xref: /openbmc/linux/kernel/audit_tree.c (revision d0e22329)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "audit.h"
3 #include <linux/fsnotify_backend.h>
4 #include <linux/namei.h>
5 #include <linux/mount.h>
6 #include <linux/kthread.h>
7 #include <linux/refcount.h>
8 #include <linux/slab.h>
9 
10 struct audit_tree;
11 struct audit_chunk;
12 
13 struct audit_tree {
14 	refcount_t count;
15 	int goner;
16 	struct audit_chunk *root;
17 	struct list_head chunks;
18 	struct list_head rules;
19 	struct list_head list;
20 	struct list_head same_root;
21 	struct rcu_head head;
22 	char pathname[];
23 };
24 
25 struct audit_chunk {
26 	struct list_head hash;
27 	unsigned long key;
28 	struct fsnotify_mark *mark;
29 	struct list_head trees;		/* with root here */
30 	int count;
31 	atomic_long_t refs;
32 	struct rcu_head head;
33 	struct node {
34 		struct list_head list;
35 		struct audit_tree *owner;
36 		unsigned index;		/* index; upper bit indicates 'will prune' */
37 	} owners[];
38 };
39 
40 struct audit_tree_mark {
41 	struct fsnotify_mark mark;
42 	struct audit_chunk *chunk;
43 };
44 
45 static LIST_HEAD(tree_list);
46 static LIST_HEAD(prune_list);
47 static struct task_struct *prune_thread;
48 
49 /*
50  * One struct chunk is attached to each inode of interest through
51  * audit_tree_mark (fsnotify mark). We replace struct chunk on tagging /
52  * untagging, the mark is stable as long as there is chunk attached. The
53  * association between mark and chunk is protected by hash_lock and
54  * audit_tree_group->mark_mutex. Thus as long as we hold
55  * audit_tree_group->mark_mutex and check that the mark is alive by
56  * FSNOTIFY_MARK_FLAG_ATTACHED flag check, we are sure the mark points to
57  * the current chunk.
58  *
59  * Rules have pointer to struct audit_tree.
60  * Rules have struct list_head rlist forming a list of rules over
61  * the same tree.
62  * References to struct chunk are collected at audit_inode{,_child}()
63  * time and used in AUDIT_TREE rule matching.
64  * These references are dropped at the same time we are calling
65  * audit_free_names(), etc.
66  *
67  * Cyclic lists galore:
68  * tree.chunks anchors chunk.owners[].list			hash_lock
69  * tree.rules anchors rule.rlist				audit_filter_mutex
70  * chunk.trees anchors tree.same_root				hash_lock
71  * chunk.hash is a hash with middle bits of watch.inode as
72  * a hash function.						RCU, hash_lock
73  *
74  * tree is refcounted; one reference for "some rules on rules_list refer to
75  * it", one for each chunk with pointer to it.
76  *
77  * chunk is refcounted by embedded .refs. Mark associated with the chunk holds
78  * one chunk reference. This reference is dropped either when a mark is going
79  * to be freed (corresponding inode goes away) or when chunk attached to the
80  * mark gets replaced. This reference must be dropped using
81  * audit_mark_put_chunk() to make sure the reference is dropped only after RCU
82  * grace period as it protects RCU readers of the hash table.
83  *
84  * node.index allows to get from node.list to containing chunk.
85  * MSB of that sucker is stolen to mark taggings that we might have to
86  * revert - several operations have very unpleasant cleanup logics and
87  * that makes a difference.  Some.
88  */
89 
90 static struct fsnotify_group *audit_tree_group;
91 static struct kmem_cache *audit_tree_mark_cachep __read_mostly;
92 
93 static struct audit_tree *alloc_tree(const char *s)
94 {
95 	struct audit_tree *tree;
96 
97 	tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
98 	if (tree) {
99 		refcount_set(&tree->count, 1);
100 		tree->goner = 0;
101 		INIT_LIST_HEAD(&tree->chunks);
102 		INIT_LIST_HEAD(&tree->rules);
103 		INIT_LIST_HEAD(&tree->list);
104 		INIT_LIST_HEAD(&tree->same_root);
105 		tree->root = NULL;
106 		strcpy(tree->pathname, s);
107 	}
108 	return tree;
109 }
110 
111 static inline void get_tree(struct audit_tree *tree)
112 {
113 	refcount_inc(&tree->count);
114 }
115 
116 static inline void put_tree(struct audit_tree *tree)
117 {
118 	if (refcount_dec_and_test(&tree->count))
119 		kfree_rcu(tree, head);
120 }
121 
122 /* to avoid bringing the entire thing in audit.h */
123 const char *audit_tree_path(struct audit_tree *tree)
124 {
125 	return tree->pathname;
126 }
127 
128 static void free_chunk(struct audit_chunk *chunk)
129 {
130 	int i;
131 
132 	for (i = 0; i < chunk->count; i++) {
133 		if (chunk->owners[i].owner)
134 			put_tree(chunk->owners[i].owner);
135 	}
136 	kfree(chunk);
137 }
138 
139 void audit_put_chunk(struct audit_chunk *chunk)
140 {
141 	if (atomic_long_dec_and_test(&chunk->refs))
142 		free_chunk(chunk);
143 }
144 
145 static void __put_chunk(struct rcu_head *rcu)
146 {
147 	struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
148 	audit_put_chunk(chunk);
149 }
150 
151 /*
152  * Drop reference to the chunk that was held by the mark. This is the reference
153  * that gets dropped after we've removed the chunk from the hash table and we
154  * use it to make sure chunk cannot be freed before RCU grace period expires.
