xref: /freebsd/sys/kern/subr_witness.c (revision 6990ffd8a95caaba6858ad44ff1b3157d1efba8f)
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 processes
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 #if defined(__i386__) && defined (SMP)
201 	{ "com", &lock_class_mtx_spin },
202 #endif
203 	{ "sio", &lock_class_mtx_spin },
204 #ifdef __i386__
205 	{ "cy", &lock_class_mtx_spin },
206 #endif
207 	{ "ng_node", &lock_class_mtx_spin },
208 	{ "ng_worklist", &lock_class_mtx_spin },
209 	{ "ithread table lock", &lock_class_mtx_spin },
210 	{ "sched lock", &lock_class_mtx_spin },
211 	{ "callout", &lock_class_mtx_spin },
212 	/*
213 	 * leaf locks
214 	 */
215 #ifdef SMP
216 	{ "ap boot", &lock_class_mtx_spin },
217 #ifdef __i386__
218 	{ "imen", &lock_class_mtx_spin },
219 #endif
220 	{ "smp rendezvous", &lock_class_mtx_spin },
221 #endif
222 	{ "clk", &lock_class_mtx_spin },
223 	{ NULL, NULL },
224 	{ NULL, NULL }
225 };
226 
227 static const char *dup_list[] = {
228 	"process lock",
229 	NULL
230 };
231 
232 /*
233  * Pairs of locks which have been blessed
234  * Don't complain about order problems with blessed locks
235  */
236 static struct witness_blessed blessed_list[] = {
237 };
238 static int blessed_count =
239 	sizeof(blessed_list) / sizeof(struct witness_blessed);
240 
241 /*
242  * List of all locks in the system.
243  */
244 STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245 
246 static struct mtx all_mtx = {
247 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248 	  "All locks list",		/* mtx_object.lo_name */
249 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
250 	  { NULL },			/* mtx_object.lo_list */
251 	  NULL },			/* mtx_object.lo_witness */
252 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
253 	0,				/* mtx_savecrit */
254 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
255 	{ NULL, NULL }			/* mtx_contested */
256 };
257 
258 /*
259  * This global is set to 0 once it becomes safe to use the witness code.
260  */
261 static int witness_cold = 1;
262 
263 /*
264  * Global variables for book keeping.
265  */
266 static int lock_cur_cnt;
267 static int lock_max_cnt;
268 
269 /*
270  * The WITNESS-enabled diagnostic code.
271  */
272 static void
273 witness_initialize(void *dummy __unused)
274 {
275 	struct lock_object *lock;
276 	struct witness_order_list_entry *order;
277 	struct witness *w, *w1;
278 	int i;
279 
280 	/*
281 	 * We have to release Giant before initializing its witness
282 	 * structure so that WITNESS doesn't get confused.
283 	 */
284 	mtx_unlock(&Giant);
285 	mtx_assert(&Giant, MA_NOTOWNED);
286 
287 	CTR0(KTR_WITNESS, __func__ ": initializing witness");
288 	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
289 	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
290 	for (i = 0; i < WITNESS_COUNT; i++)
291 		witness_free(&w_data[i]);
292 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
293 		witness_child_free(&w_childdata[i]);
294 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
295 		witness_lock_list_free(&w_locklistdata[i]);
296 
297 	/* First add in all the specified order lists. */
298 	for (order = order_lists; order->w_name != NULL; order++) {
299 		w = enroll(order->w_name, order->w_class);
300 		if (w == NULL)
301 			continue;
302 		w->w_file = "order list";
303 		for (order++; order->w_name != NULL; order++) {
304 			w1 = enroll(order->w_name, order->w_class);
305 			if (w1 == NULL)
306 				continue;
307 			w1->w_file = "order list";
308 			itismychild(w, w1);
309 			w = w1;
310 		}
311 	}
312 
313 	/* Iterate through all locks and add them to witness. */
314 	mtx_lock(&all_mtx);
315 	STAILQ_FOREACH(lock, &all_locks, lo_list) {
316 		if (lock->lo_flags & LO_WITNESS)
317 			lock->lo_witness = enroll(lock->lo_name,
318 			    lock->lo_class);
