xref: /freebsd/sys/kern/subr_witness.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
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 	{ "if_addr_lock", &lock_class_mtx_sleep },
545 	{ NULL, NULL },
546 	/*
547 	 * IPv6 multicast:
548 	 * protocol locks before interface locks, after UDP locks.
549 	 */
550 	{ "in6_multi_sx", &lock_class_sx },
551 	{ "udpinp", &lock_class_rw },
552 	{ "in6_multi_list_mtx", &lock_class_mtx_sleep },
553 	{ "mld_mtx", &lock_class_mtx_sleep },
554 	{ "if_addr_lock", &lock_class_mtx_sleep },
555 	{ NULL, NULL },
556 	/*
557 	 * UNIX Domain Sockets
558 	 */
559 	{ "unp_link_rwlock", &lock_class_rw },
560 	{ "unp_list_lock", &lock_class_mtx_sleep },
561 	{ "unp", &lock_class_mtx_sleep },
562 	{ "so_snd", &lock_class_mtx_sleep },
563 	{ NULL, NULL },
564 	/*
565 	 * UDP/IP
566 	 */
567 	{ "udpinp", &lock_class_rw },
568 	{ "udp", &lock_class_mtx_sleep },
569 	{ "so_snd", &lock_class_mtx_sleep },
570 	{ NULL, NULL },
571 	/*
572 	 * TCP/IP
573 	 */
574 	{ "tcpinp", &lock_class_rw },
575 	{ "tcp", &lock_class_mtx_sleep },
576 	{ "so_snd", &lock_class_mtx_sleep },
577 	{ NULL, NULL },
578 	/*
579 	 * BPF
580 	 */
581 	{ "bpf global lock", &lock_class_sx },
582 	{ "bpf cdev lock", &lock_class_mtx_sleep },
583 	{ NULL, NULL },
584 	/*
585 	 * NFS server
586 	 */
587 	{ "nfsd_mtx", &lock_class_mtx_sleep },
588 	{ "so_snd", &lock_class_mtx_sleep },
589 	{ NULL, NULL },
590 
591 	/*
592 	 * IEEE 802.11
593 	 */
594 	{ "802.11 com lock", &lock_class_mtx_sleep},
595 	{ NULL, NULL },
596 	/*
597 	 * Network drivers
598 	 */
599 	{ "network driver", &lock_class_mtx_sleep},
600 	{ NULL, NULL },
601 
602 	/*
603 	 * Netgraph
604 	 */
605 	{ "ng_node", &lock_class_mtx_sleep },
606 	{ "ng_worklist", &lock_class_mtx_sleep },
607 	{ NULL, NULL },
608 	/*
609 	 * CDEV
610 	 */
611 	{ "vm map (system)", &lock_class_mtx_sleep },
612 	{ "vnode interlock", &lock_class_mtx_sleep },
613 	{ "cdev", &lock_class_mtx_sleep },
614 	{ "devthrd", &lock_class_mtx_sleep },
615 	{ NULL, NULL },
616 	/*
617 	 * VM
618 	 */
619 	{ "vm map (user)", &lock_class_sx },
620 	{ "vm object", &lock_class_rw },
621 	{ "vm page", &lock_class_mtx_sleep },
622 	{ "pmap pv global", &lock_class_rw },
623 	{ "pmap", &lock_class_mtx_sleep },
624 	{ "pmap pv list", &lock_class_rw },
625 	{ "vm page free queue", &lock_class_mtx_sleep },
626 	{ "vm pagequeue", &lock_class_mtx_sleep },
627 	{ NULL, NULL },
628 	/*
629 	 * kqueue/VFS interaction
630 	 */
631 	{ "kqueue", &lock_class_mtx_sleep },
632 	{ "struct mount mtx", &lock_class_mtx_sleep },
633 	{ "vnode interlock", &lock_class_mtx_sleep },
634 	{ NULL, NULL },
635 	/*
636 	 * VFS namecache
637 	 */
638 	{ "ncvn", &lock_class_mtx_sleep },
639 	{ "ncbuc", &lock_class_mtx_sleep },
640 	{ "vnode interlock", &lock_class_mtx_sleep },
641 	{ "ncneg", &lock_class_mtx_sleep },
642 	{ NULL, NULL },
643 	/*
644 	 * ZFS locking
645 	 */
646 	{ "dn->dn_mtx", &lock_class_sx },
647 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
648 	{ "db->db_mtx", &lock_class_sx },
649 	{ NULL, NULL },
650 	/*
651 	 * TCP log locks
652 	 */
653 	{ "TCP ID tree", &lock_class_rw },
654 	{ "tcp log id bucket", &lock_class_mtx_sleep },
655 	{ "tcpinp", &lock_class_rw },
656 	{ "TCP log expireq", &lock_class_mtx_sleep },
657 	{ NULL, NULL },
658 	/*
659 	 * spin locks
660 	 */
661 #ifdef SMP
662 	{ "ap boot", &lock_class_mtx_spin },
663 #endif
664 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
665 #ifdef __i386__
666 	{ "cy", &lock_class_mtx_spin },
667 #endif
668 	{ "scc_hwmtx", &lock_class_mtx_spin },
669 	{ "uart_hwmtx", &lock_class_mtx_spin },
670 	{ "fast_taskqueue", &lock_class_mtx_spin },
671 	{ "intr table", &lock_class_mtx_spin },
672 	{ "process slock", &lock_class_mtx_spin },
673 	{ "syscons video lock", &lock_class_mtx_spin },
674 	{ "sleepq chain", &lock_class_mtx_spin },
675 	{ "rm_spinlock", &lock_class_mtx_spin },
676 	{ "turnstile chain", &lock_class_mtx_spin },
677 	{ "turnstile lock", &lock_class_mtx_spin },
678 	{ "sched lock", &lock_class_mtx_spin },
679 	{ "td_contested", &lock_class_mtx_spin },
680 	{ "callout", &lock_class_mtx_spin },
681 	{ "entropy harvest mutex", &lock_class_mtx_spin },
682 #ifdef SMP
683 	{ "smp rendezvous", &lock_class_mtx_spin },
684 #endif
685 #ifdef __powerpc__
686 	{ "tlb0", &lock_class_mtx_spin },
687 #endif
688 	{ NULL, NULL },
689 	{ "sched lock", &lock_class_mtx_spin },
690 #ifdef	HWPMC_HOOKS
691 	{ "pmc-per-proc", &lock_class_mtx_spin },
692 #endif
693 	{ NULL, NULL },
694 	/*
695 	 * leaf locks
696 	 */
697 	{ "intrcnt", &lock_class_mtx_spin },
698 	{ "icu", &lock_class_mtx_spin },
699 #ifdef __i386__
700 	{ "allpmaps", &lock_class_mtx_spin },
701 	{ "descriptor tables", &lock_class_mtx_spin },
702 #endif
703 	{ "clk", &lock_class_mtx_spin },
704 	{ "cpuset", &lock_class_mtx_spin },
705 	{ "mprof lock", &lock_class_mtx_spin },
706 	{ "zombie lock", &lock_class_mtx_spin },
707 	{ "ALD Queue", &lock_class_mtx_spin },
708 #if defined(__i386__) || defined(__amd64__)
709 	{ "pcicfg", &lock_class_mtx_spin },
710 	{ "NDIS thread lock", &lock_class_mtx_spin },
711 #endif
712 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
713 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
714 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
715 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
716 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
717 #ifdef	HWPMC_HOOKS
718 	{ "pmc-leaf", &lock_class_mtx_spin },
719 #endif
720 	{ "blocked lock", &lock_class_mtx_spin },
721 	{ NULL, NULL },
722 	{ NULL, NULL }
723 };
724 
725 /*
726  * Pairs of locks which have been blessed.  Witness does not complain about
727  * order problems with blessed lock pairs.  Please do not add an entry to the
728  * table without an explanatory comment.
729  */
730 static struct witness_blessed blessed_list[] = {
731 	/*
732 	 * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
733 	 * both lock orders, so a deadlock cannot happen as a result of this
734 	 * LOR.
735 	 */
736 	{ "dirhash",	"bufwait" },
737 
738 	/*
739 	 * A UFS vnode may be locked in vget() while a buffer belonging to the
740 	 * parent directory vnode is locked.
741 	 */
742 	{ "ufs",	"bufwait" },
743 
744 	/*
745 	 * The tarfs decompression stream vnode may be locked while a
746 	 * buffer belonging to a tarfs data vnode is locked.
747 	 */
748 	{ "tarfs",	"bufwait" },
749 };
750 
751 /*
752  * This global is set to 0 once it becomes safe to use the witness code.
753  */
754 static int witness_cold = 1;
755 
756 /*
757  * This global is set to 1 once the static lock orders have been enrolled
758  * so that a warning can be issued for any spin locks enrolled later.
759  */
760 static int witness_spin_warn = 0;
761 
762 /* Trim useless garbage from filenames. */
763 static const char *
764 fixup_filename(const char *file)
765 {
766 
767 	if (file == NULL)
768 		return (NULL);
769 	while (strncmp(file, "../", 3) == 0)
770 		file += 3;
771 	return (file);
772 }
773 
774 /*
775  * Calculate the size of early witness structures.
776  */
777 int
778 witness_startup_count(void)
779 {
780 	int sz;
781 
782 	sz = sizeof(struct witness) * witness_count;
783 	sz += sizeof(*w_rmatrix) * (witness_count + 1);
784 	sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
785 	    (witness_count + 1);
786 
787 	return (sz);
788 }
789 
790 /*
791  * The WITNESS-enabled diagnostic code.  Note that the witness code does
792  * assume that the early boot is single-threaded at least until after this
793  * routine is completed.
794  */
795 void
796 witness_startup(void *mem)
797 {
798 	struct lock_object *lock;
799 	struct witness_order_list_entry *order;
800 	struct witness *w, *w1;
801 	uintptr_t p;
802 	int i;
803 
804 	p = (uintptr_t)mem;
805 	w_data = (void *)p;
806 	p += sizeof(struct witness) * witness_count;
807 
808 	w_rmatrix = (void *)p;
809 	p += sizeof(*w_rmatrix) * (witness_count + 1);
810 
811 	for (i = 0; i < witness_count + 1; i++) {
812 		w_rmatrix[i] = (void *)p;
813 		p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
814 	}
815 	badstack_sbuf_size = witness_count * 256;
816 
817 	/*
818 	 * We have to release Giant before initializing its witness
819 	 * structure so that WITNESS doesn't get confused.
