xref: /freebsd/sys/kern/subr_witness.c (revision eacee0ff7ec955b32e09515246bd97b6edcd2b0f)
1 /*-
2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
13  *    promote products derived from this software without specific prior
14  *    written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30  * $FreeBSD$
31  */
32 
33 /*
34  * Implementation of the `witness' lock verifier.  Originally implemented for
35  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36  * classes in FreeBSD.
37  */
38 
39 /*
40  *	Main Entry: witness
41  *	Pronunciation: 'wit-n&s
42  *	Function: noun
43  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44  *	    testimony, witness, from 2wit
45  *	Date: before 12th century
46  *	1 : attestation of a fact or event : TESTIMONY
47  *	2 : one that gives evidence; specifically : one who testifies in
48  *	    a cause or before a judicial tribunal
49  *	3 : one asked to be present at a transaction so as to be able to
50  *	    testify to its having taken place
51  *	4 : one who has personal knowledge of something
52  *	5 a : something serving as evidence or proof : SIGN
53  *	  b : public affirmation by word or example of usually
54  *	      religious faith or conviction <the heroic witness to divine
55  *	      life -- Pilot>
56  *	6 capitalized : a member of the Jehovah's Witnesses
57  */
58 
59 #include "opt_ddb.h"
60 #include "opt_witness.h"
61 
62 #include <sys/param.h>
63 #include <sys/bus.h>
64 #include <sys/kernel.h>
65 #include <sys/ktr.h>
66 #include <sys/lock.h>
67 #include <sys/malloc.h>
68 #include <sys/mutex.h>
69 #include <sys/proc.h>
70 #include <sys/sysctl.h>
71 #include <sys/systm.h>
72 
73 #include <ddb/ddb.h>
74 
75 #define WITNESS_COUNT 200
76 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77 /*
78  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80  * probably be safe for the most part, but it's still a SWAG.
81  */
82 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83 
84 #define	WITNESS_NCHILDREN 6
85 
86 struct witness_child_list_entry;
87 
88 struct witness {
89 	const	char *w_name;
90 	struct	lock_class *w_class;
91 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94 	const	char *w_file;
95 	int	w_line;
96 	u_int	w_level;
97 	u_int	w_refcount;
98 	u_char	w_Giant_squawked:1;
99 	u_char	w_other_squawked:1;
100 	u_char	w_same_squawked:1;
101 };
102 
103 struct witness_child_list_entry {
104 	struct	witness_child_list_entry *wcl_next;
105 	struct	witness *wcl_children[WITNESS_NCHILDREN];
106 	u_int	wcl_count;
107 };
108 
109 STAILQ_HEAD(witness_list, witness);
110 
111 struct witness_blessed {
112 	const	char *b_lock1;
113 	const	char *b_lock2;
114 };
115 
116 struct witness_order_list_entry {
117 	const	char *w_name;
118 	struct	lock_class *w_class;
119 };
120 
121 static struct	witness *enroll(const char *description,
122 				struct lock_class *lock_class);
123 static int	itismychild(struct witness *parent, struct witness *child);
124 static void	removechild(struct witness *parent, struct witness *child);
125 static int	isitmychild(struct witness *parent, struct witness *child);
126 static int	isitmydescendant(struct witness *parent, struct witness *child);
127 static int	dup_ok(struct witness *);
128 static int	blessed(struct witness *, struct witness *);
129 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
130 				     struct witness_list *list);
131 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
132 					   struct witness *);
133 static void	witness_leveldescendents(struct witness *parent, int level);
134 static void	witness_levelall(void);
135 static struct	witness *witness_get(void);
136 static void	witness_free(struct witness *m);
137 static struct	witness_child_list_entry *witness_child_get(void);
138 static void	witness_child_free(struct witness_child_list_entry *wcl);
139 static struct	lock_list_entry *witness_lock_list_get(void);
140 static void	witness_lock_list_free(struct lock_list_entry *lle);
141 static void	witness_display(void(*)(const char *fmt, ...));
142 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
143 					     struct lock_object *lock);
144 
145 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146 
147 static int witness_watch = 1;
148 TUNABLE_INT("debug.witness_watch", &witness_watch);
149 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
150 
151 #ifdef DDB
152 /*
153  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
154  * drop into kdebug() when:
155  *	- a lock heirarchy violation occurs
156  *	- locks are held when going to sleep.
