xref: /freebsd/sys/kern/subr_witness.c (revision 39beb93c3f8bdbf72a61fda42300b5ebed7390c8)
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/sched.h>
107 #include <sys/stack.h>
108 #include <sys/sysctl.h>
109 #include <sys/systm.h>
110 
111 #ifdef DDB
112 #include <ddb/ddb.h>
113 #endif
114 
115 #include <machine/stdarg.h>
116 
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
119 #endif
120 
121 /* Note that these traces do not work with KTR_ALQ. */
122 #if 0
123 #define	KTR_WITNESS	KTR_SUBSYS
124 #else
125 #define	KTR_WITNESS	0
126 #endif
127 
128 #define	LI_RECURSEMASK	0x0000ffff	/* Recursion depth of lock instance. */
129 #define	LI_EXCLUSIVE	0x00010000	/* Exclusive lock instance. */
130 #define	LI_NORELEASE	0x00020000	/* Lock not allowed to be released. */
131 
132 /* Define this to check for blessed mutexes */
133 #undef BLESSING
134 
135 #define	WITNESS_COUNT 		1024
136 #define	WITNESS_CHILDCOUNT 	(WITNESS_COUNT * 4)
137 #define	WITNESS_HASH_SIZE	251	/* Prime, gives load factor < 2 */
138 #define	WITNESS_PENDLIST	512
139 
140 /* Allocate 256 KB of stack data space */
141 #define	WITNESS_LO_DATA_COUNT	2048
142 
143 /* Prime, gives load factor of ~2 at full load */
144 #define	WITNESS_LO_HASH_SIZE	1021
145 
146 /*
147  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
149  * probably be safe for the most part, but it's still a SWAG.
150  */
151 #define	LOCK_NCHILDREN	5
152 #define	LOCK_CHILDCOUNT	2048
153 
154 #define	MAX_W_NAME	64
155 
156 #define	BADSTACK_SBUF_SIZE	(256 * WITNESS_COUNT)
157 #define	CYCLEGRAPH_SBUF_SIZE	8192
158 #define	FULLGRAPH_SBUF_SIZE	32768
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 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 #ifdef BLESSING
280 struct witness_blessed {
281 	const char	*b_lock1;
282 	const char	*b_lock2;
283 };
284 #endif
285 
286 struct witness_pendhelp {
287 	const char		*wh_type;
288 	struct lock_object	*wh_lock;
289 };
290 
291 struct witness_order_list_entry {
292 	const char		*w_name;
293 	struct lock_class	*w_class;
294 };
295 
296 /*
297  * Returns 0 if one of the locks is a spin lock and the other is not.
298  * Returns 1 otherwise.
299  */
300 static __inline int
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 {
303 
304 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 }
307 
308 static __inline int
309 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
310 {
311 
312 	return (key->from == 0 && key->to == 0);
313 }
314 
315 static __inline int
316 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
317     const struct witness_lock_order_key *b)
318 {
319 
320 	return (a->from == b->from && a->to == b->to);
321 }
322 
323 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
324 		    const char *fname);
325 #ifdef KDB
326 static void	_witness_debugger(int cond, const char *msg);
327 #endif
328 static void	adopt(struct witness *parent, struct witness *child);
329 #ifdef BLESSING
330 static int	blessed(struct witness *, struct witness *);
331 #endif
332 static void	depart(struct witness *w);
333 static struct witness	*enroll(const char *description,
334 			    struct lock_class *lock_class);
335 static struct lock_instance	*find_instance(struct lock_list_entry *list,
336 				    struct lock_object *lock);
337 static int	isitmychild(struct witness *parent, struct witness *child);
338 static int	isitmydescendant(struct witness *parent, struct witness *child);
339 static void	itismychild(struct witness *parent, struct witness *child);
340 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
341 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
342 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
343 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
344 #ifdef DDB
345 static void	witness_ddb_compute_levels(void);
346 static void	witness_ddb_display(void(*)(const char *fmt, ...));
347 static void	witness_ddb_display_descendants(void(*)(const char *fmt, ...),
348 		    struct witness *, int indent);
349 static void	witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
350 		    struct witness_list *list);
351 static void	witness_ddb_level_descendants(struct witness *parent, int l);
352 static void	witness_ddb_list(struct thread *td);
353 #endif
354 static void	witness_free(struct witness *m);
355 static struct witness	*witness_get(void);
356 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
357 static struct witness	*witness_hash_get(const char *key);
358 static void	witness_hash_put(struct witness *w);
359 static void	witness_init_hash_tables(void);
360 static void	witness_increment_graph_generation(void);
361 static void	witness_lock_list_free(struct lock_list_entry *lle);
362 static struct lock_list_entry	*witness_lock_list_get(void);
363 static int	witness_lock_order_add(struct witness *parent,
364 		    struct witness *child);
365 static int	witness_lock_order_check(struct witness *parent,
366 		    struct witness *child);
367 static struct witness_lock_order_data	*witness_lock_order_get(
368 					    struct witness *parent,
369 					    struct witness *child);
370 static void	witness_list_lock(struct lock_instance *instance);
371 static void	witness_setflag(struct lock_object *lock, int flag, int set);
372 
373 #ifdef KDB
374 #define	witness_debugger(c)	_witness_debugger(c, __func__)
375 #else
376 #define	witness_debugger(c)
377 #endif
378 
379 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
380 
381 /*
382  * If set to 0, lock order checking is disabled.  If set to -1,
383  * witness is completely disabled.  Otherwise witness performs full
384  * lock order checking for all locks.  At runtime, lock order checking
385  * may be toggled.  However, witness cannot be reenabled once it is
386  * completely disabled.
387  */
388 static int witness_watch = 1;
389 TUNABLE_INT("debug.witness.watch", &witness_watch);
390 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
391     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
392 
393 #ifdef KDB
394 /*
395  * When KDB is enabled and witness_kdb is 1, it will cause the system
396  * to drop into kdebug() when:
397  *	- a lock hierarchy violation occurs
398  *	- locks are held when going to sleep.
399  */
400 #ifdef WITNESS_KDB
401 int	witness_kdb = 1;
402 #else
403 int	witness_kdb = 0;
404 #endif
405 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
406 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
407 
408 /*
409  * When KDB is enabled and witness_trace is 1, it will cause the system
410  * to print a stack trace:
411  *	- a lock hierarchy violation occurs
412  *	- locks are held when going to sleep.
413  */
414 int	witness_trace = 1;
415 TUNABLE_INT("debug.witness.trace", &witness_trace);
416 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
417 #endif /* KDB */
418 
419 #ifdef WITNESS_SKIPSPIN
420 int	witness_skipspin = 1;
421 #else
422 int	witness_skipspin = 0;
423 #endif
424 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
425 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
426     0, "");
427 
428 /*
429  * Call this to print out the relations between locks.
430  */
431 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
432     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
433 
434 /*
435  * Call this to print out the witness faulty stacks.
436  */
437 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
438     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
439 
440 static struct mtx w_mtx;
441 
442 /* w_list */
443 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
444 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
445 
446 /* w_typelist */
447 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
448 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
449 
450 /* lock list */
451 static struct lock_list_entry *w_lock_list_free = NULL;
452 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
453 static u_int pending_cnt;
454 
455 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
456 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
457 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
458 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
459     "");
460 
461 static struct witness *w_data;
462 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
463 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
464 static struct witness_hash w_hash;	/* The witness hash table. */
465 
466 /* The lock order data hash */
467 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
468 static struct witness_lock_order_data *w_lofree = NULL;
469 static struct witness_lock_order_hash w_lohash;
470 static int w_max_used_index = 0;
471 static unsigned int w_generation = 0;
472 static const char *w_notrunning = "Witness not running\n";
473 static const char *w_stillcold = "Witness is still cold\n";
474 
475 
476 static struct witness_order_list_entry order_lists[] = {
477 	/*
478 	 * sx locks
479 	 */
480 	{ "proctree", &lock_class_sx },
481 	{ "allproc", &lock_class_sx },
482 	{ "allprison", &lock_class_sx },
483 	{ NULL, NULL },
484 	/*
485 	 * Various mutexes
486 	 */
487 	{ "Giant", &lock_class_mtx_sleep },
488 	{ "pipe mutex", &lock_class_mtx_sleep },
489 	{ "sigio lock", &lock_class_mtx_sleep },
490 	{ "process group", &lock_class_mtx_sleep },
491 	{ "process lock", &lock_class_mtx_sleep },
492 	{ "session", &lock_class_mtx_sleep },
493 	{ "uidinfo hash", &lock_class_rw },
494 #ifdef	HWPMC_HOOKS
495 	{ "pmc-sleep", &lock_class_mtx_sleep },
496 #endif
497 	{ NULL, NULL },
498 	/*
499 	 * Sockets
500 	 */
501 	{ "accept", &lock_class_mtx_sleep },
502 	{ "so_snd", &lock_class_mtx_sleep },
503 	{ "so_rcv", &lock_class_mtx_sleep },
504 	{ "sellck", &lock_class_mtx_sleep },
505 	{ NULL, NULL },
506 	/*
507 	 * Routing
508 	 */
509 	{ "so_rcv", &lock_class_mtx_sleep },
510 	{ "radix node head", &lock_class_rw },
511 	{ "rtentry", &lock_class_mtx_sleep },
512 	{ "ifaddr", &lock_class_mtx_sleep },
513 	{ NULL, NULL },
514 	/*
515 	 * Multicast - protocol locks before interface locks, after UDP locks.
