xref: /freebsd/sys/kern/subr_witness.c (revision 0efd6615cd5f39b67cec82a7034e655f3b5801e3)
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 	 * spin locks
374 	 */
375 #ifdef SMP
376 	{ "ap boot", &lock_class_mtx_spin },
377 #endif
378 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
379 	{ "hptlock", &lock_class_mtx_spin },
380 	{ "sio", &lock_class_mtx_spin },
381 #ifdef __i386__
382 	{ "cy", &lock_class_mtx_spin },
383 #endif
384 	{ "scc_hwmtx", &lock_class_mtx_spin },
385 	{ "uart_hwmtx", &lock_class_mtx_spin },
386 	{ "zstty", &lock_class_mtx_spin },
387 	{ "ng_node", &lock_class_mtx_spin },
388 	{ "ng_worklist", &lock_class_mtx_spin },
389 	{ "fast_taskqueue", &lock_class_mtx_spin },
390 	{ "intr table", &lock_class_mtx_spin },
391 	{ "sleepq chain", &lock_class_mtx_spin },
392 	{ "sched lock", &lock_class_mtx_spin },
393 	{ "turnstile chain", &lock_class_mtx_spin },
394 	{ "td_contested", &lock_class_mtx_spin },
395 	{ "callout", &lock_class_mtx_spin },
396 	{ "entropy harvest mutex", &lock_class_mtx_spin },
397 	{ "syscons video lock", &lock_class_mtx_spin },
398 	/*
399 	 * leaf locks
400 	 */
401 	{ "allpmaps", &lock_class_mtx_spin },
402 	{ "vm page queue free mutex", &lock_class_mtx_spin },
403 	{ "icu", &lock_class_mtx_spin },
404 #ifdef SMP
405 	{ "smp rendezvous", &lock_class_mtx_spin },
406 #if defined(__i386__) || defined(__amd64__)
407 	{ "tlb", &lock_class_mtx_spin },
408 #endif
409 #ifdef __sparc64__
410 	{ "ipi", &lock_class_mtx_spin },
411 	{ "rtc_mtx", &lock_class_mtx_spin },
412 #endif
413 #endif
414 	{ "clk", &lock_class_mtx_spin },
415 	{ "mutex profiling lock", &lock_class_mtx_spin },
416 	{ "kse zombie lock", &lock_class_mtx_spin },
417 	{ "ALD Queue", &lock_class_mtx_spin },
418 #ifdef __ia64__
419 	{ "MCA spin lock", &lock_class_mtx_spin },
420 #endif
421 #if defined(__i386__) || defined(__amd64__)
422 	{ "pcicfg", &lock_class_mtx_spin },
423 	{ "NDIS thread lock", &lock_class_mtx_spin },
424 #endif
425 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
426 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
427 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
428 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
429 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
430 	{ NULL, NULL },
431 	{ NULL, NULL }
432 };
433 
434 #ifdef BLESSING
435 /*
436  * Pairs of locks which have been blessed
437  * Don't complain about order problems with blessed locks
438  */
439 static struct witness_blessed blessed_list[] = {
440 };
441 static int blessed_count =
442 	sizeof(blessed_list) / sizeof(struct witness_blessed);
443 #endif
444 
445 /*
446  * List of locks initialized prior to witness being initialized whose
447  * enrollment is currently deferred.
448  */
449 STAILQ_HEAD(, lock_object) pending_locks =
450     STAILQ_HEAD_INITIALIZER(pending_locks);
451 
452 /*
453  * This global is set to 0 once it becomes safe to use the witness code.
454  */
455 static int witness_cold = 1;
456 
457 /*
458  * This global is set to 1 once the static lock orders have been enrolled
459  * so that a warning can be issued for any spin locks enrolled later.
460  */
461 static int witness_spin_warn = 0;
462 
463 /*
464  * The WITNESS-enabled diagnostic code.  Note that the witness code does
465  * assume that the early boot is single-threaded at least until after this
466  * routine is completed.
467  */
468 static void
469 witness_initialize(void *dummy __unused)
470 {
471 	struct lock_object *lock;
472 	struct witness_order_list_entry *order;
473 	struct witness *w, *w1;
474 	int i;
475 
476 	/*
477 	 * We have to release Giant before initializing its witness
478 	 * structure so that WITNESS doesn't get confused.
479 	 */
480 	mtx_unlock(&Giant);
481 	mtx_assert(&Giant, MA_NOTOWNED);
482 
483 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
484 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
485 	    MTX_NOWITNESS | MTX_NOPROFILE);
486 	for (i = 0; i < WITNESS_COUNT; i++)
487 		witness_free(&w_data[i]);
488 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
489 		witness_child_free(&w_childdata[i]);
490 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
491 		witness_lock_list_free(&w_locklistdata[i]);
492 
493 	/* First add in all the specified order lists. */
494 	for (order = order_lists; order->w_name != NULL; order++) {
495 		w = enroll(order->w_name, order->w_class);
496 		if (w == NULL)
497 			continue;
498 		w->w_file = "order list";
499 		for (order++; order->w_name != NULL; order++) {
500 			w1 = enroll(order->w_name, order->w_class);
501 			if (w1 == NULL)
502 				continue;
503 			w1->w_file = "order list";
504 			if (!itismychild(w, w1))
505 				panic("Not enough memory for static orders!");
506 			w = w1;
507 		}
508 	}
509 	witness_spin_warn = 1;
510 
511 	/* Iterate through all locks and add them to witness. */
512 	while (!STAILQ_EMPTY(&pending_locks)) {
513 		lock = STAILQ_FIRST(&pending_locks);
514 		STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
515 		KASSERT(lock->lo_flags & LO_WITNESS,
516 		    ("%s: lock %s is on pending list but not LO_WITNESS",
517 		    __func__, lock->lo_name));
518 		lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
519 	}
520 
521 	/* Mark the witness code as being ready for use. */
522 	witness_cold = 0;
523 
524 	mtx_lock(&Giant);
525 }
526 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
527 
528 static int
529 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
530 {
531 	int error, value;
532 
533 	value = witness_watch;
534 	error = sysctl_handle_int(oidp, &value, 0, req);
535 	if (error != 0 || req->newptr == NULL)
536 		return (error);
537 	/*
538 	 * XXXRW: Why a priv check here?