157  */
158 #ifdef WITNESS_DDB
159 int	witness_ddb = 1;
160 #else
161 int	witness_ddb = 0;
162 #endif
163 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
164 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
165 #endif /* DDB */
166 
167 #ifdef WITNESS_SKIPSPIN
168 int	witness_skipspin = 1;
169 #else
170 int	witness_skipspin = 0;
171 #endif
172 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
173 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
174     "");
175 
176 static struct mtx w_mtx;
177 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
178 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
179 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
180 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
181 static struct witness_child_list_entry *w_child_free = NULL;
182 static struct lock_list_entry *w_lock_list_free = NULL;
183 static int witness_dead;	/* fatal error, probably no memory */
184 
185 static struct witness w_data[WITNESS_COUNT];
186 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
187 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188 
189 static struct witness_order_list_entry order_lists[] = {
190 	{ "Giant", &lock_class_mtx_sleep },
191 	{ "proctree", &lock_class_sx },
192 	{ "allproc", &lock_class_sx },
193 	{ "process lock", &lock_class_mtx_sleep },
194 	{ "uidinfo hash", &lock_class_mtx_sleep },
195 	{ "uidinfo struct", &lock_class_mtx_sleep },
196 	{ NULL, NULL },
197 	/*
198 	 * spin locks
199 	 */
200 #ifdef SMP
201 	{ "ap boot", &lock_class_mtx_spin },
202 #ifdef __i386__
203 	{ "com", &lock_class_mtx_spin },
204 #endif
205 #endif
206 	{ "sio", &lock_class_mtx_spin },
207 #ifdef __i386__
208 	{ "cy", &lock_class_mtx_spin },
209 #endif
210 	{ "ng_node", &lock_class_mtx_spin },
211 	{ "ng_worklist", &lock_class_mtx_spin },
212 	{ "ithread table lock", &lock_class_mtx_spin },
213 	{ "sched lock", &lock_class_mtx_spin },
214 	{ "callout", &lock_class_mtx_spin },
215 	/*
216 	 * leaf locks
217 	 */
218 	{ "allpmaps", &lock_class_mtx_spin },
219 	{ "icu", &lock_class_mtx_spin },
220 #ifdef SMP
221 	{ "smp rendezvous", &lock_class_mtx_spin },
222 #endif
223 	{ "clk", &lock_class_mtx_spin },
224 	{ NULL, NULL },
225 	{ NULL, NULL }
226 };
227 
228 static const char *dup_list[] = {
229 	"process lock",
230 	NULL
231 };
232 
233 /*
234  * Pairs of locks which have been blessed
235  * Don't complain about order problems with blessed locks
236  */
237 static struct witness_blessed blessed_list[] = {
238 };
239 static int blessed_count =
240 	sizeof(blessed_list) / sizeof(struct witness_blessed);
241 
242 /*
243  * List of all locks in the system.
244  */
245 STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
246 
247 static struct mtx all_mtx = {
248 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
249 	  "All locks list",		/* mtx_object.lo_name */
250 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
251 	  { NULL },			/* mtx_object.lo_list */
252 	  NULL },			/* mtx_object.lo_witness */
253 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
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 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
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 			CTR2(KTR_WITNESS,
386 			    "%s: marking witness %s as dead", __func__, 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 thread *td;
467 	int i, j;
468 #ifdef DDB
469 	int go_into_ddb = 0;
470 #endif /* DDB */
471 
472 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
473 	    panicstr != NULL)
474 		return;
475 	w = lock->lo_witness;
476 	class = lock->lo_class;
477 	td = curthread;
478 
479 	if (class->lc_flags & LC_SLEEPLOCK) {
480 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
481 			panic("blockable sleep lock (%s) %s @ %s:%d",
482 			    class->lc_name, lock->lo_name, file, line);
483 		lock_list = &td->td_sleeplocks;
484 	} else
485 		lock_list = PCPU_PTR(spinlocks);
486 
487 	/*
488 	 * Try locks do not block if they fail to acquire the lock, thus
489 	 * there is no danger of deadlocks or of switching while holding a
490 	 * spin lock if we acquire a lock via a try operation.
491 	 */
492 	if (flags & LOP_TRYLOCK)
493 		goto out;
494 
495 	/*
496 	 * Is this the first lock acquired?  If so, then no order checking
497 	 * is needed.
498 	 */
499 	if (*lock_list == NULL)
500 		goto out;
501 
502 	/*
503 	 * Check to see if we are recursing on a lock we already own.
504 	 */
505 	lock1 = find_instance(*lock_list, lock);
506 	if (lock1 != NULL) {
507 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
508 		    (flags & LOP_EXCLUSIVE) == 0) {
509 			printf("shared lock of (%s) %s @ %s:%d\n",
510 			    class->lc_name, lock->lo_name, file, line);
511 			printf("while exclusively locked from %s:%d\n",
512 			    lock1->li_file, lock1->li_line);
513 			panic("share->excl");
514 		}
515 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
516 		    (flags & LOP_EXCLUSIVE) != 0) {
517 			printf("exclusive lock of (%s) %s @ %s:%d\n",
518 			    class->lc_name, lock->lo_name, file, line);
519 			printf("while share locked from %s:%d\n",
520 			    lock1->li_file, lock1->li_line);
521 			panic("excl->share");
522 		}
523 		lock1->li_flags++;
524 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
525 			printf(
526 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
527 			    class->lc_name, lock->lo_name, file, line);
528 			printf("first acquired @ %s:%d\n", lock1->li_file,
529 			    lock1->li_line);
530 			panic("recurse");
531 		}
532 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
533 		    td->td_proc->p_pid, lock->lo_name,
534 		    lock1->li_flags & LI_RECURSEMASK);
535 		lock1->li_file = file;
536 		lock1->li_line = line;
537 		return;
538 	}
539 
540 	/*
541 	 * Check for duplicate locks of the same type.  Note that we only
542 	 * have to check for this on the last lock we just acquired.  Any
543 	 * other cases will be caught as lock order violations.
