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