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
   7 struct audit_tree;
   8 struct audit_chunk;
   9
  10 struct audit_tree {
  11         atomic_t count;
  12         int goner;
  13         struct audit_chunk *root;
  14         struct list_head chunks;
  15         struct list_head rules;
  16         struct list_head list;
  17         struct list_head same_root;
  18         struct rcu_head head;
  19         char pathname[];
  20 };
  21
  22 struct audit_chunk {
  23         struct list_head hash;
  24         struct inotify_watch watch;
  25         struct list_head trees;         /* with root here */
  26         int dead;
  27         int count;
  28         atomic_long_t refs;
  29         struct rcu_head head;
  30         struct node {
  31                 struct list_head list;
  32                 struct audit_tree *owner;
  33                 unsigned index;         /* index; upper bit indicates 'will prune' */
  34         } owners[];
  35 };
  36
  37 static LIST_HEAD(tree_list);
  38 static LIST_HEAD(prune_list);
  39
  40 /*
  41  * One struct chunk is attached to each inode of interest.
  42  * We replace struct chunk on tagging/untagging.
  43  * Rules have pointer to struct audit_tree.
  44  * Rules have struct list_head rlist forming a list of rules over
  45  * the same tree.
  46  * References to struct chunk are collected at audit_inode{,_child}()
  47  * time and used in AUDIT_TREE rule matching.
  48  * These references are dropped at the same time we are calling
  49  * audit_free_names(), etc.
  50  *
  51  * Cyclic lists galore:
  52  * tree.chunks anchors chunk.owners[].list                      hash_lock
  53  * tree.rules anchors rule.rlist                                audit_filter_mutex
  54  * chunk.trees anchors tree.same_root                           hash_lock
  55  * chunk.hash is a hash with middle bits of watch.inode as
  56  * a hash function.                                             RCU, hash_lock
  57  *
  58  * tree is refcounted; one reference for "some rules on rules_list refer to
  59  * it", one for each chunk with pointer to it.
  60  *
  61  * chunk is refcounted by embedded inotify_watch + .refs (non-zero refcount
  62  * of watch contributes 1 to .refs).
  63  *
  64  * node.index allows to get from node.list to containing chunk.
  65  * MSB of that sucker is stolen to mark taggings that we might have to
  66  * revert - several operations have very unpleasant cleanup logics and
  67  * that makes a difference.  Some.
  68  */
  69
  70 static struct inotify_handle *rtree_ih;
  71
  72 static struct audit_tree *alloc_tree(const char *s)
  73 {
  74         struct audit_tree *tree;
  75
  76         tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
  77         if (tree) {
  78                 atomic_set(&tree->count, 1);
  79                 tree->goner = 0;
  80                 INIT_LIST_HEAD(&tree->chunks);
  81                 INIT_LIST_HEAD(&tree->rules);
  82                 INIT_LIST_HEAD(&tree->list);
  83                 INIT_LIST_HEAD(&tree->same_root);
  84                 tree->root = NULL;
  85                 strcpy(tree->pathname, s);
  86         }
  87         return tree;
  88 }
  89
  90 static inline void get_tree(struct audit_tree *tree)
  91 {
  92         atomic_inc(&tree->count);
  93 }
  94
  95 static void __put_tree(struct rcu_head *rcu)
  96 {
  97         struct audit_tree *tree = container_of(rcu, struct audit_tree, head);
  98         kfree(tree);
  99 }
 100
 101 static inline void put_tree(struct audit_tree *tree)
 102 {
 103         if (atomic_dec_and_test(&tree->count))
 104                 call_rcu(&tree->head, __put_tree);
 105 }
 106
 107 /* to avoid bringing the entire thing in audit.h */
 108 const char *audit_tree_path(struct audit_tree *tree)
 109 {
 110         return tree->pathname;
 111 }
 112
 113 static struct audit_chunk *alloc_chunk(int count)
 114 {
 115         struct audit_chunk *chunk;
 116         size_t size;
 117         int i;
 118
