xref: /freebsd/sys/kern/subr_witness.c (revision 5521ff5a4d1929056e7ffc982fac3341ca54df7c)
1 /*-
2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
13  *    promote products derived from this software without specific prior
14  *    written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  *	from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
29  *	and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
30  * $FreeBSD$
31  */
32 
33 /*
34  * Implementation of the `witness' lock verifier.  Originally implemented for
35  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
36  * classes in FreeBSD.
37  */
38 
39 /*
40  *	Main Entry: witness
41  *	Pronunciation: 'wit-n&s
42  *	Function: noun
43  *	Etymology: Middle English witnesse, from Old English witnes knowledge,
44  *	    testimony, witness, from 2wit
45  *	Date: before 12th century
46  *	1 : attestation of a fact or event : TESTIMONY
47  *	2 : one that gives evidence; specifically : one who testifies in
48  *	    a cause or before a judicial tribunal
49  *	3 : one asked to be present at a transaction so as to be able to
50  *	    testify to its having taken place
51  *	4 : one who has personal knowledge of something
52  *	5 a : something serving as evidence or proof : SIGN
53  *	  b : public affirmation by word or example of usually
54  *	      religious faith or conviction <the heroic witness to divine
55  *	      life -- Pilot>
56  *	6 capitalized : a member of the Jehovah's Witnesses
57  */
58 
59 #include "opt_ddb.h"
60 #include "opt_witness.h"
61 
62 #include <sys/param.h>
63 #include <sys/bus.h>
64 #include <sys/kernel.h>
65 #include <sys/ktr.h>
66 #include <sys/lock.h>
67 #include <sys/malloc.h>
68 #include <sys/mutex.h>
69 #include <sys/proc.h>
70 #include <sys/sysctl.h>
71 #include <sys/systm.h>
72 
73 #include <ddb/ddb.h>
74 
75 #define WITNESS_COUNT 200
76 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
77 /*
78  * XXX: This is somewhat bogus, as we assume here that at most 1024 processes
79  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
80  * probably be safe for the most part, but it's still a SWAG.
81  */
82 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
83 
84 #define	WITNESS_NCHILDREN 6
85 
86 struct witness_child_list_entry;
87 
88 struct witness {
89 	const	char *w_name;
90 	struct	lock_class *w_class;
91 	STAILQ_ENTRY(witness) w_list;		/* List of all witnesses. */
92 	STAILQ_ENTRY(witness) w_typelist;	/* Witnesses of a type. */
93 	struct	witness_child_list_entry *w_children;	/* Great evilness... */
94 	const	char *w_file;
95 	int	w_line;
96 	u_int	w_level;
97 	u_int	w_refcount;
98 	u_char	w_Giant_squawked:1;
99 	u_char	w_other_squawked:1;
100 	u_char	w_same_squawked:1;
101 };
102 
103 struct witness_child_list_entry {
104 	struct	witness_child_list_entry *wcl_next;
105 	struct	witness *wcl_children[WITNESS_NCHILDREN];
106 	u_int	wcl_count;
107 };
108 
109 STAILQ_HEAD(witness_list, witness);
110 
111 struct witness_blessed {
112 	const	char *b_lock1;
113 	const	char *b_lock2;
114 };
115 
116 struct witness_order_list_entry {
117 	const	char *w_name;
118 	struct	lock_class *w_class;
119 };
120 
121 static struct	witness *enroll(const char *description,
122 				struct lock_class *lock_class);
123 static int	itismychild(struct witness *parent, struct witness *child);
124 static void	removechild(struct witness *parent, struct witness *child);
125 static int	isitmychild(struct witness *parent, struct witness *child);
126 static int	isitmydescendant(struct witness *parent, struct witness *child);
127 static int	dup_ok(struct witness *);
128 static int	blessed(struct witness *, struct witness *);
129 static void	witness_display_list(void(*prnt)(const char *fmt, ...),
130 				     struct witness_list *list);
131 static void	witness_displaydescendants(void(*)(const char *fmt, ...),
132 					   struct witness *);
133 static void	witness_leveldescendents(struct witness *parent, int level);
134 static void	witness_levelall(void);
135 static struct	witness *witness_get(void);
136 static void	witness_free(struct witness *m);
137 static struct	witness_child_list_entry *witness_child_get(void);
138 static void	witness_child_free(struct witness_child_list_entry *wcl);
139 static struct	lock_list_entry *witness_lock_list_get(void);
140 static void	witness_lock_list_free(struct lock_list_entry *lle);
141 static void	witness_display(void(*)(const char *fmt, ...));
142 static struct	lock_instance *find_instance(struct lock_list_entry *lock_list,
143 					     struct lock_object *lock);
144 
145 MALLOC_DEFINE(M_WITNESS, "witness", "witness structure");
146 
147 static int witness_watch = 1;
148 TUNABLE_INT("debug.witness_watch", &witness_watch);
149 SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, "");
150 
151 #ifdef DDB
152 /*
153  * When DDB is enabled and witness_ddb is set to 1, it will cause the system to
154  * drop into kdebug() when:
155  *	- a lock heirarchy violation occurs
156  *	- locks are held when going to sleep.
