xref: /freebsd/sys/kern/subr_witness.c (revision 0772266e638b0b33e816febac50777b942192b7c)
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 static 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 				    const 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 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
380     "Witness Locking");
381 
382 /*
383  * If set to 0, lock order checking is disabled.  If set to -1,
384  * witness is completely disabled.  Otherwise witness performs full
385  * lock order checking for all locks.  At runtime, lock order checking
386  * may be toggled.  However, witness cannot be reenabled once it is
387  * completely disabled.
388  */
389 static int witness_watch = 1;
390 TUNABLE_INT("debug.witness.watch", &witness_watch);
391 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
392     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
393 
394 #ifdef KDB
395 /*
396  * When KDB is enabled and witness_kdb is 1, it will cause the system
397  * to drop into kdebug() when:
398  *	- a lock hierarchy violation occurs
399  *	- locks are held when going to sleep.
400  */
401 #ifdef WITNESS_KDB
402 int	witness_kdb = 1;
403 #else
404 int	witness_kdb = 0;
405 #endif
406 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
408 
409 /*
410  * When KDB is enabled and witness_trace is 1, it will cause the system
411  * to print a stack trace:
412  *	- a lock hierarchy violation occurs
413  *	- locks are held when going to sleep.
414  */
415 int	witness_trace = 1;
416 TUNABLE_INT("debug.witness.trace", &witness_trace);
417 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
418 #endif /* KDB */
419 
420 #ifdef WITNESS_SKIPSPIN
421 int	witness_skipspin = 1;
422 #else
423 int	witness_skipspin = 0;
424 #endif
425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
427     0, "");
428 
429 /*
430  * Call this to print out the relations between locks.
431  */
432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
433     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
434 
435 /*
436  * Call this to print out the witness faulty stacks.
437  */
438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
439     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
440 
441 static struct mtx w_mtx;
442 
443 /* w_list */
444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
446 
447 /* w_typelist */
448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
450 
451 /* lock list */
452 static struct lock_list_entry *w_lock_list_free = NULL;
453 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
454 static u_int pending_cnt;
455 
456 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
457 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
458 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
459 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
460     "");
461 
462 static struct witness *w_data;
463 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
464 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
465 static struct witness_hash w_hash;	/* The witness hash table. */
466 
467 /* The lock order data hash */
468 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
469 static struct witness_lock_order_data *w_lofree = NULL;
470 static struct witness_lock_order_hash w_lohash;
471 static int w_max_used_index = 0;
472 static unsigned int w_generation = 0;
473 static const char w_notrunning[] = "Witness not running\n";
474 static const char w_stillcold[] = "Witness is still cold\n";
475 
476 
477 static struct witness_order_list_entry order_lists[] = {
478 	/*
479 	 * sx locks
480 	 */
481 	{ "proctree", &lock_class_sx },
482 	{ "allproc", &lock_class_sx },
483 	{ "allprison", &lock_class_sx },
484 	{ NULL, NULL },
485 	/*
486 	 * Various mutexes
487 	 */
488 	{ "Giant", &lock_class_mtx_sleep },
489 	{ "pipe mutex", &lock_class_mtx_sleep },
490 	{ "sigio lock", &lock_class_mtx_sleep },
491 	{ "process group", &lock_class_mtx_sleep },
492 	{ "process lock", &lock_class_mtx_sleep },
493 	{ "session", &lock_class_mtx_sleep },
494 	{ "uidinfo hash", &lock_class_rw },
495 #ifdef	HWPMC_HOOKS
496 	{ "pmc-sleep", &lock_class_mtx_sleep },
497 #endif
498 	{ "time lock", &lock_class_mtx_sleep },
499 	{ NULL, NULL },
500 	/*
501 	 * Sockets
502 	 */
503 	{ "accept", &lock_class_mtx_sleep },
504 	{ "so_snd", &lock_class_mtx_sleep },
505 	{ "so_rcv", &lock_class_mtx_sleep },
506 	{ "sellck", &lock_class_mtx_sleep },
507 	{ NULL, NULL },
508 	/*
509 	 * Routing
510 	 */
511 	{ "so_rcv", &lock_class_mtx_sleep },
512 	{ "radix node head", &lock_class_rw },
513 	{ "rtentry", &lock_class_mtx_sleep },
514 	{ "ifaddr", &lock_class_mtx_sleep },
515 	{ NULL, NULL },
516 	/*
517 	 * IPv4 multicast:
518 	 * protocol locks before interface locks, after UDP locks.
519 	 */
520 	{ "udpinp", &lock_class_rw },
521 	{ "in_multi_mtx", &lock_class_mtx_sleep },
522 	{ "igmp_mtx", &lock_class_mtx_sleep },
523 	{ "if_addr_lock", &lock_class_rw },
524 	{ NULL, NULL },
525 	/*
526 	 * IPv6 multicast:
527 	 * protocol locks before interface locks, after UDP locks.
528 	 */
529 	{ "udpinp", &lock_class_rw },
530 	{ "in6_multi_mtx", &lock_class_mtx_sleep },
531 	{ "mld_mtx", &lock_class_mtx_sleep },
532 	{ "if_addr_lock", &lock_class_rw },
533 	{ NULL, NULL },
534 	/*
535 	 * UNIX Domain Sockets
536 	 */
537 	{ "unp_global_rwlock", &lock_class_rw },
538 	{ "unp_list_lock", &lock_class_mtx_sleep },
539 	{ "unp", &lock_class_mtx_sleep },
540 	{ "so_snd", &lock_class_mtx_sleep },
541 	{ NULL, NULL },
542 	/*
543 	 * UDP/IP
544 	 */
545 	{ "udp", &lock_class_rw },
546 	{ "udpinp", &lock_class_rw },
547 	{ "so_snd", &lock_class_mtx_sleep },
548 	{ NULL, NULL },
549 	/*
550 	 * TCP/IP
551 	 */
552 	{ "tcp", &lock_class_rw },
553 	{ "tcpinp", &lock_class_rw },
554 	{ "so_snd", &lock_class_mtx_sleep },
555 	{ NULL, NULL },
556 	/*
557 	 * netatalk
558 	 */
559 	{ "ddp_list_mtx", &lock_class_mtx_sleep },
560 	{ "ddp_mtx", &lock_class_mtx_sleep },
561 	{ NULL, NULL },
562 	/*
563 	 * BPF
564 	 */
565 	{ "bpf global lock", &lock_class_mtx_sleep },
566 	{ "bpf interface lock", &lock_class_rw },
567 	{ "bpf cdev lock", &lock_class_mtx_sleep },
568 	{ NULL, NULL },
569 	/*
570 	 * NFS server
571 	 */
572 	{ "nfsd_mtx", &lock_class_mtx_sleep },
573 	{ "so_snd", &lock_class_mtx_sleep },
574 	{ NULL, NULL },
575 
576 	/*
577 	 * IEEE 802.11
578 	 */
579 	{ "802.11 com lock", &lock_class_mtx_sleep},
580 	{ NULL, NULL },
581 	/*
582 	 * Network drivers
583 	 */
584 	{ "network driver", &lock_class_mtx_sleep},
585 	{ NULL, NULL },
586 
587 	/*
588 	 * Netgraph
589 	 */
590 	{ "ng_node", &lock_class_mtx_sleep },
591 	{ "ng_worklist", &lock_class_mtx_sleep },
592 	{ NULL, NULL },
593 	/*
594 	 * CDEV
595 	 */
596 	{ "vm map (system)", &lock_class_mtx_sleep },
597 	{ "vm page queue", &lock_class_mtx_sleep },
598 	{ "vnode interlock", &lock_class_mtx_sleep },
599 	{ "cdev", &lock_class_mtx_sleep },
600 	{ NULL, NULL },
601 	/*
602 	 * VM
603 	 */
604 	{ "vm map (user)", &lock_class_sx },
605 	{ "vm object", &lock_class_rw },
606 	{ "vm page", &lock_class_mtx_sleep },
607 	{ "vm page queue", &lock_class_mtx_sleep },
608 	{ "pmap pv global", &lock_class_rw },
609 	{ "pmap", &lock_class_mtx_sleep },
610 	{ "pmap pv list", &lock_class_rw },
611 	{ "vm page free queue", &lock_class_mtx_sleep },
612 	{ NULL, NULL },
613 	/*
614 	 * kqueue/VFS interaction
615 	 */
616 	{ "kqueue", &lock_class_mtx_sleep },
617 	{ "struct mount mtx", &lock_class_mtx_sleep },
618 	{ "vnode interlock", &lock_class_mtx_sleep },
619 	{ NULL, NULL },
620 	/*
621 	 * ZFS locking
622 	 */
623 	{ "dn->dn_mtx", &lock_class_sx },
624 	{ "dr->dt.di.dr_mtx", &lock_class_sx },
625 	{ "db->db_mtx", &lock_class_sx },
626 	{ NULL, NULL },
627 	/*
628 	 * spin locks
629 	 */
630 #ifdef SMP
631 	{ "ap boot", &lock_class_mtx_spin },
632 #endif
633 	{ "rm.mutex_mtx", &lock_class_mtx_spin },
634 	{ "sio", &lock_class_mtx_spin },
635 	{ "scrlock", &lock_class_mtx_spin },
636 #ifdef __i386__
637 	{ "cy", &lock_class_mtx_spin },
638 #endif
639 #ifdef __sparc64__
640 	{ "pcib_mtx", &lock_class_mtx_spin },
641 	{ "rtc_mtx", &lock_class_mtx_spin },
642 #endif
643 	{ "scc_hwmtx", &lock_class_mtx_spin },
644 	{ "uart_hwmtx", &lock_class_mtx_spin },
645 	{ "fast_taskqueue", &lock_class_mtx_spin },
646 	{ "intr table", &lock_class_mtx_spin },
647 #ifdef	HWPMC_HOOKS
648 	{ "pmc-per-proc", &lock_class_mtx_spin },
649 #endif
650 	{ "process slock", &lock_class_mtx_spin },
651 	{ "sleepq chain", &lock_class_mtx_spin },
652 	{ "umtx lock", &lock_class_mtx_spin },
653 	{ "rm_spinlock", &lock_class_mtx_spin },
654 	{ "turnstile chain", &lock_class_mtx_spin },
655 	{ "turnstile lock", &lock_class_mtx_spin },
656 	{ "sched lock", &lock_class_mtx_spin },
657 	{ "td_contested", &lock_class_mtx_spin },
658 	{ "callout", &lock_class_mtx_spin },
659 	{ "entropy harvest mutex", &lock_class_mtx_spin },
660 	{ "syscons video lock", &lock_class_mtx_spin },
661 #ifdef SMP
662 	{ "smp rendezvous", &lock_class_mtx_spin },
663 #endif
664 #ifdef __powerpc__
665 	{ "tlb0", &lock_class_mtx_spin },
666 #endif
667 	/*
668 	 * leaf locks
669 	 */
670 	{ "intrcnt", &lock_class_mtx_spin },
671 	{ "icu", &lock_class_mtx_spin },
672 #ifdef __i386__
673 	{ "allpmaps", &lock_class_mtx_spin },
674 	{ "descriptor tables", &lock_class_mtx_spin },
675 #endif
676 	{ "clk", &lock_class_mtx_spin },
677 	{ "cpuset", &lock_class_mtx_spin },
678 	{ "mprof lock", &lock_class_mtx_spin },
679 	{ "zombie lock", &lock_class_mtx_spin },
680 	{ "ALD Queue", &lock_class_mtx_spin },
681 #ifdef __ia64__
682 	{ "MCA spin lock", &lock_class_mtx_spin },
683 #endif
684 #if defined(__i386__) || defined(__amd64__)
685 	{ "pcicfg", &lock_class_mtx_spin },
686 	{ "NDIS thread lock", &lock_class_mtx_spin },
687 #endif
688 	{ "tw_osl_io_lock", &lock_class_mtx_spin },
689 	{ "tw_osl_q_lock", &lock_class_mtx_spin },
690 	{ "tw_cl_io_lock", &lock_class_mtx_spin },
691 	{ "tw_cl_intr_lock", &lock_class_mtx_spin },
692 	{ "tw_cl_gen_lock", &lock_class_mtx_spin },
693 #ifdef	HWPMC_HOOKS
694 	{ "pmc-leaf", &lock_class_mtx_spin },
695 #endif
696 	{ "blocked lock", &lock_class_mtx_spin },
697 	{ NULL, NULL },
698 	{ NULL, NULL }
699 };
700 
701 #ifdef BLESSING
702 /*
703  * Pairs of locks which have been blessed
704  * Don't complain about order problems with blessed locks
705  */
706 static struct witness_blessed blessed_list[] = {
707 };
708 static int blessed_count =
709 	sizeof(blessed_list) / sizeof(struct witness_blessed);
710 #endif
711 
712 /*
713  * This global is set to 0 once it becomes safe to use the witness code.
