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;
227 struct audit_tree *owner;
228 int size = chunk->count - 1;
231 fsnotify_get_mark(entry);
233 spin_unlock(&hash_lock);
235 spin_lock(&entry->lock);
236 if (chunk->dead || !entry->i.inode) {
237 spin_unlock(&entry->lock);
245 spin_lock(&hash_lock);
246 list_del_init(&chunk->trees);
247 if (owner->root == chunk)
249 list_del_init(&p->list);
250 list_del_rcu(&chunk->hash);
251 spin_unlock(&hash_lock);
252 spin_unlock(&entry->lock);
253 fsnotify_destroy_mark(entry);
254 fsnotify_put_mark(entry);
258 new = alloc_chunk(size);
261 fsnotify_duplicate_mark(&new->mark, entry);
262 if (fsnotify_add_mark(&new->mark, new->mark.group, new->mark.i.inode, NULL, 1)) {
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);
297 fsnotify_put_mark(entry);
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)) {
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);
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);
354 /* the first tagged inode becomes root of tree */
355 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
357 struct fsnotify_mark *old_entry, *chunk_entry;
358 struct audit_tree *owner;
359 struct audit_chunk *chunk, *old;
363 old_entry = fsnotify_find_inode_mark(audit_tree_group, inode);
365 return create_chunk(inode, tree);
367 old = container_of(old_entry, struct audit_chunk, mark);
369 /* are we already there? */
370 spin_lock(&hash_lock);
371 for (n = 0; n < old->count; n++) {
372 if (old->owners[n].owner == tree) {
373 spin_unlock(&hash_lock);
374 fsnotify_put_mark(old_entry);
378 spin_unlock(&hash_lock);
380 chunk = alloc_chunk(old->count + 1);
382 fsnotify_put_mark(old_entry);
386 chunk_entry = &chunk->mark;
388 spin_lock(&old_entry->lock);
389 if (!old_entry->i.inode) {
390 /* old_entry is being shot, lets just lie */
391 spin_unlock(&old_entry->lock);
392 fsnotify_put_mark(old_entry);
397 fsnotify_duplicate_mark(chunk_entry, old_entry);
398 if (fsnotify_add_mark(chunk_entry, chunk_entry->group, chunk_entry->i.inode, NULL, 1)) {
399 spin_unlock(&old_entry->lock);
401 fsnotify_put_mark(old_entry);
405 /* even though we hold old_entry->lock, this is safe since chunk_entry->lock could NEVER have been grabbed before */
406 spin_lock(&chunk_entry->lock);
407 spin_lock(&hash_lock);
409 /* we now hold old_entry->lock, chunk_entry->lock, and hash_lock */
411 spin_unlock(&hash_lock);
413 spin_unlock(&chunk_entry->lock);
414 spin_unlock(&old_entry->lock);
416 fsnotify_destroy_mark(chunk_entry);
418 fsnotify_put_mark(chunk_entry);
419 fsnotify_put_mark(old_entry);
422 list_replace_init(&old->trees, &chunk->trees);
423 for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
424 struct audit_tree *s = old->owners[n].owner;
426 p->index = old->owners[n].index;
427 if (!s) /* result of fallback in untag */
430 list_replace_init(&old->owners[n].list, &p->list);
432 p->index = (chunk->count - 1) | (1U<<31);
435 list_add(&p->list, &tree->chunks);
436 list_replace_rcu(&old->hash, &chunk->hash);
437 list_for_each_entry(owner, &chunk->trees, same_root)
442 list_add(&tree->same_root, &chunk->trees);
444 spin_unlock(&hash_lock);
445 spin_unlock(&chunk_entry->lock);
446 spin_unlock(&old_entry->lock);
447 fsnotify_destroy_mark(old_entry);
448 fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
449 fsnotify_put_mark(old_entry); /* and kill it */
453 static void kill_rules(struct audit_tree *tree)
455 struct audit_krule *rule, *next;
456 struct audit_entry *entry;
457 struct audit_buffer *ab;
459 list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
460 entry = container_of(rule, struct audit_entry, rule);
462 list_del_init(&rule->rlist);
464 /* not a half-baked one */
465 ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
466 audit_log_format(ab, "op=");
467 audit_log_string(ab, "remove rule");
468 audit_log_format(ab, " dir=");
469 audit_log_untrustedstring(ab, rule->tree->pathname);
470 audit_log_key(ab, rule->filterkey);
471 audit_log_format(ab, " list=%d res=1", rule->listnr);
474 list_del_rcu(&entry->list);
475 list_del(&entry->rule.list);
476 call_rcu(&entry->rcu, audit_free_rule_rcu);
482 * finish killing struct audit_tree
484 static void prune_one(struct audit_tree *victim)
486 spin_lock(&hash_lock);
487 while (!list_empty(&victim->chunks)) {
490 p = list_entry(victim->chunks.next, struct node, list);
494 spin_unlock(&hash_lock);
498 /* trim the uncommitted chunks from tree */
500 static void trim_marked(struct audit_tree *tree)
502 struct list_head *p, *q;
503 spin_lock(&hash_lock);
505 spin_unlock(&hash_lock);
509 for (p = tree->chunks.next; p != &tree->chunks; p = q) {
510 struct node *node = list_entry(p, struct node, list);
512 if (node->index & (1U<<31)) {
514 list_add(p, &tree->chunks);
