xref: /openbmc/linux/kernel/audit_tree.c (revision e8e0929d)
1 #include "audit.h"
2 #include <linux/inotify.h>
3 #include <linux/namei.h>
4 #include <linux/mount.h>
5 #include <linux/kthread.h>
6 
7 struct audit_tree;
8 struct audit_chunk;
9 
10 struct audit_tree {
11 	atomic_t count;
12 	int goner;
13 	struct audit_chunk *root;
14 	struct list_head chunks;
15 	struct list_head rules;
16 	struct list_head list;
17 	struct list_head same_root;
18 	struct rcu_head head;
19 	char pathname[];
20 };
21 
22 struct audit_chunk {
23 	struct list_head hash;
24 	struct inotify_watch watch;
25 	struct list_head trees;		/* with root here */
26 	int dead;
27 	int count;
28 	atomic_long_t refs;
29 	struct rcu_head head;
30 	struct node {
31 		struct list_head list;
32 		struct audit_tree *owner;
33 		unsigned index;		/* index; upper bit indicates 'will prune' */
34 	} owners[];
35 };
36 
37 static LIST_HEAD(tree_list);
38 static LIST_HEAD(prune_list);
39 
40 /*
41  * One struct chunk is attached to each inode of interest.
42  * We replace struct chunk on tagging/untagging.
43  * Rules have pointer to struct audit_tree.
44  * Rules have struct list_head rlist forming a list of rules over
45  * the same tree.
46  * References to struct chunk are collected at audit_inode{,_child}()
47  * time and used in AUDIT_TREE rule matching.
48  * These references are dropped at the same time we are calling
49  * audit_free_names(), etc.
50  *
51  * Cyclic lists galore:
52  * tree.chunks anchors chunk.owners[].list			hash_lock
53  * tree.rules anchors rule.rlist				audit_filter_mutex
54  * chunk.trees anchors tree.same_root				hash_lock
55  * chunk.hash is a hash with middle bits of watch.inode as
56  * a hash function.						RCU, hash_lock
57  *
58  * tree is refcounted; one reference for "some rules on rules_list refer to
59  * it", one for each chunk with pointer to it.
60  *
61  * chunk is refcounted by embedded inotify_watch + .refs (non-zero refcount
62  * of watch contributes 1 to .refs).
63  *
64  * node.index allows to get from node.list to containing chunk.
65  * MSB of that sucker is stolen to mark taggings that we might have to
66  * revert - several operations have very unpleasant cleanup logics and
67  * that makes a difference.  Some.
68  */
69 
70 static struct inotify_handle *rtree_ih;
71 
72 static struct audit_tree *alloc_tree(const char *s)
73 {
74 	struct audit_tree *tree;
75 
76 	tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
77 	if (tree) {
78 		atomic_set(&tree->count, 1);
79 		tree->goner = 0;
80 		INIT_LIST_HEAD(&tree->chunks);
81 		INIT_LIST_HEAD(&tree->rules);
82 		INIT_LIST_HEAD(&tree->list);
83 		INIT_LIST_HEAD(&tree->same_root);
84 		tree->root = NULL;
85 		strcpy(tree->pathname, s);
86 	}
87 	return tree;
88 }
89 
90 static inline void get_tree(struct audit_tree *tree)
91 {
92 	atomic_inc(&tree->count);
93 }
94 
95 static void __put_tree(struct rcu_head *rcu)
96 {
97 	struct audit_tree *tree = container_of(rcu, struct audit_tree, head);
98 	kfree(tree);
99 }
100 
101 static inline void put_tree(struct audit_tree *tree)
102 {
103 	if (atomic_dec_and_test(&tree->count))
104 		call_rcu(&tree->head, __put_tree);
105 }
106 
107 /* to avoid bringing the entire thing in audit.h */
108 const char *audit_tree_path(struct audit_tree *tree)
109 {
110 	return tree->pathname;
111 }
112 
