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                           struct node *skip)
 282 {
 283         struct audit_tree *owner;
 284         int i, j;
 285
 286         new->key = old->key;
 287         list_splice_init(&old->trees, &new->trees);
 288         list_for_each_entry(owner, &new->trees, same_root)
 289                 owner->root = new;
 290         for (i = j = 0; j < old->count; i++, j++) {
 291                 if (&old->owners[j] == skip) {
 292                         i--;
 293                         continue;
 294                 }
 295                 owner = old->owners[j].owner;
 296                 new->owners[i].owner = owner;
 297                 new->owners[i].index = old->owners[j].index - j + i;
 298                 if (!owner) /* result of earlier fallback */
 299                         continue;
 300                 get_tree(owner);
 301                 list_replace_init(&old->owners[j].list, &new->owners[i].list);
 302         }
 303         /*
 304          * Make sure chunk is fully initialized before making it visible in the
 305          * hash. Pairs with a data dependency barrier in READ_ONCE() in
 306          * audit_tree_lookup().
 307          */
 308         smp_wmb();
 309         list_replace_rcu(&old->hash, &new->hash);
 310 }
 311
 312 static void remove_chunk_node(struct audit_chunk *chunk, struct node *p)
 313 {
 314         struct audit_tree *owner = p->owner;
 315
 316         if (owner->root == chunk) {
 317                 list_del_init(&owner->same_root);
 318                 owner->root = NULL;
 319         }
 320         list_del_init(&p->list);
 321         p->owner = NULL;
 322         put_tree(owner);
 323 }
 324
 325 static void untag_chunk(struct node *p)
 326 {
 327         struct audit_chunk *chunk = find_chunk(p);
 328         struct fsnotify_mark *entry = chunk->mark;
 329         struct audit_chunk *new = NULL;
 330         int size = chunk->count - 1;
 331
 332         remove_chunk_node(chunk, p);
 333         fsnotify_get_mark(entry);
 334         spin_unlock(&hash_lock);
 335
 336         if (size)
 337                 new = alloc_chunk(size);
 338
 339         mutex_lock(&entry->group->mark_mutex);
 340         /*
 341          * mark_mutex protects mark from getting detached and thus also from
 342          * mark->connector->obj getting NULL.
 343          */
 344         if (chunk->dead || !(entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 345                 mutex_unlock(&entry->group->mark_mutex);
 346                 if (new)
 347                         fsnotify_put_mark(new->mark);
 348                 goto out;
 349         }
 350
 351         if (!size) {
 352                 chunk->dead = 1;
 353                 spin_lock(&hash_lock);
 354                 list_del_init(&chunk->trees);
 355                 list_del_rcu(&chunk->hash);
 356                 spin_unlock(&hash_lock);
 357                 fsnotify_detach_mark(entry);
 358                 mutex_unlock(&entry->group->mark_mutex);
 359                 fsnotify_free_mark(entry);
 360                 goto out;
 361         }
 362
 363         if (!new)
 364                 goto out_mutex;
 365
 366         if (fsnotify_add_mark_locked(new->mark, entry->connector->obj,
 367                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 368                 fsnotify_put_mark(new->mark);
 369                 goto out_mutex;
 370         }
 371
 372         chunk->dead = 1;
 373         spin_lock(&hash_lock);
 374         /*
 375          * This has to go last when updating chunk as once replace_chunk() is
 376          * called, new RCU readers can see the new chunk.
