fs/btrfs/backref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/rbtree.h>
   8 #include <trace/events/btrfs.h>
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "backref.h"
  12 #include "ulist.h"
  13 #include "transaction.h"
  14 #include "delayed-ref.h"
  15 #include "locking.h"
  16 #include "misc.h"
  17
  18 /* Just an arbitrary number so we can be sure this happened */
  19 #define BACKREF_FOUND_SHARED 6
  20
  21 struct extent_inode_elem {
  22         u64 inum;
  23         u64 offset;
  24         struct extent_inode_elem *next;
  25 };
  26
  27 static int check_extent_in_eb(const struct btrfs_key *key,
  28                               const struct extent_buffer *eb,
  29                               const struct btrfs_file_extent_item *fi,
  30                               u64 extent_item_pos,
  31                               struct extent_inode_elem **eie,
  32                               bool ignore_offset)
  33 {
  34         u64 offset = 0;
  35         struct extent_inode_elem *e;
  36
  37         if (!ignore_offset &&
  38             !btrfs_file_extent_compression(eb, fi) &&
  39             !btrfs_file_extent_encryption(eb, fi) &&
  40             !btrfs_file_extent_other_encoding(eb, fi)) {
  41                 u64 data_offset;
  42                 u64 data_len;
  43
  44                 data_offset = btrfs_file_extent_offset(eb, fi);
  45                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  46
  47                 if (extent_item_pos < data_offset ||
  48                     extent_item_pos >= data_offset + data_len)
  49                         return 1;
  50                 offset = extent_item_pos - data_offset;
  51         }
  52
  53         e = kmalloc(sizeof(*e), GFP_NOFS);
  54         if (!e)
  55                 return -ENOMEM;
  56
  57         e->next = *eie;
  58         e->inum = key->objectid;
  59         e->offset = key->offset + offset;
  60         *eie = e;
  61
  62         return 0;
  63 }
  64
  65 static void free_inode_elem_list(struct extent_inode_elem *eie)
  66 {
  67         struct extent_inode_elem *eie_next;
  68
  69         for (; eie; eie = eie_next) {
  70                 eie_next = eie->next;
  71                 kfree(eie);
  72         }
  73 }
  74
  75 static int find_extent_in_eb(const struct extent_buffer *eb,
  76                              u64 wanted_disk_byte, u64 extent_item_pos,
  77                              struct extent_inode_elem **eie,
  78                              bool ignore_offset)
  79 {
  80         u64 disk_byte;
  81         struct btrfs_key key;
  82         struct btrfs_file_extent_item *fi;
  83         int slot;
  84         int nritems;
  85         int extent_type;
  86         int ret;
  87
  88         /*
  89          * from the shared data ref, we only have the leaf but we need
  90          * the key. thus, we must look into all items and see that we
  91          * find one (some) with a reference to our extent item.
  92          */
  93         nritems = btrfs_header_nritems(eb);
  94         for (slot = 0; slot < nritems; ++slot) {
  95                 btrfs_item_key_to_cpu(eb, &key, slot);
  96                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  97                         continue;
  98                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  99                 extent_type = btrfs_file_extent_type(eb, fi);
 100                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 101                         continue;
 102                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 103                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 104                 if (disk_byte != wanted_disk_byte)
 105                         continue;
 106
 107                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 108                 if (ret < 0)
 109                         return ret;
 110         }
 111
 112         return 0;
 113 }
 114
 115 struct preftree {
 116         struct rb_root_cached root;
 117         unsigned int count;
 118 };
 119
 120 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 121
 122 struct preftrees {
 123         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 124         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 125         struct preftree indirect_missing_keys;
 126 };
 127
 128 /*
 129  * Checks for a shared extent during backref search.
 130  *
 131  * The share_count tracks prelim_refs (direct and indirect) having a
 132  * ref->count >0:
 133  *  - incremented when a ref->count transitions to >0
 134  *  - decremented when a ref->count transitions to <1
 135  */
 136 struct share_check {
 137         u64 root_objectid;
 138         u64 inum;
 139         int share_count;
 140 };
 141
 142 static inline int extent_is_shared(struct share_check *sc)
 143 {
 144         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 145 }
 146
 147 static struct kmem_cache *btrfs_prelim_ref_cache;
 148
 149 int __init btrfs_prelim_ref_init(void)
 150 {
 151         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 152                                         sizeof(struct prelim_ref),
 153                                         0,
 154                                         SLAB_MEM_SPREAD,
 155                                         NULL);
 156         if (!btrfs_prelim_ref_cache)
 157                 return -ENOMEM;
 158         return 0;
 159 }
 160
 161 void __cold btrfs_prelim_ref_exit(void)
 162 {
 163         kmem_cache_destroy(btrfs_prelim_ref_cache);
 164 }
 165
 166 static void free_pref(struct prelim_ref *ref)
 167 {
 168         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 169 }
 170
 171 /*
 172  * Return 0 when both refs are for the same block (and can be merged).
 173  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 174  * indicates a 'higher' block.