155  */
156 static void audit_mark_put_chunk(struct audit_chunk *chunk)
157 {
158 	call_rcu(&chunk->head, __put_chunk);
159 }
160 
161 static inline struct audit_tree_mark *audit_mark(struct fsnotify_mark *mark)
162 {
163 	return container_of(mark, struct audit_tree_mark, mark);
164 }
165 
166 static struct audit_chunk *mark_chunk(struct fsnotify_mark *mark)
167 {
168 	return audit_mark(mark)->chunk;
169 }
170 
171 static void audit_tree_destroy_watch(struct fsnotify_mark *mark)
172 {
173 	kmem_cache_free(audit_tree_mark_cachep, audit_mark(mark));
174 }
175 
176 static struct fsnotify_mark *alloc_mark(void)
177 {
178 	struct audit_tree_mark *amark;
179 
180 	amark = kmem_cache_zalloc(audit_tree_mark_cachep, GFP_KERNEL);
181 	if (!amark)
182 		return NULL;
183 	fsnotify_init_mark(&amark->mark, audit_tree_group);
184 	amark->mark.mask = FS_IN_IGNORED;
185 	return &amark->mark;
186 }
187 
188 static struct audit_chunk *alloc_chunk(int count)
189 {
190 	struct audit_chunk *chunk;
191 	size_t size;
192 	int i;
193 
194 	size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
195 	chunk = kzalloc(size, GFP_KERNEL);
196 	if (!chunk)
197 		return NULL;
198 
199 	INIT_LIST_HEAD(&chunk->hash);
200 	INIT_LIST_HEAD(&chunk->trees);
201 	chunk->count = count;
202 	atomic_long_set(&chunk->refs, 1);
203 	for (i = 0; i < count; i++) {
204 		INIT_LIST_HEAD(&chunk->owners[i].list);
205 		chunk->owners[i].index = i;
206 	}
207 	return chunk;
208 }
209 
210 enum {HASH_SIZE = 128};
211 static struct list_head chunk_hash_heads[HASH_SIZE];
212 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
213 
214 /* Function to return search key in our hash from inode. */
215 static unsigned long inode_to_key(const struct inode *inode)
216 {
217 	/* Use address pointed to by connector->obj as the key */
218 	return (unsigned long)&inode->i_fsnotify_marks;
219 }
220 
221 static inline struct list_head *chunk_hash(unsigned long key)
222 {
223 	unsigned long n = key / L1_CACHE_BYTES;
224 	return chunk_hash_heads + n % HASH_SIZE;
225 }
226 
227 /* hash_lock & mark->group->mark_mutex is held by caller */
228 static void insert_hash(struct audit_chunk *chunk)
229 {
230 	struct list_head *list;
231 
232 	/*
233 	 * Make sure chunk is fully initialized before making it visible in the
234 	 * hash. Pairs with a data dependency barrier in READ_ONCE() in
235 	 * audit_tree_lookup().
236 	 */
237 	smp_wmb();
238 	WARN_ON_ONCE(!chunk->key);
239 	list = chunk_hash(chunk->key);
240 	list_add_rcu(&chunk->hash, list);
241 }
242 
243 /* called under rcu_read_lock */
244 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
245 {
246 	unsigned long key = inode_to_key(inode);
247 	struct list_head *list = chunk_hash(key);
248 	struct audit_chunk *p;
249 
250 	list_for_each_entry_rcu(p, list, hash) {
251 		/*
252 		 * We use a data dependency barrier in READ_ONCE() to make sure
253 		 * the chunk we see is fully initialized.