820 	 */
821 	mtx_unlock(&Giant);
822 	mtx_assert(&Giant, MA_NOTOWNED);
823 
824 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
825 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
826 	    MTX_NOWITNESS | MTX_NOPROFILE);
827 	for (i = witness_count - 1; i >= 0; i--) {
828 		w = &w_data[i];
829 		memset(w, 0, sizeof(*w));
830 		w_data[i].w_index = i;	/* Witness index never changes. */
831 		witness_free(w);
832 	}
833 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
834 	    ("%s: Invalid list of free witness objects", __func__));
835 
836 	/* Witness with index 0 is not used to aid in debugging. */
837 	STAILQ_REMOVE_HEAD(&w_free, w_list);
838 	w_free_cnt--;
839 
840 	for (i = 0; i < witness_count; i++) {
841 		memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
842 		    (witness_count + 1));
843 	}
844 
845 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
846 		witness_lock_list_free(&w_locklistdata[i]);
847 	witness_init_hash_tables();
848 
849 	/* First add in all the specified order lists. */
850 	for (order = order_lists; order->w_name != NULL; order++) {
851 		w = enroll(order->w_name, order->w_class);
852 		if (w == NULL)
853 			continue;
854 		w->w_file = "order list";
855 		for (order++; order->w_name != NULL; order++) {
856 			w1 = enroll(order->w_name, order->w_class);
857 			if (w1 == NULL)
858 				continue;
859 			w1->w_file = "order list";
860 			itismychild(w, w1);
861 			w = w1;
862 		}
863 	}
864 	witness_spin_warn = 1;
865 
866 	/* Iterate through all locks and add them to witness. */
867 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
868 		lock = pending_locks[i].wh_lock;
869 		KASSERT(lock->lo_flags & LO_WITNESS,
870 		    ("%s: lock %s is on pending list but not LO_WITNESS",
871 		    __func__, lock->lo_name));
872 		lock->lo_witness = enroll(pending_locks[i].wh_type,
873 		    LOCK_CLASS(lock));
874 	}
875 
876 	/* Mark the witness code as being ready for use. */
877 	witness_cold = 0;
878 
879 	mtx_lock(&Giant);
880 }
881 
882 void
883 witness_init(struct lock_object *lock, const char *type)
884 {
885 	struct lock_class *class;
886 
887 	/* Various sanity checks. */
888 	class = LOCK_CLASS(lock);
889 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
890 	    (class->lc_flags & LC_RECURSABLE) == 0)
891 		kassert_panic("%s: lock (%s) %s can not be recursable",
892 		    __func__, class->lc_name, lock->lo_name);
893 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
894 	    (class->lc_flags & LC_SLEEPABLE) == 0)
895 		kassert_panic("%s: lock (%s) %s can not be sleepable",
896 		    __func__, class->lc_name, lock->lo_name);
897 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
898 	    (class->lc_flags & LC_UPGRADABLE) == 0)
899 		kassert_panic("%s: lock (%s) %s can not be upgradable",
900 		    __func__, class->lc_name, lock->lo_name);
901 
902 	/*
903 	 * If we shouldn't watch this lock, then just clear lo_witness.
904 	 * Otherwise, if witness_cold is set, then it is too early to
905 	 * enroll this lock, so defer it to witness_initialize() by adding
906 	 * it to the pending_locks list.  If it is not too early, then enroll
907 	 * the lock now.
908 	 */
909 	if (witness_watch < 1 || KERNEL_PANICKED() ||
910 	    (lock->lo_flags & LO_WITNESS) == 0)
911 		lock->lo_witness = NULL;
912 	else if (witness_cold) {
913 		pending_locks[pending_cnt].wh_lock = lock;
914 		pending_locks[pending_cnt++].wh_type = type;
915 		if (pending_cnt > WITNESS_PENDLIST)
916 			panic("%s: pending locks list is too small, "
917 			    "increase WITNESS_PENDLIST\n",
918 			    __func__);
919 	} else
920 		lock->lo_witness = enroll(type, class);
921 }
922 
923 void
924 witness_destroy(struct lock_object *lock)
925 {
926 	struct lock_class *class;
927 	struct witness *w;
928 
929 	class = LOCK_CLASS(lock);
930 
931 	if (witness_cold)
932 		panic("lock (%s) %s destroyed while witness_cold",
933 		    class->lc_name, lock->lo_name);
934 
935 	/* XXX: need to verify that no one holds the lock */
936 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
937 		return;
938 	w = lock->lo_witness;
939 
940 	mtx_lock_spin(&w_mtx);
941 	MPASS(w->w_refcount > 0);
942 	w->w_refcount--;
943 
944 	if (w->w_refcount == 0)
945 		depart(w);
946 	mtx_unlock_spin(&w_mtx);
947 }
948 
949 #ifdef DDB
950 static void
951 witness_ddb_compute_levels(void)
952 {
953 	struct witness *w;
954 
955 	/*
956 	 * First clear all levels.
957 	 */
958 	STAILQ_FOREACH(w, &w_all, w_list)
959 		w->w_ddb_level = -1;
960 
961 	/*
962 	 * Look for locks with no parents and level all their descendants.
963 	 */
964 	STAILQ_FOREACH(w, &w_all, w_list) {
965 		/* If the witness has ancestors (is not a root), skip it. */
966 		if (w->w_num_ancestors > 0)
967 			continue;
968 		witness_ddb_level_descendants(w, 0);
969 	}
970 }
971 
972 static void
973 witness_ddb_level_descendants(struct witness *w, int l)
974 {
975 	int i;
976 
977 	if (w->w_ddb_level >= l)
978 		return;
979 
980 	w->w_ddb_level = l;
981 	l++;
982 
983 	for (i = 1; i <= w_max_used_index; i++) {
984 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
985 			witness_ddb_level_descendants(&w_data[i], l);
986 	}
987 }
988 
989 static void
990 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
991     struct witness *w, int indent)
992 {
993 	int i;
994 
995  	for (i = 0; i < indent; i++)
996  		prnt(" ");
997 	prnt("%s (type: %s, depth: %d, active refs: %d)",
998 	     w->w_name, w->w_class->lc_name,
999 	     w->w_ddb_level, w->w_refcount);
1000  	if (w->w_displayed) {
1001  		prnt(" -- (already displayed)\n");
1002  		return;
1003  	}
1004  	w->w_displayed = 1;
1005 	if (w->w_file != NULL && w->w_line != 0)
1006 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1007 		    w->w_line);
1008 	else
1009 		prnt(" -- never acquired\n");
1010 	indent++;
1011 	WITNESS_INDEX_ASSERT(w->w_index);
1012 	for (i = 1; i <= w_max_used_index; i++) {
1013 		if (db_pager_quit)
1014 			return;
1015 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1016 			witness_ddb_display_descendants(prnt, &w_data[i],
1017 			    indent);
1018 	}
1019 }
1020 
1021 static void
1022 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1023     struct witness_list *list)
1024 {
1025 	struct witness *w;
1026 
1027 	STAILQ_FOREACH(w, list, w_typelist) {
1028 		if (w->w_file == NULL || w->w_ddb_level > 0)
1029 			continue;
1030 
1031 		/* This lock has no anscestors - display its descendants. */
1032 		witness_ddb_display_descendants(prnt, w, 0);
1033 		if (db_pager_quit)
1034 			return;
1035 	}
1036 }
1037 
1038 static void
1039 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1040 {
1041 	struct witness *w;
1042 
1043 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1044 	witness_ddb_compute_levels();
1045 
1046 	/* Clear all the displayed flags. */
1047 	STAILQ_FOREACH(w, &w_all, w_list)
1048 		w->w_displayed = 0;
1049 
1050 	/*
1051 	 * First, handle sleep locks which have been acquired at least
1052 	 * once.
1053 	 */
1054 	prnt("Sleep locks:\n");
1055 	witness_ddb_display_list(prnt, &w_sleep);
1056 	if (db_pager_quit)
1057 		return;
1058 
1059 	/*
1060 	 * Now do spin locks which have been acquired at least once.
1061 	 */
1062 	prnt("\nSpin locks:\n");
1063 	witness_ddb_display_list(prnt, &w_spin);
1064 	if (db_pager_quit)
1065 		return;
1066 
1067 	/*
1068 	 * Finally, any locks which have not been acquired yet.
1069 	 */
1070 	prnt("\nLocks which were never acquired:\n");
1071 	STAILQ_FOREACH(w, &w_all, w_list) {
1072 		if (w->w_file != NULL || w->w_refcount == 0)
1073 			continue;
1074 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1075 		    w->w_class->lc_name, w->w_ddb_level);
1076 		if (db_pager_quit)
1077 			return;
1078 	}
1079 }
1080 #endif /* DDB */
1081 
1082 int
1083 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1084 {
1085 
1086 	if (witness_watch == -1 || KERNEL_PANICKED())
1087 		return (0);
1088 
1089 	/* Require locks that witness knows about. */
1090 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1091 	    lock2->lo_witness == NULL)
1092 		return (EINVAL);
1093 
1094 	mtx_assert(&w_mtx, MA_NOTOWNED);
1095 	mtx_lock_spin(&w_mtx);
1096 
1097 	/*
1098 	 * If we already have either an explicit or implied lock order that
1099 	 * is the other way around, then return an error.
1100 	 */
1101 	if (witness_watch &&
1102 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1103 		mtx_unlock_spin(&w_mtx);
1104 		return (EDOOFUS);
1105 	}
1106 
1107 	/* Try to add the new order. */
1108 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1109 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1110 	itismychild(lock1->lo_witness, lock2->lo_witness);
1111 	mtx_unlock_spin(&w_mtx);
1112 	return (0);
1113 }
1114 
1115 void
1116 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1117     int line, struct lock_object *interlock)
1118 {
1119 	struct lock_list_entry *lock_list, *lle;
1120 	struct lock_instance *lock1, *lock2, *plock;
1121 	struct lock_class *class, *iclass;
1122 	struct witness *w, *w1;
1123 	struct thread *td;
1124 	int i, j;
1125 
1126 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1127 	    KERNEL_PANICKED())
1128 		return;
1129 
1130 	w = lock->lo_witness;
1131 	class = LOCK_CLASS(lock);
1132 	td = curthread;
1133 
1134 	if (class->lc_flags & LC_SLEEPLOCK) {
1135 		/*
1136 		 * Since spin locks include a critical section, this check
1137 		 * implicitly enforces a lock order of all sleep locks before
1138 		 * all spin locks.
1139 		 */
1140 		if (td->td_critnest != 0 && !kdb_active)
1141 			kassert_panic("acquiring blockable sleep lock with "
1142 			    "spinlock or critical section held (%s) %s @ %s:%d",
1143 			    class->lc_name, lock->lo_name,
1144 			    fixup_filename(file), line);
1145 
1146 		/*
1147 		 * If this is the first lock acquired then just return as
1148 		 * no order checking is needed.
