Loading...
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
10struct audit_tree;
11struct audit_chunk;
12
13struct 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
25struct 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
40struct audit_tree_mark {
41 struct fsnotify_mark mark;
42 struct audit_chunk *chunk;
43};
44
45static LIST_HEAD(tree_list);
46static LIST_HEAD(prune_list);
47static 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
90static struct fsnotify_group *audit_tree_group;
91static struct kmem_cache *audit_tree_mark_cachep __read_mostly;
92
93static 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
111static inline void get_tree(struct audit_tree *tree)
112{
113 refcount_inc(&tree->count);
114}
115
116static 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 */
123const char *audit_tree_path(struct audit_tree *tree)
124{
125 return tree->pathname;
126}
127
128static 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
139void audit_put_chunk(struct audit_chunk *chunk)
140{
141 if (atomic_long_dec_and_test(&chunk->refs))
142 free_chunk(chunk);
143}
144
145static 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 */
156static void audit_mark_put_chunk(struct audit_chunk *chunk)
157{
158 call_rcu(&chunk->head, __put_chunk);
159}
160
161static inline struct audit_tree_mark *audit_mark(struct fsnotify_mark *mark)
162{
163 return container_of(mark, struct audit_tree_mark, mark);
164}
165
166static struct audit_chunk *mark_chunk(struct fsnotify_mark *mark)
167{
168 return audit_mark(mark)->chunk;
169}
170
171static void audit_tree_destroy_watch(struct fsnotify_mark *mark)
172{
173 kmem_cache_free(audit_tree_mark_cachep, audit_mark(mark));
174}
175
176static 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
188static 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
210enum {HASH_SIZE = 128};
211static struct list_head chunk_hash_heads[HASH_SIZE];
212static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
213
214/* Function to return search key in our hash from inode. */
215static 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
221static 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 */
228static 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 */
244struct 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
263bool 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
274static 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
281static 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
295static 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
327static 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
340static 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
351static 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
394out_mutex:
395 mutex_unlock(&audit_tree_group->mark_mutex);
396}
397
398/* Call with group->mark_mutex held, releases it */
399static 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 */
460static 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
527static void audit_tree_log_remove_rule(struct audit_context *context,
528 struct audit_krule *rule)
529{
530 struct audit_buffer *ab;
531
532 if (!audit_enabled)
533 return;
534 ab = audit_log_start(context, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
535 if (unlikely(!ab))
536 return;
537 audit_log_format(ab, "op=remove_rule dir=");
538 audit_log_untrustedstring(ab, rule->tree->pathname);
539 audit_log_key(ab, rule->filterkey);
540 audit_log_format(ab, " list=%d res=1", rule->listnr);
541 audit_log_end(ab);
542}
543
544static void kill_rules(struct audit_context *context, struct audit_tree *tree)
545{
546 struct audit_krule *rule, *next;
547 struct audit_entry *entry;
548
549 list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
550 entry = container_of(rule, struct audit_entry, rule);
551
552 list_del_init(&rule->rlist);
553 if (rule->tree) {
554 /* not a half-baked one */
555 audit_tree_log_remove_rule(context, rule);
556 if (entry->rule.exe)
557 audit_remove_mark(entry->rule.exe);
558 rule->tree = NULL;
559 list_del_rcu(&entry->list);
560 list_del(&entry->rule.list);
561 call_rcu(&entry->rcu, audit_free_rule_rcu);
562 }
563 }
564}
565
566/*
567 * Remove tree from chunks. If 'tagged' is set, remove tree only from tagged
568 * chunks. The function expects tagged chunks are all at the beginning of the
569 * chunks list.
570 */
571static void prune_tree_chunks(struct audit_tree *victim, bool tagged)
572{
573 spin_lock(&hash_lock);
574 while (!list_empty(&victim->chunks)) {
575 struct node *p;
576 struct audit_chunk *chunk;
577 struct fsnotify_mark *mark;
578
579 p = list_first_entry(&victim->chunks, struct node, list);
580 /* have we run out of marked? */
581 if (tagged && !(p->index & (1U<<31)))
582 break;
583 chunk = find_chunk(p);
584 mark = chunk->mark;
585 remove_chunk_node(chunk, p);
586 /* Racing with audit_tree_freeing_mark()? */
587 if (!mark)
588 continue;
589 fsnotify_get_mark(mark);
590 spin_unlock(&hash_lock);
591
592 untag_chunk(chunk, mark);
593 fsnotify_put_mark(mark);
594
595 spin_lock(&hash_lock);
596 }
597 spin_unlock(&hash_lock);
598 put_tree(victim);
599}
600
601/*
602 * finish killing struct audit_tree
603 */
604static void prune_one(struct audit_tree *victim)
605{
606 prune_tree_chunks(victim, false);
607}
608
609/* trim the uncommitted chunks from tree */
610
611static void trim_marked(struct audit_tree *tree)
612{
613 struct list_head *p, *q;
614 spin_lock(&hash_lock);
