xref: /freebsd/sys/kern/subr_witness.c (revision 02b0a160dcf4c6458cfcc86bc11aec043439bf87)
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_hwpmc_hooks.h"
89 #include "opt_witness.h"
90 
91 #include <sys/param.h>
92 #include <sys/bus.h>
93 #include <sys/kdb.h>
94 #include <sys/kernel.h>
95 #include <sys/ktr.h>
96 #include <sys/lock.h>
97 #include <sys/malloc.h>
98 #include <sys/mutex.h>
99 #include <sys/priv.h>
100 #include <sys/proc.h>
101 #include <sys/sysctl.h>
102 #include <sys/systm.h>
103 
104 #include <ddb/ddb.h>
105 
106 #include <machine/stdarg.h>
107 
108 /* Note that these traces do not work with KTR_ALQ. */
109 #if 0
110 #define	KTR_WITNESS	KTR_SUBSYS
111 #else
112 #define	KTR_WITNESS	0
113 #endif
114 
115 /* Easier to stay with the old names. */
116 #define	lo_list		lo_witness_data.lod_list
117 #define	lo_witness	lo_witness_data.lod_witness
118 
119 /* Define this to check for blessed mutexes */
120 #undef BLESSING
121 
122 #define WITNESS_COUNT 1024
123 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
124 /*
125  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
126  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
127  * probably be safe for the most part, but it's still a SWAG.
128  */
129 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
130 
131 #define	WITNESS_NCHILDREN 6
132 
133 struct witness_child_list_entry;
134 
135 struct witness {
136 	const	char *w_name;
137 	struct	lock_class *w_class;
138 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
139 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
140 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
141 	const	char *w_file;
142 	int	w_line;
143 	u_int	w_level;
144 	u_int	w_refcount;
145 	u_char	w_Giant_squawked:1;
146 	u_char	w_other_squawked:1;
147 	u_char	w_same_squawked:1;
148 	u_char	w_displayed:1;
149 };
150 
151 struct witness_child_list_entry {
152 	struct	witness_child_list_entry *wcl_next;
153 	struct	witness *wcl_children[WITNESS_NCHILDREN];
154 	u_int	wcl_count;
155 };
156 
157 STAILQ_HEAD(witness_list, witness);
158 
159 #ifdef BLESSING
160 struct witness_blessed {
161 	const	char *b_lock1;
162 	const	char *b_lock2;
163 };
164 #endif
165 
166 struct witness_order_list_entry {
167 	const	char *w_name;
168 	struct	lock_class *w_class;
169 };
170 
171 #ifdef BLESSING
172 static int	blessed(struct witness *, struct witness *);
173 #endif
174 static int	depart(struct witness *w);
175 static struct	witness *enroll(const char *description,
176 				struct lock_class *lock_class);
177 static int	insertchild(struct witness *parent, struct witness *child);
178 static int	isitmychild(struct witness *parent, struct witness *child);
179 static int	isitmydescendant(struct witness *parent, struct witness *child);
180 static int	itismychild(struct witness *parent, struct witness *child);
181 static void	removechild(struct witness *parent, struct witness *child);
182 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
183 static const char *fixup_filename(const char *file);
184 static struct	witness *witness_get(void);
185 static void	witness_free(struct witness *m);
186 static struct	witness_child_list_entry *witness_child_get(void);
187 static void	witness_child_free(struct witness_child_list_entry *wcl);
188 static struct	lock_list_entry *witness_lock_list_get(void);
189 static void	witness_lock_list_free(struct lock_list_entry *lle);
190 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
191 					     struct lock_object *lock);
192 static void	witness_list_lock(struct lock_instance *instance);
193 #ifdef DDB
194 static void	witness_leveldescendents(struct witness *parent, int level);
195 static void	witness_levelall(void);
196 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
197 					   struct witness *, int indent);
198 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
199 				     struct witness_list *list);
200 static void	witness_display(void(*)(const char *fmt, ...));
201 static void	witness_list(struct thread *td);
202 #endif
203 
204 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
205 
206 /*
207  * If set to 0, witness is disabled.  If set to a non-zero value, witness
208  * performs full lock order checking for all locks.  At runtime, this
209  * value may be set to 0 to turn off witness.  witness is not allowed be
210  * turned on once it is turned off, however.
211  */
212 static int witness_watch = 1;
213 TUNABLE_INT("debug.witness.watch", &witness_watch);
214 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
215     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
216 
217 #ifdef KDB
218 /*
219  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
220  * to drop into kdebug() when:
221  *	- a lock hierarchy violation occurs
222  *	- locks are held when going to sleep.
223  */
224 #ifdef WITNESS_KDB
225 int	witness_kdb = 1;
226 #else
227 int	witness_kdb = 0;
228 #endif
229 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
230 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
231 
232 /*
233  * When KDB is enabled and witness_trace is set to 1, it will cause the system
234  * to print a stack trace:
235  *	- a lock hierarchy violation occurs
236  *	- locks are held when going to sleep.
237  */
238 int	witness_trace = 1;
239 TUNABLE_INT("debug.witness.trace", &witness_trace);
240 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
241 #endif /* KDB */
242 
243 #ifdef WITNESS_SKIPSPIN
244 int	witness_skipspin = 1;
245 #else
246 int	witness_skipspin = 0;
247 #endif
248 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
249 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
250     &witness_skipspin, 0, "");
251 
252 static struct mtx w_mtx;
253 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
254 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
255 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
256 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
257 static struct witness_child_list_entry *w_child_free = NULL;
258 static struct lock_list_entry *w_lock_list_free = NULL;
259 
260 static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt;
261 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
262 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
263 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
264     "");
265 SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD,
266     &w_child_free_cnt, 0, "");
267 SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0,
268     "");
269 
270 static struct witness w_data[WITNESS_COUNT];
271 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
272 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
273 
274 static struct witness_order_list_entry order_lists[] = {
275 	/*
276 	 * sx locks
277 	 */
278 	{ "proctree", &lock_class_sx },
279 	{ "allproc", &lock_class_sx },
280 	{ "allprison", &lock_class_sx },
281 	{ NULL, NULL },
282 	/*
283 	 * Various mutexes
284 	 */
285 	{ "Giant", &lock_class_mtx_sleep },
286 	{ "pipe mutex", &lock_class_mtx_sleep },
287 	{ "sigio lock", &lock_class_mtx_sleep },
288 	{ "process group", &lock_class_mtx_sleep },
289 	{ "process lock", &lock_class_mtx_sleep },
290 	{ "session", &lock_class_mtx_sleep },
291 	{ "uidinfo hash", &lock_class_mtx_sleep },
292 	{ "uidinfo struct", &lock_class_mtx_sleep },
293 #ifdef	HWPMC_HOOKS
294 	{ "pmc-sleep", &lock_class_mtx_sleep },
295 #endif
296 	{ NULL, NULL },
297 	/*
298 	 * Sockets
299 	 */
300 	{ "accept", &lock_class_mtx_sleep },
301 	{ "so_snd", &lock_class_mtx_sleep },
302 	{ "so_rcv", &lock_class_mtx_sleep },
303 	{ "sellck", &lock_class_mtx_sleep },
304 	{ NULL, NULL },
305 	/*
306 	 * Routing
307 	 */
308 	{ "so_rcv", &lock_class_mtx_sleep },
309 	{ "radix node head", &lock_class_mtx_sleep },
310 	{ "rtentry", &lock_class_mtx_sleep },
311 	{ "ifaddr", &lock_class_mtx_sleep },
312 	{ NULL, NULL },
313 	/*
314 	 * Multicast - protocol locks before interface locks, after UDP locks.