113 static struct audit_chunk *alloc_chunk(int count)
114 {
115 	struct audit_chunk *chunk;
116 	size_t size;
117 	int i;
118 
119 	size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
120 	chunk = kzalloc(size, GFP_KERNEL);
121 	if (!chunk)
122 		return NULL;
123 
124 	INIT_LIST_HEAD(&chunk->hash);
125 	INIT_LIST_HEAD(&chunk->trees);
126 	chunk->count = count;
127 	atomic_long_set(&chunk->refs, 1);
128 	for (i = 0; i < count; i++) {
129 		INIT_LIST_HEAD(&chunk->owners[i].list);
130 		chunk->owners[i].index = i;
131 	}
132 	inotify_init_watch(&chunk->watch);
133 	return chunk;
134 }
135 
136 static void free_chunk(struct audit_chunk *chunk)
137 {
138 	int i;
139 
140 	for (i = 0; i < chunk->count; i++) {
141 		if (chunk->owners[i].owner)
142 			put_tree(chunk->owners[i].owner);
143 	}
144 	kfree(chunk);
145 }
146 
147 void audit_put_chunk(struct audit_chunk *chunk)
148 {
149 	if (atomic_long_dec_and_test(&chunk->refs))
150 		free_chunk(chunk);
151 }
152 
153 static void __put_chunk(struct rcu_head *rcu)
154 {
155 	struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
156 	audit_put_chunk(chunk);
157 }
158 
159 enum {HASH_SIZE = 128};
160 static struct list_head chunk_hash_heads[HASH_SIZE];
161 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
162 
163 static inline struct list_head *chunk_hash(const struct inode *inode)
164 {
165 	unsigned long n = (unsigned long)inode / L1_CACHE_BYTES;
166 	return chunk_hash_heads + n % HASH_SIZE;
167 }
168 
169 /* hash_lock is held by caller */
170 static void insert_hash(struct audit_chunk *chunk)
171 {
172 	struct list_head *list = chunk_hash(chunk->watch.inode);
173 	list_add_rcu(&chunk->hash, list);
174 }
175 
176 /* called under rcu_read_lock */
177 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
178 {
179 	struct list_head *list = chunk_hash(inode);
180 	struct audit_chunk *p;
181 
182 	list_for_each_entry_rcu(p, list, hash) {
183 		if (p->watch.inode == inode) {
184 			atomic_long_inc(&p->refs);
185 			return p;
186 		}
187 	}
188 	return NULL;
189 }
190 
191 int audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
192 {
193 	int n;
194 	for (n = 0; n < chunk->count; n++)
195 		if (chunk->owners[n].owner == tree)
196 			return 1;
197 	return 0;
198 }
199 
200 /* tagging and untagging inodes with trees */
201 
202 static struct audit_chunk *find_chunk(struct node *p)
203 {
204 	int index = p->index & ~(1U<<31);
205 	p -= index;
206 	return container_of(p, struct audit_chunk, owners[0]);
207 }
208 
209 static void untag_chunk(struct node *p)
210 {
211 	struct audit_chunk *chunk = find_chunk(p);
212 	struct audit_chunk *new;
213 	struct audit_tree *owner;
214 	int size = chunk->count - 1;
215 	int i, j;
216 
217 	if (!pin_inotify_watch(&chunk->watch)) {
218 		/*
219 		 * Filesystem is shutting down; all watches are getting
220 		 * evicted, just take it off the node list for this
221 		 * tree and let the eviction logics take care of the
222 		 * rest.
223 		 */
224 		owner = p->owner;
225 		if (owner->root == chunk) {
226 			list_del_init(&owner->same_root);
227 			owner->root = NULL;
228 		}
229 		list_del_init(&p->list);
230 		p->owner = NULL;
231 		put_tree(owner);
232 		return;
233 	}
234 
235 	spin_unlock(&hash_lock);
236 
237 	/*
238 	 * pin_inotify_watch() succeeded, so the watch won't go away
239 	 * from under us.
