umh: creds: convert call_usermodehelper_keys() to use subprocess_info->init()
[linux-2.6.git] / kernel / audit_tree.c
1 #include "audit.h"
2 #include <linux/inotify.h>
3 #include <linux/namei.h>
4 #include <linux/mount.h>
5 #include <linux/kthread.h>
6 #include <linux/slab.h>
7
8 struct audit_tree;
9 struct audit_chunk;
10
11 struct 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
23 struct audit_chunk {
24         struct list_head hash;
25         struct inotify_watch watch;
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
38 static LIST_HEAD(tree_list);
39 static 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 inotify_watch + .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
71 static struct inotify_handle *rtree_ih;
72
73 static 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
91 static inline void get_tree(struct audit_tree *tree)
92 {
93         atomic_inc(&tree->count);
94 }
95
96 static void __put_tree(struct rcu_head *rcu)
97 {
98         struct audit_tree *tree = container_of(rcu, struct audit_tree, head);
99         kfree(tree);
100 }
101
102 static inline void put_tree(struct audit_tree *tree)
103 {
104         if (atomic_dec_and_test(&tree->count))
105                 call_rcu(&tree->head, __put_tree);
106 }
107
108 /* to avoid bringing the entire thing in audit.h */
109 const char *audit_tree_path(struct audit_tree *tree)
110 {
111         return tree->pathname;
112 }
113
114 static struct audit_chunk *alloc_chunk(int count)
115 {
116         struct audit_chunk *chunk;
117         size_t size;
118         int i;
119
120         size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
121         chunk = kzalloc(size, GFP_KERNEL);
122         if (!chunk)
123                 return NULL;
124
125         INIT_LIST_HEAD(&chunk->hash);
126         INIT_LIST_HEAD(&chunk->trees);
127         chunk->count = count;
128         atomic_long_set(&chunk->refs, 1);
129         for (i = 0; i < count; i++) {
130                 INIT_LIST_HEAD(&chunk->owners[i].list);
131                 chunk->owners[i].index = i;
132         }
133         inotify_init_watch(&chunk->watch);
134         return chunk;
135 }
136
137 static void free_chunk(struct audit_chunk *chunk)
138 {
139         int i;
140
141         for (i = 0; i < chunk->count; i++) {
142                 if (chunk->owners[i].owner)
143                         put_tree(chunk->owners[i].owner);
144         }
145         kfree(chunk);
146 }
147
148 void audit_put_chunk(struct audit_chunk *chunk)
149 {
150         if (atomic_long_dec_and_test(&chunk->refs))
151                 free_chunk(chunk);
152 }
153
154 static void __put_chunk(struct rcu_head *rcu)
155 {
156         struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
157         audit_put_chunk(chunk);
158 }
159
160 enum {HASH_SIZE = 128};
161 static struct list_head chunk_hash_heads[HASH_SIZE];
162 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
163
164 static 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 is held by caller */
171 static void insert_hash(struct audit_chunk *chunk)
172 {
173         struct list_head *list = chunk_hash(chunk->watch.inode);
174         list_add_rcu(&chunk->hash, list);
175 }
176
177 /* called under rcu_read_lock */
178 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
179 {
180         struct list_head *list = chunk_hash(inode);
181         struct audit_chunk *p;
182
183         list_for_each_entry_rcu(p, list, hash) {
184                 if (p->watch.inode == inode) {
185                         atomic_long_inc(&p->refs);
186                         return p;
187                 }
188         }
189         return NULL;
190 }
191
192 int audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
193 {
194         int n;
195         for (n = 0; n < chunk->count; n++)
196                 if (chunk->owners[n].owner == tree)
197                         return 1;
198         return 0;
199 }
200
201 /* tagging and untagging inodes with trees */
202
203 static struct audit_chunk *find_chunk(struct node *p)
204 {
205         int index = p->index & ~(1U<<31);
206         p -= index;
207         return container_of(p, struct audit_chunk, owners[0]);
208 }
209
210 static void untag_chunk(struct node *p)
211 {
212         struct audit_chunk *chunk = find_chunk(p);
213         struct audit_chunk *new;
214         struct audit_tree *owner;
215         int size = chunk->count - 1;
216         int i, j;
217
218         if (!pin_inotify_watch(&chunk->watch)) {
219                 /*
220                  * Filesystem is shutting down; all watches are getting
221                  * evicted, just take it off the node list for this
222                  * tree and let the eviction logics take care of the
223                  * rest.
