xref: /freebsd/sys/kern/subr_witness.c (revision 8fa113e5fc65fe6abc757f0089f477a87ee4d185)
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 	{ "proctree", &lock_class_sx },
192 	{ "allproc", &lock_class_sx },
193 	{ "process lock", &lock_class_mtx_sleep },
194 	{ "uidinfo hash", &lock_class_mtx_sleep },
195 	{ "uidinfo struct", &lock_class_mtx_sleep },
196 	{ NULL, NULL },
197 	/*
198 	 * spin locks
199 	 */
200 #ifdef SMP
201 	{ "ap boot", &lock_class_mtx_spin },
202 #ifdef __i386__
203 	{ "com", &lock_class_mtx_spin },
204 #endif
205 #endif
206 	{ "sio", &lock_class_mtx_spin },
207 #ifdef __i386__
208 	{ "cy", &lock_class_mtx_spin },
209 #endif
210 	{ "ng_node", &lock_class_mtx_spin },
211 	{ "ng_worklist", &lock_class_mtx_spin },
212 	{ "ithread table lock", &lock_class_mtx_spin },
213 	{ "sched lock", &lock_class_mtx_spin },
214 	{ "callout", &lock_class_mtx_spin },
215 	/*
216 	 * leaf locks
217 	 */
218 #ifdef SMP
219 #ifdef __i386__
220 	{ "imen", &lock_class_mtx_spin },
221 #endif
222 	{ "smp rendezvous", &lock_class_mtx_spin },
223 #endif
224 	{ "clk", &lock_class_mtx_spin },
225 	{ NULL, NULL },
226 	{ NULL, NULL }
227 };
228 
229 static const char *dup_list[] = {
230 	"process lock",
231 	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 STAILQ_HEAD(, lock_object) all_locks = STAILQ_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 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
252 	  { NULL },			/* mtx_object.lo_list */
253 	  NULL },			/* mtx_object.lo_witness */
254 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
255 	0,				/* mtx_savecrit */
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 	CTR0(KTR_WITNESS, __func__ ": initializing witness");
290 	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
291 	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
292 	for (i = 0; i < WITNESS_COUNT; i++)
293 		witness_free(&w_data[i]);
294 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
295 		witness_child_free(&w_childdata[i]);
296 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
297 		witness_lock_list_free(&w_locklistdata[i]);
298 
299 	/* First add in all the specified order lists. */
300 	for (order = order_lists; order->w_name != NULL; order++) {
301 		w = enroll(order->w_name, order->w_class);
302 		if (w == NULL)
303 			continue;
304 		w->w_file = "order list";
305 		for (order++; order->w_name != NULL; order++) {
306 			w1 = enroll(order->w_name, order->w_class);
307 			if (w1 == NULL)
308 				continue;
309 			w1->w_file = "order list";
310 			itismychild(w, w1);
311 			w = w1;
312 		}
313 	}
314 
315 	/* Iterate through all locks and add them to witness. */
316 	mtx_lock(&all_mtx);
317 	STAILQ_FOREACH(lock, &all_locks, lo_list) {
318 		if (lock->lo_flags & LO_WITNESS)
319 			lock->lo_witness = enroll(lock->lo_name,
320 			    lock->lo_class);
321 		else
322 			lock->lo_witness = NULL;
323 	}
324 	mtx_unlock(&all_mtx);
325 
326 	/* Mark the witness code as being ready for use. */
327 	atomic_store_rel_int(&witness_cold, 0);
328 
329 	mtx_lock(&Giant);
330 }
331 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
332 
333 void
334 witness_init(struct lock_object *lock)
335 {
336 	struct lock_class *class;
337 
338 	class = lock->lo_class;
339 	if (lock->lo_flags & LO_INITIALIZED)
340 		panic("%s: lock (%s) %s is already initialized", __func__,
341 		    class->lc_name, lock->lo_name);
342 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
343 	    (class->lc_flags & LC_RECURSABLE) == 0)
344 		panic("%s: lock (%s) %s can not be recursable", __func__,
345 		    class->lc_name, lock->lo_name);
346 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
347 	    (class->lc_flags & LC_SLEEPABLE) == 0)
348 		panic("%s: lock (%s) %s can not be sleepable", __func__,
349 		    class->lc_name, lock->lo_name);
350 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
351 	    (class->lc_flags & LC_UPGRADABLE) == 0)
352 		panic("%s: lock (%s) %s can not be upgradable", __func__,
353 		    class->lc_name, lock->lo_name);
354 
355 	mtx_lock(&all_mtx);
356 	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
357 	lock->lo_flags |= LO_INITIALIZED;
358 	lock_cur_cnt++;
359 	if (lock_cur_cnt > lock_max_cnt)
360 		lock_max_cnt = lock_cur_cnt;
361 	mtx_unlock(&all_mtx);
362 	if (!witness_cold && !witness_dead && panicstr == NULL &&
363 	    (lock->lo_flags & LO_WITNESS) != 0)
