kernel/audit_tree.c

   1 #include "audit.h"
   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>
   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 fsnotify_mark mark;
  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 static struct task_struct *prune_thread;
  41
  42 /*
  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
  47  * the same tree.
  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.
  52  *
  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
  59  *
  60  * tree is refcounted; one reference for "some rules on rules_list refer to
  61  * it", one for each chunk with pointer to it.
  62  *
  63  * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
  64  * of watch contributes 1 to .refs).
  65  *
  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.
  70  */
  71
  72 static struct fsnotify_group *audit_tree_group;
  73
  74 static struct audit_tree *alloc_tree(const char *s)
  75 {
  76         struct audit_tree *tree;
  77
  78         tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
  79         if (tree) {
  80                 atomic_set(&tree->count, 1);
  81                 tree->goner = 0;
  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);
  86                 tree->root = NULL;
  87                 strcpy(tree->pathname, s);
  88         }
  89         return tree;
  90 }
  91
  92 static inline void get_tree(struct audit_tree *tree)
  93 {
  94         atomic_inc(&tree->count);
  95 }
  96
  97 static inline void put_tree(struct audit_tree *tree)
  98 {
  99         if (atomic_dec_and_test(&tree->count))
 100                 kfree_rcu(tree, head);
 101 }
 102
 103 /* to avoid bringing the entire thing in audit.h */
 104 const char *audit_tree_path(struct audit_tree *tree)
 105 {
 106         return tree->pathname;
 107 }
 108
 109 static void free_chunk(struct audit_chunk *chunk)
 110 {
 111         int i;
 112
 113         for (i = 0; i < chunk->count; i++) {
 114                 if (chunk->owners[i].owner)
 115                         put_tree(chunk->owners[i].owner);
 116         }
 117         kfree(chunk);
 118 }
 119
 120 void audit_put_chunk(struct audit_chunk *chunk)
 121 {
 122         if (atomic_long_dec_and_test(&chunk->refs))
 123                 free_chunk(chunk);
 124 }
 125
 126 static void __put_chunk(struct rcu_head *rcu)
 127 {
 128         struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
 129         audit_put_chunk(chunk);
 130 }
 131
 132 static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
 133 {
 134         struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
 135         call_rcu(&chunk->head, __put_chunk);
 136 }
 137
 138 static struct audit_chunk *alloc_chunk(int count)
 139 {
 140         struct audit_chunk *chunk;
 141         size_t size;
 142         int i;
 143
 144         size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
 145         chunk = kzalloc(size, GFP_KERNEL);
 146         if (!chunk)
 147                 return NULL;
 148
 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;
 156         }
 157         fsnotify_init_mark(&chunk->mark, audit_tree_destroy_watch);
 158         chunk->mark.mask = FS_IN_IGNORED;
 159         return chunk;
 160 }
 161
 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);
 165
 166 /* Function to return search key in our hash from inode. */
 167 static unsigned long inode_to_key(const struct inode *inode)
 168 {
 169         return (unsigned long)inode;
 170 }
 171
 172 /*
 173  * Function to return search key in our hash from chunk. Key 0 is special and
 174  * should never be present in the hash.
 175  */
 176 static unsigned long chunk_to_key(struct audit_chunk *chunk)
 177 {
 178         /*
 179          * We have a reference to the mark so it should be attached to a
 180          * connector.