 119         size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
 120         chunk = kzalloc(size, GFP_KERNEL);
 121         if (!chunk)
 122                 return NULL;
 123
 124         INIT_LIST_HEAD(&chunk->hash);
 125         INIT_LIST_HEAD(&chunk->trees);
 126         chunk->count = count;
 127         atomic_long_set(&chunk->refs, 1);
 128         for (i = 0; i < count; i++) {
 129                 INIT_LIST_HEAD(&chunk->owners[i].list);
 130                 chunk->owners[i].index = i;
 131         }
 132         inotify_init_watch(&chunk->watch);
 133         return chunk;
 134 }
 135
 136 static void free_chunk(struct audit_chunk *chunk)
 137 {
 138         int i;
 139
 140         for (i = 0; i < chunk->count; i++) {
 141                 if (chunk->owners[i].owner)
 142                         put_tree(chunk->owners[i].owner);
 143         }
 144         kfree(chunk);
 145 }
 146
 147 void audit_put_chunk(struct audit_chunk *chunk)
 148 {
 149         if (atomic_long_dec_and_test(&chunk->refs))
 150                 free_chunk(chunk);
 151 }
 152
 153 static void __put_chunk(struct rcu_head *rcu)
 154 {
 155         struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
 156         audit_put_chunk(chunk);
 157 }
 158
 159 enum {HASH_SIZE = 128};
 160 static struct list_head chunk_hash_heads[HASH_SIZE];
 161 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
 162
 163 static inline struct list_head *chunk_hash(const struct inode *inode)
 164 {
 165         unsigned long n = (unsigned long)inode / L1_CACHE_BYTES;
 166         return chunk_hash_heads + n % HASH_SIZE;
 167 }
 168
 169 /* hash_lock is held by caller */
 170 static void insert_hash(struct audit_chunk *chunk)
 171 {
 172         struct list_head *list = chunk_hash(chunk->watch.inode);
 173         list_add_rcu(&chunk->hash, list);
 174 }
 175
 176 /* called under rcu_read_lock */
 177 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
 178 {
 179         struct list_head *list = chunk_hash(inode);
 180         struct audit_chunk *p;
 181
 182         list_for_each_entry_rcu(p, list, hash) {
 183                 if (p->watch.inode == inode) {
 184                         atomic_long_inc(&p->refs);
 185                         return p;
 186                 }
 187         }
 188         return NULL;
 189 }
 190
 191 int audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
 192 {
 193         int n;
 194         for (n = 0; n < chunk->count; n++)
 195                 if (chunk->owners[n].owner == tree)
 196                         return 1;
 197         return 0;
 198 }
 199
 200 /* tagging and untagging inodes with trees */
 201
 202 static struct audit_chunk *find_chunk(struct node *p)
 203 {
 204         int index = p->index & ~(1U<<31);
 205         p -= index;
 206         return container_of(p, struct audit_chunk, owners[0]);
 207 }
 208
 209 static void untag_chunk(struct node *p)
 210 {
 211         struct audit_chunk *chunk = find_chunk(p);
 212         struct audit_chunk *new;
 213         struct audit_tree *owner;
 214         int size = chunk->count - 1;
 215         int i, j;
 216
 217         if (!pin_inotify_watch(&chunk->watch)) {
 218                 /*
 219                  * Filesystem is shutting down; all watches are getting
 220                  * evicted, just take it off the node list for this
 221                  * tree and let the eviction logics take care of the
 222                  * rest.
 223                  */
 224                 owner = p->owner;
 225                 if (owner->root == chunk) {
 226                         list_del_init(&owner->same_root);
 227                         owner->root = NULL;
 228                 }
 229                 list_del_init(&p->list);
 230                 p->owner = NULL;
 231                 put_tree(owner);
 232                 return;
 233         }
 234
 235         spin_unlock(&hash_lock);
 236
 237         /*
 238          * pin_inotify_watch() succeeded, so the watch won't go away
 239          * from under us.
