xref: /freebsd/sys/kern/subr_witness.c (revision 6ef6ba9950260f42b47499d17874d00ca9290955)
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	1024
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, *iclass;
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("excl->share");
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("share->excl");
1132 		}
1133 		return;
1134 	}
1135 
1136 	/* Warn if the interlock is not locked exactly once. */
1137 	if (interlock != NULL) {
1138 		iclass = LOCK_CLASS(interlock);
1139 		lock1 = find_instance(lock_list, interlock);
1140 		if (lock1 == NULL)
1141 			kassert_panic("interlock (%s) %s not locked @ %s:%d",
1142 			    iclass->lc_name, interlock->lo_name,
1143 			    fixup_filename(file), line);
1144 		else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1145 			kassert_panic("interlock (%s) %s recursed @ %s:%d",
1146 			    iclass->lc_name, interlock->lo_name,
1147 			    fixup_filename(file), line);
1148 	}
1149 
1150 	/*
1151 	 * Find the previously acquired lock, but ignore interlocks.
1152 	 */
1153 	plock = &lock_list->ll_children[lock_list->ll_count - 1];
1154 	if (interlock != NULL && plock->li_lock == interlock) {
1155 		if (lock_list->ll_count > 1)
1156 			plock =
1157 			    &lock_list->ll_children[lock_list->ll_count - 2];
1158 		else {
1159 			lle = lock_list->ll_next;
1160 
1161 			/*
1162 			 * The interlock is the only lock we hold, so
1163 			 * simply return.
1164 			 */
1165 			if (lle == NULL)
1166 				return;
1167 			plock = &lle->ll_children[lle->ll_count - 1];
1168 		}
1169 	}
1170 
1171 	/*
1172 	 * Try to perform most checks without a lock.  If this succeeds we
1173 	 * can skip acquiring the lock and return success.
1174 	 */
1175 	w1 = plock->li_lock->lo_witness;
1176 	if (witness_lock_order_check(w1, w))
1177 		return;
1178 
1179 	/*
1180 	 * Check for duplicate locks of the same type.  Note that we only
1181 	 * have to check for this on the last lock we just acquired.  Any
1182 	 * other cases will be caught as lock order violations.
1183 	 */
1184 	mtx_lock_spin(&w_mtx);
1185 	witness_lock_order_add(w1, w);
1186 	if (w1 == w) {
1187 		i = w->w_index;
1188 		if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1189 		    !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1190 		    w_rmatrix[i][i] |= WITNESS_REVERSAL;
1191 			w->w_reversed = 1;
1192 			mtx_unlock_spin(&w_mtx);
1193 			printf(
1194 			    "acquiring duplicate lock of same type: \"%s\"\n",
1195 			    w->w_name);
1196 			printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1197 			    fixup_filename(plock->li_file), plock->li_line);
1198 			printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1199 			    fixup_filename(file), line);
1200 			witness_debugger(1);
1201 		} else
1202 			mtx_unlock_spin(&w_mtx);
1203 		return;
1204 	}
1205 	mtx_assert(&w_mtx, MA_OWNED);
1206 
1207 	/*
1208 	 * If we know that the lock we are acquiring comes after
1209 	 * the lock we most recently acquired in the lock order tree,
1210 	 * then there is no need for any further checks.
1211 	 */
1212 	if (isitmychild(w1, w))
1213 		goto out;
1214 
1215 	for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1216 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1217 
1218 			MPASS(j < WITNESS_COUNT);
1219 			lock1 = &lle->ll_children[i];
1220 
1221 			/*
1222 			 * Ignore the interlock.
1223 			 */
1224 			if (interlock == lock1->li_lock)
1225 				continue;
1226 
1227 			/*
1228 			 * If this lock doesn't undergo witness checking,
1229 			 * then skip it.
1230 			 */
1231 			w1 = lock1->li_lock->lo_witness;
1232 			if (w1 == NULL) {
1233 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1234 				    ("lock missing witness structure"));
1235 				continue;
1236 			}
1237 
1238 			/*
1239 			 * If we are locking Giant and this is a sleepable
1240 			 * lock, then skip it.
1241 			 */
1242 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1243 			    lock == &Giant.lock_object)
1244 				continue;
1245 
1246 			/*
1247 			 * If we are locking a sleepable lock and this lock
1248 			 * is Giant, then skip it.
1249 			 */
1250 			if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1251 			    lock1->li_lock == &Giant.lock_object)
1252 				continue;
1253 
1254 			/*
1255 			 * If we are locking a sleepable lock and this lock
1256 			 * isn't sleepable, we want to treat it as a lock
1257 			 * order violation to enfore a general lock order of
1258 			 * sleepable locks before non-sleepable locks.
1259 			 */
1260 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1261 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1262 				goto reversal;
1263 
1264 			/*
1265 			 * If we are locking Giant and this is a non-sleepable
1266 			 * lock, then treat it as a reversal.
1267 			 */
1268 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1269 			    lock == &Giant.lock_object)
1270 				goto reversal;
1271 
1272 			/*
1273 			 * Check the lock order hierarchy for a reveresal.
1274 			 */
1275 			if (!isitmydescendant(w, w1))
1276 				continue;
1277 		reversal:
1278 
1279 			/*
1280 			 * We have a lock order violation, check to see if it
1281 			 * is allowed or has already been yelled about.
1282 			 */
1283 #ifdef BLESSING
1284 
1285 			/*
1286 			 * If the lock order is blessed, just bail.  We don't
1287 			 * look for other lock order violations though, which
1288 			 * may be a bug.
1289 			 */
1290 			if (blessed(w, w1))
1291 				goto out;
1292 #endif
1293 
1294 			/* Bail if this violation is known */
1295 			if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1296 				goto out;
1297 
1298 			/* Record this as a violation */
1299 			w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1300 			w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1301 			w->w_reversed = w1->w_reversed = 1;
1302 			witness_increment_graph_generation();
1303 			mtx_unlock_spin(&w_mtx);
1304 
1305 #ifdef WITNESS_NO_VNODE
1306 			/*
1307 			 * There are known LORs between VNODE locks. They are
1308 			 * not an indication of a bug. VNODE locks are flagged
1309 			 * as such (LO_IS_VNODE) and we don't yell if the LOR
1310 			 * is between 2 VNODE locks.
1311 			 */
1312 			if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1313 			    (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1314 				return;
1315 #endif
1316 
1317 			/*
1318 			 * Ok, yell about it.
1319 			 */
1320 			if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1321 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1322 				printf(
1323 		"lock order reversal: (sleepable after non-sleepable)\n");
1324 			else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1325 			    && lock == &Giant.lock_object)
1326 				printf(
1327 		"lock order reversal: (Giant after non-sleepable)\n");
1328 			else
1329 				printf("lock order reversal:\n");
1330 
1331 			/*
1332 			 * Try to locate an earlier lock with
1333 			 * witness w in our list.