539 	 */
540 	error = priv_check(req->td, PRIV_WITNESS);
541 	if (error != 0)
542 		return (error);
543 	if (value == witness_watch)
544 		return (0);
545 	if (value != 0)
546 		return (EINVAL);
547 	witness_watch = 0;
548 	return (0);
549 }
550 
551 void
552 witness_init(struct lock_object *lock)
553 {
554 	struct lock_class *class;
555 
556 	/* Various sanity checks. */
557 	class = LOCK_CLASS(lock);
558 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
559 	    (class->lc_flags & LC_RECURSABLE) == 0)
560 		panic("%s: lock (%s) %s can not be recursable", __func__,
561 		    class->lc_name, lock->lo_name);
562 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
563 	    (class->lc_flags & LC_SLEEPABLE) == 0)
564 		panic("%s: lock (%s) %s can not be sleepable", __func__,
565 		    class->lc_name, lock->lo_name);
566 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
567 	    (class->lc_flags & LC_UPGRADABLE) == 0)
568 		panic("%s: lock (%s) %s can not be upgradable", __func__,
569 		    class->lc_name, lock->lo_name);
570 
571 	/*
572 	 * If we shouldn't watch this lock, then just clear lo_witness.
573 	 * Otherwise, if witness_cold is set, then it is too early to
574 	 * enroll this lock, so defer it to witness_initialize() by adding
575 	 * it to the pending_locks list.  If it is not too early, then enroll
576 	 * the lock now.
577 	 */
578 	if (witness_watch == 0 || panicstr != NULL ||
579 	    (lock->lo_flags & LO_WITNESS) == 0)
580 		lock->lo_witness = NULL;
581 	else if (witness_cold) {
582 		STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
583 		lock->lo_flags |= LO_ENROLLPEND;
584 	} else
585 		lock->lo_witness = enroll(lock->lo_type, class);
586 }
587 
588 void
589 witness_destroy(struct lock_object *lock)
590 {
591 	struct lock_class *class;
592 	struct witness *w;
593 
594 	class = LOCK_CLASS(lock);
595 	if (witness_cold)
596 		panic("lock (%s) %s destroyed while witness_cold",
597 		    class->lc_name, lock->lo_name);
598 
599 	/* XXX: need to verify that no one holds the lock */
600 	if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
601 	    lock->lo_witness != NULL) {
602 		w = lock->lo_witness;
603 		mtx_lock_spin(&w_mtx);
604 		MPASS(w->w_refcount > 0);
605 		w->w_refcount--;
606 
607 		/*
608 		 * Lock is already released if we have an allocation failure
609 		 * and depart() fails.
610 		 */
611 		if (w->w_refcount != 0 || depart(w))
612 			mtx_unlock_spin(&w_mtx);
613 	}
614 
615 	/*
616 	 * If this lock is destroyed before witness is up and running,
617 	 * remove it from the pending list.
618 	 */
619 	if (lock->lo_flags & LO_ENROLLPEND) {
620 		STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
621 		lock->lo_flags &= ~LO_ENROLLPEND;
622 	}
623 }
624 
625 #ifdef DDB
626 static void
627 witness_levelall (void)
628 {
629 	struct witness_list *list;
630 	struct witness *w, *w1;
631 
632 	/*
633 	 * First clear all levels.
634 	 */
635 	STAILQ_FOREACH(w, &w_all, w_list) {
636 		w->w_level = 0;
637 	}
638 
639 	/*
640 	 * Look for locks with no parent and level all their descendants.
641 	 */
642 	STAILQ_FOREACH(w, &w_all, w_list) {
643 		/*
644 		 * This is just an optimization, technically we could get
645 		 * away just walking the all list each time.
646 		 */
647 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
648 			list = &w_sleep;
649 		else
650 			list = &w_spin;
651 		STAILQ_FOREACH(w1, list, w_typelist) {
652 			if (isitmychild(w1, w))
653 				goto skip;
654 		}
655 		witness_leveldescendents(w, 0);
656 	skip:
657 		;	/* silence GCC 3.x */
658 	}
659 }
660 
661 static void
662 witness_leveldescendents(struct witness *parent, int level)
663 {
664 	struct witness_child_list_entry *wcl;
665 	int i;
666 
667 	if (parent->w_level < level)
668 		parent->w_level = level;
669 	level++;
670 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
671 		for (i = 0; i < wcl->wcl_count; i++)
672 			witness_leveldescendents(wcl->wcl_children[i], level);
673 }
674 
675 static void
676 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
677 			   struct witness *parent, int indent)
678 {
679 	struct witness_child_list_entry *wcl;
680 	int i, level;
681 
682 	level = parent->w_level;
683 	prnt("%-2d", level);
684 	for (i = 0; i < indent; i++)
685 		prnt(" ");
686 	if (parent->w_refcount > 0)
687 		prnt("%s", parent->w_name);
688 	else
689 		prnt("(dead)");
690 	if (parent->w_displayed) {
691 		prnt(" -- (already displayed)\n");
692 		return;
693 	}
694 	parent->w_displayed = 1;
695 	if (parent->w_refcount > 0) {
696 		if (parent->w_file != NULL)
697 			prnt(" -- last acquired @ %s:%d", parent->w_file,
698 			    parent->w_line);
699 	}
700 	prnt("\n");
701 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
702 		for (i = 0; i < wcl->wcl_count; i++)
703 			    witness_displaydescendants(prnt,
704 				wcl->wcl_children[i], indent + 1);
705 }
706 
707 static void
708 witness_display_list(void(*prnt)(const char *fmt, ...),
709 		     struct witness_list *list)
710 {
711 	struct witness *w;
712 
713 	STAILQ_FOREACH(w, list, w_typelist) {
714 		if (w->w_file == NULL || w->w_level > 0)
715 			continue;
716 		/*
717 		 * This lock has no anscestors, display its descendants.
718 		 */
719 		witness_displaydescendants(prnt, w, 0);
720 	}
721 }
722 
723 static void
724 witness_display(void(*prnt)(const char *fmt, ...))
725 {
726 	struct witness *w;
727 
728 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
729 	witness_levelall();
730 
731 	/* Clear all the displayed flags. */
732 	STAILQ_FOREACH(w, &w_all, w_list) {
733 		w->w_displayed = 0;
734 	}
735 
736 	/*
737 	 * First, handle sleep locks which have been acquired at least
738 	 * once.
739 	 */
740 	prnt("Sleep locks:\n");
741 	witness_display_list(prnt, &w_sleep);
742 
743 	/*
744 	 * Now do spin locks which have been acquired at least once.
745 	 */
746 	prnt("\nSpin locks:\n");
747 	witness_display_list(prnt, &w_spin);
748 
749 	/*
750 	 * Finally, any locks which have not been acquired yet.