254 		 */
255 		if (READ_ONCE(p->key) == key) {
256 			atomic_long_inc(&p->refs);
257 			return p;
258 		}
259 	}
260 	return NULL;
261 }
262 
263 bool audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
264 {
265 	int n;
266 	for (n = 0; n < chunk->count; n++)
267 		if (chunk->owners[n].owner == tree)
268 			return true;
269 	return false;
270 }
271 
272 /* tagging and untagging inodes with trees */
273 
274 static struct audit_chunk *find_chunk(struct node *p)
275 {
276 	int index = p->index & ~(1U<<31);
277 	p -= index;
278 	return container_of(p, struct audit_chunk, owners[0]);
279 }
280 
281 static void replace_mark_chunk(struct fsnotify_mark *mark,
282 			       struct audit_chunk *chunk)
283 {
284 	struct audit_chunk *old;
285 
286 	assert_spin_locked(&hash_lock);
287 	old = mark_chunk(mark);
288 	audit_mark(mark)->chunk = chunk;
289 	if (chunk)
290 		chunk->mark = mark;
291 	if (old)
292 		old->mark = NULL;
293 }
294 
295 static void replace_chunk(struct audit_chunk *new, struct audit_chunk *old)
296 {
297 	struct audit_tree *owner;
298 	int i, j;
299 
300 	new->key = old->key;
301 	list_splice_init(&old->trees, &new->trees);
302 	list_for_each_entry(owner, &new->trees, same_root)
303 		owner->root = new;
304 	for (i = j = 0; j < old->count; i++, j++) {
305 		if (!old->owners[j].owner) {
306 			i--;
307 			continue;
308 		}
309 		owner = old->owners[j].owner;
310 		new->owners[i].owner = owner;
311 		new->owners[i].index = old->owners[j].index - j + i;
312 		if (!owner) /* result of earlier fallback */
313 			continue;
314 		get_tree(owner);
315 		list_replace_init(&old->owners[j].list, &new->owners[i].list);
316 	}
317 	replace_mark_chunk(old->mark, new);
318 	/*
319 	 * Make sure chunk is fully initialized before making it visible in the
320 	 * hash. Pairs with a data dependency barrier in READ_ONCE() in
321 	 * audit_tree_lookup().
322 	 */
323 	smp_wmb();
324 	list_replace_rcu(&old->hash, &new->hash);
325 }
326 
327 static void remove_chunk_node(struct audit_chunk *chunk, struct node *p)
328 {
329 	struct audit_tree *owner = p->owner;
330 
331 	if (owner->root == chunk) {
332 		list_del_init(&owner->same_root);
333 		owner->root = NULL;
334 	}
335 	list_del_init(&p->list);
336 	p->owner = NULL;
337 	put_tree(owner);
338 }
339 
340 static int chunk_count_trees(struct audit_chunk *chunk)
341 {
342 	int i;
343 	int ret = 0;
344 
345 	for (i = 0; i < chunk->count; i++)
346 		if (chunk->owners[i].owner)
347 			ret++;
348 	return ret;
349 }
350 
351 static void untag_chunk(struct audit_chunk *chunk, struct fsnotify_mark *mark)
352 {
353 	struct audit_chunk *new;
354 	int size;
355 
356 	mutex_lock(&audit_tree_group->mark_mutex);
357 	/*
358 	 * mark_mutex stabilizes chunk attached to the mark so we can check
359 	 * whether it didn't change while we've dropped hash_lock.
360 	 */
361 	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) ||
362 	    mark_chunk(mark) != chunk)
363 		goto out_mutex;
364 
365 	size = chunk_count_trees(chunk);
366 	if (!size) {
367 		spin_lock(&hash_lock);
368 		list_del_init(&chunk->trees);
369 		list_del_rcu(&chunk->hash);
370 		replace_mark_chunk(mark, NULL);
371 		spin_unlock(&hash_lock);
372 		fsnotify_detach_mark(mark);
373 		mutex_unlock(&audit_tree_group->mark_mutex);
374 		audit_mark_put_chunk(chunk);
375 		fsnotify_free_mark(mark);
376 		return;
377 	}
378 
379 	new = alloc_chunk(size);
380 	if (!new)
381 		goto out_mutex;
382 
383 	spin_lock(&hash_lock);
384 	/*
385 	 * This has to go last when updating chunk as once replace_chunk() is
386 	 * called, new RCU readers can see the new chunk.
