xref: /freebsd/sys/kern/subr_witness.c (revision 81d1ffee089aab2652954909acbe6aadd8a1a72c)
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 /*
60  * Special rules concerning Giant and lock orders:
61  *
62  * 1) Giant must be acquired before any other mutexes.  Stated another way,
63  *    no other mutex may be held when Giant is acquired.
64  *
65  * 2) Giant must be released when blocking on a sleepable lock.
66  *
67  * This rule is less obvious, but is a result of Giant providing the same
68  * semantics as spl().  Basically, when a thread sleeps, it must release
69  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
70  * 2).
71  *
72  * 3) Giant may be acquired before or after sleepable locks.
73  *
74  * This rule is also not quite as obvious.  Giant may be acquired after
75  * a sleepable lock because it is a non-sleepable lock and non-sleepable
76  * locks may always be acquired while holding a sleepable lock.  The second
77  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
78  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
79  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
80  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
81  * execute.  Thus, acquiring Giant both before and after a sleepable lock
82  * will not result in a lock order reversal.
83  */
84 
85 #include "opt_ddb.h"
86 #include "opt_witness.h"
87 
88 #include <sys/param.h>
89 #include <sys/bus.h>
90 #include <sys/kernel.h>
91 #include <sys/ktr.h>
92 #include <sys/lock.h>
93 #include <sys/malloc.h>
94 #include <sys/mutex.h>
95 #include <sys/proc.h>
96 #include <sys/sysctl.h>
97 #include <sys/systm.h>
98 
99 #include <ddb/ddb.h>
100 
101 #include <machine/stdarg.h>
102 
103 /* Define this to check for blessed mutexes */
104 #undef BLESSING
105 
106 #define WITNESS_COUNT 200
107 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
108 /*
109  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
110  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
111  * probably be safe for the most part, but it's still a SWAG.
112  */
113 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
114 
115 #define	WITNESS_NCHILDREN 6
116 
117 struct witness_child_list_entry;
118 
119 struct witness {
120 	const	char *w_name;
121 	struct	lock_class *w_class;
122 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
123 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
124 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
125 	const	char *w_file;
126 	int	w_line;
127 	u_int	w_level;
128 	u_int	w_refcount;
129 	u_char	w_Giant_squawked:1;
130 	u_char	w_other_squawked:1;
131 	u_char	w_same_squawked:1;
132 };
133 
134 struct witness_child_list_entry {
135 	struct	witness_child_list_entry *wcl_next;
136 	struct	witness *wcl_children[WITNESS_NCHILDREN];
137 	u_int	wcl_count;
138 };
139 
140 STAILQ_HEAD(witness_list, witness);
141 
142 #ifdef BLESSING
143 struct witness_blessed {
144 	const	char *b_lock1;
145 	const	char *b_lock2;
146 };
147 #endif
148 
149 struct witness_order_list_entry {
150 	const	char *w_name;
151 	struct	lock_class *w_class;
152 };
153 
154 static struct	witness *enroll(const char *description,
155 				struct lock_class *lock_class);
156 static int	itismychild(struct witness *parent, struct witness *child);
157 static void	removechild(struct witness *parent, struct witness *child);
158 static int	isitmychild(struct witness *parent, struct witness *child);
159 static int	isitmydescendant(struct witness *parent, struct witness *child);
160 #ifdef BLESSING
161 static int	blessed(struct witness *, struct witness *);
162 #endif
163 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
164 					   struct witness *);
165 static void	witness_leveldescendents(struct witness *parent, int level);
166 static void	witness_levelall(void);
167 static struct	witness *witness_get(void);
168 static void	witness_free(struct witness *m);
169 static struct	witness_child_list_entry *witness_child_get(void);
170 static void	witness_child_free(struct witness_child_list_entry *wcl);
171 static struct	lock_list_entry *witness_lock_list_get(void);
172 static void	witness_lock_list_free(struct lock_list_entry *lle);
173 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
174 					     struct lock_object *lock);
175 static int	witness_list(struct thread *td);
176 static void	witness_list_lock(struct lock_instance *instance);
177 #if defined(DDB)
178 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
179 				     struct witness_list *list);
180 static void	witness_display(void(*)(const char *fmt, ...));
181 #endif
182 
183 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
184 
185 static int witness_watch = 1;
186 TUNABLE_INT("debug.witness_watch", &witness_watch);
187 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
188 
189 #ifdef DDB
190 /*
191  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
192  * drop into kdebug() when:
193  *	- a lock heirarchy violation occurs
194  *	- locks are held when going to sleep.
195  */
196 #ifdef WITNESS_DDB
197 int	witness_ddb = 1;
198 #else
199 int	witness_ddb = 0;
200 #endif
201 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
202 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
203 
204 /*
205  * When DDB is enabled and witness_trace is set to 1, it will cause the system
206  * to print a stack trace:
207  *	- a lock heirarchy violation occurs
208  *	- locks are held when going to sleep.
