xref: /netbsd-src/sys/kern/subr_lockdebug.c (revision b2a8932dbe9fdfd3d41d60d0a04b9a3ba294763d)
1 /*	$NetBSD: subr_lockdebug.c,v 1.64 2018/03/19 08:41:21 ozaki-r Exp $	*/
2 
3 /*-
4  * Copyright (c) 2006, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Basic lock debugging code shared among lock primitives.
34  */
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: subr_lockdebug.c,v 1.64 2018/03/19 08:41:21 ozaki-r Exp $");
38 
39 #ifdef _KERNEL_OPT
40 #include "opt_ddb.h"
41 #endif
42 
43 #include <sys/param.h>
44 #include <sys/proc.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
47 #include <sys/kmem.h>
48 #include <sys/lockdebug.h>
49 #include <sys/sleepq.h>
50 #include <sys/cpu.h>
51 #include <sys/atomic.h>
52 #include <sys/lock.h>
53 #include <sys/rbtree.h>
54 #include <sys/ksyms.h>
55 
56 #include <machine/lock.h>
57 
58 unsigned int		ld_panic;
59 
60 #ifdef LOCKDEBUG
61 
62 #define	LD_BATCH_SHIFT	9
63 #define	LD_BATCH	(1 << LD_BATCH_SHIFT)
64 #define	LD_BATCH_MASK	(LD_BATCH - 1)
65 #define	LD_MAX_LOCKS	1048576
66 #define	LD_SLOP		16
67 
68 #define	LD_LOCKED	0x01
69 #define	LD_SLEEPER	0x02
70 
71 #define	LD_WRITE_LOCK	0x80000000
72 
73 typedef struct lockdebug {
74 	struct rb_node	ld_rb_node;
75 	__cpu_simple_lock_t ld_spinlock;
76 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_chain;
77 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_achain;
78 	volatile void	*ld_lock;
79 	lockops_t	*ld_lockops;
80 	struct lwp	*ld_lwp;
81 	uintptr_t	ld_locked;
82 	uintptr_t	ld_unlocked;
83 	uintptr_t	ld_initaddr;
84 	uint16_t	ld_shares;
85 	uint16_t	ld_cpu;
86 	uint8_t		ld_flags;
87 	uint8_t		ld_shwant;	/* advisory */
88 	uint8_t		ld_exwant;	/* advisory */
89 	uint8_t		ld_unused;
90 } volatile lockdebug_t;
91 
92 typedef _TAILQ_HEAD(lockdebuglist, struct lockdebug, volatile) lockdebuglist_t;
93 
94 __cpu_simple_lock_t	ld_mod_lk;
95 lockdebuglist_t		ld_free = TAILQ_HEAD_INITIALIZER(ld_free);
96 lockdebuglist_t		ld_all = TAILQ_HEAD_INITIALIZER(ld_all);
97 int			ld_nfree;
98 int			ld_freeptr;
99 int			ld_recurse;
100 bool			ld_nomore;
101 lockdebug_t		ld_prime[LD_BATCH];
102 
103 static void	lockdebug_abort1(const char *, size_t, lockdebug_t *, int,
104     const char *, bool);
105 static int	lockdebug_more(int);
106 static void	lockdebug_init(void);
107 static void	lockdebug_dump(lockdebug_t *, void (*)(const char *, ...)