544 	 */
545 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
546 	w1 = lock1->li_lock->lo_witness;
547 	if (w1 == w) {
548 		if (w->w_same_squawked || dup_ok(w))
549 			goto out;
550 		w->w_same_squawked = 1;
551 		printf("acquiring duplicate lock of same type: \"%s\"\n",
552 			lock->lo_name);
553 		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
554 		printf(" 2nd @ %s:%d\n", file, line);
555 #ifdef DDB
556 		go_into_ddb = 1;
557 #endif /* DDB */
558 		goto out;
559 	}
560 	MPASS(!mtx_owned(&w_mtx));
561 	mtx_lock_spin(&w_mtx);
562 	/*
563 	 * If we have a known higher number just say ok
564 	 */
565 	if (witness_watch > 1 && w->w_level > w1->w_level) {
566 		mtx_unlock_spin(&w_mtx);
567 		goto out;
568 	}
569 	if (isitmydescendant(w1, w)) {
570 		mtx_unlock_spin(&w_mtx);
571 		goto out;
572 	}
573 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
574 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
575 
576 			MPASS(j < WITNESS_COUNT);
577 			lock1 = &lle->ll_children[i];
578 			w1 = lock1->li_lock->lo_witness;
579 
580 			/*
581 			 * If this lock doesn't undergo witness checking,
582 			 * then skip it.
583 			 */
584 			if (w1 == NULL) {
585 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
586 				    ("lock missing witness structure"));
587 				continue;
588 			}
589 			/*
590 			 * If we are locking Giant and we slept with this
591 			 * lock, then skip it.
592 			 */
593 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
594 			    lock == &Giant.mtx_object)
595 				continue;
596 			if (!isitmydescendant(w, w1))
597 				continue;
598 			/*
599 			 * We have a lock order violation, check to see if it
600 			 * is allowed or has already been yelled about.
601 			 */
602 			mtx_unlock_spin(&w_mtx);
603 			if (blessed(w, w1))
604 				goto out;
605 			if (lock1->li_lock == &Giant.mtx_object) {
606 				if (w1->w_Giant_squawked)
607 					goto out;
608 				else
609 					w1->w_Giant_squawked = 1;
610 			} else {
611 				if (w1->w_other_squawked)
612 					goto out;
613 				else
614 					w1->w_other_squawked = 1;
615 			}
616 			/*
617 			 * Ok, yell about it.
618 			 */
619 			printf("lock order reversal\n");
620 			/*
621 			 * Try to locate an earlier lock with
622 			 * witness w in our list.
623 			 */
624 			do {
625 				lock2 = &lle->ll_children[i];
626 				MPASS(lock2->li_lock != NULL);
627 				if (lock2->li_lock->lo_witness == w)
628 					break;
629 				i--;
630 				if (i == 0 && lle->ll_next != NULL) {
631 					lle = lle->ll_next;
632 					i = lle->ll_count - 1;
633 					MPASS(i != 0);
634 				}
635 			} while (i >= 0);
636 			if (i < 0) {
637 				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
638 				    lock1->li_lock->lo_name, lock1->li_file,
639 				    lock1->li_line);
640 				printf(" 2nd %p %s @ %s:%d\n", lock,
641 				    lock->lo_name, file, line);
642 			} else {
643 				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
644 				    lock2->li_lock->lo_name, lock2->li_file,
645 				    lock2->li_line);
646 				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
647 				    lock1->li_lock->lo_name, lock1->li_file,
648 				    lock1->li_line);
649 				printf(" 3rd %p %s @ %s:%d\n", lock,
650 				    lock->lo_name, file, line);
651 			}
652 #ifdef DDB
653 			go_into_ddb = 1;
654 #endif /* DDB */
655 			goto out;
656 		}
657 	}
658 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
659 	/*
660 	 * Don't build a new relationship if we are locking Giant just
661 	 * after waking up and the previous lock in the list was acquired
662 	 * prior to blocking.
