xref: /freebsd/sys/kern/subr_witness.c (revision 69c5bce6ee1ec42997757e7f1334767d217c5d7d)
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	768
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	FULLGRAPH_SBUF_SIZE	512
158 
159 /*
160  * These flags go in the witness relationship matrix and describe the
161  * relationship between any two struct witness objects.
162  */
163 #define	WITNESS_UNRELATED        0x00    /* No lock order relation. */
164 #define	WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
165 #define	WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
166 #define	WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
167 #define	WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
168 #define	WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define	WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define	WITNESS_RELATED_MASK						\
171 	(WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define	WITNESS_REVERSAL         0x10    /* A lock order reversal has been
173 					  * observed. */
174 #define	WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
175 #define	WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
176 #define	WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
177 
178 /* Descendant to ancestor flags */
179 #define	WITNESS_DTOA(x)	(((x) & WITNESS_RELATED_MASK) >> 2)
180 
181 /* Ancestor to descendant flags */
182 #define	WITNESS_ATOD(x)	(((x) & WITNESS_RELATED_MASK) << 2)
183 
184 #define	WITNESS_INDEX_ASSERT(i)						\
185 	MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
186 
187 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
188 
189 /*
190  * Lock instances.  A lock instance is the data associated with a lock while
191  * it is held by witness.  For example, a lock instance will hold the
192  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
193  * are held in a per-cpu list while sleep locks are held in per-thread list.
194  */
195 struct lock_instance {
196 	struct lock_object	*li_lock;
197 	const char		*li_file;
198 	int			li_line;
199 	u_int			li_flags;
200 };
201 
202 /*
203  * A simple list type used to build the list of locks held by a thread
204  * or CPU.  We can't simply embed the list in struct lock_object since a
205  * lock may be held by more than one thread if it is a shared lock.  Locks
206  * are added to the head of the list, so we fill up each list entry from
207  * "the back" logically.  To ease some of the arithmetic, we actually fill
208  * in each list entry the normal way (children[0] then children[1], etc.) but
209  * when we traverse the list we read children[count-1] as the first entry
210  * down to children[0] as the final entry.
211  */
212 struct lock_list_entry {
213 	struct lock_list_entry	*ll_next;
214 	struct lock_instance	ll_children[LOCK_NCHILDREN];
215 	u_int			ll_count;
216 };
217 
218 /*
219  * The main witness structure. One of these per named lock type in the system
220  * (for example, "vnode interlock").
221  */
222 struct witness {
223 	char  			w_name[MAX_W_NAME];
224 	uint32_t 		w_index;  /* Index in the relationship matrix */
225 	struct lock_class	*w_class;
226 	STAILQ_ENTRY(witness) 	w_list;		/* List of all witnesses. */
227 	STAILQ_ENTRY(witness) 	w_typelist;	/* Witnesses of a type. */
228 	struct witness		*w_hash_next; /* Linked list in hash buckets. */
229 	const char		*w_file; /* File where last acquired */
230 	uint32_t 		w_line; /* Line where last acquired */
231 	uint32_t 		w_refcount;
232 	uint16_t 		w_num_ancestors; /* direct/indirect
233 						  * ancestor count */
234 	uint16_t 		w_num_descendants; /* direct/indirect
235 						    * descendant count */
236 	int16_t 		w_ddb_level;
237 	unsigned		w_displayed:1;
238 	unsigned		w_reversed:1;
239 };
240 
241 STAILQ_HEAD(witness_list, witness);
242 
243 /*
244  * The witness hash table. Keys are witness names (const char *), elements are
245  * witness objects (struct witness *).
246  */
247 struct witness_hash {
248 	struct witness	*wh_array[WITNESS_HASH_SIZE];
249 	uint32_t	wh_size;
250 	uint32_t	wh_count;
251 };
252 
253 /*
254  * Key type for the lock order data hash table.
255  */
256 struct witness_lock_order_key {
257 	uint16_t	from;
258 	uint16_t	to;
259 };
260 
261 struct witness_lock_order_data {
262 	struct stack			wlod_stack;
263 	struct witness_lock_order_key	wlod_key;
264 	struct witness_lock_order_data	*wlod_next;
265 };
266 
267 /*
268  * The witness lock order data hash table. Keys are witness index tuples
269  * (struct witness_lock_order_key), elements are lock order data objects
270  * (struct witness_lock_order_data).
271  */
272 struct witness_lock_order_hash {
273 	struct witness_lock_order_data	*wloh_array[WITNESS_LO_HASH_SIZE];
274 	u_int	wloh_size;
275 	u_int	wloh_count;
276 };
277 
278 #ifdef BLESSING
279 struct witness_blessed {
280 	const char	*b_lock1;
281 	const char	*b_lock2;
282 };
283 #endif
284 
285 struct witness_pendhelp {
286 	const char		*wh_type;
287 	struct lock_object	*wh_lock;
288 };
289 
290 struct witness_order_list_entry {
291 	const char		*w_name;
292 	struct lock_class	*w_class;
293 };
294 
295 /*
296  * Returns 0 if one of the locks is a spin lock and the other is not.
297  * Returns 1 otherwise.
298  */
299 static __inline int
300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
301 {
302 
303 	return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
304 		(w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
305 }
306 
307 static __inline int
308 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
309 {
310 
311 	return (key->from == 0 && key->to == 0);
312 }
313 
314 static __inline int
315 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
316     const struct witness_lock_order_key *b)
317 {
318 
319 	return (a->from == b->from && a->to == b->to);
320 }
321 
322 static int	_isitmyx(struct witness *w1, struct witness *w2, int rmask,
323 		    const char *fname);
324 #ifdef KDB
325 static void	_witness_debugger(int cond, const char *msg);
326 #endif
327 static void	adopt(struct witness *parent, struct witness *child);
328 #ifdef BLESSING
329 static int	blessed(struct witness *, struct witness *);
330 #endif
331 static void	depart(struct witness *w);
332 static struct witness	*enroll(const char *description,
333 			    struct lock_class *lock_class);
334 static struct lock_instance	*find_instance(struct lock_list_entry *list,
335 				    struct lock_object *lock);
336 static int	isitmychild(struct witness *parent, struct witness *child);
337 static int	isitmydescendant(struct witness *parent, struct witness *child);
338 static void	itismychild(struct witness *parent, struct witness *child);
339 static int	sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
340 static int	sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
341 static int	sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
342 static void	witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
343 #ifdef DDB
344 static void	witness_ddb_compute_levels(void);
345 static void	witness_ddb_display(int(*)(const char *fmt, ...));
346 static void	witness_ddb_display_descendants(int(*)(const char *fmt, ...),
347 		    struct witness *, int indent);
348 static void	witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
349 		    struct witness_list *list);
350 static void	witness_ddb_level_descendants(struct witness *parent, int l);
351 static void	witness_ddb_list(struct thread *td);
352 #endif
353 static void	witness_free(struct witness *m);
354 static struct witness	*witness_get(void);
355 static uint32_t	witness_hash_djb2(const uint8_t *key, uint32_t size);
356 static struct witness	*witness_hash_get(const char *key);
357 static void	witness_hash_put(struct witness *w);
358 static void	witness_init_hash_tables(void);
359 static void	witness_increment_graph_generation(void);
360 static void	witness_lock_list_free(struct lock_list_entry *lle);
361 static struct lock_list_entry	*witness_lock_list_get(void);
362 static int	witness_lock_order_add(struct witness *parent,
363 		    struct witness *child);
364 static int	witness_lock_order_check(struct witness *parent,
365 		    struct witness *child);
366 static struct witness_lock_order_data	*witness_lock_order_get(
367 					    struct witness *parent,
368 					    struct witness *child);
369 static void	witness_list_lock(struct lock_instance *instance,
370 		    int (*prnt)(const char *fmt, ...));
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 	{ "time lock", &lock_class_mtx_sleep },
498 	{ NULL, NULL },
499 	/*
500 	 * Sockets
501 	 */
502 	{ "accept", &lock_class_mtx_sleep },
503 	{ "so_snd", &lock_class_mtx_sleep },
504 	{ "so_rcv", &lock_class_mtx_sleep },
505 	{ "sellck", &lock_class_mtx_sleep },
506 	{ NULL, NULL },
507 	/*
508 	 * Routing
509 	 */
510 	{ "so_rcv", &lock_class_mtx_sleep },
511 	{ "radix node head", &lock_class_rw },
512 	{ "rtentry", &lock_class_mtx_sleep },
513 	{ "ifaddr", &lock_class_mtx_sleep },
514 	{ NULL, NULL },
515 	/*
516 	 * IPv4 multicast:
517 	 * protocol locks before interface locks, after UDP locks.
518 	 */
519 	{ "udpinp", &lock_class_rw },
520 	{ "in_multi_mtx", &lock_class_mtx_sleep },
521 	{ "igmp_mtx", &lock_class_mtx_sleep },
522 	{ "if_addr_mtx", &lock_class_mtx_sleep },
523 	{ NULL, NULL },
524 	/*
525 	 * IPv6 multicast:
526 	 * protocol locks before interface locks, after UDP locks.