209  */
210 int	witness_trace = 1;
211 TUNABLE_INT("debug.witness_trace", &witness_trace);
212 SYSCTL_INT(_debug, OID_AUTO, witness_trace, CTLFLAG_RW, &witness_trace, 0, "");
213 #endif /* DDB */
214 
215 #ifdef WITNESS_SKIPSPIN
216 int	witness_skipspin = 1;
217 #else
218 int	witness_skipspin = 0;
219 #endif
220 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
221 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
222     "");
223 
224 static struct mtx w_mtx;
225 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
226 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
227 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
228 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
229 static struct witness_child_list_entry *w_child_free = NULL;
230 static struct lock_list_entry *w_lock_list_free = NULL;
231 static int witness_dead;	/* fatal error, probably no memory */
232 
233 static struct witness w_data[WITNESS_COUNT];
234 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
235 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
236 
237 static struct witness_order_list_entry order_lists[] = {
238 	{ "Giant", &lock_class_mtx_sleep },
239 	{ "proctree", &lock_class_sx },
240 	{ "allproc", &lock_class_sx },
241 	{ "filedesc structure", &lock_class_mtx_sleep },
242 	{ "pipe mutex", &lock_class_mtx_sleep },
243 	{ "sigio lock", &lock_class_mtx_sleep },
244 	{ "process group", &lock_class_mtx_sleep },
245 	{ "process lock", &lock_class_mtx_sleep },
246 	{ "session", &lock_class_mtx_sleep },
247 	{ "uidinfo hash", &lock_class_mtx_sleep },
248 	{ "uidinfo struct", &lock_class_mtx_sleep },
249 	{ NULL, NULL },
250 	/*
251 	 * spin locks
252 	 */
253 #ifdef SMP
254 	{ "ap boot", &lock_class_mtx_spin },
255 #ifdef __i386__
256 	{ "com", &lock_class_mtx_spin },
257 #endif
258 #endif
259 	{ "sio", &lock_class_mtx_spin },
260 #ifdef __i386__
261 	{ "cy", &lock_class_mtx_spin },
262 #endif
263 	{ "sabtty", &lock_class_mtx_spin },
264 	{ "zstty", &lock_class_mtx_spin },
265 	{ "ng_node", &lock_class_mtx_spin },
266 	{ "ng_worklist", &lock_class_mtx_spin },
267 	{ "ithread table lock", &lock_class_mtx_spin },
268 	{ "sched lock", &lock_class_mtx_spin },
269 	{ "callout", &lock_class_mtx_spin },
270 	/*
271 	 * leaf locks
272 	 */
273 	{ "allpmaps", &lock_class_mtx_spin },
274 	{ "vm page queue free mutex", &lock_class_mtx_spin },
275 	{ "icu", &lock_class_mtx_spin },
276 #ifdef SMP
277 	{ "smp rendezvous", &lock_class_mtx_spin },
278 #if defined(__i386__) && defined(APIC_IO)
279 	{ "tlb", &lock_class_mtx_spin },
280 #endif
281 #ifdef __sparc64__
282 	{ "ipi", &lock_class_mtx_spin },
283 #endif
284 #endif
285 	{ "clk", &lock_class_mtx_spin },
286 	{ "mutex profiling lock", &lock_class_mtx_spin },
287 	{ "kse zombie lock", &lock_class_mtx_spin },
288 	{ "ALD Queue", &lock_class_mtx_spin },
289 #ifdef __ia64__
290 	{ "MCA spin lock", &lock_class_mtx_spin },
291 #endif
292 #ifdef __i386__
293 	{ "pcicfg", &lock_class_mtx_spin },
294 #endif
295 	{ NULL, NULL },
296 	{ NULL, NULL }
297 };
298 
299 #ifdef BLESSING
300 /*
301  * Pairs of locks which have been blessed
302  * Don't complain about order problems with blessed locks
303  */
304 static struct witness_blessed blessed_list[] = {
305 };
306 static int blessed_count =
307 	sizeof(blessed_list) / sizeof(struct witness_blessed);
308 #endif
309 
310 /*
311  * List of all locks in the system.
312  */
313 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
314 
315 static struct mtx all_mtx = {
316 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
317 	  "All locks list",		/* mtx_object.lo_name */
318 	  "All locks list",		/* mtx_object.lo_type */
319 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
320 	  { NULL, NULL },		/* mtx_object.lo_list */
321 	  NULL },			/* mtx_object.lo_witness */
322 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
323 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
324 	{ NULL, NULL }			/* mtx_contested */
325 };
326 
327 /*
328  * This global is set to 0 once it becomes safe to use the witness code.
329  */
330 static int witness_cold = 1;
331 
332 /*
333  * Global variables for book keeping.
334  */
335 static int lock_cur_cnt;
336 static int lock_max_cnt;
337 
338 /*
339  * The WITNESS-enabled diagnostic code.
340  */
341 static void
342 witness_initialize(void *dummy __unused)
343 {
344 	struct lock_object *lock;
345 	struct witness_order_list_entry *order;
346 	struct witness *w, *w1;
347 	int i;
348 
349 	/*
350 	 * We have to release Giant before initializing its witness
351 	 * structure so that WITNESS doesn't get confused.
352 	 */
353 	mtx_unlock(&Giant);
354 	mtx_assert(&Giant, MA_NOTOWNED);
355 
356 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
357 	TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
358 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
359 	    MTX_NOWITNESS);
360 	for (i = 0; i < WITNESS_COUNT; i++)
361 		witness_free(&w_data[i]);
362 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
363 		witness_child_free(&w_childdata[i]);
364 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
365 		witness_lock_list_free(&w_locklistdata[i]);
366 
367 	/* First add in all the specified order lists. */
368 	for (order = order_lists; order->w_name != NULL; order++) {
369 		w = enroll(order->w_name, order->w_class);
370 		if (w == NULL)
371 			continue;
372 		w->w_file = "order list";
373 		for (order++; order->w_name != NULL; order++) {
374 			w1 = enroll(order->w_name, order->w_class);
375 			if (w1 == NULL)
376 				continue;
377 			w1->w_file = "order list";
378 			itismychild(w, w1);
379 			w = w1;
380 		}
381 	}
382 
383 	/* Iterate through all locks and add them to witness. */
384 	mtx_lock(&all_mtx);
385 	TAILQ_FOREACH(lock, &all_locks, lo_list) {
386 		if (lock->lo_flags & LO_WITNESS)
387 			lock->lo_witness = enroll(lock->lo_type,
388 			    lock->lo_class);
389 		else
390 			lock->lo_witness = NULL;
391 	}
392 	mtx_unlock(&all_mtx);
393 
394 	/* Mark the witness code as being ready for use. */
395 	atomic_store_rel_int(&witness_cold, 0);
396 
397 	mtx_lock(&Giant);
398 }
399 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
400 
401 void
402 witness_init(struct lock_object *lock)
403 {
404 	struct lock_class *class;
405 
406 	class = lock->lo_class;
407 	if (lock->lo_flags & LO_INITIALIZED)
408 		panic("%s: lock (%s) %s is already initialized", __func__,
