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