xref: /freebsd/sys/kern/subr_witness.c (revision 09e8dea79366f1e5b3a73e8a271b26e4b6bf2e6a)
1 /*-
2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
13  *    promote products derived from this software without specific prior
14  *    written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30  * $FreeBSD$
31  */
32 
33 /*
34  * Implementation of the `witness' lock verifier.  Originally implemented for
35  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36  * classes in FreeBSD.
37  */
38 
39 /*
40  *	Main Entry: witness
41  *	Pronunciation: 'wit-n&s
42  *	Function: noun
43  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44  *	    testimony, witness, from 2wit
45  *	Date: before 12th century
46  *	1 : attestation of a fact or event : TESTIMONY
47  *	2 : one that gives evidence; specifically : one who testifies in
48  *	    a cause or before a judicial tribunal
49  *	3 : one asked to be present at a transaction so as to be able to
50  *	    testify to its having taken place
51  *	4 : one who has personal knowledge of something
52  *	5 a : something serving as evidence or proof : SIGN
53  *	  b : public affirmation by word or example of usually
54  *	      religious faith or conviction <the heroic witness to divine
55  *	      life -- Pilot>
56  *	6 capitalized : a member of the Jehovah's Witnesses
57  */
58 
59 #include "opt_ddb.h"
60 #include "opt_witness.h"
61 
62 #include <sys/param.h>
63 #include <sys/bus.h>
64 #include <sys/kernel.h>
65 #include <sys/ktr.h>
66 #include <sys/lock.h>
67 #include <sys/malloc.h>
68 #include <sys/mutex.h>
69 #include <sys/proc.h>
70 #include <sys/sysctl.h>
71 #include <sys/systm.h>
72 
73 #include <ddb/ddb.h>
74 
75 #define WITNESS_COUNT 200
76 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77 /*
78  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
79  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80  * probably be safe for the most part, but it's still a SWAG.
81  */
82 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83 
84 #define	WITNESS_NCHILDREN 6
85 
86 struct witness_child_list_entry;
87 
88 struct witness {
89 	const	char *w_name;
90 	struct	lock_class *w_class;
91 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94 	const	char *w_file;
95 	int	w_line;
96 	u_int	w_level;
97 	u_int	w_refcount;
98 	u_char	w_Giant_squawked:1;
99 	u_char	w_other_squawked:1;
100 	u_char	w_same_squawked:1;
101 };
102 
103 struct witness_child_list_entry {
104 	struct	witness_child_list_entry *wcl_next;
105 	struct	witness *wcl_children[WITNESS_NCHILDREN];
106 	u_int	wcl_count;
107 };
108 
109 STAILQ_HEAD(witness_list, witness);
110 
111 struct witness_blessed {
112 	const	char *b_lock1;
113 	const	char *b_lock2;
114 };
115 
116 struct witness_order_list_entry {
117 	const	char *w_name;
118 	struct	lock_class *w_class;
119 };
120 
121 static struct	witness *enroll(const char *description,
122 				struct lock_class *lock_class);
123 static int	itismychild(struct witness *parent, struct witness *child);
124 static void	removechild(struct witness *parent, struct witness *child);
125 static int	isitmychild(struct witness *parent, struct witness *child);
126 static int	isitmydescendant(struct witness *parent, struct witness *child);
127 static int	blessed(struct witness *, struct witness *);
128 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
129 				     struct witness_list *list);
130 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
131 					   struct witness *);
132 static void	witness_leveldescendents(struct witness *parent, int level);
133 static void	witness_levelall(void);
134 static struct	witness *witness_get(void);
135 static void	witness_free(struct witness *m);
136 static struct	witness_child_list_entry *witness_child_get(void);
137 static void	witness_child_free(struct witness_child_list_entry *wcl);
138 static struct	lock_list_entry *witness_lock_list_get(void);
139 static void	witness_lock_list_free(struct lock_list_entry *lle);
140 static void	witness_display(void(*)(const char *fmt, ...));
141 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
142 					     struct lock_object *lock);
143 
144 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
145 
146 static int witness_watch = 1;
147 TUNABLE_INT("debug.witness_watch", &witness_watch);
148 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
149 
150 #ifdef DDB
151 /*
152  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
153  * drop into kdebug() when:
154  *	- a lock heirarchy violation occurs
155  *	- locks are held when going to sleep.
