	if (witness_dead) {
1216 		mtx_unlock_spin(&w_mtx);
1217 		return (NULL);
1218 	}
1219 	wcl = w_child_free;
1220 	if (wcl == NULL) {
1221 		witness_dead = 1;
1222 		mtx_unlock_spin(&w_mtx);
1223 		printf("%s: witness exhausted\n", __func__);
1224 		return (NULL);
1225 	}
1226 	w_child_free = wcl->wcl_next;
1227 	bzero(wcl, sizeof(*wcl));
1228 	return (wcl);
1229 }
1230 
1231 static void
1232 witness_child_free(struct witness_child_list_entry *wcl)
1233 {
1234 
1235 	wcl->wcl_next = w_child_free;
1236 	w_child_free = wcl;
1237 }
1238 
1239 static struct lock_list_entry *
1240 witness_lock_list_get(void)
1241 {
1242 	struct lock_list_entry *lle;
1243 
1244 	if (witness_dead)
1245 		return (NULL);
1246 	mtx_lock_spin(&w_mtx);
1247 	lle = w_lock_list_free;
1248 	if (lle == NULL) {
1249 		witness_dead = 1;
1250 		mtx_unlock_spin(&w_mtx);
1251 		printf("%s: witness exhausted\n", __func__);
1252 		return (NULL);
1253 	}
1254 	w_lock_list_free = lle->ll_next;
1255 	mtx_unlock_spin(&w_mtx);
1256 	bzero(lle, sizeof(*lle));
1257 	return (lle);
1258 }
1259 
1260 static void
1261 witness_lock_list_free(struct lock_list_entry *lle)
1262 {
1263 
1264 	mtx_lock_spin(&w_mtx);
1265 	lle->ll_next = w_lock_list_free;
1266 	w_lock_list_free = lle;
1267 	mtx_unlock_spin(&w_mtx);
1268 }
1269 
1270 static struct lock_instance *
1271 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1272 {
1273 	struct lock_list_entry *lle;
1274 	struct lock_instance *instance;
1275 	int i;
1276 
1277 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1278 		for (i = lle->ll_count - 1; i >= 0; i--) {
1279 			instance = &lle->ll_children[i];
1280 			if (instance->li_lock == lock)
1281 				return (instance);
1282 		}
1283 	return (NULL);
1284 }
1285 
1286 int
1287 witness_list_locks(struct lock_list_entry **lock_list)
1288 {
1289 	struct lock_list_entry *lle;
1290 	struct lock_instance *instance;
1291 	struct lock_object *lock;
1292 	int i, nheld;
1293 
1294 	nheld = 0;
1295 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1296 		for (i = lle->ll_count - 1; i >= 0; i--) {
1297 			instance = &lle->ll_children[i];
1298 			lock = instance->li_lock;
1299 			printf("%s %s %s",
1300 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1301 			    "exclusive" : "shared",
1302 			    lock->lo_class->lc_name, lock->lo_name);
1303 			if (lock->lo_type != lock->lo_name)
1304 				printf(" (%s)", lock->lo_type);
1305 			printf(" r = %d (%p) locked @ %s:%d\n",
1306 			    instance->li_flags & LI_RECURSEMASK, lock,
1307 			    instance->li_file, instance->li_line);
1308 			nheld++;
1309 		}
1310 	return (nheld);
1311 }
1312 
1313 /*
1314  * Calling this on td != curthread is bad unless we are in ddb.
1315  */
1316 int
1317 witness_list(struct thread *td)
1318 {
1319 	int nheld;
1320 
1321 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1322 #ifdef DDB
1323 	KASSERT(td == curthread || db_active,
1324 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1325 	if (!db_active && witness_dead)
1326 		return (0);
1327 #else
1328 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1329 	if (witness_dead)
1330 		return (0);
1331 #endif
1332 	nheld = witness_list_locks(&td->td_sleeplocks);
1333 
1334 	/*
1335 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1336 	 * if td is currently executing on some other CPU and holds spin locks
1337 	 * as we won't display those locks.  If we had a MI way of getting
1338 	 * the per-cpu data for a given cpu then we could use
1339 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1340 	 * and "fix" this.
1341 	 *
1342 	 * That still wouldn't really fix this unless we locked sched_lock
1343 	 * or stopped the other CPU to make sure it wasn't changing the list
1344 	 * out from under us.  It is probably best to just not try to handle
1345 	 * threads on other CPU's for now.
