xref: /freebsd/sys/kern/subr_witness.c (revision 25cdacf79b06356c929e59d5074d26c9dac41bdf)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2008 Isilon Systems, Inc.
5  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
6  * Copyright (c) 1998 Berkeley Software Design, Inc.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Berkeley Software Design Inc's name may not be used to endorse or
18  *    promote products derived from this software without specific prior
19  *    written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35  */
36 
37 /*
38  * Implementation of the `witness' lock verifier.  Originally implemented for
39  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
40  * classes in FreeBSD.
41  */
42 
43 /*
44  *	Main Entry: witness
45  *	Pronunciation: 'wit-n&s
46  *	Function: noun
47  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
48  *	    testimony, witness, from 2wit
49  *	Date: before 12th century
50  *	1 : attestation of a fact or event : TESTIMONY
51  *	2 : one that gives evidence; specifically : one who testifies in
52  *	    a cause or before a judicial tribunal
53  *	3 : one asked to be present at a transaction so as to be able to
54  *	    testify to its having taken place
55  *	4 : one who has personal knowledge of something
56  *	5 a : something serving as evidence or proof : SIGN
57  *	  b : public affirmation by word or example of usually
58  *	      religious faith or conviction <the heroic witness to divine
59  *	      life -- Pilot>
60  *	6 capitalized : a member of the Jehovah's Witnesses
61  */
62 
63 /*
64  * Special rules concerning Giant and lock orders:
65  *
66  * 1) Giant must be acquired before any other mutexes.  Stated another way,
67  *    no other mutex may be held when Giant is acquired.
68  *
69  * 2) Giant must be released when blocking on a sleepable lock.
70  *
71  * This rule is less obvious, but is a result of Giant providing the same
72  * semantics as spl().  Basically, when a thread sleeps, it must release
73  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
74  * 2).
75  *
76  * 3) Giant may be acquired before or after sleepable locks.
77  *
78  * This rule is also not quite as obvious.  Giant may be acquired after
79  * a sleepable lock because it is a non-sleepable lock and non-sleepable
80  * locks may always be acquired while holding a sleepable lock.  The second
81  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
82  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
83  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
84  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
85  * execute.  Thus, acquiring Giant both before and after a sleepable lock
86  * will not result in a lock order reversal.
87  */
88 
89 #include <sys/cdefs.h>
90 __FBSDID("$FreeBSD$");
91 
92 #include "opt_ddb.h"
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
96 
97 #include <sys/param.h>
98 #include <sys/bus.h>
99 #include <sys/kdb.h>
100 #include <sys/kernel.h>
101 #include <sys/ktr.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
113 
114 #ifdef DDB
115 #include <ddb/ddb.h>
116 #endif
117 
118 #include <machine/stdarg.h>
119 
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
122 #endif
123 
124 /* Note that these traces do not work with KTR_ALQ. */
125 #if 0
126 #define	KTR_WITNESS	KTR_SUBSYS
127 #else
128 #define	KTR_WITNESS	0
129 #endif
130 
131 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
132 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
133 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
134 #define	LI_SLEEPABLE	0x00040000	/* Lock may be held while sleeping. */
135 
136 #ifndef WITNESS_COUNT
137 #define	WITNESS_COUNT 		1536
138 #endif
139 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
140 #define	WITNESS_PENDLIST	(512 + (MAXCPU * 4))
141 
142 /* Allocate 256 KB of stack data space */
143 #define	WITNESS_LO_DATA_COUNT	2048
144 
145 /* Prime, gives load factor of ~2 at full load */
146 #define	WITNESS_LO_HASH_SIZE	1021
147 
148 /*
149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
151  * probably be safe for the most part, but it's still a SWAG.
152  */
153 #define	LOCK_NCHILDREN	5
154 #define	LOCK_CHILDCOUNT	2048
155 
156 #define	MAX_W_NAME	64
157 
158 #define	FULLGRAPH_SBUF_SIZE	512
159 
160 /*
161  * These flags go in the witness relationship matrix and describe the
162  * relationship between any two struct witness objects.
163  */
164 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
165 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
166 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
167 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
168 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
169 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define	WITNESS_RELATED_MASK						\
172 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
174 					  * observed. */
175 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
176 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
177 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
178 
179 /* Descendant to ancestor flags */
180 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
181 
182 /* Ancestor to descendant flags */
183 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
184 
185 #define	WITNESS_INDEX_ASSERT(i)						\
186 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
187 
188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189 
190 /*
191  * Lock instances.  A lock instance is the data associated with a lock while
192  * it is held by witness.  For example, a lock instance will hold the
193  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
194  * are held in a per-cpu list while sleep locks are held in per-thread list.
195  */
196 struct lock_instance {
197 	struct lock_object	*li_lock;
198 	const char		*li_file;
199 	int			li_line;
200 	u_int			li_flags;
201 };
202 
203 /*
204  * A simple list type used to build the list of locks held by a thread
205  * or CPU.  We can't simply embed the list in struct lock_object since a
206  * lock may be held by more than one thread if it is a shared lock.  Locks
207  * are added to the head of the list, so we fill up each list entry from
208  * "the back" logically.  To ease some of the arithmetic, we actually fill
209  * in each list entry the normal way (children[0] then children[1], etc.) but
210  * when we traverse the list we read children[count-1] as the first entry
211  * down to children[0] as the final entry.
212  */
213 struct lock_list_entry {
214 	struct lock_list_entry	*ll_next;
215 	struct lock_instance	ll_children[LOCK_NCHILDREN];
216 	u_int			ll_count;
217 };
218 
219 /*
220  * The main witness structure. One of these per named lock type in the system
221  * (for example, "vnode interlock").
222  */
223 struct witness {
224 	char  			w_name[MAX_W_NAME];
225 	uint32_t 		w_index;  /* Index in the relationship matrix */
226 	struct lock_class	*w_class;
227 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
228 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
229 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
230 	const char		*w_file; /* File where last acquired */
231 	uint32_t 		w_line; /* Line where last acquired */
232 	uint32_t 		w_refcount;
233 	uint16_t 		w_num_ancestors; /* direct/indirect
234 						  * ancestor count */
235 	uint16_t 		w_num_descendants; /* direct/indirect
236 						    * descendant count */
237 	int16_t 		w_ddb_level;
238 	unsigned		w_displayed:1;
239 	unsigned		w_reversed:1;
240 };
241 
242 STAILQ_HEAD(witness_list, witness);
243 
244 /*
245  * The witness hash table. Keys are witness names (const char *), elements are
246  * witness objects (struct witness *).
247  */
248 struct witness_hash {
249 	struct witness	*wh_array[WITNESS_HASH_SIZE];
250 	uint32_t	wh_size;
251 	uint32_t	wh_count;
252 };
253 
254 /*
255  * Key type for the lock order data hash table.
256  */
257 struct witness_lock_order_key {
258 	uint16_t	from;
259 	uint16_t	to;
260 };
261 
262 struct witness_lock_order_data {
263 	struct stack			wlod_stack;
264 	struct witness_lock_order_key	wlod_key;
265 	struct witness_lock_order_data	*wlod_next;
266 };
267 
268 /*
269  * The witness lock order data hash table. Keys are witness index tuples
270  * (struct witness_lock_order_key), elements are lock order data objects
271  * (struct witness_lock_order_data).
272  */
273 struct witness_lock_order_hash {
274 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
275 	u_int	wloh_size;
276 	u_int	wloh_count;
277 };
278 
279 struct witness_blessed {
280 	const char	*b_lock1;
281 	const char	*b_lock2;
282 };
283 
284 struct witness_pendhelp {
285 	const char		*wh_type;
286 	struct lock_object	*wh_lock;
287 };
288 
289 struct witness_order_list_entry {
290 	const char		*w_name;
291 	struct lock_class	*w_class;
292 };
293 
294 /*
295  * Returns 0 if one of the locks is a spin lock and the other is not.
296  * Returns 1 otherwise.
297  */
298 static __inline int
299 witness_lock_type_equal(struct witness *w1, struct witness *w2)
300 {
301 
302 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
303 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
304 }
305 
306 static __inline int
307 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
308     const struct witness_lock_order_key *b)
309 {
310 
311 	return (a->from == b->from && a->to == b->to);
312 }
313 
314 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
315 		    const char *fname);
316 static void	adopt(struct witness *parent, struct witness *child);
317 static int	blessed(struct witness *, struct witness *);
318 static void	depart(struct witness *w);
319 static struct witness	*enroll(const char *description,
320 			    struct lock_class *lock_class);
321 static struct lock_instance	*find_instance(struct lock_list_entry *list,
322 				    const struct lock_object *lock);
323 static int	isitmychild(struct witness *parent, struct witness *child);
324 static int	isitmydescendant(struct witness *parent, struct witness *child);
325 static void	itismychild(struct witness *parent, struct witness *child);
326 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
327 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
328 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
329 static int	sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
330 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
331 #ifdef DDB
332 static void	witness_ddb_compute_levels(void);
333 static void	witness_ddb_display(int(*)(const char *fmt, ...));
334 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
335 		    struct witness *, int indent);
336 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
337 		    struct witness_list *list);
338 static void	witness_ddb_level_descendants(struct witness *parent, int l);
339 static void	witness_ddb_list(struct thread *td);
340 #endif
341 static void	witness_enter_debugger(const char *msg);
342 static void	witness_debugger(int cond, const char *msg);
343 static void	witness_free(struct witness *m);
344 static struct witness	*witness_get(void);
345 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
346 static struct witness	*witness_hash_get(const char *key);
347 static void	witness_hash_put(struct witness *w);
348 static void	witness_init_hash_tables(void);
349 static void	witness_increment_graph_generation(void);
350 static void	witness_lock_list_free(struct lock_list_entry *lle);
351 static struct lock_list_entry	*witness_lock_list_get(void);
352 static int	witness_lock_order_add(struct witness *parent,
353 		    struct witness *child);
354 static int	witness_lock_order_check(struct witness *parent,
355 		    struct witness *child);
356 static struct witness_lock_order_data	*witness_lock_order_get(
357 					    struct witness *parent,
358 					    struct witness *child);
359 static void	witness_list_lock(struct lock_instance *instance,
360 		    int (*prnt)(const char *fmt, ...));
361 static int	witness_output(const char *fmt, ...) __printflike(1, 2);
362 static int	witness_output_drain(void *arg __unused, const char *data,
363 		    int len);
364 static int	witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
365 static void	witness_setflag(struct lock_object *lock, int flag, int set);
366 
367 FEATURE(witness, "kernel has witness(9) support");
368 
369 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
370     "Witness Locking");
371 
372 /*
373  * If set to 0, lock order checking is disabled.  If set to -1,
374  * witness is completely disabled.  Otherwise witness performs full
375  * lock order checking for all locks.  At runtime, lock order checking
376  * may be toggled.  However, witness cannot be reenabled once it is
377  * completely disabled.
378  */
379 static int witness_watch = 1;
380 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
381     CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
382     sysctl_debug_witness_watch, "I",
383     "witness is watching lock operations");
384 
385 #ifdef KDB
386 /*
387  * When KDB is enabled and witness_kdb is 1, it will cause the system
388  * to drop into kdebug() when:
389  *	- a lock hierarchy violation occurs
390  *	- locks are held when going to sleep.
391  */
392 #ifdef WITNESS_KDB
393 int	witness_kdb = 1;
394 #else
395 int	witness_kdb = 0;
396 #endif
397 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
398 #endif /* KDB */
399 
400 #if defined(DDB) || defined(KDB)
401 /*
402  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
403  * to print a stack trace:
404  *	- a lock hierarchy violation occurs
405  *	- locks are held when going to sleep.
406  */
407 int	witness_trace = 1;
408 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
409 #endif /* DDB || KDB */
410 
411 #ifdef WITNESS_SKIPSPIN
412 int	witness_skipspin = 1;
413 #else
414 int	witness_skipspin = 0;
415 #endif
416 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
417 
418 int badstack_sbuf_size;
419 
420 int witness_count = WITNESS_COUNT;
421 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
422     &witness_count, 0, "");
423 
424 /*
425  * Output channel for witness messages.  By default we print to the console.
426  */
427 enum witness_channel {
428 	WITNESS_CONSOLE,
429 	WITNESS_LOG,
430 	WITNESS_NONE,
431 };
432 
433 static enum witness_channel witness_channel = WITNESS_CONSOLE;
434 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
435     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
436     sysctl_debug_witness_channel, "A",
437     "Output channel for warnings");
438 
439 /*
440  * Call this to print out the relations between locks.
