xref: /freebsd/sys/kern/subr_witness.c (revision 13014ca04aad1931d41958b56f71a2c65b9a7a2c)
1 /*-
2  * Copyright (c) 2008 Isilon Systems, Inc.
3  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4  * Copyright (c) 1998 Berkeley Software Design, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Berkeley Software Design Inc's name may not be used to endorse or
16  *    promote products derived from this software without specific prior
17  *    written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
32  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
33  */
34 
35 /*
36  * Implementation of the `witness' lock verifier.  Originally implemented for
37  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
38  * classes in FreeBSD.
39  */
40 
41 /*
42  *	Main Entry: witness
43  *	Pronunciation: 'wit-n&s
44  *	Function: noun
45  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
46  *	    testimony, witness, from 2wit
47  *	Date: before 12th century
48  *	1 : attestation of a fact or event : TESTIMONY
49  *	2 : one that gives evidence; specifically : one who testifies in
50  *	    a cause or before a judicial tribunal
51  *	3 : one asked to be present at a transaction so as to be able to
52  *	    testify to its having taken place
53  *	4 : one who has personal knowledge of something
54  *	5 a : something serving as evidence or proof : SIGN
55  *	  b : public affirmation by word or example of usually
56  *	      religious faith or conviction <the heroic witness to divine
57  *	      life -- Pilot>
58  *	6 capitalized : a member of the Jehovah's Witnesses
59  */
60 
61 /*
62  * Special rules concerning Giant and lock orders:
63  *
64  * 1) Giant must be acquired before any other mutexes.  Stated another way,
65  *    no other mutex may be held when Giant is acquired.
66  *
67  * 2) Giant must be released when blocking on a sleepable lock.
68  *
69  * This rule is less obvious, but is a result of Giant providing the same
70  * semantics as spl().  Basically, when a thread sleeps, it must release
71  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
72  * 2).
73  *
74  * 3) Giant may be acquired before or after sleepable locks.
75  *
76  * This rule is also not quite as obvious.  Giant may be acquired after
77  * a sleepable lock because it is a non-sleepable lock and non-sleepable
78  * locks may always be acquired while holding a sleepable lock.  The second
79  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
80  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
81  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
82  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
83  * execute.  Thus, acquiring Giant both before and after a sleepable lock
84  * will not result in a lock order reversal.
85  */
86 
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD$");
89 
90 #include "opt_ddb.h"
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
94 
95 #include <sys/param.h>
96 #include <sys/bus.h>
97 #include <sys/kdb.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.h>
100 #include <sys/lock.h>
101 #include <sys/malloc.h>
102 #include <sys/mutex.h>
103 #include <sys/priv.h>
104 #include <sys/proc.h>
105 #include <sys/sbuf.h>
106 #include <sys/stack.h>
107 #include <sys/sysctl.h>
108 #include <sys/systm.h>
109 
110 #ifdef DDB
111 #include <ddb/ddb.h>
112 #endif
113 
114 #include <machine/stdarg.h>
115 
116 #if !defined(DDB) && !defined(STACK)
117 #error "DDB or STACK options are required for WITNESS"
118 #endif
119 
120 /* Note that these traces do not work with KTR_ALQ. */
121 #if 0
122 #define	KTR_WITNESS	KTR_SUBSYS
123 #else
124 #define	KTR_WITNESS	0
125 #endif
126 
127 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
128 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
129 
130 /* Define this to check for blessed mutexes */
131 #undef BLESSING
132 
133 #define	WITNESS_COUNT 		1024
134 #define	WITNESS_CHILDCOUNT 	(WITNESS_COUNT * 4)
135 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
136 #define	WITNESS_PENDLIST	512
137 
138 /* Allocate 256 KB of stack data space */
139 #define	WITNESS_LO_DATA_COUNT	2048
140 
141 /* Prime, gives load factor of ~2 at full load */
142 #define	WITNESS_LO_HASH_SIZE	1021
143 
144 /*
145  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
146  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
147  * probably be safe for the most part, but it's still a SWAG.
148  */
149 #define	LOCK_NCHILDREN	5
150 #define	LOCK_CHILDCOUNT	2048
151 
152 #define	MAX_W_NAME	64
153 
154 #define	BADSTACK_SBUF_SIZE	(256 * WITNESS_COUNT)
155 #define	CYCLEGRAPH_SBUF_SIZE	8192
156 #define	FULLGRAPH_SBUF_SIZE	32768
157 
158 /*
159  * These flags go in the witness relationship matrix and describe the
160  * relationship between any two struct witness objects.
161  */
162 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
163 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
164 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
165 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
166 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
167 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
168 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
169 #define	WITNESS_RELATED_MASK						\
170 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
171 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
172 					  * observed. */
173 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
174 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
175 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
176 
177 /* Descendant to ancestor flags */
178 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
179 
180 /* Ancestor to descendant flags */
181 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
182 
183 #define	WITNESS_INDEX_ASSERT(i)						\
184 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
185 
186 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
187 
188 /*
189  * Lock instances.  A lock instance is the data associated with a lock while
190  * it is held by witness.  For example, a lock instance will hold the
191  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
192  * are held in a per-cpu list while sleep locks are held in per-thread list.
193  */
194 struct lock_instance {
195 	struct lock_object	*li_lock;
196 	const char		*li_file;
197 	int			li_line;
198 	u_int			li_flags;
199 };
200 
201 /*
202  * A simple list type used to build the list of locks held by a thread
203  * or CPU.  We can't simply embed the list in struct lock_object since a
204  * lock may be held by more than one thread if it is a shared lock.  Locks
205  * are added to the head of the list, so we fill up each list entry from
206  * "the back" logically.  To ease some of the arithmetic, we actually fill
207  * in each list entry the normal way (children[0] then children[1], etc.) but
208  * when we traverse the list we read children[count-1] as the first entry
209  * down to children[0] as the final entry.
210  */
211 struct lock_list_entry {
212 	struct lock_list_entry	*ll_next;
213 	struct lock_instance	ll_children[LOCK_NCHILDREN];
214 	u_int			ll_count;
215 };
216 
217 /*
218  * The main witness structure. One of these per named lock type in the system
219  * (for example, "vnode interlock").
220  */
221 struct witness {
222 	char  			w_name[MAX_W_NAME];
223 	uint32_t 		w_index;  /* Index in the relationship matrix */
224 	struct lock_class	*w_class;
225 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
226 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
227 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
228 	const char		*w_file; /* File where last acquired */
229 	uint32_t 		w_line; /* Line where last acquired */
230 	uint32_t 		w_refcount;
231 	uint16_t 		w_num_ancestors; /* direct/indirect
232 						  * ancestor count */
233 	uint16_t 		w_num_descendants; /* direct/indirect
234 						    * descendant count */
235 	int16_t 		w_ddb_level;
236 	int 			w_displayed:1;
237 	int 			w_reversed:1;
238 };
239 
240 STAILQ_HEAD(witness_list, witness);
241 
242 /*
243  * The witness hash table. Keys are witness names (const char *), elements are
244  * witness objects (struct witness *).
245  */
246 struct witness_hash {
247 	struct witness	*wh_array[WITNESS_HASH_SIZE];
248 	uint32_t	wh_size;
249 	uint32_t	wh_count;
250 };
251 
252 /*
253  * Key type for the lock order data hash table.
254  */
255 struct witness_lock_order_key {
256 	uint16_t	from;
257 	uint16_t	to;
258 };
259 
260 struct witness_lock_order_data {
261 	struct stack			wlod_stack;
262 	struct witness_lock_order_key	wlod_key;
263 	struct witness_lock_order_data	*wlod_next;
264 };
265 
266 /*
267  * The witness lock order data hash table. Keys are witness index tuples
268  * (struct witness_lock_order_key), elements are lock order data objects
269  * (struct witness_lock_order_data).
270  */
271 struct witness_lock_order_hash {
272 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
273 	u_int	wloh_size;
274 	u_int	wloh_count;
275 };
276 
277 #ifdef BLESSING
278 struct witness_blessed {
279 	const char	*b_lock1;
280 	const char	*b_lock2;
281 };
282 #endif
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_empty(const struct witness_lock_order_key *key)
308 {
309 
310 	return (key->from == 0 && key->to == 0);
311 }
312 
313 static __inline int
314 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
315     const struct witness_lock_order_key *b)
316 {
317 
318 	return (a->from == b->from && a->to == b->to);
319 }
320 
321 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
322 		    const char *fname);
323 #ifdef KDB
324 static void	_witness_debugger(int cond, const char *msg);
325 #endif
326 static void	adopt(struct witness *parent, struct witness *child);
327 #ifdef BLESSING
328 static int	blessed(struct witness *, struct witness *);
329 #endif
330 static void	depart(struct witness *w);
331 static struct witness	*enroll(const char *description,
332 			    struct lock_class *lock_class);
333 static struct lock_instance	*find_instance(struct lock_list_entry *list,
334 				    struct lock_object *lock);
335 static int	isitmychild(struct witness *parent, struct witness *child);
336 static int	isitmydescendant(struct witness *parent, struct witness *child);
337 static void	itismychild(struct witness *parent, struct witness *child);
338 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
339 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
340 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
341 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
342 #ifdef DDB
343 static void	witness_ddb_compute_levels(void);
344 static void	witness_ddb_display(void(*)(const char *fmt, ...));
345 static void	witness_ddb_display_descendants(void(*)(const char *fmt, ...),
346 		    struct witness *, int indent);
347 static void	witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
348 		    struct witness_list *list);
349 static void	witness_ddb_level_descendants(struct witness *parent, int l);
350 static void	witness_ddb_list(struct thread *td);
351 #endif
352 static void	witness_free(struct witness *m);
353 static struct witness	*witness_get(void);
354 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
355 static struct witness	*witness_hash_get(const char *key);
356 static void	witness_hash_put(struct witness *w);
357 static void	witness_init_hash_tables(void);
358 static void	witness_increment_graph_generation(void);
359 static void	witness_lock_list_free(struct lock_list_entry *lle);
360 static struct lock_list_entry	*witness_lock_list_get(void);
361 static int	witness_lock_order_add(struct witness *parent,
362 		    struct witness *child);
363 static int	witness_lock_order_check(struct witness *parent,
364 		    struct witness *child);
365 static struct witness_lock_order_data	*witness_lock_order_get(
366 					    struct witness *parent,
367 					    struct witness *child);
368 static void	witness_list_lock(struct lock_instance *instance);
369 
370 #ifdef KDB
371 #define	witness_debugger(c)	_witness_debugger(c, __func__)
372 #else
373 #define	witness_debugger(c)
374 #endif
375 
376 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
377 
378 /*
379  * If set to 0, lock order checking is disabled.  If set to -1,
380  * witness is completely disabled.  Otherwise witness performs full
381  * lock order checking for all locks.  At runtime, lock order checking
382  * may be toggled.  However, witness cannot be reenabled once it is
383  * completely disabled.