224                  */
225                 owner = p->owner;
226                 if (owner->root == chunk) {
227                         list_del_init(&owner->same_root);
228                         owner->root = NULL;
229                 }
230                 list_del_init(&p->list);
231                 p->owner = NULL;
232                 put_tree(owner);
233                 return;
234         }
235
236         spin_unlock(&hash_lock);
237
238         /*
239          * pin_inotify_watch() succeeded, so the watch won't go away
240          * from under us.
241          */
242         mutex_lock(&chunk->watch.inode->inotify_mutex);
243         if (chunk->dead) {
244                 mutex_unlock(&chunk->watch.inode->inotify_mutex);
245                 goto out;
246         }
247
248         owner = p->owner;
249
250         if (!size) {
251                 chunk->dead = 1;
252                 spin_lock(&hash_lock);
253                 list_del_init(&chunk->trees);
254                 if (owner->root == chunk)
255                         owner->root = NULL;
256                 list_del_init(&p->list);
257                 list_del_rcu(&chunk->hash);
258                 spin_unlock(&hash_lock);
259                 inotify_evict_watch(&chunk->watch);
260                 mutex_unlock(&chunk->watch.inode->inotify_mutex);
261                 put_inotify_watch(&chunk->watch);
262                 goto out;
263         }
264
265         new = alloc_chunk(size);
266         if (!new)
267                 goto Fallback;
268         if (inotify_clone_watch(&chunk->watch, &new->watch) < 0) {
269                 free_chunk(new);
270                 goto Fallback;
271         }
272
273         chunk->dead = 1;
274         spin_lock(&hash_lock);
275         list_replace_init(&chunk->trees, &new->trees);
276         if (owner->root == chunk) {
277                 list_del_init(&owner->same_root);
278                 owner->root = NULL;
279         }
280
281         for (i = j = 0; j <= size; i++, j++) {
282                 struct audit_tree *s;
283                 if (&chunk->owners[j] == p) {
284                         list_del_init(&p->list);
285                         i--;
286                         continue;
287                 }
288                 s = chunk->owners[j].owner;
289                 new->owners[i].owner = s;
290                 new->owners[i].index = chunk->owners[j].index - j + i;
291                 if (!s) /* result of earlier fallback */
292                         continue;
293                 get_tree(s);
294                 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
295         }
296
297         list_replace_rcu(&chunk->hash, &new->hash);
298         list_for_each_entry(owner, &new->trees, same_root)
299                 owner->root = new;
300         spin_unlock(&hash_lock);
301         inotify_evict_watch(&chunk->watch);
302         mutex_unlock(&chunk->watch.inode->inotify_mutex);
303         put_inotify_watch(&chunk->watch);
304         goto out;
305
306 Fallback:
307         // do the best we can
308         spin_lock(&hash_lock);
309         if (owner->root == chunk) {
310                 list_del_init(&owner->same_root);
311                 owner->root = NULL;
312         }
313         list_del_init(&p->list);
314         p->owner = NULL;
315         put_tree(owner);
316         spin_unlock(&hash_lock);
317         mutex_unlock(&chunk->watch.inode->inotify_mutex);
318 out:
319         unpin_inotify_watch(&chunk->watch);
320         spin_lock(&hash_lock);
321 }
322
323 static int create_chunk(struct inode *inode, struct audit_tree *tree)
324 {
325         struct audit_chunk *chunk = alloc_chunk(1);
326         if (!chunk)
327                 return -ENOMEM;
328
329         if (inotify_add_watch(rtree_ih, &chunk->watch, inode, IN_IGNORED | IN_DELETE_SELF) < 0) {
330                 free_chunk(chunk);
331                 return -ENOSPC;
332         }
333
334         mutex_lock(&inode->inotify_mutex);
335         spin_lock(&hash_lock);
336         if (tree->goner) {