108     __printflike(1, 2));
109 
110 static signed int
111 ld_rbto_compare_nodes(void *ctx, const void *n1, const void *n2)
112 {
113 	const lockdebug_t *ld1 = n1;
114 	const lockdebug_t *ld2 = n2;
115 	const uintptr_t a = (uintptr_t)ld1->ld_lock;
116 	const uintptr_t b = (uintptr_t)ld2->ld_lock;
117 
118 	if (a < b)
119 		return -1;
120 	if (a > b)
121 		return 1;
122 	return 0;
123 }
124 
125 static signed int
126 ld_rbto_compare_key(void *ctx, const void *n, const void *key)
127 {
128 	const lockdebug_t *ld = n;
129 	const uintptr_t a = (uintptr_t)ld->ld_lock;
130 	const uintptr_t b = (uintptr_t)key;
131 
132 	if (a < b)
133 		return -1;
134 	if (a > b)
135 		return 1;
136 	return 0;
137 }
138 
139 static rb_tree_t ld_rb_tree;
140 
141 static const rb_tree_ops_t ld_rb_tree_ops = {
142 	.rbto_compare_nodes = ld_rbto_compare_nodes,
143 	.rbto_compare_key = ld_rbto_compare_key,
144 	.rbto_node_offset = offsetof(lockdebug_t, ld_rb_node),
145 	.rbto_context = NULL
146 };
147 
148 static inline lockdebug_t *
149 lockdebug_lookup1(const volatile void *lock)
150 {
151 	lockdebug_t *ld;
152 	struct cpu_info *ci;
153 
154 	ci = curcpu();
155 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
156 	ld = rb_tree_find_node(&ld_rb_tree, (void *)(intptr_t)lock);
157 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
158 	if (ld == NULL) {
159 		return NULL;
160 	}
161 	__cpu_simple_lock(&ld->ld_spinlock);
162 
163 	return ld;
164 }
165 
166 static void
167 lockdebug_lock_cpus(void)
168 {
169 	CPU_INFO_ITERATOR cii;
170 	struct cpu_info *ci;
171 
172 	for (CPU_INFO_FOREACH(cii, ci)) {
173 		__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
174 	}
175 }
176 
177 static void
178 lockdebug_unlock_cpus(void)
179 {
180 	CPU_INFO_ITERATOR cii;
181 	struct cpu_info *ci;
182 
183 	for (CPU_INFO_FOREACH(cii, ci)) {
184 		__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
185 	}
186 }
187 
188 /*
189  * lockdebug_lookup:
190  *
191  *	Find a lockdebug structure by a pointer to a lock and return it locked.
192  */
193 static inline lockdebug_t *
194 lockdebug_lookup(const char *func, size_t line, const volatile void *lock,
195     uintptr_t where)
196 {
197 	lockdebug_t *ld;
198 
199 	ld = lockdebug_lookup1(lock);
200 	if (__predict_false(ld == NULL)) {
201 		panic("%s,%zu: uninitialized lock (lock=%p, from=%08"
202 		    PRIxPTR ")", func, line, lock, where);
203 	}
204 	return ld;
205 }
206 
207 /*
208  * lockdebug_init:
209  *
210  *	Initialize the lockdebug system.  Allocate an initial pool of
211  *	lockdebug structures before the VM system is up and running.
212  */
213 static void
214 lockdebug_init(void)
215 {
216 	lockdebug_t *ld;
217 	int i;
218 
219 	TAILQ_INIT(&curcpu()->ci_data.cpu_ld_locks);
220 	TAILQ_INIT(&curlwp->l_ld_locks);
221 	__cpu_simple_lock_init(&curcpu()->ci_data.cpu_ld_lock);
222 	__cpu_simple_lock_init(&ld_mod_lk);
223 
224 	rb_tree_init(&ld_rb_tree, &ld_rb_tree_ops);
225 
226 	ld = ld_prime;
227 	for (i = 1, ld++; i < LD_BATCH; i++, ld++) {
228 		__cpu_simple_lock_init(&ld->ld_spinlock);
229 		TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
230 		TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
231 	}
232 	ld_freeptr = 1;
233 	ld_nfree = LD_BATCH - 1;
234 }
235 
236 /*
237  * lockdebug_alloc:
238  *
239  *	A lock is being initialized, so allocate an associated debug
240  *	structure.
241  */
242 bool
243 lockdebug_alloc(const char *func, size_t line, volatile void *lock,
244     lockops_t *lo, uintptr_t initaddr)
245 {
246 	struct cpu_info *ci;
247 	lockdebug_t *ld;
248 	int s;
249 
250 	if (__predict_false(lo == NULL || panicstr != NULL || ld_panic))
251 		return false;
252 	if (__predict_false(ld_freeptr == 0))
253 		lockdebug_init();
254 
255 	s = splhigh();
256 	__cpu_simple_lock(&ld_mod_lk);
257 	if (__predict_false((ld = lockdebug_lookup1(lock)) != NULL)) {
258 		__cpu_simple_unlock(&ld_mod_lk);
259 		lockdebug_abort1(func, line, ld, s, "already initialized",
260 		    true);
261 		return false;
262 	}
263 
264 	/*
265 	 * Pinch a new debug structure.  We may recurse because we call
266 	 * kmem_alloc(), which may need to initialize new locks somewhere
267 	 * down the path.  If not recursing, we try to maintain at least
268 	 * LD_SLOP structures free, which should hopefully be enough to
269 	 * satisfy kmem_alloc().  If we can't provide a structure, not to
270 	 * worry: we'll just mark the lock as not having an ID.