516 	 */
517 	{ "udpinp", &lock_class_rw },
518 	{ "in_multi_mtx", &lock_class_mtx_sleep },
519 	{ "igmp_mtx", &lock_class_mtx_sleep },
520 	{ "if_addr_mtx", &lock_class_mtx_sleep },
521 	{ NULL, NULL },
522 	/*
523 	 * UNIX Domain Sockets
524 	 */
525 	{ "unp", &lock_class_mtx_sleep },
526 	{ "so_snd", &lock_class_mtx_sleep },
527 	{ NULL, NULL },
528 	/*
529 	 * UDP/IP
530 	 */
531 	{ "udp", &lock_class_rw },
532 	{ "udpinp", &lock_class_rw },
533 	{ "so_snd", &lock_class_mtx_sleep },
534 	{ NULL, NULL },
535 	/*
536 	 * TCP/IP
537 	 */
538 	{ "tcp", &lock_class_rw },
539 	{ "tcpinp", &lock_class_rw },
540 	{ "so_snd", &lock_class_mtx_sleep },
541 	{ NULL, NULL },
542 	/*
543 	 * SLIP
544 	 */
545 	{ "slip_mtx", &lock_class_mtx_sleep },
546 	{ "slip sc_mtx", &lock_class_mtx_sleep },
547 	{ NULL, NULL },
548 	/*
549 	 * netatalk
550 	 */
551 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
552 	{ "ddp_mtx", &lock_class_mtx_sleep },
553 	{ NULL, NULL },
554 	/*
555 	 * BPF
556 	 */
557 	{ "bpf global lock", &lock_class_mtx_sleep },
558 	{ "bpf interface lock", &lock_class_mtx_sleep },
559 	{ "bpf cdev lock", &lock_class_mtx_sleep },
560 	{ NULL, NULL },
561 	/*
562 	 * NFS server
563 	 */
564 	{ "nfsd_mtx", &lock_class_mtx_sleep },
565 	{ "so_snd", &lock_class_mtx_sleep },
566 	{ NULL, NULL },
567 
568 	/*
569 	 * IEEE 802.11
570 	 */
571 	{ "802.11 com lock", &lock_class_mtx_sleep},
572 	{ NULL, NULL },
573 	/*
574 	 * Network drivers
575 	 */
576 	{ "network driver", &lock_class_mtx_sleep},
577 	{ NULL, NULL },
578 
579 	/*
580 	 * Netgraph
581 	 */
582 	{ "ng_node", &lock_class_mtx_sleep },
583 	{ "ng_worklist", &lock_class_mtx_sleep },
584 	{ NULL, NULL },
585 	/*
586 	 * CDEV
587 	 */
588 	{ "system map", &lock_class_mtx_sleep },
589 	{ "vm page queue mutex", &lock_class_mtx_sleep },
590 	{ "vnode interlock", &lock_class_mtx_sleep },
591 	{ "cdev", &lock_class_mtx_sleep },
592 	{ NULL, NULL },
593 	/*
594 	 * kqueue/VFS interaction
595 	 */
596 	{ "kqueue", &lock_class_mtx_sleep },
597 	{ "struct mount mtx", &lock_class_mtx_sleep },
598 	{ "vnode interlock", &lock_class_mtx_sleep },
599 	{ NULL, NULL },
600 	/*
601 	 * spin locks
602 	 */
603 #ifdef SMP
604 	{ "ap boot", &lock_class_mtx_spin },
605 #endif
606 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
607 	{ "sio", &lock_class_mtx_spin },
608 	{ "scrlock", &lock_class_mtx_spin },
609 #ifdef __i386__
610 	{ "cy", &lock_class_mtx_spin },
611 #endif
612 #ifdef __sparc64__
613 	{ "pcib_mtx", &lock_class_mtx_spin },
614 	{ "rtc_mtx", &lock_class_mtx_spin },
615 #endif
616 	{ "scc_hwmtx", &lock_class_mtx_spin },
617 	{ "uart_hwmtx", &lock_class_mtx_spin },
618 	{ "fast_taskqueue", &lock_class_mtx_spin },
619 	{ "intr table", &lock_class_mtx_spin },
620 #ifdef	HWPMC_HOOKS
621 	{ "pmc-per-proc", &lock_class_mtx_spin },
622 #endif
623 	{ "process slock", &lock_class_mtx_spin },
624 	{ "sleepq chain", &lock_class_mtx_spin },
625 	{ "umtx lock", &lock_class_mtx_spin },
626 	{ "rm_spinlock", &lock_class_mtx_spin },
627 	{ "turnstile chain", &lock_class_mtx_spin },
628 	{ "turnstile lock", &lock_class_mtx_spin },
629 	{ "sched lock", &lock_class_mtx_spin },
630 	{ "td_contested", &lock_class_mtx_spin },
631 	{ "callout", &lock_class_mtx_spin },
632 	{ "entropy harvest mutex", &lock_class_mtx_spin },
633 	{ "syscons video lock", &lock_class_mtx_spin },
634 	{ "time lock", &lock_class_mtx_spin },
635 #ifdef SMP
636 	{ "smp rendezvous", &lock_class_mtx_spin },
637 #endif
638 #ifdef __powerpc__
639 	{ "tlb0", &lock_class_mtx_spin },
640 #endif
641 	/*
642 	 * leaf locks
643 	 */
644 	{ "intrcnt", &lock_class_mtx_spin },
645 	{ "icu", &lock_class_mtx_spin },
646 #if defined(SMP) && defined(__sparc64__)
647 	{ "ipi", &lock_class_mtx_spin },
648 #endif
649 #ifdef __i386__
650 	{ "allpmaps", &lock_class_mtx_spin },
651 	{ "descriptor tables", &lock_class_mtx_spin },
652 #endif
653 	{ "clk", &lock_class_mtx_spin },
654 	{ "cpuset", &lock_class_mtx_spin },
655 	{ "mprof lock", &lock_class_mtx_spin },
656 	{ "zombie lock", &lock_class_mtx_spin },
657 	{ "ALD Queue", &lock_class_mtx_spin },
658 #ifdef __ia64__
659 	{ "MCA spin lock", &lock_class_mtx_spin },
660 #endif
661 #if defined(__i386__) || defined(__amd64__)
662 	{ "pcicfg", &lock_class_mtx_spin },
663 	{ "NDIS thread lock", &lock_class_mtx_spin },
664 #endif
665 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
666 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
667 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
668 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
669 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
670 #ifdef	HWPMC_HOOKS
671 	{ "pmc-leaf", &lock_class_mtx_spin },
672 #endif
673 	{ "blocked lock", &lock_class_mtx_spin },
674 	{ NULL, NULL },
675 	{ NULL, NULL }
676 };
677 
678 #ifdef BLESSING
679 /*
680  * Pairs of locks which have been blessed
681  * Don't complain about order problems with blessed locks
682  */
683 static struct witness_blessed blessed_list[] = {
684 };
685 static int blessed_count =
686 	sizeof(blessed_list) / sizeof(struct witness_blessed);
687 #endif
688 
689 /*
690  * This global is set to 0 once it becomes safe to use the witness code.
691  */
692 static int witness_cold = 1;
693 
694 /*
695  * This global is set to 1 once the static lock orders have been enrolled
696  * so that a warning can be issued for any spin locks enrolled later.
697  */
698 static int witness_spin_warn = 0;
699 
700 /*
701  * The WITNESS-enabled diagnostic code.  Note that the witness code does
702  * assume that the early boot is single-threaded at least until after this
703  * routine is completed.
704  */
705 static void
706 witness_initialize(void *dummy __unused)
707 {
708 	struct lock_object *lock;
709 	struct witness_order_list_entry *order;
710 	struct witness *w, *w1;
711 	int i;
712 
713 	w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
714 	    M_NOWAIT | M_ZERO);
715 
716 	/*
717 	 * We have to release Giant before initializing its witness
718 	 * structure so that WITNESS doesn't get confused.
719 	 */
720 	mtx_unlock(&Giant);
721 	mtx_assert(&Giant, MA_NOTOWNED);
722 
723 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
724 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
725 	    MTX_NOWITNESS | MTX_NOPROFILE);
726 	for (i = WITNESS_COUNT - 1; i >= 0; i--) {
727 		w = &w_data[i];
728 		memset(w, 0, sizeof(*w));
729 		w_data[i].w_index = i;	/* Witness index never changes. */
730 		witness_free(w);
731 	}
732 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
733 	    ("%s: Invalid list of free witness objects", __func__));
734 
735 	/* Witness with index 0 is not used to aid in debugging. */
736 	STAILQ_REMOVE_HEAD(&w_free, w_list);
737 	w_free_cnt--;
738 
739 	memset(w_rmatrix, 0,
740 	    (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
741 
742 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
743 		witness_lock_list_free(&w_locklistdata[i]);
744 	witness_init_hash_tables();
745 
746 	/* First add in all the specified order lists. */
747 	for (order = order_lists; order->w_name != NULL; order++) {
748 		w = enroll(order->w_name, order->w_class);
749 		if (w == NULL)
750 			continue;
751 		w->w_file = "order list";
752 		for (order++; order->w_name != NULL; order++) {
753 			w1 = enroll(order->w_name, order->w_class);
754 			if (w1 == NULL)
755 				continue;
756 			w1->w_file = "order list";
757 			itismychild(w, w1);
758 			w = w1;
759 		}
760 	}
761 	witness_spin_warn = 1;
762 
763 	/* Iterate through all locks and add them to witness. */
764 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
765 		lock = pending_locks[i].wh_lock;
766 		KASSERT(lock->lo_flags & LO_WITNESS,
767 		    ("%s: lock %s is on pending list but not LO_WITNESS",
768 		    __func__, lock->lo_name));
769 		lock->lo_witness = enroll(pending_locks[i].wh_type,
770 		    LOCK_CLASS(lock));
771 	}
772 
773 	/* Mark the witness code as being ready for use. */
774 	witness_cold = 0;
775 
776 	mtx_lock(&Giant);
777 }
778 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
779     NULL);
780 
781 void
782 witness_init(struct lock_object *lock, const char *type)
783 {
784 	struct lock_class *class;
785 
786 	/* Various sanity checks. */
787 	class = LOCK_CLASS(lock);
788 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
789 	    (class->lc_flags & LC_RECURSABLE) == 0)
790 		panic("%s: lock (%s) %s can not be recursable", __func__,
791 		    class->lc_name, lock->lo_name);
792 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
793 	    (class->lc_flags & LC_SLEEPABLE) == 0)
794 		panic("%s: lock (%s) %s can not be sleepable", __func__,
795 		    class->lc_name, lock->lo_name);
796 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
797 	    (class->lc_flags & LC_UPGRADABLE) == 0)
798 		panic("%s: lock (%s) %s can not be upgradable", __func__,
799 		    class->lc_name, lock->lo_name);
800 
801 	/*
802 	 * If we shouldn't watch this lock, then just clear lo_witness.
803 	 * Otherwise, if witness_cold is set, then it is too early to
804 	 * enroll this lock, so defer it to witness_initialize() by adding
805 	 * it to the pending_locks list.  If it is not too early, then enroll
806 	 * the lock now.