337                 spin_unlock(&hash_lock);
338                 chunk->dead = 1;
339                 inotify_evict_watch(&chunk->watch);
340                 mutex_unlock(&inode->inotify_mutex);
341                 put_inotify_watch(&chunk->watch);
342                 return 0;
343         }
344         chunk->owners[0].index = (1U << 31);
345         chunk->owners[0].owner = tree;
346         get_tree(tree);
347         list_add(&chunk->owners[0].list, &tree->chunks);
348         if (!tree->root) {
349                 tree->root = chunk;
350                 list_add(&tree->same_root, &chunk->trees);
351         }
352         insert_hash(chunk);
353         spin_unlock(&hash_lock);
354         mutex_unlock(&inode->inotify_mutex);
355         return 0;
356 }
357
358 /* the first tagged inode becomes root of tree */
359 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
360 {
361         struct inotify_watch *watch;
362         struct audit_tree *owner;
363         struct audit_chunk *chunk, *old;
364         struct node *p;
365         int n;
366
367         if (inotify_find_watch(rtree_ih, inode, &watch) < 0)
368                 return create_chunk(inode, tree);
369
370         old = container_of(watch, struct audit_chunk, watch);
371
372         /* are we already there? */
373         spin_lock(&hash_lock);
374         for (n = 0; n < old->count; n++) {
375                 if (old->owners[n].owner == tree) {
376                         spin_unlock(&hash_lock);
377                         put_inotify_watch(&old->watch);
378                         return 0;
379                 }
380         }
381         spin_unlock(&hash_lock);
382
383         chunk = alloc_chunk(old->count + 1);
384         if (!chunk) {
385                 put_inotify_watch(&old->watch);
386                 return -ENOMEM;
387         }
388
389         mutex_lock(&inode->inotify_mutex);
390         if (inotify_clone_watch(&old->watch, &chunk->watch) < 0) {
391                 mutex_unlock(&inode->inotify_mutex);
392                 put_inotify_watch(&old->watch);
393                 free_chunk(chunk);
394                 return -ENOSPC;
395         }
396         spin_lock(&hash_lock);
397         if (tree->goner) {
398                 spin_unlock(&hash_lock);
399                 chunk->dead = 1;
400                 inotify_evict_watch(&chunk->watch);
401                 mutex_unlock(&inode->inotify_mutex);
402                 put_inotify_watch(&old->watch);
403                 put_inotify_watch(&chunk->watch);
404                 return 0;
405         }
406         list_replace_init(&old->trees, &chunk->trees);
407         for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
408                 struct audit_tree *s = old->owners[n].owner;
409                 p->owner = s;
410                 p->index = old->owners[n].index;
411                 if (!s) /* result of fallback in untag */
412                         continue;
413                 get_tree(s);
414                 list_replace_init(&old->owners[n].list, &p->list);
415         }
416         p->index = (chunk->count - 1) | (1U<<31);
417         p->owner = tree;
418         get_tree(tree);
419         list_add(&p->list, &tree->chunks);
420         list_replace_rcu(&old->hash, &chunk->hash);
421         list_for_each_entry(owner, &chunk->trees, same_root)
422                 owner->root = chunk;
423         old->dead = 1;
424         if (!tree->root) {
425                 tree->root = chunk;
426                 list_add(&tree->same_root, &chunk->trees);
427         }
428         spin_unlock(&hash_lock);
429         inotify_evict_watch(&old->watch);
430         mutex_unlock(&inode->inotify_mutex);
431         put_inotify_watch(&old->watch); /* pair to inotify_find_watch */
432         put_inotify_watch(&old->watch); /* and kill it */
433         return 0;
434 }
435
436 static void kill_rules(struct audit_tree *tree)
437 {
438         struct audit_krule *rule, *next;
439         struct audit_entry *entry;
440         struct audit_buffer *ab;