240 	 */
241 	mutex_lock(&chunk->watch.inode->inotify_mutex);
242 	if (chunk->dead) {
243 		mutex_unlock(&chunk->watch.inode->inotify_mutex);
244 		goto out;
245 	}
246 
247 	owner = p->owner;
248 
249 	if (!size) {
250 		chunk->dead = 1;
251 		spin_lock(&hash_lock);
252 		list_del_init(&chunk->trees);
253 		if (owner->root == chunk)
254 			owner->root = NULL;
255 		list_del_init(&p->list);
256 		list_del_rcu(&chunk->hash);
257 		spin_unlock(&hash_lock);
258 		inotify_evict_watch(&chunk->watch);
259 		mutex_unlock(&chunk->watch.inode->inotify_mutex);
260 		put_inotify_watch(&chunk->watch);
261 		goto out;
262 	}
263 
264 	new = alloc_chunk(size);
265 	if (!new)
266 		goto Fallback;
267 	if (inotify_clone_watch(&chunk->watch, &new->watch) < 0) {
268 		free_chunk(new);
269 		goto Fallback;
270 	}
271 
272 	chunk->dead = 1;
273 	spin_lock(&hash_lock);
274 	list_replace_init(&chunk->trees, &new->trees);
275 	if (owner->root == chunk) {
276 		list_del_init(&owner->same_root);
277 		owner->root = NULL;
278 	}
279 
280 	for (i = j = 0; i < size; i++, j++) {
281 		struct audit_tree *s;
282 		if (&chunk->owners[j] == p) {
283 			list_del_init(&p->list);
284 			i--;
285 			continue;
286 		}
287 		s = chunk->owners[j].owner;
288 		new->owners[i].owner = s;
289 		new->owners[i].index = chunk->owners[j].index - j + i;
290 		if (!s) /* result of earlier fallback */
291 			continue;
292 		get_tree(s);
293 		list_replace_init(&chunk->owners[i].list, &new->owners[j].list);
294 	}
295 
296 	list_replace_rcu(&chunk->hash, &new->hash);
297 	list_for_each_entry(owner, &new->trees, same_root)
298 		owner->root = new;
299 	spin_unlock(&hash_lock);
300 	inotify_evict_watch(&chunk->watch);
301 	mutex_unlock(&chunk->watch.inode->inotify_mutex);
302 	put_inotify_watch(&chunk->watch);
303 	goto out;
304 
305 Fallback:
306 	// do the best we can
307 	spin_lock(&hash_lock);
308 	if (owner->root == chunk) {
309 		list_del_init(&owner->same_root);
310 		owner->root = NULL;
311 	}
312 	list_del_init(&p->list);
313 	p->owner = NULL;
314 	put_tree(owner);
315 	spin_unlock(&hash_lock);
316 	mutex_unlock(&chunk->watch.inode->inotify_mutex);
317 out:
318 	unpin_inotify_watch(&chunk->watch);
319 	spin_lock(&hash_lock);
320 }
321 
322 static int create_chunk(struct inode *inode, struct audit_tree *tree)
323 {
324 	struct audit_chunk *chunk = alloc_chunk(1);
325 	if (!chunk)
326 		return -ENOMEM;
327 
328 	if (inotify_add_watch(rtree_ih, &chunk->watch, inode, IN_IGNORED | IN_DELETE_SELF) < 0) {
329 		free_chunk(chunk);
330 		return -ENOSPC;
331 	}
332 
333 	mutex_lock(&inode->inotify_mutex);
334 	spin_lock(&hash_lock);
335 	if (tree->goner) {
336 		spin_unlock(&hash_lock);
337 		chunk->dead = 1;
338 		inotify_evict_watch(&chunk->watch);
339 		mutex_unlock(&inode->inotify_mutex);
340 		put_inotify_watch(&chunk->watch);
341 		return 0;
342 	}
343 	chunk->owners[0].index = (1U << 31);
344 	chunk->owners[0].owner = tree;
345 	get_tree(tree);
346 	list_add(&chunk->owners[0].list, &tree->chunks);
347 	if (!tree->root) {
348 		tree->root = chunk;
349 		list_add(&tree->same_root, &chunk->trees);
350 	}
351 	insert_hash(chunk);