615 if (tree->goner) {
616 spin_unlock(&hash_lock);
617 return;
618 }
619 /* reorder */
620 for (p = tree->chunks.next; p != &tree->chunks; p = q) {
621 struct node *node = list_entry(p, struct node, list);
622 q = p->next;
623 if (node->index & (1U<<31)) {
624 list_del_init(p);
625 list_add(p, &tree->chunks);
626 }
627 }
628 spin_unlock(&hash_lock);
629
630 prune_tree_chunks(tree, true);
631
632 spin_lock(&hash_lock);
633 if (!tree->root && !tree->goner) {
634 tree->goner = 1;
635 spin_unlock(&hash_lock);
636 mutex_lock(&audit_filter_mutex);
637 kill_rules(audit_context(), tree);
638 list_del_init(&tree->list);
639 mutex_unlock(&audit_filter_mutex);
640 prune_one(tree);
641 } else {
642 spin_unlock(&hash_lock);
643 }
644}
645
646static void audit_schedule_prune(void);
647
648/* called with audit_filter_mutex */
649int audit_remove_tree_rule(struct audit_krule *rule)
650{
651 struct audit_tree *tree;
652 tree = rule->tree;
653 if (tree) {
654 spin_lock(&hash_lock);
655 list_del_init(&rule->rlist);
656 if (list_empty(&tree->rules) && !tree->goner) {
657 tree->root = NULL;
658 list_del_init(&tree->same_root);
659 tree->goner = 1;
660 list_move(&tree->list, &prune_list);
661 rule->tree = NULL;
662 spin_unlock(&hash_lock);
663 audit_schedule_prune();
664 return 1;
665 }
666 rule->tree = NULL;
667 spin_unlock(&hash_lock);
668 return 1;
669 }
670 return 0;
671}
672
673static int compare_root(struct vfsmount *mnt, void *arg)
674{
675 return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
676 (unsigned long)arg;
677}
678
679void audit_trim_trees(void)
680{
681 struct list_head cursor;
682
683 mutex_lock(&audit_filter_mutex);
684 list_add(&cursor, &tree_list);
685 while (cursor.next != &tree_list) {
686 struct audit_tree *tree;
687 struct path path;
688 struct vfsmount *root_mnt;
689 struct node *node;
690 int err;
691
692 tree = container_of(cursor.next, struct audit_tree, list);
693 get_tree(tree);
694 list_del(&cursor);
695 list_add(&cursor, &tree->list);
696 mutex_unlock(&audit_filter_mutex);
697
698 err = kern_path(tree->pathname, 0, &path);
699 if (err)
700 goto skip_it;
701
702 root_mnt = collect_mounts(&path);
703 path_put(&path);
704 if (IS_ERR(root_mnt))
705 goto skip_it;
706
707 spin_lock(&hash_lock);
708 list_for_each_entry(node, &tree->chunks, list) {
709 struct audit_chunk *chunk = find_chunk(node);
710 /* this could be NULL if the watch is dying else where... */
711 node->index |= 1U<<31;
712 if (iterate_mounts(compare_root,
713 (void *)(chunk->key),
714 root_mnt))
715 node->index &= ~(1U<<31);
716 }
717 spin_unlock(&hash_lock);
718 trim_marked(tree);
719 drop_collected_mounts(root_mnt);
720skip_it:
721 put_tree(tree);
722 mutex_lock(&audit_filter_mutex);
723 }
724 list_del(&cursor);
725 mutex_unlock(&audit_filter_mutex);
726}
727
728int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
729{
730
731 if (pathname[0] != '/' ||
732 rule->listnr != AUDIT_FILTER_EXIT ||
733 op != Audit_equal ||
734 rule->inode_f || rule->watch || rule->tree)
735 return -EINVAL;
736 rule->tree = alloc_tree(pathname);
737 if (!rule->tree)
738 return -ENOMEM;
739 return 0;
740}
741
742void audit_put_tree(struct audit_tree *tree)
743{
744 put_tree(tree);
745}
746
747static int tag_mount(struct vfsmount *mnt, void *arg)
748{
749 return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
750}
751
752/*
753 * That gets run when evict_chunk() ends up needing to kill audit_tree.
754 * Runs from a separate thread.
755 */
756static int prune_tree_thread(void *unused)
757{
758 for (;;) {
759 if (list_empty(&prune_list)) {
760 set_current_state(TASK_INTERRUPTIBLE);
761 schedule();
762 }
763
764 audit_ctl_lock();
765 mutex_lock(&audit_filter_mutex);
766
767 while (!list_empty(&prune_list)) {
768 struct audit_tree *victim;
769
770 victim = list_entry(prune_list.next,
771 struct audit_tree, list);
772 list_del_init(&victim->list);
773
774 mutex_unlock(&audit_filter_mutex);
775
776 prune_one(victim);
777
778 mutex_lock(&audit_filter_mutex);
779 }
780
781 mutex_unlock(&audit_filter_mutex);
782 audit_ctl_unlock();
783 }
784 return 0;
785}
786
787static int audit_launch_prune(void)
788{
789 if (prune_thread)
790 return 0;
791 prune_thread = kthread_run(prune_tree_thread, NULL,
792 "audit_prune_tree");
793 if (IS_ERR(prune_thread)) {
794 pr_err("cannot start thread audit_prune_tree");
795 prune_thread = NULL;
796 return -ENOMEM;
797 }
798 return 0;
799}
800
801/* called with audit_filter_mutex */
802int audit_add_tree_rule(struct audit_krule *rule)
803{
804 struct audit_tree *seed = rule->tree, *tree;
805 struct path path;
806 struct vfsmount *mnt;
807 int err;
808
809 rule->tree = NULL;
810 list_for_each_entry(tree, &tree_list, list) {
811 if (!strcmp(seed->pathname, tree->pathname)) {
812 put_tree(seed);
813 rule->tree = tree;
814 list_add(&rule->rlist, &tree->rules);
815 return 0;
816 }
817 }
818 tree = seed;
819 list_add(&tree->list, &tree_list);
820 list_add(&rule->rlist, &tree->rules);
821 /* do not set rule->tree yet */
822 mutex_unlock(&audit_filter_mutex);
823
824 if (unlikely(!prune_thread)) {
825 err = audit_launch_prune();
826 if (err)
827 goto Err;
828 }
829
830 err = kern_path(tree->pathname, 0, &path);
831 if (err)
832 goto Err;
833 mnt = collect_mounts(&path);
834 path_put(&path);
835 if (IS_ERR(mnt)) {
836 err = PTR_ERR(mnt);
837 goto Err;
838 }
839
840 get_tree(tree);
841 err = iterate_mounts(tag_mount, tree, mnt);
842 drop_collected_mounts(mnt);
843
844 if (!err) {
