xref: /freebsd/sys/kern/subr_witness.c (revision 2b743a9e9ddc6736208dc8ca1ce06ce64ad20a19)
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  */
31 
32 /*
33  * Implementation of the `witness' lock verifier.  Originally implemented for
34  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
35  * classes in FreeBSD.
36  */
37 
38 /*
39  *	Main Entry: witness
40  *	Pronunciation: 'wit-n&s
41  *	Function: noun
42  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
43  *	    testimony, witness, from 2wit
44  *	Date: before 12th century
45  *	1 : attestation of a fact or event : TESTIMONY
46  *	2 : one that gives evidence; specifically : one who testifies in
47  *	    a cause or before a judicial tribunal
48  *	3 : one asked to be present at a transaction so as to be able to
49  *	    testify to its having taken place
50  *	4 : one who has personal knowledge of something
51  *	5 a : something serving as evidence or proof : SIGN
52  *	  b : public affirmation by word or example of usually
53  *	      religious faith or conviction <the heroic witness to divine
54  *	      life -- Pilot>
55  *	6 capitalized : a member of the Jehovah's Witnesses
56  */
57 
58 /*
59  * Special rules concerning Giant and lock orders:
60  *
61  * 1) Giant must be acquired before any other mutexes.  Stated another way,
62  *    no other mutex may be held when Giant is acquired.
63  *
64  * 2) Giant must be released when blocking on a sleepable lock.
65  *
66  * This rule is less obvious, but is a result of Giant providing the same
67  * semantics as spl().  Basically, when a thread sleeps, it must release
68  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
69  * 2).
70  *
71  * 3) Giant may be acquired before or after sleepable locks.
72  *
73  * This rule is also not quite as obvious.  Giant may be acquired after
74  * a sleepable lock because it is a non-sleepable lock and non-sleepable
75  * locks may always be acquired while holding a sleepable lock.  The second
76  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
77  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
78  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
79  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
80  * execute.  Thus, acquiring Giant both before and after a sleepable lock
81  * will not result in a lock order reversal.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_ddb.h"
88 #include "opt_witness.h"
89 
90 #include <sys/param.h>
91 #include <sys/bus.h>
92 #include <sys/kdb.h>
93 #include <sys/kernel.h>
94 #include <sys/ktr.h>
95 #include <sys/lock.h>
96 #include <sys/malloc.h>
97 #include <sys/mutex.h>
98 #include <sys/priv.h>
99 #include <sys/proc.h>
100 #include <sys/sysctl.h>
101 #include <sys/systm.h>
102 
103 #include <ddb/ddb.h>
104 
105 #include <machine/stdarg.h>
106 
107 /* Note that these traces do not work with KTR_ALQ. */
108 #if 0
109 #define	KTR_WITNESS	KTR_SUBSYS
110 #else
111 #define	KTR_WITNESS	0
112 #endif
113 
114 /* Easier to stay with the old names. */
115 #define	lo_list		lo_witness_data.lod_list
116 #define	lo_witness	lo_witness_data.lod_witness
117 
118 /* Define this to check for blessed mutexes */
119 #undef BLESSING
120 
121 #define WITNESS_COUNT 1024
122 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
123 /*
124  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
125  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
126  * probably be safe for the most part, but it's still a SWAG.
127  */
128 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
129 
130 #define	WITNESS_NCHILDREN 6
131 
132 struct witness_child_list_entry;
133 
134 struct witness {
135 	const	char *w_name;
136 	struct	lock_class *w_class;
137 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
138 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
139 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
140 	const	char *w_file;
141 	int	w_line;
142 	u_int	w_level;
143 	u_int	w_refcount;
144 	u_char	w_Giant_squawked:1;
145 	u_char	w_other_squawked:1;
146 	u_char	w_same_squawked:1;
147 	u_char	w_displayed:1;
148 };
149 
150 struct witness_child_list_entry {
151 	struct	witness_child_list_entry *wcl_next;
152 	struct	witness *wcl_children[WITNESS_NCHILDREN];
153 	u_int	wcl_count;
154 };
155 
156 STAILQ_HEAD(witness_list, witness);
157 
158 #ifdef BLESSING
159 struct witness_blessed {
160 	const	char *b_lock1;
161 	const	char *b_lock2;
162 };
163 #endif
164 
165 struct witness_order_list_entry {
166 	const	char *w_name;
167 	struct	lock_class *w_class;
168 };
169 
170 #ifdef BLESSING
171 static int	blessed(struct witness *, struct witness *);
172 #endif
173 static int	depart(struct witness *w);
174 static struct	witness *enroll(const char *description,
175 				struct lock_class *lock_class);
176 static int	insertchild(struct witness *parent, struct witness *child);
177 static int	isitmychild(struct witness *parent, struct witness *child);
178 static int	isitmydescendant(struct witness *parent, struct witness *child);
179 static int	itismychild(struct witness *parent, struct witness *child);
180 static void	removechild(struct witness *parent, struct witness *child);
181 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
182 static const char *fixup_filename(const char *file);
183 static struct	witness *witness_get(void);
184 static void	witness_free(struct witness *m);
185 static struct	witness_child_list_entry *witness_child_get(void);
186 static void	witness_child_free(struct witness_child_list_entry *wcl);
187 static struct	lock_list_entry *witness_lock_list_get(void);
188 static void	witness_lock_list_free(struct lock_list_entry *lle);
189 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
190 					     struct lock_object *lock);
191 static void	witness_list_lock(struct lock_instance *instance);
192 #ifdef DDB
193 static void	witness_leveldescendents(struct witness *parent, int level);
194 static void	witness_levelall(void);
195 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
196 					   struct witness *, int indent);
197 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
198 				     struct witness_list *list);
199 static void	witness_display(void(*)(const char *fmt, ...));
200 static void	witness_list(struct thread *td);
201 #endif
202 
203 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
204 
205 /*
206  * If set to 0, witness is disabled.  If set to a non-zero value, witness
207  * performs full lock order checking for all locks.  At runtime, this
208  * value may be set to 0 to turn off witness.  witness is not allowed be
209  * turned on once it is turned off, however.
210  */
211 static int witness_watch = 1;
212 TUNABLE_INT("debug.witness.watch", &witness_watch);
213 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
214     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
215 
216 #ifdef KDB
217 /*
218  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
219  * to drop into kdebug() when:
220  *	- a lock hierarchy violation occurs
221  *	- locks are held when going to sleep.
222  */
223 #ifdef WITNESS_KDB
224 int	witness_kdb = 1;
225 #else
226 int	witness_kdb = 0;
227 #endif
228 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
229 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
230 
231 /*
232  * When KDB is enabled and witness_trace is set to 1, it will cause the system
233  * to print a stack trace:
234  *	- a lock hierarchy violation occurs
235  *	- locks are held when going to sleep.
236  */
237 int	witness_trace = 1;
238 TUNABLE_INT("debug.witness.trace", &witness_trace);
239 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
240 #endif /* KDB */
241 
242 #ifdef WITNESS_SKIPSPIN
243 int	witness_skipspin = 1;
244 #else
245 int	witness_skipspin = 0;
246 #endif
247 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
248 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
249     &witness_skipspin, 0, "");
250 
251 static struct mtx w_mtx;
252 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
253 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
254 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
255 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
256 static struct witness_child_list_entry *w_child_free = NULL;
257 static struct lock_list_entry *w_lock_list_free = NULL;
258 
259 static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt;
260 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
261 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
262 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
263     "");
264 SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD,
265     &w_child_free_cnt, 0, "");
266 SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0,
267     "");
268 
269 static struct witness w_data[WITNESS_COUNT];
270 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
271 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
272 
273 static struct witness_order_list_entry order_lists[] = {
274 	/*
275 	 * sx locks
276 	 */
277 	{ "proctree", &lock_class_sx },
278 	{ "allproc", &lock_class_sx },
279 	{ NULL, NULL },
280 	/*
281 	 * Various mutexes
282 	 */
283 	{ "Giant", &lock_class_mtx_sleep },
284 	{ "filedesc structure", &lock_class_mtx_sleep },
285 	{ "pipe mutex", &lock_class_mtx_sleep },
286 	{ "sigio lock", &lock_class_mtx_sleep },
287 	{ "process group", &lock_class_mtx_sleep },
288 	{ "process lock", &lock_class_mtx_sleep },
289 	{ "session", &lock_class_mtx_sleep },
290 	{ "uidinfo hash", &lock_class_mtx_sleep },
291 	{ "uidinfo struct", &lock_class_mtx_sleep },
292 	{ "allprison", &lock_class_mtx_sleep },
293 	{ NULL, NULL },
294 	/*
295 	 * Sockets
296 	 */
297 	{ "filedesc structure", &lock_class_mtx_sleep },
298 	{ "accept", &lock_class_mtx_sleep },
299 	{ "so_snd", &lock_class_mtx_sleep },
300 	{ "so_rcv", &lock_class_mtx_sleep },
301 	{ "sellck", &lock_class_mtx_sleep },
302 	{ NULL, NULL },
303 	/*
304 	 * Routing
305 	 */
306 	{ "so_rcv", &lock_class_mtx_sleep },
307 	{ "radix node head", &lock_class_mtx_sleep },
308 	{ "rtentry", &lock_class_mtx_sleep },
309 	{ "ifaddr", &lock_class_mtx_sleep },
310 	{ NULL, NULL },
311 	/*
312 	 * Multicast - protocol locks before interface locks, after UDP locks.