527 	 */
528 	{ "udpinp", &lock_class_rw },
529 	{ "in6_multi_mtx", &lock_class_mtx_sleep },
530 	{ "mld_mtx", &lock_class_mtx_sleep },
531 	{ "if_addr_mtx", &lock_class_mtx_sleep },
532 	{ NULL, NULL },
533 	/*
534 	 * UNIX Domain Sockets
535 	 */
536 	{ "unp_global_rwlock", &lock_class_rw },
537 	{ "unp_list_lock", &lock_class_mtx_sleep },
538 	{ "unp", &lock_class_mtx_sleep },
539 	{ "so_snd", &lock_class_mtx_sleep },
540 	{ NULL, NULL },
541 	/*
542 	 * UDP/IP
543 	 */
544 	{ "udp", &lock_class_rw },
545 	{ "udpinp", &lock_class_rw },
546 	{ "so_snd", &lock_class_mtx_sleep },
547 	{ NULL, NULL },
548 	/*
549 	 * TCP/IP
550 	 */
551 	{ "tcp", &lock_class_rw },
552 	{ "tcpinp", &lock_class_rw },
553 	{ "so_snd", &lock_class_mtx_sleep },
554 	{ NULL, NULL },
555 	/*
556 	 * netatalk
557 	 */
558 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
559 	{ "ddp_mtx", &lock_class_mtx_sleep },
560 	{ NULL, NULL },
561 	/*
562 	 * BPF
563 	 */
564 	{ "bpf global lock", &lock_class_mtx_sleep },
565 	{ "bpf interface lock", &lock_class_mtx_sleep },
566 	{ "bpf cdev lock", &lock_class_mtx_sleep },
567 	{ NULL, NULL },
568 	/*
569 	 * NFS server
570 	 */
571 	{ "nfsd_mtx", &lock_class_mtx_sleep },
572 	{ "so_snd", &lock_class_mtx_sleep },
573 	{ NULL, NULL },
574 
575 	/*
576 	 * IEEE 802.11
577 	 */
578 	{ "802.11 com lock", &lock_class_mtx_sleep},
579 	{ NULL, NULL },
580 	/*
581 	 * Network drivers
582 	 */
583 	{ "network driver", &lock_class_mtx_sleep},
584 	{ NULL, NULL },
585 
586 	/*
587 	 * Netgraph
588 	 */
589 	{ "ng_node", &lock_class_mtx_sleep },
590 	{ "ng_worklist", &lock_class_mtx_sleep },
591 	{ NULL, NULL },
592 	/*
593 	 * CDEV
594 	 */
595 	{ "system map", &lock_class_mtx_sleep },
596 	{ "vm page queue mutex", &lock_class_mtx_sleep },
597 	{ "vnode interlock", &lock_class_mtx_sleep },
598 	{ "cdev", &lock_class_mtx_sleep },
599 	{ NULL, NULL },
600 	/*
601 	 * VM
602 	 *
603 	 */
604 	{ "vm object", &lock_class_mtx_sleep },
605 	{ "page lock", &lock_class_mtx_sleep },
606 	{ "vm page queue mutex", &lock_class_mtx_sleep },
607 	{ "pmap", &lock_class_mtx_sleep },
608 	{ NULL, NULL },
609 	/*
610 	 * kqueue/VFS interaction
611 	 */
612 	{ "kqueue", &lock_class_mtx_sleep },
613 	{ "struct mount mtx", &lock_class_mtx_sleep },
614 	{ "vnode interlock", &lock_class_mtx_sleep },
615 	{ NULL, NULL },
616 	/*
617 	 * ZFS locking
618 	 */
619 	{ "dn->dn_mtx", &lock_class_sx },
620 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
621 	{ "db->db_mtx", &lock_class_sx },
622 	{ NULL, NULL },
623 	/*
624 	 * spin locks
625 	 */
626 #ifdef SMP
627 	{ "ap boot", &lock_class_mtx_spin },
628 #endif
629 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
630 	{ "sio", &lock_class_mtx_spin },
631 	{ "scrlock", &lock_class_mtx_spin },
632 #ifdef __i386__
633 	{ "cy", &lock_class_mtx_spin },
634 #endif
635 #ifdef __sparc64__
636 	{ "pcib_mtx", &lock_class_mtx_spin },
637 	{ "rtc_mtx", &lock_class_mtx_spin },
638 #endif
639 	{ "scc_hwmtx", &lock_class_mtx_spin },
640 	{ "uart_hwmtx", &lock_class_mtx_spin },
641 	{ "fast_taskqueue", &lock_class_mtx_spin },
642 	{ "intr table", &lock_class_mtx_spin },
643 #ifdef	HWPMC_HOOKS
644 	{ "pmc-per-proc", &lock_class_mtx_spin },
645 #endif
646 	{ "process slock", &lock_class_mtx_spin },
647 	{ "sleepq chain", &lock_class_mtx_spin },
648 	{ "umtx lock", &lock_class_mtx_spin },
649 	{ "rm_spinlock", &lock_class_mtx_spin },
650 	{ "turnstile chain", &lock_class_mtx_spin },
651 	{ "turnstile lock", &lock_class_mtx_spin },
652 	{ "sched lock", &lock_class_mtx_spin },
653 	{ "td_contested", &lock_class_mtx_spin },
654 	{ "callout", &lock_class_mtx_spin },
655 	{ "entropy harvest mutex", &lock_class_mtx_spin },
656 	{ "syscons video lock", &lock_class_mtx_spin },
657 #ifdef SMP
658 	{ "smp rendezvous", &lock_class_mtx_spin },
659 #endif
660 #ifdef __powerpc__
661 	{ "tlb0", &lock_class_mtx_spin },
662 #endif
663 	/*
664 	 * leaf locks
665 	 */
666 	{ "intrcnt", &lock_class_mtx_spin },
667 	{ "icu", &lock_class_mtx_spin },
668 #if defined(SMP) && defined(__sparc64__)
669 	{ "ipi", &lock_class_mtx_spin },
670 #endif
671 #ifdef __i386__
672 	{ "allpmaps", &lock_class_mtx_spin },
673 	{ "descriptor tables", &lock_class_mtx_spin },
674 #endif
675 	{ "clk", &lock_class_mtx_spin },
676 	{ "cpuset", &lock_class_mtx_spin },
677 	{ "mprof lock", &lock_class_mtx_spin },
678 	{ "zombie lock", &lock_class_mtx_spin },
679 	{ "ALD Queue", &lock_class_mtx_spin },
680 #ifdef __ia64__
681 	{ "MCA spin lock", &lock_class_mtx_spin },
682 #endif
683 #if defined(__i386__) || defined(__amd64__)
684 	{ "pcicfg", &lock_class_mtx_spin },
685 	{ "NDIS thread lock", &lock_class_mtx_spin },
686 #endif
687 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
688 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
689 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
690 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
691 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
692 #ifdef	HWPMC_HOOKS
693 	{ "pmc-leaf", &lock_class_mtx_spin },
694 #endif
695 	{ "blocked lock", &lock_class_mtx_spin },
696 	{ NULL, NULL },
697 	{ NULL, NULL }
698 };
699 
700 #ifdef BLESSING
701 /*
702  * Pairs of locks which have been blessed
703  * Don't complain about order problems with blessed locks
704  */
705 static struct witness_blessed blessed_list[] = {
706 };
707 static int blessed_count =
708 	sizeof(blessed_list) / sizeof(struct witness_blessed);
709 #endif
710 
711 /*
712  * This global is set to 0 once it becomes safe to use the witness code.
713  */
714 static int witness_cold = 1;
715 
716 /*
717  * This global is set to 1 once the static lock orders have been enrolled
718  * so that a warning can be issued for any spin locks enrolled later.
719  */
720 static int witness_spin_warn = 0;
721 
722 /* Trim useless garbage from filenames. */
723 static const char *
724 fixup_filename(const char *file)
725 {
726 
727 	if (file == NULL)
728 		return (NULL);
729 	while (strncmp(file, "../", 3) == 0)
730 		file += 3;
731 	return (file);
732 }
733 
734 /*
735  * The WITNESS-enabled diagnostic code.  Note that the witness code does
736  * assume that the early boot is single-threaded at least until after this
737  * routine is completed.
738  */
739 static void
740 witness_initialize(void *dummy __unused)
741 {
742 	struct lock_object *lock;
743 	struct witness_order_list_entry *order;
744 	struct witness *w, *w1;
745 	int i;
746 
747 	w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
748 	    M_NOWAIT | M_ZERO);
749 
750 	/*
751 	 * We have to release Giant before initializing its witness
752 	 * structure so that WITNESS doesn't get confused.
753 	 */
754 	mtx_unlock(&Giant);
755 	mtx_assert(&Giant, MA_NOTOWNED);
756 
757 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
758 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
759 	    MTX_NOWITNESS | MTX_NOPROFILE);
760 	for (i = WITNESS_COUNT - 1; i >= 0; i--) {
761 		w = &w_data[i];
762 		memset(w, 0, sizeof(*w));
763 		w_data[i].w_index = i;	/* Witness index never changes. */
764 		witness_free(w);
765 	}
766 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
767 	    ("%s: Invalid list of free witness objects", __func__));
768 
769 	/* Witness with index 0 is not used to aid in debugging. */
770 	STAILQ_REMOVE_HEAD(&w_free, w_list);
771 	w_free_cnt--;
772 
773 	memset(w_rmatrix, 0,
774 	    (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
775 
776 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
777 		witness_lock_list_free(&w_locklistdata[i]);
778 	witness_init_hash_tables();
779 
780 	/* First add in all the specified order lists. */
781 	for (order = order_lists; order->w_name != NULL; order++) {
782 		w = enroll(order->w_name, order->w_class);
783 		if (w == NULL)
784 			continue;
785 		w->w_file = "order list";
786 		for (order++; order->w_name != NULL; order++) {
787 			w1 = enroll(order->w_name, order->w_class);
788 			if (w1 == NULL)
789 				continue;
790 			w1->w_file = "order list";
791 			itismychild(w, w1);
792 			w = w1;
793 		}
794 	}
795 	witness_spin_warn = 1;
796 
797 	/* Iterate through all locks and add them to witness. */
798 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
799 		lock = pending_locks[i].wh_lock;
800 		KASSERT(lock->lo_flags & LO_WITNESS,
801 		    ("%s: lock %s is on pending list but not LO_WITNESS",
802 		    __func__, lock->lo_name));
803 		lock->lo_witness = enroll(pending_locks[i].wh_type,
804 		    LOCK_CLASS(lock));
805 	}
806 
807 	/* Mark the witness code as being ready for use. */
808 	witness_cold = 0;
809 
810 	mtx_lock(&Giant);
811 }
812 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
813     NULL);
814 
815 void
816 witness_init(struct lock_object *lock, const char *type)
817 {
818 	struct lock_class *class;
819 
820 	/* Various sanity checks. */
821 	class = LOCK_CLASS(lock);
822 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
823 	    (class->lc_flags & LC_RECURSABLE) == 0)
824 		panic("%s: lock (%s) %s can not be recursable", __func__,
825 		    class->lc_name, lock->lo_name);
826 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
827 	    (class->lc_flags & LC_SLEEPABLE) == 0)
828 		panic("%s: lock (%s) %s can not be sleepable", __func__,
829 		    class->lc_name, lock->lo_name);
830 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
831 	    (class->lc_flags & LC_UPGRADABLE) == 0)
832 		panic("%s: lock (%s) %s can not be upgradable", __func__,
833 		    class->lc_name, lock->lo_name);
834 
835 	/*
836 	 * If we shouldn't watch this lock, then just clear lo_witness.
837 	 * Otherwise, if witness_cold is set, then it is too early to
838 	 * enroll this lock, so defer it to witness_initialize() by adding
839 	 * it to the pending_locks list.  If it is not too early, then enroll
840 	 * the lock now.