409 		    class->lc_name, lock->lo_name);
410 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
411 	    (class->lc_flags & LC_RECURSABLE) == 0)
412 		panic("%s: lock (%s) %s can not be recursable", __func__,
413 		    class->lc_name, lock->lo_name);
414 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
415 	    (class->lc_flags & LC_SLEEPABLE) == 0)
416 		panic("%s: lock (%s) %s can not be sleepable", __func__,
417 		    class->lc_name, lock->lo_name);
418 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
419 	    (class->lc_flags & LC_UPGRADABLE) == 0)
420 		panic("%s: lock (%s) %s can not be upgradable", __func__,
421 		    class->lc_name, lock->lo_name);
422 
423 	mtx_lock(&all_mtx);
424 	TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
425 	lock->lo_flags |= LO_INITIALIZED;
426 	lock_cur_cnt++;
427 	if (lock_cur_cnt > lock_max_cnt)
428 		lock_max_cnt = lock_cur_cnt;
429 	mtx_unlock(&all_mtx);
430 	if (!witness_cold && !witness_dead && panicstr == NULL &&
431 	    (lock->lo_flags & LO_WITNESS) != 0)
432 		lock->lo_witness = enroll(lock->lo_type, class);
433 	else
434 		lock->lo_witness = NULL;
435 }
436 
437 void
438 witness_destroy(struct lock_object *lock)
439 {
440 	struct witness *w;
441 
442 	if (witness_cold)
443 		panic("lock (%s) %s destroyed while witness_cold",
444 		    lock->lo_class->lc_name, lock->lo_name);
445 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
446 		panic("%s: lock (%s) %s is not initialized", __func__,
447 		    lock->lo_class->lc_name, lock->lo_name);
448 
449 	/* XXX: need to verify that no one holds the lock */
450 	w = lock->lo_witness;
451 	if (w != NULL) {
452 		mtx_lock_spin(&w_mtx);
453 		MPASS(w->w_refcount > 0);
454 		w->w_refcount--;
455 		mtx_unlock_spin(&w_mtx);
456 	}
457 
458 	mtx_lock(&all_mtx);
459 	lock_cur_cnt--;
460 	TAILQ_REMOVE(&all_locks, lock, lo_list);
461 	lock->lo_flags &= ~LO_INITIALIZED;
462 	mtx_unlock(&all_mtx);
463 }
464 
465 #if defined(DDB)
466 static void
467 witness_display_list(void(*prnt)(const char *fmt, ...),
468 		     struct witness_list *list)
469 {
470 	struct witness *w, *w1;
471 	int found;
472 
473 	STAILQ_FOREACH(w, list, w_typelist) {
474 		if (w->w_file == NULL)
475 			continue;
476 		found = 0;
477 		STAILQ_FOREACH(w1, list, w_typelist) {
478 			if (isitmychild(w1, w)) {
479 				found++;
480 				break;
481 			}
482 		}
483 		if (found)
484 			continue;
485 		/*
486 		 * This lock has no anscestors, display its descendants.
487 		 */
488 		witness_displaydescendants(prnt, w);
489 	}
490 }
491 
492 static void
493 witness_display(void(*prnt)(const char *fmt, ...))
494 {
495 	struct witness *w;
496 
497 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
498 	witness_levelall();
499 
500 	/*
501 	 * First, handle sleep locks which have been acquired at least
502 	 * once.
503 	 */
504 	prnt("Sleep locks:\n");
505 	witness_display_list(prnt, &w_sleep);
506 
507 	/*
508 	 * Now do spin locks which have been acquired at least once.
509 	 */
510 	prnt("\nSpin locks:\n");
511 	witness_display_list(prnt, &w_spin);
512 
513 	/*
514 	 * Finally, any locks which have not been acquired yet.
515 	 */
516 	prnt("\nLocks which were never acquired:\n");
517 	STAILQ_FOREACH(w, &w_all, w_list) {
518 		if (w->w_file != NULL || w->w_refcount == 0)
519 			continue;
520 		prnt("%s\n", w->w_name);
521 	}
522 }
523 #endif
524 
525 void
526 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
527 {
528 	struct lock_list_entry **lock_list, *lle;
529 	struct lock_instance *lock1, *lock2;
530 	struct lock_class *class;
531 	struct witness *w, *w1;
532 	struct thread *td;
533 	int i, j;
534 #ifdef DDB
535 	int go_into_ddb = 0;
536 #endif /* DDB */
537 
538 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
539 	    panicstr != NULL)
540 		return;
541 	w = lock->lo_witness;
542 	class = lock->lo_class;
543 	td = curthread;
544 
545 	if (class->lc_flags & LC_SLEEPLOCK) {
546 		/*
547 		 * Since spin locks include a critical section, this check
548 		 * impliclty enforces a lock order of all sleep locks before
549 		 * all spin locks.
550 		 */
551 		if (td->td_critnest != 0 && (flags & LOP_TRYLOCK) == 0)
552 			panic("blockable sleep lock (%s) %s @ %s:%d",
553 			    class->lc_name, lock->lo_name, file, line);
554 		lock_list = &td->td_sleeplocks;
555 	} else
556 		lock_list = PCPU_PTR(spinlocks);
557 
558 	/*
559 	 * Try locks do not block if they fail to acquire the lock, thus
560 	 * there is no danger of deadlocks or of switching while holding a
561 	 * spin lock if we acquire a lock via a try operation.
562 	 */
563 	if (flags & LOP_TRYLOCK)
564 		goto out;
565 
566 	/*
567 	 * Is this the first lock acquired?  If so, then no order checking
568 	 * is needed.
569 	 */
570 	if (*lock_list == NULL)
571 		goto out;
572 
573 	/*
574 	 * Check to see if we are recursing on a lock we already own.
575 	 */
576 	lock1 = find_instance(*lock_list, lock);
577 	if (lock1 != NULL) {
578 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
579 		    (flags & LOP_EXCLUSIVE) == 0) {
580 			printf("shared lock of (%s) %s @ %s:%d\n",
581 			    class->lc_name, lock->lo_name, file, line);
582 			printf("while exclusively locked from %s:%d\n",
583 			    lock1->li_file, lock1->li_line);
584 			panic("share->excl");
585 		}
586 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
587 		    (flags & LOP_EXCLUSIVE) != 0) {
588 			printf("exclusive lock of (%s) %s @ %s:%d\n",
589 			    class->lc_name, lock->lo_name, file, line);
590 			printf("while share locked from %s:%d\n",
591 			    lock1->li_file, lock1->li_line);
592 			panic("excl->share");
593 		}
594 		lock1->li_flags++;
595 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
596 			printf(
597 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
598 			    class->lc_name, lock->lo_name, file, line);
599 			printf("first acquired @ %s:%d\n", lock1->li_file,
600 			    lock1->li_line);
601 			panic("recurse");
602 		}
603 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
604 		    td->td_proc->p_pid, lock->lo_name,
605 		    lock1->li_flags & LI_RECURSEMASK);
606 		lock1->li_file = file;
607 		lock1->li_line = line;
608 		return;
609 	}
610 
611 	/*
612 	 * Check for duplicate locks of the same type.  Note that we only
613 	 * have to check for this on the last lock we just acquired.  Any
614 	 * other cases will be caught as lock order violations.