384  */
385 static int witness_watch = 1;
386 TUNABLE_INT("debug.witness.watch", &witness_watch);
387 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
388     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
389 
390 #ifdef KDB
391 /*
392  * When KDB is enabled and witness_kdb is 1, it will cause the system
393  * to drop into kdebug() when:
394  *	- a lock hierarchy violation occurs
395  *	- locks are held when going to sleep.
396  */
397 #ifdef WITNESS_KDB
398 int	witness_kdb = 1;
399 #else
400 int	witness_kdb = 0;
401 #endif
402 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
403 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
404 
405 /*
406  * When KDB is enabled and witness_trace is 1, it will cause the system
407  * to print a stack trace:
408  *	- a lock hierarchy violation occurs
409  *	- locks are held when going to sleep.
410  */
411 int	witness_trace = 1;
412 TUNABLE_INT("debug.witness.trace", &witness_trace);
413 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
414 #endif /* KDB */
415 
416 #ifdef WITNESS_SKIPSPIN
417 int	witness_skipspin = 1;
418 #else
419 int	witness_skipspin = 0;
420 #endif
421 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
422 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
423     0, "");
424 
425 /*
426  * Call this to print out the relations between locks.
427  */
428 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
429     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
430 
431 /*
432  * Call this to print out the witness faulty stacks.
433  */
434 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
435     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
436 
437 static struct mtx w_mtx;
438 
439 /* w_list */
440 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
441 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
442 
443 /* w_typelist */
444 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
445 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
446 
447 /* lock list */
448 static struct lock_list_entry *w_lock_list_free = NULL;
449 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
450 static u_int pending_cnt;
451 
452 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
453 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
454 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
455 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
456     "");
457 
458 static struct witness *w_data;
459 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
460 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
461 static struct witness_hash w_hash;	/* The witness hash table. */
462 
463 /* The lock order data hash */
464 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
465 static struct witness_lock_order_data *w_lofree = NULL;
466 static struct witness_lock_order_hash w_lohash;
467 static int w_max_used_index = 0;
468 static unsigned int w_generation = 0;
469 static const char *w_notrunning = "Witness not running\n";
470 static const char *w_stillcold = "Witness is still cold\n";
471 
472 
473 static struct witness_order_list_entry order_lists[] = {
474 	/*
475 	 * sx locks
476 	 */
477 	{ "proctree", &lock_class_sx },
478 	{ "allproc", &lock_class_sx },
479 	{ "allprison", &lock_class_sx },
480 	{ NULL, NULL },
481 	/*
482 	 * Various mutexes
483 	 */
484 	{ "Giant", &lock_class_mtx_sleep },
485 	{ "pipe mutex", &lock_class_mtx_sleep },
486 	{ "sigio lock", &lock_class_mtx_sleep },
487 	{ "process group", &lock_class_mtx_sleep },
488 	{ "process lock", &lock_class_mtx_sleep },
489 	{ "session", &lock_class_mtx_sleep },
490 	{ "uidinfo hash", &lock_class_rw },
491 #ifdef	HWPMC_HOOKS
492 	{ "pmc-sleep", &lock_class_mtx_sleep },
493 #endif
494 	{ NULL, NULL },
495 	/*
496 	 * Sockets
497 	 */
498 	{ "accept", &lock_class_mtx_sleep },
499 	{ "so_snd", &lock_class_mtx_sleep },
500 	{ "so_rcv", &lock_class_mtx_sleep },
501 	{ "sellck", &lock_class_mtx_sleep },
502 	{ NULL, NULL },
503 	/*
504 	 * Routing
505 	 */
506 	{ "so_rcv", &lock_class_mtx_sleep },
507 	{ "radix node head", &lock_class_mtx_sleep },
508 	{ "rtentry", &lock_class_mtx_sleep },
509 	{ "ifaddr", &lock_class_mtx_sleep },
510 	{ NULL, NULL },
511 	/*
512 	 * Multicast - protocol locks before interface locks, after UDP locks.
513 	 */
514 	{ "udpinp", &lock_class_rw },
515 	{ "in_multi_mtx", &lock_class_mtx_sleep },
516 	{ "igmp_mtx", &lock_class_mtx_sleep },
517 	{ "if_addr_mtx", &lock_class_mtx_sleep },
518 	{ NULL, NULL },
519 	/*
520 	 * UNIX Domain Sockets
521 	 */
522 	{ "unp", &lock_class_mtx_sleep },
523 	{ "so_snd", &lock_class_mtx_sleep },
524 	{ NULL, NULL },
525 	/*
526 	 * UDP/IP
527 	 */
528 	{ "udp", &lock_class_rw },
529 	{ "udpinp", &lock_class_rw },
530 	{ "so_snd", &lock_class_mtx_sleep },
531 	{ NULL, NULL },
532 	/*
533 	 * TCP/IP
534 	 */
535 	{ "tcp", &lock_class_rw },
536 	{ "tcpinp", &lock_class_rw },
537 	{ "so_snd", &lock_class_mtx_sleep },
538 	{ NULL, NULL },
539 	/*
540 	 * SLIP
541 	 */
542 	{ "slip_mtx", &lock_class_mtx_sleep },
543 	{ "slip sc_mtx", &lock_class_mtx_sleep },
544 	{ NULL, NULL },
545 	/*
546 	 * netatalk
547 	 */
548 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
549 	{ "ddp_mtx", &lock_class_mtx_sleep },
550 	{ NULL, NULL },
551 	/*
552 	 * BPF
553 	 */
554 	{ "bpf global lock", &lock_class_mtx_sleep },
555 	{ "bpf interface lock", &lock_class_mtx_sleep },
556 	{ "bpf cdev lock", &lock_class_mtx_sleep },
557 	{ NULL, NULL },
558 	/*
559 	 * NFS server
560 	 */
561 	{ "nfsd_mtx", &lock_class_mtx_sleep },
562 	{ "so_snd", &lock_class_mtx_sleep },
563 	{ NULL, NULL },
564 
565 	/*
566 	 * IEEE 802.11
567 	 */
568 	{ "802.11 com lock", &lock_class_mtx_sleep},
569 	{ NULL, NULL },
570 	/*
571 	 * Network drivers
572 	 */
573 	{ "network driver", &lock_class_mtx_sleep},
574 	{ NULL, NULL },
575 
576 	/*
577 	 * Netgraph
578 	 */
579 	{ "ng_node", &lock_class_mtx_sleep },
580 	{ "ng_worklist", &lock_class_mtx_sleep },
581 	{ NULL, NULL },
582 	/*
583 	 * CDEV
584 	 */
585 	{ "system map", &lock_class_mtx_sleep },
586 	{ "vm page queue mutex", &lock_class_mtx_sleep },
587 	{ "vnode interlock", &lock_class_mtx_sleep },
588 	{ "cdev", &lock_class_mtx_sleep },
589 	{ NULL, NULL },
590 	/*
591 	 * kqueue/VFS interaction
592 	 */
593 	{ "kqueue", &lock_class_mtx_sleep },
594 	{ "struct mount mtx", &lock_class_mtx_sleep },
595 	{ "vnode interlock", &lock_class_mtx_sleep },
596 	{ NULL, NULL },
597 	/*
598 	 * spin locks
599 	 */
600 #ifdef SMP
601 	{ "ap boot", &lock_class_mtx_spin },
602 #endif
603 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
604 	{ "sio", &lock_class_mtx_spin },
605 	{ "scrlock", &lock_class_mtx_spin },
606 #ifdef __i386__
607 	{ "cy", &lock_class_mtx_spin },
608 #endif
609 #ifdef __sparc64__
610 	{ "pcib_mtx", &lock_class_mtx_spin },
611 	{ "rtc_mtx", &lock_class_mtx_spin },
612 #endif
613 	{ "scc_hwmtx", &lock_class_mtx_spin },
614 	{ "uart_hwmtx", &lock_class_mtx_spin },
615 	{ "fast_taskqueue", &lock_class_mtx_spin },
616 	{ "intr table", &lock_class_mtx_spin },
617 #ifdef	HWPMC_HOOKS
618 	{ "pmc-per-proc", &lock_class_mtx_spin },
619 #endif
620 	{ "process slock", &lock_class_mtx_spin },
621 	{ "sleepq chain", &lock_class_mtx_spin },
622 	{ "umtx lock", &lock_class_mtx_spin },
623 	{ "rm_spinlock", &lock_class_mtx_spin },
624 	{ "turnstile chain", &lock_class_mtx_spin },
625 	{ "turnstile lock", &lock_class_mtx_spin },
626 	{ "sched lock", &lock_class_mtx_spin },
627 	{ "td_contested", &lock_class_mtx_spin },
628 	{ "callout", &lock_class_mtx_spin },
629 	{ "entropy harvest mutex", &lock_class_mtx_spin },
630 	{ "syscons video lock", &lock_class_mtx_spin },
631 	{ "time lock", &lock_class_mtx_spin },
632 #ifdef SMP
633 	{ "smp rendezvous", &lock_class_mtx_spin },
634 #endif
635 #ifdef __powerpc__
636 	{ "tlb0", &lock_class_mtx_spin },
637 #endif
638 	/*
639 	 * leaf locks
640 	 */
641 	{ "intrcnt", &lock_class_mtx_spin },
642 	{ "icu", &lock_class_mtx_spin },
643 #if defined(SMP) && defined(__sparc64__)
644 	{ "ipi", &lock_class_mtx_spin },
645 #endif
646 #ifdef __i386__
647 	{ "allpmaps", &lock_class_mtx_spin },
648 	{ "descriptor tables", &lock_class_mtx_spin },
649 #endif
650 	{ "clk", &lock_class_mtx_spin },
651 	{ "cpuset", &lock_class_mtx_spin },
652 	{ "mprof lock", &lock_class_mtx_spin },
653 	{ "zombie lock", &lock_class_mtx_spin },
654 	{ "ALD Queue", &lock_class_mtx_spin },
655 #ifdef __ia64__
656 	{ "MCA spin lock", &lock_class_mtx_spin },
657 #endif
658 #if defined(__i386__) || defined(__amd64__)
659 	{ "pcicfg", &lock_class_mtx_spin },
660 	{ "NDIS thread lock", &lock_class_mtx_spin },
661 #endif
662 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
663 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
664 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
665 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
666 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
667 #ifdef	HWPMC_HOOKS
668 	{ "pmc-leaf", &lock_class_mtx_spin },
669 #endif
670 	{ "blocked lock", &lock_class_mtx_spin },
671 	{ NULL, NULL },
672 	{ NULL, NULL }
673 };
674 
675 #ifdef BLESSING
676 /*
677  * Pairs of locks which have been blessed
678  * Don't complain about order problems with blessed locks
679  */
680 static struct witness_blessed blessed_list[] = {
681 };
682 static int blessed_count =
683 	sizeof(blessed_list) / sizeof(struct witness_blessed);
684 #endif
685 
686 /*
687  * This global is set to 0 once it becomes safe to use the witness code.