841 	 */
842 	if (witness_watch < 1 || panicstr != NULL ||
843 	    (lock->lo_flags & LO_WITNESS) == 0)
844 		lock->lo_witness = NULL;
845 	else if (witness_cold) {
846 		pending_locks[pending_cnt].wh_lock = lock;
847 		pending_locks[pending_cnt++].wh_type = type;
848 		if (pending_cnt > WITNESS_PENDLIST)
849 			panic("%s: pending locks list is too small, bump it\n",
850 			    __func__);
851 	} else
852 		lock->lo_witness = enroll(type, class);
853 }
854 
855 void
856 witness_destroy(struct lock_object *lock)
857 {
858 	struct lock_class *class;
859 	struct witness *w;
860 
861 	class = LOCK_CLASS(lock);
862 
863 	if (witness_cold)
864 		panic("lock (%s) %s destroyed while witness_cold",
865 		    class->lc_name, lock->lo_name);
866 
867 	/* XXX: need to verify that no one holds the lock */
868 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
869 		return;
870 	w = lock->lo_witness;
871 
872 	mtx_lock_spin(&w_mtx);
873 	MPASS(w->w_refcount > 0);
874 	w->w_refcount--;
875 
876 	if (w->w_refcount == 0)
877 		depart(w);
878 	mtx_unlock_spin(&w_mtx);
879 }
880 
881 #ifdef DDB
882 static void
883 witness_ddb_compute_levels(void)
884 {
885 	struct witness *w;
886 
887 	/*
888 	 * First clear all levels.
889 	 */
890 	STAILQ_FOREACH(w, &w_all, w_list)
891 		w->w_ddb_level = -1;
892 
893 	/*
894 	 * Look for locks with no parents and level all their descendants.
895 	 */
896 	STAILQ_FOREACH(w, &w_all, w_list) {
897 
898 		/* If the witness has ancestors (is not a root), skip it. */
899 		if (w->w_num_ancestors > 0)
900 			continue;
901 		witness_ddb_level_descendants(w, 0);
902 	}
903 }
904 
905 static void
906 witness_ddb_level_descendants(struct witness *w, int l)
907 {
908 	int i;
909 
910 	if (w->w_ddb_level >= l)
911 		return;
912 
913 	w->w_ddb_level = l;
914 	l++;
915 
916 	for (i = 1; i <= w_max_used_index; i++) {
917 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
918 			witness_ddb_level_descendants(&w_data[i], l);
919 	}
920 }
921 
922 static void
923 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
924     struct witness *w, int indent)
925 {
926 	int i;
927 
928  	for (i = 0; i < indent; i++)
929  		prnt(" ");
930 	prnt("%s (type: %s, depth: %d, active refs: %d)",
931 	     w->w_name, w->w_class->lc_name,
932 	     w->w_ddb_level, w->w_refcount);
933  	if (w->w_displayed) {
934  		prnt(" -- (already displayed)\n");
935  		return;
936  	}
937  	w->w_displayed = 1;
938 	if (w->w_file != NULL && w->w_line != 0)
939 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
940 		    w->w_line);
941 	else
942 		prnt(" -- never acquired\n");
943 	indent++;
944 	WITNESS_INDEX_ASSERT(w->w_index);
945 	for (i = 1; i <= w_max_used_index; i++) {
946 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
947 			witness_ddb_display_descendants(prnt, &w_data[i],
948 			    indent);
949 	}
950 }
951 
952 static void
953 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
954     struct witness_list *list)
955 {
956 	struct witness *w;
957 
958 	STAILQ_FOREACH(w, list, w_typelist) {
959 		if (w->w_file == NULL || w->w_ddb_level > 0)
960 			continue;
961 
962 		/* This lock has no anscestors - display its descendants. */
963 		witness_ddb_display_descendants(prnt, w, 0);
964 	}
965 }
966 
967 static void
968 witness_ddb_display(int(*prnt)(const char *fmt, ...))
969 {
970 	struct witness *w;
971 
972 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
973 	witness_ddb_compute_levels();
974 
975 	/* Clear all the displayed flags. */
976 	STAILQ_FOREACH(w, &w_all, w_list)
977 		w->w_displayed = 0;
978 
979 	/*
980 	 * First, handle sleep locks which have been acquired at least
981 	 * once.
982 	 */
983 	prnt("Sleep locks:\n");
984 	witness_ddb_display_list(prnt, &w_sleep);
985 
986 	/*
987 	 * Now do spin locks which have been acquired at least once.
988 	 */
989 	prnt("\nSpin locks:\n");
990 	witness_ddb_display_list(prnt, &w_spin);
991 
992 	/*
993 	 * Finally, any locks which have not been acquired yet.
994 	 */
995 	prnt("\nLocks which were never acquired:\n");
996 	STAILQ_FOREACH(w, &w_all, w_list) {
997 		if (w->w_file != NULL || w->w_refcount == 0)
998 			continue;
999 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1000 		    w->w_class->lc_name, w->w_ddb_level);
1001 	}
1002 }
1003 #endif /* DDB */
1004 
1005 int
1006 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1007 {
1008 
1009 	if (witness_watch == -1 || panicstr != NULL)
1010 		return (0);
1011 
1012 	/* Require locks that witness knows about. */
1013 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1014 	    lock2->lo_witness == NULL)
1015 		return (EINVAL);
1016 
1017 	mtx_assert(&w_mtx, MA_NOTOWNED);
1018 	mtx_lock_spin(&w_mtx);
1019 
1020 	/*
1021 	 * If we already have either an explicit or implied lock order that
1022 	 * is the other way around, then return an error.
1023 	 */
1024 	if (witness_watch &&
1025 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1026 		mtx_unlock_spin(&w_mtx);
1027 		return (EDOOFUS);
1028 	}
1029 
1030 	/* Try to add the new order. */
1031 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1032 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1033 	itismychild(lock1->lo_witness, lock2->lo_witness);
1034 	mtx_unlock_spin(&w_mtx);
1035 	return (0);
1036 }
1037 
1038 void
1039 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1040     int line, struct lock_object *interlock)
1041 {
1042 	struct lock_list_entry *lock_list, *lle;
1043 	struct lock_instance *lock1, *lock2, *plock;
1044 	struct lock_class *class;
1045 	struct witness *w, *w1;
1046 	struct thread *td;
1047 	int i, j;
1048 
1049 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1050 	    panicstr != NULL)
1051 		return;
1052 
1053 	w = lock->lo_witness;
1054 	class = LOCK_CLASS(lock);
1055 	td = curthread;
1056 
1057 	if (class->lc_flags & LC_SLEEPLOCK) {
1058 
1059 		/*
1060 		 * Since spin locks include a critical section, this check
1061 		 * implicitly enforces a lock order of all sleep locks before
1062 		 * all spin locks.
1063 		 */
1064 		if (td->td_critnest != 0 && !kdb_active)
1065 			panic("blockable sleep lock (%s) %s @ %s:%d",
1066 			    class->lc_name, lock->lo_name,
1067 			    fixup_filename(file), line);
1068 
1069 		/*
1070 		 * If this is the first lock acquired then just return as
1071 		 * no order checking is needed.
1072 		 */
1073 		lock_list = td->td_sleeplocks;
1074 		if (lock_list == NULL || lock_list->ll_count == 0)
1075 			return;
1076 	} else {
1077 
1078 		/*
1079 		 * If this is the first lock, just return as no order
1080 		 * checking is needed.  Avoid problems with thread
1081 		 * migration pinning the thread while checking if
1082 		 * spinlocks are held.  If at least one spinlock is held
1083 		 * the thread is in a safe path and it is allowed to
1084 		 * unpin it.
1085 		 */
1086 		sched_pin();
1087 		lock_list = PCPU_GET(spinlocks);
1088 		if (lock_list == NULL || lock_list->ll_count == 0) {
1089 			sched_unpin();
1090 			return;
1091 		}
1092 		sched_unpin();
1093 	}
1094 
1095 	/*
1096 	 * Check to see if we are recursing on a lock we already own.  If
1097 	 * so, make sure that we don't mismatch exclusive and shared lock
1098 	 * acquires.
1099 	 */
1100 	lock1 = find_instance(lock_list, lock);
1101 	if (lock1 != NULL) {
1102 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1103 		    (flags & LOP_EXCLUSIVE) == 0) {
1104 			printf("shared lock of (%s) %s @ %s:%d\n",
1105 			    class->lc_name, lock->lo_name,
1106 			    fixup_filename(file), line);
1107 			printf("while exclusively locked from %s:%d\n",
1108 			    fixup_filename(lock1->li_file), lock1->li_line);
1109 			panic("share->excl");
1110 		}
1111 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1112 		    (flags & LOP_EXCLUSIVE) != 0) {
1113 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1114 			    class->lc_name, lock->lo_name,
1115 			    fixup_filename(file), line);
1116 			printf("while share locked from %s:%d\n",
1117 			    fixup_filename(lock1->li_file), lock1->li_line);
1118 			panic("excl->share");
1119 		}
1120 		return;
1121 	}
1122 
1123 	/*
1124 	 * Find the previously acquired lock, but ignore interlocks.
1125 	 */
1126 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1127 	if (interlock != NULL && plock->li_lock == interlock) {
1128 		if (lock_list->ll_count > 1)
1129 			plock =
1130 			    &lock_list->ll_children[lock_list->ll_count - 2];
1131 		else {
1132 			lle = lock_list->ll_next;
1133 
1134 			/*
1135 			 * The interlock is the only lock we hold, so
1136 			 * simply return.
1137 			 */
1138 			if (lle == NULL)
1139 				return;
1140 			plock = &lle->ll_children[lle->ll_count - 1];
1141 		}
1142 	}
1143 
1144 	/*
1145 	 * Try to perform most checks without a lock.  If this succeeds we
1146 	 * can skip acquiring the lock and return success.
1147 	 */
1148 	w1 = plock->li_lock->lo_witness;
1149 	if (witness_lock_order_check(w1, w))
1150 		return;
1151 
1152 	/*
1153 	 * Check for duplicate locks of the same type.  Note that we only
1154 	 * have to check for this on the last lock we just acquired.  Any
1155 	 * other cases will be caught as lock order violations.
