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);
40 static struct task_struct *prune_thread;
43 * One struct chunk is attached to each inode of interest.
44 * We replace struct chunk on tagging/untagging.
45 * Rules have pointer to struct audit_tree.
46 * Rules have struct list_head rlist forming a list of rules over
48 * References to struct chunk are collected at audit_inode{,_child}()
49 * time and used in AUDIT_TREE rule matching.
50 * These references are dropped at the same time we are calling
51 * audit_free_names(), etc.
53 * Cyclic lists galore:
54 * tree.chunks anchors chunk.owners[].list hash_lock
55 * tree.rules anchors rule.rlist audit_filter_mutex
56 * chunk.trees anchors tree.same_root hash_lock
57 * chunk.hash is a hash with middle bits of watch.inode as
58 * a hash function. RCU, hash_lock
60 * tree is refcounted; one reference for "some rules on rules_list refer to
61 * it", one for each chunk with pointer to it.
63 * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
64 * of watch contributes 1 to .refs).
66 * node.index allows to get from node.list to containing chunk.
67 * MSB of that sucker is stolen to mark taggings that we might have to
68 * revert - several operations have very unpleasant cleanup logics and
69 * that makes a difference. Some.
72 static struct fsnotify_group *audit_tree_group;
74 static struct audit_tree *alloc_tree(const char *s)
76 struct audit_tree *tree;
78 tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
80 atomic_set(&tree->count, 1);
82 INIT_LIST_HEAD(&tree->chunks);
83 INIT_LIST_HEAD(&tree->rules);
84 INIT_LIST_HEAD(&tree->list);
85 INIT_LIST_HEAD(&tree->same_root);
87 strcpy(tree->pathname, s);
92 static inline void get_tree(struct audit_tree *tree)
94 atomic_inc(&tree->count);
97 static inline void put_tree(struct audit_tree *tree)
99 if (atomic_dec_and_test(&tree->count))
100 kfree_rcu(tree, head);
103 /* to avoid bringing the entire thing in audit.h */
104 const char *audit_tree_path(struct audit_tree *tree)
106 return tree->pathname;
109 static void free_chunk(struct audit_chunk *chunk)
113 for (i = 0; i < chunk->count; i++) {
114 if (chunk->owners[i].owner)
115 put_tree(chunk->owners[i].owner);
120 void audit_put_chunk(struct audit_chunk *chunk)
122 if (atomic_long_dec_and_test(&chunk->refs))
126 static void __put_chunk(struct rcu_head *rcu)
128 struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
129 audit_put_chunk(chunk);
132 static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
134 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
135 call_rcu(&chunk->head, __put_chunk);
138 static struct audit_chunk *alloc_chunk(int count)
140 struct audit_chunk *chunk;
144 size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
145 chunk = kzalloc(size, GFP_KERNEL);
149 INIT_LIST_HEAD(&chunk->hash);
150 INIT_LIST_HEAD(&chunk->trees);
151 chunk->count = count;
152 atomic_long_set(&chunk->refs, 1);
153 for (i = 0; i < count; i++) {
154 INIT_LIST_HEAD(&chunk->owners[i].list);
155 chunk->owners[i].index = i;
157 fsnotify_init_mark(&chunk->mark, audit_tree_destroy_watch);
158 chunk->mark.mask = FS_IN_IGNORED;
162 enum {HASH_SIZE = 128};
163 static struct list_head chunk_hash_heads[HASH_SIZE];
164 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
166 static inline struct list_head *chunk_hash(const struct inode *inode)
168 unsigned long n = (unsigned long)inode / L1_CACHE_BYTES;
169 return chunk_hash_heads + n % HASH_SIZE;
172 /* hash_lock & entry->lock is held by caller */
173 static void insert_hash(struct audit_chunk *chunk)
175 struct fsnotify_mark *entry = &chunk->mark;
176 struct list_head *list;
180 list = chunk_hash(entry->inode);
181 list_add_rcu(&chunk->hash, list);
184 /* called under rcu_read_lock */
185 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
