xref: /freebsd/sys/kern/subr_witness.c (revision c4f6a2a9e1b1879b618c436ab4f56ff75c73a0f5)
1 /*-
2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
13  *    promote products derived from this software without specific prior
14  *    written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30  * $FreeBSD$
31  */
32 
33 /*
34  * Implementation of the `witness' lock verifier.  Originally implemented for
35  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36  * classes in FreeBSD.
37  */
38 
39 /*
40  *	Main Entry: witness
41  *	Pronunciation: 'wit-n&s
42  *	Function: noun
43  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44  *	    testimony, witness, from 2wit
45  *	Date: before 12th century
46  *	1 : attestation of a fact or event : TESTIMONY
47  *	2 : one that gives evidence; specifically : one who testifies in
48  *	    a cause or before a judicial tribunal
49  *	3 : one asked to be present at a transaction so as to be able to
50  *	    testify to its having taken place
51  *	4 : one who has personal knowledge of something
52  *	5 a : something serving as evidence or proof : SIGN
53  *	  b : public affirmation by word or example of usually
54  *	      religious faith or conviction <the heroic witness to divine
55  *	      life -- Pilot>
56  *	6 capitalized : a member of the Jehovah's Witnesses
57  */
58 
59 #include "opt_ddb.h"
60 #include "opt_witness.h"
61 
62 #include <sys/param.h>
63 #include <sys/bus.h>
64 #include <sys/kernel.h>
65 #include <sys/ktr.h>
66 #include <sys/lock.h>
67 #include <sys/malloc.h>
68 #include <sys/mutex.h>
69 #include <sys/proc.h>
70 #include <sys/sysctl.h>
71 #include <sys/systm.h>
72 
73 #include <ddb/ddb.h>
74 
75 #define WITNESS_COUNT 200
76 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77 /*
78  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80  * probably be safe for the most part, but it's still a SWAG.
81  */
82 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83 
84 #define	WITNESS_NCHILDREN 6
85 
86 struct witness_child_list_entry;
87 
88 struct witness {
89 	const	char *w_name;
90 	struct	lock_class *w_class;
91 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94 	const	char *w_file;
95 	int	w_line;
96 	u_int	w_level;
97 	u_int	w_refcount;
98 	u_char	w_Giant_squawked:1;
99 	u_char	w_other_squawked:1;
100 	u_char	w_same_squawked:1;
101 };
102 
103 struct witness_child_list_entry {
104 	struct	witness_child_list_entry *wcl_next;
105 	struct	witness *wcl_children[WITNESS_NCHILDREN];
106 	u_int	wcl_count;
107 };
108 
109 STAILQ_HEAD(witness_list, witness);
110 
111 struct witness_blessed {
112 	const	char *b_lock1;
113 	const	char *b_lock2;
114 };
115 
116 struct witness_order_list_entry {
117 	const	char *w_name;
118 	struct	lock_class *w_class;
119 };
120 
121 static struct	witness *enroll(const char *description,
122 				struct lock_class *lock_class);
123 static int	itismychild(struct witness *parent, struct witness *child);
124 static void	removechild(struct witness *parent, struct witness *child);
125 static int	isitmychild(struct witness *parent, struct witness *child);
126 static int	isitmydescendant(struct witness *parent, struct witness *child);
127 static int	blessed(struct witness *, struct witness *);
128 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
129 					   struct witness *);
130 static void	witness_leveldescendents(struct witness *parent, int level);
131 static void	witness_levelall(void);
132 static struct	witness *witness_get(void);
133 static void	witness_free(struct witness *m);
134 static struct	witness_child_list_entry *witness_child_get(void);
135 static void	witness_child_free(struct witness_child_list_entry *wcl);
136 static struct	lock_list_entry *witness_lock_list_get(void);
137 static void	witness_lock_list_free(struct lock_list_entry *lle);
138 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
139 					     struct lock_object *lock);
140 #if defined(DDB)
141 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
142 				     struct witness_list *list);
143 static void	witness_display(void(*)(const char *fmt, ...));
144 #endif
145 
146 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
147 
148 static int witness_watch = 1;
149 TUNABLE_INT("debug.witness_watch", &witness_watch);
150 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
151 
152 #ifdef DDB
153 /*
154  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
155  * drop into kdebug() when:
156  *	- a lock heirarchy violation occurs
157  *	- locks are held when going to sleep.
158  */
159 #ifdef WITNESS_DDB
160 int	witness_ddb = 1;
161 #else
162 int	witness_ddb = 0;
163 #endif
164 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
165 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
166 #endif /* DDB */
167 
168 #ifdef WITNESS_SKIPSPIN
169 int	witness_skipspin = 1;
170 #else
171 int	witness_skipspin = 0;
172 #endif
173 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
174 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
175     "");
176 
177 static struct mtx w_mtx;
178 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
179 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
180 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
181 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
182 static struct witness_child_list_entry *w_child_free = NULL;
183 static struct lock_list_entry *w_lock_list_free = NULL;
184 static int witness_dead;	/* fatal error, probably no memory */
185 
186 static struct witness w_data[WITNESS_COUNT];
187 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
188 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
189 
190 static struct witness_order_list_entry order_lists[] = {
191 	{ "Giant", &lock_class_mtx_sleep },
192 	{ "proctree", &lock_class_sx },
193 	{ "allproc", &lock_class_sx },
194 	{ "sigio lock", &lock_class_mtx_sleep },
195 	{ "process group", &lock_class_mtx_sleep },
196 	{ "process lock", &lock_class_mtx_sleep },
197 	{ "session", &lock_class_mtx_sleep },
198 	{ "uidinfo hash", &lock_class_mtx_sleep },
199 	{ "uidinfo struct", &lock_class_mtx_sleep },
200 	{ NULL, NULL },
201 	/*
202 	 * spin locks
203 	 */
204 #ifdef SMP
205 	{ "ap boot", &lock_class_mtx_spin },
206 #ifdef __i386__
207 	{ "com", &lock_class_mtx_spin },
208 #endif
209 #endif
210 	{ "sio", &lock_class_mtx_spin },
211 #ifdef __i386__
212 	{ "cy", &lock_class_mtx_spin },
213 #endif
214 	{ "ng_node", &lock_class_mtx_spin },
215 	{ "ng_worklist", &lock_class_mtx_spin },
216 	{ "ithread table lock", &lock_class_mtx_spin },
217 	{ "sched lock", &lock_class_mtx_spin },
218 	{ "callout", &lock_class_mtx_spin },
219 	/*
220 	 * leaf locks
221 	 */
222 	{ "allpmaps", &lock_class_mtx_spin },
223 	{ "vm page buckets mutex", &lock_class_mtx_spin },
224 	{ "vm page queue free mutex", &lock_class_mtx_spin },
225 	{ "icu", &lock_class_mtx_spin },
226 #ifdef SMP
227 	{ "smp rendezvous", &lock_class_mtx_spin },
228 #if defined(__i386__) && defined(APIC_IO)
229 	{ "tlb", &lock_class_mtx_spin },
230 #endif
231 #endif
232 	{ "clk", &lock_class_mtx_spin },
233 	{ "mutex profiling lock", &lock_class_mtx_spin },
234 	{ "zombie_thread_lock", &lock_class_mtx_spin },
235 	{ NULL, NULL },
236 	{ NULL, NULL }
237 };
238 
239 /*
240  * Pairs of locks which have been blessed
241  * Don't complain about order problems with blessed locks
242  */
243 static struct witness_blessed blessed_list[] = {
244 };
245 static int blessed_count =
246 	sizeof(blessed_list) / sizeof(struct witness_blessed);
247 
248 /*
249  * List of all locks in the system.