157  */
158 #ifdef WITNESS_DDB
159 int	witness_ddb = 1;
160 #else
161 int	witness_ddb = 0;
162 #endif
163 TUNABLE_INT("debug.witness_ddb", &witness_ddb);
164 SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, "");
165 #endif /* DDB */
166 
167 #ifdef WITNESS_SKIPSPIN
168 int	witness_skipspin = 1;
169 #else
170 int	witness_skipspin = 0;
171 #endif
172 TUNABLE_INT("debug.witness_skipspin", &witness_skipspin);
173 SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0,
174     "");
175 
176 static struct mtx w_mtx;
177 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
178 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
179 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
180 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
181 static struct witness_child_list_entry *w_child_free = NULL;
182 static struct lock_list_entry *w_lock_list_free = NULL;
183 static int witness_dead;	/* fatal error, probably no memory */
184 
185 static struct witness w_data[WITNESS_COUNT];
186 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
187 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
188 
189 static struct witness_order_list_entry order_lists[] = {
190 	{ "Giant", &lock_class_mtx_sleep },
191 	{ "proctree", &lock_class_sx },
192 	{ "allproc", &lock_class_sx },
193 	{ "process lock", &lock_class_mtx_sleep },
194 	{ "uidinfo hash", &lock_class_mtx_sleep },
195 	{ "uidinfo struct", &lock_class_mtx_sleep },
196 	{ NULL, NULL },
197 	/*
198 	 * spin locks
199 	 */
200 #if defined(__i386__) && defined (SMP)
201 	{ "com", &lock_class_mtx_spin },
202 #endif
203 	{ "sio", &lock_class_mtx_spin },
204 #ifdef __i386__
205 	{ "cy", &lock_class_mtx_spin },
206 #endif
207 	{ "ng_node", &lock_class_mtx_spin },
208 	{ "ng_worklist", &lock_class_mtx_spin },
209 	{ "ithread table lock", &lock_class_mtx_spin },
210 	{ "sched lock", &lock_class_mtx_spin },
211 	{ "callout", &lock_class_mtx_spin },
212 	/*
213 	 * leaf locks
214 	 */
215 #ifdef SMP
216 	{ "ap boot", &lock_class_mtx_spin },
217 #ifdef __i386__
218 	{ "imen", &lock_class_mtx_spin },
219 #endif
220 	{ "smp rendezvous", &lock_class_mtx_spin },
221 #endif
222 	{ "clk", &lock_class_mtx_spin },
223 	{ NULL, NULL },
224 	{ NULL, NULL }
225 };
226 
227 static const char *dup_list[] = {
228 	"process lock",
229 	NULL
230 };
231 
232 /*
233  * Pairs of locks which have been blessed
234  * Don't complain about order problems with blessed locks
235  */
236 static struct witness_blessed blessed_list[] = {
237 };
238 static int blessed_count =
239 	sizeof(blessed_list) / sizeof(struct witness_blessed);
240 
241 /*
242  * List of all locks in the system.
243  */
244 STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks);
245 
246 static struct mtx all_mtx = {
247 	{ &lock_class_mtx_sleep,	/* mtx_object.lo_class */
248 	  "All locks list",		/* mtx_object.lo_name */
249 	  LO_INITIALIZED,		/* mtx_object.lo_flags */
250 	  { NULL },			/* mtx_object.lo_list */
251 	  NULL },			/* mtx_object.lo_witness */
252 	MTX_UNOWNED, 0,			/* mtx_lock, mtx_recurse */
253 	0,				/* mtx_savecrit */
254 	TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked),
255 	{ NULL, NULL }			/* mtx_contested */
256 };
257 
258 /*
259  * This global is set to 0 once it becomes safe to use the witness code.
260  */
261 static int witness_cold = 1;
262 
263 /*
264  * Global variables for book keeping.
265  */
266 static int lock_cur_cnt;
267 static int lock_max_cnt;
268 
269 /*
270  * The WITNESS-enabled diagnostic code.
271  */
272 static void
273 witness_initialize(void *dummy __unused)
274 {
275 	struct lock_object *lock;
276 	struct witness_order_list_entry *order;
277 	struct witness *w, *w1;
278 	int i;
279 
280 	/*
281 	 * We have to release Giant before initializing its witness
282 	 * structure so that WITNESS doesn't get confused.
283 	 */
284 	mtx_unlock(&Giant);
285 	mtx_assert(&Giant, MA_NOTOWNED);
286 
287 	CTR0(KTR_WITNESS, __func__ ": initializing witness");
288 	STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
289 	mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS);
290 	for (i = 0; i < WITNESS_COUNT; i++)
291 		witness_free(&w_data[i]);
292 	for (i = 0; i < WITNESS_CHILDCOUNT; i++)
293 		witness_child_free(&w_childdata[i]);
294 	for (i = 0; i < LOCK_CHILDCOUNT; i++)
295 		witness_lock_list_free(&w_locklistdata[i]);
296 
297 	/* First add in all the specified order lists. */
298 	for (order = order_lists; order->w_name != NULL; order++) {
299 		w = enroll(order->w_name, order->w_class);
300 		if (w == NULL)
301 			continue;
302 		w->w_file = "order list";
303 		for (order++; order->w_name != NULL; order++) {
304 			w1 = enroll(order->w_name, order->w_class);
305 			if (w1 == NULL)
306 				continue;
307 			w1->w_file = "order list";
308 			itismychild(w, w1);
309 			w = w1;
310 		}
311 	}
312 
313 	/* Iterate through all locks and add them to witness. */
314 	mtx_lock(&all_mtx);
315 	STAILQ_FOREACH(lock, &all_locks, lo_list) {
316 		if (lock->lo_flags & LO_WITNESS)
317 			lock->lo_witness = enroll(lock->lo_name,
318 			    lock->lo_class);
319 		else
320 			lock->lo_witness = NULL;
321 	}
322 	mtx_unlock(&all_mtx);
323 
324 	/* Mark the witness code as being ready for use. */
325 	atomic_store_rel_int(&witness_cold, 0);
326 
327 	mtx_lock(&Giant);
328 }
329 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
330 
331 void
332 witness_init(struct lock_object *lock)
333 {
334 	struct lock_class *class;
335 
336 	class = lock->lo_class;
337 	if (lock->lo_flags & LO_INITIALIZED)
338 		panic("%s: lock (%s) %s is already initialized!\n", __func__,
339 		    class->lc_name, lock->lo_name);
340 
341 	if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
342 	    (class->lc_flags & LC_RECURSABLE) == 0)
343 		panic("%s: lock (%s) %s can not be recursable!\n", __func__,
344 		    class->lc_name, lock->lo_name);
345 
346 	if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
347 	    (class->lc_flags & LC_SLEEPABLE) == 0)
348 		panic("%s: lock (%s) %s can not be sleepable!\n", __func__,
349 		    class->lc_name, lock->lo_name);
350 
351 	mtx_lock(&all_mtx);
352 	STAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
353 	lock->lo_flags |= LO_INITIALIZED;
354 	lock_cur_cnt++;
355 	if (lock_cur_cnt > lock_max_cnt)
356 		lock_max_cnt = lock_cur_cnt;
357 	mtx_unlock(&all_mtx);
358 	if (!witness_cold && !witness_dead &&
359 	    (lock->lo_flags & LO_WITNESS) != 0)
360 		lock->lo_witness = enroll(lock->lo_name, class);
361 	else
362 		lock->lo_witness = NULL;
363 }
364 
365 void
366 witness_destroy(struct lock_object *lock)
367 {
368 	struct witness *w;
369 
370 	if (witness_cold)
371 		panic("lock (%s) %s destroyed while witness_cold",
372 		    lock->lo_class->lc_name, lock->lo_name);
373 
374 	if ((lock->lo_flags & LO_INITIALIZED) == 0)
375 		panic("%s: lock (%s) %s is not initialized!\n", __func__,
376 		    lock->lo_class->lc_name, lock->lo_name);
377 
378 	/* XXX: need to verify that no one holds the lock */
379 	w = lock->lo_witness;
380 	if (w != NULL) {
381 		mtx_lock_spin(&w_mtx);
382 		w->w_refcount--;
383 		if (w->w_refcount == 0) {
384 			CTR1(KTR_WITNESS,
385 			    __func__ ": marking witness %s as dead", w->w_name);
386 			w->w_name = "(dead)";
387 			w->w_file = "(dead)";
388 			w->w_line = 0;
389 		}
390 		mtx_unlock_spin(&w_mtx);
391 	}
392 
393 	mtx_lock(&all_mtx);
394 	lock_cur_cnt--;
395 	STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list);
396 	lock->lo_flags &= LO_INITIALIZED;
397 	mtx_unlock(&all_mtx);
398 }
399 
400 static void
401 witness_display_list(void(*prnt)(const char *fmt, ...),
402 		     struct witness_list *list)
403 {
404 	struct witness *w, *w1;
405 	int found;
406 
407 	STAILQ_FOREACH(w, list, w_typelist) {
408 		if (w->w_file == NULL)
409 			continue;
410 		found = 0;
411 		STAILQ_FOREACH(w1, list, w_typelist) {
412 			if (isitmychild(w1, w)) {
413 				found++;
414 				break;
415 			}
416 		}
417 		if (found)
418 			continue;
419 		/*
420 		 * This lock has no anscestors, display its descendants.