 175  */
 176 static int prelim_ref_compare(struct prelim_ref *ref1,
 177                               struct prelim_ref *ref2)
 178 {
 179         if (ref1->level < ref2->level)
 180                 return -1;
 181         if (ref1->level > ref2->level)
 182                 return 1;
 183         if (ref1->root_id < ref2->root_id)
 184                 return -1;
 185         if (ref1->root_id > ref2->root_id)
 186                 return 1;
 187         if (ref1->key_for_search.type < ref2->key_for_search.type)
 188                 return -1;
 189         if (ref1->key_for_search.type > ref2->key_for_search.type)
 190                 return 1;
 191         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 192                 return -1;
 193         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 194                 return 1;
 195         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 196                 return -1;
 197         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 198                 return 1;
 199         if (ref1->parent < ref2->parent)
 200                 return -1;
 201         if (ref1->parent > ref2->parent)
 202                 return 1;
 203
 204         return 0;
 205 }
 206
 207 static void update_share_count(struct share_check *sc, int oldcount,
 208                                int newcount)
 209 {
 210         if ((!sc) || (oldcount == 0 && newcount < 1))
 211                 return;
 212
 213         if (oldcount > 0 && newcount < 1)
 214                 sc->share_count--;
 215         else if (oldcount < 1 && newcount > 0)
 216                 sc->share_count++;
 217 }
 218
 219 /*
 220  * Add @newref to the @root rbtree, merging identical refs.
 221  *
 222  * Callers should assume that newref has been freed after calling.
 223  */
 224 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 225                               struct preftree *preftree,
 226                               struct prelim_ref *newref,
 227                               struct share_check *sc)
 228 {
 229         struct rb_root_cached *root;
 230         struct rb_node **p;
 231         struct rb_node *parent = NULL;
 232         struct prelim_ref *ref;
 233         int result;
 234         bool leftmost = true;
 235
 236         root = &preftree->root;
 237         p = &root->rb_root.rb_node;
 238
 239         while (*p) {
 240                 parent = *p;
 241                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 242                 result = prelim_ref_compare(ref, newref);
 243                 if (result < 0) {
 244                         p = &(*p)->rb_left;
 245                 } else if (result > 0) {
 246                         p = &(*p)->rb_right;
 247                         leftmost = false;
 248                 } else {
 249                         /* Identical refs, merge them and free @newref */
 250                         struct extent_inode_elem *eie = ref->inode_list;
 251
 252                         while (eie && eie->next)
 253                                 eie = eie->next;
 254
 255                         if (!eie)
 256                                 ref->inode_list = newref->inode_list;
 257                         else
 258                                 eie->next = newref->inode_list;
 259                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 260                                                      preftree->count);
 261                         /*
 262                          * A delayed ref can have newref->count < 0.
 263                          * The ref->count is updated to follow any
 264                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 265                          */
 266                         update_share_count(sc, ref->count,
 267                                            ref->count + newref->count);
 268                         ref->count += newref->count;
 269                         free_pref(newref);
 270                         return;
 271                 }
 272         }
 273
 274         update_share_count(sc, 0, newref->count);
 275         preftree->count++;
 276         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 277         rb_link_node(&newref->rbnode, parent, p);
 278         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 279 }
 280
 281 /*
 282  * Release the entire tree.  We don't care about internal consistency so
 283  * just free everything and then reset the tree root.
 284  */
 285 static void prelim_release(struct preftree *preftree)
 286 {
 287         struct prelim_ref *ref, *next_ref;
 288
 289         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 290                                              &preftree->root.rb_root, rbnode)
 291                 free_pref(ref);
 292
 293         preftree->root = RB_ROOT_CACHED;
 294         preftree->count = 0;
 295 }
 296
 297 /*
 298  * the rules for all callers of this function are:
 299  * - obtaining the parent is the goal
 300  * - if you add a key, you must know that it is a correct key
 301  * - if you cannot add the parent or a correct key, then we will look into the
 302  *   block later to set a correct key
 303  *
 304  * delayed refs
 305  * ============
 306  *        backref type | shared | indirect | shared | indirect
 307  * information         |   tree |     tree |   data |     data
 308  * --------------------+--------+----------+--------+----------
 309  *      parent logical |    y   |     -    |    -   |     -
 310  *      key to resolve |    -   |     y    |    y   |     y
 311  *  tree block logical |    -   |     -    |    -   |     -
 312  *  root for resolving |    y   |     y    |    y   |     y
 313  *
 314  * - column 1:       we've the parent -> done
 315  * - column 2, 3, 4: we use the key to find the parent
 316  *
 317  * on disk refs (inline or keyed)
 318  * ==============================
 319  *        backref type | shared | indirect | shared | indirect
 320  * information         |   tree |     tree |   data |     data
 321  * --------------------+--------+----------+--------+----------
 322  *      parent logical |    y   |     -    |    y   |     -
 323  *      key to resolve |    -   |     -    |    -   |     y
 324  *  tree block logical |    y   |     y    |    y   |     y
 325  *  root for resolving |    -   |     y    |    y   |     y
 326  *
 327  * - column 1, 3: we've the parent -> done
 328  * - column 2:    we take the first key from the block to find the parent
 329  *                (see add_missing_keys)
 330  * - column 4:    we use the key to find the parent
 331  *
 332  * additional information that's available but not required to find the parent
 333  * block might help in merging entries to gain some speed.