1149 		 */
1150 		lock_list = td->td_sleeplocks;
1151 		if (lock_list == NULL || lock_list->ll_count == 0)
1152 			return;
1153 	} else {
1154 		/*
1155 		 * If this is the first lock, just return as no order
1156 		 * checking is needed.  Avoid problems with thread
1157 		 * migration pinning the thread while checking if
1158 		 * spinlocks are held.  If at least one spinlock is held
1159 		 * the thread is in a safe path and it is allowed to
1160 		 * unpin it.
1161 		 */
1162 		sched_pin();
1163 		lock_list = PCPU_GET(spinlocks);
1164 		if (lock_list == NULL || lock_list->ll_count == 0) {
1165 			sched_unpin();
1166 			return;
1167 		}
1168 		sched_unpin();
1169 	}
1170 
1171 	/*
1172 	 * Check to see if we are recursing on a lock we already own.  If
1173 	 * so, make sure that we don't mismatch exclusive and shared lock
1174 	 * acquires.
1175 	 */
1176 	lock1 = find_instance(lock_list, lock);
1177 	if (lock1 != NULL) {
1178 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1179 		    (flags & LOP_EXCLUSIVE) == 0) {
1180 			witness_output("shared lock of (%s) %s @ %s:%d\n",
1181 			    class->lc_name, lock->lo_name,
1182 			    fixup_filename(file), line);
1183 			witness_output("while exclusively locked from %s:%d\n",
1184 			    fixup_filename(lock1->li_file), lock1->li_line);
1185 			kassert_panic("excl->share");
1186 		}
1187 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1188 		    (flags & LOP_EXCLUSIVE) != 0) {
1189 			witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1190 			    class->lc_name, lock->lo_name,
1191 			    fixup_filename(file), line);
1192 			witness_output("while share locked from %s:%d\n",
1193 			    fixup_filename(lock1->li_file), lock1->li_line);
1194 			kassert_panic("share->excl");
1195 		}
1196 		return;
1197 	}
1198 
1199 	/* Warn if the interlock is not locked exactly once. */
1200 	if (interlock != NULL) {
1201 		iclass = LOCK_CLASS(interlock);
1202 		lock1 = find_instance(lock_list, interlock);
1203 		if (lock1 == NULL)
1204 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1205 			    iclass->lc_name, interlock->lo_name,
1206 			    fixup_filename(file), line);
1207 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1208 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1209 			    iclass->lc_name, interlock->lo_name,
1210 			    fixup_filename(file), line);
1211 	}
1212 
1213 	/*
1214 	 * Find the previously acquired lock, but ignore interlocks.
1215 	 */
1216 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1217 	if (interlock != NULL && plock->li_lock == interlock) {
1218 		if (lock_list->ll_count > 1)
1219 			plock =
1220 			    &lock_list->ll_children[lock_list->ll_count - 2];
1221 		else {
1222 			lle = lock_list->ll_next;
1223 
1224 			/*
1225 			 * The interlock is the only lock we hold, so
1226 			 * simply return.
1227 			 */
1228 			if (lle == NULL)
1229 				return;
1230 			plock = &lle->ll_children[lle->ll_count - 1];
1231 		}
1232 	}
1233 
1234 	/*
1235 	 * Try to perform most checks without a lock.  If this succeeds we
1236 	 * can skip acquiring the lock and return success.  Otherwise we redo
1237 	 * the check with the lock held to handle races with concurrent updates.
1238 	 */
1239 	w1 = plock->li_lock->lo_witness;
1240 	if (witness_lock_order_check(w1, w))
1241 		return;
1242 
1243 	mtx_lock_spin(&w_mtx);
1244 	if (witness_lock_order_check(w1, w)) {
1245 		mtx_unlock_spin(&w_mtx);
1246 		return;
1247 	}
1248 	witness_lock_order_add(w1, w);
1249 
1250 	/*
1251 	 * Check for duplicate locks of the same type.  Note that we only
1252 	 * have to check for this on the last lock we just acquired.  Any
1253 	 * other cases will be caught as lock order violations.
1254 	 */
1255 	if (w1 == w) {
1256 		i = w->w_index;
1257 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1258 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1259 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1260 			w->w_reversed = 1;
1261 			mtx_unlock_spin(&w_mtx);
1262 			witness_output(
1263 			    "acquiring duplicate lock of same type: \"%s\"\n",
1264 			    w->w_name);
1265 			witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1266 			    fixup_filename(plock->li_file), plock->li_line);
1267 			witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1268 			    fixup_filename(file), line);
1269 			witness_debugger(1, __func__);
1270 		} else
1271 			mtx_unlock_spin(&w_mtx);
1272 		return;
1273 	}
1274 	mtx_assert(&w_mtx, MA_OWNED);
1275 
1276 	/*
1277 	 * If we know that the lock we are acquiring comes after
1278 	 * the lock we most recently acquired in the lock order tree,
1279 	 * then there is no need for any further checks.
1280 	 */
1281 	if (isitmychild(w1, w))
1282 		goto out;
1283 
1284 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1285 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1286 			struct stack pstack;
1287 			bool pstackv, trace;
1288 
1289 			MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1290 			lock1 = &lle->ll_children[i];
1291 
1292 			/*
1293 			 * Ignore the interlock.
1294 			 */
1295 			if (interlock == lock1->li_lock)
1296 				continue;
1297 
1298 			/*
1299 			 * If this lock doesn't undergo witness checking,
1300 			 * then skip it.
1301 			 */
1302 			w1 = lock1->li_lock->lo_witness;
1303 			if (w1 == NULL) {
1304 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1305 				    ("lock missing witness structure"));
1306 				continue;
1307 			}
1308 
1309 			/*
1310 			 * If we are locking Giant and this is a sleepable
1311 			 * lock, then skip it.
1312 			 */
1313 			if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1314 			    lock == &Giant.lock_object)
1315 				continue;
1316 
1317 			/*
1318 			 * If we are locking a sleepable lock and this lock
1319 			 * is Giant, then skip it.
1320 			 */
1321 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1322 			    (flags & LOP_NOSLEEP) == 0 &&
1323 			    lock1->li_lock == &Giant.lock_object)
1324 				continue;
1325 
1326 			/*
1327 			 * If we are locking a sleepable lock and this lock
1328 			 * isn't sleepable, we want to treat it as a lock
1329 			 * order violation to enfore a general lock order of
1330 			 * sleepable locks before non-sleepable locks.
1331 			 */
1332 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1333 			    (flags & LOP_NOSLEEP) == 0 &&
1334 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1335 				goto reversal;
1336 
1337 			/*
1338 			 * If we are locking Giant and this is a non-sleepable
1339 			 * lock, then treat it as a reversal.
1340 			 */
1341 			if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1342 			    lock == &Giant.lock_object)
1343 				goto reversal;
1344 
1345 			/*
1346 			 * Check the lock order hierarchy for a reveresal.
1347 			 */
1348 			if (!isitmydescendant(w, w1))
1349 				continue;
1350 		reversal:
1351 
1352 			/*
1353 			 * We have a lock order violation, check to see if it
1354 			 * is allowed or has already been yelled about.
1355 			 */
1356 
1357 			/* Bail if this violation is known */
1358 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1359 				goto out;
1360 
1361 			/* Record this as a violation */
1362 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1363 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1364 			w->w_reversed = w1->w_reversed = 1;
1365 			witness_increment_graph_generation();
1366 
1367 			/*
1368 			 * If the lock order is blessed, bail before logging
1369 			 * anything.  We don't look for other lock order
1370 			 * violations though, which may be a bug.
1371 			 */
1372 			if (blessed(w, w1))
1373 				goto out;
1374 
1375 			trace = atomic_load_int(&witness_trace);
1376 			if (trace) {
1377 				struct witness_lock_order_data *data;
1378 
1379 				pstackv = false;
1380 				data = witness_lock_order_get(w, w1);
1381 				if (data != NULL) {
1382 					stack_copy(&data->wlod_stack,
1383 					    &pstack);
1384 					pstackv = true;
1385 				}
1386 			}
1387 			mtx_unlock_spin(&w_mtx);
1388 
1389 #ifdef WITNESS_NO_VNODE
1390 			/*
1391 			 * There are known LORs between VNODE locks. They are
1392 			 * not an indication of a bug. VNODE locks are flagged
1393 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1394 			 * is between 2 VNODE locks.
1395 			 */
1396 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1397 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1398 				return;
1399 #endif
1400 
1401 			/*
1402 			 * Ok, yell about it.
1403 			 */
1404 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1405 			    (flags & LOP_NOSLEEP) == 0 &&
1406 			    (lock1->li_flags & LI_SLEEPABLE) == 0)
1407 				witness_output(
1408 		"lock order reversal: (sleepable after non-sleepable)\n");
1409 			else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1410 			    && lock == &Giant.lock_object)
1411 				witness_output(
1412 		"lock order reversal: (Giant after non-sleepable)\n");
1413 			else
1414 				witness_output("lock order reversal:\n");
1415 
1416 			/*
1417 			 * Try to locate an earlier lock with
1418 			 * witness w in our list.
1419 			 */
1420 			do {
1421 				lock2 = &lle->ll_children[i];
1422 				MPASS(lock2->li_lock != NULL);
1423 				if (lock2->li_lock->lo_witness == w)
1424 					break;
1425 				if (i == 0 && lle->ll_next != NULL) {
1426 					lle = lle->ll_next;
1427 					i = lle->ll_count - 1;
1428 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1429 				} else
1430 					i--;
1431 			} while (i >= 0);
1432 			if (i < 0) {
1433 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1434 				    lock1->li_lock, lock1->li_lock->lo_name,
1435 				    w1->w_name, w1->w_class->lc_name,
1436 				    fixup_filename(lock1->li_file),
1437 				    lock1->li_line);
1438 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1439 				    lock, lock->lo_name, w->w_name,
1440 				    w->w_class->lc_name, fixup_filename(file),
1441 				    line);
1442 			} else {
1443 				struct witness *w2 = lock2->li_lock->lo_witness;
1444 
1445 				witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1446 				    lock2->li_lock, lock2->li_lock->lo_name,
1447 				    w2->w_name, w2->w_class->lc_name,
1448 				    fixup_filename(lock2->li_file),
1449 				    lock2->li_line);
1450 				witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1451 				    lock1->li_lock, lock1->li_lock->lo_name,
1452 				    w1->w_name, w1->w_class->lc_name,
1453 				    fixup_filename(lock1->li_file),
1454 				    lock1->li_line);
1455 				witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1456 				    lock->lo_name, w->w_name,
1457 				    w->w_class->lc_name, fixup_filename(file),
1458 				    line);
1459 			}
1460 			if (trace) {
1461 				char buf[64];
1462 				struct sbuf sb;
1463 
1464 				sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1465 				sbuf_set_drain(&sb, witness_output_drain,
1466 				    NULL);
1467 
1468 				if (pstackv) {
1469 					sbuf_printf(&sb,
1470 				    "lock order %s -> %s established at:\n",
1471 					    w->w_name, w1->w_name);
1472 					stack_sbuf_print_flags(&sb, &pstack,
1473 					    M_NOWAIT, STACK_SBUF_FMT_LONG);
1474 				}
1475 
1476 				sbuf_printf(&sb,
1477 				    "lock order %s -> %s attempted at:\n",
1478 				    w1->w_name, w->w_name);
1479 				stack_save(&pstack);
1480 				stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1481 				    STACK_SBUF_FMT_LONG);
1482 
1483 				sbuf_finish(&sb);
1484 				sbuf_delete(&sb);
1485 			}
1486 			witness_enter_debugger(__func__);
1487 			return;
1488 		}
1489 	}
1490 
1491 	/*
1492 	 * If requested, build a new lock order.  However, don't build a new
1493 	 * relationship between a sleepable lock and Giant if it is in the
1494 	 * wrong direction.  The correct lock order is that sleepable locks
1495 	 * always come before Giant.