421 		 */
422 		witness_displaydescendants(prnt, w);
423 	}
424 }
425 
426 static void
427 witness_display(void(*prnt)(const char *fmt, ...))
428 {
429 	struct witness *w;
430 
431 	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
432 	witness_levelall();
433 
434 	/*
435 	 * First, handle sleep locks which have been acquired at least
436 	 * once.
437 	 */
438 	prnt("Sleep locks:\n");
439 	witness_display_list(prnt, &w_sleep);
440 
441 	/*
442 	 * Now do spin locks which have been acquired at least once.
443 	 */
444 	prnt("\nSpin locks:\n");
445 	witness_display_list(prnt, &w_spin);
446 
447 	/*
448 	 * Finally, any locks which have not been acquired yet.
449 	 */
450 	prnt("\nLocks which were never acquired:\n");
451 	STAILQ_FOREACH(w, &w_all, w_list) {
452 		if (w->w_file != NULL)
453 			continue;
454 		prnt("%s\n", w->w_name);
455 	}
456 }
457 
458 void
459 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
460 {
461 	struct lock_list_entry **lock_list, *lle;
462 	struct lock_instance *lock1, *lock2;
463 	struct lock_class *class;
464 	struct witness *w, *w1;
465 	struct proc *p;
466 	int i, j;
467 #ifdef DDB
468 	int go_into_ddb = 0;
469 #endif /* DDB */
470 
471 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
472 	    panicstr)
473 		return;
474 	w = lock->lo_witness;
475 	class = lock->lo_class;
476 	p = curproc;
477 
478 	/*
479 	 * We have to hold a spinlock to keep lock_list valid across the check
480 	 * in the LC_SLEEPLOCK case.  In the LC_SPINLOCK case, it is already
481 	 * protected by the spinlock we are currently performing the witness
482 	 * checks on, so it is ok to release the lock after performing this
483 	 * check.  All we have to protect is the LC_SLEEPLOCK case when no
484 	 * spinlocks are held as we may get preempted during this check and
485 	 * lock_list could end up pointing to some other CPU's spinlock list.
486 	 */
487 	mtx_lock_spin(&w_mtx);
488 	lock_list = PCPU_PTR(spinlocks);
489 	if (class->lc_flags & LC_SLEEPLOCK) {
490 		if (*lock_list != NULL && (flags & LOP_TRYLOCK) == 0) {
491 			mtx_unlock_spin(&w_mtx);
492 			panic("blockable sleep lock (%s) %s @ %s:%d",
493 			    class->lc_name, lock->lo_name, file, line);
494 		}
495 		lock_list = &p->p_sleeplocks;
496 	}
497 	mtx_unlock_spin(&w_mtx);
498 
499 	/*
500 	 * Try locks do not block if they fail to acquire the lock, thus
501 	 * there is no danger of deadlocks or of switching while holding a
502 	 * spin lock if we acquire a lock via a try operation.
503 	 */
504 	if (flags & LOP_TRYLOCK)
505 		goto out;
506 
507 	/*
508 	 * Is this the first lock acquired?  If so, then no order checking
509 	 * is needed.
510 	 */
511 	if (*lock_list == NULL)
512 		goto out;
513 
514 	/*
515 	 * Check to see if we are recursing on a lock we already own.
516 	 */
517 	lock1 = find_instance(*lock_list, lock);
518 	if (lock1 != NULL) {
519 		if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
520 		    (flags & LOP_EXCLUSIVE) == 0) {
521 			printf("shared lock of (%s) %s @ %s:%d\n",
522 			    class->lc_name, lock->lo_name, file, line);
523 			printf("while exclusively locked from %s:%d\n",
524 			    lock1->li_file, lock1->li_line);
525 			panic("share->excl");
526 		}
527 		if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
528 		    (flags & LOP_EXCLUSIVE) != 0) {
529 			printf("exclusive lock of (%s) %s @ %s:%d\n",
530 			    class->lc_name, lock->lo_name, file, line);
531 			printf("while share locked from %s:%d\n",
532 			    lock1->li_file, lock1->li_line);
533 			panic("excl->share");
534 		}
535 		lock1->li_flags++;
536 		if ((lock->lo_flags & LO_RECURSABLE) == 0) {
537 			printf(
538 			"recursed on non-recursive lock (%s) %s @ %s:%d\n",
539 			    class->lc_name, lock->lo_name, file, line);
540 			printf("first acquired @ %s:%d\n", lock1->li_file,
541 			    lock1->li_line);
542 			panic("recurse");
543 		}
544 		CTR3(KTR_WITNESS, __func__ ": pid %d recursed on %s r=%d",
545 		    curproc->p_pid, lock->lo_name,
546 		    lock1->li_flags & LI_RECURSEMASK);
547 		lock1->li_file = file;
548 		lock1->li_line = line;
549 		return;
550 	}
551 
552 	/*
553 	 * Check for duplicate locks of the same type.  Note that we only
554 	 * have to check for this on the last lock we just acquired.  Any
555 	 * other cases will be caught as lock order violations.