1334 			 */
1335 			do {
1336 				lock2 = &lle->ll_children[i];
1337 				MPASS(lock2->li_lock != NULL);
1338 				if (lock2->li_lock->lo_witness == w)
1339 					break;
1340 				if (i == 0 && lle->ll_next != NULL) {
1341 					lle = lle->ll_next;
1342 					i = lle->ll_count - 1;
1343 					MPASS(i >= 0 && i < LOCK_NCHILDREN);
1344 				} else
1345 					i--;
1346 			} while (i >= 0);
1347 			if (i < 0) {
1348 				printf(" 1st %p %s (%s) @ %s:%d\n",
1349 				    lock1->li_lock, lock1->li_lock->lo_name,
1350 				    w1->w_name, fixup_filename(lock1->li_file),
1351 				    lock1->li_line);
1352 				printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1353 				    lock->lo_name, w->w_name,
1354 				    fixup_filename(file), line);
1355 			} else {
1356 				printf(" 1st %p %s (%s) @ %s:%d\n",
1357 				    lock2->li_lock, lock2->li_lock->lo_name,
1358 				    lock2->li_lock->lo_witness->w_name,
1359 				    fixup_filename(lock2->li_file),
1360 				    lock2->li_line);
1361 				printf(" 2nd %p %s (%s) @ %s:%d\n",
1362 				    lock1->li_lock, lock1->li_lock->lo_name,
1363 				    w1->w_name, fixup_filename(lock1->li_file),
1364 				    lock1->li_line);
1365 				printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1366 				    lock->lo_name, w->w_name,
1367 				    fixup_filename(file), line);
1368 			}
1369 			witness_debugger(1);
1370 			return;
1371 		}
1372 	}
1373 
1374 	/*
1375 	 * If requested, build a new lock order.  However, don't build a new
1376 	 * relationship between a sleepable lock and Giant if it is in the
1377 	 * wrong direction.  The correct lock order is that sleepable locks
1378 	 * always come before Giant.
1379 	 */
1380 	if (flags & LOP_NEWORDER &&
1381 	    !(plock->li_lock == &Giant.lock_object &&
1382 	    (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1383 		CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1384 		    w->w_name, plock->li_lock->lo_witness->w_name);
1385 		itismychild(plock->li_lock->lo_witness, w);
1386 	}
1387 out:
1388 	mtx_unlock_spin(&w_mtx);
1389 }
1390 
1391 void
1392 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1393 {
1394 	struct lock_list_entry **lock_list, *lle;
1395 	struct lock_instance *instance;
1396 	struct witness *w;
1397 	struct thread *td;
1398 
1399 	if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1400 	    panicstr != NULL)
1401 		return;
1402 	w = lock->lo_witness;
1403 	td = curthread;
1404 
1405 	/* Determine lock list for this lock. */
1406 	if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1407 		lock_list = &td->td_sleeplocks;
1408 	else
1409 		lock_list = PCPU_PTR(spinlocks);
1410 
1411 	/* Check to see if we are recursing on a lock we already own. */
1412 	instance = find_instance(*lock_list, lock);
1413 	if (instance != NULL) {
1414 		instance->li_flags++;
1415 		CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1416 		    td->td_proc->p_pid, lock->lo_name,
1417 		    instance->li_flags & LI_RECURSEMASK);
1418 		instance->li_file = file;
1419 		instance->li_line = line;
1420 		return;
1421 	}
1422 
1423 	/* Update per-witness last file and line acquire. */
1424 	w->w_file = file;
1425 	w->w_line = line;
1426 
1427 	/* Find the next open lock instance in the list and fill it. */
1428 	lle = *lock_list;
1429 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1430 		lle = witness_lock_list_get();
1431 		if (lle == NULL)
1432 			return;
1433 		lle->ll_next = *lock_list;
1434 		CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1435 		    td->td_proc->p_pid, lle);
1436 		*lock_list = lle;
1437 	}
1438 	instance = &lle->ll_children[lle->ll_count++];
1439 	instance->li_lock = lock;
1440 	instance->li_line = line;
1441 	instance->li_file = file;
1442 	if ((flags & LOP_EXCLUSIVE) != 0)
1443 		instance->li_flags = LI_EXCLUSIVE;
1444 	else
1445 		instance->li_flags = 0;
1446 	CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1447 	    td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1448 }
1449 
1450 void
1451 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1452 {
1453 	struct lock_instance *instance;
1454 	struct lock_class *class;
1455 
1456 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1457 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1458 		return;
1459 	class = LOCK_CLASS(lock);
1460 	if (witness_watch) {
1461 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1462 			kassert_panic(
1463 			    "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1464 			    class->lc_name, lock->lo_name,
1465 			    fixup_filename(file), line);
1466 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1467 			kassert_panic(
1468 			    "upgrade of non-sleep lock (%s) %s @ %s:%d",
1469 			    class->lc_name, lock->lo_name,
1470 			    fixup_filename(file), line);
1471 	}
1472 	instance = find_instance(curthread->td_sleeplocks, lock);
1473 	if (instance == NULL) {
1474 		kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1475 		    class->lc_name, lock->lo_name,
1476 		    fixup_filename(file), line);
1477 		return;
1478 	}
1479 	if (witness_watch) {
1480 		if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1481 			kassert_panic(
1482 			    "upgrade of exclusive lock (%s) %s @ %s:%d",
1483 			    class->lc_name, lock->lo_name,
1484 			    fixup_filename(file), line);
1485 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1486 			kassert_panic(
1487 			    "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1488 			    class->lc_name, lock->lo_name,
1489 			    instance->li_flags & LI_RECURSEMASK,
1490 			    fixup_filename(file), line);
1491 	}
1492 	instance->li_flags |= LI_EXCLUSIVE;
1493 }
1494 
1495 void
1496 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1497     int line)
1498 {
1499 	struct lock_instance *instance;
1500 	struct lock_class *class;
1501 
1502 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1503 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1504 		return;
1505 	class = LOCK_CLASS(lock);
1506 	if (witness_watch) {
1507 		if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1508 			kassert_panic(
1509 			    "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1510 			    class->lc_name, lock->lo_name,
1511 			    fixup_filename(file), line);
1512 		if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1513 			kassert_panic(
1514 			    "downgrade of non-sleep lock (%s) %s @ %s:%d",
1515 			    class->lc_name, lock->lo_name,
1516 			    fixup_filename(file), line);
1517 	}
1518 	instance = find_instance(curthread->td_sleeplocks, lock);
1519 	if (instance == NULL) {
1520 		kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1521 		    class->lc_name, lock->lo_name,
1522 		    fixup_filename(file), line);
1523 		return;
1524 	}
1525 	if (witness_watch) {
1526 		if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1527 			kassert_panic(
1528 			    "downgrade of shared lock (%s) %s @ %s:%d",
1529 			    class->lc_name, lock->lo_name,
1530 			    fixup_filename(file), line);
1531 		if ((instance->li_flags & LI_RECURSEMASK) != 0)
1532 			kassert_panic(
1533 			    "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1534 			    class->lc_name, lock->lo_name,
1535 			    instance->li_flags & LI_RECURSEMASK,
1536 			    fixup_filename(file), line);
1537 	}
1538 	instance->li_flags &= ~LI_EXCLUSIVE;
1539 }
1540 
1541 void
1542 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1543 {
1544 	struct lock_list_entry **lock_list, *lle;
1545 	struct lock_instance *instance;
1546 	struct lock_class *class;
1547 	struct thread *td;
1548 	register_t s;
1549 	int i, j;
1550 
1551 	if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1552 		return;
1553 	td = curthread;
1554 	class = LOCK_CLASS(lock);
1555 
1556 	/* Find lock instance associated with this lock. */
1557 	if (class->lc_flags & LC_SLEEPLOCK)
1558 		lock_list = &td->td_sleeplocks;
1559 	else
1560 		lock_list = PCPU_PTR(spinlocks);
1561 	lle = *lock_list;
1562 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1563 		for (i = 0; i < (*lock_list)->ll_count; i++) {
1564 			instance = &(*lock_list)->ll_children[i];
1565 			if (instance->li_lock == lock)
1566 				goto found;
1567 		}
1568 
1569 	/*
1570 	 * When disabling WITNESS through witness_watch we could end up in
1571 	 * having registered locks in the td_sleeplocks queue.