 377          */
 378         replace_chunk(new, chunk, p);
 379         spin_unlock(&hash_lock);
 380         fsnotify_detach_mark(entry);
 381         mutex_unlock(&entry->group->mark_mutex);
 382         fsnotify_free_mark(entry);
 383         fsnotify_put_mark(new->mark);   /* drop initial reference */
 384         goto out;
 385
 386 out_mutex:
 387         mutex_unlock(&entry->group->mark_mutex);
 388 out:
 389         fsnotify_put_mark(entry);
 390         spin_lock(&hash_lock);
 391 }
 392
 393 /* Call with group->mark_mutex held, releases it */
 394 static int create_chunk(struct inode *inode, struct audit_tree *tree)
 395 {
 396         struct fsnotify_mark *entry;
 397         struct audit_chunk *chunk = alloc_chunk(1);
 398
 399         if (!chunk) {
 400                 mutex_unlock(&audit_tree_group->mark_mutex);
 401                 return -ENOMEM;
 402         }
 403
 404         entry = chunk->mark;
 405         if (fsnotify_add_inode_mark_locked(entry, inode, 0)) {
 406                 mutex_unlock(&audit_tree_group->mark_mutex);
 407                 fsnotify_put_mark(entry);
 408                 return -ENOSPC;
 409         }
 410
 411         spin_lock(&hash_lock);
 412         if (tree->goner) {
 413                 spin_unlock(&hash_lock);
 414                 chunk->dead = 1;
 415                 fsnotify_detach_mark(entry);
 416                 mutex_unlock(&audit_tree_group->mark_mutex);
 417                 fsnotify_free_mark(entry);
 418                 fsnotify_put_mark(entry);
 419                 return 0;
 420         }
 421         chunk->owners[0].index = (1U << 31);
 422         chunk->owners[0].owner = tree;
 423         get_tree(tree);
 424         list_add(&chunk->owners[0].list, &tree->chunks);
 425         if (!tree->root) {
 426                 tree->root = chunk;
 427                 list_add(&tree->same_root, &chunk->trees);
 428         }
 429         chunk->key = inode_to_key(inode);
 430         /*
 431          * Inserting into the hash table has to go last as once we do that RCU
 432          * readers can see the chunk.
 433          */
 434         insert_hash(chunk);
 435         spin_unlock(&hash_lock);
 436         mutex_unlock(&audit_tree_group->mark_mutex);
 437         fsnotify_put_mark(entry);       /* drop initial reference */
 438         return 0;
 439 }
 440
 441 /* the first tagged inode becomes root of tree */
 442 static int tag_chunk(struct inode *inode, struct audit_tree *tree)
 443 {
 444         struct fsnotify_mark *old_entry, *chunk_entry;
 445         struct audit_chunk *chunk, *old;
 446         struct node *p;
 447         int n;
 448
 449         mutex_lock(&audit_tree_group->mark_mutex);
 450         old_entry = fsnotify_find_mark(&inode->i_fsnotify_marks,
 451                                        audit_tree_group);
 452         if (!old_entry)
 453                 return create_chunk(inode, tree);
 454
 455         old = mark_chunk(old_entry);
 456
 457         /* are we already there? */
 458         spin_lock(&hash_lock);
 459         for (n = 0; n < old->count; n++) {
 460                 if (old->owners[n].owner == tree) {
 461                         spin_unlock(&hash_lock);
 462                         mutex_unlock(&audit_tree_group->mark_mutex);
 463                         fsnotify_put_mark(old_entry);
 464                         return 0;
 465                 }
 466         }
 467         spin_unlock(&hash_lock);
 468
 469         chunk = alloc_chunk(old->count + 1);
 470         if (!chunk) {
 471                 mutex_unlock(&audit_tree_group->mark_mutex);
 472                 fsnotify_put_mark(old_entry);
 473                 return -ENOMEM;
 474         }
 475
 476         chunk_entry = chunk->mark;
 477
 478         /*
 479          * mark_mutex protects mark from getting detached and thus also from
 480          * mark->connector->obj getting NULL.