751 	 */
752 	prnt("\nLocks which were never acquired:\n");
753 	STAILQ_FOREACH(w, &w_all, w_list) {
754 		if (w->w_file != NULL || w->w_refcount == 0)
755 			continue;
756 		prnt("%s\n", w->w_name);
757 	}
758 }
759 #endif /* DDB */
760 
761 /* Trim useless garbage from filenames. */
762 static const char *
763 fixup_filename(const char *file)
764 {
765 
766 	if (file == NULL)
767 		return (NULL);
768 	while (strncmp(file, "../", 3) == 0)
769 		file += 3;
770 	return (file);
771 }
772 
773 int
774 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
775 {
776 
777 	if (witness_watch == 0 || panicstr != NULL)
778 		return (0);
779 
780 	/* Require locks that witness knows about. */
781 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
782 	    lock2->lo_witness == NULL)
783 		return (EINVAL);
784 
785 	MPASS(!mtx_owned(&w_mtx));
786 	mtx_lock_spin(&w_mtx);
787 
788 	/*
789 	 * If we already have either an explicit or implied lock order that
790 	 * is the other way around, then return an error.
791 	 */
792 	if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
793 		mtx_unlock_spin(&w_mtx);
794 		return (EDOOFUS);
795 	}
796 
797 	/* Try to add the new order. */
798 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
799 	    lock2->lo_type, lock1->lo_type);
800 	if (!itismychild(lock1->lo_witness, lock2->lo_witness))
801 		return (ENOMEM);
802 	mtx_unlock_spin(&w_mtx);
803 	return (0);
804 }
805 
806 void
807 witness_checkorder(struct lock_object *lock, int flags, const char *file,
808     int line)
809 {
810 	struct lock_list_entry **lock_list, *lle;
811 	struct lock_instance *lock1, *lock2;
812 	struct lock_class *class;
813 	struct witness *w, *w1;
814 	struct thread *td;
815 	int i, j;
816 
817 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
818 	    panicstr != NULL)
819 		return;
820 
821 	/*
822 	 * Try locks do not block if they fail to acquire the lock, thus
823 	 * there is no danger of deadlocks or of switching while holding a
824 	 * spin lock if we acquire a lock via a try operation.  This
825 	 * function shouldn't even be called for try locks, so panic if
826 	 * that happens.
827 	 */
828 	if (flags & LOP_TRYLOCK)
829 		panic("%s should not be called for try lock operations",
830 		    __func__);
831 
832 	w = lock->lo_witness;
833 	class = LOCK_CLASS(lock);
834 	td = curthread;
835 	file = fixup_filename(file);
836 
837 	if (class->lc_flags & LC_SLEEPLOCK) {
838 		/*
839 		 * Since spin locks include a critical section, this check
840 		 * implicitly enforces a lock order of all sleep locks before
841 		 * all spin locks.
842 		 */
843 		if (td->td_critnest != 0 && !kdb_active)
844 			panic("blockable sleep lock (%s) %s @ %s:%d",
845 			    class->lc_name, lock->lo_name, file, line);
846 
847 		/*
848 		 * If this is the first lock acquired then just return as
849 		 * no order checking is needed.
850 		 */
851 		if (td->td_sleeplocks == NULL)
852 			return;
853 		lock_list = &td->td_sleeplocks;
854 	} else {
855 		/*
856 		 * If this is the first lock, just return as no order
857 		 * checking is needed.  We check this in both if clauses
858 		 * here as unifying the check would require us to use a
859 		 * critical section to ensure we don't migrate while doing
860 		 * the check.  Note that if this is not the first lock, we
861 		 * are already in a critical section and are safe for the
862 		 * rest of the check.
863 		 */
864 		if (PCPU_GET(spinlocks) == NULL)
865 			return;
866 		lock_list = PCPU_PTR(spinlocks);
867 	}
868 
869 	/*
870 	 * Check to see if we are recursing on a lock we already own.  If
871 	 * so, make sure that we don't mismatch exclusive and shared lock
872 	 * acquires.
873 	 */
874 	lock1 = find_instance(*lock_list, lock);
875 	if (lock1 != NULL) {
876 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
877 		    (flags & LOP_EXCLUSIVE) == 0) {
878 			printf("shared lock of (%s) %s @ %s:%d\n",
879 			    class->lc_name, lock->lo_name, file, line);
880 			printf("while exclusively locked from %s:%d\n",
881 			    lock1->li_file, lock1->li_line);
882 			panic("share->excl");
883 		}
884 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
885 		    (flags & LOP_EXCLUSIVE) != 0) {
886 			printf("exclusive lock of (%s) %s @ %s:%d\n",
887 			    class->lc_name, lock->lo_name, file, line);
888 			printf("while share locked from %s:%d\n",
889 			    lock1->li_file, lock1->li_line);
890 			panic("excl->share");
891 		}
892 		return;
893 	}
894 
895 	/*
896 	 * Try locks do not block if they fail to acquire the lock, thus
897 	 * there is no danger of deadlocks or of switching while holding a
898 	 * spin lock if we acquire a lock via a try operation.
899 	 */
900 	if (flags & LOP_TRYLOCK)
901 		return;
902 
903 	/*
904 	 * Check for duplicate locks of the same type.  Note that we only
905 	 * have to check for this on the last lock we just acquired.  Any
906 	 * other cases will be caught as lock order violations.
907 	 */
908 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
909 	w1 = lock1->li_lock->lo_witness;
910 	if (w1 == w) {
911 		if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
912 		    (flags & LOP_DUPOK))
913 			return;
914 		w->w_same_squawked = 1;
915 		printf("acquiring duplicate lock of same type: \"%s\"\n",
916 			lock->lo_type);
917 		printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
918 		    lock1->li_file, lock1->li_line);
919 		printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
920 #ifdef KDB
921 		goto debugger;
922 #else
923 		return;
924 #endif
925 	}
926 	MPASS(!mtx_owned(&w_mtx));
927 	mtx_lock_spin(&w_mtx);
928 	/*
929 	 * If we know that the the lock we are acquiring comes after
930 	 * the lock we most recently acquired in the lock order tree,
931 	 * then there is no need for any further checks.
932 	 */
933 	if (isitmychild(w1, w)) {
934 		mtx_unlock_spin(&w_mtx);
935 		return;
936 	}
937 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
938 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
939 
940 			MPASS(j < WITNESS_COUNT);
941 			lock1 = &lle->ll_children[i];
942 			w1 = lock1->li_lock->lo_witness;
943 
944 			/*
945 			 * If this lock doesn't undergo witness checking,
946 			 * then skip it.
947 			 */
948 			if (w1 == NULL) {
949 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
950 				    ("lock missing witness structure"));
951 				continue;
952 			}
953 			/*
954 			 * If we are locking Giant and this is a sleepable
955 			 * lock, then skip it.
956 			 */
957 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
958 			    lock == &Giant.mtx_object)
959 				continue;
960 			/*
961 			 * If we are locking a sleepable lock and this lock
962 			 * is Giant, then skip it.