 181          */
 182         if (WARN_ON_ONCE(!chunk->mark.connector))
 183                 return 0;
 184         return (unsigned long)chunk->mark.connector->inode;
 185 }
 186
 187 static inline struct list_head *chunk_hash(unsigned long key)
 188 {
 189         unsigned long n = key / L1_CACHE_BYTES;
 190         return chunk_hash_heads + n % HASH_SIZE;
 191 }
 192
 193 /* hash_lock & entry->lock is held by caller */
 194 static void insert_hash(struct audit_chunk *chunk)
 195 {
 196         unsigned long key = chunk_to_key(chunk);
 197         struct list_head *list;
 198
 199         if (!(chunk->mark.flags & FSNOTIFY_MARK_FLAG_ATTACHED))
 200                 return;
 201         list = chunk_hash(key);
 202         list_add_rcu(&chunk->hash, list);
 203 }
 204
 205 /* called under rcu_read_lock */
 206 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
 207 {
 208         unsigned long key = inode_to_key(inode);
 209         struct list_head *list = chunk_hash(key);
 210         struct audit_chunk *p;
 211
 212         list_for_each_entry_rcu(p, list, hash) {
 213                 if (chunk_to_key(p) == key) {
 214                         atomic_long_inc(&p->refs);
 215                         return p;
 216                 }
 217         }
 218         return NULL;
 219 }
 220
 221 bool audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
 222 {
 223         int n;
 224         for (n = 0; n < chunk->count; n++)
 225                 if (chunk->owners[n].owner == tree)
 226                         return true;
 227         return false;
 228 }
 229
 230 /* tagging and untagging inodes with trees */
 231
 232 static struct audit_chunk *find_chunk(struct node *p)
 233 {
 234         int index = p->index & ~(1U<<31);
 235         p -= index;
 236         return container_of(p, struct audit_chunk, owners[0]);
 237 }
 238
 239 static void untag_chunk(struct node *p)
 240 {
 241         struct audit_chunk *chunk = find_chunk(p);
 242         struct fsnotify_mark *entry = &chunk->mark;
 243         struct audit_chunk *new = NULL;
 244         struct audit_tree *owner;
 245         int size = chunk->count - 1;
 246         int i, j;
 247
 248         fsnotify_get_mark(entry);
 249
 250         spin_unlock(&hash_lock);
 251
 252         if (size)
 253                 new = alloc_chunk(size);
 254
 255         mutex_lock(&entry->group->mark_mutex);
 256         spin_lock(&entry->lock);
 257         /*
 258          * mark_mutex protects mark from getting detached and thus also from
 259          * mark->connector->inode getting NULL.
 260          */
 261         if (chunk->dead || !(entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 262                 spin_unlock(&entry->lock);
 263                 mutex_unlock(&entry->group->mark_mutex);
 264                 if (new)
 265                         free_chunk(new);
 266                 goto out;
 267         }
 268
 269         owner = p->owner;
 270
 271         if (!size) {
 272                 chunk->dead = 1;
 273                 spin_lock(&hash_lock);
 274                 list_del_init(&chunk->trees);
 275                 if (owner->root == chunk)
 276                         owner->root = NULL;
 277                 list_del_init(&p->list);
 278                 list_del_rcu(&chunk->hash);
 279                 spin_unlock(&hash_lock);
 280                 spin_unlock(&entry->lock);
 281                 mutex_unlock(&entry->group->mark_mutex);
 282                 fsnotify_destroy_mark(entry, audit_tree_group);
 283                 goto out;
 284         }
 285
 286         if (!new)
 287                 goto Fallback;
 288
 289         if (fsnotify_add_mark_locked(&new->mark, entry->group,
 290                                      entry->connector->inode, NULL, 1)) {
 291                 fsnotify_put_mark(&new->mark);
 292                 goto Fallback;
 293         }
 294
 295         chunk->dead = 1;
 296         spin_lock(&hash_lock);
 297         list_replace_init(&chunk->trees, &new->trees);
 298         if (owner->root == chunk) {
 299                 list_del_init(&owner->same_root);
 300                 owner->root = NULL;
 301         }
 302
 303         for (i = j = 0; j <= size; i++, j++) {
 304                 struct audit_tree *s;
 305                 if (&chunk->owners[j] == p) {
 306                         list_del_init(&p->list);