364 		lock->lo_witness = enroll(lock->lo_name, class);
365 	else
366 		lock->lo_witness = NULL;
367 }
368 
369 void
370 witness_destroy(struct lock_object *lock)
371 {
372 	struct witness *w;
373 
374 	if (witness_cold)
375 		panic("lock (%s) %s destroyed while witness_cold",
376 		    lock->lo_class->lc_name, lock->lo_name);
377 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
378 		panic("%s: lock (%s) %s is not initialized", __func__,
379 		    lock->lo_class->lc_name, lock->lo_name);
380 
381 	/* XXX: need to verify that no one holds the lock */
382 	w = lock->lo_witness;
383 	if (w != NULL) {
384 		mtx_lock_spin(&w_mtx);
385 		w->w_refcount--;
386 		if (w->w_refcount == 0) {
387 			CTR1(KTR_WITNESS,
388 			    __func__ ": marking witness %s as dead", w->w_name);
389 			w->w_name = "(dead)";
390 			w->w_file = "(dead)";
391 			w->w_line = 0;
392 		}
393 		mtx_unlock_spin(&w_mtx);
394 	}
395 
396 	mtx_lock(&all_mtx);
397 	lock_cur_cnt--;
398 	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
399 	lock->lo_flags &= ~LO_INITIALIZED;
400 	mtx_unlock(&all_mtx);
401 }
402 
403 static void
404 witness_display_list(void(*prnt)(const char *fmt, ...),
405 		     struct witness_list *list)
406 {
407 	struct witness *w, *w1;
408 	int found;
409 
410 	STAILQ_FOREACH(w, list, w_typelist) {
411 		if (w->w_file == NULL)
412 			continue;
413 		found = 0;
414 		STAILQ_FOREACH(w1, list, w_typelist) {
415 			if (isitmychild(w1, w)) {
416 				found++;
417 				break;
418 			}
419 		}
420 		if (found)
421 			continue;
422 		/*
423 		 * This lock has no anscestors, display its descendants.
424 		 */
425 		witness_displaydescendants(prnt, w);
426 	}
427 }
428 
429 static void
430 witness_display(void(*prnt)(const char *fmt, ...))
431 {
432 	struct witness *w;
433 
434 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
435 	witness_levelall();
436 
437 	/*
438 	 * First, handle sleep locks which have been acquired at least
439 	 * once.
440 	 */
441 	prnt("Sleep locks:\n");
442 	witness_display_list(prnt, &w_sleep);
443 
444 	/*
445 	 * Now do spin locks which have been acquired at least once.
446 	 */
447 	prnt("\nSpin locks:\n");
448 	witness_display_list(prnt, &w_spin);
449 
450 	/*
451 	 * Finally, any locks which have not been acquired yet.
452 	 */
453 	prnt("\nLocks which were never acquired:\n");
454 	STAILQ_FOREACH(w, &w_all, w_list) {
455 		if (w->w_file != NULL)
456 			continue;
457 		prnt("%s\n", w->w_name);
458 	}
459 }
460 
461 void
462 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
463 {
464 	struct lock_list_entry **lock_list, *lle;
465 	struct lock_instance *lock1, *lock2;
466 	struct lock_class *class;
467 	struct witness *w, *w1;
468 	struct thread *td;
469 	int i, j;
470 #ifdef DDB
471 	int go_into_ddb = 0;
472 #endif /* DDB */
473 
474 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
475 	    panicstr != NULL)
476 		return;
477 	w = lock->lo_witness;
478 	class = lock->lo_class;
479 	td = curthread;
480 
481 	/*
482 	 * We have to hold a spinlock to keep lock_list valid across the check
483 	 * in the LC_SLEEPLOCK case.  In the LC_SPINLOCK case, it is already
484 	 * protected by the spinlock we are currently performing the witness
485 	 * checks on, so it is ok to release the lock after performing this
486 	 * check.  All we have to protect is the LC_SLEEPLOCK case when no
487 	 * spinlocks are held as we may get preempted during this check and
488 	 * lock_list could end up pointing to some other CPU's spinlock list.
489 	 */
490 	mtx_lock_spin(&w_mtx);
491 	lock_list = PCPU_PTR(spinlocks);
492 	if (class->lc_flags & LC_SLEEPLOCK) {
493 		if (*lock_list != NULL && (flags & LOP_TRYLOCK) == 0) {
494 			mtx_unlock_spin(&w_mtx);
495 			panic("blockable sleep lock (%s) %s @ %s:%d",
496 			    class->lc_name, lock->lo_name, file, line);
497 		}
498 		lock_list = &td->td_sleeplocks;
499 	}
500 	mtx_unlock_spin(&w_mtx);
501 
502 	/*
503 	 * Try locks do not block if they fail to acquire the lock, thus
504 	 * there is no danger of deadlocks or of switching while holding a
505 	 * spin lock if we acquire a lock via a try operation.
506 	 */
507 	if (flags & LOP_TRYLOCK)
508 		goto out;
509 
510 	/*
511 	 * Is this the first lock acquired?  If so, then no order checking
512 	 * is needed.
513 	 */
514 	if (*lock_list == NULL)
515 		goto out;
516 
517 	/*
518 	 * Check to see if we are recursing on a lock we already own.