1572 	 * We have to make sure we flush these queues, so just search for
1573 	 * eventual register locks and remove them.
1574 	 */
1575 	if (witness_watch > 0) {
1576 		kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1577 		    lock->lo_name, fixup_filename(file), line);
1578 		return;
1579 	} else {
1580 		return;
1581 	}
1582 found:
1583 
1584 	/* First, check for shared/exclusive mismatches. */
1585 	if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1586 	    (flags & LOP_EXCLUSIVE) == 0) {
1587 		printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1588 		    lock->lo_name, fixup_filename(file), line);
1589 		printf("while exclusively locked from %s:%d\n",
1590 		    fixup_filename(instance->li_file), instance->li_line);
1591 		kassert_panic("excl->ushare");
1592 	}
1593 	if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1594 	    (flags & LOP_EXCLUSIVE) != 0) {
1595 		printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1596 		    lock->lo_name, fixup_filename(file), line);
1597 		printf("while share locked from %s:%d\n",
1598 		    fixup_filename(instance->li_file),
1599 		    instance->li_line);
1600 		kassert_panic("share->uexcl");
1601 	}
1602 	/* If we are recursed, unrecurse. */
1603 	if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1604 		CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1605 		    td->td_proc->p_pid, instance->li_lock->lo_name,
1606 		    instance->li_flags);
1607 		instance->li_flags--;
1608 		return;
1609 	}
1610 	/* The lock is now being dropped, check for NORELEASE flag */
1611 	if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1612 		printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1613 		    lock->lo_name, fixup_filename(file), line);
1614 		kassert_panic("lock marked norelease");
1615 	}
1616 
1617 	/* Otherwise, remove this item from the list. */
1618 	s = intr_disable();
1619 	CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1620 	    td->td_proc->p_pid, instance->li_lock->lo_name,
1621 	    (*lock_list)->ll_count - 1);
1622 	for (j = i; j < (*lock_list)->ll_count - 1; j++)
1623 		(*lock_list)->ll_children[j] =
1624 		    (*lock_list)->ll_children[j + 1];
1625 	(*lock_list)->ll_count--;
1626 	intr_restore(s);
1627 
1628 	/*
1629 	 * In order to reduce contention on w_mtx, we want to keep always an
1630 	 * head object into lists so that frequent allocation from the
1631 	 * free witness pool (and subsequent locking) is avoided.
1632 	 * In order to maintain the current code simple, when the head
1633 	 * object is totally unloaded it means also that we do not have
1634 	 * further objects in the list, so the list ownership needs to be
1635 	 * hand over to another object if the current head needs to be freed.
1636 	 */
1637 	if ((*lock_list)->ll_count == 0) {
1638 		if (*lock_list == lle) {
1639 			if (lle->ll_next == NULL)
1640 				return;
1641 		} else
1642 			lle = *lock_list;
1643 		*lock_list = lle->ll_next;
1644 		CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1645 		    td->td_proc->p_pid, lle);
1646 		witness_lock_list_free(lle);
1647 	}
1648 }
1649 
1650 void
1651 witness_thread_exit(struct thread *td)
1652 {
1653 	struct lock_list_entry *lle;
1654 	int i, n;
1655 
1656 	lle = td->td_sleeplocks;
1657 	if (lle == NULL || panicstr != NULL)
1658 		return;
1659 	if (lle->ll_count != 0) {
1660 		for (n = 0; lle != NULL; lle = lle->ll_next)
1661 			for (i = lle->ll_count - 1; i >= 0; i--) {
1662 				if (n == 0)
1663 		printf("Thread %p exiting with the following locks held:\n",
1664 					    td);
1665 				n++;
1666 				witness_list_lock(&lle->ll_children[i], printf);
1667 
1668 			}
1669 		kassert_panic(
1670 		    "Thread %p cannot exit while holding sleeplocks\n", td);
1671 	}
1672 	witness_lock_list_free(lle);
1673 }
1674 
1675 /*
1676  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1677  * exempt Giant and sleepable locks from the checks as well.  If any
1678  * non-exempt locks are held, then a supplied message is printed to the
1679  * console along with a list of the offending locks.  If indicated in the
1680  * flags then a failure results in a panic as well.
1681  */
1682 int
1683 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1684 {
1685 	struct lock_list_entry *lock_list, *lle;
1686 	struct lock_instance *lock1;
1687 	struct thread *td;
1688 	va_list ap;
1689 	int i, n;
1690 
1691 	if (witness_cold || witness_watch < 1 || panicstr != NULL)
1692 		return (0);
1693 	n = 0;
1694 	td = curthread;
1695 	for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1696 		for (i = lle->ll_count - 1; i >= 0; i--) {
1697 			lock1 = &lle->ll_children[i];
1698 			if (lock1->li_lock == lock)
1699 				continue;
1700 			if (flags & WARN_GIANTOK &&
1701 			    lock1->li_lock == &Giant.lock_object)
1702 				continue;
1703 			if (flags & WARN_SLEEPOK &&
1704 			    (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1705 				continue;
1706 			if (n == 0) {
1707 				va_start(ap, fmt);
1708 				vprintf(fmt, ap);
1709 				va_end(ap);
1710 				printf(" with the following");
1711 				if (flags & WARN_SLEEPOK)
1712 					printf(" non-sleepable");
1713 				printf(" locks held:\n");
1714 			}
1715 			n++;
1716 			witness_list_lock(lock1, printf);
1717 		}
1718 
1719 	/*
1720 	 * Pin the thread in order to avoid problems with thread migration.
1721 	 * Once that all verifies are passed about spinlocks ownership,
1722 	 * the thread is in a safe path and it can be unpinned.
1723 	 */
1724 	sched_pin();
1725 	lock_list = PCPU_GET(spinlocks);
1726 	if (lock_list != NULL && lock_list->ll_count != 0) {
1727 		sched_unpin();
1728 
1729 		/*
1730 		 * We should only have one spinlock and as long as
1731 		 * the flags cannot match for this locks class,
1732 		 * check if the first spinlock is the one curthread
1733 		 * should hold.