807 	 */
808 	if (witness_watch < 1 || panicstr != NULL ||
809 	    (lock->lo_flags & LO_WITNESS) == 0)
810 		lock->lo_witness = NULL;
811 	else if (witness_cold) {
812 		pending_locks[pending_cnt].wh_lock = lock;
813 		pending_locks[pending_cnt++].wh_type = type;
814 		if (pending_cnt > WITNESS_PENDLIST)
815 			panic("%s: pending locks list is too small, bump it\n",
816 			    __func__);
817 	} else
818 		lock->lo_witness = enroll(type, class);
819 }
820 
821 void
822 witness_destroy(struct lock_object *lock)
823 {
824 	struct lock_class *class;
825 	struct witness *w;
826 
827 	class = LOCK_CLASS(lock);
828 
829 	if (witness_cold)
830 		panic("lock (%s) %s destroyed while witness_cold",
831 		    class->lc_name, lock->lo_name);
832 
833 	/* XXX: need to verify that no one holds the lock */
834 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
835 		return;
836 	w = lock->lo_witness;
837 
838 	mtx_lock_spin(&w_mtx);
839 	MPASS(w->w_refcount > 0);
840 	w->w_refcount--;
841 
842 	if (w->w_refcount == 0)
843 		depart(w);
844 	mtx_unlock_spin(&w_mtx);
845 }
846 
847 #ifdef DDB
848 static void
849 witness_ddb_compute_levels(void)
850 {
851 	struct witness *w;
852 
853 	/*
854 	 * First clear all levels.
855 	 */
856 	STAILQ_FOREACH(w, &w_all, w_list)
857 		w->w_ddb_level = -1;
858 
859 	/*
860 	 * Look for locks with no parents and level all their descendants.
861 	 */
862 	STAILQ_FOREACH(w, &w_all, w_list) {
863 
864 		/* If the witness has ancestors (is not a root), skip it. */
865 		if (w->w_num_ancestors > 0)
866 			continue;
867 		witness_ddb_level_descendants(w, 0);
868 	}
869 }
870 
871 static void
872 witness_ddb_level_descendants(struct witness *w, int l)
873 {
874 	int i;
875 
876 	if (w->w_ddb_level >= l)
877 		return;
878 
879 	w->w_ddb_level = l;
880 	l++;
881 
882 	for (i = 1; i <= w_max_used_index; i++) {
883 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
884 			witness_ddb_level_descendants(&w_data[i], l);
885 	}
886 }
887 
888 static void
889 witness_ddb_display_descendants(void(*prnt)(const char *fmt, ...),
890     struct witness *w, int indent)
891 {
892 	int i;
893 
894  	for (i = 0; i < indent; i++)
895  		prnt(" ");
896 	prnt("%s (type: %s, depth: %d, active refs: %d)",
897 	     w->w_name, w->w_class->lc_name,
898 	     w->w_ddb_level, w->w_refcount);
899  	if (w->w_displayed) {
900  		prnt(" -- (already displayed)\n");
901  		return;
902  	}
903  	w->w_displayed = 1;
904 	if (w->w_file != NULL && w->w_line != 0)
905 		prnt(" -- last acquired @ %s:%d\n", w->w_file,
906 		    w->w_line);
907 	else
908 		prnt(" -- never acquired\n");
909 	indent++;
910 	WITNESS_INDEX_ASSERT(w->w_index);
911 	for (i = 1; i <= w_max_used_index; i++) {
912 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
913 			witness_ddb_display_descendants(prnt, &w_data[i],
914 			    indent);
915 	}
916 }
917 
918 static void
919 witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
920     struct witness_list *list)
921 {
922 	struct witness *w;
923 
924 	STAILQ_FOREACH(w, list, w_typelist) {
925 		if (w->w_file == NULL || w->w_ddb_level > 0)
926 			continue;
927 
928 		/* This lock has no anscestors - display its descendants. */
929 		witness_ddb_display_descendants(prnt, w, 0);
930 	}
931 }
932 
933 static void
934 witness_ddb_display(void(*prnt)(const char *fmt, ...))
935 {
936 	struct witness *w;
937 
938 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
939 	witness_ddb_compute_levels();
940 
941 	/* Clear all the displayed flags. */
942 	STAILQ_FOREACH(w, &w_all, w_list)
943 		w->w_displayed = 0;
944 
945 	/*
946 	 * First, handle sleep locks which have been acquired at least
947 	 * once.
948 	 */
949 	prnt("Sleep locks:\n");
950 	witness_ddb_display_list(prnt, &w_sleep);
951 
952 	/*
953 	 * Now do spin locks which have been acquired at least once.
954 	 */
955 	prnt("\nSpin locks:\n");
956 	witness_ddb_display_list(prnt, &w_spin);
957 
958 	/*
959 	 * Finally, any locks which have not been acquired yet.
960 	 */
961 	prnt("\nLocks which were never acquired:\n");
962 	STAILQ_FOREACH(w, &w_all, w_list) {
963 		if (w->w_file != NULL || w->w_refcount == 0)
964 			continue;
965 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
966 		    w->w_class->lc_name, w->w_ddb_level);
967 	}
968 }
969 #endif /* DDB */
970 
971 /* Trim useless garbage from filenames. */
972 static const char *
973 fixup_filename(const char *file)
974 {
975 
976 	if (file == NULL)
977 		return (NULL);
978 	while (strncmp(file, "../", 3) == 0)
979 		file += 3;
980 	return (file);
981 }
982 
983 int
984 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
985 {
986 
987 	if (witness_watch == -1 || panicstr != NULL)
988 		return (0);
989 
990 	/* Require locks that witness knows about. */
991 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
992 	    lock2->lo_witness == NULL)
993 		return (EINVAL);
994 
995 	mtx_assert(&w_mtx, MA_NOTOWNED);
996 	mtx_lock_spin(&w_mtx);
997 
998 	/*
999 	 * If we already have either an explicit or implied lock order that
1000 	 * is the other way around, then return an error.
1001 	 */
1002 	if (witness_watch &&
1003 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1004 		mtx_unlock_spin(&w_mtx);
1005 		return (EDOOFUS);
1006 	}
1007 
1008 	/* Try to add the new order. */
1009 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1010 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1011 	itismychild(lock1->lo_witness, lock2->lo_witness);
1012 	mtx_unlock_spin(&w_mtx);
1013 	return (0);
1014 }
1015 
1016 void
1017 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1018     int line, struct lock_object *interlock)
1019 {
1020 	struct lock_list_entry *lock_list, *lle;
1021 	struct lock_instance *lock1, *lock2, *plock;
1022 	struct lock_class *class;
1023 	struct witness *w, *w1;
1024 	struct thread *td;
1025 	int i, j;
1026 
1027 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1028 	    panicstr != NULL)
1029 		return;
1030 
1031 	w = lock->lo_witness;
1032 	class = LOCK_CLASS(lock);
1033 	td = curthread;
1034 	file = fixup_filename(file);
1035 
1036 	if (class->lc_flags & LC_SLEEPLOCK) {
1037 
1038 		/*
1039 		 * Since spin locks include a critical section, this check
1040 		 * implicitly enforces a lock order of all sleep locks before
1041 		 * all spin locks.
1042 		 */
1043 		if (td->td_critnest != 0 && !kdb_active)
1044 			panic("blockable sleep lock (%s) %s @ %s:%d",
1045 			    class->lc_name, lock->lo_name, file, line);
1046 
1047 		/*
1048 		 * If this is the first lock acquired then just return as
1049 		 * no order checking is needed.
1050 		 */
1051 		lock_list = td->td_sleeplocks;
1052 		if (lock_list == NULL || lock_list->ll_count == 0)
1053 			return;
1054 	} else {
1055 
1056 		/*
1057 		 * If this is the first lock, just return as no order
1058 		 * checking is needed.  Avoid problems with thread
1059 		 * migration pinning the thread while checking if
1060 		 * spinlocks are held.  If at least one spinlock is held
1061 		 * the thread is in a safe path and it is allowed to
1062 		 * unpin it.
1063 		 */
1064 		sched_pin();
1065 		lock_list = PCPU_GET(spinlocks);
1066 		if (lock_list == NULL || lock_list->ll_count == 0) {
1067 			sched_unpin();
1068 			return;
1069 		}
1070 		sched_unpin();
1071 	}
1072 
1073 	/*
1074 	 * Check to see if we are recursing on a lock we already own.  If
1075 	 * so, make sure that we don't mismatch exclusive and shared lock
1076 	 * acquires.
1077 	 */
1078 	lock1 = find_instance(lock_list, lock);
1079 	if (lock1 != NULL) {
1080 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1081 		    (flags & LOP_EXCLUSIVE) == 0) {
1082 			printf("shared lock of (%s) %s @ %s:%d\n",
1083 			    class->lc_name, lock->lo_name, file, line);
1084 			printf("while exclusively locked from %s:%d\n",
1085 			    lock1->li_file, lock1->li_line);
1086 			panic("share->excl");
1087 		}
1088 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1089 		    (flags & LOP_EXCLUSIVE) != 0) {
1090 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1091 			    class->lc_name, lock->lo_name, file, line);
1092 			printf("while share locked from %s:%d\n",
1093 			    lock1->li_file, lock1->li_line);
1094 			panic("excl->share");
1095 		}
1096 		return;
1097 	}
1098 
1099 	/*
1100 	 * Find the previously acquired lock, but ignore interlocks.
1101 	 */
1102 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1103 	if (interlock != NULL && plock->li_lock == interlock) {
1104 		if (lock_list->ll_count > 1)
1105 			plock =
1106 			    &lock_list->ll_children[lock_list->ll_count - 2];
1107 		else {
1108 			lle = lock_list->ll_next;
1109 
1110 			/*
1111 			 * The interlock is the only lock we hold, so
1112 			 * simply return.
1113 			 */
1114 			if (lle == NULL)
1115 				return;
1116 			plock = &lle->ll_children[lle->ll_count - 1];
1117 		}
1118 	}
1119 
1120 	/*
1121 	 * Try to perform most checks without a lock.  If this succeeds we
1122 	 * can skip acquiring the lock and return success.
1123 	 */
1124 	w1 = plock->li_lock->lo_witness;
1125 	if (witness_lock_order_check(w1, w))
1126 		return;
1127 
1128 	/*
1129 	 * Check for duplicate locks of the same type.  Note that we only
1130 	 * have to check for this on the last lock we just acquired.  Any
1131 	 * other cases will be caught as lock order violations.