313 	 */
314 	{ "udpinp", &lock_class_mtx_sleep },
315 	{ "in_multi_mtx", &lock_class_mtx_sleep },
316 	{ "igmp_mtx", &lock_class_mtx_sleep },
317 	{ "if_addr_mtx", &lock_class_mtx_sleep },
318 	{ NULL, NULL },
319 	/*
320 	 * UNIX Domain Sockets
321 	 */
322 	{ "unp", &lock_class_mtx_sleep },
323 	{ "so_snd", &lock_class_mtx_sleep },
324 	{ NULL, NULL },
325 	/*
326 	 * UDP/IP
327 	 */
328 	{ "udp", &lock_class_mtx_sleep },
329 	{ "udpinp", &lock_class_mtx_sleep },
330 	{ "so_snd", &lock_class_mtx_sleep },
331 	{ NULL, NULL },
332 	/*
333 	 * TCP/IP
334 	 */
335 	{ "tcp", &lock_class_mtx_sleep },
336 	{ "tcpinp", &lock_class_mtx_sleep },
337 	{ "so_snd", &lock_class_mtx_sleep },
338 	{ NULL, NULL },
339 	/*
340 	 * SLIP
341 	 */
342 	{ "slip_mtx", &lock_class_mtx_sleep },
343 	{ "slip sc_mtx", &lock_class_mtx_sleep },
344 	{ NULL, NULL },
345 	/*
346 	 * netatalk
347 	 */
348 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
349 	{ "ddp_mtx", &lock_class_mtx_sleep },
350 	{ NULL, NULL },
351 	/*
352 	 * BPF
353 	 */
354 	{ "bpf global lock", &lock_class_mtx_sleep },
355 	{ "bpf interface lock", &lock_class_mtx_sleep },
356 	{ "bpf cdev lock", &lock_class_mtx_sleep },
357 	{ NULL, NULL },
358 	/*
359 	 * NFS server
360 	 */
361 	{ "nfsd_mtx", &lock_class_mtx_sleep },
362 	{ "so_snd", &lock_class_mtx_sleep },
363 	{ NULL, NULL },
364 	/*
365 	 * CDEV
366 	 */
367 	{ "system map", &lock_class_mtx_sleep },
368 	{ "vm page queue mutex", &lock_class_mtx_sleep },
369 	{ "vnode interlock", &lock_class_mtx_sleep },
370 	{ "cdev", &lock_class_mtx_sleep },
371 	{ NULL, NULL },
372 	/*
373 	 * kqueue/VFS interaction
374 	 */
375 	{ "kqueue", &lock_class_mtx_sleep },
376 	{ "struct mount mtx", &lock_class_mtx_sleep },
377 	{ "vnode interlock", &lock_class_mtx_sleep },
378 	{ NULL, NULL },
379 	/*
380 	 * spin locks
381 	 */
382 #ifdef SMP
383 	{ "ap boot", &lock_class_mtx_spin },
384 #endif
385 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
386 	{ "sio", &lock_class_mtx_spin },
387 #ifdef __i386__
388 	{ "cy", &lock_class_mtx_spin },
389 #endif
390 	{ "scc_hwmtx", &lock_class_mtx_spin },
391 	{ "uart_hwmtx", &lock_class_mtx_spin },
392 	{ "zstty", &lock_class_mtx_spin },
393 	{ "ng_node", &lock_class_mtx_spin },
394 	{ "ng_worklist", &lock_class_mtx_spin },
395 	{ "fast_taskqueue", &lock_class_mtx_spin },
396 	{ "intr table", &lock_class_mtx_spin },
397 	{ "sleepq chain", &lock_class_mtx_spin },
398 	{ "sched lock", &lock_class_mtx_spin },
399 	{ "turnstile chain", &lock_class_mtx_spin },
400 	{ "td_contested", &lock_class_mtx_spin },
401 	{ "callout", &lock_class_mtx_spin },
402 	{ "entropy harvest mutex", &lock_class_mtx_spin },
403 	{ "syscons video lock", &lock_class_mtx_spin },
404 	/*
405 	 * leaf locks
406 	 */
407 	{ "allpmaps", &lock_class_mtx_spin },
408 	{ "icu", &lock_class_mtx_spin },
409 #ifdef SMP
410 	{ "smp rendezvous", &lock_class_mtx_spin },
411 #if defined(__i386__) || defined(__amd64__)
412 	{ "tlb", &lock_class_mtx_spin },
413 #endif
414 #ifdef __sparc64__
415 	{ "ipi", &lock_class_mtx_spin },
416 	{ "rtc_mtx", &lock_class_mtx_spin },
417 #endif
418 #endif
419 	{ "clk", &lock_class_mtx_spin },
420 	{ "mutex profiling lock", &lock_class_mtx_spin },
421 	{ "kse zombie lock", &lock_class_mtx_spin },
422 	{ "ALD Queue", &lock_class_mtx_spin },
423 #ifdef __ia64__
424 	{ "MCA spin lock", &lock_class_mtx_spin },
425 #endif
426 #if defined(__i386__) || defined(__amd64__)
427 	{ "pcicfg", &lock_class_mtx_spin },
428 	{ "NDIS thread lock", &lock_class_mtx_spin },
429 #endif
430 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
431 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
432 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
433 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
434 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
435 	{ NULL, NULL },
436 	{ NULL, NULL }
437 };
438 
439 #ifdef BLESSING
440 /*
441  * Pairs of locks which have been blessed
442  * Don't complain about order problems with blessed locks
443  */
444 static struct witness_blessed blessed_list[] = {
445 };
446 static int blessed_count =
447 	sizeof(blessed_list) / sizeof(struct witness_blessed);
448 #endif
449 
450 /*
451  * List of locks initialized prior to witness being initialized whose
452  * enrollment is currently deferred.
453  */
454 STAILQ_HEAD(, lock_object) pending_locks =
455     STAILQ_HEAD_INITIALIZER(pending_locks);
456 
457 /*
458  * This global is set to 0 once it becomes safe to use the witness code.
459  */
460 static int witness_cold = 1;
461 
462 /*
463  * This global is set to 1 once the static lock orders have been enrolled
464  * so that a warning can be issued for any spin locks enrolled later.
465  */
466 static int witness_spin_warn = 0;
467 
468 /*
469  * The WITNESS-enabled diagnostic code.  Note that the witness code does
470  * assume that the early boot is single-threaded at least until after this
471  * routine is completed.
472  */
473 static void
474 witness_initialize(void *dummy __unused)
475 {
476 	struct lock_object *lock;
477 	struct witness_order_list_entry *order;
478 	struct witness *w, *w1;
479 	int i;
480 
481 	/*
482 	 * We have to release Giant before initializing its witness
483 	 * structure so that WITNESS doesn't get confused.