187 struct list_head *list = chunk_hash(inode);
188 struct audit_chunk *p;
190 list_for_each_entry_rcu(p, list, hash) {
191 /* mark.inode may have gone NULL, but who cares? */
192 if (p->mark.inode == inode) {
193 atomic_long_inc(&p->refs);
200 bool audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
203 for (n = 0; n < chunk->count; n++)
204 if (chunk->owners[n].owner == tree)
209 /* tagging and untagging inodes with trees */
211 static struct audit_chunk *find_chunk(struct node *p)
213 int index = p->index & ~(1U<<31);
215 return container_of(p, struct audit_chunk, owners[0]);
218 static void untag_chunk(struct node *p)
220 struct audit_chunk *chunk = find_chunk(p);
221 struct fsnotify_mark *entry = &chunk->mark;
222 struct audit_chunk *new = NULL;
223 struct audit_tree *owner;
224 int size = chunk->count - 1;
227 fsnotify_get_mark(entry);
229 spin_unlock(&hash_lock);
232 new = alloc_chunk(size);
234 spin_lock(&entry->lock);
235 if (chunk->dead || !entry->inode) {
236 spin_unlock(&entry->lock);
246 spin_lock(&hash_lock);
247 list_del_init(&chunk->trees);
248 if (owner->root == chunk)
250 list_del_init(&p->list);
251 list_del_rcu(&chunk->hash);
252 spin_unlock(&hash_lock);
253 spin_unlock(&entry->lock);
254 fsnotify_destroy_mark(entry, audit_tree_group);
261 fsnotify_duplicate_mark(&new->mark, entry);
262 if (fsnotify_add_mark(&new->mark, new->mark.group, new->mark.inode, NULL, 1)) {
263 fsnotify_put_mark(&new->mark);
268 spin_lock(&hash_lock);
269 list_replace_init(&chunk->trees, &new->trees);
270 if (owner->root == chunk) {
271 list_del_init(&owner->same_root);
275 for (i = j = 0; j <= size; i++, j++) {
276 struct audit_tree *s;
277 if (&chunk->owners[j] == p) {
278 list_del_init(&p->list);
282 s = chunk->owners[j].owner;
283 new->owners[i].owner = s;
284 new->owners[i].index = chunk->owners[j].index - j + i;
285 if (!s) /* result of earlier fallback */
288 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
291 list_replace_rcu(&chunk->hash, &new->hash);
292 list_for_each_entry(owner, &new->trees, same_root)
294 spin_unlock(&hash_lock);
295 spin_unlock(&entry->lock);
296 fsnotify_destroy_mark(entry, audit_tree_group);
297 fsnotify_put_mark(&new->mark); /* drop initial reference */
301 // do the best we can
302 spin_lock(&hash_lock);
303 if (owner->root == chunk) {
304 list_del_init(&owner->same_root);
307 list_del_init(&p->list);
310 spin_unlock(&hash_lock);
311 spin_unlock(&entry->lock);
313 fsnotify_put_mark(entry);
314 spin_lock(&hash_lock);
317 static int create_chunk(struct inode *inode, struct audit_tree *tree)
319 struct fsnotify_mark *entry;
320 struct audit_chunk *chunk = alloc_chunk(1);
324 entry = &chunk->mark;
325 if (fsnotify_add_mark(entry, audit_tree_group, inode, NULL, 0)) {
326 fsnotify_put_mark(entry);
330 spin_lock(&entry->lock);
331 spin_lock(&hash_lock);
333 spin_unlock(&hash_lock);
335 spin_unlock(&entry->lock);
336 fsnotify_destroy_mark(entry, audit_tree_group);
337 fsnotify_put_mark(entry);
340 chunk->owners[0].index = (1U << 31);
341 chunk->owners[0].owner = tree;
343 list_add(&chunk->owners[0].list, &tree->chunks);
346 list_add(&tree->same_root, &chunk->trees);
349 spin_unlock(&hash_lock);
350 spin_unlock(&entry->lock);
351 fsnotify_put_mark(entry); /* drop initial reference */
355 /* the first tagged inode becomes root of tree */
356 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
358 struct fsnotify_mark *old_entry, *chunk_entry;
359 struct audit_tree *owner;
360 struct audit_chunk *chunk, *old;
364 old_entry = fsnotify_find_inode_mark(audit_tree_group, inode);
366 return create_chunk(inode, tree);
368 old = container_of(old_entry, struct audit_chunk, mark);
370 /* are we already there? */
371 spin_lock(&hash_lock);
372 for (n = 0; n < old->count; n++) {
373 if (old->owners[n].owner == tree) {
374 spin_unlock(&hash_lock);
