kernel/audit_tree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "audit.h"
   3 #include <linux/fsnotify_backend.h>
   4 #include <linux/namei.h>
   5 #include <linux/mount.h>
   6 #include <linux/kthread.h>
   7 #include <linux/refcount.h>
   8 #include <linux/slab.h>
   9
  10 struct audit_tree;
  11 struct audit_chunk;
  12
  13 struct audit_tree {
  14         refcount_t count;
  15         int goner;
  16         struct audit_chunk *root;
  17         struct list_head chunks;
  18         struct list_head rules;
  19         struct list_head list;
  20         struct list_head same_root;
  21         struct rcu_head head;
  22         char pathname[];
  23 };
  24
  25 struct audit_chunk {
  26         struct list_head hash;
  27         unsigned long key;
  28         struct fsnotify_mark mark;
  29         struct list_head trees;         /* with root here */
  30         int dead;
  31         int count;
  32         atomic_long_t refs;
  33         struct rcu_head head;
  34         struct node {
  35                 struct list_head list;
  36                 struct audit_tree *owner;
  37                 unsigned index;         /* index; upper bit indicates 'will prune' */
  38         } owners[];
  39 };
  40
  41 static LIST_HEAD(tree_list);
  42 static LIST_HEAD(prune_list);
  43 static struct task_struct *prune_thread;
  44
  45 /*
  46  * One struct chunk is attached to each inode of interest.
  47  * We replace struct chunk on tagging/untagging.
  48  * Rules have pointer to struct audit_tree.
  49  * Rules have struct list_head rlist forming a list of rules over
  50  * the same tree.
  51  * References to struct chunk are collected at audit_inode{,_child}()
  52  * time and used in AUDIT_TREE rule matching.
  53  * These references are dropped at the same time we are calling
  54  * audit_free_names(), etc.
  55  *
  56  * Cyclic lists galore:
  57  * tree.chunks anchors chunk.owners[].list                      hash_lock
  58  * tree.rules anchors rule.rlist                                audit_filter_mutex
  59  * chunk.trees anchors tree.same_root                           hash_lock
  60  * chunk.hash is a hash with middle bits of watch.inode as
  61  * a hash function.                                             RCU, hash_lock
  62  *
  63  * tree is refcounted; one reference for "some rules on rules_list refer to
  64  * it", one for each chunk with pointer to it.
  65  *
  66  * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
  67  * of watch contributes 1 to .refs).
  68  *
  69  * node.index allows to get from node.list to containing chunk.
  70  * MSB of that sucker is stolen to mark taggings that we might have to
  71  * revert - several operations have very unpleasant cleanup logics and
  72  * that makes a difference.  Some.
  73  */
  74
  75 static struct fsnotify_group *audit_tree_group;
  76
  77 static struct audit_tree *alloc_tree(const char *s)
  78 {
  79         struct audit_tree *tree;
  80
  81         tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
  82         if (tree) {
  83                 refcount_set(&tree->count, 1);
  84                 tree->goner = 0;
  85                 INIT_LIST_HEAD(&tree->chunks);
  86                 INIT_LIST_HEAD(&tree->rules);
  87                 INIT_LIST_HEAD(&tree->list);
  88                 INIT_LIST_HEAD(&tree->same_root);
  89                 tree->root = NULL;
  90                 strcpy(tree->pathname, s);
  91         }
  92         return tree;
  93 }
  94
  95 static inline void get_tree(struct audit_tree *tree)
  96 {
  97         refcount_inc(&tree->count);
  98 }
  99
 100 static inline void put_tree(struct audit_tree *tree)
 101 {
 102         if (refcount_dec_and_test(&tree->count))
 103                 kfree_rcu(tree, head);
 104 }
 105
 106 /* to avoid bringing the entire thing in audit.h */
 107 const char *audit_tree_path(struct audit_tree *tree)
 108 {
 109         return tree->pathname;
 110 }
 111
 112 static void free_chunk(struct audit_chunk *chunk)
 113 {
 114         int i;
 115
 116         for (i = 0; i < chunk->count; i++) {
 117                 if (chunk->owners[i].owner)
 118                         put_tree(chunk->owners[i].owner);
 119         }
 120         kfree(chunk);
 121 }
 122
 123 void audit_put_chunk(struct audit_chunk *chunk)
 124 {
 125         if (atomic_long_dec_and_test(&chunk->refs))
 126                 free_chunk(chunk);
 127 }
 128
 129 static void __put_chunk(struct rcu_head *rcu)
 130 {
 131         struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
 132         audit_put_chunk(chunk);
 133 }
 134
 135 /*
 136  * Drop reference to the chunk that was held by the mark. This is the reference
 137  * that gets dropped after we've removed the chunk from the hash table and we
 138  * use it to make sure chunk cannot be freed before RCU grace period expires.