387 	 */
388 	replace_chunk(new, chunk);
389 	spin_unlock(&hash_lock);
390 	mutex_unlock(&audit_tree_group->mark_mutex);
391 	audit_mark_put_chunk(chunk);
392 	return;
393 
394 out_mutex:
395 	mutex_unlock(&audit_tree_group->mark_mutex);
396 }
397 
398 /* Call with group->mark_mutex held, releases it */
399 static int create_chunk(struct inode *inode, struct audit_tree *tree)
400 {
401 	struct fsnotify_mark *mark;
402 	struct audit_chunk *chunk = alloc_chunk(1);
403 
404 	if (!chunk) {
405 		mutex_unlock(&audit_tree_group->mark_mutex);
406 		return -ENOMEM;
407 	}
408 
409 	mark = alloc_mark();
410 	if (!mark) {
411 		mutex_unlock(&audit_tree_group->mark_mutex);
412 		kfree(chunk);
413 		return -ENOMEM;
414 	}
415 
416 	if (fsnotify_add_inode_mark_locked(mark, inode, 0)) {
417 		mutex_unlock(&audit_tree_group->mark_mutex);
418 		fsnotify_put_mark(mark);
419 		kfree(chunk);
420 		return -ENOSPC;
421 	}
422 
423 	spin_lock(&hash_lock);
424 	if (tree->goner) {
425 		spin_unlock(&hash_lock);
426 		fsnotify_detach_mark(mark);
427 		mutex_unlock(&audit_tree_group->mark_mutex);
428 		fsnotify_free_mark(mark);
429 		fsnotify_put_mark(mark);
430 		kfree(chunk);
431 		return 0;
432 	}
433 	replace_mark_chunk(mark, chunk);
434 	chunk->owners[0].index = (1U << 31);
435 	chunk->owners[0].owner = tree;
436 	get_tree(tree);
437 	list_add(&chunk->owners[0].list, &tree->chunks);
438 	if (!tree->root) {
439 		tree->root = chunk;
440 		list_add(&tree->same_root, &chunk->trees);
441 	}
442 	chunk->key = inode_to_key(inode);
443 	/*
444 	 * Inserting into the hash table has to go last as once we do that RCU
445 	 * readers can see the chunk.
446 	 */
447 	insert_hash(chunk);
448 	spin_unlock(&hash_lock);
449 	mutex_unlock(&audit_tree_group->mark_mutex);
450 	/*
451 	 * Drop our initial reference. When mark we point to is getting freed,
452 	 * we get notification through ->freeing_mark callback and cleanup
453 	 * chunk pointing to this mark.
454 	 */
455 	fsnotify_put_mark(mark);
456 	return 0;
457 }
458 
459 /* the first tagged inode becomes root of tree */
460 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
461 {
462 	struct fsnotify_mark *mark;
463 	struct audit_chunk *chunk, *old;
464 	struct node *p;
465 	int n;
466 
467 	mutex_lock(&audit_tree_group->mark_mutex);
468 	mark = fsnotify_find_mark(&inode->i_fsnotify_marks, audit_tree_group);
469 	if (!mark)
470 		return create_chunk(inode, tree);
471 
472 	/*
473 	 * Found mark is guaranteed to be attached and mark_mutex protects mark
474 	 * from getting detached and thus it makes sure there is chunk attached
475 	 * to the mark.
476 	 */
477 	/* are we already there? */
478 	spin_lock(&hash_lock);
479 	old = mark_chunk(mark);
480 	for (n = 0; n < old->count; n++) {
481 		if (old->owners[n].owner == tree) {
482 			spin_unlock(&hash_lock);
483 			mutex_unlock(&audit_tree_group->mark_mutex);
484 			fsnotify_put_mark(mark);
485 			return 0;
486 		}
487 	}
488 	spin_unlock(&hash_lock);
489 
490 	chunk = alloc_chunk(old->count + 1);
491 	if (!chunk) {
492 		mutex_unlock(&audit_tree_group->mark_mutex);
493 		fsnotify_put_mark(mark);
494 		return -ENOMEM;
495 	}
496 
497 	spin_lock(&hash_lock);
498 	if (tree->goner) {
499 		spin_unlock(&hash_lock);
500 		mutex_unlock(&audit_tree_group->mark_mutex);
501 		fsnotify_put_mark(mark);
502 		kfree(chunk);
503 		return 0;
504 	}
505 	p = &chunk->owners[chunk->count - 1];
506 	p->index = (chunk->count - 1) | (1U<<31);
507 	p->owner = tree;
508 	get_tree(tree);
509 	list_add(&p->list, &tree->chunks);
510 	if (!tree->root) {
511 		tree->root = chunk;
512 		list_add(&tree->same_root, &chunk->trees);
513 	}
514 	/*
515 	 * This has to go last when updating chunk as once replace_chunk() is
516 	 * called, new RCU readers can see the new chunk.