556 	 */
557 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
558 	w1 = lock1->li_lock->lo_witness;
559 	if (w1 == w) {
560 		if (w->w_same_squawked || dup_ok(w))
561 			goto out;
562 		w->w_same_squawked = 1;
563 		printf("acquiring duplicate lock of same type: \"%s\"\n",
564 			lock->lo_name);
565 		printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line);
566 		printf(" 2nd @ %s:%d\n", file, line);
567 #ifdef DDB
568 		go_into_ddb = 1;
569 #endif /* DDB */
570 		goto out;
571 	}
572 	MPASS(!mtx_owned(&w_mtx));
573 	mtx_lock_spin(&w_mtx);
574 	/*
575 	 * If we have a known higher number just say ok
576 	 */
577 	if (witness_watch > 1 && w->w_level > w1->w_level) {
578 		mtx_unlock_spin(&w_mtx);
579 		goto out;
580 	}
581 	if (isitmydescendant(w1, w)) {
582 		mtx_unlock_spin(&w_mtx);
583 		goto out;
584 	}
585 	for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
586 		for (i = lle->ll_count - 1; i >= 0; i--, j++) {
587 
588 			MPASS(j < WITNESS_COUNT);
589 			lock1 = &lle->ll_children[i];
590 			w1 = lock1->li_lock->lo_witness;
591 
592 			/*
593 			 * If this lock doesn't undergo witness checking,
594 			 * then skip it.
595 			 */
596 			if (w1 == NULL) {
597 				KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
598 				    ("lock missing witness structure"));
599 				continue;
600 			}
601 			/*
602 			 * If we are locking Giant and we slept with this
603 			 * lock, then skip it.
604 			 */
605 			if ((lock1->li_flags & LI_SLEPT) != 0 &&
606 			    lock == &Giant.mtx_object)
607 				continue;
608 			if (!isitmydescendant(w, w1))
609 				continue;
610 			/*
611 			 * We have a lock order violation, check to see if it
612 			 * is allowed or has already been yelled about.
613 			 */
614 			mtx_unlock_spin(&w_mtx);
615 			if (blessed(w, w1))
616 				goto out;
617 			if (lock1->li_lock == &Giant.mtx_object) {
618 				if (w1->w_Giant_squawked)
619 					goto out;
620 				else
621 					w1->w_Giant_squawked = 1;
622 			} else {
623 				if (w1->w_other_squawked)
624 					goto out;
625 				else
626 					w1->w_other_squawked = 1;
627 			}
628 			/*
629 			 * Ok, yell about it.
630 			 */
631 			printf("lock order reversal\n");
632 			/*
633 			 * Try to locate an earlier lock with
634 			 * witness w in our list.
635 			 */
636 			do {
637 				lock2 = &lle->ll_children[i];
638 				MPASS(lock2->li_lock != NULL);
639 				if (lock2->li_lock->lo_witness == w)
640 					break;
641 				i--;
642 				if (i == 0 && lle->ll_next != NULL) {
643 					lle = lle->ll_next;
644 					i = lle->ll_count - 1;
645 					MPASS(i != 0);
646 				}
647 			} while (i >= 0);
648 			if (i < 0) {
649 				printf(" 1st %p %s @ %s:%d\n", lock1->li_lock,
650 				    lock1->li_lock->lo_name, lock1->li_file,
651 				    lock1->li_line);
652 				printf(" 2nd %p %s @ %s:%d\n", lock,
653 				    lock->lo_name, file, line);
654 			} else {
655 				printf(" 1st %p %s @ %s:%d\n", lock2->li_lock,
656 				    lock2->li_lock->lo_name, lock2->li_file,
657 				    lock2->li_line);
658 				printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock,
659 				    lock1->li_lock->lo_name, lock1->li_file,
660 				    lock1->li_line);
661 				printf(" 3rd %p %s @ %s:%d\n", lock,
662 				    lock->lo_name, file, line);
663 			}
664 #ifdef DDB
665 			go_into_ddb = 1;
666 #endif /* DDB */
667 			goto out;
668 		}
669 	}
670 	lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
671 	/*
672 	 * Don't build a new relationship if we are locking Giant just
673 	 * after waking up and the previous lock in the list was acquired
674 	 * prior to blocking.