1132 	 */
1133 	mtx_lock_spin(&w_mtx);
1134 	witness_lock_order_add(w1, w);
1135 	if (w1 == w) {
1136 		i = w->w_index;
1137 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1138 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1139 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1140 			w->w_reversed = 1;
1141 			mtx_unlock_spin(&w_mtx);
1142 			printf(
1143 			    "acquiring duplicate lock of same type: \"%s\"\n",
1144 			    w->w_name);
1145 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1146 			       plock->li_file, plock->li_line);
1147 			printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1148 			witness_debugger(1);
1149 		    } else
1150 			    mtx_unlock_spin(&w_mtx);
1151 		return;
1152 	}
1153 	mtx_assert(&w_mtx, MA_OWNED);
1154 
1155 	/*
1156 	 * If we know that the the lock we are acquiring comes after
1157 	 * the lock we most recently acquired in the lock order tree,
1158 	 * then there is no need for any further checks.
1159 	 */
1160 	if (isitmychild(w1, w))
1161 		goto out;
1162 
1163 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1164 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1165 
1166 			MPASS(j < WITNESS_COUNT);
1167 			lock1 = &lle->ll_children[i];
1168 
1169 			/*
1170 			 * Ignore the interlock the first time we see it.
1171 			 */
1172 			if (interlock != NULL && interlock == lock1->li_lock) {
1173 				interlock = NULL;
1174 				continue;
1175 			}
1176 
1177 			/*
1178 			 * If this lock doesn't undergo witness checking,
1179 			 * then skip it.
1180 			 */
1181 			w1 = lock1->li_lock->lo_witness;
1182 			if (w1 == NULL) {
1183 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1184 				    ("lock missing witness structure"));
1185 				continue;
1186 			}
1187 
1188 			/*
1189 			 * If we are locking Giant and this is a sleepable
1190 			 * lock, then skip it.
1191 			 */
1192 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1193 			    lock == &Giant.lock_object)
1194 				continue;
1195 
1196 			/*
1197 			 * If we are locking a sleepable lock and this lock
1198 			 * is Giant, then skip it.
1199 			 */
1200 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1201 			    lock1->li_lock == &Giant.lock_object)
1202 				continue;
1203 
1204 			/*
1205 			 * If we are locking a sleepable lock and this lock
1206 			 * isn't sleepable, we want to treat it as a lock
1207 			 * order violation to enfore a general lock order of
1208 			 * sleepable locks before non-sleepable locks.
1209 			 */
1210 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1211 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1212 				goto reversal;
1213 
1214 			/*
1215 			 * If we are locking Giant and this is a non-sleepable
1216 			 * lock, then treat it as a reversal.
1217 			 */
1218 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1219 			    lock == &Giant.lock_object)
1220 				goto reversal;
1221 
1222 			/*
1223 			 * Check the lock order hierarchy for a reveresal.
1224 			 */
1225 			if (!isitmydescendant(w, w1))
1226 				continue;
1227 		reversal:
1228 
1229 			/*
1230 			 * We have a lock order violation, check to see if it
1231 			 * is allowed or has already been yelled about.
1232 			 */
1233 #ifdef BLESSING
1234 
1235 			/*
1236 			 * If the lock order is blessed, just bail.  We don't
1237 			 * look for other lock order violations though, which
1238 			 * may be a bug.
1239 			 */
1240 			if (blessed(w, w1))
1241 				goto out;
1242 #endif
1243 
1244 			/* Bail if this violation is known */
1245 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1246 				goto out;
1247 
1248 			/* Record this as a violation */
1249 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1250 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1251 			w->w_reversed = w1->w_reversed = 1;
1252 			witness_increment_graph_generation();
1253 			mtx_unlock_spin(&w_mtx);
1254 
1255 			/*
1256 			 * Ok, yell about it.
1257 			 */
1258 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1259 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1260 				printf(
1261 		"lock order reversal: (sleepable after non-sleepable)\n");
1262 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1263 			    && lock == &Giant.lock_object)
1264 				printf(
1265 		"lock order reversal: (Giant after non-sleepable)\n");
1266 			else
1267 				printf("lock order reversal:\n");
1268 
1269 			/*
1270 			 * Try to locate an earlier lock with
1271 			 * witness w in our list.
1272 			 */
1273 			do {
1274 				lock2 = &lle->ll_children[i];
1275 				MPASS(lock2->li_lock != NULL);
1276 				if (lock2->li_lock->lo_witness == w)
1277 					break;
1278 				if (i == 0 && lle->ll_next != NULL) {
1279 					lle = lle->ll_next;
1280 					i = lle->ll_count - 1;
1281 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1282 				} else
1283 					i--;
1284 			} while (i >= 0);
1285 			if (i < 0) {
1286 				printf(" 1st %p %s (%s) @ %s:%d\n",
1287 				    lock1->li_lock, lock1->li_lock->lo_name,
1288 				    w1->w_name, lock1->li_file, lock1->li_line);
1289 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1290 				    lock->lo_name, w->w_name, file, line);
1291 			} else {
1292 				printf(" 1st %p %s (%s) @ %s:%d\n",
1293 				    lock2->li_lock, lock2->li_lock->lo_name,
1294 				    lock2->li_lock->lo_witness->w_name,
1295 				    lock2->li_file, lock2->li_line);
1296 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1297 				    lock1->li_lock, lock1->li_lock->lo_name,
1298 				    w1->w_name, lock1->li_file, lock1->li_line);
1299 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1300 				    lock->lo_name, w->w_name, file, line);
1301 			}
1302 			witness_debugger(1);
1303 			return;
1304 		}
1305 	}
1306 
1307 	/*
1308 	 * If requested, build a new lock order.  However, don't build a new
1309 	 * relationship between a sleepable lock and Giant if it is in the
1310 	 * wrong direction.  The correct lock order is that sleepable locks
1311 	 * always come before Giant.
1312 	 */
1313 	if (flags & LOP_NEWORDER &&
1314 	    !(plock->li_lock == &Giant.lock_object &&
1315 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1316 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1317 		    w->w_name, plock->li_lock->lo_witness->w_name);
1318 		itismychild(plock->li_lock->lo_witness, w);
1319 	}
1320 out:
1321 	mtx_unlock_spin(&w_mtx);
1322 }
1323 
1324 void
1325 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1326 {
1327 	struct lock_list_entry **lock_list, *lle;
1328 	struct lock_instance *instance;
1329 	struct witness *w;
1330 	struct thread *td;
1331 
1332 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1333 	    panicstr != NULL)
1334 		return;
1335 	w = lock->lo_witness;
1336 	td = curthread;
1337 	file = fixup_filename(file);
1338 
1339 	/* Determine lock list for this lock. */
1340 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1341 		lock_list = &td->td_sleeplocks;
1342 	else
1343 		lock_list = PCPU_PTR(spinlocks);
1344 
1345 	/* Check to see if we are recursing on a lock we already own. */
1346 	instance = find_instance(*lock_list, lock);
1347 	if (instance != NULL) {
1348 		instance->li_flags++;
1349 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1350 		    td->td_proc->p_pid, lock->lo_name,
1351 		    instance->li_flags & LI_RECURSEMASK);
1352 		instance->li_file = file;
1353 		instance->li_line = line;
1354 		return;
1355 	}
1356 
1357 	/* Update per-witness last file and line acquire. */
1358 	w->w_file = file;
1359 	w->w_line = line;
1360 
1361 	/* Find the next open lock instance in the list and fill it. */
1362 	lle = *lock_list;
1363 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1364 		lle = witness_lock_list_get();
1365 		if (lle == NULL)
1366 			return;
1367 		lle->ll_next = *lock_list;
1368 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1369 		    td->td_proc->p_pid, lle);
1370 		*lock_list = lle;
1371 	}
1372 	instance = &lle->ll_children[lle->ll_count++];
1373 	instance->li_lock = lock;
1374 	instance->li_line = line;
1375 	instance->li_file = file;
1376 	if ((flags & LOP_EXCLUSIVE) != 0)
1377 		instance->li_flags = LI_EXCLUSIVE;
1378 	else
1379 		instance->li_flags = 0;
1380 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1381 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1382 }
1383 
1384 void
1385 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1386 {
1387 	struct lock_instance *instance;
1388 	struct lock_class *class;
1389 
1390 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1391 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1392 		return;
1393 	class = LOCK_CLASS(lock);
1394 	file = fixup_filename(file);
1395 	if (witness_watch) {
1396 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1397 			panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1398 			    class->lc_name, lock->lo_name, file, line);
1399 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1400 			panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1401 			    class->lc_name, lock->lo_name, file, line);
1402 	}
1403 	instance = find_instance(curthread->td_sleeplocks, lock);
1404 	if (instance == NULL)
1405 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1406 		    class->lc_name, lock->lo_name, file, line);
1407 	if (witness_watch) {
1408 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1409 			panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1410 			    class->lc_name, lock->lo_name, file, line);
1411 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1412 			panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1413 			    class->lc_name, lock->lo_name,
1414 			    instance->li_flags & LI_RECURSEMASK, file, line);
1415 	}
1416 	instance->li_flags |= LI_EXCLUSIVE;
1417 }
1418 
1419 void
1420 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1421     int line)
1422 {
1423 	struct lock_instance *instance;
1424 	struct lock_class *class;
1425 
1426 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1427 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1428 		return;
1429 	class = LOCK_CLASS(lock);
1430 	file = fixup_filename(file);
1431 	if (witness_watch) {
1432 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1433 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1434 			    class->lc_name, lock->lo_name, file, line);
1435 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1436 			panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1437 			    class->lc_name, lock->lo_name, file, line);
1438 	}
1439 	instance = find_instance(curthread->td_sleeplocks, lock);
1440 	if (instance == NULL)
1441 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1442 		    class->lc_name, lock->lo_name, file, line);
1443 	if (witness_watch) {
1444 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1445 			panic("downgrade of shared lock (%s) %s @ %s:%d",
1446 			    class->lc_name, lock->lo_name, file, line);
1447 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1448 			panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1449 			    class->lc_name, lock->lo_name,
1450 			    instance->li_flags & LI_RECURSEMASK, file, line);
1451 	}
1452 	instance->li_flags &= ~LI_EXCLUSIVE;
1453 }
1454 
1455 void
1456 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1457 {
1458 	struct lock_list_entry **lock_list, *lle;
1459 	struct lock_instance *instance;
1460 	struct lock_class *class;
1461 	struct thread *td;
1462 	register_t s;
1463 	int i, j;
1464 
1465 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1466 		return;
1467 	td = curthread;
1468 	class = LOCK_CLASS(lock);
1469 	file = fixup_filename(file);
1470 
1471 	/* Find lock instance associated with this lock. */
1472 	if (class->lc_flags & LC_SLEEPLOCK)
1473 		lock_list = &td->td_sleeplocks;
1474 	else
1475 		lock_list = PCPU_PTR(spinlocks);
1476 	lle = *lock_list;
1477 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1478 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1479 			instance = &(*lock_list)->ll_children[i];
1480 			if (instance->li_lock == lock)
1481 				goto found;
1482 		}
1483 
1484 	/*
1485 	 * When disabling WITNESS through witness_watch we could end up in
1486 	 * having registered locks in the td_sleeplocks queue.
