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 struct audit_tree_mark {
  42         struct fsnotify_mark mark;
  43         struct audit_chunk *chunk;
  44 };
  45
  46 static LIST_HEAD(tree_list);
  47 static LIST_HEAD(prune_list);
  48 static struct task_struct *prune_thread;
  49
  50 /*
  51  * One struct chunk is attached to each inode of interest.
  52  * We replace struct chunk on tagging/untagging.
  53  * Rules have pointer to struct audit_tree.
  54  * Rules have struct list_head rlist forming a list of rules over
  55  * the same tree.
  56  * References to struct chunk are collected at audit_inode{,_child}()
  57  * time and used in AUDIT_TREE rule matching.
  58  * These references are dropped at the same time we are calling
  59  * audit_free_names(), etc.
  60  *
  61  * Cyclic lists galore:
  62  * tree.chunks anchors chunk.owners[].list                      hash_lock
  63  * tree.rules anchors rule.rlist                                audit_filter_mutex
  64  * chunk.trees anchors tree.same_root                           hash_lock
  65  * chunk.hash is a hash with middle bits of watch.inode as
  66  * a hash function.                                             RCU, hash_lock
  67  *
  68  * tree is refcounted; one reference for "some rules on rules_list refer to
  69  * it", one for each chunk with pointer to it.
  70  *
  71  * chunk is refcounted by embedded fsnotify_mark + .refs (non-zero refcount
  72  * of watch contributes 1 to .refs).
  73  *
  74  * node.index allows to get from node.list to containing chunk.
  75  * MSB of that sucker is stolen to mark taggings that we might have to
  76  * revert - several operations have very unpleasant cleanup logics and
  77  * that makes a difference.  Some.
  78  */
  79
  80 static struct fsnotify_group *audit_tree_group;
  81 static struct kmem_cache *audit_tree_mark_cachep __read_mostly;
  82
  83 static struct audit_tree *alloc_tree(const char *s)
  84 {
  85         struct audit_tree *tree;
  86
  87         tree = kmalloc(sizeof(struct audit_tree) + strlen(s) + 1, GFP_KERNEL);
  88         if (tree) {
  89                 refcount_set(&tree->count, 1);
  90                 tree->goner = 0;
  91                 INIT_LIST_HEAD(&tree->chunks);
  92                 INIT_LIST_HEAD(&tree->rules);
  93                 INIT_LIST_HEAD(&tree->list);
  94                 INIT_LIST_HEAD(&tree->same_root);
  95                 tree->root = NULL;
  96                 strcpy(tree->pathname, s);
  97         }
  98         return tree;
  99 }
 100
 101 static inline void get_tree(struct audit_tree *tree)
 102 {
 103         refcount_inc(&tree->count);
 104 }
 105
 106 static inline void put_tree(struct audit_tree *tree)
 107 {
 108         if (refcount_dec_and_test(&tree->count))
 109                 kfree_rcu(tree, head);
 110 }
 111
 112 /* to avoid bringing the entire thing in audit.h */
 113 const char *audit_tree_path(struct audit_tree *tree)
 114 {
 115         return tree->pathname;
 116 }
 117
 118 static void free_chunk(struct audit_chunk *chunk)
 119 {
 120         int i;
 121
 122         for (i = 0; i < chunk->count; i++) {
 123                 if (chunk->owners[i].owner)
 124                         put_tree(chunk->owners[i].owner);
 125         }
 126         kfree(chunk);
 127 }
 128
 129 void audit_put_chunk(struct audit_chunk *chunk)
 130 {
 131         if (atomic_long_dec_and_test(&chunk->refs))
 132                 free_chunk(chunk);
 133 }
 134
 135 static void __put_chunk(struct rcu_head *rcu)
 136 {
 137         struct audit_chunk *chunk = container_of(rcu, struct audit_chunk, head);
 138         audit_put_chunk(chunk);
 139 }
 140
 141 /*
 142  * Drop reference to the chunk that was held by the mark. This is the reference
 143  * that gets dropped after we've removed the chunk from the hash table and we
 144  * use it to make sure chunk cannot be freed before RCU grace period expires.