519 	 */
520 	lock1 = find_instance(*lock_list, lock);
521 	if (lock1 != NULL) {
522 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
523 		    (flags & LOP_EXCLUSIVE) == 0) {
524 			printf("shared lock of (%s) %s @ %s:%d\n",
525 			    class->lc_name, lock->lo_name, file, line);
526 			printf("while exclusively locked from %s:%d\n",
527 			    lock1->li_file, lock1->li_line);
528 			panic("share->excl");
529 		}
530 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
531 		    (flags & LOP_EXCLUSIVE) != 0) {
532 			printf("exclusive lock of (%s) %s @ %s:%d\n",
533 			    class->lc_name, lock->lo_name, file, line);
534 			printf("while share locked from %s:%d\n",
535 			    lock1->li_file, lock1->li_line);
536 			panic("excl->share");
537 		}
538 		lock1->li_flags++;
539 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
540 			printf(
541 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
542 			    class->lc_name, lock->lo_name, file, line);
543 			printf("first acquired @ %s:%d\n", lock1->li_file,
544 			    lock1->li_line);
545 			panic("recurse");
546 		}
547 		CTR3(KTR_WITNESS, __func__ ": pid %d recursed on %s r=%d",
548 		    td->td_proc->p_pid, lock->lo_name,
549 		    lock1->li_flags & LI_RECURSEMASK);
550 		lock1->li_file = file;
551 		lock1->li_line = line;
552 		return;
553 	}
554 
555 	/*
556 	 * Check for duplicate locks of the same type.  Note that we only
557 	 * have to check for this on the last lock we just acquired.  Any
558 	 * other cases will be caught as lock order violations.
559 	 */
560 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
561 	w1 = lock1->li_lock->lo_witness;
562 	if (w1 == w) {
563 		if (w->w_same_squawked || dup_ok(w))
564 			goto out;
565 		w->w_same_squawked = 1;
566 		printf("acquiring duplicate lock of same type: \"%s\"\n",
567 			lock->lo_name);
568 		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
569 		printf(" 2nd @ %s:%d\n", file, line);
570 #ifdef DDB
571 		go_into_ddb = 1;
572 #endif /* DDB */
573 		goto out;
574 	}
575 	MPASS(!mtx_owned(&w_mtx));
576 	mtx_lock_spin(&w_mtx);
577 	/*
578 	 * If we have a known higher number just say ok
579 	 */
580 	if (witness_watch > 1 && w->w_level > w1->w_level) {
581 		mtx_unlock_spin(&w_mtx);
582 		goto out;
583 	}
584 	if (isitmydescendant(w1, w)) {
585 		mtx_unlock_spin(&w_mtx);
586 		goto out;
587 	}
588 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
589 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
590 
591 			MPASS(j < WITNESS_COUNT);
592 			lock1 = &lle->ll_children[i];
593 			w1 = lock1->li_lock->lo_witness;
594 
595 			/*
596 			 * If this lock doesn't undergo witness checking,
597 			 * then skip it.
598 			 */
599 			if (w1 == NULL) {
600 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
601 				    ("lock missing witness structure"));
602 				continue;
603 			}
604 			/*
605 			 * If we are locking Giant and we slept with this
606 			 * lock, then skip it.
607 			 */
608 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
609 			    lock == &Giant.mtx_object)
610 				continue;
611 			if (!isitmydescendant(w, w1))
612 				continue;
613 			/*
614 			 * We have a lock order violation, check to see if it
615 			 * is allowed or has already been yelled about.
616 			 */
617 			mtx_unlock_spin(&w_mtx);
618 			if (blessed(w, w1))
619 				goto out;
620 			if (lock1->li_lock == &Giant.mtx_object) {
621 				if (w1->w_Giant_squawked)
622 					goto out;
623 				else
624 					w1->w_Giant_squawked = 1;
625 			} else {
626 				if (w1->w_other_squawked)
627 					goto out;
628 				else
629 					w1->w_other_squawked = 1;
630 			}
631 			/*
632 			 * Ok, yell about it.
633 			 */
634 			printf("lock order reversal\n");
635 			/*
636 			 * Try to locate an earlier lock with
637 			 * witness w in our list.
638 			 */
639 			do {
640 				lock2 = &lle->ll_children[i];
641 				MPASS(lock2->li_lock != NULL);
642 				if (lock2->li_lock->lo_witness == w)
643 					break;
644 				i--;
645 				if (i == 0 && lle->ll_next != NULL) {
646 					lle = lle->ll_next;
647 					i = lle->ll_count - 1;
648 					MPASS(i != 0);
649 				}
650 			} while (i >= 0);
651 			if (i < 0) {
652 				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
653 				    lock1->li_lock->lo_name, lock1->li_file,
654 				    lock1->li_line);
655 				printf(" 2nd %p %s @ %s:%d\n", lock,
656 				    lock->lo_name, file, line);
657 			} else {
658 				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
659 				    lock2->li_lock->lo_name, lock2->li_file,
660 				    lock2->li_line);
661 				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
662 				    lock1->li_lock->lo_name, lock1->li_file,
663 				    lock1->li_line);
664 				printf(" 3rd %p %s @ %s:%d\n", lock,
665 				    lock->lo_name, file, line);
666 			}
667 #ifdef DDB
668 			go_into_ddb = 1;
669 #endif /* DDB */
670 			goto out;
671 		}
672 	}
673 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
674 	/*
675 	 * Don't build a new relationship if we are locking Giant just
676 	 * after waking up and the previous lock in the list was acquired
677 	 * prior to blocking.