714  */
715 static int witness_cold = 1;
716 
717 /*
718  * This global is set to 1 once the static lock orders have been enrolled
719  * so that a warning can be issued for any spin locks enrolled later.
720  */
721 static int witness_spin_warn = 0;
722 
723 /* Trim useless garbage from filenames. */
724 static const char *
725 fixup_filename(const char *file)
726 {
727 
728 	if (file == NULL)
729 		return (NULL);
730 	while (strncmp(file, "../", 3) == 0)
731 		file += 3;
732 	return (file);
733 }
734 
735 /*
736  * The WITNESS-enabled diagnostic code.  Note that the witness code does
737  * assume that the early boot is single-threaded at least until after this
738  * routine is completed.
739  */
740 static void
741 witness_initialize(void *dummy __unused)
742 {
743 	struct lock_object *lock;
744 	struct witness_order_list_entry *order;
745 	struct witness *w, *w1;
746 	int i;
747 
748 	w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
749 	    M_NOWAIT | M_ZERO);
750 
751 	/*
752 	 * We have to release Giant before initializing its witness
753 	 * structure so that WITNESS doesn't get confused.
754 	 */
755 	mtx_unlock(&Giant);
756 	mtx_assert(&Giant, MA_NOTOWNED);
757 
758 	CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
759 	mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
760 	    MTX_NOWITNESS | MTX_NOPROFILE);
761 	for (i = WITNESS_COUNT - 1; i >= 0; i--) {
762 		w = &w_data[i];
763 		memset(w, 0, sizeof(*w));
764 		w_data[i].w_index = i;	/* Witness index never changes. */
765 		witness_free(w);
766 	}
767 	KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
768 	    ("%s: Invalid list of free witness objects", __func__));
769 
770 	/* Witness with index 0 is not used to aid in debugging. */
771 	STAILQ_REMOVE_HEAD(&w_free, w_list);
772 	w_free_cnt--;
773 
774 	memset(w_rmatrix, 0,
775 	    (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
776 
777 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
778 		witness_lock_list_free(&w_locklistdata[i]);
779 	witness_init_hash_tables();
780 
781 	/* First add in all the specified order lists. */
782 	for (order = order_lists; order->w_name != NULL; order++) {
783 		w = enroll(order->w_name, order->w_class);
784 		if (w == NULL)
785 			continue;
786 		w->w_file = "order list";
787 		for (order++; order->w_name != NULL; order++) {
788 			w1 = enroll(order->w_name, order->w_class);
789 			if (w1 == NULL)
790 				continue;
791 			w1->w_file = "order list";
792 			itismychild(w, w1);
793 			w = w1;
794 		}
795 	}
796 	witness_spin_warn = 1;
797 
798 	/* Iterate through all locks and add them to witness. */
799 	for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
800 		lock = pending_locks[i].wh_lock;
801 		KASSERT(lock->lo_flags & LO_WITNESS,
802 		    ("%s: lock %s is on pending list but not LO_WITNESS",
803 		    __func__, lock->lo_name));
804 		lock->lo_witness = enroll(pending_locks[i].wh_type,
805 		    LOCK_CLASS(lock));
806 	}
807 
808 	/* Mark the witness code as being ready for use. */
809 	witness_cold = 0;
810 
811 	mtx_lock(&Giant);
812 }
813 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
814     NULL);
815 
816 void
817 witness_init(struct lock_object *lock, const char *type)
818 {
819 	struct lock_class *class;
820 
821 	/* Various sanity checks. */
822 	class = LOCK_CLASS(lock);
823 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
824 	    (class->lc_flags & LC_RECURSABLE) == 0)
825 		kassert_panic("%s: lock (%s) %s can not be recursable",
826 		    __func__, class->lc_name, lock->lo_name);
827 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
828 	    (class->lc_flags & LC_SLEEPABLE) == 0)
829 		kassert_panic("%s: lock (%s) %s can not be sleepable",
830 		    __func__, class->lc_name, lock->lo_name);
831 	if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
832 	    (class->lc_flags & LC_UPGRADABLE) == 0)
833 		kassert_panic("%s: lock (%s) %s can not be upgradable",
834 		    __func__, class->lc_name, lock->lo_name);
835 
836 	/*
837 	 * If we shouldn't watch this lock, then just clear lo_witness.
838 	 * Otherwise, if witness_cold is set, then it is too early to
839 	 * enroll this lock, so defer it to witness_initialize() by adding
840 	 * it to the pending_locks list.  If it is not too early, then enroll
841 	 * the lock now.
842 	 */
843 	if (witness_watch < 1 || panicstr != NULL ||
844 	    (lock->lo_flags & LO_WITNESS) == 0)
845 		lock->lo_witness = NULL;
846 	else if (witness_cold) {
847 		pending_locks[pending_cnt].wh_lock = lock;
848 		pending_locks[pending_cnt++].wh_type = type;
849 		if (pending_cnt > WITNESS_PENDLIST)
850 			panic("%s: pending locks list is too small, "
851 			    "increase WITNESS_PENDLIST\n",
852 			    __func__);
853 	} else
854 		lock->lo_witness = enroll(type, class);
855 }
856 
857 void
858 witness_destroy(struct lock_object *lock)
859 {
860 	struct lock_class *class;
861 	struct witness *w;
862 
863 	class = LOCK_CLASS(lock);
864 
865 	if (witness_cold)
866 		panic("lock (%s) %s destroyed while witness_cold",
867 		    class->lc_name, lock->lo_name);
868 
869 	/* XXX: need to verify that no one holds the lock */
870 	if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
871 		return;
872 	w = lock->lo_witness;
873 
874 	mtx_lock_spin(&w_mtx);
875 	MPASS(w->w_refcount > 0);
876 	w->w_refcount--;
877 
878 	if (w->w_refcount == 0)
879 		depart(w);
880 	mtx_unlock_spin(&w_mtx);
881 }
882 
883 #ifdef DDB
884 static void
885 witness_ddb_compute_levels(void)
886 {
887 	struct witness *w;
888 
889 	/*
890 	 * First clear all levels.
891 	 */
892 	STAILQ_FOREACH(w, &w_all, w_list)
893 		w->w_ddb_level = -1;
894 
895 	/*
896 	 * Look for locks with no parents and level all their descendants.
897 	 */
898 	STAILQ_FOREACH(w, &w_all, w_list) {
899 
900 		/* If the witness has ancestors (is not a root), skip it. */
901 		if (w->w_num_ancestors > 0)
902 			continue;
903 		witness_ddb_level_descendants(w, 0);
904 	}
905 }
906 
907 static void
908 witness_ddb_level_descendants(struct witness *w, int l)
909 {
910 	int i;
911 
912 	if (w->w_ddb_level >= l)
913 		return;
914 
915 	w->w_ddb_level = l;
916 	l++;
917 
918 	for (i = 1; i <= w_max_used_index; i++) {
919 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
920 			witness_ddb_level_descendants(&w_data[i], l);
921 	}
922 }
923 
924 static void
925 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
926     struct witness *w, int indent)
927 {
928 	int i;
929 
930  	for (i = 0; i < indent; i++)
931  		prnt(" ");
932 	prnt("%s (type: %s, depth: %d, active refs: %d)",
933 	     w->w_name, w->w_class->lc_name,
934 	     w->w_ddb_level, w->w_refcount);
935  	if (w->w_displayed) {
936  		prnt(" -- (already displayed)\n");
937  		return;
938  	}
939  	w->w_displayed = 1;
940 	if (w->w_file != NULL && w->w_line != 0)
941 		prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
942 		    w->w_line);
943 	else
944 		prnt(" -- never acquired\n");
945 	indent++;
946 	WITNESS_INDEX_ASSERT(w->w_index);
947 	for (i = 1; i <= w_max_used_index; i++) {
948 		if (db_pager_quit)
949 			return;
950 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
951 			witness_ddb_display_descendants(prnt, &w_data[i],
952 			    indent);
953 	}
954 }
955 
956 static void
957 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
958     struct witness_list *list)
959 {
960 	struct witness *w;
961 
962 	STAILQ_FOREACH(w, list, w_typelist) {
963 		if (w->w_file == NULL || w->w_ddb_level > 0)
964 			continue;
965 
966 		/* This lock has no anscestors - display its descendants. */
967 		witness_ddb_display_descendants(prnt, w, 0);
968 		if (db_pager_quit)
969 			return;
970 	}
971 }
972 
973 static void
974 witness_ddb_display(int(*prnt)(const char *fmt, ...))