 145  */
 146 static void audit_mark_put_chunk(struct audit_chunk *chunk)
 147 {
 148         call_rcu(&chunk->head, __put_chunk);
 149 }
 150
 151 static inline struct audit_tree_mark *audit_mark(struct fsnotify_mark *entry)
 152 {
 153         return container_of(entry, struct audit_tree_mark, mark);
 154 }
 155
 156 static struct audit_chunk *mark_chunk(struct fsnotify_mark *mark)
 157 {
 158         return audit_mark(mark)->chunk;
 159 }
 160
 161 static void audit_tree_destroy_watch(struct fsnotify_mark *entry)
 162 {
 163         struct audit_chunk *chunk = mark_chunk(entry);
 164         audit_mark_put_chunk(chunk);
 165         kmem_cache_free(audit_tree_mark_cachep, audit_mark(entry));
 166 }
 167
 168 static struct fsnotify_mark *alloc_mark(void)
 169 {
 170         struct audit_tree_mark *amark;
 171
 172         amark = kmem_cache_zalloc(audit_tree_mark_cachep, GFP_KERNEL);
 173         if (!amark)
 174                 return NULL;
 175         fsnotify_init_mark(&amark->mark, audit_tree_group);
 176         amark->mark.mask = FS_IN_IGNORED;
 177         return &amark->mark;
 178 }
 179
 180 static struct audit_chunk *alloc_chunk(int count)
 181 {
 182         struct audit_chunk *chunk;
 183         size_t size;
 184         int i;
 185
 186         size = offsetof(struct audit_chunk, owners) + count * sizeof(struct node);
 187         chunk = kzalloc(size, GFP_KERNEL);
 188         if (!chunk)
 189                 return NULL;
 190
 191         chunk->mark = alloc_mark();
 192         if (!chunk->mark) {
 193                 kfree(chunk);
 194                 return NULL;
 195         }
 196         audit_mark(chunk->mark)->chunk = chunk;
 197
 198         INIT_LIST_HEAD(&chunk->hash);
 199         INIT_LIST_HEAD(&chunk->trees);
 200         chunk->count = count;
 201         atomic_long_set(&chunk->refs, 1);
 202         for (i = 0; i < count; i++) {
 203                 INIT_LIST_HEAD(&chunk->owners[i].list);
 204                 chunk->owners[i].index = i;
 205         }
 206         return chunk;
 207 }
 208
 209 enum {HASH_SIZE = 128};
 210 static struct list_head chunk_hash_heads[HASH_SIZE];
 211 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(hash_lock);
 212
 213 /* Function to return search key in our hash from inode. */
 214 static unsigned long inode_to_key(const struct inode *inode)
 215 {
 216         /* Use address pointed to by connector->obj as the key */
 217         return (unsigned long)&inode->i_fsnotify_marks;
 218 }
 219
 220 static inline struct list_head *chunk_hash(unsigned long key)
 221 {
 222         unsigned long n = key / L1_CACHE_BYTES;
 223         return chunk_hash_heads + n % HASH_SIZE;
 224 }
 225
 226 /* hash_lock & entry->group->mark_mutex is held by caller */
 227 static void insert_hash(struct audit_chunk *chunk)
 228 {
 229         struct list_head *list;
 230
 231         /*
 232          * Make sure chunk is fully initialized before making it visible in the
 233          * hash. Pairs with a data dependency barrier in READ_ONCE() in
 234          * audit_tree_lookup().
 235          */
 236         smp_wmb();
 237         WARN_ON_ONCE(!chunk->key);
 238         list = chunk_hash(chunk->key);
 239         list_add_rcu(&chunk->hash, list);
 240 }
 241
 242 /* called under rcu_read_lock */
 243 struct audit_chunk *audit_tree_lookup(const struct inode *inode)
 244 {
 245         unsigned long key = inode_to_key(inode);
 246         struct list_head *list = chunk_hash(key);
 247         struct audit_chunk *p;
 248
 249         list_for_each_entry_rcu(p, list, hash) {
 250                 /*
 251                  * We use a data dependency barrier in READ_ONCE() to make sure
 252                  * the chunk we see is fully initialized.
 253                  */
 254                 if (READ_ONCE(p->key) == key) {
 255                         atomic_long_inc(&p->refs);
 256                         return p;
 257                 }
 258         }
 259         return NULL;
 260 }
 261
 262 bool audit_tree_match(struct audit_chunk *chunk, struct audit_tree *tree)
 263 {
 264         int n;
 265         for (n = 0; n < chunk->count; n++)
 266                 if (chunk->owners[n].owner == tree)
 267                         return true;
 268         return false;
 269 }
 270
 271 /* tagging and untagging inodes with trees */
 272
 273 static struct audit_chunk *find_chunk(struct node *p)
 274 {
 275         int index = p->index & ~(1U<<31);
 276         p -= index;
 277         return container_of(p, struct audit_chunk, owners[0]);
 278 }
 279
 280 static void replace_chunk(struct audit_chunk *new, struct audit_chunk *old)
 281 {
 282         struct audit_tree *owner;
 283         int i, j;
 284
 285         new->key = old->key;
 286         list_splice_init(&old->trees, &new->trees);
 287         list_for_each_entry(owner, &new->trees, same_root)
 288                 owner->root = new;
 289         for (i = j = 0; j < old->count; i++, j++) {
 290                 if (!old->owners[j].owner) {
 291                         i--;
 292                         continue;
 293                 }
 294                 owner = old->owners[j].owner;
 295                 new->owners[i].owner = owner;
 296                 new->owners[i].index = old->owners[j].index - j + i;
 297                 if (!owner) /* result of earlier fallback */
 298                         continue;
 299                 get_tree(owner);
 300                 list_replace_init(&old->owners[j].list, &new->owners[i].list);
 301         }
 302         /*
 303          * Make sure chunk is fully initialized before making it visible in the
 304          * hash. Pairs with a data dependency barrier in READ_ONCE() in
 305          * audit_tree_lookup().