484 	 */
485 	mtx_unlock(&Giant);
486 	mtx_assert(&Giant, MA_NOTOWNED);
487 
488 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
489 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
490 	    MTX_NOWITNESS | MTX_NOPROFILE);
491 	for (i = 0; i < WITNESS_COUNT; i++)
492 		witness_free(&w_data[i]);
493 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
494 		witness_child_free(&w_childdata[i]);
495 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
496 		witness_lock_list_free(&w_locklistdata[i]);
497 
498 	/* First add in all the specified order lists. */
499 	for (order = order_lists; order->w_name != NULL; order++) {
500 		w = enroll(order->w_name, order->w_class);
501 		if (w == NULL)
502 			continue;
503 		w->w_file = "order list";
504 		for (order++; order->w_name != NULL; order++) {
505 			w1 = enroll(order->w_name, order->w_class);
506 			if (w1 == NULL)
507 				continue;
508 			w1->w_file = "order list";
509 			if (!itismychild(w, w1))
510 				panic("Not enough memory for static orders!");
511 			w = w1;
512 		}
513 	}
514 	witness_spin_warn = 1;
515 
516 	/* Iterate through all locks and add them to witness. */
517 	while (!STAILQ_EMPTY(&pending_locks)) {
518 		lock = STAILQ_FIRST(&pending_locks);
519 		STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
520 		KASSERT(lock->lo_flags & LO_WITNESS,
521 		    ("%s: lock %s is on pending list but not LO_WITNESS",
522 		    __func__, lock->lo_name));
523 		lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
524 	}
525 
526 	/* Mark the witness code as being ready for use. */
527 	witness_cold = 0;
528 
529 	mtx_lock(&Giant);
530 }
531 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
532 
533 static int
534 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
535 {
536 	int error, value;
537 
538 	value = witness_watch;
539 	error = sysctl_handle_int(oidp, &value, 0, req);
540 	if (error != 0 || req->newptr == NULL)
541 		return (error);
542 	if (value == witness_watch)
543 		return (0);
544 	if (value != 0)
545 		return (EINVAL);
546 	witness_watch = 0;
547 	return (0);
548 }
549 
550 void
551 witness_init(struct lock_object *lock)
552 {
553 	struct lock_class *class;
554 
555 	/* Various sanity checks. */
556 	class = LOCK_CLASS(lock);
557 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
558 	    (class->lc_flags & LC_RECURSABLE) == 0)
559 		panic("%s: lock (%s) %s can not be recursable", __func__,
560 		    class->lc_name, lock->lo_name);
561 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
562 	    (class->lc_flags & LC_SLEEPABLE) == 0)
563 		panic("%s: lock (%s) %s can not be sleepable", __func__,
564 		    class->lc_name, lock->lo_name);
565 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
566 	    (class->lc_flags & LC_UPGRADABLE) == 0)
567 		panic("%s: lock (%s) %s can not be upgradable", __func__,
568 		    class->lc_name, lock->lo_name);
569 
570 	/*
571 	 * If we shouldn't watch this lock, then just clear lo_witness.
572 	 * Otherwise, if witness_cold is set, then it is too early to
573 	 * enroll this lock, so defer it to witness_initialize() by adding
574 	 * it to the pending_locks list.  If it is not too early, then enroll
575 	 * the lock now.
576 	 */
577 	if (witness_watch == 0 || panicstr != NULL ||
578 	    (lock->lo_flags & LO_WITNESS) == 0)
579 		lock->lo_witness = NULL;
580 	else if (witness_cold) {
581 		STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
582 		lock->lo_flags |= LO_ENROLLPEND;
583 	} else
584 		lock->lo_witness = enroll(lock->lo_type, class);
585 }
586 
587 void
588 witness_destroy(struct lock_object *lock)
589 {
590 	struct lock_class *class;
591 	struct witness *w;
592 
593 	class = LOCK_CLASS(lock);
594 	if (witness_cold)
595 		panic("lock (%s) %s destroyed while witness_cold",
596 		    class->lc_name, lock->lo_name);
597 
598 	/* XXX: need to verify that no one holds the lock */
599 	if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
600 	    lock->lo_witness != NULL) {
601 		w = lock->lo_witness;
602 		mtx_lock_spin(&w_mtx);
603 		MPASS(w->w_refcount > 0);
604 		w->w_refcount--;
605 
606 		/*
607 		 * Lock is already released if we have an allocation failure
608 		 * and depart() fails.
609 		 */
610 		if (w->w_refcount != 0 || depart(w))
611 			mtx_unlock_spin(&w_mtx);
612 	}
613 
614 	/*
615 	 * If this lock is destroyed before witness is up and running,
616 	 * remove it from the pending list.
617 	 */
618 	if (lock->lo_flags & LO_ENROLLPEND) {
619 		STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
620 		lock->lo_flags &= ~LO_ENROLLPEND;
621 	}
622 }
623 
624 #ifdef DDB
625 static void
626 witness_levelall (void)
627 {
628 	struct witness_list *list;
629 	struct witness *w, *w1;
630 
631 	/*
632 	 * First clear all levels.
633 	 */
634 	STAILQ_FOREACH(w, &w_all, w_list) {
635 		w->w_level = 0;
636 	}
637 
638 	/*
639 	 * Look for locks with no parent and level all their descendants.
640 	 */
641 	STAILQ_FOREACH(w, &w_all, w_list) {
642 		/*
643 		 * This is just an optimization, technically we could get
644 		 * away just walking the all list each time.
645 		 */
646 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
647 			list = &w_sleep;
648 		else
649 			list = &w_spin;
650 		STAILQ_FOREACH(w1, list, w_typelist) {
651 			if (isitmychild(w1, w))
652 				goto skip;
653 		}
654 		witness_leveldescendents(w, 0);
655 	skip:
656 		;	/* silence GCC 3.x */
657 	}
658 }
659 
660 static void
661 witness_leveldescendents(struct witness *parent, int level)
662 {
663 	struct witness_child_list_entry *wcl;
664 	int i;
665 
666 	if (parent->w_level < level)
667 		parent->w_level = level;
668 	level++;
669 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
670 		for (i = 0; i < wcl->wcl_count; i++)
671 			witness_leveldescendents(wcl->wcl_children[i], level);
672 }
673 
674 static void
675 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
676 			   struct witness *parent, int indent)
677 {
678 	struct witness_child_list_entry *wcl;
679 	int i, level;
680 
681 	level = parent->w_level;
682 	prnt("%-2d", level);
683 	for (i = 0; i < indent; i++)
684 		prnt(" ");
685 	if (parent->w_refcount > 0)
686 		prnt("%s", parent->w_name);
687 	else
688 		prnt("(dead)");
689 	if (parent->w_displayed) {
690 		prnt(" -- (already displayed)\n");
691 		return;
692 	}
693 	parent->w_displayed = 1;
694 	if (parent->w_refcount > 0) {
695 		if (parent->w_file != NULL)
696 			prnt(" -- last acquired @ %s:%d", parent->w_file,
697 			    parent->w_line);
698 	}
699 	prnt("\n");
700 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
701 		for (i = 0; i < wcl->wcl_count; i++)
702 			    witness_displaydescendants(prnt,
703 				wcl->wcl_children[i], indent + 1);
704 }
705 
706 static void
707 witness_display_list(void(*prnt)(const char *fmt, ...),
708 		     struct witness_list *list)
709 {
710 	struct witness *w;
711 
712 	STAILQ_FOREACH(w, list, w_typelist) {
713 		if (w->w_file == NULL || w->w_level > 0)
714 			continue;
715 		/*
716 		 * This lock has no anscestors, display its descendants.
717 		 */
718 		witness_displaydescendants(prnt, w, 0);
719 	}
720 }
721 
722 static void
723 witness_display(void(*prnt)(const char *fmt, ...))
724 {
725 	struct witness *w;
726 
727 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
728 	witness_levelall();
729 
730 	/* Clear all the displayed flags. */
731 	STAILQ_FOREACH(w, &w_all, w_list) {
732 		w->w_displayed = 0;
733 	}
734 
735 	/*
736 	 * First, handle sleep locks which have been acquired at least
737 	 * once.