 139  */
 140 static void audit_mark_put_chunk(struct audit_chunk *chunk)
 141 {
 142         call_rcu(&chunk->head, __put_chunk);
 143 }
 144
 145 static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
 146 {
 147         struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
 148         audit_mark_put_chunk(chunk);
 149 }
 150
 151 static struct audit_chunk *alloc_chunk(int count)
 152 {
 153         struct audit_chunk *chunk;
 154         size_t size;
 155         int i;
 156
 157         size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
 158         chunk = kzalloc(size, GFP_KERNEL);
 159         if (!chunk)
 160                 return NULL;
 161
 162         INIT_LIST_HEAD(&chunk->hash);
 163         INIT_LIST_HEAD(&chunk->trees);
 164         chunk->count = count;
 165         atomic_long_set(&chunk->refs, 1);
 166         for (i = 0; i < count; i++) {
 167                 INIT_LIST_HEAD(&chunk->owners[i].list);
 168                 chunk->owners[i].index = i;
 169         }
 170         fsnotify_init_mark(&chunk->mark, audit_tree_group);
 171         chunk->mark.mask = FS_IN_IGNORED;
 172         return chunk;
 173 }
 174
 175 enum {HASH_SIZE = 128};
 176 static struct list_head chunk_hash_heads[HASH_SIZE];
 177 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
 178
 179 /* Function to return search key in our hash from inode. */
 180 static unsigned long inode_to_key(const struct inode *inode)
 181 {
 182         /* Use address pointed to by connector->obj as the key */
 183         return (unsigned long)&inode->i_fsnotify_marks;
 184 }
 185
 186 static inline struct list_head *chunk_hash(unsigned long key)
 187 {
 188         unsigned long n = key / L1_CACHE_BYTES;
 189         return chunk_hash_heads + n % HASH_SIZE;
 190 }
 191
 192 /* hash_lock & entry->group->mark_mutex is held by caller */
 193 static void insert_hash(struct audit_chunk *chunk)
 194 {
 195         struct list_head *list;
 196
 197         /*
 198          * Make sure chunk is fully initialized before making it visible in the
 199          * hash. Pairs with a data dependency barrier in READ_ONCE() in
 200          * audit_tree_lookup().
 201          */
 202         smp_wmb();
 203         WARN_ON_ONCE(!chunk->key);
 204         list = chunk_hash(chunk->key);
 205         list_add_rcu(&chunk->hash, list);
 206 }
 207
 208 /* called under rcu_read_lock */
 209 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
 210 {
 211         unsigned long key = inode_to_key(inode);
 212         struct list_head *list = chunk_hash(key);
 213         struct audit_chunk *p;
 214
 215         list_for_each_entry_rcu(p, list, hash) {
 216                 /*
 217                  * We use a data dependency barrier in READ_ONCE() to make sure
 218                  * the chunk we see is fully initialized.
 219                  */
 220                 if (READ_ONCE(p->key) == key) {
 221                         atomic_long_inc(&p->refs);
 222                         return p;
 223                 }
 224         }
 225         return NULL;
 226 }
 227
 228 bool audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
 229 {
 230         int n;
 231         for (n = 0; n < chunk->count; n++)
 232                 if (chunk->owners[n].owner == tree)
 233                         return true;
 234         return false;
 235 }
 236
 237 /* tagging and untagging inodes with trees */
 238
 239 static struct audit_chunk *find_chunk(struct node *p)
 240 {
 241         int index = p->index & ~(1U<<31);
 242         p -= index;
 243         return container_of(p, struct audit_chunk, owners[0]);
 244 }
 245
 246 static void replace_chunk(struct audit_chunk *new, struct audit_chunk *old,
 247                           struct node *skip)
 248 {
 249         struct audit_tree *owner;
 250         int i, j;
 251
 252         new->key = old->key;
 253         list_splice_init(&old->trees, &new->trees);
 254         list_for_each_entry(owner, &new->trees, same_root)
 255                 owner->root = new;
 256         for (i = j = 0; j < old->count; i++, j++) {
 257                 if (&old->owners[j] == skip) {
 258                         i--;
 259                         continue;
 260                 }
 261                 owner = old->owners[j].owner;
 262                 new->owners[i].owner = owner;
 263                 new->owners[i].index = old->owners[j].index - j + i;
 264                 if (!owner) /* result of earlier fallback */
 265                         continue;
 266                 get_tree(owner);
 267                 list_replace_init(&old->owners[j].list, &new->owners[i].list);
 268         }
 269         /*
 270          * Make sure chunk is fully initialized before making it visible in the
 271          * hash. Pairs with a data dependency barrier in READ_ONCE() in
 272          * audit_tree_lookup().