315 	 */
316 	{ "udpinp", &lock_class_mtx_sleep },
317 	{ "in_multi_mtx", &lock_class_mtx_sleep },
318 	{ "igmp_mtx", &lock_class_mtx_sleep },
319 	{ "if_addr_mtx", &lock_class_mtx_sleep },
320 	{ NULL, NULL },
321 	/*
322 	 * UNIX Domain Sockets
323 	 */
324 	{ "unp", &lock_class_mtx_sleep },
325 	{ "so_snd", &lock_class_mtx_sleep },
326 	{ NULL, NULL },
327 	/*
328 	 * UDP/IP
329 	 */
330 	{ "udp", &lock_class_mtx_sleep },
331 	{ "udpinp", &lock_class_mtx_sleep },
332 	{ "so_snd", &lock_class_mtx_sleep },
333 	{ NULL, NULL },
334 	/*
335 	 * TCP/IP
336 	 */
337 	{ "tcp", &lock_class_mtx_sleep },
338 	{ "tcpinp", &lock_class_mtx_sleep },
339 	{ "so_snd", &lock_class_mtx_sleep },
340 	{ NULL, NULL },
341 	/*
342 	 * SLIP
343 	 */
344 	{ "slip_mtx", &lock_class_mtx_sleep },
345 	{ "slip sc_mtx", &lock_class_mtx_sleep },
346 	{ NULL, NULL },
347 	/*
348 	 * netatalk
349 	 */
350 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
351 	{ "ddp_mtx", &lock_class_mtx_sleep },
352 	{ NULL, NULL },
353 	/*
354 	 * BPF
355 	 */
356 	{ "bpf global lock", &lock_class_mtx_sleep },
357 	{ "bpf interface lock", &lock_class_mtx_sleep },
358 	{ "bpf cdev lock", &lock_class_mtx_sleep },
359 	{ NULL, NULL },
360 	/*
361 	 * NFS server
362 	 */
363 	{ "nfsd_mtx", &lock_class_mtx_sleep },
364 	{ "so_snd", &lock_class_mtx_sleep },
365 	{ NULL, NULL },
366 	/*
367 	 * Netgraph
368 	 */
369 	{ "ng_node", &lock_class_mtx_sleep },
370 	{ "ng_worklist", &lock_class_mtx_sleep },
371 	{ NULL, NULL },
372 	/*
373 	 * CDEV
374 	 */
375 	{ "system map", &lock_class_mtx_sleep },
376 	{ "vm page queue mutex", &lock_class_mtx_sleep },
377 	{ "vnode interlock", &lock_class_mtx_sleep },
378 	{ "cdev", &lock_class_mtx_sleep },
379 	{ NULL, NULL },
380 	/*
381 	 * kqueue/VFS interaction
382 	 */
383 	{ "kqueue", &lock_class_mtx_sleep },
384 	{ "struct mount mtx", &lock_class_mtx_sleep },
385 	{ "vnode interlock", &lock_class_mtx_sleep },
386 	{ NULL, NULL },
387 	/*
388 	 * spin locks
389 	 */
390 #ifdef SMP
391 	{ "ap boot", &lock_class_mtx_spin },
392 #endif
393 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
394 	{ "sio", &lock_class_mtx_spin },
395 #ifdef __i386__
396 	{ "cy", &lock_class_mtx_spin },
397 	{ "descriptor tables", &lock_class_mtx_spin },
398 #endif
399 	{ "scc_hwmtx", &lock_class_mtx_spin },
400 	{ "uart_hwmtx", &lock_class_mtx_spin },
401 	{ "zstty", &lock_class_mtx_spin },
402 	{ "fast_taskqueue", &lock_class_mtx_spin },
403 	{ "intr table", &lock_class_mtx_spin },
404 #ifdef	HWPMC_HOOKS
405 	{ "pmc-per-proc", &lock_class_mtx_spin },
406 #endif
407 	{ "sleepq chain", &lock_class_mtx_spin },
408 	{ "sched lock", &lock_class_mtx_spin },
409 	{ "turnstile chain", &lock_class_mtx_spin },
410 	{ "td_contested", &lock_class_mtx_spin },
411 	{ "callout", &lock_class_mtx_spin },
412 	{ "entropy harvest mutex", &lock_class_mtx_spin },
413 	{ "syscons video lock", &lock_class_mtx_spin },
414 	{ "time lock", &lock_class_mtx_spin },
415 	/*
416 	 * leaf locks
417 	 */
418 	{ "allpmaps", &lock_class_mtx_spin },
419 	{ "icu", &lock_class_mtx_spin },
420 #ifdef SMP
421 	{ "smp rendezvous", &lock_class_mtx_spin },
422 #if defined(__i386__) || defined(__amd64__)
423 	{ "tlb", &lock_class_mtx_spin },
424 #endif
425 #ifdef __sparc64__
426 	{ "ipi", &lock_class_mtx_spin },
427 	{ "rtc_mtx", &lock_class_mtx_spin },
428 #endif
429 #endif
430 	{ "clk", &lock_class_mtx_spin },
431 	{ "mutex profiling lock", &lock_class_mtx_spin },
432 	{ "kse zombie lock", &lock_class_mtx_spin },
433 	{ "ALD Queue", &lock_class_mtx_spin },
434 #ifdef __ia64__
435 	{ "MCA spin lock", &lock_class_mtx_spin },
436 #endif
437 #if defined(__i386__) || defined(__amd64__)
438 	{ "pcicfg", &lock_class_mtx_spin },
439 	{ "NDIS thread lock", &lock_class_mtx_spin },
440 #endif
441 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
442 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
443 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
444 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
445 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
446 #ifdef	HWPMC_HOOKS
447 	{ "pmc-leaf", &lock_class_mtx_spin },
448 #endif
449 	{ NULL, NULL },
450 	{ NULL, NULL }
451 };
452 
453 #ifdef BLESSING
454 /*
455  * Pairs of locks which have been blessed
456  * Don't complain about order problems with blessed locks
457  */
458 static struct witness_blessed blessed_list[] = {
459 };
460 static int blessed_count =
461 	sizeof(blessed_list) / sizeof(struct witness_blessed);
462 #endif
463 
464 /*
465  * List of locks initialized prior to witness being initialized whose
466  * enrollment is currently deferred.
467  */
468 STAILQ_HEAD(, lock_object) pending_locks =
469     STAILQ_HEAD_INITIALIZER(pending_locks);
470 
471 /*
472  * This global is set to 0 once it becomes safe to use the witness code.
473  */
474 static int witness_cold = 1;
475 
476 /*
477  * This global is set to 1 once the static lock orders have been enrolled
478  * so that a warning can be issued for any spin locks enrolled later.
479  */
480 static int witness_spin_warn = 0;
481 
482 /*
483  * The WITNESS-enabled diagnostic code.  Note that the witness code does
484  * assume that the early boot is single-threaded at least until after this
485  * routine is completed.
486  */
487 static void
488 witness_initialize(void *dummy __unused)
489 {
490 	struct lock_object *lock;
491 	struct witness_order_list_entry *order;
492 	struct witness *w, *w1;
493 	int i;
494 
495 	/*
496 	 * We have to release Giant before initializing its witness
497 	 * structure so that WITNESS doesn't get confused.