738 	 */
739 	prnt("Sleep locks:\n");
740 	witness_display_list(prnt, &w_sleep);
741 
742 	/*
743 	 * Now do spin locks which have been acquired at least once.
744 	 */
745 	prnt("\nSpin locks:\n");
746 	witness_display_list(prnt, &w_spin);
747 
748 	/*
749 	 * Finally, any locks which have not been acquired yet.
750 	 */
751 	prnt("\nLocks which were never acquired:\n");
752 	STAILQ_FOREACH(w, &w_all, w_list) {
753 		if (w->w_file != NULL || w->w_refcount == 0)
754 			continue;
755 		prnt("%s\n", w->w_name);
756 	}
757 }
758 #endif /* DDB */
759 
760 /* Trim useless garbage from filenames. */
761 static const char *
762 fixup_filename(const char *file)
763 {
764 
765 	if (file == NULL)
766 		return (NULL);
767 	while (strncmp(file, "../", 3) == 0)
768 		file += 3;
769 	return (file);
770 }
771 
772 int
773 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
774 {
775 
776 	if (witness_watch == 0 || panicstr != NULL)
777 		return (0);
778 
779 	/* Require locks that witness knows about. */
780 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
781 	    lock2->lo_witness == NULL)
782 		return (EINVAL);
783 
784 	MPASS(!mtx_owned(&w_mtx));
785 	mtx_lock_spin(&w_mtx);
786 
787 	/*
788 	 * If we already have either an explicit or implied lock order that
789 	 * is the other way around, then return an error.
790 	 */
791 	if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
792 		mtx_unlock_spin(&w_mtx);
793 		return (EDOOFUS);
794 	}
795 
796 	/* Try to add the new order. */
797 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
798 	    lock2->lo_type, lock1->lo_type);
799 	if (!itismychild(lock1->lo_witness, lock2->lo_witness))
800 		return (ENOMEM);
801 	mtx_unlock_spin(&w_mtx);
802 	return (0);
803 }
804 
805 void
806 witness_checkorder(struct lock_object *lock, int flags, const char *file,
807     int line)
808 {
809 	struct lock_list_entry **lock_list, *lle;
810 	struct lock_instance *lock1, *lock2;
811 	struct lock_class *class;
812 	struct witness *w, *w1;
813 	struct thread *td;
814 	int i, j;
815 
816 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
817 	    panicstr != NULL)
818 		return;
819 
820 	/*
821 	 * Try locks do not block if they fail to acquire the lock, thus
822 	 * there is no danger of deadlocks or of switching while holding a
823 	 * spin lock if we acquire a lock via a try operation.  This
824 	 * function shouldn't even be called for try locks, so panic if
825 	 * that happens.
826 	 */
827 	if (flags & LOP_TRYLOCK)
828 		panic("%s should not be called for try lock operations",
829 		    __func__);
830 
831 	w = lock->lo_witness;
832 	class = LOCK_CLASS(lock);
833 	td = curthread;
834 	file = fixup_filename(file);
835 
836 	if (class->lc_flags & LC_SLEEPLOCK) {
837 		/*
838 		 * Since spin locks include a critical section, this check
839 		 * implicitly enforces a lock order of all sleep locks before
840 		 * all spin locks.
841 		 */
842 		if (td->td_critnest != 0 && !kdb_active)
843 			panic("blockable sleep lock (%s) %s @ %s:%d",
844 			    class->lc_name, lock->lo_name, file, line);
845 
846 		/*
847 		 * If this is the first lock acquired then just return as
848 		 * no order checking is needed.
849 		 */
850 		if (td->td_sleeplocks == NULL)
851 			return;
852 		lock_list = &td->td_sleeplocks;
853 	} else {
854 		/*
855 		 * If this is the first lock, just return as no order
856 		 * checking is needed.  We check this in both if clauses
857 		 * here as unifying the check would require us to use a
858 		 * critical section to ensure we don't migrate while doing
859 		 * the check.  Note that if this is not the first lock, we
860 		 * are already in a critical section and are safe for the
861 		 * rest of the check.
862 		 */
863 		if (PCPU_GET(spinlocks) == NULL)
864 			return;
865 		lock_list = PCPU_PTR(spinlocks);
866 	}
867 
868 	/*
869 	 * Check to see if we are recursing on a lock we already own.  If
870 	 * so, make sure that we don't mismatch exclusive and shared lock
871 	 * acquires.
872 	 */
873 	lock1 = find_instance(*lock_list, lock);
874 	if (lock1 != NULL) {
875 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
876 		    (flags & LOP_EXCLUSIVE) == 0) {
877 			printf("shared lock of (%s) %s @ %s:%d\n",
878 			    class->lc_name, lock->lo_name, file, line);
879 			printf("while exclusively locked from %s:%d\n",
880 			    lock1->li_file, lock1->li_line);
881 			panic("share->excl");
882 		}
883 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
884 		    (flags & LOP_EXCLUSIVE) != 0) {
885 			printf("exclusive lock of (%s) %s @ %s:%d\n",
886 			    class->lc_name, lock->lo_name, file, line);
887 			printf("while share locked from %s:%d\n",
888 			    lock1->li_file, lock1->li_line);
889 			panic("excl->share");
890 		}
891 		return;
892 	}
893 
894 	/*
895 	 * Try locks do not block if they fail to acquire the lock, thus
896 	 * there is no danger of deadlocks or of switching while holding a
897 	 * spin lock if we acquire a lock via a try operation.
898 	 */
899 	if (flags & LOP_TRYLOCK)
900 		return;
901 
902 	/*
903 	 * Check for duplicate locks of the same type.  Note that we only
904 	 * have to check for this on the last lock we just acquired.  Any
905 	 * other cases will be caught as lock order violations.
906 	 */
907 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
908 	w1 = lock1->li_lock->lo_witness;
909 	if (w1 == w) {
910 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
911 		    (flags & LOP_DUPOK))
912 			return;
913 		w->w_same_squawked = 1;
914 		printf("acquiring duplicate lock of same type: \"%s\"\n",
915 			lock->lo_type);
916 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
917 		    lock1->li_file, lock1->li_line);
918 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
919 #ifdef KDB
920 		goto debugger;
921 #else
922 		return;
923 #endif
924 	}
925 	MPASS(!mtx_owned(&w_mtx));
926 	mtx_lock_spin(&w_mtx);
927 	/*
928 	 * If we know that the the lock we are acquiring comes after
929 	 * the lock we most recently acquired in the lock order tree,
930 	 * then there is no need for any further checks.
931 	 */
932 	if (isitmychild(w1, w)) {
933 		mtx_unlock_spin(&w_mtx);
934 		return;
935 	}
936 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
937 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
938 
939 			MPASS(j < WITNESS_COUNT);
940 			lock1 = &lle->ll_children[i];
941 			w1 = lock1->li_lock->lo_witness;
942 
943 			/*
944 			 * If this lock doesn't undergo witness checking,
945 			 * then skip it.
946 			 */
947 			if (w1 == NULL) {
948 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
949 				    ("lock missing witness structure"));
950 				continue;
951 			}
952 			/*
953 			 * If we are locking Giant and this is a sleepable
954 			 * lock, then skip it.
955 			 */
956 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
957 			    lock == &Giant.mtx_object)
958 				continue;
959 			/*
960 			 * If we are locking a sleepable lock and this lock
961 			 * is Giant, then skip it.
962 			 */
963 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
964 			    lock1->li_lock == &Giant.mtx_object)
965 				continue;
966 			/*
967 			 * If we are locking a sleepable lock and this lock
968 			 * isn't sleepable, we want to treat it as a lock
969 			 * order violation to enfore a general lock order of
970 			 * sleepable locks before non-sleepable locks.
971 			 */
972 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
973 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
974 				goto reversal;
975 			/*
976 			 * If we are locking Giant and this is a non-sleepable
977 			 * lock, then treat it as a reversal.
978 			 */
979 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
980 			    lock == &Giant.mtx_object)
981 				goto reversal;
982 			/*
983 			 * Check the lock order hierarchy for a reveresal.
984 			 */
985 			if (!isitmydescendant(w, w1))
986 				continue;
987 		reversal:
988 			/*
989 			 * We have a lock order violation, check to see if it
990 			 * is allowed or has already been yelled about.