 307                         i--;
 308                         continue;
 309                 }
 310                 s = chunk->owners[j].owner;
 311                 new->owners[i].owner = s;
 312                 new->owners[i].index = chunk->owners[j].index - j + i;
 313                 if (!s) /* result of earlier fallback */
 314                         continue;
 315                 get_tree(s);
 316                 list_replace_init(&chunk->owners[j].list, &new->owners[i].list);
 317         }
 318
 319         list_replace_rcu(&chunk->hash, &new->hash);
 320         list_for_each_entry(owner, &new->trees, same_root)
 321                 owner->root = new;
 322         spin_unlock(&hash_lock);
 323         spin_unlock(&entry->lock);
 324         mutex_unlock(&entry->group->mark_mutex);
 325         fsnotify_destroy_mark(entry, audit_tree_group);
 326         fsnotify_put_mark(&new->mark);  /* drop initial reference */
 327         goto out;
 328
 329 Fallback:
 330         // do the best we can
 331         spin_lock(&hash_lock);
 332         if (owner->root == chunk) {
 333                 list_del_init(&owner->same_root);
 334                 owner->root = NULL;
 335         }
 336         list_del_init(&p->list);
 337         p->owner = NULL;
 338         put_tree(owner);
 339         spin_unlock(&hash_lock);
 340         spin_unlock(&entry->lock);
 341         mutex_unlock(&entry->group->mark_mutex);
 342 out:
 343         fsnotify_put_mark(entry);
 344         spin_lock(&hash_lock);
 345 }
 346
 347 static int create_chunk(struct inode *inode, struct audit_tree *tree)
 348 {
 349         struct fsnotify_mark *entry;
 350         struct audit_chunk *chunk = alloc_chunk(1);
 351         if (!chunk)
 352                 return -ENOMEM;
 353
 354         entry = &chunk->mark;
 355         if (fsnotify_add_mark(entry, audit_tree_group, inode, NULL, 0)) {
 356                 fsnotify_put_mark(entry);
 357                 return -ENOSPC;
 358         }
 359
 360         spin_lock(&entry->lock);
 361         spin_lock(&hash_lock);
 362         if (tree->goner) {
 363                 spin_unlock(&hash_lock);
 364                 chunk->dead = 1;
 365                 spin_unlock(&entry->lock);
 366                 fsnotify_destroy_mark(entry, audit_tree_group);
 367                 fsnotify_put_mark(entry);
 368                 return 0;
 369         }
 370         chunk->owners[0].index = (1U << 31);
 371         chunk->owners[0].owner = tree;
 372         get_tree(tree);
 373         list_add(&chunk->owners[0].list, &tree->chunks);
 374         if (!tree->root) {
 375                 tree->root = chunk;
 376                 list_add(&tree->same_root, &chunk->trees);
 377         }
 378         insert_hash(chunk);
 379         spin_unlock(&hash_lock);
 380         spin_unlock(&entry->lock);
 381         fsnotify_put_mark(entry);       /* drop initial reference */
 382         return 0;
 383 }
 384
 385 /* the first tagged inode becomes root of tree */
 386 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
 387 {
 388         struct fsnotify_mark *old_entry, *chunk_entry;
 389         struct audit_tree *owner;
 390         struct audit_chunk *chunk, *old;
 391         struct node *p;
 392         int n;
 393
 394         old_entry = fsnotify_find_inode_mark(audit_tree_group, inode);
 395         if (!old_entry)
 396                 return create_chunk(inode, tree);
 397
 398         old = container_of(old_entry, struct audit_chunk, mark);
 399
 400         /* are we already there? */
 401         spin_lock(&hash_lock);
 402         for (n = 0; n < old->count; n++) {
 403                 if (old->owners[n].owner == tree) {
 404                         spin_unlock(&hash_lock);
 405                         fsnotify_put_mark(old_entry);
 406                         return 0;
 407                 }
 408         }
 409         spin_unlock(&hash_lock);
 410
 411         chunk = alloc_chunk(old->count + 1);
 412         if (!chunk) {
 413                 fsnotify_put_mark(old_entry);
 414                 return -ENOMEM;
 415         }
 416
 417         chunk_entry = &chunk->mark;
 418
 419         mutex_lock(&old_entry->group->mark_mutex);
 420         spin_lock(&old_entry->lock);
 421         /*
 422          * mark_mutex protects mark from getting detached and thus also from
 423          * mark->connector->inode getting NULL.