441  */
442 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
443     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
444     sysctl_debug_witness_fullgraph, "A",
445     "Show locks relation graphs");
446 
447 /*
448  * Call this to print out the witness faulty stacks.
449  */
450 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
451     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
452     sysctl_debug_witness_badstacks, "A",
453     "Show bad witness stacks");
454 
455 static struct mtx w_mtx;
456 
457 /* w_list */
458 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
459 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
460 
461 /* w_typelist */
462 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
463 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
464 
465 /* lock list */
466 static struct lock_list_entry *w_lock_list_free = NULL;
467 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
468 static u_int pending_cnt;
469 
470 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
471 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
472 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
473 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
474     "");
475 
476 static struct witness *w_data;
477 static uint8_t **w_rmatrix;
478 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
479 static struct witness_hash w_hash;	/* The witness hash table. */
480 
481 /* The lock order data hash */
482 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
483 static struct witness_lock_order_data *w_lofree = NULL;
484 static struct witness_lock_order_hash w_lohash;
485 static int w_max_used_index = 0;
486 static unsigned int w_generation = 0;
487 static const char w_notrunning[] = "Witness not running\n";
488 static const char w_stillcold[] = "Witness is still cold\n";
489 #ifdef __i386__
490 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
491 #endif
492 
493 static struct witness_order_list_entry order_lists[] = {
494 	/*
495 	 * sx locks
496 	 */
497 	{ "proctree", &lock_class_sx },
498 	{ "allproc", &lock_class_sx },
499 	{ "allprison", &lock_class_sx },
500 	{ NULL, NULL },
501 	/*
502 	 * Various mutexes
503 	 */
504 	{ "Giant", &lock_class_mtx_sleep },
505 	{ "pipe mutex", &lock_class_mtx_sleep },
506 	{ "sigio lock", &lock_class_mtx_sleep },
507 	{ "process group", &lock_class_mtx_sleep },
508 #ifdef	HWPMC_HOOKS
509 	{ "pmc-sleep", &lock_class_mtx_sleep },
510 #endif
511 	{ "process lock", &lock_class_mtx_sleep },
512 	{ "session", &lock_class_mtx_sleep },
513 	{ "uidinfo hash", &lock_class_rw },
514 	{ "time lock", &lock_class_mtx_sleep },
515 	{ NULL, NULL },
516 	/*
517 	 * umtx
518 	 */
519 	{ "umtx lock", &lock_class_mtx_sleep },
520 	{ NULL, NULL },
521 	/*
522 	 * Sockets
523 	 */
524 	{ "accept", &lock_class_mtx_sleep },
525 	{ "so_snd", &lock_class_mtx_sleep },
526 	{ "so_rcv", &lock_class_mtx_sleep },
527 	{ "sellck", &lock_class_mtx_sleep },
528 	{ NULL, NULL },
529 	/*
530 	 * Routing
531 	 */
532 	{ "so_rcv", &lock_class_mtx_sleep },
533 	{ "radix node head", &lock_class_rm },
534 	{ "ifaddr", &lock_class_mtx_sleep },
535 	{ NULL, NULL },
536 	/*
537 	 * IPv4 multicast:
538 	 * protocol locks before interface locks, after UDP locks.
539 	 */
540 	{ "in_multi_sx", &lock_class_sx },
541 	{ "udpinp", &lock_class_rw },
542 	{ "in_multi_list_mtx", &lock_class_mtx_sleep },
543 	{ "igmp_mtx", &lock_class_mtx_sleep },
544 	{ "ifnet_rw", &lock_class_rw },
545 	{ "if_addr_lock", &lock_class_mtx_sleep },
546 	{ NULL, NULL },
547 	/*
548 	 * IPv6 multicast:
549 	 * protocol locks before interface locks, after UDP locks.
550 	 */
551 	{ "in6_multi_sx", &lock_class_sx },
552 	{ "udpinp", &lock_class_rw },
553 	{ "in6_multi_list_mtx", &lock_class_mtx_sleep },
554 	{ "mld_mtx", &lock_class_mtx_sleep },
555 	{ "ifnet_rw", &lock_class_rw },
556 	{ "if_addr_lock", &lock_class_mtx_sleep },
557 	{ NULL, NULL },
558 	/*
559 	 * UNIX Domain Sockets
560 	 */
561 	{ "unp_link_rwlock", &lock_class_rw },
562 	{ "unp_list_lock", &lock_class_mtx_sleep },
563 	{ "unp", &lock_class_mtx_sleep },
564 	{ "so_snd", &lock_class_mtx_sleep },
565 	{ NULL, NULL },
566 	/*
567 	 * UDP/IP
568 	 */
569 	{ "udp", &lock_class_mtx_sleep },
570 	{ "udpinp", &lock_class_rw },
571 	{ "so_snd", &lock_class_mtx_sleep },
572 	{ NULL, NULL },
573 	/*
574 	 * TCP/IP
575 	 */
576 	{ "tcp", &lock_class_mtx_sleep },
577 	{ "tcpinp", &lock_class_rw },
578 	{ "so_snd", &lock_class_mtx_sleep },
579 	{ NULL, NULL },
580 	/*
581 	 * BPF
582 	 */
583 	{ "bpf global lock", &lock_class_sx },
584 	{ "bpf cdev lock", &lock_class_mtx_sleep },
585 	{ NULL, NULL },
586 	/*
587 	 * NFS server
588 	 */
589 	{ "nfsd_mtx", &lock_class_mtx_sleep },
590 	{ "so_snd", &lock_class_mtx_sleep },
591 	{ NULL, NULL },
592 
593 	/*
594 	 * IEEE 802.11
595 	 */
596 	{ "802.11 com lock", &lock_class_mtx_sleep},
597 	{ NULL, NULL },
598 	/*
599 	 * Network drivers
600 	 */
601 	{ "network driver", &lock_class_mtx_sleep},
602 	{ NULL, NULL },
603 
604 	/*
605 	 * Netgraph
606 	 */
607 	{ "ng_node", &lock_class_mtx_sleep },
608 	{ "ng_worklist", &lock_class_mtx_sleep },
609 	{ NULL, NULL },
610 	/*
611 	 * CDEV
612 	 */
613 	{ "vm map (system)", &lock_class_mtx_sleep },
614 	{ "vnode interlock", &lock_class_mtx_sleep },
615 	{ "cdev", &lock_class_mtx_sleep },
616 	{ "devthrd", &lock_class_mtx_sleep },
617 	{ NULL, NULL },
618 	/*
619 	 * VM
620 	 */
621 	{ "vm map (user)", &lock_class_sx },
622 	{ "vm object", &lock_class_rw },
623 	{ "vm page", &lock_class_mtx_sleep },
624 	{ "pmap pv global", &lock_class_rw },
625 	{ "pmap", &lock_class_mtx_sleep },
626 	{ "pmap pv list", &lock_class_rw },
627 	{ "vm page free queue", &lock_class_mtx_sleep },
628 	{ "vm pagequeue", &lock_class_mtx_sleep },
629 	{ NULL, NULL },
630 	/*
631 	 * kqueue/VFS interaction
632 	 */
633 	{ "kqueue", &lock_class_mtx_sleep },
634 	{ "struct mount mtx", &lock_class_mtx_sleep },
635 	{ "vnode interlock", &lock_class_mtx_sleep },
636 	{ NULL, NULL },
637 	/*
638 	 * VFS namecache
639 	 */
640 	{ "ncvn", &lock_class_mtx_sleep },
641 	{ "ncbuc", &lock_class_mtx_sleep },
642 	{ "vnode interlock", &lock_class_mtx_sleep },
643 	{ "ncneg", &lock_class_mtx_sleep },
644 	{ NULL, NULL },
645 	/*
646 	 * ZFS locking
647 	 */
648 	{ "dn->dn_mtx", &lock_class_sx },
649 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
650 	{ "db->db_mtx", &lock_class_sx },
651 	{ NULL, NULL },
652 	/*
653 	 * TCP log locks
654 	 */
655 	{ "TCP ID tree", &lock_class_rw },
656 	{ "tcp log id bucket", &lock_class_mtx_sleep },
657 	{ "tcpinp", &lock_class_rw },
658 	{ "TCP log expireq", &lock_class_mtx_sleep },
659 	{ NULL, NULL },
660 	/*
661 	 * spin locks
662 	 */
663 #ifdef SMP
664 	{ "ap boot", &lock_class_mtx_spin },
665 #endif
666 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
667 	{ "sio", &lock_class_mtx_spin },
668 #ifdef __i386__
669 	{ "cy", &lock_class_mtx_spin },
670 #endif
671 	{ "scc_hwmtx", &lock_class_mtx_spin },
672 	{ "uart_hwmtx", &lock_class_mtx_spin },
673 	{ "fast_taskqueue", &lock_class_mtx_spin },
674 	{ "intr table", &lock_class_mtx_spin },
675 	{ "process slock", &lock_class_mtx_spin },
676 	{ "syscons video lock", &lock_class_mtx_spin },
677 	{ "sleepq chain", &lock_class_mtx_spin },
678 	{ "rm_spinlock", &lock_class_mtx_spin },
679 	{ "turnstile chain", &lock_class_mtx_spin },
680 	{ "turnstile lock", &lock_class_mtx_spin },
681 	{ "sched lock", &lock_class_mtx_spin },
682 	{ "td_contested", &lock_class_mtx_spin },
683 	{ "callout", &lock_class_mtx_spin },
684 	{ "entropy harvest mutex", &lock_class_mtx_spin },
685 #ifdef SMP
686 	{ "smp rendezvous", &lock_class_mtx_spin },
687 #endif
688 #ifdef __powerpc__
689 	{ "tlb0", &lock_class_mtx_spin },
690 #endif
691 	{ NULL, NULL },
692 	{ "sched lock", &lock_class_mtx_spin },
693 #ifdef	HWPMC_HOOKS
694 	{ "pmc-per-proc", &lock_class_mtx_spin },
695 #endif
696 	{ NULL, NULL },
697 	/*
698 	 * leaf locks
699 	 */
700 	{ "intrcnt", &lock_class_mtx_spin },
701 	{ "icu", &lock_class_mtx_spin },
702 #ifdef __i386__
703 	{ "allpmaps", &lock_class_mtx_spin },
704 	{ "descriptor tables", &lock_class_mtx_spin },
705 #endif
706 	{ "clk", &lock_class_mtx_spin },
707 	{ "cpuset", &lock_class_mtx_spin },
708 	{ "mprof lock", &lock_class_mtx_spin },
709 	{ "zombie lock", &lock_class_mtx_spin },
710 	{ "ALD Queue", &lock_class_mtx_spin },
711 #if defined(__i386__) || defined(__amd64__)
712 	{ "pcicfg", &lock_class_mtx_spin },
713 	{ "NDIS thread lock", &lock_class_mtx_spin },
714 #endif
715 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
716 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
717 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
718 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
719 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
720 #ifdef	HWPMC_HOOKS
721 	{ "pmc-leaf", &lock_class_mtx_spin },
722 #endif
723 	{ "blocked lock", &lock_class_mtx_spin },
724 	{ NULL, NULL },
725 	{ NULL, NULL }
726 };
727 
728 /*
729  * Pairs of locks which have been blessed.  Witness does not complain about
730  * order problems with blessed lock pairs.  Please do not add an entry to the
731  * table without an explanatory comment.
732  */
733 static struct witness_blessed blessed_list[] = {
734 	/*
735 	 * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
736 	 * both lock orders, so a deadlock cannot happen as a result of this
737 	 * LOR.
738 	 */
739 	{ "dirhash",	"bufwait" },
740 
741 	/*
742 	 * A UFS vnode may be locked in vget() while a buffer belonging to the
743 	 * parent directory vnode is locked.
744 	 */
745 	{ "ufs",	"bufwait" },
746 };
747 
748 /*
749  * This global is set to 0 once it becomes safe to use the witness code.
750  */
751 static int witness_cold = 1;
752 
753 /*
754  * This global is set to 1 once the static lock orders have been enrolled
755  * so that a warning can be issued for any spin locks enrolled later.
756  */
757 static int witness_spin_warn = 0;
758 
759 /* Trim useless garbage from filenames. */
760 static const char *
761 fixup_filename(const char *file)
762 {
763 
764 	if (file == NULL)
765 		return (NULL);
766 	while (strncmp(file, "../", 3) == 0)
767 		file += 3;
768 	return (file);
769 }
770 
771 /*
772  * Calculate the size of early witness structures.