975 {
976 	struct witness *w;
977 
978 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
979 	witness_ddb_compute_levels();
980 
981 	/* Clear all the displayed flags. */
982 	STAILQ_FOREACH(w, &w_all, w_list)
983 		w->w_displayed = 0;
984 
985 	/*
986 	 * First, handle sleep locks which have been acquired at least
987 	 * once.
988 	 */
989 	prnt("Sleep locks:\n");
990 	witness_ddb_display_list(prnt, &w_sleep);
991 	if (db_pager_quit)
992 		return;
993 
994 	/*
995 	 * Now do spin locks which have been acquired at least once.
996 	 */
997 	prnt("\nSpin locks:\n");
998 	witness_ddb_display_list(prnt, &w_spin);
999 	if (db_pager_quit)
1000 		return;
1001 
1002 	/*
1003 	 * Finally, any locks which have not been acquired yet.
1004 	 */
1005 	prnt("\nLocks which were never acquired:\n");
1006 	STAILQ_FOREACH(w, &w_all, w_list) {
1007 		if (w->w_file != NULL || w->w_refcount == 0)
1008 			continue;
1009 		prnt("%s (type: %s, depth: %d)\n", w->w_name,
1010 		    w->w_class->lc_name, w->w_ddb_level);
1011 		if (db_pager_quit)
1012 			return;
1013 	}
1014 }
1015 #endif /* DDB */
1016 
1017 int
1018 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1019 {
1020 
1021 	if (witness_watch == -1 || panicstr != NULL)
1022 		return (0);
1023 
1024 	/* Require locks that witness knows about. */
1025 	if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1026 	    lock2->lo_witness == NULL)
1027 		return (EINVAL);
1028 
1029 	mtx_assert(&w_mtx, MA_NOTOWNED);
1030 	mtx_lock_spin(&w_mtx);
1031 
1032 	/*
1033 	 * If we already have either an explicit or implied lock order that
1034 	 * is the other way around, then return an error.
1035 	 */
1036 	if (witness_watch &&
1037 	    isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1038 		mtx_unlock_spin(&w_mtx);
1039 		return (EDOOFUS);
1040 	}
1041 
1042 	/* Try to add the new order. */
1043 	CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1044 	    lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1045 	itismychild(lock1->lo_witness, lock2->lo_witness);
1046 	mtx_unlock_spin(&w_mtx);
1047 	return (0);
1048 }
1049 
1050 void
1051 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1052     int line, struct lock_object *interlock)
1053 {
1054 	struct lock_list_entry *lock_list, *lle;
1055 	struct lock_instance *lock1, *lock2, *plock;
1056 	struct lock_class *class;
1057 	struct witness *w, *w1;
1058 	struct thread *td;
1059 	int i, j;
1060 
1061 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1062 	    panicstr != NULL)
1063 		return;
1064 
1065 	w = lock->lo_witness;
1066 	class = LOCK_CLASS(lock);
1067 	td = curthread;
1068 
1069 	if (class->lc_flags & LC_SLEEPLOCK) {
1070 
1071 		/*
1072 		 * Since spin locks include a critical section, this check
1073 		 * implicitly enforces a lock order of all sleep locks before
1074 		 * all spin locks.
1075 		 */
1076 		if (td->td_critnest != 0 && !kdb_active)
1077 			kassert_panic("acquiring blockable sleep lock with "
1078 			    "spinlock or critical section held (%s) %s @ %s:%d",
1079 			    class->lc_name, lock->lo_name,
1080 			    fixup_filename(file), line);
1081 
1082 		/*
1083 		 * If this is the first lock acquired then just return as
1084 		 * no order checking is needed.
1085 		 */
1086 		lock_list = td->td_sleeplocks;
1087 		if (lock_list == NULL || lock_list->ll_count == 0)
1088 			return;
1089 	} else {
1090 
1091 		/*
1092 		 * If this is the first lock, just return as no order
1093 		 * checking is needed.  Avoid problems with thread
1094 		 * migration pinning the thread while checking if
1095 		 * spinlocks are held.  If at least one spinlock is held
1096 		 * the thread is in a safe path and it is allowed to
1097 		 * unpin it.
1098 		 */
1099 		sched_pin();
1100 		lock_list = PCPU_GET(spinlocks);
1101 		if (lock_list == NULL || lock_list->ll_count == 0) {
1102 			sched_unpin();
1103 			return;
1104 		}
1105 		sched_unpin();
1106 	}
1107 
1108 	/*
1109 	 * Check to see if we are recursing on a lock we already own.  If
1110 	 * so, make sure that we don't mismatch exclusive and shared lock
1111 	 * acquires.
1112 	 */
1113 	lock1 = find_instance(lock_list, lock);
1114 	if (lock1 != NULL) {
1115 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1116 		    (flags & LOP_EXCLUSIVE) == 0) {
1117 			printf("shared lock of (%s) %s @ %s:%d\n",
1118 			    class->lc_name, lock->lo_name,
1119 			    fixup_filename(file), line);
1120 			printf("while exclusively locked from %s:%d\n",
1121 			    fixup_filename(lock1->li_file), lock1->li_line);
1122 			kassert_panic("share->excl");
1123 		}
1124 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1125 		    (flags & LOP_EXCLUSIVE) != 0) {
1126 			printf("exclusive lock of (%s) %s @ %s:%d\n",
1127 			    class->lc_name, lock->lo_name,
1128 			    fixup_filename(file), line);
1129 			printf("while share locked from %s:%d\n",
1130 			    fixup_filename(lock1->li_file), lock1->li_line);
1131 			kassert_panic("excl->share");
1132 		}
1133 		return;
1134 	}
1135 
1136 	/*
1137 	 * Find the previously acquired lock, but ignore interlocks.
1138 	 */
1139 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1140 	if (interlock != NULL && plock->li_lock == interlock) {
1141 		if (lock_list->ll_count > 1)
1142 			plock =
1143 			    &lock_list->ll_children[lock_list->ll_count - 2];
1144 		else {
1145 			lle = lock_list->ll_next;
1146 
1147 			/*
1148 			 * The interlock is the only lock we hold, so
1149 			 * simply return.
1150 			 */
1151 			if (lle == NULL)
1152 				return;
1153 			plock = &lle->ll_children[lle->ll_count - 1];
1154 		}
1155 	}
1156 
1157 	/*
1158 	 * Try to perform most checks without a lock.  If this succeeds we
1159 	 * can skip acquiring the lock and return success.
1160 	 */
1161 	w1 = plock->li_lock->lo_witness;
1162 	if (witness_lock_order_check(w1, w))
1163 		return;
1164 
1165 	/*
1166 	 * Check for duplicate locks of the same type.  Note that we only
1167 	 * have to check for this on the last lock we just acquired.  Any
1168 	 * other cases will be caught as lock order violations.
1169 	 */
1170 	mtx_lock_spin(&w_mtx);
1171 	witness_lock_order_add(w1, w);
1172 	if (w1 == w) {
1173 		i = w->w_index;
1174 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1175 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1176 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1177 			w->w_reversed = 1;
1178 			mtx_unlock_spin(&w_mtx);
1179 			printf(
1180 			    "acquiring duplicate lock of same type: \"%s\"\n",
1181 			    w->w_name);
1182 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1183 			    fixup_filename(plock->li_file), plock->li_line);
1184 			printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1185 			    fixup_filename(file), line);
1186 			witness_debugger(1);
1187 		} else
1188 			mtx_unlock_spin(&w_mtx);
1189 		return;
1190 	}
1191 	mtx_assert(&w_mtx, MA_OWNED);
1192 
1193 	/*
1194 	 * If we know that the lock we are acquiring comes after
1195 	 * the lock we most recently acquired in the lock order tree,
1196 	 * then there is no need for any further checks.
1197 	 */
1198 	if (isitmychild(w1, w))
1199 		goto out;
1200 
1201 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1202 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1203 
1204 			MPASS(j < WITNESS_COUNT);
1205 			lock1 = &lle->ll_children[i];
1206 
1207 			/*
1208 			 * Ignore the interlock the first time we see it.
1209 			 */
1210 			if (interlock != NULL && interlock == lock1->li_lock) {
1211 				interlock = NULL;
1212 				continue;
1213 			}
1214 
1215 			/*
1216 			 * If this lock doesn't undergo witness checking,
1217 			 * then skip it.
1218 			 */
1219 			w1 = lock1->li_lock->lo_witness;
1220 			if (w1 == NULL) {
1221 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1222 				    ("lock missing witness structure"));
1223 				continue;
1224 			}
1225 
1226 			/*
1227 			 * If we are locking Giant and this is a sleepable
1228 			 * lock, then skip it.
1229 			 */
1230 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1231 			    lock == &Giant.lock_object)
1232 				continue;
1233 
1234 			/*
1235 			 * If we are locking a sleepable lock and this lock
1236 			 * is Giant, then skip it.
1237 			 */
1238 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1239 			    lock1->li_lock == &Giant.lock_object)
1240 				continue;
1241 
1242 			/*
1243 			 * If we are locking a sleepable lock and this lock
1244 			 * isn't sleepable, we want to treat it as a lock
1245 			 * order violation to enfore a general lock order of
1246 			 * sleepable locks before non-sleepable locks.