991 			 */
992 			mtx_unlock_spin(&w_mtx);
993 #ifdef BLESSING
994 			/*
995 			 * If the lock order is blessed, just bail.  We don't
996 			 * look for other lock order violations though, which
997 			 * may be a bug.
998 			 */
999 			if (blessed(w, w1))
1000 				return;
1001 #endif
1002 			if (lock1->li_lock == &Giant.mtx_object) {
1003 				if (w1->w_Giant_squawked)
1004 					return;
1005 				else
1006 					w1->w_Giant_squawked = 1;
1007 			} else {
1008 				if (w1->w_other_squawked)
1009 					return;
1010 				else
1011 					w1->w_other_squawked = 1;
1012 			}
1013 			/*
1014 			 * Ok, yell about it.
1015 			 */
1016 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1017 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1018 				printf(
1019 		"lock order reversal: (sleepable after non-sleepable)\n");
1020 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1021 			    && lock == &Giant.mtx_object)
1022 				printf(
1023 		"lock order reversal: (Giant after non-sleepable)\n");
1024 			else
1025 				printf("lock order reversal:\n");
1026 			/*
1027 			 * Try to locate an earlier lock with
1028 			 * witness w in our list.
1029 			 */
1030 			do {
1031 				lock2 = &lle->ll_children[i];
1032 				MPASS(lock2->li_lock != NULL);
1033 				if (lock2->li_lock->lo_witness == w)
1034 					break;
1035 				if (i == 0 && lle->ll_next != NULL) {
1036 					lle = lle->ll_next;
1037 					i = lle->ll_count - 1;
1038 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1039 				} else
1040 					i--;
1041 			} while (i >= 0);
1042 			if (i < 0) {
1043 				printf(" 1st %p %s (%s) @ %s:%d\n",
1044 				    lock1->li_lock, lock1->li_lock->lo_name,
1045 				    lock1->li_lock->lo_type, lock1->li_file,
1046 				    lock1->li_line);
1047 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1048 				    lock->lo_name, lock->lo_type, file, line);
1049 			} else {
1050 				printf(" 1st %p %s (%s) @ %s:%d\n",
1051 				    lock2->li_lock, lock2->li_lock->lo_name,
1052 				    lock2->li_lock->lo_type, lock2->li_file,
1053 				    lock2->li_line);
1054 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1055 				    lock1->li_lock, lock1->li_lock->lo_name,
1056 				    lock1->li_lock->lo_type, lock1->li_file,
1057 				    lock1->li_line);
1058 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1059 				    lock->lo_name, lock->lo_type, file, line);
1060 			}
1061 #ifdef KDB
1062 			goto debugger;
1063 #else
1064 			return;
1065 #endif
1066 		}
1067 	}
1068 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1069 	/*
1070 	 * If requested, build a new lock order.  However, don't build a new
1071 	 * relationship between a sleepable lock and Giant if it is in the
1072 	 * wrong direction.  The correct lock order is that sleepable locks
1073 	 * always come before Giant.
1074 	 */
1075 	if (flags & LOP_NEWORDER &&
1076 	    !(lock1->li_lock == &Giant.mtx_object &&
1077 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1078 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1079 		    lock->lo_type, lock1->li_lock->lo_type);
1080 		if (!itismychild(lock1->li_lock->lo_witness, w))
1081 			/* Witness is dead. */
1082 			return;
1083 	}
1084 	mtx_unlock_spin(&w_mtx);
1085 	return;
1086 
1087 #ifdef KDB
1088 debugger:
1089 	if (witness_trace)
1090 		kdb_backtrace();
1091 	if (witness_kdb)
1092 		kdb_enter(__func__);
1093 #endif
1094 }
1095 
1096 void
1097 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1098 {
1099 	struct lock_list_entry **lock_list, *lle;
1100 	struct lock_instance *instance;
1101 	struct witness *w;
1102 	struct thread *td;
1103 
1104 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1105 	    panicstr != NULL)
1106 		return;
1107 	w = lock->lo_witness;
1108 	td = curthread;
1109 	file = fixup_filename(file);
1110 
1111 	/* Determine lock list for this lock. */
1112 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1113 		lock_list = &td->td_sleeplocks;
1114 	else
1115 		lock_list = PCPU_PTR(spinlocks);
1116 
1117 	/* Check to see if we are recursing on a lock we already own. */
1118 	instance = find_instance(*lock_list, lock);
1119 	if (instance != NULL) {
1120 		instance->li_flags++;
1121 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1122 		    td->td_proc->p_pid, lock->lo_name,
1123 		    instance->li_flags & LI_RECURSEMASK);
1124 		instance->li_file = file;
1125 		instance->li_line = line;
1126 		return;
1127 	}
1128 
1129 	/* Update per-witness last file and line acquire. */
1130 	w->w_file = file;
1131 	w->w_line = line;
1132 
1133 	/* Find the next open lock instance in the list and fill it. */
1134 	lle = *lock_list;
1135 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1136 		lle = witness_lock_list_get();
1137 		if (lle == NULL)
1138 			return;
1139 		lle->ll_next = *lock_list;
1140 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1141 		    td->td_proc->p_pid, lle);
1142 		*lock_list = lle;
1143 	}
1144 	instance = &lle->ll_children[lle->ll_count++];
1145 	instance->li_lock = lock;
1146 	instance->li_line = line;
1147 	instance->li_file = file;
1148 	if ((flags & LOP_EXCLUSIVE) != 0)
1149 		instance->li_flags = LI_EXCLUSIVE;
1150 	else
1151 		instance->li_flags = 0;
1152 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1153 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1154 }
1155 
1156 void
1157 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1158 {
1159 	struct lock_instance *instance;
1160 	struct lock_class *class;
1161 
1162 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1163 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1164 		return;
1165 	class = LOCK_CLASS(lock);
1166 	file = fixup_filename(file);
1167 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1168 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1169 		    class->lc_name, lock->lo_name, file, line);
1170 	if ((flags & LOP_TRYLOCK) == 0)
1171 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1172 		    lock->lo_name, file, line);
1173 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1174 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1175 		    class->lc_name, lock->lo_name, file, line);
1176 	instance = find_instance(curthread->td_sleeplocks, lock);
1177 	if (instance == NULL)
1178 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1179 		    class->lc_name, lock->lo_name, file, line);
1180 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1181 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1182 		    class->lc_name, lock->lo_name, file, line);
1183 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1184 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1185 		    class->lc_name, lock->lo_name,
1186 		    instance->li_flags & LI_RECURSEMASK, file, line);
1187 	instance->li_flags |= LI_EXCLUSIVE;
1188 }
1189 
1190 void
1191 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1192     int line)
1193 {
1194 	struct lock_instance *instance;
1195 	struct lock_class *class;
1196 
1197 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1198 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1199 		return;
1200 	class = LOCK_CLASS(lock);
1201 	file = fixup_filename(file);
1202 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1203 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1204 		    class->lc_name, lock->lo_name, file, line);
1205 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1206 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1207 		    class->lc_name, lock->lo_name, file, line);
1208 	instance = find_instance(curthread->td_sleeplocks, lock);
1209 	if (instance == NULL)
1210 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1211 		    class->lc_name, lock->lo_name, file, line);
1212 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1213 		panic("downgrade of shared lock (%s) %s @ %s:%d",
1214 		    class->lc_name, lock->lo_name, file, line);
1215 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1216 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1217 		    class->lc_name, lock->lo_name,
1218 		    instance->li_flags & LI_RECURSEMASK, file, line);
1219 	instance->li_flags &= ~LI_EXCLUSIVE;
1220 }
1221 
1222 void
1223 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1224 {
1225 	struct lock_list_entry **lock_list, *lle;
1226 	struct lock_instance *instance;
1227 	struct lock_class *class;
1228 	struct thread *td;
1229 	register_t s;
1230 	int i, j;
1231 
1232 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1233 	    panicstr != NULL)
1234 		return;
1235 	td = curthread;
1236 	class = LOCK_CLASS(lock);
1237 	file = fixup_filename(file);
1238 
1239 	/* Find lock instance associated with this lock. */
1240 	if (class->lc_flags & LC_SLEEPLOCK)
1241 		lock_list = &td->td_sleeplocks;
1242 	else
1243 		lock_list = PCPU_PTR(spinlocks);
1244 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1245 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1246 			instance = &(*lock_list)->ll_children[i];
1247 			if (instance->li_lock == lock)
1248 				goto found;
1249 		}
1250 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1251 	    file, line);
1252 found:
1253 
1254 	/* First, check for shared/exclusive mismatches. */
1255 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1256 	    (flags & LOP_EXCLUSIVE) == 0) {
1257 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1258 		    lock->lo_name, file, line);
1259 		printf("while exclusively locked from %s:%d\n",
1260 		    instance->li_file, instance->li_line);
1261 		panic("excl->ushare");
1262 	}
1263 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1264 	    (flags & LOP_EXCLUSIVE) != 0) {
1265 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1266 		    lock->lo_name, file, line);
1267 		printf("while share locked from %s:%d\n", instance->li_file,
1268 		    instance->li_line);
1269 		panic("share->uexcl");
1270 	}
1271 
1272 	/* If we are recursed, unrecurse. */
1273 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1274 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1275 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1276 		    instance->li_flags);
1277 		instance->li_flags--;
1278 		return;
1279 	}
1280 
1281 	/* Otherwise, remove this item from the list. */
1282 	s = intr_disable();
1283 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1284 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1285 	    (*lock_list)->ll_count - 1);
1286 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1287 		(*lock_list)->ll_children[j] =
1288 		    (*lock_list)->ll_children[j + 1];
1289 	(*lock_list)->ll_count--;
1290 	intr_restore(s);
1291 
1292 	/* If this lock list entry is now empty, free it. */
1293 	if ((*lock_list)->ll_count == 0) {
1294 		lle = *lock_list;
1295 		*lock_list = lle->ll_next;
1296 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1297 		    td->td_proc->p_pid, lle);
1298 		witness_lock_list_free(lle);
1299 	}
1300 }
1301 
1302 /*
1303  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1304  * exempt Giant and sleepable locks from the checks as well.  If any
1305  * non-exempt locks are held, then a supplied message is printed to the
1306  * console along with a list of the offending locks.  If indicated in the
1307  * flags then a failure results in a panic as well.