250  */
251 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
252 
253 static struct mtx all_mtx = {
254 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
255 	  "All locks list",		/* mtx_object.lo_name */
256 	  "All locks list",		/* mtx_object.lo_type */
257 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
258 	  { NULL, NULL },		/* mtx_object.lo_list */
259 	  NULL },			/* mtx_object.lo_witness */
260 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
261 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
262 	{ NULL, NULL }			/* mtx_contested */
263 };
264 
265 /*
266  * This global is set to 0 once it becomes safe to use the witness code.
267  */
268 static int witness_cold = 1;
269 
270 /*
271  * Global variables for book keeping.
272  */
273 static int lock_cur_cnt;
274 static int lock_max_cnt;
275 
276 /*
277  * The WITNESS-enabled diagnostic code.
278  */
279 static void
280 witness_initialize(void *dummy __unused)
281 {
282 	struct lock_object *lock;
283 	struct witness_order_list_entry *order;
284 	struct witness *w, *w1;
285 	int i;
286 
287 	/*
288 	 * We have to release Giant before initializing its witness
289 	 * structure so that WITNESS doesn't get confused.
290 	 */
291 	mtx_unlock(&Giant);
292 	mtx_assert(&Giant, MA_NOTOWNED);
293 
294 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
295 	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
296 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
297 	    MTX_NOWITNESS);
298 	for (i = 0; i < WITNESS_COUNT; i++)
299 		witness_free(&w_data[i]);
300 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
301 		witness_child_free(&w_childdata[i]);
302 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
303 		witness_lock_list_free(&w_locklistdata[i]);
304 
305 	/* First add in all the specified order lists. */
306 	for (order = order_lists; order->w_name != NULL; order++) {
307 		w = enroll(order->w_name, order->w_class);
308 		if (w == NULL)
309 			continue;
310 		w->w_file = "order list";
311 		for (order++; order->w_name != NULL; order++) {
312 			w1 = enroll(order->w_name, order->w_class);
313 			if (w1 == NULL)
314 				continue;
315 			w1->w_file = "order list";
316 			itismychild(w, w1);
317 			w = w1;
318 		}
319 	}
320 
321 	/* Iterate through all locks and add them to witness. */
322 	mtx_lock(&all_mtx);
323 	TAILQ_FOREACH(lock, &all_locks, lo_list) {
324 		if (lock->lo_flags & LO_WITNESS)
325 			lock->lo_witness = enroll(lock->lo_type,
326 			    lock->lo_class);
327 		else
328 			lock->lo_witness = NULL;
329 	}
330 	mtx_unlock(&all_mtx);
331 
332 	/* Mark the witness code as being ready for use. */
333 	atomic_store_rel_int(&witness_cold, 0);
334 
335 	mtx_lock(&Giant);
336 }
337 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
338 
339 void
340 witness_init(struct lock_object *lock)
341 {
342 	struct lock_class *class;
343 
344 	class = lock->lo_class;
345 	if (lock->lo_flags & LO_INITIALIZED)
346 		panic("%s: lock (%s) %s is already initialized", __func__,
347 		    class->lc_name, lock->lo_name);
348 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
349 	    (class->lc_flags & LC_RECURSABLE) == 0)
350 		panic("%s: lock (%s) %s can not be recursable", __func__,
351 		    class->lc_name, lock->lo_name);
352 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
353 	    (class->lc_flags & LC_SLEEPABLE) == 0)
354 		panic("%s: lock (%s) %s can not be sleepable", __func__,
355 		    class->lc_name, lock->lo_name);
356 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
357 	    (class->lc_flags & LC_UPGRADABLE) == 0)
358 		panic("%s: lock (%s) %s can not be upgradable", __func__,
359 		    class->lc_name, lock->lo_name);
360 
361 	mtx_lock(&all_mtx);
362 	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
363 	lock->lo_flags |= LO_INITIALIZED;
364 	lock_cur_cnt++;
365 	if (lock_cur_cnt > lock_max_cnt)
366 		lock_max_cnt = lock_cur_cnt;
367 	mtx_unlock(&all_mtx);
368 	if (!witness_cold && !witness_dead && panicstr == NULL &&
369 	    (lock->lo_flags & LO_WITNESS) != 0)
370 		lock->lo_witness = enroll(lock->lo_type, class);
371 	else
372 		lock->lo_witness = NULL;
373 }
374 
375 void
376 witness_destroy(struct lock_object *lock)
377 {
378 	struct witness *w;
379 
380 	if (witness_cold)
381 		panic("lock (%s) %s destroyed while witness_cold",
382 		    lock->lo_class->lc_name, lock->lo_name);
383 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
384 		panic("%s: lock (%s) %s is not initialized", __func__,
385 		    lock->lo_class->lc_name, lock->lo_name);
386 
387 	/* XXX: need to verify that no one holds the lock */
388 	w = lock->lo_witness;
389 	if (w != NULL) {
390 		mtx_lock_spin(&w_mtx);
391 		MPASS(w->w_refcount > 0);
392 		w->w_refcount--;
393 		mtx_unlock_spin(&w_mtx);
394 	}
395 
396 	mtx_lock(&all_mtx);
397 	lock_cur_cnt--;
398 	TAILQ_REMOVE(&all_locks, lock, lo_list);
399 	lock->lo_flags &= ~LO_INITIALIZED;
400 	mtx_unlock(&all_mtx);
401 }
402 
403 #if defined(DDB)
404 static void
405 witness_display_list(void(*prnt)(const char *fmt, ...),
406 		     struct witness_list *list)
407 {
408 	struct witness *w, *w1;
409 	int found;
410 
411 	STAILQ_FOREACH(w, list, w_typelist) {
412 		if (w->w_file == NULL)
413 			continue;
414 		found = 0;
415 		STAILQ_FOREACH(w1, list, w_typelist) {
416 			if (isitmychild(w1, w)) {
417 				found++;
418 				break;
419 			}
420 		}
421 		if (found)
422 			continue;
423 		/*
424 		 * This lock has no anscestors, display its descendants.