156  */
157 #ifdef WITNESS_DDB
158 int	witness_ddb = 1;
159 #else
160 int	witness_ddb = 0;
161 #endif
162 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
163 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
164 #endif /* DDB */
165 
166 #ifdef WITNESS_SKIPSPIN
167 int	witness_skipspin = 1;
168 #else
169 int	witness_skipspin = 0;
170 #endif
171 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
172 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
173     "");
174 
175 static struct mtx w_mtx;
176 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
177 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
178 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
179 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
180 static struct witness_child_list_entry *w_child_free = NULL;
181 static struct lock_list_entry *w_lock_list_free = NULL;
182 static int witness_dead;	/* fatal error, probably no memory */
183 
184 static struct witness w_data[WITNESS_COUNT];
185 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
186 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
187 
188 static struct witness_order_list_entry order_lists[] = {
189 	{ "Giant", &lock_class_mtx_sleep },
190 	{ "proctree", &lock_class_sx },
191 	{ "allproc", &lock_class_sx },
192 	{ "sigio lock", &lock_class_mtx_sleep },
193 	{ "process group", &lock_class_mtx_sleep },
194 	{ "process lock", &lock_class_mtx_sleep },
195 	{ "session", &lock_class_mtx_sleep },
196 	{ "uidinfo hash", &lock_class_mtx_sleep },
197 	{ "uidinfo struct", &lock_class_mtx_sleep },
198 	{ NULL, NULL },
199 	/*
200 	 * spin locks
201 	 */
202 #ifdef SMP
203 	{ "ap boot", &lock_class_mtx_spin },
204 #ifdef __i386__
205 	{ "com", &lock_class_mtx_spin },
206 #endif
207 #endif
208 	{ "sio", &lock_class_mtx_spin },
209 #ifdef __i386__
210 	{ "cy", &lock_class_mtx_spin },
211 #endif
212 	{ "ng_node", &lock_class_mtx_spin },
213 	{ "ng_worklist", &lock_class_mtx_spin },
214 	{ "ithread table lock", &lock_class_mtx_spin },
215 	{ "sched lock", &lock_class_mtx_spin },
216 	{ "callout", &lock_class_mtx_spin },
217 	/*
218 	 * leaf locks
219 	 */
220 	{ "allpmaps", &lock_class_mtx_spin },
221 	{ "vm page buckets mutex", &lock_class_mtx_spin },
222 	{ "icu", &lock_class_mtx_spin },
223 #ifdef SMP
224 	{ "smp rendezvous", &lock_class_mtx_spin },
225 #endif
226 	{ "clk", &lock_class_mtx_spin },
227 	{ "mutex profiling lock", &lock_class_mtx_spin },
228 	{ NULL, NULL },
229 	{ NULL, NULL }
230 };
231 
232 /*
233  * Pairs of locks which have been blessed
234  * Don't complain about order problems with blessed locks
235  */
236 static struct witness_blessed blessed_list[] = {
237 };
238 static int blessed_count =
239 	sizeof(blessed_list) / sizeof(struct witness_blessed);
240 
241 /*
242  * List of all locks in the system.
243  */
244 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
245 
246 static struct mtx all_mtx = {
247 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248 	  "All locks list",		/* mtx_object.lo_name */
249 	  "All locks list",		/* mtx_object.lo_type */
250 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
251 	  { NULL, 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 	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
289 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
290 	    MTX_NOWITNESS);
291 	for (i = 0; i < WITNESS_COUNT; i++)
292 		witness_free(&w_data[i]);
293 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
294 		witness_child_free(&w_childdata[i]);
295 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
296 		witness_lock_list_free(&w_locklistdata[i]);
297 
298 	/* First add in all the specified order lists. */
299 	for (order = order_lists; order->w_name != NULL; order++) {
300 		w = enroll(order->w_name, order->w_class);
301 		if (w == NULL)
302 			continue;
303 		w->w_file = "order list";
304 		for (order++; order->w_name != NULL; order++) {
305 			w1 = enroll(order->w_name, order->w_class);
306 			if (w1 == NULL)
307 				continue;
308 			w1->w_file = "order list";
309 			itismychild(w, w1);
310 			w = w1;
311 		}
312 	}
313 
314 	/* Iterate through all locks and add them to witness. */
315 	mtx_lock(&all_mtx);
316 	TAILQ_FOREACH(lock, &all_locks, lo_list) {
317 		if (lock->lo_flags & LO_WITNESS)
318 			lock->lo_witness = enroll(lock->lo_type,
319 			    lock->lo_class);
320 		else
321 			lock->lo_witness = NULL;
322 	}
323 	mtx_unlock(&all_mtx);
324 
325 	/* Mark the witness code as being ready for use. */
326 	atomic_store_rel_int(&witness_cold, 0);
327 
328 	mtx_lock(&Giant);
329 }
330 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
331 
332 void
333 witness_init(struct lock_object *lock)
334 {
335 	struct lock_class *class;
336 
337 	class = lock->lo_class;
338 	if (lock->lo_flags & LO_INITIALIZED)
339 		panic("%s: lock (%s) %s is already initialized", __func__,
340 		    class->lc_name, lock->lo_name);
341 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
342 	    (class->lc_flags & LC_RECURSABLE) == 0)
343 		panic("%s: lock (%s) %s can not be recursable", __func__,
344 		    class->lc_name, lock->lo_name);
345 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
346 	    (class->lc_flags & LC_SLEEPABLE) == 0)
347 		panic("%s: lock (%s) %s can not be sleepable", __func__,
348 		    class->lc_name, lock->lo_name);
349 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
350 	    (class->lc_flags & LC_UPGRADABLE) == 0)
351 		panic("%s: lock (%s) %s can not be upgradable", __func__,
352 		    class->lc_name, lock->lo_name);
353 
354 	mtx_lock(&all_mtx);
355 	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
356 	lock->lo_flags |= LO_INITIALIZED;
357 	lock_cur_cnt++;
358 	if (lock_cur_cnt > lock_max_cnt)
359 		lock_max_cnt = lock_cur_cnt;
360 	mtx_unlock(&all_mtx);
361 	if (!witness_cold && !witness_dead && panicstr == NULL &&
362 	    (lock->lo_flags & LO_WITNESS) != 0)
363 		lock->lo_witness = enroll(lock->lo_type, class);
364 	else
365 		lock->lo_witness = NULL;
366 }
367 
368 void
369 witness_destroy(struct lock_object *lock)
370 {
371 	struct witness *w;
372 
373 	if (witness_cold)
374 		panic("lock (%s) %s destroyed while witness_cold",
375 		    lock->lo_class->lc_name, lock->lo_name);
376 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
377 		panic("%s: lock (%s) %s is not initialized", __func__,
378 		    lock->lo_class->lc_name, lock->lo_name);
379 
380 	/* XXX: need to verify that no one holds the lock */
381 	w = lock->lo_witness;
382 	if (w != NULL) {
383 		mtx_lock_spin(&w_mtx);
384 		MPASS(w->w_refcount > 0);
385 		w->w_refcount--;
386 		mtx_unlock_spin(&w_mtx);
387 	}
388 
389 	mtx_lock(&all_mtx);
390 	lock_cur_cnt--;
391 	TAILQ_REMOVE(&all_locks, lock, lo_list);
392 	lock->lo_flags &= ~LO_INITIALIZED;
393 	mtx_unlock(&all_mtx);
394 }
395 
396 static void
397 witness_display_list(void(*prnt)(const char *fmt, ...),
398 		     struct witness_list *list)
399 {
400 	struct witness *w, *w1;
401 	int found;
402 
403 	STAILQ_FOREACH(w, list, w_typelist) {
404 		if (w->w_file == NULL)
405 			continue;
406 		found = 0;
407 		STAILQ_FOREACH(w1, list, w_typelist) {
408 			if (isitmychild(w1, w)) {
409 				found++;
410 				break;
411 			}
412 		}
413 		if (found)
414 			continue;
415 		/*
416 		 * This lock has no anscestors, display its descendants.