1734 		 */
1735 		lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1736 		if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1737 		    lock1->li_lock == lock && n == 0)
1738 			return (0);
1739 
1740 		va_start(ap, fmt);
1741 		vprintf(fmt, ap);
1742 		va_end(ap);
1743 		printf(" with the following");
1744 		if (flags & WARN_SLEEPOK)
1745 			printf(" non-sleepable");
1746 		printf(" locks held:\n");
1747 		n += witness_list_locks(&lock_list, printf);
1748 	} else
1749 		sched_unpin();
1750 	if (flags & WARN_PANIC && n)
1751 		kassert_panic("%s", __func__);
1752 	else
1753 		witness_debugger(n);
1754 	return (n);
1755 }
1756 
1757 const char *
1758 witness_file(struct lock_object *lock)
1759 {
1760 	struct witness *w;
1761 
1762 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1763 		return ("?");
1764 	w = lock->lo_witness;
1765 	return (w->w_file);
1766 }
1767 
1768 int
1769 witness_line(struct lock_object *lock)
1770 {
1771 	struct witness *w;
1772 
1773 	if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1774 		return (0);
1775 	w = lock->lo_witness;
1776 	return (w->w_line);
1777 }
1778 
1779 static struct witness *
1780 enroll(const char *description, struct lock_class *lock_class)
1781 {
1782 	struct witness *w;
1783 	struct witness_list *typelist;
1784 
1785 	MPASS(description != NULL);
1786 
1787 	if (witness_watch == -1 || panicstr != NULL)
1788 		return (NULL);
1789 	if ((lock_class->lc_flags & LC_SPINLOCK)) {
1790 		if (witness_skipspin)
1791 			return (NULL);
1792 		else
1793 			typelist = &w_spin;
1794 	} else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1795 		typelist = &w_sleep;
1796 	} else {
1797 		kassert_panic("lock class %s is not sleep or spin",
1798 		    lock_class->lc_name);
1799 		return (NULL);
1800 	}
1801 
1802 	mtx_lock_spin(&w_mtx);
1803 	w = witness_hash_get(description);
1804 	if (w)
1805 		goto found;
1806 	if ((w = witness_get()) == NULL)
1807 		return (NULL);
1808 	MPASS(strlen(description) < MAX_W_NAME);
1809 	strcpy(w->w_name, description);
1810 	w->w_class = lock_class;
1811 	w->w_refcount = 1;
1812 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
1813 	if (lock_class->lc_flags & LC_SPINLOCK) {
1814 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1815 		w_spin_cnt++;
1816 	} else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1817 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1818 		w_sleep_cnt++;
1819 	}
1820 
1821 	/* Insert new witness into the hash */
1822 	witness_hash_put(w);
1823 	witness_increment_graph_generation();
1824 	mtx_unlock_spin(&w_mtx);
1825 	return (w);
1826 found:
1827 	w->w_refcount++;
1828 	mtx_unlock_spin(&w_mtx);
1829 	if (lock_class != w->w_class)
1830 		kassert_panic(
1831 			"lock (%s) %s does not match earlier (%s) lock",
1832 			description, lock_class->lc_name,
1833 			w->w_class->lc_name);
1834 	return (w);
1835 }
1836 
1837 static void
1838 depart(struct witness *w)
1839 {
1840 	struct witness_list *list;
1841 
1842 	MPASS(w->w_refcount == 0);
1843 	if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1844 		list = &w_sleep;
1845 		w_sleep_cnt--;
1846 	} else {
1847 		list = &w_spin;
1848 		w_spin_cnt--;
1849 	}
1850 	/*
1851 	 * Set file to NULL as it may point into a loadable module.
1852 	 */
1853 	w->w_file = NULL;
1854 	w->w_line = 0;
1855 	witness_increment_graph_generation();
1856 }
1857 
1858 
1859 static void
1860 adopt(struct witness *parent, struct witness *child)
1861 {
1862 	int pi, ci, i, j;
1863 
1864 	if (witness_cold == 0)
1865 		mtx_assert(&w_mtx, MA_OWNED);
1866 
1867 	/* If the relationship is already known, there's no work to be done. */
1868 	if (isitmychild(parent, child))
1869 		return;
1870 
1871 	/* When the structure of the graph changes, bump up the generation. */
1872 	witness_increment_graph_generation();
1873 
1874 	/*
1875 	 * The hard part ... create the direct relationship, then propagate all
1876 	 * indirect relationships.
1877 	 */
1878 	pi = parent->w_index;
1879 	ci = child->w_index;
1880 	WITNESS_INDEX_ASSERT(pi);
1881 	WITNESS_INDEX_ASSERT(ci);
1882 	MPASS(pi != ci);
1883 	w_rmatrix[pi][ci] |= WITNESS_PARENT;
1884 	w_rmatrix[ci][pi] |= WITNESS_CHILD;
1885 
1886 	/*
1887 	 * If parent was not already an ancestor of child,
1888 	 * then we increment the descendant and ancestor counters.
1889 	 */
1890 	if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1891 		parent->w_num_descendants++;
1892 		child->w_num_ancestors++;
1893 	}
1894 
1895 	/*
1896 	 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1897 	 * an ancestor of 'pi' during this loop.
1898 	 */
1899 	for (i = 1; i <= w_max_used_index; i++) {
1900 		if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1901 		    (i != pi))
1902 			continue;
1903 
1904 		/* Find each descendant of 'i' and mark it as a descendant. */
1905 		for (j = 1; j <= w_max_used_index; j++) {
1906 
1907 			/*
1908 			 * Skip children that are already marked as
1909 			 * descendants of 'i'.
1910 			 */
1911 			if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1912 				continue;
1913 
1914 			/*
1915 			 * We are only interested in descendants of 'ci'. Note
1916 			 * that 'ci' itself is counted as a descendant of 'ci'.
1917 			 */
1918 			if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1919 			    (j != ci))
1920 				continue;
1921 			w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1922 			w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1923 			w_data[i].w_num_descendants++;
1924 			w_data[j].w_num_ancestors++;
1925 
1926 			/*
1927 			 * Make sure we aren't marking a node as both an
1928 			 * ancestor and descendant. We should have caught
1929 			 * this as a lock order reversal earlier.
1930 			 */
1931 			if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1932 			    (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1933 				printf("witness rmatrix paradox! [%d][%d]=%d "
1934 				    "both ancestor and descendant\n",
1935 				    i, j, w_rmatrix[i][j]);
1936 				kdb_backtrace();
1937 				printf("Witness disabled.\n");
1938 				witness_watch = -1;
1939 			}
1940 			if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1941 			    (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1942 				printf("witness rmatrix paradox! [%d][%d]=%d "
1943 				    "both ancestor and descendant\n",
1944 				    j, i, w_rmatrix[j][i]);
1945 				kdb_backtrace();
1946 				printf("Witness disabled.\n");
1947 				witness_watch = -1;
1948 			}
1949 		}
1950 	}
1951 }
1952 
1953 static void
1954 itismychild(struct witness *parent, struct witness *child)
1955 {
1956 	int unlocked;
1957 
1958 	MPASS(child != NULL && parent != NULL);
1959 	if (witness_cold == 0)
1960 		mtx_assert(&w_mtx, MA_OWNED);
1961 
1962 	if (!witness_lock_type_equal(parent, child)) {
1963 		if (witness_cold == 0) {
1964 			unlocked = 1;
1965 			mtx_unlock_spin(&w_mtx);
1966 		} else {
1967 			unlocked = 0;
1968 		}
1969 		kassert_panic(
1970 		    "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1971 		    "the same lock type", __func__, parent->w_name,
1972 		    parent->w_class->lc_name, child->w_name,
1973 		    child->w_class->lc_name);
1974 		if (unlocked)
1975 			mtx_lock_spin(&w_mtx);
1976 	}
1977 	adopt(parent, child);
1978 }
1979 
1980 /*
1981  * Generic code for the isitmy*() functions. The rmask parameter is the
1982  * expected relationship of w1 to w2.
1983  */
1984 static int
1985 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1986 {
1987 	unsigned char r1, r2;
1988 	int i1, i2;
1989 
1990 	i1 = w1->w_index;
1991 	i2 = w2->w_index;
1992 	WITNESS_INDEX_ASSERT(i1);
1993 	WITNESS_INDEX_ASSERT(i2);
1994 	r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1995 	r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1996 
1997 	/* The flags on one better be the inverse of the flags on the other */
1998 	if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1999 		(WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2000 		printf("%s: rmatrix mismatch between %s (index %d) and %s "
2001 		    "(index %d): w_rmatrix[%d][%d] == %hhx but "
2002 		    "w_rmatrix[%d][%d] == %hhx\n",
2003 		    fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2004 		    i2, i1, r2);
2005 		kdb_backtrace();
2006 		printf("Witness disabled.\n");
2007 		witness_watch = -1;
2008 	}
2009 	return (r1 & rmask);
2010 }
2011 
2012 /*
2013  * Checks if @child is a direct child of @parent.