518 while (!list_empty(&tree->chunks)) {
521 node = list_entry(tree->chunks.next, struct node, list);
523 /* have we run out of marked? */
524 if (!(node->index & (1U<<31)))
529 if (!tree->root && !tree->goner) {
531 spin_unlock(&hash_lock);
532 mutex_lock(&audit_filter_mutex);
534 list_del_init(&tree->list);
535 mutex_unlock(&audit_filter_mutex);
538 spin_unlock(&hash_lock);
542 static void audit_schedule_prune(void);
544 /* called with audit_filter_mutex */
545 int audit_remove_tree_rule(struct audit_krule *rule)
547 struct audit_tree *tree;
550 spin_lock(&hash_lock);
551 list_del_init(&rule->rlist);
552 if (list_empty(&tree->rules) && !tree->goner) {
554 list_del_init(&tree->same_root);
556 list_move(&tree->list, &prune_list);
558 spin_unlock(&hash_lock);
559 audit_schedule_prune();
563 spin_unlock(&hash_lock);
569 static int compare_root(struct vfsmount *mnt, void *arg)
571 return mnt->mnt_root->d_inode == arg;
574 void audit_trim_trees(void)
576 struct list_head cursor;
578 mutex_lock(&audit_filter_mutex);
579 list_add(&cursor, &tree_list);
580 while (cursor.next != &tree_list) {
581 struct audit_tree *tree;
583 struct vfsmount *root_mnt;
587 tree = container_of(cursor.next, struct audit_tree, list);
590 list_add(&cursor, &tree->list);
591 mutex_unlock(&audit_filter_mutex);
593 err = kern_path(tree->pathname, 0, &path);
597 root_mnt = collect_mounts(&path);
602 spin_lock(&hash_lock);
603 list_for_each_entry(node, &tree->chunks, list) {
604 struct audit_chunk *chunk = find_chunk(node);
605 /* this could be NULL if the watch is dieing else where... */
606 struct inode *inode = chunk->mark.i.inode;
607 node->index |= 1U<<31;
608 if (iterate_mounts(compare_root, inode, root_mnt))
609 node->index &= ~(1U<<31);
611 spin_unlock(&hash_lock);
614 drop_collected_mounts(root_mnt);
616 mutex_lock(&audit_filter_mutex);
619 mutex_unlock(&audit_filter_mutex);
622 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
625 if (pathname[0] != '/' ||
626 rule->listnr != AUDIT_FILTER_EXIT ||
628 rule->inode_f || rule->watch || rule->tree)
630 rule->tree = alloc_tree(pathname);
636 void audit_put_tree(struct audit_tree *tree)
641 static int tag_mount(struct vfsmount *mnt, void *arg)
643 return tag_chunk(mnt->mnt_root->d_inode, arg);
646 /* called with audit_filter_mutex */
647 int audit_add_tree_rule(struct audit_krule *rule)
649 struct audit_tree *seed = rule->tree, *tree;
651 struct vfsmount *mnt;
654 list_for_each_entry(tree, &tree_list, list) {
655 if (!strcmp(seed->pathname, tree->pathname)) {
658 list_add(&rule->rlist, &tree->rules);
663 list_add(&tree->list, &tree_list);
664 list_add(&rule->rlist, &tree->rules);
665 /* do not set rule->tree yet */
666 mutex_unlock(&audit_filter_mutex);
668 err = kern_path(tree->pathname, 0, &path);
671 mnt = collect_mounts(&path);
679 err = iterate_mounts(tag_mount, tree, mnt);
680 drop_collected_mounts(mnt);
684 spin_lock(&hash_lock);
685 list_for_each_entry(node, &tree->chunks, list)
686 node->index &= ~(1U<<31);
687 spin_unlock(&hash_lock);
693 mutex_lock(&audit_filter_mutex);
694 if (list_empty(&rule->rlist)) {
703 mutex_lock(&audit_filter_mutex);
704 list_del_init(&tree->list);
705 list_del_init(&tree->rules);
710 int audit_tag_tree(char *old, char *new)
712 struct list_head cursor, barrier;
714 struct path path1, path2;
715 struct vfsmount *tagged;
718 err = kern_path(new, 0, &path2);
721 tagged = collect_mounts(&path2);
726 err = kern_path(old, 0, &path1);
728 drop_collected_mounts(tagged);
732 mutex_lock(&audit_filter_mutex);
733 list_add(&barrier, &tree_list);
734 list_add(&cursor, &barrier);
736 while (cursor.next != &tree_list) {
737 struct audit_tree *tree;
740 tree = container_of(cursor.next, struct audit_tree, list);
743 list_add(&cursor, &tree->list);
744 mutex_unlock(&audit_filter_mutex);
746 err = kern_path(tree->pathname, 0, &path2);
748 good_one = path_is_under(&path1, &path2);
754 mutex_lock(&audit_filter_mutex);
758 failed = iterate_mounts(tag_mount, tree, tagged);
761 mutex_lock(&audit_filter_mutex);
765 mutex_lock(&audit_filter_mutex);
766 spin_lock(&hash_lock);
768 list_del(&tree->list);
769 list_add(&tree->list, &tree_list);
771 spin_unlock(&hash_lock);
775 while (barrier.prev != &tree_list) {
776 struct audit_tree *tree;
778 tree = container_of(barrier.prev, struct audit_tree, list);
780 list_del(&tree->list);
781 list_add(&tree->list, &barrier);
782 mutex_unlock(&audit_filter_mutex);
786 spin_lock(&hash_lock);
787 list_for_each_entry(node, &tree->chunks, list)
788 node->index &= ~(1U<<31);
789 spin_unlock(&hash_lock);
795 mutex_lock(&audit_filter_mutex);
799 mutex_unlock(&audit_filter_mutex);
801 drop_collected_mounts(tagged);
806 * That gets run when evict_chunk() ends up needing to kill audit_tree.