615 	 */
616 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
617 	w1 = lock1->li_lock->lo_witness;
618 	if (w1 == w) {
619 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
620 			goto out;
621 		w->w_same_squawked = 1;
622 		printf("acquiring duplicate lock of same type: \"%s\"\n",
623 			lock->lo_type);
624 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
625 		    lock1->li_file, lock1->li_line);
626 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
627 #ifdef DDB
628 		go_into_ddb = 1;
629 #endif /* DDB */
630 		goto out;
631 	}
632 	MPASS(!mtx_owned(&w_mtx));
633 	mtx_lock_spin(&w_mtx);
634 	/*
635 	 * If we have a known higher number just say ok
636 	 */
637 	if (witness_watch > 1 && w->w_level > w1->w_level) {
638 		mtx_unlock_spin(&w_mtx);
639 		goto out;
640 	}
641 	/*
642 	 * If we know that the the lock we are acquiring comes after
643 	 * the lock we most recently acquired in the lock order tree,
644 	 * then there is no need for any further checks.
645 	 */
646 	if (isitmydescendant(w1, w)) {
647 		mtx_unlock_spin(&w_mtx);
648 		goto out;
649 	}
650 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
651 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
652 
653 			MPASS(j < WITNESS_COUNT);
654 			lock1 = &lle->ll_children[i];
655 			w1 = lock1->li_lock->lo_witness;
656 
657 			/*
658 			 * If this lock doesn't undergo witness checking,
659 			 * then skip it.
660 			 */
661 			if (w1 == NULL) {
662 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
663 				    ("lock missing witness structure"));
664 				continue;
665 			}
666 			/*
667 			 * If we are locking Giant and this is a sleepable
668 			 * lock, then skip it.
669 			 */
670 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
671 			    lock == &Giant.mtx_object)
672 				continue;
673 			/*
674 			 * If we are locking a sleepable lock and this lock
675 			 * is Giant, then skip it.
676 			 */
677 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
678 			    lock1->li_lock == &Giant.mtx_object)
679 				continue;
680 			/*
681 			 * If we are locking a sleepable lock and this lock
682 			 * isn't sleepable, we want to treat it as a lock
683 			 * order violation to enfore a general lock order of
684 			 * sleepable locks before non-sleepable locks.
685 			 */
686 			if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
687 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
688 			    /*
689 			     * Check the lock order hierarchy for a reveresal.
690 			     */
691 			    if (!isitmydescendant(w, w1))
692 				continue;
693 			/*
694 			 * We have a lock order violation, check to see if it
695 			 * is allowed or has already been yelled about.
696 			 */
697 			mtx_unlock_spin(&w_mtx);
698 #ifdef BLESSING
699 			if (blessed(w, w1))
700 				goto out;
701 #endif
702 			if (lock1->li_lock == &Giant.mtx_object) {
703 				if (w1->w_Giant_squawked)
704 					goto out;
705 				else
706 					w1->w_Giant_squawked = 1;
707 			} else {
708 				if (w1->w_other_squawked)
709 					goto out;
710 				else
711 					w1->w_other_squawked = 1;
712 			}
713 			/*
714 			 * Ok, yell about it.
715 			 */
716 			printf("lock order reversal\n");
717 			/*
718 			 * Try to locate an earlier lock with
719 			 * witness w in our list.
720 			 */
721 			do {
722 				lock2 = &lle->ll_children[i];
723 				MPASS(lock2->li_lock != NULL);
724 				if (lock2->li_lock->lo_witness == w)
725 					break;
726 				i--;
727 				if (i == 0 && lle->ll_next != NULL) {
728 					lle = lle->ll_next;
729 					i = lle->ll_count - 1;
730 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
731 				}
732 			} while (i >= 0);
733 			if (i < 0) {
734 				printf(" 1st %p %s (%s) @ %s:%d\n",
735 				    lock1->li_lock, lock1->li_lock->lo_name,
736 				    lock1->li_lock->lo_type, lock1->li_file,
737 				    lock1->li_line);
738 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
739 				    lock->lo_name, lock->lo_type, file, line);
740 			} else {
741 				printf(" 1st %p %s (%s) @ %s:%d\n",
742 				    lock2->li_lock, lock2->li_lock->lo_name,
743 				    lock2->li_lock->lo_type, lock2->li_file,
744 				    lock2->li_line);
745 				printf(" 2nd %p %s (%s) @ %s:%d\n",
746 				    lock1->li_lock, lock1->li_lock->lo_name,
747 				    lock1->li_lock->lo_type, lock1->li_file,
748 				    lock1->li_line);
749 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
750 				    lock->lo_name, lock->lo_type, file, line);
751 			}
752 #ifdef DDB
753 			go_into_ddb = 1;
754 #endif /* DDB */
755 			goto out;
756 		}
757 	}
758 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
759 	/*
760 	 * Don't build a new relationship between a sleepable lock and
761 	 * Giant if it is the wrong direction.  The real lock order is that
762 	 * sleepable locks come before Giant.