663 	 */
664 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
665 		mtx_unlock_spin(&w_mtx);
666 	else {
667 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
668 		    lock->lo_name, lock1->li_lock->lo_name);
669 		if (!itismychild(lock1->li_lock->lo_witness, w))
670 			mtx_unlock_spin(&w_mtx);
671 	}
672 
673 out:
674 #ifdef DDB
675 	if (witness_ddb && go_into_ddb)
676 		Debugger(__func__);
677 #endif /* DDB */
678 	w->w_file = file;
679 	w->w_line = line;
680 
681 	lle = *lock_list;
682 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
683 		lle = witness_lock_list_get();
684 		if (lle == NULL)
685 			return;
686 		lle->ll_next = *lock_list;
687 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
688 		    td->td_proc->p_pid, lle);
689 		*lock_list = lle;
690 	}
691 	lock1 = &lle->ll_children[lle->ll_count++];
692 	lock1->li_lock = lock;
693 	lock1->li_line = line;
694 	lock1->li_file = file;
695 	if ((flags & LOP_EXCLUSIVE) != 0)
696 		lock1->li_flags = LI_EXCLUSIVE;
697 	else
698 		lock1->li_flags = 0;
699 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
700 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
701 }
702 
703 void
704 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
705 {
706 	struct lock_instance *instance;
707 	struct lock_class *class;
708 
709 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
710 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
711 		return;
712 	class = lock->lo_class;
713 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
714 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
715 		    class->lc_name, lock->lo_name, file, line);
716 	if ((flags & LOP_TRYLOCK) == 0)
717 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
718 		    lock->lo_name, file, line);
719 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
720 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
721 		    class->lc_name, lock->lo_name, file, line);
722 	instance = find_instance(curthread->td_sleeplocks, lock);
723 	if (instance == NULL)
724 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
725 		    class->lc_name, lock->lo_name, file, line);
726 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
727 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
728 		    class->lc_name, lock->lo_name, file, line);
729 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
730 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
731 		    class->lc_name, lock->lo_name,
732 		    instance->li_flags & LI_RECURSEMASK, file, line);
733 	instance->li_flags |= LI_EXCLUSIVE;
734 }
735 
736 void
737 witness_downgrade(struct lock_object *lock, int flags, const char *file,
738     int line)
739 {
740 	struct lock_instance *instance;
741 	struct lock_class *class;
742 
743 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
744 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
745 		return;
746 	class = lock->lo_class;
747 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
748 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
749 		    class->lc_name, lock->lo_name, file, line);
750 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
751 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
752 		    class->lc_name, lock->lo_name, file, line);
753 	instance = find_instance(curthread->td_sleeplocks, lock);
754 	if (instance == NULL)
755 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
756 		    class->lc_name, lock->lo_name, file, line);
757 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
758 		panic("downgrade of shared lock (%s) %s @ %s:%d",
759 		    class->lc_name, lock->lo_name, file, line);
760 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
761 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
762 		    class->lc_name, lock->lo_name,
763 		    instance->li_flags & LI_RECURSEMASK, file, line);
764 	instance->li_flags &= ~LI_EXCLUSIVE;
765 }
766 
767 void
768 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
769 {
770 	struct lock_list_entry **lock_list, *lle;
771 	struct lock_instance *instance;
772 	struct lock_class *class;
773 	struct thread *td;
774 	critical_t s;
775 	int i, j;
776 
777 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
778 	    panicstr != NULL)
779 		return;
780 	td = curthread;
781 	class = lock->lo_class;
782 	if (class->lc_flags & LC_SLEEPLOCK)
783 		lock_list = &td->td_sleeplocks;
784 	else
785 		lock_list = PCPU_PTR(spinlocks);
786 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
787 		for (i = 0; i < (*lock_list)->ll_count; i++) {
788 			instance = &(*lock_list)->ll_children[i];
789 			if (instance->li_lock == lock) {
790 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
791 				    (flags & LOP_EXCLUSIVE) == 0) {
792 					printf(
793 					"shared unlock of (%s) %s @ %s:%d\n",
794 					    class->lc_name, lock->lo_name,
795 					    file, line);
796 					printf(
797 					"while exclusively locked from %s:%d\n",
798 					    instance->li_file,
799 					    instance->li_line);
800 					panic("excl->ushare");
801 				}
802 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
803 				    (flags & LOP_EXCLUSIVE) != 0) {
804 					printf(
805 					"exclusive unlock of (%s) %s @ %s:%d\n",
806 					    class->lc_name, lock->lo_name,
807 					    file, line);
808 					printf(
809 					"while share locked from %s:%d\n",
810 					    instance->li_file,
811 					    instance->li_line);
812 					panic("share->uexcl");
813 				}
814 				/* If we are recursed, unrecurse. */
815 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
816 					CTR4(KTR_WITNESS,
817 				    "%s: pid %d unrecursed on %s r=%d", __func__,
818 					    td->td_proc->p_pid,
819 					    instance->li_lock->lo_name,
820 					    instance->li_flags);
821 					instance->li_flags--;
822 					return;
823 				}
824 				s = cpu_critical_enter();
825 				CTR4(KTR_WITNESS,
826 				    "%s: pid %d removed %s from lle[%d]", __func__,
827 				    td->td_proc->p_pid,
828 				    instance->li_lock->lo_name,
829 				    (*lock_list)->ll_count - 1);
830 				(*lock_list)->ll_count--;
831 				for (j = i; j < (*lock_list)->ll_count; j++)
832 					(*lock_list)->ll_children[j] =
833 					    (*lock_list)->ll_children[j + 1];
834 				cpu_critical_exit(s);
835 				if ((*lock_list)->ll_count == 0) {
836 					lle = *lock_list;
837 					*lock_list = lle->ll_next;
838 					CTR3(KTR_WITNESS,
839 					    "%s: pid %d removed lle %p", __func__,
840 					    td->td_proc->p_pid, lle);
841 					witness_lock_list_free(lle);
842 				}
843 				return;
844 			}
845 		}
846 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
847 	    file, line);
848 }
849 
850 /*
851  * Warn if any held locks are not sleepable.  Note that Giant and the lock
852  * passed in are both special cases since they are both released during the
853  * sleep process and aren't actually held while the thread is asleep.