2014  */
2015 static int
2016 isitmychild(struct witness *parent, struct witness *child)
2017 {
2018 
2019 	return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2020 }
2021 
2022 /*
2023  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2024  */
2025 static int
2026 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2027 {
2028 
2029 	return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2030 	    __func__));
2031 }
2032 
2033 #ifdef BLESSING
2034 static int
2035 blessed(struct witness *w1, struct witness *w2)
2036 {
2037 	int i;
2038 	struct witness_blessed *b;
2039 
2040 	for (i = 0; i < blessed_count; i++) {
2041 		b = &blessed_list[i];
2042 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
2043 			if (strcmp(w2->w_name, b->b_lock2) == 0)
2044 				return (1);
2045 			continue;
2046 		}
2047 		if (strcmp(w1->w_name, b->b_lock2) == 0)
2048 			if (strcmp(w2->w_name, b->b_lock1) == 0)
2049 				return (1);
2050 	}
2051 	return (0);
2052 }
2053 #endif
2054 
2055 static struct witness *
2056 witness_get(void)
2057 {
2058 	struct witness *w;
2059 	int index;
2060 
2061 	if (witness_cold == 0)
2062 		mtx_assert(&w_mtx, MA_OWNED);
2063 
2064 	if (witness_watch == -1) {
2065 		mtx_unlock_spin(&w_mtx);
2066 		return (NULL);
2067 	}
2068 	if (STAILQ_EMPTY(&w_free)) {
2069 		witness_watch = -1;
2070 		mtx_unlock_spin(&w_mtx);
2071 		printf("WITNESS: unable to allocate a new witness object\n");
2072 		return (NULL);
2073 	}
2074 	w = STAILQ_FIRST(&w_free);
2075 	STAILQ_REMOVE_HEAD(&w_free, w_list);
2076 	w_free_cnt--;
2077 	index = w->w_index;
2078 	MPASS(index > 0 && index == w_max_used_index+1 &&
2079 	    index < WITNESS_COUNT);
2080 	bzero(w, sizeof(*w));
2081 	w->w_index = index;
2082 	if (index > w_max_used_index)
2083 		w_max_used_index = index;
2084 	return (w);
2085 }
2086 
2087 static void
2088 witness_free(struct witness *w)
2089 {
2090 
2091 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
2092 	w_free_cnt++;
2093 }
2094 
2095 static struct lock_list_entry *
2096 witness_lock_list_get(void)
2097 {
2098 	struct lock_list_entry *lle;
2099 
2100 	if (witness_watch == -1)
2101 		return (NULL);
2102 	mtx_lock_spin(&w_mtx);
2103 	lle = w_lock_list_free;
2104 	if (lle == NULL) {
2105 		witness_watch = -1;
2106 		mtx_unlock_spin(&w_mtx);
2107 		printf("%s: witness exhausted\n", __func__);
2108 		return (NULL);
2109 	}
2110 	w_lock_list_free = lle->ll_next;
2111 	mtx_unlock_spin(&w_mtx);
2112 	bzero(lle, sizeof(*lle));
2113 	return (lle);
2114 }
2115 
2116 static void
2117 witness_lock_list_free(struct lock_list_entry *lle)
2118 {
2119 
2120 	mtx_lock_spin(&w_mtx);
2121 	lle->ll_next = w_lock_list_free;
2122 	w_lock_list_free = lle;
2123 	mtx_unlock_spin(&w_mtx);
2124 }
2125 
2126 static struct lock_instance *
2127 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2128 {
2129 	struct lock_list_entry *lle;
2130 	struct lock_instance *instance;
2131 	int i;
2132 
2133 	for (lle = list; lle != NULL; lle = lle->ll_next)
2134 		for (i = lle->ll_count - 1; i >= 0; i--) {
2135 			instance = &lle->ll_children[i];
2136 			if (instance->li_lock == lock)
2137 				return (instance);
2138 		}
2139 	return (NULL);
2140 }
2141 
2142 static void
2143 witness_list_lock(struct lock_instance *instance,
2144     int (*prnt)(const char *fmt, ...))
2145 {
2146 	struct lock_object *lock;
2147 
2148 	lock = instance->li_lock;
2149 	prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2150 	    "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2151 	if (lock->lo_witness->w_name != lock->lo_name)
2152 		prnt(" (%s)", lock->lo_witness->w_name);
2153 	prnt(" r = %d (%p) locked @ %s:%d\n",
2154 	    instance->li_flags & LI_RECURSEMASK, lock,
2155 	    fixup_filename(instance->li_file), instance->li_line);
2156 }
2157 
2158 #ifdef DDB
2159 static int
2160 witness_thread_has_locks(struct thread *td)
2161 {
2162 
2163 	if (td->td_sleeplocks == NULL)
2164 		return (0);
2165 	return (td->td_sleeplocks->ll_count != 0);
2166 }
2167 
2168 static int
2169 witness_proc_has_locks(struct proc *p)
2170 {
2171 	struct thread *td;
2172 
2173 	FOREACH_THREAD_IN_PROC(p, td) {
2174 		if (witness_thread_has_locks(td))
2175 			return (1);
2176 	}
2177 	return (0);
2178 }
2179 #endif
2180 
2181 int
2182 witness_list_locks(struct lock_list_entry **lock_list,
2183     int (*prnt)(const char *fmt, ...))
2184 {
2185 	struct lock_list_entry *lle;
2186 	int i, nheld;
2187 
2188 	nheld = 0;
2189 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2190 		for (i = lle->ll_count - 1; i >= 0; i--) {
2191 			witness_list_lock(&lle->ll_children[i], prnt);
2192 			nheld++;
2193 		}
2194 	return (nheld);
2195 }
2196 
2197 /*
2198  * This is a bit risky at best.  We call this function when we have timed
2199  * out acquiring a spin lock, and we assume that the other CPU is stuck
2200  * with this lock held.  So, we go groveling around in the other CPU's
2201  * per-cpu data to try to find the lock instance for this spin lock to
2202  * see when it was last acquired.
2203  */
2204 void
2205 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2206     int (*prnt)(const char *fmt, ...))
2207 {
2208 	struct lock_instance *instance;
2209 	struct pcpu *pc;
2210 
2211 	if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2212 		return;
2213 	pc = pcpu_find(owner->td_oncpu);
2214 	instance = find_instance(pc->pc_spinlocks, lock);
2215 	if (instance != NULL)
2216 		witness_list_lock(instance, prnt);
2217 }
2218 
2219 void
2220 witness_save(struct lock_object *lock, const char **filep, int *linep)
2221 {
2222 	struct lock_list_entry *lock_list;
2223 	struct lock_instance *instance;
2224 	struct lock_class *class;
2225 
2226 	/*
2227 	 * This function is used independently in locking code to deal with
2228 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2229 	 * is gone.