 334  */
 335 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 336                           struct preftree *preftree, u64 root_id,
 337                           const struct btrfs_key *key, int level, u64 parent,
 338                           u64 wanted_disk_byte, int count,
 339                           struct share_check *sc, gfp_t gfp_mask)
 340 {
 341         struct prelim_ref *ref;
 342
 343         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 344                 return 0;
 345
 346         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 347         if (!ref)
 348                 return -ENOMEM;
 349
 350         ref->root_id = root_id;
 351         if (key)
 352                 ref->key_for_search = *key;
 353         else
 354                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 355
 356         ref->inode_list = NULL;
 357         ref->level = level;
 358         ref->count = count;
 359         ref->parent = parent;
 360         ref->wanted_disk_byte = wanted_disk_byte;
 361         prelim_ref_insert(fs_info, preftree, ref, sc);
 362         return extent_is_shared(sc);
 363 }
 364
 365 /* direct refs use root == 0, key == NULL */
 366 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 367                           struct preftrees *preftrees, int level, u64 parent,
 368                           u64 wanted_disk_byte, int count,
 369                           struct share_check *sc, gfp_t gfp_mask)
 370 {
 371         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 372                               parent, wanted_disk_byte, count, sc, gfp_mask);
 373 }
 374
 375 /* indirect refs use parent == 0 */
 376 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 377                             struct preftrees *preftrees, u64 root_id,
 378                             const struct btrfs_key *key, int level,
 379                             u64 wanted_disk_byte, int count,
 380                             struct share_check *sc, gfp_t gfp_mask)
 381 {
 382         struct preftree *tree = &preftrees->indirect;
 383
 384         if (!key)
 385                 tree = &preftrees->indirect_missing_keys;
 386         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 387                               wanted_disk_byte, count, sc, gfp_mask);
 388 }
 389
 390 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 391 {
 392         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 393         struct rb_node *parent = NULL;
 394         struct prelim_ref *ref = NULL;
 395         struct prelim_ref target = {};
 396         int result;
 397
 398         target.parent = bytenr;
 399
 400         while (*p) {
 401                 parent = *p;
 402                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 403                 result = prelim_ref_compare(ref, &target);
 404
 405                 if (result < 0)
 406                         p = &(*p)->rb_left;
 407                 else if (result > 0)
 408                         p = &(*p)->rb_right;
 409                 else
 410                         return 1;
 411         }
 412         return 0;
 413 }
 414
 415 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 416                            struct ulist *parents,
 417                            struct preftrees *preftrees, struct prelim_ref *ref,
 418                            int level, u64 time_seq, const u64 *extent_item_pos,
 419                            bool ignore_offset)
 420 {
 421         int ret = 0;
 422         int slot;
 423         struct extent_buffer *eb;
 424         struct btrfs_key key;
 425         struct btrfs_key *key_for_search = &ref->key_for_search;
 426         struct btrfs_file_extent_item *fi;
 427         struct extent_inode_elem *eie = NULL, *old = NULL;
 428         u64 disk_byte;
 429         u64 wanted_disk_byte = ref->wanted_disk_byte;
 430         u64 count = 0;
 431         u64 data_offset;
 432
 433         if (level != 0) {
 434                 eb = path->nodes[level];
 435                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 436                 if (ret < 0)
 437                         return ret;
 438                 return 0;
 439         }
 440
 441         /*
 442          * 1. We normally enter this function with the path already pointing to
 443          *    the first item to check. But sometimes, we may enter it with
 444          *    slot == nritems.
 445          * 2. We are searching for normal backref but bytenr of this leaf
 446          *    matches shared data backref
 447          * 3. The leaf owner is not equal to the root we are searching
 448          *
 449          * For these cases, go to the next leaf before we continue.
 450          */
 451         eb = path->nodes[0];
 452         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 453             is_shared_data_backref(preftrees, eb->start) ||
 454             ref->root_id != btrfs_header_owner(eb)) {
 455                 if (time_seq == SEQ_LAST)
 456                         ret = btrfs_next_leaf(root, path);
 457                 else
 458                         ret = btrfs_next_old_leaf(root, path, time_seq);
 459         }
 460
 461         while (!ret && count < ref->count) {
 462                 eb = path->nodes[0];
 463                 slot = path->slots[0];
 464
 465                 btrfs_item_key_to_cpu(eb, &key, slot);
 466
 467                 if (key.objectid != key_for_search->objectid ||
 468                     key.type != BTRFS_EXTENT_DATA_KEY)
 469                         break;
 470
 471                 /*
 472                  * We are searching for normal backref but bytenr of this leaf
 473                  * matches shared data backref, OR
 474                  * the leaf owner is not equal to the root we are searching for
 475                  */
 476                 if (slot == 0 &&
 477                     (is_shared_data_backref(preftrees, eb->start) ||
 478                      ref->root_id != btrfs_header_owner(eb))) {
 479                         if (time_seq == SEQ_LAST)
 480                                 ret = btrfs_next_leaf(root, path);
 481                         else
 482                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 483                         continue;
 484                 }
 485                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 486                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 487                 data_offset = btrfs_file_extent_offset(eb, fi);
 488
 489                 if (disk_byte == wanted_disk_byte) {
 490                         eie = NULL;
 491                         old = NULL;
 492                         if (ref->key_for_search.offset == key.offset - data_offset)
 493                                 count++;