963 			 */
964 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
965 			    lock1->li_lock == &Giant.mtx_object)
966 				continue;
967 			/*
968 			 * If we are locking a sleepable lock and this lock
969 			 * isn't sleepable, we want to treat it as a lock
970 			 * order violation to enfore a general lock order of
971 			 * sleepable locks before non-sleepable locks.
972 			 */
973 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
974 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
975 				goto reversal;
976 			/*
977 			 * If we are locking Giant and this is a non-sleepable
978 			 * lock, then treat it as a reversal.
979 			 */
980 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
981 			    lock == &Giant.mtx_object)
982 				goto reversal;
983 			/*
984 			 * Check the lock order hierarchy for a reveresal.
985 			 */
986 			if (!isitmydescendant(w, w1))
987 				continue;
988 		reversal:
989 			/*
990 			 * We have a lock order violation, check to see if it
991 			 * is allowed or has already been yelled about.
992 			 */
993 			mtx_unlock_spin(&w_mtx);
994 #ifdef BLESSING
995 			/*
996 			 * If the lock order is blessed, just bail.  We don't
997 			 * look for other lock order violations though, which
998 			 * may be a bug.
999 			 */
1000 			if (blessed(w, w1))
1001 				return;
1002 #endif
1003 			if (lock1->li_lock == &Giant.mtx_object) {
1004 				if (w1->w_Giant_squawked)
1005 					return;
1006 				else
1007 					w1->w_Giant_squawked = 1;
1008 			} else {
1009 				if (w1->w_other_squawked)
1010 					return;
1011 				else
1012 					w1->w_other_squawked = 1;
1013 			}
1014 			/*
1015 			 * Ok, yell about it.
1016 			 */
1017 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1018 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1019 				printf(
1020 		"lock order reversal: (sleepable after non-sleepable)\n");
1021 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1022 			    && lock == &Giant.mtx_object)
1023 				printf(
1024 		"lock order reversal: (Giant after non-sleepable)\n");
1025 			else
1026 				printf("lock order reversal:\n");
1027 			/*
1028 			 * Try to locate an earlier lock with
1029 			 * witness w in our list.
1030 			 */
1031 			do {
1032 				lock2 = &lle->ll_children[i];
1033 				MPASS(lock2->li_lock != NULL);
1034 				if (lock2->li_lock->lo_witness == w)
1035 					break;
1036 				if (i == 0 && lle->ll_next != NULL) {
1037 					lle = lle->ll_next;
1038 					i = lle->ll_count - 1;
1039 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1040 				} else
1041 					i--;
1042 			} while (i >= 0);
1043 			if (i < 0) {
1044 				printf(" 1st %p %s (%s) @ %s:%d\n",
1045 				    lock1->li_lock, lock1->li_lock->lo_name,
1046 				    lock1->li_lock->lo_type, lock1->li_file,
1047 				    lock1->li_line);
1048 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1049 				    lock->lo_name, lock->lo_type, file, line);
1050 			} else {
1051 				printf(" 1st %p %s (%s) @ %s:%d\n",
1052 				    lock2->li_lock, lock2->li_lock->lo_name,
1053 				    lock2->li_lock->lo_type, lock2->li_file,
1054 				    lock2->li_line);
1055 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1056 				    lock1->li_lock, lock1->li_lock->lo_name,
1057 				    lock1->li_lock->lo_type, lock1->li_file,
1058 				    lock1->li_line);
1059 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1060 				    lock->lo_name, lock->lo_type, file, line);
1061 			}
1062 #ifdef KDB
1063 			goto debugger;
1064 #else
1065 			return;
1066 #endif
1067 		}
1068 	}
1069 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1070 	/*
1071 	 * If requested, build a new lock order.  However, don't build a new
1072 	 * relationship between a sleepable lock and Giant if it is in the
1073 	 * wrong direction.  The correct lock order is that sleepable locks
1074 	 * always come before Giant.
1075 	 */
1076 	if (flags & LOP_NEWORDER &&
1077 	    !(lock1->li_lock == &Giant.mtx_object &&
1078 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1079 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1080 		    lock->lo_type, lock1->li_lock->lo_type);
1081 		if (!itismychild(lock1->li_lock->lo_witness, w))
1082 			/* Witness is dead. */
1083 			return;
1084 	}
1085 	mtx_unlock_spin(&w_mtx);
1086 	return;
1087 
1088 #ifdef KDB
1089 debugger:
1090 	if (witness_trace)
1091 		kdb_backtrace();
1092 	if (witness_kdb)
1093 		kdb_enter(__func__);
1094 #endif
1095 }
1096 
1097 void
1098 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1099 {
1100 	struct lock_list_entry **lock_list, *lle;
1101 	struct lock_instance *instance;
1102 	struct witness *w;
1103 	struct thread *td;
1104 
1105 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1106 	    panicstr != NULL)
1107 		return;
1108 	w = lock->lo_witness;
1109 	td = curthread;
1110 	file = fixup_filename(file);
1111 
1112 	/* Determine lock list for this lock. */
1113 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1114 		lock_list = &td->td_sleeplocks;
1115 	else
1116 		lock_list = PCPU_PTR(spinlocks);
1117 
1118 	/* Check to see if we are recursing on a lock we already own. */
1119 	instance = find_instance(*lock_list, lock);
1120 	if (instance != NULL) {
1121 		instance->li_flags++;
1122 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1123 		    td->td_proc->p_pid, lock->lo_name,
1124 		    instance->li_flags & LI_RECURSEMASK);
1125 		instance->li_file = file;
1126 		instance->li_line = line;
1127 		return;
1128 	}
1129 
1130 	/* Update per-witness last file and line acquire. */
1131 	w->w_file = file;
1132 	w->w_line = line;
1133 
1134 	/* Find the next open lock instance in the list and fill it. */
1135 	lle = *lock_list;
1136 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1137 		lle = witness_lock_list_get();
1138 		if (lle == NULL)
1139 			return;
1140 		lle->ll_next = *lock_list;
1141 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1142 		    td->td_proc->p_pid, lle);
1143 		*lock_list = lle;
1144 	}
1145 	instance = &lle->ll_children[lle->ll_count++];
1146 	instance->li_lock = lock;
1147 	instance->li_line = line;
1148 	instance->li_file = file;
1149 	if ((flags & LOP_EXCLUSIVE) != 0)
1150 		instance->li_flags = LI_EXCLUSIVE;
1151 	else
1152 		instance->li_flags = 0;
1153 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1154 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1155 }
1156 
1157 void
1158 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1159 {
1160 	struct lock_instance *instance;
1161 	struct lock_class *class;
1162 
1163 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1164 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1165 		return;
1166 	class = LOCK_CLASS(lock);
1167 	file = fixup_filename(file);