 306          */
 307         smp_wmb();
 308         list_replace_rcu(&old->hash, &new->hash);
 309 }
 310
 311 static void remove_chunk_node(struct audit_chunk *chunk, struct node *p)
 312 {
 313         struct audit_tree *owner = p->owner;
 314
 315         if (owner->root == chunk) {
 316                 list_del_init(&owner->same_root);
 317                 owner->root = NULL;
 318         }
 319         list_del_init(&p->list);
 320         p->owner = NULL;
 321         put_tree(owner);
 322 }
 323
 324 static int chunk_count_trees(struct audit_chunk *chunk)
 325 {
 326         int i;
 327         int ret = 0;
 328
 329         for (i = 0; i < chunk->count; i++)
 330                 if (chunk->owners[i].owner)
 331                         ret++;
 332         return ret;
 333 }
 334
 335 static void untag_chunk(struct audit_chunk *chunk, struct fsnotify_mark *entry)
 336 {
 337         struct audit_chunk *new;
 338         int size;
 339
 340         mutex_lock(&audit_tree_group->mark_mutex);
 341         /*
 342          * mark_mutex protects mark from getting detached and thus also from
 343          * mark->connector->obj getting NULL.
 344          */
 345         if (chunk->dead || !(entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED))
 346                 goto out_mutex;
 347
 348         size = chunk_count_trees(chunk);
 349         if (!size) {
 350                 chunk->dead = 1;
 351                 spin_lock(&hash_lock);
 352                 list_del_init(&chunk->trees);
 353                 list_del_rcu(&chunk->hash);
 354                 spin_unlock(&hash_lock);
 355                 fsnotify_detach_mark(entry);
 356                 mutex_unlock(&audit_tree_group->mark_mutex);
 357                 fsnotify_free_mark(entry);
 358                 return;
 359         }
 360
 361         new = alloc_chunk(size);
 362         if (!new)
 363                 goto out_mutex;
 364
 365         if (fsnotify_add_mark_locked(new->mark, entry->connector->obj,
 366                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 367                 fsnotify_put_mark(new->mark);
 368                 goto out_mutex;
 369         }
 370
 371         chunk->dead = 1;
 372         spin_lock(&hash_lock);
 373         /*
 374          * This has to go last when updating chunk as once replace_chunk() is
 375          * called, new RCU readers can see the new chunk.
 376          */
 377         replace_chunk(new, chunk);
 378         spin_unlock(&hash_lock);
 379         fsnotify_detach_mark(entry);
 380         mutex_unlock(&audit_tree_group->mark_mutex);
 381         fsnotify_free_mark(entry);
 382         fsnotify_put_mark(new->mark);   /* drop initial reference */
 383         return;
 384
 385 out_mutex:
 386         mutex_unlock(&audit_tree_group->mark_mutex);
 387 }
 388
 389 /* Call with group->mark_mutex held, releases it */
 390 static int create_chunk(struct inode *inode, struct audit_tree *tree)
 391 {
 392         struct fsnotify_mark *entry;
 393         struct audit_chunk *chunk = alloc_chunk(1);
 394
 395         if (!chunk) {
 396                 mutex_unlock(&audit_tree_group->mark_mutex);
 397                 return -ENOMEM;
 398         }
 399
 400         entry = chunk->mark;
 401         if (fsnotify_add_inode_mark_locked(entry, inode, 0)) {
 402                 mutex_unlock(&audit_tree_group->mark_mutex);
 403                 fsnotify_put_mark(entry);
 404                 return -ENOSPC;
 405         }
 406
 407         spin_lock(&hash_lock);
 408         if (tree->goner) {
 409                 spin_unlock(&hash_lock);
 410                 chunk->dead = 1;
 411                 fsnotify_detach_mark(entry);
 412                 mutex_unlock(&audit_tree_group->mark_mutex);
 413                 fsnotify_free_mark(entry);
 414                 fsnotify_put_mark(entry);
 415                 return 0;
 416         }
 417         chunk->owners[0].index = (1U << 31);
 418         chunk->owners[0].owner = tree;
 419         get_tree(tree);
 420         list_add(&chunk->owners[0].list, &tree->chunks);
 421         if (!tree->root) {
 422                 tree->root = chunk;
 423                 list_add(&tree->same_root, &chunk->trees);
 424         }
 425         chunk->key = inode_to_key(inode);
 426         /*
 427          * Inserting into the hash table has to go last as once we do that RCU
 428          * readers can see the chunk.