319 		else
320 			lock->lo_witness = NULL;
321 	}
322 	mtx_unlock(&all_mtx);
323 
324 	/* Mark the witness code as being ready for use. */
325 	atomic_store_rel_int(&witness_cold, 0);
326 
327 	mtx_lock(&Giant);
328 }
329 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
330 
331 void
332 witness_init(struct lock_object *lock)
333 {
334 	struct lock_class *class;
335 
336 	class = lock->lo_class;
337 	if (lock->lo_flags & LO_INITIALIZED)
338 		panic("%s: lock (%s) %s is already initialized", __func__,
339 		    class->lc_name, lock->lo_name);
340 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
341 	    (class->lc_flags & LC_RECURSABLE) == 0)
342 		panic("%s: lock (%s) %s can not be recursable", __func__,
343 		    class->lc_name, lock->lo_name);
344 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
345 	    (class->lc_flags & LC_SLEEPABLE) == 0)
346 		panic("%s: lock (%s) %s can not be sleepable", __func__,
347 		    class->lc_name, lock->lo_name);
348 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
349 	    (class->lc_flags & LC_UPGRADABLE) == 0)
350 		panic("%s: lock (%s) %s can not be upgradable", __func__,
351 		    class->lc_name, lock->lo_name);
352 
353 	mtx_lock(&all_mtx);
354 	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
355 	lock->lo_flags |= LO_INITIALIZED;
356 	lock_cur_cnt++;
357 	if (lock_cur_cnt > lock_max_cnt)
358 		lock_max_cnt = lock_cur_cnt;
359 	mtx_unlock(&all_mtx);
360 	if (!witness_cold && !witness_dead && panicstr == NULL &&
361 	    (lock->lo_flags & LO_WITNESS) != 0)
362 		lock->lo_witness = enroll(lock->lo_name, class);
363 	else
364 		lock->lo_witness = NULL;
365 }
366 
367 void
368 witness_destroy(struct lock_object *lock)
369 {
370 	struct witness *w;
371 
372 	if (witness_cold)
373 		panic("lock (%s) %s destroyed while witness_cold",
374 		    lock->lo_class->lc_name, lock->lo_name);
375 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
376 		panic("%s: lock (%s) %s is not initialized", __func__,
377 		    lock->lo_class->lc_name, lock->lo_name);
378 
379 	/* XXX: need to verify that no one holds the lock */
380 	w = lock->lo_witness;
381 	if (w != NULL) {
382 		mtx_lock_spin(&w_mtx);
383 		w->w_refcount--;
384 		if (w->w_refcount == 0) {
385 			CTR1(KTR_WITNESS,
386 			    __func__ ": marking witness %s as dead", w->w_name);
387 			w->w_name = "(dead)";
388 			w->w_file = "(dead)";
389 			w->w_line = 0;
390 		}
391 		mtx_unlock_spin(&w_mtx);
392 	}
393 
394 	mtx_lock(&all_mtx);
395 	lock_cur_cnt--;
396 	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
397 	lock->lo_flags &= ~LO_INITIALIZED;
398 	mtx_unlock(&all_mtx);
399 }
400 
401 static void
402 witness_display_list(void(*prnt)(const char *fmt, ...),
403 		     struct witness_list *list)
404 {
405 	struct witness *w, *w1;
406 	int found;
407 
408 	STAILQ_FOREACH(w, list, w_typelist) {
409 		if (w->w_file == NULL)
410 			continue;
411 		found = 0;
412 		STAILQ_FOREACH(w1, list, w_typelist) {
413 			if (isitmychild(w1, w)) {
414 				found++;
415 				break;
416 			}
417 		}
418 		if (found)
419 			continue;
420 		/*
421 		 * This lock has no anscestors, display its descendants.
422 		 */
423 		witness_displaydescendants(prnt, w);
424 	}
425 }
426 
427 static void
428 witness_display(void(*prnt)(const char *fmt, ...))
429 {
430 	struct witness *w;
431 
432 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
433 	witness_levelall();
434 
435 	/*
436 	 * First, handle sleep locks which have been acquired at least
437 	 * once.
438 	 */
439 	prnt("Sleep locks:\n");
440 	witness_display_list(prnt, &w_sleep);
441 
442 	/*
443 	 * Now do spin locks which have been acquired at least once.
444 	 */
445 	prnt("\nSpin locks:\n");
446 	witness_display_list(prnt, &w_spin);
447 
448 	/*
449 	 * Finally, any locks which have not been acquired yet.