1247 			 */
1248 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1249 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1250 				goto reversal;
1251 
1252 			/*
1253 			 * If we are locking Giant and this is a non-sleepable
1254 			 * lock, then treat it as a reversal.
1255 			 */
1256 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1257 			    lock == &Giant.lock_object)
1258 				goto reversal;
1259 
1260 			/*
1261 			 * Check the lock order hierarchy for a reveresal.
1262 			 */
1263 			if (!isitmydescendant(w, w1))
1264 				continue;
1265 		reversal:
1266 
1267 			/*
1268 			 * We have a lock order violation, check to see if it
1269 			 * is allowed or has already been yelled about.
1270 			 */
1271 #ifdef BLESSING
1272 
1273 			/*
1274 			 * If the lock order is blessed, just bail.  We don't
1275 			 * look for other lock order violations though, which
1276 			 * may be a bug.
1277 			 */
1278 			if (blessed(w, w1))
1279 				goto out;
1280 #endif
1281 
1282 			/* Bail if this violation is known */
1283 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1284 				goto out;
1285 
1286 			/* Record this as a violation */
1287 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1288 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1289 			w->w_reversed = w1->w_reversed = 1;
1290 			witness_increment_graph_generation();
1291 			mtx_unlock_spin(&w_mtx);
1292 
1293 			/*
1294 			 * Ok, yell about it.
1295 			 */
1296 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1297 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1298 				printf(
1299 		"lock order reversal: (sleepable after non-sleepable)\n");
1300 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1301 			    && lock == &Giant.lock_object)
1302 				printf(
1303 		"lock order reversal: (Giant after non-sleepable)\n");
1304 			else
1305 				printf("lock order reversal:\n");
1306 
1307 			/*
1308 			 * Try to locate an earlier lock with
1309 			 * witness w in our list.
1310 			 */
1311 			do {
1312 				lock2 = &lle->ll_children[i];
1313 				MPASS(lock2->li_lock != NULL);
1314 				if (lock2->li_lock->lo_witness == w)
1315 					break;
1316 				if (i == 0 && lle->ll_next != NULL) {
1317 					lle = lle->ll_next;
1318 					i = lle->ll_count - 1;
1319 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1320 				} else
1321 					i--;
1322 			} while (i >= 0);
1323 			if (i < 0) {
1324 				printf(" 1st %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(" 2nd %p %s (%s) @ %s:%d\n", lock,
1329 				    lock->lo_name, w->w_name,
1330 				    fixup_filename(file), line);
1331 			} else {
1332 				printf(" 1st %p %s (%s) @ %s:%d\n",
1333 				    lock2->li_lock, lock2->li_lock->lo_name,
1334 				    lock2->li_lock->lo_witness->w_name,
1335 				    fixup_filename(lock2->li_file),
1336 				    lock2->li_line);
1337 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1338 				    lock1->li_lock, lock1->li_lock->lo_name,
1339 				    w1->w_name, fixup_filename(lock1->li_file),
1340 				    lock1->li_line);
1341 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1342 				    lock->lo_name, w->w_name,
1343 				    fixup_filename(file), line);
1344 			}
1345 			witness_debugger(1);
1346 			return;
1347 		}
1348 	}
1349 
1350 	/*
1351 	 * If requested, build a new lock order.  However, don't build a new
1352 	 * relationship between a sleepable lock and Giant if it is in the
1353 	 * wrong direction.  The correct lock order is that sleepable locks
1354 	 * always come before Giant.
1355 	 */
1356 	if (flags & LOP_NEWORDER &&
1357 	    !(plock->li_lock == &Giant.lock_object &&
1358 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1359 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1360 		    w->w_name, plock->li_lock->lo_witness->w_name);
1361 		itismychild(plock->li_lock->lo_witness, w);
1362 	}
1363 out:
1364 	mtx_unlock_spin(&w_mtx);
1365 }
1366 
1367 void
1368 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1369 {
1370 	struct lock_list_entry **lock_list, *lle;
1371 	struct lock_instance *instance;
1372 	struct witness *w;
1373 	struct thread *td;
1374 
1375 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1376 	    panicstr != NULL)
1377 		return;
1378 	w = lock->lo_witness;
1379 	td = curthread;
1380 
1381 	/* Determine lock list for this lock. */
1382 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1383 		lock_list = &td->td_sleeplocks;
1384 	else
1385 		lock_list = PCPU_PTR(spinlocks);
1386 
1387 	/* Check to see if we are recursing on a lock we already own. */
1388 	instance = find_instance(*lock_list, lock);
1389 	if (instance != NULL) {
1390 		instance->li_flags++;
1391 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1392 		    td->td_proc->p_pid, lock->lo_name,
1393 		    instance->li_flags & LI_RECURSEMASK);
1394 		instance->li_file = file;
1395 		instance->li_line = line;
1396 		return;
1397 	}
1398 
1399 	/* Update per-witness last file and line acquire. */
1400 	w->w_file = file;
1401 	w->w_line = line;
1402 
1403 	/* Find the next open lock instance in the list and fill it. */
1404 	lle = *lock_list;
1405 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1406 		lle = witness_lock_list_get();
1407 		if (lle == NULL)
1408 			return;
1409 		lle->ll_next = *lock_list;
1410 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1411 		    td->td_proc->p_pid, lle);
1412 		*lock_list = lle;
1413 	}
1414 	instance = &lle->ll_children[lle->ll_count++];
1415 	instance->li_lock = lock;
1416 	instance->li_line = line;
1417 	instance->li_file = file;
1418 	if ((flags & LOP_EXCLUSIVE) != 0)
1419 		instance->li_flags = LI_EXCLUSIVE;
1420 	else
1421 		instance->li_flags = 0;
1422 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1423 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1424 }
1425 
1426 void
1427 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1428 {
1429 	struct lock_instance *instance;
1430 	struct lock_class *class;
1431 
1432 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1433 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1434 		return;
1435 	class = LOCK_CLASS(lock);
1436 	if (witness_watch) {
1437 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1438 			kassert_panic(
1439 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1440 			    class->lc_name, lock->lo_name,
1441 			    fixup_filename(file), line);
1442 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1443 			kassert_panic(
1444 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1445 			    class->lc_name, lock->lo_name,
1446 			    fixup_filename(file), line);
1447 	}
1448 	instance = find_instance(curthread->td_sleeplocks, lock);
1449 	if (instance == NULL) {
1450 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1451 		    class->lc_name, lock->lo_name,
1452 		    fixup_filename(file), line);
1453 		return;
1454 	}
1455 	if (witness_watch) {
1456 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1457 			kassert_panic(
1458 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1459 			    class->lc_name, lock->lo_name,
1460 			    fixup_filename(file), line);
1461 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1462 			kassert_panic(
1463 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1464 			    class->lc_name, lock->lo_name,
1465 			    instance->li_flags & LI_RECURSEMASK,
1466 			    fixup_filename(file), line);
1467 	}
1468 	instance->li_flags |= LI_EXCLUSIVE;
1469 }
1470 
1471 void
1472 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1473     int line)
1474 {
1475 	struct lock_instance *instance;
1476 	struct lock_class *class;
1477 
1478 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1479 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1480 		return;
1481 	class = LOCK_CLASS(lock);
1482 	if (witness_watch) {
1483 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1484 			kassert_panic(
1485 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1486 			    class->lc_name, lock->lo_name,
1487 			    fixup_filename(file), line);
1488 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1489 			kassert_panic(
1490 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1491 			    class->lc_name, lock->lo_name,
1492 			    fixup_filename(file), line);
1493 	}
1494 	instance = find_instance(curthread->td_sleeplocks, lock);
1495 	if (instance == NULL) {
1496 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1497 		    class->lc_name, lock->lo_name,
1498 		    fixup_filename(file), line);
1499 		return;
1500 	}
1501 	if (witness_watch) {
1502 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1503 			kassert_panic(
1504 			    "downgrade of shared lock (%s) %s @ %s:%d",
1505 			    class->lc_name, lock->lo_name,
1506 			    fixup_filename(file), line);
1507 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1508 			kassert_panic(
1509 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1510 			    class->lc_name, lock->lo_name,
1511 			    instance->li_flags & LI_RECURSEMASK,
1512 			    fixup_filename(file), line);
1513 	}
1514 	instance->li_flags &= ~LI_EXCLUSIVE;
1515 }
1516 
1517 void
1518 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1519 {
1520 	struct lock_list_entry **lock_list, *lle;
1521 	struct lock_instance *instance;
1522 	struct lock_class *class;
1523 	struct thread *td;
1524 	register_t s;
1525 	int i, j;
1526 
1527 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1528 		return;
1529 	td = curthread;
1530 	class = LOCK_CLASS(lock);
1531 
1532 	/* Find lock instance associated with this lock. */
1533 	if (class->lc_flags & LC_SLEEPLOCK)
1534 		lock_list = &td->td_sleeplocks;
1535 	else
1536 		lock_list = PCPU_PTR(spinlocks);
1537 	lle = *lock_list;
1538 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1539 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1540 			instance = &(*lock_list)->ll_children[i];
1541 			if (instance->li_lock == lock)
1542 				goto found;
1543 		}
1544 
1545 	/*
1546 	 * When disabling WITNESS through witness_watch we could end up in
1547 	 * having registered locks in the td_sleeplocks queue.
1548 	 * We have to make sure we flush these queues, so just search for
1549 	 * eventual register locks and remove them.