688  */
689 static int witness_cold = 1;
690 
691 /*
692  * This global is set to 1 once the static lock orders have been enrolled
693  * so that a warning can be issued for any spin locks enrolled later.
694  */
695 static int witness_spin_warn = 0;
696 
697 /*
698  * The WITNESS-enabled diagnostic code.  Note that the witness code does
699  * assume that the early boot is single-threaded at least until after this
700  * routine is completed.
701  */
702 static void
703 witness_initialize(void *dummy __unused)
704 {
705 	struct lock_object *lock;
706 	struct witness_order_list_entry *order;
707 	struct witness *w, *w1;
708 	int i;
709 
710 	MALLOC(w_data, struct witness *,
711 	    sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
712 	    M_NOWAIT | M_ZERO);
713 
714 	/*
715 	 * We have to release Giant before initializing its witness
716 	 * structure so that WITNESS doesn't get confused.
717 	 */
718 	mtx_unlock(&Giant);
719 	mtx_assert(&Giant, MA_NOTOWNED);
720 
721 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
722 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
723 	    MTX_NOWITNESS | MTX_NOPROFILE);
724 	for (i = WITNESS_COUNT - 1; i >= 0; i--) {
725 		w = &w_data[i];
726 		memset(w, 0, sizeof(*w));
727 		w_data[i].w_index = i;	/* Witness index never changes. */
728 		witness_free(w);
729 	}
730 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
731 	    ("%s: Invalid list of free witness objects", __func__));
732 
733 	/* Witness with index 0 is not used to aid in debugging. */
734 	STAILQ_REMOVE_HEAD(&w_free, w_list);
735 	w_free_cnt--;
736 
737 	memset(w_rmatrix, 0,
738 	    (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
739 
740 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
741 		witness_lock_list_free(&w_locklistdata[i]);
742 	witness_init_hash_tables();
743 
744 	/* First add in all the specified order lists. */
745 	for (order = order_lists; order->w_name != NULL; order++) {
746 		w = enroll(order->w_name, order->w_class);
747 		if (w == NULL)
748 			continue;
749 		w->w_file = "order list";
750 		for (order++; order->w_name != NULL; order++) {
751 			w1 = enroll(order->w_name, order->w_class);
752 			if (w1 == NULL)
753 				continue;
754 			w1->w_file = "order list";
755 			itismychild(w, w1);
756 			w = w1;
757 		}
758 	}
759 	witness_spin_warn = 1;
760 
761 	/* Iterate through all locks and add them to witness. */
762 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
763 		lock = pending_locks[i].wh_lock;
764 		KASSERT(lock->lo_flags & LO_WITNESS,
765 		    ("%s: lock %s is on pending list but not LO_WITNESS",
766 		    __func__, lock->lo_name));
767 		lock->lo_witness = enroll(pending_locks[i].wh_type,
768 		    LOCK_CLASS(lock));
769 	}
770 
771 	/* Mark the witness code as being ready for use. */
772 	witness_cold = 0;
773 
774 	mtx_lock(&Giant);
775 }
776 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
777     NULL);
778 
779 void
780 witness_init(struct lock_object *lock, const char *type)
781 {
782 	struct lock_class *class;
783 
784 	/* Various sanity checks. */
785 	class = LOCK_CLASS(lock);
786 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
787 	    (class->lc_flags & LC_RECURSABLE) == 0)
788 		panic("%s: lock (%s) %s can not be recursable", __func__,
789 		    class->lc_name, lock->lo_name);
790 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
791 	    (class->lc_flags & LC_SLEEPABLE) == 0)
792 		panic("%s: lock (%s) %s can not be sleepable", __func__,
793 		    class->lc_name, lock->lo_name);
794 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
795 	    (class->lc_flags & LC_UPGRADABLE) == 0)
796 		panic("%s: lock (%s) %s can not be upgradable", __func__,
797 		    class->lc_name, lock->lo_name);
798 
799 	/*
800 	 * If we shouldn't watch this lock, then just clear lo_witness.
801 	 * Otherwise, if witness_cold is set, then it is too early to
802 	 * enroll this lock, so defer it to witness_initialize() by adding
803 	 * it to the pending_locks list.  If it is not too early, then enroll
804 	 * the lock now.
805 	 */
806 	if (witness_watch < 1 || panicstr != NULL ||
807 	    (lock->lo_flags & LO_WITNESS) == 0)
808 		lock->lo_witness = NULL;
809 	else if (witness_cold) {
810 		pending_locks[pending_cnt].wh_lock = lock;
811 		pending_locks[pending_cnt++].wh_type = type;
812 		if (pending_cnt > WITNESS_PENDLIST)
813 			panic("%s: pending locks list is too small, bump it\n",
814 			    __func__);
815 	} else
816 		lock->lo_witness = enroll(type, class);
817 }
818 
819 void
820 witness_destroy(struct lock_object *lock)
821 {
822 	struct lock_class *class;
823 	struct witness *w;
824 
825 	class = LOCK_CLASS(lock);
826 
827 	if (witness_cold)
828 		panic("lock (%s) %s destroyed while witness_cold",
829 		    class->lc_name, lock->lo_name);
830 
831 	/* XXX: need to verify that no one holds the lock */
832 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
833 		return;
834 	w = lock->lo_witness;
835 
836 	mtx_lock_spin(&w_mtx);
837 	MPASS(w->w_refcount > 0);
838 	w->w_refcount--;
839 
840 	if (w->w_refcount == 0)
841 		depart(w);
842 	mtx_unlock_spin(&w_mtx);
843 }
844 
845 #ifdef DDB
846 static void
847 witness_ddb_compute_levels(void)
848 {
849 	struct witness *w;
850 
851 	/*
852 	 * First clear all levels.
853 	 */
854 	STAILQ_FOREACH(w, &w_all, w_list)
855 		w->w_ddb_level = -1;
856 
857 	/*
858 	 * Look for locks with no parents and level all their descendants.
859 	 */
860 	STAILQ_FOREACH(w, &w_all, w_list) {
861 
862 		/* If the witness has ancestors (is not a root), skip it. */
863 		if (w->w_num_ancestors > 0)
864 			continue;
865 		witness_ddb_level_descendants(w, 0);
866 	}
867 }
868 
869 static void
870 witness_ddb_level_descendants(struct witness *w, int l)
871 {
872 	int i;
873 
874 	if (w->w_ddb_level >= l)
875 		return;
876 
877 	w->w_ddb_level = l;
878 	l++;
879 
880 	for (i = 1; i <= w_max_used_index; i++) {
881 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
882 			witness_ddb_level_descendants(&w_data[i], l);
883 	}
884 }
885 
886 static void
887 witness_ddb_display_descendants(void(*prnt)(const char *fmt, ...),
888     struct witness *w, int indent)
889 {
890 	int i;
891 
892  	for (i = 0; i < indent; i++)
893  		prnt(" ");
894 	prnt("%s (type: %s, depth: %d, active refs: %d)",
895 	     w->w_name, w->w_class->lc_name,
896 	     w->w_ddb_level, w->w_refcount);
897  	if (w->w_displayed) {
898  		prnt(" -- (already displayed)\n");
899  		return;
900  	}
901  	w->w_displayed = 1;
902 	if (w->w_file != NULL && w->w_line != 0)
903 		prnt(" -- last acquired @ %s:%d\n", w->w_file,
904 		    w->w_line);
905 	else
906 		prnt(" -- never acquired\n");
907 	indent++;
908 	WITNESS_INDEX_ASSERT(w->w_index);
909 	for (i = 1; i <= w_max_used_index; i++) {
910 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
911 			witness_ddb_display_descendants(prnt, &w_data[i],
912 			    indent);
913 	}
914 }
915 
916 static void
917 witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
918     struct witness_list *list)
919 {
920 	struct witness *w;
921 
922 	STAILQ_FOREACH(w, list, w_typelist) {
923 		if (w->w_file == NULL || w->w_ddb_level > 0)
924 			continue;
925 
926 		/* This lock has no anscestors - display its descendants. */
927 		witness_ddb_display_descendants(prnt, w, 0);
928 	}
929 }
930 
931 static void
932 witness_ddb_display(void(*prnt)(const char *fmt, ...))
933 {
934 	struct witness *w;
935 
936 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
937 	witness_ddb_compute_levels();
938 
939 	/* Clear all the displayed flags. */
940 	STAILQ_FOREACH(w, &w_all, w_list)
941 		w->w_displayed = 0;
942 
943 	/*
944 	 * First, handle sleep locks which have been acquired at least
945 	 * once.
946 	 */
947 	prnt("Sleep locks:\n");
948 	witness_ddb_display_list(prnt, &w_sleep);
949 
950 	/*
951 	 * Now do spin locks which have been acquired at least once.
952 	 */
953 	prnt("\nSpin locks:\n");
954 	witness_ddb_display_list(prnt, &w_spin);
955 
956 	/*
957 	 * Finally, any locks which have not been acquired yet.
958 	 */
959 	prnt("\nLocks which were never acquired:\n");
960 	STAILQ_FOREACH(w, &w_all, w_list) {
961 		if (w->w_file != NULL || w->w_refcount == 0)
962 			continue;
963 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
964 		    w->w_class->lc_name, w->w_ddb_level);
965 	}
966 }
967 #endif /* DDB */
968 
969 /* Trim useless garbage from filenames. */
970 static const char *
971 fixup_filename(const char *file)
972 {
973 
974 	if (file == NULL)
975 		return (NULL);
976 	while (strncmp(file, "../", 3) == 0)
977 		file += 3;
978 	return (file);
979 }
980 
981 int
982 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
983 {
984 
985 	if (witness_watch == -1 || panicstr != NULL)
986 		return (0);
987 
988 	/* Require locks that witness knows about. */
989 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
990 	    lock2->lo_witness == NULL)
991 		return (EINVAL);
992 
993 	mtx_assert(&w_mtx, MA_NOTOWNED);
994 	mtx_lock_spin(&w_mtx);
995 
996 	/*
997 	 * If we already have either an explicit or implied lock order that
998 	 * is the other way around, then return an error.