271 	 */
272 	ci = curcpu();
273 	ci->ci_lkdebug_recurse++;
274 	if (TAILQ_EMPTY(&ld_free)) {
275 		if (ci->ci_lkdebug_recurse > 1 || ld_nomore) {
276 			ci->ci_lkdebug_recurse--;
277 			__cpu_simple_unlock(&ld_mod_lk);
278 			splx(s);
279 			return false;
280 		}
281 		s = lockdebug_more(s);
282 	} else if (ci->ci_lkdebug_recurse == 1 && ld_nfree < LD_SLOP) {
283 		s = lockdebug_more(s);
284 	}
285 	if (__predict_false((ld = TAILQ_FIRST(&ld_free)) == NULL)) {
286 		__cpu_simple_unlock(&ld_mod_lk);
287 		splx(s);
288 		return false;
289 	}
290 	TAILQ_REMOVE(&ld_free, ld, ld_chain);
291 	ld_nfree--;
292 	ci->ci_lkdebug_recurse--;
293 
294 	if (__predict_false(ld->ld_lock != NULL)) {
295 		panic("%s,%zu: corrupt table ld %p", func, line, ld);
296 	}
297 
298 	/* Initialise the structure. */
299 	ld->ld_lock = lock;
300 	ld->ld_lockops = lo;
301 	ld->ld_locked = 0;
302 	ld->ld_unlocked = 0;
303 	ld->ld_lwp = NULL;
304 	ld->ld_initaddr = initaddr;
305 	ld->ld_flags = (lo->lo_type == LOCKOPS_SLEEP ? LD_SLEEPER : 0);
306 	lockdebug_lock_cpus();
307 	(void)rb_tree_insert_node(&ld_rb_tree, __UNVOLATILE(ld));
308 	lockdebug_unlock_cpus();
309 	__cpu_simple_unlock(&ld_mod_lk);
310 
311 	splx(s);
312 	return true;
313 }
314 
315 /*
316  * lockdebug_free:
317  *
318  *	A lock is being destroyed, so release debugging resources.
319  */
320 void
321 lockdebug_free(const char *func, size_t line, volatile void *lock)
322 {
323 	lockdebug_t *ld;
324 	int s;
325 
326 	if (__predict_false(panicstr != NULL || ld_panic))
327 		return;
328 
329 	s = splhigh();
330 	__cpu_simple_lock(&ld_mod_lk);
331 	ld = lockdebug_lookup(func, line, lock,
332 	    (uintptr_t) __builtin_return_address(0));
333 	if (__predict_false(ld == NULL)) {
334 		__cpu_simple_unlock(&ld_mod_lk);
335 		panic("%s,%zu: destroying uninitialized object %p"
336 		    "(ld_lock=%p)", func, line, lock, ld->ld_lock);
337 		return;
338 	}
339 	if (__predict_false((ld->ld_flags & LD_LOCKED) != 0 ||
340 	    ld->ld_shares != 0)) {
341 		__cpu_simple_unlock(&ld_mod_lk);
342 		lockdebug_abort1(func, line, ld, s, "is locked or in use",
343 		    true);
344 		return;
345 	}
346 	lockdebug_lock_cpus();
347 	rb_tree_remove_node(&ld_rb_tree, __UNVOLATILE(ld));
348 	lockdebug_unlock_cpus();
349 	ld->ld_lock = NULL;
350 	TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
351 	ld_nfree++;
352 	__cpu_simple_unlock(&ld->ld_spinlock);
353 	__cpu_simple_unlock(&ld_mod_lk);
354 	splx(s);
355 }
356 
357 /*
358  * lockdebug_more:
359  *
360  *	Allocate a batch of debug structures and add to the free list.
361  *	Must be called with ld_mod_lk held.
362  */
363 static int
364 lockdebug_more(int s)
365 {
366 	lockdebug_t *ld;
367 	void *block;
368 	int i, base, m;
369 
370 	/*
371 	 * Can't call kmem_alloc() if in interrupt context.  XXX We could
372 	 * deadlock, because we don't know which locks the caller holds.