417 		 */
418 		witness_displaydescendants(prnt, w);
419 	}
420 }
421 
422 static void
423 witness_display(void(*prnt)(const char *fmt, ...))
424 {
425 	struct witness *w;
426 
427 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
428 	witness_levelall();
429 
430 	/*
431 	 * First, handle sleep locks which have been acquired at least
432 	 * once.
433 	 */
434 	prnt("Sleep locks:\n");
435 	witness_display_list(prnt, &w_sleep);
436 
437 	/*
438 	 * Now do spin locks which have been acquired at least once.
439 	 */
440 	prnt("\nSpin locks:\n");
441 	witness_display_list(prnt, &w_spin);
442 
443 	/*
444 	 * Finally, any locks which have not been acquired yet.
445 	 */
446 	prnt("\nLocks which were never acquired:\n");
447 	STAILQ_FOREACH(w, &w_all, w_list) {
448 		if (w->w_file != NULL || w->w_refcount == 0)
449 			continue;
450 		prnt("%s\n", w->w_name);
451 	}
452 }
453 
454 void
455 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
456 {
457 	struct lock_list_entry **lock_list, *lle;
458 	struct lock_instance *lock1, *lock2;
459 	struct lock_class *class;
460 	struct witness *w, *w1;
461 	struct thread *td;
462 	int i, j;
463 #ifdef DDB
464 	int go_into_ddb = 0;
465 #endif /* DDB */
466 
467 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
468 	    panicstr != NULL)
469 		return;
470 	w = lock->lo_witness;
471 	class = lock->lo_class;
472 	td = curthread;
473 
474 	if (class->lc_flags & LC_SLEEPLOCK) {
475 		/*
476 		 * Since spin locks include a critical section, this check
477 		 * impliclty enforces a lock order of all sleep locks before
478 		 * all spin locks.
479 		 */
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 || (lock->lo_flags & LO_DUPOK))
549 			goto out;
550 		w->w_same_squawked = 1;
551 		printf("acquiring duplicate lock of same type: \"%s\"\n",
552 			lock->lo_type);
553 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
554 		    lock1->li_file, lock1->li_line);
555 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
556 #ifdef DDB
557 		go_into_ddb = 1;
558 #endif /* DDB */
559 		goto out;
560 	}
561 	MPASS(!mtx_owned(&w_mtx));
562 	mtx_lock_spin(&w_mtx);
563 	/*
564 	 * If we have a known higher number just say ok
565 	 */
566 	if (witness_watch > 1 && w->w_level > w1->w_level) {
567 		mtx_unlock_spin(&w_mtx);
568 		goto out;
569 	}
570 	if (isitmydescendant(w1, w)) {
571 		mtx_unlock_spin(&w_mtx);
572 		goto out;
573 	}
574 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
575 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
576 
577 			MPASS(j < WITNESS_COUNT);
578 			lock1 = &lle->ll_children[i];
579 			w1 = lock1->li_lock->lo_witness;
580 
581 			/*
582 			 * If this lock doesn't undergo witness checking,
583 			 * then skip it.
584 			 */
585 			if (w1 == NULL) {
586 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
587 				    ("lock missing witness structure"));
588 				continue;
589 			}
590 			/*
591 			 * If we are locking Giant and we slept with this
592 			 * lock, then skip it.
593 			 */
594 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
595 			    lock == &Giant.mtx_object)
596 				continue;
597 			/*
598 			 * If we are locking a sleepable lock and this lock
599 			 * isn't sleepable and isn't Giant, we want to treat
600 			 * it as a lock order violation to enfore a general
601 			 * lock order of sleepable locks before non-sleepable
602 			 * locks.  Thus, we only bother checking the lock
603 			 * order hierarchy if we pass the initial test.
604 			 */
605 			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
606 			    ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
607 			    lock1->li_lock != &Giant.mtx_object)) &&
608 			    !isitmydescendant(w, w1))
609 				continue;
610 			/*
611 			 * We have a lock order violation, check to see if it
612 			 * is allowed or has already been yelled about.
613 			 */
614 			mtx_unlock_spin(&w_mtx);
615 			if (blessed(w, w1))
616 				goto out;
617 			if (lock1->li_lock == &Giant.mtx_object) {
618 				if (w1->w_Giant_squawked)
619 					goto out;
620 				else
621 					w1->w_Giant_squawked = 1;
622 			} else {
623 				if (w1->w_other_squawked)
624 					goto out;
625 				else
626 					w1->w_other_squawked = 1;
627 			}
628 			/*
629 			 * Ok, yell about it.