498 	 */
499 	mtx_unlock(&Giant);
500 	mtx_assert(&Giant, MA_NOTOWNED);
501 
502 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
503 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
504 	    MTX_NOWITNESS | MTX_NOPROFILE);
505 	for (i = 0; i < WITNESS_COUNT; i++)
506 		witness_free(&w_data[i]);
507 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
508 		witness_child_free(&w_childdata[i]);
509 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
510 		witness_lock_list_free(&w_locklistdata[i]);
511 
512 	/* First add in all the specified order lists. */
513 	for (order = order_lists; order->w_name != NULL; order++) {
514 		w = enroll(order->w_name, order->w_class);
515 		if (w == NULL)
516 			continue;
517 		w->w_file = "order list";
518 		for (order++; order->w_name != NULL; order++) {
519 			w1 = enroll(order->w_name, order->w_class);
520 			if (w1 == NULL)
521 				continue;
522 			w1->w_file = "order list";
523 			if (!itismychild(w, w1))
524 				panic("Not enough memory for static orders!");
525 			w = w1;
526 		}
527 	}
528 	witness_spin_warn = 1;
529 
530 	/* Iterate through all locks and add them to witness. */
531 	while (!STAILQ_EMPTY(&pending_locks)) {
532 		lock = STAILQ_FIRST(&pending_locks);
533 		STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
534 		KASSERT(lock->lo_flags & LO_WITNESS,
535 		    ("%s: lock %s is on pending list but not LO_WITNESS",
536 		    __func__, lock->lo_name));
537 		lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
538 	}
539 
540 	/* Mark the witness code as being ready for use. */
541 	witness_cold = 0;
542 
543 	mtx_lock(&Giant);
544 }
545 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
546 
547 static int
548 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
549 {
550 	int error, value;
551 
552 	value = witness_watch;
553 	error = sysctl_handle_int(oidp, &value, 0, req);
554 	if (error != 0 || req->newptr == NULL)
555 		return (error);
556 	if (value == witness_watch)
557 		return (0);
558 	if (value != 0)
559 		return (EINVAL);
560 	witness_watch = 0;
561 	return (0);
562 }
563 
564 void
565 witness_init(struct lock_object *lock)
566 {
567 	struct lock_class *class;
568 
569 	/* Various sanity checks. */
570 	class = LOCK_CLASS(lock);
571 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
572 	    (class->lc_flags & LC_RECURSABLE) == 0)
573 		panic("%s: lock (%s) %s can not be recursable", __func__,
574 		    class->lc_name, lock->lo_name);
575 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
576 	    (class->lc_flags & LC_SLEEPABLE) == 0)
577 		panic("%s: lock (%s) %s can not be sleepable", __func__,
578 		    class->lc_name, lock->lo_name);
579 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
580 	    (class->lc_flags & LC_UPGRADABLE) == 0)
581 		panic("%s: lock (%s) %s can not be upgradable", __func__,
582 		    class->lc_name, lock->lo_name);
583 
584 	/*
585 	 * If we shouldn't watch this lock, then just clear lo_witness.
586 	 * Otherwise, if witness_cold is set, then it is too early to
587 	 * enroll this lock, so defer it to witness_initialize() by adding
588 	 * it to the pending_locks list.  If it is not too early, then enroll
589 	 * the lock now.
590 	 */
591 	if (witness_watch == 0 || panicstr != NULL ||
592 	    (lock->lo_flags & LO_WITNESS) == 0)
593 		lock->lo_witness = NULL;
594 	else if (witness_cold) {
595 		STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
596 		lock->lo_flags |= LO_ENROLLPEND;
597 	} else
598 		lock->lo_witness = enroll(lock->lo_type, class);
599 }
600 
601 void
602 witness_destroy(struct lock_object *lock)
603 {
604 	struct lock_class *class;
605 	struct witness *w;
606 
607 	class = LOCK_CLASS(lock);
608 	if (witness_cold)
609 		panic("lock (%s) %s destroyed while witness_cold",
610 		    class->lc_name, lock->lo_name);
611 
612 	/* XXX: need to verify that no one holds the lock */
613 	if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
614 	    lock->lo_witness != NULL) {
615 		w = lock->lo_witness;
616 		mtx_lock_spin(&w_mtx);
617 		MPASS(w->w_refcount > 0);
618 		w->w_refcount--;
619 
620 		/*
621 		 * Lock is already released if we have an allocation failure
622 		 * and depart() fails.
623 		 */
624 		if (w->w_refcount != 0 || depart(w))
625 			mtx_unlock_spin(&w_mtx);
626 	}
627 
628 	/*
629 	 * If this lock is destroyed before witness is up and running,
630 	 * remove it from the pending list.
631 	 */
632 	if (lock->lo_flags & LO_ENROLLPEND) {
633 		STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
634 		lock->lo_flags &= ~LO_ENROLLPEND;
635 	}
636 }
637 
638 #ifdef DDB
639 static void
640 witness_levelall (void)
641 {
642 	struct witness_list *list;
643 	struct witness *w, *w1;
644 
645 	/*
646 	 * First clear all levels.
647 	 */
648 	STAILQ_FOREACH(w, &w_all, w_list) {
649 		w->w_level = 0;
650 	}
651 
652 	/*
653 	 * Look for locks with no parent and level all their descendants.
654 	 */
655 	STAILQ_FOREACH(w, &w_all, w_list) {
656 		/*
657 		 * This is just an optimization, technically we could get
658 		 * away just walking the all list each time.
659 		 */
660 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
661 			list = &w_sleep;
662 		else
663 			list = &w_spin;
664 		STAILQ_FOREACH(w1, list, w_typelist) {
665 			if (isitmychild(w1, w))
666 				goto skip;
667 		}
668 		witness_leveldescendents(w, 0);
669 	skip:
670 		;	/* silence GCC 3.x */
671 	}
672 }
673 
674 static void
675 witness_leveldescendents(struct witness *parent, int level)
676 {
677 	struct witness_child_list_entry *wcl;
678 	int i;
679 
680 	if (parent->w_level < level)
681 		parent->w_level = level;
682 	level++;
683 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
684 		for (i = 0; i < wcl->wcl_count; i++)
685 			witness_leveldescendents(wcl->wcl_children[i], level);
686 }
687 
688 static void
689 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
690 			   struct witness *parent, int indent)
691 {
692 	struct witness_child_list_entry *wcl;
693 	int i, level;
694 
695 	level = parent->w_level;
696 	prnt("%-2d", level);
697 	for (i = 0; i < indent; i++)
698 		prnt(" ");
699 	if (parent->w_refcount > 0)
700 		prnt("%s", parent->w_name);
701 	else
702 		prnt("(dead)");
703 	if (parent->w_displayed) {
704 		prnt(" -- (already displayed)\n");
705 		return;
706 	}
707 	parent->w_displayed = 1;
708 	if (parent->w_refcount > 0) {
709 		if (parent->w_file != NULL)
710 			prnt(" -- last acquired @ %s:%d", parent->w_file,
711 			    parent->w_line);
712 	}
713 	prnt("\n");
714 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
715 		for (i = 0; i < wcl->wcl_count; i++)
716 			    witness_displaydescendants(prnt,
717 				wcl->wcl_children[i], indent + 1);
718 }
719 
720 static void
721 witness_display_list(void(*prnt)(const char *fmt, ...),
722 		     struct witness_list *list)
723 {
724 	struct witness *w;
725 
726 	STAILQ_FOREACH(w, list, w_typelist) {
727 		if (w->w_file == NULL || w->w_level > 0)
728 			continue;
729 		/*
730 		 * This lock has no anscestors, display its descendants.
731 		 */
732 		witness_displaydescendants(prnt, w, 0);
733 	}
734 }
735 
736 static void
737 witness_display(void(*prnt)(const char *fmt, ...))
738 {
739 	struct witness *w;
740 
741 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
742 	witness_levelall();
743 
744 	/* Clear all the displayed flags. */
745 	STAILQ_FOREACH(w, &w_all, w_list) {
746 		w->w_displayed = 0;
747 	}
748 
749 	/*
750 	 * First, handle sleep locks which have been acquired at least
751 	 * once.
752 	 */
753 	prnt("Sleep locks:\n");
754 	witness_display_list(prnt, &w_sleep);
755 
756 	/*
757 	 * Now do spin locks which have been acquired at least once.
758 	 */
759 	prnt("\nSpin locks:\n");
760 	witness_display_list(prnt, &w_spin);
761 
762 	/*
763 	 * Finally, any locks which have not been acquired yet.