1168 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1169 		panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1170 		    class->lc_name, lock->lo_name, file, line);
1171 	if ((flags & LOP_TRYLOCK) == 0)
1172 		panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1173 		    lock->lo_name, file, line);
1174 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1175 		panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1176 		    class->lc_name, lock->lo_name, file, line);
1177 	instance = find_instance(curthread->td_sleeplocks, lock);
1178 	if (instance == NULL)
1179 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1180 		    class->lc_name, lock->lo_name, file, line);
1181 	if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1182 		panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1183 		    class->lc_name, lock->lo_name, file, line);
1184 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1185 		panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1186 		    class->lc_name, lock->lo_name,
1187 		    instance->li_flags & LI_RECURSEMASK, file, line);
1188 	instance->li_flags |= LI_EXCLUSIVE;
1189 }
1190 
1191 void
1192 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1193     int line)
1194 {
1195 	struct lock_instance *instance;
1196 	struct lock_class *class;
1197 
1198 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1199 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1200 		return;
1201 	class = LOCK_CLASS(lock);
1202 	file = fixup_filename(file);
1203 	if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1204 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1205 		    class->lc_name, lock->lo_name, file, line);
1206 	if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1207 		panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1208 		    class->lc_name, lock->lo_name, file, line);
1209 	instance = find_instance(curthread->td_sleeplocks, lock);
1210 	if (instance == NULL)
1211 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1212 		    class->lc_name, lock->lo_name, file, line);
1213 	if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1214 		panic("downgrade of shared lock (%s) %s @ %s:%d",
1215 		    class->lc_name, lock->lo_name, file, line);
1216 	if ((instance->li_flags & LI_RECURSEMASK) != 0)
1217 		panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1218 		    class->lc_name, lock->lo_name,
1219 		    instance->li_flags & LI_RECURSEMASK, file, line);
1220 	instance->li_flags &= ~LI_EXCLUSIVE;
1221 }
1222 
1223 void
1224 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1225 {
1226 	struct lock_list_entry **lock_list, *lle;
1227 	struct lock_instance *instance;
1228 	struct lock_class *class;
1229 	struct thread *td;
1230 	register_t s;
1231 	int i, j;
1232 
1233 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1234 	    panicstr != NULL)
1235 		return;
1236 	td = curthread;
1237 	class = LOCK_CLASS(lock);
1238 	file = fixup_filename(file);
1239 
1240 	/* Find lock instance associated with this lock. */
1241 	if (class->lc_flags & LC_SLEEPLOCK)
1242 		lock_list = &td->td_sleeplocks;
1243 	else
1244 		lock_list = PCPU_PTR(spinlocks);
1245 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1246 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1247 			instance = &(*lock_list)->ll_children[i];
1248 			if (instance->li_lock == lock)
1249 				goto found;
1250 		}
1251 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1252 	    file, line);
1253 found:
1254 
1255 	/* First, check for shared/exclusive mismatches. */
1256 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1257 	    (flags & LOP_EXCLUSIVE) == 0) {
1258 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1259 		    lock->lo_name, file, line);
1260 		printf("while exclusively locked from %s:%d\n",
1261 		    instance->li_file, instance->li_line);
1262 		panic("excl->ushare");
1263 	}
1264 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1265 	    (flags & LOP_EXCLUSIVE) != 0) {
1266 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1267 		    lock->lo_name, file, line);
1268 		printf("while share locked from %s:%d\n", instance->li_file,
1269 		    instance->li_line);
1270 		panic("share->uexcl");
1271 	}
1272 
1273 	/* If we are recursed, unrecurse. */
1274 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1275 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1276 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1277 		    instance->li_flags);
1278 		instance->li_flags--;
1279 		return;
1280 	}
1281 
1282 	/* Otherwise, remove this item from the list. */
1283 	s = intr_disable();
1284 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1285 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1286 	    (*lock_list)->ll_count - 1);
1287 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1288 		(*lock_list)->ll_children[j] =
1289 		    (*lock_list)->ll_children[j + 1];
1290 	(*lock_list)->ll_count--;
1291 	intr_restore(s);
1292 
1293 	/* If this lock list entry is now empty, free it. */
1294 	if ((*lock_list)->ll_count == 0) {
1295 		lle = *lock_list;
1296 		*lock_list = lle->ll_next;
1297 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1298 		    td->td_proc->p_pid, lle);
1299 		witness_lock_list_free(lle);
1300 	}
1301 }
1302 
1303 /*
1304  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1305  * exempt Giant and sleepable locks from the checks as well.  If any
1306  * non-exempt locks are held, then a supplied message is printed to the
1307  * console along with a list of the offending locks.  If indicated in the
1308  * flags then a failure results in a panic as well.
1309  */
1310 int
1311 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1312 {
1313 	struct lock_list_entry *lle;
1314 	struct lock_instance *lock1;
1315 	struct thread *td;
1316 	va_list ap;
1317 	int i, n;
1318 
1319 	if (witness_cold || witness_watch == 0 || panicstr != NULL)
1320 		return (0);
1321 	n = 0;
1322 	td = curthread;
1323 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1324 		for (i = lle->ll_count - 1; i >= 0; i--) {
1325 			lock1 = &lle->ll_children[i];
1326 			if (lock1->li_lock == lock)
1327 				continue;
1328 			if (flags & WARN_GIANTOK &&
1329 			    lock1->li_lock == &Giant.mtx_object)
1330 				continue;
1331 			if (flags & WARN_SLEEPOK &&
1332 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1333 				continue;
1334 			if (n == 0) {
1335 				va_start(ap, fmt);
1336 				vprintf(fmt, ap);
1337 				va_end(ap);
1338 				printf(" with the following");
1339 				if (flags & WARN_SLEEPOK)
1340 					printf(" non-sleepable");
1341 				printf(" locks held:\n");
1342 			}
1343 			n++;
1344 			witness_list_lock(lock1);
1345 		}
1346 	if (PCPU_GET(spinlocks) != NULL) {
1347 		/*
1348 		 * Since we already hold a spinlock preemption is
1349 		 * already blocked.