630 			 */
631 			printf("lock order reversal\n");
632 			/*
633 			 * Try to locate an earlier lock with
634 			 * witness w in our list.
635 			 */
636 			do {
637 				lock2 = &lle->ll_children[i];
638 				MPASS(lock2->li_lock != NULL);
639 				if (lock2->li_lock->lo_witness == w)
640 					break;
641 				i--;
642 				if (i == 0 && lle->ll_next != NULL) {
643 					lle = lle->ll_next;
644 					i = lle->ll_count - 1;
645 					MPASS(i != 0);
646 				}
647 			} while (i >= 0);
648 			if (i < 0) {
649 				printf(" 1st %p %s (%s) @ %s:%d\n",
650 				    lock1->li_lock, lock1->li_lock->lo_name,
651 				    lock1->li_lock->lo_type, lock1->li_file,
652 				    lock1->li_line);
653 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
654 				    lock->lo_name, lock->lo_type, file, line);
655 			} else {
656 				printf(" 1st %p %s (%s) @ %s:%d\n",
657 				    lock2->li_lock, lock2->li_lock->lo_name,
658 				    lock2->li_lock->lo_type, lock2->li_file,
659 				    lock2->li_line);
660 				printf(" 2nd %p %s (%s) @ %s:%d\n",
661 				    lock1->li_lock, lock1->li_lock->lo_name,
662 				    lock1->li_lock->lo_type, lock1->li_file,
663 				    lock1->li_line);
664 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
665 				    lock->lo_name, lock->lo_type, file, line);
666 			}
667 #ifdef DDB
668 			go_into_ddb = 1;
669 #endif /* DDB */
670 			goto out;
671 		}
672 	}
673 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
674 	/*
675 	 * Don't build a new relationship if we are locking Giant just
676 	 * after waking up and the previous lock in the list was acquired
677 	 * prior to blocking.
678 	 */
679 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
680 		mtx_unlock_spin(&w_mtx);
681 	else {
682 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
683 		    lock->lo_type, lock1->li_lock->lo_type);
684 		if (!itismychild(lock1->li_lock->lo_witness, w))
685 			mtx_unlock_spin(&w_mtx);
686 	}
687 
688 out:
689 #ifdef DDB
690 	if (witness_ddb && go_into_ddb)
691 		Debugger(__func__);
692 #endif /* DDB */
693 	w->w_file = file;
694 	w->w_line = line;
695 
696 	lle = *lock_list;
697 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
698 		lle = witness_lock_list_get();
699 		if (lle == NULL)
700 			return;
701 		lle->ll_next = *lock_list;
702 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
703 		    td->td_proc->p_pid, lle);
704 		*lock_list = lle;
705 	}
706 	lock1 = &lle->ll_children[lle->ll_count++];
707 	lock1->li_lock = lock;
708 	lock1->li_line = line;
709 	lock1->li_file = file;
710 	if ((flags & LOP_EXCLUSIVE) != 0)
711 		lock1->li_flags = LI_EXCLUSIVE;
712 	else
713 		lock1->li_flags = 0;
714 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
715 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
716 }
717 
718 void
719 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
720 {
721 	struct lock_instance *instance;
722 	struct lock_class *class;
723 
724 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
725 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
726 		return;
727 	class = lock->lo_class;
728 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
729 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
730 		    class->lc_name, lock->lo_name, file, line);
731 	if ((flags & LOP_TRYLOCK) == 0)
732 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
733 		    lock->lo_name, file, line);
734 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
735 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
736 		    class->lc_name, lock->lo_name, file, line);
737 	instance = find_instance(curthread->td_sleeplocks, lock);
738 	if (instance == NULL)
739 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
740 		    class->lc_name, lock->lo_name, file, line);
741 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
742 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
743 		    class->lc_name, lock->lo_name, file, line);
744 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
745 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
746 		    class->lc_name, lock->lo_name,
747 		    instance->li_flags & LI_RECURSEMASK, file, line);
748 	instance->li_flags |= LI_EXCLUSIVE;
749 }
750 
751 void
752 witness_downgrade(struct lock_object *lock, int flags, const char *file,
753     int line)
754 {
755 	struct lock_instance *instance;
756 	struct lock_class *class;
757 
758 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
759 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
760 		return;
761 	class = lock->lo_class;
762 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
763 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
764 		    class->lc_name, lock->lo_name, file, line);
765 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
766 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
767 		    class->lc_name, lock->lo_name, file, line);
768 	instance = find_instance(curthread->td_sleeplocks, lock);
769 	if (instance == NULL)
770 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
771 		    class->lc_name, lock->lo_name, file, line);
772 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
773 		panic("downgrade of shared lock (%s) %s @ %s:%d",
774 		    class->lc_name, lock->lo_name, file, line);
775 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
776 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
777 		    class->lc_name, lock->lo_name,
778 		    instance->li_flags & LI_RECURSEMASK, file, line);
779 	instance->li_flags &= ~LI_EXCLUSIVE;