 494                         else
 495                                 goto next;
 496                         if (extent_item_pos) {
 497                                 ret = check_extent_in_eb(&key, eb, fi,
 498                                                 *extent_item_pos,
 499                                                 &eie, ignore_offset);
 500                                 if (ret < 0)
 501                                         break;
 502                         }
 503                         if (ret > 0)
 504                                 goto next;
 505                         ret = ulist_add_merge_ptr(parents, eb->start,
 506                                                   eie, (void **)&old, GFP_NOFS);
 507                         if (ret < 0)
 508                                 break;
 509                         if (!ret && extent_item_pos) {
 510                                 while (old->next)
 511                                         old = old->next;
 512                                 old->next = eie;
 513                         }
 514                         eie = NULL;
 515                 }
 516 next:
 517                 if (time_seq == SEQ_LAST)
 518                         ret = btrfs_next_item(root, path);
 519                 else
 520                         ret = btrfs_next_old_item(root, path, time_seq);
 521         }
 522
 523         if (ret > 0)
 524                 ret = 0;
 525         else if (ret < 0)
 526                 free_inode_elem_list(eie);
 527         return ret;
 528 }
 529
 530 /*
 531  * resolve an indirect backref in the form (root_id, key, level)
 532  * to a logical address
 533  */
 534 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 535                                 struct btrfs_path *path, u64 time_seq,
 536                                 struct preftrees *preftrees,
 537                                 struct prelim_ref *ref, struct ulist *parents,
 538                                 const u64 *extent_item_pos, bool ignore_offset)
 539 {
 540         struct btrfs_root *root;
 541         struct extent_buffer *eb;
 542         int ret = 0;
 543         int root_level;
 544         int level = ref->level;
 545         struct btrfs_key search_key = ref->key_for_search;
 546
 547         /*
 548          * If we're search_commit_root we could possibly be holding locks on
 549          * other tree nodes.  This happens when qgroups does backref walks when
 550          * adding new delayed refs.  To deal with this we need to look in cache
 551          * for the root, and if we don't find it then we need to search the
 552          * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
 553          * here.
 554          */
 555         if (path->search_commit_root)
 556                 root = btrfs_get_fs_root_commit_root(fs_info, path, ref->root_id);
 557         else
 558                 root = btrfs_get_fs_root(fs_info, ref->root_id, false);
 559         if (IS_ERR(root)) {
 560                 ret = PTR_ERR(root);
 561                 goto out_free;
 562         }
 563
 564         if (!path->search_commit_root &&
 565             test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 566                 ret = -ENOENT;
 567                 goto out;
 568         }
 569
 570         if (btrfs_is_testing(fs_info)) {
 571                 ret = -ENOENT;
 572                 goto out;
 573         }
 574
 575         if (path->search_commit_root)
 576                 root_level = btrfs_header_level(root->commit_root);
 577         else if (time_seq == SEQ_LAST)
 578                 root_level = btrfs_header_level(root->node);
 579         else
 580                 root_level = btrfs_old_root_level(root, time_seq);
 581
 582         if (root_level + 1 == level)
 583                 goto out;
 584
 585         /*
 586          * We can often find data backrefs with an offset that is too large
 587          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 588          * subtracting a file's offset with the data offset of its
 589          * corresponding extent data item. This can happen for example in the
 590          * clone ioctl.
 591          *
 592          * So if we detect such case we set the search key's offset to zero to
 593          * make sure we will find the matching file extent item at
 594          * add_all_parents(), otherwise we will miss it because the offset
 595          * taken form the backref is much larger then the offset of the file
 596          * extent item. This can make us scan a very large number of file
 597          * extent items, but at least it will not make us miss any.
 598          *
 599          * This is an ugly workaround for a behaviour that should have never
 600          * existed, but it does and a fix for the clone ioctl would touch a lot
 601          * of places, cause backwards incompatibility and would not fix the
 602          * problem for extents cloned with older kernels.
 603          */
 604         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 605             search_key.offset >= LLONG_MAX)
 606                 search_key.offset = 0;
 607         path->lowest_level = level;
 608         if (time_seq == SEQ_LAST)
 609                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 610         else
 611                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 612
 613         btrfs_debug(fs_info,
 614                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 615                  ref->root_id, level, ref->count, ret,
 616                  ref->key_for_search.objectid, ref->key_for_search.type,
 617                  ref->key_for_search.offset);
 618         if (ret < 0)
 619                 goto out;
 620
 621         eb = path->nodes[level];
 622         while (!eb) {
 623                 if (WARN_ON(!level)) {
 624                         ret = 1;
 625                         goto out;
 626                 }
 627                 level--;
 628                 eb = path->nodes[level];
 629         }
 630
 631         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 632                               time_seq, extent_item_pos, ignore_offset);
 633 out:
 634         btrfs_put_root(root);
 635 out_free:
 636         path->lowest_level = 0;
 637         btrfs_release_path(path);
 638         return ret;
 639 }
 640
 641 static struct extent_inode_elem *
 642 unode_aux_to_inode_list(struct ulist_node *node)
 643 {
 644         if (!node)
 645                 return NULL;
 646         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 647 }
 648
 649 /*
 650  * We maintain three separate rbtrees: one for direct refs, one for
 651  * indirect refs which have a key, and one for indirect refs which do not
 652  * have a key. Each tree does merge on insertion.