517 	 */
518 	replace_chunk(chunk, old);
519 	spin_unlock(&hash_lock);
520 	mutex_unlock(&audit_tree_group->mark_mutex);
521 	fsnotify_put_mark(mark); /* pair to fsnotify_find_mark */
522 	audit_mark_put_chunk(old);
523 
524 	return 0;
525 }
526 
527 static void audit_tree_log_remove_rule(struct audit_krule *rule)
528 {
529 	struct audit_buffer *ab;
530 
531 	if (!audit_enabled)
532 		return;
533 	ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
534 	if (unlikely(!ab))
535 		return;
536 	audit_log_format(ab, "op=remove_rule dir=");
537 	audit_log_untrustedstring(ab, rule->tree->pathname);
538 	audit_log_key(ab, rule->filterkey);
539 	audit_log_format(ab, " list=%d res=1", rule->listnr);
540 	audit_log_end(ab);
541 }
542 
543 static void kill_rules(struct audit_tree *tree)
544 {
545 	struct audit_krule *rule, *next;
546 	struct audit_entry *entry;
547 
548 	list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
549 		entry = container_of(rule, struct audit_entry, rule);
550 
551 		list_del_init(&rule->rlist);
552 		if (rule->tree) {
553 			/* not a half-baked one */
554 			audit_tree_log_remove_rule(rule);
555 			if (entry->rule.exe)
556 				audit_remove_mark(entry->rule.exe);
557 			rule->tree = NULL;
558 			list_del_rcu(&entry->list);
559 			list_del(&entry->rule.list);
560 			call_rcu(&entry->rcu, audit_free_rule_rcu);
561 		}
562 	}
563 }
564 
565 /*
566  * Remove tree from chunks. If 'tagged' is set, remove tree only from tagged
567  * chunks. The function expects tagged chunks are all at the beginning of the
568  * chunks list.
569  */
570 static void prune_tree_chunks(struct audit_tree *victim, bool tagged)
571 {
572 	spin_lock(&hash_lock);
573 	while (!list_empty(&victim->chunks)) {
574 		struct node *p;
575 		struct audit_chunk *chunk;
576 		struct fsnotify_mark *mark;
577 
578 		p = list_first_entry(&victim->chunks, struct node, list);
579 		/* have we run out of marked? */
580 		if (tagged && !(p->index & (1U<<31)))
581 			break;
582 		chunk = find_chunk(p);
583 		mark = chunk->mark;
584 		remove_chunk_node(chunk, p);
585 		/* Racing with audit_tree_freeing_mark()? */
586 		if (!mark)
587 			continue;
588 		fsnotify_get_mark(mark);
589 		spin_unlock(&hash_lock);
590 
591 		untag_chunk(chunk, mark);
592 		fsnotify_put_mark(mark);
593 
594 		spin_lock(&hash_lock);
595 	}
596 	spin_unlock(&hash_lock);
597 	put_tree(victim);
598 }
599 
600 /*
601  * finish killing struct audit_tree
602  */
603 static void prune_one(struct audit_tree *victim)
604 {
605 	prune_tree_chunks(victim, false);
606 }
607 
608 /* trim the uncommitted chunks from tree */
609 
610 static void trim_marked(struct audit_tree *tree)
611 {
612 	struct list_head *p, *q;
613 	spin_lock(&hash_lock);
614 	if (tree->goner) {
615 		spin_unlock(&hash_lock);
616 		return;
617 	}
618 	/* reorder */
619 	for (p = tree->chunks.next; p != &tree->chunks; p = q) {
620 		struct node *node = list_entry(p, struct node, list);
621 		q = p->next;
622 		if (node->index & (1U<<31)) {
623 			list_del_init(p);
624 			list_add(p, &tree->chunks);
625 		}
626 	}
627 	spin_unlock(&hash_lock);
628 
629 	prune_tree_chunks(tree, true);
630 
631 	spin_lock(&hash_lock);
632 	if (!tree->root && !tree->goner) {
633 		tree->goner = 1;
634 		spin_unlock(&hash_lock);
635 		mutex_lock(&audit_filter_mutex);
636 		kill_rules(tree);
637 		list_del_init(&tree->list);
638 		mutex_unlock(&audit_filter_mutex);
639 		prune_one(tree);
640 	} else {
641 		spin_unlock(&hash_lock);
642 	}
643 }
644 
645 static void audit_schedule_prune(void);
646 
647 /* called with audit_filter_mutex */
648 int audit_remove_tree_rule(struct audit_krule *rule)
649 {
650 	struct audit_tree *tree;
651 	tree = rule->tree;
652 	if (tree) {
653 		spin_lock(&hash_lock);
654 		list_del_init(&rule->rlist);
655 		if (list_empty(&tree->rules) && !tree->goner) {
656 			tree->root = NULL;
657 			list_del_init(&tree->same_root);
658 			tree->goner = 1;