854  */
855 int
856 witness_sleep(int check_only, struct lock_object *lock, const char *file,
857 	      int line)
858 {
859 	struct lock_list_entry **lock_list, *lle;
860 	struct lock_instance *lock1;
861 	struct thread *td;
862 	critical_t savecrit;
863 	int i, n;
864 
865 	if (witness_dead || panicstr != NULL)
866 		return (0);
867 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
868 	n = 0;
869 	/*
870 	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
871 	 */
872 	savecrit = cpu_critical_enter();
873 	td = curthread;
874 	lock_list = &td->td_sleeplocks;
875 again:
876 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
877 		for (i = lle->ll_count - 1; i >= 0; i--) {
878 			lock1 = &lle->ll_children[i];
879 			if (lock1->li_lock == lock ||
880 			    lock1->li_lock == &Giant.mtx_object)
881 				continue;
882 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
883 				if (check_only == 0) {
884 					CTR3(KTR_WITNESS,
885 				    "pid %d: sleeping with lock (%s) %s held",
886 					    td->td_proc->p_pid,
887 					    lock1->li_lock->lo_class->lc_name,
888 					    lock1->li_lock->lo_name);
889 					lock1->li_flags |= LI_SLEPT;
890 				}
891 				continue;
892 			}
893 			n++;
894 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
895 			    file, line, check_only ? "could sleep" : "sleeping",
896 			    lock1->li_lock->lo_name, lock1->li_file,
897 			    lock1->li_line);
898 		}
899 	if (lock_list == &td->td_sleeplocks) {
900 		lock_list = PCPU_PTR(spinlocks);
901 		goto again;
902 	}
903 #ifdef DDB
904 	if (witness_ddb && n)
905 		Debugger(__func__);
906 #endif /* DDB */
907 	cpu_critical_exit(savecrit);
908 	return (n);
909 }
910 
911 static struct witness *
912 enroll(const char *description, struct lock_class *lock_class)
913 {
914 	struct witness *w;
915 
916 	if (!witness_watch || witness_dead || panicstr != NULL)
917 		return (NULL);
918 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
919 		return (NULL);
920 	mtx_lock_spin(&w_mtx);
921 	STAILQ_FOREACH(w, &w_all, w_list) {
922 		if (strcmp(description, w->w_name) == 0) {
923 			w->w_refcount++;
924 			mtx_unlock_spin(&w_mtx);
925 			if (lock_class != w->w_class)
926 				panic(
927 				"lock (%s) %s does not match earlier (%s) lock",
928 				    description, lock_class->lc_name,
929 				    w->w_class->lc_name);
930 			return (w);
931 		}
932 	}
933 	/*
934 	 * This isn't quite right, as witness_cold is still 0 while we
935 	 * enroll all the locks initialized before witness_initialize().
936 	 */
937 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
938 		mtx_unlock_spin(&w_mtx);
939 		panic("spin lock %s not in order list", description);
940 	}
941 	if ((w = witness_get()) == NULL)
942 		return (NULL);
943 	w->w_name = description;
944 	w->w_class = lock_class;
945 	w->w_refcount = 1;
946 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
947 	if (lock_class->lc_flags & LC_SPINLOCK)
948 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
949 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
950 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
951 	else {
952 		mtx_unlock_spin(&w_mtx);
953 		panic("lock class %s is not sleep or spin",
954 		    lock_class->lc_name);
955 	}
956 	mtx_unlock_spin(&w_mtx);
957 	return (w);
958 }
959 
960 static int
961 itismychild(struct witness *parent, struct witness *child)
962 {
963 	static int recursed;
964 	struct witness_child_list_entry **wcl;
965 	struct witness_list *list;
966 
967 	MPASS(child != NULL && parent != NULL);
968 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
969 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
970 		panic(
971 		"%s: parent (%s) and child (%s) are not the same lock type",
972 		    __func__, parent->w_class->lc_name,
973 		    child->w_class->lc_name);
974 
975 	/*
976 	 * Insert "child" after "parent"
977 	 */
978 	wcl = &parent->w_children;
979 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
980 		wcl = &(*wcl)->wcl_next;
981 	if (*wcl == NULL) {
982 		*wcl = witness_child_get();
983 		if (*wcl == NULL)
984 			return (1);
985 	}
986 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
987 
988 	/*
989 	 * Now prune whole tree.  We look for cases where a lock is now
990 	 * both a descendant and a direct child of a given lock.  In that
991 	 * case, we want to remove the direct child link from the tree.