999 	 */
1000 	if (witness_watch &&
1001 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1002 		mtx_unlock_spin(&w_mtx);
1003 		return (EDOOFUS);
1004 	}
1005 
1006 	/* Try to add the new order. */
1007 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1008 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1009 	itismychild(lock1->lo_witness, lock2->lo_witness);
1010 	mtx_unlock_spin(&w_mtx);
1011 	return (0);
1012 }
1013 
1014 void
1015 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1016     int line, struct lock_object *interlock)
1017 {
1018 	struct lock_list_entry **lock_list, *lle;
1019 	struct lock_instance *lock1, *lock2, *plock;
1020 	struct lock_class *class;
1021 	struct witness *w, *w1;
1022 	struct thread *td;
1023 	int i, j;
1024 
1025 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1026 	    panicstr != NULL)
1027 		return;
1028 
1029 	w = lock->lo_witness;
1030 	class = LOCK_CLASS(lock);
1031 	td = curthread;
1032 	file = fixup_filename(file);
1033 
1034 	if (class->lc_flags & LC_SLEEPLOCK) {
1035 
1036 		/*
1037 		 * Since spin locks include a critical section, this check
1038 		 * implicitly enforces a lock order of all sleep locks before
1039 		 * all spin locks.
1040 		 */
1041 		if (td->td_critnest != 0 && !kdb_active)
1042 			panic("blockable sleep lock (%s) %s @ %s:%d",
1043 			    class->lc_name, lock->lo_name, file, line);
1044 
1045 		/*
1046 		 * If this is the first lock acquired then just return as
1047 		 * no order checking is needed.
1048 		 */
1049 		if (td->td_sleeplocks == NULL)
1050 			return;
1051 		lock_list = &td->td_sleeplocks;
1052 	} else {
1053 
1054 		/*
1055 		 * If this is the first lock, just return as no order
1056 		 * checking is needed.  We check this in both if clauses
1057 		 * here as unifying the check would require us to use a
1058 		 * critical section to ensure we don't migrate while doing
1059 		 * the check.  Note that if this is not the first lock, we
1060 		 * are already in a critical section and are safe for the
1061 		 * rest of the check.
1062 		 */
1063 		if (PCPU_GET(spinlocks) == NULL)
1064 			return;
1065 		lock_list = PCPU_PTR(spinlocks);
1066 	}
1067 
1068 	/* Empty list? */
1069 	if ((*lock_list)->ll_count == 0)
1070 		return;
1071 
1072 	/*
1073 	 * Check to see if we are recursing on a lock we already own.  If
1074 	 * so, make sure that we don't mismatch exclusive and shared lock
1075 	 * acquires.
1076 	 */
1077 	lock1 = find_instance(*lock_list, lock);
1078 	if (lock1 != NULL) {
1079 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1080 		    (flags & LOP_EXCLUSIVE) == 0) {
1081 			printf("shared lock of (%s) %s @ %s:%d\n",
1082 			    class->lc_name, lock->lo_name, file, line);
1083 			printf("while exclusively locked from %s:%d\n",
1084 			    lock1->li_file, lock1->li_line);
1085 			panic("share->excl");
1086 		}
1087 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1088 		    (flags & LOP_EXCLUSIVE) != 0) {
1089 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1090 			    class->lc_name, lock->lo_name, file, line);
1091 			printf("while share locked from %s:%d\n",
1092 			    lock1->li_file, lock1->li_line);
1093 			panic("excl->share");
1094 		}
1095 		return;
1096 	}
1097 
1098 	/*
1099 	 * Find the previously acquired lock, but ignore interlocks.
1100 	 */
1101 	plock = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1102 	if (interlock != NULL && plock->li_lock == interlock) {
1103 		if ((*lock_list)->ll_count == 1) {
1104 
1105 			/*
1106 			 * The interlock is the only lock we hold, so
1107 			 * nothing to do.
1108 			 */
1109 			return;
1110 		}
1111 		plock = &(*lock_list)->ll_children[(*lock_list)->ll_count - 2];
1112 	}
1113 
1114 	/*
1115 	 * Try to perform most checks without a lock.  If this succeeds we
1116 	 * can skip acquiring the lock and return success.
1117 	 */
1118 	w1 = plock->li_lock->lo_witness;
1119 	if (witness_lock_order_check(w1, w))
1120 		return;
1121 
1122 	/*
1123 	 * Check for duplicate locks of the same type.  Note that we only
1124 	 * have to check for this on the last lock we just acquired.  Any
1125 	 * other cases will be caught as lock order violations.
1126 	 */
1127 	mtx_lock_spin(&w_mtx);
1128 	witness_lock_order_add(w1, w);
1129 	if (w1 == w) {
1130 		i = w->w_index;
1131 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1132 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1133 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1134 			w->w_reversed = 1;
1135 			mtx_unlock_spin(&w_mtx);
1136 			printf(
1137 			    "acquiring duplicate lock of same type: \"%s\"\n",
1138 			    w->w_name);
1139 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1140 			       plock->li_file, plock->li_line);
1141 			printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1142 			witness_debugger(1);
1143 		    } else
1144 			    mtx_unlock_spin(&w_mtx);
1145 		return;
1146 	}
1147 	mtx_assert(&w_mtx, MA_OWNED);
1148 
1149 	/*
1150 	 * If we know that the the lock we are acquiring comes after
1151 	 * the lock we most recently acquired in the lock order tree,
1152 	 * then there is no need for any further checks.
1153 	 */
1154 	if (isitmychild(w1, w))
1155 		goto out;
1156 
1157 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
1158 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1159 
1160 			MPASS(j < WITNESS_COUNT);
1161 			lock1 = &lle->ll_children[i];
1162 
1163 			/*
1164 			 * Ignore the interlock the first time we see it.
1165 			 */
1166 			if (interlock != NULL && interlock == lock1->li_lock) {
1167 				interlock = NULL;
1168 				continue;
1169 			}
1170 
1171 			/*
1172 			 * If this lock doesn't undergo witness checking,
1173 			 * then skip it.
1174 			 */
1175 			w1 = lock1->li_lock->lo_witness;
1176 			if (w1 == NULL) {
1177 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1178 				    ("lock missing witness structure"));
1179 				continue;
1180 			}
1181 
1182 			/*
1183 			 * If we are locking Giant and this is a sleepable
1184 			 * lock, then skip it.
1185 			 */
1186 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1187 			    lock == &Giant.lock_object)
1188 				continue;
1189 
1190 			/*
1191 			 * If we are locking a sleepable lock and this lock
1192 			 * is Giant, then skip it.
1193 			 */
1194 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1195 			    lock1->li_lock == &Giant.lock_object)
1196 				continue;
1197 
1198 			/*
1199 			 * If we are locking a sleepable lock and this lock
1200 			 * isn't sleepable, we want to treat it as a lock
1201 			 * order violation to enfore a general lock order of
1202 			 * sleepable locks before non-sleepable locks.
1203 			 */
1204 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1205 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1206 				goto reversal;
1207 
1208 			/*
1209 			 * If we are locking Giant and this is a non-sleepable
1210 			 * lock, then treat it as a reversal.
1211 			 */
1212 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1213 			    lock == &Giant.lock_object)
1214 				goto reversal;
1215 
1216 			/*
1217 			 * Check the lock order hierarchy for a reveresal.
1218 			 */
1219 			if (!isitmydescendant(w, w1))
1220 				continue;
1221 		reversal:
1222 
1223 			/*
1224 			 * We have a lock order violation, check to see if it
1225 			 * is allowed or has already been yelled about.
1226 			 */
1227 #ifdef BLESSING
1228 
1229 			/*
1230 			 * If the lock order is blessed, just bail.  We don't
1231 			 * look for other lock order violations though, which
1232 			 * may be a bug.
1233 			 */
1234 			if (blessed(w, w1))
1235 				goto out;
1236 #endif
1237 
1238 			/* Bail if this violation is known */
1239 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1240 				goto out;
1241 
1242 			/* Record this as a violation */
1243 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1244 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1245 			w->w_reversed = w1->w_reversed = 1;
1246 			witness_increment_graph_generation();
1247 			mtx_unlock_spin(&w_mtx);
1248 
1249 			/*
1250 			 * Ok, yell about it.
1251 			 */
1252 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1253 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1254 				printf(
1255 		"lock order reversal: (sleepable after non-sleepable)\n");
1256 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1257 			    && lock == &Giant.lock_object)
1258 				printf(
1259 		"lock order reversal: (Giant after non-sleepable)\n");
1260 			else
1261 				printf("lock order reversal:\n");
1262 
1263 			/*
1264 			 * Try to locate an earlier lock with
1265 			 * witness w in our list.
1266 			 */
1267 			do {
1268 				lock2 = &lle->ll_children[i];
1269 				MPASS(lock2->li_lock != NULL);
1270 				if (lock2->li_lock->lo_witness == w)
1271 					break;
1272 				if (i == 0 && lle->ll_next != NULL) {
1273 					lle = lle->ll_next;
1274 					i = lle->ll_count - 1;
1275 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1276 				} else
1277 					i--;
1278 			} while (i >= 0);
1279 			if (i < 0) {
1280 				printf(" 1st %p %s (%s) @ %s:%d\n",
1281 				    lock1->li_lock, lock1->li_lock->lo_name,
1282 				    w1->w_name, lock1->li_file, lock1->li_line);
1283 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1284 				    lock->lo_name, w->w_name, file, line);
1285 			} else {
1286 				printf(" 1st %p %s (%s) @ %s:%d\n",
1287 				    lock2->li_lock, lock2->li_lock->lo_name,
1288 				    lock2->li_lock->lo_witness->w_name,
1289 				    lock2->li_file, lock2->li_line);
1290 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1291 				    lock1->li_lock, lock1->li_lock->lo_name,
1292 				    w1->w_name, lock1->li_file, lock1->li_line);
1293 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1294 				    lock->lo_name, w->w_name, file, line);
1295 			}
1296 			witness_debugger(1);
1297 			return;
1298 		}
1299 	}
1300 
1301 	/*
1302 	 * If requested, build a new lock order.  However, don't build a new
1303 	 * relationship between a sleepable lock and Giant if it is in the
1304 	 * wrong direction.  The correct lock order is that sleepable locks
1305 	 * always come before Giant.