 273          */
 274         smp_wmb();
 275         list_replace_rcu(&old->hash, &new->hash);
 276 }
 277
 278 static void untag_chunk(struct node *p)
 279 {
 280         struct audit_chunk *chunk = find_chunk(p);
 281         struct fsnotify_mark *entry = &chunk->mark;
 282         struct audit_chunk *new = NULL;
 283         struct audit_tree *owner;
 284         int size = chunk->count - 1;
 285
 286         fsnotify_get_mark(entry);
 287
 288         spin_unlock(&hash_lock);
 289
 290         if (size)
 291                 new = alloc_chunk(size);
 292
 293         mutex_lock(&entry->group->mark_mutex);
 294         /*
 295          * mark_mutex protects mark from getting detached and thus also from
 296          * mark->connector->obj getting NULL.
 297          */
 298         if (chunk->dead || !(entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 299                 mutex_unlock(&entry->group->mark_mutex);
 300                 if (new)
 301                         fsnotify_put_mark(&new->mark);
 302                 goto out;
 303         }
 304
 305         owner = p->owner;
 306
 307         if (!size) {
 308                 chunk->dead = 1;
 309                 spin_lock(&hash_lock);
 310                 list_del_init(&chunk->trees);
 311                 if (owner->root == chunk)
 312                         owner->root = NULL;
 313                 list_del_init(&p->list);
 314                 list_del_rcu(&chunk->hash);
 315                 spin_unlock(&hash_lock);
 316                 fsnotify_detach_mark(entry);
 317                 mutex_unlock(&entry->group->mark_mutex);
 318                 fsnotify_free_mark(entry);
 319                 goto out;
 320         }
 321
 322         if (!new)
 323                 goto Fallback;
 324
 325         if (fsnotify_add_mark_locked(&new->mark, entry->connector->obj,
 326                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 327                 fsnotify_put_mark(&new->mark);
 328                 goto Fallback;
 329         }
 330
 331         chunk->dead = 1;
 332         spin_lock(&hash_lock);
 333         if (owner->root == chunk) {
 334                 list_del_init(&owner->same_root);
 335                 owner->root = NULL;
 336         }
 337         list_del_init(&p->list);
 338         /*
 339          * This has to go last when updating chunk as once replace_chunk() is
 340          * called, new RCU readers can see the new chunk.
 341          */
 342         replace_chunk(new, chunk, p);
 343         spin_unlock(&hash_lock);
 344         fsnotify_detach_mark(entry);
 345         mutex_unlock(&entry->group->mark_mutex);
 346         fsnotify_free_mark(entry);
 347         fsnotify_put_mark(&new->mark);  /* drop initial reference */
 348         goto out;
 349
 350 Fallback:
 351         // do the best we can
 352         spin_lock(&hash_lock);
 353         if (owner->root == chunk) {
 354                 list_del_init(&owner->same_root);
 355                 owner->root = NULL;
 356         }
 357         list_del_init(&p->list);
 358         p->owner = NULL;
 359         put_tree(owner);
 360         spin_unlock(&hash_lock);
 361         mutex_unlock(&entry->group->mark_mutex);
 362 out:
 363         fsnotify_put_mark(entry);
 364         spin_lock(&hash_lock);
 365 }
 366
 367 /* Call with group->mark_mutex held, releases it */
 368 static int create_chunk(struct inode *inode, struct audit_tree *tree)
 369 {
 370         struct fsnotify_mark *entry;
 371         struct audit_chunk *chunk = alloc_chunk(1);
 372
 373         if (!chunk) {
 374                 mutex_unlock(&audit_tree_group->mark_mutex);
 375                 return -ENOMEM;
 376         }
 377
 378         entry = &chunk->mark;
 379         if (fsnotify_add_inode_mark_locked(entry, inode, 0)) {
 380                 mutex_unlock(&audit_tree_group->mark_mutex);
 381                 fsnotify_put_mark(entry);
 382                 return -ENOSPC;
 383         }
 384
 385         spin_lock(&hash_lock);
 386         if (tree->goner) {
 387                 spin_unlock(&hash_lock);
 388                 chunk->dead = 1;
 389                 fsnotify_detach_mark(entry);
 390                 mutex_unlock(&audit_tree_group->mark_mutex);
 391                 fsnotify_free_mark(entry);
 392                 fsnotify_put_mark(entry);
 393                 return 0;
 394         }
 395         chunk->owners[0].index = (1U << 31);
 396         chunk->owners[0].owner = tree;
 397         get_tree(tree);
 398         list_add(&chunk->owners[0].list, &tree->chunks);
 399         if (!tree->root) {
 400                 tree->root = chunk;
 401                 list_add(&tree->same_root, &chunk->trees);
 402         }
 403         chunk->key = inode_to_key(inode);
 404         /*
 405          * Inserting into the hash table has to go last as once we do that RCU
 406          * readers can see the chunk.