441
442         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
443                 entry = container_of(rule, struct audit_entry, rule);
444
445                 list_del_init(&rule->rlist);
446                 if (rule->tree) {
447                         /* not a half-baked one */
448                         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
449                         audit_log_format(ab, "op=");
450                         audit_log_string(ab, "remove rule");
451                         audit_log_format(ab, " dir=");
452                         audit_log_untrustedstring(ab, rule->tree->pathname);
453                         audit_log_key(ab, rule->filterkey);
454                         audit_log_format(ab, " list=%d res=1", rule->listnr);
455                         audit_log_end(ab);
456                         rule->tree = NULL;
457                         list_del_rcu(&entry->list);
458                         list_del(&entry->rule.list);
459                         call_rcu(&entry->rcu, audit_free_rule_rcu);
460                 }
461         }
462 }
463
464 /*
465  * finish killing struct audit_tree
466  */
467 static void prune_one(struct audit_tree *victim)
468 {
469         spin_lock(&hash_lock);
470         while (!list_empty(&victim->chunks)) {
471                 struct node *p;
472
473                 p = list_entry(victim->chunks.next, struct node, list);
474
475                 untag_chunk(p);
476         }
477         spin_unlock(&hash_lock);
478         put_tree(victim);
479 }
480
481 /* trim the uncommitted chunks from tree */
482
483 static void trim_marked(struct audit_tree *tree)
484 {
485         struct list_head *p, *q;
486         spin_lock(&hash_lock);
487         if (tree->goner) {
488                 spin_unlock(&hash_lock);
489                 return;
490         }
491         /* reorder */
492         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
493                 struct node *node = list_entry(p, struct node, list);
494                 q = p->next;
495                 if (node->index & (1U<<31)) {
496                         list_del_init(p);
497                         list_add(p, &tree->chunks);
498                 }
499         }
500
501         while (!list_empty(&tree->chunks)) {
502                 struct node *node;
503
504                 node = list_entry(tree->chunks.next, struct node, list);
505
506                 /* have we run out of marked? */
507                 if (!(node->index & (1U<<31)))
508                         break;
509
510                 untag_chunk(node);
511         }
512         if (!tree->root && !tree->goner) {
513                 tree->goner = 1;
514                 spin_unlock(&hash_lock);
515                 mutex_lock(&audit_filter_mutex);
516                 kill_rules(tree);
517                 list_del_init(&tree->list);
518                 mutex_unlock(&audit_filter_mutex);
519                 prune_one(tree);
520         } else {
521                 spin_unlock(&hash_lock);
522         }
523 }
524
525 static void audit_schedule_prune(void);
526
527 /* called with audit_filter_mutex */
528 int audit_remove_tree_rule(struct audit_krule *rule)
529 {
530         struct audit_tree *tree;
531         tree = rule->tree;
532         if (tree) {
533                 spin_lock(&hash_lock);
534                 list_del_init(&rule->rlist);
535                 if (list_empty(&tree->rules) && !tree->goner) {
536                         tree->root = NULL;
537                         list_del_init(&tree->same_root);
538                         tree->goner = 1;
539                         list_move(&tree->list, &prune_list);
540                         rule->tree = NULL;
541                         spin_unlock(&hash_lock);
542                         audit_schedule_prune();
543                         return 1;
544                 }
545                 rule->tree = NULL;
546                 spin_unlock(&hash_lock);
547                 return 1;
548         }
549         return 0;