659 			list_move(&tree->list, &prune_list);
660 			rule->tree = NULL;
661 			spin_unlock(&hash_lock);
662 			audit_schedule_prune();
663 			return 1;
664 		}
665 		rule->tree = NULL;
666 		spin_unlock(&hash_lock);
667 		return 1;
668 	}
669 	return 0;
670 }
671 
672 static int compare_root(struct vfsmount *mnt, void *arg)
673 {
674 	return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
675 	       (unsigned long)arg;
676 }
677 
678 void audit_trim_trees(void)
679 {
680 	struct list_head cursor;
681 
682 	mutex_lock(&audit_filter_mutex);
683 	list_add(&cursor, &tree_list);
684 	while (cursor.next != &tree_list) {
685 		struct audit_tree *tree;
686 		struct path path;
687 		struct vfsmount *root_mnt;
688 		struct node *node;
689 		int err;
690 
691 		tree = container_of(cursor.next, struct audit_tree, list);
692 		get_tree(tree);
693 		list_del(&cursor);
694 		list_add(&cursor, &tree->list);
695 		mutex_unlock(&audit_filter_mutex);
696 
697 		err = kern_path(tree->pathname, 0, &path);
698 		if (err)
699 			goto skip_it;
700 
701 		root_mnt = collect_mounts(&path);
702 		path_put(&path);
703 		if (IS_ERR(root_mnt))
704 			goto skip_it;
705 
706 		spin_lock(&hash_lock);
707 		list_for_each_entry(node, &tree->chunks, list) {
708 			struct audit_chunk *chunk = find_chunk(node);
709 			/* this could be NULL if the watch is dying else where... */
710 			node->index |= 1U<<31;
711 			if (iterate_mounts(compare_root,
712 					   (void *)(chunk->key),
713 					   root_mnt))
714 				node->index &= ~(1U<<31);
715 		}
716 		spin_unlock(&hash_lock);
717 		trim_marked(tree);
718 		drop_collected_mounts(root_mnt);
719 skip_it:
720 		put_tree(tree);
721 		mutex_lock(&audit_filter_mutex);
722 	}
723 	list_del(&cursor);
724 	mutex_unlock(&audit_filter_mutex);
725 }
726 
727 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
728 {
729 
730 	if (pathname[0] != '/' ||
731 	    rule->listnr != AUDIT_FILTER_EXIT ||
732 	    op != Audit_equal ||
733 	    rule->inode_f || rule->watch || rule->tree)
734 		return -EINVAL;
735 	rule->tree = alloc_tree(pathname);
736 	if (!rule->tree)
737 		return -ENOMEM;
738 	return 0;
739 }
740 
741 void audit_put_tree(struct audit_tree *tree)
742 {
743 	put_tree(tree);
744 }
745 
746 static int tag_mount(struct vfsmount *mnt, void *arg)
747 {
748 	return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
749 }
750 
751 /*
752  * That gets run when evict_chunk() ends up needing to kill audit_tree.
753  * Runs from a separate thread.
754  */
755 static int prune_tree_thread(void *unused)
756 {
757 	for (;;) {
758 		if (list_empty(&prune_list)) {
759 			set_current_state(TASK_INTERRUPTIBLE);
760 			schedule();
761 		}
762 
763 		audit_ctl_lock();
764 		mutex_lock(&audit_filter_mutex);
765 
766 		while (!list_empty(&prune_list)) {
767 			struct audit_tree *victim;
768 
769 			victim = list_entry(prune_list.next,
770 					struct audit_tree, list);
771 			list_del_init(&victim->list);
772 
773 			mutex_unlock(&audit_filter_mutex);
774 
775 			prune_one(victim);
776 
777 			mutex_lock(&audit_filter_mutex);
778 		}
779 
780 		mutex_unlock(&audit_filter_mutex);
781 		audit_ctl_unlock();
782 	}
783 	return 0;
784 }
785 
786 static int audit_launch_prune(void)
787 {
788 	if (prune_thread)
789 		return 0;
790 	prune_thread = kthread_run(prune_tree_thread, NULL,
791 				"audit_prune_tree");
792 	if (IS_ERR(prune_thread)) {
793 		pr_err("cannot start thread audit_prune_tree");
794 		prune_thread = NULL;
795 		return -ENOMEM;
796 	}
797 	return 0;
798 }
799 
800 /* called with audit_filter_mutex */
801 int audit_add_tree_rule(struct audit_krule *rule)
802 {
803 	struct audit_tree *seed = rule->tree, *tree;
804 	struct path path;
805 	struct vfsmount *mnt;
806 	int err;
807 
808 	rule->tree = NULL;
809 	list_for_each_entry(tree, &tree_list, list) {
810 		if (!strcmp(seed->pathname, tree->pathname)) {
811 			put_tree(seed);