 424          */
 425         if (!(old_entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 426                 /* old_entry is being shot, lets just lie */
 427                 spin_unlock(&old_entry->lock);
 428                 mutex_unlock(&old_entry->group->mark_mutex);
 429                 fsnotify_put_mark(old_entry);
 430                 free_chunk(chunk);
 431                 return -ENOENT;
 432         }
 433
 434         if (fsnotify_add_mark_locked(chunk_entry, old_entry->group,
 435                              old_entry->connector->inode, NULL, 1)) {
 436                 spin_unlock(&old_entry->lock);
 437                 mutex_unlock(&old_entry->group->mark_mutex);
 438                 fsnotify_put_mark(chunk_entry);
 439                 fsnotify_put_mark(old_entry);
 440                 return -ENOSPC;
 441         }
 442
 443         /* even though we hold old_entry->lock, this is safe since chunk_entry->lock could NEVER have been grabbed before */
 444         spin_lock(&chunk_entry->lock);
 445         spin_lock(&hash_lock);
 446
 447         /* we now hold old_entry->lock, chunk_entry->lock, and hash_lock */
 448         if (tree->goner) {
 449                 spin_unlock(&hash_lock);
 450                 chunk->dead = 1;
 451                 spin_unlock(&chunk_entry->lock);
 452                 spin_unlock(&old_entry->lock);
 453                 mutex_unlock(&old_entry->group->mark_mutex);
 454
 455                 fsnotify_destroy_mark(chunk_entry, audit_tree_group);
 456
 457                 fsnotify_put_mark(chunk_entry);
 458                 fsnotify_put_mark(old_entry);
 459                 return 0;
 460         }
 461         list_replace_init(&old->trees, &chunk->trees);
 462         for (n = 0, p = chunk->owners; n < old->count; n++, p++) {
 463                 struct audit_tree *s = old->owners[n].owner;
 464                 p->owner = s;
 465                 p->index = old->owners[n].index;
 466                 if (!s) /* result of fallback in untag */
 467                         continue;
 468                 get_tree(s);
 469                 list_replace_init(&old->owners[n].list, &p->list);
 470         }
 471         p->index = (chunk->count - 1) | (1U<<31);
 472         p->owner = tree;
 473         get_tree(tree);
 474         list_add(&p->list, &tree->chunks);
 475         list_replace_rcu(&old->hash, &chunk->hash);
 476         list_for_each_entry(owner, &chunk->trees, same_root)
 477                 owner->root = chunk;
 478         old->dead = 1;
 479         if (!tree->root) {
 480                 tree->root = chunk;
 481                 list_add(&tree->same_root, &chunk->trees);
 482         }
 483         spin_unlock(&hash_lock);
 484         spin_unlock(&chunk_entry->lock);
 485         spin_unlock(&old_entry->lock);
 486         mutex_unlock(&old_entry->group->mark_mutex);
 487         fsnotify_destroy_mark(old_entry, audit_tree_group);
 488         fsnotify_put_mark(chunk_entry); /* drop initial reference */
 489         fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
 490         return 0;
 491 }
 492
 493 static void audit_tree_log_remove_rule(struct audit_krule *rule)
 494 {
 495         struct audit_buffer *ab;
 496
 497         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
 498         if (unlikely(!ab))
 499                 return;
 500         audit_log_format(ab, "op=remove_rule");
 501         audit_log_format(ab, " dir=");
 502         audit_log_untrustedstring(ab, rule->tree->pathname);
 503         audit_log_key(ab, rule->filterkey);
 504         audit_log_format(ab, " list=%d res=1", rule->listnr);
 505         audit_log_end(ab);
 506 }
 507
 508 static void kill_rules(struct audit_tree *tree)
 509 {
 510         struct audit_krule *rule, *next;
 511         struct audit_entry *entry;
 512
 513         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
 514                 entry = container_of(rule, struct audit_entry, rule);
 515
 516                 list_del_init(&rule->rlist);
 517                 if (rule->tree) {
 518                         /* not a half-baked one */
 519                         audit_tree_log_remove_rule(rule);
 520                         if (entry->rule.exe)
 521                                 audit_remove_mark(entry->rule.exe);
 522                         rule->tree = NULL;
 523                         list_del_rcu(&entry->list);
 524                         list_del(&entry->rule.list);
 525                         call_rcu(&entry->rcu, audit_free_rule_rcu);
 526                 }