678 	 */
679 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
680 		mtx_unlock_spin(&w_mtx);
681 	else {
682 		CTR2(KTR_WITNESS, __func__ ": adding %s as a child of %s",
683 		    lock->lo_name, lock1->li_lock->lo_name);
684 		if (!itismychild(lock1->li_lock->lo_witness, w))
685 			mtx_unlock_spin(&w_mtx);
686 	}
687 
688 out:
689 #ifdef DDB
690 	if (witness_ddb && go_into_ddb)
691 		Debugger(__func__);
692 #endif /* DDB */
693 	w->w_file = file;
694 	w->w_line = line;
695 
696 	lle = *lock_list;
697 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
698 		lle = witness_lock_list_get();
699 		if (lle == NULL)
700 			return;
701 		lle->ll_next = *lock_list;
702 		CTR2(KTR_WITNESS, __func__ ": pid %d added lle %p",
703 		    td->td_proc->p_pid, lle);
704 		*lock_list = lle;
705 	}
706 	lock1 = &lle->ll_children[lle->ll_count++];
707 	lock1->li_lock = lock;
708 	lock1->li_line = line;
709 	lock1->li_file = file;
710 	if ((flags & LOP_EXCLUSIVE) != 0)
711 		lock1->li_flags = LI_EXCLUSIVE;
712 	else
713 		lock1->li_flags = 0;
714 	CTR3(KTR_WITNESS, __func__ ": pid %d added %s as lle[%d]",
715 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
716 }
717 
718 void
719 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
720 {
721 	struct lock_instance *instance;
722 	struct lock_class *class;
723 
724 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
725 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
726 		return;
727 	class = lock->lo_class;
728 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
729 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
730 		    class->lc_name, lock->lo_name, file, line);
731 	if ((flags & LOP_TRYLOCK) == 0)
732 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
733 		    lock->lo_name, file, line);
734 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
735 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
736 		    class->lc_name, lock->lo_name, file, line);
737 	instance = find_instance(curthread->td_sleeplocks, lock);
738 	if (instance == NULL)
739 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
740 		    class->lc_name, lock->lo_name, file, line);
741 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
742 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
743 		    class->lc_name, lock->lo_name, file, line);
744 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
745 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
746 		    class->lc_name, lock->lo_name,
747 		    instance->li_flags & LI_RECURSEMASK, file, line);
748 	instance->li_flags |= LI_EXCLUSIVE;
749 }
750 
751 void
752 witness_downgrade(struct lock_object *lock, int flags, const char *file,
753     int line)
754 {
755 	struct lock_instance *instance;
756 	struct lock_class *class;
757 
758 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
759 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
760 		return;
761 	class = lock->lo_class;
762 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
763 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
764 		    class->lc_name, lock->lo_name, file, line);
765 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
766 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
767 		    class->lc_name, lock->lo_name, file, line);
768 	instance = find_instance(curthread->td_sleeplocks, lock);
769 	if (instance == NULL)
770 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
771 		    class->lc_name, lock->lo_name, file, line);
772 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
773 		panic("downgrade of shared lock (%s) %s @ %s:%d",
774 		    class->lc_name, lock->lo_name, file, line);
775 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
776 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
777 		    class->lc_name, lock->lo_name,
778 		    instance->li_flags & LI_RECURSEMASK, file, line);
779 	instance->li_flags &= ~LI_EXCLUSIVE;
780 }
781 
782 void
783 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
784 {
785 	struct lock_list_entry **lock_list, *lle;
786 	struct lock_instance *instance;
787 	struct lock_class *class;
788 	struct thread *td;
789 	critical_t s;
790 	int i, j;
791 
792 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
793 	    panicstr != NULL)
794 		return;
795 	td = curthread;
796 	class = lock->lo_class;
797 	if (class->lc_flags & LC_SLEEPLOCK)
798 		lock_list = &td->td_sleeplocks;
799 	else
800 		lock_list = PCPU_PTR(spinlocks);
801 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
802 		for (i = 0; i < (*lock_list)->ll_count; i++) {
803 			instance = &(*lock_list)->ll_children[i];
804 			if (instance->li_lock == lock) {
805 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
806 				    (flags & LOP_EXCLUSIVE) == 0) {
807 					printf(
808 					"shared unlock of (%s) %s @ %s:%d\n",
809 					    class->lc_name, lock->lo_name,
810 					    file, line);
811 					printf(
812 					"while exclusively locked from %s:%d\n",
813 					    instance->li_file,
814 					    instance->li_line);
815 					panic("excl->ushare");
816 				}
817 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
818 				    (flags & LOP_EXCLUSIVE) != 0) {
819 					printf(
820 					"exclusive unlock of (%s) %s @ %s:%d\n",
821 					    class->lc_name, lock->lo_name,
822 					    file, line);
823 					printf(
824 					"while share locked from %s:%d\n",
825 					    instance->li_file,
826 					    instance->li_line);
827 					panic("share->uexcl");
828 				}
829 				/* If we are recursed, unrecurse. */
830 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
831 					CTR3(KTR_WITNESS,
832 				    __func__ ": pid %d unrecursed on %s r=%d",
833 					    td->td_proc->p_pid,
834 					    instance->li_lock->lo_name,
835 					    instance->li_flags);
836 					instance->li_flags--;
837 					goto out;
838 				}
839 				s = critical_enter();
840 				CTR3(KTR_WITNESS,
841 				    __func__ ": pid %d removed %s from lle[%d]",
842 				    td->td_proc->p_pid,
843 				    instance->li_lock->lo_name,
844 				    (*lock_list)->ll_count - 1);
845 				(*lock_list)->ll_count--;
846 				for (j = i; j < (*lock_list)->ll_count; j++)
847 					(*lock_list)->ll_children[j] =
848 					    (*lock_list)->ll_children[j + 1];
849 				critical_exit(s);
850 				if ((*lock_list)->ll_count == 0) {
851 					lle = *lock_list;
852 					*lock_list = lle->ll_next;
853 					CTR2(KTR_WITNESS,
854 					    __func__ ": pid %d removed lle %p",
855 					    td->td_proc->p_pid, lle);
856 					witness_lock_list_free(lle);
857 				}
858 				goto out;
859 			}
860 		}
861 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
862 	    file, line);
863 out:
864 	/*
865 	 * We don't need to protect this PCPU_GET() here against preemption
866 	 * because if we hold any spinlocks then we are already protected,
867 	 * and if we don't we will get NULL if we hold no spinlocks even if
868 	 * we switch CPU's while reading it.