 240          */
 241         mutex_lock(&chunk->watch.inode->inotify_mutex);
 242         if (chunk->dead) {
 243                 mutex_unlock(&chunk->watch.inode->inotify_mutex);
 244                 goto out;
 245         }
 246
 247         owner = p->owner;
 248
 249         if (!size) {
 250                 chunk->dead = 1;
 251                 spin_lock(&hash_lock);
 252                 list_del_init(&chunk->trees);
 253                 if (owner->root == chunk)
 254                         owner->root = NULL;
 255                 list_del_init(&p->list);
 256                 list_del_rcu(&chunk->hash);
 257                 spin_unlock(&hash_lock);
 258                 inotify_evict_watch(&chunk->watch);
 259                 mutex_unlock(&chunk->watch.inode->inotify_mutex);
 260                 put_inotify_watch(&chunk->watch);
 261                 goto out;
 262         }
 263
 264         new = alloc_chunk(size);
 265         if (!new)
 266                 goto Fallback;
 267         if (inotify_clone_watch(&chunk->watch, &new->watch) < 0) {
 268                 free_chunk(new);
 269                 goto Fallback;
 270         }
 271
 272         chunk->dead = 1;
 273         spin_lock(&hash_lock);
 274         list_replace_init(&chunk->trees, &new->trees);
 275         if (owner->root == chunk) {
 276                 list_del_init(&owner->same_root);
 277                 owner->root = NULL;
 278         }
 279
 280         for (i = j = 0; j <= size; i++, j++) {
 281                 struct audit_tree *s;
 282                 if (&chunk->owners[j] == p) {
 283                         list_del_init(&p->list);
 284                         i--;
 285                         continue;
 286                 }
 287                 s = chunk->owners[j].owner;
 288                 new->owners[i].owner = s;
 289                 new->owners[i].index = chunk->owners[j].index - j + i;
 290                 if (!s) /* result of earlier fallback */
 291                         continue;
 292                 get_tree(s);
 293                 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
 294         }
 295
 296         list_replace_rcu(&chunk->hash, &new->hash);
 297         list_for_each_entry(owner, &new->trees, same_root)
 298                 owner->root = new;
 299         spin_unlock(&hash_lock);
 300         inotify_evict_watch(&chunk->watch);
 301         mutex_unlock(&chunk->watch.inode->inotify_mutex);
 302         put_inotify_watch(&chunk->watch);
 303         goto out;
 304
 305 Fallback:
 306         // do the best we can
 307         spin_lock(&hash_lock);
 308         if (owner->root == chunk) {
 309                 list_del_init(&owner->same_root);
 310                 owner->root = NULL;
 311         }
 312         list_del_init(&p->list);
 313         p->owner = NULL;
 314         put_tree(owner);
 315         spin_unlock(&hash_lock);
 316         mutex_unlock(&chunk->watch.inode->inotify_mutex);
 317 out:
 318         unpin_inotify_watch(&chunk->watch);
 319         spin_lock(&hash_lock);
 320 }
 321
 322 static int create_chunk(struct inode *inode, struct audit_tree *tree)
 323 {
 324         struct audit_chunk *chunk = alloc_chunk(1);
 325         if (!chunk)
 326                 return -ENOMEM;
 327
 328         if (inotify_add_watch(rtree_ih, &chunk->watch, inode, IN_IGNORED | IN_DELETE_SELF) < 0) {
 329                 free_chunk(chunk);
 330                 return -ENOSPC;
 331         }
 332
 333         mutex_lock(&inode->inotify_mutex);
 334         spin_lock(&hash_lock);
 335         if (tree->goner) {
 336                 spin_unlock(&hash_lock);
 337                 chunk->dead = 1;
 338                 inotify_evict_watch(&chunk->watch);
 339                 mutex_unlock(&inode->inotify_mutex);
 340                 put_inotify_watch(&chunk->watch);
 341                 return 0;
 342         }
 343         chunk->owners[0].index = (1U << 31);
 344         chunk->owners[0].owner = tree;
 345         get_tree(tree);
 346         list_add(&chunk->owners[0].list, &tree->chunks);
 347         if (!tree->root) {
 348                 tree->root = chunk;
 349                 list_add(&tree->same_root, &chunk->trees);
 350         }