1306 	 */
1307 	if (flags & LOP_NEWORDER &&
1308 	    !(plock->li_lock == &Giant.lock_object &&
1309 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1310 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1311 		    w->w_name, plock->li_lock->lo_witness->w_name);
1312 		itismychild(plock->li_lock->lo_witness, w);
1313 	}
1314 out:
1315 	mtx_unlock_spin(&w_mtx);
1316 }
1317 
1318 void
1319 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1320 {
1321 	struct lock_list_entry **lock_list, *lle;
1322 	struct lock_instance *instance;
1323 	struct witness *w;
1324 	struct thread *td;
1325 
1326 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1327 	    panicstr != NULL)
1328 		return;
1329 	w = lock->lo_witness;
1330 	td = curthread;
1331 	file = fixup_filename(file);
1332 
1333 	/* Determine lock list for this lock. */
1334 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1335 		lock_list = &td->td_sleeplocks;
1336 	else
1337 		lock_list = PCPU_PTR(spinlocks);
1338 
1339 	/* Check to see if we are recursing on a lock we already own. */
1340 	instance = find_instance(*lock_list, lock);
1341 	if (instance != NULL) {
1342 		instance->li_flags++;
1343 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1344 		    td->td_proc->p_pid, lock->lo_name,
1345 		    instance->li_flags & LI_RECURSEMASK);
1346 		instance->li_file = file;
1347 		instance->li_line = line;
1348 		return;
1349 	}
1350 
1351 	/* Update per-witness last file and line acquire. */
1352 	w->w_file = file;
1353 	w->w_line = line;
1354 
1355 	/* Find the next open lock instance in the list and fill it. */
1356 	lle = *lock_list;
1357 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1358 		lle = witness_lock_list_get();
1359 		if (lle == NULL)
1360 			return;
1361 		lle->ll_next = *lock_list;
1362 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1363 		    td->td_proc->p_pid, lle);
1364 		*lock_list = lle;
1365 	}
1366 	instance = &lle->ll_children[lle->ll_count++];
1367 	instance->li_lock = lock;
1368 	instance->li_line = line;
1369 	instance->li_file = file;
1370 	if ((flags & LOP_EXCLUSIVE) != 0)
1371 		instance->li_flags = LI_EXCLUSIVE;
1372 	else
1373 		instance->li_flags = 0;
1374 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1375 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1376 }
1377 
1378 void
1379 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1380 {
1381 	struct lock_instance *instance;
1382 	struct lock_class *class;
1383 
1384 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1385 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1386 		return;
1387 	class = LOCK_CLASS(lock);
1388 	file = fixup_filename(file);
1389 	if (witness_watch) {
1390 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1391 			panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1392 			    class->lc_name, lock->lo_name, file, line);
1393 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1394 			panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1395 			    class->lc_name, lock->lo_name, file, line);
1396 	}
1397 	instance = find_instance(curthread->td_sleeplocks, lock);
1398 	if (instance == NULL)
1399 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1400 		    class->lc_name, lock->lo_name, file, line);
1401 	if (witness_watch) {
1402 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1403 			panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1404 			    class->lc_name, lock->lo_name, file, line);
1405 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1406 			panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1407 			    class->lc_name, lock->lo_name,
1408 			    instance->li_flags & LI_RECURSEMASK, file, line);
1409 	}
1410 	instance->li_flags |= LI_EXCLUSIVE;
1411 }
1412 
1413 void
1414 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1415     int line)
1416 {
1417 	struct lock_instance *instance;
1418 	struct lock_class *class;
1419 
1420 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1421 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1422 		return;
1423 	class = LOCK_CLASS(lock);
1424 	file = fixup_filename(file);
1425 	if (witness_watch) {
1426 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1427 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1428 			    class->lc_name, lock->lo_name, file, line);
1429 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1430 			panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1431 			    class->lc_name, lock->lo_name, file, line);
1432 	}
1433 	instance = find_instance(curthread->td_sleeplocks, lock);
1434 	if (instance == NULL)
1435 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1436 		    class->lc_name, lock->lo_name, file, line);
1437 	if (witness_watch) {
1438 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1439 			panic("downgrade of shared lock (%s) %s @ %s:%d",
1440 			    class->lc_name, lock->lo_name, file, line);
1441 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1442 			panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1443 			    class->lc_name, lock->lo_name,
1444 			    instance->li_flags & LI_RECURSEMASK, file, line);
1445 	}
1446 	instance->li_flags &= ~LI_EXCLUSIVE;
1447 }
1448 
1449 void
1450 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1451 {
1452 	struct lock_list_entry **lock_list, *lle;
1453 	struct lock_instance *instance;
1454 	struct lock_class *class;
1455 	struct thread *td;
1456 	register_t s;
1457 	int i, j;
1458 
1459 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1460 		return;
1461 	td = curthread;
1462 	class = LOCK_CLASS(lock);
1463 	file = fixup_filename(file);
1464 
1465 	/* Find lock instance associated with this lock. */
1466 	if (class->lc_flags & LC_SLEEPLOCK)
1467 		lock_list = &td->td_sleeplocks;
1468 	else
1469 		lock_list = PCPU_PTR(spinlocks);
1470 	lle = *lock_list;
1471 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1472 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1473 			instance = &(*lock_list)->ll_children[i];
1474 			if (instance->li_lock == lock)
1475 				goto found;
1476 		}
1477 
1478 	/*
1479 	 * When disabling WITNESS through witness_watch we could end up in
1480 	 * having registered locks in the td_sleeplocks queue.
1481 	 * We have to make sure we flush these queues, so just search for
1482 	 * eventual register locks and remove them.
1483 	 */
1484 	if (witness_watch > 0)
1485 		panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1486 		    lock->lo_name, file, line);
1487 	else
1488 		return;
1489 found:
1490 
1491 	/* First, check for shared/exclusive mismatches. */
1492 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1493 	    (flags & LOP_EXCLUSIVE) == 0) {
1494 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1495 		    lock->lo_name, file, line);
1496 		printf("while exclusively locked from %s:%d\n",
1497 		    instance->li_file, instance->li_line);
1498 		panic("excl->ushare");
1499 	}
1500 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1501 	    (flags & LOP_EXCLUSIVE) != 0) {
1502 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1503 		    lock->lo_name, file, line);
1504 		printf("while share locked from %s:%d\n", instance->li_file,
1505 		    instance->li_line);
1506 		panic("share->uexcl");
1507 	}
1508 
1509 	/* If we are recursed, unrecurse. */
1510 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1511 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1512 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1513 		    instance->li_flags);
1514 		instance->li_flags--;
1515 		return;
1516 	}
1517 
1518 	/* Otherwise, remove this item from the list. */
1519 	s = intr_disable();
1520 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1521 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1522 	    (*lock_list)->ll_count - 1);
1523 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1524 		(*lock_list)->ll_children[j] =
1525 		    (*lock_list)->ll_children[j + 1];
1526 	(*lock_list)->ll_count--;
1527 	intr_restore(s);
1528 
1529 	/*
1530 	 * In order to reduce contention on w_mtx, we want to keep always an
1531 	 * head object into lists so that frequent allocation from the
1532 	 * free witness pool (and subsequent locking) is avoided.
1533 	 * In order to maintain the current code simple, when the head
1534 	 * object is totally unloaded it means also that we do not have
1535 	 * further objects in the list, so the list ownership needs to be
1536 	 * hand over to another object if the current head needs to be freed.
1537 	 */
1538 	if ((*lock_list)->ll_count == 0) {
1539 		if (*lock_list == lle) {
1540 			if (lle->ll_next == NULL)
1541 				return;
1542 		} else
1543 			lle = *lock_list;
1544 		*lock_list = lle->ll_next;
1545 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1546 		    td->td_proc->p_pid, lle);
1547 		witness_lock_list_free(lle);
1548 	}
1549 }
1550 
1551 void
1552 witness_thread_exit(struct thread *td)
1553 {
1554 	struct lock_list_entry *lle;
1555 	int i, n;
1556 
1557 	lle = td->td_sleeplocks;
1558 	if (lle == NULL || panicstr != NULL)
1559 		return;
1560 	if (lle->ll_count != 0) {
1561 		for (n = 0; lle != NULL; lle = lle->ll_next)
1562 			for (i = lle->ll_count - 1; i >= 0; i--) {
1563 				if (n == 0)
1564 		printf("Thread %p exiting with the following locks held:\n",
1565 					    td);
1566 				n++;
1567 				witness_list_lock(&lle->ll_children[i]);
1568 
1569 			}
1570 		panic("Thread %p cannot exit while holding sleeplocks\n", td);
1571 	}
1572 	witness_lock_list_free(lle);
1573 }
1574 
1575 /*
1576  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1577  * exempt Giant and sleepable locks from the checks as well.  If any
1578  * non-exempt locks are held, then a supplied message is printed to the
1579  * console along with a list of the offending locks.  If indicated in the
1580  * flags then a failure results in a panic as well.
1581  */
1582 int
1583 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1584 {
1585 	struct lock_list_entry **lock_list, *lle;
1586 	struct lock_instance *lock1;
1587 	struct thread *td;
1588 	va_list ap;
1589 	int i, n;
1590 
1591 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1592 		return (0);
1593 	n = 0;
1594 	td = curthread;
1595 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1596 		for (i = lle->ll_count - 1; i >= 0; i--) {
1597 			lock1 = &lle->ll_children[i];
1598 			if (lock1->li_lock == lock)
1599 				continue;
1600 			if (flags & WARN_GIANTOK &&
1601 			    lock1->li_lock == &Giant.lock_object)
1602 				continue;
1603 			if (flags & WARN_SLEEPOK &&
1604 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1605 				continue;
1606 			if (n == 0) {
1607 				va_start(ap, fmt);
1608 				vprintf(fmt, ap);
1609 				va_end(ap);
1610 				printf(" with the following");
1611 				if (flags & WARN_SLEEPOK)
1612 					printf(" non-sleepable");
1613 				printf(" locks held:\n");
1614 			}
1615 			n++;
1616 			witness_list_lock(lock1);
1617 		}
1618 	if (PCPU_GET(spinlocks) != NULL) {
1619 		lock_list = PCPU_PTR(spinlocks);
1620 
1621 		/* Empty list? */
1622 		if ((*lock_list)->ll_count == 0)
1623 			return (n);
1624 
1625 		/*
1626 		 * Since we already hold a spinlock preemption is
1627 		 * already blocked.