550 }
551
552 static int compare_root(struct vfsmount *mnt, void *arg)
553 {
554         return mnt->mnt_root->d_inode == arg;
555 }
556
557 void audit_trim_trees(void)
558 {
559         struct list_head cursor;
560
561         mutex_lock(&audit_filter_mutex);
562         list_add(&cursor, &tree_list);
563         while (cursor.next != &tree_list) {
564                 struct audit_tree *tree;
565                 struct path path;
566                 struct vfsmount *root_mnt;
567                 struct node *node;
568                 int err;
569
570                 tree = container_of(cursor.next, struct audit_tree, list);
571                 get_tree(tree);
572                 list_del(&cursor);
573                 list_add(&cursor, &tree->list);
574                 mutex_unlock(&audit_filter_mutex);
575
576                 err = kern_path(tree->pathname, 0, &path);
577                 if (err)
578                         goto skip_it;
579
580                 root_mnt = collect_mounts(&path);
581                 path_put(&path);
582                 if (!root_mnt)
583                         goto skip_it;
584
585                 spin_lock(&hash_lock);
586                 list_for_each_entry(node, &tree->chunks, list) {
587                         struct inode *inode = find_chunk(node)->watch.inode;
588                         node->index |= 1U<<31;
589                         if (iterate_mounts(compare_root, inode, root_mnt))
590                                 node->index &= ~(1U<<31);
591                 }
592                 spin_unlock(&hash_lock);
593                 trim_marked(tree);
594                 put_tree(tree);
595                 drop_collected_mounts(root_mnt);
596 skip_it:
597                 mutex_lock(&audit_filter_mutex);
598         }
599         list_del(&cursor);
600         mutex_unlock(&audit_filter_mutex);
601 }
602
603 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
604 {
605
606         if (pathname[0] != '/' ||
607             rule->listnr != AUDIT_FILTER_EXIT ||
608             op != Audit_equal ||
609             rule->inode_f || rule->watch || rule->tree)
610                 return -EINVAL;
611         rule->tree = alloc_tree(pathname);
612         if (!rule->tree)
613                 return -ENOMEM;
614         return 0;
615 }
616
617 void audit_put_tree(struct audit_tree *tree)
618 {
619         put_tree(tree);
620 }
621
622 static int tag_mount(struct vfsmount *mnt, void *arg)
623 {
624         return tag_chunk(mnt->mnt_root->d_inode, arg);
625 }
626
627 /* called with audit_filter_mutex */
628 int audit_add_tree_rule(struct audit_krule *rule)
629 {
630         struct audit_tree *seed = rule->tree, *tree;
631         struct path path;
632         struct vfsmount *mnt;
633         int err;
634
635         list_for_each_entry(tree, &tree_list, list) {
636                 if (!strcmp(seed->pathname, tree->pathname)) {
637                         put_tree(seed);
638                         rule->tree = tree;
639                         list_add(&rule->rlist, &tree->rules);
640                         return 0;
641                 }
642         }
643         tree = seed;
644         list_add(&tree->list, &tree_list);
645         list_add(&rule->rlist, &tree->rules);
646         /* do not set rule->tree yet */
647         mutex_unlock(&audit_filter_mutex);
648
649         err = kern_path(tree->pathname, 0, &path);
650         if (err)
651                 goto Err;
652         mnt = collect_mounts(&path);
653         path_put(&path);
654         if (!mnt) {
655                 err = -ENOMEM;
656                 goto Err;
657         }
658
659         get_tree(tree);
660         err = iterate_mounts(tag_mount, tree, mnt);
661         drop_collected_mounts(mnt);
662
663         if (!err) {
664                 struct node *node;
665                 spin_lock(&hash_lock);
666                 list_for_each_entry(node, &tree->chunks, list)
667                         node->index &= ~(1U<<31);
668                 spin_unlock(&hash_lock);