1346 	 */
1347 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1348 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1349 
1350 	return (nheld);
1351 }
1352 
1353 void
1354 witness_save(struct lock_object *lock, const char **filep, int *linep)
1355 {
1356 	struct lock_instance *instance;
1357 
1358 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1359 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1360 		return;
1361 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1362 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1363 		    lock->lo_class->lc_name, lock->lo_name);
1364 	instance = find_instance(curthread->td_sleeplocks, lock);
1365 	if (instance == NULL)
1366 		panic("%s: lock (%s) %s not locked", __func__,
1367 		    lock->lo_class->lc_name, lock->lo_name);
1368 	*filep = instance->li_file;
1369 	*linep = instance->li_line;
1370 }
1371 
1372 void
1373 witness_restore(struct lock_object *lock, const char *file, int line)
1374 {
1375 	struct lock_instance *instance;
1376 
1377 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1378 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1379 		return;
1380 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1381 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1382 		    lock->lo_class->lc_name, lock->lo_name);
1383 	instance = find_instance(curthread->td_sleeplocks, lock);
1384 	if (instance == NULL)
1385 		panic("%s: lock (%s) %s not locked", __func__,
1386 		    lock->lo_class->lc_name, lock->lo_name);
1387 	lock->lo_witness->w_file = file;
1388 	lock->lo_witness->w_line = line;
1389 	instance->li_file = file;
1390 	instance->li_line = line;
1391 }
1392 
1393 void
1394 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1395 {
1396 #ifdef INVARIANT_SUPPORT
1397 	struct lock_instance *instance;
1398 
1399 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1400 		return;
1401 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1402 		instance = find_instance(curthread->td_sleeplocks, lock);
1403 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1404 		instance = find_instance(PCPU_GET(spinlocks), lock);
1405 	else {
1406 		panic("Lock (%s) %s is not sleep or spin!",
1407 		    lock->lo_class->lc_name, lock->lo_name);
1408 		return;
1409 	}
1410 	switch (flags) {
1411 	case LA_UNLOCKED:
1412 		if (instance != NULL)
1413 			panic("Lock (%s) %s locked @ %s:%d.",
1414 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1415 		break;
1416 	case LA_LOCKED:
1417 	case LA_LOCKED | LA_RECURSED:
1418 	case LA_LOCKED | LA_NOTRECURSED:
1419 	case LA_SLOCKED:
1420 	case LA_SLOCKED | LA_RECURSED:
1421 	case LA_SLOCKED | LA_NOTRECURSED:
1422 	case LA_XLOCKED:
1423 	case LA_XLOCKED | LA_RECURSED:
1424 	case LA_XLOCKED | LA_NOTRECURSED:
1425 		if (instance == NULL) {
1426 			panic("Lock (%s) %s not locked @ %s:%d.",
1427 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1428 			break;
1429 		}
1430 		if ((flags & LA_XLOCKED) != 0 &&
1431 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1432 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1433 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1434 		if ((flags & LA_SLOCKED) != 0 &&
1435 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1436 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1437 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1438 		if ((flags & LA_RECURSED) != 0 &&
1439 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1440 			panic("Lock (%s) %s not recursed @ %s:%d.",
1441 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1442 		if ((flags & LA_NOTRECURSED) != 0 &&
1443 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1444 			panic("Lock (%s) %s recursed @ %s:%d.",
1445 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1446 		break;
1447 	default:
1448 		panic("Invalid lock assertion at %s:%d.", file, line);
1449 
1450 	}
1451 #endif	/* INVARIANT_SUPPORT */
1452 }
1453 
1454 #ifdef DDB
1455 
1456 DB_SHOW_COMMAND(locks, db_witness_list)
1457 {
1458 	struct thread *td;
1459 	pid_t pid;
1460 	struct proc *p;
1461 
1462 	if (have_addr) {
1463 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1464 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1465 		    ((addr >> 16) % 16) * 10000;
1466 		/* sx_slock(&allproc_lock); */
1467 		FOREACH_PROC_IN_SYSTEM(p) {
1468 			if (p->p_pid == pid)
1469 				break;
1470 		}
1471 		/* sx_sunlock(&allproc_lock); */
1472 		if (p == NULL) {
1473 			db_printf("pid %d not found\n", pid);
1474 			return;
1475 		}
1476 		FOREACH_THREAD_IN_PROC(p, td) {
1477 			witness_list(td);
1478 		}
1479 	} else {
1480 		td = curthread;
1481 		witness_list(td);
1482 	}
1483 }
1484 
1485 DB_SHOW_COMMAND(witness, db_witness_display)
1486 {
1487 
1488 	witness_display(db_printf);
1489 }
1490 #endif
1491