1496 	 */
1497 	if (flags & LOP_NEWORDER &&
1498 	    !(plock->li_lock == &Giant.lock_object &&
1499 	    (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1500 	    (flags & LOP_NOSLEEP) == 0)) {
1501 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1502 		    w->w_name, plock->li_lock->lo_witness->w_name);
1503 		itismychild(plock->li_lock->lo_witness, w);
1504 	}
1505 out:
1506 	mtx_unlock_spin(&w_mtx);
1507 }
1508 
1509 void
1510 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1511 {
1512 	struct lock_list_entry **lock_list, *lle;
1513 	struct lock_instance *instance;
1514 	struct witness *w;
1515 	struct thread *td;
1516 
1517 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1518 	    KERNEL_PANICKED())
1519 		return;
1520 	w = lock->lo_witness;
1521 	td = curthread;
1522 
1523 	/* Determine lock list for this lock. */
1524 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1525 		lock_list = &td->td_sleeplocks;
1526 	else
1527 		lock_list = PCPU_PTR(spinlocks);
1528 
1529 	/* Check to see if we are recursing on a lock we already own. */
1530 	instance = find_instance(*lock_list, lock);
1531 	if (instance != NULL) {
1532 		instance->li_flags++;
1533 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1534 		    td->td_proc->p_pid, lock->lo_name,
1535 		    instance->li_flags & LI_RECURSEMASK);
1536 		instance->li_file = file;
1537 		instance->li_line = line;
1538 		return;
1539 	}
1540 
1541 	/* Update per-witness last file and line acquire. */
1542 	w->w_file = file;
1543 	w->w_line = line;
1544 
1545 	/* Find the next open lock instance in the list and fill it. */
1546 	lle = *lock_list;
1547 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1548 		lle = witness_lock_list_get();
1549 		if (lle == NULL)
1550 			return;
1551 		lle->ll_next = *lock_list;
1552 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1553 		    td->td_proc->p_pid, lle);
1554 		*lock_list = lle;
1555 	}
1556 	instance = &lle->ll_children[lle->ll_count++];
1557 	instance->li_lock = lock;
1558 	instance->li_line = line;
1559 	instance->li_file = file;
1560 	instance->li_flags = 0;
1561 	if ((flags & LOP_EXCLUSIVE) != 0)
1562 		instance->li_flags |= LI_EXCLUSIVE;
1563 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1564 		instance->li_flags |= LI_SLEEPABLE;
1565 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1566 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1567 }
1568 
1569 void
1570 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1571 {
1572 	struct lock_instance *instance;
1573 	struct lock_class *class;
1574 
1575 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1576 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1577 		return;
1578 	class = LOCK_CLASS(lock);
1579 	if (witness_watch) {
1580 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1581 			kassert_panic(
1582 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1583 			    class->lc_name, lock->lo_name,
1584 			    fixup_filename(file), line);
1585 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1586 			kassert_panic(
1587 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1588 			    class->lc_name, lock->lo_name,
1589 			    fixup_filename(file), line);
1590 	}
1591 	instance = find_instance(curthread->td_sleeplocks, lock);
1592 	if (instance == NULL) {
1593 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1594 		    class->lc_name, lock->lo_name,
1595 		    fixup_filename(file), line);
1596 		return;
1597 	}
1598 	if (witness_watch) {
1599 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1600 			kassert_panic(
1601 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1602 			    class->lc_name, lock->lo_name,
1603 			    fixup_filename(file), line);
1604 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1605 			kassert_panic(
1606 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1607 			    class->lc_name, lock->lo_name,
1608 			    instance->li_flags & LI_RECURSEMASK,
1609 			    fixup_filename(file), line);
1610 	}
1611 	instance->li_flags |= LI_EXCLUSIVE;
1612 }
1613 
1614 void
1615 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1616     int line)
1617 {
1618 	struct lock_instance *instance;
1619 	struct lock_class *class;
1620 
1621 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1622 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1623 		return;
1624 	class = LOCK_CLASS(lock);
1625 	if (witness_watch) {
1626 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1627 			kassert_panic(
1628 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1629 			    class->lc_name, lock->lo_name,
1630 			    fixup_filename(file), line);
1631 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1632 			kassert_panic(
1633 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1634 			    class->lc_name, lock->lo_name,
1635 			    fixup_filename(file), line);
1636 	}
1637 	instance = find_instance(curthread->td_sleeplocks, lock);
1638 	if (instance == NULL) {
1639 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1640 		    class->lc_name, lock->lo_name,
1641 		    fixup_filename(file), line);
1642 		return;
1643 	}
1644 	if (witness_watch) {
1645 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1646 			kassert_panic(
1647 			    "downgrade of shared lock (%s) %s @ %s:%d",
1648 			    class->lc_name, lock->lo_name,
1649 			    fixup_filename(file), line);
1650 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1651 			kassert_panic(
1652 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1653 			    class->lc_name, lock->lo_name,
1654 			    instance->li_flags & LI_RECURSEMASK,
1655 			    fixup_filename(file), line);
1656 	}
1657 	instance->li_flags &= ~LI_EXCLUSIVE;
1658 }
1659 
1660 void
1661 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1662 {
1663 	struct lock_list_entry **lock_list, *lle;
1664 	struct lock_instance *instance;
1665 	struct lock_class *class;
1666 	struct thread *td;
1667 	register_t s;
1668 	int i, j;
1669 
1670 	if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1671 		return;
1672 	td = curthread;
1673 	class = LOCK_CLASS(lock);
1674 
1675 	/* Find lock instance associated with this lock. */
1676 	if (class->lc_flags & LC_SLEEPLOCK)
1677 		lock_list = &td->td_sleeplocks;
1678 	else
1679 		lock_list = PCPU_PTR(spinlocks);
1680 	lle = *lock_list;
1681 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1682 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1683 			instance = &(*lock_list)->ll_children[i];
1684 			if (instance->li_lock == lock)
1685 				goto found;
1686 		}
1687 
1688 	/*
1689 	 * When disabling WITNESS through witness_watch we could end up in
1690 	 * having registered locks in the td_sleeplocks queue.
1691 	 * We have to make sure we flush these queues, so just search for
1692 	 * eventual register locks and remove them.
1693 	 */
1694 	if (witness_watch > 0) {
1695 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1696 		    lock->lo_name, fixup_filename(file), line);
1697 		return;
1698 	} else {
1699 		return;
1700 	}
1701 found:
1702 
1703 	/* First, check for shared/exclusive mismatches. */
1704 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1705 	    (flags & LOP_EXCLUSIVE) == 0) {
1706 		witness_output("shared unlock of (%s) %s @ %s:%d\n",
1707 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1708 		witness_output("while exclusively locked from %s:%d\n",
1709 		    fixup_filename(instance->li_file), instance->li_line);
1710 		kassert_panic("excl->ushare");
1711 	}
1712 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1713 	    (flags & LOP_EXCLUSIVE) != 0) {
1714 		witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1715 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1716 		witness_output("while share locked from %s:%d\n",
1717 		    fixup_filename(instance->li_file),
1718 		    instance->li_line);
1719 		kassert_panic("share->uexcl");
1720 	}
1721 	/* If we are recursed, unrecurse. */
1722 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1723 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1724 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1725 		    instance->li_flags);
1726 		instance->li_flags--;
1727 		return;
1728 	}
1729 	/* The lock is now being dropped, check for NORELEASE flag */
1730 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1731 		witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1732 		    class->lc_name, lock->lo_name, fixup_filename(file), line);
1733 		kassert_panic("lock marked norelease");
1734 	}
1735 
1736 	/* Otherwise, remove this item from the list. */
1737 	s = intr_disable();
1738 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1739 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1740 	    (*lock_list)->ll_count - 1);
1741 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1742 		(*lock_list)->ll_children[j] =
1743 		    (*lock_list)->ll_children[j + 1];
1744 	(*lock_list)->ll_count--;
1745 	intr_restore(s);
1746 
1747 	/*
1748 	 * In order to reduce contention on w_mtx, we want to keep always an
1749 	 * head object into lists so that frequent allocation from the
1750 	 * free witness pool (and subsequent locking) is avoided.
1751 	 * In order to maintain the current code simple, when the head
1752 	 * object is totally unloaded it means also that we do not have
1753 	 * further objects in the list, so the list ownership needs to be
1754 	 * hand over to another object if the current head needs to be freed.
1755 	 */
1756 	if ((*lock_list)->ll_count == 0) {
1757 		if (*lock_list == lle) {
1758 			if (lle->ll_next == NULL)
1759 				return;
1760 		} else
1761 			lle = *lock_list;
1762 		*lock_list = lle->ll_next;
1763 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1764 		    td->td_proc->p_pid, lle);
1765 		witness_lock_list_free(lle);
1766 	}
1767 }
1768 
1769 void
1770 witness_thread_exit(struct thread *td)
1771 {
1772 	struct lock_list_entry *lle;
1773 	int i, n;
1774 
1775 	lle = td->td_sleeplocks;
1776 	if (lle == NULL || KERNEL_PANICKED())
1777 		return;
1778 	if (lle->ll_count != 0) {
1779 		for (n = 0; lle != NULL; lle = lle->ll_next)
1780 			for (i = lle->ll_count - 1; i >= 0; i--) {
1781 				if (n == 0)
1782 					witness_output(
1783 		    "Thread %p exiting with the following locks held:\n", td);
1784 				n++;
1785 				witness_list_lock(&lle->ll_children[i],
1786 				    witness_output);
1787 
1788 			}
1789 		kassert_panic(
1790 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1791 	}
1792 	witness_lock_list_free(lle);
1793 }
1794 
1795 /*
1796  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1797  * exempt Giant and sleepable locks from the checks as well.  If any
1798  * non-exempt locks are held, then a supplied message is printed to the
1799  * output channel along with a list of the offending locks.  If indicated in the
1800  * flags then a failure results in a panic as well.