992 	 */
993 	if (recursed)
994 		return (0);
995 	recursed = 1;
996 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
997 		list = &w_sleep;
998 	else
999 		list = &w_spin;
1000 	STAILQ_FOREACH(child, list, w_typelist) {
1001 		STAILQ_FOREACH(parent, list, w_typelist) {
1002 			if (!isitmychild(parent, child))
1003 				continue;
1004 			removechild(parent, child);
1005 			if (isitmydescendant(parent, child))
1006 				continue;
1007 			itismychild(parent, child);
1008 		}
1009 	}
1010 	recursed = 0;
1011 	witness_levelall();
1012 	return (0);
1013 }
1014 
1015 static void
1016 removechild(struct witness *parent, struct witness *child)
1017 {
1018 	struct witness_child_list_entry **wcl, *wcl1;
1019 	int i;
1020 
1021 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1022 		for (i = 0; i < (*wcl)->wcl_count; i++)
1023 			if ((*wcl)->wcl_children[i] == child)
1024 				goto found;
1025 	return;
1026 found:
1027 	(*wcl)->wcl_count--;
1028 	if ((*wcl)->wcl_count > i)
1029 		(*wcl)->wcl_children[i] =
1030 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1031 	MPASS((*wcl)->wcl_children[i] != NULL);
1032 	if ((*wcl)->wcl_count != 0)
1033 		return;
1034 	wcl1 = *wcl;
1035 	*wcl = wcl1->wcl_next;
1036 	witness_child_free(wcl1);
1037 }
1038 
1039 static int
1040 isitmychild(struct witness *parent, struct witness *child)
1041 {
1042 	struct witness_child_list_entry *wcl;
1043 	int i;
1044 
1045 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1046 		for (i = 0; i < wcl->wcl_count; i++) {
1047 			if (wcl->wcl_children[i] == child)
1048 				return (1);
1049 		}
1050 	}
1051 	return (0);
1052 }
1053 
1054 static int
1055 isitmydescendant(struct witness *parent, struct witness *child)
1056 {
1057 	struct witness_child_list_entry *wcl;
1058 	int i, j;
1059 
1060 	if (isitmychild(parent, child))
1061 		return (1);
1062 	j = 0;
1063 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1064 		MPASS(j < 1000);
1065 		for (i = 0; i < wcl->wcl_count; i++) {
1066 			if (isitmydescendant(wcl->wcl_children[i], child))
1067 				return (1);
1068 		}
1069 		j++;
1070 	}
1071 	return (0);
1072 }
1073 
1074 void
1075 witness_levelall (void)
1076 {
1077 	struct witness_list *list;
1078 	struct witness *w, *w1;
1079 
1080 	/*
1081 	 * First clear all levels.
1082 	 */
1083 	STAILQ_FOREACH(w, &w_all, w_list) {
1084 		w->w_level = 0;
1085 	}
1086 
1087 	/*
1088 	 * Look for locks with no parent and level all their descendants.
1089 	 */
1090 	STAILQ_FOREACH(w, &w_all, w_list) {
1091 		/*
1092 		 * This is just an optimization, technically we could get
1093 		 * away just walking the all list each time.
1094 		 */
1095 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1096 			list = &w_sleep;
1097 		else
1098 			list = &w_spin;
1099 		STAILQ_FOREACH(w1, list, w_typelist) {
1100 			if (isitmychild(w1, w))
1101 				goto skip;
1102 		}
1103 		witness_leveldescendents(w, 0);
1104 	skip:
1105 	}
1106 }
1107 
1108 static void
1109 witness_leveldescendents(struct witness *parent, int level)
1110 {
1111 	struct witness_child_list_entry *wcl;
1112 	int i;
1113 
1114 	if (parent->w_level < level)
1115 		parent->w_level = level;
1116 	level++;
1117 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1118 		for (i = 0; i < wcl->wcl_count; i++)
1119 			witness_leveldescendents(wcl->wcl_children[i], level);
1120 }
1121 
1122 static void
1123 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1124 			   struct witness *parent)
1125 {
1126 	struct witness_child_list_entry *wcl;
1127 	int i, level;
1128 
1129 	level =  parent->w_level;
1130 	prnt("%-2d", level);
1131 	for (i = 0; i < level; i++)
1132 		prnt(" ");
1133 	prnt("%s", parent->w_name);
1134 	if (parent->w_file != NULL)
1135 		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1136 		    parent->w_line);
1137 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1138 		for (i = 0; i < wcl->wcl_count; i++)
1139 			    witness_displaydescendants(prnt,
1140 				wcl->wcl_children[i]);
1141 }
1142 
1143 static int
1144 dup_ok(struct witness *w)
1145 {
1146 	const char **dup;
1147 
1148 	for (dup = dup_list; *dup != NULL; dup++)
1149 		if (strcmp(w->w_name, *dup) == 0)
1150 			return (1);
1151 	return (0);
1152 }
1153 
1154 static int
1155 blessed(struct witness *w1, struct witness *w2)
1156 {
1157 	int i;
1158 	struct witness_blessed *b;
1159 
1160 	for (i = 0; i < blessed_count; i++) {
1161 		b = &blessed_list[i];