425 		 */
426 		witness_displaydescendants(prnt, w);
427 	}
428 }
429 
430 static void
431 witness_display(void(*prnt)(const char *fmt, ...))
432 {
433 	struct witness *w;
434 
435 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
436 	witness_levelall();
437 
438 	/*
439 	 * First, handle sleep locks which have been acquired at least
440 	 * once.
441 	 */
442 	prnt("Sleep locks:\n");
443 	witness_display_list(prnt, &w_sleep);
444 
445 	/*
446 	 * Now do spin locks which have been acquired at least once.
447 	 */
448 	prnt("\nSpin locks:\n");
449 	witness_display_list(prnt, &w_spin);
450 
451 	/*
452 	 * Finally, any locks which have not been acquired yet.
453 	 */
454 	prnt("\nLocks which were never acquired:\n");
455 	STAILQ_FOREACH(w, &w_all, w_list) {
456 		if (w->w_file != NULL || w->w_refcount == 0)
457 			continue;
458 		prnt("%s\n", w->w_name);
459 	}
460 }
461 #endif
462 
463 void
464 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
465 {
466 	struct lock_list_entry **lock_list, *lle;
467 	struct lock_instance *lock1, *lock2;
468 	struct lock_class *class;
469 	struct witness *w, *w1;
470 	struct thread *td;
471 	int i, j;
472 #ifdef DDB
473 	int go_into_ddb = 0;
474 #endif /* DDB */
475 
476 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
477 	    panicstr != NULL)
478 		return;
479 	w = lock->lo_witness;
480 	class = lock->lo_class;
481 	td = curthread;
482 
483 	if (class->lc_flags & LC_SLEEPLOCK) {
484 		/*
485 		 * Since spin locks include a critical section, this check
486 		 * impliclty enforces a lock order of all sleep locks before
487 		 * all spin locks.
488 		 */
489 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
490 			panic("blockable sleep lock (%s) %s @ %s:%d",
491 			    class->lc_name, lock->lo_name, file, line);
492 		lock_list = &td->td_sleeplocks;
493 	} else
494 		lock_list = PCPU_PTR(spinlocks);
495 
496 	/*
497 	 * Try locks do not block if they fail to acquire the lock, thus
498 	 * there is no danger of deadlocks or of switching while holding a
499 	 * spin lock if we acquire a lock via a try operation.
500 	 */
501 	if (flags & LOP_TRYLOCK)
502 		goto out;
503 
504 	/*
505 	 * Is this the first lock acquired?  If so, then no order checking
506 	 * is needed.
507 	 */
508 	if (*lock_list == NULL)
509 		goto out;
510 
511 	/*
512 	 * Check to see if we are recursing on a lock we already own.
513 	 */
514 	lock1 = find_instance(*lock_list, lock);
515 	if (lock1 != NULL) {
516 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
517 		    (flags & LOP_EXCLUSIVE) == 0) {
518 			printf("shared lock of (%s) %s @ %s:%d\n",
519 			    class->lc_name, lock->lo_name, file, line);
520 			printf("while exclusively locked from %s:%d\n",
521 			    lock1->li_file, lock1->li_line);
522 			panic("share->excl");
523 		}
524 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
525 		    (flags & LOP_EXCLUSIVE) != 0) {
526 			printf("exclusive lock of (%s) %s @ %s:%d\n",
527 			    class->lc_name, lock->lo_name, file, line);
528 			printf("while share locked from %s:%d\n",
529 			    lock1->li_file, lock1->li_line);
530 			panic("excl->share");
531 		}
532 		lock1->li_flags++;
533 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
534 			printf(
535 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
536 			    class->lc_name, lock->lo_name, file, line);
537 			printf("first acquired @ %s:%d\n", lock1->li_file,
538 			    lock1->li_line);
539 			panic("recurse");
540 		}
541 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
542 		    td->td_proc->p_pid, lock->lo_name,
543 		    lock1->li_flags & LI_RECURSEMASK);
544 		lock1->li_file = file;
545 		lock1->li_line = line;
546 		return;
547 	}
548 
549 	/*
550 	 * Check for duplicate locks of the same type.  Note that we only
551 	 * have to check for this on the last lock we just acquired.  Any
552 	 * other cases will be caught as lock order violations.