675 	 */
676 	if (lock == &Giant.mtx_object && (lock1->li_flags & LI_SLEPT) != 0)
677 		mtx_unlock_spin(&w_mtx);
678 	else {
679 		CTR2(KTR_WITNESS, __func__ ": adding %s as a child of %s",
680 		    lock->lo_name, lock1->li_lock->lo_name);
681 		if (!itismychild(lock1->li_lock->lo_witness, w))
682 			mtx_unlock_spin(&w_mtx);
683 	}
684 
685 out:
686 #ifdef DDB
687 	if (witness_ddb && go_into_ddb)
688 		Debugger(__func__);
689 #endif /* DDB */
690 	w->w_file = file;
691 	w->w_line = line;
692 
693 	lle = *lock_list;
694 	if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
695 		lle = witness_lock_list_get();
696 		if (lle == NULL)
697 			return;
698 		lle->ll_next = *lock_list;
699 		CTR2(KTR_WITNESS, __func__ ": pid %d added lle %p",
700 		    curproc->p_pid, lle);
701 		*lock_list = lle;
702 	}
703 	lock1 = &lle->ll_children[lle->ll_count++];
704 	lock1->li_lock = lock;
705 	lock1->li_line = line;
706 	lock1->li_file = file;
707 	if ((flags & LOP_EXCLUSIVE) != 0)
708 		lock1->li_flags = LI_EXCLUSIVE;
709 	else
710 		lock1->li_flags = 0;
711 	CTR3(KTR_WITNESS, __func__ ": pid %d added %s as lle[%d]",
712 	    curproc->p_pid, lock->lo_name, lle->ll_count - 1);
713 }
714 
715 void
716 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
717 {
718 	struct lock_list_entry **lock_list, *lle;
719 	struct lock_instance *instance;
720 	struct lock_class *class;
721 	struct proc *p;
722 	critical_t s;
723 	int i, j;
724 
725 	if (witness_cold || witness_dead || lock->lo_witness == NULL ||
726 	    panicstr)
727 		return;
728 	p = curproc;
729 	class = lock->lo_class;
730 	if (class->lc_flags & LC_SLEEPLOCK)
731 		lock_list = &p->p_sleeplocks;
732 	else
733 		lock_list = PCPU_PTR(spinlocks);
734 	for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
735 		for (i = 0; i < (*lock_list)->ll_count; i++) {
736 			instance = &(*lock_list)->ll_children[i];
737 			if (instance->li_lock == lock) {
738 				if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
739 				    (flags & LOP_EXCLUSIVE) == 0) {
740 					printf(
741 					"shared unlock of (%s) %s @ %s:%d\n",
742 					    class->lc_name, lock->lo_name,
743 					    file, line);
744 					printf(
745 					"while exclusively locked from %s:%d\n",
746 					    instance->li_file,
747 					    instance->li_line);
748 					panic("excl->ushare");
749 				}
750 				if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
751 				    (flags & LOP_EXCLUSIVE) != 0) {
752 					printf(
753 					"exclusive unlock of (%s) %s @ %s:%d\n",
754 					    class->lc_name, lock->lo_name,
755 					    file, line);
756 					printf(
757 					"while share locked from %s:%d\n",
758 					    instance->li_file,
759 					    instance->li_line);
760 					panic("share->uexcl");
761 				}
762 				/* If we are recursed, unrecurse. */
763 				if ((instance->li_flags & LI_RECURSEMASK) > 0) {
764 					CTR3(KTR_WITNESS,
765 				    __func__ ": pid %d unrecursed on %s r=%d",
766 					    curproc->p_pid,
767 					    instance->li_lock->lo_name,
768 					    instance->li_flags);
769 					instance->li_flags--;
770 					goto out;
771 				}
772 				s = critical_enter();
773 				CTR3(KTR_WITNESS,
774 				    __func__ ": pid %d removed %s from lle[%d]",
775 				    curproc->p_pid, instance->li_lock->lo_name,
776 				    (*lock_list)->ll_count - 1);
777 				(*lock_list)->ll_count--;
778 				for (j = i; j < (*lock_list)->ll_count; j++)
779 					(*lock_list)->ll_children[j] =
780 					    (*lock_list)->ll_children[j + 1];
781 				critical_exit(s);
782 				if ((*lock_list)->ll_count == 0) {
783 					lle = *lock_list;
784 					*lock_list = lle->ll_next;
785 					CTR2(KTR_WITNESS,
786 					    __func__ ": pid %d removed lle %p",
787 					    curproc->p_pid, lle);
788 					witness_lock_list_free(lle);
789 				}
790 				goto out;
791 			}
792 		}
793 	panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
794 	    file, line);
795 out:
796 	/*
797 	 * We don't need to protect this PCPU_GET() here against preemption
798 	 * because if we hold any spinlocks then we are already protected,
799 	 * and if we don't we will get NULL if we hold no spinlocks even if
800 	 * we switch CPU's while reading it.
801 	 */
802 	if (class->lc_flags & LC_SLEEPLOCK) {
803 		if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL)
804 			panic("switchable sleep unlock (%s) %s @ %s:%d",
805 			    class->lc_name, lock->lo_name, file, line);
806 	}
807 }
808 
809 /*
810  * Warn if any held locks are not sleepable.  Note that Giant and the lock
811  * passed in are both special cases since they are both released during the
812  * sleep process and aren't actually held while the process is asleep.
813  */
814 int
815 witness_sleep(int check_only, struct lock_object *lock, const char *file,
816 	      int line)
817 {
818 	struct lock_list_entry **lock_list, *lle;
819 	struct lock_instance *lock1;
820 	struct proc *p;
821 	critical_t savecrit;
822 	int i, n;
823 
824 	if (witness_dead || panicstr)
825 		return (0);
826 	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
827 	n = 0;
828 	/*
829 	 * Preemption bad because we need PCPU_PTR(spinlocks) to not change.
830 	 */
831 	savecrit = critical_enter();
832 	p = curproc;
833 	lock_list = &p->p_sleeplocks;
834 again:
835 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
836 		for (i = lle->ll_count - 1; i >= 0; i--) {
837 			lock1 = &lle->ll_children[i];
838 			if (lock1->li_lock == lock ||
839 			    lock1->li_lock == &Giant.mtx_object)
840 				continue;
841 			if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) {
842 				if (check_only == 0) {
843 					CTR3(KTR_WITNESS,
844 				    "pid %d: sleeping with lock (%s) %s held",
845 					    curproc->p_pid,
846 					    lock1->li_lock->lo_class->lc_name,
847 					    lock1->li_lock->lo_name);
848 					lock1->li_flags |= LI_SLEPT;
849 				}
850 				continue;
851 			}
852 			n++;
853 			printf("%s:%d: %s with \"%s\" locked from %s:%d\n",
854 			    file, line, check_only ? "could sleep" : "sleeping",
855 			    lock1->li_lock->lo_name, lock1->li_file,
856 			    lock1->li_line);
857 		}
858 	if (lock_list == &p->p_sleeplocks) {
859 		lock_list = PCPU_PTR(spinlocks);
860 		goto again;
861 	}
862 #ifdef DDB
863 	if (witness_ddb && n)
864 		Debugger(__func__);
865 #endif /* DDB */
866 	critical_exit(savecrit);
867 	return (n);
868 }
869 
870 static struct witness *
871 enroll(const char *description, struct lock_class *lock_class)
872 {
873 	struct witness *w;
874 
875 	if (!witness_watch || witness_dead)
876 		return (NULL);
877 
878 	if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
879 		return (NULL);
880 	mtx_lock_spin(&w_mtx);
881 	STAILQ_FOREACH(w, &w_all, w_list) {
882 		if (strcmp(description, w->w_name) == 0) {
883 			w->w_refcount++;
884 			mtx_unlock_spin(&w_mtx);
885 			if (lock_class != w->w_class)
886 				panic(
887 				"lock (%s) %s does not match earlier (%s) lock",
888 				    description, lock_class->lc_name,
889 				    w->w_class->lc_name);
890 			return (w);
891 		}
892 	}
893 	/*
894 	 * This isn't quite right, as witness_cold is still 0 while we
895 	 * enroll all the locks initialized before witness_initialize().