450 	 */
451 	prnt("\nLocks which were never acquired:\n");
452 	STAILQ_FOREACH(w, &w_all, w_list) {
453 		if (w->w_file != NULL)
454 			continue;
455 		prnt("%s\n", w->w_name);
456 	}
457 }
458 
459 void
460 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
461 {
462 	struct lock_list_entry **lock_list, *lle;
463 	struct lock_instance *lock1, *lock2;
464 	struct lock_class *class;
465 	struct witness *w, *w1;
466 	struct proc *p;
467 	struct thread *td;
468 	int i, j;
469 #ifdef DDB
470 	int go_into_ddb = 0;
471 #endif /* DDB */
472 
473 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
474 	    panicstr != NULL)
475 		return;
476 	w = lock->lo_witness;
477 	class = lock->lo_class;
478 	td = curthread;
479 	p = td->td_proc;
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 		    curproc->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 		    curproc->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 	    curproc->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 proc *p;
789 	struct thread *td;
790 	critical_t s;
791 	int i, j;
792 
793 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
794 	    panicstr != NULL)
795 		return;
796 	td = curthread;
797 	p = td->td_proc;
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 					CTR3(KTR_WITNESS,
834 				    __func__ ": pid %d unrecursed on %s r=%d",
835 					    curproc->p_pid,
836 					    instance->li_lock->lo_name,
837 					    instance->li_flags);
838 					instance->li_flags--;
839 					goto out;
840 				}
841 				s = critical_enter();
842 				CTR3(KTR_WITNESS,
843 				    __func__ ": pid %d removed %s from lle[%d]",
844 				    curproc->p_pid, instance->li_lock->lo_name,
845 				    (*lock_list)->ll_count - 1);
846 				(*lock_list)->ll_count--;
847 				for (j = i; j < (*lock_list)->ll_count; j++)
848 					(*lock_list)->ll_children[j] =
849 					    (*lock_list)->ll_children[j + 1];
850 				critical_exit(s);
851 				if ((*lock_list)->ll_count == 0) {
852 					lle = *lock_list;
853 					*lock_list = lle->ll_next;
854 					CTR2(KTR_WITNESS,
855 					    __func__ ": pid %d removed lle %p",
856 					    curproc->p_pid, lle);
857 					witness_lock_list_free(lle);
858 				}
859 				goto out;
860 			}
861 		}
862 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
863 	    file, line);
864 out:
865 	/*
866 	 * We don't need to protect this PCPU_GET() here against preemption
867 	 * because if we hold any spinlocks then we are already protected,
868 	 * and if we don't we will get NULL if we hold no spinlocks even if
869 	 * we switch CPU's while reading it.
870 	 */
871 	if (class->lc_flags & LC_SLEEPLOCK) {
872 		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
873 			panic("switchable sleep unlock (%s) %s @ %s:%d",
874 			    class->lc_name, lock->lo_name, file, line);
875 	}
876 }
877 
878 /*
879  * Warn if any held locks are not sleepable.  Note that Giant and the lock
880  * passed in are both special cases since they are both released during the
881  * sleep process and aren't actually held while the process is asleep.