764 	 */
765 	prnt("\nLocks which were never acquired:\n");
766 	STAILQ_FOREACH(w, &w_all, w_list) {
767 		if (w->w_file != NULL || w->w_refcount == 0)
768 			continue;
769 		prnt("%s\n", w->w_name);
770 	}
771 }
772 #endif /* DDB */
773 
774 /* Trim useless garbage from filenames. */
775 static const char *
776 fixup_filename(const char *file)
777 {
778 
779 	if (file == NULL)
780 		return (NULL);
781 	while (strncmp(file, "../", 3) == 0)
782 		file += 3;
783 	return (file);
784 }
785 
786 int
787 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
788 {
789 
790 	if (witness_watch == 0 || panicstr != NULL)
791 		return (0);
792 
793 	/* Require locks that witness knows about. */
794 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
795 	    lock2->lo_witness == NULL)
796 		return (EINVAL);
797 
798 	MPASS(!mtx_owned(&w_mtx));
799 	mtx_lock_spin(&w_mtx);
800 
801 	/*
802 	 * If we already have either an explicit or implied lock order that
803 	 * is the other way around, then return an error.
804 	 */
805 	if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
806 		mtx_unlock_spin(&w_mtx);
807 		return (EDOOFUS);
808 	}
809 
810 	/* Try to add the new order. */
811 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
812 	    lock2->lo_type, lock1->lo_type);
813 	if (!itismychild(lock1->lo_witness, lock2->lo_witness))
814 		return (ENOMEM);
815 	mtx_unlock_spin(&w_mtx);
816 	return (0);
817 }
818 
819 void
820 witness_checkorder(struct lock_object *lock, int flags, const char *file,
821     int line)
822 {
823 	struct lock_list_entry **lock_list, *lle;
824 	struct lock_instance *lock1, *lock2;
825 	struct lock_class *class;
826 	struct witness *w, *w1;
827 	struct thread *td;
828 	int i, j;
829 
830 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
831 	    panicstr != NULL)
832 		return;
833 
834 	/*
835 	 * Try locks do not block if they fail to acquire the lock, thus
836 	 * there is no danger of deadlocks or of switching while holding a
837 	 * spin lock if we acquire a lock via a try operation.  This
838 	 * function shouldn't even be called for try locks, so panic if
839 	 * that happens.
840 	 */
841 	if (flags & LOP_TRYLOCK)
842 		panic("%s should not be called for try lock operations",
843 		    __func__);
844 
845 	w = lock->lo_witness;
846 	class = LOCK_CLASS(lock);
847 	td = curthread;
848 	file = fixup_filename(file);
849 
850 	if (class->lc_flags & LC_SLEEPLOCK) {
851 		/*
852 		 * Since spin locks include a critical section, this check
853 		 * implicitly enforces a lock order of all sleep locks before
854 		 * all spin locks.
855 		 */
856 		if (td->td_critnest != 0 && !kdb_active)
857 			panic("blockable sleep lock (%s) %s @ %s:%d",
858 			    class->lc_name, lock->lo_name, file, line);
859 
860 		/*
861 		 * If this is the first lock acquired then just return as
862 		 * no order checking is needed.
863 		 */
864 		if (td->td_sleeplocks == NULL)
865 			return;
866 		lock_list = &td->td_sleeplocks;
867 	} else {
868 		/*
869 		 * If this is the first lock, just return as no order
870 		 * checking is needed.  We check this in both if clauses
871 		 * here as unifying the check would require us to use a
872 		 * critical section to ensure we don't migrate while doing
873 		 * the check.  Note that if this is not the first lock, we
874 		 * are already in a critical section and are safe for the
875 		 * rest of the check.
876 		 */
877 		if (PCPU_GET(spinlocks) == NULL)
878 			return;
879 		lock_list = PCPU_PTR(spinlocks);
880 	}
881 
882 	/*
883 	 * Check to see if we are recursing on a lock we already own.  If
884 	 * so, make sure that we don't mismatch exclusive and shared lock
885 	 * acquires.
886 	 */
887 	lock1 = find_instance(*lock_list, lock);
888 	if (lock1 != NULL) {
889 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
890 		    (flags & LOP_EXCLUSIVE) == 0) {
891 			printf("shared lock of (%s) %s @ %s:%d\n",
892 			    class->lc_name, lock->lo_name, file, line);
893 			printf("while exclusively locked from %s:%d\n",
894 			    lock1->li_file, lock1->li_line);
895 			panic("share->excl");
896 		}
897 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
898 		    (flags & LOP_EXCLUSIVE) != 0) {
899 			printf("exclusive lock of (%s) %s @ %s:%d\n",
900 			    class->lc_name, lock->lo_name, file, line);
901 			printf("while share locked from %s:%d\n",
902 			    lock1->li_file, lock1->li_line);
903 			panic("excl->share");
904 		}
905 		return;
906 	}
907 
908 	/*
909 	 * Try locks do not block if they fail to acquire the lock, thus
910 	 * there is no danger of deadlocks or of switching while holding a
911 	 * spin lock if we acquire a lock via a try operation.
912 	 */
913 	if (flags & LOP_TRYLOCK)
914 		return;
915 
916 	/*
917 	 * Check for duplicate locks of the same type.  Note that we only
918 	 * have to check for this on the last lock we just acquired.  Any
919 	 * other cases will be caught as lock order violations.
920 	 */
921 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
922 	w1 = lock1->li_lock->lo_witness;
923 	if (w1 == w) {
924 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
925 		    (flags & LOP_DUPOK))
926 			return;
927 		w->w_same_squawked = 1;
928 		printf("acquiring duplicate lock of same type: \"%s\"\n",
929 			lock->lo_type);
930 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
931 		    lock1->li_file, lock1->li_line);
932 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
933 #ifdef KDB
934 		goto debugger;
935 #else
936 		return;
937 #endif
938 	}
939 	MPASS(!mtx_owned(&w_mtx));
940 	mtx_lock_spin(&w_mtx);
941 	/*
942 	 * If we know that the the lock we are acquiring comes after
943 	 * the lock we most recently acquired in the lock order tree,
944 	 * then there is no need for any further checks.
945 	 */
946 	if (isitmychild(w1, w)) {
947 		mtx_unlock_spin(&w_mtx);
948 		return;
949 	}
950 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
951 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
952 
953 			MPASS(j < WITNESS_COUNT);
954 			lock1 = &lle->ll_children[i];
955 			w1 = lock1->li_lock->lo_witness;
956 
957 			/*
958 			 * If this lock doesn't undergo witness checking,
959 			 * then skip it.
960 			 */
961 			if (w1 == NULL) {
962 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
963 				    ("lock missing witness structure"));
964 				continue;
965 			}
966 			/*
967 			 * If we are locking Giant and this is a sleepable
968 			 * lock, then skip it.
969 			 */
970 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
971 			    lock == &Giant.lock_object)
972 				continue;
973 			/*
974 			 * If we are locking a sleepable lock and this lock
975 			 * is Giant, then skip it.
976 			 */
977 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
978 			    lock1->li_lock == &Giant.lock_object)
979 				continue;
980 			/*
981 			 * If we are locking a sleepable lock and this lock
982 			 * isn't sleepable, we want to treat it as a lock
983 			 * order violation to enfore a general lock order of
984 			 * sleepable locks before non-sleepable locks.
985 			 */
986 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
987 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
988 				goto reversal;
989 			/*
990 			 * If we are locking Giant and this is a non-sleepable
991 			 * lock, then treat it as a reversal.
992 			 */
993 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
994 			    lock == &Giant.lock_object)
995 				goto reversal;
996 			/*
997 			 * Check the lock order hierarchy for a reveresal.
998 			 */
999 			if (!isitmydescendant(w, w1))
1000 				continue;
1001 		reversal:
1002 			/*
1003 			 * We have a lock order violation, check to see if it
1004 			 * is allowed or has already been yelled about.
1005 			 */
1006 			mtx_unlock_spin(&w_mtx);
1007 #ifdef BLESSING
1008 			/*
1009 			 * If the lock order is blessed, just bail.  We don't
1010 			 * look for other lock order violations though, which
1011 			 * may be a bug.
1012 			 */
1013 			if (blessed(w, w1))
1014 				return;
1015 #endif
1016 			if (lock1->li_lock == &Giant.lock_object) {
1017 				if (w1->w_Giant_squawked)
1018 					return;
1019 				else
1020 					w1->w_Giant_squawked = 1;
1021 			} else {
1022 				if (w1->w_other_squawked)
1023 					return;
1024 				else
1025 					w1->w_other_squawked = 1;
1026 			}
1027 			/*
1028 			 * Ok, yell about it.