1350 		 */
1351 		if (n == 0) {
1352 			va_start(ap, fmt);
1353 			vprintf(fmt, ap);
1354 			va_end(ap);
1355 			printf(" with the following");
1356 			if (flags & WARN_SLEEPOK)
1357 				printf(" non-sleepable");
1358 			printf(" locks held:\n");
1359 		}
1360 		n += witness_list_locks(PCPU_PTR(spinlocks));
1361 	}
1362 	if (flags & WARN_PANIC && n)
1363 		panic("witness_warn");
1364 #ifdef KDB
1365 	else if (witness_kdb && n)
1366 		kdb_enter(__func__);
1367 	else if (witness_trace && n)
1368 		kdb_backtrace();
1369 #endif
1370 	return (n);
1371 }
1372 
1373 const char *
1374 witness_file(struct lock_object *lock)
1375 {
1376 	struct witness *w;
1377 
1378 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1379 		return ("?");
1380 	w = lock->lo_witness;
1381 	return (w->w_file);
1382 }
1383 
1384 int
1385 witness_line(struct lock_object *lock)
1386 {
1387 	struct witness *w;
1388 
1389 	if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1390 		return (0);
1391 	w = lock->lo_witness;
1392 	return (w->w_line);
1393 }
1394 
1395 static struct witness *
1396 enroll(const char *description, struct lock_class *lock_class)
1397 {
1398 	struct witness *w;
1399 
1400 	if (witness_watch == 0 || panicstr != NULL)
1401 		return (NULL);
1402 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1403 		return (NULL);
1404 	mtx_lock_spin(&w_mtx);
1405 	STAILQ_FOREACH(w, &w_all, w_list) {
1406 		if (w->w_name == description || (w->w_refcount > 0 &&
1407 		    strcmp(description, w->w_name) == 0)) {
1408 			w->w_refcount++;
1409 			mtx_unlock_spin(&w_mtx);
1410 			if (lock_class != w->w_class)
1411 				panic(
1412 				"lock (%s) %s does not match earlier (%s) lock",
1413 				    description, lock_class->lc_name,
1414 				    w->w_class->lc_name);
1415 			return (w);
1416 		}
1417 	}
1418 	if ((w = witness_get()) == NULL)
1419 		goto out;
1420 	w->w_name = description;
1421 	w->w_class = lock_class;
1422 	w->w_refcount = 1;
1423 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1424 	if (lock_class->lc_flags & LC_SPINLOCK) {
1425 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1426 		w_spin_cnt++;
1427 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1428 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1429 		w_sleep_cnt++;
1430 	} else {
1431 		mtx_unlock_spin(&w_mtx);
1432 		panic("lock class %s is not sleep or spin",
1433 		    lock_class->lc_name);
1434 	}
1435 	mtx_unlock_spin(&w_mtx);
1436 out:
1437 	/*
1438 	 * We issue a warning for any spin locks not defined in the static
1439 	 * order list as a way to discourage their use (folks should really
1440 	 * be using non-spin mutexes most of the time).  However, several
1441 	 * 3rd part device drivers use spin locks because that is all they
1442 	 * have available on Windows and Linux and they think that normal
1443 	 * mutexes are insufficient.
1444 	 */
1445 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1446 		printf("WITNESS: spin lock %s not in order list\n",
1447 		    description);
1448 	return (w);
1449 }
1450 
1451 /* Don't let the door bang you on the way out... */
1452 static int
1453 depart(struct witness *w)
1454 {
1455 	struct witness_child_list_entry *wcl, *nwcl;
1456 	struct witness_list *list;
1457 	struct witness *parent;
1458 
1459 	MPASS(w->w_refcount == 0);
1460 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1461 		list = &w_sleep;
1462 		w_sleep_cnt--;
1463 	} else {
1464 		list = &w_spin;
1465 		w_spin_cnt--;
1466 	}
1467 	/*
1468 	 * First, we run through the entire tree looking for any
1469 	 * witnesses that the outgoing witness is a child of.  For
1470 	 * each parent that we find, we reparent all the direct
1471 	 * children of the outgoing witness to its parent.
1472 	 */
1473 	STAILQ_FOREACH(parent, list, w_typelist) {
1474 		if (!isitmychild(parent, w))
1475 			continue;
1476 		removechild(parent, w);
1477 	}
1478 
1479 	/*
1480 	 * Now we go through and free up the child list of the
1481 	 * outgoing witness.
1482 	 */
1483 	for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1484 		nwcl = wcl->wcl_next;
1485         	w_child_cnt--;
1486 		witness_child_free(wcl);
1487 	}
1488 
1489 	/*
1490 	 * Detach from various lists and free.
1491 	 */
1492 	STAILQ_REMOVE(list, w, witness, w_typelist);
1493 	STAILQ_REMOVE(&w_all, w, witness, w_list);
1494 	witness_free(w);
1495 
1496 	return (1);
1497 }
1498 
1499 /*
1500  * Add "child" as a direct child of "parent".  Returns false if
1501  * we fail due to out of memory.
1502  */
1503 static int
1504 insertchild(struct witness *parent, struct witness *child)
1505 {
1506 	struct witness_child_list_entry **wcl;
1507 
1508 	MPASS(child != NULL && parent != NULL);
1509 
1510 	/*
1511 	 * Insert "child" after "parent"
1512 	 */
1513 	wcl = &parent->w_children;
1514 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1515 		wcl = &(*wcl)->wcl_next;
1516 	if (*wcl == NULL) {
1517 		*wcl = witness_child_get();
1518 		if (*wcl == NULL)
1519 			return (0);
1520         	w_child_cnt++;
1521 	}
1522 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1523 
1524 	return (1);
1525 }
1526 
1527 
1528 static int
1529 itismychild(struct witness *parent, struct witness *child)
1530 {
1531 	struct witness_list *list;
1532 
1533 	MPASS(child != NULL && parent != NULL);
1534 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1535 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1536 		panic(
1537 		"%s: parent (%s) and child (%s) are not the same lock type",
1538 		    __func__, parent->w_class->lc_name,
1539 		    child->w_class->lc_name);
1540 
1541 	if (!insertchild(parent, child))
1542 		return (0);
1543 
1544 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1545 		list = &w_sleep;
1546 	else
1547 		list = &w_spin;
1548 	return (1);
1549 }
1550 
1551 static void
1552 removechild(struct witness *parent, struct witness *child)
1553 {
1554 	struct witness_child_list_entry **wcl, *wcl1;
1555 	int i;
1556 
1557 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1558 		for (i = 0; i < (*wcl)->wcl_count; i++)
1559 			if ((*wcl)->wcl_children[i] == child)