669         } else {
670                 trim_marked(tree);
671                 goto Err;
672         }
673
674         mutex_lock(&audit_filter_mutex);
675         if (list_empty(&rule->rlist)) {
676                 put_tree(tree);
677                 return -ENOENT;
678         }
679         rule->tree = tree;
680         put_tree(tree);
681
682         return 0;
683 Err:
684         mutex_lock(&audit_filter_mutex);
685         list_del_init(&tree->list);
686         list_del_init(&tree->rules);
687         put_tree(tree);
688         return err;
689 }
690
691 int audit_tag_tree(char *old, char *new)
692 {
693         struct list_head cursor, barrier;
694         int failed = 0;
695         struct path path1, path2;
696         struct vfsmount *tagged;
697         int err;
698
699         err = kern_path(new, 0, &path2);
700         if (err)
701                 return err;
702         tagged = collect_mounts(&path2);
703         path_put(&path2);
704         if (!tagged)
705                 return -ENOMEM;
706
707         err = kern_path(old, 0, &path1);
708         if (err) {
709                 drop_collected_mounts(tagged);
710                 return err;
711         }
712
713         mutex_lock(&audit_filter_mutex);
714         list_add(&barrier, &tree_list);
715         list_add(&cursor, &barrier);
716
717         while (cursor.next != &tree_list) {
718                 struct audit_tree *tree;
719                 int good_one = 0;
720
721                 tree = container_of(cursor.next, struct audit_tree, list);
722                 get_tree(tree);
723                 list_del(&cursor);
724                 list_add(&cursor, &tree->list);
725                 mutex_unlock(&audit_filter_mutex);
726
727                 err = kern_path(tree->pathname, 0, &path2);
728                 if (!err) {
729                         good_one = path_is_under(&path1, &path2);
730                         path_put(&path2);
731                 }
732
733                 if (!good_one) {
734                         put_tree(tree);
735                         mutex_lock(&audit_filter_mutex);
736                         continue;
737                 }
738
739                 failed = iterate_mounts(tag_mount, tree, tagged);
740                 if (failed) {
741                         put_tree(tree);
742                         mutex_lock(&audit_filter_mutex);
743                         break;
744                 }
745
746                 mutex_lock(&audit_filter_mutex);
747                 spin_lock(&hash_lock);
748                 if (!tree->goner) {
749                         list_del(&tree->list);
750                         list_add(&tree->list, &tree_list);
751                 }
752                 spin_unlock(&hash_lock);
753                 put_tree(tree);
754         }
755
756         while (barrier.prev != &tree_list) {
757                 struct audit_tree *tree;
758
759                 tree = container_of(barrier.prev, struct audit_tree, list);
760                 get_tree(tree);
761                 list_del(&tree->list);
762                 list_add(&tree->list, &barrier);
763                 mutex_unlock(&audit_filter_mutex);
764
765                 if (!failed) {
766                         struct node *node;
767                         spin_lock(&hash_lock);
768                         list_for_each_entry(node, &tree->chunks, list)
769                                 node->index &= ~(1U<<31);
770                         spin_unlock(&hash_lock);
771                 } else {
772                         trim_marked(tree);
773                 }
774
775                 put_tree(tree);
776                 mutex_lock(&audit_filter_mutex);
777         }
778         list_del(&barrier);
779         list_del(&cursor);
780         mutex_unlock(&audit_filter_mutex);
781         path_put(&path1);
782         drop_collected_mounts(tagged);
783         return failed;
784 }
785
786 /*
787  * That gets run when evict_chunk() ends up needing to kill audit_tree.
788  * Runs from a separate thread.