773  */
774 int
775 witness_startup_count(void)
776 {
777 	int sz;
778 
779 	sz = sizeof(struct witness) * witness_count;
780 	sz += sizeof(*w_rmatrix) * (witness_count + 1);
781 	sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
782 	    (witness_count + 1);
783 
784 	return (sz);
785 }
786 
787 /*
788  * The WITNESS-enabled diagnostic code.  Note that the witness code does
789  * assume that the early boot is single-threaded at least until after this
790  * routine is completed.
791  */
792 void
793 witness_startup(void *mem)
794 {
795 	struct lock_object *lock;
796 	struct witness_order_list_entry *order;
797 	struct witness *w, *w1;
798 	uintptr_t p;
799 	int i;
800 
801 	p = (uintptr_t)mem;
802 	w_data = (void *)p;
803 	p += sizeof(struct witness) * witness_count;
804 
805 	w_rmatrix = (void *)p;
806 	p += sizeof(*w_rmatrix) * (witness_count + 1);
807 
808 	for (i = 0; i < witness_count + 1; i++) {
809 		w_rmatrix[i] = (void *)p;
810 		p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
811 	}
812 	badstack_sbuf_size = witness_count * 256;
813 
814 	/*
815 	 * We have to release Giant before initializing its witness
816 	 * structure so that WITNESS doesn't get confused.
817 	 */
818 	mtx_unlock(&Giant);
819 	mtx_assert(&Giant, MA_NOTOWNED);
820 
821 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
822 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
823 	    MTX_NOWITNESS | MTX_NOPROFILE);
824 	for (i = witness_count - 1; i >= 0; i--) {
825 		w = &w_data[i];
826 		memset(w, 0, sizeof(*w));
827 		w_data[i].w_index = i;	/* Witness index never changes. */
828 		witness_free(w);
829 	}
830 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
831 	    ("%s: Invalid list of free witness objects", __func__));
832 
833 	/* Witness with index 0 is not used to aid in debugging. */
834 	STAILQ_REMOVE_HEAD(&w_free, w_list);
835 	w_free_cnt--;
836 
837 	for (i = 0; i < witness_count; i++) {
838 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
839 		    (witness_count + 1));
840 	}
841 
842 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
843 		witness_lock_list_free(&w_locklistdata[i]);
844 	witness_init_hash_tables();
845 
846 	/* First add in all the specified order lists. */
847 	for (order = order_lists; order->w_name != NULL; order++) {
848 		w = enroll(order->w_name, order->w_class);
849 		if (w == NULL)
850 			continue;
851 		w->w_file = "order list";
852 		for (order++; order->w_name != NULL; order++) {
853 			w1 = enroll(order->w_name, order->w_class);
854 			if (w1 == NULL)
855 				continue;
856 			w1->w_file = "order list";
857 			itismychild(w, w1);
858 			w = w1;
859 		}
860 	}
861 	witness_spin_warn = 1;
862 
863 	/* Iterate through all locks and add them to witness. */
864 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
865 		lock = pending_locks[i].wh_lock;
866 		KASSERT(lock->lo_flags & LO_WITNESS,
867 		    ("%s: lock %s is on pending list but not LO_WITNESS",
868 		    __func__, lock->lo_name));
869 		lock->lo_witness = enroll(pending_locks[i].wh_type,
870 		    LOCK_CLASS(lock));
871 	}
872 
873 	/* Mark the witness code as being ready for use. */
874 	witness_cold = 0;
875 
876 	mtx_lock(&Giant);
877 }
878 
879 void
880 witness_init(struct lock_object *lock, const char *type)
881 {
882 	struct lock_class *class;
883 
884 	/* Various sanity checks. */
885 	class = LOCK_CLASS(lock);
886 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
887 	    (class->lc_flags & LC_RECURSABLE) == 0)
888 		kassert_panic("%s: lock (%s) %s can not be recursable",
889 		    __func__, class->lc_name, lock->lo_name);
890 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
891 	    (class->lc_flags & LC_SLEEPABLE) == 0)
892 		kassert_panic("%s: lock (%s) %s can not be sleepable",
893 		    __func__, class->lc_name, lock->lo_name);
894 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
895 	    (class->lc_flags & LC_UPGRADABLE) == 0)
896 		kassert_panic("%s: lock (%s) %s can not be upgradable",
897 		    __func__, class->lc_name, lock->lo_name);
898 
899 	/*
900 	 * If we shouldn't watch this lock, then just clear lo_witness.
901 	 * Otherwise, if witness_cold is set, then it is too early to
902 	 * enroll this lock, so defer it to witness_initialize() by adding
903 	 * it to the pending_locks list.  If it is not too early, then enroll
904 	 * the lock now.
905 	 */
906 	if (witness_watch < 1 || KERNEL_PANICKED() ||
907 	    (lock->lo_flags & LO_WITNESS) == 0)
908 		lock->lo_witness = NULL;
909 	else if (witness_cold) {
910 		pending_locks[pending_cnt].wh_lock = lock;
911 		pending_locks[pending_cnt++].wh_type = type;
912 		if (pending_cnt > WITNESS_PENDLIST)
913 			panic("%s: pending locks list is too small, "
914 			    "increase WITNESS_PENDLIST\n",
915 			    __func__);
916 	} else
917 		lock->lo_witness = enroll(type, class);
918 }
919 
920 void
921 witness_destroy(struct lock_object *lock)
922 {
923 	struct lock_class *class;
924 	struct witness *w;
925 
926 	class = LOCK_CLASS(lock);
927 
928 	if (witness_cold)
929 		panic("lock (%s) %s destroyed while witness_cold",
930 		    class->lc_name, lock->lo_name);
931 
932 	/* XXX: need to verify that no one holds the lock */
933 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
934 		return;
935 	w = lock->lo_witness;
936 
937 	mtx_lock_spin(&w_mtx);
938 	MPASS(w->w_refcount > 0);
939 	w->w_refcount--;
940 
941 	if (w->w_refcount == 0)
942 		depart(w);
943 	mtx_unlock_spin(&w_mtx);
944 }
945 
946 #ifdef DDB
947 static void
948 witness_ddb_compute_levels(void)
949 {
950 	struct witness *w;
951 
952 	/*
953 	 * First clear all levels.
954 	 */
955 	STAILQ_FOREACH(w, &w_all, w_list)
956 		w->w_ddb_level = -1;
957 
958 	/*
959 	 * Look for locks with no parents and level all their descendants.
960 	 */
961 	STAILQ_FOREACH(w, &w_all, w_list) {
962 		/* If the witness has ancestors (is not a root), skip it. */
963 		if (w->w_num_ancestors > 0)
964 			continue;
965 		witness_ddb_level_descendants(w, 0);
966 	}
967 }
968 
969 static void
970 witness_ddb_level_descendants(struct witness *w, int l)
971 {
972 	int i;
973 
974 	if (w->w_ddb_level >= l)
975 		return;
976 
977 	w->w_ddb_level = l;
978 	l++;
979 
980 	for (i = 1; i <= w_max_used_index; i++) {
981 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
982 			witness_ddb_level_descendants(&w_data[i], l);
983 	}
984 }
985 
986 static void
987 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
988     struct witness *w, int indent)
989 {
990 	int i;
991 
992  	for (i = 0; i < indent; i++)
993  		prnt(" ");
994 	prnt("%s (type: %s, depth: %d, active refs: %d)",
995 	     w->w_name, w->w_class->lc_name,
996 	     w->w_ddb_level, w->w_refcount);
997  	if (w->w_displayed) {
998  		prnt(" -- (already displayed)\n");
999  		return;
1000  	}
1001  	w->w_displayed = 1;
1002 	if (w->w_file != NULL && w->w_line != 0)
1003 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1004 		    w->w_line);
1005 	else
1006 		prnt(" -- never acquired\n");
1007 	indent++;
1008 	WITNESS_INDEX_ASSERT(w->w_index);
1009 	for (i = 1; i <= w_max_used_index; i++) {
1010 		if (db_pager_quit)
1011 			return;
1012 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1013 			witness_ddb_display_descendants(prnt, &w_data[i],
1014 			    indent);
1015 	}
1016 }
1017 
1018 static void
1019 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1020     struct witness_list *list)
1021 {
1022 	struct witness *w;
1023 
1024 	STAILQ_FOREACH(w, list, w_typelist) {
1025 		if (w->w_file == NULL || w->w_ddb_level > 0)
1026 			continue;
1027 
1028 		/* This lock has no anscestors - display its descendants. */
1029 		witness_ddb_display_descendants(prnt, w, 0);
1030 		if (db_pager_quit)
1031 			return;
1032 	}
1033 }
1034 
1035 static void
1036 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1037 {
1038 	struct witness *w;
1039 
1040 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1041 	witness_ddb_compute_levels();
1042 
1043 	/* Clear all the displayed flags. */
1044 	STAILQ_FOREACH(w, &w_all, w_list)
1045 		w->w_displayed = 0;
1046 
1047 	/*
1048 	 * First, handle sleep locks which have been acquired at least
1049 	 * once.
1050 	 */
1051 	prnt("Sleep locks:\n");
1052 	witness_ddb_display_list(prnt, &w_sleep);
1053 	if (db_pager_quit)
1054 		return;
1055 
1056 	/*
1057 	 * Now do spin locks which have been acquired at least once.
1058 	 */
1059 	prnt("\nSpin locks:\n");
1060 	witness_ddb_display_list(prnt, &w_spin);
1061 	if (db_pager_quit)
1062 		return;
1063 
1064 	/*
1065 	 * Finally, any locks which have not been acquired yet.
1066 	 */
1067 	prnt("\nLocks which were never acquired:\n");
1068 	STAILQ_FOREACH(w, &w_all, w_list) {
1069 		if (w->w_file != NULL || w->w_refcount == 0)
1070 			continue;
1071 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1072 		    w->w_class->lc_name, w->w_ddb_level);
1073 		if (db_pager_quit)
1074 			return;
1075 	}
1076 }
1077 #endif /* DDB */
1078 
1079 int
1080 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1081 {
1082 
1083 	if (witness_watch == -1 || KERNEL_PANICKED())
1084 		return (0);
1085 
1086 	/* Require locks that witness knows about. */
1087 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1088 	    lock2->lo_witness == NULL)
1089 		return (EINVAL);
1090 
1091 	mtx_assert(&w_mtx, MA_NOTOWNED);
1092 	mtx_lock_spin(&w_mtx);
1093 
1094 	/*
1095 	 * If we already have either an explicit or implied lock order that
1096 	 * is the other way around, then return an error.
1097 	 */
1098 	if (witness_watch &&
1099 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1100 		mtx_unlock_spin(&w_mtx);
1101 		return (EDOOFUS);
1102 	}
1103 
1104 	/* Try to add the new order. */
1105 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1106 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1107 	itismychild(lock1->lo_witness, lock2->lo_witness);
1108 	mtx_unlock_spin(&w_mtx);
1109 	return (0);
1110 }
1111 
1112 void
1113 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1114     int line, struct lock_object *interlock)
1115 {
1116 	struct lock_list_entry *lock_list, *lle;
1117 	struct lock_instance *lock1, *lock2, *plock;
1118 	struct lock_class *class, *iclass;
1119 	struct witness *w, *w1;
1120 	struct thread *td;
1121 	int i, j;
1122 
1123 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1124 	    KERNEL_PANICKED())
1125 		return;
1126 
1127 	w = lock->lo_witness;
1128 	class = LOCK_CLASS(lock);
1129 	td = curthread;
1130 
1131 	if (class->lc_flags & LC_SLEEPLOCK) {
1132 		/*
1133 		 * Since spin locks include a critical section, this check
1134 		 * implicitly enforces a lock order of all sleep locks before
1135 		 * all spin locks.
1136 		 */
1137 		if (td->td_critnest != 0 && !kdb_active)
1138 			kassert_panic("acquiring blockable sleep lock with "
1139 			    "spinlock or critical section held (%s) %s @ %s:%d",
1140 			    class->lc_name, lock->lo_name,
1141 			    fixup_filename(file), line);
1142 
1143 		/*
1144 		 * If this is the first lock acquired then just return as
1145 		 * no order checking is needed.