553 	 */
554 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
555 	w1 = lock1->li_lock->lo_witness;
556 	if (w1 == w) {
557 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
558 			goto out;
559 		w->w_same_squawked = 1;
560 		printf("acquiring duplicate lock of same type: \"%s\"\n",
561 			lock->lo_type);
562 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
563 		    lock1->li_file, lock1->li_line);
564 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
565 #ifdef DDB
566 		go_into_ddb = 1;
567 #endif /* DDB */
568 		goto out;
569 	}
570 	MPASS(!mtx_owned(&w_mtx));
571 	mtx_lock_spin(&w_mtx);
572 	/*
573 	 * If we have a known higher number just say ok
574 	 */
575 	if (witness_watch > 1 && w->w_level > w1->w_level) {
576 		mtx_unlock_spin(&w_mtx);
577 		goto out;
578 	}
579 	if (isitmydescendant(w1, w)) {
580 		mtx_unlock_spin(&w_mtx);
581 		goto out;
582 	}
583 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
584 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
585 
586 			MPASS(j < WITNESS_COUNT);
587 			lock1 = &lle->ll_children[i];
588 			w1 = lock1->li_lock->lo_witness;
589 
590 			/*
591 			 * If this lock doesn't undergo witness checking,
592 			 * then skip it.
593 			 */
594 			if (w1 == NULL) {
595 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
596 				    ("lock missing witness structure"));
597 				continue;
598 			}
599 			/*
600 			 * If we are locking Giant and we slept with this
601 			 * lock, then skip it.
602 			 */
603 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
604 			    lock == &Giant.mtx_object)
605 				continue;
606 			/*
607 			 * If we are locking a sleepable lock and this lock
608 			 * isn't sleepable and isn't Giant, we want to treat
609 			 * it as a lock order violation to enfore a general
610 			 * lock order of sleepable locks before non-sleepable
611 			 * locks.  Thus, we only bother checking the lock
612 			 * order hierarchy if we pass the initial test.
613 			 */
614 			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
615 			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
616 			    lock1->li_lock != &Giant.mtx_object)) &&
617 			    !isitmydescendant(w, w1))
618 				continue;
619 			/*
620 			 * We have a lock order violation, check to see if it
621 			 * is allowed or has already been yelled about.
622 			 */
623 			mtx_unlock_spin(&w_mtx);
624 			if (blessed(w, w1))
625 				goto out;
626 			if (lock1->li_lock == &Giant.mtx_object) {
627 				if (w1->w_Giant_squawked)
628 					goto out;
629 				else
630 					w1->w_Giant_squawked = 1;
631 			} else {
632 				if (w1->w_other_squawked)
633 					goto out;
634 				else
635 					w1->w_other_squawked = 1;
636 			}
637 			/*
638 			 * Ok, yell about it.
639 			 */
640 			printf("lock order reversal\n");
641 			/*
642 			 * Try to locate an earlier lock with
643 			 * witness w in our list.
644 			 */
645 			do {
646 				lock2 = &lle->ll_children[i];
647 				MPASS(lock2->li_lock != NULL);
648 				if (lock2->li_lock->lo_witness == w)
649 					break;
650 				i--;
651 				if (i == 0 && lle->ll_next != NULL) {
652 					lle = lle->ll_next;
653 					i = lle->ll_count - 1;
654 					MPASS(i != 0);
655 				}
656 			} while (i >= 0);
657 			if (i < 0) {
658 				printf(" 1st %p %s (%s) @ %s:%d\n",
659 				    lock1->li_lock, lock1->li_lock->lo_name,
660 				    lock1->li_lock->lo_type, lock1->li_file,
661 				    lock1->li_line);
662 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
663 				    lock->lo_name, lock->lo_type, file, line);
664 			} else {
665 				printf(" 1st %p %s (%s) @ %s:%d\n",
666 				    lock2->li_lock, lock2->li_lock->lo_name,
667 				    lock2->li_lock->lo_type, lock2->li_file,
668 				    lock2->li_line);
669 				printf(" 2nd %p %s (%s) @ %s:%d\n",
670 				    lock1->li_lock, lock1->li_lock->lo_name,
671 				    lock1->li_lock->lo_type, lock1->li_file,
672 				    lock1->li_line);
673 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
674 				    lock->lo_name, lock->lo_type, file, line);
675 			}
676 #ifdef DDB
677 			go_into_ddb = 1;
678 #endif /* DDB */
679 			goto out;
680 		}
681 	}
682 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
683 	/*
684 	 * Don't build a new relationship if we are locking Giant just
685 	 * after waking up and the previous lock in the list was acquired
686 	 * prior to blocking.