1628 		 */
1629 		if (n == 0) {
1630 			va_start(ap, fmt);
1631 			vprintf(fmt, ap);
1632 			va_end(ap);
1633 			printf(" with the following");
1634 			if (flags & WARN_SLEEPOK)
1635 				printf(" non-sleepable");
1636 			printf(" locks held:\n");
1637 		}
1638 		n += witness_list_locks(PCPU_PTR(spinlocks));
1639 	}
1640 	if (flags & WARN_PANIC && n)
1641 		panic("%s", __func__);
1642 	else
1643 		witness_debugger(n);
1644 	return (n);
1645 }
1646 
1647 const char *
1648 witness_file(struct lock_object *lock)
1649 {
1650 	struct witness *w;
1651 
1652 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1653 		return ("?");
1654 	w = lock->lo_witness;
1655 	return (w->w_file);
1656 }
1657 
1658 int
1659 witness_line(struct lock_object *lock)
1660 {
1661 	struct witness *w;
1662 
1663 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1664 		return (0);
1665 	w = lock->lo_witness;
1666 	return (w->w_line);
1667 }
1668 
1669 static struct witness *
1670 enroll(const char *description, struct lock_class *lock_class)
1671 {
1672 	struct witness *w;
1673 	struct witness_list *typelist;
1674 
1675 	MPASS(description != NULL);
1676 
1677 	if (witness_watch == -1 || panicstr != NULL)
1678 		return (NULL);
1679 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1680 		if (witness_skipspin)
1681 			return (NULL);
1682 		else
1683 			typelist = &w_spin;
1684 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1685 		typelist = &w_sleep;
1686 	else
1687 		panic("lock class %s is not sleep or spin",
1688 		    lock_class->lc_name);
1689 
1690 	mtx_lock_spin(&w_mtx);
1691 	w = witness_hash_get(description);
1692 	if (w)
1693 		goto found;
1694 	if ((w = witness_get()) == NULL)
1695 		return (NULL);
1696 	MPASS(strlen(description) < MAX_W_NAME);
1697 	strcpy(w->w_name, description);
1698 	w->w_class = lock_class;
1699 	w->w_refcount = 1;
1700 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1701 	if (lock_class->lc_flags & LC_SPINLOCK) {
1702 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1703 		w_spin_cnt++;
1704 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1705 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1706 		w_sleep_cnt++;
1707 	}
1708 
1709 	/* Insert new witness into the hash */
1710 	witness_hash_put(w);
1711 	witness_increment_graph_generation();
1712 	mtx_unlock_spin(&w_mtx);
1713 	return (w);
1714 found:
1715 	w->w_refcount++;
1716 	mtx_unlock_spin(&w_mtx);
1717 	if (lock_class != w->w_class)
1718 		panic(
1719 			"lock (%s) %s does not match earlier (%s) lock",
1720 			description, lock_class->lc_name,
1721 			w->w_class->lc_name);
1722 	return (w);
1723 }
1724 
1725 static void
1726 depart(struct witness *w)
1727 {
1728 	struct witness_list *list;
1729 
1730 	MPASS(w->w_refcount == 0);
1731 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1732 		list = &w_sleep;
1733 		w_sleep_cnt--;
1734 	} else {
1735 		list = &w_spin;
1736 		w_spin_cnt--;
1737 	}
1738 	/*
1739 	 * Set file to NULL as it may point into a loadable module.
1740 	 */
1741 	w->w_file = NULL;
1742 	w->w_line = 0;
1743 	witness_increment_graph_generation();
1744 }
1745 
1746 
1747 static void
1748 adopt(struct witness *parent, struct witness *child)
1749 {
1750 	int pi, ci, i, j;
1751 
1752 	if (witness_cold == 0)
1753 		mtx_assert(&w_mtx, MA_OWNED);
1754 
1755 	/* If the relationship is already known, there's no work to be done. */
1756 	if (isitmychild(parent, child))
1757 		return;
1758 
1759 	/* When the structure of the graph changes, bump up the generation. */
1760 	witness_increment_graph_generation();
1761 
1762 	/*
1763 	 * The hard part ... create the direct relationship, then propagate all
1764 	 * indirect relationships.
1765 	 */
1766 	pi = parent->w_index;
1767 	ci = child->w_index;
1768 	WITNESS_INDEX_ASSERT(pi);
1769 	WITNESS_INDEX_ASSERT(ci);
1770 	MPASS(pi != ci);
1771 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1772 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1773 
1774 	/*
1775 	 * If parent was not already an ancestor of child,
1776 	 * then we increment the descendant and ancestor counters.
1777 	 */
1778 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1779 		parent->w_num_descendants++;
1780 		child->w_num_ancestors++;
1781 	}
1782 
1783 	/*
1784 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1785 	 * an ancestor of 'pi' during this loop.
1786 	 */
1787 	for (i = 1; i <= w_max_used_index; i++) {
1788 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1789 		    (i != pi))
1790 			continue;
1791 
1792 		/* Find each descendant of 'i' and mark it as a descendant. */
1793 		for (j = 1; j <= w_max_used_index; j++) {
1794 
1795 			/*
1796 			 * Skip children that are already marked as
1797 			 * descendants of 'i'.
1798 			 */
1799 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1800 				continue;
1801 
1802 			/*
1803 			 * We are only interested in descendants of 'ci'. Note
1804 			 * that 'ci' itself is counted as a descendant of 'ci'.
1805 			 */
1806 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1807 			    (j != ci))
1808 				continue;
1809 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1810 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1811 			w_data[i].w_num_descendants++;
1812 			w_data[j].w_num_ancestors++;
1813 
1814 			/*
1815 			 * Make sure we aren't marking a node as both an
1816 			 * ancestor and descendant. We should have caught
1817 			 * this as a lock order reversal earlier.
1818 			 */
1819 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1820 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1821 				printf("witness rmatrix paradox! [%d][%d]=%d "
1822 				    "both ancestor and descendant\n",
1823 				    i, j, w_rmatrix[i][j]);
1824 				kdb_backtrace();
1825 				printf("Witness disabled.\n");
1826 				witness_watch = -1;
1827 			}
1828 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1829 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1830 				printf("witness rmatrix paradox! [%d][%d]=%d "
1831 				    "both ancestor and descendant\n",
1832 				    j, i, w_rmatrix[j][i]);
1833 				kdb_backtrace();
1834 				printf("Witness disabled.\n");
1835 				witness_watch = -1;
1836 			}
1837 		}
1838 	}
1839 }
1840 
1841 static void
1842 itismychild(struct witness *parent, struct witness *child)
1843 {
1844 
1845 	MPASS(child != NULL && parent != NULL);
1846 	if (witness_cold == 0)
1847 		mtx_assert(&w_mtx, MA_OWNED);
1848 
1849 	if (!witness_lock_type_equal(parent, child)) {
1850 		if (witness_cold == 0)
1851 			mtx_unlock_spin(&w_mtx);
1852 		panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1853 		    "the same lock type", __func__, parent->w_name,
1854 		    parent->w_class->lc_name, child->w_name,
1855 		    child->w_class->lc_name);
1856 	}
1857 	adopt(parent, child);
1858 }
1859 
1860 /*
1861  * Generic code for the isitmy*() functions. The rmask parameter is the
1862  * expected relationship of w1 to w2.
1863  */
1864 static int
1865 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1866 {
1867 	unsigned char r1, r2;
1868 	int i1, i2;
1869 
1870 	i1 = w1->w_index;
1871 	i2 = w2->w_index;
1872 	WITNESS_INDEX_ASSERT(i1);
1873 	WITNESS_INDEX_ASSERT(i2);
1874 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1875 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1876 
1877 	/* The flags on one better be the inverse of the flags on the other */
1878 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1879 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1880 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
1881 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
1882 		    "w_rmatrix[%d][%d] == %hhx\n",
1883 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1884 		    i2, i1, r2);
1885 		kdb_backtrace();
1886 		printf("Witness disabled.\n");
1887 		witness_watch = -1;
1888 	}
1889 	return (r1 & rmask);
1890 }
1891 
1892 /*
1893  * Checks if @child is a direct child of @parent.
1894  */
1895 static int
1896 isitmychild(struct witness *parent, struct witness *child)
1897 {
1898 
1899 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1900 }
1901 
1902 /*
1903  * Checks if @descendant is a direct or inderect descendant of @ancestor.
1904  */
1905 static int
1906 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1907 {
1908 
1909 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1910 	    __func__));
1911 }
1912 
1913 #ifdef BLESSING
1914 static int
1915 blessed(struct witness *w1, struct witness *w2)
1916 {
1917 	int i;
1918 	struct witness_blessed *b;
1919 
1920 	for (i = 0; i < blessed_count; i++) {
1921 		b = &blessed_list[i];
1922 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1923 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1924 				return (1);
1925 			continue;
1926 		}
1927 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1928 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1929 				return (1);
1930 	}
1931 	return (0);
1932 }
1933 #endif
1934 
1935 static struct witness *
1936 witness_get(void)
1937 {
1938 	struct witness *w;
1939 	int index;
1940 
1941 	if (witness_cold == 0)
1942 		mtx_assert(&w_mtx, MA_OWNED);
1943 
1944 	if (witness_watch == -1) {
1945 		mtx_unlock_spin(&w_mtx);
1946 		return (NULL);
1947 	}
1948 	if (STAILQ_EMPTY(&w_free)) {
1949 		witness_watch = -1;
1950 		mtx_unlock_spin(&w_mtx);
1951 		printf("WITNESS: unable to allocate a new witness object\n");
1952 		return (NULL);
1953 	}
1954 	w = STAILQ_FIRST(&w_free);
1955 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1956 	w_free_cnt--;
1957 	index = w->w_index;
1958 	MPASS(index > 0 && index == w_max_used_index+1 &&
1959 	    index < WITNESS_COUNT);
1960 	bzero(w, sizeof(*w));
1961 	w->w_index = index;
1962 	if (index > w_max_used_index)
1963 		w_max_used_index = index;
1964 	return (w);
1965 }
1966 
1967 static void
1968 witness_free(struct witness *w)
1969 {
1970 
1971 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1972 	w_free_cnt++;
1973 }
1974 
1975 static struct lock_list_entry *
1976 witness_lock_list_get(void)
1977 {
1978 	struct lock_list_entry *lle;
1979 
1980 	if (witness_watch == -1)
1981 		return (NULL);
1982 	mtx_lock_spin(&w_mtx);
1983 	lle = w_lock_list_free;
1984 	if (lle == NULL) {
1985 		witness_watch = -1;
1986 		mtx_unlock_spin(&w_mtx);
1987 		printf("%s: witness exhausted\n", __func__);
1988 		return (NULL);
1989 	}
1990 	w_lock_list_free = lle->ll_next;
1991 	mtx_unlock_spin(&w_mtx);
1992 	bzero(lle, sizeof(*lle));
1993 	return (lle);
1994 }
1995 
1996 static void
1997 witness_lock_list_free(struct lock_list_entry *lle)
1998 {
1999 
2000 	mtx_lock_spin(&w_mtx);
2001 	lle->ll_next = w_lock_list_free;
2002 	w_lock_list_free = lle;
2003 	mtx_unlock_spin(&w_mtx);
2004 }
2005 
2006 static struct lock_instance *
2007 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2008 {
2009 	struct lock_list_entry *lle;
2010 	struct lock_instance *instance;
2011 	int i;
2012 
2013 	for (lle = list; lle != NULL; lle = lle->ll_next)
2014 		for (i = lle->ll_count - 1; i >= 0; i--) {
2015 			instance = &lle->ll_children[i];
2016 			if (instance->li_lock == lock)
2017 				return (instance);
2018 		}
2019 	return (NULL);
2020 }
2021 
2022 static void
2023 witness_list_lock(struct lock_instance *instance)
2024 {
2025 	struct lock_object *lock;
2026 
2027 	lock = instance->li_lock;
2028 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2029 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2030 	if (lock->lo_witness->w_name != lock->lo_name)
2031 		printf(" (%s)", lock->lo_witness->w_name);
2032 	printf(" r = %d (%p) locked @ %s:%d\n",
2033 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2034 	    instance->li_line);
2035 }
2036 
2037 #ifdef DDB
2038 static int
2039 witness_thread_has_locks(struct thread *td)
2040 {
2041 
2042 	if (td->td_sleeplocks == NULL)
2043 		return (0);
2044 	return (td->td_sleeplocks->ll_count != 0);
2045 }
2046 
2047 static int
2048 witness_proc_has_locks(struct proc *p)
2049 {
2050 	struct thread *td;
2051 
2052 	FOREACH_THREAD_IN_PROC(p, td) {
2053 		if (witness_thread_has_locks(td))
2054 			return (1);
2055 	}
2056 	return (0);
2057 }
2058 #endif
2059 
2060 int
2061 witness_list_locks(struct lock_list_entry **lock_list)
2062 {
2063 	struct lock_list_entry *lle;
2064 	int i, nheld;
2065 
2066 	nheld = 0;
2067 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2068 		for (i = lle->ll_count - 1; i >= 0; i--) {
2069 			witness_list_lock(&lle->ll_children[i]);
2070 			nheld++;
2071 		}
2072 	return (nheld);
2073 }
2074 
2075 /*
2076  * This is a bit risky at best.  We call this function when we have timed
2077  * out acquiring a spin lock, and we assume that the other CPU is stuck
2078  * with this lock held.  So, we go groveling around in the other CPU's
2079  * per-cpu data to try to find the lock instance for this spin lock to
2080  * see when it was last acquired.