2230 	 */
2231 	if (SCHEDULER_STOPPED())
2232 		return;
2233 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2234 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2235 		return;
2236 	class = LOCK_CLASS(lock);
2237 	if (class->lc_flags & LC_SLEEPLOCK)
2238 		lock_list = curthread->td_sleeplocks;
2239 	else {
2240 		if (witness_skipspin)
2241 			return;
2242 		lock_list = PCPU_GET(spinlocks);
2243 	}
2244 	instance = find_instance(lock_list, lock);
2245 	if (instance == NULL) {
2246 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2247 		    class->lc_name, lock->lo_name);
2248 		return;
2249 	}
2250 	*filep = instance->li_file;
2251 	*linep = instance->li_line;
2252 }
2253 
2254 void
2255 witness_restore(struct lock_object *lock, const char *file, int line)
2256 {
2257 	struct lock_list_entry *lock_list;
2258 	struct lock_instance *instance;
2259 	struct lock_class *class;
2260 
2261 	/*
2262 	 * This function is used independently in locking code to deal with
2263 	 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2264 	 * is gone.
2265 	 */
2266 	if (SCHEDULER_STOPPED())
2267 		return;
2268 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2269 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2270 		return;
2271 	class = LOCK_CLASS(lock);
2272 	if (class->lc_flags & LC_SLEEPLOCK)
2273 		lock_list = curthread->td_sleeplocks;
2274 	else {
2275 		if (witness_skipspin)
2276 			return;
2277 		lock_list = PCPU_GET(spinlocks);
2278 	}
2279 	instance = find_instance(lock_list, lock);
2280 	if (instance == NULL)
2281 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2282 		    class->lc_name, lock->lo_name);
2283 	lock->lo_witness->w_file = file;
2284 	lock->lo_witness->w_line = line;
2285 	if (instance == NULL)
2286 		return;
2287 	instance->li_file = file;
2288 	instance->li_line = line;
2289 }
2290 
2291 void
2292 witness_assert(const struct lock_object *lock, int flags, const char *file,
2293     int line)
2294 {
2295 #ifdef INVARIANT_SUPPORT
2296 	struct lock_instance *instance;
2297 	struct lock_class *class;
2298 
2299 	if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2300 		return;
2301 	class = LOCK_CLASS(lock);
2302 	if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2303 		instance = find_instance(curthread->td_sleeplocks, lock);
2304 	else if ((class->lc_flags & LC_SPINLOCK) != 0)
2305 		instance = find_instance(PCPU_GET(spinlocks), lock);
2306 	else {
2307 		kassert_panic("Lock (%s) %s is not sleep or spin!",
2308 		    class->lc_name, lock->lo_name);
2309 		return;
2310 	}
2311 	switch (flags) {
2312 	case LA_UNLOCKED:
2313 		if (instance != NULL)
2314 			kassert_panic("Lock (%s) %s locked @ %s:%d.",
2315 			    class->lc_name, lock->lo_name,
2316 			    fixup_filename(file), line);
2317 		break;
2318 	case LA_LOCKED:
2319 	case LA_LOCKED | LA_RECURSED:
2320 	case LA_LOCKED | LA_NOTRECURSED:
2321 	case LA_SLOCKED:
2322 	case LA_SLOCKED | LA_RECURSED:
2323 	case LA_SLOCKED | LA_NOTRECURSED:
2324 	case LA_XLOCKED:
2325 	case LA_XLOCKED | LA_RECURSED:
2326 	case LA_XLOCKED | LA_NOTRECURSED:
2327 		if (instance == NULL) {
2328 			kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2329 			    class->lc_name, lock->lo_name,
2330 			    fixup_filename(file), line);
2331 			break;
2332 		}
2333 		if ((flags & LA_XLOCKED) != 0 &&
2334 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
2335 			kassert_panic(
2336 			    "Lock (%s) %s not exclusively locked @ %s:%d.",
2337 			    class->lc_name, lock->lo_name,
2338 			    fixup_filename(file), line);
2339 		if ((flags & LA_SLOCKED) != 0 &&
2340 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
2341 			kassert_panic(
2342 			    "Lock (%s) %s exclusively locked @ %s:%d.",
2343 			    class->lc_name, lock->lo_name,
2344 			    fixup_filename(file), line);
2345 		if ((flags & LA_RECURSED) != 0 &&
2346 		    (instance->li_flags & LI_RECURSEMASK) == 0)
2347 			kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2348 			    class->lc_name, lock->lo_name,
2349 			    fixup_filename(file), line);
2350 		if ((flags & LA_NOTRECURSED) != 0 &&
2351 		    (instance->li_flags & LI_RECURSEMASK) != 0)
2352 			kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2353 			    class->lc_name, lock->lo_name,
2354 			    fixup_filename(file), line);
2355 		break;
2356 	default:
2357 		kassert_panic("Invalid lock assertion at %s:%d.",
2358 		    fixup_filename(file), line);
2359 
2360 	}
2361 #endif	/* INVARIANT_SUPPORT */
2362 }
2363 
2364 static void
2365 witness_setflag(struct lock_object *lock, int flag, int set)
2366 {
2367 	struct lock_list_entry *lock_list;
2368 	struct lock_instance *instance;
2369 	struct lock_class *class;
2370 
2371 	if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2372 		return;
2373 	class = LOCK_CLASS(lock);
2374 	if (class->lc_flags & LC_SLEEPLOCK)
2375 		lock_list = curthread->td_sleeplocks;
2376 	else {
2377 		if (witness_skipspin)
2378 			return;
2379 		lock_list = PCPU_GET(spinlocks);
2380 	}
2381 	instance = find_instance(lock_list, lock);
2382 	if (instance == NULL) {
2383 		kassert_panic("%s: lock (%s) %s not locked", __func__,
2384 		    class->lc_name, lock->lo_name);
2385 		return;
2386 	}
2387 
2388 	if (set)
2389 		instance->li_flags |= flag;
2390 	else
2391 		instance->li_flags &= ~flag;
2392 }
2393 
2394 void
2395 witness_norelease(struct lock_object *lock)
2396 {
2397 
2398 	witness_setflag(lock, LI_NORELEASE, 1);
2399 }
2400 
2401 void
2402 witness_releaseok(struct lock_object *lock)
2403 {
2404 
2405 	witness_setflag(lock, LI_NORELEASE, 0);
2406 }
2407 
2408 #ifdef DDB
2409 static void
2410 witness_ddb_list(struct thread *td)
2411 {
2412 
2413 	KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2414 	KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2415 
2416 	if (witness_watch < 1)
2417 		return;
2418 
2419 	witness_list_locks(&td->td_sleeplocks, db_printf);
2420 
2421 	/*
2422 	 * We only handle spinlocks if td == curthread.  This is somewhat broken
2423 	 * if td is currently executing on some other CPU and holds spin locks
2424 	 * as we won't display those locks.  If we had a MI way of getting
2425 	 * the per-cpu data for a given cpu then we could use
2426 	 * td->td_oncpu to get the list of spinlocks for this thread
2427 	 * and "fix" this.
2428 	 *
2429 	 * That still wouldn't really fix this unless we locked the scheduler
2430 	 * lock or stopped the other CPU to make sure it wasn't changing the
2431 	 * list out from under us.  It is probably best to just not try to
2432 	 * handle threads on other CPU's for now.
2433 	 */
2434 	if (td == curthread && PCPU_GET(spinlocks) != NULL)
2435 		witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2436 }
2437 
2438 DB_SHOW_COMMAND(locks, db_witness_list)
2439 {
2440 	struct thread *td;
2441 
2442 	if (have_addr)
2443 		td = db_lookup_thread(addr, TRUE);
2444 	else
2445 		td = kdb_thread;
2446 	witness_ddb_list(td);
2447 }
2448 
2449 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2450 {
2451 	struct thread *td;
2452 	struct proc *p;
2453 
2454 	/*
2455 	 * It would be nice to list only threads and processes that actually
2456 	 * held sleep locks, but that information is currently not exported
2457 	 * by WITNESS.