 407          */
 408         insert_hash(chunk);
 409         spin_unlock(&hash_lock);
 410         mutex_unlock(&audit_tree_group->mark_mutex);
 411         fsnotify_put_mark(entry);       /* drop initial reference */
 412         return 0;
 413 }
 414
 415 /* the first tagged inode becomes root of tree */
 416 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
 417 {
 418         struct fsnotify_mark *old_entry, *chunk_entry;
 419         struct audit_chunk *chunk, *old;
 420         struct node *p;
 421         int n;
 422
 423         mutex_lock(&audit_tree_group->mark_mutex);
 424         old_entry = fsnotify_find_mark(&inode->i_fsnotify_marks,
 425                                        audit_tree_group);
 426         if (!old_entry)
 427                 return create_chunk(inode, tree);
 428
 429         old = container_of(old_entry, struct audit_chunk, mark);
 430
 431         /* are we already there? */
 432         spin_lock(&hash_lock);
 433         for (n = 0; n < old->count; n++) {
 434                 if (old->owners[n].owner == tree) {
 435                         spin_unlock(&hash_lock);
 436                         mutex_unlock(&audit_tree_group->mark_mutex);
 437                         fsnotify_put_mark(old_entry);
 438                         return 0;
 439                 }
 440         }
 441         spin_unlock(&hash_lock);
 442
 443         chunk = alloc_chunk(old->count + 1);
 444         if (!chunk) {
 445                 mutex_unlock(&audit_tree_group->mark_mutex);
 446                 fsnotify_put_mark(old_entry);
 447                 return -ENOMEM;
 448         }
 449
 450         chunk_entry = &chunk->mark;
 451
 452         /*
 453          * mark_mutex protects mark from getting detached and thus also from
 454          * mark->connector->obj getting NULL.
 455          */
 456         if (!(old_entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 457                 /* old_entry is being shot, lets just lie */
 458                 mutex_unlock(&audit_tree_group->mark_mutex);
 459                 fsnotify_put_mark(old_entry);
 460                 fsnotify_put_mark(&chunk->mark);
 461                 return -ENOENT;
 462         }
 463
 464         if (fsnotify_add_mark_locked(chunk_entry, old_entry->connector->obj,
 465                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 466                 mutex_unlock(&audit_tree_group->mark_mutex);
 467                 fsnotify_put_mark(chunk_entry);
 468                 fsnotify_put_mark(old_entry);
 469                 return -ENOSPC;
 470         }
 471
 472         spin_lock(&hash_lock);
 473         if (tree->goner) {
 474                 spin_unlock(&hash_lock);
 475                 chunk->dead = 1;
 476                 fsnotify_detach_mark(chunk_entry);
 477                 mutex_unlock(&audit_tree_group->mark_mutex);
 478                 fsnotify_free_mark(chunk_entry);
 479                 fsnotify_put_mark(chunk_entry);
 480                 fsnotify_put_mark(old_entry);
 481                 return 0;
 482         }
 483         p = &chunk->owners[chunk->count - 1];
 484         p->index = (chunk->count - 1) | (1U<<31);
 485         p->owner = tree;
 486         get_tree(tree);
 487         list_add(&p->list, &tree->chunks);
 488         old->dead = 1;
 489         if (!tree->root) {
 490                 tree->root = chunk;
 491                 list_add(&tree->same_root, &chunk->trees);
 492         }
 493         /*
 494          * This has to go last when updating chunk as once replace_chunk() is
 495          * called, new RCU readers can see the new chunk.