845 struct node *node;
846 spin_lock(&hash_lock);
847 list_for_each_entry(node, &tree->chunks, list)
848 node->index &= ~(1U<<31);
849 spin_unlock(&hash_lock);
850 } else {
851 trim_marked(tree);
852 goto Err;
853 }
854
855 mutex_lock(&audit_filter_mutex);
856 if (list_empty(&rule->rlist)) {
857 put_tree(tree);
858 return -ENOENT;
859 }
860 rule->tree = tree;
861 put_tree(tree);
862
863 return 0;
864Err:
865 mutex_lock(&audit_filter_mutex);
866 list_del_init(&tree->list);
867 list_del_init(&tree->rules);
868 put_tree(tree);
869 return err;
870}
871
872int audit_tag_tree(char *old, char *new)
873{
874 struct list_head cursor, barrier;
875 int failed = 0;
876 struct path path1, path2;
877 struct vfsmount *tagged;
878 int err;
879
880 err = kern_path(new, 0, &path2);
881 if (err)
882 return err;
883 tagged = collect_mounts(&path2);
884 path_put(&path2);
885 if (IS_ERR(tagged))
886 return PTR_ERR(tagged);
887
888 err = kern_path(old, 0, &path1);
889 if (err) {
890 drop_collected_mounts(tagged);
891 return err;
892 }
893
894 mutex_lock(&audit_filter_mutex);
895 list_add(&barrier, &tree_list);
896 list_add(&cursor, &barrier);
897
898 while (cursor.next != &tree_list) {
899 struct audit_tree *tree;
900 int good_one = 0;
901
902 tree = container_of(cursor.next, struct audit_tree, list);
903 get_tree(tree);
904 list_del(&cursor);
905 list_add(&cursor, &tree->list);
906 mutex_unlock(&audit_filter_mutex);
907
908 err = kern_path(tree->pathname, 0, &path2);
909 if (!err) {
910 good_one = path_is_under(&path1, &path2);
911 path_put(&path2);
912 }
913
914 if (!good_one) {
915 put_tree(tree);
916 mutex_lock(&audit_filter_mutex);
917 continue;
918 }
919
920 failed = iterate_mounts(tag_mount, tree, tagged);
921 if (failed) {
922 put_tree(tree);
923 mutex_lock(&audit_filter_mutex);
924 break;
925 }
926
927 mutex_lock(&audit_filter_mutex);
928 spin_lock(&hash_lock);
929 if (!tree->goner) {
930 list_del(&tree->list);
931 list_add(&tree->list, &tree_list);
932 }
933 spin_unlock(&hash_lock);
934 put_tree(tree);
935 }
936
937 while (barrier.prev != &tree_list) {
938 struct audit_tree *tree;
939
940 tree = container_of(barrier.prev, struct audit_tree, list);
941 get_tree(tree);
942 list_del(&tree->list);
943 list_add(&tree->list, &barrier);
944 mutex_unlock(&audit_filter_mutex);
945
946 if (!failed) {
947 struct node *node;
948 spin_lock(&hash_lock);
949 list_for_each_entry(node, &tree->chunks, list)
950 node->index &= ~(1U<<31);
951 spin_unlock(&hash_lock);
952 } else {
953 trim_marked(tree);
954 }
955
956 put_tree(tree);
957 mutex_lock(&audit_filter_mutex);
958 }
959 list_del(&barrier);
960 list_del(&cursor);
961 mutex_unlock(&audit_filter_mutex);
962 path_put(&path1);
963 drop_collected_mounts(tagged);
964 return failed;
965}
966
967
968static void audit_schedule_prune(void)
969{
970 wake_up_process(prune_thread);
971}
972
973/*
974 * ... and that one is done if evict_chunk() decides to delay until the end
975 * of syscall. Runs synchronously.
976 */
977void audit_kill_trees(struct audit_context *context)
978{
979 struct list_head *list = &context->killed_trees;
980
981 audit_ctl_lock();
982 mutex_lock(&audit_filter_mutex);
983
984 while (!list_empty(list)) {
985 struct audit_tree *victim;
986
987 victim = list_entry(list->next, struct audit_tree, list);
988 kill_rules(context, victim);
989 list_del_init(&victim->list);
990
991 mutex_unlock(&audit_filter_mutex);
992
993 prune_one(victim);
994
995 mutex_lock(&audit_filter_mutex);
996 }
997
998 mutex_unlock(&audit_filter_mutex);
999 audit_ctl_unlock();
1000}
1001
1002/*
1003 * Here comes the stuff asynchronous to auditctl operations
1004 */
1005
1006static void evict_chunk(struct audit_chunk *chunk)
1007{
1008 struct audit_tree *owner;
1009 struct list_head *postponed = audit_killed_trees();
1010 int need_prune = 0;
1011 int n;
1012
1013 mutex_lock(&audit_filter_mutex);
1014 spin_lock(&hash_lock);
1015 while (!list_empty(&chunk->trees)) {
1016 owner = list_entry(chunk->trees.next,
1017 struct audit_tree, same_root);
1018 owner->goner = 1;
1019 owner->root = NULL;
1020 list_del_init(&owner->same_root);
1021 spin_unlock(&hash_lock);
1022 if (!postponed) {
1023 kill_rules(audit_context(), owner);
1024 list_move(&owner->list, &prune_list);
1025 need_prune = 1;
1026 } else {
1027 list_move(&owner->list, postponed);
1028 }
1029 spin_lock(&hash_lock);
1030 }
1031 list_del_rcu(&chunk->hash);
1032 for (n = 0; n < chunk->count; n++)
1033 list_del_init(&chunk->owners[n].list);
1034 spin_unlock(&hash_lock);
1035 mutex_unlock(&audit_filter_mutex);
1036 if (need_prune)
1037 audit_schedule_prune();
1038}
1039
1040static int audit_tree_handle_event(struct fsnotify_group *group,
1041 struct inode *to_tell,
1042 u32 mask, const void *data, int data_type,
1043 const struct qstr *file_name, u32 cookie,
1044 struct fsnotify_iter_info *iter_info)
1045{
1046 return 0;
1047}
1048
1049static void audit_tree_freeing_mark(struct fsnotify_mark *mark,
1050 struct fsnotify_group *group)
1051{
1052 struct audit_chunk *chunk;
1053
1054 mutex_lock(&mark->group->mark_mutex);
1055 spin_lock(&hash_lock);
1056 chunk = mark_chunk(mark);
1057 replace_mark_chunk(mark, NULL);
1058 spin_unlock(&hash_lock);
1059 mutex_unlock(&mark->group->mark_mutex);
1060 if (chunk) {
1061 evict_chunk(chunk);
1062 audit_mark_put_chunk(chunk);
1063 }
1064
1065 /*
1066 * We are guaranteed to have at least one reference to the mark from
1067 * either the inode or the caller of fsnotify_destroy_mark().