882  */
883 int
884 witness_sleep(int check_only, struct lock_object *lock, const char *file,
885 	      int line)
886 {
887 	struct lock_list_entry **lock_list, *lle;
888 	struct lock_instance *lock1;
889 	struct proc *p;
890 	struct thread *td;
891 	critical_t savecrit;
892 	int i, n;
893 
894 	if (witness_dead || panicstr != NULL)
895 		return (0);
896 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
897 	n = 0;
898 	/*
899 	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
900 	 */
901 	savecrit = critical_enter();
902 	td = curthread;
903 	p = td->td_proc;
904 	lock_list = &td->td_sleeplocks;
905 again:
906 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
907 		for (i = lle->ll_count - 1; i >= 0; i--) {
908 			lock1 = &lle->ll_children[i];
909 			if (lock1->li_lock == lock ||
910 			    lock1->li_lock == &Giant.mtx_object)
911 				continue;
912 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
913 				if (check_only == 0) {
914 					CTR3(KTR_WITNESS,
915 				    "pid %d: sleeping with lock (%s) %s held",
916 					    curproc->p_pid,
917 					    lock1->li_lock->lo_class->lc_name,
918 					    lock1->li_lock->lo_name);
919 					lock1->li_flags |= LI_SLEPT;
920 				}
921 				continue;
922 			}
923 			n++;
924 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
925 			    file, line, check_only ? "could sleep" : "sleeping",
926 			    lock1->li_lock->lo_name, lock1->li_file,
927 			    lock1->li_line);
928 		}
929 	if (lock_list == &td->td_sleeplocks) {
930 		lock_list = PCPU_PTR(spinlocks);
931 		goto again;
932 	}
933 #ifdef DDB
934 	if (witness_ddb && n)
935 		Debugger(__func__);
936 #endif /* DDB */
937 	critical_exit(savecrit);
938 	return (n);
939 }
940 
941 static struct witness *
942 enroll(const char *description, struct lock_class *lock_class)
943 {
944 	struct witness *w;
945 
946 	if (!witness_watch || witness_dead || panicstr != NULL)
947 		return (NULL);
948 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
949 		return (NULL);
950 	mtx_lock_spin(&w_mtx);
951 	STAILQ_FOREACH(w, &w_all, w_list) {
952 		if (strcmp(description, w->w_name) == 0) {
953 			w->w_refcount++;
954 			mtx_unlock_spin(&w_mtx);
955 			if (lock_class != w->w_class)
956 				panic(
957 				"lock (%s) %s does not match earlier (%s) lock",
958 				    description, lock_class->lc_name,
959 				    w->w_class->lc_name);
960 			return (w);
961 		}
962 	}
963 	/*
964 	 * This isn't quite right, as witness_cold is still 0 while we
965 	 * enroll all the locks initialized before witness_initialize().
966 	 */
967 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
968 		mtx_unlock_spin(&w_mtx);
969 		panic("spin lock %s not in order list", description);
970 	}
971 	if ((w = witness_get()) == NULL)
972 		return (NULL);
973 	w->w_name = description;
974 	w->w_class = lock_class;
975 	w->w_refcount = 1;
976 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
977 	if (lock_class->lc_flags & LC_SPINLOCK)
978 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
979 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
980 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
981 	else {
982 		mtx_unlock_spin(&w_mtx);
983 		panic("lock class %s is not sleep or spin",
984 		    lock_class->lc_name);
985 	}
986 	mtx_unlock_spin(&w_mtx);
987 	return (w);
988 }
989 
990 static int
991 itismychild(struct witness *parent, struct witness *child)
992 {
993 	static int recursed;
994 	struct witness_child_list_entry **wcl;
995 	struct witness_list *list;
996 
997 	MPASS(child != NULL && parent != NULL);
998 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
999 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1000 		panic(
1001 		"%s: parent (%s) and child (%s) are not the same lock type",
1002 		    __func__, parent->w_class->lc_name,
1003 		    child->w_class->lc_name);
1004 
1005 	/*
1006 	 * Insert "child" after "parent"
1007 	 */
1008 	wcl = &parent->w_children;
1009 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1010 		wcl = &(*wcl)->wcl_next;
1011 	if (*wcl == NULL) {
1012 		*wcl = witness_child_get();
1013 		if (*wcl == NULL)
1014 			return (1);
1015 	}
1016 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1017 
1018 	/*
1019 	 * Now prune whole tree.  We look for cases where a lock is now
1020 	 * both a descendant and a direct child of a given lock.  In that
1021 	 * case, we want to remove the direct child link from the tree.