 496          */
 497         replace_chunk(chunk, old, NULL);
 498         spin_unlock(&hash_lock);
 499         fsnotify_detach_mark(old_entry);
 500         mutex_unlock(&audit_tree_group->mark_mutex);
 501         fsnotify_free_mark(old_entry);
 502         fsnotify_put_mark(chunk_entry); /* drop initial reference */
 503         fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
 504         return 0;
 505 }
 506
 507 static void audit_tree_log_remove_rule(struct audit_krule *rule)
 508 {
 509         struct audit_buffer *ab;
 510
 511         if (!audit_enabled)
 512                 return;
 513         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
 514         if (unlikely(!ab))
 515                 return;
 516         audit_log_format(ab, "op=remove_rule");
 517         audit_log_format(ab, " dir=");
 518         audit_log_untrustedstring(ab, rule->tree->pathname);
 519         audit_log_key(ab, rule->filterkey);
 520         audit_log_format(ab, " list=%d res=1", rule->listnr);
 521         audit_log_end(ab);
 522 }
 523
 524 static void kill_rules(struct audit_tree *tree)
 525 {
 526         struct audit_krule *rule, *next;
 527         struct audit_entry *entry;
 528
 529         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
 530                 entry = container_of(rule, struct audit_entry, rule);
 531
 532                 list_del_init(&rule->rlist);
 533                 if (rule->tree) {
 534                         /* not a half-baked one */
 535                         audit_tree_log_remove_rule(rule);
 536                         if (entry->rule.exe)
 537                                 audit_remove_mark(entry->rule.exe);
 538                         rule->tree = NULL;
 539                         list_del_rcu(&entry->list);
 540                         list_del(&entry->rule.list);
 541                         call_rcu(&entry->rcu, audit_free_rule_rcu);
 542                 }
 543         }
 544 }
 545
 546 /*
 547  * finish killing struct audit_tree
 548  */
 549 static void prune_one(struct audit_tree *victim)
 550 {
 551         spin_lock(&hash_lock);
 552         while (!list_empty(&victim->chunks)) {
 553                 struct node *p;
 554
 555                 p = list_entry(victim->chunks.next, struct node, list);
 556
 557                 untag_chunk(p);
 558         }
 559         spin_unlock(&hash_lock);
 560         put_tree(victim);
 561 }
 562
 563 /* trim the uncommitted chunks from tree */
 564
 565 static void trim_marked(struct audit_tree *tree)
 566 {
 567         struct list_head *p, *q;
 568         spin_lock(&hash_lock);
 569         if (tree->goner) {
 570                 spin_unlock(&hash_lock);
 571                 return;
 572         }
 573         /* reorder */
 574         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
 575                 struct node *node = list_entry(p, struct node, list);
 576                 q = p->next;
 577                 if (node->index & (1U<<31)) {
 578                         list_del_init(p);
 579                         list_add(p, &tree->chunks);
 580                 }
 581         }
 582
 583         while (!list_empty(&tree->chunks)) {
 584                 struct node *node;
 585
 586                 node = list_entry(tree->chunks.next, struct node, list);
 587
 588                 /* have we run out of marked? */
 589                 if (!(node->index & (1U<<31)))
 590                         break;
 591
 592                 untag_chunk(node);
 593         }
 594         if (!tree->root && !tree->goner) {
 595                 tree->goner = 1;
 596                 spin_unlock(&hash_lock);
 597                 mutex_lock(&audit_filter_mutex);
 598                 kill_rules(tree);
 599                 list_del_init(&tree->list);
 600                 mutex_unlock(&audit_filter_mutex);
 601                 prune_one(tree);
 602         } else {
 603                 spin_unlock(&hash_lock);
 604         }
 605 }
 606
 607 static void audit_schedule_prune(void);
 608
 609 /* called with audit_filter_mutex */
 610 int audit_remove_tree_rule(struct audit_krule *rule)
 611 {
 612         struct audit_tree *tree;
 613         tree = rule->tree;
 614         if (tree) {
 615                 spin_lock(&hash_lock);
 616                 list_del_init(&rule->rlist);
 617                 if (list_empty(&tree->rules) && !tree->goner) {
 618                         tree->root = NULL;
 619                         list_del_init(&tree->same_root);
 620                         tree->goner = 1;
 621                         list_move(&tree->list, &prune_list);
 622                         rule->tree = NULL;
 623                         spin_unlock(&hash_lock);
 624                         audit_schedule_prune();
 625                         return 1;
 626                 }
 627                 rule->tree = NULL;
 628                 spin_unlock(&hash_lock);
 629                 return 1;
 630         }
 631         return 0;
 632 }
 633
 634 static int compare_root(struct vfsmount *mnt, void *arg)
 635 {
 636         return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
 637                (unsigned long)arg;
 638 }
 639