373 	 */
374 	if (cpu_intr_p() || cpu_softintr_p()) {
375 		return s;
376 	}
377 
378 	while (ld_nfree < LD_SLOP) {
379 		__cpu_simple_unlock(&ld_mod_lk);
380 		splx(s);
381 		block = kmem_zalloc(LD_BATCH * sizeof(lockdebug_t), KM_SLEEP);
382 		s = splhigh();
383 		__cpu_simple_lock(&ld_mod_lk);
384 
385 		if (ld_nfree > LD_SLOP) {
386 			/* Somebody beat us to it. */
387 			__cpu_simple_unlock(&ld_mod_lk);
388 			splx(s);
389 			kmem_free(block, LD_BATCH * sizeof(lockdebug_t));
390 			s = splhigh();
391 			__cpu_simple_lock(&ld_mod_lk);
392 			continue;
393 		}
394 
395 		base = ld_freeptr;
396 		ld_nfree += LD_BATCH;
397 		ld = block;
398 		base <<= LD_BATCH_SHIFT;
399 		m = min(LD_MAX_LOCKS, base + LD_BATCH);
400 
401 		if (m == LD_MAX_LOCKS)
402 			ld_nomore = true;
403 
404 		for (i = base; i < m; i++, ld++) {
405 			__cpu_simple_lock_init(&ld->ld_spinlock);
406 			TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
407 			TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
408 		}
409 
410 		membar_producer();
411 	}
412 
413 	return s;
414 }
415 
416 /*
417  * lockdebug_wantlock:
418  *
419  *	Process the preamble to a lock acquire.  The "shared"
420  *	parameter controls which ld_{ex,sh}want counter is
421  *	updated; a negative value of shared updates neither.
422  */
423 void
424 lockdebug_wantlock(const char *func, size_t line,
425     const volatile void *lock, uintptr_t where, int shared)
426 {
427 	struct lwp *l = curlwp;
428 	lockdebug_t *ld;
429 	bool recurse;
430 	int s;
431 
432 	(void)shared;
433 	recurse = false;
434 
435 	if (__predict_false(panicstr != NULL || ld_panic))
436 		return;
437 
438 	s = splhigh();
439 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
440 		splx(s);
441 		return;
442 	}
443 	if ((ld->ld_flags & LD_LOCKED) != 0 || ld->ld_shares != 0) {
444 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
445 			if (ld->ld_lwp == l)
446 				recurse = true;
447 		} else if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
448 			recurse = true;
449 	}
450 	if (cpu_intr_p()) {
451 		if (__predict_false((ld->ld_flags & LD_SLEEPER) != 0)) {
452 			lockdebug_abort1(func, line, ld, s,
453 			    "acquiring sleep lock from interrupt context",
454 			    true);
455 			return;
456 		}
457 	}
458 	if (shared > 0)
459 		ld->ld_shwant++;
460 	else if (shared == 0)
461 		ld->ld_exwant++;
462 	if (__predict_false(recurse)) {
463 		lockdebug_abort1(func, line, ld, s, "locking against myself",
464 		    true);
465 		return;
466 	}
467 	__cpu_simple_unlock(&ld->ld_spinlock);
468 	splx(s);
469 }
470 
471 /*
472  * lockdebug_locked:
473  *
474  *	Process a lock acquire operation.
475  */
476 void
477 lockdebug_locked(const char *func, size_t line,
478     volatile void *lock, void *cvlock, uintptr_t where, int shared)
479 {
480 	struct lwp *l = curlwp;
481 	lockdebug_t *ld;
482 	int s;
483 
484 	if (__predict_false(panicstr != NULL || ld_panic))
485 		return;
486 
487 	s = splhigh();
488 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
489 		splx(s);
490 		return;
491 	}
492 	if (cvlock) {
493 		KASSERT(ld->ld_lockops->lo_type == LOCKOPS_CV);
494 		if (lock == (void *)&lbolt) {
495 			/* nothing */
496 		} else if (ld->ld_shares++ == 0) {
497 			ld->ld_locked = (uintptr_t)cvlock;
498 		} else if (__predict_false(cvlock != (void *)ld->ld_locked)) {
499 			lockdebug_abort1(func, line, ld, s,
500 			    "multiple locks used with condition variable",
501 			    true);
502 			return;
503 		}
504 	} else if (shared) {
505 		l->l_shlocks++;
506 		ld->ld_locked = where;
507 		ld->ld_shares++;
508 		ld->ld_shwant--;
509 	} else {
510 		if (__predict_false((ld->ld_flags & LD_LOCKED) != 0)) {
511 			lockdebug_abort1(func, line, ld, s, "already locked",
512 			    true);
513 			return;
514 		}
515 		ld->ld_flags |= LD_LOCKED;
516 		ld->ld_locked = where;
517 		ld->ld_exwant--;
518 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
519 			TAILQ_INSERT_TAIL(&l->l_ld_locks, ld, ld_chain);
520 		} else {
521 			TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_ld_locks,
522 			    ld, ld_chain);
523 		}
524 	}
525 	ld->ld_cpu = (uint16_t)cpu_index(curcpu());
526 	ld->ld_lwp = l;
527 	__cpu_simple_unlock(&ld->ld_spinlock);
528 	splx(s);
529 }
530 
531 /*
532  * lockdebug_unlocked:
533  *
534  *	Process a lock release operation.