 429          */
 430         insert_hash(chunk);
 431         spin_unlock(&hash_lock);
 432         mutex_unlock(&audit_tree_group->mark_mutex);
 433         fsnotify_put_mark(entry);       /* drop initial reference */
 434         return 0;
 435 }
 436
 437 /* the first tagged inode becomes root of tree */
 438 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
 439 {
 440         struct fsnotify_mark *old_entry, *chunk_entry;
 441         struct audit_chunk *chunk, *old;
 442         struct node *p;
 443         int n;
 444
 445         mutex_lock(&audit_tree_group->mark_mutex);
 446         old_entry = fsnotify_find_mark(&inode->i_fsnotify_marks,
 447                                        audit_tree_group);
 448         if (!old_entry)
 449                 return create_chunk(inode, tree);
 450
 451         old = mark_chunk(old_entry);
 452
 453         /* are we already there? */
 454         spin_lock(&hash_lock);
 455         for (n = 0; n < old->count; n++) {
 456                 if (old->owners[n].owner == tree) {
 457                         spin_unlock(&hash_lock);
 458                         mutex_unlock(&audit_tree_group->mark_mutex);
 459                         fsnotify_put_mark(old_entry);
 460                         return 0;
 461                 }
 462         }
 463         spin_unlock(&hash_lock);
 464
 465         chunk = alloc_chunk(old->count + 1);
 466         if (!chunk) {
 467                 mutex_unlock(&audit_tree_group->mark_mutex);
 468                 fsnotify_put_mark(old_entry);
 469                 return -ENOMEM;
 470         }
 471
 472         chunk_entry = chunk->mark;
 473
 474         /*
 475          * mark_mutex protects mark from getting detached and thus also from
 476          * mark->connector->obj getting NULL.
 477          */
 478         if (!(old_entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 479                 /* old_entry is being shot, lets just lie */
 480                 mutex_unlock(&audit_tree_group->mark_mutex);
 481                 fsnotify_put_mark(old_entry);
 482                 fsnotify_put_mark(chunk->mark);
 483                 return -ENOENT;
 484         }
 485
 486         if (fsnotify_add_mark_locked(chunk_entry, old_entry->connector->obj,
 487                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 488                 mutex_unlock(&audit_tree_group->mark_mutex);
 489                 fsnotify_put_mark(chunk_entry);
 490                 fsnotify_put_mark(old_entry);
 491                 return -ENOSPC;
 492         }
 493
 494         spin_lock(&hash_lock);
 495         if (tree->goner) {
 496                 spin_unlock(&hash_lock);
 497                 chunk->dead = 1;
 498                 fsnotify_detach_mark(chunk_entry);
 499                 mutex_unlock(&audit_tree_group->mark_mutex);
 500                 fsnotify_free_mark(chunk_entry);
 501                 fsnotify_put_mark(chunk_entry);
 502                 fsnotify_put_mark(old_entry);
 503                 return 0;
 504         }
 505         p = &chunk->owners[chunk->count - 1];
 506         p->index = (chunk->count - 1) | (1U<<31);
 507         p->owner = tree;
 508         get_tree(tree);
 509         list_add(&p->list, &tree->chunks);
 510         old->dead = 1;
 511         if (!tree->root) {
 512                 tree->root = chunk;
 513                 list_add(&tree->same_root, &chunk->trees);
 514         }
 515         /*
 516          * This has to go last when updating chunk as once replace_chunk() is
 517          * called, new RCU readers can see the new chunk.