780 }
781 
782 void
783 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
784 {
785 	struct lock_list_entry **lock_list, *lle;
786 	struct lock_instance *instance;
787 	struct lock_class *class;
788 	struct thread *td;
789 	register_t s;
790 	int i, j;
791 
792 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
793 	    panicstr != NULL)
794 		return;
795 	td = curthread;
796 	class = lock->lo_class;
797 	if (class->lc_flags & LC_SLEEPLOCK)
798 		lock_list = &td->td_sleeplocks;
799 	else
800 		lock_list = PCPU_PTR(spinlocks);
801 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
802 		for (i = 0; i < (*lock_list)->ll_count; i++) {
803 			instance = &(*lock_list)->ll_children[i];
804 			if (instance->li_lock == lock) {
805 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
806 				    (flags & LOP_EXCLUSIVE) == 0) {
807 					printf(
808 					"shared unlock of (%s) %s @ %s:%d\n",
809 					    class->lc_name, lock->lo_name,
810 					    file, line);
811 					printf(
812 					"while exclusively locked from %s:%d\n",
813 					    instance->li_file,
814 					    instance->li_line);
815 					panic("excl->ushare");
816 				}
817 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
818 				    (flags & LOP_EXCLUSIVE) != 0) {
819 					printf(
820 					"exclusive unlock of (%s) %s @ %s:%d\n",
821 					    class->lc_name, lock->lo_name,
822 					    file, line);
823 					printf(
824 					"while share locked from %s:%d\n",
825 					    instance->li_file,
826 					    instance->li_line);
827 					panic("share->uexcl");
828 				}
829 				/* If we are recursed, unrecurse. */
830 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
831 					CTR4(KTR_WITNESS,
832 				    "%s: pid %d unrecursed on %s r=%d", __func__,
833 					    td->td_proc->p_pid,
834 					    instance->li_lock->lo_name,
835 					    instance->li_flags);
836 					instance->li_flags--;
837 					return;
838 				}
839 				s = intr_disable();
840 				CTR4(KTR_WITNESS,
841 				    "%s: pid %d removed %s from lle[%d]", __func__,
842 				    td->td_proc->p_pid,
843 				    instance->li_lock->lo_name,
844 				    (*lock_list)->ll_count - 1);
845 				for (j = i; j < (*lock_list)->ll_count - 1; j++)
846 					(*lock_list)->ll_children[j] =
847 					    (*lock_list)->ll_children[j + 1];
848 				(*lock_list)->ll_count--;
849 				intr_restore(s);
850 				if ((*lock_list)->ll_count == 0) {
851 					lle = *lock_list;
852 					*lock_list = lle->ll_next;
853 					CTR3(KTR_WITNESS,
854 					    "%s: pid %d removed lle %p", __func__,
855 					    td->td_proc->p_pid, lle);
856 					witness_lock_list_free(lle);
857 				}
858 				return;
859 			}
860 		}
861 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
862 	    file, line);
863 }
864 
865 /*
866  * Warn if any held locks are not sleepable.  Note that Giant and the lock
867  * passed in are both special cases since they are both released during the
868  * sleep process and aren't actually held while the thread is asleep.
869  */
870 int
871 witness_sleep(int check_only, struct lock_object *lock, const char *file,
872 	      int line)
873 {
874 	struct lock_list_entry **lock_list, *lle;
875 	struct lock_instance *lock1;
876 	struct thread *td;
877 	int i, n;
878 
879 	if (witness_cold || witness_dead || panicstr != NULL)
880 		return (0);
881 	n = 0;
882 	td = curthread;
883 	lock_list = &td->td_sleeplocks;
884 again:
885 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
886 		for (i = lle->ll_count - 1; i >= 0; i--) {
887 			lock1 = &lle->ll_children[i];
888 			if (lock1->li_lock == lock ||
889 			    lock1->li_lock == &Giant.mtx_object)
890 				continue;
891 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
892 				if (check_only == 0) {
893 					CTR3(KTR_WITNESS,
894 				    "pid %d: sleeping with lock (%s) %s held",
895 					    td->td_proc->p_pid,
896 					    lock1->li_lock->lo_class->lc_name,
897 					    lock1->li_lock->lo_name);
898 					lock1->li_flags |= LI_SLEPT;
899 				}
900 				continue;
901 			}
902 			n++;
903 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
904 			    file, line, check_only ? "could sleep" : "sleeping",
905 			    lock1->li_lock->lo_name, lock1->li_file,
906 			    lock1->li_line);
907 		}
908 	if (lock_list == &td->td_sleeplocks && PCPU_GET(spinlocks) != NULL) {
909 		/*
910 		 * Since we already hold a spinlock preemption is
911 		 * already blocked.
912 		 */
913 		lock_list = PCPU_PTR(spinlocks);
914 		goto again;
915 	}
916 #ifdef DDB
917 	if (witness_ddb && n)
918 		Debugger(__func__);
919 #endif /* DDB */
920 	return (n);
921 }
922 
923 static struct witness *
924 enroll(const char *description, struct lock_class *lock_class)
925 {
926 	struct witness *w;
927 
928 	if (!witness_watch || witness_dead || panicstr != NULL)
929 		return (NULL);
930 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
931 		return (NULL);
932 	mtx_lock_spin(&w_mtx);
933 	STAILQ_FOREACH(w, &w_all, w_list) {
934 		if (w->w_name == description || (w->w_refcount > 0 &&
935 		    strcmp(description, w->w_name) == 0)) {
936 			w->w_refcount++;
937 			mtx_unlock_spin(&w_mtx);
938 			if (lock_class != w->w_class)
939 				panic(
940 				"lock (%s) %s does not match earlier (%s) lock",
941 				    description, lock_class->lc_name,
942 				    w->w_class->lc_name);
943 			return (w);
944 		}
945 	}
946 	/*
947 	 * This isn't quite right, as witness_cold is still 0 while we
948 	 * enroll all the locks initialized before witness_initialize().