1029 			 */
1030 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1031 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1032 				printf(
1033 		"lock order reversal: (sleepable after non-sleepable)\n");
1034 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1035 			    && lock == &Giant.lock_object)
1036 				printf(
1037 		"lock order reversal: (Giant after non-sleepable)\n");
1038 			else
1039 				printf("lock order reversal:\n");
1040 			/*
1041 			 * Try to locate an earlier lock with
1042 			 * witness w in our list.
1043 			 */
1044 			do {
1045 				lock2 = &lle->ll_children[i];
1046 				MPASS(lock2->li_lock != NULL);
1047 				if (lock2->li_lock->lo_witness == w)
1048 					break;
1049 				if (i == 0 && lle->ll_next != NULL) {
1050 					lle = lle->ll_next;
1051 					i = lle->ll_count - 1;
1052 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1053 				} else
1054 					i--;
1055 			} while (i >= 0);
1056 			if (i < 0) {
1057 				printf(" 1st %p %s (%s) @ %s:%d\n",
1058 				    lock1->li_lock, lock1->li_lock->lo_name,
1059 				    lock1->li_lock->lo_type, lock1->li_file,
1060 				    lock1->li_line);
1061 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1062 				    lock->lo_name, lock->lo_type, file, line);
1063 			} else {
1064 				printf(" 1st %p %s (%s) @ %s:%d\n",
1065 				    lock2->li_lock, lock2->li_lock->lo_name,
1066 				    lock2->li_lock->lo_type, lock2->li_file,
1067 				    lock2->li_line);
1068 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1069 				    lock1->li_lock, lock1->li_lock->lo_name,
1070 				    lock1->li_lock->lo_type, lock1->li_file,
1071 				    lock1->li_line);
1072 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1073 				    lock->lo_name, lock->lo_type, file, line);
1074 			}
1075 #ifdef KDB
1076 			goto debugger;
1077 #else
1078 			return;
1079 #endif
1080 		}
1081 	}
1082 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1083 	/*
1084 	 * If requested, build a new lock order.  However, don't build a new
1085 	 * relationship between a sleepable lock and Giant if it is in the
1086 	 * wrong direction.  The correct lock order is that sleepable locks
1087 	 * always come before Giant.
1088 	 */
1089 	if (flags & LOP_NEWORDER &&
1090 	    !(lock1->li_lock == &Giant.lock_object &&
1091 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1092 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1093 		    lock->lo_type, lock1->li_lock->lo_type);
1094 		if (!itismychild(lock1->li_lock->lo_witness, w))
1095 			/* Witness is dead. */
1096 			return;
1097 	}
1098 	mtx_unlock_spin(&w_mtx);
1099 	return;
1100 
1101 #ifdef KDB
1102 debugger:
1103 	if (witness_trace)
1104 		kdb_backtrace();
1105 	if (witness_kdb)
1106 		kdb_enter(__func__);
1107 #endif
1108 }
1109 
1110 void
1111 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1112 {
1113 	struct lock_list_entry **lock_list, *lle;
1114 	struct lock_instance *instance;
1115 	struct witness *w;
1116 	struct thread *td;
1117 
1118 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1119 	    panicstr != NULL)
1120 		return;
1121 	w = lock->lo_witness;
1122 	td = curthread;
1123 	file = fixup_filename(file);
1124 
1125 	/* Determine lock list for this lock. */
1126 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1127 		lock_list = &td->td_sleeplocks;
1128 	else
1129 		lock_list = PCPU_PTR(spinlocks);
1130 
1131 	/* Check to see if we are recursing on a lock we already own. */
1132 	instance = find_instance(*lock_list, lock);
1133 	if (instance != NULL) {
1134 		instance->li_flags++;
1135 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1136 		    td->td_proc->p_pid, lock->lo_name,
1137 		    instance->li_flags & LI_RECURSEMASK);
1138 		instance->li_file = file;
1139 		instance->li_line = line;
1140 		return;
1141 	}
1142 
1143 	/* Update per-witness last file and line acquire. */
1144 	w->w_file = file;
1145 	w->w_line = line;
1146 
1147 	/* Find the next open lock instance in the list and fill it. */
1148 	lle = *lock_list;
1149 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1150 		lle = witness_lock_list_get();
1151 		if (lle == NULL)
1152 			return;
1153 		lle->ll_next = *lock_list;
1154 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1155 		    td->td_proc->p_pid, lle);
1156 		*lock_list = lle;
1157 	}
1158 	instance = &lle->ll_children[lle->ll_count++];
1159 	instance->li_lock = lock;
1160 	instance->li_line = line;
1161 	instance->li_file = file;
1162 	if ((flags & LOP_EXCLUSIVE) != 0)
1163 		instance->li_flags = LI_EXCLUSIVE;
1164 	else
1165 		instance->li_flags = 0;
1166 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1167 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1168 }
1169 
1170 void
1171 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1172 {
1173 	struct lock_instance *instance;
1174 	struct lock_class *class;
1175 
1176 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1177 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1178 		return;
1179 	class = LOCK_CLASS(lock);
1180 	file = fixup_filename(file);
1181 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1182 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1183 		    class->lc_name, lock->lo_name, file, line);
1184 	if ((flags & LOP_TRYLOCK) == 0)
1185 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1186 		    lock->lo_name, file, line);
1187 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1188 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1189 		    class->lc_name, lock->lo_name, file, line);
1190 	instance = find_instance(curthread->td_sleeplocks, lock);
1191 	if (instance == NULL)
1192 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1193 		    class->lc_name, lock->lo_name, file, line);
1194 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1195 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1196 		    class->lc_name, lock->lo_name, file, line);
1197 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1198 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1199 		    class->lc_name, lock->lo_name,
1200 		    instance->li_flags & LI_RECURSEMASK, file, line);
1201 	instance->li_flags |= LI_EXCLUSIVE;
1202 }
1203 
1204 void
1205 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1206     int line)
1207 {
1208 	struct lock_instance *instance;
1209 	struct lock_class *class;
1210 
1211 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1212 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1213 		return;
1214 	class = LOCK_CLASS(lock);
1215 	file = fixup_filename(file);
1216 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1217 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1218 		    class->lc_name, lock->lo_name, file, line);
1219 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1220 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1221 		    class->lc_name, lock->lo_name, file, line);
1222 	instance = find_instance(curthread->td_sleeplocks, lock);
1223 	if (instance == NULL)
1224 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1225 		    class->lc_name, lock->lo_name, file, line);
1226 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1227 		panic("downgrade of shared lock (%s) %s @ %s:%d",
1228 		    class->lc_name, lock->lo_name, file, line);
1229 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1230 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1231 		    class->lc_name, lock->lo_name,
1232 		    instance->li_flags & LI_RECURSEMASK, file, line);
1233 	instance->li_flags &= ~LI_EXCLUSIVE;
1234 }
1235 
1236 void
1237 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1238 {
1239 	struct lock_list_entry **lock_list, *lle;
1240 	struct lock_instance *instance;
1241 	struct lock_class *class;
1242 	struct thread *td;
1243 	register_t s;
1244 	int i, j;
1245 
1246 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1247 	    panicstr != NULL)
1248 		return;
1249 	td = curthread;
1250 	class = LOCK_CLASS(lock);
1251 	file = fixup_filename(file);
1252 
1253 	/* Find lock instance associated with this lock. */
1254 	if (class->lc_flags & LC_SLEEPLOCK)
1255 		lock_list = &td->td_sleeplocks;
1256 	else
1257 		lock_list = PCPU_PTR(spinlocks);
1258 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1259 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1260 			instance = &(*lock_list)->ll_children[i];
1261 			if (instance->li_lock == lock)
1262 				goto found;
1263 		}
1264 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1265 	    file, line);
1266 found:
1267 
1268 	/* First, check for shared/exclusive mismatches. */
1269 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1270 	    (flags & LOP_EXCLUSIVE) == 0) {
1271 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1272 		    lock->lo_name, file, line);
1273 		printf("while exclusively locked from %s:%d\n",
1274 		    instance->li_file, instance->li_line);
1275 		panic("excl->ushare");
1276 	}
1277 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1278 	    (flags & LOP_EXCLUSIVE) != 0) {
1279 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1280 		    lock->lo_name, file, line);
1281 		printf("while share locked from %s:%d\n", instance->li_file,
1282 		    instance->li_line);
1283 		panic("share->uexcl");
1284 	}
1285 
1286 	/* If we are recursed, unrecurse. */
1287 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1288 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1289 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1290 		    instance->li_flags);
1291 		instance->li_flags--;
1292 		return;
1293 	}
1294 
1295 	/* Otherwise, remove this item from the list. */
1296 	s = intr_disable();
1297 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1298 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1299 	    (*lock_list)->ll_count - 1);
1300 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1301 		(*lock_list)->ll_children[j] =
1302 		    (*lock_list)->ll_children[j + 1];
1303 	(*lock_list)->ll_count--;
1304 	intr_restore(s);
1305 
1306 	/* If this lock list entry is now empty, free it. */
1307 	if ((*lock_list)->ll_count == 0) {
1308 		lle = *lock_list;
1309 		*lock_list = lle->ll_next;
1310 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1311 		    td->td_proc->p_pid, lle);
1312 		witness_lock_list_free(lle);
1313 	}
1314 }
1315 
1316 /*
1317  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1318  * exempt Giant and sleepable locks from the checks as well.  If any
1319  * non-exempt locks are held, then a supplied message is printed to the
1320  * console along with a list of the offending locks.  If indicated in the
1321  * flags then a failure results in a panic as well.
