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>
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;
24 struct list_head hash;
25 struct fsnotify_mark mark;
26 struct list_head trees; /* with root here */
32 struct list_head list;
33 struct audit_tree *owner;
34 unsigned index; /* index; upper bit indicates 'will prune' */
38 static LIST_HEAD(tree_list);
39 static LIST_HEAD(prune_list);
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
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.
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
59 * tree is refcounted; one reference for "some rules on rules_list refer to
60 * it", one for each chunk with pointer to it.
62 * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
63 * of watch contributes 1 to .refs).
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.
71 static struct fsnotify_group *audit_tree_group;
73 static struct audit_tree *alloc_tree(const char *s)
75 struct audit_tree *tree;
77 tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
79 atomic_set(&tree->count, 1);
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);
86 strcpy(tree->pathname, s);
91 static inline void get_tree(struct audit_tree *tree)
93 atomic_inc(&tree->count);
96 static void __put_tree(struct rcu_head *rcu)
98 struct audit_tree *tree = container_of(rcu, struct audit_tree, head);
102 static inline void put_tree(struct audit_tree *tree)
104 if (atomic_dec_and_test(&tree->count))
105 call_rcu(&tree->head, __put_tree);
108 /* to avoid bringing the entire thing in audit.h */
109 const char *audit_tree_path(struct audit_tree *tree)
111 return tree->pathname;
114 static void free_chunk(struct audit_chunk *chunk)
118 for (i = 0; i < chunk->count; i++) {
119 if (chunk->owners[i].owner)
120 put_tree(chunk->owners[i].owner);
125 void audit_put_chunk(struct audit_chunk *chunk)
127 if (atomic_long_dec_and_test(&chunk->refs))
131 static void __put_chunk(struct rcu_head *rcu)
133 struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
134 audit_put_chunk(chunk);
137 static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
139 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
140 call_rcu(&chunk->head, __put_chunk);
143 static struct audit_chunk *alloc_chunk(int count)
145 struct audit_chunk *chunk;
149 size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
150 chunk = kzalloc(size, GFP_KERNEL);
154 INIT_LIST_HEAD(&chunk->hash);
155 INIT_LIST_HEAD(&chunk->trees);
156 chunk->count = count;
157 atomic_long_set(&chunk->refs, 1);
158 for (i = 0; i < count; i++) {
159 INIT_LIST_HEAD(&chunk->owners[i].list);
160 chunk->owners[i].index = i;
162 fsnotify_init_mark(&chunk->mark, audit_tree_destroy_watch);
166 enum {HASH_SIZE = 128};
167 static struct list_head chunk_hash_heads[HASH_SIZE];
168 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
170 static inline struct list_head *chunk_hash(const struct inode *inode)
172 unsigned long n = (unsigned long)inode / L1_CACHE_BYTES;
173 return chunk_hash_heads + n % HASH_SIZE;
176 /* hash_lock & entry->lock is held by caller */
177 static void insert_hash(struct audit_chunk *chunk)
179 struct fsnotify_mark *entry = &chunk->mark;
180 struct list_head *list;
184 list = chunk_hash(entry->i.inode);
185 list_add_rcu(&chunk->hash, list);
188 /* called under rcu_read_lock */
189 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
191 struct list_head *list = chunk_hash(inode);
192 struct audit_chunk *p;
194 list_for_each_entry_rcu(p, list, hash) {
195 /* mark.inode may have gone NULL, but who cares? */
196 if (p->mark.i.inode == inode) {
197 atomic_long_inc(&p->refs);
204 int audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
207 for (n = 0; n < chunk->count; n++)
208 if (chunk->owners[n].owner == tree)