2081  */
2082 void
2083 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
2084 {
2085 	struct lock_instance *instance;
2086 	struct pcpu *pc;
2087 
2088 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2089 		return;
2090 	pc = pcpu_find(owner->td_oncpu);
2091 	instance = find_instance(pc->pc_spinlocks, lock);
2092 	if (instance != NULL)
2093 		witness_list_lock(instance);
2094 }
2095 
2096 void
2097 witness_save(struct lock_object *lock, const char **filep, int *linep)
2098 {
2099 	struct lock_list_entry *lock_list;
2100 	struct lock_instance *instance;
2101 	struct lock_class *class;
2102 
2103 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2104 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2105 		return;
2106 	class = LOCK_CLASS(lock);
2107 	if (class->lc_flags & LC_SLEEPLOCK)
2108 		lock_list = curthread->td_sleeplocks;
2109 	else {
2110 		if (witness_skipspin)
2111 			return;
2112 		lock_list = PCPU_GET(spinlocks);
2113 	}
2114 	instance = find_instance(lock_list, lock);
2115 	if (instance == NULL)
2116 		panic("%s: lock (%s) %s not locked", __func__,
2117 		    class->lc_name, lock->lo_name);
2118 	*filep = instance->li_file;
2119 	*linep = instance->li_line;
2120 }
2121 
2122 void
2123 witness_restore(struct lock_object *lock, const char *file, int line)
2124 {
2125 	struct lock_list_entry *lock_list;
2126 	struct lock_instance *instance;
2127 	struct lock_class *class;
2128 
2129 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2130 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2131 		return;
2132 	class = LOCK_CLASS(lock);
2133 	if (class->lc_flags & LC_SLEEPLOCK)
2134 		lock_list = curthread->td_sleeplocks;
2135 	else {
2136 		if (witness_skipspin)
2137 			return;
2138 		lock_list = PCPU_GET(spinlocks);
2139 	}
2140 	instance = find_instance(lock_list, lock);
2141 	if (instance == NULL)
2142 		panic("%s: lock (%s) %s not locked", __func__,
2143 		    class->lc_name, lock->lo_name);
2144 	lock->lo_witness->w_file = file;
2145 	lock->lo_witness->w_line = line;
2146 	instance->li_file = file;
2147 	instance->li_line = line;
2148 }
2149 
2150 void
2151 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2152 {
2153 #ifdef INVARIANT_SUPPORT
2154 	struct lock_instance *instance;
2155 	struct lock_class *class;
2156 
2157 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2158 		return;
2159 	class = LOCK_CLASS(lock);
2160 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2161 		instance = find_instance(curthread->td_sleeplocks, lock);
2162 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2163 		instance = find_instance(PCPU_GET(spinlocks), lock);
2164 	else {
2165 		panic("Lock (%s) %s is not sleep or spin!",
2166 		    class->lc_name, lock->lo_name);
2167 	}
2168 	file = fixup_filename(file);
2169 	switch (flags) {
2170 	case LA_UNLOCKED:
2171 		if (instance != NULL)
2172 			panic("Lock (%s) %s locked @ %s:%d.",
2173 			    class->lc_name, lock->lo_name, file, line);
2174 		break;
2175 	case LA_LOCKED:
2176 	case LA_LOCKED | LA_RECURSED:
2177 	case LA_LOCKED | LA_NOTRECURSED:
2178 	case LA_SLOCKED:
2179 	case LA_SLOCKED | LA_RECURSED:
2180 	case LA_SLOCKED | LA_NOTRECURSED:
2181 	case LA_XLOCKED:
2182 	case LA_XLOCKED | LA_RECURSED:
2183 	case LA_XLOCKED | LA_NOTRECURSED:
2184 		if (instance == NULL) {
2185 			panic("Lock (%s) %s not locked @ %s:%d.",
2186 			    class->lc_name, lock->lo_name, file, line);
2187 			break;
2188 		}
2189 		if ((flags & LA_XLOCKED) != 0 &&
2190 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2191 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2192 			    class->lc_name, lock->lo_name, file, line);
2193 		if ((flags & LA_SLOCKED) != 0 &&
2194 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2195 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
2196 			    class->lc_name, lock->lo_name, file, line);
2197 		if ((flags & LA_RECURSED) != 0 &&
2198 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2199 			panic("Lock (%s) %s not recursed @ %s:%d.",
2200 			    class->lc_name, lock->lo_name, file, line);
2201 		if ((flags & LA_NOTRECURSED) != 0 &&
2202 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2203 			panic("Lock (%s) %s recursed @ %s:%d.",
2204 			    class->lc_name, lock->lo_name, file, line);
2205 		break;
2206 	default:
2207 		panic("Invalid lock assertion at %s:%d.", file, line);
2208 
2209 	}
2210 #endif	/* INVARIANT_SUPPORT */
2211 }
2212 
2213 #ifdef DDB
2214 static void
2215 witness_ddb_list(struct thread *td)
2216 {
2217 
2218 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2219 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2220 
2221 	if (witness_watch < 1)
2222 		return;
2223 
2224 	witness_list_locks(&td->td_sleeplocks);
2225 
2226 	/*
2227 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2228 	 * if td is currently executing on some other CPU and holds spin locks
2229 	 * as we won't display those locks.  If we had a MI way of getting
2230 	 * the per-cpu data for a given cpu then we could use
2231 	 * td->td_oncpu to get the list of spinlocks for this thread
2232 	 * and "fix" this.
2233 	 *
2234 	 * That still wouldn't really fix this unless we locked the scheduler
2235 	 * lock or stopped the other CPU to make sure it wasn't changing the
2236 	 * list out from under us.  It is probably best to just not try to
2237 	 * handle threads on other CPU's for now.
2238 	 */
2239 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2240 		witness_list_locks(PCPU_PTR(spinlocks));
2241 }
2242 
2243 DB_SHOW_COMMAND(locks, db_witness_list)
2244 {
2245 	struct thread *td;
2246 
2247 	if (have_addr)
2248 		td = db_lookup_thread(addr, TRUE);
2249 	else
2250 		td = kdb_thread;
2251 	witness_ddb_list(td);
2252 }
2253 
2254 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2255 {
2256 	struct thread *td;
2257 	struct proc *p;
2258 
2259 	/*
2260 	 * It would be nice to list only threads and processes that actually
2261 	 * held sleep locks, but that information is currently not exported
2262 	 * by WITNESS.
2263 	 */
2264 	FOREACH_PROC_IN_SYSTEM(p) {
2265 		if (!witness_proc_has_locks(p))
2266 			continue;
2267 		FOREACH_THREAD_IN_PROC(p, td) {
2268 			if (!witness_thread_has_locks(td))
2269 				continue;
2270 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2271 			    p->p_comm, td, td->td_tid);
2272 			witness_ddb_list(td);
2273 		}
2274 	}
2275 }
2276 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2277 
2278 DB_SHOW_COMMAND(witness, db_witness_display)
2279 {
2280 
2281 	witness_ddb_display(db_printf);
2282 }
2283 #endif
2284 
2285 static int
2286 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2287 {
2288 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2289 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2290 	struct sbuf *sb;
2291 	u_int w_rmatrix1, w_rmatrix2;
2292 	int error, generation, i, j;
2293 
2294 	tmp_data1 = NULL;
2295 	tmp_data2 = NULL;
2296 	tmp_w1 = NULL;
2297 	tmp_w2 = NULL;
2298 	if (witness_watch < 1) {
2299 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2300 		return (error);
2301 	}
2302 	if (witness_cold) {
2303 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2304 		return (error);
2305 	}
2306 	error = 0;
2307 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2308 	if (sb == NULL)
2309 		return (ENOMEM);
2310 
2311 	/* Allocate and init temporary storage space. */
2312 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2313 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2314 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2315 	    M_WAITOK | M_ZERO);
2316 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2317 	    M_WAITOK | M_ZERO);
2318 	stack_zero(&tmp_data1->wlod_stack);
2319 	stack_zero(&tmp_data2->wlod_stack);
2320 
2321 restart:
2322 	mtx_lock_spin(&w_mtx);
2323 	generation = w_generation;
2324 	mtx_unlock_spin(&w_mtx);
2325 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2326 	    w_lohash.wloh_count);
2327 	for (i = 1; i < w_max_used_index; i++) {
2328 		mtx_lock_spin(&w_mtx);
2329 		if (generation != w_generation) {
2330 			mtx_unlock_spin(&w_mtx);
2331 
2332 			/* The graph has changed, try again. */
2333 			req->oldidx = 0;
2334 			sbuf_clear(sb);
2335 			goto restart;
2336 		}
2337 
2338 		w1 = &w_data[i];
2339 		if (w1->w_reversed == 0) {
2340 			mtx_unlock_spin(&w_mtx);
2341 			continue;
2342 		}
2343 
2344 		/* Copy w1 locally so we can release the spin lock. */
2345 		*tmp_w1 = *w1;
2346 		mtx_unlock_spin(&w_mtx);
2347 
2348 		if (tmp_w1->w_reversed == 0)
2349 			continue;
2350 		for (j = 1; j < w_max_used_index; j++) {
2351 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2352 				continue;
2353 
2354 			mtx_lock_spin(&w_mtx);
2355 			if (generation != w_generation) {
2356 				mtx_unlock_spin(&w_mtx);
2357 
2358 				/* The graph has changed, try again. */
2359 				req->oldidx = 0;
2360 				sbuf_clear(sb);
2361 				goto restart;
2362 			}
2363 
2364 			w2 = &w_data[j];
2365 			data1 = witness_lock_order_get(w1, w2);
2366 			data2 = witness_lock_order_get(w2, w1);
2367 
2368 			/*
2369 			 * Copy information locally so we can release the
2370 			 * spin lock.