896 	 */
897 	if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
898 		mtx_unlock_spin(&w_mtx);
899 		panic("spin lock %s not in order list", description);
900 	}
901 	if ((w = witness_get()) == NULL)
902 		return (NULL);
903 	w->w_name = description;
904 	w->w_class = lock_class;
905 	w->w_refcount = 1;
906 	STAILQ_INSERT_HEAD(&w_all, w, w_list);
907 	if (lock_class->lc_flags & LC_SPINLOCK)
908 		STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
909 	else if (lock_class->lc_flags & LC_SLEEPLOCK)
910 		STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
911 	else {
912 		mtx_unlock_spin(&w_mtx);
913 		panic("lock class %s is not sleep or spin",
914 		    lock_class->lc_name);
915 	}
916 	mtx_unlock_spin(&w_mtx);
917 
918 	return (w);
919 }
920 
921 static int
922 itismychild(struct witness *parent, struct witness *child)
923 {
924 	static int recursed;
925 	struct witness_child_list_entry **wcl;
926 	struct witness_list *list;
927 
928 	MPASS(child != NULL && parent != NULL);
929 	if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
930 	    (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
931 		panic(
932 		"%s: parent (%s) and child (%s) are not the same lock type",
933 		    __func__, parent->w_class->lc_name,
934 		    child->w_class->lc_name);
935 
936 	/*
937 	 * Insert "child" after "parent"
938 	 */
939 	wcl = &parent->w_children;
940 	while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
941 		wcl = &(*wcl)->wcl_next;
942 
943 	if (*wcl == NULL) {
944 		*wcl = witness_child_get();
945 		if (*wcl == NULL)
946 			return (1);
947 	}
948 
949 	(*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
950 
951 	/*
952 	 * Now prune whole tree.  We look for cases where a lock is now
953 	 * both a descendant and a direct child of a given lock.  In that
954 	 * case, we want to remove the direct child link from the tree.
955 	 */
956 	if (recursed)
957 		return (0);
958 	recursed = 1;
959 	if (parent->w_class->lc_flags & LC_SLEEPLOCK)
960 		list = &w_sleep;
961 	else
962 		list = &w_spin;
963 	STAILQ_FOREACH(child, list, w_typelist) {
964 		STAILQ_FOREACH(parent, list, w_typelist) {
965 			if (!isitmychild(parent, child))
966 				continue;
967 			removechild(parent, child);
968 			if (isitmydescendant(parent, child))
969 				continue;
970 			itismychild(parent, child);
971 		}
972 	}
973 	recursed = 0;
974 	witness_levelall();
975 	return (0);
976 }
977 
978 static void
979 removechild(struct witness *parent, struct witness *child)
980 {
981 	struct witness_child_list_entry **wcl, *wcl1;
982 	int i;
983 
984 	for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
985 		for (i = 0; i < (*wcl)->wcl_count; i++)
986 			if ((*wcl)->wcl_children[i] == child)
987 				goto found;
988 	return;
989 found:
990 	(*wcl)->wcl_count--;
991 	if ((*wcl)->wcl_count > i)
992 		(*wcl)->wcl_children[i] =
993 		    (*wcl)->wcl_children[(*wcl)->wcl_count];
994 	MPASS((*wcl)->wcl_children[i] != NULL);
995 
996 	if ((*wcl)->wcl_count != 0)
997 		return;
998 
999 	wcl1 = *wcl;
1000 	*wcl = wcl1->wcl_next;
1001 	witness_child_free(wcl1);
1002 }
1003 
1004 static int
1005 isitmychild(struct witness *parent, struct witness *child)
1006 {
1007 	struct witness_child_list_entry *wcl;
1008 	int i;
1009 
1010 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1011 		for (i = 0; i < wcl->wcl_count; i++) {
1012 			if (wcl->wcl_children[i] == child)
1013 				return (1);
1014 		}
1015 	}
1016 	return (0);
1017 }
1018 
1019 static int
1020 isitmydescendant(struct witness *parent, struct witness *child)
1021 {
1022 	struct witness_child_list_entry *wcl;
1023 	int i, j;
1024 
1025 	if (isitmychild(parent, child))
1026 		return (1);
1027 	j = 0;
1028 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1029 		MPASS(j < 1000);
1030 		for (i = 0; i < wcl->wcl_count; i++) {
1031 			if (isitmydescendant(wcl->wcl_children[i], child))
1032 				return (1);
1033 		}
1034 		j++;
1035 	}
1036 	return (0);
1037 }
1038 
1039 void
1040 witness_levelall (void)
1041 {
1042 	struct witness_list *list;
1043 	struct witness *w, *w1;
1044 
1045 	/*
1046 	 * First clear all levels.
1047 	 */
1048 	STAILQ_FOREACH(w, &w_all, w_list) {
1049 		w->w_level = 0;
1050 	}
1051 
1052 	/*
1053 	 * Look for locks with no parent and level all their descendants.
1054 	 */
1055 	STAILQ_FOREACH(w, &w_all, w_list) {
1056 		/*
1057 		 * This is just an optimization, technically we could get
1058 		 * away just walking the all list each time.