1022 	 */
1023 	if (recursed)
1024 		return (0);
1025 	recursed = 1;
1026 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1027 		list = &w_sleep;
1028 	else
1029 		list = &w_spin;
1030 	STAILQ_FOREACH(child, list, w_typelist) {
1031 		STAILQ_FOREACH(parent, list, w_typelist) {
1032 			if (!isitmychild(parent, child))
1033 				continue;
1034 			removechild(parent, child);
1035 			if (isitmydescendant(parent, child))
1036 				continue;
1037 			itismychild(parent, child);
1038 		}
1039 	}
1040 	recursed = 0;
1041 	witness_levelall();
1042 	return (0);
1043 }
1044 
1045 static void
1046 removechild(struct witness *parent, struct witness *child)
1047 {
1048 	struct witness_child_list_entry **wcl, *wcl1;
1049 	int i;
1050 
1051 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1052 		for (i = 0; i < (*wcl)->wcl_count; i++)
1053 			if ((*wcl)->wcl_children[i] == child)
1054 				goto found;
1055 	return;
1056 found:
1057 	(*wcl)->wcl_count--;
1058 	if ((*wcl)->wcl_count > i)
1059 		(*wcl)->wcl_children[i] =
1060 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1061 	MPASS((*wcl)->wcl_children[i] != NULL);
1062 	if ((*wcl)->wcl_count != 0)
1063 		return;
1064 	wcl1 = *wcl;
1065 	*wcl = wcl1->wcl_next;
1066 	witness_child_free(wcl1);
1067 }
1068 
1069 static int
1070 isitmychild(struct witness *parent, struct witness *child)
1071 {
1072 	struct witness_child_list_entry *wcl;
1073 	int i;
1074 
1075 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1076 		for (i = 0; i < wcl->wcl_count; i++) {
1077 			if (wcl->wcl_children[i] == child)
1078 				return (1);
1079 		}
1080 	}
1081 	return (0);
1082 }
1083 
1084 static int
1085 isitmydescendant(struct witness *parent, struct witness *child)
1086 {
1087 	struct witness_child_list_entry *wcl;
1088 	int i, j;
1089 
1090 	if (isitmychild(parent, child))
1091 		return (1);
1092 	j = 0;
1093 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1094 		MPASS(j < 1000);
1095 		for (i = 0; i < wcl->wcl_count; i++) {
1096 			if (isitmydescendant(wcl->wcl_children[i], child))
1097 				return (1);
1098 		}
1099 		j++;
1100 	}
1101 	return (0);
1102 }
1103 
1104 void
1105 witness_levelall (void)
1106 {
1107 	struct witness_list *list;
1108 	struct witness *w, *w1;
1109 
1110 	/*
1111 	 * First clear all levels.
1112 	 */
1113 	STAILQ_FOREACH(w, &w_all, w_list) {
1114 		w->w_level = 0;
1115 	}
1116 
1117 	/*
1118 	 * Look for locks with no parent and level all their descendants.
1119 	 */
1120 	STAILQ_FOREACH(w, &w_all, w_list) {
1121 		/*
1122 		 * This is just an optimization, technically we could get
1123 		 * away just walking the all list each time.
1124 		 */
1125 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1126 			list = &w_sleep;
1127 		else
1128 			list = &w_spin;
1129 		STAILQ_FOREACH(w1, list, w_typelist) {
1130 			if (isitmychild(w1, w))
1131 				goto skip;
1132 		}
1133 		witness_leveldescendents(w, 0);
1134 	skip:
1135 	}
1136 }
1137 
1138 static void
1139 witness_leveldescendents(struct witness *parent, int level)
1140 {
1141 	struct witness_child_list_entry *wcl;
1142 	int i;
1143 
1144 	if (parent->w_level < level)
1145 		parent->w_level = level;
1146 	level++;
1147 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1148 		for (i = 0; i < wcl->wcl_count; i++)
1149 			witness_leveldescendents(wcl->wcl_children[i], level);
1150 }
1151 
1152 static void
1153 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1154 			   struct witness *parent)
1155 {
1156 	struct witness_child_list_entry *wcl;
1157 	int i, level;
1158 
1159 	level =  parent->w_level;
1160 	prnt("%-2d", level);
1161 	for (i = 0; i < level; i++)
1162 		prnt(" ");
1163 	prnt("%s", parent->w_name);
1164 	if (parent->w_file != NULL)
1165 		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1166 		    parent->w_line);
1167 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1168 		for (i = 0; i < wcl->wcl_count; i++)
1169 			    witness_displaydescendants(prnt,
1170 				wcl->wcl_children[i]);
1171 }
1172 
1173 static int
1174 dup_ok(struct witness *w)