352 	spin_unlock(&hash_lock);
353 	mutex_unlock(&inode->inotify_mutex);
354 	return 0;
355 }
356 
357 /* the first tagged inode becomes root of tree */
358 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
359 {
360 	struct inotify_watch *watch;
361 	struct audit_tree *owner;
362 	struct audit_chunk *chunk, *old;
363 	struct node *p;
364 	int n;
365 
366 	if (inotify_find_watch(rtree_ih, inode, &watch) < 0)
367 		return create_chunk(inode, tree);
368 
369 	old = container_of(watch, struct audit_chunk, watch);
370 
371 	/* are we already there? */
372 	spin_lock(&hash_lock);
373 	for (n = 0; n < old->count; n++) {
374 		if (old->owners[n].owner == tree) {
375 			spin_unlock(&hash_lock);
376 			put_inotify_watch(watch);
377 			return 0;
378 		}
379 	}
380 	spin_unlock(&hash_lock);
381 
382 	chunk = alloc_chunk(old->count + 1);
383 	if (!chunk)
384 		return -ENOMEM;
385 
386 	mutex_lock(&inode->inotify_mutex);
387 	if (inotify_clone_watch(&old->watch, &chunk->watch) < 0) {
388 		mutex_unlock(&inode->inotify_mutex);
389 		put_inotify_watch(&old->watch);
390 		free_chunk(chunk);
391 		return -ENOSPC;
392 	}
393 	spin_lock(&hash_lock);
394 	if (tree->goner) {
395 		spin_unlock(&hash_lock);
396 		chunk->dead = 1;
397 		inotify_evict_watch(&chunk->watch);
398 		mutex_unlock(&inode->inotify_mutex);
399 		put_inotify_watch(&old->watch);
400 		put_inotify_watch(&chunk->watch);
401 		return 0;
402 	}
403 	list_replace_init(&old->trees, &chunk->trees);
404 	for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
405 		struct audit_tree *s = old->owners[n].owner;
406 		p->owner = s;
407 		p->index = old->owners[n].index;
408 		if (!s) /* result of fallback in untag */
409 			continue;
410 		get_tree(s);
411 		list_replace_init(&old->owners[n].list, &p->list);
412 	}
413 	p->index = (chunk->count - 1) | (1U<<31);
414 	p->owner = tree;
415 	get_tree(tree);
416 	list_add(&p->list, &tree->chunks);
417 	list_replace_rcu(&old->hash, &chunk->hash);
418 	list_for_each_entry(owner, &chunk->trees, same_root)
419 		owner->root = chunk;
420 	old->dead = 1;
421 	if (!tree->root) {
422 		tree->root = chunk;
423 		list_add(&tree->same_root, &chunk->trees);
424 	}
425 	spin_unlock(&hash_lock);
426 	inotify_evict_watch(&old->watch);
427 	mutex_unlock(&inode->inotify_mutex);
428 	put_inotify_watch(&old->watch);
429 	return 0;
430 }
431 
432 static void kill_rules(struct audit_tree *tree)
433 {
434 	struct audit_krule *rule, *next;
435 	struct audit_entry *entry;
436 	struct audit_buffer *ab;
437 
438 	list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
439 		entry = container_of(rule, struct audit_entry, rule);
440 
441 		list_del_init(&rule->rlist);
442 		if (rule->tree) {
443 			/* not a half-baked one */
444 			ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
445 			audit_log_format(ab, "op=");
446 			audit_log_string(ab, "remove rule");
447 			audit_log_format(ab, " dir=");
448 			audit_log_untrustedstring(ab, rule->tree->pathname);
449 			audit_log_key(ab, rule->filterkey);
450 			audit_log_format(ab, " list=%d res=1", rule->listnr);
451 			audit_log_end(ab);
452 			rule->tree = NULL;