213 /* tagging and untagging inodes with trees */
215 static struct audit_chunk *find_chunk(struct node *p)
217 int index = p->index & ~(1U<<31);
219 return container_of(p, struct audit_chunk, owners[0]);
222 static void untag_chunk(struct node *p)
224 struct audit_chunk *chunk = find_chunk(p);
225 struct fsnotify_mark *entry = &chunk->mark;
226 struct audit_chunk *new = NULL;
227 struct audit_tree *owner;
228 int size = chunk->count - 1;
231 fsnotify_get_mark(entry);
233 spin_unlock(&hash_lock);
236 new = alloc_chunk(size);
238 spin_lock(&entry->lock);
239 if (chunk->dead || !entry->i.inode) {
240 spin_unlock(&entry->lock);
250 spin_lock(&hash_lock);
251 list_del_init(&chunk->trees);
252 if (owner->root == chunk)
254 list_del_init(&p->list);
255 list_del_rcu(&chunk->hash);
256 spin_unlock(&hash_lock);
257 spin_unlock(&entry->lock);
258 fsnotify_destroy_mark(entry);
265 fsnotify_duplicate_mark(&new->mark, entry);
266 if (fsnotify_add_mark(&new->mark, new->mark.group, new->mark.i.inode, NULL, 1)) {
267 fsnotify_put_mark(&new->mark);
272 spin_lock(&hash_lock);
273 list_replace_init(&chunk->trees, &new->trees);
274 if (owner->root == chunk) {
275 list_del_init(&owner->same_root);
279 for (i = j = 0; j <= size; i++, j++) {
280 struct audit_tree *s;
281 if (&chunk->owners[j] == p) {
282 list_del_init(&p->list);
286 s = chunk->owners[j].owner;
287 new->owners[i].owner = s;
288 new->owners[i].index = chunk->owners[j].index - j + i;
289 if (!s) /* result of earlier fallback */
292 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
295 list_replace_rcu(&chunk->hash, &new->hash);
296 list_for_each_entry(owner, &new->trees, same_root)
298 spin_unlock(&hash_lock);
299 spin_unlock(&entry->lock);
300 fsnotify_destroy_mark(entry);
304 // do the best we can
305 spin_lock(&hash_lock);
306 if (owner->root == chunk) {
307 list_del_init(&owner->same_root);
310 list_del_init(&p->list);
313 spin_unlock(&hash_lock);
314 spin_unlock(&entry->lock);
316 fsnotify_put_mark(entry);
317 spin_lock(&hash_lock);
320 static int create_chunk(struct inode *inode, struct audit_tree *tree)
322 struct fsnotify_mark *entry;
323 struct audit_chunk *chunk = alloc_chunk(1);
327 entry = &chunk->mark;
328 if (fsnotify_add_mark(entry, audit_tree_group, inode, NULL, 0)) {
329 fsnotify_put_mark(entry);
333 spin_lock(&entry->lock);
334 spin_lock(&hash_lock);
336 spin_unlock(&hash_lock);
338 spin_unlock(&entry->lock);
339 fsnotify_get_mark(entry);
340 fsnotify_destroy_mark(entry);
341 fsnotify_put_mark(entry);
344 chunk->owners[0].index = (1U << 31);
345 chunk->owners[0].owner = tree;
347 list_add(&chunk->owners[0].list, &tree->chunks);
350 list_add(&tree->same_root, &chunk->trees);
353 spin_unlock(&hash_lock);
354 spin_unlock(&entry->lock);
358 /* the first tagged inode becomes root of tree */
359 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
361 struct fsnotify_mark *old_entry, *chunk_entry;
362 struct audit_tree *owner;
363 struct audit_chunk *chunk, *old;
367 old_entry = fsnotify_find_inode_mark(audit_tree_group, inode);
369 return create_chunk(inode, tree);
371 old = container_of(old_entry, struct audit_chunk, mark);
373 /* are we already there? */
374 spin_lock(&hash_lock);
375 for (n = 0; n < old->count; n++) {
376 if (old->owners[n].owner == tree) {
377 spin_unlock(&hash_lock);
378 fsnotify_put_mark(old_entry);
382 spin_unlock(&hash_lock);
384 chunk = alloc_chunk(old->count + 1);
386 fsnotify_put_mark(old_entry);
390 chunk_entry = &chunk->mark;
392 spin_lock(&old_entry->lock);
393 if (!old_entry->i.inode) {
394 /* old_entry is being shot, lets just lie */
395 spin_unlock(&old_entry->lock);
396 fsnotify_put_mark(old_entry);
401 fsnotify_duplicate_mark(chunk_entry, old_entry);