812 			rule->tree = tree;
813 			list_add(&rule->rlist, &tree->rules);
814 			return 0;
815 		}
816 	}
817 	tree = seed;
818 	list_add(&tree->list, &tree_list);
819 	list_add(&rule->rlist, &tree->rules);
820 	/* do not set rule->tree yet */
821 	mutex_unlock(&audit_filter_mutex);
822 
823 	if (unlikely(!prune_thread)) {
824 		err = audit_launch_prune();
825 		if (err)
826 			goto Err;
827 	}
828 
829 	err = kern_path(tree->pathname, 0, &path);
830 	if (err)
831 		goto Err;
832 	mnt = collect_mounts(&path);
833 	path_put(&path);
834 	if (IS_ERR(mnt)) {
835 		err = PTR_ERR(mnt);
836 		goto Err;
837 	}
838 
839 	get_tree(tree);
840 	err = iterate_mounts(tag_mount, tree, mnt);
841 	drop_collected_mounts(mnt);
842 
843 	if (!err) {
844 		struct node *node;
845 		spin_lock(&hash_lock);
846 		list_for_each_entry(node, &tree->chunks, list)
847 			node->index &= ~(1U<<31);
848 		spin_unlock(&hash_lock);
849 	} else {
850 		trim_marked(tree);
851 		goto Err;
852 	}
853 
854 	mutex_lock(&audit_filter_mutex);
855 	if (list_empty(&rule->rlist)) {
856 		put_tree(tree);
857 		return -ENOENT;
858 	}
859 	rule->tree = tree;
860 	put_tree(tree);
861 
862 	return 0;
863 Err:
864 	mutex_lock(&audit_filter_mutex);
865 	list_del_init(&tree->list);
866 	list_del_init(&tree->rules);
867 	put_tree(tree);
868 	return err;
869 }
870 
871 int audit_tag_tree(char *old, char *new)
872 {
873 	struct list_head cursor, barrier;
874 	int failed = 0;
875 	struct path path1, path2;
876 	struct vfsmount *tagged;
877 	int err;
878 
879 	err = kern_path(new, 0, &path2);
880 	if (err)
881 		return err;
882 	tagged = collect_mounts(&path2);
883 	path_put(&path2);
884 	if (IS_ERR(tagged))
885 		return PTR_ERR(tagged);
886 
887 	err = kern_path(old, 0, &path1);
888 	if (err) {
889 		drop_collected_mounts(tagged);
890 		return err;
891 	}
892 
893 	mutex_lock(&audit_filter_mutex);
894 	list_add(&barrier, &tree_list);
895 	list_add(&cursor, &barrier);
896 
897 	while (cursor.next != &tree_list) {
898 		struct audit_tree *tree;
899 		int good_one = 0;
900 
901 		tree = container_of(cursor.next, struct audit_tree, list);
902 		get_tree(tree);
903 		list_del(&cursor);
904 		list_add(&cursor, &tree->list);
905 		mutex_unlock(&audit_filter_mutex);
906 
907 		err = kern_path(tree->pathname, 0, &path2);
908 		if (!err) {
909 			good_one = path_is_under(&path1, &path2);
910 			path_put(&path2);
911 		}
912 
913 		if (!good_one) {
914 			put_tree(tree);
915 			mutex_lock(&audit_filter_mutex);
916 			continue;
917 		}
918 
919 		failed = iterate_mounts(tag_mount, tree, tagged);
920 		if (failed) {
921 			put_tree(tree);
922 			mutex_lock(&audit_filter_mutex);
923 			break;
924 		}
925 
926 		mutex_lock(&audit_filter_mutex);
927 		spin_lock(&hash_lock);
928 		if (!tree->goner) {
929 			list_del(&tree->list);
930 			list_add(&tree->list, &tree_list);
931 		}
932 		spin_unlock(&hash_lock);
933 		put_tree(tree);
934 	}
935 
936 	while (barrier.prev != &tree_list) {
937 		struct audit_tree *tree;
938 
939 		tree = container_of(barrier.prev, struct audit_tree, list);
940 		get_tree(tree);
941 		list_del(&tree->list);
942 		list_add(&tree->list, &barrier);
943 		mutex_unlock(&audit_filter_mutex);
944 
945 		if (!failed) {
946 			struct node *node;
947 			spin_lock(&hash_lock);
948 			list_for_each_entry(node, &tree->chunks, list)
949 				node->index &= ~(1U<<31);
950 			spin_unlock(&hash_lock);
951 		} else {
952 			trim_marked(tree);
953 		}
954 
955 		put_tree(tree);
956 		mutex_lock(&audit_filter_mutex);
957 	}
958 	list_del(&barrier);
959 	list_del(&cursor);
960 	mutex_unlock(&audit_filter_mutex);
961 	path_put(&path1);
962 	drop_collected_mounts(tagged);
963 	return failed;
964 }
965 
966 
967 static void audit_schedule_prune(void)
968 {
969 	wake_up_process(prune_thread);
970 }
971 
972 /*
973  * ... and that one is done if evict_chunk() decides to delay until the end
974  * of syscall.  Runs synchronously.