 481          */
 482         if (!(old_entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
 483                 /* old_entry is being shot, lets just lie */
 484                 mutex_unlock(&audit_tree_group->mark_mutex);
 485                 fsnotify_put_mark(old_entry);
 486                 fsnotify_put_mark(chunk->mark);
 487                 return -ENOENT;
 488         }
 489
 490         if (fsnotify_add_mark_locked(chunk_entry, old_entry->connector->obj,
 491                                      FSNOTIFY_OBJ_TYPE_INODE, 1)) {
 492                 mutex_unlock(&audit_tree_group->mark_mutex);
 493                 fsnotify_put_mark(chunk_entry);
 494                 fsnotify_put_mark(old_entry);
 495                 return -ENOSPC;
 496         }
 497
 498         spin_lock(&hash_lock);
 499         if (tree->goner) {
 500                 spin_unlock(&hash_lock);
 501                 chunk->dead = 1;
 502                 fsnotify_detach_mark(chunk_entry);
 503                 mutex_unlock(&audit_tree_group->mark_mutex);
 504                 fsnotify_free_mark(chunk_entry);
 505                 fsnotify_put_mark(chunk_entry);
 506                 fsnotify_put_mark(old_entry);
 507                 return 0;
 508         }
 509         p = &chunk->owners[chunk->count - 1];
 510         p->index = (chunk->count - 1) | (1U<<31);
 511         p->owner = tree;
 512         get_tree(tree);
 513         list_add(&p->list, &tree->chunks);
 514         old->dead = 1;
 515         if (!tree->root) {
 516                 tree->root = chunk;
 517                 list_add(&tree->same_root, &chunk->trees);
 518         }
 519         /*
 520          * This has to go last when updating chunk as once replace_chunk() is
 521          * called, new RCU readers can see the new chunk.
 522          */
 523         replace_chunk(chunk, old, NULL);
 524         spin_unlock(&hash_lock);
 525         fsnotify_detach_mark(old_entry);
 526         mutex_unlock(&audit_tree_group->mark_mutex);
 527         fsnotify_free_mark(old_entry);
 528         fsnotify_put_mark(chunk_entry); /* drop initial reference */
 529         fsnotify_put_mark(old_entry); /* pair to fsnotify_find mark_entry */
 530         return 0;
 531 }
 532
 533 static void audit_tree_log_remove_rule(struct audit_krule *rule)
 534 {
 535         struct audit_buffer *ab;
 536
 537         if (!audit_enabled)
 538                 return;
 539         ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
 540         if (unlikely(!ab))
 541                 return;
 542         audit_log_format(ab, "op=remove_rule");
 543         audit_log_format(ab, " dir=");
 544         audit_log_untrustedstring(ab, rule->tree->pathname);
 545         audit_log_key(ab, rule->filterkey);
 546         audit_log_format(ab, " list=%d res=1", rule->listnr);
 547         audit_log_end(ab);
 548 }
 549
 550 static void kill_rules(struct audit_tree *tree)
 551 {
 552         struct audit_krule *rule, *next;
 553         struct audit_entry *entry;
 554
 555         list_for_each_entry_safe(rule, next, &tree->rules, rlist) {
 556                 entry = container_of(rule, struct audit_entry, rule);
 557
 558                 list_del_init(&rule->rlist);
 559                 if (rule->tree) {
 560                         /* not a half-baked one */
 561                         audit_tree_log_remove_rule(rule);
 562                         if (entry->rule.exe)
 563                                 audit_remove_mark(entry->rule.exe);
 564                         rule->tree = NULL;
 565                         list_del_rcu(&entry->list);
 566                         list_del(&entry->rule.list);
 567                         call_rcu(&entry->rcu, audit_free_rule_rcu);
 568                 }
 569         }
 570 }
 571
 572 /*
 573  * finish killing struct audit_tree
 574  */
 575 static void prune_one(struct audit_tree *victim)
 576 {
 577         spin_lock(&hash_lock);
 578         while (!list_empty(&victim->chunks)) {
 579                 struct node *p;
 580
 581                 p = list_entry(victim->chunks.next, struct node, list);
 582
 583                 untag_chunk(p);
 584         }
 585         spin_unlock(&hash_lock);
 586         put_tree(victim);
 587 }
 588
 589 /* trim the uncommitted chunks from tree */
 590
 591 static void trim_marked(struct audit_tree *tree)
 592 {
 593         struct list_head *p, *q;
 594         spin_lock(&hash_lock);
 595         if (tree->goner) {
 596                 spin_unlock(&hash_lock);
 597                 return;
 598         }
 599         /* reorder */
 600         for (p = tree->chunks.next; p != &tree->chunks; p = q) {
 601                 struct node *node = list_entry(p, struct node, list);
 602                 q = p->next;
 603                 if (node->index & (1U<<31)) {
 604                         list_del_init(p);