402 if (fsnotify_add_mark(chunk_entry, chunk_entry->group, chunk_entry->i.inode, NULL, 1)) {
403 spin_unlock(&old_entry->lock);
404 fsnotify_put_mark(chunk_entry);
405 fsnotify_put_mark(old_entry);
409 /* even though we hold old_entry->lock, this is safe since chunk_entry->lock could NEVER have been grabbed before */
410 spin_lock(&chunk_entry->lock);
411 spin_lock(&hash_lock);
413 /* we now hold old_entry->lock, chunk_entry->lock, and hash_lock */
415 spin_unlock(&hash_lock);
417 spin_unlock(&chunk_entry->lock);
418 spin_unlock(&old_entry->lock);
420 fsnotify_get_mark(chunk_entry);
421 fsnotify_destroy_mark(chunk_entry);
423 fsnotify_put_mark(chunk_entry);
424 fsnotify_put_mark(old_entry);
427 list_replace_init(&old->trees, &chunk->trees);
428 for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
429 struct audit_tree *s = old->owners[n].owner;
431 p->index = old->owners[n].index;
432 if (!s) /* result of fallback in untag */
435 list_replace_init(&old->owners[n].list, &p->list);
437 p->index = (chunk->count - 1) | (1U<<31);
440 list_add(&p->list, &tree->chunks);
441 list_replace_rcu(&old->hash, &chunk->hash);
442 list_for_each_entry(owner, &chunk->trees, same_root)
447 list_add(&tree->same_root, &chunk->trees);
449 spin_unlock(&hash_lock);
450 spin_unlock(&chunk_entry->lock);
451 spin_unlock(&old_entry->lock);
452 fsnotify_destroy_mark(old_entry);
453 fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
457 static void kill_rules(struct audit_tree *tree)
459 struct audit_krule *rule, *next;
460 struct audit_entry *entry;
461 struct audit_buffer *ab;
463 list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
464 entry = container_of(rule, struct audit_entry, rule);
466 list_del_init(&rule->rlist);
468 /* not a half-baked one */
469 ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
470 audit_log_format(ab, "op=");
471 audit_log_string(ab, "remove rule");
472 audit_log_format(ab, " dir=");
473 audit_log_untrustedstring(ab, rule->tree->pathname);
474 audit_log_key(ab, rule->filterkey);
475 audit_log_format(ab, " list=%d res=1", rule->listnr);
478 list_del_rcu(&entry->list);
479 list_del(&entry->rule.list);
480 call_rcu(&entry->rcu, audit_free_rule_rcu);
486 * finish killing struct audit_tree
488 static void prune_one(struct audit_tree *victim)
490 spin_lock(&hash_lock);
491 while (!list_empty(&victim->chunks)) {
494 p = list_entry(victim->chunks.next, struct node, list);
498 spin_unlock(&hash_lock);
502 /* trim the uncommitted chunks from tree */
504 static void trim_marked(struct audit_tree *tree)
506 struct list_head *p, *q;
507 spin_lock(&hash_lock);
509 spin_unlock(&hash_lock);
513 for (p = tree->chunks.next; p != &tree->chunks; p = q) {
514 struct node *node = list_entry(p, struct node, list);
516 if (node->index & (1U<<31)) {
518 list_add(p, &tree->chunks);
522 while (!list_empty(&tree->chunks)) {
525 node = list_entry(tree->chunks.next, struct node, list);
527 /* have we run out of marked? */
528 if (!(node->index & (1U<<31)))
533 if (!tree->root && !tree->goner) {
535 spin_unlock(&hash_lock);
536 mutex_lock(&audit_filter_mutex);
538 list_del_init(&tree->list);
539 mutex_unlock(&audit_filter_mutex);
542 spin_unlock(&hash_lock);
546 static void audit_schedule_prune(void);
548 /* called with audit_filter_mutex */
549 int audit_remove_tree_rule(struct audit_krule *rule)
551 struct audit_tree *tree;
554 spin_lock(&hash_lock);
555 list_del_init(&rule->rlist);
556 if (list_empty(&tree->rules) && !tree->goner) {
558 list_del_init(&tree->same_root);
560 list_move(&tree->list, &prune_list);
562 spin_unlock(&hash_lock);
563 audit_schedule_prune();
567 spin_unlock(&hash_lock);
573 static int compare_root(struct vfsmount *mnt, void *arg)
575 return mnt->mnt_root->d_inode == arg;
578 void audit_trim_trees(void)
580 struct list_head cursor;