1146 		 */
1147 		lock_list = td->td_sleeplocks;
1148 		if (lock_list == NULL || lock_list->ll_count == 0)
1149 			return;
1150 	} else {
1151 		/*
1152 		 * If this is the first lock, just return as no order
1153 		 * checking is needed.  Avoid problems with thread
1154 		 * migration pinning the thread while checking if
1155 		 * spinlocks are held.  If at least one spinlock is held
1156 		 * the thread is in a safe path and it is allowed to
1157 		 * unpin it.
1158 		 */
1159 		sched_pin();
1160 		lock_list = PCPU_GET(spinlocks);
1161 		if (lock_list == NULL || lock_list->ll_count == 0) {
1162 			sched_unpin();
1163 			return;
1164 		}
1165 		sched_unpin();
1166 	}
1167 
1168 	/*
1169 	 * Check to see if we are recursing on a lock we already own.  If
1170 	 * so, make sure that we don't mismatch exclusive and shared lock
1171 	 * acquires.
1172 	 */
1173 	lock1 = find_instance(lock_list, lock);
1174 	if (lock1 != NULL) {
1175 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1176 		    (flags & LOP_EXCLUSIVE) == 0) {
1177 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1178 			    class->lc_name, lock->lo_name,
1179 			    fixup_filename(file), line);
1180 			witness_output("while exclusively locked from %s:%d\n",
1181 			    fixup_filename(lock1->li_file), lock1->li_line);
1182 			kassert_panic("excl->share");
1183 		}
1184 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1185 		    (flags & LOP_EXCLUSIVE) != 0) {
1186 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1187 			    class->lc_name, lock->lo_name,
1188 			    fixup_filename(file), line);
1189 			witness_output("while share locked from %s:%d\n",
1190 			    fixup_filename(lock1->li_file), lock1->li_line);
1191 			kassert_panic("share->excl");
1192 		}
1193 		return;
1194 	}
1195 
1196 	/* Warn if the interlock is not locked exactly once. */
1197 	if (interlock != NULL) {
1198 		iclass = LOCK_CLASS(interlock);
1199 		lock1 = find_instance(lock_list, interlock);
1200 		if (lock1 == NULL)
1201 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1202 			    iclass->lc_name, interlock->lo_name,
1203 			    fixup_filename(file), line);
1204 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1205 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1206 			    iclass->lc_name, interlock->lo_name,
1207 			    fixup_filename(file), line);
1208 	}
1209 
1210 	/*
1211 	 * Find the previously acquired lock, but ignore interlocks.
1212 	 */
1213 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1214 	if (interlock != NULL && plock->li_lock == interlock) {
1215 		if (lock_list->ll_count > 1)
1216 			plock =
1217 			    &lock_list->ll_children[lock_list->ll_count - 2];
1218 		else {
1219 			lle = lock_list->ll_next;
1220 
1221 			/*
1222 			 * The interlock is the only lock we hold, so
1223 			 * simply return.
1224 			 */
1225 			if (lle == NULL)
1226 				return;
1227 			plock = &lle->ll_children[lle->ll_count - 1];
1228 		}
1229 	}
1230 
1231 	/*
1232 	 * Try to perform most checks without a lock.  If this succeeds we
1233 	 * can skip acquiring the lock and return success.  Otherwise we redo
1234 	 * the check with the lock held to handle races with concurrent updates.
1235 	 */
1236 	w1 = plock->li_lock->lo_witness;
1237 	if (witness_lock_order_check(w1, w))
1238 		return;
1239 
1240 	mtx_lock_spin(&w_mtx);
1241 	if (witness_lock_order_check(w1, w)) {
1242 		mtx_unlock_spin(&w_mtx);
1243 		return;
1244 	}
1245 	witness_lock_order_add(w1, w);
1246 
1247 	/*
1248 	 * Check for duplicate locks of the same type.  Note that we only
1249 	 * have to check for this on the last lock we just acquired.  Any
1250 	 * other cases will be caught as lock order violations.
1251 	 */
1252 	if (w1 == w) {
1253 		i = w->w_index;
1254 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1255 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1256 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1257 			w->w_reversed = 1;
1258 			mtx_unlock_spin(&w_mtx);
1259 			witness_output(
1260 			    "acquiring duplicate lock of same type: \"%s\"\n",
1261 			    w->w_name);
1262 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1263 			    fixup_filename(plock->li_file), plock->li_line);
1264 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1265 			    fixup_filename(file), line);
1266 			witness_debugger(1, __func__);
1267 		} else
1268 			mtx_unlock_spin(&w_mtx);
1269 		return;
1270 	}
1271 	mtx_assert(&w_mtx, MA_OWNED);
1272 
1273 	/*
1274 	 * If we know that the lock we are acquiring comes after
1275 	 * the lock we most recently acquired in the lock order tree,
1276 	 * then there is no need for any further checks.
1277 	 */
1278 	if (isitmychild(w1, w))
1279 		goto out;
1280 
1281 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1282 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1283 			struct stack pstack;
1284 			bool pstackv, trace;
1285 
1286 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1287 			lock1 = &lle->ll_children[i];
1288 
1289 			/*
1290 			 * Ignore the interlock.
1291 			 */
1292 			if (interlock == lock1->li_lock)
1293 				continue;
1294 
1295 			/*
1296 			 * If this lock doesn't undergo witness checking,
1297 			 * then skip it.
1298 			 */
1299 			w1 = lock1->li_lock->lo_witness;
1300 			if (w1 == NULL) {
1301 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1302 				    ("lock missing witness structure"));
1303 				continue;
1304 			}
1305 
1306 			/*
1307 			 * If we are locking Giant and this is a sleepable
1308 			 * lock, then skip it.
1309 			 */
1310 			if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1311 			    lock == &Giant.lock_object)
1312 				continue;
1313 
1314 			/*
1315 			 * If we are locking a sleepable lock and this lock
1316 			 * is Giant, then skip it.
1317 			 */
1318 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1319 			    (flags & LOP_NOSLEEP) == 0 &&
1320 			    lock1->li_lock == &Giant.lock_object)
1321 				continue;
1322 
1323 			/*
1324 			 * If we are locking a sleepable lock and this lock
1325 			 * isn't sleepable, we want to treat it as a lock
1326 			 * order violation to enfore a general lock order of
1327 			 * sleepable locks before non-sleepable locks.
1328 			 */
1329 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1330 			    (flags & LOP_NOSLEEP) == 0 &&
1331 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1332 				goto reversal;
1333 
1334 			/*
1335 			 * If we are locking Giant and this is a non-sleepable
1336 			 * lock, then treat it as a reversal.
1337 			 */
1338 			if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1339 			    lock == &Giant.lock_object)
1340 				goto reversal;
1341 
1342 			/*
1343 			 * Check the lock order hierarchy for a reveresal.
1344 			 */
1345 			if (!isitmydescendant(w, w1))
1346 				continue;
1347 		reversal:
1348 
1349 			/*
1350 			 * We have a lock order violation, check to see if it
1351 			 * is allowed or has already been yelled about.
1352 			 */
1353 
1354 			/* Bail if this violation is known */
1355 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1356 				goto out;
1357 
1358 			/* Record this as a violation */
1359 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1360 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1361 			w->w_reversed = w1->w_reversed = 1;
1362 			witness_increment_graph_generation();
1363 
1364 			/*
1365 			 * If the lock order is blessed, bail before logging
1366 			 * anything.  We don't look for other lock order
1367 			 * violations though, which may be a bug.
1368 			 */
1369 			if (blessed(w, w1))
1370 				goto out;
1371 
1372 			trace = atomic_load_int(&witness_trace);
1373 			if (trace) {
1374 				struct witness_lock_order_data *data;
1375 
1376 				pstackv = false;
1377 				data = witness_lock_order_get(w, w1);
1378 				if (data != NULL) {
1379 					stack_copy(&data->wlod_stack,
1380 					    &pstack);
1381 					pstackv = true;
1382 				}
1383 			}
1384 			mtx_unlock_spin(&w_mtx);
1385 
1386 #ifdef WITNESS_NO_VNODE
1387 			/*
1388 			 * There are known LORs between VNODE locks. They are
1389 			 * not an indication of a bug. VNODE locks are flagged
1390 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1391 			 * is between 2 VNODE locks.
1392 			 */
1393 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1394 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1395 				return;
1396 #endif
1397 
1398 			/*
1399 			 * Ok, yell about it.
1400 			 */
1401 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1402 			    (flags & LOP_NOSLEEP) == 0 &&
1403 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1404 				witness_output(
1405 		"lock order reversal: (sleepable after non-sleepable)\n");
1406 			else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1407 			    && lock == &Giant.lock_object)
1408 				witness_output(
1409 		"lock order reversal: (Giant after non-sleepable)\n");
1410 			else
1411 				witness_output("lock order reversal:\n");
1412 
1413 			/*
1414 			 * Try to locate an earlier lock with
1415 			 * witness w in our list.
1416 			 */
1417 			do {
1418 				lock2 = &lle->ll_children[i];
1419 				MPASS(lock2->li_lock != NULL);
1420 				if (lock2->li_lock->lo_witness == w)
1421 					break;
1422 				if (i == 0 && lle->ll_next != NULL) {
1423 					lle = lle->ll_next;
1424 					i = lle->ll_count - 1;
1425 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1426 				} else
1427 					i--;
1428 			} while (i >= 0);
1429 			if (i < 0) {
1430 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1431 				    lock1->li_lock, lock1->li_lock->lo_name,
1432 				    w1->w_name, w1->w_class->lc_name,
1433 				    fixup_filename(lock1->li_file),
1434 				    lock1->li_line);
1435 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1436 				    lock, lock->lo_name, w->w_name,
1437 				    w->w_class->lc_name, fixup_filename(file),
1438 				    line);
1439 			} else {
1440 				struct witness *w2 = lock2->li_lock->lo_witness;
1441 
1442 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1443 				    lock2->li_lock, lock2->li_lock->lo_name,
1444 				    w2->w_name, w2->w_class->lc_name,
1445 				    fixup_filename(lock2->li_file),
1446 				    lock2->li_line);
1447 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1448 				    lock1->li_lock, lock1->li_lock->lo_name,
1449 				    w1->w_name, w1->w_class->lc_name,
1450 				    fixup_filename(lock1->li_file),
1451 				    lock1->li_line);
1452 				witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1453 				    lock->lo_name, w->w_name,
1454 				    w->w_class->lc_name, fixup_filename(file),
1455 				    line);
1456 			}
1457 			if (trace) {
1458 				char buf[64];
1459 				struct sbuf sb;
1460 
1461 				sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1462 				sbuf_set_drain(&sb, witness_output_drain,
1463 				    NULL);
1464 
1465 				if (pstackv) {
1466 					sbuf_printf(&sb,
1467 				    "lock order %s -> %s established at:\n",
1468 					    w->w_name, w1->w_name);
1469 					stack_sbuf_print_flags(&sb, &pstack,
1470 					    M_NOWAIT, STACK_SBUF_FMT_LONG);
1471 				}
1472 
1473 				sbuf_printf(&sb,
1474 				    "lock order %s -> %s attempted at:\n",
1475 				    w1->w_name, w->w_name);
1476 				stack_save(&pstack);
1477 				stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1478 				    STACK_SBUF_FMT_LONG);
1479 
1480 				sbuf_finish(&sb);
1481 				sbuf_delete(&sb);
1482 			}
1483 			witness_enter_debugger(__func__);
1484 			return;
1485 		}
1486 	}
1487 
1488 	/*
1489 	 * If requested, build a new lock order.  However, don't build a new
1490 	 * relationship between a sleepable lock and Giant if it is in the
1491 	 * wrong direction.  The correct lock order is that sleepable locks
1492 	 * always come before Giant.