1068 */
1069 BUG_ON(refcount_read(&mark->refcnt) < 1);
1070}
1071
1072static const struct fsnotify_ops audit_tree_ops = {
1073 .handle_event = audit_tree_handle_event,
1074 .freeing_mark = audit_tree_freeing_mark,
1075 .free_mark = audit_tree_destroy_watch,
1076};
1077
1078static int __init audit_tree_init(void)
1079{
1080 int i;
1081
1082 audit_tree_mark_cachep = KMEM_CACHE(audit_tree_mark, SLAB_PANIC);
1083
1084 audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1085 if (IS_ERR(audit_tree_group))
1086 audit_panic("cannot initialize fsnotify group for rectree watches");
1087
1088 for (i = 0; i < HASH_SIZE; i++)
1089 INIT_LIST_HEAD(&chunk_hash_heads[i]);
1090
1091 return 0;
1092}
1093__initcall(audit_tree_init);
1#include "audit.h"
2#include <linux/fsnotify_backend.h>
3#include <linux/namei.h>
4#include <linux/mount.h>
5#include <linux/kthread.h>
6#include <linux/slab.h>
7
8struct audit_tree;
9struct audit_chunk;
10
11struct audit_tree {
12 atomic_t count;
13 int goner;
14 struct audit_chunk *root;
15 struct list_head chunks;
16 struct list_head rules;
17 struct list_head list;
18 struct list_head same_root;
19 struct rcu_head head;
20 char pathname[];
21};
22
23struct audit_chunk {
24 struct list_head hash;
25 struct fsnotify_mark mark;
26 struct list_head trees; /* with root here */
27 int dead;
28 int count;
29 atomic_long_t refs;
30 struct rcu_head head;
31 struct node {
32 struct list_head list;
33 struct audit_tree *owner;
34 unsigned index; /* index; upper bit indicates 'will prune' */
35 } owners[];
36};
37
38static LIST_HEAD(tree_list);
39static LIST_HEAD(prune_list);
40
41/*
42 * One struct chunk is attached to each inode of interest.
43 * We replace struct chunk on tagging/untagging.
44 * Rules have pointer to struct audit_tree.
45 * Rules have struct list_head rlist forming a list of rules over
46 * the same tree.
47 * References to struct chunk are collected at audit_inode{,_child}()
48 * time and used in AUDIT_TREE rule matching.
49 * These references are dropped at the same time we are calling
50 * audit_free_names(), etc.
51 *
52 * Cyclic lists galore:
53 * tree.chunks anchors chunk.owners[].list hash_lock
54 * tree.rules anchors rule.rlist audit_filter_mutex
55 * chunk.trees anchors tree.same_root hash_lock
56 * chunk.hash is a hash with middle bits of watch.inode as
57 * a hash function. RCU, hash_lock
58 *
59 * tree is refcounted; one reference for "some rules on rules_list refer to
60 * it", one for each chunk with pointer to it.
61 *
62 * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
63 * of watch contributes 1 to .refs).
64 *
65 * node.index allows to get from node.list to containing chunk.
66 * MSB of that sucker is stolen to mark taggings that we might have to
67 * revert - several operations have very unpleasant cleanup logics and
68 * that makes a difference. Some.
69 */
70
71static struct fsnotify_group *audit_tree_group;
72
73static struct audit_tree *alloc_tree(const char *s)
74{
75 struct audit_tree *tree;
76
77 tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
78 if (tree) {
79 atomic_set(&tree->count, 1);
80 tree->goner = 0;
81 INIT_LIST_HEAD(&tree->chunks);
82 INIT_LIST_HEAD(&tree->rules);
83 INIT_LIST_HEAD(&tree->list);
84 INIT_LIST_HEAD(&tree->same_root);
85 tree->root = NULL;
86 strcpy(tree->pathname, s);
87 }
88 return tree;
89}
90
91static inline void get_tree(struct audit_tree *tree)
92{
93 atomic_inc(&tree->count);
94}
95
96static inline void put_tree(struct audit_tree *tree)
97{
98 if (atomic_dec_and_test(&tree->count))
99 kfree_rcu(tree, head);
100}
101
102/* to avoid bringing the entire thing in audit.h */
103const char *audit_tree_path(struct audit_tree *tree)
104{
105 return tree->pathname;
106}
107
108static void free_chunk(struct audit_chunk *chunk)
109{
110 int i;
111
112 for (i = 0; i < chunk->count; i++) {
113 if (chunk->owners[i].owner)
114 put_tree(chunk->owners[i].owner);
115 }
116 kfree(chunk);
117}
118
119void audit_put_chunk(struct audit_chunk *chunk)
120{
121 if (atomic_long_dec_and_test(&chunk->refs))
122 free_chunk(chunk);
123}
124
125static void __put_chunk(struct rcu_head *rcu)
126{
127 struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
128 audit_put_chunk(chunk);
129}
130
131static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
132{
133 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
134 call_rcu(&chunk->head, __put_chunk);
135}
136
137static struct audit_chunk *alloc_chunk(int count)
138{
139 struct audit_chunk *chunk;
140 size_t size;
141 int i;
142
143 size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
144 chunk = kzalloc(size, GFP_KERNEL);
145 if (!chunk)
146 return NULL;
147
148 INIT_LIST_HEAD(&chunk->hash);
149 INIT_LIST_HEAD(&chunk->trees);
150 chunk->count = count;
151 atomic_long_set(&chunk->refs, 1);
152 for (i = 0; i < count; i++) {
153 INIT_LIST_HEAD(&chunk->owners[i].list);
154 chunk->owners[i].index = i;
155 }
156 fsnotify_init_mark(&chunk->mark, audit_tree_destroy_watch);
157 return chunk;
158}