1162 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1163 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1164 				return (1);
1165 			continue;
1166 		}
1167 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1168 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1169 				return (1);
1170 	}
1171 	return (0);
1172 }
1173 
1174 static struct witness *
1175 witness_get(void)
1176 {
1177 	struct witness *w;
1178 
1179 	if (witness_dead) {
1180 		mtx_unlock_spin(&w_mtx);
1181 		return (NULL);
1182 	}
1183 	if (STAILQ_EMPTY(&w_free)) {
1184 		witness_dead = 1;
1185 		mtx_unlock_spin(&w_mtx);
1186 		printf("%s: witness exhausted\n", __func__);
1187 		return (NULL);
1188 	}
1189 	w = STAILQ_FIRST(&w_free);
1190 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1191 	bzero(w, sizeof(*w));
1192 	return (w);
1193 }
1194 
1195 static void
1196 witness_free(struct witness *w)
1197 {
1198 
1199 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1200 }
1201 
1202 static struct witness_child_list_entry *
1203 witness_child_get(void)
1204 {
1205 	struct witness_child_list_entry *wcl;
1206 
1207 	if (witness_dead) {
1208 		mtx_unlock_spin(&w_mtx);
1209 		return (NULL);
1210 	}
1211 	wcl = w_child_free;
1212 	if (wcl == NULL) {
1213 		witness_dead = 1;
1214 		mtx_unlock_spin(&w_mtx);
1215 		printf("%s: witness exhausted\n", __func__);
1216 		return (NULL);
1217 	}
1218 	w_child_free = wcl->wcl_next;
1219 	bzero(wcl, sizeof(*wcl));
1220 	return (wcl);
1221 }
1222 
1223 static void
1224 witness_child_free(struct witness_child_list_entry *wcl)
1225 {
1226 
1227 	wcl->wcl_next = w_child_free;
1228 	w_child_free = wcl;
1229 }
1230 
1231 static struct lock_list_entry *
1232 witness_lock_list_get(void)
1233 {
1234 	struct lock_list_entry *lle;
1235 
1236 	if (witness_dead)
1237 		return (NULL);
1238 	mtx_lock_spin(&w_mtx);
1239 	lle = w_lock_list_free;
1240 	if (lle == NULL) {
1241 		witness_dead = 1;
1242 		mtx_unlock_spin(&w_mtx);
1243 		printf("%s: witness exhausted\n", __func__);
1244 		return (NULL);
1245 	}
1246 	w_lock_list_free = lle->ll_next;
1247 	mtx_unlock_spin(&w_mtx);
1248 	bzero(lle, sizeof(*lle));
1249 	return (lle);
1250 }
1251 
1252 static void
1253 witness_lock_list_free(struct lock_list_entry *lle)
1254 {
1255 
1256 	mtx_lock_spin(&w_mtx);
1257 	lle->ll_next = w_lock_list_free;
1258 	w_lock_list_free = lle;
1259 	mtx_unlock_spin(&w_mtx);
1260 }
1261 
1262 static struct lock_instance *
1263 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1264 {
1265 	struct lock_list_entry *lle;
1266 	struct lock_instance *instance;
1267 	int i;
1268 
1269 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1270 		for (i = lle->ll_count - 1; i >= 0; i--) {
1271 			instance = &lle->ll_children[i];
1272 			if (instance->li_lock == lock)
1273 				return (instance);
1274 		}
1275 	return (NULL);
1276 }
1277 
1278 int
1279 witness_list_locks(struct lock_list_entry **lock_list)
1280 {
1281 	struct lock_list_entry *lle;
1282 	struct lock_instance *instance;
1283 	struct lock_object *lock;
1284 	int i, nheld;
1285 
1286 	nheld = 0;
1287 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1288 		for (i = lle->ll_count - 1; i >= 0; i--) {
1289 			instance = &lle->ll_children[i];
1290 			lock = instance->li_lock;
1291 			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1292 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1293 			    "exclusive" : "shared",
1294 			    lock->lo_class->lc_name, lock->lo_name, lock,
1295 			    instance->li_file, instance->li_line);
1296 			nheld++;
1297 		}
1298 	return (nheld);
1299 }
1300 
1301 /*
1302  * Calling this on td != curthread is bad unless we are in ddb.
1303  */
1304 int
1305 witness_list(struct thread *td)
1306 {
1307 	critical_t savecrit;
1308 	int nheld;
1309 
1310 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1311 #ifdef DDB
1312 	KASSERT(td == curthread || db_active,
1313 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1314 	if (!db_active && witness_dead)
1315 		return (0);
1316 #else
1317 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1318 	if (witness_dead)
1319 		return (0);
1320 #endif
1321 	nheld = witness_list_locks(&td->td_sleeplocks);
1322 
1323 	/*
1324 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1325 	 * if td is currently executing on some other CPU and holds spin locks
1326 	 * as we won't display those locks.  If we had a MI way of getting
1327 	 * the per-cpu data for a given cpu then we could use
1328 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1329 	 * and "fix" this.