375 fsnotify_put_mark(old_entry);
379 spin_unlock(&hash_lock);
381 chunk = alloc_chunk(old->count + 1);
383 fsnotify_put_mark(old_entry);
387 chunk_entry = &chunk->mark;
389 spin_lock(&old_entry->lock);
390 if (!old_entry->inode) {
391 /* old_entry is being shot, lets just lie */
392 spin_unlock(&old_entry->lock);
393 fsnotify_put_mark(old_entry);
398 fsnotify_duplicate_mark(chunk_entry, old_entry);
399 if (fsnotify_add_mark(chunk_entry, chunk_entry->group, chunk_entry->inode, NULL, 1)) {
400 spin_unlock(&old_entry->lock);
401 fsnotify_put_mark(chunk_entry);
402 fsnotify_put_mark(old_entry);
406 /* even though we hold old_entry->lock, this is safe since chunk_entry->lock could NEVER have been grabbed before */
407 spin_lock(&chunk_entry->lock);
408 spin_lock(&hash_lock);
410 /* we now hold old_entry->lock, chunk_entry->lock, and hash_lock */
412 spin_unlock(&hash_lock);
414 spin_unlock(&chunk_entry->lock);
415 spin_unlock(&old_entry->lock);
417 fsnotify_destroy_mark(chunk_entry, audit_tree_group);
419 fsnotify_put_mark(chunk_entry);
420 fsnotify_put_mark(old_entry);
423 list_replace_init(&old->trees, &chunk->trees);
424 for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
425 struct audit_tree *s = old->owners[n].owner;
427 p->index = old->owners[n].index;
428 if (!s) /* result of fallback in untag */
431 list_replace_init(&old->owners[n].list, &p->list);
433 p->index = (chunk->count - 1) | (1U<<31);
436 list_add(&p->list, &tree->chunks);
437 list_replace_rcu(&old->hash, &chunk->hash);
438 list_for_each_entry(owner, &chunk->trees, same_root)
443 list_add(&tree->same_root, &chunk->trees);
445 spin_unlock(&hash_lock);
446 spin_unlock(&chunk_entry->lock);
447 spin_unlock(&old_entry->lock);
448 fsnotify_destroy_mark(old_entry, audit_tree_group);
449 fsnotify_put_mark(chunk_entry); /* drop initial reference */
450 fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
454 static void audit_tree_log_remove_rule(struct audit_krule *rule)
456 struct audit_buffer *ab;
458 ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
461 audit_log_format(ab, "op=remove_rule");
462 audit_log_format(ab, " dir=");
463 audit_log_untrustedstring(ab, rule->tree->pathname);
464 audit_log_key(ab, rule->filterkey);
465 audit_log_format(ab, " list=%d res=1", rule->listnr);
469 static void kill_rules(struct audit_tree *tree)
471 struct audit_krule *rule, *next;
472 struct audit_entry *entry;
474 list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
475 entry = container_of(rule, struct audit_entry, rule);
477 list_del_init(&rule->rlist);
479 /* not a half-baked one */
480 audit_tree_log_remove_rule(rule);
482 audit_remove_mark(entry->rule.exe);
484 list_del_rcu(&entry->list);
485 list_del(&entry->rule.list);
486 call_rcu(&entry->rcu, audit_free_rule_rcu);
492 * finish killing struct audit_tree
494 static void prune_one(struct audit_tree *victim)
496 spin_lock(&hash_lock);
497 while (!list_empty(&victim->chunks)) {
500 p = list_entry(victim->chunks.next, struct node, list);
504 spin_unlock(&hash_lock);
508 /* trim the uncommitted chunks from tree */
510 static void trim_marked(struct audit_tree *tree)
512 struct list_head *p, *q;
513 spin_lock(&hash_lock);
515 spin_unlock(&hash_lock);
519 for (p = tree->chunks.next; p != &tree->chunks; p = q) {
520 struct node *node = list_entry(p, struct node, list);
522 if (node->index & (1U<<31)) {
524 list_add(p, &tree->chunks);
528 while (!list_empty(&tree->chunks)) {
531 node = list_entry(tree->chunks.next, struct node, list);
533 /* have we run out of marked? */
534 if (!(node->index & (1U<<31)))
539 if (!tree->root && !tree->goner) {
541 spin_unlock(&hash_lock);
542 mutex_lock(&audit_filter_mutex);
544 list_del_init(&tree->list);
545 mutex_unlock(&audit_filter_mutex);
548 spin_unlock(&hash_lock);
552 static void audit_schedule_prune(void);