159
160enum {HASH_SIZE = 128};
161static struct list_head chunk_hash_heads[HASH_SIZE];
162static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
163
164static inline struct list_head *chunk_hash(const struct inode *inode)
165{
166 unsigned long n = (unsigned long)inode / L1_CACHE_BYTES;
167 return chunk_hash_heads + n % HASH_SIZE;
168}
169
170/* hash_lock & entry->lock is held by caller */
171static void insert_hash(struct audit_chunk *chunk)
172{
173 struct fsnotify_mark *entry = &chunk->mark;
174 struct list_head *list;
175
176 if (!entry->i.inode)
177 return;
178 list = chunk_hash(entry->i.inode);
179 list_add_rcu(&chunk->hash, list);
180}
181
182/* called under rcu_read_lock */
183struct audit_chunk *audit_tree_lookup(const struct inode *inode)
184{
185 struct list_head *list = chunk_hash(inode);
186 struct audit_chunk *p;
187
188 list_for_each_entry_rcu(p, list, hash) {
189 /* mark.inode may have gone NULL, but who cares? */
190 if (p->mark.i.inode == inode) {
191 atomic_long_inc(&p->refs);
192 return p;
193 }
194 }
195 return NULL;
196}
197
198int audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
199{
200 int n;
201 for (n = 0; n < chunk->count; n++)
202 if (chunk->owners[n].owner == tree)
203 return 1;
204 return 0;
205}
206
207/* tagging and untagging inodes with trees */
208
209static struct audit_chunk *find_chunk(struct node *p)
210{
211 int index = p->index & ~(1U<<31);
212 p -= index;
213 return container_of(p, struct audit_chunk, owners[0]);
214}
215
216static void untag_chunk(struct node *p)
217{
218 struct audit_chunk *chunk = find_chunk(p);
219 struct fsnotify_mark *entry = &chunk->mark;
220 struct audit_chunk *new = NULL;
221 struct audit_tree *owner;
222 int size = chunk->count - 1;
223 int i, j;
224
225 fsnotify_get_mark(entry);
226
227 spin_unlock(&hash_lock);
228
229 if (size)
230 new = alloc_chunk(size);
231
232 spin_lock(&entry->lock);
233 if (chunk->dead || !entry->i.inode) {
234 spin_unlock(&entry->lock);
235 if (new)
236 free_chunk(new);
237 goto out;
238 }
239
240 owner = p->owner;
241
242 if (!size) {
243 chunk->dead = 1;
244 spin_lock(&hash_lock);
245 list_del_init(&chunk->trees);
246 if (owner->root == chunk)
247 owner->root = NULL;
248 list_del_init(&p->list);
249 list_del_rcu(&chunk->hash);
250 spin_unlock(&hash_lock);
251 spin_unlock(&entry->lock);
252 fsnotify_destroy_mark(entry, audit_tree_group);
253 goto out;
254 }
255
256 if (!new)
257 goto Fallback;
258
259 fsnotify_duplicate_mark(&new->mark, entry);
260 if (fsnotify_add_mark(&new->mark, new->mark.group, new->mark.i.inode, NULL, 1)) {
261 fsnotify_put_mark(&new->mark);
262 goto Fallback;
263 }
264
265 chunk->dead = 1;
266 spin_lock(&hash_lock);
267 list_replace_init(&chunk->trees, &new->trees);
268 if (owner->root == chunk) {
269 list_del_init(&owner->same_root);
270 owner->root = NULL;
271 }
272
273 for (i = j = 0; j <= size; i++, j++) {
274 struct audit_tree *s;
275 if (&chunk->owners[j] == p) {
276 list_del_init(&p->list);
277 i--;
278 continue;
279 }
280 s = chunk->owners[j].owner;
281 new->owners[i].owner = s;
282 new->owners[i].index = chunk->owners[j].index - j + i;
283 if (!s) /* result of earlier fallback */
284 continue;
285 get_tree(s);
286 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
287 }
288
289 list_replace_rcu(&chunk->hash, &new->hash);
290 list_for_each_entry(owner, &new->trees, same_root)
291 owner->root = new;
292 spin_unlock(&hash_lock);
293 spin_unlock(&entry->lock);
294 fsnotify_destroy_mark(entry, audit_tree_group);
295 fsnotify_put_mark(&new->mark); /* drop initial reference */
296 goto out;
297
298Fallback:
299 // do the best we can
300 spin_lock(&hash_lock);
301 if (owner->root == chunk) {
302 list_del_init(&owner->same_root);
303 owner->root = NULL;
304 }
305 list_del_init(&p->list);
306 p->owner = NULL;
307 put_tree(owner);
308 spin_unlock(&hash_lock);
309 spin_unlock(&entry->lock);
310out:
311 fsnotify_put_mark(entry);
312 spin_lock(&hash_lock);
313}
314
315static int create_chunk(struct inode *inode, struct audit_tree *tree)
316{
317 struct fsnotify_mark *entry;
318 struct audit_chunk *chunk = alloc_chunk(1);
319 if (!chunk)
320 return -ENOMEM;
321
322 entry = &chunk->mark;
323 if (fsnotify_add_mark(entry, audit_tree_group, inode, NULL, 0)) {
324 fsnotify_put_mark(entry);
325 return -ENOSPC;
326 }
327
328 spin_lock(&entry->lock);
329 spin_lock(&hash_lock);
330 if (tree->goner) {
331 spin_unlock(&hash_lock);
332 chunk->dead = 1;
333 spin_unlock(&entry->lock);
334 fsnotify_destroy_mark(entry, audit_tree_group);
335 fsnotify_put_mark(entry);
336 return 0;
337 }
338 chunk->owners[0].index = (1U << 31);
339 chunk->owners[0].owner = tree;
340 get_tree(tree);
341 list_add(&chunk->owners[0].list, &tree->chunks);
342 if (!tree->root) {
343 tree->root = chunk;
344 list_add(&tree->same_root, &chunk->trees);
345 }
346 insert_hash(chunk);
347 spin_unlock(&hash_lock);
348 spin_unlock(&entry->lock);
349 fsnotify_put_mark(entry); /* drop initial reference */
350 return 0;
351}
352