 518          */
 519         replace_chunk(chunk, old);
 520         spin_unlock(&hash_lock);
 521         fsnotify_detach_mark(old_entry);
 522         mutex_unlock(&audit_tree_group->mark_mutex);
 523         fsnotify_free_mark(old_entry);
 524         fsnotify_put_mark(chunk_entry); /* drop initial reference */
 525         fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
 526         return 0;
 527 }
 528
 529 static void audit_tree_log_remove_rule(struct audit_krule *rule)
 530 {
 531         struct audit_buffer *ab;
 532
 533         if (!audit_enabled)
 534                 return;
 535         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
 536         if (unlikely(!ab))
 537                 return;
 538         audit_log_format(ab, "op=remove_rule");
 539         audit_log_format(ab, " dir=");
 540         audit_log_untrustedstring(ab, rule->tree->pathname);
 541         audit_log_key(ab, rule->filterkey);
 542         audit_log_format(ab, " list=%d res=1", rule->listnr);
 543         audit_log_end(ab);
 544 }
 545
 546 static void kill_rules(struct audit_tree *tree)
 547 {
 548         struct audit_krule *rule, *next;
 549         struct audit_entry *entry;
 550
 551         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
 552                 entry = container_of(rule, struct audit_entry, rule);
 553
 554                 list_del_init(&rule->rlist);
 555                 if (rule->tree) {
 556                         /* not a half-baked one */
 557                         audit_tree_log_remove_rule(rule);
 558                         if (entry->rule.exe)
 559                                 audit_remove_mark(entry->rule.exe);
 560                         rule->tree = NULL;
 561                         list_del_rcu(&entry->list);
 562                         list_del(&entry->rule.list);
 563                         call_rcu(&entry->rcu, audit_free_rule_rcu);
 564                 }
 565         }
 566 }
 567
 568 /*
 569  * Remove tree from chunks. If 'tagged' is set, remove tree only from tagged
 570  * chunks. The function expects tagged chunks are all at the beginning of the
 571  * chunks list.
 572  */
 573 static void prune_tree_chunks(struct audit_tree *victim, bool tagged)
 574 {
 575         spin_lock(&hash_lock);
 576         while (!list_empty(&victim->chunks)) {
 577                 struct node *p;
 578                 struct audit_chunk *chunk;
 579                 struct fsnotify_mark *mark;
 580
 581                 p = list_first_entry(&victim->chunks, struct node, list);
 582                 /* have we run out of marked? */
 583                 if (tagged && !(p->index & (1U<<31)))
 584                         break;
 585                 chunk = find_chunk(p);
 586                 mark = chunk->mark;
 587                 remove_chunk_node(chunk, p);
 588                 fsnotify_get_mark(mark);
 589                 spin_unlock(&hash_lock);
 590
 591                 untag_chunk(chunk, mark);
 592                 fsnotify_put_mark(mark);
 593
 594                 spin_lock(&hash_lock);
 595         }
 596         spin_unlock(&hash_lock);
 597         put_tree(victim);
 598 }
 599
 600 /*
 601  * finish killing struct audit_tree
 602  */
 603 static void prune_one(struct audit_tree *victim)
 604 {
 605         prune_tree_chunks(victim, false);
 606 }
 607
 608 /* trim the uncommitted chunks from tree */
 609
 610 static void trim_marked(struct audit_tree *tree)
 611 {
 612         struct list_head *p, *q;
 613         spin_lock(&hash_lock);
 614         if (tree->goner) {
 615                 spin_unlock(&hash_lock);
 616                 return;
 617         }
 618         /* reorder */
 619         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
 620                 struct node *node = list_entry(p, struct node, list);
 621                 q = p->next;
 622                 if (node->index & (1U<<31)) {
 623                         list_del_init(p);
 624                         list_add(p, &tree->chunks);
 625                 }
 626         }
 627         spin_unlock(&hash_lock);
 628
 629         prune_tree_chunks(tree, true);
 630
 631         spin_lock(&hash_lock);
 632         if (!tree->root && !tree->goner) {
 633                 tree->goner = 1;
 634                 spin_unlock(&hash_lock);
 635                 mutex_lock(&audit_filter_mutex);
 636                 kill_rules(tree);
 637                 list_del_init(&tree->list);
 638                 mutex_unlock(&audit_filter_mutex);
 639                 prune_one(tree);
 640         } else {
 641                 spin_unlock(&hash_lock);
 642         }
 643 }
 644
 645 static void audit_schedule_prune(void);
 646
 647 /* called with audit_filter_mutex */
 648 int audit_remove_tree_rule(struct audit_krule *rule)
 649 {
 650         struct audit_tree *tree;
 651         tree = rule->tree;
 652         if (tree) {
 653                 spin_lock(&hash_lock);
 654                 list_del_init(&rule->rlist);
 655                 if (list_empty(&tree->rules) && !tree->goner) {
 656                         tree->root = NULL;
 657                         list_del_init(&tree->same_root);
 658                         tree->goner = 1;
 659                         list_move(&tree->list, &prune_list);
 660                         rule->tree = NULL;
 661                         spin_unlock(&hash_lock);
 662                         audit_schedule_prune();