1156 	 */
1157 	mtx_lock_spin(&w_mtx);
1158 	witness_lock_order_add(w1, w);
1159 	if (w1 == w) {
1160 		i = w->w_index;
1161 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1162 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1163 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1164 			w->w_reversed = 1;
1165 			mtx_unlock_spin(&w_mtx);
1166 			printf(
1167 			    "acquiring duplicate lock of same type: \"%s\"\n",
1168 			    w->w_name);
1169 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1170 			    fixup_filename(plock->li_file), plock->li_line);
1171 			printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1172 			    fixup_filename(file), line);
1173 			witness_debugger(1);
1174 		} else
1175 			mtx_unlock_spin(&w_mtx);
1176 		return;
1177 	}
1178 	mtx_assert(&w_mtx, MA_OWNED);
1179 
1180 	/*
1181 	 * If we know that the lock we are acquiring comes after
1182 	 * the lock we most recently acquired in the lock order tree,
1183 	 * then there is no need for any further checks.
1184 	 */
1185 	if (isitmychild(w1, w))
1186 		goto out;
1187 
1188 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1189 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1190 
1191 			MPASS(j < WITNESS_COUNT);
1192 			lock1 = &lle->ll_children[i];
1193 
1194 			/*
1195 			 * Ignore the interlock the first time we see it.
1196 			 */
1197 			if (interlock != NULL && interlock == lock1->li_lock) {
1198 				interlock = NULL;
1199 				continue;
1200 			}
1201 
1202 			/*
1203 			 * If this lock doesn't undergo witness checking,
1204 			 * then skip it.
1205 			 */
1206 			w1 = lock1->li_lock->lo_witness;
1207 			if (w1 == NULL) {
1208 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1209 				    ("lock missing witness structure"));
1210 				continue;
1211 			}
1212 
1213 			/*
1214 			 * If we are locking Giant and this is a sleepable
1215 			 * lock, then skip it.
1216 			 */
1217 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1218 			    lock == &Giant.lock_object)
1219 				continue;
1220 
1221 			/*
1222 			 * If we are locking a sleepable lock and this lock
1223 			 * is Giant, then skip it.
1224 			 */
1225 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1226 			    lock1->li_lock == &Giant.lock_object)
1227 				continue;
1228 
1229 			/*
1230 			 * If we are locking a sleepable lock and this lock
1231 			 * isn't sleepable, we want to treat it as a lock
1232 			 * order violation to enfore a general lock order of
1233 			 * sleepable locks before non-sleepable locks.
1234 			 */
1235 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1236 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1237 				goto reversal;
1238 
1239 			/*
1240 			 * If we are locking Giant and this is a non-sleepable
1241 			 * lock, then treat it as a reversal.
1242 			 */
1243 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1244 			    lock == &Giant.lock_object)
1245 				goto reversal;
1246 
1247 			/*
1248 			 * Check the lock order hierarchy for a reveresal.
1249 			 */
1250 			if (!isitmydescendant(w, w1))
1251 				continue;
1252 		reversal:
1253 
1254 			/*
1255 			 * We have a lock order violation, check to see if it
1256 			 * is allowed or has already been yelled about.
1257 			 */
1258 #ifdef BLESSING
1259 
1260 			/*
1261 			 * If the lock order is blessed, just bail.  We don't
1262 			 * look for other lock order violations though, which
1263 			 * may be a bug.
1264 			 */
1265 			if (blessed(w, w1))
1266 				goto out;
1267 #endif
1268 
1269 			/* Bail if this violation is known */
1270 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1271 				goto out;
1272 
1273 			/* Record this as a violation */
1274 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1275 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1276 			w->w_reversed = w1->w_reversed = 1;
1277 			witness_increment_graph_generation();
1278 			mtx_unlock_spin(&w_mtx);
1279 
1280 			/*
1281 			 * Ok, yell about it.
1282 			 */
1283 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1284 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1285 				printf(
1286 		"lock order reversal: (sleepable after non-sleepable)\n");
1287 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1288 			    && lock == &Giant.lock_object)
1289 				printf(
1290 		"lock order reversal: (Giant after non-sleepable)\n");
1291 			else
1292 				printf("lock order reversal:\n");
1293 
1294 			/*
1295 			 * Try to locate an earlier lock with
1296 			 * witness w in our list.
1297 			 */
1298 			do {
1299 				lock2 = &lle->ll_children[i];
1300 				MPASS(lock2->li_lock != NULL);
1301 				if (lock2->li_lock->lo_witness == w)
1302 					break;
1303 				if (i == 0 && lle->ll_next != NULL) {
1304 					lle = lle->ll_next;
1305 					i = lle->ll_count - 1;
1306 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1307 				} else
1308 					i--;
1309 			} while (i >= 0);
1310 			if (i < 0) {
1311 				printf(" 1st %p %s (%s) @ %s:%d\n",
1312 				    lock1->li_lock, lock1->li_lock->lo_name,
1313 				    w1->w_name, fixup_filename(lock1->li_file),
1314 				    lock1->li_line);
1315 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1316 				    lock->lo_name, w->w_name,
1317 				    fixup_filename(file), line);
1318 			} else {
1319 				printf(" 1st %p %s (%s) @ %s:%d\n",
1320 				    lock2->li_lock, lock2->li_lock->lo_name,
1321 				    lock2->li_lock->lo_witness->w_name,
1322 				    fixup_filename(lock2->li_file),
1323 				    lock2->li_line);
1324 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1325 				    lock1->li_lock, lock1->li_lock->lo_name,
1326 				    w1->w_name, fixup_filename(lock1->li_file),
1327 				    lock1->li_line);
1328 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1329 				    lock->lo_name, w->w_name,
1330 				    fixup_filename(file), line);
1331 			}
1332 			witness_debugger(1);
1333 			return;
1334 		}
1335 	}
1336 
1337 	/*
1338 	 * If requested, build a new lock order.  However, don't build a new
1339 	 * relationship between a sleepable lock and Giant if it is in the
1340 	 * wrong direction.  The correct lock order is that sleepable locks
1341 	 * always come before Giant.
1342 	 */
1343 	if (flags & LOP_NEWORDER &&
1344 	    !(plock->li_lock == &Giant.lock_object &&
1345 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1346 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1347 		    w->w_name, plock->li_lock->lo_witness->w_name);
1348 		itismychild(plock->li_lock->lo_witness, w);
1349 	}
1350 out:
1351 	mtx_unlock_spin(&w_mtx);
1352 }
1353 
1354 void
1355 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1356 {
1357 	struct lock_list_entry **lock_list, *lle;
1358 	struct lock_instance *instance;
1359 	struct witness *w;
1360 	struct thread *td;
1361 
1362 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1363 	    panicstr != NULL)
1364 		return;
1365 	w = lock->lo_witness;
1366 	td = curthread;
1367 
1368 	/* Determine lock list for this lock. */
1369 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1370 		lock_list = &td->td_sleeplocks;
1371 	else
1372 		lock_list = PCPU_PTR(spinlocks);
1373 
1374 	/* Check to see if we are recursing on a lock we already own. */
1375 	instance = find_instance(*lock_list, lock);
1376 	if (instance != NULL) {
1377 		instance->li_flags++;
1378 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1379 		    td->td_proc->p_pid, lock->lo_name,
1380 		    instance->li_flags & LI_RECURSEMASK);
1381 		instance->li_file = file;
1382 		instance->li_line = line;
1383 		return;
1384 	}
1385 
1386 	/* Update per-witness last file and line acquire. */
1387 	w->w_file = file;
1388 	w->w_line = line;
1389 
1390 	/* Find the next open lock instance in the list and fill it. */
1391 	lle = *lock_list;
1392 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1393 		lle = witness_lock_list_get();
1394 		if (lle == NULL)
1395 			return;
1396 		lle->ll_next = *lock_list;
1397 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1398 		    td->td_proc->p_pid, lle);
1399 		*lock_list = lle;
1400 	}
1401 	instance = &lle->ll_children[lle->ll_count++];
1402 	instance->li_lock = lock;
1403 	instance->li_line = line;
1404 	instance->li_file = file;
1405 	if ((flags & LOP_EXCLUSIVE) != 0)
1406 		instance->li_flags = LI_EXCLUSIVE;
1407 	else
1408 		instance->li_flags = 0;
1409 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1410 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1411 }
1412 
1413 void
1414 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1415 {
1416 	struct lock_instance *instance;
1417 	struct lock_class *class;
1418 
1419 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1420 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1421 		return;
1422 	class = LOCK_CLASS(lock);
1423 	if (witness_watch) {
1424 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1425 			panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1426 			    class->lc_name, lock->lo_name,
1427 			    fixup_filename(file), line);
1428 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1429 			panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1430 			    class->lc_name, lock->lo_name,
1431 			    fixup_filename(file), line);
1432 	}
1433 	instance = find_instance(curthread->td_sleeplocks, lock);
1434 	if (instance == NULL)
1435 		panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1436 		    class->lc_name, lock->lo_name,
1437 		    fixup_filename(file), line);
1438 	if (witness_watch) {
1439 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1440 			panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1441 			    class->lc_name, lock->lo_name,
1442 			    fixup_filename(file), line);
1443 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1444 			panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1445 			    class->lc_name, lock->lo_name,
1446 			    instance->li_flags & LI_RECURSEMASK,
1447 			    fixup_filename(file), line);
1448 	}
1449 	instance->li_flags |= LI_EXCLUSIVE;
1450 }
1451 
1452 void
1453 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1454     int line)
1455 {
1456 	struct lock_instance *instance;
1457 	struct lock_class *class;
1458 
1459 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1460 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1461 		return;
1462 	class = LOCK_CLASS(lock);
1463 	if (witness_watch) {
1464 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1465 		panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1466 			    class->lc_name, lock->lo_name,
1467 			    fixup_filename(file), line);
1468 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1469 			panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1470 			    class->lc_name, lock->lo_name,
1471 			    fixup_filename(file), line);
1472 	}
1473 	instance = find_instance(curthread->td_sleeplocks, lock);
1474 	if (instance == NULL)
1475 		panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1476 		    class->lc_name, lock->lo_name,
1477 		    fixup_filename(file), line);
1478 	if (witness_watch) {
1479 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1480 			panic("downgrade of shared lock (%s) %s @ %s:%d",
1481 			    class->lc_name, lock->lo_name,
1482 			    fixup_filename(file), line);
1483 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1484 			panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1485 			    class->lc_name, lock->lo_name,
1486 			    instance->li_flags & LI_RECURSEMASK,
1487 			    fixup_filename(file), line);
1488 	}
1489 	instance->li_flags &= ~LI_EXCLUSIVE;
1490 }
1491 
1492 void
1493 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1494 {
1495 	struct lock_list_entry **lock_list, *lle;
1496 	struct lock_instance *instance;
1497 	struct lock_class *class;
1498 	struct thread *td;
1499 	register_t s;
1500 	int i, j;
1501 
1502 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1503 		return;
1504 	td = curthread;
1505 	class = LOCK_CLASS(lock);
1506 
1507 	/* Find lock instance associated with this lock. */
1508 	if (class->lc_flags & LC_SLEEPLOCK)
1509 		lock_list = &td->td_sleeplocks;
1510 	else
1511 		lock_list = PCPU_PTR(spinlocks);
1512 	lle = *lock_list;
1513 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1514 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1515 			instance = &(*lock_list)->ll_children[i];
1516 			if (instance->li_lock == lock)
1517 				goto found;
1518 		}
1519 
1520 	/*
1521 	 * When disabling WITNESS through witness_watch we could end up in
1522 	 * having registered locks in the td_sleeplocks queue.