1493 	 */
1494 	if (flags & LOP_NEWORDER &&
1495 	    !(plock->li_lock == &Giant.lock_object &&
1496 	    (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1497 	    (flags & LOP_NOSLEEP) == 0)) {
1498 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1499 		    w->w_name, plock->li_lock->lo_witness->w_name);
1500 		itismychild(plock->li_lock->lo_witness, w);
1501 	}
1502 out:
1503 	mtx_unlock_spin(&w_mtx);
1504 }
1505 
1506 void
1507 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1508 {
1509 	struct lock_list_entry **lock_list, *lle;
1510 	struct lock_instance *instance;
1511 	struct witness *w;
1512 	struct thread *td;
1513 
1514 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1515 	    KERNEL_PANICKED())
1516 		return;
1517 	w = lock->lo_witness;
1518 	td = curthread;
1519 
1520 	/* Determine lock list for this lock. */
1521 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1522 		lock_list = &td->td_sleeplocks;
1523 	else
1524 		lock_list = PCPU_PTR(spinlocks);
1525 
1526 	/* Check to see if we are recursing on a lock we already own. */
1527 	instance = find_instance(*lock_list, lock);
1528 	if (instance != NULL) {
1529 		instance->li_flags++;
1530 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1531 		    td->td_proc->p_pid, lock->lo_name,
1532 		    instance->li_flags & LI_RECURSEMASK);
1533 		instance->li_file = file;
1534 		instance->li_line = line;
1535 		return;
1536 	}
1537 
1538 	/* Update per-witness last file and line acquire. */
1539 	w->w_file = file;
1540 	w->w_line = line;
1541 
1542 	/* Find the next open lock instance in the list and fill it. */
1543 	lle = *lock_list;
1544 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1545 		lle = witness_lock_list_get();
1546 		if (lle == NULL)
1547 			return;
1548 		lle->ll_next = *lock_list;
1549 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1550 		    td->td_proc->p_pid, lle);
1551 		*lock_list = lle;
1552 	}
1553 	instance = &lle->ll_children[lle->ll_count++];
1554 	instance->li_lock = lock;
1555 	instance->li_line = line;
1556 	instance->li_file = file;
1557 	instance->li_flags = 0;
1558 	if ((flags & LOP_EXCLUSIVE) != 0)
1559 		instance->li_flags |= LI_EXCLUSIVE;
1560 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1561 		instance->li_flags |= LI_SLEEPABLE;
1562 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1563 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1564 }
1565 
1566 void
1567 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1568 {
1569 	struct lock_instance *instance;
1570 	struct lock_class *class;
1571 
1572 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1573 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1574 		return;
1575 	class = LOCK_CLASS(lock);
1576 	if (witness_watch) {
1577 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1578 			kassert_panic(
1579 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1580 			    class->lc_name, lock->lo_name,
1581 			    fixup_filename(file), line);
1582 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1583 			kassert_panic(
1584 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1585 			    class->lc_name, lock->lo_name,
1586 			    fixup_filename(file), line);
1587 	}
1588 	instance = find_instance(curthread->td_sleeplocks, lock);
1589 	if (instance == NULL) {
1590 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1591 		    class->lc_name, lock->lo_name,
1592 		    fixup_filename(file), line);
1593 		return;
1594 	}
1595 	if (witness_watch) {
1596 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1597 			kassert_panic(
1598 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1599 			    class->lc_name, lock->lo_name,
1600 			    fixup_filename(file), line);
1601 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1602 			kassert_panic(
1603 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1604 			    class->lc_name, lock->lo_name,
1605 			    instance->li_flags & LI_RECURSEMASK,
1606 			    fixup_filename(file), line);
1607 	}
1608 	instance->li_flags |= LI_EXCLUSIVE;
1609 }
1610 
1611 void
1612 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1613     int line)
1614 {
1615 	struct lock_instance *instance;
1616 	struct lock_class *class;
1617 
1618 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1619 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1620 		return;
1621 	class = LOCK_CLASS(lock);
1622 	if (witness_watch) {
1623 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1624 			kassert_panic(
1625 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1626 			    class->lc_name, lock->lo_name,
1627 			    fixup_filename(file), line);
1628 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1629 			kassert_panic(
1630 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1631 			    class->lc_name, lock->lo_name,
1632 			    fixup_filename(file), line);
1633 	}
1634 	instance = find_instance(curthread->td_sleeplocks, lock);
1635 	if (instance == NULL) {
1636 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1637 		    class->lc_name, lock->lo_name,
1638 		    fixup_filename(file), line);
1639 		return;
1640 	}
1641 	if (witness_watch) {
1642 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1643 			kassert_panic(
1644 			    "downgrade of shared lock (%s) %s @ %s:%d",
1645 			    class->lc_name, lock->lo_name,
1646 			    fixup_filename(file), line);
1647 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1648 			kassert_panic(
1649 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1650 			    class->lc_name, lock->lo_name,
1651 			    instance->li_flags & LI_RECURSEMASK,
1652 			    fixup_filename(file), line);
1653 	}
1654 	instance->li_flags &= ~LI_EXCLUSIVE;
1655 }
1656 
1657 void
1658 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1659 {
1660 	struct lock_list_entry **lock_list, *lle;
1661 	struct lock_instance *instance;
1662 	struct lock_class *class;
1663 	struct thread *td;
1664 	register_t s;
1665 	int i, j;
1666 
1667 	if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1668 		return;
1669 	td = curthread;
1670 	class = LOCK_CLASS(lock);
1671 
1672 	/* Find lock instance associated with this lock. */
1673 	if (class->lc_flags & LC_SLEEPLOCK)
1674 		lock_list = &td->td_sleeplocks;
1675 	else
1676 		lock_list = PCPU_PTR(spinlocks);
1677 	lle = *lock_list;
1678 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1679 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1680 			instance = &(*lock_list)->ll_children[i];
1681 			if (instance->li_lock == lock)
1682 				goto found;
1683 		}
1684 
1685 	/*
1686 	 * When disabling WITNESS through witness_watch we could end up in
1687 	 * having registered locks in the td_sleeplocks queue.
1688 	 * We have to make sure we flush these queues, so just search for
1689 	 * eventual register locks and remove them.
1690 	 */
1691 	if (witness_watch > 0) {
1692 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1693 		    lock->lo_name, fixup_filename(file), line);
1694 		return;
1695 	} else {
1696 		return;
1697 	}
1698 found:
1699 
1700 	/* First, check for shared/exclusive mismatches. */
1701 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1702 	    (flags & LOP_EXCLUSIVE) == 0) {
1703 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1704 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1705 		witness_output("while exclusively locked from %s:%d\n",
1706 		    fixup_filename(instance->li_file), instance->li_line);
1707 		kassert_panic("excl->ushare");
1708 	}
1709 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1710 	    (flags & LOP_EXCLUSIVE) != 0) {
1711 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1712 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1713 		witness_output("while share locked from %s:%d\n",
1714 		    fixup_filename(instance->li_file),
1715 		    instance->li_line);
1716 		kassert_panic("share->uexcl");
1717 	}
1718 	/* If we are recursed, unrecurse. */
1719 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1720 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1721 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1722 		    instance->li_flags);
1723 		instance->li_flags--;
1724 		return;
1725 	}
1726 	/* The lock is now being dropped, check for NORELEASE flag */
1727 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1728 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1729 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1730 		kassert_panic("lock marked norelease");
1731 	}
1732 
1733 	/* Otherwise, remove this item from the list. */
1734 	s = intr_disable();
1735 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1736 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1737 	    (*lock_list)->ll_count - 1);
1738 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1739 		(*lock_list)->ll_children[j] =
1740 		    (*lock_list)->ll_children[j + 1];
1741 	(*lock_list)->ll_count--;
1742 	intr_restore(s);
1743 
1744 	/*
1745 	 * In order to reduce contention on w_mtx, we want to keep always an
1746 	 * head object into lists so that frequent allocation from the
1747 	 * free witness pool (and subsequent locking) is avoided.
1748 	 * In order to maintain the current code simple, when the head
1749 	 * object is totally unloaded it means also that we do not have
1750 	 * further objects in the list, so the list ownership needs to be
1751 	 * hand over to another object if the current head needs to be freed.
1752 	 */
1753 	if ((*lock_list)->ll_count == 0) {
1754 		if (*lock_list == lle) {
1755 			if (lle->ll_next == NULL)
1756 				return;
1757 		} else
1758 			lle = *lock_list;
1759 		*lock_list = lle->ll_next;
1760 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1761 		    td->td_proc->p_pid, lle);
1762 		witness_lock_list_free(lle);
1763 	}
1764 }
1765 
1766 void
1767 witness_thread_exit(struct thread *td)
1768 {
1769 	struct lock_list_entry *lle;
1770 	int i, n;
1771 
1772 	lle = td->td_sleeplocks;
1773 	if (lle == NULL || KERNEL_PANICKED())
1774 		return;
1775 	if (lle->ll_count != 0) {
1776 		for (n = 0; lle != NULL; lle = lle->ll_next)
1777 			for (i = lle->ll_count - 1; i >= 0; i--) {
1778 				if (n == 0)
1779 					witness_output(
1780 		    "Thread %p exiting with the following locks held:\n", td);
1781 				n++;
1782 				witness_list_lock(&lle->ll_children[i],
1783 				    witness_output);
1784 
1785 			}
1786 		kassert_panic(
1787 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1788 	}
1789 	witness_lock_list_free(lle);
1790 }
1791 
1792 /*
1793  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1794  * exempt Giant and sleepable locks from the checks as well.  If any
1795  * non-exempt locks are held, then a supplied message is printed to the
1796  * output channel along with a list of the offending locks.  If indicated in the
1797  * flags then a failure results in a panic as well.
1798  */
1799 int
1800 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1801 {
1802 	struct lock_list_entry *lock_list, *lle;
1803 	struct lock_instance *lock1;
1804 	struct thread *td;
1805 	va_list ap;
1806 	int i, n;
1807 
1808 	if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1809 		return (0);
1810 	n = 0;
1811 	td = curthread;
1812 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1813 		for (i = lle->ll_count - 1; i >= 0; i--) {
1814 			lock1 = &lle->ll_children[i];
1815 			if (lock1->li_lock == lock)
1816 				continue;
1817 			if (flags & WARN_GIANTOK &&
1818 			    lock1->li_lock == &Giant.lock_object)
1819 				continue;
1820 			if (flags & WARN_SLEEPOK &&
1821 			    (lock1->li_flags & LI_SLEEPABLE) != 0)
1822 				continue;
1823 			if (n == 0) {
1824 				va_start(ap, fmt);
1825 				vprintf(fmt, ap);
1826 				va_end(ap);
1827 				printf(" with the following %slocks held:\n",
1828 				    (flags & WARN_SLEEPOK) != 0 ?
1829 				    "non-sleepable " : "");
1830 			}
1831 			n++;
1832 			witness_list_lock(lock1, printf);
1833 		}
1834 
1835 	/*
1836 	 * Pin the thread in order to avoid problems with thread migration.
1837 	 * Once that all verifies are passed about spinlocks ownership,
1838 	 * the thread is in a safe path and it can be unpinned.
1839 	 */
1840 	sched_pin();
1841 	lock_list = PCPU_GET(spinlocks);
1842 	if (lock_list != NULL && lock_list->ll_count != 0) {
1843 		sched_unpin();
1844 
1845 		/*
1846 		 * We should only have one spinlock and as long as
1847 		 * the flags cannot match for this locks class,
1848 		 * check if the first spinlock is the one curthread
1849 		 * should hold.
1850 		 */
1851 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1852 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1853 		    lock1->li_lock == lock && n == 0)
1854 			return (0);
1855 
1856 		va_start(ap, fmt);
1857 		vprintf(fmt, ap);
1858 		va_end(ap);
1859 		printf(" with the following %slocks held:\n",
1860 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1861 		n += witness_list_locks(&lock_list, printf);
1862 	} else
1863 		sched_unpin();
1864 	if (flags & WARN_PANIC && n)
1865 		kassert_panic("%s", __func__);
1866 	else
1867 		witness_debugger(n, __func__);
1868 	return (n);
1869 }
1870 
1871 const char *
1872 witness_file(struct lock_object *lock)
1873 {
1874 	struct witness *w;
1875 
1876 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1877 		return ("?");
1878 	w = lock->lo_witness;
1879 	return (w->w_file);
1880 }
1881 
1882 int
1883 witness_line(struct lock_object *lock)
1884 {
1885 	struct witness *w;
1886 
1887 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1888 		return (0);
1889 	w = lock->lo_witness;
1890 	return (w->w_line);
1891 }
1892 
1893 static struct witness *
1894 enroll(const char *description, struct lock_class *lock_class)
1895 {
1896 	struct witness *w;
1897 
1898 	MPASS(description != NULL);
1899 
1900 	if (witness_watch == -1 || KERNEL_PANICKED())
1901 		return (NULL);
1902 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1903 		if (witness_skipspin)
1904 			return (NULL);
1905 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1906 		kassert_panic("lock class %s is not sleep or spin",
1907 		    lock_class->lc_name);
1908 		return (NULL);
1909 	}
1910 
1911 	mtx_lock_spin(&w_mtx);
1912 	w = witness_hash_get(description);
1913 	if (w)
1914 		goto found;
1915 	if ((w = witness_get()) == NULL)
1916 		return (NULL);
1917 	MPASS(strlen(description) < MAX_W_NAME);
1918 	strcpy(w->w_name, description);
1919 	w->w_class = lock_class;
1920 	w->w_refcount = 1;
1921 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1922 	if (lock_class->lc_flags & LC_SPINLOCK) {
1923 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1924 		w_spin_cnt++;
1925 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1926 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1927 		w_sleep_cnt++;
1928 	}
1929 
1930 	/* Insert new witness into the hash */
1931 	witness_hash_put(w);
1932 	witness_increment_graph_generation();
1933 	mtx_unlock_spin(&w_mtx);
1934 	return (w);
1935 found:
1936 	w->w_refcount++;
1937 	if (w->w_refcount == 1)
1938 		w->w_class = lock_class;
1939 	mtx_unlock_spin(&w_mtx);
1940 	if (lock_class != w->w_class)
1941 		kassert_panic(
1942 		    "lock (%s) %s does not match earlier (%s) lock",
1943 		    description, lock_class->lc_name,
1944 		    w->w_class->lc_name);
1945 	return (w);
1946 }
1947 
1948 static void
1949 depart(struct witness *w)
1950 {
1951 
1952 	MPASS(w->w_refcount == 0);
1953 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1954 		w_sleep_cnt--;
1955 	} else {
1956 		w_spin_cnt--;
1957 	}
1958 	/*
1959 	 * Set file to NULL as it may point into a loadable module.