 640 void audit_trim_trees(void)
 641 {
 642         struct list_head cursor;
 643
 644         mutex_lock(&audit_filter_mutex);
 645         list_add(&cursor, &tree_list);
 646         while (cursor.next != &tree_list) {
 647                 struct audit_tree *tree;
 648                 struct path path;
 649                 struct vfsmount *root_mnt;
 650                 struct node *node;
 651                 int err;
 652
 653                 tree = container_of(cursor.next, struct audit_tree, list);
 654                 get_tree(tree);
 655                 list_del(&cursor);
 656                 list_add(&cursor, &tree->list);
 657                 mutex_unlock(&audit_filter_mutex);
 658
 659                 err = kern_path(tree->pathname, 0, &path);
 660                 if (err)
 661                         goto skip_it;
 662
 663                 root_mnt = collect_mounts(&path);
 664                 path_put(&path);
 665                 if (IS_ERR(root_mnt))
 666                         goto skip_it;
 667
 668                 spin_lock(&hash_lock);
 669                 list_for_each_entry(node, &tree->chunks, list) {
 670                         struct audit_chunk *chunk = find_chunk(node);
 671                         /* this could be NULL if the watch is dying else where... */
 672                         node->index |= 1U<<31;
 673                         if (iterate_mounts(compare_root,
 674                                            (void *)(chunk->key),
 675                                            root_mnt))
 676                                 node->index &= ~(1U<<31);
 677                 }
 678                 spin_unlock(&hash_lock);
 679                 trim_marked(tree);
 680                 drop_collected_mounts(root_mnt);
 681 skip_it:
 682                 put_tree(tree);
 683                 mutex_lock(&audit_filter_mutex);
 684         }
 685         list_del(&cursor);
 686         mutex_unlock(&audit_filter_mutex);
 687 }
 688
 689 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
 690 {
 691
 692         if (pathname[0] != '/' ||
 693             rule->listnr != AUDIT_FILTER_EXIT ||
 694             op != Audit_equal ||
 695             rule->inode_f || rule->watch || rule->tree)
 696                 return -EINVAL;
 697         rule->tree = alloc_tree(pathname);
 698         if (!rule->tree)
 699                 return -ENOMEM;
 700         return 0;
 701 }
 702
 703 void audit_put_tree(struct audit_tree *tree)
 704 {
 705         put_tree(tree);
 706 }
 707
 708 static int tag_mount(struct vfsmount *mnt, void *arg)
 709 {
 710         return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
 711 }
 712
 713 /*
 714  * That gets run when evict_chunk() ends up needing to kill audit_tree.
 715  * Runs from a separate thread.
 716  */
 717 static int prune_tree_thread(void *unused)
 718 {
 719         for (;;) {
 720                 if (list_empty(&prune_list)) {
 721                         set_current_state(TASK_INTERRUPTIBLE);
 722                         schedule();
 723                 }
 724
 725                 audit_ctl_lock();
 726                 mutex_lock(&audit_filter_mutex);
 727
 728                 while (!list_empty(&prune_list)) {
 729                         struct audit_tree *victim;
 730
 731                         victim = list_entry(prune_list.next,
 732                                         struct audit_tree, list);
 733                         list_del_init(&victim->list);
 734
 735                         mutex_unlock(&audit_filter_mutex);
 736
 737                         prune_one(victim);
 738
 739                         mutex_lock(&audit_filter_mutex);
 740                 }
 741
 742                 mutex_unlock(&audit_filter_mutex);
 743                 audit_ctl_unlock();
 744         }
 745         return 0;
 746 }
 747
 748 static int audit_launch_prune(void)
 749 {
 750         if (prune_thread)
 751                 return 0;
 752         prune_thread = kthread_run(prune_tree_thread, NULL,
 753                                 "audit_prune_tree");
 754         if (IS_ERR(prune_thread)) {
 755                 pr_err("cannot start thread audit_prune_tree");
 756                 prune_thread = NULL;
 757                 return -ENOMEM;
 758         }
 759         return 0;
 760 }
 761
 762 /* called with audit_filter_mutex */
 763 int audit_add_tree_rule(struct audit_krule *rule)
 764 {
 765         struct audit_tree *seed = rule->tree, *tree;
 766         struct path path;
 767         struct vfsmount *mnt;
 768         int err;
 769
 770         rule->tree = NULL;
 771         list_for_each_entry(tree, &tree_list, list) {
 772                 if (!strcmp(seed->pathname, tree->pathname)) {
 773                         put_tree(seed);
 774                         rule->tree = tree;
 775                         list_add(&rule->rlist, &tree->rules);
 776                         return 0;
 777                 }
 778         }
 779         tree = seed;
 780         list_add(&tree->list, &tree_list);
 781         list_add(&rule->rlist, &tree->rules);
 782         /* do not set rule->tree yet */
 783         mutex_unlock(&audit_filter_mutex);
 784