 605                         list_add(p, &tree->chunks);
 606                 }
 607         }
 608
 609         while (!list_empty(&tree->chunks)) {
 610                 struct node *node;
 611
 612                 node = list_entry(tree->chunks.next, struct node, list);
 613
 614                 /* have we run out of marked? */
 615                 if (!(node->index & (1U<<31)))
 616                         break;
 617
 618                 untag_chunk(node);
 619         }
 620         if (!tree->root && !tree->goner) {
 621                 tree->goner = 1;
 622                 spin_unlock(&hash_lock);
 623                 mutex_lock(&audit_filter_mutex);
 624                 kill_rules(tree);
 625                 list_del_init(&tree->list);
 626                 mutex_unlock(&audit_filter_mutex);
 627                 prune_one(tree);
 628         } else {
 629                 spin_unlock(&hash_lock);
 630         }
 631 }
 632
 633 static void audit_schedule_prune(void);
 634
 635 /* called with audit_filter_mutex */
 636 int audit_remove_tree_rule(struct audit_krule *rule)
 637 {
 638         struct audit_tree *tree;
 639         tree = rule->tree;
 640         if (tree) {
 641                 spin_lock(&hash_lock);
 642                 list_del_init(&rule->rlist);
 643                 if (list_empty(&tree->rules) && !tree->goner) {
 644                         tree->root = NULL;
 645                         list_del_init(&tree->same_root);
 646                         tree->goner = 1;
 647                         list_move(&tree->list, &prune_list);
 648                         rule->tree = NULL;
 649                         spin_unlock(&hash_lock);
 650                         audit_schedule_prune();
 651                         return 1;
 652                 }
 653                 rule->tree = NULL;
 654                 spin_unlock(&hash_lock);
 655                 return 1;
 656         }
 657         return 0;
 658 }
 659
 660 static int compare_root(struct vfsmount *mnt, void *arg)
 661 {
 662         return inode_to_key(d_backing_inode(mnt->mnt_root)) ==
 663                (unsigned long)arg;
 664 }
 665
 666 void audit_trim_trees(void)
 667 {
 668         struct list_head cursor;
 669
 670         mutex_lock(&audit_filter_mutex);
 671         list_add(&cursor, &tree_list);
 672         while (cursor.next != &tree_list) {
 673                 struct audit_tree *tree;
 674                 struct path path;
 675                 struct vfsmount *root_mnt;
 676                 struct node *node;
 677                 int err;
 678
 679                 tree = container_of(cursor.next, struct audit_tree, list);
 680                 get_tree(tree);
 681                 list_del(&cursor);
 682                 list_add(&cursor, &tree->list);
 683                 mutex_unlock(&audit_filter_mutex);
 684
 685                 err = kern_path(tree->pathname, 0, &path);
 686                 if (err)
 687                         goto skip_it;
 688
 689                 root_mnt = collect_mounts(&path);
 690                 path_put(&path);
 691                 if (IS_ERR(root_mnt))
 692                         goto skip_it;
 693
 694                 spin_lock(&hash_lock);
 695                 list_for_each_entry(node, &tree->chunks, list) {
 696                         struct audit_chunk *chunk = find_chunk(node);
 697                         /* this could be NULL if the watch is dying else where... */
 698                         node->index |= 1U<<31;
 699                         if (iterate_mounts(compare_root,
 700                                            (void *)(chunk->key),
 701                                            root_mnt))
 702                                 node->index &= ~(1U<<31);
 703                 }
 704                 spin_unlock(&hash_lock);
 705                 trim_marked(tree);
 706                 drop_collected_mounts(root_mnt);
 707 skip_it:
 708                 put_tree(tree);
 709                 mutex_lock(&audit_filter_mutex);
 710         }
 711         list_del(&cursor);
 712         mutex_unlock(&audit_filter_mutex);
 713 }
 714
 715 int audit_make_tree(struct audit_krule *rule, char *pathname, u32 op)
 716 {
 717
 718         if (pathname[0] != '/' ||
 719             rule->listnr != AUDIT_FILTER_EXIT ||
 720             op != Audit_equal ||
 721             rule->inode_f || rule->watch || rule->tree)
 722                 return -EINVAL;
 723         rule->tree = alloc_tree(pathname);
 724         if (!rule->tree)
 725                 return -ENOMEM;
 726         return 0;
 727 }
 728
 729 void audit_put_tree(struct audit_tree *tree)
 730 {
 731         put_tree(tree);
 732 }
 733
 734 static int tag_mount(struct vfsmount *mnt, void *arg)
 735 {
 736         return tag_chunk(d_backing_inode(mnt->mnt_root), arg);
 737 }
 738
 739 /*
 740  * That gets run when evict_chunk() ends up needing to kill audit_tree.