1960 	 */
1961 	w->w_file = NULL;
1962 	w->w_line = 0;
1963 	witness_increment_graph_generation();
1964 }
1965 
1966 static void
1967 adopt(struct witness *parent, struct witness *child)
1968 {
1969 	int pi, ci, i, j;
1970 
1971 	if (witness_cold == 0)
1972 		mtx_assert(&w_mtx, MA_OWNED);
1973 
1974 	/* If the relationship is already known, there's no work to be done. */
1975 	if (isitmychild(parent, child))
1976 		return;
1977 
1978 	/* When the structure of the graph changes, bump up the generation. */
1979 	witness_increment_graph_generation();
1980 
1981 	/*
1982 	 * The hard part ... create the direct relationship, then propagate all
1983 	 * indirect relationships.
1984 	 */
1985 	pi = parent->w_index;
1986 	ci = child->w_index;
1987 	WITNESS_INDEX_ASSERT(pi);
1988 	WITNESS_INDEX_ASSERT(ci);
1989 	MPASS(pi != ci);
1990 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1991 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1992 
1993 	/*
1994 	 * If parent was not already an ancestor of child,
1995 	 * then we increment the descendant and ancestor counters.
1996 	 */
1997 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1998 		parent->w_num_descendants++;
1999 		child->w_num_ancestors++;
2000 	}
2001 
2002 	/*
2003 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2004 	 * an ancestor of 'pi' during this loop.
2005 	 */
2006 	for (i = 1; i <= w_max_used_index; i++) {
2007 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2008 		    (i != pi))
2009 			continue;
2010 
2011 		/* Find each descendant of 'i' and mark it as a descendant. */
2012 		for (j = 1; j <= w_max_used_index; j++) {
2013 			/*
2014 			 * Skip children that are already marked as
2015 			 * descendants of 'i'.
2016 			 */
2017 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2018 				continue;
2019 
2020 			/*
2021 			 * We are only interested in descendants of 'ci'. Note
2022 			 * that 'ci' itself is counted as a descendant of 'ci'.
2023 			 */
2024 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2025 			    (j != ci))
2026 				continue;
2027 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2028 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2029 			w_data[i].w_num_descendants++;
2030 			w_data[j].w_num_ancestors++;
2031 
2032 			/*
2033 			 * Make sure we aren't marking a node as both an
2034 			 * ancestor and descendant. We should have caught
2035 			 * this as a lock order reversal earlier.
2036 			 */
2037 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2038 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2039 				printf("witness rmatrix paradox! [%d][%d]=%d "
2040 				    "both ancestor and descendant\n",
2041 				    i, j, w_rmatrix[i][j]);
2042 				kdb_backtrace();
2043 				printf("Witness disabled.\n");
2044 				witness_watch = -1;
2045 			}
2046 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2047 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2048 				printf("witness rmatrix paradox! [%d][%d]=%d "
2049 				    "both ancestor and descendant\n",
2050 				    j, i, w_rmatrix[j][i]);
2051 				kdb_backtrace();
2052 				printf("Witness disabled.\n");
2053 				witness_watch = -1;
2054 			}
2055 		}
2056 	}
2057 }
2058 
2059 static void
2060 itismychild(struct witness *parent, struct witness *child)
2061 {
2062 	int unlocked;
2063 
2064 	MPASS(child != NULL && parent != NULL);
2065 	if (witness_cold == 0)
2066 		mtx_assert(&w_mtx, MA_OWNED);
2067 
2068 	if (!witness_lock_type_equal(parent, child)) {
2069 		if (witness_cold == 0) {
2070 			unlocked = 1;
2071 			mtx_unlock_spin(&w_mtx);
2072 		} else {
2073 			unlocked = 0;
2074 		}
2075 		kassert_panic(
2076 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2077 		    "the same lock type", __func__, parent->w_name,
2078 		    parent->w_class->lc_name, child->w_name,
2079 		    child->w_class->lc_name);
2080 		if (unlocked)
2081 			mtx_lock_spin(&w_mtx);
2082 	}
2083 	adopt(parent, child);
2084 }
2085 
2086 /*
2087  * Generic code for the isitmy*() functions. The rmask parameter is the
2088  * expected relationship of w1 to w2.
2089  */
2090 static int
2091 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2092 {
2093 	unsigned char r1, r2;
2094 	int i1, i2;
2095 
2096 	i1 = w1->w_index;
2097 	i2 = w2->w_index;
2098 	WITNESS_INDEX_ASSERT(i1);
2099 	WITNESS_INDEX_ASSERT(i2);
2100 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2101 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2102 
2103 	/* The flags on one better be the inverse of the flags on the other */
2104 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2105 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2106 		/* Don't squawk if we're potentially racing with an update. */
2107 		if (!mtx_owned(&w_mtx))
2108 			return (0);
2109 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2110 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2111 		    "w_rmatrix[%d][%d] == %hhx\n",
2112 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2113 		    i2, i1, r2);
2114 		kdb_backtrace();
2115 		printf("Witness disabled.\n");
2116 		witness_watch = -1;
2117 	}
2118 	return (r1 & rmask);
2119 }
2120 
2121 /*
2122  * Checks if @child is a direct child of @parent.
2123  */
2124 static int
2125 isitmychild(struct witness *parent, struct witness *child)
2126 {
2127 
2128 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2129 }
2130 
2131 /*
2132  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2133  */
2134 static int
2135 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2136 {
2137 
2138 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2139 	    __func__));
2140 }
2141 
2142 static int
2143 blessed(struct witness *w1, struct witness *w2)
2144 {
2145 	int i;
2146 	struct witness_blessed *b;
2147 
2148 	for (i = 0; i < nitems(blessed_list); i++) {
2149 		b = &blessed_list[i];
2150 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2151 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2152 				return (1);
2153 			continue;
2154 		}
2155 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2156 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2157 				return (1);
2158 	}
2159 	return (0);
2160 }
2161 
2162 static struct witness *
2163 witness_get(void)
2164 {
2165 	struct witness *w;
2166 	int index;
2167 
2168 	if (witness_cold == 0)
2169 		mtx_assert(&w_mtx, MA_OWNED);
2170 
2171 	if (witness_watch == -1) {
2172 		mtx_unlock_spin(&w_mtx);
2173 		return (NULL);
2174 	}
2175 	if (STAILQ_EMPTY(&w_free)) {
2176 		witness_watch = -1;
2177 		mtx_unlock_spin(&w_mtx);
2178 		printf("WITNESS: unable to allocate a new witness object\n");
2179 		return (NULL);
2180 	}
2181 	w = STAILQ_FIRST(&w_free);
2182 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2183 	w_free_cnt--;
2184 	index = w->w_index;
2185 	MPASS(index > 0 && index == w_max_used_index+1 &&
2186 	    index < witness_count);
2187 	bzero(w, sizeof(*w));
2188 	w->w_index = index;
2189 	if (index > w_max_used_index)
2190 		w_max_used_index = index;
2191 	return (w);
2192 }
2193 
2194 static void
2195 witness_free(struct witness *w)
2196 {
2197 
2198 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2199 	w_free_cnt++;
2200 }
2201 
2202 static struct lock_list_entry *
2203 witness_lock_list_get(void)
2204 {
2205 	struct lock_list_entry *lle;
2206 
2207 	if (witness_watch == -1)
2208 		return (NULL);
2209 	mtx_lock_spin(&w_mtx);
2210 	lle = w_lock_list_free;
2211 	if (lle == NULL) {
2212 		witness_watch = -1;
2213 		mtx_unlock_spin(&w_mtx);
2214 		printf("%s: witness exhausted\n", __func__);
2215 		return (NULL);
2216 	}
2217 	w_lock_list_free = lle->ll_next;
2218 	mtx_unlock_spin(&w_mtx);
2219 	bzero(lle, sizeof(*lle));
2220 	return (lle);
2221 }
2222 
2223 static void
2224 witness_lock_list_free(struct lock_list_entry *lle)
2225 {
2226 
2227 	mtx_lock_spin(&w_mtx);
2228 	lle->ll_next = w_lock_list_free;
2229 	w_lock_list_free = lle;
2230 	mtx_unlock_spin(&w_mtx);
2231 }
2232 
2233 static struct lock_instance *
2234 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2235 {
2236 	struct lock_list_entry *lle;
2237 	struct lock_instance *instance;
2238 	int i;
2239 
2240 	for (lle = list; lle != NULL; lle = lle->ll_next)
2241 		for (i = lle->ll_count - 1; i >= 0; i--) {
2242 			instance = &lle->ll_children[i];
2243 			if (instance->li_lock == lock)
2244 				return (instance);
2245 		}
2246 	return (NULL);
2247 }
2248 
2249 static void
2250 witness_list_lock(struct lock_instance *instance,
2251     int (*prnt)(const char *fmt, ...))
2252 {
2253 	struct lock_object *lock;
2254 
2255 	lock = instance->li_lock;
2256 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2257 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2258 	if (lock->lo_witness->w_name != lock->lo_name)
2259 		prnt(" (%s)", lock->lo_witness->w_name);
2260 	prnt(" r = %d (%p) locked @ %s:%d\n",
2261 	    instance->li_flags & LI_RECURSEMASK, lock,
2262 	    fixup_filename(instance->li_file), instance->li_line);
2263 }
2264 
2265 static int
2266 witness_output(const char *fmt, ...)
2267 {
2268 	va_list ap;
2269 	int ret;
2270 
2271 	va_start(ap, fmt);
2272 	ret = witness_voutput(fmt, ap);
2273 	va_end(ap);
2274 	return (ret);
2275 }
2276 
2277 static int
2278 witness_voutput(const char *fmt, va_list ap)
2279 {
2280 	int ret;
2281 
2282 	ret = 0;
2283 	switch (witness_channel) {
2284 	case WITNESS_CONSOLE:
2285 		ret = vprintf(fmt, ap);
2286 		break;
2287 	case WITNESS_LOG:
2288 		vlog(LOG_NOTICE, fmt, ap);
2289 		break;
2290 	case WITNESS_NONE:
2291 		break;
2292 	}
2293 	return (ret);
2294 }
2295 
2296 #ifdef DDB
2297 static int
2298 witness_thread_has_locks(struct thread *td)
2299 {
2300 
2301 	if (td->td_sleeplocks == NULL)
2302 		return (0);
2303 	return (td->td_sleeplocks->ll_count != 0);
2304 }
2305 
2306 static int
2307 witness_proc_has_locks(struct proc *p)
2308 {
2309 	struct thread *td;
2310 
2311 	FOREACH_THREAD_IN_PROC(p, td) {
2312 		if (witness_thread_has_locks(td))
2313 			return (1);
2314 	}
2315 	return (0);
2316 }
2317 #endif
2318 
2319 int
2320 witness_list_locks(struct lock_list_entry **lock_list,
2321     int (*prnt)(const char *fmt, ...))