1801  */
1802 int
1803 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1804 {
1805 	struct lock_list_entry *lock_list, *lle;
1806 	struct lock_instance *lock1;
1807 	struct thread *td;
1808 	va_list ap;
1809 	int i, n;
1810 
1811 	if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1812 		return (0);
1813 	n = 0;
1814 	td = curthread;
1815 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1816 		for (i = lle->ll_count - 1; i >= 0; i--) {
1817 			lock1 = &lle->ll_children[i];
1818 			if (lock1->li_lock == lock)
1819 				continue;
1820 			if (flags & WARN_GIANTOK &&
1821 			    lock1->li_lock == &Giant.lock_object)
1822 				continue;
1823 			if (flags & WARN_SLEEPOK &&
1824 			    (lock1->li_flags & LI_SLEEPABLE) != 0)
1825 				continue;
1826 			if (n == 0) {
1827 				va_start(ap, fmt);
1828 				vprintf(fmt, ap);
1829 				va_end(ap);
1830 				printf(" with the following %slocks held:\n",
1831 				    (flags & WARN_SLEEPOK) != 0 ?
1832 				    "non-sleepable " : "");
1833 			}
1834 			n++;
1835 			witness_list_lock(lock1, printf);
1836 		}
1837 
1838 	/*
1839 	 * Pin the thread in order to avoid problems with thread migration.
1840 	 * Once that all verifies are passed about spinlocks ownership,
1841 	 * the thread is in a safe path and it can be unpinned.
1842 	 */
1843 	sched_pin();
1844 	lock_list = PCPU_GET(spinlocks);
1845 	if (lock_list != NULL && lock_list->ll_count != 0) {
1846 		sched_unpin();
1847 
1848 		/*
1849 		 * We should only have one spinlock and as long as
1850 		 * the flags cannot match for this locks class,
1851 		 * check if the first spinlock is the one curthread
1852 		 * should hold.
1853 		 */
1854 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1855 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1856 		    lock1->li_lock == lock && n == 0)
1857 			return (0);
1858 
1859 		va_start(ap, fmt);
1860 		vprintf(fmt, ap);
1861 		va_end(ap);
1862 		printf(" with the following %slocks held:\n",
1863 		    (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1864 		n += witness_list_locks(&lock_list, printf);
1865 	} else
1866 		sched_unpin();
1867 	if (flags & WARN_PANIC && n)
1868 		kassert_panic("%s", __func__);
1869 	else
1870 		witness_debugger(n, __func__);
1871 	return (n);
1872 }
1873 
1874 const char *
1875 witness_file(struct lock_object *lock)
1876 {
1877 	struct witness *w;
1878 
1879 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1880 		return ("?");
1881 	w = lock->lo_witness;
1882 	return (w->w_file);
1883 }
1884 
1885 int
1886 witness_line(struct lock_object *lock)
1887 {
1888 	struct witness *w;
1889 
1890 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1891 		return (0);
1892 	w = lock->lo_witness;
1893 	return (w->w_line);
1894 }
1895 
1896 static struct witness *
1897 enroll(const char *description, struct lock_class *lock_class)
1898 {
1899 	struct witness *w;
1900 
1901 	MPASS(description != NULL);
1902 
1903 	if (witness_watch == -1 || KERNEL_PANICKED())
1904 		return (NULL);
1905 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1906 		if (witness_skipspin)
1907 			return (NULL);
1908 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1909 		kassert_panic("lock class %s is not sleep or spin",
1910 		    lock_class->lc_name);
1911 		return (NULL);
1912 	}
1913 
1914 	mtx_lock_spin(&w_mtx);
1915 	w = witness_hash_get(description);
1916 	if (w)
1917 		goto found;
1918 	if ((w = witness_get()) == NULL)
1919 		return (NULL);
1920 	MPASS(strlen(description) < MAX_W_NAME);
1921 	strcpy(w->w_name, description);
1922 	w->w_class = lock_class;
1923 	w->w_refcount = 1;
1924 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1925 	if (lock_class->lc_flags & LC_SPINLOCK) {
1926 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1927 		w_spin_cnt++;
1928 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1929 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1930 		w_sleep_cnt++;
1931 	}
1932 
1933 	/* Insert new witness into the hash */
1934 	witness_hash_put(w);
1935 	witness_increment_graph_generation();
1936 	mtx_unlock_spin(&w_mtx);
1937 	return (w);
1938 found:
1939 	w->w_refcount++;
1940 	if (w->w_refcount == 1)
1941 		w->w_class = lock_class;
1942 	mtx_unlock_spin(&w_mtx);
1943 	if (lock_class != w->w_class)
1944 		kassert_panic(
1945 		    "lock (%s) %s does not match earlier (%s) lock",
1946 		    description, lock_class->lc_name,
1947 		    w->w_class->lc_name);
1948 	return (w);
1949 }
1950 
1951 static void
1952 depart(struct witness *w)
1953 {
1954 
1955 	MPASS(w->w_refcount == 0);
1956 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1957 		w_sleep_cnt--;
1958 	} else {
1959 		w_spin_cnt--;
1960 	}
1961 	/*
1962 	 * Set file to NULL as it may point into a loadable module.
1963 	 */
1964 	w->w_file = NULL;
1965 	w->w_line = 0;
1966 	witness_increment_graph_generation();
1967 }
1968 
1969 static void
1970 adopt(struct witness *parent, struct witness *child)
1971 {
1972 	int pi, ci, i, j;
1973 
1974 	if (witness_cold == 0)
1975 		mtx_assert(&w_mtx, MA_OWNED);
1976 
1977 	/* If the relationship is already known, there's no work to be done. */
1978 	if (isitmychild(parent, child))
1979 		return;
1980 
1981 	/* When the structure of the graph changes, bump up the generation. */
1982 	witness_increment_graph_generation();
1983 
1984 	/*
1985 	 * The hard part ... create the direct relationship, then propagate all
1986 	 * indirect relationships.
1987 	 */
1988 	pi = parent->w_index;
1989 	ci = child->w_index;
1990 	WITNESS_INDEX_ASSERT(pi);
1991 	WITNESS_INDEX_ASSERT(ci);
1992 	MPASS(pi != ci);
1993 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1994 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1995 
1996 	/*
1997 	 * If parent was not already an ancestor of child,
1998 	 * then we increment the descendant and ancestor counters.
1999 	 */
2000 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
2001 		parent->w_num_descendants++;
2002 		child->w_num_ancestors++;
2003 	}
2004 
2005 	/*
2006 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2007 	 * an ancestor of 'pi' during this loop.
2008 	 */
2009 	for (i = 1; i <= w_max_used_index; i++) {
2010 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2011 		    (i != pi))
2012 			continue;
2013 
2014 		/* Find each descendant of 'i' and mark it as a descendant. */
2015 		for (j = 1; j <= w_max_used_index; j++) {
2016 			/*
2017 			 * Skip children that are already marked as
2018 			 * descendants of 'i'.
2019 			 */
2020 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2021 				continue;
2022 
2023 			/*
2024 			 * We are only interested in descendants of 'ci'. Note
2025 			 * that 'ci' itself is counted as a descendant of 'ci'.
2026 			 */
2027 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2028 			    (j != ci))
2029 				continue;
2030 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2031 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2032 			w_data[i].w_num_descendants++;
2033 			w_data[j].w_num_ancestors++;
2034 
2035 			/*
2036 			 * Make sure we aren't marking a node as both an
2037 			 * ancestor and descendant. We should have caught
2038 			 * this as a lock order reversal earlier.
2039 			 */
2040 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2041 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2042 				printf("witness rmatrix paradox! [%d][%d]=%d "
2043 				    "both ancestor and descendant\n",
2044 				    i, j, w_rmatrix[i][j]);
2045 				kdb_backtrace();
2046 				printf("Witness disabled.\n");
2047 				witness_watch = -1;
2048 			}
2049 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2050 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2051 				printf("witness rmatrix paradox! [%d][%d]=%d "
2052 				    "both ancestor and descendant\n",
2053 				    j, i, w_rmatrix[j][i]);
2054 				kdb_backtrace();
2055 				printf("Witness disabled.\n");
2056 				witness_watch = -1;
2057 			}
2058 		}
2059 	}
2060 }
2061 
2062 static void
2063 itismychild(struct witness *parent, struct witness *child)
2064 {
2065 	int unlocked;
2066 
2067 	MPASS(child != NULL && parent != NULL);
2068 	if (witness_cold == 0)
2069 		mtx_assert(&w_mtx, MA_OWNED);
2070 
2071 	if (!witness_lock_type_equal(parent, child)) {
2072 		if (witness_cold == 0) {
2073 			unlocked = 1;
2074 			mtx_unlock_spin(&w_mtx);
2075 		} else {
2076 			unlocked = 0;
2077 		}
2078 		kassert_panic(
2079 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2080 		    "the same lock type", __func__, parent->w_name,
2081 		    parent->w_class->lc_name, child->w_name,
2082 		    child->w_class->lc_name);
2083 		if (unlocked)
2084 			mtx_lock_spin(&w_mtx);
2085 	}
2086 	adopt(parent, child);
2087 }
2088 
2089 /*
2090  * Generic code for the isitmy*() functions. The rmask parameter is the
2091  * expected relationship of w1 to w2.
2092  */
2093 static int
2094 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2095 {
2096 	unsigned char r1, r2;
2097 	int i1, i2;
2098 
2099 	i1 = w1->w_index;
2100 	i2 = w2->w_index;
2101 	WITNESS_INDEX_ASSERT(i1);
2102 	WITNESS_INDEX_ASSERT(i2);
2103 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2104 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2105 
2106 	/* The flags on one better be the inverse of the flags on the other */
2107 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2108 	    (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2109 		/* Don't squawk if we're potentially racing with an update. */
2110 		if (!mtx_owned(&w_mtx))
2111 			return (0);
2112 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2113 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2114 		    "w_rmatrix[%d][%d] == %hhx\n",
2115 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2116 		    i2, i1, r2);
2117 		kdb_backtrace();
2118 		printf("Witness disabled.\n");
2119 		witness_watch = -1;
2120 	}
2121 	return (r1 & rmask);
2122 }
2123 
2124 /*
2125  * Checks if @child is a direct child of @parent.