 351         insert_hash(chunk);
 352         spin_unlock(&hash_lock);
 353         mutex_unlock(&inode->inotify_mutex);
 354         return 0;
 355 }
 356
 357 /* the first tagged inode becomes root of tree */
 358 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
 359 {
 360         struct inotify_watch *watch;
 361         struct audit_tree *owner;
 362         struct audit_chunk *chunk, *old;
 363         struct node *p;
 364         int n;
 365
 366         if (inotify_find_watch(rtree_ih, inode, &watch) < 0)
 367                 return create_chunk(inode, tree);
 368
 369         old = container_of(watch, struct audit_chunk, watch);
 370
 371         /* are we already there? */
 372         spin_lock(&hash_lock);
 373         for (n = 0; n < old->count; n++) {
 374                 if (old->owners[n].owner == tree) {
 375                         spin_unlock(&hash_lock);
 376                         put_inotify_watch(&old->watch);
 377                         return 0;
 378                 }
 379         }
 380         spin_unlock(&hash_lock);
 381
 382         chunk = alloc_chunk(old->count + 1);
 383         if (!chunk) {
 384                 put_inotify_watch(&old->watch);
 385                 return -ENOMEM;
 386         }
 387
 388         mutex_lock(&inode->inotify_mutex);
 389         if (inotify_clone_watch(&old->watch, &chunk->watch) < 0) {
 390                 mutex_unlock(&inode->inotify_mutex);
 391                 put_inotify_watch(&old->watch);
 392                 free_chunk(chunk);
 393                 return -ENOSPC;
 394         }
 395         spin_lock(&hash_lock);
 396         if (tree->goner) {
 397                 spin_unlock(&hash_lock);
 398                 chunk->dead = 1;
 399                 inotify_evict_watch(&chunk->watch);
 400                 mutex_unlock(&inode->inotify_mutex);
 401                 put_inotify_watch(&old->watch);
 402                 put_inotify_watch(&chunk->watch);
 403                 return 0;
 404         }
 405         list_replace_init(&old->trees, &chunk->trees);
 406         for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
 407                 struct audit_tree *s = old->owners[n].owner;
 408                 p->owner = s;
 409                 p->index = old->owners[n].index;
 410                 if (!s) /* result of fallback in untag */
 411                         continue;
 412                 get_tree(s);
 413                 list_replace_init(&old->owners[n].list, &p->list);
 414         }
 415         p->index = (chunk->count - 1) | (1U<<31);
 416         p->owner = tree;
 417         get_tree(tree);
 418         list_add(&p->list, &tree->chunks);
 419         list_replace_rcu(&old->hash, &chunk->hash);
 420         list_for_each_entry(owner, &chunk->trees, same_root)
 421                 owner->root = chunk;
 422         old->dead = 1;
 423         if (!tree->root) {
 424                 tree->root = chunk;
 425                 list_add(&tree->same_root, &chunk->trees);
 426         }
 427         spin_unlock(&hash_lock);
 428         inotify_evict_watch(&old->watch);
 429         mutex_unlock(&inode->inotify_mutex);
 430         put_inotify_watch(&old->watch); /* pair to inotify_find_watch */
 431         put_inotify_watch(&old->watch); /* and kill it */
 432         return 0;
 433 }
 434
 435 static void kill_rules(struct audit_tree *tree)
 436 {
 437         struct audit_krule *rule, *next;
 438         struct audit_entry *entry;
 439         struct audit_buffer *ab;
 440
 441         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
 442                 entry = container_of(rule, struct audit_entry, rule);
 443
 444                 list_del_init(&rule->rlist);
 445                 if (rule->tree) {
 446                         /* not a half-baked one */
 447                         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
 448                         audit_log_format(ab, "op=");
 449                         audit_log_string(ab, "remove rule");
 450                         audit_log_format(ab, " dir=");
 451                         audit_log_untrustedstring(ab, rule->tree->pathname);
 452                         audit_log_key(ab, rule->filterkey);