 527         }
 528 }
 529
 530 /*
 531  * finish killing struct audit_tree
 532  */
 533 static void prune_one(struct audit_tree *victim)
 534 {
 535         spin_lock(&hash_lock);
 536         while (!list_empty(&victim->chunks)) {
 537                 struct node *p;
 538
 539                 p = list_entry(victim->chunks.next, struct node, list);
 540
 541                 untag_chunk(p);
 542         }
 543         spin_unlock(&hash_lock);
 544         put_tree(victim);
 545 }
 546
 547 /* trim the uncommitted chunks from tree */
 548
 549 static void trim_marked(struct audit_tree *tree)
 550 {
 551         struct list_head *p, *q;
 552         spin_lock(&hash_lock);
 553         if (tree->goner) {
 554                 spin_unlock(&hash_lock);
 555                 return;
 556         }
 557         /* reorder */
 558         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
 559                 struct node *node = list_entry(p, struct node, list);
 560                 q = p->next;
 561                 if (node->index & (1U<<31)) {
 562                         list_del_init(p);
 563                         list_add(p, &tree->chunks);
 564                 }
 565         }
 566
 567         while (!list_empty(&tree->chunks)) {
 568                 struct node *node;
 569
 570                 node = list_entry(tree->chunks.next, struct node, list);
 571
 572                 /* have we run out of marked? */
 573                 if (!(node->index & (1U<<31)))
 574                         break;
 575
 576                 untag_chunk(node);
 577         }
 578         if (!tree->root && !tree->goner) {
 579                 tree->goner = 1;
 580                 spin_unlock(&hash_lock);
 581                 mutex_lock(&audit_filter_mutex);
 582                 kill_rules(tree);
 583                 list_del_init(&tree->list);
 584                 mutex_unlock(&audit_filter_mutex);
 585                 prune_one(tree);
 586         } else {
 587                 spin_unlock(&hash_lock);
 588         }
 589 }
 590
 591 static void audit_schedule_prune(void);
 592
 593 /* called with audit_filter_mutex */
 594 int audit_remove_tree_rule(struct audit_krule *rule)
 595 {
 596         struct audit_tree *tree;
 597         tree = rule->tree;
 598         if (tree) {
 599                 spin_lock(&hash_lock);
 600                 list_del_init(&rule->rlist);
 601                 if (list_empty(&tree->rules) && !tree->goner) {
 602                         tree->root = NULL;
 603                         list_del_init(&tree->same_root);
 604                         tree->goner = 1;
 605                         list_move(&tree->list, &prune_list);
 606                         rule->tree = NULL;
 607                         spin_unlock(&hash_lock);
 608                         audit_schedule_prune();
 609                         return 1;
 610                 }
 611                 rule->tree = NULL;
 612                 spin_unlock(&hash_lock);
 613                 return 1;
 614         }
 615         return 0;
 616 }
 617
 618 static int compare_root(struct vfsmount *mnt, void *arg)
 619 {
 620         return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
 621                (unsigned long)arg;
 622 }
 623
 624 void audit_trim_trees(void)
 625 {
 626         struct list_head cursor;
 627
 628         mutex_lock(&audit_filter_mutex);
 629         list_add(&cursor, &tree_list);
 630         while (cursor.next != &tree_list) {
 631                 struct audit_tree *tree;
 632                 struct path path;
 633                 struct vfsmount *root_mnt;
 634                 struct node *node;
 635                 int err;
 636
 637                 tree = container_of(cursor.next, struct audit_tree, list);
 638                 get_tree(tree);
 639                 list_del(&cursor);
 640                 list_add(&cursor, &tree->list);
 641                 mutex_unlock(&audit_filter_mutex);
 642
 643                 err = kern_path(tree->pathname, 0, &path);
 644                 if (err)
 645                         goto skip_it;
 646
 647                 root_mnt = collect_mounts(&path);
 648                 path_put(&path);
 649                 if (IS_ERR(root_mnt))
 650                         goto skip_it;
 651
 652                 spin_lock(&hash_lock);
 653                 list_for_each_entry(node, &tree->chunks, list) {
 654                         struct audit_chunk *chunk = find_chunk(node);
 655                         /* this could be NULL if the watch is dying else where... */