 785         if (unlikely(!prune_thread)) {
 786                 err = audit_launch_prune();
 787                 if (err)
 788                         goto Err;
 789         }
 790
 791         err = kern_path(tree->pathname, 0, &path);
 792         if (err)
 793                 goto Err;
 794         mnt = collect_mounts(&path);
 795         path_put(&path);
 796         if (IS_ERR(mnt)) {
 797                 err = PTR_ERR(mnt);
 798                 goto Err;
 799         }
 800
 801         get_tree(tree);
 802         err = iterate_mounts(tag_mount, tree, mnt);
 803         drop_collected_mounts(mnt);
 804
 805         if (!err) {
 806                 struct node *node;
 807                 spin_lock(&hash_lock);
 808                 list_for_each_entry(node, &tree->chunks, list)
 809                         node->index &= ~(1U<<31);
 810                 spin_unlock(&hash_lock);
 811         } else {
 812                 trim_marked(tree);
 813                 goto Err;
 814         }
 815
 816         mutex_lock(&audit_filter_mutex);
 817         if (list_empty(&rule->rlist)) {
 818                 put_tree(tree);
 819                 return -ENOENT;
 820         }
 821         rule->tree = tree;
 822         put_tree(tree);
 823
 824         return 0;
 825 Err:
 826         mutex_lock(&audit_filter_mutex);
 827         list_del_init(&tree->list);
 828         list_del_init(&tree->rules);
 829         put_tree(tree);
 830         return err;
 831 }
 832
 833 int audit_tag_tree(char *old, char *new)
 834 {
 835         struct list_head cursor, barrier;
 836         int failed = 0;
 837         struct path path1, path2;
 838         struct vfsmount *tagged;
 839         int err;
 840
 841         err = kern_path(new, 0, &path2);
 842         if (err)
 843                 return err;
 844         tagged = collect_mounts(&path2);
 845         path_put(&path2);
 846         if (IS_ERR(tagged))
 847                 return PTR_ERR(tagged);
 848
 849         err = kern_path(old, 0, &path1);
 850         if (err) {
 851                 drop_collected_mounts(tagged);
 852                 return err;
 853         }
 854
 855         mutex_lock(&audit_filter_mutex);
 856         list_add(&barrier, &tree_list);
 857         list_add(&cursor, &barrier);
 858
 859         while (cursor.next != &tree_list) {
 860                 struct audit_tree *tree;
 861                 int good_one = 0;
 862
 863                 tree = container_of(cursor.next, struct audit_tree, list);
 864                 get_tree(tree);
 865                 list_del(&cursor);
 866                 list_add(&cursor, &tree->list);
 867                 mutex_unlock(&audit_filter_mutex);
 868
 869                 err = kern_path(tree->pathname, 0, &path2);
 870                 if (!err) {
 871                         good_one = path_is_under(&path1, &path2);
 872                         path_put(&path2);
 873                 }
 874
 875                 if (!good_one) {
 876                         put_tree(tree);
 877                         mutex_lock(&audit_filter_mutex);
 878                         continue;
 879                 }
 880
 881                 failed = iterate_mounts(tag_mount, tree, tagged);
 882                 if (failed) {
 883                         put_tree(tree);
 884                         mutex_lock(&audit_filter_mutex);
 885                         break;
 886                 }
 887
 888                 mutex_lock(&audit_filter_mutex);
 889                 spin_lock(&hash_lock);
 890                 if (!tree->goner) {
 891                         list_del(&tree->list);
 892                         list_add(&tree->list, &tree_list);
 893                 }
 894                 spin_unlock(&hash_lock);
 895                 put_tree(tree);
 896         }
 897
 898         while (barrier.prev != &tree_list) {
 899                 struct audit_tree *tree;
 900
 901                 tree = container_of(barrier.prev, struct audit_tree, list);
 902                 get_tree(tree);
 903                 list_del(&tree->list);
 904                 list_add(&tree->list, &barrier);
 905                 mutex_unlock(&audit_filter_mutex);
 906
 907                 if (!failed) {
 908                         struct node *node;
 909                         spin_lock(&hash_lock);
 910                         list_for_each_entry(node, &tree->chunks, list)
 911                                 node->index &= ~(1U<<31);
 912                         spin_unlock(&hash_lock);
 913                 } else {
 914                         trim_marked(tree);
 915                 }
 916
 917                 put_tree(tree);
 918                 mutex_lock(&audit_filter_mutex);
 919         }
 920         list_del(&barrier);
 921         list_del(&cursor);
 922         mutex_unlock(&audit_filter_mutex);
 923         path_put(&path1);
 924         drop_collected_mounts(tagged);
 925         return failed;
 926 }
 927
 928
 929 static void audit_schedule_prune(void)
 930 {
 931         wake_up_process(prune_thread);
 932 }
 933
 934 /*
 935  * ... and that one is done if evict_chunk() decides to delay until the end
 936  * of syscall.  Runs synchronously.