 453                         audit_log_format(ab, " list=%d res=1", rule->listnr);
 454                         audit_log_end(ab);
 455                         rule->tree = NULL;
 456                         list_del_rcu(&entry->list);
 457                         list_del(&entry->rule.list);
 458                         call_rcu(&entry->rcu, audit_free_rule_rcu);
 459                 }
 460         }
 461 }
 462
 463 /*
 464  * finish killing struct audit_tree
 465  */
 466 static void prune_one(struct audit_tree *victim)
 467 {
 468         spin_lock(&hash_lock);
 469         while (!list_empty(&victim->chunks)) {
 470                 struct node *p;
 471
 472                 p = list_entry(victim->chunks.next, struct node, list);
 473
 474                 untag_chunk(p);
 475         }
 476         spin_unlock(&hash_lock);
 477         put_tree(victim);
 478 }
 479
 480 /* trim the uncommitted chunks from tree */
 481
 482 static void trim_marked(struct audit_tree *tree)
 483 {
 484         struct list_head *p, *q;
 485         spin_lock(&hash_lock);
 486         if (tree->goner) {
 487                 spin_unlock(&hash_lock);
 488                 return;
 489         }
 490         /* reorder */
 491         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
 492                 struct node *node = list_entry(p, struct node, list);
 493                 q = p->next;
 494                 if (node->index & (1U<<31)) {
 495                         list_del_init(p);
 496                         list_add(p, &tree->chunks);
 497                 }
 498         }
 499
 500         while (!list_empty(&tree->chunks)) {
 501                 struct node *node;
 502
 503                 node = list_entry(tree->chunks.next, struct node, list);
 504
 505                 /* have we run out of marked? */
 506                 if (!(node->index & (1U<<31)))
 507                         break;
 508
 509                 untag_chunk(node);
 510         }
 511         if (!tree->root && !tree->goner) {
 512                 tree->goner = 1;
 513                 spin_unlock(&hash_lock);
 514                 mutex_lock(&audit_filter_mutex);
 515                 kill_rules(tree);
 516                 list_del_init(&tree->list);
 517                 mutex_unlock(&audit_filter_mutex);
 518                 prune_one(tree);
 519         } else {
 520                 spin_unlock(&hash_lock);
 521         }
 522 }
 523
 524 static void audit_schedule_prune(void);
 525
 526 /* called with audit_filter_mutex */
 527 int audit_remove_tree_rule(struct audit_krule *rule)
 528 {
 529         struct audit_tree *tree;
 530         tree = rule->tree;
 531         if (tree) {
 532                 spin_lock(&hash_lock);
 533                 list_del_init(&rule->rlist);
 534                 if (list_empty(&tree->rules) && !tree->goner) {
 535                         tree->root = NULL;
 536                         list_del_init(&tree->same_root);
 537                         tree->goner = 1;
 538                         list_move(&tree->list, &prune_list);
 539                         rule->tree = NULL;
 540                         spin_unlock(&hash_lock);
 541                         audit_schedule_prune();
 542                         return 1;
 543                 }
 544                 rule->tree = NULL;
 545                 spin_unlock(&hash_lock);
 546                 return 1;
 547         }
 548         return 0;
 549 }
 550
 551 static int compare_root(struct vfsmount *mnt, void *arg)
 552 {
 553         return mnt->mnt_root->d_inode == arg;
 554 }
 555
 556 void audit_trim_trees(void)
 557 {
 558         struct list_head cursor;
 559
 560         mutex_lock(&audit_filter_mutex);
 561         list_add(&cursor, &tree_list);
 562         while (cursor.next != &tree_list) {
 563                 struct audit_tree *tree;
 564                 struct path path;
 565                 struct vfsmount *root_mnt;
 566                 struct node *node;
 567                 int err;
 568
 569                 tree = container_of(cursor.next, struct audit_tree, list);
 570                 get_tree(tree);
 571                 list_del(&cursor);
 572                 list_add(&cursor, &tree->list);
 573                 mutex_unlock(&audit_filter_mutex);