1487 	 * We have to make sure we flush these queues, so just search for
1488 	 * eventual register locks and remove them.
1489 	 */
1490 	if (witness_watch > 0)
1491 		panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1492 		    lock->lo_name, file, line);
1493 	else
1494 		return;
1495 found:
1496 
1497 	/* First, check for shared/exclusive mismatches. */
1498 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1499 	    (flags & LOP_EXCLUSIVE) == 0) {
1500 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1501 		    lock->lo_name, file, line);
1502 		printf("while exclusively locked from %s:%d\n",
1503 		    instance->li_file, instance->li_line);
1504 		panic("excl->ushare");
1505 	}
1506 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1507 	    (flags & LOP_EXCLUSIVE) != 0) {
1508 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1509 		    lock->lo_name, file, line);
1510 		printf("while share locked from %s:%d\n", instance->li_file,
1511 		    instance->li_line);
1512 		panic("share->uexcl");
1513 	}
1514 	/* If we are recursed, unrecurse. */
1515 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1516 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1517 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1518 		    instance->li_flags);
1519 		instance->li_flags--;
1520 		return;
1521 	}
1522 	/* The lock is now being dropped, check for NORELEASE flag */
1523 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1524 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1525 		    lock->lo_name, file, line);
1526 		panic("lock marked norelease");
1527 	}
1528 
1529 	/* Otherwise, remove this item from the list. */
1530 	s = intr_disable();
1531 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1532 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1533 	    (*lock_list)->ll_count - 1);
1534 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1535 		(*lock_list)->ll_children[j] =
1536 		    (*lock_list)->ll_children[j + 1];
1537 	(*lock_list)->ll_count--;
1538 	intr_restore(s);
1539 
1540 	/*
1541 	 * In order to reduce contention on w_mtx, we want to keep always an
1542 	 * head object into lists so that frequent allocation from the
1543 	 * free witness pool (and subsequent locking) is avoided.
1544 	 * In order to maintain the current code simple, when the head
1545 	 * object is totally unloaded it means also that we do not have
1546 	 * further objects in the list, so the list ownership needs to be
1547 	 * hand over to another object if the current head needs to be freed.
1548 	 */
1549 	if ((*lock_list)->ll_count == 0) {
1550 		if (*lock_list == lle) {
1551 			if (lle->ll_next == NULL)
1552 				return;
1553 		} else
1554 			lle = *lock_list;
1555 		*lock_list = lle->ll_next;
1556 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1557 		    td->td_proc->p_pid, lle);
1558 		witness_lock_list_free(lle);
1559 	}
1560 }
1561 
1562 void
1563 witness_thread_exit(struct thread *td)
1564 {
1565 	struct lock_list_entry *lle;
1566 	int i, n;
1567 
1568 	lle = td->td_sleeplocks;
1569 	if (lle == NULL || panicstr != NULL)
1570 		return;
1571 	if (lle->ll_count != 0) {
1572 		for (n = 0; lle != NULL; lle = lle->ll_next)
1573 			for (i = lle->ll_count - 1; i >= 0; i--) {
1574 				if (n == 0)
1575 		printf("Thread %p exiting with the following locks held:\n",
1576 					    td);
1577 				n++;
1578 				witness_list_lock(&lle->ll_children[i]);
1579 
1580 			}
1581 		panic("Thread %p cannot exit while holding sleeplocks\n", td);
1582 	}
1583 	witness_lock_list_free(lle);
1584 }
1585 
1586 /*
1587  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1588  * exempt Giant and sleepable locks from the checks as well.  If any
1589  * non-exempt locks are held, then a supplied message is printed to the
1590  * console along with a list of the offending locks.  If indicated in the
1591  * flags then a failure results in a panic as well.
1592  */
1593 int
1594 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1595 {
1596 	struct lock_list_entry *lock_list, *lle;
1597 	struct lock_instance *lock1;
1598 	struct thread *td;
1599 	va_list ap;
1600 	int i, n;
1601 
1602 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1603 		return (0);
1604 	n = 0;
1605 	td = curthread;
1606 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1607 		for (i = lle->ll_count - 1; i >= 0; i--) {
1608 			lock1 = &lle->ll_children[i];
1609 			if (lock1->li_lock == lock)
1610 				continue;
1611 			if (flags & WARN_GIANTOK &&
1612 			    lock1->li_lock == &Giant.lock_object)
1613 				continue;
1614 			if (flags & WARN_SLEEPOK &&
1615 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1616 				continue;
1617 			if (n == 0) {
1618 				va_start(ap, fmt);
1619 				vprintf(fmt, ap);
1620 				va_end(ap);
1621 				printf(" with the following");
1622 				if (flags & WARN_SLEEPOK)
1623 					printf(" non-sleepable");
1624 				printf(" locks held:\n");
1625 			}
1626 			n++;
1627 			witness_list_lock(lock1);
1628 		}
1629 
1630 	/*
1631 	 * Pin the thread in order to avoid problems with thread migration.
1632 	 * Once that all verifies are passed about spinlocks ownership,
1633 	 * the thread is in a safe path and it can be unpinned.
1634 	 */
1635 	sched_pin();
1636 	lock_list = PCPU_GET(spinlocks);
1637 	if (lock_list != NULL && lock_list->ll_count != 0) {
1638 		sched_unpin();
1639 
1640 		/*
1641 		 * We should only have one spinlock and as long as
1642 		 * the flags cannot match for this locks class,
1643 		 * check if the first spinlock is the one curthread
1644 		 * should hold.
1645 		 */
1646 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1647 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1648 		    lock1->li_lock == lock && n == 0)
1649 			return (0);
1650 
1651 		va_start(ap, fmt);
1652 		vprintf(fmt, ap);
1653 		va_end(ap);
1654 		printf(" with the following");
1655 		if (flags & WARN_SLEEPOK)
1656 			printf(" non-sleepable");
1657 		printf(" locks held:\n");
1658 		n += witness_list_locks(&lock_list);
1659 	} else
1660 		sched_unpin();
1661 	if (flags & WARN_PANIC && n)
1662 		panic("%s", __func__);
1663 	else
1664 		witness_debugger(n);
1665 	return (n);
1666 }
1667 
1668 const char *
1669 witness_file(struct lock_object *lock)
1670 {
1671 	struct witness *w;
1672 
1673 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1674 		return ("?");
1675 	w = lock->lo_witness;
1676 	return (w->w_file);
1677 }
1678 
1679 int
1680 witness_line(struct lock_object *lock)
1681 {
1682 	struct witness *w;
1683 
1684 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1685 		return (0);
1686 	w = lock->lo_witness;
1687 	return (w->w_line);
1688 }
1689 
1690 static struct witness *
1691 enroll(const char *description, struct lock_class *lock_class)
1692 {
1693 	struct witness *w;
1694 	struct witness_list *typelist;
1695 
1696 	MPASS(description != NULL);
1697 
1698 	if (witness_watch == -1 || panicstr != NULL)
1699 		return (NULL);
1700 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1701 		if (witness_skipspin)
1702 			return (NULL);
1703 		else
1704 			typelist = &w_spin;
1705 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1706 		typelist = &w_sleep;
1707 	else
1708 		panic("lock class %s is not sleep or spin",
1709 		    lock_class->lc_name);
1710 
1711 	mtx_lock_spin(&w_mtx);
1712 	w = witness_hash_get(description);
1713 	if (w)
1714 		goto found;
1715 	if ((w = witness_get()) == NULL)
1716 		return (NULL);
1717 	MPASS(strlen(description) < MAX_W_NAME);
1718 	strcpy(w->w_name, description);
1719 	w->w_class = lock_class;
1720 	w->w_refcount = 1;
1721 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1722 	if (lock_class->lc_flags & LC_SPINLOCK) {
1723 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1724 		w_spin_cnt++;
1725 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1726 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1727 		w_sleep_cnt++;
1728 	}
1729 
1730 	/* Insert new witness into the hash */
1731 	witness_hash_put(w);
1732 	witness_increment_graph_generation();
1733 	mtx_unlock_spin(&w_mtx);
1734 	return (w);
1735 found:
1736 	w->w_refcount++;
1737 	mtx_unlock_spin(&w_mtx);
1738 	if (lock_class != w->w_class)
1739 		panic(
1740 			"lock (%s) %s does not match earlier (%s) lock",
1741 			description, lock_class->lc_name,
1742 			w->w_class->lc_name);
1743 	return (w);
1744 }
1745 
1746 static void
1747 depart(struct witness *w)
1748 {
1749 	struct witness_list *list;
1750 
1751 	MPASS(w->w_refcount == 0);
1752 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1753 		list = &w_sleep;
1754 		w_sleep_cnt--;
1755 	} else {
1756 		list = &w_spin;
1757 		w_spin_cnt--;
1758 	}
1759 	/*
1760 	 * Set file to NULL as it may point into a loadable module.
1761 	 */
1762 	w->w_file = NULL;
1763 	w->w_line = 0;
1764 	witness_increment_graph_generation();
1765 }
1766 
1767 
1768 static void
1769 adopt(struct witness *parent, struct witness *child)
1770 {
1771 	int pi, ci, i, j;
1772 
1773 	if (witness_cold == 0)
1774 		mtx_assert(&w_mtx, MA_OWNED);
1775 
1776 	/* If the relationship is already known, there's no work to be done. */
1777 	if (isitmychild(parent, child))
1778 		return;
1779 
1780 	/* When the structure of the graph changes, bump up the generation. */
1781 	witness_increment_graph_generation();
1782 
1783 	/*
1784 	 * The hard part ... create the direct relationship, then propagate all
1785 	 * indirect relationships.