535  */
536 void
537 lockdebug_unlocked(const char *func, size_t line,
538     volatile void *lock, uintptr_t where, int shared)
539 {
540 	struct lwp *l = curlwp;
541 	lockdebug_t *ld;
542 	int s;
543 
544 	if (__predict_false(panicstr != NULL || ld_panic))
545 		return;
546 
547 	s = splhigh();
548 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
549 		splx(s);
550 		return;
551 	}
552 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
553 		if (lock == (void *)&lbolt) {
554 			/* nothing */
555 		} else {
556 			ld->ld_shares--;
557 		}
558 	} else if (shared) {
559 		if (__predict_false(l->l_shlocks == 0)) {
560 			lockdebug_abort1(func, line, ld, s,
561 			    "no shared locks held by LWP", true);
562 			return;
563 		}
564 		if (__predict_false(ld->ld_shares == 0)) {
565 			lockdebug_abort1(func, line, ld, s,
566 			    "no shared holds on this lock", true);
567 			return;
568 		}
569 		l->l_shlocks--;
570 		ld->ld_shares--;
571 		if (ld->ld_lwp == l) {
572 			ld->ld_unlocked = where;
573 			ld->ld_lwp = NULL;
574 		}
575 		if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
576 			ld->ld_cpu = (uint16_t)-1;
577 	} else {
578 		if (__predict_false((ld->ld_flags & LD_LOCKED) == 0)) {
579 			lockdebug_abort1(func, line, ld, s, "not locked", true);
580 			return;
581 		}
582 
583 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
584 			if (__predict_false(ld->ld_lwp != curlwp)) {
585 				lockdebug_abort1(func, line, ld, s,
586 				    "not held by current LWP", true);
587 				return;
588 			}
589 			TAILQ_REMOVE(&l->l_ld_locks, ld, ld_chain);
590 		} else {
591 			uint16_t idx = (uint16_t)cpu_index(curcpu());
592 			if (__predict_false(ld->ld_cpu != idx)) {
593 				lockdebug_abort1(func, line, ld, s,
594 				    "not held by current CPU", true);
595 				return;
596 			}
597 			TAILQ_REMOVE(&curcpu()->ci_data.cpu_ld_locks, ld,
598 			    ld_chain);
599 		}
600 		ld->ld_flags &= ~LD_LOCKED;
601 		ld->ld_unlocked = where;
602 		ld->ld_lwp = NULL;
603 	}
604 	__cpu_simple_unlock(&ld->ld_spinlock);
605 	splx(s);
606 }
607 
608 /*
609  * lockdebug_wakeup:
610  *
611  *	Process a wakeup on a condition variable.
612  */
613 void
614 lockdebug_wakeup(const char *func, size_t line, volatile void *lock,
615     uintptr_t where)
616 {
617 	lockdebug_t *ld;
618 	int s;
619 
620 	if (__predict_false(panicstr != NULL || ld_panic || lock == (void *)&lbolt))
621 		return;
622 
623 	s = splhigh();
624 	/* Find the CV... */
625 	if ((ld = lockdebug_lookup(func, line, lock, where)) == NULL) {
626 		splx(s);
627 		return;
628 	}
629 	/*
630 	 * If it has any waiters, ensure that they are using the
631 	 * same interlock.
632 	 */
633 	if (__predict_false(ld->ld_shares != 0 &&
634 	    !mutex_owned((kmutex_t *)ld->ld_locked))) {
635 		lockdebug_abort1(func, line, ld, s, "interlocking mutex not "
636 		    "held during wakeup", true);
637 		return;
638 	}
639 	__cpu_simple_unlock(&ld->ld_spinlock);
640 	splx(s);
641 }
642 
643 /*
644  * lockdebug_barrier:
645  *
646  *	Panic if we hold more than one specified spin lock, and optionally,
647  *	if we hold sleep locks.