 741  * Runs from a separate thread.
 742  */
 743 static int prune_tree_thread(void *unused)
 744 {
 745         for (;;) {
 746                 if (list_empty(&prune_list)) {
 747                         set_current_state(TASK_INTERRUPTIBLE);
 748                         schedule();
 749                 }
 750
 751                 audit_ctl_lock();
 752                 mutex_lock(&audit_filter_mutex);
 753
 754                 while (!list_empty(&prune_list)) {
 755                         struct audit_tree *victim;
 756
 757                         victim = list_entry(prune_list.next,
 758                                         struct audit_tree, list);
 759                         list_del_init(&victim->list);
 760
 761                         mutex_unlock(&audit_filter_mutex);
 762
 763                         prune_one(victim);
 764
 765                         mutex_lock(&audit_filter_mutex);
 766                 }
 767
 768                 mutex_unlock(&audit_filter_mutex);
 769                 audit_ctl_unlock();
 770         }
 771         return 0;
 772 }
 773
 774 static int audit_launch_prune(void)
 775 {
 776         if (prune_thread)
 777                 return 0;
 778         prune_thread = kthread_run(prune_tree_thread, NULL,
 779                                 "audit_prune_tree");
 780         if (IS_ERR(prune_thread)) {
 781                 pr_err("cannot start thread audit_prune_tree");
 782                 prune_thread = NULL;
 783                 return -ENOMEM;
 784         }
 785         return 0;
 786 }
 787
 788 /* called with audit_filter_mutex */
 789 int audit_add_tree_rule(struct audit_krule *rule)
 790 {
 791         struct audit_tree *seed = rule->tree, *tree;
 792         struct path path;
 793         struct vfsmount *mnt;
 794         int err;
 795
 796         rule->tree = NULL;
 797         list_for_each_entry(tree, &tree_list, list) {
 798                 if (!strcmp(seed->pathname, tree->pathname)) {
 799                         put_tree(seed);
 800                         rule->tree = tree;
 801                         list_add(&rule->rlist, &tree->rules);
 802                         return 0;
 803                 }
 804         }
 805         tree = seed;
 806         list_add(&tree->list, &tree_list);
 807         list_add(&rule->rlist, &tree->rules);
 808         /* do not set rule->tree yet */
 809         mutex_unlock(&audit_filter_mutex);
 810
 811         if (unlikely(!prune_thread)) {
 812                 err = audit_launch_prune();
 813                 if (err)
 814                         goto Err;
 815         }
 816
 817         err = kern_path(tree->pathname, 0, &path);
 818         if (err)
 819                 goto Err;
 820         mnt = collect_mounts(&path);
 821         path_put(&path);
 822         if (IS_ERR(mnt)) {
 823                 err = PTR_ERR(mnt);
 824                 goto Err;
 825         }
 826
 827         get_tree(tree);
 828         err = iterate_mounts(tag_mount, tree, mnt);
 829         drop_collected_mounts(mnt);
 830
 831         if (!err) {
 832                 struct node *node;
 833                 spin_lock(&hash_lock);
 834                 list_for_each_entry(node, &tree->chunks, list)
 835                         node->index &= ~(1U<<31);
 836                 spin_unlock(&hash_lock);
 837         } else {
 838                 trim_marked(tree);
 839                 goto Err;
 840         }
 841
 842         mutex_lock(&audit_filter_mutex);
 843         if (list_empty(&rule->rlist)) {
 844                 put_tree(tree);
 845                 return -ENOENT;
 846         }
 847         rule->tree = tree;
 848         put_tree(tree);
 849
 850         return 0;
 851 Err:
 852         mutex_lock(&audit_filter_mutex);
 853         list_del_init(&tree->list);