869 	 */
870 	if (class->lc_flags & LC_SLEEPLOCK) {
871 		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
872 			panic("switchable sleep unlock (%s) %s @ %s:%d",
873 			    class->lc_name, lock->lo_name, file, line);
874 	}
875 }
876 
877 /*
878  * Warn if any held locks are not sleepable.  Note that Giant and the lock
879  * passed in are both special cases since they are both released during the
880  * sleep process and aren't actually held while the thread is asleep.
881  */
882 int
883 witness_sleep(int check_only, struct lock_object *lock, const char *file,
884 	      int line)
885 {
886 	struct lock_list_entry **lock_list, *lle;
887 	struct lock_instance *lock1;
888 	struct thread *td;
889 	critical_t savecrit;
890 	int i, n;
891 
892 	if (witness_dead || panicstr != NULL)
893 		return (0);
894 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
895 	n = 0;
896 	/*
897 	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
898 	 */
899 	savecrit = critical_enter();
900 	td = curthread;
901 	lock_list = &td->td_sleeplocks;
902 again:
903 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
904 		for (i = lle->ll_count - 1; i >= 0; i--) {
905 			lock1 = &lle->ll_children[i];
906 			if (lock1->li_lock == lock ||
907 			    lock1->li_lock == &Giant.mtx_object)
908 				continue;
909 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
910 				if (check_only == 0) {
911 					CTR3(KTR_WITNESS,
912 				    "pid %d: sleeping with lock (%s) %s held",
913 					    td->td_proc->p_pid,
914 					    lock1->li_lock->lo_class->lc_name,
915 					    lock1->li_lock->lo_name);
916 					lock1->li_flags |= LI_SLEPT;
917 				}
918 				continue;
919 			}
920 			n++;
921 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
922 			    file, line, check_only ? "could sleep" : "sleeping",
923 			    lock1->li_lock->lo_name, lock1->li_file,
924 			    lock1->li_line);
925 		}
926 	if (lock_list == &td->td_sleeplocks) {
927 		lock_list = PCPU_PTR(spinlocks);
928 		goto again;
929 	}
930 #ifdef DDB
931 	if (witness_ddb && n)
932 		Debugger(__func__);
933 #endif /* DDB */
934 	critical_exit(savecrit);
935 	return (n);
936 }
937 
938 static struct witness *
939 enroll(const char *description, struct lock_class *lock_class)
940 {
941 	struct witness *w;
942 
943 	if (!witness_watch || witness_dead || panicstr != NULL)
944 		return (NULL);
945 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
946 		return (NULL);
947 	mtx_lock_spin(&w_mtx);
948 	STAILQ_FOREACH(w, &w_all, w_list) {
949 		if (strcmp(description, w->w_name) == 0) {
950 			w->w_refcount++;
951 			mtx_unlock_spin(&w_mtx);
952 			if (lock_class != w->w_class)
953 				panic(
954 				"lock (%s) %s does not match earlier (%s) lock",
955 				    description, lock_class->lc_name,
956 				    w->w_class->lc_name);
957 			return (w);
958 		}
959 	}
960 	/*
961 	 * This isn't quite right, as witness_cold is still 0 while we
962 	 * enroll all the locks initialized before witness_initialize().