1308  */
1309 int
1310 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1311 {
1312 	struct lock_list_entry *lle;
1313 	struct lock_instance *lock1;
1314 	struct thread *td;
1315 	va_list ap;
1316 	int i, n;
1317 
1318 	if (witness_cold || witness_watch == 0 || panicstr != NULL)
1319 		return (0);
1320 	n = 0;
1321 	td = curthread;
1322 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1323 		for (i = lle->ll_count - 1; i >= 0; i--) {
1324 			lock1 = &lle->ll_children[i];
1325 			if (lock1->li_lock == lock)
1326 				continue;
1327 			if (flags & WARN_GIANTOK &&
1328 			    lock1->li_lock == &Giant.mtx_object)
1329 				continue;
1330 			if (flags & WARN_SLEEPOK &&
1331 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1332 				continue;
1333 			if (n == 0) {
1334 				va_start(ap, fmt);
1335 				vprintf(fmt, ap);
1336 				va_end(ap);
1337 				printf(" with the following");
1338 				if (flags & WARN_SLEEPOK)
1339 					printf(" non-sleepable");
1340 				printf(" locks held:\n");
1341 			}
1342 			n++;
1343 			witness_list_lock(lock1);
1344 		}
1345 	if (PCPU_GET(spinlocks) != NULL) {
1346 		/*
1347 		 * Since we already hold a spinlock preemption is
1348 		 * already blocked.
1349 		 */
1350 		if (n == 0) {
1351 			va_start(ap, fmt);
1352 			vprintf(fmt, ap);
1353 			va_end(ap);
1354 			printf(" with the following");
1355 			if (flags & WARN_SLEEPOK)
1356 				printf(" non-sleepable");
1357 			printf(" locks held:\n");
1358 		}
1359 		n += witness_list_locks(PCPU_PTR(spinlocks));
1360 	}
1361 	if (flags & WARN_PANIC && n)
1362 		panic("witness_warn");
1363 #ifdef KDB
1364 	else if (witness_kdb && n)
1365 		kdb_enter(__func__);
1366 	else if (witness_trace && n)
1367 		kdb_backtrace();
1368 #endif
1369 	return (n);
1370 }
1371 
1372 const char *
1373 witness_file(struct lock_object *lock)
1374 {
1375 	struct witness *w;
1376 
1377 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1378 		return ("?");
1379 	w = lock->lo_witness;
1380 	return (w->w_file);
1381 }
1382 
1383 int
1384 witness_line(struct lock_object *lock)
1385 {
1386 	struct witness *w;
1387 
1388 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1389 		return (0);
1390 	w = lock->lo_witness;
1391 	return (w->w_line);
1392 }
1393 
1394 static struct witness *
1395 enroll(const char *description, struct lock_class *lock_class)
1396 {
1397 	struct witness *w;
1398 
1399 	if (witness_watch == 0 || panicstr != NULL)
1400 		return (NULL);
1401 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1402 		return (NULL);
1403 	mtx_lock_spin(&w_mtx);
1404 	STAILQ_FOREACH(w, &w_all, w_list) {
1405 		if (w->w_name == description || (w->w_refcount > 0 &&
1406 		    strcmp(description, w->w_name) == 0)) {
1407 			w->w_refcount++;
1408 			mtx_unlock_spin(&w_mtx);
1409 			if (lock_class != w->w_class)
1410 				panic(
1411 				"lock (%s) %s does not match earlier (%s) lock",
1412 				    description, lock_class->lc_name,
1413 				    w->w_class->lc_name);
1414 			return (w);
1415 		}
1416 	}
1417 	if ((w = witness_get()) == NULL)
1418 		goto out;
1419 	w->w_name = description;
1420 	w->w_class = lock_class;
1421 	w->w_refcount = 1;
1422 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1423 	if (lock_class->lc_flags & LC_SPINLOCK) {
1424 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1425 		w_spin_cnt++;
1426 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1427 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1428 		w_sleep_cnt++;
1429 	} else {
1430 		mtx_unlock_spin(&w_mtx);
1431 		panic("lock class %s is not sleep or spin",
1432 		    lock_class->lc_name);
1433 	}
1434 	mtx_unlock_spin(&w_mtx);
1435 out:
1436 	/*
1437 	 * We issue a warning for any spin locks not defined in the static
1438 	 * order list as a way to discourage their use (folks should really
1439 	 * be using non-spin mutexes most of the time).  However, several
1440 	 * 3rd part device drivers use spin locks because that is all they
1441 	 * have available on Windows and Linux and they think that normal
1442 	 * mutexes are insufficient.
1443 	 */
1444 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1445 		printf("WITNESS: spin lock %s not in order list\n",
1446 		    description);
1447 	return (w);
1448 }
1449 
1450 /* Don't let the door bang you on the way out... */
1451 static int
1452 depart(struct witness *w)
1453 {
1454 	struct witness_child_list_entry *wcl, *nwcl;
1455 	struct witness_list *list;
1456 	struct witness *parent;
1457 
1458 	MPASS(w->w_refcount == 0);
1459 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1460 		list = &w_sleep;
1461 		w_sleep_cnt--;
1462 	} else {
1463 		list = &w_spin;
1464 		w_spin_cnt--;
1465 	}
1466 	/*
1467 	 * First, we run through the entire tree looking for any
1468 	 * witnesses that the outgoing witness is a child of.  For
1469 	 * each parent that we find, we reparent all the direct
1470 	 * children of the outgoing witness to its parent.
1471 	 */
1472 	STAILQ_FOREACH(parent, list, w_typelist) {
1473 		if (!isitmychild(parent, w))
1474 			continue;
1475 		removechild(parent, w);
1476 	}
1477 
1478 	/*
1479 	 * Now we go through and free up the child list of the
1480 	 * outgoing witness.