1550 	 */
1551 	if (witness_watch > 0) {
1552 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1553 		    lock->lo_name, fixup_filename(file), line);
1554 		return;
1555 	} else {
1556 		return;
1557 	}
1558 found:
1559 
1560 	/* First, check for shared/exclusive mismatches. */
1561 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1562 	    (flags & LOP_EXCLUSIVE) == 0) {
1563 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1564 		    lock->lo_name, fixup_filename(file), line);
1565 		printf("while exclusively locked from %s:%d\n",
1566 		    fixup_filename(instance->li_file), instance->li_line);
1567 		kassert_panic("excl->ushare");
1568 	}
1569 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1570 	    (flags & LOP_EXCLUSIVE) != 0) {
1571 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1572 		    lock->lo_name, fixup_filename(file), line);
1573 		printf("while share locked from %s:%d\n",
1574 		    fixup_filename(instance->li_file),
1575 		    instance->li_line);
1576 		kassert_panic("share->uexcl");
1577 	}
1578 	/* If we are recursed, unrecurse. */
1579 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1580 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1581 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1582 		    instance->li_flags);
1583 		instance->li_flags--;
1584 		return;
1585 	}
1586 	/* The lock is now being dropped, check for NORELEASE flag */
1587 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1588 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1589 		    lock->lo_name, fixup_filename(file), line);
1590 		kassert_panic("lock marked norelease");
1591 	}
1592 
1593 	/* Otherwise, remove this item from the list. */
1594 	s = intr_disable();
1595 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1596 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1597 	    (*lock_list)->ll_count - 1);
1598 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1599 		(*lock_list)->ll_children[j] =
1600 		    (*lock_list)->ll_children[j + 1];
1601 	(*lock_list)->ll_count--;
1602 	intr_restore(s);
1603 
1604 	/*
1605 	 * In order to reduce contention on w_mtx, we want to keep always an
1606 	 * head object into lists so that frequent allocation from the
1607 	 * free witness pool (and subsequent locking) is avoided.
1608 	 * In order to maintain the current code simple, when the head
1609 	 * object is totally unloaded it means also that we do not have
1610 	 * further objects in the list, so the list ownership needs to be
1611 	 * hand over to another object if the current head needs to be freed.
1612 	 */
1613 	if ((*lock_list)->ll_count == 0) {
1614 		if (*lock_list == lle) {
1615 			if (lle->ll_next == NULL)
1616 				return;
1617 		} else
1618 			lle = *lock_list;
1619 		*lock_list = lle->ll_next;
1620 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1621 		    td->td_proc->p_pid, lle);
1622 		witness_lock_list_free(lle);
1623 	}
1624 }
1625 
1626 void
1627 witness_thread_exit(struct thread *td)
1628 {
1629 	struct lock_list_entry *lle;
1630 	int i, n;
1631 
1632 	lle = td->td_sleeplocks;
1633 	if (lle == NULL || panicstr != NULL)
1634 		return;
1635 	if (lle->ll_count != 0) {
1636 		for (n = 0; lle != NULL; lle = lle->ll_next)
1637 			for (i = lle->ll_count - 1; i >= 0; i--) {
1638 				if (n == 0)
1639 		printf("Thread %p exiting with the following locks held:\n",
1640 					    td);
1641 				n++;
1642 				witness_list_lock(&lle->ll_children[i], printf);
1643 
1644 			}
1645 		kassert_panic(
1646 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1647 	}
1648 	witness_lock_list_free(lle);
1649 }
1650 
1651 /*
1652  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1653  * exempt Giant and sleepable locks from the checks as well.  If any
1654  * non-exempt locks are held, then a supplied message is printed to the
1655  * console along with a list of the offending locks.  If indicated in the
1656  * flags then a failure results in a panic as well.
1657  */
1658 int
1659 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1660 {
1661 	struct lock_list_entry *lock_list, *lle;
1662 	struct lock_instance *lock1;
1663 	struct thread *td;
1664 	va_list ap;
1665 	int i, n;
1666 
1667 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1668 		return (0);
1669 	n = 0;
1670 	td = curthread;
1671 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1672 		for (i = lle->ll_count - 1; i >= 0; i--) {
1673 			lock1 = &lle->ll_children[i];
1674 			if (lock1->li_lock == lock)
1675 				continue;
1676 			if (flags & WARN_GIANTOK &&
1677 			    lock1->li_lock == &Giant.lock_object)
1678 				continue;
1679 			if (flags & WARN_SLEEPOK &&
1680 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1681 				continue;
1682 			if (n == 0) {
1683 				va_start(ap, fmt);
1684 				vprintf(fmt, ap);
1685 				va_end(ap);
1686 				printf(" with the following");
1687 				if (flags & WARN_SLEEPOK)
1688 					printf(" non-sleepable");
1689 				printf(" locks held:\n");
1690 			}
1691 			n++;
1692 			witness_list_lock(lock1, printf);
1693 		}
1694 
1695 	/*
1696 	 * Pin the thread in order to avoid problems with thread migration.
1697 	 * Once that all verifies are passed about spinlocks ownership,
1698 	 * the thread is in a safe path and it can be unpinned.
1699 	 */
1700 	sched_pin();
1701 	lock_list = PCPU_GET(spinlocks);
1702 	if (lock_list != NULL && lock_list->ll_count != 0) {
1703 		sched_unpin();
1704 
1705 		/*
1706 		 * We should only have one spinlock and as long as
1707 		 * the flags cannot match for this locks class,
1708 		 * check if the first spinlock is the one curthread
1709 		 * should hold.
1710 		 */
1711 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1712 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1713 		    lock1->li_lock == lock && n == 0)
1714 			return (0);
1715 
1716 		va_start(ap, fmt);
1717 		vprintf(fmt, ap);
1718 		va_end(ap);
1719 		printf(" with the following");
1720 		if (flags & WARN_SLEEPOK)
1721 			printf(" non-sleepable");
1722 		printf(" locks held:\n");
1723 		n += witness_list_locks(&lock_list, printf);
1724 	} else
1725 		sched_unpin();
1726 	if (flags & WARN_PANIC && n)
1727 		kassert_panic("%s", __func__);
1728 	else
1729 		witness_debugger(n);
1730 	return (n);
1731 }
1732 
1733 const char *
1734 witness_file(struct lock_object *lock)
1735 {
1736 	struct witness *w;
1737 
1738 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1739 		return ("?");
1740 	w = lock->lo_witness;
1741 	return (w->w_file);
1742 }
1743 
1744 int
1745 witness_line(struct lock_object *lock)
1746 {
1747 	struct witness *w;
1748 
1749 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1750 		return (0);
1751 	w = lock->lo_witness;
1752 	return (w->w_line);
1753 }
1754 
1755 static struct witness *
1756 enroll(const char *description, struct lock_class *lock_class)
1757 {
1758 	struct witness *w;
1759 	struct witness_list *typelist;
1760 
1761 	MPASS(description != NULL);
1762 
1763 	if (witness_watch == -1 || panicstr != NULL)
1764 		return (NULL);
1765 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1766 		if (witness_skipspin)
1767 			return (NULL);
1768 		else
1769 			typelist = &w_spin;
1770 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1771 		typelist = &w_sleep;
1772 	} else {
1773 		kassert_panic("lock class %s is not sleep or spin",
1774 		    lock_class->lc_name);
1775 		return (NULL);
1776 	}
1777 
1778 	mtx_lock_spin(&w_mtx);
1779 	w = witness_hash_get(description);
1780 	if (w)
1781 		goto found;
1782 	if ((w = witness_get()) == NULL)
1783 		return (NULL);
1784 	MPASS(strlen(description) < MAX_W_NAME);
1785 	strcpy(w->w_name, description);
1786 	w->w_class = lock_class;
1787 	w->w_refcount = 1;
1788 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1789 	if (lock_class->lc_flags & LC_SPINLOCK) {
1790 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1791 		w_spin_cnt++;
1792 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1793 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1794 		w_sleep_cnt++;
1795 	}
1796 
1797 	/* Insert new witness into the hash */
1798 	witness_hash_put(w);
1799 	witness_increment_graph_generation();
1800 	mtx_unlock_spin(&w_mtx);
1801 	return (w);
1802 found:
1803 	w->w_refcount++;
1804 	mtx_unlock_spin(&w_mtx);
1805 	if (lock_class != w->w_class)
1806 		kassert_panic(
1807 			"lock (%s) %s does not match earlier (%s) lock",
1808 			description, lock_class->lc_name,
1809 			w->w_class->lc_name);
1810 	return (w);
1811 }
1812 
1813 static void
1814 depart(struct witness *w)
1815 {
1816 	struct witness_list *list;
1817 
1818 	MPASS(w->w_refcount == 0);
1819 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1820 		list = &w_sleep;
1821 		w_sleep_cnt--;
1822 	} else {
1823 		list = &w_spin;
1824 		w_spin_cnt--;
1825 	}
1826 	/*
1827 	 * Set file to NULL as it may point into a loadable module.
1828 	 */
1829 	w->w_file = NULL;
1830 	w->w_line = 0;
1831 	witness_increment_graph_generation();
1832 }
1833 
1834 
1835 static void
1836 adopt(struct witness *parent, struct witness *child)
1837 {
1838 	int pi, ci, i, j;
1839 
1840 	if (witness_cold == 0)
1841 		mtx_assert(&w_mtx, MA_OWNED);
1842 
1843 	/* If the relationship is already known, there's no work to be done. */
1844 	if (isitmychild(parent, child))
1845 		return;
1846 
1847 	/* When the structure of the graph changes, bump up the generation. */
1848 	witness_increment_graph_generation();
1849 
1850 	/*
1851 	 * The hard part ... create the direct relationship, then propagate all
1852 	 * indirect relationships.
1853 	 */
1854 	pi = parent->w_index;
1855 	ci = child->w_index;
1856 	WITNESS_INDEX_ASSERT(pi);
1857 	WITNESS_INDEX_ASSERT(ci);
1858 	MPASS(pi != ci);
1859 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1860 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1861 
1862 	/*
1863 	 * If parent was not already an ancestor of child,
1864 	 * then we increment the descendant and ancestor counters.
1865 	 */
1866 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1867 		parent->w_num_descendants++;
1868 		child->w_num_ancestors++;
1869 	}
1870 
1871 	/*
1872 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1873 	 * an ancestor of 'pi' during this loop.
1874 	 */
1875 	for (i = 1; i <= w_max_used_index; i++) {
1876 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1877 		    (i != pi))
1878 			continue;
1879 
1880 		/* Find each descendant of 'i' and mark it as a descendant. */
1881 		for (j = 1; j <= w_max_used_index; j++) {
1882 
1883 			/*
1884 			 * Skip children that are already marked as
1885 			 * descendants of 'i'.
1886 			 */
1887 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1888 				continue;
1889 
1890 			/*
1891 			 * We are only interested in descendants of 'ci'. Note
1892 			 * that 'ci' itself is counted as a descendant of 'ci'.