963 	 */
964 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
965 		mtx_unlock_spin(&w_mtx);
966 		panic("spin lock %s not in order list", description);
967 	}
968 	if ((w = witness_get()) == NULL)
969 		return (NULL);
970 	w->w_name = description;
971 	w->w_class = lock_class;
972 	w->w_refcount = 1;
973 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
974 	if (lock_class->lc_flags & LC_SPINLOCK)
975 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
976 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
977 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
978 	else {
979 		mtx_unlock_spin(&w_mtx);
980 		panic("lock class %s is not sleep or spin",
981 		    lock_class->lc_name);
982 	}
983 	mtx_unlock_spin(&w_mtx);
984 	return (w);
985 }
986 
987 static int
988 itismychild(struct witness *parent, struct witness *child)
989 {
990 	static int recursed;
991 	struct witness_child_list_entry **wcl;
992 	struct witness_list *list;
993 
994 	MPASS(child != NULL && parent != NULL);
995 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
996 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
997 		panic(
998 		"%s: parent (%s) and child (%s) are not the same lock type",
999 		    __func__, parent->w_class->lc_name,
1000 		    child->w_class->lc_name);
1001 
1002 	/*
1003 	 * Insert "child" after "parent"
1004 	 */
1005 	wcl = &parent->w_children;
1006 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1007 		wcl = &(*wcl)->wcl_next;
1008 	if (*wcl == NULL) {
1009 		*wcl = witness_child_get();
1010 		if (*wcl == NULL)
1011 			return (1);
1012 	}
1013 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1014 
1015 	/*
1016 	 * Now prune whole tree.  We look for cases where a lock is now
1017 	 * both a descendant and a direct child of a given lock.  In that
1018 	 * case, we want to remove the direct child link from the tree.
1019 	 */
1020 	if (recursed)
1021 		return (0);
1022 	recursed = 1;
1023 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1024 		list = &w_sleep;
1025 	else
1026 		list = &w_spin;
1027 	STAILQ_FOREACH(child, list, w_typelist) {
1028 		STAILQ_FOREACH(parent, list, w_typelist) {
1029 			if (!isitmychild(parent, child))
1030 				continue;
1031 			removechild(parent, child);
1032 			if (isitmydescendant(parent, child))
1033 				continue;
1034 			itismychild(parent, child);
1035 		}
1036 	}
1037 	recursed = 0;
1038 	witness_levelall();
1039 	return (0);
1040 }
1041 
1042 static void
1043 removechild(struct witness *parent, struct witness *child)
1044 {
1045 	struct witness_child_list_entry **wcl, *wcl1;
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 				goto found;
1052 	return;
1053 found:
1054 	(*wcl)->wcl_count--;
1055 	if ((*wcl)->wcl_count > i)
1056 		(*wcl)->wcl_children[i] =
1057 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1058 	MPASS((*wcl)->wcl_children[i] != NULL);
1059 	if ((*wcl)->wcl_count != 0)
1060 		return;
1061 	wcl1 = *wcl;
1062 	*wcl = wcl1->wcl_next;
1063 	witness_child_free(wcl1);
1064 }
1065 
1066 static int
1067 isitmychild(struct witness *parent, struct witness *child)
1068 {
1069 	struct witness_child_list_entry *wcl;
1070 	int i;
1071 
1072 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1073 		for (i = 0; i < wcl->wcl_count; i++) {
1074 			if (wcl->wcl_children[i] == child)
1075 				return (1);
1076 		}
1077 	}
1078 	return (0);
1079 }
1080 
1081 static int
1082 isitmydescendant(struct witness *parent, struct witness *child)
1083 {
1084 	struct witness_child_list_entry *wcl;
1085 	int i, j;
1086 
1087 	if (isitmychild(parent, child))
1088 		return (1);
1089 	j = 0;
1090 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1091 		MPASS(j < 1000);
1092 		for (i = 0; i < wcl->wcl_count; i++) {
1093 			if (isitmydescendant(wcl->wcl_children[i], child))
1094 				return (1);
1095 		}
1096 		j++;
1097 	}
1098 	return (0);
1099 }
1100 
1101 void
1102 witness_levelall (void)
1103 {
1104 	struct witness_list *list;
1105 	struct witness *w, *w1;
1106 
1107 	/*
1108 	 * First clear all levels.
1109 	 */
1110 	STAILQ_FOREACH(w, &w_all, w_list) {
1111 		w->w_level = 0;
1112 	}
1113 
1114 	/*
1115 	 * Look for locks with no parent and level all their descendants.
1116 	 */
1117 	STAILQ_FOREACH(w, &w_all, w_list) {
1118 		/*
1119 		 * This is just an optimization, technically we could get
1120 		 * away just walking the all list each time.