1059 		 */
1060 		if (w->w_class->lc_flags & LC_SLEEPLOCK)
1061 			list = &w_sleep;
1062 		else
1063 			list = &w_spin;
1064 		STAILQ_FOREACH(w1, list, w_typelist) {
1065 			if (isitmychild(w1, w))
1066 				goto skip;
1067 		}
1068 		witness_leveldescendents(w, 0);
1069 	skip:
1070 	}
1071 }
1072 
1073 static void
1074 witness_leveldescendents(struct witness *parent, int level)
1075 {
1076 	struct witness_child_list_entry *wcl;
1077 	int i;
1078 
1079 	if (parent->w_level < level)
1080 		parent->w_level = level;
1081 	level++;
1082 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1083 		for (i = 0; i < wcl->wcl_count; i++)
1084 			witness_leveldescendents(wcl->wcl_children[i], level);
1085 }
1086 
1087 static void
1088 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1089 			   struct witness *parent)
1090 {
1091 	struct witness_child_list_entry *wcl;
1092 	int i, level;
1093 
1094 	level =  parent->w_level;
1095 
1096 	prnt("%-2d", level);
1097 	for (i = 0; i < level; i++)
1098 		prnt(" ");
1099 	prnt("%s", parent->w_name);
1100 	if (parent->w_file != NULL)
1101 		prnt(" -- last acquired @ %s:%d\n", parent->w_file,
1102 		    parent->w_line);
1103 
1104 	for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1105 		for (i = 0; i < wcl->wcl_count; i++)
1106 			    witness_displaydescendants(prnt,
1107 				wcl->wcl_children[i]);
1108 }
1109 
1110 static int
1111 dup_ok(struct witness *w)
1112 {
1113 	const char **dup;
1114 
1115 	for (dup = dup_list; *dup != NULL; dup++)
1116 		if (strcmp(w->w_name, *dup) == 0)
1117 			return (1);
1118 	return (0);
1119 }
1120 
1121 static int
1122 blessed(struct witness *w1, struct witness *w2)
1123 {
1124 	int i;
1125 	struct witness_blessed *b;
1126 
1127 	for (i = 0; i < blessed_count; i++) {
1128 		b = &blessed_list[i];
1129 		if (strcmp(w1->w_name, b->b_lock1) == 0) {
1130 			if (strcmp(w2->w_name, b->b_lock2) == 0)
1131 				return (1);
1132 			continue;
1133 		}
1134 		if (strcmp(w1->w_name, b->b_lock2) == 0)
1135 			if (strcmp(w2->w_name, b->b_lock1) == 0)
1136 				return (1);
1137 	}
1138 	return (0);
1139 }
1140 
1141 static struct witness *
1142 witness_get(void)
1143 {
1144 	struct witness *w;
1145 
1146 	if (witness_dead) {
1147 		mtx_unlock_spin(&w_mtx);
1148 		return (NULL);
1149 	}
1150 	if (STAILQ_EMPTY(&w_free)) {
1151 		witness_dead = 1;
1152 		mtx_unlock_spin(&w_mtx);
1153 		printf("%s: witness exhausted\n", __func__);
1154 		return (NULL);
1155 	}
1156 	w = STAILQ_FIRST(&w_free);
1157 	STAILQ_REMOVE_HEAD(&w_free, w_list);
1158 	bzero(w, sizeof(*w));
1159 	return (w);
1160 }
1161 
1162 static void
1163 witness_free(struct witness *w)
1164 {
1165 
1166 	STAILQ_INSERT_HEAD(&w_free, w, w_list);
1167 }
1168 
1169 static struct witness_child_list_entry *
1170 witness_child_get(void)
1171 {
1172 	struct witness_child_list_entry *wcl;
1173 
1174 	if (witness_dead) {
1175 		mtx_unlock_spin(&w_mtx);
1176 		return (NULL);
1177 	}
1178 	wcl = w_child_free;
1179 	if (wcl == NULL) {
1180 		witness_dead = 1;
1181 		mtx_unlock_spin(&w_mtx);
1182 		printf("%s: witness exhausted\n", __func__);
1183 		return (NULL);
1184 	}
1185 	w_child_free = wcl->wcl_next;
1186 	bzero(wcl, sizeof(*wcl));
1187 	return (wcl);
1188 }
1189 
1190 static void
1191 witness_child_free(struct witness_child_list_entry *wcl)
1192 {
1193 
1194 	wcl->wcl_next = w_child_free;
1195 	w_child_free = wcl;
1196 }
1197 
1198 static struct lock_list_entry *
1199 witness_lock_list_get(void)
1200 {
1201 	struct lock_list_entry *lle;
1202 
1203 	if (witness_dead)
1204 		return (NULL);
1205 	mtx_lock_spin(&w_mtx);
1206 	lle = w_lock_list_free;
1207 	if (lle == NULL) {
1208 		witness_dead = 1;
1209 		mtx_unlock_spin(&w_mtx);
1210 		printf("%s: witness exhausted\n", __func__);
1211 		return (NULL);
1212 	}
1213 	w_lock_list_free = lle->ll_next;
1214 	mtx_unlock_spin(&w_mtx);
1215 	bzero(lle, sizeof(*lle));
1216 	return (lle);
1217 }
1218 
1219 static void
1220 witness_lock_list_free(struct lock_list_entry *lle)
1221 {
1222 
1223 	mtx_lock_spin(&w_mtx);
1224 	lle->ll_next = w_lock_list_free;
1225 	w_lock_list_free = lle;
1226 	mtx_unlock_spin(&w_mtx);
1227 }
1228 
1229 static struct lock_instance *
1230 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1231 {
1232 	struct lock_list_entry *lle;
1233 	struct lock_instance *instance;
1234 	int i;
1235 
1236 	for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1237 		for (i = lle->ll_count - 1; i >= 0; i--) {
1238 			instance = &lle->ll_children[i];
1239 			if (instance->li_lock == lock)
1240 				return (instance);
1241 		}
1242 	return (NULL);
1243 }
1244 
1245 int
1246 witness_list_locks(struct lock_list_entry **lock_list)
1247 {
1248 	struct lock_list_entry *lle;
1249 	struct lock_instance *instance;
1250 	struct lock_object *lock;
1251 	int i, nheld;
1252 
1253 	nheld = 0;
1254 	for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1255 		for (i = lle->ll_count - 1; i >= 0; i--) {
1256 			instance = &lle->ll_children[i];
1257 			lock = instance->li_lock;
1258 			printf("%s (%s) %s (%p) locked @ %s:%d\n",
1259 			    (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1260 			    "exclusive" : "shared",
1261 			    lock->lo_class->lc_name, lock->lo_name, lock,
1262 			    instance->li_file, instance->li_line);
1263 			nheld++;
1264 		}
1265 	return (nheld);
1266 }
1267 
1268 /*
1269  * Calling this on p != curproc is bad unless we are in ddb.