763 	 */
764 	if (lock1->li_lock == &Giant.mtx_object &&
765 	    (lock->lo_flags & LO_SLEEPABLE) != 0)
766 		mtx_unlock_spin(&w_mtx);
767 	else {
768 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
769 		    lock->lo_type, lock1->li_lock->lo_type);
770 		if (!itismychild(lock1->li_lock->lo_witness, w))
771 			mtx_unlock_spin(&w_mtx);
772 	}
773 
774 out:
775 #ifdef DDB
776 	if (go_into_ddb) {
777 		if (witness_trace)
778 			backtrace();
779 		if (witness_ddb)
780 			Debugger(__func__);
781 	}
782 #endif /* DDB */
783 	w->w_file = file;
784 	w->w_line = line;
785 
786 	lle = *lock_list;
787 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
788 		lle = witness_lock_list_get();
789 		if (lle == NULL)
790 			return;
791 		lle->ll_next = *lock_list;
792 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
793 		    td->td_proc->p_pid, lle);
794 		*lock_list = lle;
795 	}
796 	lock1 = &lle->ll_children[lle->ll_count++];
797 	lock1->li_lock = lock;
798 	lock1->li_line = line;
799 	lock1->li_file = file;
800 	if ((flags & LOP_EXCLUSIVE) != 0)
801 		lock1->li_flags = LI_EXCLUSIVE;
802 	else
803 		lock1->li_flags = 0;
804 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
805 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
806 }
807 
808 void
809 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
810 {
811 	struct lock_instance *instance;
812 	struct lock_class *class;
813 
814 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
815 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
816 		return;
817 	class = lock->lo_class;
818 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
819 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
820 		    class->lc_name, lock->lo_name, file, line);
821 	if ((flags & LOP_TRYLOCK) == 0)
822 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
823 		    lock->lo_name, file, line);
824 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
825 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
826 		    class->lc_name, lock->lo_name, file, line);
827 	instance = find_instance(curthread->td_sleeplocks, lock);
828 	if (instance == NULL)
829 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
830 		    class->lc_name, lock->lo_name, file, line);
831 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
832 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
833 		    class->lc_name, lock->lo_name, file, line);
834 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
835 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
836 		    class->lc_name, lock->lo_name,
837 		    instance->li_flags & LI_RECURSEMASK, file, line);
838 	instance->li_flags |= LI_EXCLUSIVE;
839 }
840 
841 void
842 witness_downgrade(struct lock_object *lock, int flags, const char *file,
843     int line)
844 {
845 	struct lock_instance *instance;
846 	struct lock_class *class;
847 
848 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
849 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
850 		return;
851 	class = lock->lo_class;
852 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
853 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
854 		    class->lc_name, lock->lo_name, file, line);
855 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
856 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
857 		    class->lc_name, lock->lo_name, file, line);
858 	instance = find_instance(curthread->td_sleeplocks, lock);
859 	if (instance == NULL)
860 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
861 		    class->lc_name, lock->lo_name, file, line);
862 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
863 		panic("downgrade of shared lock (%s) %s @ %s:%d",
864 		    class->lc_name, lock->lo_name, file, line);
865 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
866 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
867 		    class->lc_name, lock->lo_name,
868 		    instance->li_flags & LI_RECURSEMASK, file, line);
869 	instance->li_flags &= ~LI_EXCLUSIVE;
870 }
871 
872 void
873 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
874 {
875 	struct lock_list_entry **lock_list, *lle;
876 	struct lock_instance *instance;
877 	struct lock_class *class;
878 	struct thread *td;
879 	register_t s;
880 	int i, j;
881 
882 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
883 	    panicstr != NULL)
884 		return;
885 	td = curthread;
886 	class = lock->lo_class;
887 	if (class->lc_flags & LC_SLEEPLOCK)
888 		lock_list = &td->td_sleeplocks;
889 	else
890 		lock_list = PCPU_PTR(spinlocks);
891 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
892 		for (i = 0; i < (*lock_list)->ll_count; i++) {
893 			instance = &(*lock_list)->ll_children[i];
894 			if (instance->li_lock == lock) {
895 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
896 				    (flags & LOP_EXCLUSIVE) == 0) {
897 					printf(
898 					"shared unlock of (%s) %s @ %s:%d\n",
899 					    class->lc_name, lock->lo_name,
900 					    file, line);
901 					printf(
902 					"while exclusively locked from %s:%d\n",
903 					    instance->li_file,
904 					    instance->li_line);
905 					panic("excl->ushare");
906 				}
907 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
908 				    (flags & LOP_EXCLUSIVE) != 0) {
909 					printf(
910 					"exclusive unlock of (%s) %s @ %s:%d\n",
911 					    class->lc_name, lock->lo_name,
912 					    file, line);
913 					printf(
914 					"while share locked from %s:%d\n",
915 					    instance->li_file,
916 					    instance->li_line);
917 					panic("share->uexcl");
918 				}
919 				/* If we are recursed, unrecurse. */
920 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
921 					CTR4(KTR_WITNESS,
922 				    "%s: pid %d unrecursed on %s r=%d", __func__,
923 					    td->td_proc->p_pid,
924 					    instance->li_lock->lo_name,
925 					    instance->li_flags);
926 					instance->li_flags--;
927 					return;
928 				}
929 				s = intr_disable();
930 				CTR4(KTR_WITNESS,
931 				    "%s: pid %d removed %s from lle[%d]", __func__,
932 				    td->td_proc->p_pid,
933 				    instance->li_lock->lo_name,
934 				    (*lock_list)->ll_count - 1);
935 				for (j = i; j < (*lock_list)->ll_count - 1; j++)
936 					(*lock_list)->ll_children[j] =
937 					    (*lock_list)->ll_children[j + 1];
938 				(*lock_list)->ll_count--;
939 				intr_restore(s);
940 				if ((*lock_list)->ll_count == 0) {
941 					lle = *lock_list;
942 					*lock_list = lle->ll_next;
943 					CTR3(KTR_WITNESS,
944 					    "%s: pid %d removed lle %p", __func__,
945 					    td->td_proc->p_pid, lle);
946 					witness_lock_list_free(lle);
947 				}
948 				return;
949 			}
950 		}
951 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
952 	    file, line);
953 }
954 
955 /*
956  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
957  * exempt Giant and sleepable locks from the checks as well.  If any
958  * non-exempt locks are held, then a supplied message is printed to the
959  * console along with a list of the offending locks.  If indicated in the
960  * flags then a failure results in a panic as well.
961  */
962 int
963 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
964 {
965 	struct lock_list_entry *lle;
966 	struct lock_instance *lock1;
967 	struct thread *td;
968 	va_list ap;
969 	int i, n;
970 
971 	if (witness_cold || witness_dead || panicstr != NULL)
972 		return (0);
973 	n = 0;
974 	td = curthread;
975 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
976 		for (i = lle->ll_count - 1; i >= 0; i--) {
977 			lock1 = &lle->ll_children[i];
978 			if (lock1->li_lock == lock)
979 				continue;
980 			if (flags & WARN_GIANTOK &&
981 			    lock1->li_lock == &Giant.mtx_object)
982 				continue;
983 			if (flags & WARN_SLEEPOK &&
984 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
985 				continue;
986 			if (n == 0) {
987 				va_start(ap, fmt);
988 				vprintf(fmt, ap);
989 				va_end(ap);
990 				printf(" with the following");
991 				if (flags & WARN_SLEEPOK)
992 					printf(" non-sleepable");
993 				printf("locks held:\n");
994 			}
995 			n++;
996 			witness_list_lock(lock1);
997 		}
998 	if (PCPU_GET(spinlocks) != NULL) {
999 		/*
1000 		 * Since we already hold a spinlock preemption is
1001 		 * already blocked.