687 	 */
688 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
689 		mtx_unlock_spin(&w_mtx);
690 	else {
691 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
692 		    lock->lo_type, lock1->li_lock->lo_type);
693 		if (!itismychild(lock1->li_lock->lo_witness, w))
694 			mtx_unlock_spin(&w_mtx);
695 	}
696 
697 out:
698 #ifdef DDB
699 	if (witness_ddb && go_into_ddb)
700 		Debugger(__func__);
701 #endif /* DDB */
702 	w->w_file = file;
703 	w->w_line = line;
704 
705 	lle = *lock_list;
706 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
707 		lle = witness_lock_list_get();
708 		if (lle == NULL)
709 			return;
710 		lle->ll_next = *lock_list;
711 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
712 		    td->td_proc->p_pid, lle);
713 		*lock_list = lle;
714 	}
715 	lock1 = &lle->ll_children[lle->ll_count++];
716 	lock1->li_lock = lock;
717 	lock1->li_line = line;
718 	lock1->li_file = file;
719 	if ((flags & LOP_EXCLUSIVE) != 0)
720 		lock1->li_flags = LI_EXCLUSIVE;
721 	else
722 		lock1->li_flags = 0;
723 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
724 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
725 }
726 
727 void
728 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
729 {
730 	struct lock_instance *instance;
731 	struct lock_class *class;
732 
733 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
734 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
735 		return;
736 	class = lock->lo_class;
737 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
738 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
739 		    class->lc_name, lock->lo_name, file, line);
740 	if ((flags & LOP_TRYLOCK) == 0)
741 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
742 		    lock->lo_name, file, line);
743 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
744 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
745 		    class->lc_name, lock->lo_name, file, line);
746 	instance = find_instance(curthread->td_sleeplocks, lock);
747 	if (instance == NULL)
748 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
749 		    class->lc_name, lock->lo_name, file, line);
750 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
751 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
752 		    class->lc_name, lock->lo_name, file, line);
753 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
754 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
755 		    class->lc_name, lock->lo_name,
756 		    instance->li_flags & LI_RECURSEMASK, file, line);
757 	instance->li_flags |= LI_EXCLUSIVE;
758 }
759 
760 void
761 witness_downgrade(struct lock_object *lock, int flags, const char *file,
762     int line)
763 {
764 	struct lock_instance *instance;
765 	struct lock_class *class;
766 
767 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
768 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
769 		return;
770 	class = lock->lo_class;
771 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
772 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
773 		    class->lc_name, lock->lo_name, file, line);
774 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
775 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
776 		    class->lc_name, lock->lo_name, file, line);
777 	instance = find_instance(curthread->td_sleeplocks, lock);
778 	if (instance == NULL)
779 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
780 		    class->lc_name, lock->lo_name, file, line);
781 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
782 		panic("downgrade of shared lock (%s) %s @ %s:%d",
783 		    class->lc_name, lock->lo_name, file, line);
784 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
785 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
786 		    class->lc_name, lock->lo_name,
787 		    instance->li_flags & LI_RECURSEMASK, file, line);
788 	instance->li_flags &= ~LI_EXCLUSIVE;
789 }
790 
791 void
792 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
793 {
794 	struct lock_list_entry **lock_list, *lle;
795 	struct lock_instance *instance;
796 	struct lock_class *class;
797 	struct thread *td;
798 	register_t s;
799 	int i, j;
800 
801 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
802 	    panicstr != NULL)
803 		return;
804 	td = curthread;
805 	class = lock->lo_class;
806 	if (class->lc_flags & LC_SLEEPLOCK)
807 		lock_list = &td->td_sleeplocks;
808 	else
809 		lock_list = PCPU_PTR(spinlocks);
810 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
811 		for (i = 0; i < (*lock_list)->ll_count; i++) {
812 			instance = &(*lock_list)->ll_children[i];
813 			if (instance->li_lock == lock) {
814 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
815 				    (flags & LOP_EXCLUSIVE) == 0) {
816 					printf(
817 					"shared unlock of (%s) %s @ %s:%d\n",
818 					    class->lc_name, lock->lo_name,
819 					    file, line);
820 					printf(
821 					"while exclusively locked from %s:%d\n",
822 					    instance->li_file,
823 					    instance->li_line);
824 					panic("excl->ushare");
825 				}
826 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
827 				    (flags & LOP_EXCLUSIVE) != 0) {
828 					printf(
829 					"exclusive unlock of (%s) %s @ %s:%d\n",
830 					    class->lc_name, lock->lo_name,
831 					    file, line);
832 					printf(
833 					"while share locked from %s:%d\n",
834 					    instance->li_file,
835 					    instance->li_line);
836 					panic("share->uexcl");
837 				}
838 				/* If we are recursed, unrecurse. */
839 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
840 					CTR4(KTR_WITNESS,
841 				    "%s: pid %d unrecursed on %s r=%d", __func__,
842 					    td->td_proc->p_pid,
843 					    instance->li_lock->lo_name,
844 					    instance->li_flags);
845 					instance->li_flags--;
846 					return;
847 				}
848 				s = intr_disable();
849 				CTR4(KTR_WITNESS,
850 				    "%s: pid %d removed %s from lle[%d]", __func__,
851 				    td->td_proc->p_pid,
852 				    instance->li_lock->lo_name,
853 				    (*lock_list)->ll_count - 1);
854 				for (j = i; j < (*lock_list)->ll_count - 1; j++)
855 					(*lock_list)->ll_children[j] =
856 					    (*lock_list)->ll_children[j + 1];
857 				(*lock_list)->ll_count--;
858 				intr_restore(s);
859 				if ((*lock_list)->ll_count == 0) {
860 					lle = *lock_list;
861 					*lock_list = lle->ll_next;
862 					CTR3(KTR_WITNESS,
863 					    "%s: pid %d removed lle %p", __func__,
864 					    td->td_proc->p_pid, lle);
865 					witness_lock_list_free(lle);
866 				}
867 				return;
868 			}
869 		}
870 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
871 	    file, line);
872 }
873 
874 /*
875  * Warn if any held locks are not sleepable.  Note that Giant and the lock
876  * passed in are both special cases since they are both released during the
877  * sleep process and aren't actually held while the thread is asleep.
878  */
879 int
880 witness_sleep(int check_only, struct lock_object *lock, const char *file,
881 	      int line)
882 {
883 	struct lock_list_entry **lock_list, *lle;
884 	struct lock_instance *lock1;
885 	struct thread *td;
886 	int i, n;
887 
888 	if (witness_cold || witness_dead || panicstr != NULL)
889 		return (0);
890 	n = 0;
891 	td = curthread;
892 	lock_list = &td->td_sleeplocks;
893 again:
894 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
895 		for (i = lle->ll_count - 1; i >= 0; i--) {
896 			lock1 = &lle->ll_children[i];
897 			if (lock1->li_lock == lock ||
898 			    lock1->li_lock == &Giant.mtx_object)
899 				continue;
900 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
901 				if (check_only == 0) {
902 					CTR3(KTR_WITNESS,
903 				    "pid %d: sleeping with lock (%s) %s held",
904 					    td->td_proc->p_pid,
905 					    lock1->li_lock->lo_class->lc_name,
906 					    lock1->li_lock->lo_name);
907 					lock1->li_flags |= LI_SLEPT;
908 				}
909 				continue;
910 			}
911 			n++;
912 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
913 			    file, line, check_only ? "could sleep" : "sleeping",
914 			    lock1->li_lock->lo_name, lock1->li_file,
915 			    lock1->li_line);
916 		}
917 	if (lock_list == &td->td_sleeplocks && PCPU_GET(spinlocks) != NULL) {
918 		/*
919 		 * Since we already hold a spinlock preemption is
920 		 * already blocked.