949 	 */
950 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
951 		mtx_unlock_spin(&w_mtx);
952 		panic("spin lock %s not in order list", description);
953 	}
954 	if ((w = witness_get()) == NULL)
955 		return (NULL);
956 	w->w_name = description;
957 	w->w_class = lock_class;
958 	w->w_refcount = 1;
959 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
960 	if (lock_class->lc_flags & LC_SPINLOCK)
961 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
962 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
963 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
964 	else {
965 		mtx_unlock_spin(&w_mtx);
966 		panic("lock class %s is not sleep or spin",
967 		    lock_class->lc_name);
968 	}
969 	mtx_unlock_spin(&w_mtx);
970 	return (w);
971 }
972 
973 static int
974 itismychild(struct witness *parent, struct witness *child)
975 {
976 	static int recursed;
977 	struct witness_child_list_entry **wcl;
978 	struct witness_list *list;
979 
980 	MPASS(child != NULL && parent != NULL);
981 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
982 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
983 		panic(
984 		"%s: parent (%s) and child (%s) are not the same lock type",
985 		    __func__, parent->w_class->lc_name,
986 		    child->w_class->lc_name);
987 
988 	/*
989 	 * Insert "child" after "parent"
990 	 */
991 	wcl = &parent->w_children;
992 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
993 		wcl = &(*wcl)->wcl_next;
994 	if (*wcl == NULL) {
995 		*wcl = witness_child_get();
996 		if (*wcl == NULL)
997 			return (1);
998 	}
999 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1000 
1001 	/*
1002 	 * Now prune whole tree.  We look for cases where a lock is now
1003 	 * both a descendant and a direct child of a given lock.  In that
1004 	 * case, we want to remove the direct child link from the tree.
1005 	 */
1006 	if (recursed)
1007 		return (0);
1008 	recursed = 1;
1009 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1010 		list = &w_sleep;
1011 	else
1012 		list = &w_spin;
1013 	STAILQ_FOREACH(child, list, w_typelist) {
1014 		STAILQ_FOREACH(parent, list, w_typelist) {
1015 			if (!isitmychild(parent, child))
1016 				continue;
1017 			removechild(parent, child);
1018 			if (isitmydescendant(parent, child))
1019 				continue;
1020 			itismychild(parent, child);
1021 		}
1022 	}
1023 	recursed = 0;
1024 	witness_levelall();
1025 	return (0);
1026 }
1027 
1028 static void
1029 removechild(struct witness *parent, struct witness *child)
1030 {
1031 	struct witness_child_list_entry **wcl, *wcl1;
1032 	int i;
1033 
1034 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1035 		for (i = 0; i < (*wcl)->wcl_count; i++)
1036 			if ((*wcl)->wcl_children[i] == child)
1037 				goto found;
1038 	return;
1039 found:
1040 	(*wcl)->wcl_count--;
1041 	if ((*wcl)->wcl_count > i)
1042 		(*wcl)->wcl_children[i] =
1043 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1044 	MPASS((*wcl)->wcl_children[i] != NULL);
1045 	if ((*wcl)->wcl_count != 0)
1046 		return;
1047 	wcl1 = *wcl;
1048 	*wcl = wcl1->wcl_next;
1049 	witness_child_free(wcl1);
1050 }
1051 
1052 static int
1053 isitmychild(struct witness *parent, struct witness *child)
1054 {
1055 	struct witness_child_list_entry *wcl;
1056 	int i;
1057 
1058 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1059 		for (i = 0; i < wcl->wcl_count; i++) {
1060 			if (wcl->wcl_children[i] == child)
1061 				return (1);
1062 		}
1063 	}
1064 	return (0);
1065 }
1066 
1067 static int
1068 isitmydescendant(struct witness *parent, struct witness *child)
1069 {
1070 	struct witness_child_list_entry *wcl;
1071 	int i, j;
1072 
1073 	if (isitmychild(parent, child))
1074 		return (1);
1075 	j = 0;
1076 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1077 		MPASS(j < 1000);
1078 		for (i = 0; i < wcl->wcl_count; i++) {
1079 			if (isitmydescendant(wcl->wcl_children[i], child))
1080 				return (1);
1081 		}
1082 		j++;
1083 	}
1084 	return (0);
1085 }
1086 
1087 void
1088 witness_levelall (void)
1089 {
1090 	struct witness_list *list;
1091 	struct witness *w, *w1;
1092 
1093 	/*
1094 	 * First clear all levels.
1095 	 */
1096 	STAILQ_FOREACH(w, &w_all, w_list) {
1097 		w->w_level = 0;
1098 	}
1099 
1100 	/*
1101 	 * Look for locks with no parent and level all their descendants.
1102 	 */
1103 	STAILQ_FOREACH(w, &w_all, w_list) {
1104 		/*
1105 		 * This is just an optimization, technically we could get
1106 		 * away just walking the all list each time.