1893 			 */
1894 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1895 			    (j != ci))
1896 				continue;
1897 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1898 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1899 			w_data[i].w_num_descendants++;
1900 			w_data[j].w_num_ancestors++;
1901 
1902 			/*
1903 			 * Make sure we aren't marking a node as both an
1904 			 * ancestor and descendant. We should have caught
1905 			 * this as a lock order reversal earlier.
1906 			 */
1907 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1908 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1909 				printf("witness rmatrix paradox! [%d][%d]=%d "
1910 				    "both ancestor and descendant\n",
1911 				    i, j, w_rmatrix[i][j]);
1912 				kdb_backtrace();
1913 				printf("Witness disabled.\n");
1914 				witness_watch = -1;
1915 			}
1916 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1917 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1918 				printf("witness rmatrix paradox! [%d][%d]=%d "
1919 				    "both ancestor and descendant\n",
1920 				    j, i, w_rmatrix[j][i]);
1921 				kdb_backtrace();
1922 				printf("Witness disabled.\n");
1923 				witness_watch = -1;
1924 			}
1925 		}
1926 	}
1927 }
1928 
1929 static void
1930 itismychild(struct witness *parent, struct witness *child)
1931 {
1932 	int unlocked;
1933 
1934 	MPASS(child != NULL && parent != NULL);
1935 	if (witness_cold == 0)
1936 		mtx_assert(&w_mtx, MA_OWNED);
1937 
1938 	if (!witness_lock_type_equal(parent, child)) {
1939 		if (witness_cold == 0) {
1940 			unlocked = 1;
1941 			mtx_unlock_spin(&w_mtx);
1942 		} else {
1943 			unlocked = 0;
1944 		}
1945 		kassert_panic(
1946 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1947 		    "the same lock type", __func__, parent->w_name,
1948 		    parent->w_class->lc_name, child->w_name,
1949 		    child->w_class->lc_name);
1950 		if (unlocked)
1951 			mtx_lock_spin(&w_mtx);
1952 	}
1953 	adopt(parent, child);
1954 }
1955 
1956 /*
1957  * Generic code for the isitmy*() functions. The rmask parameter is the
1958  * expected relationship of w1 to w2.
1959  */
1960 static int
1961 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1962 {
1963 	unsigned char r1, r2;
1964 	int i1, i2;
1965 
1966 	i1 = w1->w_index;
1967 	i2 = w2->w_index;
1968 	WITNESS_INDEX_ASSERT(i1);
1969 	WITNESS_INDEX_ASSERT(i2);
1970 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1971 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1972 
1973 	/* The flags on one better be the inverse of the flags on the other */
1974 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1975 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1976 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
1977 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
1978 		    "w_rmatrix[%d][%d] == %hhx\n",
1979 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1980 		    i2, i1, r2);
1981 		kdb_backtrace();
1982 		printf("Witness disabled.\n");
1983 		witness_watch = -1;
1984 	}
1985 	return (r1 & rmask);
1986 }
1987 
1988 /*
1989  * Checks if @child is a direct child of @parent.
1990  */
1991 static int
1992 isitmychild(struct witness *parent, struct witness *child)
1993 {
1994 
1995 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1996 }
1997 
1998 /*
1999  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2000  */
2001 static int
2002 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2003 {
2004 
2005 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2006 	    __func__));
2007 }
2008 
2009 #ifdef BLESSING
2010 static int
2011 blessed(struct witness *w1, struct witness *w2)
2012 {
2013 	int i;
2014 	struct witness_blessed *b;
2015 
2016 	for (i = 0; i < blessed_count; i++) {
2017 		b = &blessed_list[i];
2018 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2019 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2020 				return (1);
2021 			continue;
2022 		}
2023 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2024 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2025 				return (1);
2026 	}
2027 	return (0);
2028 }
2029 #endif
2030 
2031 static struct witness *
2032 witness_get(void)
2033 {
2034 	struct witness *w;
2035 	int index;
2036 
2037 	if (witness_cold == 0)
2038 		mtx_assert(&w_mtx, MA_OWNED);
2039 
2040 	if (witness_watch == -1) {
2041 		mtx_unlock_spin(&w_mtx);
2042 		return (NULL);
2043 	}
2044 	if (STAILQ_EMPTY(&w_free)) {
2045 		witness_watch = -1;
2046 		mtx_unlock_spin(&w_mtx);
2047 		printf("WITNESS: unable to allocate a new witness object\n");
2048 		return (NULL);
2049 	}
2050 	w = STAILQ_FIRST(&w_free);
2051 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2052 	w_free_cnt--;
2053 	index = w->w_index;
2054 	MPASS(index > 0 && index == w_max_used_index+1 &&
2055 	    index < WITNESS_COUNT);
2056 	bzero(w, sizeof(*w));
2057 	w->w_index = index;
2058 	if (index > w_max_used_index)
2059 		w_max_used_index = index;
2060 	return (w);
2061 }
2062 
2063 static void
2064 witness_free(struct witness *w)
2065 {
2066 
2067 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2068 	w_free_cnt++;
2069 }
2070 
2071 static struct lock_list_entry *
2072 witness_lock_list_get(void)
2073 {
2074 	struct lock_list_entry *lle;
2075 
2076 	if (witness_watch == -1)
2077 		return (NULL);
2078 	mtx_lock_spin(&w_mtx);
2079 	lle = w_lock_list_free;
2080 	if (lle == NULL) {
2081 		witness_watch = -1;
2082 		mtx_unlock_spin(&w_mtx);
2083 		printf("%s: witness exhausted\n", __func__);
2084 		return (NULL);
2085 	}
2086 	w_lock_list_free = lle->ll_next;
2087 	mtx_unlock_spin(&w_mtx);
2088 	bzero(lle, sizeof(*lle));
2089 	return (lle);
2090 }
2091 
2092 static void
2093 witness_lock_list_free(struct lock_list_entry *lle)
2094 {
2095 
2096 	mtx_lock_spin(&w_mtx);
2097 	lle->ll_next = w_lock_list_free;
2098 	w_lock_list_free = lle;
2099 	mtx_unlock_spin(&w_mtx);
2100 }
2101 
2102 static struct lock_instance *
2103 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2104 {
2105 	struct lock_list_entry *lle;
2106 	struct lock_instance *instance;
2107 	int i;
2108 
2109 	for (lle = list; lle != NULL; lle = lle->ll_next)
2110 		for (i = lle->ll_count - 1; i >= 0; i--) {
2111 			instance = &lle->ll_children[i];
2112 			if (instance->li_lock == lock)
2113 				return (instance);
2114 		}
2115 	return (NULL);
2116 }
2117 
2118 static void
2119 witness_list_lock(struct lock_instance *instance,
2120     int (*prnt)(const char *fmt, ...))
2121 {
2122 	struct lock_object *lock;
2123 
2124 	lock = instance->li_lock;
2125 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2126 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2127 	if (lock->lo_witness->w_name != lock->lo_name)
2128 		prnt(" (%s)", lock->lo_witness->w_name);
2129 	prnt(" r = %d (%p) locked @ %s:%d\n",
2130 	    instance->li_flags & LI_RECURSEMASK, lock,
2131 	    fixup_filename(instance->li_file), instance->li_line);
2132 }
2133 
2134 #ifdef DDB
2135 static int
2136 witness_thread_has_locks(struct thread *td)
2137 {
2138 
2139 	if (td->td_sleeplocks == NULL)
2140 		return (0);
2141 	return (td->td_sleeplocks->ll_count != 0);
2142 }
2143 
2144 static int
2145 witness_proc_has_locks(struct proc *p)
2146 {
2147 	struct thread *td;
2148 
2149 	FOREACH_THREAD_IN_PROC(p, td) {
2150 		if (witness_thread_has_locks(td))
2151 			return (1);
2152 	}
2153 	return (0);
2154 }
2155 #endif
2156 
2157 int
2158 witness_list_locks(struct lock_list_entry **lock_list,
2159     int (*prnt)(const char *fmt, ...))
2160 {
2161 	struct lock_list_entry *lle;
2162 	int i, nheld;
2163 
2164 	nheld = 0;
2165 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2166 		for (i = lle->ll_count - 1; i >= 0; i--) {
2167 			witness_list_lock(&lle->ll_children[i], prnt);
2168 			nheld++;
2169 		}
2170 	return (nheld);
2171 }
2172 
2173 /*
2174  * This is a bit risky at best.  We call this function when we have timed
2175  * out acquiring a spin lock, and we assume that the other CPU is stuck
2176  * with this lock held.  So, we go groveling around in the other CPU's
2177  * per-cpu data to try to find the lock instance for this spin lock to
2178  * see when it was last acquired.
2179  */
2180 void
2181 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2182     int (*prnt)(const char *fmt, ...))
2183 {
2184 	struct lock_instance *instance;
2185 	struct pcpu *pc;
2186 
2187 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2188 		return;
2189 	pc = pcpu_find(owner->td_oncpu);
2190 	instance = find_instance(pc->pc_spinlocks, lock);
2191 	if (instance != NULL)
2192 		witness_list_lock(instance, prnt);
2193 }
2194 
2195 void
2196 witness_save(struct lock_object *lock, const char **filep, int *linep)
2197 {
2198 	struct lock_list_entry *lock_list;
2199 	struct lock_instance *instance;
2200 	struct lock_class *class;
2201 
2202 	/*
2203 	 * This function is used independently in locking code to deal with
2204 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2205 	 * is gone.
2206 	 */
2207 	if (SCHEDULER_STOPPED())
2208 		return;
2209 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2210 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2211 		return;
2212 	class = LOCK_CLASS(lock);
2213 	if (class->lc_flags & LC_SLEEPLOCK)
2214 		lock_list = curthread->td_sleeplocks;
2215 	else {
2216 		if (witness_skipspin)
2217 			return;
2218 		lock_list = PCPU_GET(spinlocks);
2219 	}
2220 	instance = find_instance(lock_list, lock);
2221 	if (instance == NULL) {
2222 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2223 		    class->lc_name, lock->lo_name);
2224 		return;
2225 	}
2226 	*filep = instance->li_file;
2227 	*linep = instance->li_line;
2228 }
2229 
2230 void
2231 witness_restore(struct lock_object *lock, const char *file, int line)
2232 {
2233 	struct lock_list_entry *lock_list;
2234 	struct lock_instance *instance;
2235 	struct lock_class *class;
2236 
2237 	/*
2238 	 * This function is used independently in locking code to deal with
2239 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2240 	 * is gone.