1523 	 * We have to make sure we flush these queues, so just search for
1524 	 * eventual register locks and remove them.
1525 	 */
1526 	if (witness_watch > 0)
1527 		panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1528 		    lock->lo_name, fixup_filename(file), line);
1529 	else
1530 		return;
1531 found:
1532 
1533 	/* First, check for shared/exclusive mismatches. */
1534 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1535 	    (flags & LOP_EXCLUSIVE) == 0) {
1536 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1537 		    lock->lo_name, fixup_filename(file), line);
1538 		printf("while exclusively locked from %s:%d\n",
1539 		    fixup_filename(instance->li_file), instance->li_line);
1540 		panic("excl->ushare");
1541 	}
1542 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1543 	    (flags & LOP_EXCLUSIVE) != 0) {
1544 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1545 		    lock->lo_name, fixup_filename(file), line);
1546 		printf("while share locked from %s:%d\n",
1547 		    fixup_filename(instance->li_file),
1548 		    instance->li_line);
1549 		panic("share->uexcl");
1550 	}
1551 	/* If we are recursed, unrecurse. */
1552 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1553 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1554 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1555 		    instance->li_flags);
1556 		instance->li_flags--;
1557 		return;
1558 	}
1559 	/* The lock is now being dropped, check for NORELEASE flag */
1560 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1561 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1562 		    lock->lo_name, fixup_filename(file), line);
1563 		panic("lock marked norelease");
1564 	}
1565 
1566 	/* Otherwise, remove this item from the list. */
1567 	s = intr_disable();
1568 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1569 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1570 	    (*lock_list)->ll_count - 1);
1571 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1572 		(*lock_list)->ll_children[j] =
1573 		    (*lock_list)->ll_children[j + 1];
1574 	(*lock_list)->ll_count--;
1575 	intr_restore(s);
1576 
1577 	/*
1578 	 * In order to reduce contention on w_mtx, we want to keep always an
1579 	 * head object into lists so that frequent allocation from the
1580 	 * free witness pool (and subsequent locking) is avoided.
1581 	 * In order to maintain the current code simple, when the head
1582 	 * object is totally unloaded it means also that we do not have
1583 	 * further objects in the list, so the list ownership needs to be
1584 	 * hand over to another object if the current head needs to be freed.
1585 	 */
1586 	if ((*lock_list)->ll_count == 0) {
1587 		if (*lock_list == lle) {
1588 			if (lle->ll_next == NULL)
1589 				return;
1590 		} else
1591 			lle = *lock_list;
1592 		*lock_list = lle->ll_next;
1593 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1594 		    td->td_proc->p_pid, lle);
1595 		witness_lock_list_free(lle);
1596 	}
1597 }
1598 
1599 void
1600 witness_thread_exit(struct thread *td)
1601 {
1602 	struct lock_list_entry *lle;
1603 	int i, n;
1604 
1605 	lle = td->td_sleeplocks;
1606 	if (lle == NULL || panicstr != NULL)
1607 		return;
1608 	if (lle->ll_count != 0) {
1609 		for (n = 0; lle != NULL; lle = lle->ll_next)
1610 			for (i = lle->ll_count - 1; i >= 0; i--) {
1611 				if (n == 0)
1612 		printf("Thread %p exiting with the following locks held:\n",
1613 					    td);
1614 				n++;
1615 				witness_list_lock(&lle->ll_children[i], printf);
1616 
1617 			}
1618 		panic("Thread %p cannot exit while holding sleeplocks\n", td);
1619 	}
1620 	witness_lock_list_free(lle);
1621 }
1622 
1623 /*
1624  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1625  * exempt Giant and sleepable locks from the checks as well.  If any
1626  * non-exempt locks are held, then a supplied message is printed to the
1627  * console along with a list of the offending locks.  If indicated in the
1628  * flags then a failure results in a panic as well.
1629  */
1630 int
1631 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1632 {
1633 	struct lock_list_entry *lock_list, *lle;
1634 	struct lock_instance *lock1;
1635 	struct thread *td;
1636 	va_list ap;
1637 	int i, n;
1638 
1639 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1640 		return (0);
1641 	n = 0;
1642 	td = curthread;
1643 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1644 		for (i = lle->ll_count - 1; i >= 0; i--) {
1645 			lock1 = &lle->ll_children[i];
1646 			if (lock1->li_lock == lock)
1647 				continue;
1648 			if (flags & WARN_GIANTOK &&
1649 			    lock1->li_lock == &Giant.lock_object)
1650 				continue;
1651 			if (flags & WARN_SLEEPOK &&
1652 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1653 				continue;
1654 			if (n == 0) {
1655 				va_start(ap, fmt);
1656 				vprintf(fmt, ap);
1657 				va_end(ap);
1658 				printf(" with the following");
1659 				if (flags & WARN_SLEEPOK)
1660 					printf(" non-sleepable");
1661 				printf(" locks held:\n");
1662 			}
1663 			n++;
1664 			witness_list_lock(lock1, printf);
1665 		}
1666 
1667 	/*
1668 	 * Pin the thread in order to avoid problems with thread migration.
1669 	 * Once that all verifies are passed about spinlocks ownership,
1670 	 * the thread is in a safe path and it can be unpinned.
1671 	 */
1672 	sched_pin();
1673 	lock_list = PCPU_GET(spinlocks);
1674 	if (lock_list != NULL && lock_list->ll_count != 0) {
1675 		sched_unpin();
1676 
1677 		/*
1678 		 * We should only have one spinlock and as long as
1679 		 * the flags cannot match for this locks class,
1680 		 * check if the first spinlock is the one curthread
1681 		 * should hold.
1682 		 */
1683 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1684 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1685 		    lock1->li_lock == lock && n == 0)
1686 			return (0);
1687 
1688 		va_start(ap, fmt);
1689 		vprintf(fmt, ap);
1690 		va_end(ap);
1691 		printf(" with the following");
1692 		if (flags & WARN_SLEEPOK)
1693 			printf(" non-sleepable");
1694 		printf(" locks held:\n");
1695 		n += witness_list_locks(&lock_list, printf);
1696 	} else
1697 		sched_unpin();
1698 	if (flags & WARN_PANIC && n)
1699 		panic("%s", __func__);
1700 	else
1701 		witness_debugger(n);
1702 	return (n);
1703 }
1704 
1705 const char *
1706 witness_file(struct lock_object *lock)
1707 {
1708 	struct witness *w;
1709 
1710 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1711 		return ("?");
1712 	w = lock->lo_witness;
1713 	return (w->w_file);
1714 }
1715 
1716 int
1717 witness_line(struct lock_object *lock)
1718 {
1719 	struct witness *w;
1720 
1721 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1722 		return (0);
1723 	w = lock->lo_witness;
1724 	return (w->w_line);
1725 }
1726 
1727 static struct witness *
1728 enroll(const char *description, struct lock_class *lock_class)
1729 {
1730 	struct witness *w;
1731 	struct witness_list *typelist;
1732 
1733 	MPASS(description != NULL);
1734 
1735 	if (witness_watch == -1 || panicstr != NULL)
1736 		return (NULL);
1737 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1738 		if (witness_skipspin)
1739 			return (NULL);
1740 		else
1741 			typelist = &w_spin;
1742 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1743 		typelist = &w_sleep;
1744 	else
1745 		panic("lock class %s is not sleep or spin",
1746 		    lock_class->lc_name);
1747 
1748 	mtx_lock_spin(&w_mtx);
1749 	w = witness_hash_get(description);
1750 	if (w)
1751 		goto found;
1752 	if ((w = witness_get()) == NULL)
1753 		return (NULL);
1754 	MPASS(strlen(description) < MAX_W_NAME);
1755 	strcpy(w->w_name, description);
1756 	w->w_class = lock_class;
1757 	w->w_refcount = 1;
1758 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1759 	if (lock_class->lc_flags & LC_SPINLOCK) {
1760 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1761 		w_spin_cnt++;
1762 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1763 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1764 		w_sleep_cnt++;
1765 	}
1766 
1767 	/* Insert new witness into the hash */
1768 	witness_hash_put(w);
1769 	witness_increment_graph_generation();
1770 	mtx_unlock_spin(&w_mtx);
1771 	return (w);
1772 found:
1773 	w->w_refcount++;
1774 	mtx_unlock_spin(&w_mtx);
1775 	if (lock_class != w->w_class)
1776 		panic(
1777 			"lock (%s) %s does not match earlier (%s) lock",
1778 			description, lock_class->lc_name,
1779 			w->w_class->lc_name);
1780 	return (w);
1781 }
1782 
1783 static void
1784 depart(struct witness *w)
1785 {
1786 	struct witness_list *list;
1787 
1788 	MPASS(w->w_refcount == 0);
1789 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1790 		list = &w_sleep;
1791 		w_sleep_cnt--;
1792 	} else {
1793 		list = &w_spin;
1794 		w_spin_cnt--;
1795 	}
1796 	/*
1797 	 * Set file to NULL as it may point into a loadable module.
1798 	 */
1799 	w->w_file = NULL;
1800 	w->w_line = 0;
1801 	witness_increment_graph_generation();
1802 }
1803 
1804 
1805 static void
1806 adopt(struct witness *parent, struct witness *child)
1807 {
1808 	int pi, ci, i, j;
1809 
1810 	if (witness_cold == 0)
1811 		mtx_assert(&w_mtx, MA_OWNED);
1812 
1813 	/* If the relationship is already known, there's no work to be done. */
1814 	if (isitmychild(parent, child))
1815 		return;
1816 
1817 	/* When the structure of the graph changes, bump up the generation. */
1818 	witness_increment_graph_generation();
1819 
1820 	/*
1821 	 * The hard part ... create the direct relationship, then propagate all
1822 	 * indirect relationships.