353/* the first tagged inode becomes root of tree */
354static int tag_chunk(struct inode *inode, struct audit_tree *tree)
355{
356 struct fsnotify_mark *old_entry, *chunk_entry;
357 struct audit_tree *owner;
358 struct audit_chunk *chunk, *old;
359 struct node *p;
360 int n;
361
362 old_entry = fsnotify_find_inode_mark(audit_tree_group, inode);
363 if (!old_entry)
364 return create_chunk(inode, tree);
365
366 old = container_of(old_entry, struct audit_chunk, mark);
367
368 /* are we already there? */
369 spin_lock(&hash_lock);
370 for (n = 0; n < old->count; n++) {
371 if (old->owners[n].owner == tree) {
372 spin_unlock(&hash_lock);
373 fsnotify_put_mark(old_entry);
374 return 0;
375 }
376 }
377 spin_unlock(&hash_lock);
378
379 chunk = alloc_chunk(old->count + 1);
380 if (!chunk) {
381 fsnotify_put_mark(old_entry);
382 return -ENOMEM;
383 }
384
385 chunk_entry = &chunk->mark;
386
387 spin_lock(&old_entry->lock);
388 if (!old_entry->i.inode) {
389 /* old_entry is being shot, lets just lie */
390 spin_unlock(&old_entry->lock);
391 fsnotify_put_mark(old_entry);
392 free_chunk(chunk);
393 return -ENOENT;
394 }
395
396 fsnotify_duplicate_mark(chunk_entry, old_entry);
397 if (fsnotify_add_mark(chunk_entry, chunk_entry->group, chunk_entry->i.inode, NULL, 1)) {
398 spin_unlock(&old_entry->lock);
399 fsnotify_put_mark(chunk_entry);
400 fsnotify_put_mark(old_entry);
401 return -ENOSPC;
402 }
403
404 /* even though we hold old_entry->lock, this is safe since chunk_entry->lock could NEVER have been grabbed before */
405 spin_lock(&chunk_entry->lock);
406 spin_lock(&hash_lock);
407
408 /* we now hold old_entry->lock, chunk_entry->lock, and hash_lock */
409 if (tree->goner) {
410 spin_unlock(&hash_lock);
411 chunk->dead = 1;
412 spin_unlock(&chunk_entry->lock);
413 spin_unlock(&old_entry->lock);
414
415 fsnotify_destroy_mark(chunk_entry, audit_tree_group);
416
417 fsnotify_put_mark(chunk_entry);
418 fsnotify_put_mark(old_entry);
419 return 0;
420 }
421 list_replace_init(&old->trees, &chunk->trees);
422 for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
423 struct audit_tree *s = old->owners[n].owner;
424 p->owner = s;
425 p->index = old->owners[n].index;
426 if (!s) /* result of fallback in untag */
427 continue;
428 get_tree(s);
429 list_replace_init(&old->owners[n].list, &p->list);
430 }
431 p->index = (chunk->count - 1) | (1U<<31);
432 p->owner = tree;
433 get_tree(tree);
434 list_add(&p->list, &tree->chunks);
435 list_replace_rcu(&old->hash, &chunk->hash);
436 list_for_each_entry(owner, &chunk->trees, same_root)
437 owner->root = chunk;
438 old->dead = 1;
439 if (!tree->root) {
440 tree->root = chunk;
441 list_add(&tree->same_root, &chunk->trees);
442 }
443 spin_unlock(&hash_lock);
444 spin_unlock(&chunk_entry->lock);
445 spin_unlock(&old_entry->lock);
446 fsnotify_destroy_mark(old_entry, audit_tree_group);
447 fsnotify_put_mark(chunk_entry); /* drop initial reference */
448 fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
449 return 0;
450}
451
452static void audit_log_remove_rule(struct audit_krule *rule)
453{
454 struct audit_buffer *ab;
455
456 ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
457 if (unlikely(!ab))
458 return;
459 audit_log_format(ab, "op=");
460 audit_log_string(ab, "remove rule");
461 audit_log_format(ab, " dir=");
462 audit_log_untrustedstring(ab, rule->tree->pathname);
463 audit_log_key(ab, rule->filterkey);
464 audit_log_format(ab, " list=%d res=1", rule->listnr);
465 audit_log_end(ab);
466}
467
468static void kill_rules(struct audit_tree *tree)
469{
470 struct audit_krule *rule, *next;
471 struct audit_entry *entry;
472
473 list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
474 entry = container_of(rule, struct audit_entry, rule);
475
476 list_del_init(&rule->rlist);
477 if (rule->tree) {
478 /* not a half-baked one */
479 audit_log_remove_rule(rule);
480 rule->tree = NULL;
481 list_del_rcu(&entry->list);
482 list_del(&entry->rule.list);
483 call_rcu(&entry->rcu, audit_free_rule_rcu);
484 }
485 }
486}
487
488/*
489 * finish killing struct audit_tree
490 */
491static void prune_one(struct audit_tree *victim)
492{
493 spin_lock(&hash_lock);
494 while (!list_empty(&victim->chunks)) {
495 struct node *p;
496
497 p = list_entry(victim->chunks.next, struct node, list);
498
499 untag_chunk(p);
500 }
501 spin_unlock(&hash_lock);
502 put_tree(victim);
503}
504
505/* trim the uncommitted chunks from tree */
506
507static void trim_marked(struct audit_tree *tree)
508{
509 struct list_head *p, *q;
510 spin_lock(&hash_lock);
511 if (tree->goner) {
512 spin_unlock(&hash_lock);
513 return;
514 }
515 /* reorder */
516 for (p = tree->chunks.next; p != &tree->chunks; p = q) {
517 struct node *node = list_entry(p, struct node, list);
518 q = p->next;
519 if (node->index & (1U<<31)) {
520 list_del_init(p);
521 list_add(p, &tree->chunks);
522 }
523 }
524
525 while (!list_empty(&tree->chunks)) {
526 struct node *node;
527
528 node = list_entry(tree->chunks.next, struct node, list);
529