554 /* called with audit_filter_mutex */
555 int audit_remove_tree_rule(struct audit_krule *rule)
557 struct audit_tree *tree;
560 spin_lock(&hash_lock);
561 list_del_init(&rule->rlist);
562 if (list_empty(&tree->rules) && !tree->goner) {
564 list_del_init(&tree->same_root);
566 list_move(&tree->list, &prune_list);
568 spin_unlock(&hash_lock);
569 audit_schedule_prune();
573 spin_unlock(&hash_lock);
579 static int compare_root(struct vfsmount *mnt, void *arg)
581 return d_backing_inode(mnt->mnt_root) == arg;
584 void audit_trim_trees(void)
586 struct list_head cursor;
588 mutex_lock(&audit_filter_mutex);
589 list_add(&cursor, &tree_list);
590 while (cursor.next != &tree_list) {
591 struct audit_tree *tree;
593 struct vfsmount *root_mnt;
597 tree = container_of(cursor.next, struct audit_tree, list);
600 list_add(&cursor, &tree->list);
601 mutex_unlock(&audit_filter_mutex);
603 err = kern_path(tree->pathname, 0, &path);
607 root_mnt = collect_mounts(&path);
609 if (IS_ERR(root_mnt))
612 spin_lock(&hash_lock);
613 list_for_each_entry(node, &tree->chunks, list) {
614 struct audit_chunk *chunk = find_chunk(node);
615 /* this could be NULL if the watch is dying else where... */
616 struct inode *inode = chunk->mark.inode;
617 node->index |= 1U<<31;
618 if (iterate_mounts(compare_root, inode, root_mnt))
619 node->index &= ~(1U<<31);
621 spin_unlock(&hash_lock);
623 drop_collected_mounts(root_mnt);
626 mutex_lock(&audit_filter_mutex);
629 mutex_unlock(&audit_filter_mutex);
632 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
635 if (pathname[0] != '/' ||
636 rule->listnr != AUDIT_FILTER_EXIT ||
638 rule->inode_f || rule->watch || rule->tree)
640 rule->tree = alloc_tree(pathname);
646 void audit_put_tree(struct audit_tree *tree)
651 static int tag_mount(struct vfsmount *mnt, void *arg)
653 return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
657 * That gets run when evict_chunk() ends up needing to kill audit_tree.
658 * Runs from a separate thread.
660 static int prune_tree_thread(void *unused)
663 if (list_empty(&prune_list)) {
664 set_current_state(TASK_INTERRUPTIBLE);
668 mutex_lock(&audit_cmd_mutex);
669 mutex_lock(&audit_filter_mutex);
671 while (!list_empty(&prune_list)) {
672 struct audit_tree *victim;
674 victim = list_entry(prune_list.next,
675 struct audit_tree, list);
676 list_del_init(&victim->list);
678 mutex_unlock(&audit_filter_mutex);
682 mutex_lock(&audit_filter_mutex);
685 mutex_unlock(&audit_filter_mutex);
686 mutex_unlock(&audit_cmd_mutex);
691 static int audit_launch_prune(void)
695 prune_thread = kthread_run(prune_tree_thread, NULL,
697 if (IS_ERR(prune_thread)) {
698 pr_err("cannot start thread audit_prune_tree");
705 /* called with audit_filter_mutex */
706 int audit_add_tree_rule(struct audit_krule *rule)
708 struct audit_tree *seed = rule->tree, *tree;
710 struct vfsmount *mnt;
714 list_for_each_entry(tree, &tree_list, list) {
715 if (!strcmp(seed->pathname, tree->pathname)) {
718 list_add(&rule->rlist, &tree->rules);
723 list_add(&tree->list, &tree_list);
724 list_add(&rule->rlist, &tree->rules);
725 /* do not set rule->tree yet */
726 mutex_unlock(&audit_filter_mutex);
728 if (unlikely(!prune_thread)) {
729 err = audit_launch_prune();
734 err = kern_path(tree->pathname, 0, &path);
737 mnt = collect_mounts(&path);
745 err = iterate_mounts(tag_mount, tree, mnt);
746 drop_collected_mounts(mnt);
750 spin_lock(&hash_lock);
751 list_for_each_entry(node, &tree->chunks, list)
752 node->index &= ~(1U<<31);
753 spin_unlock(&hash_lock);
759 mutex_lock(&audit_filter_mutex);
760 if (list_empty(&rule->rlist)) {
769 mutex_lock(&audit_filter_mutex);
770 list_del_init(&tree->list);
771 list_del_init(&tree->rules);
776 int audit_tag_tree(char *old, char *new)
778 struct list_head cursor, barrier;
780 struct path path1, path2;