1823 	 */
1824 	pi = parent->w_index;
1825 	ci = child->w_index;
1826 	WITNESS_INDEX_ASSERT(pi);
1827 	WITNESS_INDEX_ASSERT(ci);
1828 	MPASS(pi != ci);
1829 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1830 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1831 
1832 	/*
1833 	 * If parent was not already an ancestor of child,
1834 	 * then we increment the descendant and ancestor counters.
1835 	 */
1836 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1837 		parent->w_num_descendants++;
1838 		child->w_num_ancestors++;
1839 	}
1840 
1841 	/*
1842 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1843 	 * an ancestor of 'pi' during this loop.
1844 	 */
1845 	for (i = 1; i <= w_max_used_index; i++) {
1846 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1847 		    (i != pi))
1848 			continue;
1849 
1850 		/* Find each descendant of 'i' and mark it as a descendant. */
1851 		for (j = 1; j <= w_max_used_index; j++) {
1852 
1853 			/*
1854 			 * Skip children that are already marked as
1855 			 * descendants of 'i'.
1856 			 */
1857 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1858 				continue;
1859 
1860 			/*
1861 			 * We are only interested in descendants of 'ci'. Note
1862 			 * that 'ci' itself is counted as a descendant of 'ci'.
1863 			 */
1864 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1865 			    (j != ci))
1866 				continue;
1867 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1868 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1869 			w_data[i].w_num_descendants++;
1870 			w_data[j].w_num_ancestors++;
1871 
1872 			/*
1873 			 * Make sure we aren't marking a node as both an
1874 			 * ancestor and descendant. We should have caught
1875 			 * this as a lock order reversal earlier.
1876 			 */
1877 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1878 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1879 				printf("witness rmatrix paradox! [%d][%d]=%d "
1880 				    "both ancestor and descendant\n",
1881 				    i, j, w_rmatrix[i][j]);
1882 				kdb_backtrace();
1883 				printf("Witness disabled.\n");
1884 				witness_watch = -1;
1885 			}
1886 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1887 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1888 				printf("witness rmatrix paradox! [%d][%d]=%d "
1889 				    "both ancestor and descendant\n",
1890 				    j, i, w_rmatrix[j][i]);
1891 				kdb_backtrace();
1892 				printf("Witness disabled.\n");
1893 				witness_watch = -1;
1894 			}
1895 		}
1896 	}
1897 }
1898 
1899 static void
1900 itismychild(struct witness *parent, struct witness *child)
1901 {
1902 
1903 	MPASS(child != NULL && parent != NULL);
1904 	if (witness_cold == 0)
1905 		mtx_assert(&w_mtx, MA_OWNED);
1906 
1907 	if (!witness_lock_type_equal(parent, child)) {
1908 		if (witness_cold == 0)
1909 			mtx_unlock_spin(&w_mtx);
1910 		panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1911 		    "the same lock type", __func__, parent->w_name,
1912 		    parent->w_class->lc_name, child->w_name,
1913 		    child->w_class->lc_name);
1914 	}
1915 	adopt(parent, child);
1916 }
1917 
1918 /*
1919  * Generic code for the isitmy*() functions. The rmask parameter is the
1920  * expected relationship of w1 to w2.
1921  */
1922 static int
1923 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1924 {
1925 	unsigned char r1, r2;
1926 	int i1, i2;
1927 
1928 	i1 = w1->w_index;
1929 	i2 = w2->w_index;
1930 	WITNESS_INDEX_ASSERT(i1);
1931 	WITNESS_INDEX_ASSERT(i2);
1932 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1933 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1934 
1935 	/* The flags on one better be the inverse of the flags on the other */
1936 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1937 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1938 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
1939 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
1940 		    "w_rmatrix[%d][%d] == %hhx\n",
1941 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1942 		    i2, i1, r2);
1943 		kdb_backtrace();
1944 		printf("Witness disabled.\n");
1945 		witness_watch = -1;
1946 	}
1947 	return (r1 & rmask);
1948 }
1949 
1950 /*
1951  * Checks if @child is a direct child of @parent.
1952  */
1953 static int
1954 isitmychild(struct witness *parent, struct witness *child)
1955 {
1956 
1957 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1958 }
1959 
1960 /*
1961  * Checks if @descendant is a direct or inderect descendant of @ancestor.
1962  */
1963 static int
1964 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1965 {
1966 
1967 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1968 	    __func__));
1969 }
1970 
1971 #ifdef BLESSING
1972 static int
1973 blessed(struct witness *w1, struct witness *w2)
1974 {
1975 	int i;
1976 	struct witness_blessed *b;
1977 
1978 	for (i = 0; i < blessed_count; i++) {
1979 		b = &blessed_list[i];
1980 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1981 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1982 				return (1);
1983 			continue;
1984 		}
1985 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1986 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1987 				return (1);
1988 	}
1989 	return (0);
1990 }
1991 #endif
1992 
1993 static struct witness *
1994 witness_get(void)
1995 {
1996 	struct witness *w;
1997 	int index;
1998 
1999 	if (witness_cold == 0)
2000 		mtx_assert(&w_mtx, MA_OWNED);
2001 
2002 	if (witness_watch == -1) {
2003 		mtx_unlock_spin(&w_mtx);
2004 		return (NULL);
2005 	}
2006 	if (STAILQ_EMPTY(&w_free)) {
2007 		witness_watch = -1;
2008 		mtx_unlock_spin(&w_mtx);
2009 		printf("WITNESS: unable to allocate a new witness object\n");
2010 		return (NULL);
2011 	}
2012 	w = STAILQ_FIRST(&w_free);
2013 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2014 	w_free_cnt--;
2015 	index = w->w_index;
2016 	MPASS(index > 0 && index == w_max_used_index+1 &&
2017 	    index < WITNESS_COUNT);
2018 	bzero(w, sizeof(*w));
2019 	w->w_index = index;
2020 	if (index > w_max_used_index)
2021 		w_max_used_index = index;
2022 	return (w);
2023 }
2024 
2025 static void
2026 witness_free(struct witness *w)
2027 {
2028 
2029 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2030 	w_free_cnt++;
2031 }
2032 
2033 static struct lock_list_entry *
2034 witness_lock_list_get(void)
2035 {
2036 	struct lock_list_entry *lle;
2037 
2038 	if (witness_watch == -1)
2039 		return (NULL);
2040 	mtx_lock_spin(&w_mtx);
2041 	lle = w_lock_list_free;
2042 	if (lle == NULL) {
2043 		witness_watch = -1;
2044 		mtx_unlock_spin(&w_mtx);
2045 		printf("%s: witness exhausted\n", __func__);
2046 		return (NULL);
2047 	}
2048 	w_lock_list_free = lle->ll_next;
2049 	mtx_unlock_spin(&w_mtx);
2050 	bzero(lle, sizeof(*lle));
2051 	return (lle);
2052 }
2053 
2054 static void
2055 witness_lock_list_free(struct lock_list_entry *lle)
2056 {
2057 
2058 	mtx_lock_spin(&w_mtx);
2059 	lle->ll_next = w_lock_list_free;
2060 	w_lock_list_free = lle;
2061 	mtx_unlock_spin(&w_mtx);
2062 }
2063 
2064 static struct lock_instance *
2065 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2066 {
2067 	struct lock_list_entry *lle;
2068 	struct lock_instance *instance;
2069 	int i;
2070 
2071 	for (lle = list; lle != NULL; lle = lle->ll_next)
2072 		for (i = lle->ll_count - 1; i >= 0; i--) {
2073 			instance = &lle->ll_children[i];
2074 			if (instance->li_lock == lock)
2075 				return (instance);
2076 		}
2077 	return (NULL);
2078 }
2079 
2080 static void
2081 witness_list_lock(struct lock_instance *instance,
2082     int (*prnt)(const char *fmt, ...))
2083 {
2084 	struct lock_object *lock;
2085 
2086 	lock = instance->li_lock;
2087 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2088 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2089 	if (lock->lo_witness->w_name != lock->lo_name)
2090 		prnt(" (%s)", lock->lo_witness->w_name);
2091 	prnt(" r = %d (%p) locked @ %s:%d\n",
2092 	    instance->li_flags & LI_RECURSEMASK, lock,
2093 	    fixup_filename(instance->li_file), instance->li_line);
2094 }
2095 
2096 #ifdef DDB
2097 static int
2098 witness_thread_has_locks(struct thread *td)
2099 {
2100 
2101 	if (td->td_sleeplocks == NULL)
2102 		return (0);
2103 	return (td->td_sleeplocks->ll_count != 0);
2104 }
2105 
2106 static int
2107 witness_proc_has_locks(struct proc *p)
2108 {
2109 	struct thread *td;
2110 
2111 	FOREACH_THREAD_IN_PROC(p, td) {
2112 		if (witness_thread_has_locks(td))
2113 			return (1);
2114 	}
2115 	return (0);
2116 }
2117 #endif
2118 
2119 int
2120 witness_list_locks(struct lock_list_entry **lock_list,
2121     int (*prnt)(const char *fmt, ...))
2122 {
2123 	struct lock_list_entry *lle;
2124 	int i, nheld;
2125 
2126 	nheld = 0;
2127 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2128 		for (i = lle->ll_count - 1; i >= 0; i--) {
2129 			witness_list_lock(&lle->ll_children[i], prnt);
2130 			nheld++;
2131 		}
2132 	return (nheld);
2133 }
2134 
2135 /*
2136  * This is a bit risky at best.  We call this function when we have timed
2137  * out acquiring a spin lock, and we assume that the other CPU is stuck
2138  * with this lock held.  So, we go groveling around in the other CPU's
2139  * per-cpu data to try to find the lock instance for this spin lock to
2140  * see when it was last acquired.
2141  */
2142 void
2143 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2144     int (*prnt)(const char *fmt, ...))