1107 		 */
1108 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1109 			list = &w_sleep;
1110 		else
1111 			list = &w_spin;
1112 		STAILQ_FOREACH(w1, list, w_typelist) {
1113 			if (isitmychild(w1, w))
1114 				goto skip;
1115 		}
1116 		witness_leveldescendents(w, 0);
1117 	skip:
1118 		;	/* silence GCC 3.x */
1119 	}
1120 }
1121 
1122 static void
1123 witness_leveldescendents(struct witness *parent, int level)
1124 {
1125 	struct witness_child_list_entry *wcl;
1126 	int i;
1127 
1128 	if (parent->w_level < level)
1129 		parent->w_level = level;
1130 	level++;
1131 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1132 		for (i = 0; i < wcl->wcl_count; i++)
1133 			witness_leveldescendents(wcl->wcl_children[i], level);
1134 }
1135 
1136 static void
1137 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1138 			   struct witness *parent)
1139 {
1140 	struct witness_child_list_entry *wcl;
1141 	int i, level;
1142 
1143 	level = parent->w_level;
1144 	prnt("%-2d", level);
1145 	for (i = 0; i < level; i++)
1146 		prnt(" ");
1147 	if (parent->w_refcount > 0) {
1148 		prnt("%s", parent->w_name);
1149 		if (parent->w_file != NULL)
1150 			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1151 			    parent->w_line);
1152 	} else
1153 		prnt("(dead)\n");
1154 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1155 		for (i = 0; i < wcl->wcl_count; i++)
1156 			    witness_displaydescendants(prnt,
1157 				wcl->wcl_children[i]);
1158 }
1159 
1160 static int
1161 blessed(struct witness *w1, struct witness *w2)
1162 {
1163 	int i;
1164 	struct witness_blessed *b;
1165 
1166 	for (i = 0; i < blessed_count; i++) {
1167 		b = &blessed_list[i];
1168 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1169 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1170 				return (1);
1171 			continue;
1172 		}
1173 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1174 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1175 				return (1);
1176 	}
1177 	return (0);
1178 }
1179 
1180 static struct witness *
1181 witness_get(void)
1182 {
1183 	struct witness *w;
1184 
1185 	if (witness_dead) {
1186 		mtx_unlock_spin(&w_mtx);
1187 		return (NULL);
1188 	}
1189 	if (STAILQ_EMPTY(&w_free)) {
1190 		witness_dead = 1;
1191 		mtx_unlock_spin(&w_mtx);
1192 		printf("%s: witness exhausted\n", __func__);
1193 		return (NULL);
1194 	}
1195 	w = STAILQ_FIRST(&w_free);
1196 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1197 	bzero(w, sizeof(*w));
1198 	return (w);
1199 }
1200 
1201 static void
1202 witness_free(struct witness *w)
1203 {
1204 
1205 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1206 }
1207 
1208 static struct witness_child_list_entry *
1209 witness_child_get(void)
1210 {
1211 	struct witness_child_list_entry *wcl;
1212 
1213 	if (witness_dead) {
1214 		mtx_unlock_spin(&w_mtx);
1215 		return (NULL);
1216 	}
1217 	wcl = w_child_free;
1218 	if (wcl == NULL) {
1219 		witness_dead = 1;
1220 		mtx_unlock_spin(&w_mtx);
1221 		printf("%s: witness exhausted\n", __func__);
1222 		return (NULL);
1223 	}
1224 	w_child_free = wcl->wcl_next;
1225 	bzero(wcl, sizeof(*wcl));
1226 	return (wcl);
1227 }
1228 
1229 static void
1230 witness_child_free(struct witness_child_list_entry *wcl)
1231 {
1232 
1233 	wcl->wcl_next = w_child_free;
1234 	w_child_free = wcl;
1235 }
1236 
1237 static struct lock_list_entry *
1238 witness_lock_list_get(void)
1239 {
1240 	struct lock_list_entry *lle;
1241 
1242 	if (witness_dead)
1243 		return (NULL);
1244 	mtx_lock_spin(&w_mtx);
1245 	lle = w_lock_list_free;
1246 	if (lle == NULL) {
1247 		witness_dead = 1;
1248 		mtx_unlock_spin(&w_mtx);
1249 		printf("%s: witness exhausted\n", __func__);
1250 		return (NULL);
1251 	}
1252 	w_lock_list_free = lle->ll_next;
1253 	mtx_unlock_spin(&w_mtx);
1254 	bzero(lle, sizeof(*lle));
1255 	return (lle);
1256 }
1257 
1258 static void
1259 witness_lock_list_free(struct lock_list_entry *lle)
1260 {
1261 
1262 	mtx_lock_spin(&w_mtx);
1263 	lle->ll_next = w_lock_list_free;
1264 	w_lock_list_free = lle;
1265 	mtx_unlock_spin(&w_mtx);
1266 }
1267 
1268 static struct lock_instance *
1269 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1270 {
1271 	struct lock_list_entry *lle;
1272 	struct lock_instance *instance;
1273 	int i;
1274 
1275 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1276 		for (i = lle->ll_count - 1; i >= 0; i--) {
1277 			instance = &lle->ll_children[i];
1278 			if (instance->li_lock == lock)
1279 				return (instance);
1280 		}
1281 	return (NULL);
1282 }
1283 
1284 int
1285 witness_list_locks(struct lock_list_entry **lock_list)
1286 {
1287 	struct lock_list_entry *lle;
1288 	struct lock_instance *instance;
1289 	struct lock_object *lock;
1290 	int i, nheld;
1291 
1292 	nheld = 0;
1293 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1294 		for (i = lle->ll_count - 1; i >= 0; i--) {
1295 			instance = &lle->ll_children[i];
1296 			lock = instance->li_lock;
1297 			printf("%s %s %s",
1298 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1299 			    "exclusive" : "shared",
1300 			    lock->lo_class->lc_name, lock->lo_name);
1301 			if (lock->lo_type != lock->lo_name)
1302 				printf(" (%s)", lock->lo_type);
1303 			printf(" r = %d (%p) locked @ %s:%d\n",
1304 			    instance->li_flags & LI_RECURSEMASK, lock,
1305 			    instance->li_file, instance->li_line);
1306 			nheld++;
1307 		}
1308 	return (nheld);
1309 }
1310 
1311 /*
1312  * Calling this on td != curthread is bad unless we are in ddb.