789  */
790 static int prune_tree_thread(void *unused)
791 {
792         mutex_lock(&audit_cmd_mutex);
793         mutex_lock(&audit_filter_mutex);
794
795         while (!list_empty(&prune_list)) {
796                 struct audit_tree *victim;
797
798                 victim = list_entry(prune_list.next, struct audit_tree, list);
799                 list_del_init(&victim->list);
800
801                 mutex_unlock(&audit_filter_mutex);
802
803                 prune_one(victim);
804
805                 mutex_lock(&audit_filter_mutex);
806         }
807
808         mutex_unlock(&audit_filter_mutex);
809         mutex_unlock(&audit_cmd_mutex);
810         return 0;
811 }
812
813 static void audit_schedule_prune(void)
814 {
815         kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
816 }
817
818 /*
819  * ... and that one is done if evict_chunk() decides to delay until the end
820  * of syscall.  Runs synchronously.
821  */
822 void audit_kill_trees(struct list_head *list)
823 {
824         mutex_lock(&audit_cmd_mutex);
825         mutex_lock(&audit_filter_mutex);
826
827         while (!list_empty(list)) {
828                 struct audit_tree *victim;
829
830                 victim = list_entry(list->next, struct audit_tree, list);
831                 kill_rules(victim);
832                 list_del_init(&victim->list);
833
834                 mutex_unlock(&audit_filter_mutex);
835
836                 prune_one(victim);
837
838                 mutex_lock(&audit_filter_mutex);
839         }
840
841         mutex_unlock(&audit_filter_mutex);
842         mutex_unlock(&audit_cmd_mutex);
843 }
844
845 /*
846  *  Here comes the stuff asynchronous to auditctl operations
847  */
848
849 /* inode->inotify_mutex is locked */
850 static void evict_chunk(struct audit_chunk *chunk)
851 {
852         struct audit_tree *owner;
853         struct list_head *postponed = audit_killed_trees();
854         int need_prune = 0;
855         int n;
856
857         if (chunk->dead)
858                 return;
859
860         chunk->dead = 1;
861         mutex_lock(&audit_filter_mutex);
862         spin_lock(&hash_lock);
863         while (!list_empty(&chunk->trees)) {
864                 owner = list_entry(chunk->trees.next,
865                                    struct audit_tree, same_root);
866                 owner->goner = 1;
867                 owner->root = NULL;
868                 list_del_init(&owner->same_root);
869                 spin_unlock(&hash_lock);
870                 if (!postponed) {
871                         kill_rules(owner);
872                         list_move(&owner->list, &prune_list);
873                         need_prune = 1;
874                 } else {
875                         list_move(&owner->list, postponed);
876                 }
877                 spin_lock(&hash_lock);
878         }
879         list_del_rcu(&chunk->hash);
880         for (n = 0; n < chunk->count; n++)
881                 list_del_init(&chunk->owners[n].list);
882         spin_unlock(&hash_lock);
883         if (need_prune)
884                 audit_schedule_prune();
885         mutex_unlock(&audit_filter_mutex);
886 }
887
888 static void handle_event(struct inotify_watch *watch, u32 wd, u32 mask,
889                          u32 cookie, const char *dname, struct inode *inode)
890 {
891         struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
892
893         if (mask & IN_IGNORED) {
894                 evict_chunk(chunk);
895                 put_inotify_watch(watch);
896         }
897 }
898
899 static void destroy_watch(struct inotify_watch *watch)
900 {
901         struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
902         call_rcu(&chunk->head, __put_chunk);
903 }
904
905 static const struct inotify_operations rtree_inotify_ops = {
906         .handle_event   = handle_event,
907         .destroy_watch  = destroy_watch,
908 };
909
910 static int __init audit_tree_init(void)
911 {
912         int i;
913
914         rtree_ih = inotify_init(&rtree_inotify_ops);
915         if (IS_ERR(rtree_ih))
916                 audit_panic("cannot initialize inotify handle for rectree watches");
917
918         for (i = 0; i < HASH_SIZE; i++)
919                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
920
921         return 0;
922 }
923 __initcall(audit_tree_init);