921 		 */
922 		lock_list = PCPU_PTR(spinlocks);
923 		goto again;
924 	}
925 #ifdef DDB
926 	if (witness_ddb && n)
927 		Debugger(__func__);
928 #endif /* DDB */
929 	return (n);
930 }
931 
932 static struct witness *
933 enroll(const char *description, struct lock_class *lock_class)
934 {
935 	struct witness *w;
936 
937 	if (!witness_watch || witness_dead || panicstr != NULL)
938 		return (NULL);
939 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
940 		return (NULL);
941 	mtx_lock_spin(&w_mtx);
942 	STAILQ_FOREACH(w, &w_all, w_list) {
943 		if (w->w_name == description || (w->w_refcount > 0 &&
944 		    strcmp(description, w->w_name) == 0)) {
945 			w->w_refcount++;
946 			mtx_unlock_spin(&w_mtx);
947 			if (lock_class != w->w_class)
948 				panic(
949 				"lock (%s) %s does not match earlier (%s) lock",
950 				    description, lock_class->lc_name,
951 				    w->w_class->lc_name);
952 			return (w);
953 		}
954 	}
955 	/*
956 	 * This isn't quite right, as witness_cold is still 0 while we
957 	 * enroll all the locks initialized before witness_initialize().
958 	 */
959 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
960 		mtx_unlock_spin(&w_mtx);
961 		panic("spin lock %s not in order list", description);
962 	}
963 	if ((w = witness_get()) == NULL)
964 		return (NULL);
965 	w->w_name = description;
966 	w->w_class = lock_class;
967 	w->w_refcount = 1;
968 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
969 	if (lock_class->lc_flags & LC_SPINLOCK)
970 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
971 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
972 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
973 	else {
974 		mtx_unlock_spin(&w_mtx);
975 		panic("lock class %s is not sleep or spin",
976 		    lock_class->lc_name);
977 	}
978 	mtx_unlock_spin(&w_mtx);
979 	return (w);
980 }
981 
982 static int
983 itismychild(struct witness *parent, struct witness *child)
984 {
985 	static int recursed;
986 	struct witness_child_list_entry **wcl;
987 	struct witness_list *list;
988 
989 	MPASS(child != NULL && parent != NULL);
990 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
991 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
992 		panic(
993 		"%s: parent (%s) and child (%s) are not the same lock type",
994 		    __func__, parent->w_class->lc_name,
995 		    child->w_class->lc_name);
996 
997 	/*
998 	 * Insert "child" after "parent"
999 	 */
1000 	wcl = &parent->w_children;
1001 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1002 		wcl = &(*wcl)->wcl_next;
1003 	if (*wcl == NULL) {
1004 		*wcl = witness_child_get();
1005 		if (*wcl == NULL)
1006 			return (1);
1007 	}
1008 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1009 
1010 	/*
1011 	 * Now prune whole tree.  We look for cases where a lock is now
1012 	 * both a descendant and a direct child of a given lock.  In that
1013 	 * case, we want to remove the direct child link from the tree.
1014 	 */
1015 	if (recursed)
1016 		return (0);
1017 	recursed = 1;
1018 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1019 		list = &w_sleep;
1020 	else
1021 		list = &w_spin;
1022 	STAILQ_FOREACH(child, list, w_typelist) {
1023 		STAILQ_FOREACH(parent, list, w_typelist) {
1024 			if (!isitmychild(parent, child))
1025 				continue;
1026 			removechild(parent, child);
1027 			if (isitmydescendant(parent, child))
1028 				continue;
1029 			itismychild(parent, child);
1030 		}
1031 	}
1032 	recursed = 0;
1033 	witness_levelall();
1034 	return (0);
1035 }
1036 
1037 static void
1038 removechild(struct witness *parent, struct witness *child)
1039 {
1040 	struct witness_child_list_entry **wcl, *wcl1;
1041 	int i;
1042 
1043 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1044 		for (i = 0; i < (*wcl)->wcl_count; i++)
1045 			if ((*wcl)->wcl_children[i] == child)
1046 				goto found;
1047 	return;
1048 found:
1049 	(*wcl)->wcl_count--;
1050 	if ((*wcl)->wcl_count > i)
1051 		(*wcl)->wcl_children[i] =
1052 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1053 	MPASS((*wcl)->wcl_children[i] != NULL);
1054 	if ((*wcl)->wcl_count != 0)
1055 		return;
1056 	wcl1 = *wcl;
1057 	*wcl = wcl1->wcl_next;
1058 	witness_child_free(wcl1);
1059 }
1060 
1061 static int
1062 isitmychild(struct witness *parent, struct witness *child)
1063 {
1064 	struct witness_child_list_entry *wcl;
1065 	int i;
1066 
1067 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1068 		for (i = 0; i < wcl->wcl_count; i++) {
1069 			if (wcl->wcl_children[i] == child)
1070 				return (1);
1071 		}
1072 	}
1073 	return (0);
1074 }
1075 
1076 static int
1077 isitmydescendant(struct witness *parent, struct witness *child)
1078 {
1079 	struct witness_child_list_entry *wcl;
1080 	int i, j;
1081 
1082 	if (isitmychild(parent, child))
1083 		return (1);
1084 	j = 0;
1085 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1086 		MPASS(j < 1000);
1087 		for (i = 0; i < wcl->wcl_count; i++) {
1088 			if (isitmydescendant(wcl->wcl_children[i], child))
1089 				return (1);
1090 		}
1091 		j++;
1092 	}
1093 	return (0);
1094 }
1095 
1096 void
1097 witness_levelall (void)
1098 {
1099 	struct witness_list *list;
1100 	struct witness *w, *w1;
1101 
1102 	/*
1103 	 * First clear all levels.
1104 	 */
1105 	STAILQ_FOREACH(w, &w_all, w_list) {
1106 		w->w_level = 0;
1107 	}
1108 
1109 	/*
1110 	 * Look for locks with no parent and level all their descendants.