2371 			 */
2372 			*tmp_w2 = *w2;
2373 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2374 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2375 
2376 			if (data1) {
2377 				stack_zero(&tmp_data1->wlod_stack);
2378 				stack_copy(&data1->wlod_stack,
2379 				    &tmp_data1->wlod_stack);
2380 			}
2381 			if (data2 && data2 != data1) {
2382 				stack_zero(&tmp_data2->wlod_stack);
2383 				stack_copy(&data2->wlod_stack,
2384 				    &tmp_data2->wlod_stack);
2385 			}
2386 			mtx_unlock_spin(&w_mtx);
2387 
2388 			sbuf_printf(sb,
2389 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2390 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2391 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2392 #if 0
2393  			sbuf_printf(sb,
2394 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2395  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2396  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2397 #endif
2398 			if (data1) {
2399 				sbuf_printf(sb,
2400 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2401 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2402 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2403 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2404 				sbuf_printf(sb, "\n");
2405 			}
2406 			if (data2 && data2 != data1) {
2407 				sbuf_printf(sb,
2408 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2409 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2410 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2411 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2412 				sbuf_printf(sb, "\n");
2413 			}
2414 		}
2415 	}
2416 	mtx_lock_spin(&w_mtx);
2417 	if (generation != w_generation) {
2418 		mtx_unlock_spin(&w_mtx);
2419 
2420 		/*
2421 		 * The graph changed while we were printing stack data,
2422 		 * try again.
2423 		 */
2424 		req->oldidx = 0;
2425 		sbuf_clear(sb);
2426 		goto restart;
2427 	}
2428 	mtx_unlock_spin(&w_mtx);
2429 
2430 	/* Free temporary storage space. */
2431 	free(tmp_data1, M_TEMP);
2432 	free(tmp_data2, M_TEMP);
2433 	free(tmp_w1, M_TEMP);
2434 	free(tmp_w2, M_TEMP);
2435 
2436 	sbuf_finish(sb);
2437 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2438 	sbuf_delete(sb);
2439 
2440 	return (error);
2441 }
2442 
2443 static int
2444 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2445 {
2446 	struct witness *w;
2447 	struct sbuf *sb;
2448 	int error;
2449 
2450 	if (witness_watch < 1) {
2451 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2452 		return (error);
2453 	}
2454 	if (witness_cold) {
2455 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2456 		return (error);
2457 	}
2458 	error = 0;
2459 	sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2460 	if (sb == NULL)
2461 		return (ENOMEM);
2462 	sbuf_printf(sb, "\n");
2463 
2464 	mtx_lock_spin(&w_mtx);
2465 	STAILQ_FOREACH(w, &w_all, w_list)
2466 		w->w_displayed = 0;
2467 	STAILQ_FOREACH(w, &w_all, w_list)
2468 		witness_add_fullgraph(sb, w);
2469 	mtx_unlock_spin(&w_mtx);
2470 
2471 	/*
2472 	 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2473 	 */
2474 	if (sbuf_overflowed(sb)) {
2475 		sbuf_delete(sb);
2476 		panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2477 		    __func__);
2478 	}
2479 
2480 	/*
2481 	 * Close the sbuf and return to userland.
2482 	 */
2483 	sbuf_finish(sb);
2484 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2485 	sbuf_delete(sb);
2486 
2487 	return (error);
2488 }
2489 
2490 static int
2491 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2492 {
2493 	int error, value;
2494 
2495 	value = witness_watch;
2496 	error = sysctl_handle_int(oidp, &value, 0, req);
2497 	if (error != 0 || req->newptr == NULL)
2498 		return (error);
2499 	if (value > 1 || value < -1 ||
2500 	    (witness_watch == -1 && value != witness_watch))
2501 		return (EINVAL);
2502 	witness_watch = value;
2503 	return (0);
2504 }
2505 
2506 static void
2507 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2508 {
2509 	int i;
2510 
2511 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2512 		return;
2513 	w->w_displayed = 1;
2514 
2515 	WITNESS_INDEX_ASSERT(w->w_index);
2516 	for (i = 1; i <= w_max_used_index; i++) {
2517 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2518 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2519 			    w_data[i].w_name);
2520 			witness_add_fullgraph(sb, &w_data[i]);
2521 		}
2522 	}
2523 }
2524 
2525 /*
2526  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2527  * interprets the key as a string and reads until the null
2528  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2529  * hash value computed from the key.
2530  */
2531 static uint32_t
2532 witness_hash_djb2(const uint8_t *key, uint32_t size)
2533 {
2534 	unsigned int hash = 5381;
2535 	int i;
2536 
2537 	/* hash = hash * 33 + key[i] */
2538 	if (size)
2539 		for (i = 0; i < size; i++)
2540 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2541 	else
2542 		for (i = 0; key[i] != 0; i++)
2543 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2544 
2545 	return (hash);
2546 }
2547 
2548 
2549 /*
2550  * Initializes the two witness hash tables. Called exactly once from
2551  * witness_initialize().
2552  */
2553 static void
2554 witness_init_hash_tables(void)
2555 {
2556 	int i;
2557 
2558 	MPASS(witness_cold);
2559 
2560 	/* Initialize the hash tables. */
2561 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2562 		w_hash.wh_array[i] = NULL;
2563 
2564 	w_hash.wh_size = WITNESS_HASH_SIZE;
2565 	w_hash.wh_count = 0;
2566 
2567 	/* Initialize the lock order data hash. */
2568 	w_lofree = NULL;
2569 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2570 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2571 		w_lodata[i].wlod_next = w_lofree;
2572 		w_lofree = &w_lodata[i];
2573 	}
2574 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2575 	w_lohash.wloh_count = 0;
2576 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2577 		w_lohash.wloh_array[i] = NULL;
2578 }
2579 
2580 static struct witness *
2581 witness_hash_get(const char *key)
2582 {
2583 	struct witness *w;
2584 	uint32_t hash;
2585 
2586 	MPASS(key != NULL);
2587 	if (witness_cold == 0)
2588 		mtx_assert(&w_mtx, MA_OWNED);
2589 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2590 	w = w_hash.wh_array[hash];
2591 	while (w != NULL) {
2592 		if (strcmp(w->w_name, key) == 0)
2593 			goto out;
2594 		w = w->w_hash_next;
2595 	}
2596 
2597 out:
2598 	return (w);
2599 }
2600 
2601 static void
2602 witness_hash_put(struct witness *w)
2603 {
2604 	uint32_t hash;
2605 
2606 	MPASS(w != NULL);
2607 	MPASS(w->w_name != NULL);
2608 	if (witness_cold == 0)
2609 		mtx_assert(&w_mtx, MA_OWNED);
2610 	KASSERT(witness_hash_get(w->w_name) == NULL,
2611 	    ("%s: trying to add a hash entry that already exists!", __func__));
2612 	KASSERT(w->w_hash_next == NULL,
2613 	    ("%s: w->w_hash_next != NULL", __func__));
2614 
2615 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2616 	w->w_hash_next = w_hash.wh_array[hash];
2617 	w_hash.wh_array[hash] = w;
2618 	w_hash.wh_count++;
2619 }
2620 
2621 
2622 static struct witness_lock_order_data *
2623 witness_lock_order_get(struct witness *parent, struct witness *child)
2624 {
2625 	struct witness_lock_order_data *data = NULL;
2626 	struct witness_lock_order_key key;
2627 	unsigned int hash;
2628 
2629 	MPASS(parent != NULL && child != NULL);
2630 	key.from = parent->w_index;
2631 	key.to = child->w_index;
2632 	WITNESS_INDEX_ASSERT(key.from);
2633 	WITNESS_INDEX_ASSERT(key.to);
2634 	if ((w_rmatrix[parent->w_index][child->w_index]
2635 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2636 		goto out;
2637 
2638 	hash = witness_hash_djb2((const char*)&key,
2639 	    sizeof(key)) % w_lohash.wloh_size;
2640 	data = w_lohash.wloh_array[hash];
2641 	while (data != NULL) {
2642 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2643 			break;
2644 		data = data->wlod_next;
2645 	}
2646 
2647 out:
2648 	return (data);
2649 }
2650 
2651 /*
2652  * Verify that parent and child have a known relationship, are not the same,
2653  * and child is actually a child of parent.  This is done without w_mtx
2654  * to avoid contention in the common case.
2655  */
2656 static int
2657 witness_lock_order_check(struct witness *parent, struct witness *child)
2658 {
2659 
2660 	if (parent != child &&
2661 	    w_rmatrix[parent->w_index][child->w_index]
2662 	    & WITNESS_LOCK_ORDER_KNOWN &&
2663 	    isitmychild(parent, child))
2664 		return (1);
2665 
2666 	return (0);
2667 }
2668 
2669 static int
2670 witness_lock_order_add(struct witness *parent, struct witness *child)
2671 {
2672 	struct witness_lock_order_data *data = NULL;
2673 	struct witness_lock_order_key key;
2674 	unsigned int hash;
2675 
2676 	MPASS(parent != NULL && child != NULL);
2677 	key.from = parent->w_index;
2678 	key.to = child->w_index;
2679 	WITNESS_INDEX_ASSERT(key.from);
2680 	WITNESS_INDEX_ASSERT(key.to);
2681 	if (w_rmatrix[parent->w_index][child->w_index]
2682 	    & WITNESS_LOCK_ORDER_KNOWN)
2683 		return (1);
2684 
2685 	hash = witness_hash_djb2((const char*)&key,
2686 	    sizeof(key)) % w_lohash.wloh_size;
2687 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2688 	data = w_lofree;
2689 	if (data == NULL)
2690 		return (0);
2691 	w_lofree = data->wlod_next;
2692 	data->wlod_next = w_lohash.wloh_array[hash];
2693 	data->wlod_key = key;
2694 	w_lohash.wloh_array[hash] = data;
2695 	w_lohash.wloh_count++;
2696 	stack_zero(&data->wlod_stack);
2697 	stack_save(&data->wlod_stack);
2698 	return (1);
2699 }
2700 
2701 /* Call this whenver the structure of the witness graph changes. */
2702 static void
2703 witness_increment_graph_generation(void)
2704 {
2705 
2706 	if (witness_cold == 0)
2707 		mtx_assert(&w_mtx, MA_OWNED);
2708 	w_generation++;
2709 }
2710 
2711 #ifdef KDB
2712 static void
2713 _witness_debugger(int cond, const char *msg)
2714 {
2715 
2716 	if (witness_trace && cond)
2717 		kdb_backtrace();
2718 	if (witness_kdb && cond)
2719 		kdb_enter(KDB_WHY_WITNESS, msg);
2720 }
2721 #endif
2722