 656                         node->index |= 1U<<31;
 657                         if (iterate_mounts(compare_root,
 658                                            (void *)chunk_to_key(chunk),
 659                                            root_mnt))
 660                                 node->index &= ~(1U<<31);
 661                 }
 662                 spin_unlock(&hash_lock);
 663                 trim_marked(tree);
 664                 drop_collected_mounts(root_mnt);
 665 skip_it:
 666                 put_tree(tree);
 667                 mutex_lock(&audit_filter_mutex);
 668         }
 669         list_del(&cursor);
 670         mutex_unlock(&audit_filter_mutex);
 671 }
 672
 673 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
 674 {
 675
 676         if (pathname[0] != '/' ||
 677             rule->listnr != AUDIT_FILTER_EXIT ||
 678             op != Audit_equal ||
 679             rule->inode_f || rule->watch || rule->tree)
 680                 return -EINVAL;
 681         rule->tree = alloc_tree(pathname);
 682         if (!rule->tree)
 683                 return -ENOMEM;
 684         return 0;
 685 }
 686
 687 void audit_put_tree(struct audit_tree *tree)
 688 {
 689         put_tree(tree);
 690 }
 691
 692 static int tag_mount(struct vfsmount *mnt, void *arg)
 693 {
 694         return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
 695 }
 696
 697 /*
 698  * That gets run when evict_chunk() ends up needing to kill audit_tree.
 699  * Runs from a separate thread.
 700  */
 701 static int prune_tree_thread(void *unused)
 702 {
 703         for (;;) {
 704                 if (list_empty(&prune_list)) {
 705                         set_current_state(TASK_INTERRUPTIBLE);
 706                         schedule();
 707                 }
 708
 709                 mutex_lock(&audit_cmd_mutex);
 710                 mutex_lock(&audit_filter_mutex);
 711
 712                 while (!list_empty(&prune_list)) {
 713                         struct audit_tree *victim;
 714
 715                         victim = list_entry(prune_list.next,
 716                                         struct audit_tree, list);
 717                         list_del_init(&victim->list);
 718
 719                         mutex_unlock(&audit_filter_mutex);
 720
 721                         prune_one(victim);
 722
 723                         mutex_lock(&audit_filter_mutex);
 724                 }
 725
 726                 mutex_unlock(&audit_filter_mutex);
 727                 mutex_unlock(&audit_cmd_mutex);
 728         }
 729         return 0;
 730 }
 731
 732 static int audit_launch_prune(void)
 733 {
 734         if (prune_thread)
 735                 return 0;
 736         prune_thread = kthread_run(prune_tree_thread, NULL,
 737                                 "audit_prune_tree");
 738         if (IS_ERR(prune_thread)) {
 739                 pr_err("cannot start thread audit_prune_tree");
 740                 prune_thread = NULL;
 741                 return -ENOMEM;
 742         }
 743         return 0;
 744 }
 745
 746 /* called with audit_filter_mutex */
 747 int audit_add_tree_rule(struct audit_krule *rule)
 748 {
 749         struct audit_tree *seed = rule->tree, *tree;
 750         struct path path;
 751         struct vfsmount *mnt;
 752         int err;
 753
 754         rule->tree = NULL;
 755         list_for_each_entry(tree, &tree_list, list) {
 756                 if (!strcmp(seed->pathname, tree->pathname)) {
 757                         put_tree(seed);
 758                         rule->tree = tree;
 759                         list_add(&rule->rlist, &tree->rules);
 760                         return 0;
 761                 }
 762         }
 763         tree = seed;
 764         list_add(&tree->list, &tree_list);
 765         list_add(&rule->rlist, &tree->rules);
 766         /* do not set rule->tree yet */
 767         mutex_unlock(&audit_filter_mutex);
 768
 769         if (unlikely(!prune_thread)) {
 770                 err = audit_launch_prune();
 771                 if (err)
 772                         goto Err;
 773         }
 774
 775         err = kern_path(tree->pathname, 0, &path);
 776         if (err)
 777                 goto Err;
 778         mnt = collect_mounts(&path);
 779         path_put(&path);
 780         if (IS_ERR(mnt)) {
 781                 err = PTR_ERR(mnt);
 782                 goto Err;
 783         }
 784
 785         get_tree(tree);
 786         err = iterate_mounts(tag_mount, tree, mnt);
 787         drop_collected_mounts(mnt);
 788
 789         if (!err) {