1270  */
1271 int
1272 witness_list(struct proc *p)
1273 {
1274 	critical_t savecrit;
1275 	int nheld;
1276 
1277 	KASSERT(p == curproc || db_active,
1278 	    ("%s: p != curproc and we aren't in the debugger", __func__));
1279 	KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1280 
1281 	if (!db_active && witness_dead)
1282 		return (0);
1283 
1284 	nheld = witness_list_locks(&p->p_sleeplocks);
1285 
1286 	/*
1287 	 * We only handle spinlocks if p == curproc.  This is somewhat broken
1288 	 * if p is currently executing on some other CPU and holds spin locks
1289 	 * as we won't display those locks.  If we had a MI way of getting
1290 	 * the per-cpu data for a given cpu then we could use p->p_oncpu to
1291 	 * get the list of spinlocks for this process and "fix" this.
1292 	 */
1293 	if (p == curproc) {
1294 		/*
1295 		 * Preemption bad because we need PCPU_PTR(spinlocks) to not
1296 		 * change.
1297 		 */
1298 		savecrit = critical_enter();
1299 		nheld += witness_list_locks(PCPU_PTR(spinlocks));
1300 		critical_exit(savecrit);
1301 	}
1302 
1303 	return (nheld);
1304 }
1305 
1306 void
1307 witness_save(struct lock_object *lock, const char **filep, int *linep)
1308 {
1309 	struct lock_instance *instance;
1310 
1311 	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
1312 	if (lock->lo_witness == NULL || witness_dead)
1313 		return;
1314 
1315 	KASSERT(lock->lo_class->lc_flags & LC_SLEEPLOCK,
1316 	    ("%s: lock (%s) %s is not a sleep lock", __func__,
1317 	    lock->lo_class->lc_name, lock->lo_name));
1318 	instance = find_instance(curproc->p_sleeplocks, lock);
1319 	KASSERT(instance != NULL, ("%s: lock (%s) %s not locked", __func__,
1320 	    lock->lo_class->lc_name, lock->lo_name));
1321 
1322 	*filep = instance->li_file;
1323 	*linep = instance->li_line;
1324 }
1325 
1326 void
1327 witness_restore(struct lock_object *lock, const char *file, int line)
1328 {
1329 	struct lock_instance *instance;
1330 
1331 	KASSERT(!witness_cold, ("%s: witness_cold\n", __func__));
1332 	if (lock->lo_witness == NULL || witness_dead)
1333 		return;
1334 
1335 	KASSERT(lock->lo_class->lc_flags & LC_SLEEPLOCK,
1336 	    ("%s: lock (%s) %s is not a sleep lock", __func__,
1337 	    lock->lo_class->lc_name, lock->lo_name));
1338 	instance = find_instance(curproc->p_sleeplocks, lock);
1339 	KASSERT(instance != NULL, ("%s: lock (%s) %s not locked", __func__,
1340 	    lock->lo_class->lc_name, lock->lo_name));
1341 
1342 	lock->lo_witness->w_file = file;
1343 	lock->lo_witness->w_line = line;
1344 	instance->li_file = file;
1345 	instance->li_line = line;
1346 }
1347 
1348 void
1349 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1350 {
1351 #ifdef INVARIANT_SUPPORT
1352 	struct lock_instance *instance;
1353 
1354 	if (lock->lo_witness == NULL || witness_dead)
1355 		return;
1356 
1357 	if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1358 		instance = find_instance(curproc->p_sleeplocks, lock);
1359 	else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1360 		instance = find_instance(PCPU_GET(spinlocks), lock);
1361 	else
1362 		panic("Lock (%s) %s is not sleep or spin!",
1363 		    lock->lo_class->lc_name, lock->lo_name);
1364 	switch (flags) {
1365 	case LA_UNLOCKED:
1366 		if (instance != NULL)
1367 			panic("Lock (%s) %s locked @ %s:%d.",
1368 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1369 		break;
1370 	case LA_LOCKED:
1371 	case LA_LOCKED | LA_RECURSED:
1372 	case LA_LOCKED | LA_NOTRECURSED:
1373 	case LA_SLOCKED:
1374 	case LA_SLOCKED | LA_RECURSED:
1375 	case LA_SLOCKED | LA_NOTRECURSED:
1376 	case LA_XLOCKED:
1377 	case LA_XLOCKED | LA_RECURSED:
1378 	case LA_XLOCKED | LA_NOTRECURSED:
1379 		if (instance == NULL)
1380 			panic("Lock (%s) %s not locked @ %s:%d.",
1381 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1382 		if ((flags & LA_XLOCKED) != 0 &&
1383 		    (instance->li_flags & LI_EXCLUSIVE) == 0)
1384 			panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1385 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1386 		if ((flags & LA_SLOCKED) != 0 &&
1387 		    (instance->li_flags & LI_EXCLUSIVE) != 0)
1388 			panic("Lock (%s) %s exclusively locked @ %s:%d.",
1389 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1390 		if ((flags & LA_RECURSED) != 0 &&
1391 		    (instance->li_flags & LI_RECURSEMASK) == 0)
1392 			panic("Lock (%s) %s not recursed @ %s:%d.",
1393 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1394 		if ((flags & LA_NOTRECURSED) != 0 &&
1395 		    (instance->li_flags & LI_RECURSEMASK) != 0)
1396 			panic("Lock (%s) %s recursed @ %s:%d.",
1397 			    lock->lo_class->lc_name, lock->lo_name, file, line);
1398 		break;
1399 	default:
1400 		panic("Invalid lock assertion at %s:%d.", file, line);
1401 
1402 	}
1403 #endif	/* INVARIANT_SUPPORT */
1404 }
1405 
1406 #ifdef DDB
1407 
1408 DB_SHOW_COMMAND(locks, db_witness_list)
1409 {
1410 	struct proc *p;
1411 	pid_t pid;
1412 
1413 	if (have_addr) {
1414 		pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1415 		    ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1416 		    ((addr >> 16) % 16) * 10000;
1417 
1418 		/* sx_slock(&allproc_lock); */
1419 		LIST_FOREACH(p, &allproc, p_list) {
1420 			if (p->p_pid == pid)
1421 				break;
1422 		}
1423 		/* sx_sunlock(&allproc_lock); */
1424 		if (p == NULL) {
1425 			db_printf("pid %d not found\n", pid);
1426 			return;
1427 		}
1428 	} else
1429 		p = curproc;
1430 
1431 	witness_list(p);
1432 }
1433 
1434 DB_SHOW_COMMAND(witness, db_witness_display)
1435 {
1436 
1437 	witness_display(db_printf);
1438 }
1439 #endif
1440