 574
 575                 err = kern_path(tree->pathname, 0, &path);
 576                 if (err)
 577                         goto skip_it;
 578
 579                 root_mnt = collect_mounts(&path);
 580                 path_put(&path);
 581                 if (!root_mnt)
 582                         goto skip_it;
 583
 584                 spin_lock(&hash_lock);
 585                 list_for_each_entry(node, &tree->chunks, list) {
 586                         struct inode *inode = find_chunk(node)->watch.inode;
 587                         node->index |= 1U<<31;
 588                         if (iterate_mounts(compare_root, inode, root_mnt))
 589                                 node->index &= ~(1U<<31);
 590                 }
 591                 spin_unlock(&hash_lock);
 592                 trim_marked(tree);
 593                 put_tree(tree);
 594                 drop_collected_mounts(root_mnt);
 595 skip_it:
 596                 mutex_lock(&audit_filter_mutex);
 597         }
 598         list_del(&cursor);
 599         mutex_unlock(&audit_filter_mutex);
 600 }
 601
 602 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
 603 {
 604
 605         if (pathname[0] != '/' ||
 606             rule->listnr != AUDIT_FILTER_EXIT ||
 607             op != Audit_equal ||
 608             rule->inode_f || rule->watch || rule->tree)
 609                 return -EINVAL;
 610         rule->tree = alloc_tree(pathname);
 611         if (!rule->tree)
 612                 return -ENOMEM;
 613         return 0;
 614 }
 615
 616 void audit_put_tree(struct audit_tree *tree)
 617 {
 618         put_tree(tree);
 619 }
 620
 621 static int tag_mount(struct vfsmount *mnt, void *arg)
 622 {
 623         return tag_chunk(mnt->mnt_root->d_inode, arg);
 624 }
 625
 626 /* called with audit_filter_mutex */
 627 int audit_add_tree_rule(struct audit_krule *rule)
 628 {
 629         struct audit_tree *seed = rule->tree, *tree;
 630         struct path path;
 631         struct vfsmount *mnt;
 632         int err;
 633
 634         list_for_each_entry(tree, &tree_list, list) {
 635                 if (!strcmp(seed->pathname, tree->pathname)) {
 636                         put_tree(seed);
 637                         rule->tree = tree;
 638                         list_add(&rule->rlist, &tree->rules);
 639                         return 0;
 640                 }
 641         }
 642         tree = seed;
 643         list_add(&tree->list, &tree_list);
 644         list_add(&rule->rlist, &tree->rules);
 645         /* do not set rule->tree yet */
 646         mutex_unlock(&audit_filter_mutex);
 647
 648         err = kern_path(tree->pathname, 0, &path);
 649         if (err)
 650                 goto Err;
 651         mnt = collect_mounts(&path);
 652         path_put(&path);
 653         if (!mnt) {
 654                 err = -ENOMEM;
 655                 goto Err;
 656         }
 657
 658         get_tree(tree);
 659         err = iterate_mounts(tag_mount, tree, mnt);
 660         drop_collected_mounts(mnt);
 661
 662         if (!err) {
 663                 struct node *node;
 664                 spin_lock(&hash_lock);
 665                 list_for_each_entry(node, &tree->chunks, list)
 666                         node->index &= ~(1U<<31);
 667                 spin_unlock(&hash_lock);
 668         } else {
 669                 trim_marked(tree);
 670                 goto Err;
 671         }
 672
 673         mutex_lock(&audit_filter_mutex);
 674         if (list_empty(&rule->rlist)) {
 675                 put_tree(tree);
 676                 return -ENOENT;
 677         }
 678         rule->tree = tree;
 679         put_tree(tree);
 680
 681         return 0;
 682 Err:
 683         mutex_lock(&audit_filter_mutex);
 684         list_del_init(&tree->list);
 685         list_del_init(&tree->rules);
 686         put_tree(tree);
 687         return err;
 688 }
 689
 690 int audit_tag_tree(char *old, char *new)
 691 {
 692         struct list_head cursor, barrier;
 693         int failed = 0;
 694         struct path path1, path2;
 695         struct vfsmount *tagged;
 696         int err;
 697
 698         err = kern_path(new, 0, &path2);
 699         if (err)