582 mutex_lock(&audit_filter_mutex);
583 list_add(&cursor, &tree_list);
584 while (cursor.next != &tree_list) {
585 struct audit_tree *tree;
587 struct vfsmount *root_mnt;
591 tree = container_of(cursor.next, struct audit_tree, list);
594 list_add(&cursor, &tree->list);
595 mutex_unlock(&audit_filter_mutex);
597 err = kern_path(tree->pathname, 0, &path);
601 root_mnt = collect_mounts(&path);
606 spin_lock(&hash_lock);
607 list_for_each_entry(node, &tree->chunks, list) {
608 struct audit_chunk *chunk = find_chunk(node);
609 /* this could be NULL if the watch is dying else where... */
610 struct inode *inode = chunk->mark.i.inode;
611 node->index |= 1U<<31;
612 if (iterate_mounts(compare_root, inode, root_mnt))
613 node->index &= ~(1U<<31);
615 spin_unlock(&hash_lock);
618 drop_collected_mounts(root_mnt);
620 mutex_lock(&audit_filter_mutex);
623 mutex_unlock(&audit_filter_mutex);
626 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
629 if (pathname[0] != '/' ||
630 rule->listnr != AUDIT_FILTER_EXIT ||
632 rule->inode_f || rule->watch || rule->tree)
634 rule->tree = alloc_tree(pathname);
640 void audit_put_tree(struct audit_tree *tree)
645 static int tag_mount(struct vfsmount *mnt, void *arg)
647 return tag_chunk(mnt->mnt_root->d_inode, arg);
650 /* called with audit_filter_mutex */
651 int audit_add_tree_rule(struct audit_krule *rule)
653 struct audit_tree *seed = rule->tree, *tree;
655 struct vfsmount *mnt;
658 list_for_each_entry(tree, &tree_list, list) {
659 if (!strcmp(seed->pathname, tree->pathname)) {
662 list_add(&rule->rlist, &tree->rules);
667 list_add(&tree->list, &tree_list);
668 list_add(&rule->rlist, &tree->rules);
669 /* do not set rule->tree yet */
670 mutex_unlock(&audit_filter_mutex);
672 err = kern_path(tree->pathname, 0, &path);
675 mnt = collect_mounts(&path);
683 err = iterate_mounts(tag_mount, tree, mnt);
684 drop_collected_mounts(mnt);
688 spin_lock(&hash_lock);
689 list_for_each_entry(node, &tree->chunks, list)
690 node->index &= ~(1U<<31);
691 spin_unlock(&hash_lock);
697 mutex_lock(&audit_filter_mutex);
698 if (list_empty(&rule->rlist)) {
707 mutex_lock(&audit_filter_mutex);
708 list_del_init(&tree->list);
709 list_del_init(&tree->rules);
714 int audit_tag_tree(char *old, char *new)
716 struct list_head cursor, barrier;
718 struct path path1, path2;
719 struct vfsmount *tagged;
722 err = kern_path(new, 0, &path2);
725 tagged = collect_mounts(&path2);
730 err = kern_path(old, 0, &path1);
732 drop_collected_mounts(tagged);
736 mutex_lock(&audit_filter_mutex);
737 list_add(&barrier, &tree_list);
738 list_add(&cursor, &barrier);
740 while (cursor.next != &tree_list) {
741 struct audit_tree *tree;
744 tree = container_of(cursor.next, struct audit_tree, list);
747 list_add(&cursor, &tree->list);
748 mutex_unlock(&audit_filter_mutex);
750 err = kern_path(tree->pathname, 0, &path2);
752 good_one = path_is_under(&path1, &path2);
758 mutex_lock(&audit_filter_mutex);
762 failed = iterate_mounts(tag_mount, tree, tagged);
765 mutex_lock(&audit_filter_mutex);
769 mutex_lock(&audit_filter_mutex);
770 spin_lock(&hash_lock);
772 list_del(&tree->list);
773 list_add(&tree->list, &tree_list);
775 spin_unlock(&hash_lock);
779 while (barrier.prev != &tree_list) {
780 struct audit_tree *tree;
782 tree = container_of(barrier.prev, struct audit_tree, list);
784 list_del(&tree->list);
785 list_add(&tree->list, &barrier);
786 mutex_unlock(&audit_filter_mutex);
790 spin_lock(&hash_lock);
791 list_for_each_entry(node, &tree->chunks, list)
792 node->index &= ~(1U<<31);
793 spin_unlock(&hash_lock);
799 mutex_lock(&audit_filter_mutex);
803 mutex_unlock(&audit_filter_mutex);
805 drop_collected_mounts(tagged);
810 * That gets run when evict_chunk() ends up needing to kill audit_tree.