530 /* have we run out of marked? */
531 if (!(node->index & (1U<<31)))
532 break;
533
534 untag_chunk(node);
535 }
536 if (!tree->root && !tree->goner) {
537 tree->goner = 1;
538 spin_unlock(&hash_lock);
539 mutex_lock(&audit_filter_mutex);
540 kill_rules(tree);
541 list_del_init(&tree->list);
542 mutex_unlock(&audit_filter_mutex);
543 prune_one(tree);
544 } else {
545 spin_unlock(&hash_lock);
546 }
547}
548
549static void audit_schedule_prune(void);
550
551/* called with audit_filter_mutex */
552int audit_remove_tree_rule(struct audit_krule *rule)
553{
554 struct audit_tree *tree;
555 tree = rule->tree;
556 if (tree) {
557 spin_lock(&hash_lock);
558 list_del_init(&rule->rlist);
559 if (list_empty(&tree->rules) && !tree->goner) {
560 tree->root = NULL;
561 list_del_init(&tree->same_root);
562 tree->goner = 1;
563 list_move(&tree->list, &prune_list);
564 rule->tree = NULL;
565 spin_unlock(&hash_lock);
566 audit_schedule_prune();
567 return 1;
568 }
569 rule->tree = NULL;
570 spin_unlock(&hash_lock);
571 return 1;
572 }
573 return 0;
574}
575
576static int compare_root(struct vfsmount *mnt, void *arg)
577{
578 return mnt->mnt_root->d_inode == arg;
579}
580
581void audit_trim_trees(void)
582{
583 struct list_head cursor;
584
585 mutex_lock(&audit_filter_mutex);
586 list_add(&cursor, &tree_list);
587 while (cursor.next != &tree_list) {
588 struct audit_tree *tree;
589 struct path path;
590 struct vfsmount *root_mnt;
591 struct node *node;
592 int err;
593
594 tree = container_of(cursor.next, struct audit_tree, list);
595 get_tree(tree);
596 list_del(&cursor);
597 list_add(&cursor, &tree->list);
598 mutex_unlock(&audit_filter_mutex);
599
600 err = kern_path(tree->pathname, 0, &path);
601 if (err)
602 goto skip_it;
603
604 root_mnt = collect_mounts(&path);
605 path_put(&path);
606 if (IS_ERR(root_mnt))
607 goto skip_it;
608
609 spin_lock(&hash_lock);
610 list_for_each_entry(node, &tree->chunks, list) {
611 struct audit_chunk *chunk = find_chunk(node);
612 /* this could be NULL if the watch is dying else where... */
613 struct inode *inode = chunk->mark.i.inode;
614 node->index |= 1U<<31;
615 if (iterate_mounts(compare_root, inode, root_mnt))
616 node->index &= ~(1U<<31);
617 }
618 spin_unlock(&hash_lock);
619 trim_marked(tree);
620 drop_collected_mounts(root_mnt);
621skip_it:
622 put_tree(tree);
623 mutex_lock(&audit_filter_mutex);
624 }
625 list_del(&cursor);
626 mutex_unlock(&audit_filter_mutex);
627}
628
629int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
630{
631
632 if (pathname[0] != '/' ||
633 rule->listnr != AUDIT_FILTER_EXIT ||
634 op != Audit_equal ||
635 rule->inode_f || rule->watch || rule->tree)
636 return -EINVAL;
637 rule->tree = alloc_tree(pathname);
638 if (!rule->tree)
639 return -ENOMEM;
640 return 0;
641}
642
643void audit_put_tree(struct audit_tree *tree)
644{
645 put_tree(tree);
646}
647
648static int tag_mount(struct vfsmount *mnt, void *arg)
649{
650 return tag_chunk(mnt->mnt_root->d_inode, arg);
651}
652
653/* called with audit_filter_mutex */
654int audit_add_tree_rule(struct audit_krule *rule)
655{
656 struct audit_tree *seed = rule->tree, *tree;
657 struct path path;
658 struct vfsmount *mnt;
659 int err;
660
661 rule->tree = NULL;
662 list_for_each_entry(tree, &tree_list, list) {
663 if (!strcmp(seed->pathname, tree->pathname)) {
664 put_tree(seed);
665 rule->tree = tree;
666 list_add(&rule->rlist, &tree->rules);
667 return 0;
668 }
669 }
670 tree = seed;
671 list_add(&tree->list, &tree_list);
672 list_add(&rule->rlist, &tree->rules);
673 /* do not set rule->tree yet */
674 mutex_unlock(&audit_filter_mutex);
675
676 err = kern_path(tree->pathname, 0, &path);
677 if (err)
678 goto Err;
679 mnt = collect_mounts(&path);
680 path_put(&path);
681 if (IS_ERR(mnt)) {
682 err = PTR_ERR(mnt);
683 goto Err;
684 }
685
686 get_tree(tree);
687 err = iterate_mounts(tag_mount, tree, mnt);
688 drop_collected_mounts(mnt);
689
690 if (!err) {
691 struct node *node;
692 spin_lock(&hash_lock);
693 list_for_each_entry(node, &tree->chunks, list)
694 node->index &= ~(1U<<31);
695 spin_unlock(&hash_lock);
696 } else {
697 trim_marked(tree);
698 goto Err;
699 }
700
701 mutex_lock(&audit_filter_mutex);
702 if (list_empty(&rule->rlist)) {
703 put_tree(tree);
704 return -ENOENT;
705 }
706 rule->tree = tree;
707 put_tree(tree);
708
709 return 0;
710Err:
711 mutex_lock(&audit_filter_mutex);
712 list_del_init(&tree->list);
713 list_del_init(&tree->rules);
714 put_tree(tree);
715 return err;
716}
717
718int audit_tag_tree(char *old, char *new)
719{
720 struct list_head cursor, barrier;
721 int failed = 0;
722 struct path path1, path2;
723 struct vfsmount *tagged;
724 int err;
725
726 err = kern_path(new, 0, &path2);
727 if (err)
728 return err;
729 tagged = collect_mounts(&path2);
730 path_put(&path2);
731 if (IS_ERR(tagged))
732 return PTR_ERR(tagged);
733
734 err = kern_path(old, 0, &path1);
735 if (err) {
736 drop_collected_mounts(tagged);
737 return err;
738 }
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 int good_one = 0;