2126  */
2127 static int
2128 isitmychild(struct witness *parent, struct witness *child)
2129 {
2130 
2131 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2132 }
2133 
2134 /*
2135  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2136  */
2137 static int
2138 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2139 {
2140 
2141 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2142 	    __func__));
2143 }
2144 
2145 static int
2146 blessed(struct witness *w1, struct witness *w2)
2147 {
2148 	int i;
2149 	struct witness_blessed *b;
2150 
2151 	for (i = 0; i < nitems(blessed_list); i++) {
2152 		b = &blessed_list[i];
2153 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2154 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2155 				return (1);
2156 			continue;
2157 		}
2158 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2159 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2160 				return (1);
2161 	}
2162 	return (0);
2163 }
2164 
2165 static struct witness *
2166 witness_get(void)
2167 {
2168 	struct witness *w;
2169 	int index;
2170 
2171 	if (witness_cold == 0)
2172 		mtx_assert(&w_mtx, MA_OWNED);
2173 
2174 	if (witness_watch == -1) {
2175 		mtx_unlock_spin(&w_mtx);
2176 		return (NULL);
2177 	}
2178 	if (STAILQ_EMPTY(&w_free)) {
2179 		witness_watch = -1;
2180 		mtx_unlock_spin(&w_mtx);
2181 		printf("WITNESS: unable to allocate a new witness object\n");
2182 		return (NULL);
2183 	}
2184 	w = STAILQ_FIRST(&w_free);
2185 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2186 	w_free_cnt--;
2187 	index = w->w_index;
2188 	MPASS(index > 0 && index == w_max_used_index+1 &&
2189 	    index < witness_count);
2190 	bzero(w, sizeof(*w));
2191 	w->w_index = index;
2192 	if (index > w_max_used_index)
2193 		w_max_used_index = index;
2194 	return (w);
2195 }
2196 
2197 static void
2198 witness_free(struct witness *w)
2199 {
2200 
2201 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2202 	w_free_cnt++;
2203 }
2204 
2205 static struct lock_list_entry *
2206 witness_lock_list_get(void)
2207 {
2208 	struct lock_list_entry *lle;
2209 
2210 	if (witness_watch == -1)
2211 		return (NULL);
2212 	mtx_lock_spin(&w_mtx);
2213 	lle = w_lock_list_free;
2214 	if (lle == NULL) {
2215 		witness_watch = -1;
2216 		mtx_unlock_spin(&w_mtx);
2217 		printf("%s: witness exhausted\n", __func__);
2218 		return (NULL);
2219 	}
2220 	w_lock_list_free = lle->ll_next;
2221 	mtx_unlock_spin(&w_mtx);
2222 	bzero(lle, sizeof(*lle));
2223 	return (lle);
2224 }
2225 
2226 static void
2227 witness_lock_list_free(struct lock_list_entry *lle)
2228 {
2229 
2230 	mtx_lock_spin(&w_mtx);
2231 	lle->ll_next = w_lock_list_free;
2232 	w_lock_list_free = lle;
2233 	mtx_unlock_spin(&w_mtx);
2234 }
2235 
2236 static struct lock_instance *
2237 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2238 {
2239 	struct lock_list_entry *lle;
2240 	struct lock_instance *instance;
2241 	int i;
2242 
2243 	for (lle = list; lle != NULL; lle = lle->ll_next)
2244 		for (i = lle->ll_count - 1; i >= 0; i--) {
2245 			instance = &lle->ll_children[i];
2246 			if (instance->li_lock == lock)
2247 				return (instance);
2248 		}
2249 	return (NULL);
2250 }
2251 
2252 static void
2253 witness_list_lock(struct lock_instance *instance,
2254     int (*prnt)(const char *fmt, ...))
2255 {
2256 	struct lock_object *lock;
2257 
2258 	lock = instance->li_lock;
2259 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2260 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2261 	if (lock->lo_witness->w_name != lock->lo_name)
2262 		prnt(" (%s)", lock->lo_witness->w_name);
2263 	prnt(" r = %d (%p) locked @ %s:%d\n",
2264 	    instance->li_flags & LI_RECURSEMASK, lock,
2265 	    fixup_filename(instance->li_file), instance->li_line);
2266 }
2267 
2268 static int
2269 witness_output(const char *fmt, ...)
2270 {
2271 	va_list ap;
2272 	int ret;
2273 
2274 	va_start(ap, fmt);
2275 	ret = witness_voutput(fmt, ap);
2276 	va_end(ap);
2277 	return (ret);
2278 }
2279 
2280 static int
2281 witness_voutput(const char *fmt, va_list ap)
2282 {
2283 	int ret;
2284 
2285 	ret = 0;
2286 	switch (witness_channel) {
2287 	case WITNESS_CONSOLE:
2288 		ret = vprintf(fmt, ap);
2289 		break;
2290 	case WITNESS_LOG:
2291 		vlog(LOG_NOTICE, fmt, ap);
2292 		break;
2293 	case WITNESS_NONE:
2294 		break;
2295 	}
2296 	return (ret);
2297 }
2298 
2299 #ifdef DDB
2300 static int
2301 witness_thread_has_locks(struct thread *td)
2302 {
2303 
2304 	if (td->td_sleeplocks == NULL)
2305 		return (0);
2306 	return (td->td_sleeplocks->ll_count != 0);
2307 }
2308 
2309 static int
2310 witness_proc_has_locks(struct proc *p)
2311 {
2312 	struct thread *td;
2313 
2314 	FOREACH_THREAD_IN_PROC(p, td) {
2315 		if (witness_thread_has_locks(td))
2316 			return (1);
2317 	}
2318 	return (0);
2319 }
2320 #endif
2321 
2322 int
2323 witness_list_locks(struct lock_list_entry **lock_list,
2324     int (*prnt)(const char *fmt, ...))
2325 {
2326 	struct lock_list_entry *lle;
2327 	int i, nheld;
2328 
2329 	nheld = 0;
2330 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2331 		for (i = lle->ll_count - 1; i >= 0; i--) {
2332 			witness_list_lock(&lle->ll_children[i], prnt);
2333 			nheld++;
2334 		}
2335 	return (nheld);
2336 }
2337 
2338 /*
2339  * This is a bit risky at best.  We call this function when we have timed
2340  * out acquiring a spin lock, and we assume that the other CPU is stuck
2341  * with this lock held.  So, we go groveling around in the other CPU's
2342  * per-cpu data to try to find the lock instance for this spin lock to
2343  * see when it was last acquired.
2344  */
2345 void
2346 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2347     int (*prnt)(const char *fmt, ...))
2348 {
2349 	struct lock_instance *instance;
2350 	struct pcpu *pc;
2351 
2352 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2353 		return;
2354 	pc = pcpu_find(owner->td_oncpu);
2355 	instance = find_instance(pc->pc_spinlocks, lock);
2356 	if (instance != NULL)
2357 		witness_list_lock(instance, prnt);
2358 }
2359 
2360 void
2361 witness_save(struct lock_object *lock, const char **filep, int *linep)
2362 {
2363 	struct lock_list_entry *lock_list;
2364 	struct lock_instance *instance;
2365 	struct lock_class *class;
2366 
2367 	/*
2368 	 * This function is used independently in locking code to deal with
2369 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2370 	 * is gone.
2371 	 */
2372 	if (SCHEDULER_STOPPED())
2373 		return;
2374 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2375 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2376 		return;
2377 	class = LOCK_CLASS(lock);
2378 	if (class->lc_flags & LC_SLEEPLOCK)
2379 		lock_list = curthread->td_sleeplocks;
2380 	else {
2381 		if (witness_skipspin)
2382 			return;
2383 		lock_list = PCPU_GET(spinlocks);
2384 	}
2385 	instance = find_instance(lock_list, lock);
2386 	if (instance == NULL) {
2387 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2388 		    class->lc_name, lock->lo_name);
2389 		return;
2390 	}
2391 	*filep = instance->li_file;
2392 	*linep = instance->li_line;
2393 }
2394 
2395 void
2396 witness_restore(struct lock_object *lock, const char *file, int line)
2397 {
2398 	struct lock_list_entry *lock_list;
2399 	struct lock_instance *instance;
2400 	struct lock_class *class;
2401 
2402 	/*
2403 	 * This function is used independently in locking code to deal with
2404 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2405 	 * is gone.
2406 	 */
2407 	if (SCHEDULER_STOPPED())
2408 		return;
2409 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2410 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2411 		return;
2412 	class = LOCK_CLASS(lock);
2413 	if (class->lc_flags & LC_SLEEPLOCK)
2414 		lock_list = curthread->td_sleeplocks;
2415 	else {
2416 		if (witness_skipspin)
2417 			return;
2418 		lock_list = PCPU_GET(spinlocks);
2419 	}
2420 	instance = find_instance(lock_list, lock);
2421 	if (instance == NULL)
2422 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2423 		    class->lc_name, lock->lo_name);
2424 	lock->lo_witness->w_file = file;
2425 	lock->lo_witness->w_line = line;
2426 	if (instance == NULL)
2427 		return;
2428 	instance->li_file = file;
2429 	instance->li_line = line;
2430 }
2431 
2432 static bool
2433 witness_find_instance(const struct lock_object *lock,
2434     struct lock_instance **instance)
2435 {
2436 #ifdef INVARIANT_SUPPORT
2437 	struct lock_class *class;
2438 
2439 	if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2440 		return (false);
2441 	class = LOCK_CLASS(lock);
2442 	if ((class->lc_flags & LC_SLEEPLOCK) != 0) {
2443 		*instance = find_instance(curthread->td_sleeplocks, lock);
2444 		return (true);
2445 	} else if ((class->lc_flags & LC_SPINLOCK) != 0) {
2446 		*instance = find_instance(PCPU_GET(spinlocks), lock);
2447 		return (true);
2448 	} else {
2449 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2450 		    class->lc_name, lock->lo_name);
2451 		return (false);
2452 	}
2453 #else
2454 	return (false);
2455 #endif
2456 }
2457 
2458 void
2459 witness_assert(const struct lock_object *lock, int flags, const char *file,
2460     int line)
2461 {
2462 #ifdef INVARIANT_SUPPORT
2463 	struct lock_instance *instance;
2464 	struct lock_class *class;
2465 
2466 	if (!witness_find_instance(lock, &instance))
2467 		return;
2468 	class = LOCK_CLASS(lock);
2469 	switch (flags) {
2470 	case LA_UNLOCKED:
2471 		if (instance != NULL)
2472 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2473 			    class->lc_name, lock->lo_name,
2474 			    fixup_filename(file), line);
2475 		break;
2476 	case LA_LOCKED:
2477 	case LA_LOCKED | LA_RECURSED:
2478 	case LA_LOCKED | LA_NOTRECURSED:
2479 	case LA_SLOCKED:
2480 	case LA_SLOCKED | LA_RECURSED:
2481 	case LA_SLOCKED | LA_NOTRECURSED:
2482 	case LA_XLOCKED:
2483 	case LA_XLOCKED | LA_RECURSED:
2484 	case LA_XLOCKED | LA_NOTRECURSED:
2485 		if (instance == NULL) {
2486 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2487 			    class->lc_name, lock->lo_name,
2488 			    fixup_filename(file), line);
2489 			break;
2490 		}
2491 		if ((flags & LA_XLOCKED) != 0 &&
2492 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2493 			kassert_panic(
2494 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2495 			    class->lc_name, lock->lo_name,
2496 			    fixup_filename(file), line);
2497 		if ((flags & LA_SLOCKED) != 0 &&
2498 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2499 			kassert_panic(
2500 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2501 			    class->lc_name, lock->lo_name,
2502 			    fixup_filename(file), line);
2503 		if ((flags & LA_RECURSED) != 0 &&
2504 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2505 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2506 			    class->lc_name, lock->lo_name,
2507 			    fixup_filename(file), line);
2508 		if ((flags & LA_NOTRECURSED) != 0 &&
2509 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2510 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2511 			    class->lc_name, lock->lo_name,
2512 			    fixup_filename(file), line);
2513 		break;
2514 	default:
2515 		kassert_panic("Invalid lock assertion at %s:%d.",
2516 		    fixup_filename(file), line);
2517 	}
2518 #endif	/* INVARIANT_SUPPORT */
2519 }
2520 
2521 /*
2522  * Checks the ownership of the lock by curthread, consulting the witness list.