1322  */
1323 int
1324 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1325 {
1326 	struct lock_list_entry *lle;
1327 	struct lock_instance *lock1;
1328 	struct thread *td;
1329 	va_list ap;
1330 	int i, n;
1331 
1332 	if (witness_cold || witness_watch == 0 || panicstr != NULL)
1333 		return (0);
1334 	n = 0;
1335 	td = curthread;
1336 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1337 		for (i = lle->ll_count - 1; i >= 0; i--) {
1338 			lock1 = &lle->ll_children[i];
1339 			if (lock1->li_lock == lock)
1340 				continue;
1341 			if (flags & WARN_GIANTOK &&
1342 			    lock1->li_lock == &Giant.lock_object)
1343 				continue;
1344 			if (flags & WARN_SLEEPOK &&
1345 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1346 				continue;
1347 			if (n == 0) {
1348 				va_start(ap, fmt);
1349 				vprintf(fmt, ap);
1350 				va_end(ap);
1351 				printf(" with the following");
1352 				if (flags & WARN_SLEEPOK)
1353 					printf(" non-sleepable");
1354 				printf(" locks held:\n");
1355 			}
1356 			n++;
1357 			witness_list_lock(lock1);
1358 		}
1359 	if (PCPU_GET(spinlocks) != NULL) {
1360 		/*
1361 		 * Since we already hold a spinlock preemption is
1362 		 * already blocked.
1363 		 */
1364 		if (n == 0) {
1365 			va_start(ap, fmt);
1366 			vprintf(fmt, ap);
1367 			va_end(ap);
1368 			printf(" with the following");
1369 			if (flags & WARN_SLEEPOK)
1370 				printf(" non-sleepable");
1371 			printf(" locks held:\n");
1372 		}
1373 		n += witness_list_locks(PCPU_PTR(spinlocks));
1374 	}
1375 	if (flags & WARN_PANIC && n)
1376 		panic("witness_warn");
1377 #ifdef KDB
1378 	else if (witness_kdb && n)
1379 		kdb_enter(__func__);
1380 	else if (witness_trace && n)
1381 		kdb_backtrace();
1382 #endif
1383 	return (n);
1384 }
1385 
1386 const char *
1387 witness_file(struct lock_object *lock)
1388 {
1389 	struct witness *w;
1390 
1391 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1392 		return ("?");
1393 	w = lock->lo_witness;
1394 	return (w->w_file);
1395 }
1396 
1397 int
1398 witness_line(struct lock_object *lock)
1399 {
1400 	struct witness *w;
1401 
1402 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1403 		return (0);
1404 	w = lock->lo_witness;
1405 	return (w->w_line);
1406 }
1407 
1408 static struct witness *
1409 enroll(const char *description, struct lock_class *lock_class)
1410 {
1411 	struct witness *w;
1412 
1413 	if (witness_watch == 0 || panicstr != NULL)
1414 		return (NULL);
1415 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1416 		return (NULL);
1417 	mtx_lock_spin(&w_mtx);
1418 	STAILQ_FOREACH(w, &w_all, w_list) {
1419 		if (w->w_name == description || (w->w_refcount > 0 &&
1420 		    strcmp(description, w->w_name) == 0)) {
1421 			w->w_refcount++;
1422 			mtx_unlock_spin(&w_mtx);
1423 			if (lock_class != w->w_class)
1424 				panic(
1425 				"lock (%s) %s does not match earlier (%s) lock",
1426 				    description, lock_class->lc_name,
1427 				    w->w_class->lc_name);
1428 			return (w);
1429 		}
1430 	}
1431 	if ((w = witness_get()) == NULL)
1432 		goto out;
1433 	w->w_name = description;
1434 	w->w_class = lock_class;
1435 	w->w_refcount = 1;
1436 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1437 	if (lock_class->lc_flags & LC_SPINLOCK) {
1438 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1439 		w_spin_cnt++;
1440 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1441 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1442 		w_sleep_cnt++;
1443 	} else {
1444 		mtx_unlock_spin(&w_mtx);
1445 		panic("lock class %s is not sleep or spin",
1446 		    lock_class->lc_name);
1447 	}
1448 	mtx_unlock_spin(&w_mtx);
1449 out:
1450 	/*
1451 	 * We issue a warning for any spin locks not defined in the static
1452 	 * order list as a way to discourage their use (folks should really
1453 	 * be using non-spin mutexes most of the time).  However, several
1454 	 * 3rd part device drivers use spin locks because that is all they
1455 	 * have available on Windows and Linux and they think that normal
1456 	 * mutexes are insufficient.
1457 	 */
1458 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1459 		printf("WITNESS: spin lock %s not in order list\n",
1460 		    description);
1461 	return (w);
1462 }
1463 
1464 /* Don't let the door bang you on the way out... */
1465 static int
1466 depart(struct witness *w)
1467 {
1468 	struct witness_child_list_entry *wcl, *nwcl;
1469 	struct witness_list *list;
1470 	struct witness *parent;
1471 
1472 	MPASS(w->w_refcount == 0);
1473 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1474 		list = &w_sleep;
1475 		w_sleep_cnt--;
1476 	} else {
1477 		list = &w_spin;
1478 		w_spin_cnt--;
1479 	}
1480 	/*
1481 	 * First, we run through the entire tree looking for any
1482 	 * witnesses that the outgoing witness is a child of.  For
1483 	 * each parent that we find, we reparent all the direct
1484 	 * children of the outgoing witness to its parent.
1485 	 */
1486 	STAILQ_FOREACH(parent, list, w_typelist) {
1487 		if (!isitmychild(parent, w))
1488 			continue;
1489 		removechild(parent, w);
1490 	}
1491 
1492 	/*
1493 	 * Now we go through and free up the child list of the
1494 	 * outgoing witness.
1495 	 */
1496 	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1497 		nwcl = wcl->wcl_next;
1498         	w_child_cnt--;
1499 		witness_child_free(wcl);
1500 	}
1501 
1502 	/*
1503 	 * Detach from various lists and free.