648  */
649 void
650 lockdebug_barrier(const char *func, size_t line, volatile void *spinlock,
651     int slplocks)
652 {
653 	struct lwp *l = curlwp;
654 	lockdebug_t *ld;
655 	int s;
656 
657 	if (__predict_false(panicstr != NULL || ld_panic))
658 		return;
659 
660 	s = splhigh();
661 	if ((l->l_pflag & LP_INTR) == 0) {
662 		TAILQ_FOREACH(ld, &curcpu()->ci_data.cpu_ld_locks, ld_chain) {
663 			if (ld->ld_lock == spinlock) {
664 				continue;
665 			}
666 			__cpu_simple_lock(&ld->ld_spinlock);
667 			lockdebug_abort1(func, line, ld, s,
668 			    "spin lock held", true);
669 			return;
670 		}
671 	}
672 	if (slplocks) {
673 		splx(s);
674 		return;
675 	}
676 	ld = TAILQ_FIRST(&l->l_ld_locks);
677 	if (__predict_false(ld != NULL)) {
678 		__cpu_simple_lock(&ld->ld_spinlock);
679 		lockdebug_abort1(func, line, ld, s, "sleep lock held", true);
680 		return;
681 	}
682 	splx(s);
683 	if (l->l_shlocks != 0) {
684 		TAILQ_FOREACH(ld, &ld_all, ld_achain) {
685 			if (ld->ld_lockops->lo_type == LOCKOPS_CV)
686 				continue;
687 			if (ld->ld_lwp == l)
688 				lockdebug_dump(ld, printf);
689 		}
690 		panic("%s,%zu: holding %d shared locks", func, line,
691 		    l->l_shlocks);
692 	}
693 }
694 
695 /*
696  * lockdebug_mem_check:
697  *
698  *	Check for in-use locks within a memory region that is
699  *	being freed.
700  */
701 void
702 lockdebug_mem_check(const char *func, size_t line, void *base, size_t sz)
703 {
704 	lockdebug_t *ld;
705 	struct cpu_info *ci;
706 	int s;
707 
708 	if (__predict_false(panicstr != NULL || ld_panic))
709 		return;
710 
711 	s = splhigh();
712 	ci = curcpu();
713 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
714 	ld = (lockdebug_t *)rb_tree_find_node_geq(&ld_rb_tree, base);
715 	if (ld != NULL) {
716 		const uintptr_t lock = (uintptr_t)ld->ld_lock;
717 
718 		if (__predict_false((uintptr_t)base > lock))
719 			panic("%s,%zu: corrupt tree ld=%p, base=%p, sz=%zu",
720 			    func, line, ld, base, sz);
721 		if (lock >= (uintptr_t)base + sz)
722 			ld = NULL;
723 	}
724 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
725 	if (__predict_false(ld != NULL)) {
726 		__cpu_simple_lock(&ld->ld_spinlock);
727 		lockdebug_abort1(func, line, ld, s,
728 		    "allocation contains active lock", !cold);
729 		return;
730 	}
731 	splx(s);
732 }
733 
734 /*
735  * lockdebug_dump:
736  *
737  *	Dump information about a lock on panic, or for DDB.
738  */
739 static void
740 lockdebug_dump(lockdebug_t *ld, void (*pr)(const char *, ...)
741     __printflike(1, 2))
742 {
743 	int sleeper = (ld->ld_flags & LD_SLEEPER);
744 
745 	(*pr)(
746 	    "lock address : %#018lx type     : %18s\n"
747 	    "initialized  : %#018lx",
748 	    (long)ld->ld_lock, (sleeper ? "sleep/adaptive" : "spin"),
749 	    (long)ld->ld_initaddr);
750 
751 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
752 		(*pr)(" interlock: %#018lx\n", (long)ld->ld_locked);
753 	} else {
754 		(*pr)("\n"
755 		    "shared holds : %18u exclusive: %18u\n"
756 		    "shares wanted: %18u exclusive: %18u\n"
757 		    "current cpu  : %18u last held: %18u\n"
758 		    "current lwp  : %#018lx last held: %#018lx\n"
759 		    "last locked%c : %#018lx unlocked%c: %#018lx\n",
760 		    (unsigned)ld->ld_shares, ((ld->ld_flags & LD_LOCKED) != 0),
761 		    (unsigned)ld->ld_shwant, (unsigned)ld->ld_exwant,
762 		    (unsigned)cpu_index(curcpu()), (unsigned)ld->ld_cpu,
763 		    (long)curlwp, (long)ld->ld_lwp,
764 		    ((ld->ld_flags & LD_LOCKED) ? '*' : ' '),
765 		    (long)ld->ld_locked,
766 		    ((ld->ld_flags & LD_LOCKED) ? ' ' : '*'),
767 		    (long)ld->ld_unlocked);
768 	}
769 
770 	if (ld->ld_lockops->lo_dump != NULL)
771 		(*ld->ld_lockops->lo_dump)(ld->ld_lock);
772 
773 	if (sleeper) {
774 		(*pr)("\n");
775 		turnstile_print(ld->ld_lock, pr);
776 	}
777 }
778 
779 /*
780  * lockdebug_abort1:
781  *
782  *	An error has been trapped - dump lock info and panic.