2322 {
2323 	struct lock_list_entry *lle;
2324 	int i, nheld;
2325 
2326 	nheld = 0;
2327 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2328 		for (i = lle->ll_count - 1; i >= 0; i--) {
2329 			witness_list_lock(&lle->ll_children[i], prnt);
2330 			nheld++;
2331 		}
2332 	return (nheld);
2333 }
2334 
2335 /*
2336  * This is a bit risky at best.  We call this function when we have timed
2337  * out acquiring a spin lock, and we assume that the other CPU is stuck
2338  * with this lock held.  So, we go groveling around in the other CPU's
2339  * per-cpu data to try to find the lock instance for this spin lock to
2340  * see when it was last acquired.
2341  */
2342 void
2343 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2344     int (*prnt)(const char *fmt, ...))
2345 {
2346 	struct lock_instance *instance;
2347 	struct pcpu *pc;
2348 
2349 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2350 		return;
2351 	pc = pcpu_find(owner->td_oncpu);
2352 	instance = find_instance(pc->pc_spinlocks, lock);
2353 	if (instance != NULL)
2354 		witness_list_lock(instance, prnt);
2355 }
2356 
2357 void
2358 witness_save(struct lock_object *lock, const char **filep, int *linep)
2359 {
2360 	struct lock_list_entry *lock_list;
2361 	struct lock_instance *instance;
2362 	struct lock_class *class;
2363 
2364 	/*
2365 	 * This function is used independently in locking code to deal with
2366 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2367 	 * is gone.
2368 	 */
2369 	if (SCHEDULER_STOPPED())
2370 		return;
2371 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2372 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2373 		return;
2374 	class = LOCK_CLASS(lock);
2375 	if (class->lc_flags & LC_SLEEPLOCK)
2376 		lock_list = curthread->td_sleeplocks;
2377 	else {
2378 		if (witness_skipspin)
2379 			return;
2380 		lock_list = PCPU_GET(spinlocks);
2381 	}
2382 	instance = find_instance(lock_list, lock);
2383 	if (instance == NULL) {
2384 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2385 		    class->lc_name, lock->lo_name);
2386 		return;
2387 	}
2388 	*filep = instance->li_file;
2389 	*linep = instance->li_line;
2390 }
2391 
2392 void
2393 witness_restore(struct lock_object *lock, const char *file, int line)
2394 {
2395 	struct lock_list_entry *lock_list;
2396 	struct lock_instance *instance;
2397 	struct lock_class *class;
2398 
2399 	/*
2400 	 * This function is used independently in locking code to deal with
2401 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2402 	 * is gone.
2403 	 */
2404 	if (SCHEDULER_STOPPED())
2405 		return;
2406 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2407 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2408 		return;
2409 	class = LOCK_CLASS(lock);
2410 	if (class->lc_flags & LC_SLEEPLOCK)
2411 		lock_list = curthread->td_sleeplocks;
2412 	else {
2413 		if (witness_skipspin)
2414 			return;
2415 		lock_list = PCPU_GET(spinlocks);
2416 	}
2417 	instance = find_instance(lock_list, lock);
2418 	if (instance == NULL)
2419 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2420 		    class->lc_name, lock->lo_name);
2421 	lock->lo_witness->w_file = file;
2422 	lock->lo_witness->w_line = line;
2423 	if (instance == NULL)
2424 		return;
2425 	instance->li_file = file;
2426 	instance->li_line = line;
2427 }
2428 
2429 void
2430 witness_assert(const struct lock_object *lock, int flags, const char *file,
2431     int line)
2432 {
2433 #ifdef INVARIANT_SUPPORT
2434 	struct lock_instance *instance;
2435 	struct lock_class *class;
2436 
2437 	if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2438 		return;
2439 	class = LOCK_CLASS(lock);
2440 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2441 		instance = find_instance(curthread->td_sleeplocks, lock);
2442 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2443 		instance = find_instance(PCPU_GET(spinlocks), lock);
2444 	else {
2445 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2446 		    class->lc_name, lock->lo_name);
2447 		return;
2448 	}
2449 	switch (flags) {
2450 	case LA_UNLOCKED:
2451 		if (instance != NULL)
2452 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2453 			    class->lc_name, lock->lo_name,
2454 			    fixup_filename(file), line);
2455 		break;
2456 	case LA_LOCKED:
2457 	case LA_LOCKED | LA_RECURSED:
2458 	case LA_LOCKED | LA_NOTRECURSED:
2459 	case LA_SLOCKED:
2460 	case LA_SLOCKED | LA_RECURSED:
2461 	case LA_SLOCKED | LA_NOTRECURSED:
2462 	case LA_XLOCKED:
2463 	case LA_XLOCKED | LA_RECURSED:
2464 	case LA_XLOCKED | LA_NOTRECURSED:
2465 		if (instance == NULL) {
2466 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2467 			    class->lc_name, lock->lo_name,
2468 			    fixup_filename(file), line);
2469 			break;
2470 		}
2471 		if ((flags & LA_XLOCKED) != 0 &&
2472 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2473 			kassert_panic(
2474 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2475 			    class->lc_name, lock->lo_name,
2476 			    fixup_filename(file), line);
2477 		if ((flags & LA_SLOCKED) != 0 &&
2478 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2479 			kassert_panic(
2480 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2481 			    class->lc_name, lock->lo_name,
2482 			    fixup_filename(file), line);
2483 		if ((flags & LA_RECURSED) != 0 &&
2484 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2485 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2486 			    class->lc_name, lock->lo_name,
2487 			    fixup_filename(file), line);
2488 		if ((flags & LA_NOTRECURSED) != 0 &&
2489 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2490 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2491 			    class->lc_name, lock->lo_name,
2492 			    fixup_filename(file), line);
2493 		break;
2494 	default:
2495 		kassert_panic("Invalid lock assertion at %s:%d.",
2496 		    fixup_filename(file), line);
2497 	}
2498 #endif	/* INVARIANT_SUPPORT */
2499 }
2500 
2501 static void
2502 witness_setflag(struct lock_object *lock, int flag, int set)
2503 {
2504 	struct lock_list_entry *lock_list;
2505 	struct lock_instance *instance;
2506 	struct lock_class *class;
2507 
2508 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2509 		return;
2510 	class = LOCK_CLASS(lock);
2511 	if (class->lc_flags & LC_SLEEPLOCK)
2512 		lock_list = curthread->td_sleeplocks;
2513 	else {
2514 		if (witness_skipspin)
2515 			return;
2516 		lock_list = PCPU_GET(spinlocks);
2517 	}
2518 	instance = find_instance(lock_list, lock);
2519 	if (instance == NULL) {
2520 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2521 		    class->lc_name, lock->lo_name);
2522 		return;
2523 	}
2524 
2525 	if (set)
2526 		instance->li_flags |= flag;
2527 	else
2528 		instance->li_flags &= ~flag;
2529 }
2530 
2531 void
2532 witness_norelease(struct lock_object *lock)
2533 {
2534 
2535 	witness_setflag(lock, LI_NORELEASE, 1);
2536 }
2537 
2538 void
2539 witness_releaseok(struct lock_object *lock)
2540 {
2541 
2542 	witness_setflag(lock, LI_NORELEASE, 0);
2543 }
2544 
2545 #ifdef DDB
2546 static void
2547 witness_ddb_list(struct thread *td)
2548 {
2549 
2550 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2551 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2552 
2553 	if (witness_watch < 1)
2554 		return;
2555 
2556 	witness_list_locks(&td->td_sleeplocks, db_printf);
2557 
2558 	/*
2559 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2560 	 * if td is currently executing on some other CPU and holds spin locks
2561 	 * as we won't display those locks.  If we had a MI way of getting
2562 	 * the per-cpu data for a given cpu then we could use
2563 	 * td->td_oncpu to get the list of spinlocks for this thread
2564 	 * and "fix" this.
2565 	 *
2566 	 * That still wouldn't really fix this unless we locked the scheduler
2567 	 * lock or stopped the other CPU to make sure it wasn't changing the
2568 	 * list out from under us.  It is probably best to just not try to
2569 	 * handle threads on other CPU's for now.
2570 	 */
2571 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2572 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2573 }
2574 
2575 DB_SHOW_COMMAND(locks, db_witness_list)
2576 {
2577 	struct thread *td;
2578 
2579 	if (have_addr)
2580 		td = db_lookup_thread(addr, true);
2581 	else
2582 		td = kdb_thread;
2583 	witness_ddb_list(td);
2584 }
2585 
2586 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2587 {
2588 	struct thread *td;
2589 	struct proc *p;
2590 
2591 	/*
2592 	 * It would be nice to list only threads and processes that actually
2593 	 * held sleep locks, but that information is currently not exported
2594 	 * by WITNESS.
2595 	 */
2596 	FOREACH_PROC_IN_SYSTEM(p) {
2597 		if (!witness_proc_has_locks(p))
2598 			continue;
2599 		FOREACH_THREAD_IN_PROC(p, td) {
2600 			if (!witness_thread_has_locks(td))
2601 				continue;
2602 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2603 			    p->p_comm, td, td->td_tid);
2604 			witness_ddb_list(td);
2605 			if (db_pager_quit)
2606 				return;
2607 		}
2608 	}
2609 }
2610 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2611 
2612 DB_SHOW_COMMAND(witness, db_witness_display)
2613 {
2614 
2615 	witness_ddb_display(db_printf);
2616 }
2617 #endif
2618 
2619 static void
2620 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2621 {
2622 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2623 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2624 	int generation, i, j;
2625 
2626 	tmp_data1 = NULL;
2627 	tmp_data2 = NULL;
2628 	tmp_w1 = NULL;
2629 	tmp_w2 = NULL;
2630 
2631 	/* Allocate and init temporary storage space. */
2632 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2633 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2634 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2635 	    M_WAITOK | M_ZERO);
2636 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2637 	    M_WAITOK | M_ZERO);
2638 	stack_zero(&tmp_data1->wlod_stack);
2639 	stack_zero(&tmp_data2->wlod_stack);
2640 
2641 restart:
2642 	mtx_lock_spin(&w_mtx);
2643 	generation = w_generation;
2644 	mtx_unlock_spin(&w_mtx);
2645 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2646 	    w_lohash.wloh_count);
2647 	for (i = 1; i < w_max_used_index; i++) {
2648 		mtx_lock_spin(&w_mtx);
2649 		if (generation != w_generation) {
2650 			mtx_unlock_spin(&w_mtx);
2651 
2652 			/* The graph has changed, try again. */
2653 			*oldidx = 0;
2654 			sbuf_clear(sb);
2655 			goto restart;
2656 		}
2657 
2658 		w1 = &w_data[i];
2659 		if (w1->w_reversed == 0) {
2660 			mtx_unlock_spin(&w_mtx);
2661 			continue;
2662 		}
2663 
2664 		/* Copy w1 locally so we can release the spin lock. */
2665 		*tmp_w1 = *w1;
2666 		mtx_unlock_spin(&w_mtx);
2667 
2668 		if (tmp_w1->w_reversed == 0)
2669 			continue;
2670 		for (j = 1; j < w_max_used_index; j++) {
2671 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2672 				continue;
2673 
2674 			mtx_lock_spin(&w_mtx);
2675 			if (generation != w_generation) {
2676 				mtx_unlock_spin(&w_mtx);
2677 
2678 				/* The graph has changed, try again. */
2679 				*oldidx = 0;
2680 				sbuf_clear(sb);
2681 				goto restart;
2682 			}
2683 
2684 			w2 = &w_data[j];
2685 			data1 = witness_lock_order_get(w1, w2);
2686 			data2 = witness_lock_order_get(w2, w1);
2687 
2688 			/*
2689 			 * Copy information locally so we can release the
2690 			 * spin lock.
2691 			 */
2692 			*tmp_w2 = *w2;
2693 
2694 			if (data1) {
2695 				stack_zero(&tmp_data1->wlod_stack);
2696 				stack_copy(&data1->wlod_stack,
2697 				    &tmp_data1->wlod_stack);
2698 			}
2699 			if (data2 && data2 != data1) {
2700 				stack_zero(&tmp_data2->wlod_stack);
2701 				stack_copy(&data2->wlod_stack,
2702 				    &tmp_data2->wlod_stack);
2703 			}
2704 			mtx_unlock_spin(&w_mtx);
2705 
2706 			if (blessed(tmp_w1, tmp_w2))
2707 				continue;
2708 
2709 			sbuf_printf(sb,
2710 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2711 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2712 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2713 			if (data1) {
2714 				sbuf_printf(sb,
2715 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2716 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2717 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2718 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2719 				sbuf_printf(sb, "\n");
2720 			}
2721 			if (data2 && data2 != data1) {
2722 				sbuf_printf(sb,
2723 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2724 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2725 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2726 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2727 				sbuf_printf(sb, "\n");
2728 			}
2729 		}
2730 	}
2731 	mtx_lock_spin(&w_mtx);
2732 	if (generation != w_generation) {
2733 		mtx_unlock_spin(&w_mtx);
2734 
2735 		/*
2736 		 * The graph changed while we were printing stack data,
2737 		 * try again.