 937  */
 938 void audit_kill_trees(struct list_head *list)
 939 {
 940         audit_ctl_lock();
 941         mutex_lock(&audit_filter_mutex);
 942
 943         while (!list_empty(list)) {
 944                 struct audit_tree *victim;
 945
 946                 victim = list_entry(list->next, struct audit_tree, list);
 947                 kill_rules(victim);
 948                 list_del_init(&victim->list);
 949
 950                 mutex_unlock(&audit_filter_mutex);
 951
 952                 prune_one(victim);
 953
 954                 mutex_lock(&audit_filter_mutex);
 955         }
 956
 957         mutex_unlock(&audit_filter_mutex);
 958         audit_ctl_unlock();
 959 }
 960
 961 /*
 962  *  Here comes the stuff asynchronous to auditctl operations
 963  */
 964
 965 static void evict_chunk(struct audit_chunk *chunk)
 966 {
 967         struct audit_tree *owner;
 968         struct list_head *postponed = audit_killed_trees();
 969         int need_prune = 0;
 970         int n;
 971
 972         if (chunk->dead)
 973                 return;
 974
 975         chunk->dead = 1;
 976         mutex_lock(&audit_filter_mutex);
 977         spin_lock(&hash_lock);
 978         while (!list_empty(&chunk->trees)) {
 979                 owner = list_entry(chunk->trees.next,
 980                                    struct audit_tree, same_root);
 981                 owner->goner = 1;
 982                 owner->root = NULL;
 983                 list_del_init(&owner->same_root);
 984                 spin_unlock(&hash_lock);
 985                 if (!postponed) {
 986                         kill_rules(owner);
 987                         list_move(&owner->list, &prune_list);
 988                         need_prune = 1;
 989                 } else {
 990                         list_move(&owner->list, postponed);
 991                 }
 992                 spin_lock(&hash_lock);
 993         }
 994         list_del_rcu(&chunk->hash);
 995         for (n = 0; n < chunk->count; n++)
 996                 list_del_init(&chunk->owners[n].list);
 997         spin_unlock(&hash_lock);
 998         mutex_unlock(&audit_filter_mutex);
 999         if (need_prune)
1000                 audit_schedule_prune();
1001 }
1002
1003 static int audit_tree_handle_event(struct fsnotify_group *group,
1004                                    struct inode *to_tell,
1005                                    u32 mask, const void *data, int data_type,
1006                                    const unsigned char *file_name, u32 cookie,
1007                                    struct fsnotify_iter_info *iter_info)
1008 {
1009         return 0;
1010 }
1011
1012 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
1013 {
1014         struct audit_chunk *chunk = container_of(entry, struct audit_chunk, mark);
1015
1016         evict_chunk(chunk);
1017
1018         /*
1019          * We are guaranteed to have at least one reference to the mark from
1020          * either the inode or the caller of fsnotify_destroy_mark().
1021          */
1022         BUG_ON(refcount_read(&entry->refcnt) < 1);
1023 }
1024
1025 static const struct fsnotify_ops audit_tree_ops = {
1026         .handle_event = audit_tree_handle_event,
1027         .freeing_mark = audit_tree_freeing_mark,
1028         .free_mark = audit_tree_destroy_watch,
1029 };
1030
1031 static int __init audit_tree_init(void)
1032 {
1033         int i;
1034
1035         audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1036         if (IS_ERR(audit_tree_group))
1037                 audit_panic("cannot initialize fsnotify group for rectree watches");
1038
1039         for (i = 0; i < HASH_SIZE; i++)
1040                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
1041
1042         return 0;
1043 }
1044 __initcall(audit_tree_init);