2458 	 */
2459 	FOREACH_PROC_IN_SYSTEM(p) {
2460 		if (!witness_proc_has_locks(p))
2461 			continue;
2462 		FOREACH_THREAD_IN_PROC(p, td) {
2463 			if (!witness_thread_has_locks(td))
2464 				continue;
2465 			db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2466 			    p->p_comm, td, td->td_tid);
2467 			witness_ddb_list(td);
2468 			if (db_pager_quit)
2469 				return;
2470 		}
2471 	}
2472 }
2473 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2474 
2475 DB_SHOW_COMMAND(witness, db_witness_display)
2476 {
2477 
2478 	witness_ddb_display(db_printf);
2479 }
2480 #endif
2481 
2482 static int
2483 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2484 {
2485 	struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2486 	struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2487 	struct sbuf *sb;
2488 	u_int w_rmatrix1, w_rmatrix2;
2489 	int error, generation, i, j;
2490 
2491 	tmp_data1 = NULL;
2492 	tmp_data2 = NULL;
2493 	tmp_w1 = NULL;
2494 	tmp_w2 = NULL;
2495 	if (witness_watch < 1) {
2496 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2497 		return (error);
2498 	}
2499 	if (witness_cold) {
2500 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2501 		return (error);
2502 	}
2503 	error = 0;
2504 	sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2505 	if (sb == NULL)
2506 		return (ENOMEM);
2507 
2508 	/* Allocate and init temporary storage space. */
2509 	tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2510 	tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2511 	tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2512 	    M_WAITOK | M_ZERO);
2513 	tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2514 	    M_WAITOK | M_ZERO);
2515 	stack_zero(&tmp_data1->wlod_stack);
2516 	stack_zero(&tmp_data2->wlod_stack);
2517 
2518 restart:
2519 	mtx_lock_spin(&w_mtx);
2520 	generation = w_generation;
2521 	mtx_unlock_spin(&w_mtx);
2522 	sbuf_printf(sb, "Number of known direct relationships is %d\n",
2523 	    w_lohash.wloh_count);
2524 	for (i = 1; i < w_max_used_index; i++) {
2525 		mtx_lock_spin(&w_mtx);
2526 		if (generation != w_generation) {
2527 			mtx_unlock_spin(&w_mtx);
2528 
2529 			/* The graph has changed, try again. */
2530 			req->oldidx = 0;
2531 			sbuf_clear(sb);
2532 			goto restart;
2533 		}
2534 
2535 		w1 = &w_data[i];
2536 		if (w1->w_reversed == 0) {
2537 			mtx_unlock_spin(&w_mtx);
2538 			continue;
2539 		}
2540 
2541 		/* Copy w1 locally so we can release the spin lock. */
2542 		*tmp_w1 = *w1;
2543 		mtx_unlock_spin(&w_mtx);
2544 
2545 		if (tmp_w1->w_reversed == 0)
2546 			continue;
2547 		for (j = 1; j < w_max_used_index; j++) {
2548 			if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2549 				continue;
2550 
2551 			mtx_lock_spin(&w_mtx);
2552 			if (generation != w_generation) {
2553 				mtx_unlock_spin(&w_mtx);
2554 
2555 				/* The graph has changed, try again. */
2556 				req->oldidx = 0;
2557 				sbuf_clear(sb);
2558 				goto restart;
2559 			}
2560 
2561 			w2 = &w_data[j];
2562 			data1 = witness_lock_order_get(w1, w2);
2563 			data2 = witness_lock_order_get(w2, w1);
2564 
2565 			/*
2566 			 * Copy information locally so we can release the
2567 			 * spin lock.
2568 			 */
2569 			*tmp_w2 = *w2;
2570 			w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2571 			w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2572 
2573 			if (data1) {
2574 				stack_zero(&tmp_data1->wlod_stack);
2575 				stack_copy(&data1->wlod_stack,
2576 				    &tmp_data1->wlod_stack);
2577 			}
2578 			if (data2 && data2 != data1) {
2579 				stack_zero(&tmp_data2->wlod_stack);
2580 				stack_copy(&data2->wlod_stack,
2581 				    &tmp_data2->wlod_stack);
2582 			}
2583 			mtx_unlock_spin(&w_mtx);
2584 
2585 			sbuf_printf(sb,
2586 	    "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2587 			    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2588 			    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2589 #if 0
2590  			sbuf_printf(sb,
2591 			"w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2592  			    tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2593  			    tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2594 #endif
2595 			if (data1) {
2596 				sbuf_printf(sb,
2597 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2598 				    tmp_w1->w_name, tmp_w1->w_class->lc_name,
2599 				    tmp_w2->w_name, tmp_w2->w_class->lc_name);
2600 				stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2601 				sbuf_printf(sb, "\n");
2602 			}
2603 			if (data2 && data2 != data1) {
2604 				sbuf_printf(sb,
2605 			"Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2606 				    tmp_w2->w_name, tmp_w2->w_class->lc_name,
2607 				    tmp_w1->w_name, tmp_w1->w_class->lc_name);
2608 				stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2609 				sbuf_printf(sb, "\n");
2610 			}
2611 		}
2612 	}
2613 	mtx_lock_spin(&w_mtx);
2614 	if (generation != w_generation) {
2615 		mtx_unlock_spin(&w_mtx);
2616 
2617 		/*
2618 		 * The graph changed while we were printing stack data,
2619 		 * try again.
2620 		 */
2621 		req->oldidx = 0;
2622 		sbuf_clear(sb);
2623 		goto restart;
2624 	}
2625 	mtx_unlock_spin(&w_mtx);
2626 
2627 	/* Free temporary storage space. */
2628 	free(tmp_data1, M_TEMP);
2629 	free(tmp_data2, M_TEMP);
2630 	free(tmp_w1, M_TEMP);
2631 	free(tmp_w2, M_TEMP);
2632 
2633 	sbuf_finish(sb);
2634 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2635 	sbuf_delete(sb);
2636 
2637 	return (error);
2638 }
2639 
2640 static int
2641 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2642 {
2643 	struct witness *w;
2644 	struct sbuf *sb;
2645 	int error;
2646 
2647 	if (witness_watch < 1) {
2648 		error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2649 		return (error);
2650 	}
2651 	if (witness_cold) {
2652 		error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2653 		return (error);
2654 	}
2655 	error = 0;
2656 
2657 	error = sysctl_wire_old_buffer(req, 0);
2658 	if (error != 0)
2659 		return (error);
2660 	sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2661 	if (sb == NULL)
2662 		return (ENOMEM);
2663 	sbuf_printf(sb, "\n");
2664 
2665 	mtx_lock_spin(&w_mtx);
2666 	STAILQ_FOREACH(w, &w_all, w_list)
2667 		w->w_displayed = 0;
2668 	STAILQ_FOREACH(w, &w_all, w_list)
2669 		witness_add_fullgraph(sb, w);
2670 	mtx_unlock_spin(&w_mtx);
2671 
2672 	/*
2673 	 * Close the sbuf and return to userland.