2738 		 */
2739 		*oldidx = 0;
2740 		sbuf_clear(sb);
2741 		goto restart;
2742 	}
2743 	mtx_unlock_spin(&w_mtx);
2744 
2745 	/* Free temporary storage space. */
2746 	free(tmp_data1, M_TEMP);
2747 	free(tmp_data2, M_TEMP);
2748 	free(tmp_w1, M_TEMP);
2749 	free(tmp_w2, M_TEMP);
2750 }
2751 
2752 static int
2753 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2754 {
2755 	struct sbuf *sb;
2756 	int error;
2757 
2758 	if (witness_watch < 1) {
2759 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2760 		return (error);
2761 	}
2762 	if (witness_cold) {
2763 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2764 		return (error);
2765 	}
2766 	error = 0;
2767 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2768 	if (sb == NULL)
2769 		return (ENOMEM);
2770 
2771 	sbuf_print_witness_badstacks(sb, &req->oldidx);
2772 
2773 	sbuf_finish(sb);
2774 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2775 	sbuf_delete(sb);
2776 
2777 	return (error);
2778 }
2779 
2780 #ifdef DDB
2781 static int
2782 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2783 {
2784 
2785 	return (db_printf("%.*s", len, data));
2786 }
2787 
2788 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2789 {
2790 	struct sbuf sb;
2791 	char buffer[128];
2792 	size_t dummy;
2793 
2794 	sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2795 	sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2796 	sbuf_print_witness_badstacks(&sb, &dummy);
2797 	sbuf_finish(&sb);
2798 }
2799 #endif
2800 
2801 static int
2802 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2803 {
2804 	static const struct {
2805 		enum witness_channel channel;
2806 		const char *name;
2807 	} channels[] = {
2808 		{ WITNESS_CONSOLE, "console" },
2809 		{ WITNESS_LOG, "log" },
2810 		{ WITNESS_NONE, "none" },
2811 	};
2812 	char buf[16];
2813 	u_int i;
2814 	int error;
2815 
2816 	buf[0] = '\0';
2817 	for (i = 0; i < nitems(channels); i++)
2818 		if (witness_channel == channels[i].channel) {
2819 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2820 			break;
2821 		}
2822 
2823 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2824 	if (error != 0 || req->newptr == NULL)
2825 		return (error);
2826 
2827 	error = EINVAL;
2828 	for (i = 0; i < nitems(channels); i++)
2829 		if (strcmp(channels[i].name, buf) == 0) {
2830 			witness_channel = channels[i].channel;
2831 			error = 0;
2832 			break;
2833 		}
2834 	return (error);
2835 }
2836 
2837 static int
2838 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2839 {
2840 	struct witness *w;
2841 	struct sbuf *sb;
2842 	int error;
2843 
2844 #ifdef __i386__
2845 	error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2846 	return (error);
2847 #endif
2848 
2849 	if (witness_watch < 1) {
2850 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2851 		return (error);
2852 	}
2853 	if (witness_cold) {
2854 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2855 		return (error);
2856 	}
2857 	error = 0;
2858 
2859 	error = sysctl_wire_old_buffer(req, 0);
2860 	if (error != 0)
2861 		return (error);
2862 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2863 	if (sb == NULL)
2864 		return (ENOMEM);
2865 	sbuf_printf(sb, "\n");
2866 
2867 	mtx_lock_spin(&w_mtx);
2868 	STAILQ_FOREACH(w, &w_all, w_list)
2869 		w->w_displayed = 0;
2870 	STAILQ_FOREACH(w, &w_all, w_list)
2871 		witness_add_fullgraph(sb, w);
2872 	mtx_unlock_spin(&w_mtx);
2873 
2874 	/*
2875 	 * Close the sbuf and return to userland.
2876 	 */
2877 	error = sbuf_finish(sb);
2878 	sbuf_delete(sb);
2879 
2880 	return (error);
2881 }
2882 
2883 static int
2884 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2885 {
2886 	int error, value;
2887 
2888 	value = witness_watch;
2889 	error = sysctl_handle_int(oidp, &value, 0, req);
2890 	if (error != 0 || req->newptr == NULL)
2891 		return (error);
2892 	if (value > 1 || value < -1 ||
2893 	    (witness_watch == -1 && value != witness_watch))
2894 		return (EINVAL);
2895 	witness_watch = value;
2896 	return (0);
2897 }
2898 
2899 static void
2900 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2901 {
2902 	int i;
2903 
2904 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2905 		return;
2906 	w->w_displayed = 1;
2907 
2908 	WITNESS_INDEX_ASSERT(w->w_index);
2909 	for (i = 1; i <= w_max_used_index; i++) {
2910 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2911 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2912 			    w_data[i].w_name);
2913 			witness_add_fullgraph(sb, &w_data[i]);
2914 		}
2915 	}
2916 }
2917 
2918 /*
2919  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2920  * interprets the key as a string and reads until the null
2921  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2922  * hash value computed from the key.
2923  */
2924 static uint32_t
2925 witness_hash_djb2(const uint8_t *key, uint32_t size)
2926 {
2927 	unsigned int hash = 5381;
2928 	int i;
2929 
2930 	/* hash = hash * 33 + key[i] */
2931 	if (size)
2932 		for (i = 0; i < size; i++)
2933 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2934 	else
2935 		for (i = 0; key[i] != 0; i++)
2936 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2937 
2938 	return (hash);
2939 }
2940 
2941 /*
2942  * Initializes the two witness hash tables. Called exactly once from
2943  * witness_initialize().
2944  */
2945 static void
2946 witness_init_hash_tables(void)
2947 {
2948 	int i;
2949 
2950 	MPASS(witness_cold);
2951 
2952 	/* Initialize the hash tables. */
2953 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2954 		w_hash.wh_array[i] = NULL;
2955 
2956 	w_hash.wh_size = WITNESS_HASH_SIZE;
2957 	w_hash.wh_count = 0;
2958 
2959 	/* Initialize the lock order data hash. */
2960 	w_lofree = NULL;
2961 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2962 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2963 		w_lodata[i].wlod_next = w_lofree;
2964 		w_lofree = &w_lodata[i];
2965 	}
2966 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2967 	w_lohash.wloh_count = 0;
2968 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2969 		w_lohash.wloh_array[i] = NULL;
2970 }
2971 
2972 static struct witness *
2973 witness_hash_get(const char *key)
2974 {
2975 	struct witness *w;
2976 	uint32_t hash;
2977 
2978 	MPASS(key != NULL);
2979 	if (witness_cold == 0)
2980 		mtx_assert(&w_mtx, MA_OWNED);
2981 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2982 	w = w_hash.wh_array[hash];
2983 	while (w != NULL) {
2984 		if (strcmp(w->w_name, key) == 0)
2985 			goto out;
2986 		w = w->w_hash_next;
2987 	}
2988 
2989 out:
2990 	return (w);
2991 }
2992 
2993 static void
2994 witness_hash_put(struct witness *w)
2995 {
2996 	uint32_t hash;
2997 
2998 	MPASS(w != NULL);
2999 	MPASS(w->w_name != NULL);
3000 	if (witness_cold == 0)
3001 		mtx_assert(&w_mtx, MA_OWNED);
3002 	KASSERT(witness_hash_get(w->w_name) == NULL,
3003 	    ("%s: trying to add a hash entry that already exists!", __func__));
3004 	KASSERT(w->w_hash_next == NULL,
3005 	    ("%s: w->w_hash_next != NULL", __func__));
3006 
3007 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3008 	w->w_hash_next = w_hash.wh_array[hash];
3009 	w_hash.wh_array[hash] = w;
3010 	w_hash.wh_count++;
3011 }
3012 
3013 static struct witness_lock_order_data *
3014 witness_lock_order_get(struct witness *parent, struct witness *child)
3015 {
3016 	struct witness_lock_order_data *data = NULL;
3017 	struct witness_lock_order_key key;
3018 	unsigned int hash;
3019 
3020 	MPASS(parent != NULL && child != NULL);
3021 	key.from = parent->w_index;
3022 	key.to = child->w_index;
3023 	WITNESS_INDEX_ASSERT(key.from);
3024 	WITNESS_INDEX_ASSERT(key.to);
3025 	if ((w_rmatrix[parent->w_index][child->w_index]
3026 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
3027 		goto out;
3028 
3029 	hash = witness_hash_djb2((const char*)&key,
3030 	    sizeof(key)) % w_lohash.wloh_size;
3031 	data = w_lohash.wloh_array[hash];
3032 	while (data != NULL) {
3033 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
3034 			break;
3035 		data = data->wlod_next;
3036 	}
3037 
3038 out:
3039 	return (data);
3040 }
3041 
3042 /*
3043  * Verify that parent and child have a known relationship, are not the same,
3044  * and child is actually a child of parent.  This is done without w_mtx
3045  * to avoid contention in the common case.
3046  */
3047 static int
3048 witness_lock_order_check(struct witness *parent, struct witness *child)
3049 {
3050 
3051 	if (parent != child &&
3052 	    w_rmatrix[parent->w_index][child->w_index]
3053 	    & WITNESS_LOCK_ORDER_KNOWN &&
3054 	    isitmychild(parent, child))
3055 		return (1);
3056 
3057 	return (0);
3058 }
3059 
3060 static int
3061 witness_lock_order_add(struct witness *parent, struct witness *child)
3062 {
3063 	struct witness_lock_order_data *data = NULL;
3064 	struct witness_lock_order_key key;
3065 	unsigned int hash;
3066 
3067 	MPASS(parent != NULL && child != NULL);
3068 	key.from = parent->w_index;
3069 	key.to = child->w_index;
3070 	WITNESS_INDEX_ASSERT(key.from);
3071 	WITNESS_INDEX_ASSERT(key.to);
3072 	if (w_rmatrix[parent->w_index][child->w_index]
3073 	    & WITNESS_LOCK_ORDER_KNOWN)
3074 		return (1);
3075 
3076 	hash = witness_hash_djb2((const char*)&key,
3077 	    sizeof(key)) % w_lohash.wloh_size;
3078 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3079 	data = w_lofree;
3080 	if (data == NULL)
3081 		return (0);
3082 	w_lofree = data->wlod_next;
3083 	data->wlod_next = w_lohash.wloh_array[hash];
3084 	data->wlod_key = key;
3085 	w_lohash.wloh_array[hash] = data;
3086 	w_lohash.wloh_count++;
3087 	stack_zero(&data->wlod_stack);
3088 	stack_save(&data->wlod_stack);
3089 	return (1);
3090 }
3091 
3092 /* Call this whenever the structure of the witness graph changes. */
3093 static void
3094 witness_increment_graph_generation(void)
3095 {
3096 
3097 	if (witness_cold == 0)
3098 		mtx_assert(&w_mtx, MA_OWNED);
3099 	w_generation++;
3100 }
3101 
3102 static int
3103 witness_output_drain(void *arg __unused, const char *data, int len)
3104 {
3105 
3106 	witness_output("%.*s", len, data);
3107 	return (len);
3108 }
3109 
3110 static void
3111 witness_debugger(int cond, const char *msg)
3112 {
3113 	char buf[32];
3114 	struct sbuf sb;
3115 	struct stack st;
3116 
3117 	if (!cond)
3118 		return;
3119 
3120 	if (witness_trace) {
3121 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3122 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3123 
3124 		stack_zero(&st);
3125 		stack_save(&st);
3126 		witness_output("stack backtrace:\n");
3127 		stack_sbuf_print_ddb(&sb, &st);
3128 
3129 		sbuf_finish(&sb);
3130 	}
3131 
3132 	witness_enter_debugger(msg);
3133 }
3134 
3135 static void
3136 witness_enter_debugger(const char *msg)
3137 {
3138 #ifdef KDB
3139 	if (witness_kdb)
3140 		kdb_enter(KDB_WHY_WITNESS, msg);
3141 #endif
3142 }
3143