2145 {
2146 	struct lock_instance *instance;
2147 	struct pcpu *pc;
2148 
2149 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2150 		return;
2151 	pc = pcpu_find(owner->td_oncpu);
2152 	instance = find_instance(pc->pc_spinlocks, lock);
2153 	if (instance != NULL)
2154 		witness_list_lock(instance, prnt);
2155 }
2156 
2157 void
2158 witness_save(struct lock_object *lock, const char **filep, int *linep)
2159 {
2160 	struct lock_list_entry *lock_list;
2161 	struct lock_instance *instance;
2162 	struct lock_class *class;
2163 
2164 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2165 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2166 		return;
2167 	class = LOCK_CLASS(lock);
2168 	if (class->lc_flags & LC_SLEEPLOCK)
2169 		lock_list = curthread->td_sleeplocks;
2170 	else {
2171 		if (witness_skipspin)
2172 			return;
2173 		lock_list = PCPU_GET(spinlocks);
2174 	}
2175 	instance = find_instance(lock_list, lock);
2176 	if (instance == NULL)
2177 		panic("%s: lock (%s) %s not locked", __func__,
2178 		    class->lc_name, lock->lo_name);
2179 	*filep = instance->li_file;
2180 	*linep = instance->li_line;
2181 }
2182 
2183 void
2184 witness_restore(struct lock_object *lock, const char *file, int line)
2185 {
2186 	struct lock_list_entry *lock_list;
2187 	struct lock_instance *instance;
2188 	struct lock_class *class;
2189 
2190 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2191 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2192 		return;
2193 	class = LOCK_CLASS(lock);
2194 	if (class->lc_flags & LC_SLEEPLOCK)
2195 		lock_list = curthread->td_sleeplocks;
2196 	else {
2197 		if (witness_skipspin)
2198 			return;
2199 		lock_list = PCPU_GET(spinlocks);
2200 	}
2201 	instance = find_instance(lock_list, lock);
2202 	if (instance == NULL)
2203 		panic("%s: lock (%s) %s not locked", __func__,
2204 		    class->lc_name, lock->lo_name);
2205 	lock->lo_witness->w_file = file;
2206 	lock->lo_witness->w_line = line;
2207 	instance->li_file = file;
2208 	instance->li_line = line;
2209 }
2210 
2211 void
2212 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2213 {
2214 #ifdef INVARIANT_SUPPORT
2215 	struct lock_instance *instance;
2216 	struct lock_class *class;
2217 
2218 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2219 		return;
2220 	class = LOCK_CLASS(lock);
2221 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2222 		instance = find_instance(curthread->td_sleeplocks, lock);
2223 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2224 		instance = find_instance(PCPU_GET(spinlocks), lock);
2225 	else {
2226 		panic("Lock (%s) %s is not sleep or spin!",
2227 		    class->lc_name, lock->lo_name);
2228 	}
2229 	switch (flags) {
2230 	case LA_UNLOCKED:
2231 		if (instance != NULL)
2232 			panic("Lock (%s) %s locked @ %s:%d.",
2233 			    class->lc_name, lock->lo_name,
2234 			    fixup_filename(file), line);
2235 		break;
2236 	case LA_LOCKED:
2237 	case LA_LOCKED | LA_RECURSED:
2238 	case LA_LOCKED | LA_NOTRECURSED:
2239 	case LA_SLOCKED:
2240 	case LA_SLOCKED | LA_RECURSED:
2241 	case LA_SLOCKED | LA_NOTRECURSED:
2242 	case LA_XLOCKED:
2243 	case LA_XLOCKED | LA_RECURSED:
2244 	case LA_XLOCKED | LA_NOTRECURSED:
2245 		if (instance == NULL) {
2246 			panic("Lock (%s) %s not locked @ %s:%d.",
2247 			    class->lc_name, lock->lo_name,
2248 			    fixup_filename(file), line);
2249 			break;
2250 		}
2251 		if ((flags & LA_XLOCKED) != 0 &&
2252 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2253 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2254 			    class->lc_name, lock->lo_name,
2255 			    fixup_filename(file), line);
2256 		if ((flags & LA_SLOCKED) != 0 &&
2257 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2258 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
2259 			    class->lc_name, lock->lo_name,
2260 			    fixup_filename(file), line);
2261 		if ((flags & LA_RECURSED) != 0 &&
2262 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2263 			panic("Lock (%s) %s not recursed @ %s:%d.",
2264 			    class->lc_name, lock->lo_name,
2265 			    fixup_filename(file), line);
2266 		if ((flags & LA_NOTRECURSED) != 0 &&
2267 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2268 			panic("Lock (%s) %s recursed @ %s:%d.",
2269 			    class->lc_name, lock->lo_name,
2270 			    fixup_filename(file), line);
2271 		break;
2272 	default:
2273 		panic("Invalid lock assertion at %s:%d.",
2274 		    fixup_filename(file), line);
2275 
2276 	}
2277 #endif	/* INVARIANT_SUPPORT */
2278 }
2279 
2280 static void
2281 witness_setflag(struct lock_object *lock, int flag, int set)
2282 {
2283 	struct lock_list_entry *lock_list;
2284 	struct lock_instance *instance;
2285 	struct lock_class *class;
2286 
2287 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2288 		return;
2289 	class = LOCK_CLASS(lock);
2290 	if (class->lc_flags & LC_SLEEPLOCK)
2291 		lock_list = curthread->td_sleeplocks;
2292 	else {
2293 		if (witness_skipspin)
2294 			return;
2295 		lock_list = PCPU_GET(spinlocks);
2296 	}
2297 	instance = find_instance(lock_list, lock);
2298 	if (instance == NULL)
2299 		panic("%s: lock (%s) %s not locked", __func__,
2300 		    class->lc_name, lock->lo_name);
2301 
2302 	if (set)
2303 		instance->li_flags |= flag;
2304 	else
2305 		instance->li_flags &= ~flag;
2306 }
2307 
2308 void
2309 witness_norelease(struct lock_object *lock)
2310 {
2311 
2312 	witness_setflag(lock, LI_NORELEASE, 1);
2313 }
2314 
2315 void
2316 witness_releaseok(struct lock_object *lock)
2317 {
2318 
2319 	witness_setflag(lock, LI_NORELEASE, 0);
2320 }
2321 
2322 #ifdef DDB
2323 static void
2324 witness_ddb_list(struct thread *td)
2325 {
2326 
2327 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2328 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2329 
2330 	if (witness_watch < 1)
2331 		return;
2332 
2333 	witness_list_locks(&td->td_sleeplocks, db_printf);
2334 
2335 	/*
2336 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2337 	 * if td is currently executing on some other CPU and holds spin locks
2338 	 * as we won't display those locks.  If we had a MI way of getting
2339 	 * the per-cpu data for a given cpu then we could use
2340 	 * td->td_oncpu to get the list of spinlocks for this thread
2341 	 * and "fix" this.
2342 	 *
2343 	 * That still wouldn't really fix this unless we locked the scheduler
2344 	 * lock or stopped the other CPU to make sure it wasn't changing the
2345 	 * list out from under us.  It is probably best to just not try to
2346 	 * handle threads on other CPU's for now.
2347 	 */
2348 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2349 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2350 }
2351 
2352 DB_SHOW_COMMAND(locks, db_witness_list)
2353 {
2354 	struct thread *td;
2355 
2356 	if (have_addr)
2357 		td = db_lookup_thread(addr, TRUE);
2358 	else
2359 		td = kdb_thread;
2360 	witness_ddb_list(td);
2361 }
2362 
2363 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2364 {
2365 	struct thread *td;
2366 	struct proc *p;
2367 
2368 	/*
2369 	 * It would be nice to list only threads and processes that actually
2370 	 * held sleep locks, but that information is currently not exported
2371 	 * by WITNESS.
2372 	 */
2373 	FOREACH_PROC_IN_SYSTEM(p) {
2374 		if (!witness_proc_has_locks(p))
2375 			continue;
2376 		FOREACH_THREAD_IN_PROC(p, td) {
2377 			if (!witness_thread_has_locks(td))
2378 				continue;
2379 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2380 			    p->p_comm, td, td->td_tid);
2381 			witness_ddb_list(td);
2382 		}
2383 	}
2384 }
2385 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2386 
2387 DB_SHOW_COMMAND(witness, db_witness_display)
2388 {
2389 
2390 	witness_ddb_display(db_printf);
2391 }
2392 #endif
2393 
2394 static int
2395 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2396 {
2397 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2398 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2399 	struct sbuf *sb;
2400 	u_int w_rmatrix1, w_rmatrix2;
2401 	int error, generation, i, j;
2402 
2403 	tmp_data1 = NULL;
2404 	tmp_data2 = NULL;
2405 	tmp_w1 = NULL;
2406 	tmp_w2 = NULL;
2407 	if (witness_watch < 1) {
2408 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2409 		return (error);
2410 	}
2411 	if (witness_cold) {
2412 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2413 		return (error);
2414 	}
2415 	error = 0;
2416 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2417 	if (sb == NULL)
2418 		return (ENOMEM);
2419 
2420 	/* Allocate and init temporary storage space. */
2421 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2422 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2423 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2424 	    M_WAITOK | M_ZERO);
2425 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2426 	    M_WAITOK | M_ZERO);
2427 	stack_zero(&tmp_data1->wlod_stack);
2428 	stack_zero(&tmp_data2->wlod_stack);
2429 
2430 restart:
2431 	mtx_lock_spin(&w_mtx);
2432 	generation = w_generation;
2433 	mtx_unlock_spin(&w_mtx);
2434 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2435 	    w_lohash.wloh_count);
2436 	for (i = 1; i < w_max_used_index; i++) {
2437 		mtx_lock_spin(&w_mtx);
2438 		if (generation != w_generation) {
2439 			mtx_unlock_spin(&w_mtx);
2440 
2441 			/* The graph has changed, try again. */
2442 			req->oldidx = 0;
2443 			sbuf_clear(sb);
2444 			goto restart;
2445 		}
2446 
2447 		w1 = &w_data[i];
2448 		if (w1->w_reversed == 0) {
2449 			mtx_unlock_spin(&w_mtx);
2450 			continue;
2451 		}
2452 
2453 		/* Copy w1 locally so we can release the spin lock. */
2454 		*tmp_w1 = *w1;
2455 		mtx_unlock_spin(&w_mtx);
2456 
2457 		if (tmp_w1->w_reversed == 0)
2458 			continue;
2459 		for (j = 1; j < w_max_used_index; j++) {
2460 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2461 				continue;
2462 
2463 			mtx_lock_spin(&w_mtx);
2464 			if (generation != w_generation) {
2465 				mtx_unlock_spin(&w_mtx);
2466 
2467 				/* The graph has changed, try again. */
2468 				req->oldidx = 0;
2469 				sbuf_clear(sb);
2470 				goto restart;
2471 			}
2472 
2473 			w2 = &w_data[j];
2474 			data1 = witness_lock_order_get(w1, w2);
2475 			data2 = witness_lock_order_get(w2, w1);
2476 
2477 			/*
2478 			 * Copy information locally so we can release the
2479 			 * spin lock.