 700                 return err;
 701         tagged = collect_mounts(&path2);
 702         path_put(&path2);
 703         if (!tagged)
 704                 return -ENOMEM;
 705
 706         err = kern_path(old, 0, &path1);
 707         if (err) {
 708                 drop_collected_mounts(tagged);
 709                 return err;
 710         }
 711
 712         mutex_lock(&audit_filter_mutex);
 713         list_add(&barrier, &tree_list);
 714         list_add(&cursor, &barrier);
 715
 716         while (cursor.next != &tree_list) {
 717                 struct audit_tree *tree;
 718                 int good_one = 0;
 719
 720                 tree = container_of(cursor.next, struct audit_tree, list);
 721                 get_tree(tree);
 722                 list_del(&cursor);
 723                 list_add(&cursor, &tree->list);
 724                 mutex_unlock(&audit_filter_mutex);
 725
 726                 err = kern_path(tree->pathname, 0, &path2);
 727                 if (!err) {
 728                         good_one = path_is_under(&path1, &path2);
 729                         path_put(&path2);
 730                 }
 731
 732                 if (!good_one) {
 733                         put_tree(tree);
 734                         mutex_lock(&audit_filter_mutex);
 735                         continue;
 736                 }
 737
 738                 failed = iterate_mounts(tag_mount, tree, tagged);
 739                 if (failed) {
 740                         put_tree(tree);
 741                         mutex_lock(&audit_filter_mutex);
 742                         break;
 743                 }
 744
 745                 mutex_lock(&audit_filter_mutex);
 746                 spin_lock(&hash_lock);
 747                 if (!tree->goner) {
 748                         list_del(&tree->list);
 749                         list_add(&tree->list, &tree_list);
 750                 }
 751                 spin_unlock(&hash_lock);
 752                 put_tree(tree);
 753         }
 754
 755         while (barrier.prev != &tree_list) {
 756                 struct audit_tree *tree;
 757
 758                 tree = container_of(barrier.prev, struct audit_tree, list);
 759                 get_tree(tree);
 760                 list_del(&tree->list);
 761                 list_add(&tree->list, &barrier);
 762                 mutex_unlock(&audit_filter_mutex);
 763
 764                 if (!failed) {
 765                         struct node *node;
 766                         spin_lock(&hash_lock);
 767                         list_for_each_entry(node, &tree->chunks, list)
 768                                 node->index &= ~(1U<<31);
 769                         spin_unlock(&hash_lock);
 770                 } else {
 771                         trim_marked(tree);
 772                 }
 773
 774                 put_tree(tree);
 775                 mutex_lock(&audit_filter_mutex);
 776         }
 777         list_del(&barrier);
 778         list_del(&cursor);
 779         mutex_unlock(&audit_filter_mutex);
 780         path_put(&path1);
 781         drop_collected_mounts(tagged);
 782         return failed;
 783 }
 784
 785 /*
 786  * That gets run when evict_chunk() ends up needing to kill audit_tree.
 787  * Runs from a separate thread.
 788  */
 789 static int prune_tree_thread(void *unused)
 790 {
 791         mutex_lock(&audit_cmd_mutex);
 792         mutex_lock(&audit_filter_mutex);
 793
 794         while (!list_empty(&prune_list)) {
 795                 struct audit_tree *victim;
 796
 797                 victim = list_entry(prune_list.next, struct audit_tree, list);
 798                 list_del_init(&victim->list);
 799
 800                 mutex_unlock(&audit_filter_mutex);
 801
 802                 prune_one(victim);
 803
 804                 mutex_lock(&audit_filter_mutex);
 805         }
 806
 807         mutex_unlock(&audit_filter_mutex);
 808         mutex_unlock(&audit_cmd_mutex);
 809         return 0;
 810 }
 811
 812 static void audit_schedule_prune(void)
 813 {
 814         kthread_run(prune_tree_thread, NULL, "audit_prune_tree");
 815 }
 816
 817 /*
 818  * ... and that one is done if evict_chunk() decides to delay until the end
 819  * of syscall.  Runs synchronously.