 663                         return 1;
 664                 }
 665                 rule->tree = NULL;
 666                 spin_unlock(&hash_lock);
 667                 return 1;
 668         }
 669         return 0;
 670 }
 671
 672 static int compare_root(struct vfsmount *mnt, void *arg)
 673 {
 674         return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
 675                (unsigned long)arg;
 676 }
 677
 678 void audit_trim_trees(void)
 679 {
 680         struct list_head cursor;
 681
 682         mutex_lock(&audit_filter_mutex);
 683         list_add(&cursor, &tree_list);
 684         while (cursor.next != &tree_list) {
 685                 struct audit_tree *tree;
 686                 struct path path;
 687                 struct vfsmount *root_mnt;
 688                 struct node *node;
 689                 int err;
 690
 691                 tree = container_of(cursor.next, struct audit_tree, list);
 692                 get_tree(tree);
 693                 list_del(&cursor);
 694                 list_add(&cursor, &tree->list);
 695                 mutex_unlock(&audit_filter_mutex);
 696
 697                 err = kern_path(tree->pathname, 0, &path);
 698                 if (err)
 699                         goto skip_it;
 700
 701                 root_mnt = collect_mounts(&path);
 702                 path_put(&path);
 703                 if (IS_ERR(root_mnt))
 704                         goto skip_it;
 705
 706                 spin_lock(&hash_lock);
 707                 list_for_each_entry(node, &tree->chunks, list) {
 708                         struct audit_chunk *chunk = find_chunk(node);
 709                         /* this could be NULL if the watch is dying else where... */
 710                         node->index |= 1U<<31;
 711                         if (iterate_mounts(compare_root,
 712                                            (void *)(chunk->key),
 713                                            root_mnt))
 714                                 node->index &= ~(1U<<31);
 715                 }
 716                 spin_unlock(&hash_lock);
 717                 trim_marked(tree);
 718                 drop_collected_mounts(root_mnt);
 719 skip_it:
 720                 put_tree(tree);
 721                 mutex_lock(&audit_filter_mutex);
 722         }
 723         list_del(&cursor);
 724         mutex_unlock(&audit_filter_mutex);
 725 }
 726
 727 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
 728 {
 729
 730         if (pathname[0] != '/' ||
 731             rule->listnr != AUDIT_FILTER_EXIT ||
 732             op != Audit_equal ||
 733             rule->inode_f || rule->watch || rule->tree)
 734                 return -EINVAL;
 735         rule->tree = alloc_tree(pathname);
 736         if (!rule->tree)
 737                 return -ENOMEM;
 738         return 0;
 739 }
 740
 741 void audit_put_tree(struct audit_tree *tree)
 742 {
 743         put_tree(tree);
 744 }
 745
 746 static int tag_mount(struct vfsmount *mnt, void *arg)
 747 {
 748         return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
 749 }
 750
 751 /*
 752  * That gets run when evict_chunk() ends up needing to kill audit_tree.
 753  * Runs from a separate thread.
 754  */
 755 static int prune_tree_thread(void *unused)
 756 {
 757         for (;;) {
 758                 if (list_empty(&prune_list)) {
 759                         set_current_state(TASK_INTERRUPTIBLE);
 760                         schedule();
 761                 }
 762
 763                 audit_ctl_lock();
 764                 mutex_lock(&audit_filter_mutex);
 765
 766                 while (!list_empty(&prune_list)) {
 767                         struct audit_tree *victim;
 768
 769                         victim = list_entry(prune_list.next,
 770                                         struct audit_tree, list);
 771                         list_del_init(&victim->list);
 772
 773                         mutex_unlock(&audit_filter_mutex);
 774
 775                         prune_one(victim);
 776
 777                         mutex_lock(&audit_filter_mutex);
 778                 }
 779
 780                 mutex_unlock(&audit_filter_mutex);
 781                 audit_ctl_unlock();
 782         }
 783         return 0;
 784 }
 785
 786 static int audit_launch_prune(void)
 787 {
 788         if (prune_thread)
 789                 return 0;
 790         prune_thread = kthread_run(prune_tree_thread, NULL,
 791                                 "audit_prune_tree");
 792         if (IS_ERR(prune_thread)) {
 793                 pr_err("cannot start thread audit_prune_tree");
 794                 prune_thread = NULL;
 795                 return -ENOMEM;
 796         }
 797         return 0;
 798 }
 799
 800 /* called with audit_filter_mutex */
 801 int audit_add_tree_rule(struct audit_krule *rule)
 802 {
 803         struct audit_tree *seed = rule->tree, *tree;
 804         struct path path;
 805         struct vfsmount *mnt;
 806         int err;
 807
 808         rule->tree = NULL;
 809         list_for_each_entry(tree, &tree_list, list) {
 810                 if (!strcmp(seed->pathname, tree->pathname)) {
 811                         put_tree(seed);
 812                         rule->tree = tree;
 813                         list_add(&rule->rlist, &tree->rules);
 814                         return 0;
 815                 }
 816         }