1786 	 */
1787 	pi = parent->w_index;
1788 	ci = child->w_index;
1789 	WITNESS_INDEX_ASSERT(pi);
1790 	WITNESS_INDEX_ASSERT(ci);
1791 	MPASS(pi != ci);
1792 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1793 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1794 
1795 	/*
1796 	 * If parent was not already an ancestor of child,
1797 	 * then we increment the descendant and ancestor counters.
1798 	 */
1799 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1800 		parent->w_num_descendants++;
1801 		child->w_num_ancestors++;
1802 	}
1803 
1804 	/*
1805 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1806 	 * an ancestor of 'pi' during this loop.
1807 	 */
1808 	for (i = 1; i <= w_max_used_index; i++) {
1809 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1810 		    (i != pi))
1811 			continue;
1812 
1813 		/* Find each descendant of 'i' and mark it as a descendant. */
1814 		for (j = 1; j <= w_max_used_index; j++) {
1815 
1816 			/*
1817 			 * Skip children that are already marked as
1818 			 * descendants of 'i'.
1819 			 */
1820 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1821 				continue;
1822 
1823 			/*
1824 			 * We are only interested in descendants of 'ci'. Note
1825 			 * that 'ci' itself is counted as a descendant of 'ci'.
1826 			 */
1827 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1828 			    (j != ci))
1829 				continue;
1830 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1831 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1832 			w_data[i].w_num_descendants++;
1833 			w_data[j].w_num_ancestors++;
1834 
1835 			/*
1836 			 * Make sure we aren't marking a node as both an
1837 			 * ancestor and descendant. We should have caught
1838 			 * this as a lock order reversal earlier.
1839 			 */
1840 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1841 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1842 				printf("witness rmatrix paradox! [%d][%d]=%d "
1843 				    "both ancestor and descendant\n",
1844 				    i, j, w_rmatrix[i][j]);
1845 				kdb_backtrace();
1846 				printf("Witness disabled.\n");
1847 				witness_watch = -1;
1848 			}
1849 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1850 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1851 				printf("witness rmatrix paradox! [%d][%d]=%d "
1852 				    "both ancestor and descendant\n",
1853 				    j, i, w_rmatrix[j][i]);
1854 				kdb_backtrace();
1855 				printf("Witness disabled.\n");
1856 				witness_watch = -1;
1857 			}
1858 		}
1859 	}
1860 }
1861 
1862 static void
1863 itismychild(struct witness *parent, struct witness *child)
1864 {
1865 
1866 	MPASS(child != NULL && parent != NULL);
1867 	if (witness_cold == 0)
1868 		mtx_assert(&w_mtx, MA_OWNED);
1869 
1870 	if (!witness_lock_type_equal(parent, child)) {
1871 		if (witness_cold == 0)
1872 			mtx_unlock_spin(&w_mtx);
1873 		panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1874 		    "the same lock type", __func__, parent->w_name,
1875 		    parent->w_class->lc_name, child->w_name,
1876 		    child->w_class->lc_name);
1877 	}
1878 	adopt(parent, child);
1879 }
1880 
1881 /*
1882  * Generic code for the isitmy*() functions. The rmask parameter is the
1883  * expected relationship of w1 to w2.
1884  */
1885 static int
1886 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1887 {
1888 	unsigned char r1, r2;
1889 	int i1, i2;
1890 
1891 	i1 = w1->w_index;
1892 	i2 = w2->w_index;
1893 	WITNESS_INDEX_ASSERT(i1);
1894 	WITNESS_INDEX_ASSERT(i2);
1895 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1896 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1897 
1898 	/* The flags on one better be the inverse of the flags on the other */
1899 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1900 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1901 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
1902 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
1903 		    "w_rmatrix[%d][%d] == %hhx\n",
1904 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1905 		    i2, i1, r2);
1906 		kdb_backtrace();
1907 		printf("Witness disabled.\n");
1908 		witness_watch = -1;
1909 	}
1910 	return (r1 & rmask);
1911 }
1912 
1913 /*
1914  * Checks if @child is a direct child of @parent.
1915  */
1916 static int
1917 isitmychild(struct witness *parent, struct witness *child)
1918 {
1919 
1920 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1921 }
1922 
1923 /*
1924  * Checks if @descendant is a direct or inderect descendant of @ancestor.
1925  */
1926 static int
1927 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1928 {
1929 
1930 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1931 	    __func__));
1932 }
1933 
1934 #ifdef BLESSING
1935 static int
1936 blessed(struct witness *w1, struct witness *w2)
1937 {
1938 	int i;
1939 	struct witness_blessed *b;
1940 
1941 	for (i = 0; i < blessed_count; i++) {
1942 		b = &blessed_list[i];
1943 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1944 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1945 				return (1);
1946 			continue;
1947 		}
1948 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1949 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1950 				return (1);
1951 	}
1952 	return (0);
1953 }
1954 #endif
1955 
1956 static struct witness *
1957 witness_get(void)
1958 {
1959 	struct witness *w;
1960 	int index;
1961 
1962 	if (witness_cold == 0)
1963 		mtx_assert(&w_mtx, MA_OWNED);
1964 
1965 	if (witness_watch == -1) {
1966 		mtx_unlock_spin(&w_mtx);
1967 		return (NULL);
1968 	}
1969 	if (STAILQ_EMPTY(&w_free)) {
1970 		witness_watch = -1;
1971 		mtx_unlock_spin(&w_mtx);
1972 		printf("WITNESS: unable to allocate a new witness object\n");
1973 		return (NULL);
1974 	}
1975 	w = STAILQ_FIRST(&w_free);
1976 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1977 	w_free_cnt--;
1978 	index = w->w_index;
1979 	MPASS(index > 0 && index == w_max_used_index+1 &&
1980 	    index < WITNESS_COUNT);
1981 	bzero(w, sizeof(*w));
1982 	w->w_index = index;
1983 	if (index > w_max_used_index)
1984 		w_max_used_index = index;
1985 	return (w);
1986 }
1987 
1988 static void
1989 witness_free(struct witness *w)
1990 {
1991 
1992 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1993 	w_free_cnt++;
1994 }
1995 
1996 static struct lock_list_entry *
1997 witness_lock_list_get(void)
1998 {
1999 	struct lock_list_entry *lle;
2000 
2001 	if (witness_watch == -1)
2002 		return (NULL);
2003 	mtx_lock_spin(&w_mtx);
2004 	lle = w_lock_list_free;
2005 	if (lle == NULL) {
2006 		witness_watch = -1;
2007 		mtx_unlock_spin(&w_mtx);
2008 		printf("%s: witness exhausted\n", __func__);
2009 		return (NULL);
2010 	}
2011 	w_lock_list_free = lle->ll_next;
2012 	mtx_unlock_spin(&w_mtx);
2013 	bzero(lle, sizeof(*lle));
2014 	return (lle);
2015 }
2016 
2017 static void
2018 witness_lock_list_free(struct lock_list_entry *lle)
2019 {
2020 
2021 	mtx_lock_spin(&w_mtx);
2022 	lle->ll_next = w_lock_list_free;
2023 	w_lock_list_free = lle;
2024 	mtx_unlock_spin(&w_mtx);
2025 }
2026 
2027 static struct lock_instance *
2028 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2029 {
2030 	struct lock_list_entry *lle;
2031 	struct lock_instance *instance;
2032 	int i;
2033 
2034 	for (lle = list; lle != NULL; lle = lle->ll_next)
2035 		for (i = lle->ll_count - 1; i >= 0; i--) {
2036 			instance = &lle->ll_children[i];
2037 			if (instance->li_lock == lock)
2038 				return (instance);
2039 		}
2040 	return (NULL);
2041 }
2042 
2043 static void
2044 witness_list_lock(struct lock_instance *instance)
2045 {
2046 	struct lock_object *lock;
2047 
2048 	lock = instance->li_lock;
2049 	printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2050 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2051 	if (lock->lo_witness->w_name != lock->lo_name)
2052 		printf(" (%s)", lock->lo_witness->w_name);
2053 	printf(" r = %d (%p) locked @ %s:%d\n",
2054 	    instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2055 	    instance->li_line);
2056 }
2057 
2058 #ifdef DDB
2059 static int
2060 witness_thread_has_locks(struct thread *td)
2061 {
2062 
2063 	if (td->td_sleeplocks == NULL)
2064 		return (0);
2065 	return (td->td_sleeplocks->ll_count != 0);
2066 }
2067 
2068 static int
2069 witness_proc_has_locks(struct proc *p)
2070 {
2071 	struct thread *td;
2072 
2073 	FOREACH_THREAD_IN_PROC(p, td) {
2074 		if (witness_thread_has_locks(td))
2075 			return (1);
2076 	}
2077 	return (0);
2078 }
2079 #endif
2080 
2081 int
2082 witness_list_locks(struct lock_list_entry **lock_list)
2083 {
2084 	struct lock_list_entry *lle;
2085 	int i, nheld;
2086 
2087 	nheld = 0;
2088 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2089 		for (i = lle->ll_count - 1; i >= 0; i--) {
2090 			witness_list_lock(&lle->ll_children[i]);
2091 			nheld++;
2092 		}
2093 	return (nheld);
2094 }
2095 
2096 /*
2097  * This is a bit risky at best.  We call this function when we have timed
2098  * out acquiring a spin lock, and we assume that the other CPU is stuck
2099  * with this lock held.  So, we go groveling around in the other CPU's
2100  * per-cpu data to try to find the lock instance for this spin lock to
2101  * see when it was last acquired.