 790                 struct node *node;
 791                 spin_lock(&hash_lock);
 792                 list_for_each_entry(node, &tree->chunks, list)
 793                         node->index &= ~(1U<<31);
 794                 spin_unlock(&hash_lock);
 795         } else {
 796                 trim_marked(tree);
 797                 goto Err;
 798         }
 799
 800         mutex_lock(&audit_filter_mutex);
 801         if (list_empty(&rule->rlist)) {
 802                 put_tree(tree);
 803                 return -ENOENT;
 804         }
 805         rule->tree = tree;
 806         put_tree(tree);
 807
 808         return 0;
 809 Err:
 810         mutex_lock(&audit_filter_mutex);
 811         list_del_init(&tree->list);
 812         list_del_init(&tree->rules);
 813         put_tree(tree);
 814         return err;
 815 }
 816
 817 int audit_tag_tree(char *old, char *new)
 818 {
 819         struct list_head cursor, barrier;
 820         int failed = 0;
 821         struct path path1, path2;
 822         struct vfsmount *tagged;
 823         int err;
 824
 825         err = kern_path(new, 0, &path2);
 826         if (err)
 827                 return err;
 828         tagged = collect_mounts(&path2);
 829         path_put(&path2);
 830         if (IS_ERR(tagged))
 831                 return PTR_ERR(tagged);
 832
 833         err = kern_path(old, 0, &path1);
 834         if (err) {
 835                 drop_collected_mounts(tagged);
 836                 return err;
 837         }
 838
 839         mutex_lock(&audit_filter_mutex);
 840         list_add(&barrier, &tree_list);
 841         list_add(&cursor, &barrier);
 842
 843         while (cursor.next != &tree_list) {
 844                 struct audit_tree *tree;
 845                 int good_one = 0;
 846
 847                 tree = container_of(cursor.next, struct audit_tree, list);
 848                 get_tree(tree);
 849                 list_del(&cursor);
 850                 list_add(&cursor, &tree->list);
 851                 mutex_unlock(&audit_filter_mutex);
 852
 853                 err = kern_path(tree->pathname, 0, &path2);
 854                 if (!err) {
 855                         good_one = path_is_under(&path1, &path2);
 856                         path_put(&path2);
 857                 }
 858
 859                 if (!good_one) {
 860                         put_tree(tree);
 861                         mutex_lock(&audit_filter_mutex);
 862                         continue;
 863                 }
 864
 865                 failed = iterate_mounts(tag_mount, tree, tagged);
 866                 if (failed) {
 867                         put_tree(tree);
 868                         mutex_lock(&audit_filter_mutex);
 869                         break;
 870                 }
 871
 872                 mutex_lock(&audit_filter_mutex);
 873                 spin_lock(&hash_lock);
 874                 if (!tree->goner) {
 875                         list_del(&tree->list);
 876                         list_add(&tree->list, &tree_list);
 877                 }
 878                 spin_unlock(&hash_lock);
 879                 put_tree(tree);
 880         }
 881
 882         while (barrier.prev != &tree_list) {
 883                 struct audit_tree *tree;
 884
 885                 tree = container_of(barrier.prev, struct audit_tree, list);
 886                 get_tree(tree);
 887                 list_del(&tree->list);
 888                 list_add(&tree->list, &barrier);
 889                 mutex_unlock(&audit_filter_mutex);
 890
 891                 if (!failed) {
 892                         struct node *node;
 893                         spin_lock(&hash_lock);
 894                         list_for_each_entry(node, &tree->chunks, list)
 895                                 node->index &= ~(1U<<31);
 896                         spin_unlock(&hash_lock);
 897                 } else {
 898                         trim_marked(tree);
 899                 }
 900
 901                 put_tree(tree);
 902                 mutex_lock(&audit_filter_mutex);
 903         }
 904         list_del(&barrier);
 905         list_del(&cursor);
 906         mutex_unlock(&audit_filter_mutex);
 907         path_put(&path1);
 908         drop_collected_mounts(tagged);
 909         return failed;
 910 }
 911
 912
 913 static void audit_schedule_prune(void)
 914 {
 915         wake_up_process(prune_thread);
 916 }
 917
 918 /*
 919  * ... and that one is done if evict_chunk() decides to delay until the end
 920  * of syscall.  Runs synchronously.