1175 {
1176 	const char **dup;
1177 
1178 	for (dup = dup_list; *dup != NULL; dup++)
1179 		if (strcmp(w->w_name, *dup) == 0)
1180 			return (1);
1181 	return (0);
1182 }
1183 
1184 static int
1185 blessed(struct witness *w1, struct witness *w2)
1186 {
1187 	int i;
1188 	struct witness_blessed *b;
1189 
1190 	for (i = 0; i < blessed_count; i++) {
1191 		b = &blessed_list[i];
1192 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1193 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1194 				return (1);
1195 			continue;
1196 		}
1197 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1198 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1199 				return (1);
1200 	}
1201 	return (0);
1202 }
1203 
1204 static struct witness *
1205 witness_get(void)
1206 {
1207 	struct witness *w;
1208 
1209 	if (witness_dead) {
1210 		mtx_unlock_spin(&w_mtx);
1211 		return (NULL);
1212 	}
1213 	if (STAILQ_EMPTY(&w_free)) {
1214 		witness_dead = 1;
1215 		mtx_unlock_spin(&w_mtx);
1216 		printf("%s: witness exhausted\n", __func__);
1217 		return (NULL);
1218 	}
1219 	w = STAILQ_FIRST(&w_free);
1220 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1221 	bzero(w, sizeof(*w));
1222 	return (w);
1223 }
1224 
1225 static void
1226 witness_free(struct witness *w)
1227 {
1228 
1229 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1230 }
1231 
1232 static struct witness_child_list_entry *
1233 witness_child_get(void)
1234 {
1235 	struct witness_child_list_entry *wcl;
1236 
1237 	if (witness_dead) {
1238 		mtx_unlock_spin(&w_mtx);
1239 		return (NULL);
1240 	}
1241 	wcl = w_child_free;
1242 	if (wcl == NULL) {
1243 		witness_dead = 1;
1244 		mtx_unlock_spin(&w_mtx);
1245 		printf("%s: witness exhausted\n", __func__);
1246 		return (NULL);
1247 	}
1248 	w_child_free = wcl->wcl_next;
1249 	bzero(wcl, sizeof(*wcl));
1250 	return (wcl);
1251 }
1252 
1253 static void
1254 witness_child_free(struct witness_child_list_entry *wcl)
1255 {
1256 
1257 	wcl->wcl_next = w_child_free;
1258 	w_child_free = wcl;
1259 }
1260 
1261 static struct lock_list_entry *
1262 witness_lock_list_get(void)
1263 {
1264 	struct lock_list_entry *lle;
1265 
1266 	if (witness_dead)
1267 		return (NULL);
1268 	mtx_lock_spin(&w_mtx);
1269 	lle = w_lock_list_free;
1270 	if (lle == NULL) {
1271 		witness_dead = 1;
1272 		mtx_unlock_spin(&w_mtx);
1273 		printf("%s: witness exhausted\n", __func__);
1274 		return (NULL);
1275 	}
1276 	w_lock_list_free = lle->ll_next;
1277 	mtx_unlock_spin(&w_mtx);
1278 	bzero(lle, sizeof(*lle));
1279 	return (lle);
1280 }
1281 
1282 static void
1283 witness_lock_list_free(struct lock_list_entry *lle)
1284 {
1285 
1286 	mtx_lock_spin(&w_mtx);
1287 	lle->ll_next = w_lock_list_free;
1288 	w_lock_list_free = lle;
1289 	mtx_unlock_spin(&w_mtx);
1290 }
1291 
1292 static struct lock_instance *
1293 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1294 {
1295 	struct lock_list_entry *lle;
1296 	struct lock_instance *instance;
1297 	int i;
1298 
1299 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1300 		for (i = lle->ll_count - 1; i >= 0; i--) {
1301 			instance = &lle->ll_children[i];
1302 			if (instance->li_lock == lock)
1303 				return (instance);
1304 		}
1305 	return (NULL);
1306 }
1307 
1308 int
1309 witness_list_locks(struct lock_list_entry **lock_list)
1310 {
1311 	struct lock_list_entry *lle;
1312 	struct lock_instance *instance;
1313 	struct lock_object *lock;
1314 	int i, nheld;
1315 
1316 	nheld = 0;
1317 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1318 		for (i = lle->ll_count - 1; i >= 0; i--) {
1319 			instance = &lle->ll_children[i];
1320 			lock = instance->li_lock;
1321 			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1322 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1323 			    "exclusive" : "shared",
1324 			    lock->lo_class->lc_name, lock->lo_name, lock,
1325 			    instance->li_file, instance->li_line);
1326 			nheld++;
1327 		}
1328 	return (nheld);
1329 }
1330 
1331 /*
1332  * Calling this on td != curthread is bad unless we are in ddb.