1481 	 */
1482 	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1483 		nwcl = wcl->wcl_next;
1484         	w_child_cnt--;
1485 		witness_child_free(wcl);
1486 	}
1487 
1488 	/*
1489 	 * Detach from various lists and free.
1490 	 */
1491 	STAILQ_REMOVE(list, w, witness, w_typelist);
1492 	STAILQ_REMOVE(&w_all, w, witness, w_list);
1493 	witness_free(w);
1494 
1495 	return (1);
1496 }
1497 
1498 /*
1499  * Add "child" as a direct child of "parent".  Returns false if
1500  * we fail due to out of memory.
1501  */
1502 static int
1503 insertchild(struct witness *parent, struct witness *child)
1504 {
1505 	struct witness_child_list_entry **wcl;
1506 
1507 	MPASS(child != NULL && parent != NULL);
1508 
1509 	/*
1510 	 * Insert "child" after "parent"
1511 	 */
1512 	wcl = &parent->w_children;
1513 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1514 		wcl = &(*wcl)->wcl_next;
1515 	if (*wcl == NULL) {
1516 		*wcl = witness_child_get();
1517 		if (*wcl == NULL)
1518 			return (0);
1519         	w_child_cnt++;
1520 	}
1521 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1522 
1523 	return (1);
1524 }
1525 
1526 
1527 static int
1528 itismychild(struct witness *parent, struct witness *child)
1529 {
1530 	struct witness_list *list;
1531 
1532 	MPASS(child != NULL && parent != NULL);
1533 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1534 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1535 		panic(
1536 		"%s: parent (%s) and child (%s) are not the same lock type",
1537 		    __func__, parent->w_class->lc_name,
1538 		    child->w_class->lc_name);
1539 
1540 	if (!insertchild(parent, child))
1541 		return (0);
1542 
1543 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1544 		list = &w_sleep;
1545 	else
1546 		list = &w_spin;
1547 	return (1);
1548 }
1549 
1550 static void
1551 removechild(struct witness *parent, struct witness *child)
1552 {
1553 	struct witness_child_list_entry **wcl, *wcl1;
1554 	int i;
1555 
1556 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1557 		for (i = 0; i < (*wcl)->wcl_count; i++)
1558 			if ((*wcl)->wcl_children[i] == child)
1559 				goto found;
1560 	return;
1561 found:
1562 	(*wcl)->wcl_count--;
1563 	if ((*wcl)->wcl_count > i)
1564 		(*wcl)->wcl_children[i] =
1565 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1566 	MPASS((*wcl)->wcl_children[i] != NULL);
1567 	if ((*wcl)->wcl_count != 0)
1568 		return;
1569 	wcl1 = *wcl;
1570 	*wcl = wcl1->wcl_next;
1571 	w_child_cnt--;
1572 	witness_child_free(wcl1);
1573 }
1574 
1575 static int
1576 isitmychild(struct witness *parent, struct witness *child)
1577 {
1578 	struct witness_child_list_entry *wcl;
1579 	int i;
1580 
1581 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1582 		for (i = 0; i < wcl->wcl_count; i++) {
1583 			if (wcl->wcl_children[i] == child)
1584 				return (1);
1585 		}
1586 	}
1587 	return (0);
1588 }
1589 
1590 static int
1591 isitmydescendant(struct witness *parent, struct witness *child)
1592 {
1593 	struct witness_child_list_entry *wcl;
1594 	int i, j;
1595 
1596 	if (isitmychild(parent, child))
1597 		return (1);
1598 	j = 0;
1599 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1600 		MPASS(j < 1000);
1601 		for (i = 0; i < wcl->wcl_count; i++) {
1602 			if (isitmydescendant(wcl->wcl_children[i], child))
1603 				return (1);
1604 		}
1605 		j++;
1606 	}
1607 	return (0);
1608 }
1609 
1610 #ifdef BLESSING
1611 static int
1612 blessed(struct witness *w1, struct witness *w2)
1613 {
1614 	int i;
1615 	struct witness_blessed *b;
1616 
1617 	for (i = 0; i < blessed_count; i++) {
1618 		b = &blessed_list[i];
1619 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1620 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1621 				return (1);
1622 			continue;
1623 		}
1624 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1625 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1626 				return (1);
1627 	}
1628 	return (0);
1629 }
1630 #endif
1631 
1632 static struct witness *
1633 witness_get(void)
1634 {
1635 	struct witness *w;
1636 
1637 	if (witness_watch == 0) {
1638 		mtx_unlock_spin(&w_mtx);
1639 		return (NULL);
1640 	}
1641 	if (STAILQ_EMPTY(&w_free)) {
1642 		witness_watch = 0;
1643 		mtx_unlock_spin(&w_mtx);
1644 		printf("%s: witness exhausted\n", __func__);
1645 		return (NULL);
1646 	}
1647 	w = STAILQ_FIRST(&w_free);
1648 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1649 	w_free_cnt--;
1650 	bzero(w, sizeof(*w));
1651 	return (w);
1652 }
1653 
1654 static void
1655 witness_free(struct witness *w)
1656 {
1657 
1658 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1659 	w_free_cnt++;
1660 }
1661 
1662 static struct witness_child_list_entry *
1663 witness_child_get(void)
1664 {
1665 	struct witness_child_list_entry *wcl;
1666 
1667 	if (witness_watch == 0) {
1668 		mtx_unlock_spin(&w_mtx);
1669 		return (NULL);
1670 	}
1671 	wcl = w_child_free;
1672 	if (wcl == NULL) {
1673 		witness_watch = 0;
1674 		mtx_unlock_spin(&w_mtx);
1675 		printf("%s: witness exhausted\n", __func__);
1676 		return (NULL);
1677 	}
1678 	w_child_free = wcl->wcl_next;
1679 	w_child_free_cnt--;
1680 	bzero(wcl, sizeof(*wcl));
1681 	return (wcl);
1682 }
1683 
1684 static void
1685 witness_child_free(struct witness_child_list_entry *wcl)
1686 {
1687 
1688 	wcl->wcl_next = w_child_free;
1689 	w_child_free = wcl;
1690 	w_child_free_cnt++;
1691 }
1692 
1693 static struct lock_list_entry *
1694 witness_lock_list_get(void)
1695 {
1696 	struct lock_list_entry *lle;
1697 
1698 	if (witness_watch == 0)
1699 		return (NULL);
1700 	mtx_lock_spin(&w_mtx);
1701 	lle = w_lock_list_free;
1702 	if (lle == NULL) {
1703 		witness_watch = 0;
1704 		mtx_unlock_spin(&w_mtx);
1705 		printf("%s: witness exhausted\n", __func__);
1706 		return (NULL);
1707 	}
1708 	w_lock_list_free = lle->ll_next;
1709 	mtx_unlock_spin(&w_mtx);
1710 	bzero(lle, sizeof(*lle));
1711 	return (lle);
1712 }
1713 
1714 static void
1715 witness_lock_list_free(struct lock_list_entry *lle)
1716 {
1717 
1718 	mtx_lock_spin(&w_mtx);
1719 	lle->ll_next = w_lock_list_free;
1720 	w_lock_list_free = lle;
1721 	mtx_unlock_spin(&w_mtx);
1722 }
1723 
1724 static struct lock_instance *
1725 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1726 {
1727 	struct lock_list_entry *lle;
1728 	struct lock_instance *instance;
1729 	int i;
1730 
1731 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1732 		for (i = lle->ll_count - 1; i >= 0; i--) {
1733 			instance = &lle->ll_children[i];
1734 			if (instance->li_lock == lock)
1735 				return (instance);
1736 		}
1737 	return (NULL);
1738 }
1739 
1740 static void
1741 witness_list_lock(struct lock_instance *instance)
1742 {
1743 	struct lock_object *lock;
1744 
1745 	lock = instance->li_lock;
1746 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1747 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1748 	if (lock->lo_type != lock->lo_name)
1749 		printf(" (%s)", lock->lo_type);
1750 	printf(" r = %d (%p) locked @ %s:%d\n",
1751 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1752 	    instance->li_line);
1753 }
1754 
1755 #ifdef DDB
1756 static int
1757 witness_thread_has_locks(struct thread *td)
1758 {
1759 
1760 	return (td->td_sleeplocks != NULL);
1761 }
1762 
1763 static int
1764 witness_proc_has_locks(struct proc *p)
1765 {
1766 	struct thread *td;
1767 
1768 	FOREACH_THREAD_IN_PROC(p, td) {
1769 		if (witness_thread_has_locks(td))
1770 			return (1);
1771 	}
1772 	return (0);
1773 }
1774 #endif
1775 
1776 int
1777 witness_list_locks(struct lock_list_entry **lock_list)
1778 {
1779 	struct lock_list_entry *lle;
1780 	int i, nheld;
1781 
1782 	nheld = 0;
1783 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1784 		for (i = lle->ll_count - 1; i >= 0; i--) {
1785 			witness_list_lock(&lle->ll_children[i]);
1786 			nheld++;
1787 		}
1788 	return (nheld);
1789 }
1790 
1791 /*
1792  * This is a bit risky at best.  We call this function when we have timed
1793  * out acquiring a spin lock, and we assume that the other CPU is stuck
1794  * with this lock held.  So, we go groveling around in the other CPU's
1795  * per-cpu data to try to find the lock instance for this spin lock to
1796  * see when it was last acquired.