1121 		 */
1122 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1123 			list = &w_sleep;
1124 		else
1125 			list = &w_spin;
1126 		STAILQ_FOREACH(w1, list, w_typelist) {
1127 			if (isitmychild(w1, w))
1128 				goto skip;
1129 		}
1130 		witness_leveldescendents(w, 0);
1131 	skip:
1132 	}
1133 }
1134 
1135 static void
1136 witness_leveldescendents(struct witness *parent, int level)
1137 {
1138 	struct witness_child_list_entry *wcl;
1139 	int i;
1140 
1141 	if (parent->w_level < level)
1142 		parent->w_level = level;
1143 	level++;
1144 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1145 		for (i = 0; i < wcl->wcl_count; i++)
1146 			witness_leveldescendents(wcl->wcl_children[i], level);
1147 }
1148 
1149 static void
1150 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1151 			   struct witness *parent)
1152 {
1153 	struct witness_child_list_entry *wcl;
1154 	int i, level;
1155 
1156 	level =  parent->w_level;
1157 	prnt("%-2d", level);
1158 	for (i = 0; i < level; i++)
1159 		prnt(" ");
1160 	prnt("%s", parent->w_name);
1161 	if (parent->w_file != NULL)
1162 		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1163 		    parent->w_line);
1164 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1165 		for (i = 0; i < wcl->wcl_count; i++)
1166 			    witness_displaydescendants(prnt,
1167 				wcl->wcl_children[i]);
1168 }
1169 
1170 static int
1171 dup_ok(struct witness *w)
1172 {
1173 	const char **dup;
1174 
1175 	for (dup = dup_list; *dup != NULL; dup++)
1176 		if (strcmp(w->w_name, *dup) == 0)
1177 			return (1);
1178 	return (0);
1179 }
1180 
1181 static int
1182 blessed(struct witness *w1, struct witness *w2)
1183 {
1184 	int i;
1185 	struct witness_blessed *b;
1186 
1187 	for (i = 0; i < blessed_count; i++) {
1188 		b = &blessed_list[i];
1189 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1190 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1191 				return (1);
1192 			continue;
1193 		}
1194 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1195 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1196 				return (1);
1197 	}
1198 	return (0);
1199 }
1200 
1201 static struct witness *
1202 witness_get(void)
1203 {
1204 	struct witness *w;
1205 
1206 	if (witness_dead) {
1207 		mtx_unlock_spin(&w_mtx);
1208 		return (NULL);
1209 	}
1210 	if (STAILQ_EMPTY(&w_free)) {
1211 		witness_dead = 1;
1212 		mtx_unlock_spin(&w_mtx);
1213 		printf("%s: witness exhausted\n", __func__);
1214 		return (NULL);
1215 	}
1216 	w = STAILQ_FIRST(&w_free);
1217 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1218 	bzero(w, sizeof(*w));
1219 	return (w);
1220 }
1221 
1222 static void
1223 witness_free(struct witness *w)
1224 {
1225 
1226 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1227 }
1228 
1229 static struct witness_child_list_entry *
1230 witness_child_get(void)
1231 {
1232 	struct witness_child_list_entry *wcl;
1233 
1234 	if (witness_dead) {
1235 		mtx_unlock_spin(&w_mtx);
1236 		return (NULL);
1237 	}
1238 	wcl = w_child_free;
1239 	if (wcl == NULL) {
1240 		witness_dead = 1;
1241 		mtx_unlock_spin(&w_mtx);
1242 		printf("%s: witness exhausted\n", __func__);
1243 		return (NULL);
1244 	}
1245 	w_child_free = wcl->wcl_next;
1246 	bzero(wcl, sizeof(*wcl));
1247 	return (wcl);
1248 }
1249 
1250 static void
1251 witness_child_free(struct witness_child_list_entry *wcl)
1252 {
1253 
1254 	wcl->wcl_next = w_child_free;
1255 	w_child_free = wcl;
1256 }
1257 
1258 static struct lock_list_entry *
1259 witness_lock_list_get(void)
1260 {
1261 	struct lock_list_entry *lle;
1262 
1263 	if (witness_dead)
1264 		return (NULL);
1265 	mtx_lock_spin(&w_mtx);
1266 	lle = w_lock_list_free;
1267 	if (lle == NULL) {
1268 		witness_dead = 1;
1269 		mtx_unlock_spin(&w_mtx);
1270 		printf("%s: witness exhausted\n", __func__);
1271 		return (NULL);
1272 	}
1273 	w_lock_list_free = lle->ll_next;
1274 	mtx_unlock_spin(&w_mtx);
1275 	bzero(lle, sizeof(*lle));
1276 	return (lle);
1277 }
1278 
1279 static void
1280 witness_lock_list_free(struct lock_list_entry *lle)
1281 {
1282 
1283 	mtx_lock_spin(&w_mtx);
1284 	lle->ll_next = w_lock_list_free;
1285 	w_lock_list_free = lle;
1286 	mtx_unlock_spin(&w_mtx);
1287 }
1288 
1289 static struct lock_instance *
1290 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1291 {
1292 	struct lock_list_entry *lle;
1293 	struct lock_instance *instance;
1294 	int i;
1295 
1296 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1297 		for (i = lle->ll_count - 1; i >= 0; i--) {
1298 			instance = &lle->ll_children[i];
1299 			if (instance->li_lock == lock)
1300 				return (instance);
1301 		}
1302 	return (NULL);
1303 }
1304 
1305 int
1306 witness_list_locks(struct lock_list_entry **lock_list)
1307 {
1308 	struct lock_list_entry *lle;
1309 	struct lock_instance *instance;
1310 	struct lock_object *lock;
1311 	int i, nheld;
1312 
1313 	nheld = 0;
1314 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1315 		for (i = lle->ll_count - 1; i >= 0; i--) {
1316 			instance = &lle->ll_children[i];
1317 			lock = instance->li_lock;
1318 			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1319 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1320 			    "exclusive" : "shared",
1321 			    lock->lo_class->lc_name, lock->lo_name, lock,
1322 			    instance->li_file, instance->li_line);
1323 			nheld++;
1324 		}
1325 	return (nheld);
1326 }
1327 
1328 /*
1329  * Calling this on td != curthread is bad unless we are in ddb.