 921  */
 922 void audit_kill_trees(struct list_head *list)
 923 {
 924         mutex_lock(&audit_cmd_mutex);
 925         mutex_lock(&audit_filter_mutex);
 926
 927         while (!list_empty(list)) {
 928                 struct audit_tree *victim;
 929
 930                 victim = list_entry(list->next, struct audit_tree, list);
 931                 kill_rules(victim);
 932                 list_del_init(&victim->list);
 933
 934                 mutex_unlock(&audit_filter_mutex);
 935
 936                 prune_one(victim);
 937
 938                 mutex_lock(&audit_filter_mutex);
 939         }
 940
 941         mutex_unlock(&audit_filter_mutex);
 942         mutex_unlock(&audit_cmd_mutex);
 943 }
 944
 945 /*
 946  *  Here comes the stuff asynchronous to auditctl operations
 947  */
 948
 949 static void evict_chunk(struct audit_chunk *chunk)
 950 {
 951         struct audit_tree *owner;
 952         struct list_head *postponed = audit_killed_trees();
 953         int need_prune = 0;
 954         int n;
 955
 956         if (chunk->dead)
 957                 return;
 958
 959         chunk->dead = 1;
 960         mutex_lock(&audit_filter_mutex);
 961         spin_lock(&hash_lock);
 962         while (!list_empty(&chunk->trees)) {
 963                 owner = list_entry(chunk->trees.next,
 964                                    struct audit_tree, same_root);
 965                 owner->goner = 1;
 966                 owner->root = NULL;
 967                 list_del_init(&owner->same_root);
 968                 spin_unlock(&hash_lock);
 969                 if (!postponed) {
 970                         kill_rules(owner);
 971                         list_move(&owner->list, &prune_list);
 972                         need_prune = 1;
 973                 } else {
 974                         list_move(&owner->list, postponed);
 975                 }
 976                 spin_lock(&hash_lock);
 977         }
 978         list_del_rcu(&chunk->hash);
 979         for (n = 0; n < chunk->count; n++)
 980                 list_del_init(&chunk->owners[n].list);
 981         spin_unlock(&hash_lock);
 982         mutex_unlock(&audit_filter_mutex);
 983         if (need_prune)
 984                 audit_schedule_prune();
 985 }
 986
 987 static int audit_tree_handle_event(struct fsnotify_group *group,
 988                                    struct inode *to_tell,
 989                                    struct fsnotify_mark *inode_mark,
 990                                    struct fsnotify_mark *vfsmount_mark,
 991                                    u32 mask, const void *data, int data_type,
 992                                    const unsigned char *file_name, u32 cookie,
 993                                    struct fsnotify_iter_info *iter_info)
 994 {
 995         return 0;
 996 }
 997
 998 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
 999 {
1000         struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
1001
1002         evict_chunk(chunk);
1003
1004         /*
1005          * We are guaranteed to have at least one reference to the mark from
1006          * either the inode or the caller of fsnotify_destroy_mark().
1007          */
1008         BUG_ON(atomic_read(&entry->refcnt) < 1);
1009 }
1010
1011 static const struct fsnotify_ops audit_tree_ops = {
1012         .handle_event = audit_tree_handle_event,
1013         .freeing_mark = audit_tree_freeing_mark,
1014 };
1015
1016 static int __init audit_tree_init(void)
1017 {
1018         int i;
1019
1020         audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1021         if (IS_ERR(audit_tree_group))
1022                 audit_panic("cannot initialize fsnotify group for rectree watches");
1023
1024         for (i = 0; i < HASH_SIZE; i++)
1025                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
1026
1027         return 0;
1028 }
1029 __initcall(audit_tree_init);