1560 				goto found;
1561 	return;
1562 found:
1563 	(*wcl)->wcl_count--;
1564 	if ((*wcl)->wcl_count > i)
1565 		(*wcl)->wcl_children[i] =
1566 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
1567 	MPASS((*wcl)->wcl_children[i] != NULL);
1568 	if ((*wcl)->wcl_count != 0)
1569 		return;
1570 	wcl1 = *wcl;
1571 	*wcl = wcl1->wcl_next;
1572 	w_child_cnt--;
1573 	witness_child_free(wcl1);
1574 }
1575 
1576 static int
1577 isitmychild(struct witness *parent, struct witness *child)
1578 {
1579 	struct witness_child_list_entry *wcl;
1580 	int i;
1581 
1582 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1583 		for (i = 0; i < wcl->wcl_count; i++) {
1584 			if (wcl->wcl_children[i] == child)
1585 				return (1);
1586 		}
1587 	}
1588 	return (0);
1589 }
1590 
1591 static int
1592 isitmydescendant(struct witness *parent, struct witness *child)
1593 {
1594 	struct witness_child_list_entry *wcl;
1595 	int i, j;
1596 
1597 	if (isitmychild(parent, child))
1598 		return (1);
1599 	j = 0;
1600 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1601 		MPASS(j < 1000);
1602 		for (i = 0; i < wcl->wcl_count; i++) {
1603 			if (isitmydescendant(wcl->wcl_children[i], child))
1604 				return (1);
1605 		}
1606 		j++;
1607 	}
1608 	return (0);
1609 }
1610 
1611 #ifdef BLESSING
1612 static int
1613 blessed(struct witness *w1, struct witness *w2)
1614 {
1615 	int i;
1616 	struct witness_blessed *b;
1617 
1618 	for (i = 0; i < blessed_count; i++) {
1619 		b = &blessed_list[i];
1620 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1621 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1622 				return (1);
1623 			continue;
1624 		}
1625 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1626 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1627 				return (1);
1628 	}
1629 	return (0);
1630 }
1631 #endif
1632 
1633 static struct witness *
1634 witness_get(void)
1635 {
1636 	struct witness *w;
1637 
1638 	if (witness_watch == 0) {
1639 		mtx_unlock_spin(&w_mtx);
1640 		return (NULL);
1641 	}
1642 	if (STAILQ_EMPTY(&w_free)) {
1643 		witness_watch = 0;
1644 		mtx_unlock_spin(&w_mtx);
1645 		printf("%s: witness exhausted\n", __func__);
1646 		return (NULL);
1647 	}
1648 	w = STAILQ_FIRST(&w_free);
1649 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1650 	w_free_cnt--;
1651 	bzero(w, sizeof(*w));
1652 	return (w);
1653 }
1654 
1655 static void
1656 witness_free(struct witness *w)
1657 {
1658 
1659 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1660 	w_free_cnt++;
1661 }
1662 
1663 static struct witness_child_list_entry *
1664 witness_child_get(void)
1665 {
1666 	struct witness_child_list_entry *wcl;
1667 
1668 	if (witness_watch == 0) {
1669 		mtx_unlock_spin(&w_mtx);
1670 		return (NULL);
1671 	}
1672 	wcl = w_child_free;
1673 	if (wcl == NULL) {
1674 		witness_watch = 0;
1675 		mtx_unlock_spin(&w_mtx);
1676 		printf("%s: witness exhausted\n", __func__);
1677 		return (NULL);
1678 	}
1679 	w_child_free = wcl->wcl_next;
1680 	w_child_free_cnt--;
1681 	bzero(wcl, sizeof(*wcl));
1682 	return (wcl);
1683 }
1684 
1685 static void
1686 witness_child_free(struct witness_child_list_entry *wcl)
1687 {
1688 
1689 	wcl->wcl_next = w_child_free;
1690 	w_child_free = wcl;
1691 	w_child_free_cnt++;
1692 }
1693 
1694 static struct lock_list_entry *
1695 witness_lock_list_get(void)
1696 {
1697 	struct lock_list_entry *lle;
1698 
1699 	if (witness_watch == 0)
1700 		return (NULL);
1701 	mtx_lock_spin(&w_mtx);
1702 	lle = w_lock_list_free;
1703 	if (lle == NULL) {
1704 		witness_watch = 0;
1705 		mtx_unlock_spin(&w_mtx);
1706 		printf("%s: witness exhausted\n", __func__);
1707 		return (NULL);
1708 	}
1709 	w_lock_list_free = lle->ll_next;
1710 	mtx_unlock_spin(&w_mtx);
1711 	bzero(lle, sizeof(*lle));
1712 	return (lle);
1713 }
1714 
1715 static void
1716 witness_lock_list_free(struct lock_list_entry *lle)
1717 {
1718 
1719 	mtx_lock_spin(&w_mtx);
1720 	lle->ll_next = w_lock_list_free;
1721 	w_lock_list_free = lle;
1722 	mtx_unlock_spin(&w_mtx);
1723 }
1724 
1725 static struct lock_instance *
1726 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1727 {
1728 	struct lock_list_entry *lle;
1729 	struct lock_instance *instance;
1730 	int i;
1731 
1732 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1733 		for (i = lle->ll_count - 1; i >= 0; i--) {
1734 			instance = &lle->ll_children[i];
1735 			if (instance->li_lock == lock)
1736 				return (instance);
1737 		}
1738 	return (NULL);
1739 }
1740 
1741 static void
1742 witness_list_lock(struct lock_instance *instance)
1743 {
1744 	struct lock_object *lock;
1745 
1746 	lock = instance->li_lock;
1747 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1748 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1749 	if (lock->lo_type != lock->lo_name)
1750 		printf(" (%s)", lock->lo_type);
1751 	printf(" r = %d (%p) locked @ %s:%d\n",
1752 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1753 	    instance->li_line);
1754 }
1755 
1756 #ifdef DDB
1757 static int
1758 witness_thread_has_locks(struct thread *td)
1759 {
1760 
1761 	return (td->td_sleeplocks != NULL);
1762 }
1763 
1764 static int
1765 witness_proc_has_locks(struct proc *p)
1766 {
1767 	struct thread *td;
1768 
1769 	FOREACH_THREAD_IN_PROC(p, td) {
1770 		if (witness_thread_has_locks(td))
1771 			return (1);
1772 	}
1773 	return (0);
1774 }
1775 #endif
1776 
1777 int
1778 witness_list_locks(struct lock_list_entry **lock_list)
1779 {
1780 	struct lock_list_entry *lle;
1781 	int i, nheld;
1782 
1783 	nheld = 0;
1784 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1785 		for (i = lle->ll_count - 1; i >= 0; i--) {
1786 			witness_list_lock(&lle->ll_children[i]);
1787 			nheld++;
1788 		}
1789 	return (nheld);
1790 }
1791 
1792 /*
1793  * This is a bit risky at best.  We call this function when we have timed
1794  * out acquiring a spin lock, and we assume that the other CPU is stuck
1795  * with this lock held.  So, we go groveling around in the other CPU's
1796  * per-cpu data to try to find the lock instance for this spin lock to
1797  * see when it was last acquired.