 653  *
 654  * Once all of the references are located, we iterate over the tree of
 655  * indirect refs with missing keys. An appropriate key is located and
 656  * the ref is moved onto the tree for indirect refs. After all missing
 657  * keys are thus located, we iterate over the indirect ref tree, resolve
 658  * each reference, and then insert the resolved reference onto the
 659  * direct tree (merging there too).
 660  *
 661  * New backrefs (i.e., for parent nodes) are added to the appropriate
 662  * rbtree as they are encountered. The new backrefs are subsequently
 663  * resolved as above.
 664  */
 665 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 666                                  struct btrfs_path *path, u64 time_seq,
 667                                  struct preftrees *preftrees,
 668                                  const u64 *extent_item_pos,
 669                                  struct share_check *sc, bool ignore_offset)
 670 {
 671         int err;
 672         int ret = 0;
 673         struct ulist *parents;
 674         struct ulist_node *node;
 675         struct ulist_iterator uiter;
 676         struct rb_node *rnode;
 677
 678         parents = ulist_alloc(GFP_NOFS);
 679         if (!parents)
 680                 return -ENOMEM;
 681
 682         /*
 683          * We could trade memory usage for performance here by iterating
 684          * the tree, allocating new refs for each insertion, and then
 685          * freeing the entire indirect tree when we're done.  In some test
 686          * cases, the tree can grow quite large (~200k objects).
 687          */
 688         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 689                 struct prelim_ref *ref;
 690
 691                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 692                 if (WARN(ref->parent,
 693                          "BUG: direct ref found in indirect tree")) {
 694                         ret = -EINVAL;
 695                         goto out;
 696                 }
 697
 698                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 699                 preftrees->indirect.count--;
 700
 701                 if (ref->count == 0) {
 702                         free_pref(ref);
 703                         continue;
 704                 }
 705
 706                 if (sc && sc->root_objectid &&
 707                     ref->root_id != sc->root_objectid) {
 708                         free_pref(ref);
 709                         ret = BACKREF_FOUND_SHARED;
 710                         goto out;
 711                 }
 712                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 713                                            ref, parents, extent_item_pos,
 714                                            ignore_offset);
 715                 /*
 716                  * we can only tolerate ENOENT,otherwise,we should catch error
 717                  * and return directly.
 718                  */
 719                 if (err == -ENOENT) {
 720                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 721                                           NULL);
 722                         continue;
 723                 } else if (err) {
 724                         free_pref(ref);
 725                         ret = err;
 726                         goto out;
 727                 }
 728
 729                 /* we put the first parent into the ref at hand */
 730                 ULIST_ITER_INIT(&uiter);
 731                 node = ulist_next(parents, &uiter);
 732                 ref->parent = node ? node->val : 0;
 733                 ref->inode_list = unode_aux_to_inode_list(node);
 734
 735                 /* Add a prelim_ref(s) for any other parent(s). */
 736                 while ((node = ulist_next(parents, &uiter))) {
 737                         struct prelim_ref *new_ref;
 738
 739                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 740                                                    GFP_NOFS);
 741                         if (!new_ref) {
 742                                 free_pref(ref);
 743                                 ret = -ENOMEM;
 744                                 goto out;
 745                         }
 746                         memcpy(new_ref, ref, sizeof(*ref));
 747                         new_ref->parent = node->val;
 748                         new_ref->inode_list = unode_aux_to_inode_list(node);
 749                         prelim_ref_insert(fs_info, &preftrees->direct,
 750                                           new_ref, NULL);
 751                 }
 752
 753                 /*
 754                  * Now it's a direct ref, put it in the direct tree. We must
 755                  * do this last because the ref could be merged/freed here.
 756                  */
 757                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 758
 759                 ulist_reinit(parents);
 760                 cond_resched();
 761         }
 762 out:
 763         ulist_free(parents);
 764         return ret;
 765 }
 766
 767 /*
 768  * read tree blocks and add keys where required.