807 * Runs from a separate thread.
809 static int prune_tree_thread(void *unused)
811 mutex_lock(&audit_cmd_mutex);
812 mutex_lock(&audit_filter_mutex);
814 while (!list_empty(&prune_list)) {
815 struct audit_tree *victim;
817 victim = list_entry(prune_list.next, struct audit_tree, list);
818 list_del_init(&victim->list);
820 mutex_unlock(&audit_filter_mutex);
824 mutex_lock(&audit_filter_mutex);
827 mutex_unlock(&audit_filter_mutex);
828 mutex_unlock(&audit_cmd_mutex);
832 static void audit_schedule_prune(void)
834 kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
838 * ... and that one is done if evict_chunk() decides to delay until the end
839 * of syscall. Runs synchronously.
841 void audit_kill_trees(struct list_head *list)
843 mutex_lock(&audit_cmd_mutex);
844 mutex_lock(&audit_filter_mutex);
846 while (!list_empty(list)) {
847 struct audit_tree *victim;
849 victim = list_entry(list->next, struct audit_tree, list);
851 list_del_init(&victim->list);
853 mutex_unlock(&audit_filter_mutex);
857 mutex_lock(&audit_filter_mutex);
860 mutex_unlock(&audit_filter_mutex);
861 mutex_unlock(&audit_cmd_mutex);
865 * Here comes the stuff asynchronous to auditctl operations
868 static void evict_chunk(struct audit_chunk *chunk)
870 struct audit_tree *owner;
871 struct list_head *postponed = audit_killed_trees();
879 mutex_lock(&audit_filter_mutex);
880 spin_lock(&hash_lock);
881 while (!list_empty(&chunk->trees)) {
882 owner = list_entry(chunk->trees.next,
883 struct audit_tree, same_root);
886 list_del_init(&owner->same_root);
887 spin_unlock(&hash_lock);
890 list_move(&owner->list, &prune_list);
893 list_move(&owner->list, postponed);
895 spin_lock(&hash_lock);
897 list_del_rcu(&chunk->hash);
898 for (n = 0; n < chunk->count; n++)
899 list_del_init(&chunk->owners[n].list);
900 spin_unlock(&hash_lock);
902 audit_schedule_prune();
903 mutex_unlock(&audit_filter_mutex);
906 static int audit_tree_handle_event(struct fsnotify_group *group,
907 struct fsnotify_mark *inode_mark,
908 struct fsnotify_mark *vfsmonut_mark,
909 struct fsnotify_event *event)
915 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
917 struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
920 fsnotify_put_mark(entry);
923 static bool audit_tree_send_event(struct fsnotify_group *group, struct inode *inode,
924 struct fsnotify_mark *inode_mark,
925 struct fsnotify_mark *vfsmount_mark,
926 __u32 mask, void *data, int data_type)
931 static const struct fsnotify_ops audit_tree_ops = {
932 .handle_event = audit_tree_handle_event,
933 .should_send_event = audit_tree_send_event,
934 .free_group_priv = NULL,
935 .free_event_priv = NULL,
936 .freeing_mark = audit_tree_freeing_mark,
939 static int __init audit_tree_init(void)
943 audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
944 if (IS_ERR(audit_tree_group))
945 audit_panic("cannot initialize fsnotify group for rectree watches");
947 for (i = 0; i < HASH_SIZE; i++)
948 INIT_LIST_HEAD(&chunk_hash_heads[i]);
952 __initcall(audit_tree_init);