783  */
784 static void
785 lockdebug_abort1(const char *func, size_t line, lockdebug_t *ld, int s,
786 		 const char *msg, bool dopanic)
787 {
788 
789 	/*
790 	 * Don't make the situation worse if the system is already going
791 	 * down in flames.  Once a panic is triggered, lockdebug state
792 	 * becomes stale and cannot be trusted.
793 	 */
794 	if (atomic_inc_uint_nv(&ld_panic) != 1) {
795 		__cpu_simple_unlock(&ld->ld_spinlock);
796 		splx(s);
797 		return;
798 	}
799 
800 	printf_nolog("%s error: %s,%zu: %s\n\n", ld->ld_lockops->lo_name,
801 	    func, line, msg);
802 	lockdebug_dump(ld, printf_nolog);
803 	__cpu_simple_unlock(&ld->ld_spinlock);
804 	splx(s);
805 	printf_nolog("\n");
806 	if (dopanic)
807 		panic("LOCKDEBUG: %s error: %s,%zu: %s",
808 		    ld->ld_lockops->lo_name, func, line, msg);
809 }
810 
811 #endif	/* LOCKDEBUG */
812 
813 /*
814  * lockdebug_lock_print:
815  *
816  *	Handle the DDB 'show lock' command.
817  */
818 #ifdef DDB
819 #include <machine/db_machdep.h>
820 #include <ddb/db_interface.h>
821 
822 void
823 lockdebug_lock_print(void *addr, void (*pr)(const char *, ...))
824 {
825 #ifdef LOCKDEBUG
826 	lockdebug_t *ld;
827 
828 	TAILQ_FOREACH(ld, &ld_all, ld_achain) {
829 		if (ld->ld_lock == NULL)
830 			continue;
831 		if (addr == NULL || ld->ld_lock == addr) {
832 			lockdebug_dump(ld, pr);
833 			if (addr != NULL)
834 				return;
835 		}
836 	}
837 	if (addr != NULL) {
838 		(*pr)("Sorry, no record of a lock with address %p found.\n",
839 		    addr);
840 	}
841 #else
842 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
843 #endif	/* LOCKDEBUG */
844 }
845 
846 #ifdef LOCKDEBUG
847 static void
848 lockdebug_show_all_locks_lwp(void (*pr)(const char *, ...), bool show_trace)
849 {
850 	struct proc *p;
851 
852 	LIST_FOREACH(p, &allproc, p_list) {
853 		struct lwp *l;
854 		LIST_FOREACH(l, &p->p_lwps, l_sibling) {
855 			lockdebug_t *ld;
856 			const char *sym;
857 			int i = 0;
858 			if (TAILQ_EMPTY(&l->l_ld_locks))
859 				continue;
860 			(*pr)("Locks held by an LWP (%s):\n",
861 			    l->l_name ? l->l_name : p->p_comm);
862 			TAILQ_FOREACH(ld, &l->l_ld_locks, ld_chain) {
863 				ksyms_getname(NULL, &sym,
864 				    (vaddr_t)ld->ld_initaddr,
865 				    KSYMS_CLOSEST|KSYMS_PROC|KSYMS_ANY);
866 				(*pr)("Lock %d (initialized at %s)\n", i++, sym);
867 				lockdebug_dump(ld, pr);
868 			}
869 			if (show_trace) {
870 				db_stack_trace_print((db_expr_t)(intptr_t)l,
871 				    true,
872 				    32 /* Limit just in case */,
873 				    "a", pr);
874 			}
875 			(*pr)("\n");
876 		}
877 	}
878 }
879 
880 static void
881 lockdebug_show_all_locks_cpu(void (*pr)(const char *, ...), bool show_trace)
882 {
883 	lockdebug_t *ld;
884 	CPU_INFO_ITERATOR cii;
885 	struct cpu_info *ci;
886 	const char *sym;
887 
888 	for (CPU_INFO_FOREACH(cii, ci)) {
889 		int i = 0;
890 		if (TAILQ_EMPTY(&ci->ci_data.cpu_ld_locks))
891 			continue;
892 		(*pr)("Locks held on CPU %u:\n", ci->ci_index);
893 		TAILQ_FOREACH(ld, &ci->ci_data.cpu_ld_locks, ld_chain) {
894 			ksyms_getname(NULL, &sym,
895 			    (vaddr_t)ld->ld_initaddr,
896 			    KSYMS_CLOSEST|KSYMS_PROC|KSYMS_ANY);
897 			(*pr)("Lock %d (initialized at %s)\n", i++, sym);
898 			lockdebug_dump(ld, pr);
899 			if (show_trace) {
900 				db_stack_trace_print(
901 				    (db_expr_t)(intptr_t)ci->ci_curlwp,
902 				    true,
903 				    32 /* Limit just in case */,
904 				    "a", pr);
905 			}
906 			(*pr)("\n");
907 		}
908 	}
909 }
910 #endif	/* LOCKDEBUG */
911 
912 void
913 lockdebug_show_all_locks(void (*pr)(const char *, ...), const char *modif)
914 {
915 #ifdef LOCKDEBUG
916 	bool show_trace = false;
917 	if (modif[0] == 't')
918 		show_trace = true;
919 
920 	(*pr)("[Locks tracked through LWPs]\n");
921 	lockdebug_show_all_locks_lwp(pr, show_trace);
922 	(*pr)("\n");
923 
924 	(*pr)("[Locks tracked through CPUs]\n");
925 	lockdebug_show_all_locks_cpu(pr, show_trace);
926 	(*pr)("\n");
927 #else
928 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
929 #endif	/* LOCKDEBUG */
930 }
931 
932 void
933 lockdebug_show_lockstats(void (*pr)(const char *, ...))