1111 	 */
1112 	STAILQ_FOREACH(w, &w_all, w_list) {
1113 		/*
1114 		 * This is just an optimization, technically we could get
1115 		 * away just walking the all list each time.
1116 		 */
1117 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1118 			list = &w_sleep;
1119 		else
1120 			list = &w_spin;
1121 		STAILQ_FOREACH(w1, list, w_typelist) {
1122 			if (isitmychild(w1, w))
1123 				goto skip;
1124 		}
1125 		witness_leveldescendents(w, 0);
1126 	skip:
1127 		;	/* silence GCC 3.x */
1128 	}
1129 }
1130 
1131 static void
1132 witness_leveldescendents(struct witness *parent, int level)
1133 {
1134 	struct witness_child_list_entry *wcl;
1135 	int i;
1136 
1137 	if (parent->w_level < level)
1138 		parent->w_level = level;
1139 	level++;
1140 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1141 		for (i = 0; i < wcl->wcl_count; i++)
1142 			witness_leveldescendents(wcl->wcl_children[i], level);
1143 }
1144 
1145 static void
1146 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1147 			   struct witness *parent)
1148 {
1149 	struct witness_child_list_entry *wcl;
1150 	int i, level;
1151 
1152 	level = parent->w_level;
1153 	prnt("%-2d", level);
1154 	for (i = 0; i < level; i++)
1155 		prnt(" ");
1156 	if (parent->w_refcount > 0) {
1157 		prnt("%s", parent->w_name);
1158 		if (parent->w_file != NULL)
1159 			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1160 			    parent->w_line);
1161 	} else
1162 		prnt("(dead)\n");
1163 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1164 		for (i = 0; i < wcl->wcl_count; i++)
1165 			    witness_displaydescendants(prnt,
1166 				wcl->wcl_children[i]);
1167 }
1168 
1169 static int
1170 blessed(struct witness *w1, struct witness *w2)
1171 {
1172 	int i;
1173 	struct witness_blessed *b;
1174 
1175 	for (i = 0; i < blessed_count; i++) {
1176 		b = &blessed_list[i];
1177 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1178 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1179 				return (1);
1180 			continue;
1181 		}
1182 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1183 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1184 				return (1);
1185 	}
1186 	return (0);
1187 }
1188 
1189 static struct witness *
1190 witness_get(void)
1191 {
1192 	struct witness *w;
1193 
1194 	if (witness_dead) {
1195 		mtx_unlock_spin(&w_mtx);
1196 		return (NULL);
1197 	}
1198 	if (STAILQ_EMPTY(&w_free)) {
1199 		witness_dead = 1;
1200 		mtx_unlock_spin(&w_mtx);
1201 		printf("%s: witness exhausted\n", __func__);
1202 		return (NULL);
1203 	}
1204 	w = STAILQ_FIRST(&w_free);
1205 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1206 	bzero(w, sizeof(*w));
1207 	return (w);
1208 }
1209 
1210 static void
1211 witness_free(struct witness *w)
1212 {
1213 
1214 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1215 }
1216 
1217 static struct witness_child_list_entry *
1218 witness_child_get(void)
1219 {
1220 	struct witness_child_list_entry *wcl;
1221 
1222 	if (witness_dead) {
1223 		mtx_unlock_spin(&w_mtx);
1224 		return (NULL);
1225 	}
1226 	wcl = w_child_free;
1227 	if (wcl == NULL) {
1228 		witness_dead = 1;
1229 		mtx_unlock_spin(&w_mtx);
1230 		printf("%s: witness exhausted\n", __func__);
1231 		return (NULL);
1232 	}
1233 	w_child_free = wcl->wcl_next;
1234 	bzero(wcl, sizeof(*wcl));
1235 	return (wcl);
1236 }
1237 
1238 static void
1239 witness_child_free(struct witness_child_list_entry *wcl)
1240 {
1241 
1242 	wcl->wcl_next = w_child_free;
1243 	w_child_free = wcl;
1244 }
1245 
1246 static struct lock_list_entry *
1247 witness_lock_list_get(void)
1248 {
1249 	struct lock_list_entry *lle;
1250 
1251 	if (witness_dead)
1252 		return (NULL);
1253 	mtx_lock_spin(&w_mtx);
1254 	lle = w_lock_list_free;
1255 	if (lle == NULL) {
1256 		witness_dead = 1;
1257 		mtx_unlock_spin(&w_mtx);
1258 		printf("%s: witness exhausted\n", __func__);
1259 		return (NULL);
1260 	}
1261 	w_lock_list_free = lle->ll_next;
1262 	mtx_unlock_spin(&w_mtx);
1263 	bzero(lle, sizeof(*lle));
1264 	return (lle);
1265 }
1266 
1267 static void
1268 witness_lock_list_free(struct lock_list_entry *lle)
1269 {
1270 
1271 	mtx_lock_spin(&w_mtx);
1272 	lle->ll_next = w_lock_list_free;
1273 	w_lock_list_free = lle;
1274 	mtx_unlock_spin(&w_mtx);
1275 }
1276 
1277 static struct lock_instance *
1278 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1279 {
1280 	struct lock_list_entry *lle;
1281 	struct lock_instance *instance;
1282 	int i;
1283 
1284 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1285 		for (i = lle->ll_count - 1; i >= 0; i--) {
1286 			instance = &lle->ll_children[i];
1287 			if (instance->li_lock == lock)
1288 				return (instance);
1289 		}
1290 	return (NULL);
1291 }
1292 
1293 int
1294 witness_list_locks(struct lock_list_entry **lock_list)
1295 {
1296 	struct lock_list_entry *lle;
1297 	struct lock_instance *instance;
1298 	struct lock_object *lock;
1299 	int i, nheld;
1300 
1301 	nheld = 0;
1302 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1303 		for (i = lle->ll_count - 1; i >= 0; i--) {
1304 			instance = &lle->ll_children[i];
1305 			lock = instance->li_lock;
1306 			printf("%s %s %s",
1307 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1308 			    "exclusive" : "shared",
1309 			    lock->lo_class->lc_name, lock->lo_name);
1310 			if (lock->lo_type != lock->lo_name)
1311 				printf(" (%s)", lock->lo_type);
1312 			printf(" r = %d (%p) locked @ %s:%d\n",
1313 			    instance->li_flags & LI_RECURSEMASK, lock,
1314 			    instance->li_file, instance->li_line);
1315 			nheld++;
1316 		}
1317 	return (nheld);
1318 }
1319 
1320 /*
1321  * Calling this on td != curthread is bad unless we are in ddb.