2523  * Returns:
2524  *   0  if witness is disabled or did not work
2525  *   -1 if not owned
2526  *   1  if owned
2527  */
2528 int
2529 witness_is_owned(const struct lock_object *lock)
2530 {
2531 #ifdef INVARIANT_SUPPORT
2532 	struct lock_instance *instance;
2533 
2534 	if (!witness_find_instance(lock, &instance))
2535 		return (0);
2536 	return (instance == NULL ? -1 : 1);
2537 #else
2538 	return (0);
2539 #endif
2540 }
2541 
2542 static void
2543 witness_setflag(struct lock_object *lock, int flag, int set)
2544 {
2545 	struct lock_list_entry *lock_list;
2546 	struct lock_instance *instance;
2547 	struct lock_class *class;
2548 
2549 	if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2550 		return;
2551 	class = LOCK_CLASS(lock);
2552 	if (class->lc_flags & LC_SLEEPLOCK)
2553 		lock_list = curthread->td_sleeplocks;
2554 	else {
2555 		if (witness_skipspin)
2556 			return;
2557 		lock_list = PCPU_GET(spinlocks);
2558 	}
2559 	instance = find_instance(lock_list, lock);
2560 	if (instance == NULL) {
2561 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2562 		    class->lc_name, lock->lo_name);
2563 		return;
2564 	}
2565 
2566 	if (set)
2567 		instance->li_flags |= flag;
2568 	else
2569 		instance->li_flags &= ~flag;
2570 }
2571 
2572 void
2573 witness_norelease(struct lock_object *lock)
2574 {
2575 
2576 	witness_setflag(lock, LI_NORELEASE, 1);
2577 }
2578 
2579 void
2580 witness_releaseok(struct lock_object *lock)
2581 {
2582 
2583 	witness_setflag(lock, LI_NORELEASE, 0);
2584 }
2585 
2586 #ifdef DDB
2587 static void
2588 witness_ddb_list(struct thread *td)
2589 {
2590 
2591 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2592 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2593 
2594 	if (witness_watch < 1)
2595 		return;
2596 
2597 	witness_list_locks(&td->td_sleeplocks, db_printf);
2598 
2599 	/*
2600 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2601 	 * if td is currently executing on some other CPU and holds spin locks
2602 	 * as we won't display those locks.  If we had a MI way of getting
2603 	 * the per-cpu data for a given cpu then we could use
2604 	 * td->td_oncpu to get the list of spinlocks for this thread
2605 	 * and "fix" this.
2606 	 *
2607 	 * That still wouldn't really fix this unless we locked the scheduler
2608 	 * lock or stopped the other CPU to make sure it wasn't changing the
2609 	 * list out from under us.  It is probably best to just not try to
2610 	 * handle threads on other CPU's for now.
2611 	 */
2612 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2613 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2614 }
2615 
2616 DB_SHOW_COMMAND(locks, db_witness_list)
2617 {
2618 	struct thread *td;
2619 
2620 	if (have_addr)
2621 		td = db_lookup_thread(addr, true);
2622 	else
2623 		td = kdb_thread;
2624 	witness_ddb_list(td);
2625 }
2626 
2627 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2628 {
2629 	struct thread *td;
2630 	struct proc *p;
2631 
2632 	/*
2633 	 * It would be nice to list only threads and processes that actually
2634 	 * held sleep locks, but that information is currently not exported
2635 	 * by WITNESS.
2636 	 */
2637 	FOREACH_PROC_IN_SYSTEM(p) {
2638 		if (!witness_proc_has_locks(p))
2639 			continue;
2640 		FOREACH_THREAD_IN_PROC(p, td) {
2641 			if (!witness_thread_has_locks(td))
2642 				continue;
2643 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2644 			    p->p_comm, td, td->td_tid);
2645 			witness_ddb_list(td);
2646 			if (db_pager_quit)
2647 				return;
2648 		}
2649 	}
2650 }
2651 DB_SHOW_ALIAS_FLAGS(alllocks, db_witness_list_all, DB_CMD_MEMSAFE);
2652 
2653 DB_SHOW_COMMAND_FLAGS(witness, db_witness_display, DB_CMD_MEMSAFE)
2654 {
2655 
2656 	witness_ddb_display(db_printf);
2657 }
2658 #endif
2659 
2660 static void
2661 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2662 {
2663 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2664 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2665 	int generation, i, j;
2666 
2667 	tmp_data1 = NULL;
2668 	tmp_data2 = NULL;
2669 	tmp_w1 = NULL;
2670 	tmp_w2 = NULL;
2671 
2672 	/* Allocate and init temporary storage space. */
2673 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2674 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2675 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2676 	    M_WAITOK | M_ZERO);
2677 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2678 	    M_WAITOK | M_ZERO);
2679 	stack_zero(&tmp_data1->wlod_stack);
2680 	stack_zero(&tmp_data2->wlod_stack);
2681 
2682 restart:
2683 	mtx_lock_spin(&w_mtx);
2684 	generation = w_generation;
2685 	mtx_unlock_spin(&w_mtx);
2686 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2687 	    w_lohash.wloh_count);
2688 	for (i = 1; i < w_max_used_index; i++) {
2689 		mtx_lock_spin(&w_mtx);
2690 		if (generation != w_generation) {
2691 			mtx_unlock_spin(&w_mtx);
2692 
2693 			/* The graph has changed, try again. */
2694 			*oldidx = 0;
2695 			sbuf_clear(sb);
2696 			goto restart;
2697 		}
2698 
2699 		w1 = &w_data[i];
2700 		if (w1->w_reversed == 0) {
2701 			mtx_unlock_spin(&w_mtx);
2702 			continue;
2703 		}
2704 
2705 		/* Copy w1 locally so we can release the spin lock. */
2706 		*tmp_w1 = *w1;
2707 		mtx_unlock_spin(&w_mtx);
2708 
2709 		if (tmp_w1->w_reversed == 0)
2710 			continue;
2711 		for (j = 1; j < w_max_used_index; j++) {
2712 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2713 				continue;
2714 
2715 			mtx_lock_spin(&w_mtx);
2716 			if (generation != w_generation) {
2717 				mtx_unlock_spin(&w_mtx);
2718 
2719 				/* The graph has changed, try again. */
2720 				*oldidx = 0;
2721 				sbuf_clear(sb);
2722 				goto restart;
2723 			}
2724 
2725 			w2 = &w_data[j];
2726 			data1 = witness_lock_order_get(w1, w2);
2727 			data2 = witness_lock_order_get(w2, w1);
2728 
2729 			/*
2730 			 * Copy information locally so we can release the
2731 			 * spin lock.
2732 			 */
2733 			*tmp_w2 = *w2;
2734 
2735 			if (data1) {
2736 				stack_zero(&tmp_data1->wlod_stack);
2737 				stack_copy(&data1->wlod_stack,
2738 				    &tmp_data1->wlod_stack);
2739 			}
2740 			if (data2 && data2 != data1) {
2741 				stack_zero(&tmp_data2->wlod_stack);
2742 				stack_copy(&data2->wlod_stack,
2743 				    &tmp_data2->wlod_stack);
2744 			}
2745 			mtx_unlock_spin(&w_mtx);
2746 
2747 			if (blessed(tmp_w1, tmp_w2))
2748 				continue;
2749 
2750 			sbuf_printf(sb,
2751 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2752 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2753 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2754 			if (data1) {
2755 				sbuf_printf(sb,
2756 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2757 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2758 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2759 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2760 				sbuf_printf(sb, "\n");
2761 			}
2762 			if (data2 && data2 != data1) {
2763 				sbuf_printf(sb,
2764 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2765 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2766 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2767 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2768 				sbuf_printf(sb, "\n");
2769 			}
2770 		}
2771 	}
2772 	mtx_lock_spin(&w_mtx);
2773 	if (generation != w_generation) {
2774 		mtx_unlock_spin(&w_mtx);
2775 
2776 		/*
2777 		 * The graph changed while we were printing stack data,
2778 		 * try again.