934 {
935 #ifdef LOCKDEBUG
936 	lockdebug_t *ld;
937 	void *_ld;
938 	uint32_t n_null = 0;
939 	uint32_t n_spin_mutex = 0;
940 	uint32_t n_adaptive_mutex = 0;
941 	uint32_t n_rwlock = 0;
942 	uint32_t n_cv = 0;
943 	uint32_t n_others = 0;
944 
945 	RB_TREE_FOREACH(_ld, &ld_rb_tree) {
946 		ld = _ld;
947 		if (ld->ld_lock == NULL) {
948 			n_null++;
949 			continue;
950 		}
951 		if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
952 			n_cv++;
953 			continue;
954 		}
955 		if (ld->ld_lockops->lo_name[0] == 'M') {
956 			if (ld->ld_lockops->lo_type == LOCKOPS_SLEEP)
957 				n_adaptive_mutex++;
958 			else
959 				n_spin_mutex++;
960 			continue;
961 		}
962 		if (ld->ld_lockops->lo_name[0] == 'R') {
963 			n_rwlock++;
964 			continue;
965 		}
966 		n_others++;
967 	}
968 	(*pr)(
969 	    "condvar: %u\n"
970 	    "spin mutex: %u\n"
971 	    "adaptive mutex: %u\n"
972 	    "rwlock: %u\n"
973 	    "null locks: %u\n"
974 	    "others: %u\n",
975 	    n_cv,  n_spin_mutex, n_adaptive_mutex, n_rwlock,
976 	    n_null, n_others);
977 #else
978 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
979 #endif	/* LOCKDEBUG */
980 }
981 #endif	/* DDB */
982 
983 /*
984  * lockdebug_abort:
985  *
986  *	An error has been trapped - dump lock info and call panic().
987  */
988 void
989 lockdebug_abort(const char *func, size_t line, const volatile void *lock,
990     lockops_t *ops, const char *msg)
991 {
992 #ifdef LOCKDEBUG
993 	lockdebug_t *ld;
994 	int s;
995 
996 	s = splhigh();
997 	if ((ld = lockdebug_lookup(func, line, lock,
998 			(uintptr_t) __builtin_return_address(0))) != NULL) {
999 		lockdebug_abort1(func, line, ld, s, msg, true);
1000 		return;
1001 	}
1002 	splx(s);
1003 #endif	/* LOCKDEBUG */
1004 
1005 	/*
1006 	 * Complain first on the occurrance only.  Otherwise proceeed to
1007 	 * panic where we will `rendezvous' with other CPUs if the machine
1008 	 * is going down in flames.
1009 	 */
1010 	if (atomic_inc_uint_nv(&ld_panic) == 1) {
1011 		printf_nolog("%s error: %s,%zu: %s\n\n"
1012 		    "lock address : %#018lx\n"
1013 		    "current cpu  : %18d\n"
1014 		    "current lwp  : %#018lx\n",
1015 		    ops->lo_name, func, line, msg, (long)lock,
1016 		    (int)cpu_index(curcpu()), (long)curlwp);
1017 		(*ops->lo_dump)(lock);
1018 		printf_nolog("\n");
1019 	}
1020 
1021 	panic("lock error: %s: %s,%zu: %s: lock %p cpu %d lwp %p",
1022 	    ops->lo_name, func, line, msg, lock, cpu_index(curcpu()), curlwp);
1023 }
1024