2241 	 */
2242 	if (SCHEDULER_STOPPED())
2243 		return;
2244 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2245 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2246 		return;
2247 	class = LOCK_CLASS(lock);
2248 	if (class->lc_flags & LC_SLEEPLOCK)
2249 		lock_list = curthread->td_sleeplocks;
2250 	else {
2251 		if (witness_skipspin)
2252 			return;
2253 		lock_list = PCPU_GET(spinlocks);
2254 	}
2255 	instance = find_instance(lock_list, lock);
2256 	if (instance == NULL)
2257 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2258 		    class->lc_name, lock->lo_name);
2259 	lock->lo_witness->w_file = file;
2260 	lock->lo_witness->w_line = line;
2261 	if (instance == NULL)
2262 		return;
2263 	instance->li_file = file;
2264 	instance->li_line = line;
2265 }
2266 
2267 void
2268 witness_assert(const struct lock_object *lock, int flags, const char *file,
2269     int line)
2270 {
2271 #ifdef INVARIANT_SUPPORT
2272 	struct lock_instance *instance;
2273 	struct lock_class *class;
2274 
2275 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2276 		return;
2277 	class = LOCK_CLASS(lock);
2278 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2279 		instance = find_instance(curthread->td_sleeplocks, lock);
2280 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2281 		instance = find_instance(PCPU_GET(spinlocks), lock);
2282 	else {
2283 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2284 		    class->lc_name, lock->lo_name);
2285 		return;
2286 	}
2287 	switch (flags) {
2288 	case LA_UNLOCKED:
2289 		if (instance != NULL)
2290 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2291 			    class->lc_name, lock->lo_name,
2292 			    fixup_filename(file), line);
2293 		break;
2294 	case LA_LOCKED:
2295 	case LA_LOCKED | LA_RECURSED:
2296 	case LA_LOCKED | LA_NOTRECURSED:
2297 	case LA_SLOCKED:
2298 	case LA_SLOCKED | LA_RECURSED:
2299 	case LA_SLOCKED | LA_NOTRECURSED:
2300 	case LA_XLOCKED:
2301 	case LA_XLOCKED | LA_RECURSED:
2302 	case LA_XLOCKED | LA_NOTRECURSED:
2303 		if (instance == NULL) {
2304 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2305 			    class->lc_name, lock->lo_name,
2306 			    fixup_filename(file), line);
2307 			break;
2308 		}
2309 		if ((flags & LA_XLOCKED) != 0 &&
2310 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2311 			kassert_panic(
2312 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2313 			    class->lc_name, lock->lo_name,
2314 			    fixup_filename(file), line);
2315 		if ((flags & LA_SLOCKED) != 0 &&
2316 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2317 			kassert_panic(
2318 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2319 			    class->lc_name, lock->lo_name,
2320 			    fixup_filename(file), line);
2321 		if ((flags & LA_RECURSED) != 0 &&
2322 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2323 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2324 			    class->lc_name, lock->lo_name,
2325 			    fixup_filename(file), line);
2326 		if ((flags & LA_NOTRECURSED) != 0 &&
2327 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2328 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2329 			    class->lc_name, lock->lo_name,
2330 			    fixup_filename(file), line);
2331 		break;
2332 	default:
2333 		kassert_panic("Invalid lock assertion at %s:%d.",
2334 		    fixup_filename(file), line);
2335 
2336 	}
2337 #endif	/* INVARIANT_SUPPORT */
2338 }
2339 
2340 static void
2341 witness_setflag(struct lock_object *lock, int flag, int set)
2342 {
2343 	struct lock_list_entry *lock_list;
2344 	struct lock_instance *instance;
2345 	struct lock_class *class;
2346 
2347 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2348 		return;
2349 	class = LOCK_CLASS(lock);
2350 	if (class->lc_flags & LC_SLEEPLOCK)
2351 		lock_list = curthread->td_sleeplocks;
2352 	else {
2353 		if (witness_skipspin)
2354 			return;
2355 		lock_list = PCPU_GET(spinlocks);
2356 	}
2357 	instance = find_instance(lock_list, lock);
2358 	if (instance == NULL) {
2359 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2360 		    class->lc_name, lock->lo_name);
2361 		return;
2362 	}
2363 
2364 	if (set)
2365 		instance->li_flags |= flag;
2366 	else
2367 		instance->li_flags &= ~flag;
2368 }
2369 
2370 void
2371 witness_norelease(struct lock_object *lock)
2372 {
2373 
2374 	witness_setflag(lock, LI_NORELEASE, 1);
2375 }
2376 
2377 void
2378 witness_releaseok(struct lock_object *lock)
2379 {
2380 
2381 	witness_setflag(lock, LI_NORELEASE, 0);
2382 }
2383 
2384 #ifdef DDB
2385 static void
2386 witness_ddb_list(struct thread *td)
2387 {
2388 
2389 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2390 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2391 
2392 	if (witness_watch < 1)
2393 		return;
2394 
2395 	witness_list_locks(&td->td_sleeplocks, db_printf);
2396 
2397 	/*
2398 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2399 	 * if td is currently executing on some other CPU and holds spin locks
2400 	 * as we won't display those locks.  If we had a MI way of getting
2401 	 * the per-cpu data for a given cpu then we could use
2402 	 * td->td_oncpu to get the list of spinlocks for this thread
2403 	 * and "fix" this.
2404 	 *
2405 	 * That still wouldn't really fix this unless we locked the scheduler
2406 	 * lock or stopped the other CPU to make sure it wasn't changing the
2407 	 * list out from under us.  It is probably best to just not try to
2408 	 * handle threads on other CPU's for now.
2409 	 */
2410 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2411 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2412 }
2413 
2414 DB_SHOW_COMMAND(locks, db_witness_list)
2415 {
2416 	struct thread *td;
2417 
2418 	if (have_addr)
2419 		td = db_lookup_thread(addr, TRUE);
2420 	else
2421 		td = kdb_thread;
2422 	witness_ddb_list(td);
2423 }
2424 
2425 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2426 {
2427 	struct thread *td;
2428 	struct proc *p;
2429 
2430 	/*
2431 	 * It would be nice to list only threads and processes that actually
2432 	 * held sleep locks, but that information is currently not exported
2433 	 * by WITNESS.
2434 	 */
2435 	FOREACH_PROC_IN_SYSTEM(p) {
2436 		if (!witness_proc_has_locks(p))
2437 			continue;
2438 		FOREACH_THREAD_IN_PROC(p, td) {
2439 			if (!witness_thread_has_locks(td))
2440 				continue;
2441 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2442 			    p->p_comm, td, td->td_tid);
2443 			witness_ddb_list(td);
2444 			if (db_pager_quit)
2445 				return;
2446 		}
2447 	}
2448 }
2449 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2450 
2451 DB_SHOW_COMMAND(witness, db_witness_display)
2452 {
2453 
2454 	witness_ddb_display(db_printf);
2455 }
2456 #endif
2457 
2458 static int
2459 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2460 {
2461 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2462 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2463 	struct sbuf *sb;
2464 	u_int w_rmatrix1, w_rmatrix2;
2465 	int error, generation, i, j;
2466 
2467 	tmp_data1 = NULL;
2468 	tmp_data2 = NULL;
2469 	tmp_w1 = NULL;
2470 	tmp_w2 = NULL;
2471 	if (witness_watch < 1) {
2472 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2473 		return (error);
2474 	}
2475 	if (witness_cold) {
2476 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2477 		return (error);
2478 	}
2479 	error = 0;
2480 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2481 	if (sb == NULL)
2482 		return (ENOMEM);
2483 
2484 	/* Allocate and init temporary storage space. */
2485 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2486 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2487 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2488 	    M_WAITOK | M_ZERO);
2489 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2490 	    M_WAITOK | M_ZERO);
2491 	stack_zero(&tmp_data1->wlod_stack);
2492 	stack_zero(&tmp_data2->wlod_stack);
2493 
2494 restart:
2495 	mtx_lock_spin(&w_mtx);
2496 	generation = w_generation;
2497 	mtx_unlock_spin(&w_mtx);
2498 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2499 	    w_lohash.wloh_count);
2500 	for (i = 1; i < w_max_used_index; i++) {
2501 		mtx_lock_spin(&w_mtx);
2502 		if (generation != w_generation) {
2503 			mtx_unlock_spin(&w_mtx);
2504 
2505 			/* The graph has changed, try again. */
2506 			req->oldidx = 0;
2507 			sbuf_clear(sb);
2508 			goto restart;
2509 		}
2510 
2511 		w1 = &w_data[i];
2512 		if (w1->w_reversed == 0) {
2513 			mtx_unlock_spin(&w_mtx);
2514 			continue;
2515 		}
2516 
2517 		/* Copy w1 locally so we can release the spin lock. */
2518 		*tmp_w1 = *w1;
2519 		mtx_unlock_spin(&w_mtx);
2520 
2521 		if (tmp_w1->w_reversed == 0)
2522 			continue;
2523 		for (j = 1; j < w_max_used_index; j++) {
2524 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2525 				continue;
2526 
2527 			mtx_lock_spin(&w_mtx);
2528 			if (generation != w_generation) {
2529 				mtx_unlock_spin(&w_mtx);
2530 
2531 				/* The graph has changed, try again. */
2532 				req->oldidx = 0;
2533 				sbuf_clear(sb);
2534 				goto restart;
2535 			}
2536 
2537 			w2 = &w_data[j];
2538 			data1 = witness_lock_order_get(w1, w2);
2539 			data2 = witness_lock_order_get(w2, w1);
2540 
2541 			/*
2542 			 * Copy information locally so we can release the
2543 			 * spin lock.