453 			list_del_rcu(&entry->list);
454 			list_del(&entry->rule.list);
455 			call_rcu(&entry->rcu, audit_free_rule_rcu);
456 		}
457 	}
458 }
459 
460 /*
461  * finish killing struct audit_tree
462  */
463 static void prune_one(struct audit_tree *victim)
464 {
465 	spin_lock(&hash_lock);
466 	while (!list_empty(&victim->chunks)) {
467 		struct node *p;
468 
469 		p = list_entry(victim->chunks.next, struct node, list);
470 
471 		untag_chunk(p);
472 	}
473 	spin_unlock(&hash_lock);
474 	put_tree(victim);
475 }
476 
477 /* trim the uncommitted chunks from tree */
478 
479 static void trim_marked(struct audit_tree *tree)
480 {
481 	struct list_head *p, *q;
482 	spin_lock(&hash_lock);
483 	if (tree->goner) {
484 		spin_unlock(&hash_lock);
485 		return;
486 	}
487 	/* reorder */
488 	for (p = tree->chunks.next; p != &tree->chunks; p = q) {
489 		struct node *node = list_entry(p, struct node, list);
490 		q = p->next;
491 		if (node->index & (1U<<31)) {
492 			list_del_init(p);
493 			list_add(p, &tree->chunks);
494 		}
495 	}
496 
497 	while (!list_empty(&tree->chunks)) {
498 		struct node *node;
499 
500 		node = list_entry(tree->chunks.next, struct node, list);
501 
502 		/* have we run out of marked? */
503 		if (!(node->index & (1U<<31)))
504 			break;
505 
506 		untag_chunk(node);
507 	}
508 	if (!tree->root && !tree->goner) {
509 		tree->goner = 1;
510 		spin_unlock(&hash_lock);
511 		mutex_lock(&audit_filter_mutex);
512 		kill_rules(tree);
513 		list_del_init(&tree->list);
514 		mutex_unlock(&audit_filter_mutex);
515 		prune_one(tree);
516 	} else {
517 		spin_unlock(&hash_lock);
518 	}
519 }
520 
521 static void audit_schedule_prune(void);
522 
523 /* called with audit_filter_mutex */
524 int audit_remove_tree_rule(struct audit_krule *rule)
525 {
526 	struct audit_tree *tree;
527 	tree = rule->tree;
528 	if (tree) {
529 		spin_lock(&hash_lock);
530 		list_del_init(&rule->rlist);
531 		if (list_empty(&tree->rules) && !tree->goner) {
532 			tree->root = NULL;
533 			list_del_init(&tree->same_root);
534 			tree->goner = 1;
535 			list_move(&tree->list, &prune_list);
536 			rule->tree = NULL;
537 			spin_unlock(&hash_lock);
538 			audit_schedule_prune();
539 			return 1;
540 		}
541 		rule->tree = NULL;
542 		spin_unlock(&hash_lock);
543 		return 1;
544 	}
545 	return 0;
546 }
547 
548 void audit_trim_trees(void)
549 {
550 	struct list_head cursor;
551 
552 	mutex_lock(&audit_filter_mutex);
553 	list_add(&cursor, &tree_list);
554 	while (cursor.next != &tree_list) {
555 		struct audit_tree *tree;
556 		struct path path;
557 		struct vfsmount *root_mnt;
558 		struct node *node;
559 		struct list_head list;
560 		int err;
561 
562 		tree = container_of(cursor.next, struct audit_tree, list);
563 		get_tree(tree);
564 		list_del(&cursor);
565 		list_add(&cursor, &tree->list);
566 		mutex_unlock(&audit_filter_mutex);
567 
568 		err = kern_path(tree->pathname, 0, &path);
569 		if (err)
570 			goto skip_it;
571 
572 		root_mnt = collect_mounts(&path);
573 		path_put(&path);
574 		if (!root_mnt)
575 			goto skip_it;
576 
577 		list_add_tail(&list, &root_mnt->mnt_list);