2480 			 */
2481 			*tmp_w2 = *w2;
2482 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2483 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2484 
2485 			if (data1) {
2486 				stack_zero(&tmp_data1->wlod_stack);
2487 				stack_copy(&data1->wlod_stack,
2488 				    &tmp_data1->wlod_stack);
2489 			}
2490 			if (data2 && data2 != data1) {
2491 				stack_zero(&tmp_data2->wlod_stack);
2492 				stack_copy(&data2->wlod_stack,
2493 				    &tmp_data2->wlod_stack);
2494 			}
2495 			mtx_unlock_spin(&w_mtx);
2496 
2497 			sbuf_printf(sb,
2498 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2499 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2500 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2501 #if 0
2502  			sbuf_printf(sb,
2503 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2504  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2505  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2506 #endif
2507 			if (data1) {
2508 				sbuf_printf(sb,
2509 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2510 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2511 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2512 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2513 				sbuf_printf(sb, "\n");
2514 			}
2515 			if (data2 && data2 != data1) {
2516 				sbuf_printf(sb,
2517 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2518 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2519 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2520 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2521 				sbuf_printf(sb, "\n");
2522 			}
2523 		}
2524 	}
2525 	mtx_lock_spin(&w_mtx);
2526 	if (generation != w_generation) {
2527 		mtx_unlock_spin(&w_mtx);
2528 
2529 		/*
2530 		 * The graph changed while we were printing stack data,
2531 		 * try again.
2532 		 */
2533 		req->oldidx = 0;
2534 		sbuf_clear(sb);
2535 		goto restart;
2536 	}
2537 	mtx_unlock_spin(&w_mtx);
2538 
2539 	/* Free temporary storage space. */
2540 	free(tmp_data1, M_TEMP);
2541 	free(tmp_data2, M_TEMP);
2542 	free(tmp_w1, M_TEMP);
2543 	free(tmp_w2, M_TEMP);
2544 
2545 	sbuf_finish(sb);
2546 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2547 	sbuf_delete(sb);
2548 
2549 	return (error);
2550 }
2551 
2552 static int
2553 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2554 {
2555 	struct witness *w;
2556 	struct sbuf *sb;
2557 	int error;
2558 
2559 	if (witness_watch < 1) {
2560 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2561 		return (error);
2562 	}
2563 	if (witness_cold) {
2564 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2565 		return (error);
2566 	}
2567 	error = 0;
2568 
2569 	error = sysctl_wire_old_buffer(req, 0);
2570 	if (error != 0)
2571 		return (error);
2572 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2573 	if (sb == NULL)
2574 		return (ENOMEM);
2575 	sbuf_printf(sb, "\n");
2576 
2577 	mtx_lock_spin(&w_mtx);
2578 	STAILQ_FOREACH(w, &w_all, w_list)
2579 		w->w_displayed = 0;
2580 	STAILQ_FOREACH(w, &w_all, w_list)
2581 		witness_add_fullgraph(sb, w);
2582 	mtx_unlock_spin(&w_mtx);
2583 
2584 	/*
2585 	 * Close the sbuf and return to userland.
2586 	 */
2587 	error = sbuf_finish(sb);
2588 	sbuf_delete(sb);
2589 
2590 	return (error);
2591 }
2592 
2593 static int
2594 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2595 {
2596 	int error, value;
2597 
2598 	value = witness_watch;
2599 	error = sysctl_handle_int(oidp, &value, 0, req);
2600 	if (error != 0 || req->newptr == NULL)
2601 		return (error);
2602 	if (value > 1 || value < -1 ||
2603 	    (witness_watch == -1 && value != witness_watch))
2604 		return (EINVAL);
2605 	witness_watch = value;
2606 	return (0);
2607 }
2608 
2609 static void
2610 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2611 {
2612 	int i;
2613 
2614 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2615 		return;
2616 	w->w_displayed = 1;
2617 
2618 	WITNESS_INDEX_ASSERT(w->w_index);
2619 	for (i = 1; i <= w_max_used_index; i++) {
2620 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2621 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2622 			    w_data[i].w_name);
2623 			witness_add_fullgraph(sb, &w_data[i]);
2624 		}
2625 	}
2626 }
2627 
2628 /*
2629  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2630  * interprets the key as a string and reads until the null
2631  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2632  * hash value computed from the key.
2633  */
2634 static uint32_t
2635 witness_hash_djb2(const uint8_t *key, uint32_t size)
2636 {
2637 	unsigned int hash = 5381;
2638 	int i;
2639 
2640 	/* hash = hash * 33 + key[i] */
2641 	if (size)
2642 		for (i = 0; i < size; i++)
2643 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2644 	else
2645 		for (i = 0; key[i] != 0; i++)
2646 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2647 
2648 	return (hash);
2649 }
2650 
2651 
2652 /*
2653  * Initializes the two witness hash tables. Called exactly once from
2654  * witness_initialize().
2655  */
2656 static void
2657 witness_init_hash_tables(void)
2658 {
2659 	int i;
2660 
2661 	MPASS(witness_cold);
2662 
2663 	/* Initialize the hash tables. */
2664 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2665 		w_hash.wh_array[i] = NULL;
2666 
2667 	w_hash.wh_size = WITNESS_HASH_SIZE;
2668 	w_hash.wh_count = 0;
2669 
2670 	/* Initialize the lock order data hash. */
2671 	w_lofree = NULL;
2672 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2673 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2674 		w_lodata[i].wlod_next = w_lofree;
2675 		w_lofree = &w_lodata[i];
2676 	}
2677 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2678 	w_lohash.wloh_count = 0;
2679 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2680 		w_lohash.wloh_array[i] = NULL;
2681 }
2682 
2683 static struct witness *
2684 witness_hash_get(const char *key)
2685 {
2686 	struct witness *w;
2687 	uint32_t hash;
2688 
2689 	MPASS(key != NULL);
2690 	if (witness_cold == 0)
2691 		mtx_assert(&w_mtx, MA_OWNED);
2692 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2693 	w = w_hash.wh_array[hash];
2694 	while (w != NULL) {
2695 		if (strcmp(w->w_name, key) == 0)
2696 			goto out;
2697 		w = w->w_hash_next;
2698 	}
2699 
2700 out:
2701 	return (w);
2702 }
2703 
2704 static void
2705 witness_hash_put(struct witness *w)
2706 {
2707 	uint32_t hash;
2708 
2709 	MPASS(w != NULL);
2710 	MPASS(w->w_name != NULL);
2711 	if (witness_cold == 0)
2712 		mtx_assert(&w_mtx, MA_OWNED);
2713 	KASSERT(witness_hash_get(w->w_name) == NULL,
2714 	    ("%s: trying to add a hash entry that already exists!", __func__));
2715 	KASSERT(w->w_hash_next == NULL,
2716 	    ("%s: w->w_hash_next != NULL", __func__));
2717 
2718 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2719 	w->w_hash_next = w_hash.wh_array[hash];
2720 	w_hash.wh_array[hash] = w;
2721 	w_hash.wh_count++;
2722 }
2723 
2724 
2725 static struct witness_lock_order_data *
2726 witness_lock_order_get(struct witness *parent, struct witness *child)
2727 {
2728 	struct witness_lock_order_data *data = NULL;
2729 	struct witness_lock_order_key key;
2730 	unsigned int hash;
2731 
2732 	MPASS(parent != NULL && child != NULL);
2733 	key.from = parent->w_index;
2734 	key.to = child->w_index;
2735 	WITNESS_INDEX_ASSERT(key.from);
2736 	WITNESS_INDEX_ASSERT(key.to);
2737 	if ((w_rmatrix[parent->w_index][child->w_index]
2738 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2739 		goto out;
2740 
2741 	hash = witness_hash_djb2((const char*)&key,
2742 	    sizeof(key)) % w_lohash.wloh_size;
2743 	data = w_lohash.wloh_array[hash];
2744 	while (data != NULL) {
2745 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2746 			break;
2747 		data = data->wlod_next;
2748 	}
2749 
2750 out:
2751 	return (data);
2752 }
2753 
2754 /*
2755  * Verify that parent and child have a known relationship, are not the same,
2756  * and child is actually a child of parent.  This is done without w_mtx
2757  * to avoid contention in the common case.
2758  */
2759 static int
2760 witness_lock_order_check(struct witness *parent, struct witness *child)
2761 {
2762 
2763 	if (parent != child &&
2764 	    w_rmatrix[parent->w_index][child->w_index]
2765 	    & WITNESS_LOCK_ORDER_KNOWN &&
2766 	    isitmychild(parent, child))
2767 		return (1);
2768 
2769 	return (0);
2770 }
2771 
2772 static int
2773 witness_lock_order_add(struct witness *parent, struct witness *child)
2774 {
2775 	struct witness_lock_order_data *data = NULL;
2776 	struct witness_lock_order_key key;
2777 	unsigned int hash;
2778 
2779 	MPASS(parent != NULL && child != NULL);
2780 	key.from = parent->w_index;
2781 	key.to = child->w_index;
2782 	WITNESS_INDEX_ASSERT(key.from);
2783 	WITNESS_INDEX_ASSERT(key.to);
2784 	if (w_rmatrix[parent->w_index][child->w_index]
2785 	    & WITNESS_LOCK_ORDER_KNOWN)
2786 		return (1);
2787 
2788 	hash = witness_hash_djb2((const char*)&key,
2789 	    sizeof(key)) % w_lohash.wloh_size;
2790 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2791 	data = w_lofree;
2792 	if (data == NULL)
2793 		return (0);
2794 	w_lofree = data->wlod_next;
2795 	data->wlod_next = w_lohash.wloh_array[hash];
2796 	data->wlod_key = key;
2797 	w_lohash.wloh_array[hash] = data;
2798 	w_lohash.wloh_count++;
2799 	stack_zero(&data->wlod_stack);
2800 	stack_save(&data->wlod_stack);
2801 	return (1);
2802 }
2803 
2804 /* Call this whenver the structure of the witness graph changes. */
2805 static void
2806 witness_increment_graph_generation(void)
2807 {
2808 
2809 	if (witness_cold == 0)
2810 		mtx_assert(&w_mtx, MA_OWNED);
2811 	w_generation++;
2812 }
2813 
2814 #ifdef KDB
2815 static void
2816 _witness_debugger(int cond, const char *msg)
2817 {
2818 
2819 	if (witness_trace && cond)
2820 		kdb_backtrace();
2821 	if (witness_kdb && cond)
2822 		kdb_enter(KDB_WHY_WITNESS, msg);
2823 }
2824 #endif
2825