781 struct vfsmount *tagged;
784 err = kern_path(new, 0, &path2);
787 tagged = collect_mounts(&path2);
790 return PTR_ERR(tagged);
792 err = kern_path(old, 0, &path1);
794 drop_collected_mounts(tagged);
798 mutex_lock(&audit_filter_mutex);
799 list_add(&barrier, &tree_list);
800 list_add(&cursor, &barrier);
802 while (cursor.next != &tree_list) {
803 struct audit_tree *tree;
806 tree = container_of(cursor.next, struct audit_tree, list);
809 list_add(&cursor, &tree->list);
810 mutex_unlock(&audit_filter_mutex);
812 err = kern_path(tree->pathname, 0, &path2);
814 good_one = path_is_under(&path1, &path2);
820 mutex_lock(&audit_filter_mutex);
824 failed = iterate_mounts(tag_mount, tree, tagged);
827 mutex_lock(&audit_filter_mutex);
831 mutex_lock(&audit_filter_mutex);
832 spin_lock(&hash_lock);
834 list_del(&tree->list);
835 list_add(&tree->list, &tree_list);
837 spin_unlock(&hash_lock);
841 while (barrier.prev != &tree_list) {
842 struct audit_tree *tree;
844 tree = container_of(barrier.prev, struct audit_tree, list);
846 list_del(&tree->list);
847 list_add(&tree->list, &barrier);
848 mutex_unlock(&audit_filter_mutex);
852 spin_lock(&hash_lock);
853 list_for_each_entry(node, &tree->chunks, list)
854 node->index &= ~(1U<<31);
855 spin_unlock(&hash_lock);
861 mutex_lock(&audit_filter_mutex);
865 mutex_unlock(&audit_filter_mutex);
867 drop_collected_mounts(tagged);
872 static void audit_schedule_prune(void)
874 wake_up_process(prune_thread);
878 * ... and that one is done if evict_chunk() decides to delay until the end
879 * of syscall. Runs synchronously.
881 void audit_kill_trees(struct list_head *list)
883 mutex_lock(&audit_cmd_mutex);
884 mutex_lock(&audit_filter_mutex);
886 while (!list_empty(list)) {
887 struct audit_tree *victim;
889 victim = list_entry(list->next, struct audit_tree, list);
891 list_del_init(&victim->list);
893 mutex_unlock(&audit_filter_mutex);
897 mutex_lock(&audit_filter_mutex);
900 mutex_unlock(&audit_filter_mutex);
901 mutex_unlock(&audit_cmd_mutex);
905 * Here comes the stuff asynchronous to auditctl operations
908 static void evict_chunk(struct audit_chunk *chunk)
910 struct audit_tree *owner;
911 struct list_head *postponed = audit_killed_trees();
919 mutex_lock(&audit_filter_mutex);
920 spin_lock(&hash_lock);
921 while (!list_empty(&chunk->trees)) {
922 owner = list_entry(chunk->trees.next,
923 struct audit_tree, same_root);
926 list_del_init(&owner->same_root);
927 spin_unlock(&hash_lock);
930 list_move(&owner->list, &prune_list);
933 list_move(&owner->list, postponed);
935 spin_lock(&hash_lock);
937 list_del_rcu(&chunk->hash);
938 for (n = 0; n < chunk->count; n++)
939 list_del_init(&chunk->owners[n].list);
940 spin_unlock(&hash_lock);
941 mutex_unlock(&audit_filter_mutex);
943 audit_schedule_prune();
946 static int audit_tree_handle_event(struct fsnotify_group *group,
947 struct inode *to_tell,
948 struct fsnotify_mark *inode_mark,
949 struct fsnotify_mark *vfsmount_mark,
950 u32 mask, const void *data, int data_type,
951 const unsigned char *file_name, u32 cookie)
956 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
958 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
963 * We are guaranteed to have at least one reference to the mark from
964 * either the inode or the caller of fsnotify_destroy_mark().
966 BUG_ON(atomic_read(&entry->refcnt) < 1);
969 static const struct fsnotify_ops audit_tree_ops = {
970 .handle_event = audit_tree_handle_event,
971 .freeing_mark = audit_tree_freeing_mark,
974 static int __init audit_tree_init(void)
978 audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
979 if (IS_ERR(audit_tree_group))
980 audit_panic("cannot initialize fsnotify group for rectree watches");
982 for (i = 0; i < HASH_SIZE; i++)
983 INIT_LIST_HEAD(&chunk_hash_heads[i]);
987 __initcall(audit_tree_init);