2779 		 */
2780 		*oldidx = 0;
2781 		sbuf_clear(sb);
2782 		goto restart;
2783 	}
2784 	mtx_unlock_spin(&w_mtx);
2785 
2786 	/* Free temporary storage space. */
2787 	free(tmp_data1, M_TEMP);
2788 	free(tmp_data2, M_TEMP);
2789 	free(tmp_w1, M_TEMP);
2790 	free(tmp_w2, M_TEMP);
2791 }
2792 
2793 static int
2794 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2795 {
2796 	struct sbuf *sb;
2797 	int error;
2798 
2799 	if (witness_watch < 1) {
2800 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2801 		return (error);
2802 	}
2803 	if (witness_cold) {
2804 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2805 		return (error);
2806 	}
2807 	error = 0;
2808 	sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2809 	if (sb == NULL)
2810 		return (ENOMEM);
2811 
2812 	sbuf_print_witness_badstacks(sb, &req->oldidx);
2813 
2814 	sbuf_finish(sb);
2815 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2816 	sbuf_delete(sb);
2817 
2818 	return (error);
2819 }
2820 
2821 #ifdef DDB
2822 static int
2823 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2824 {
2825 
2826 	return (db_printf("%.*s", len, data));
2827 }
2828 
2829 DB_SHOW_COMMAND_FLAGS(badstacks, db_witness_badstacks, DB_CMD_MEMSAFE)
2830 {
2831 	struct sbuf sb;
2832 	char buffer[128];
2833 	size_t dummy;
2834 
2835 	sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2836 	sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2837 	sbuf_print_witness_badstacks(&sb, &dummy);
2838 	sbuf_finish(&sb);
2839 }
2840 #endif
2841 
2842 static int
2843 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2844 {
2845 	static const struct {
2846 		enum witness_channel channel;
2847 		const char *name;
2848 	} channels[] = {
2849 		{ WITNESS_CONSOLE, "console" },
2850 		{ WITNESS_LOG, "log" },
2851 		{ WITNESS_NONE, "none" },
2852 	};
2853 	char buf[16];
2854 	u_int i;
2855 	int error;
2856 
2857 	buf[0] = '\0';
2858 	for (i = 0; i < nitems(channels); i++)
2859 		if (witness_channel == channels[i].channel) {
2860 			snprintf(buf, sizeof(buf), "%s", channels[i].name);
2861 			break;
2862 		}
2863 
2864 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2865 	if (error != 0 || req->newptr == NULL)
2866 		return (error);
2867 
2868 	error = EINVAL;
2869 	for (i = 0; i < nitems(channels); i++)
2870 		if (strcmp(channels[i].name, buf) == 0) {
2871 			witness_channel = channels[i].channel;
2872 			error = 0;
2873 			break;
2874 		}
2875 	return (error);
2876 }
2877 
2878 static int
2879 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2880 {
2881 	struct witness *w;
2882 	struct sbuf *sb;
2883 	int error;
2884 
2885 #ifdef __i386__
2886 	error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2887 	return (error);
2888 #endif
2889 
2890 	if (witness_watch < 1) {
2891 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2892 		return (error);
2893 	}
2894 	if (witness_cold) {
2895 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2896 		return (error);
2897 	}
2898 	error = 0;
2899 
2900 	error = sysctl_wire_old_buffer(req, 0);
2901 	if (error != 0)
2902 		return (error);
2903 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2904 	if (sb == NULL)
2905 		return (ENOMEM);
2906 	sbuf_printf(sb, "\n");
2907 
2908 	mtx_lock_spin(&w_mtx);
2909 	STAILQ_FOREACH(w, &w_all, w_list)
2910 		w->w_displayed = 0;
2911 	STAILQ_FOREACH(w, &w_all, w_list)
2912 		witness_add_fullgraph(sb, w);
2913 	mtx_unlock_spin(&w_mtx);
2914 
2915 	/*
2916 	 * Close the sbuf and return to userland.
2917 	 */
2918 	error = sbuf_finish(sb);
2919 	sbuf_delete(sb);
2920 
2921 	return (error);
2922 }
2923 
2924 static int
2925 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2926 {
2927 	int error, value;
2928 
2929 	value = witness_watch;
2930 	error = sysctl_handle_int(oidp, &value, 0, req);
2931 	if (error != 0 || req->newptr == NULL)
2932 		return (error);
2933 	if (value > 1 || value < -1 ||
2934 	    (witness_watch == -1 && value != witness_watch))
2935 		return (EINVAL);
2936 	witness_watch = value;
2937 	return (0);
2938 }
2939 
2940 static void
2941 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2942 {
2943 	int i;
2944 
2945 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2946 		return;
2947 	w->w_displayed = 1;
2948 
2949 	WITNESS_INDEX_ASSERT(w->w_index);
2950 	for (i = 1; i <= w_max_used_index; i++) {
2951 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2952 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2953 			    w_data[i].w_name);
2954 			witness_add_fullgraph(sb, &w_data[i]);
2955 		}
2956 	}
2957 }
2958 
2959 /*
2960  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2961  * interprets the key as a string and reads until the null
2962  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2963  * hash value computed from the key.
2964  */
2965 static uint32_t
2966 witness_hash_djb2(const uint8_t *key, uint32_t size)
2967 {
2968 	unsigned int hash = 5381;
2969 	int i;
2970 
2971 	/* hash = hash * 33 + key[i] */
2972 	if (size)
2973 		for (i = 0; i < size; i++)
2974 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2975 	else
2976 		for (i = 0; key[i] != 0; i++)
2977 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2978 
2979 	return (hash);
2980 }
2981 
2982 /*
2983  * Initializes the two witness hash tables. Called exactly once from
2984  * witness_initialize().
2985  */
2986 static void
2987 witness_init_hash_tables(void)
2988 {
2989 	int i;
2990 
2991 	MPASS(witness_cold);
2992 
2993 	/* Initialize the hash tables. */
2994 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2995 		w_hash.wh_array[i] = NULL;
2996 
2997 	w_hash.wh_size = WITNESS_HASH_SIZE;
2998 	w_hash.wh_count = 0;
2999 
3000 	/* Initialize the lock order data hash. */
3001 	w_lofree = NULL;
3002 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
3003 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
3004 		w_lodata[i].wlod_next = w_lofree;
3005 		w_lofree = &w_lodata[i];
3006 	}
3007 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
3008 	w_lohash.wloh_count = 0;
3009 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
3010 		w_lohash.wloh_array[i] = NULL;
3011 }
3012 
3013 static struct witness *
3014 witness_hash_get(const char *key)
3015 {
3016 	struct witness *w;
3017 	uint32_t hash;
3018 
3019 	MPASS(key != NULL);
3020 	if (witness_cold == 0)
3021 		mtx_assert(&w_mtx, MA_OWNED);
3022 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
3023 	w = w_hash.wh_array[hash];
3024 	while (w != NULL) {
3025 		if (strcmp(w->w_name, key) == 0)
3026 			goto out;
3027 		w = w->w_hash_next;
3028 	}
3029 
3030 out:
3031 	return (w);
3032 }
3033 
3034 static void
3035 witness_hash_put(struct witness *w)
3036 {
3037 	uint32_t hash;
3038 
3039 	MPASS(w != NULL);
3040 	MPASS(w->w_name != NULL);
3041 	if (witness_cold == 0)
3042 		mtx_assert(&w_mtx, MA_OWNED);
3043 	KASSERT(witness_hash_get(w->w_name) == NULL,
3044 	    ("%s: trying to add a hash entry that already exists!", __func__));
3045 	KASSERT(w->w_hash_next == NULL,
3046 	    ("%s: w->w_hash_next != NULL", __func__));
3047 
3048 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3049 	w->w_hash_next = w_hash.wh_array[hash];
3050 	w_hash.wh_array[hash] = w;
3051 	w_hash.wh_count++;
3052 }
3053 
3054 static struct witness_lock_order_data *
3055 witness_lock_order_get(struct witness *parent, struct witness *child)
3056 {
3057 	struct witness_lock_order_data *data = NULL;
3058 	struct witness_lock_order_key key;
3059 	unsigned int hash;
3060 
3061 	MPASS(parent != NULL && child != NULL);
3062 	key.from = parent->w_index;
3063 	key.to = child->w_index;
3064 	WITNESS_INDEX_ASSERT(key.from);
3065 	WITNESS_INDEX_ASSERT(key.to);
3066 	if ((w_rmatrix[parent->w_index][child->w_index]
3067 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
3068 		goto out;
3069 
3070 	hash = witness_hash_djb2((const char*)&key,
3071 	    sizeof(key)) % w_lohash.wloh_size;
3072 	data = w_lohash.wloh_array[hash];
3073 	while (data != NULL) {
3074 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
3075 			break;
3076 		data = data->wlod_next;
3077 	}
3078 
3079 out:
3080 	return (data);
3081 }
3082 
3083 /*
3084  * Verify that parent and child have a known relationship, are not the same,
3085  * and child is actually a child of parent.  This is done without w_mtx
3086  * to avoid contention in the common case.
3087  */
3088 static int
3089 witness_lock_order_check(struct witness *parent, struct witness *child)
3090 {
3091 
3092 	if (parent != child &&
3093 	    w_rmatrix[parent->w_index][child->w_index]
3094 	    & WITNESS_LOCK_ORDER_KNOWN &&
3095 	    isitmychild(parent, child))
3096 		return (1);
3097 
3098 	return (0);
3099 }
3100 
3101 static int
3102 witness_lock_order_add(struct witness *parent, struct witness *child)
3103 {
3104 	struct witness_lock_order_data *data = NULL;
3105 	struct witness_lock_order_key key;
3106 	unsigned int hash;
3107 
3108 	MPASS(parent != NULL && child != NULL);
3109 	key.from = parent->w_index;
3110 	key.to = child->w_index;
3111 	WITNESS_INDEX_ASSERT(key.from);
3112 	WITNESS_INDEX_ASSERT(key.to);
3113 	if (w_rmatrix[parent->w_index][child->w_index]
3114 	    & WITNESS_LOCK_ORDER_KNOWN)
3115 		return (1);
3116 
3117 	hash = witness_hash_djb2((const char*)&key,
3118 	    sizeof(key)) % w_lohash.wloh_size;
3119 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3120 	data = w_lofree;
3121 	if (data == NULL)
3122 		return (0);
3123 	w_lofree = data->wlod_next;
3124 	data->wlod_next = w_lohash.wloh_array[hash];
3125 	data->wlod_key = key;
3126 	w_lohash.wloh_array[hash] = data;
3127 	w_lohash.wloh_count++;
3128 	stack_save(&data->wlod_stack);
3129 	return (1);
3130 }
3131 
3132 /* Call this whenever the structure of the witness graph changes. */
3133 static void
3134 witness_increment_graph_generation(void)
3135 {
3136 
3137 	if (witness_cold == 0)
3138 		mtx_assert(&w_mtx, MA_OWNED);
3139 	w_generation++;
3140 }
3141 
3142 static int
3143 witness_output_drain(void *arg __unused, const char *data, int len)
3144 {
3145 
3146 	witness_output("%.*s", len, data);
3147 	return (len);
3148 }
3149 
3150 static void
3151 witness_debugger(int cond, const char *msg)
3152 {
3153 	char buf[32];
3154 	struct sbuf sb;
3155 	struct stack st;
3156 
3157 	if (!cond)
3158 		return;
3159 
3160 	if (witness_trace) {
3161 		sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3162 		sbuf_set_drain(&sb, witness_output_drain, NULL);
3163 
3164 		stack_save(&st);
3165 		witness_output("stack backtrace:\n");
3166 		stack_sbuf_print_ddb(&sb, &st);
3167 
3168 		sbuf_finish(&sb);
3169 	}
3170 
3171 	witness_enter_debugger(msg);
3172 }
3173 
3174 static void
3175 witness_enter_debugger(const char *msg)
3176 {
3177 #ifdef KDB
3178 	if (witness_kdb)
3179 		kdb_enter(KDB_WHY_WITNESS, msg);
3180 #endif
3181 }
3182