 854         list_del_init(&tree->rules);
 855         put_tree(tree);
 856         return err;
 857 }
 858
 859 int audit_tag_tree(char *old, char *new)
 860 {
 861         struct list_head cursor, barrier;
 862         int failed = 0;
 863         struct path path1, path2;
 864         struct vfsmount *tagged;
 865         int err;
 866
 867         err = kern_path(new, 0, &path2);
 868         if (err)
 869                 return err;
 870         tagged = collect_mounts(&path2);
 871         path_put(&path2);
 872         if (IS_ERR(tagged))
 873                 return PTR_ERR(tagged);
 874
 875         err = kern_path(old, 0, &path1);
 876         if (err) {
 877                 drop_collected_mounts(tagged);
 878                 return err;
 879         }
 880
 881         mutex_lock(&audit_filter_mutex);
 882         list_add(&barrier, &tree_list);
 883         list_add(&cursor, &barrier);
 884
 885         while (cursor.next != &tree_list) {
 886                 struct audit_tree *tree;
 887                 int good_one = 0;
 888
 889                 tree = container_of(cursor.next, struct audit_tree, list);
 890                 get_tree(tree);
 891                 list_del(&cursor);
 892                 list_add(&cursor, &tree->list);
 893                 mutex_unlock(&audit_filter_mutex);
 894
 895                 err = kern_path(tree->pathname, 0, &path2);
 896                 if (!err) {
 897                         good_one = path_is_under(&path1, &path2);
 898                         path_put(&path2);
 899                 }
 900
 901                 if (!good_one) {
 902                         put_tree(tree);
 903                         mutex_lock(&audit_filter_mutex);
 904                         continue;
 905                 }
 906
 907                 failed = iterate_mounts(tag_mount, tree, tagged);
 908                 if (failed) {
 909                         put_tree(tree);
 910                         mutex_lock(&audit_filter_mutex);
 911                         break;
 912                 }
 913
 914                 mutex_lock(&audit_filter_mutex);
 915                 spin_lock(&hash_lock);
 916                 if (!tree->goner) {
 917                         list_del(&tree->list);
 918                         list_add(&tree->list, &tree_list);
 919                 }
 920                 spin_unlock(&hash_lock);
 921                 put_tree(tree);
 922         }
 923
 924         while (barrier.prev != &tree_list) {
 925                 struct audit_tree *tree;
 926
 927                 tree = container_of(barrier.prev, struct audit_tree, list);
 928                 get_tree(tree);
 929                 list_del(&tree->list);
 930                 list_add(&tree->list, &barrier);
 931                 mutex_unlock(&audit_filter_mutex);
 932
 933                 if (!failed) {
 934                         struct node *node;
 935                         spin_lock(&hash_lock);
 936                         list_for_each_entry(node, &tree->chunks, list)
 937                                 node->index &= ~(1U<<31);
 938                         spin_unlock(&hash_lock);
 939                 } else {
 940                         trim_marked(tree);
 941                 }
 942
 943                 put_tree(tree);
 944                 mutex_lock(&audit_filter_mutex);
 945         }
 946         list_del(&barrier);
 947         list_del(&cursor);
 948         mutex_unlock(&audit_filter_mutex);
 949         path_put(&path1);
 950         drop_collected_mounts(tagged);
 951         return failed;
 952 }
 953
 954
 955 static void audit_schedule_prune(void)
 956 {
 957         wake_up_process(prune_thread);
 958 }
 959
 960 /*
 961  * ... and that one is done if evict_chunk() decides to delay until the end
 962  * of syscall.  Runs synchronously.