 820  */
 821 void audit_kill_trees(struct list_head *list)
 822 {
 823         mutex_lock(&audit_cmd_mutex);
 824         mutex_lock(&audit_filter_mutex);
 825
 826         while (!list_empty(list)) {
 827                 struct audit_tree *victim;
 828
 829                 victim = list_entry(list->next, struct audit_tree, list);
 830                 kill_rules(victim);
 831                 list_del_init(&victim->list);
 832
 833                 mutex_unlock(&audit_filter_mutex);
 834
 835                 prune_one(victim);
 836
 837                 mutex_lock(&audit_filter_mutex);
 838         }
 839
 840         mutex_unlock(&audit_filter_mutex);
 841         mutex_unlock(&audit_cmd_mutex);
 842 }
 843
 844 /*
 845  *  Here comes the stuff asynchronous to auditctl operations
 846  */
 847
 848 /* inode->inotify_mutex is locked */
 849 static void evict_chunk(struct audit_chunk *chunk)
 850 {
 851         struct audit_tree *owner;
 852         struct list_head *postponed = audit_killed_trees();
 853         int need_prune = 0;
 854         int n;
 855
 856         if (chunk->dead)
 857                 return;
 858
 859         chunk->dead = 1;
 860         mutex_lock(&audit_filter_mutex);
 861         spin_lock(&hash_lock);
 862         while (!list_empty(&chunk->trees)) {
 863                 owner = list_entry(chunk->trees.next,
 864                                    struct audit_tree, same_root);
 865                 owner->goner = 1;
 866                 owner->root = NULL;
 867                 list_del_init(&owner->same_root);
 868                 spin_unlock(&hash_lock);
 869                 if (!postponed) {
 870                         kill_rules(owner);
 871                         list_move(&owner->list, &prune_list);
 872                         need_prune = 1;
 873                 } else {
 874                         list_move(&owner->list, postponed);
 875                 }
 876                 spin_lock(&hash_lock);
 877         }
 878         list_del_rcu(&chunk->hash);
 879         for (n = 0; n < chunk->count; n++)
 880                 list_del_init(&chunk->owners[n].list);
 881         spin_unlock(&hash_lock);
 882         if (need_prune)
 883                 audit_schedule_prune();
 884         mutex_unlock(&audit_filter_mutex);
 885 }
 886
 887 static void handle_event(struct inotify_watch *watch, u32 wd, u32 mask,
 888                          u32 cookie, const char *dname, struct inode *inode)
 889 {
 890         struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
 891
 892         if (mask & IN_IGNORED) {
 893                 evict_chunk(chunk);
 894                 put_inotify_watch(watch);
 895         }
 896 }
 897
 898 static void destroy_watch(struct inotify_watch *watch)
 899 {
 900         struct audit_chunk *chunk = container_of(watch, struct audit_chunk, watch);
 901         call_rcu(&chunk->head, __put_chunk);
 902 }
 903
 904 static const struct inotify_operations rtree_inotify_ops = {
 905         .handle_event   = handle_event,
 906         .destroy_watch  = destroy_watch,
 907 };
 908
 909 static int __init audit_tree_init(void)
 910 {
 911         int i;
 912
 913         rtree_ih = inotify_init(&rtree_inotify_ops);
 914         if (IS_ERR(rtree_ih))
 915                 audit_panic("cannot initialize inotify handle for rectree watches");
 916
 917         for (i = 0; i < HASH_SIZE; i++)
 918                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
 919
 920         return 0;
 921 }
 922 __initcall(audit_tree_init);