578 		spin_lock(&hash_lock);
579 		list_for_each_entry(node, &tree->chunks, list) {
580 			struct audit_chunk *chunk = find_chunk(node);
581 			struct inode *inode = chunk->watch.inode;
582 			struct vfsmount *mnt;
583 			node->index |= 1U<<31;
584 			list_for_each_entry(mnt, &list, mnt_list) {
585 				if (mnt->mnt_root->d_inode == inode) {
586 					node->index &= ~(1U<<31);
587 					break;
588 				}
589 			}
590 		}
591 		spin_unlock(&hash_lock);
592 		trim_marked(tree);
593 		put_tree(tree);
594 		list_del_init(&list);
595 		drop_collected_mounts(root_mnt);
596 skip_it:
597 		mutex_lock(&audit_filter_mutex);
598 	}
599 	list_del(&cursor);
600 	mutex_unlock(&audit_filter_mutex);
601 }
602 
603 static int is_under(struct vfsmount *mnt, struct dentry *dentry,
604 		    struct path *path)
605 {
606 	if (mnt != path->mnt) {
607 		for (;;) {
608 			if (mnt->mnt_parent == mnt)
609 				return 0;
610 			if (mnt->mnt_parent == path->mnt)
611 					break;
612 			mnt = mnt->mnt_parent;
613 		}
614 		dentry = mnt->mnt_mountpoint;
615 	}
616 	return is_subdir(dentry, path->dentry);
617 }
618 
619 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
620 {
621 
622 	if (pathname[0] != '/' ||
623 	    rule->listnr != AUDIT_FILTER_EXIT ||
624 	    op != Audit_equal ||
625 	    rule->inode_f || rule->watch || rule->tree)
626 		return -EINVAL;
627 	rule->tree = alloc_tree(pathname);
628 	if (!rule->tree)
629 		return -ENOMEM;
630 	return 0;
631 }
632 
633 void audit_put_tree(struct audit_tree *tree)
634 {
635 	put_tree(tree);
636 }
637 
638 /* called with audit_filter_mutex */
639 int audit_add_tree_rule(struct audit_krule *rule)
640 {
641 	struct audit_tree *seed = rule->tree, *tree;
642 	struct path path;
643 	struct vfsmount *mnt, *p;
644 	struct list_head list;
645 	int err;
646 
647 	list_for_each_entry(tree, &tree_list, list) {
648 		if (!strcmp(seed->pathname, tree->pathname)) {
649 			put_tree(seed);
650 			rule->tree = tree;
651 			list_add(&rule->rlist, &tree->rules);
652 			return 0;
653 		}
654 	}
655 	tree = seed;
656 	list_add(&tree->list, &tree_list);
657 	list_add(&rule->rlist, &tree->rules);
658 	/* do not set rule->tree yet */
659 	mutex_unlock(&audit_filter_mutex);
660 
661 	err = kern_path(tree->pathname, 0, &path);
662 	if (err)
663 		goto Err;
664 	mnt = collect_mounts(&path);
665 	path_put(&path);
666 	if (!mnt) {
667 		err = -ENOMEM;
668 		goto Err;
669 	}
670 	list_add_tail(&list, &mnt->mnt_list);
671 
672 	get_tree(tree);
673 	list_for_each_entry(p, &list, mnt_list) {
674 		err = tag_chunk(p->mnt_root->d_inode, tree);
675 		if (err)
676 			break;
677 	}
678 
679 	list_del(&list);
680 	drop_collected_mounts(mnt);
681 
682 	if (!err) {
683 		struct node *node;
684 		spin_lock(&hash_lock);
685 		list_for_each_entry(node, &tree->chunks, list)
686 			node->index &= ~(1U<<31);
687 		spin_unlock(&hash_lock);
688 	} else {
689 		trim_marked(tree);
690 		goto Err;
691 	}
692 
693 	mutex_lock(&audit_filter_mutex);
694 	if (list_empty(&rule->rlist)) {
695 		put_tree(tree);
696 		return -ENOENT;
697 	}
698 	rule->tree = tree;
699 	put_tree(tree);
700 
701 	return 0;
702 Err:
703 	mutex_lock(&audit_filter_mutex);
704 	list_del_init(&tree->list);
705 	list_del_init(&tree->rules);
706 	put_tree(tree);
707 	return err;
708 }
709 
710 int audit_tag_tree(char *old, char *new)
711 {
712 	struct list_head cursor, barrier;
713 	int failed = 0;
714 	struct path path;
715 	struct vfsmount *tagged;
716 	struct list_head list;
717 	struct vfsmount *mnt;
718 	struct dentry *dentry;
719 	int err;
720 
721 	err = kern_path(new, 0, &path);
722 	if (err)
723 		return err;
724 	tagged = collect_mounts(&path);
725 	path_put(&path);
726 	if (!tagged)
727 		return -ENOMEM;
728 
729 	err = kern_path(old, 0, &path);
730 	if (err) {
731 		drop_collected_mounts(tagged);
732 		return err;
733 	}
734 	mnt = mntget(path.mnt);
735 	dentry = dget(path.dentry);
736 	path_put(&path);
737 
738 	list_add_tail(&list, &tagged->mnt_list);
739 
740 	mutex_lock(&audit_filter_mutex);
741 	list_add(&barrier, &tree_list);
742 	list_add(&cursor, &barrier);
743 
744 	while (cursor.next != &tree_list) {
745 		struct audit_tree *tree;
746 		struct vfsmount *p;
747 
748 		tree = container_of(cursor.next, struct audit_tree, list);
749 		get_tree(tree);
750 		list_del(&cursor);
751 		list_add(&cursor, &tree->list);
752 		mutex_unlock(&audit_filter_mutex);
753 
754 		err = kern_path(tree->pathname, 0, &path);
755 		if (err) {
756 			put_tree(tree);
757 			mutex_lock(&audit_filter_mutex);
758 			continue;
759 		}
760 
761 		spin_lock(&vfsmount_lock);
762 		if (!is_under(mnt, dentry, &path)) {
763 			spin_unlock(&vfsmount_lock);
764 			path_put(&path);
765 			put_tree(tree);
766 			mutex_lock(&audit_filter_mutex);
767 			continue;
768 		}
769 		spin_unlock(&vfsmount_lock);
770 		path_put(&path);
771 
772 		list_for_each_entry(p, &list, mnt_list) {
773 			failed = tag_chunk(p->mnt_root->d_inode, tree);
774 			if (failed)
775 				break;
776 		}
777 
778 		if (failed) {
779 			put_tree(tree);
780 			mutex_lock(&audit_filter_mutex);
781 			break;
782 		}
783 
784 		mutex_lock(&audit_filter_mutex);
785 		spin_lock(&hash_lock);
786 		if (!tree->goner) {
787 			list_del(&tree->list);
788 			list_add(&tree->list, &tree_list);
789 		}
790 		spin_unlock(&hash_lock);
791 		put_tree(tree);
792 	}
793 
794 	while (barrier.prev != &tree_list) {
795 		struct audit_tree *tree;
796 
797 		tree = container_of(barrier.prev, struct audit_tree, list);
798 		get_tree(tree);
799 		list_del(&tree->list);
800 		list_add(&tree->list, &barrier);
801 		mutex_unlock(&audit_filter_mutex);
802 
803 		if (!failed) {
804 			struct node *node;
805 			spin_lock(&hash_lock);
806 			list_for_each_entry(node, &tree->chunks, list)
807 				node->index &= ~(1U<<31);
808 			spin_unlock(&hash_lock);
809 		} else {
810 			trim_marked(tree);
811 		}
812 
813 		put_tree(tree);
814 		mutex_lock(&audit_filter_mutex);
815 	}
816 	list_del(&barrier);
817 	list_del(&cursor);
818 	list_del(&list);
819 	mutex_unlock(&audit_filter_mutex);
820 	dput(dentry);
821 	mntput(mnt);
822 	drop_collected_mounts(tagged);
823 	return failed;
824 }
825 
826 /*
827  * That gets run when evict_chunk() ends up needing to kill audit_tree.
828  * Runs from a separate thread.