1330 	 */
1331 	if (td == curthread) {
1332 		/*
1333 		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1334 		 * change.
1335 		 */
1336 		savecrit = cpu_critical_enter();
1337 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1338 		cpu_critical_exit(savecrit);
1339 	}
1340 	return (nheld);
1341 }
1342 
1343 void
1344 witness_save(struct lock_object *lock, const char **filep, int *linep)
1345 {
1346 	struct lock_instance *instance;
1347 
1348 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1349 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1350 		return;
1351 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1352 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1353 		    lock->lo_class->lc_name, lock->lo_name);
1354 	instance = find_instance(curthread->td_sleeplocks, lock);
1355 	if (instance == NULL)
1356 		panic("%s: lock (%s) %s not locked", __func__,
1357 		    lock->lo_class->lc_name, lock->lo_name);
1358 	*filep = instance->li_file;
1359 	*linep = instance->li_line;
1360 }
1361 
1362 void
1363 witness_restore(struct lock_object *lock, const char *file, int line)
1364 {
1365 	struct lock_instance *instance;
1366 
1367 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1368 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1369 		return;
1370 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1371 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1372 		    lock->lo_class->lc_name, lock->lo_name);
1373 	instance = find_instance(curthread->td_sleeplocks, lock);
1374 	if (instance == NULL)
1375 		panic("%s: lock (%s) %s not locked", __func__,
1376 		    lock->lo_class->lc_name, lock->lo_name);
1377 	lock->lo_witness->w_file = file;
1378 	lock->lo_witness->w_line = line;
1379 	instance->li_file = file;
1380 	instance->li_line = line;
1381 }
1382 
1383 void
1384 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1385 {
1386 #ifdef INVARIANT_SUPPORT
1387 	struct lock_instance *instance;
1388 
1389 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1390 		return;
1391 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1392 		instance = find_instance(curthread->td_sleeplocks, lock);
1393 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1394 		instance = find_instance(PCPU_GET(spinlocks), lock);
1395 	else {
1396 		panic("Lock (%s) %s is not sleep or spin!",
1397 		    lock->lo_class->lc_name, lock->lo_name);
1398 		return;
1399 	}
1400 	switch (flags) {
1401 	case LA_UNLOCKED:
1402 		if (instance != NULL)
1403 			panic("Lock (%s) %s locked @ %s:%d.",
1404 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1405 		break;
1406 	case LA_LOCKED:
1407 	case LA_LOCKED | LA_RECURSED:
1408 	case LA_LOCKED | LA_NOTRECURSED:
1409 	case LA_SLOCKED:
1410 	case LA_SLOCKED | LA_RECURSED:
1411 	case LA_SLOCKED | LA_NOTRECURSED:
1412 	case LA_XLOCKED:
1413 	case LA_XLOCKED | LA_RECURSED:
1414 	case LA_XLOCKED | LA_NOTRECURSED:
1415 		if (instance == NULL) {
1416 			panic("Lock (%s) %s not locked @ %s:%d.",
1417 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1418 			break;
1419 		}
1420 		if ((flags & LA_XLOCKED) != 0 &&
1421 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1422 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1423 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1424 		if ((flags & LA_SLOCKED) != 0 &&
1425 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1426 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1427 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1428 		if ((flags & LA_RECURSED) != 0 &&
1429 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1430 			panic("Lock (%s) %s not recursed @ %s:%d.",
1431 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1432 		if ((flags & LA_NOTRECURSED) != 0 &&
1433 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1434 			panic("Lock (%s) %s recursed @ %s:%d.",
1435 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1436 		break;
1437 	default:
1438 		panic("Invalid lock assertion at %s:%d.", file, line);
1439 
1440 	}
1441 #endif	/* INVARIANT_SUPPORT */
1442 }
1443 
1444 #ifdef DDB
1445 
1446 DB_SHOW_COMMAND(locks, db_witness_list)
1447 {
1448 	struct thread *td;
1449 	pid_t pid;
1450 	struct proc *p;
1451 
1452 	if (have_addr) {
1453 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1454 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1455 		    ((addr >> 16) % 16) * 10000;
1456 		/* sx_slock(&allproc_lock); */
1457 		FOREACH_PROC_IN_SYSTEM(p) {
1458 			if (p->p_pid == pid)
1459 				break;
1460 		}
1461 		/* sx_sunlock(&allproc_lock); */
1462 		if (p == NULL) {
1463 			db_printf("pid %d not found\n", pid);
1464 			return;
1465 		}
1466 		FOREACH_THREAD_IN_PROC(p, td) {
1467 			witness_list(td);
1468 		}
1469 	} else {
1470 		td = curthread;
1471 		witness_list(td);
1472 	}
1473 }
1474 
1475 DB_SHOW_COMMAND(witness, db_witness_display)
1476 {
1477 
1478 	witness_display(db_printf);
1479 }
1480 #endif
1481