1797  */
1798 void
1799 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1800 {
1801 	struct lock_instance *instance;
1802 	struct pcpu *pc;
1803 
1804 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1805 		return;
1806 	pc = pcpu_find(owner->td_oncpu);
1807 	instance = find_instance(pc->pc_spinlocks, lock);
1808 	if (instance != NULL)
1809 		witness_list_lock(instance);
1810 }
1811 
1812 void
1813 witness_save(struct lock_object *lock, const char **filep, int *linep)
1814 {
1815 	struct lock_list_entry *lock_list;
1816 	struct lock_instance *instance;
1817 	struct lock_class *class;
1818 
1819 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1820 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1821 		return;
1822 	class = LOCK_CLASS(lock);
1823 	if (class->lc_flags & LC_SLEEPLOCK)
1824 		lock_list = curthread->td_sleeplocks;
1825 	else {
1826 		if (witness_skipspin)
1827 			return;
1828 		lock_list = PCPU_GET(spinlocks);
1829 	}
1830 	instance = find_instance(lock_list, lock);
1831 	if (instance == NULL)
1832 		panic("%s: lock (%s) %s not locked", __func__,
1833 		    class->lc_name, lock->lo_name);
1834 	*filep = instance->li_file;
1835 	*linep = instance->li_line;
1836 }
1837 
1838 void
1839 witness_restore(struct lock_object *lock, const char *file, int line)
1840 {
1841 	struct lock_list_entry *lock_list;
1842 	struct lock_instance *instance;
1843 	struct lock_class *class;
1844 
1845 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1846 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1847 		return;
1848 	class = LOCK_CLASS(lock);
1849 	if (class->lc_flags & LC_SLEEPLOCK)
1850 		lock_list = curthread->td_sleeplocks;
1851 	else {
1852 		if (witness_skipspin)
1853 			return;
1854 		lock_list = PCPU_GET(spinlocks);
1855 	}
1856 	instance = find_instance(lock_list, lock);
1857 	if (instance == NULL)
1858 		panic("%s: lock (%s) %s not locked", __func__,
1859 		    class->lc_name, lock->lo_name);
1860 	lock->lo_witness->w_file = file;
1861 	lock->lo_witness->w_line = line;
1862 	instance->li_file = file;
1863 	instance->li_line = line;
1864 }
1865 
1866 void
1867 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1868 {
1869 #ifdef INVARIANT_SUPPORT
1870 	struct lock_instance *instance;
1871 	struct lock_class *class;
1872 
1873 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1874 		return;
1875 	class = LOCK_CLASS(lock);
1876 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1877 		instance = find_instance(curthread->td_sleeplocks, lock);
1878 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
1879 		instance = find_instance(PCPU_GET(spinlocks), lock);
1880 	else {
1881 		panic("Lock (%s) %s is not sleep or spin!",
1882 		    class->lc_name, lock->lo_name);
1883 	}
1884 	file = fixup_filename(file);
1885 	switch (flags) {
1886 	case LA_UNLOCKED:
1887 		if (instance != NULL)
1888 			panic("Lock (%s) %s locked @ %s:%d.",
1889 			    class->lc_name, lock->lo_name, file, line);
1890 		break;
1891 	case LA_LOCKED:
1892 	case LA_LOCKED | LA_RECURSED:
1893 	case LA_LOCKED | LA_NOTRECURSED:
1894 	case LA_SLOCKED:
1895 	case LA_SLOCKED | LA_RECURSED:
1896 	case LA_SLOCKED | LA_NOTRECURSED:
1897 	case LA_XLOCKED:
1898 	case LA_XLOCKED | LA_RECURSED:
1899 	case LA_XLOCKED | LA_NOTRECURSED:
1900 		if (instance == NULL) {
1901 			panic("Lock (%s) %s not locked @ %s:%d.",
1902 			    class->lc_name, lock->lo_name, file, line);
1903 			break;
1904 		}
1905 		if ((flags & LA_XLOCKED) != 0 &&
1906 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1907 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1908 			    class->lc_name, lock->lo_name, file, line);
1909 		if ((flags & LA_SLOCKED) != 0 &&
1910 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1911 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1912 			    class->lc_name, lock->lo_name, file, line);
1913 		if ((flags & LA_RECURSED) != 0 &&
1914 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1915 			panic("Lock (%s) %s not recursed @ %s:%d.",
1916 			    class->lc_name, lock->lo_name, file, line);
1917 		if ((flags & LA_NOTRECURSED) != 0 &&
1918 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1919 			panic("Lock (%s) %s recursed @ %s:%d.",
1920 			    class->lc_name, lock->lo_name, file, line);
1921 		break;
1922 	default:
1923 		panic("Invalid lock assertion at %s:%d.", file, line);
1924 
1925 	}
1926 #endif	/* INVARIANT_SUPPORT */
1927 }
1928 
1929 #ifdef DDB
1930 static void
1931 witness_list(struct thread *td)
1932 {
1933 
1934 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1935 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1936 
1937 	if (witness_watch == 0)
1938 		return;
1939 
1940 	witness_list_locks(&td->td_sleeplocks);
1941 
1942 	/*
1943 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1944 	 * if td is currently executing on some other CPU and holds spin locks
1945 	 * as we won't display those locks.  If we had a MI way of getting
1946 	 * the per-cpu data for a given cpu then we could use
1947 	 * td->td_oncpu to get the list of spinlocks for this thread
1948 	 * and "fix" this.
1949 	 *
1950 	 * That still wouldn't really fix this unless we locked sched_lock
1951 	 * or stopped the other CPU to make sure it wasn't changing the list
1952 	 * out from under us.  It is probably best to just not try to handle
1953 	 * threads on other CPU's for now.
1954 	 */
1955 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1956 		witness_list_locks(PCPU_PTR(spinlocks));
1957 }
1958 
1959 DB_SHOW_COMMAND(locks, db_witness_list)
1960 {
1961 	struct thread *td;
1962 
1963 	if (have_addr)
1964 		td = db_lookup_thread(addr, TRUE);
1965 	else
1966 		td = kdb_thread;
1967 	witness_list(td);
1968 }
1969 
1970 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1971 {
1972 	struct thread *td;
1973 	struct proc *p;
1974 
1975 	/*
1976 	 * It would be nice to list only threads and processes that actually
1977 	 * held sleep locks, but that information is currently not exported
1978 	 * by WITNESS.
1979 	 */
1980 	FOREACH_PROC_IN_SYSTEM(p) {
1981 		if (!witness_proc_has_locks(p))
1982 			continue;
1983 		FOREACH_THREAD_IN_PROC(p, td) {
1984 			if (!witness_thread_has_locks(td))
1985 				continue;
1986 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1987 			    p->p_comm, td, td->td_tid);
1988 			witness_list(td);
1989 		}
1990 	}
1991 }
1992 
1993 DB_SHOW_COMMAND(witness, db_witness_display)
1994 {
1995 
1996 	witness_display(db_printf);
1997 }
1998 #endif
1999