 769  */
 770 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 771                             struct preftrees *preftrees, bool lock)
 772 {
 773         struct prelim_ref *ref;
 774         struct extent_buffer *eb;
 775         struct preftree *tree = &preftrees->indirect_missing_keys;
 776         struct rb_node *node;
 777
 778         while ((node = rb_first_cached(&tree->root))) {
 779                 ref = rb_entry(node, struct prelim_ref, rbnode);
 780                 rb_erase_cached(node, &tree->root);
 781
 782                 BUG_ON(ref->parent);    /* should not be a direct ref */
 783                 BUG_ON(ref->key_for_search.type);
 784                 BUG_ON(!ref->wanted_disk_byte);
 785
 786                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 787                                      ref->level - 1, NULL);
 788                 if (IS_ERR(eb)) {
 789                         free_pref(ref);
 790                         return PTR_ERR(eb);
 791                 } else if (!extent_buffer_uptodate(eb)) {
 792                         free_pref(ref);
 793                         free_extent_buffer(eb);
 794                         return -EIO;
 795                 }
 796                 if (lock)
 797                         btrfs_tree_read_lock(eb);
 798                 if (btrfs_header_level(eb) == 0)
 799                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 800                 else
 801                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 802                 if (lock)
 803                         btrfs_tree_read_unlock(eb);
 804                 free_extent_buffer(eb);
 805                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 806                 cond_resched();
 807         }
 808         return 0;
 809 }
 810
 811 /*
 812  * add all currently queued delayed refs from this head whose seq nr is
 813  * smaller or equal that seq to the list
 814  */
 815 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 816                             struct btrfs_delayed_ref_head *head, u64 seq,
 817                             struct preftrees *preftrees, struct share_check *sc)
 818 {
 819         struct btrfs_delayed_ref_node *node;
 820         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 821         struct btrfs_key key;
 822         struct btrfs_key tmp_op_key;
 823         struct rb_node *n;
 824         int count;
 825         int ret = 0;
 826
 827         if (extent_op && extent_op->update_key)
 828                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 829
 830         spin_lock(&head->lock);
 831         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 832                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 833                                 ref_node);
 834                 if (node->seq > seq)
 835                         continue;
 836
 837                 switch (node->action) {
 838                 case BTRFS_ADD_DELAYED_EXTENT:
 839                 case BTRFS_UPDATE_DELAYED_HEAD:
 840                         WARN_ON(1);
 841                         continue;
 842                 case BTRFS_ADD_DELAYED_REF:
 843                         count = node->ref_mod;
 844                         break;
 845                 case BTRFS_DROP_DELAYED_REF:
 846                         count = node->ref_mod * -1;
 847                         break;
 848                 default:
 849                         BUG();
 850                 }
 851                 switch (node->type) {
 852                 case BTRFS_TREE_BLOCK_REF_KEY: {
 853                         /* NORMAL INDIRECT METADATA backref */
 854                         struct btrfs_delayed_tree_ref *ref;
 855
 856                         ref = btrfs_delayed_node_to_tree_ref(node);
 857                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 858                                                &tmp_op_key, ref->level + 1,
 859                                                node->bytenr, count, sc,
 860                                                GFP_ATOMIC);
 861                         break;
 862                 }
 863                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 864                         /* SHARED DIRECT METADATA backref */
 865                         struct btrfs_delayed_tree_ref *ref;
 866
 867                         ref = btrfs_delayed_node_to_tree_ref(node);
 868
 869                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 870                                              ref->parent, node->bytenr, count,
 871                                              sc, GFP_ATOMIC);
 872                         break;
 873                 }
 874                 case BTRFS_EXTENT_DATA_REF_KEY: {
 875                         /* NORMAL INDIRECT DATA backref */
 876                         struct btrfs_delayed_data_ref *ref;
 877                         ref = btrfs_delayed_node_to_data_ref(node);
 878
 879                         key.objectid = ref->objectid;
 880                         key.type = BTRFS_EXTENT_DATA_KEY;
 881                         key.offset = ref->offset;
 882
 883                         /*
 884                          * Found a inum that doesn't match our known inum, we
 885                          * know it's shared.
 886                          */
 887                         if (sc && sc->inum && ref->objectid != sc->inum) {
 888                                 ret = BACKREF_FOUND_SHARED;
 889                                 goto out;
 890                         }
 891
 892                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 893                                                &key, 0, node->bytenr, count, sc,
 894                                                GFP_ATOMIC);
 895                         break;
 896                 }
 897                 case BTRFS_SHARED_DATA_REF_KEY: {
 898                         /* SHARED DIRECT FULL backref */
 899                         struct btrfs_delayed_data_ref *ref;
 900
 901                         ref = btrfs_delayed_node_to_data_ref(node);
 902
 903                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 904                                              node->bytenr, count, sc,
 905                                              GFP_ATOMIC);
 906                         break;
 907                 }
 908                 default:
 909                         WARN_ON(1);
 910                 }
 911                 /*
 912                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 913                  * refs have been checked.
 914                  */
 915                 if (ret && (ret != BACKREF_FOUND_SHARED))
 916                         break;
 917         }
 918         if (!ret)
 919                 ret = extent_is_shared(sc);
 920 out:
 921         spin_unlock(&head->lock);
 922         return ret;
 923 }
 924
 925 /*
 926  * add all inline backrefs for bytenr to the list
 927  *
 928  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 929  */
 930 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 931                            struct btrfs_path *path, u64 bytenr,
 932                            int *info_level, struct preftrees *preftrees,
 933                            struct share_check *sc)
 934 {
 935         int ret = 0;
 936         int slot;
 937         struct extent_buffer *leaf;
 938         struct btrfs_key key;
 939         struct btrfs_key found_key;
 940         unsigned long ptr;
 941         unsigned long end;
 942         struct btrfs_extent_item *ei;
 943         u64 flags;
 944         u64 item_size;
 945
 946         /*
 947          * enumerate all inline refs
 948          */
 949         leaf = path->nodes[0];
 950         slot = path->slots[0];
 951
 952         item_size = btrfs_item_size_nr(leaf, slot);
 953         BUG_ON(item_size < sizeof(*ei));
 954
 955         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 956         flags = btrfs_extent_flags(leaf, ei);
 957         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 958
 959         ptr = (unsigned long)(ei + 1);
 960         end = (unsigned long)ei + item_size;
 961
 962         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 963             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 964                 struct btrfs_tree_block_info *info;