1002 		 */
1003 		if (n == 0) {
1004 			va_start(ap, fmt);
1005 			vprintf(fmt, ap);
1006 			va_end(ap);
1007 			printf(" with the following");
1008 			if (flags & WARN_SLEEPOK)
1009 				printf(" non-sleepable");
1010 			printf("locks held:\n");
1011 		}
1012 		n += witness_list_locks(PCPU_PTR(spinlocks));
1013 	}
1014 	if (flags & WARN_PANIC && n)
1015 		panic("witness_warn");
1016 #ifdef DDB
1017 	else if (witness_ddb && n)
1018 		Debugger(__func__);
1019 #endif
1020 	return (n);
1021 }
1022 
1023 const char *
1024 witness_file(struct lock_object *lock)
1025 {
1026 	struct witness *w;
1027 
1028 	if (witness_cold || witness_dead || lock->lo_witness == NULL)
1029 		return ("?");
1030 	w = lock->lo_witness;
1031 	return (w->w_file);
1032 }
1033 
1034 int
1035 witness_line(struct lock_object *lock)
1036 {
1037 	struct witness *w;
1038 
1039 	if (witness_cold || witness_dead || lock->lo_witness == NULL)
1040 		return (0);
1041 	w = lock->lo_witness;
1042 	return (w->w_line);
1043 }
1044 
1045 static struct witness *
1046 enroll(const char *description, struct lock_class *lock_class)
1047 {
1048 	struct witness *w;
1049 
1050 	if (!witness_watch || witness_dead || panicstr != NULL)
1051 		return (NULL);
1052 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1053 		return (NULL);
1054 	mtx_lock_spin(&w_mtx);
1055 	STAILQ_FOREACH(w, &w_all, w_list) {
1056 		if (w->w_name == description || (w->w_refcount > 0 &&
1057 		    strcmp(description, w->w_name) == 0)) {
1058 			w->w_refcount++;
1059 			mtx_unlock_spin(&w_mtx);
1060 			if (lock_class != w->w_class)
1061 				panic(
1062 				"lock (%s) %s does not match earlier (%s) lock",
1063 				    description, lock_class->lc_name,
1064 				    w->w_class->lc_name);
1065 			return (w);
1066 		}
1067 	}
1068 	/*
1069 	 * This isn't quite right, as witness_cold is still 0 while we
1070 	 * enroll all the locks initialized before witness_initialize().
1071 	 */
1072 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1073 		mtx_unlock_spin(&w_mtx);
1074 		panic("spin lock %s not in order list", description);
1075 	}
1076 	if ((w = witness_get()) == NULL)
1077 		return (NULL);
1078 	w->w_name = description;
1079 	w->w_class = lock_class;
1080 	w->w_refcount = 1;
1081 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1082 	if (lock_class->lc_flags & LC_SPINLOCK)
1083 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1084 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
1085 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1086 	else {
1087 		mtx_unlock_spin(&w_mtx);
1088 		panic("lock class %s is not sleep or spin",
1089 		    lock_class->lc_name);
1090 	}
1091 	mtx_unlock_spin(&w_mtx);
1092 	return (w);
1093 }
1094 
1095 static int
1096 itismychild(struct witness *parent, struct witness *child)
1097 {
1098 	static int recursed;
1099 	struct witness_child_list_entry **wcl;
1100 	struct witness_list *list;
1101 
1102 	MPASS(child != NULL && parent != NULL);
1103 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1104 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1105 		panic(
1106 		"%s: parent (%s) and child (%s) are not the same lock type",
1107 		    __func__, parent->w_class->lc_name,
1108 		    child->w_class->lc_name);
1109 
1110 	/*
1111 	 * Insert "child" after "parent"
1112 	 */
1113 	wcl = &parent->w_children;
1114 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1115 		wcl = &(*wcl)->wcl_next;
1116 	if (*wcl == NULL) {
1117 		*wcl = witness_child_get();
1118 		if (*wcl == NULL)
1119 			return (1);
1120 	}
1121 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1122 
1123 	/*
1124 	 * Now prune whole tree.  We look for cases where a lock is now
1125 	 * both a descendant and a direct child of a given lock.  In that
1126 	 * case, we want to remove the direct child link from the tree.
1127 	 */
1128 	if (recursed)
1129 		return (0);
1130 	recursed = 1;
1131 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1132 		list = &w_sleep;
1133 	else
1134 		list = &w_spin;
1135 	STAILQ_FOREACH(child, list, w_typelist) {
1136 		STAILQ_FOREACH(parent, list, w_typelist) {
1137 			if (!isitmychild(parent, child))
1138 				continue;
1139 			removechild(parent, child);
1140 			if (isitmydescendant(parent, child))
1141 				continue;
1142 			itismychild(parent, child);
1143 		}
1144 	}
1145 	recursed = 0;
1146 	witness_levelall();
1147 	return (0);
1148 }
1149 
1150 static void
1151 removechild(struct witness *parent, struct witness *child)
1152 {
1153 	struct witness_child_list_entry **wcl, *wcl1;
1154 	int i;
1155 
1156 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1157 		for (i = 0; i < (*wcl)->wcl_count; i++)
1158 			if ((*wcl)->wcl_children[i] == child)
1159 				goto found;
1160 	return;
1161 found:
1162 	(*wcl)->wcl_count--;
1163 	if ((*wcl)->wcl_count > i)
1164 		(*wcl)->wcl_children[i] =
1165 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1166 	MPASS((*wcl)->wcl_children[i] != NULL);
1167 	if ((*wcl)->wcl_count != 0)
1168 		return;
1169 	wcl1 = *wcl;
1170 	*wcl = wcl1->wcl_next;
1171 	witness_child_free(wcl1);
1172 }
1173 
1174 static int
1175 isitmychild(struct witness *parent, struct witness *child)
1176 {
1177 	struct witness_child_list_entry *wcl;
1178 	int i;
1179 
1180 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1181 		for (i = 0; i < wcl->wcl_count; i++) {
1182 			if (wcl->wcl_children[i] == child)
1183 				return (1);
1184 		}
1185 	}
1186 	return (0);
1187 }
1188 
1189 static int
1190 isitmydescendant(struct witness *parent, struct witness *child)
1191 {
1192 	struct witness_child_list_entry *wcl;
1193 	int i, j;
1194 
1195 	if (isitmychild(parent, child))
1196 		return (1);
1197 	j = 0;
1198 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1199 		MPASS(j < 1000);
1200 		for (i = 0; i < wcl->wcl_count; i++) {
1201 			if (isitmydescendant(wcl->wcl_children[i], child))
1202 				return (1);
1203 		}
1204 		j++;
1205 	}
1206 	return (0);
1207 }
1208 
1209 static void
1210 witness_levelall (void)
1211 {
1212 	struct witness_list *list;
1213 	struct witness *w, *w1;
1214 
1215 	/*
1216 	 * First clear all levels.