1322  */
1323 int
1324 witness_list(struct thread *td)
1325 {
1326 	int nheld;
1327 
1328 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1329 #ifdef DDB
1330 	KASSERT(td == curthread || db_active,
1331 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1332 	if (!db_active && witness_dead)
1333 		return (0);
1334 #else
1335 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1336 	if (witness_dead)
1337 		return (0);
1338 #endif
1339 	nheld = witness_list_locks(&td->td_sleeplocks);
1340 
1341 	/*
1342 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1343 	 * if td is currently executing on some other CPU and holds spin locks
1344 	 * as we won't display those locks.  If we had a MI way of getting
1345 	 * the per-cpu data for a given cpu then we could use
1346 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1347 	 * and "fix" this.
1348 	 *
1349 	 * That still wouldn't really fix this unless we locked sched_lock
1350 	 * or stopped the other CPU to make sure it wasn't changing the list
1351 	 * out from under us.  It is probably best to just not try to handle
1352 	 * threads on other CPU's for now.
1353 	 */
1354 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1355 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1356 
1357 	return (nheld);
1358 }
1359 
1360 void
1361 witness_save(struct lock_object *lock, const char **filep, int *linep)
1362 {
1363 	struct lock_instance *instance;
1364 
1365 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1366 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1367 		return;
1368 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1369 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1370 		    lock->lo_class->lc_name, lock->lo_name);
1371 	instance = find_instance(curthread->td_sleeplocks, lock);
1372 	if (instance == NULL)
1373 		panic("%s: lock (%s) %s not locked", __func__,
1374 		    lock->lo_class->lc_name, lock->lo_name);
1375 	*filep = instance->li_file;
1376 	*linep = instance->li_line;
1377 }
1378 
1379 void
1380 witness_restore(struct lock_object *lock, const char *file, int line)
1381 {
1382 	struct lock_instance *instance;
1383 
1384 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1385 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1386 		return;
1387 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1388 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1389 		    lock->lo_class->lc_name, lock->lo_name);
1390 	instance = find_instance(curthread->td_sleeplocks, lock);
1391 	if (instance == NULL)
1392 		panic("%s: lock (%s) %s not locked", __func__,
1393 		    lock->lo_class->lc_name, lock->lo_name);
1394 	lock->lo_witness->w_file = file;
1395 	lock->lo_witness->w_line = line;
1396 	instance->li_file = file;
1397 	instance->li_line = line;
1398 }
1399 
1400 void
1401 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1402 {
1403 #ifdef INVARIANT_SUPPORT
1404 	struct lock_instance *instance;
1405 
1406 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1407 		return;
1408 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1409 		instance = find_instance(curthread->td_sleeplocks, lock);
1410 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1411 		instance = find_instance(PCPU_GET(spinlocks), lock);
1412 	else {
1413 		panic("Lock (%s) %s is not sleep or spin!",
1414 		    lock->lo_class->lc_name, lock->lo_name);
1415 		return;
1416 	}
1417 	switch (flags) {
1418 	case LA_UNLOCKED:
1419 		if (instance != NULL)
1420 			panic("Lock (%s) %s locked @ %s:%d.",
1421 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1422 		break;
1423 	case LA_LOCKED:
1424 	case LA_LOCKED | LA_RECURSED:
1425 	case LA_LOCKED | LA_NOTRECURSED:
1426 	case LA_SLOCKED:
1427 	case LA_SLOCKED | LA_RECURSED:
1428 	case LA_SLOCKED | LA_NOTRECURSED:
1429 	case LA_XLOCKED:
1430 	case LA_XLOCKED | LA_RECURSED:
1431 	case LA_XLOCKED | LA_NOTRECURSED:
1432 		if (instance == NULL) {
1433 			panic("Lock (%s) %s not locked @ %s:%d.",
1434 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1435 			break;
1436 		}
1437 		if ((flags & LA_XLOCKED) != 0 &&
1438 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1439 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1440 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1441 		if ((flags & LA_SLOCKED) != 0 &&
1442 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1443 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1444 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1445 		if ((flags & LA_RECURSED) != 0 &&
1446 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1447 			panic("Lock (%s) %s not recursed @ %s:%d.",
1448 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1449 		if ((flags & LA_NOTRECURSED) != 0 &&
1450 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1451 			panic("Lock (%s) %s recursed @ %s:%d.",
1452 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1453 		break;
1454 	default:
1455 		panic("Invalid lock assertion at %s:%d.", file, line);
1456 
1457 	}
1458 #endif	/* INVARIANT_SUPPORT */
1459 }
1460 
1461 #ifdef DDB
1462 
1463 DB_SHOW_COMMAND(locks, db_witness_list)
1464 {
1465 	struct thread *td;
1466 	pid_t pid;
1467 	struct proc *p;
1468 
1469 	if (have_addr) {
1470 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1471 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1472 		    ((addr >> 16) % 16) * 10000;
1473 		/* sx_slock(&allproc_lock); */
1474 		FOREACH_PROC_IN_SYSTEM(p) {
1475 			if (p->p_pid == pid)
1476 				break;
1477 		}
1478 		/* sx_sunlock(&allproc_lock); */
1479 		if (p == NULL) {
1480 			db_printf("pid %d not found\n", pid);
1481 			return;
1482 		}
1483 		FOREACH_THREAD_IN_PROC(p, td) {
1484 			witness_list(td);
1485 		}
1486 	} else {
1487 		td = curthread;
1488 		witness_list(td);
1489 	}
1490 }
1491 
1492 DB_SHOW_COMMAND(witness, db_witness_display)
1493 {
1494 
1495 	witness_display(db_printf);
1496 }
1497 #endif
1498