2102  */
2103 void
2104 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
2105 {
2106 	struct lock_instance *instance;
2107 	struct pcpu *pc;
2108 
2109 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2110 		return;
2111 	pc = pcpu_find(owner->td_oncpu);
2112 	instance = find_instance(pc->pc_spinlocks, lock);
2113 	if (instance != NULL)
2114 		witness_list_lock(instance);
2115 }
2116 
2117 void
2118 witness_save(struct lock_object *lock, const char **filep, int *linep)
2119 {
2120 	struct lock_list_entry *lock_list;
2121 	struct lock_instance *instance;
2122 	struct lock_class *class;
2123 
2124 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2125 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2126 		return;
2127 	class = LOCK_CLASS(lock);
2128 	if (class->lc_flags & LC_SLEEPLOCK)
2129 		lock_list = curthread->td_sleeplocks;
2130 	else {
2131 		if (witness_skipspin)
2132 			return;
2133 		lock_list = PCPU_GET(spinlocks);
2134 	}
2135 	instance = find_instance(lock_list, lock);
2136 	if (instance == NULL)
2137 		panic("%s: lock (%s) %s not locked", __func__,
2138 		    class->lc_name, lock->lo_name);
2139 	*filep = instance->li_file;
2140 	*linep = instance->li_line;
2141 }
2142 
2143 void
2144 witness_restore(struct lock_object *lock, const char *file, int line)
2145 {
2146 	struct lock_list_entry *lock_list;
2147 	struct lock_instance *instance;
2148 	struct lock_class *class;
2149 
2150 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2151 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2152 		return;
2153 	class = LOCK_CLASS(lock);
2154 	if (class->lc_flags & LC_SLEEPLOCK)
2155 		lock_list = curthread->td_sleeplocks;
2156 	else {
2157 		if (witness_skipspin)
2158 			return;
2159 		lock_list = PCPU_GET(spinlocks);
2160 	}
2161 	instance = find_instance(lock_list, lock);
2162 	if (instance == NULL)
2163 		panic("%s: lock (%s) %s not locked", __func__,
2164 		    class->lc_name, lock->lo_name);
2165 	lock->lo_witness->w_file = file;
2166 	lock->lo_witness->w_line = line;
2167 	instance->li_file = file;
2168 	instance->li_line = line;
2169 }
2170 
2171 void
2172 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2173 {
2174 #ifdef INVARIANT_SUPPORT
2175 	struct lock_instance *instance;
2176 	struct lock_class *class;
2177 
2178 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2179 		return;
2180 	class = LOCK_CLASS(lock);
2181 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2182 		instance = find_instance(curthread->td_sleeplocks, lock);
2183 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2184 		instance = find_instance(PCPU_GET(spinlocks), lock);
2185 	else {
2186 		panic("Lock (%s) %s is not sleep or spin!",
2187 		    class->lc_name, lock->lo_name);
2188 	}
2189 	file = fixup_filename(file);
2190 	switch (flags) {
2191 	case LA_UNLOCKED:
2192 		if (instance != NULL)
2193 			panic("Lock (%s) %s locked @ %s:%d.",
2194 			    class->lc_name, lock->lo_name, file, line);
2195 		break;
2196 	case LA_LOCKED:
2197 	case LA_LOCKED | LA_RECURSED:
2198 	case LA_LOCKED | LA_NOTRECURSED:
2199 	case LA_SLOCKED:
2200 	case LA_SLOCKED | LA_RECURSED:
2201 	case LA_SLOCKED | LA_NOTRECURSED:
2202 	case LA_XLOCKED:
2203 	case LA_XLOCKED | LA_RECURSED:
2204 	case LA_XLOCKED | LA_NOTRECURSED:
2205 		if (instance == NULL) {
2206 			panic("Lock (%s) %s not locked @ %s:%d.",
2207 			    class->lc_name, lock->lo_name, file, line);
2208 			break;
2209 		}
2210 		if ((flags & LA_XLOCKED) != 0 &&
2211 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2212 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2213 			    class->lc_name, lock->lo_name, file, line);
2214 		if ((flags & LA_SLOCKED) != 0 &&
2215 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2216 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
2217 			    class->lc_name, lock->lo_name, file, line);
2218 		if ((flags & LA_RECURSED) != 0 &&
2219 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2220 			panic("Lock (%s) %s not recursed @ %s:%d.",
2221 			    class->lc_name, lock->lo_name, file, line);
2222 		if ((flags & LA_NOTRECURSED) != 0 &&
2223 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2224 			panic("Lock (%s) %s recursed @ %s:%d.",
2225 			    class->lc_name, lock->lo_name, file, line);
2226 		break;
2227 	default:
2228 		panic("Invalid lock assertion at %s:%d.", file, line);
2229 
2230 	}
2231 #endif	/* INVARIANT_SUPPORT */
2232 }
2233 
2234 static void
2235 witness_setflag(struct lock_object *lock, int flag, int set)
2236 {
2237 	struct lock_list_entry *lock_list;
2238 	struct lock_instance *instance;
2239 	struct lock_class *class;
2240 
2241 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2242 		return;
2243 	class = LOCK_CLASS(lock);
2244 	if (class->lc_flags & LC_SLEEPLOCK)
2245 		lock_list = curthread->td_sleeplocks;
2246 	else {
2247 		if (witness_skipspin)
2248 			return;
2249 		lock_list = PCPU_GET(spinlocks);
2250 	}
2251 	instance = find_instance(lock_list, lock);
2252 	if (instance == NULL)
2253 		panic("%s: lock (%s) %s not locked", __func__,
2254 		    class->lc_name, lock->lo_name);
2255 
2256 	if (set)
2257 		instance->li_flags |= flag;
2258 	else
2259 		instance->li_flags &= ~flag;
2260 }
2261 
2262 void
2263 witness_norelease(struct lock_object *lock)
2264 {
2265 
2266 	witness_setflag(lock, LI_NORELEASE, 1);
2267 }
2268 
2269 void
2270 witness_releaseok(struct lock_object *lock)
2271 {
2272 
2273 	witness_setflag(lock, LI_NORELEASE, 0);
2274 }
2275 
2276 #ifdef DDB
2277 static void
2278 witness_ddb_list(struct thread *td)
2279 {
2280 
2281 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2282 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2283 
2284 	if (witness_watch < 1)
2285 		return;
2286 
2287 	witness_list_locks(&td->td_sleeplocks);
2288 
2289 	/*
2290 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2291 	 * if td is currently executing on some other CPU and holds spin locks
2292 	 * as we won't display those locks.  If we had a MI way of getting
2293 	 * the per-cpu data for a given cpu then we could use
2294 	 * td->td_oncpu to get the list of spinlocks for this thread
2295 	 * and "fix" this.
2296 	 *
2297 	 * That still wouldn't really fix this unless we locked the scheduler
2298 	 * lock or stopped the other CPU to make sure it wasn't changing the
2299 	 * list out from under us.  It is probably best to just not try to
2300 	 * handle threads on other CPU's for now.
2301 	 */
2302 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2303 		witness_list_locks(PCPU_PTR(spinlocks));
2304 }
2305 
2306 DB_SHOW_COMMAND(locks, db_witness_list)
2307 {
2308 	struct thread *td;
2309 
2310 	if (have_addr)
2311 		td = db_lookup_thread(addr, TRUE);
2312 	else
2313 		td = kdb_thread;
2314 	witness_ddb_list(td);
2315 }
2316 
2317 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2318 {
2319 	struct thread *td;
2320 	struct proc *p;
2321 
2322 	/*
2323 	 * It would be nice to list only threads and processes that actually
2324 	 * held sleep locks, but that information is currently not exported
2325 	 * by WITNESS.
2326 	 */
2327 	FOREACH_PROC_IN_SYSTEM(p) {
2328 		if (!witness_proc_has_locks(p))
2329 			continue;
2330 		FOREACH_THREAD_IN_PROC(p, td) {
2331 			if (!witness_thread_has_locks(td))
2332 				continue;
2333 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2334 			    p->p_comm, td, td->td_tid);
2335 			witness_ddb_list(td);
2336 		}
2337 	}
2338 }
2339 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2340 
2341 DB_SHOW_COMMAND(witness, db_witness_display)
2342 {
2343 
2344 	witness_ddb_display(db_printf);
2345 }
2346 #endif
2347 
2348 static int
2349 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2350 {
2351 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2352 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2353 	struct sbuf *sb;
2354 	u_int w_rmatrix1, w_rmatrix2;
2355 	int error, generation, i, j;
2356 
2357 	tmp_data1 = NULL;
2358 	tmp_data2 = NULL;
2359 	tmp_w1 = NULL;
2360 	tmp_w2 = NULL;
2361 	if (witness_watch < 1) {
2362 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2363 		return (error);
2364 	}
2365 	if (witness_cold) {
2366 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2367 		return (error);
2368 	}
2369 	error = 0;
2370 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2371 	if (sb == NULL)
2372 		return (ENOMEM);
2373 
2374 	/* Allocate and init temporary storage space. */
2375 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2376 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2377 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2378 	    M_WAITOK | M_ZERO);
2379 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2380 	    M_WAITOK | M_ZERO);
2381 	stack_zero(&tmp_data1->wlod_stack);
2382 	stack_zero(&tmp_data2->wlod_stack);
2383 
2384 restart:
2385 	mtx_lock_spin(&w_mtx);
2386 	generation = w_generation;
2387 	mtx_unlock_spin(&w_mtx);
2388 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2389 	    w_lohash.wloh_count);
2390 	for (i = 1; i < w_max_used_index; i++) {
2391 		mtx_lock_spin(&w_mtx);
2392 		if (generation != w_generation) {
2393 			mtx_unlock_spin(&w_mtx);
2394 
2395 			/* The graph has changed, try again. */
2396 			req->oldidx = 0;
2397 			sbuf_clear(sb);
2398 			goto restart;
2399 		}
2400 
2401 		w1 = &w_data[i];
2402 		if (w1->w_reversed == 0) {
2403 			mtx_unlock_spin(&w_mtx);
2404 			continue;
2405 		}
2406 
2407 		/* Copy w1 locally so we can release the spin lock. */
2408 		*tmp_w1 = *w1;
2409 		mtx_unlock_spin(&w_mtx);
2410 
2411 		if (tmp_w1->w_reversed == 0)
2412 			continue;
2413 		for (j = 1; j < w_max_used_index; j++) {
2414 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2415 				continue;
2416 
2417 			mtx_lock_spin(&w_mtx);
2418 			if (generation != w_generation) {
2419 				mtx_unlock_spin(&w_mtx);
2420 
2421 				/* The graph has changed, try again. */
2422 				req->oldidx = 0;
2423 				sbuf_clear(sb);
2424 				goto restart;
2425 			}
2426 
2427 			w2 = &w_data[j];
2428 			data1 = witness_lock_order_get(w1, w2);
2429 			data2 = witness_lock_order_get(w2, w1);
2430 
2431 			/*
2432 			 * Copy information locally so we can release the
2433 			 * spin lock.