 963  */
 964 void audit_kill_trees(struct list_head *list)
 965 {
 966         audit_ctl_lock();
 967         mutex_lock(&audit_filter_mutex);
 968
 969         while (!list_empty(list)) {
 970                 struct audit_tree *victim;
 971
 972                 victim = list_entry(list->next, struct audit_tree, list);
 973                 kill_rules(victim);
 974                 list_del_init(&victim->list);
 975
 976                 mutex_unlock(&audit_filter_mutex);
 977
 978                 prune_one(victim);
 979
 980                 mutex_lock(&audit_filter_mutex);
 981         }
 982
 983         mutex_unlock(&audit_filter_mutex);
 984         audit_ctl_unlock();
 985 }
 986
 987 /*
 988  *  Here comes the stuff asynchronous to auditctl operations
 989  */
 990
 991 static void evict_chunk(struct audit_chunk *chunk)
 992 {
 993         struct audit_tree *owner;
 994         struct list_head *postponed = audit_killed_trees();
 995         int need_prune = 0;
 996         int n;
 997
 998         if (chunk->dead)
 999                 return;
1000
1001         chunk->dead = 1;
1002         mutex_lock(&audit_filter_mutex);
1003         spin_lock(&hash_lock);
1004         while (!list_empty(&chunk->trees)) {
1005                 owner = list_entry(chunk->trees.next,
1006                                    struct audit_tree, same_root);
1007                 owner->goner = 1;
1008                 owner->root = NULL;
1009                 list_del_init(&owner->same_root);
1010                 spin_unlock(&hash_lock);
1011                 if (!postponed) {
1012                         kill_rules(owner);
1013                         list_move(&owner->list, &prune_list);
1014                         need_prune = 1;
1015                 } else {
1016                         list_move(&owner->list, postponed);
1017                 }
1018                 spin_lock(&hash_lock);
1019         }
1020         list_del_rcu(&chunk->hash);
1021         for (n = 0; n < chunk->count; n++)
1022                 list_del_init(&chunk->owners[n].list);
1023         spin_unlock(&hash_lock);
1024         mutex_unlock(&audit_filter_mutex);
1025         if (need_prune)
1026                 audit_schedule_prune();
1027 }
1028
1029 static int audit_tree_handle_event(struct fsnotify_group *group,
1030                                    struct inode *to_tell,
1031                                    u32 mask, const void *data, int data_type,
1032                                    const unsigned char *file_name, u32 cookie,
1033                                    struct fsnotify_iter_info *iter_info)
1034 {
1035         return 0;
1036 }
1037
1038 static void audit_tree_freeing_mark(struct fsnotify_mark *entry, struct fsnotify_group *group)
1039 {
1040         struct audit_chunk *chunk = mark_chunk(entry);
1041
1042         evict_chunk(chunk);
1043
1044         /*
1045          * We are guaranteed to have at least one reference to the mark from
1046          * either the inode or the caller of fsnotify_destroy_mark().
1047          */
1048         BUG_ON(refcount_read(&entry->refcnt) < 1);
1049 }
1050
1051 static const struct fsnotify_ops audit_tree_ops = {
1052         .handle_event = audit_tree_handle_event,
1053         .freeing_mark = audit_tree_freeing_mark,
1054         .free_mark = audit_tree_destroy_watch,
1055 };
1056
1057 static int __init audit_tree_init(void)
1058 {
1059         int i;
1060
1061         audit_tree_mark_cachep = KMEM_CACHE(audit_tree_mark, SLAB_PANIC);
1062
1063         audit_tree_group = fsnotify_alloc_group(&audit_tree_ops);
1064         if (IS_ERR(audit_tree_group))
1065                 audit_panic("cannot initialize fsnotify group for rectree watches");
1066
1067         for (i = 0; i < HASH_SIZE; i++)
1068                 INIT_LIST_HEAD(&chunk_hash_heads[i]);
1069
1070         return 0;
1071 }
1072 __initcall(audit_tree_init);