1217 	 */
1218 	STAILQ_FOREACH(w, &w_all, w_list) {
1219 		w->w_level = 0;
1220 	}
1221 
1222 	/*
1223 	 * Look for locks with no parent and level all their descendants.
1224 	 */
1225 	STAILQ_FOREACH(w, &w_all, w_list) {
1226 		/*
1227 		 * This is just an optimization, technically we could get
1228 		 * away just walking the all list each time.
1229 		 */
1230 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1231 			list = &w_sleep;
1232 		else
1233 			list = &w_spin;
1234 		STAILQ_FOREACH(w1, list, w_typelist) {
1235 			if (isitmychild(w1, w))
1236 				goto skip;
1237 		}
1238 		witness_leveldescendents(w, 0);
1239 	skip:
1240 		;	/* silence GCC 3.x */
1241 	}
1242 }
1243 
1244 static void
1245 witness_leveldescendents(struct witness *parent, int level)
1246 {
1247 	struct witness_child_list_entry *wcl;
1248 	int i;
1249 
1250 	if (parent->w_level < level)
1251 		parent->w_level = level;
1252 	level++;
1253 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1254 		for (i = 0; i < wcl->wcl_count; i++)
1255 			witness_leveldescendents(wcl->wcl_children[i], level);
1256 }
1257 
1258 static void
1259 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1260 			   struct witness *parent)
1261 {
1262 	struct witness_child_list_entry *wcl;
1263 	int i, level;
1264 
1265 	level = parent->w_level;
1266 	prnt("%-2d", level);
1267 	for (i = 0; i < level; i++)
1268 		prnt(" ");
1269 	if (parent->w_refcount > 0) {
1270 		prnt("%s", parent->w_name);
1271 		if (parent->w_file != NULL)
1272 			prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1273 			    parent->w_line);
1274 	} else
1275 		prnt("(dead)\n");
1276 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1277 		for (i = 0; i < wcl->wcl_count; i++)
1278 			    witness_displaydescendants(prnt,
1279 				wcl->wcl_children[i]);
1280 }
1281 
1282 #ifdef BLESSING
1283 static int
1284 blessed(struct witness *w1, struct witness *w2)
1285 {
1286 	int i;
1287 	struct witness_blessed *b;
1288 
1289 	for (i = 0; i < blessed_count; i++) {
1290 		b = &blessed_list[i];
1291 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1292 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1293 				return (1);
1294 			continue;
1295 		}
1296 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1297 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1298 				return (1);
1299 	}
1300 	return (0);
1301 }
1302 #endif
1303 
1304 static struct witness *
1305 witness_get(void)
1306 {
1307 	struct witness *w;
1308 
1309 	if (witness_dead) {
1310 		mtx_unlock_spin(&w_mtx);
1311 		return (NULL);
1312 	}
1313 	if (STAILQ_EMPTY(&w_free)) {
1314 		witness_dead = 1;
1315 		mtx_unlock_spin(&w_mtx);
1316 		printf("%s: witness exhausted\n", __func__);
1317 		return (NULL);
1318 	}
1319 	w = STAILQ_FIRST(&w_free);
1320 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1321 	bzero(w, sizeof(*w));
1322 	return (w);
1323 }
1324 
1325 static void
1326 witness_free(struct witness *w)
1327 {
1328 
1329 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1330 }
1331 
1332 static struct witness_child_list_entry *
1333 witness_child_get(void)
1334 {
1335 	struct witness_child_list_entry *wcl;
1336 
1337 	if (witness_dead) {
1338 		mtx_unlock_spin(&w_mtx);
1339 		return (NULL);
1340 	}
1341 	wcl = w_child_free;
1342 	if (wcl == NULL) {
1343 		witness_dead = 1;
1344 		mtx_unlock_spin(&w_mtx);
1345 		printf("%s: witness exhausted\n", __func__);
1346 		return (NULL);
1347 	}
1348 	w_child_free = wcl->wcl_next;
1349 	bzero(wcl, sizeof(*wcl));
1350 	return (wcl);
1351 }
1352 
1353 static void
1354 witness_child_free(struct witness_child_list_entry *wcl)
1355 {
1356 
1357 	wcl->wcl_next = w_child_free;
1358 	w_child_free = wcl;
1359 }
1360 
1361 static struct lock_list_entry *
1362 witness_lock_list_get(void)
1363 {
1364 	struct lock_list_entry *lle;
1365 
1366 	if (witness_dead)
1367 		return (NULL);
1368 	mtx_lock_spin(&w_mtx);
1369 	lle = w_lock_list_free;
1370 	if (lle == NULL) {
1371 		witness_dead = 1;
1372 		mtx_unlock_spin(&w_mtx);
1373 		printf("%s: witness exhausted\n", __func__);
1374 		return (NULL);
1375 	}
1376 	w_lock_list_free = lle->ll_next;
1377 	mtx_unlock_spin(&w_mtx);
1378 	bzero(lle, sizeof(*lle));
1379 	return (lle);
1380 }
1381 
1382 static void
1383 witness_lock_list_free(struct lock_list_entry *lle)
1384 {
1385 
1386 	mtx_lock_spin(&w_mtx);
1387 	lle->ll_next = w_lock_list_free;
1388 	w_lock_list_free = lle;
1389 	mtx_unlock_spin(&w_mtx);
1390 }
1391 
1392 static struct lock_instance *
1393 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1394 {
1395 	struct lock_list_entry *lle;
1396 	struct lock_instance *instance;
1397 	int i;
1398 
1399 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1400 		for (i = lle->ll_count - 1; i >= 0; i--) {
1401 			instance = &lle->ll_children[i];
1402 			if (instance->li_lock == lock)
1403 				return (instance);
1404 		}
1405 	return (NULL);
1406 }
1407 
1408 static void
1409 witness_list_lock(struct lock_instance *instance)
1410 {
1411 	struct lock_object *lock;
1412 
1413 	lock = instance->li_lock;
1414 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1415 	    "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1416 	if (lock->lo_type != lock->lo_name)
1417 		printf(" (%s)", lock->lo_type);
1418 	printf(" r = %d (%p) locked @ %s:%d\n",
1419 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1420 	    instance->li_line);
1421 }
1422 
1423 int
1424 witness_list_locks(struct lock_list_entry **lock_list)
1425 {
1426 	struct lock_list_entry *lle;
1427 	int i, nheld;
1428 
1429 	nheld = 0;
1430 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1431 		for (i = lle->ll_count - 1; i >= 0; i--) {
1432 			witness_list_lock(&lle->ll_children[i]);
1433 			nheld++;
1434 		}
1435 	return (nheld);
1436 }
1437 
1438 /*
1439  * Calling this on td != curthread is bad unless we are in ddb.