2434 			 */
2435 			*tmp_w2 = *w2;
2436 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2437 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2438 
2439 			if (data1) {
2440 				stack_zero(&tmp_data1->wlod_stack);
2441 				stack_copy(&data1->wlod_stack,
2442 				    &tmp_data1->wlod_stack);
2443 			}
2444 			if (data2 && data2 != data1) {
2445 				stack_zero(&tmp_data2->wlod_stack);
2446 				stack_copy(&data2->wlod_stack,
2447 				    &tmp_data2->wlod_stack);
2448 			}
2449 			mtx_unlock_spin(&w_mtx);
2450 
2451 			sbuf_printf(sb,
2452 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2453 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2454 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2455 #if 0
2456  			sbuf_printf(sb,
2457 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2458  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2459  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2460 #endif
2461 			if (data1) {
2462 				sbuf_printf(sb,
2463 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2464 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2465 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2466 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2467 				sbuf_printf(sb, "\n");
2468 			}
2469 			if (data2 && data2 != data1) {
2470 				sbuf_printf(sb,
2471 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2472 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2473 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2474 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2475 				sbuf_printf(sb, "\n");
2476 			}
2477 		}
2478 	}
2479 	mtx_lock_spin(&w_mtx);
2480 	if (generation != w_generation) {
2481 		mtx_unlock_spin(&w_mtx);
2482 
2483 		/*
2484 		 * The graph changed while we were printing stack data,
2485 		 * try again.
2486 		 */
2487 		req->oldidx = 0;
2488 		sbuf_clear(sb);
2489 		goto restart;
2490 	}
2491 	mtx_unlock_spin(&w_mtx);
2492 
2493 	/* Free temporary storage space. */
2494 	free(tmp_data1, M_TEMP);
2495 	free(tmp_data2, M_TEMP);
2496 	free(tmp_w1, M_TEMP);
2497 	free(tmp_w2, M_TEMP);
2498 
2499 	sbuf_finish(sb);
2500 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2501 	sbuf_delete(sb);
2502 
2503 	return (error);
2504 }
2505 
2506 static int
2507 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2508 {
2509 	struct witness *w;
2510 	struct sbuf *sb;
2511 	int error;
2512 
2513 	if (witness_watch < 1) {
2514 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2515 		return (error);
2516 	}
2517 	if (witness_cold) {
2518 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2519 		return (error);
2520 	}
2521 	error = 0;
2522 	sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2523 	if (sb == NULL)
2524 		return (ENOMEM);
2525 	sbuf_printf(sb, "\n");
2526 
2527 	mtx_lock_spin(&w_mtx);
2528 	STAILQ_FOREACH(w, &w_all, w_list)
2529 		w->w_displayed = 0;
2530 	STAILQ_FOREACH(w, &w_all, w_list)
2531 		witness_add_fullgraph(sb, w);
2532 	mtx_unlock_spin(&w_mtx);
2533 
2534 	/*
2535 	 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2536 	 */
2537 	if (sbuf_overflowed(sb)) {
2538 		sbuf_delete(sb);
2539 		panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2540 		    __func__);
2541 	}
2542 
2543 	/*
2544 	 * Close the sbuf and return to userland.
2545 	 */
2546 	sbuf_finish(sb);
2547 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2548 	sbuf_delete(sb);
2549 
2550 	return (error);
2551 }
2552 
2553 static int
2554 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2555 {
2556 	int error, value;
2557 
2558 	value = witness_watch;
2559 	error = sysctl_handle_int(oidp, &value, 0, req);
2560 	if (error != 0 || req->newptr == NULL)
2561 		return (error);
2562 	if (value > 1 || value < -1 ||
2563 	    (witness_watch == -1 && value != witness_watch))
2564 		return (EINVAL);
2565 	witness_watch = value;
2566 	return (0);
2567 }
2568 
2569 static void
2570 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2571 {
2572 	int i;
2573 
2574 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2575 		return;
2576 	w->w_displayed = 1;
2577 
2578 	WITNESS_INDEX_ASSERT(w->w_index);
2579 	for (i = 1; i <= w_max_used_index; i++) {
2580 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2581 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2582 			    w_data[i].w_name);
2583 			witness_add_fullgraph(sb, &w_data[i]);
2584 		}
2585 	}
2586 }
2587 
2588 /*
2589  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2590  * interprets the key as a string and reads until the null
2591  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2592  * hash value computed from the key.
2593  */
2594 static uint32_t
2595 witness_hash_djb2(const uint8_t *key, uint32_t size)
2596 {
2597 	unsigned int hash = 5381;
2598 	int i;
2599 
2600 	/* hash = hash * 33 + key[i] */
2601 	if (size)
2602 		for (i = 0; i < size; i++)
2603 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2604 	else
2605 		for (i = 0; key[i] != 0; i++)
2606 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2607 
2608 	return (hash);
2609 }
2610 
2611 
2612 /*
2613  * Initializes the two witness hash tables. Called exactly once from
2614  * witness_initialize().
2615  */
2616 static void
2617 witness_init_hash_tables(void)
2618 {
2619 	int i;
2620 
2621 	MPASS(witness_cold);
2622 
2623 	/* Initialize the hash tables. */
2624 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2625 		w_hash.wh_array[i] = NULL;
2626 
2627 	w_hash.wh_size = WITNESS_HASH_SIZE;
2628 	w_hash.wh_count = 0;
2629 
2630 	/* Initialize the lock order data hash. */
2631 	w_lofree = NULL;
2632 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2633 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2634 		w_lodata[i].wlod_next = w_lofree;
2635 		w_lofree = &w_lodata[i];
2636 	}
2637 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2638 	w_lohash.wloh_count = 0;
2639 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2640 		w_lohash.wloh_array[i] = NULL;
2641 }
2642 
2643 static struct witness *
2644 witness_hash_get(const char *key)
2645 {
2646 	struct witness *w;
2647 	uint32_t hash;
2648 
2649 	MPASS(key != NULL);
2650 	if (witness_cold == 0)
2651 		mtx_assert(&w_mtx, MA_OWNED);
2652 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2653 	w = w_hash.wh_array[hash];
2654 	while (w != NULL) {
2655 		if (strcmp(w->w_name, key) == 0)
2656 			goto out;
2657 		w = w->w_hash_next;
2658 	}
2659 
2660 out:
2661 	return (w);
2662 }
2663 
2664 static void
2665 witness_hash_put(struct witness *w)
2666 {
2667 	uint32_t hash;
2668 
2669 	MPASS(w != NULL);
2670 	MPASS(w->w_name != NULL);
2671 	if (witness_cold == 0)
2672 		mtx_assert(&w_mtx, MA_OWNED);
2673 	KASSERT(witness_hash_get(w->w_name) == NULL,
2674 	    ("%s: trying to add a hash entry that already exists!", __func__));
2675 	KASSERT(w->w_hash_next == NULL,
2676 	    ("%s: w->w_hash_next != NULL", __func__));
2677 
2678 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2679 	w->w_hash_next = w_hash.wh_array[hash];
2680 	w_hash.wh_array[hash] = w;
2681 	w_hash.wh_count++;
2682 }
2683 
2684 
2685 static struct witness_lock_order_data *
2686 witness_lock_order_get(struct witness *parent, struct witness *child)
2687 {
2688 	struct witness_lock_order_data *data = NULL;
2689 	struct witness_lock_order_key key;
2690 	unsigned int hash;
2691 
2692 	MPASS(parent != NULL && child != NULL);
2693 	key.from = parent->w_index;
2694 	key.to = child->w_index;
2695 	WITNESS_INDEX_ASSERT(key.from);
2696 	WITNESS_INDEX_ASSERT(key.to);
2697 	if ((w_rmatrix[parent->w_index][child->w_index]
2698 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2699 		goto out;
2700 
2701 	hash = witness_hash_djb2((const char*)&key,
2702 	    sizeof(key)) % w_lohash.wloh_size;
2703 	data = w_lohash.wloh_array[hash];
2704 	while (data != NULL) {
2705 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2706 			break;
2707 		data = data->wlod_next;
2708 	}
2709 
2710 out:
2711 	return (data);
2712 }
2713 
2714 /*
2715  * Verify that parent and child have a known relationship, are not the same,
2716  * and child is actually a child of parent.  This is done without w_mtx
2717  * to avoid contention in the common case.
2718  */
2719 static int
2720 witness_lock_order_check(struct witness *parent, struct witness *child)
2721 {
2722 
2723 	if (parent != child &&
2724 	    w_rmatrix[parent->w_index][child->w_index]
2725 	    & WITNESS_LOCK_ORDER_KNOWN &&
2726 	    isitmychild(parent, child))
2727 		return (1);
2728 
2729 	return (0);
2730 }
2731 
2732 static int
2733 witness_lock_order_add(struct witness *parent, struct witness *child)
2734 {
2735 	struct witness_lock_order_data *data = NULL;
2736 	struct witness_lock_order_key key;
2737 	unsigned int hash;
2738 
2739 	MPASS(parent != NULL && child != NULL);
2740 	key.from = parent->w_index;
2741 	key.to = child->w_index;
2742 	WITNESS_INDEX_ASSERT(key.from);
2743 	WITNESS_INDEX_ASSERT(key.to);
2744 	if (w_rmatrix[parent->w_index][child->w_index]
2745 	    & WITNESS_LOCK_ORDER_KNOWN)
2746 		return (1);
2747 
2748 	hash = witness_hash_djb2((const char*)&key,
2749 	    sizeof(key)) % w_lohash.wloh_size;
2750 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2751 	data = w_lofree;
2752 	if (data == NULL)
2753 		return (0);
2754 	w_lofree = data->wlod_next;
2755 	data->wlod_next = w_lohash.wloh_array[hash];
2756 	data->wlod_key = key;
2757 	w_lohash.wloh_array[hash] = data;
2758 	w_lohash.wloh_count++;
2759 	stack_zero(&data->wlod_stack);
2760 	stack_save(&data->wlod_stack);
2761 	return (1);
2762 }
2763 
2764 /* Call this whenver the structure of the witness graph changes. */
2765 static void
2766 witness_increment_graph_generation(void)
2767 {
2768 
2769 	if (witness_cold == 0)
2770 		mtx_assert(&w_mtx, MA_OWNED);
2771 	w_generation++;
2772 }
2773 
2774 #ifdef KDB
2775 static void
2776 _witness_debugger(int cond, const char *msg)
2777 {
2778 
2779 	if (witness_trace && cond)
2780 		kdb_backtrace();
2781 	if (witness_kdb && cond)
2782 		kdb_enter(KDB_WHY_WITNESS, msg);
2783 }
2784 #endif
2785