 817         tree = seed;
 818         list_add(&tree->list, &tree_list);
 819         list_add(&rule->rlist, &tree->rules);
 820         /* do not set rule->tree yet */
 821         mutex_unlock(&audit_filter_mutex);
 822
 823         if (unlikely(!prune_thread)) {
 824                 err = audit_launch_prune();
 825                 if (err)
 826                         goto Err;
 827         }
 828
 829         err = kern_path(tree->pathname, 0, &path);
 830         if (err)
 831                 goto Err;
 832         mnt = collect_mounts(&path);
 833         path_put(&path);
 834         if (IS_ERR(mnt)) {
 835                 err = PTR_ERR(mnt);
 836                 goto Err;
 837         }
 838
 839         get_tree(tree);
 840         err = iterate_mounts(tag_mount, tree, mnt);
 841         drop_collected_mounts(mnt);
 842
 843         if (!err) {
 844                 struct node *node;
 845                 spin_lock(&hash_lock);
 846                 list_for_each_entry(node, &tree->chunks, list)
 847                         node->index &= ~(1U<<31);
 848                 spin_unlock(&hash_lock);
 849         } else {
 850                 trim_marked(tree);
 851                 goto Err;
 852         }
 853
 854         mutex_lock(&audit_filter_mutex);
 855         if (list_empty(&rule->rlist)) {
 856                 put_tree(tree);
 857                 return -ENOENT;
 858         }
 859         rule->tree = tree;
 860         put_tree(tree);
 861
 862         return 0;
 863 Err:
 864         mutex_lock(&audit_filter_mutex);
 865         list_del_init(&tree->list);
 866         list_del_init(&tree->rules);
 867         put_tree(tree);
 868         return err;
 869 }
 870
 871 int audit_tag_tree(char *old, char *new)
 872 {
 873         struct list_head cursor, barrier;
 874         int failed = 0;
 875         struct path path1, path2;
 876         struct vfsmount *tagged;
 877         int err;
 878
 879         err = kern_path(new, 0, &path2);
 880         if (err)
 881                 return err;
 882         tagged = collect_mounts(&path2);
 883         path_put(&path2);
 884         if (IS_ERR(tagged))
 885                 return PTR_ERR(tagged);
 886
 887         err = kern_path(old, 0, &path1);
 888         if (err) {
 889                 drop_collected_mounts(tagged);
 890                 return err;
 891         }
 892
 893         mutex_lock(&audit_filter_mutex);
 894         list_add(&barrier, &tree_list);
 895         list_add(&cursor, &barrier);
 896
 897         while (cursor.next != &tree_list) {
 898                 struct audit_tree *tree;
 899                 int good_one = 0;
 900
 901                 tree = container_of(cursor.next, struct audit_tree, list);
 902                 get_tree(tree);
 903                 list_del(&cursor);
 904                 list_add(&cursor, &tree->list);
 905                 mutex_unlock(&audit_filter_mutex);
 906
 907                 err = kern_path(tree->pathname, 0, &path2);
 908                 if (!err) {
 909                         good_one = path_is_under(&path1, &path2);
 910                         path_put(&path2);
 911                 }
 912
 913                 if (!good_one) {
 914                         put_tree(tree);
 915                         mutex_lock(&audit_filter_mutex);
 916                         continue;
 917                 }
 918
 919                 failed = iterate_mounts(tag_mount, tree, tagged);
 920                 if (failed) {
 921                         put_tree(tree);
 922                         mutex_lock(&audit_filter_mutex);
 923                         break;
 924                 }
 925
 926                 mutex_lock(&audit_filter_mutex);
 927                 spin_lock(&hash_lock);
 928                 if (!tree->goner) {
 929                         list_del(&tree->list);
 930                         list_add(&tree->list, &tree_list);
 931                 }
 932                 spin_unlock(&hash_lock);
 933                 put_tree(tree);
 934         }
 935
 936         while (barrier.prev != &tree_list) {
 937                 struct audit_tree *tree;
 938
 939                 tree = container_of(barrier.prev, struct audit_tree, list);
 940                 get_tree(tree);
 941                 list_del(&tree->list);
 942                 list_add(&tree->list, &barrier);
 943                 mutex_unlock(&audit_filter_mutex);
 944
 945                 if (!failed) {
 946                         struct node *node;
 947                         spin_lock(&hash_lock);
 948                         list_for_each_entry(node, &tree->chunks, list)
 949                                 node->index &= ~(1U<<31);
 950                         spin_unlock(&hash_lock);
 951                 } else {
 952                         trim_marked(tree);
 953                 }
 954
 955                 put_tree(tree);
 956                 mutex_lock(&audit_filter_mutex);
 957         }
 958         list_del(&barrier);
 959         list_del(&cursor);
 960         mutex_unlock(&audit_filter_mutex);
 961         path_put(&path1);
 962         drop_collected_mounts(tagged);
 963         return failed;
 964 }
 965
 966
 967 static void audit_schedule_prune(void)
 968 {
 969         wake_up_process(prune_thread);
 970 }
 971
 972 /*
 973  * ... and that one is done if evict_chunk() decides to delay until the end
 974  * of syscall.  Runs synchronously.