1440  */
1441 static int
1442 witness_list(struct thread *td)
1443 {
1444 	int nheld;
1445 
1446 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1447 #ifdef DDB
1448 	KASSERT(td == curthread || db_active,
1449 	    ("%s: td != curthread and we aren't in the debugger", __func__));
1450 	if (!db_active && witness_dead)
1451 		return (0);
1452 #else
1453 	KASSERT(td == curthread, ("%s: p != curthread", __func__));
1454 	if (witness_dead)
1455 		return (0);
1456 #endif
1457 	nheld = witness_list_locks(&td->td_sleeplocks);
1458 
1459 	/*
1460 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1461 	 * if td is currently executing on some other CPU and holds spin locks
1462 	 * as we won't display those locks.  If we had a MI way of getting
1463 	 * the per-cpu data for a given cpu then we could use
1464 	 * td->td_kse->ke_oncpu to get the list of spinlocks for this thread
1465 	 * and "fix" this.
1466 	 *
1467 	 * That still wouldn't really fix this unless we locked sched_lock
1468 	 * or stopped the other CPU to make sure it wasn't changing the list
1469 	 * out from under us.  It is probably best to just not try to handle
1470 	 * threads on other CPU's for now.
1471 	 */
1472 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1473 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1474 
1475 	return (nheld);
1476 }
1477 
1478 void
1479 witness_save(struct lock_object *lock, const char **filep, int *linep)
1480 {
1481 	struct lock_instance *instance;
1482 
1483 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1484 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1485 		return;
1486 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1487 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1488 		    lock->lo_class->lc_name, lock->lo_name);
1489 	instance = find_instance(curthread->td_sleeplocks, lock);
1490 	if (instance == NULL)
1491 		panic("%s: lock (%s) %s not locked", __func__,
1492 		    lock->lo_class->lc_name, lock->lo_name);
1493 	*filep = instance->li_file;
1494 	*linep = instance->li_line;
1495 }
1496 
1497 void
1498 witness_restore(struct lock_object *lock, const char *file, int line)
1499 {
1500 	struct lock_instance *instance;
1501 
1502 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1503 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1504 		return;
1505 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1506 		panic("%s: lock (%s) %s is not a sleep lock", __func__,
1507 		    lock->lo_class->lc_name, lock->lo_name);
1508 	instance = find_instance(curthread->td_sleeplocks, lock);
1509 	if (instance == NULL)
1510 		panic("%s: lock (%s) %s not locked", __func__,
1511 		    lock->lo_class->lc_name, lock->lo_name);
1512 	lock->lo_witness->w_file = file;
1513 	lock->lo_witness->w_line = line;
1514 	instance->li_file = file;
1515 	instance->li_line = line;
1516 }
1517 
1518 void
1519 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1520 {
1521 #ifdef INVARIANT_SUPPORT
1522 	struct lock_instance *instance;
1523 
1524 	if (lock->lo_witness == NULL || witness_dead || panicstr != NULL)
1525 		return;
1526 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1527 		instance = find_instance(curthread->td_sleeplocks, lock);
1528 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1529 		instance = find_instance(PCPU_GET(spinlocks), lock);
1530 	else {
1531 		panic("Lock (%s) %s is not sleep or spin!",
1532 		    lock->lo_class->lc_name, lock->lo_name);
1533 		return;
1534 	}
1535 	switch (flags) {
1536 	case LA_UNLOCKED:
1537 		if (instance != NULL)
1538 			panic("Lock (%s) %s locked @ %s:%d.",
1539 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1540 		break;
1541 	case LA_LOCKED:
1542 	case LA_LOCKED | LA_RECURSED:
1543 	case LA_LOCKED | LA_NOTRECURSED:
1544 	case LA_SLOCKED:
1545 	case LA_SLOCKED | LA_RECURSED:
1546 	case LA_SLOCKED | LA_NOTRECURSED:
1547 	case LA_XLOCKED:
1548 	case LA_XLOCKED | LA_RECURSED:
1549 	case LA_XLOCKED | LA_NOTRECURSED:
1550 		if (instance == NULL) {
1551 			panic("Lock (%s) %s not locked @ %s:%d.",
1552 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1553 			break;
1554 		}
1555 		if ((flags & LA_XLOCKED) != 0 &&
1556 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1557 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1558 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1559 		if ((flags & LA_SLOCKED) != 0 &&
1560 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1561 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1562 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1563 		if ((flags & LA_RECURSED) != 0 &&
1564 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1565 			panic("Lock (%s) %s not recursed @ %s:%d.",
1566 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1567 		if ((flags & LA_NOTRECURSED) != 0 &&
1568 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1569 			panic("Lock (%s) %s recursed @ %s:%d.",
1570 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1571 		break;
1572 	default:
1573 		panic("Invalid lock assertion at %s:%d.", file, line);
1574 
1575 	}
1576 #endif	/* INVARIANT_SUPPORT */
1577 }
1578 
1579 #ifdef DDB
1580 
1581 DB_SHOW_COMMAND(locks, db_witness_list)
1582 {
1583 	struct thread *td;
1584 	pid_t pid;
1585 	struct proc *p;
1586 
1587 	if (have_addr) {
1588 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1589 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1590 		    ((addr >> 16) % 16) * 10000;
1591 		/* sx_slock(&allproc_lock); */
1592 		FOREACH_PROC_IN_SYSTEM(p) {
1593 			if (p->p_pid == pid)
1594 				break;
1595 		}
1596 		/* sx_sunlock(&allproc_lock); */
1597 		if (p == NULL) {
1598 			db_printf("pid %d not found\n", pid);
1599 			return;
1600 		}
1601 		FOREACH_THREAD_IN_PROC(p, td) {
1602 			witness_list(td);
1603 		}
1604 	} else {
1605 		td = curthread;
1606 		witness_list(td);
1607 	}
1608 }
1609 
1610 DB_SHOW_COMMAND(witness, db_witness_display)
1611 {
1612 
1613 	witness_display(db_printf);
1614 }
1615 #endif
1616