 965
 966                 info = (struct btrfs_tree_block_info *)ptr;
 967                 *info_level = btrfs_tree_block_level(leaf, info);
 968                 ptr += sizeof(struct btrfs_tree_block_info);
 969                 BUG_ON(ptr > end);
 970         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 971                 *info_level = found_key.offset;
 972         } else {
 973                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 974         }
 975
 976         while (ptr < end) {
 977                 struct btrfs_extent_inline_ref *iref;
 978                 u64 offset;
 979                 int type;
 980
 981                 iref = (struct btrfs_extent_inline_ref *)ptr;
 982                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 983                                                         BTRFS_REF_TYPE_ANY);
 984                 if (type == BTRFS_REF_TYPE_INVALID)
 985                         return -EUCLEAN;
 986
 987                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 988
 989                 switch (type) {
 990                 case BTRFS_SHARED_BLOCK_REF_KEY:
 991                         ret = add_direct_ref(fs_info, preftrees,
 992                                              *info_level + 1, offset,
 993                                              bytenr, 1, NULL, GFP_NOFS);
 994                         break;
 995                 case BTRFS_SHARED_DATA_REF_KEY: {
 996                         struct btrfs_shared_data_ref *sdref;
 997                         int count;
 998
 999                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1000                         count = btrfs_shared_data_ref_count(leaf, sdref);
1001
1002                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1003                                              bytenr, count, sc, GFP_NOFS);
1004                         break;
1005                 }
1006                 case BTRFS_TREE_BLOCK_REF_KEY:
1007                         ret = add_indirect_ref(fs_info, preftrees, offset,
1008                                                NULL, *info_level + 1,
1009                                                bytenr, 1, NULL, GFP_NOFS);
1010                         break;
1011                 case BTRFS_EXTENT_DATA_REF_KEY: {
1012                         struct btrfs_extent_data_ref *dref;
1013                         int count;
1014                         u64 root;
1015
1016                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1017                         count = btrfs_extent_data_ref_count(leaf, dref);
1018                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1019                                                                       dref);
1020                         key.type = BTRFS_EXTENT_DATA_KEY;
1021                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1022
1023                         if (sc && sc->inum && key.objectid != sc->inum) {
1024                                 ret = BACKREF_FOUND_SHARED;
1025                                 break;
1026                         }
1027
1028                         root = btrfs_extent_data_ref_root(leaf, dref);
1029
1030                         ret = add_indirect_ref(fs_info, preftrees, root,
1031                                                &key, 0, bytenr, count,
1032                                                sc, GFP_NOFS);
1033                         break;
1034                 }
1035                 default:
1036                         WARN_ON(1);
1037                 }
1038                 if (ret)
1039                         return ret;
1040                 ptr += btrfs_extent_inline_ref_size(type);
1041         }
1042
1043         return 0;
1044 }
1045
1046 /*
1047  * add all non-inline backrefs for bytenr to the list
1048  *
1049  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1050  */
1051 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1052                           struct btrfs_path *path, u64 bytenr,
1053                           int info_level, struct preftrees *preftrees,
1054                           struct share_check *sc)
1055 {
1056         struct btrfs_root *extent_root = fs_info->extent_root;
1057         int ret;
1058         int slot;
1059         struct extent_buffer *leaf;
1060         struct btrfs_key key;
1061
1062         while (1) {
1063                 ret = btrfs_next_item(extent_root, path);
1064                 if (ret < 0)
1065                         break;
1066                 if (ret) {
1067                         ret = 0;
1068                         break;
1069                 }
1070
1071                 slot = path->slots[0];
1072                 leaf = path->nodes[0];
1073                 btrfs_item_key_to_cpu(leaf, &key, slot);
1074
1075                 if (key.objectid != bytenr)
1076                         break;
1077                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1078                         continue;
1079                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1080                         break;
1081
1082                 switch (key.type) {
1083                 case BTRFS_SHARED_BLOCK_REF_KEY:
1084                         /* SHARED DIRECT METADATA backref */
1085                         ret = add_direct_ref(fs_info, preftrees,
1086                                              info_level + 1, key.offset,
1087                                              bytenr, 1, NULL, GFP_NOFS);
1088                         break;
1089                 case BTRFS_SHARED_DATA_REF_KEY: {
1090                         /* SHARED DIRECT FULL backref */
1091                         struct btrfs_shared_data_ref *sdref;
1092                         int count;
1093
1094                         sdref = btrfs_item_ptr(leaf, slot,
1095                                               struct btrfs_shared_data_ref);
1096                         count = btrfs_shared_data_ref_count(leaf, sdref);
1097                         ret = add_direct_ref(fs_info, preftrees, 0,
1098                                              key.offset, bytenr, count,
1099                                              sc, GFP_NOFS);
1100                         break;
1101                 }
1102                 case BTRFS_TREE_BLOCK_REF_KEY:
1103                         /* NORMAL INDIRECT METADATA backref */
1104                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1105                                                NULL, info_level + 1, bytenr,
1106                                                1, NULL, GFP_NOFS);
1107                         break;
1108                 case BTRFS_EXTENT_DATA_REF_KEY: {
1109                         /* NORMAL INDIRECT DATA backref */
1110                         struct btrfs_extent_data_ref *dref;
1111                         int count;
1112                         u64 root;
1113
1114                         dref = btrfs_item_ptr(leaf, slot,
1115                                               struct btrfs_extent_data_ref);
1116                         count = btrfs_extent_data_ref_count(leaf, dref);
1117                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1118                                                                       dref);
1119                         key.type = BTRFS_EXTENT_DATA_KEY;
1120                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1121
1122                         if (sc && sc->inum && key.objectid != sc->inum) {
1123                                 ret = BACKREF_FOUND_SHARED;
1124                                 break;
1125                         }
1126
1127                         root = btrfs_extent_data_ref_root(leaf, dref);
1128                         ret = add_indirect_ref(fs_info, preftrees, root,
1129                                                &key, 0, bytenr, count,
1130                                                sc, GFP_NOFS);
1131                         break;
1132                 }
1133                 default:
1134                         WARN_ON(1);
1135                 }
1136                 if (ret)
1137                         return ret;
1138
1139         }
1140
1141         return ret;
1142 }
1143
1144 /*
1145  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1146  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1147  * indirect refs to their parent bytenr.