1504 	 */
1505 	STAILQ_REMOVE(list, w, witness, w_typelist);
1506 	STAILQ_REMOVE(&w_all, w, witness, w_list);
1507 	witness_free(w);
1508 
1509 	return (1);
1510 }
1511 
1512 /*
1513  * Add "child" as a direct child of "parent".  Returns false if
1514  * we fail due to out of memory.
1515  */
1516 static int
1517 insertchild(struct witness *parent, struct witness *child)
1518 {
1519 	struct witness_child_list_entry **wcl;
1520 
1521 	MPASS(child != NULL && parent != NULL);
1522 
1523 	/*
1524 	 * Insert "child" after "parent"
1525 	 */
1526 	wcl = &parent->w_children;
1527 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1528 		wcl = &(*wcl)->wcl_next;
1529 	if (*wcl == NULL) {
1530 		*wcl = witness_child_get();
1531 		if (*wcl == NULL)
1532 			return (0);
1533         	w_child_cnt++;
1534 	}
1535 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1536 
1537 	return (1);
1538 }
1539 
1540 
1541 static int
1542 itismychild(struct witness *parent, struct witness *child)
1543 {
1544 	struct witness_list *list;
1545 
1546 	MPASS(child != NULL && parent != NULL);
1547 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1548 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1549 		panic(
1550 		"%s: parent (%s) and child (%s) are not the same lock type",
1551 		    __func__, parent->w_class->lc_name,
1552 		    child->w_class->lc_name);
1553 
1554 	if (!insertchild(parent, child))
1555 		return (0);
1556 
1557 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1558 		list = &w_sleep;
1559 	else
1560 		list = &w_spin;
1561 	return (1);
1562 }
1563 
1564 static void
1565 removechild(struct witness *parent, struct witness *child)
1566 {
1567 	struct witness_child_list_entry **wcl, *wcl1;
1568 	int i;
1569 
1570 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1571 		for (i = 0; i < (*wcl)->wcl_count; i++)
1572 			if ((*wcl)->wcl_children[i] == child)
1573 				goto found;
1574 	return;
1575 found:
1576 	(*wcl)->wcl_count--;
1577 	if ((*wcl)->wcl_count > i)
1578 		(*wcl)->wcl_children[i] =
1579 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1580 	MPASS((*wcl)->wcl_children[i] != NULL);
1581 	if ((*wcl)->wcl_count != 0)
1582 		return;
1583 	wcl1 = *wcl;
1584 	*wcl = wcl1->wcl_next;
1585 	w_child_cnt--;
1586 	witness_child_free(wcl1);
1587 }
1588 
1589 static int
1590 isitmychild(struct witness *parent, struct witness *child)
1591 {
1592 	struct witness_child_list_entry *wcl;
1593 	int i;
1594 
1595 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1596 		for (i = 0; i < wcl->wcl_count; i++) {
1597 			if (wcl->wcl_children[i] == child)
1598 				return (1);
1599 		}
1600 	}
1601 	return (0);
1602 }
1603 
1604 static int
1605 isitmydescendant(struct witness *parent, struct witness *child)
1606 {
1607 	struct witness_child_list_entry *wcl;
1608 	int i, j;
1609 
1610 	if (isitmychild(parent, child))
1611 		return (1);
1612 	j = 0;
1613 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1614 		MPASS(j < 1000);
1615 		for (i = 0; i < wcl->wcl_count; i++) {
1616 			if (isitmydescendant(wcl->wcl_children[i], child))
1617 				return (1);
1618 		}
1619 		j++;
1620 	}
1621 	return (0);
1622 }
1623 
1624 #ifdef BLESSING
1625 static int
1626 blessed(struct witness *w1, struct witness *w2)
1627 {
1628 	int i;
1629 	struct witness_blessed *b;
1630 
1631 	for (i = 0; i < blessed_count; i++) {
1632 		b = &blessed_list[i];
1633 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1634 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1635 				return (1);
1636 			continue;
1637 		}
1638 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1639 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1640 				return (1);
1641 	}
1642 	return (0);
1643 }
1644 #endif
1645 
1646 static struct witness *
1647 witness_get(void)
1648 {
1649 	struct witness *w;
1650 
1651 	if (witness_watch == 0) {
1652 		mtx_unlock_spin(&w_mtx);
1653 		return (NULL);
1654 	}
1655 	if (STAILQ_EMPTY(&w_free)) {
1656 		witness_watch = 0;
1657 		mtx_unlock_spin(&w_mtx);
1658 		printf("%s: witness exhausted\n", __func__);
1659 		return (NULL);
1660 	}
1661 	w = STAILQ_FIRST(&w_free);
1662 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1663 	w_free_cnt--;
1664 	bzero(w, sizeof(*w));
1665 	return (w);
1666 }
1667 
1668 static void
1669 witness_free(struct witness *w)
1670 {
1671 
1672 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1673 	w_free_cnt++;
1674 }
1675 
1676 static struct witness_child_list_entry *
1677 witness_child_get(void)
1678 {
1679 	struct witness_child_list_entry *wcl;
1680 
1681 	if (witness_watch == 0) {
1682 		mtx_unlock_spin(&w_mtx);
1683 		return (NULL);
1684 	}
1685 	wcl = w_child_free;
1686 	if (wcl == NULL) {
1687 		witness_watch = 0;
1688 		mtx_unlock_spin(&w_mtx);
1689 		printf("%s: witness exhausted\n", __func__);
1690 		return (NULL);
1691 	}
1692 	w_child_free = wcl->wcl_next;
1693 	w_child_free_cnt--;
1694 	bzero(wcl, sizeof(*wcl));
1695 	return (wcl);
1696 }
1697 
1698 static void
1699 witness_child_free(struct witness_child_list_entry *wcl)
1700 {
1701 
1702 	wcl->wcl_next = w_child_free;
1703 	w_child_free = wcl;
1704 	w_child_free_cnt++;
1705 }
1706 
1707 static struct lock_list_entry *
1708 witness_lock_list_get(void)
1709 {
1710 	struct lock_list_entry *lle;
1711 
1712 	if (witness_watch == 0)
1713 		return (NULL);
1714 	mtx_lock_spin(&w_mtx);
1715 	lle = w_lock_list_free;
1716 	if (lle == NULL) {
1717 		witness_watch = 0;
1718 		mtx_unlock_spin(&w_mtx);
1719 		printf("%s: witness exhausted\n", __func__);
1720 		return (NULL);
1721 	}
1722 	w_lock_list_free = lle->ll_next;
1723 	mtx_unlock_spin(&w_mtx);
1724 	bzero(lle, sizeof(*lle));
1725 	return (lle);
1726 }
1727 
1728 static void
1729 witness_lock_list_free(struct lock_list_entry *lle)
1730 {
1731 
1732 	mtx_lock_spin(&w_mtx);
1733 	lle->ll_next = w_lock_list_free;
1734 	w_lock_list_free = lle;
1735 	mtx_unlock_spin(&w_mtx);
1736 }
1737 
1738 static struct lock_instance *
1739 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1740 {
1741 	struct lock_list_entry *lle;
1742 	struct lock_instance *instance;
1743 	int i;
1744 
1745 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1746 		for (i = lle->ll_count - 1; i >= 0; i--) {
1747 			instance = &lle->ll_children[i];
1748 			if (instance->li_lock == lock)
1749 				return (instance);
1750 		}
1751 	return (NULL);
1752 }
1753 
1754 static void
1755 witness_list_lock(struct lock_instance *instance)
1756 {
1757 	struct lock_object *lock;
1758 
1759 	lock = instance->li_lock;
1760 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1761 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1762 	if (lock->lo_type != lock->lo_name)
1763 		printf(" (%s)", lock->lo_type);
1764 	printf(" r = %d (%p) locked @ %s:%d\n",
1765 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1766 	    instance->li_line);
1767 }
1768 
1769 #ifdef DDB
1770 static int
1771 witness_thread_has_locks(struct thread *td)
1772 {
1773 
1774 	return (td->td_sleeplocks != NULL);
1775 }
1776 
1777 static int
1778 witness_proc_has_locks(struct proc *p)
1779 {
1780 	struct thread *td;
1781 
1782 	FOREACH_THREAD_IN_PROC(p, td) {
1783 		if (witness_thread_has_locks(td))
1784 			return (1);
1785 	}
1786 	return (0);
1787 }
1788 #endif
1789 
1790 int
1791 witness_list_locks(struct lock_list_entry **lock_list)
1792 {
1793 	struct lock_list_entry *lle;
1794 	int i, nheld;
1795 
1796 	nheld = 0;
1797 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1798 		for (i = lle->ll_count - 1; i >= 0; i--) {
1799 			witness_list_lock(&lle->ll_children[i]);
1800 			nheld++;
1801 		}
1802 	return (nheld);
1803 }
1804 
1805 /*
1806  * This is a bit risky at best.  We call this function when we have timed
1807  * out acquiring a spin lock, and we assume that the other CPU is stuck
1808  * with this lock held.  So, we go groveling around in the other CPU's
1809  * per-cpu data to try to find the lock instance for this spin lock to
1810  * see when it was last acquired.