829  */
830 static int prune_tree_thread(void *unused)
831 {
832 	mutex_lock(&audit_cmd_mutex);
833 	mutex_lock(&audit_filter_mutex);
834 
835 	while (!list_empty(&prune_list)) {
836 		struct audit_tree *victim;
837 
838 		victim = list_entry(prune_list.next, struct audit_tree, list);
839 		list_del_init(&victim->list);
840 
841 		mutex_unlock(&audit_filter_mutex);
842 
843 		prune_one(victim);
844 
845 		mutex_lock(&audit_filter_mutex);
846 	}
847 
848 	mutex_unlock(&audit_filter_mutex);
849 	mutex_unlock(&audit_cmd_mutex);
850 	return 0;
851 }
852 
853 static void audit_schedule_prune(void)
854 {
855 	kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
856 }
857 
858 /*
859  * ... and that one is done if evict_chunk() decides to delay until the end
860  * of syscall.  Runs synchronously.
861  */
862 void audit_kill_trees(struct list_head *list)
863 {
864 	mutex_lock(&audit_cmd_mutex);
865 	mutex_lock(&audit_filter_mutex);
866 
867 	while (!list_empty(list)) {
868 		struct audit_tree *victim;
869 
870 		victim = list_entry(list->next, struct audit_tree, list);
871 		kill_rules(victim);
872 		list_del_init(&victim->list);
873 
874 		mutex_unlock(&audit_filter_mutex);
875 
876 		prune_one(victim);
877 
878 		mutex_lock(&audit_filter_mutex);
879 	}
880 
881 	mutex_unlock(&audit_filter_mutex);
882 	mutex_unlock(&audit_cmd_mutex);
883 }
884 
885 /*
886  *  Here comes the stuff asynchronous to auditctl operations
887  */
888 
889 /* inode->inotify_mutex is locked */
890 static void evict_chunk(struct audit_chunk *chunk)
891 {
892 	struct audit_tree *owner;
893 	struct list_head *postponed = audit_killed_trees();
894 	int need_prune = 0;
895 	int n;
896 
897 	if (chunk->dead)
898 		return;
899 
900 	chunk->dead = 1;
901 	mutex_lock(&audit_filter_mutex);
902 	spin_lock(&hash_lock);
903 	while (!list_empty(&chunk->trees)) {
904 		owner = list_entry(chunk->trees.next,
905 				   struct audit_tree, same_root);
906 		owner->goner = 1;
907 		owner->root = NULL;
908 		list_del_init(&owner->same_root);
909 		spin_unlock(&hash_lock);
910 		if (!postponed) {
911 			kill_rules(owner);
912 			list_move(&owner->list, &prune_list);
913 			need_prune = 1;
914 		} else {
915 			list_move(&owner->list, postponed);
916 		}
917 		spin_lock(&hash_lock);
918 	}
919 	list_del_rcu(&chunk->hash);
920 	for (n = 0; n < chunk->count; n++)
921 		list_del_init(&chunk->owners[n].list);
922 	spin_unlock(&hash_lock);
923 	if (need_prune)
924 		audit_schedule_prune();
925 	mutex_unlock(&audit_filter_mutex);
926 }
927 
928 static void handle_event(struct inotify_watch *watch, u32 wd, u32 mask,
929                          u32 cookie, const char *dname, struct inode *inode)
930 {
931 	struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
932 
933 	if (mask & IN_IGNORED) {
934 		evict_chunk(chunk);
935 		put_inotify_watch(watch);
936 	}
937 }
938 
939 static void destroy_watch(struct inotify_watch *watch)
940 {
941 	struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
942 	call_rcu(&chunk->head, __put_chunk);
943 }
944 
945 static const struct inotify_operations rtree_inotify_ops = {
946 	.handle_event	= handle_event,
947 	.destroy_watch	= destroy_watch,
948 };
949 
950 static int __init audit_tree_init(void)
951 {
952 	int i;
953 
954 	rtree_ih = inotify_init(&rtree_inotify_ops);
955 	if (IS_ERR(rtree_ih))
956 		audit_panic("cannot initialize inotify handle for rectree watches");
957 
958 	for (i = 0; i < HASH_SIZE; i++)
959 		INIT_LIST_HEAD(&chunk_hash_heads[i]);
960 
961 	return 0;
962 }
963 __initcall(audit_tree_init);
964