 975  */
 976 void audit_kill_trees(struct list_head *list)
 977 {
 978         audit_ctl_lock();
 979         mutex_lock(&audit_filter_mutex);
 980
 981         while (!list_empty(list)) {
 982                 struct audit_tree *victim;
 983
 984                 victim = list_entry(list->next, struct audit_tree, list);
 985                 kill_rules(victim);
 986                 list_del_init(&victim->list);
 987
 988                 mutex_unlock(&audit_filter_mutex);
 989
 990                 prune_one(victim);
 991
 992                 mutex_lock(&audit_filter_mutex);
 993         }
 994
 995         mutex_unlock(&audit_filter_mutex);
 996         audit_ctl_unlock();
 997 }
 998
 999 /*
1000  *  Here comes the stuff asynchronous to auditctl operations
1001  */
1002
1003 static void evict_chunk(struct audit_chunk *chunk)
1004 {
1005         struct audit_tree *owner;
1006         struct list_head *postponed = audit_killed_trees();
1007         int need_prune = 0;
1008         int n;
1009
1010         if (chunk->dead)
1011                 return;
1012
1013         chunk->dead = 1;
1014         mutex_lock(&audit_filter_mutex);
1015         spin_lock(&hash_lock);
1016         while (!list_empty(&chunk->trees)) {
1017                 owner = list_entry(chunk->trees.next,
1018                                    struct audit_tree, same_root);
1019                 owner->goner = 1;
1020                 owner->root = NULL;
1021                 list_del_init(&owner->same_root);
1022                 spin_unlock(&hash_lock);
1023                 if (!postponed) {
1024                         kill_rules(owner);
1025                         list_move(&owner->list, &prune_list);
1026                         need_prune = 1;
1027                 } else {
1028                         list_move(&owner->list, postponed);
1029                 }
1030                 spin_lock(&hash_lock);
1031         }
1032         list_del_rcu(&chunk->hash);
1033         for (n = 0; n < chunk->count; n++)
1034                 list_del_init(&chunk->owners[n].list);
1035         spin_unlock(&hash_lock);
1036         mutex_unlock(&audit_filter_mutex);
1037         if (need_prune)
1038                 audit_schedule_prune();
1039 }
1040
1041 static int audit_tree_handle_event(struct fsnotify_group *group,
1042                                    struct inode *to_tell,
1043                                    u32 mask, const void *data, int data_type,
1044                                    const unsigned char *file_name, u32 cookie,
1045                                    struct fsnotify_iter_info *iter_info)
1046 {
1047         return 0;
1048 }
1049
1050 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
1051 {
1052         struct audit_chunk *chunk = mark_chunk(entry);
1053
1054         evict_chunk(chunk);
1055
1056         /*
1057          * We are guaranteed to have at least one reference to the mark from
1058          * either the inode or the caller of fsnotify_destroy_mark().
1059          */
1060         BUG_ON(refcount_read(&entry->refcnt) < 1);
1061 }
1062
1063 static const struct fsnotify_ops audit_tree_ops = {
1064         .handle_event = audit_tree_handle_event,
1065         .freeing_mark = audit_tree_freeing_mark,
1066         .free_mark = audit_tree_destroy_watch,
1067 };
1068
1069 static int __init audit_tree_init(void)
1070 {
1071         int i;
1072
1073         audit_tree_mark_cachep = KMEM_CACHE(audit_tree_mark, SLAB_PANIC);
1074
1075         audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1076         if (IS_ERR(audit_tree_group))
1077                 audit_panic("cannot initialize fsnotify group for rectree watches");
1078
1079         for (i = 0; i < HASH_SIZE; i++)
1080                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
1081
1082         return 0;
1083 }
1084 __initcall(audit_tree_init);