1330  */
1331 int
1332 witness_list(struct thread *td)
1333 {
1334 	critical_t savecrit;
1335 	int nheld;
1336 
1337 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1338 #ifdef DDB
1339 	KASSERT(td == curthread || db_active,
1340 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1341 	if (!db_active && witness_dead)
1342 		return (0);
1343 #else
1344 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1345 	if (witness_dead)
1346 		return (0);
1347 #endif
1348 	nheld = witness_list_locks(&td->td_sleeplocks);
1349 
1350 	/*
1351 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1352 	 * if td is currently executing on some other CPU and holds spin locks
1353 	 * as we won't display those locks.  If we had a MI way of getting
1354 	 * the per-cpu data for a given cpu then we could use
1355 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1356 	 * and "fix" this.
1357 	 */
1358 	if (td == curthread) {
1359 		/*
1360 		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1361 		 * change.
1362 		 */
1363 		savecrit = critical_enter();
1364 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1365 		critical_exit(savecrit);
1366 	}
1367 	return (nheld);
1368 }
1369 
1370 void
1371 witness_save(struct lock_object *lock, const char **filep, int *linep)
1372 {
1373 	struct lock_instance *instance;
1374 
1375 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1376 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1377 		return;
1378 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1379 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1380 		    lock->lo_class->lc_name, lock->lo_name);
1381 	instance = find_instance(curthread->td_sleeplocks, lock);
1382 	if (instance == NULL)
1383 		panic("%s: lock (%s) %s not locked", __func__,
1384 		    lock->lo_class->lc_name, lock->lo_name);
1385 	*filep = instance->li_file;
1386 	*linep = instance->li_line;
1387 }
1388 
1389 void
1390 witness_restore(struct lock_object *lock, const char *file, int line)
1391 {
1392 	struct lock_instance *instance;
1393 
1394 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1395 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1396 		return;
1397 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1398 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1399 		    lock->lo_class->lc_name, lock->lo_name);
1400 	instance = find_instance(curthread->td_sleeplocks, lock);
1401 	if (instance == NULL)
1402 		panic("%s: lock (%s) %s not locked", __func__,
1403 		    lock->lo_class->lc_name, lock->lo_name);
1404 	lock->lo_witness->w_file = file;
1405 	lock->lo_witness->w_line = line;
1406 	instance->li_file = file;
1407 	instance->li_line = line;
1408 }
1409 
1410 void
1411 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1412 {
1413 #ifdef INVARIANT_SUPPORT
1414 	struct lock_instance *instance;
1415 
1416 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1417 		return;
1418 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1419 		instance = find_instance(curthread->td_sleeplocks, lock);
1420 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1421 		instance = find_instance(PCPU_GET(spinlocks), lock);
1422 	else {
1423 		panic("Lock (%s) %s is not sleep or spin!",
1424 		    lock->lo_class->lc_name, lock->lo_name);
1425 		return;
1426 	}
1427 	switch (flags) {
1428 	case LA_UNLOCKED:
1429 		if (instance != NULL)
1430 			panic("Lock (%s) %s locked @ %s:%d.",
1431 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1432 		break;
1433 	case LA_LOCKED:
1434 	case LA_LOCKED | LA_RECURSED:
1435 	case LA_LOCKED | LA_NOTRECURSED:
1436 	case LA_SLOCKED:
1437 	case LA_SLOCKED | LA_RECURSED:
1438 	case LA_SLOCKED | LA_NOTRECURSED:
1439 	case LA_XLOCKED:
1440 	case LA_XLOCKED | LA_RECURSED:
1441 	case LA_XLOCKED | LA_NOTRECURSED:
1442 		if (instance == NULL) {
1443 			panic("Lock (%s) %s not locked @ %s:%d.",
1444 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1445 			break;
1446 		}
1447 		if ((flags & LA_XLOCKED) != 0 &&
1448 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1449 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1450 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1451 		if ((flags & LA_SLOCKED) != 0 &&
1452 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1453 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1454 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1455 		if ((flags & LA_RECURSED) != 0 &&
1456 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1457 			panic("Lock (%s) %s not recursed @ %s:%d.",
1458 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1459 		if ((flags & LA_NOTRECURSED) != 0 &&
1460 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1461 			panic("Lock (%s) %s recursed @ %s:%d.",
1462 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1463 		break;
1464 	default:
1465 		panic("Invalid lock assertion at %s:%d.", file, line);
1466 
1467 	}
1468 #endif	/* INVARIANT_SUPPORT */
1469 }
1470 
1471 #ifdef DDB
1472 
1473 DB_SHOW_COMMAND(locks, db_witness_list)
1474 {
1475 	struct thread *td;
1476 	pid_t pid;
1477 	struct proc *p;
1478 
1479 	if (have_addr) {
1480 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1481 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1482 		    ((addr >> 16) % 16) * 10000;
1483 		/* sx_slock(&allproc_lock); */
1484 		FOREACH_PROC_IN_SYSTEM(p) {
1485 			if (p->p_pid == pid)
1486 				break;
1487 		}
1488 		/* sx_sunlock(&allproc_lock); */
1489 		if (p == NULL) {
1490 			db_printf("pid %d not found\n", pid);
1491 			return;
1492 		}
1493 		td = &p->p_thread; /* XXXKSE */
1494 	} else {
1495 		td = curthread;
1496 	}
1497 	witness_list(td);
1498 }
1499 
1500 DB_SHOW_COMMAND(witness, db_witness_display)
1501 {
1502 
1503 	witness_display(db_printf);
1504 }
1505 #endif
1506