1333  */
1334 int
1335 witness_list(struct thread *td)
1336 {
1337 	critical_t savecrit;
1338 	int nheld;
1339 
1340 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1341 #ifdef DDB
1342 	KASSERT(td == curthread || db_active,
1343 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1344 	if (!db_active && witness_dead)
1345 		return (0);
1346 #else
1347 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1348 	if (witness_dead)
1349 		return (0);
1350 #endif
1351 	nheld = witness_list_locks(&td->td_sleeplocks);
1352 
1353 	/*
1354 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1355 	 * if p is currently executing on some other CPU and holds spin locks
1356 	 * as we won't display those locks.  If we had a MI way of getting
1357 	 * the per-cpu data for a given cpu then we could use p->p_oncpu to
1358 	 * get the list of spinlocks for this process and "fix" this.
1359 	 */
1360 	if (td == curthread) {
1361 		/*
1362 		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1363 		 * change.
1364 		 */
1365 		savecrit = critical_enter();
1366 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1367 		critical_exit(savecrit);
1368 	}
1369 	return (nheld);
1370 }
1371 
1372 void
1373 witness_save(struct lock_object *lock, const char **filep, int *linep)
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 	*filep = instance->li_file;
1388 	*linep = instance->li_line;
1389 }
1390 
1391 void
1392 witness_restore(struct lock_object *lock, const char *file, int line)
1393 {
1394 	struct lock_instance *instance;
1395 
1396 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1397 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1398 		return;
1399 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1400 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1401 		    lock->lo_class->lc_name, lock->lo_name);
1402 	instance = find_instance(curthread->td_sleeplocks, lock);
1403 	if (instance == NULL)
1404 		panic("%s: lock (%s) %s not locked", __func__,
1405 		    lock->lo_class->lc_name, lock->lo_name);
1406 	lock->lo_witness->w_file = file;
1407 	lock->lo_witness->w_line = line;
1408 	instance->li_file = file;
1409 	instance->li_line = line;
1410 }
1411 
1412 void
1413 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1414 {
1415 #ifdef INVARIANT_SUPPORT
1416 	struct lock_instance *instance;
1417 
1418 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1419 		return;
1420 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1421 		instance = find_instance(curthread->td_sleeplocks, lock);
1422 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1423 		instance = find_instance(PCPU_GET(spinlocks), lock);
1424 	else
1425 		panic("Lock (%s) %s is not sleep or spin!",
1426 		    lock->lo_class->lc_name, lock->lo_name);
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 		if ((flags & LA_XLOCKED) != 0 &&
1446 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1447 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1448 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1449 		if ((flags & LA_SLOCKED) != 0 &&
1450 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1451 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1452 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1453 		if ((flags & LA_RECURSED) != 0 &&
1454 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1455 			panic("Lock (%s) %s not recursed @ %s:%d.",
1456 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1457 		if ((flags & LA_NOTRECURSED) != 0 &&
1458 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1459 			panic("Lock (%s) %s recursed @ %s:%d.",
1460 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1461 		break;
1462 	default:
1463 		panic("Invalid lock assertion at %s:%d.", file, line);
1464 
1465 	}
1466 #endif	/* INVARIANT_SUPPORT */
1467 }
1468 
1469 #ifdef DDB
1470 
1471 DB_SHOW_COMMAND(locks, db_witness_list)
1472 {
1473 	struct thread *td;
1474 	pid_t pid;
1475 	struct proc *p;
1476 
1477 	if (have_addr) {
1478 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1479 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1480 		    ((addr >> 16) % 16) * 10000;
1481 		/* sx_slock(&allproc_lock); */
1482 		FOREACH_PROC_IN_SYSTEM(p) {
1483 			if (p->p_pid == pid)
1484 				break;
1485 		}
1486 		/* sx_sunlock(&allproc_lock); */
1487 		if (p == NULL) {
1488 			db_printf("pid %d not found\n", pid);
1489 			return;
1490 		}
1491 		td = &p->p_thread; /* XXXKSE */
1492 	} else {
1493 		td = curthread;
1494 	}
1495 	witness_list(td);
1496 }
1497 
1498 DB_SHOW_COMMAND(witness, db_witness_display)
1499 {
1500 
1501 	witness_display(db_printf);
1502 }
1503 #endif
1504