2674 	 */
2675 	error = sbuf_finish(sb);
2676 	sbuf_delete(sb);
2677 
2678 	return (error);
2679 }
2680 
2681 static int
2682 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2683 {
2684 	int error, value;
2685 
2686 	value = witness_watch;
2687 	error = sysctl_handle_int(oidp, &value, 0, req);
2688 	if (error != 0 || req->newptr == NULL)
2689 		return (error);
2690 	if (value > 1 || value < -1 ||
2691 	    (witness_watch == -1 && value != witness_watch))
2692 		return (EINVAL);
2693 	witness_watch = value;
2694 	return (0);
2695 }
2696 
2697 static void
2698 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2699 {
2700 	int i;
2701 
2702 	if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2703 		return;
2704 	w->w_displayed = 1;
2705 
2706 	WITNESS_INDEX_ASSERT(w->w_index);
2707 	for (i = 1; i <= w_max_used_index; i++) {
2708 		if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2709 			sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2710 			    w_data[i].w_name);
2711 			witness_add_fullgraph(sb, &w_data[i]);
2712 		}
2713 	}
2714 }
2715 
2716 /*
2717  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2718  * interprets the key as a string and reads until the null
2719  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2720  * hash value computed from the key.
2721  */
2722 static uint32_t
2723 witness_hash_djb2(const uint8_t *key, uint32_t size)
2724 {
2725 	unsigned int hash = 5381;
2726 	int i;
2727 
2728 	/* hash = hash * 33 + key[i] */
2729 	if (size)
2730 		for (i = 0; i < size; i++)
2731 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2732 	else
2733 		for (i = 0; key[i] != 0; i++)
2734 			hash = ((hash << 5) + hash) + (unsigned int)key[i];
2735 
2736 	return (hash);
2737 }
2738 
2739 
2740 /*
2741  * Initializes the two witness hash tables. Called exactly once from
2742  * witness_initialize().
2743  */
2744 static void
2745 witness_init_hash_tables(void)
2746 {
2747 	int i;
2748 
2749 	MPASS(witness_cold);
2750 
2751 	/* Initialize the hash tables. */
2752 	for (i = 0; i < WITNESS_HASH_SIZE; i++)
2753 		w_hash.wh_array[i] = NULL;
2754 
2755 	w_hash.wh_size = WITNESS_HASH_SIZE;
2756 	w_hash.wh_count = 0;
2757 
2758 	/* Initialize the lock order data hash. */
2759 	w_lofree = NULL;
2760 	for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2761 		memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2762 		w_lodata[i].wlod_next = w_lofree;
2763 		w_lofree = &w_lodata[i];
2764 	}
2765 	w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2766 	w_lohash.wloh_count = 0;
2767 	for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2768 		w_lohash.wloh_array[i] = NULL;
2769 }
2770 
2771 static struct witness *
2772 witness_hash_get(const char *key)
2773 {
2774 	struct witness *w;
2775 	uint32_t hash;
2776 
2777 	MPASS(key != NULL);
2778 	if (witness_cold == 0)
2779 		mtx_assert(&w_mtx, MA_OWNED);
2780 	hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2781 	w = w_hash.wh_array[hash];
2782 	while (w != NULL) {
2783 		if (strcmp(w->w_name, key) == 0)
2784 			goto out;
2785 		w = w->w_hash_next;
2786 	}
2787 
2788 out:
2789 	return (w);
2790 }
2791 
2792 static void
2793 witness_hash_put(struct witness *w)
2794 {
2795 	uint32_t hash;
2796 
2797 	MPASS(w != NULL);
2798 	MPASS(w->w_name != NULL);
2799 	if (witness_cold == 0)
2800 		mtx_assert(&w_mtx, MA_OWNED);
2801 	KASSERT(witness_hash_get(w->w_name) == NULL,
2802 	    ("%s: trying to add a hash entry that already exists!", __func__));
2803 	KASSERT(w->w_hash_next == NULL,
2804 	    ("%s: w->w_hash_next != NULL", __func__));
2805 
2806 	hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2807 	w->w_hash_next = w_hash.wh_array[hash];
2808 	w_hash.wh_array[hash] = w;
2809 	w_hash.wh_count++;
2810 }
2811 
2812 
2813 static struct witness_lock_order_data *
2814 witness_lock_order_get(struct witness *parent, struct witness *child)
2815 {
2816 	struct witness_lock_order_data *data = NULL;
2817 	struct witness_lock_order_key key;
2818 	unsigned int hash;
2819 
2820 	MPASS(parent != NULL && child != NULL);
2821 	key.from = parent->w_index;
2822 	key.to = child->w_index;
2823 	WITNESS_INDEX_ASSERT(key.from);
2824 	WITNESS_INDEX_ASSERT(key.to);
2825 	if ((w_rmatrix[parent->w_index][child->w_index]
2826 	    & WITNESS_LOCK_ORDER_KNOWN) == 0)
2827 		goto out;
2828 
2829 	hash = witness_hash_djb2((const char*)&key,
2830 	    sizeof(key)) % w_lohash.wloh_size;
2831 	data = w_lohash.wloh_array[hash];
2832 	while (data != NULL) {
2833 		if (witness_lock_order_key_equal(&data->wlod_key, &key))
2834 			break;
2835 		data = data->wlod_next;
2836 	}
2837 
2838 out:
2839 	return (data);
2840 }
2841 
2842 /*
2843  * Verify that parent and child have a known relationship, are not the same,
2844  * and child is actually a child of parent.  This is done without w_mtx
2845  * to avoid contention in the common case.
2846  */
2847 static int
2848 witness_lock_order_check(struct witness *parent, struct witness *child)
2849 {
2850 
2851 	if (parent != child &&
2852 	    w_rmatrix[parent->w_index][child->w_index]
2853 	    & WITNESS_LOCK_ORDER_KNOWN &&
2854 	    isitmychild(parent, child))
2855 		return (1);
2856 
2857 	return (0);
2858 }
2859 
2860 static int
2861 witness_lock_order_add(struct witness *parent, struct witness *child)
2862 {
2863 	struct witness_lock_order_data *data = NULL;
2864 	struct witness_lock_order_key key;
2865 	unsigned int hash;
2866 
2867 	MPASS(parent != NULL && child != NULL);
2868 	key.from = parent->w_index;
2869 	key.to = child->w_index;
2870 	WITNESS_INDEX_ASSERT(key.from);
2871 	WITNESS_INDEX_ASSERT(key.to);
2872 	if (w_rmatrix[parent->w_index][child->w_index]
2873 	    & WITNESS_LOCK_ORDER_KNOWN)
2874 		return (1);
2875 
2876 	hash = witness_hash_djb2((const char*)&key,
2877 	    sizeof(key)) % w_lohash.wloh_size;
2878 	w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2879 	data = w_lofree;
2880 	if (data == NULL)
2881 		return (0);
2882 	w_lofree = data->wlod_next;
2883 	data->wlod_next = w_lohash.wloh_array[hash];
2884 	data->wlod_key = key;
2885 	w_lohash.wloh_array[hash] = data;
2886 	w_lohash.wloh_count++;
2887 	stack_zero(&data->wlod_stack);
2888 	stack_save(&data->wlod_stack);
2889 	return (1);
2890 }
2891 
2892 /* Call this whenver the structure of the witness graph changes. */
2893 static void
2894 witness_increment_graph_generation(void)
2895 {
2896 
2897 	if (witness_cold == 0)
2898 		mtx_assert(&w_mtx, MA_OWNED);
2899 	w_generation++;
2900 }
2901 
2902 #ifdef KDB
2903 static void
2904 _witness_debugger(int cond, const char *msg)
2905 {
2906 
2907 	if (witness_trace && cond)
2908 		kdb_backtrace();
2909 	if (witness_kdb && cond)
2910 		kdb_enter(KDB_WHY_WITNESS, msg);
2911 }
2912 #endif
2913