811 * Runs from a separate thread.
813 static int prune_tree_thread(void *unused)
815 mutex_lock(&audit_cmd_mutex);
816 mutex_lock(&audit_filter_mutex);
818 while (!list_empty(&prune_list)) {
819 struct audit_tree *victim;
821 victim = list_entry(prune_list.next, struct audit_tree, list);
822 list_del_init(&victim->list);
824 mutex_unlock(&audit_filter_mutex);
828 mutex_lock(&audit_filter_mutex);
831 mutex_unlock(&audit_filter_mutex);
832 mutex_unlock(&audit_cmd_mutex);
836 static void audit_schedule_prune(void)
838 kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
842 * ... and that one is done if evict_chunk() decides to delay until the end
843 * of syscall. Runs synchronously.
845 void audit_kill_trees(struct list_head *list)
847 mutex_lock(&audit_cmd_mutex);
848 mutex_lock(&audit_filter_mutex);
850 while (!list_empty(list)) {
851 struct audit_tree *victim;
853 victim = list_entry(list->next, struct audit_tree, list);
855 list_del_init(&victim->list);
857 mutex_unlock(&audit_filter_mutex);
861 mutex_lock(&audit_filter_mutex);
864 mutex_unlock(&audit_filter_mutex);
865 mutex_unlock(&audit_cmd_mutex);
869 * Here comes the stuff asynchronous to auditctl operations
872 static void evict_chunk(struct audit_chunk *chunk)
874 struct audit_tree *owner;
875 struct list_head *postponed = audit_killed_trees();
883 mutex_lock(&audit_filter_mutex);
884 spin_lock(&hash_lock);
885 while (!list_empty(&chunk->trees)) {
886 owner = list_entry(chunk->trees.next,
887 struct audit_tree, same_root);
890 list_del_init(&owner->same_root);
891 spin_unlock(&hash_lock);
894 list_move(&owner->list, &prune_list);
897 list_move(&owner->list, postponed);
899 spin_lock(&hash_lock);
901 list_del_rcu(&chunk->hash);
902 for (n = 0; n < chunk->count; n++)
903 list_del_init(&chunk->owners[n].list);
904 spin_unlock(&hash_lock);
906 audit_schedule_prune();
907 mutex_unlock(&audit_filter_mutex);
910 static int audit_tree_handle_event(struct fsnotify_group *group,
911 struct fsnotify_mark *inode_mark,
912 struct fsnotify_mark *vfsmonut_mark,
913 struct fsnotify_event *event)
919 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
921 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
924 fsnotify_put_mark(entry);
927 static bool audit_tree_send_event(struct fsnotify_group *group, struct inode *inode,
928 struct fsnotify_mark *inode_mark,
929 struct fsnotify_mark *vfsmount_mark,
930 __u32 mask, void *data, int data_type)
935 static const struct fsnotify_ops audit_tree_ops = {
936 .handle_event = audit_tree_handle_event,
937 .should_send_event = audit_tree_send_event,
938 .free_group_priv = NULL,
939 .free_event_priv = NULL,
940 .freeing_mark = audit_tree_freeing_mark,
943 static int __init audit_tree_init(void)
947 audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
948 if (IS_ERR(audit_tree_group))
949 audit_panic("cannot initialize fsnotify group for rectree watches");
951 for (i = 0; i < HASH_SIZE; i++)
952 INIT_LIST_HEAD(&chunk_hash_heads[i]);
956 __initcall(audit_tree_init);