975  */
976 void audit_kill_trees(struct list_head *list)
977 {
978 	audit_ctl_lock();
979 	mutex_lock(&audit_filter_mutex);
980 
981 	while (!list_empty(list)) {
982 		struct audit_tree *victim;
983 
984 		victim = list_entry(list->next, struct audit_tree, list);
985 		kill_rules(victim);
986 		list_del_init(&victim->list);
987 
988 		mutex_unlock(&audit_filter_mutex);
989 
990 		prune_one(victim);
991 
992 		mutex_lock(&audit_filter_mutex);
993 	}
994 
995 	mutex_unlock(&audit_filter_mutex);
996 	audit_ctl_unlock();
997 }
998 
999 /*
1000  *  Here comes the stuff asynchronous to auditctl operations
1001  */
1002 
1003 static void evict_chunk(struct audit_chunk *chunk)
1004 {
1005 	struct audit_tree *owner;
1006 	struct list_head *postponed = audit_killed_trees();
1007 	int need_prune = 0;
1008 	int n;
1009 
1010 	mutex_lock(&audit_filter_mutex);
1011 	spin_lock(&hash_lock);
1012 	while (!list_empty(&chunk->trees)) {
1013 		owner = list_entry(chunk->trees.next,
1014 				   struct audit_tree, same_root);
1015 		owner->goner = 1;
1016 		owner->root = NULL;
1017 		list_del_init(&owner->same_root);
1018 		spin_unlock(&hash_lock);
1019 		if (!postponed) {
1020 			kill_rules(owner);
1021 			list_move(&owner->list, &prune_list);
1022 			need_prune = 1;
1023 		} else {
1024 			list_move(&owner->list, postponed);
1025 		}
1026 		spin_lock(&hash_lock);
1027 	}
1028 	list_del_rcu(&chunk->hash);
1029 	for (n = 0; n < chunk->count; n++)
1030 		list_del_init(&chunk->owners[n].list);
1031 	spin_unlock(&hash_lock);
1032 	mutex_unlock(&audit_filter_mutex);
1033 	if (need_prune)
1034 		audit_schedule_prune();
1035 }
1036 
1037 static int audit_tree_handle_event(struct fsnotify_group *group,
1038 				   struct inode *to_tell,
1039 				   u32 mask, const void *data, int data_type,
1040 				   const unsigned char *file_name, u32 cookie,
1041 				   struct fsnotify_iter_info *iter_info)
1042 {
1043 	return 0;
1044 }
1045 
1046 static void audit_tree_freeing_mark(struct fsnotify_mark *mark,
1047 				    struct fsnotify_group *group)
1048 {
1049 	struct audit_chunk *chunk;
1050 
1051 	mutex_lock(&mark->group->mark_mutex);
1052 	spin_lock(&hash_lock);
1053 	chunk = mark_chunk(mark);
1054 	replace_mark_chunk(mark, NULL);
1055 	spin_unlock(&hash_lock);
1056 	mutex_unlock(&mark->group->mark_mutex);
1057 	if (chunk) {
1058 		evict_chunk(chunk);
1059 		audit_mark_put_chunk(chunk);
1060 	}
1061 
1062 	/*
1063 	 * We are guaranteed to have at least one reference to the mark from
1064 	 * either the inode or the caller of fsnotify_destroy_mark().
1065 	 */
1066 	BUG_ON(refcount_read(&mark->refcnt) < 1);
1067 }
1068 
1069 static const struct fsnotify_ops audit_tree_ops = {
1070 	.handle_event = audit_tree_handle_event,
1071 	.freeing_mark = audit_tree_freeing_mark,
1072 	.free_mark = audit_tree_destroy_watch,
1073 };
1074 
1075 static int __init audit_tree_init(void)
1076 {
1077 	int i;
1078 
1079 	audit_tree_mark_cachep = KMEM_CACHE(audit_tree_mark, SLAB_PANIC);
1080 
1081 	audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1082 	if (IS_ERR(audit_tree_group))
1083 		audit_panic("cannot initialize fsnotify group for rectree watches");
1084 
1085 	for (i = 0; i < HASH_SIZE; i++)
1086 		INIT_LIST_HEAD(&chunk_hash_heads[i]);
1087 
1088 	return 0;
1089 }
1090 __initcall(audit_tree_init);
1091