2544 			 */
2545 			*tmp_w2 = *w2;
2546 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2547 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2548 
2549 			if (data1) {
2550 				stack_zero(&tmp_data1->wlod_stack);
2551 				stack_copy(&data1->wlod_stack,
2552 				    &tmp_data1->wlod_stack);
2553 			}
2554 			if (data2 && data2 != data1) {
2555 				stack_zero(&tmp_data2->wlod_stack);
2556 				stack_copy(&data2->wlod_stack,
2557 				    &tmp_data2->wlod_stack);
2558 			}
2559 			mtx_unlock_spin(&w_mtx);
2560 
2561 			sbuf_printf(sb,
2562 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2563 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2564 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2565 #if 0
2566  			sbuf_printf(sb,
2567 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2568  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2569  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2570 #endif
2571 			if (data1) {
2572 				sbuf_printf(sb,
2573 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2574 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2575 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2576 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2577 				sbuf_printf(sb, "\n");
2578 			}
2579 			if (data2 && data2 != data1) {
2580 				sbuf_printf(sb,
2581 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2582 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2583 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2584 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2585 				sbuf_printf(sb, "\n");
2586 			}
2587 		}
2588 	}
2589 	mtx_lock_spin(&w_mtx);
2590 	if (generation != w_generation) {
2591 		mtx_unlock_spin(&w_mtx);
2592 
2593 		/*
2594 		 * The graph changed while we were printing stack data,
2595 		 * try again.
2596 		 */
2597 		req->oldidx = 0;
2598 		sbuf_clear(sb);
2599 		goto restart;
2600 	}
2601 	mtx_unlock_spin(&w_mtx);
2602 
2603 	/* Free temporary storage space. */
2604 	free(tmp_data1, M_TEMP);
2605 	free(tmp_data2, M_TEMP);
2606 	free(tmp_w1, M_TEMP);
2607 	free(tmp_w2, M_TEMP);
2608 
2609 	sbuf_finish(sb);
2610 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2611 	sbuf_delete(sb);
2612 
2613 	return (error);
2614 }
2615 
2616 static int
2617 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2618 {
2619 	struct witness *w;
2620 	struct sbuf *sb;
2621 	int error;
2622 
2623 	if (witness_watch < 1) {
2624 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2625 		return (error);
2626 	}
2627 	if (witness_cold) {
2628 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2629 		return (error);
2630 	}
2631 	error = 0;
2632 
2633 	error = sysctl_wire_old_buffer(req, 0);
2634 	if (error != 0)
2635 		return (error);
2636 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2637 	if (sb == NULL)
2638 		return (ENOMEM);
2639 	sbuf_printf(sb, "\n");
2640 
2641 	mtx_lock_spin(&w_mtx);
2642 	STAILQ_FOREACH(w, &w_all, w_list)
2643 		w->w_displayed = 0;
2644 	STAILQ_FOREACH(w, &w_all, w_list)
2645 		witness_add_fullgraph(sb, w);
2646 	mtx_unlock_spin(&w_mtx);
2647 
2648 	/*
2649 	 * Close the sbuf and return to userland.
2650 	 */
2651 	error = sbuf_finish(sb);
2652 	sbuf_delete(sb);
2653 
2654 	return (error);
2655 }
2656 
2657 static int
2658 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2659 {
2660 	int error, value;
2661 
2662 	value = witness_watch;
2663 	error = sysctl_handle_int(oidp, &value, 0, req);
2664 	if (error != 0 || req->newptr == NULL)
2665 		return (error);
2666 	if (value > 1 || value < -1 ||
2667 	    (witness_watch == -1 && value != witness_watch))
2668 		return (EINVAL);
2669 	witness_watch = value;
2670 	return (0);
2671 }
2672 
2673 static void
2674 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2675 {
2676 	int i;
2677 
2678 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2679 		return;
2680 	w->w_displayed = 1;
2681 
2682 	WITNESS_INDEX_ASSERT(w->w_index);
2683 	for (i = 1; i <= w_max_used_index; i++) {
2684 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2685 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2686 			    w_data[i].w_name);
2687 			witness_add_fullgraph(sb, &w_data[i]);
2688 		}
2689 	}
2690 }
2691 
2692 /*
2693  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2694  * interprets the key as a string and reads until the null
2695  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2696  * hash value computed from the key.
2697  */
2698 static uint32_t
2699 witness_hash_djb2(const uint8_t *key, uint32_t size)
2700 {
2701 	unsigned int hash = 5381;
2702 	int i;
2703 
2704 	/* hash = hash * 33 + key[i] */
2705 	if (size)
2706 		for (i = 0; i < size; i++)
2707 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2708 	else
2709 		for (i = 0; key[i] != 0; i++)
2710 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2711 
2712 	return (hash);
2713 }
2714 
2715 
2716 /*
2717  * Initializes the two witness hash tables. Called exactly once from
2718  * witness_initialize().
2719  */
2720 static void
2721 witness_init_hash_tables(void)
2722 {
2723 	int i;
2724 
2725 	MPASS(witness_cold);
2726 
2727 	/* Initialize the hash tables. */
2728 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2729 		w_hash.wh_array[i] = NULL;
2730 
2731 	w_hash.wh_size = WITNESS_HASH_SIZE;
2732 	w_hash.wh_count = 0;
2733 
2734 	/* Initialize the lock order data hash. */
2735 	w_lofree = NULL;
2736 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2737 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2738 		w_lodata[i].wlod_next = w_lofree;
2739 		w_lofree = &w_lodata[i];
2740 	}
2741 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2742 	w_lohash.wloh_count = 0;
2743 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2744 		w_lohash.wloh_array[i] = NULL;
2745 }
2746 
2747 static struct witness *
2748 witness_hash_get(const char *key)
2749 {
2750 	struct witness *w;
2751 	uint32_t hash;
2752 
2753 	MPASS(key != NULL);
2754 	if (witness_cold == 0)
2755 		mtx_assert(&w_mtx, MA_OWNED);
2756 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2757 	w = w_hash.wh_array[hash];
2758 	while (w != NULL) {
2759 		if (strcmp(w->w_name, key) == 0)
2760 			goto out;
2761 		w = w->w_hash_next;
2762 	}
2763 
2764 out:
2765 	return (w);
2766 }
2767 
2768 static void
2769 witness_hash_put(struct witness *w)
2770 {
2771 	uint32_t hash;
2772 
2773 	MPASS(w != NULL);
2774 	MPASS(w->w_name != NULL);
2775 	if (witness_cold == 0)
2776 		mtx_assert(&w_mtx, MA_OWNED);
2777 	KASSERT(witness_hash_get(w->w_name) == NULL,
2778 	    ("%s: trying to add a hash entry that already exists!", __func__));
2779 	KASSERT(w->w_hash_next == NULL,
2780 	    ("%s: w->w_hash_next != NULL", __func__));
2781 
2782 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2783 	w->w_hash_next = w_hash.wh_array[hash];
2784 	w_hash.wh_array[hash] = w;
2785 	w_hash.wh_count++;
2786 }
2787 
2788 
2789 static struct witness_lock_order_data *
2790 witness_lock_order_get(struct witness *parent, struct witness *child)
2791 {
2792 	struct witness_lock_order_data *data = NULL;
2793 	struct witness_lock_order_key key;
2794 	unsigned int hash;
2795 
2796 	MPASS(parent != NULL && child != NULL);
2797 	key.from = parent->w_index;
2798 	key.to = child->w_index;
2799 	WITNESS_INDEX_ASSERT(key.from);
2800 	WITNESS_INDEX_ASSERT(key.to);
2801 	if ((w_rmatrix[parent->w_index][child->w_index]
2802 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2803 		goto out;
2804 
2805 	hash = witness_hash_djb2((const char*)&key,
2806 	    sizeof(key)) % w_lohash.wloh_size;
2807 	data = w_lohash.wloh_array[hash];
2808 	while (data != NULL) {
2809 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2810 			break;
2811 		data = data->wlod_next;
2812 	}
2813 
2814 out:
2815 	return (data);
2816 }
2817 
2818 /*
2819  * Verify that parent and child have a known relationship, are not the same,
2820  * and child is actually a child of parent.  This is done without w_mtx
2821  * to avoid contention in the common case.
2822  */
2823 static int
2824 witness_lock_order_check(struct witness *parent, struct witness *child)
2825 {
2826 
2827 	if (parent != child &&
2828 	    w_rmatrix[parent->w_index][child->w_index]
2829 	    & WITNESS_LOCK_ORDER_KNOWN &&
2830 	    isitmychild(parent, child))
2831 		return (1);
2832 
2833 	return (0);
2834 }
2835 
2836 static int
2837 witness_lock_order_add(struct witness *parent, struct witness *child)
2838 {
2839 	struct witness_lock_order_data *data = NULL;
2840 	struct witness_lock_order_key key;
2841 	unsigned int hash;
2842 
2843 	MPASS(parent != NULL && child != NULL);
2844 	key.from = parent->w_index;
2845 	key.to = child->w_index;
2846 	WITNESS_INDEX_ASSERT(key.from);
2847 	WITNESS_INDEX_ASSERT(key.to);
2848 	if (w_rmatrix[parent->w_index][child->w_index]
2849 	    & WITNESS_LOCK_ORDER_KNOWN)
2850 		return (1);
2851 
2852 	hash = witness_hash_djb2((const char*)&key,
2853 	    sizeof(key)) % w_lohash.wloh_size;
2854 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2855 	data = w_lofree;
2856 	if (data == NULL)
2857 		return (0);
2858 	w_lofree = data->wlod_next;
2859 	data->wlod_next = w_lohash.wloh_array[hash];
2860 	data->wlod_key = key;
2861 	w_lohash.wloh_array[hash] = data;
2862 	w_lohash.wloh_count++;
2863 	stack_zero(&data->wlod_stack);
2864 	stack_save(&data->wlod_stack);
2865 	return (1);
2866 }
2867 
2868 /* Call this whenver the structure of the witness graph changes. */
2869 static void
2870 witness_increment_graph_generation(void)
2871 {
2872 
2873 	if (witness_cold == 0)
2874 		mtx_assert(&w_mtx, MA_OWNED);
2875 	w_generation++;
2876 }
2877 
2878 #ifdef KDB
2879 static void
2880 _witness_debugger(int cond, const char *msg)
2881 {
2882 
2883 	if (witness_trace && cond)
2884 		kdb_backtrace();
2885 	if (witness_kdb && cond)
2886 		kdb_enter(KDB_WHY_WITNESS, msg);
2887 }
2888 #endif
2889