1313  */
1314 int
1315 witness_list(struct thread *td)
1316 {
1317 	int nheld;
1318 
1319 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1320 #ifdef DDB
1321 	KASSERT(td == curthread || db_active,
1322 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1323 	if (!db_active && witness_dead)
1324 		return (0);
1325 #else
1326 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1327 	if (witness_dead)
1328 		return (0);
1329 #endif
1330 	nheld = witness_list_locks(&td->td_sleeplocks);
1331 
1332 	/*
1333 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1334 	 * if td is currently executing on some other CPU and holds spin locks
1335 	 * as we won't display those locks.  If we had a MI way of getting
1336 	 * the per-cpu data for a given cpu then we could use
1337 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1338 	 * and "fix" this.
1339 	 *
1340 	 * That still wouldn't really fix this unless we locked sched_lock
1341 	 * or stopped the other CPU to make sure it wasn't changing the list
1342 	 * out from under us.  It is probably best to just not try to handle
1343 	 * threads on other CPU's for now.
1344 	 */
1345 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1346 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1347 
1348 	return (nheld);
1349 }
1350 
1351 void
1352 witness_save(struct lock_object *lock, const char **filep, int *linep)
1353 {
1354 	struct lock_instance *instance;
1355 
1356 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1357 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1358 		return;
1359 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1360 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1361 		    lock->lo_class->lc_name, lock->lo_name);
1362 	instance = find_instance(curthread->td_sleeplocks, lock);
1363 	if (instance == NULL)
1364 		panic("%s: lock (%s) %s not locked", __func__,
1365 		    lock->lo_class->lc_name, lock->lo_name);
1366 	*filep = instance->li_file;
1367 	*linep = instance->li_line;
1368 }
1369 
1370 void
1371 witness_restore(struct lock_object *lock, const char *file, int line)
1372 {
1373 	struct lock_instance *instance;
1374 
1375 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1376 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1377 		return;
1378 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1379 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1380 		    lock->lo_class->lc_name, lock->lo_name);
1381 	instance = find_instance(curthread->td_sleeplocks, lock);
1382 	if (instance == NULL)
1383 		panic("%s: lock (%s) %s not locked", __func__,
1384 		    lock->lo_class->lc_name, lock->lo_name);
1385 	lock->lo_witness->w_file = file;
1386 	lock->lo_witness->w_line = line;
1387 	instance->li_file = file;
1388 	instance->li_line = line;
1389 }
1390 
1391 void
1392 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1393 {
1394 #ifdef INVARIANT_SUPPORT
1395 	struct lock_instance *instance;
1396 
1397 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1398 		return;
1399 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1400 		instance = find_instance(curthread->td_sleeplocks, lock);
1401 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1402 		instance = find_instance(PCPU_GET(spinlocks), lock);
1403 	else {
1404 		panic("Lock (%s) %s is not sleep or spin!",
1405 		    lock->lo_class->lc_name, lock->lo_name);
1406 		return;
1407 	}
1408 	switch (flags) {
1409 	case LA_UNLOCKED:
1410 		if (instance != NULL)
1411 			panic("Lock (%s) %s locked @ %s:%d.",
1412 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1413 		break;
1414 	case LA_LOCKED:
1415 	case LA_LOCKED | LA_RECURSED:
1416 	case LA_LOCKED | LA_NOTRECURSED:
1417 	case LA_SLOCKED:
1418 	case LA_SLOCKED | LA_RECURSED:
1419 	case LA_SLOCKED | LA_NOTRECURSED:
1420 	case LA_XLOCKED:
1421 	case LA_XLOCKED | LA_RECURSED:
1422 	case LA_XLOCKED | LA_NOTRECURSED:
1423 		if (instance == NULL) {
1424 			panic("Lock (%s) %s not locked @ %s:%d.",
1425 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1426 			break;
1427 		}
1428 		if ((flags & LA_XLOCKED) != 0 &&
1429 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1430 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1431 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1432 		if ((flags & LA_SLOCKED) != 0 &&
1433 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1434 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1435 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1436 		if ((flags & LA_RECURSED) != 0 &&
1437 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1438 			panic("Lock (%s) %s not recursed @ %s:%d.",
1439 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1440 		if ((flags & LA_NOTRECURSED) != 0 &&
1441 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1442 			panic("Lock (%s) %s recursed @ %s:%d.",
1443 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1444 		break;
1445 	default:
1446 		panic("Invalid lock assertion at %s:%d.", file, line);
1447 
1448 	}
1449 #endif	/* INVARIANT_SUPPORT */
1450 }
1451 
1452 #ifdef DDB
1453 
1454 DB_SHOW_COMMAND(locks, db_witness_list)
1455 {
1456 	struct thread *td;
1457 	pid_t pid;
1458 	struct proc *p;
1459 
1460 	if (have_addr) {
1461 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1462 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1463 		    ((addr >> 16) % 16) * 10000;
1464 		/* sx_slock(&allproc_lock); */
1465 		FOREACH_PROC_IN_SYSTEM(p) {
1466 			if (p->p_pid == pid)
1467 				break;
1468 		}
1469 		/* sx_sunlock(&allproc_lock); */
1470 		if (p == NULL) {
1471 			db_printf("pid %d not found\n", pid);
1472 			return;
1473 		}
1474 		FOREACH_THREAD_IN_PROC(p, td) {
1475 			witness_list(td);
1476 		}
1477 	} else {
1478 		td = curthread;
1479 		witness_list(td);
1480 	}
1481 }
1482 
1483 DB_SHOW_COMMAND(witness, db_witness_display)
1484 {
1485 
1486 	witness_display(db_printf);
1487 }
1488 #endif
1489