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, &path2);
755 if (!err) {
756 good_one = path_is_under(&path1, &path2);
757 path_put(&path2);
758 }
759
760 if (!good_one) {
761 put_tree(tree);
762 mutex_lock(&audit_filter_mutex);
763 continue;
764 }
765
766 failed = iterate_mounts(tag_mount, tree, tagged);
767 if (failed) {
768 put_tree(tree);
769 mutex_lock(&audit_filter_mutex);
770 break;
771 }
772
773 mutex_lock(&audit_filter_mutex);
774 spin_lock(&hash_lock);
775 if (!tree->goner) {
776 list_del(&tree->list);
777 list_add(&tree->list, &tree_list);
778 }
779 spin_unlock(&hash_lock);
780 put_tree(tree);
781 }
782
783 while (barrier.prev != &tree_list) {
784 struct audit_tree *tree;
785
786 tree = container_of(barrier.prev, struct audit_tree, list);
787 get_tree(tree);
788 list_del(&tree->list);
789 list_add(&tree->list, &barrier);
790 mutex_unlock(&audit_filter_mutex);
791
792 if (!failed) {
793 struct node *node;
794 spin_lock(&hash_lock);
795 list_for_each_entry(node, &tree->chunks, list)
796 node->index &= ~(1U<<31);
797 spin_unlock(&hash_lock);
798 } else {
799 trim_marked(tree);
800 }
801
802 put_tree(tree);
803 mutex_lock(&audit_filter_mutex);
804 }
805 list_del(&barrier);
806 list_del(&cursor);
807 mutex_unlock(&audit_filter_mutex);
808 path_put(&path1);
809 drop_collected_mounts(tagged);
810 return failed;
811}
812
813/*
814 * That gets run when evict_chunk() ends up needing to kill audit_tree.
815 * Runs from a separate thread.
816 */
817static int prune_tree_thread(void *unused)
818{
819 mutex_lock(&audit_cmd_mutex);
820 mutex_lock(&audit_filter_mutex);
821
822 while (!list_empty(&prune_list)) {
823 struct audit_tree *victim;
824
825 victim = list_entry(prune_list.next, struct audit_tree, list);
826 list_del_init(&victim->list);
827
828 mutex_unlock(&audit_filter_mutex);
829
830 prune_one(victim);
831
832 mutex_lock(&audit_filter_mutex);
833 }
834
835 mutex_unlock(&audit_filter_mutex);
836 mutex_unlock(&audit_cmd_mutex);
837 return 0;
838}
839
840static void audit_schedule_prune(void)
841{
842 kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
843}
844
845/*
846 * ... and that one is done if evict_chunk() decides to delay until the end
847 * of syscall. Runs synchronously.
848 */
849void audit_kill_trees(struct list_head *list)
850{
851 mutex_lock(&audit_cmd_mutex);
852 mutex_lock(&audit_filter_mutex);
853
854 while (!list_empty(list)) {
855 struct audit_tree *victim;
856
857 victim = list_entry(list->next, struct audit_tree, list);
858 kill_rules(victim);
859 list_del_init(&victim->list);
860
861 mutex_unlock(&audit_filter_mutex);
862
863 prune_one(victim);
864
865 mutex_lock(&audit_filter_mutex);
866 }
867
868 mutex_unlock(&audit_filter_mutex);
869 mutex_unlock(&audit_cmd_mutex);
870}
871
872/*
873 * Here comes the stuff asynchronous to auditctl operations
874 */
875
876static void evict_chunk(struct audit_chunk *chunk)
877{
878 struct audit_tree *owner;
879 struct list_head *postponed = audit_killed_trees();
880 int need_prune = 0;
881 int n;
882
883 if (chunk->dead)
884 return;
885
886 chunk->dead = 1;
887 mutex_lock(&audit_filter_mutex);
888 spin_lock(&hash_lock);
889 while (!list_empty(&chunk->trees)) {
890 owner = list_entry(chunk->trees.next,
891 struct audit_tree, same_root);
892 owner->goner = 1;
893 owner->root = NULL;
894 list_del_init(&owner->same_root);
895 spin_unlock(&hash_lock);
896 if (!postponed) {
897 kill_rules(owner);
898 list_move(&owner->list, &prune_list);
899 need_prune = 1;
900 } else {
901 list_move(&owner->list, postponed);
902 }
903 spin_lock(&hash_lock);
904 }
905 list_del_rcu(&chunk->hash);
906 for (n = 0; n < chunk->count; n++)
907 list_del_init(&chunk->owners[n].list);
908 spin_unlock(&hash_lock);
909 if (need_prune)
910 audit_schedule_prune();
911 mutex_unlock(&audit_filter_mutex);
912}
913
914static int audit_tree_handle_event(struct fsnotify_group *group,
915 struct inode *to_tell,
916 struct fsnotify_mark *inode_mark,
917 struct fsnotify_mark *vfsmount_mark,
918 u32 mask, void *data, int data_type,
919 const unsigned char *file_name, u32 cookie)
920{
921 return 0;
922}
923
924static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
925{
926 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
927
928 evict_chunk(chunk);
929
930 /*
931 * We are guaranteed to have at least one reference to the mark from
932 * either the inode or the caller of fsnotify_destroy_mark().
933 */
934 BUG_ON(atomic_read(&entry->refcnt) < 1);
935}
936
937static const struct fsnotify_ops audit_tree_ops = {
938 .handle_event = audit_tree_handle_event,
939 .freeing_mark = audit_tree_freeing_mark,
940};
941
942static int __init audit_tree_init(void)
943{
944 int i;
945
946 audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
947 if (IS_ERR(audit_tree_group))
948 audit_panic("cannot initialize fsnotify group for rectree watches");
949
950 for (i = 0; i < HASH_SIZE; i++)
951 INIT_LIST_HEAD(&chunk_hash_heads[i]);
952
953 return 0;
954}
955__initcall(audit_tree_init);