1811  */
1812 void
1813 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1814 {
1815 	struct lock_instance *instance;
1816 	struct pcpu *pc;
1817 
1818 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1819 		return;
1820 	pc = pcpu_find(owner->td_oncpu);
1821 	instance = find_instance(pc->pc_spinlocks, lock);
1822 	if (instance != NULL)
1823 		witness_list_lock(instance);
1824 }
1825 
1826 void
1827 witness_save(struct lock_object *lock, const char **filep, int *linep)
1828 {
1829 	struct lock_list_entry *lock_list;
1830 	struct lock_instance *instance;
1831 	struct lock_class *class;
1832 
1833 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1834 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1835 		return;
1836 	class = LOCK_CLASS(lock);
1837 	if (class->lc_flags & LC_SLEEPLOCK)
1838 		lock_list = curthread->td_sleeplocks;
1839 	else {
1840 		if (witness_skipspin)
1841 			return;
1842 		lock_list = PCPU_GET(spinlocks);
1843 	}
1844 	instance = find_instance(lock_list, lock);
1845 	if (instance == NULL)
1846 		panic("%s: lock (%s) %s not locked", __func__,
1847 		    class->lc_name, lock->lo_name);
1848 	*filep = instance->li_file;
1849 	*linep = instance->li_line;
1850 }
1851 
1852 void
1853 witness_restore(struct lock_object *lock, const char *file, int line)
1854 {
1855 	struct lock_list_entry *lock_list;
1856 	struct lock_instance *instance;
1857 	struct lock_class *class;
1858 
1859 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1860 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1861 		return;
1862 	class = LOCK_CLASS(lock);
1863 	if (class->lc_flags & LC_SLEEPLOCK)
1864 		lock_list = curthread->td_sleeplocks;
1865 	else {
1866 		if (witness_skipspin)
1867 			return;
1868 		lock_list = PCPU_GET(spinlocks);
1869 	}
1870 	instance = find_instance(lock_list, lock);
1871 	if (instance == NULL)
1872 		panic("%s: lock (%s) %s not locked", __func__,
1873 		    class->lc_name, lock->lo_name);
1874 	lock->lo_witness->w_file = file;
1875 	lock->lo_witness->w_line = line;
1876 	instance->li_file = file;
1877 	instance->li_line = line;
1878 }
1879 
1880 void
1881 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1882 {
1883 #ifdef INVARIANT_SUPPORT
1884 	struct lock_instance *instance;
1885 	struct lock_class *class;
1886 
1887 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1888 		return;
1889 	class = LOCK_CLASS(lock);
1890 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1891 		instance = find_instance(curthread->td_sleeplocks, lock);
1892 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
1893 		instance = find_instance(PCPU_GET(spinlocks), lock);
1894 	else {
1895 		panic("Lock (%s) %s is not sleep or spin!",
1896 		    class->lc_name, lock->lo_name);
1897 	}
1898 	file = fixup_filename(file);
1899 	switch (flags) {
1900 	case LA_UNLOCKED:
1901 		if (instance != NULL)
1902 			panic("Lock (%s) %s locked @ %s:%d.",
1903 			    class->lc_name, lock->lo_name, file, line);
1904 		break;
1905 	case LA_LOCKED:
1906 	case LA_LOCKED | LA_RECURSED:
1907 	case LA_LOCKED | LA_NOTRECURSED:
1908 	case LA_SLOCKED:
1909 	case LA_SLOCKED | LA_RECURSED:
1910 	case LA_SLOCKED | LA_NOTRECURSED:
1911 	case LA_XLOCKED:
1912 	case LA_XLOCKED | LA_RECURSED:
1913 	case LA_XLOCKED | LA_NOTRECURSED:
1914 		if (instance == NULL) {
1915 			panic("Lock (%s) %s not locked @ %s:%d.",
1916 			    class->lc_name, lock->lo_name, file, line);
1917 			break;
1918 		}
1919 		if ((flags & LA_XLOCKED) != 0 &&
1920 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1921 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1922 			    class->lc_name, lock->lo_name, file, line);
1923 		if ((flags & LA_SLOCKED) != 0 &&
1924 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1925 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1926 			    class->lc_name, lock->lo_name, file, line);
1927 		if ((flags & LA_RECURSED) != 0 &&
1928 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1929 			panic("Lock (%s) %s not recursed @ %s:%d.",
1930 			    class->lc_name, lock->lo_name, file, line);
1931 		if ((flags & LA_NOTRECURSED) != 0 &&
1932 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1933 			panic("Lock (%s) %s recursed @ %s:%d.",
1934 			    class->lc_name, lock->lo_name, file, line);
1935 		break;
1936 	default:
1937 		panic("Invalid lock assertion at %s:%d.", file, line);
1938 
1939 	}
1940 #endif	/* INVARIANT_SUPPORT */
1941 }
1942 
1943 #ifdef DDB
1944 static void
1945 witness_list(struct thread *td)
1946 {
1947 
1948 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1949 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1950 
1951 	if (witness_watch == 0)
1952 		return;
1953 
1954 	witness_list_locks(&td->td_sleeplocks);
1955 
1956 	/*
1957 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1958 	 * if td is currently executing on some other CPU and holds spin locks
1959 	 * as we won't display those locks.  If we had a MI way of getting
1960 	 * the per-cpu data for a given cpu then we could use
1961 	 * td->td_oncpu to get the list of spinlocks for this thread
1962 	 * and "fix" this.
1963 	 *
1964 	 * That still wouldn't really fix this unless we locked sched_lock
1965 	 * or stopped the other CPU to make sure it wasn't changing the list
1966 	 * out from under us.  It is probably best to just not try to handle
1967 	 * threads on other CPU's for now.
1968 	 */
1969 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1970 		witness_list_locks(PCPU_PTR(spinlocks));
1971 }
1972 
1973 DB_SHOW_COMMAND(locks, db_witness_list)
1974 {
1975 	struct thread *td;
1976 
1977 	if (have_addr)
1978 		td = db_lookup_thread(addr, TRUE);
1979 	else
1980 		td = kdb_thread;
1981 	witness_list(td);
1982 }
1983 
1984 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1985 {
1986 	struct thread *td;
1987 	struct proc *p;
1988 
1989 	/*
1990 	 * It would be nice to list only threads and processes that actually
1991 	 * held sleep locks, but that information is currently not exported
1992 	 * by WITNESS.
1993 	 */
1994 	FOREACH_PROC_IN_SYSTEM(p) {
1995 		if (!witness_proc_has_locks(p))
1996 			continue;
1997 		FOREACH_THREAD_IN_PROC(p, td) {
1998 			if (!witness_thread_has_locks(td))
1999 				continue;
2000 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2001 			    p->p_comm, td, td->td_tid);
2002 			witness_list(td);
2003 		}
2004 	}
2005 }
2006 
2007 DB_SHOW_COMMAND(witness, db_witness_display)
2008 {
2009 
2010 	witness_display(db_printf);
2011 }
2012 #endif
2013