1798  */
1799 void
1800 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1801 {
1802 	struct lock_instance *instance;
1803 	struct pcpu *pc;
1804 
1805 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1806 		return;
1807 	pc = pcpu_find(owner->td_oncpu);
1808 	instance = find_instance(pc->pc_spinlocks, lock);
1809 	if (instance != NULL)
1810 		witness_list_lock(instance);
1811 }
1812 
1813 void
1814 witness_save(struct lock_object *lock, const char **filep, int *linep)
1815 {
1816 	struct lock_list_entry *lock_list;
1817 	struct lock_instance *instance;
1818 	struct lock_class *class;
1819 
1820 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1821 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1822 		return;
1823 	class = LOCK_CLASS(lock);
1824 	if (class->lc_flags & LC_SLEEPLOCK)
1825 		lock_list = curthread->td_sleeplocks;
1826 	else {
1827 		if (witness_skipspin)
1828 			return;
1829 		lock_list = PCPU_GET(spinlocks);
1830 	}
1831 	instance = find_instance(lock_list, lock);
1832 	if (instance == NULL)
1833 		panic("%s: lock (%s) %s not locked", __func__,
1834 		    class->lc_name, lock->lo_name);
1835 	*filep = instance->li_file;
1836 	*linep = instance->li_line;
1837 }
1838 
1839 void
1840 witness_restore(struct lock_object *lock, const char *file, int line)
1841 {
1842 	struct lock_list_entry *lock_list;
1843 	struct lock_instance *instance;
1844 	struct lock_class *class;
1845 
1846 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1847 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1848 		return;
1849 	class = LOCK_CLASS(lock);
1850 	if (class->lc_flags & LC_SLEEPLOCK)
1851 		lock_list = curthread->td_sleeplocks;
1852 	else {
1853 		if (witness_skipspin)
1854 			return;
1855 		lock_list = PCPU_GET(spinlocks);
1856 	}
1857 	instance = find_instance(lock_list, lock);
1858 	if (instance == NULL)
1859 		panic("%s: lock (%s) %s not locked", __func__,
1860 		    class->lc_name, lock->lo_name);
1861 	lock->lo_witness->w_file = file;
1862 	lock->lo_witness->w_line = line;
1863 	instance->li_file = file;
1864 	instance->li_line = line;
1865 }
1866 
1867 void
1868 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1869 {
1870 #ifdef INVARIANT_SUPPORT
1871 	struct lock_instance *instance;
1872 	struct lock_class *class;
1873 
1874 	if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1875 		return;
1876 	class = LOCK_CLASS(lock);
1877 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1878 		instance = find_instance(curthread->td_sleeplocks, lock);
1879 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
1880 		instance = find_instance(PCPU_GET(spinlocks), lock);
1881 	else {
1882 		panic("Lock (%s) %s is not sleep or spin!",
1883 		    class->lc_name, lock->lo_name);
1884 	}
1885 	file = fixup_filename(file);
1886 	switch (flags) {
1887 	case LA_UNLOCKED:
1888 		if (instance != NULL)
1889 			panic("Lock (%s) %s locked @ %s:%d.",
1890 			    class->lc_name, lock->lo_name, file, line);
1891 		break;
1892 	case LA_LOCKED:
1893 	case LA_LOCKED | LA_RECURSED:
1894 	case LA_LOCKED | LA_NOTRECURSED:
1895 	case LA_SLOCKED:
1896 	case LA_SLOCKED | LA_RECURSED:
1897 	case LA_SLOCKED | LA_NOTRECURSED:
1898 	case LA_XLOCKED:
1899 	case LA_XLOCKED | LA_RECURSED:
1900 	case LA_XLOCKED | LA_NOTRECURSED:
1901 		if (instance == NULL) {
1902 			panic("Lock (%s) %s not locked @ %s:%d.",
1903 			    class->lc_name, lock->lo_name, file, line);
1904 			break;
1905 		}
1906 		if ((flags & LA_XLOCKED) != 0 &&
1907 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1908 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1909 			    class->lc_name, lock->lo_name, file, line);
1910 		if ((flags & LA_SLOCKED) != 0 &&
1911 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1912 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1913 			    class->lc_name, lock->lo_name, file, line);
1914 		if ((flags & LA_RECURSED) != 0 &&
1915 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1916 			panic("Lock (%s) %s not recursed @ %s:%d.",
1917 			    class->lc_name, lock->lo_name, file, line);
1918 		if ((flags & LA_NOTRECURSED) != 0 &&
1919 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1920 			panic("Lock (%s) %s recursed @ %s:%d.",
1921 			    class->lc_name, lock->lo_name, file, line);
1922 		break;
1923 	default:
1924 		panic("Invalid lock assertion at %s:%d.", file, line);
1925 
1926 	}
1927 #endif	/* INVARIANT_SUPPORT */
1928 }
1929 
1930 #ifdef DDB
1931 static void
1932 witness_list(struct thread *td)
1933 {
1934 
1935 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1936 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1937 
1938 	if (witness_watch == 0)
1939 		return;
1940 
1941 	witness_list_locks(&td->td_sleeplocks);
1942 
1943 	/*
1944 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
1945 	 * if td is currently executing on some other CPU and holds spin locks
1946 	 * as we won't display those locks.  If we had a MI way of getting
1947 	 * the per-cpu data for a given cpu then we could use
1948 	 * td->td_oncpu to get the list of spinlocks for this thread
1949 	 * and "fix" this.
1950 	 *
1951 	 * That still wouldn't really fix this unless we locked sched_lock
1952 	 * or stopped the other CPU to make sure it wasn't changing the list
1953 	 * out from under us.  It is probably best to just not try to handle
1954 	 * threads on other CPU's for now.
1955 	 */
1956 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
1957 		witness_list_locks(PCPU_PTR(spinlocks));
1958 }
1959 
1960 DB_SHOW_COMMAND(locks, db_witness_list)
1961 {
1962 	struct thread *td;
1963 
1964 	if (have_addr)
1965 		td = db_lookup_thread(addr, TRUE);
1966 	else
1967 		td = kdb_thread;
1968 	witness_list(td);
1969 }
1970 
1971 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1972 {
1973 	struct thread *td;
1974 	struct proc *p;
1975 
1976 	/*
1977 	 * It would be nice to list only threads and processes that actually
1978 	 * held sleep locks, but that information is currently not exported
1979 	 * by WITNESS.
1980 	 */
1981 	FOREACH_PROC_IN_SYSTEM(p) {
1982 		if (!witness_proc_has_locks(p))
1983 			continue;
1984 		FOREACH_THREAD_IN_PROC(p, td) {
1985 			if (!witness_thread_has_locks(td))
1986 				continue;
1987 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1988 			    p->p_comm, td, td->td_tid);
1989 			witness_list(td);
1990 		}
1991 	}
1992 }
1993 
1994 DB_SHOW_COMMAND(witness, db_witness_display)
1995 {
1996 
1997 	witness_display(db_printf);
1998 }
1999 #endif
2000