1148  * When roots are found, they're added to the roots list
1149  *
1150  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1151  * much like trans == NULL case, the difference only lies in it will not
1152  * commit root.
1153  * The special case is for qgroup to search roots in commit_transaction().
1154  *
1155  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1156  * shared extent is detected.
1157  *
1158  * Otherwise this returns 0 for success and <0 for an error.
1159  *
1160  * If ignore_offset is set to false, only extent refs whose offsets match
1161  * extent_item_pos are returned.  If true, every extent ref is returned
1162  * and extent_item_pos is ignored.
1163  *
1164  * FIXME some caching might speed things up
1165  */
1166 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1167                              struct btrfs_fs_info *fs_info, u64 bytenr,
1168                              u64 time_seq, struct ulist *refs,
1169                              struct ulist *roots, const u64 *extent_item_pos,
1170                              struct share_check *sc, bool ignore_offset)
1171 {
1172         struct btrfs_key key;
1173         struct btrfs_path *path;
1174         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1175         struct btrfs_delayed_ref_head *head;
1176         int info_level = 0;
1177         int ret;
1178         struct prelim_ref *ref;
1179         struct rb_node *node;
1180         struct extent_inode_elem *eie = NULL;
1181         struct preftrees preftrees = {
1182                 .direct = PREFTREE_INIT,
1183                 .indirect = PREFTREE_INIT,
1184                 .indirect_missing_keys = PREFTREE_INIT
1185         };
1186
1187         key.objectid = bytenr;
1188         key.offset = (u64)-1;
1189         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1190                 key.type = BTRFS_METADATA_ITEM_KEY;
1191         else
1192                 key.type = BTRFS_EXTENT_ITEM_KEY;
1193
1194         path = btrfs_alloc_path();
1195         if (!path)
1196                 return -ENOMEM;
1197         if (!trans) {
1198                 path->search_commit_root = 1;
1199                 path->skip_locking = 1;
1200         }
1201
1202         if (time_seq == SEQ_LAST)
1203                 path->skip_locking = 1;
1204
1205         /*
1206          * grab both a lock on the path and a lock on the delayed ref head.
1207          * We need both to get a consistent picture of how the refs look
1208          * at a specified point in time
1209          */
1210 again:
1211         head = NULL;
1212
1213         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1214         if (ret < 0)
1215                 goto out;
1216         BUG_ON(ret == 0);
1217
1218 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1219         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1220             time_seq != SEQ_LAST) {
1221 #else
1222         if (trans && time_seq != SEQ_LAST) {
1223 #endif
1224                 /*
1225                  * look if there are updates for this ref queued and lock the
1226                  * head
1227                  */
1228                 delayed_refs = &trans->transaction->delayed_refs;
1229                 spin_lock(&delayed_refs->lock);
1230                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1231                 if (head) {
1232                         if (!mutex_trylock(&head->mutex)) {
1233                                 refcount_inc(&head->refs);
1234                                 spin_unlock(&delayed_refs->lock);
1235
1236                                 btrfs_release_path(path);
1237
1238                                 /*
1239                                  * Mutex was contended, block until it's
1240                                  * released and try again
1241                                  */
1242                                 mutex_lock(&head->mutex);
1243                                 mutex_unlock(&head->mutex);
1244                                 btrfs_put_delayed_ref_head(head);
1245                                 goto again;
1246                         }
1247                         spin_unlock(&delayed_refs->lock);
1248                         ret = add_delayed_refs(fs_info, head, time_seq,
1249                                                &preftrees, sc);
1250                         mutex_unlock(&head->mutex);
1251                         if (ret)
1252                                 goto out;
1253                 } else {
1254                         spin_unlock(&delayed_refs->lock);
1255                 }
1256         }
1257
1258         if (path->slots[0]) {
1259                 struct extent_buffer *leaf;
1260                 int slot;
1261
1262                 path->slots[0]--;
1263                 leaf = path->nodes[0];
1264                 slot = path->slots[0];
1265                 btrfs_item_key_to_cpu(leaf, &key, slot);
1266                 if (key.objectid == bytenr &&
1267                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1268                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1269                         ret = add_inline_refs(fs_info, path, bytenr,
1270                                               &info_level, &preftrees, sc);
1271                         if (ret)
1272                                 goto out;
1273                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1274                                              &preftrees, sc);
1275                         if (ret)
1276                                 goto out;
1277                 }
1278         }
1279
1280         btrfs_release_path(path);
1281
1282         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1283         if (ret)
1284                 goto out;
1285
1286         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1287
1288         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1289                                     extent_item_pos, sc, ignore_offset);
1290         if (ret)
1291                 goto out;
1292
1293         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1294
1295         /*
1296          * This walks the tree of merged and resolved refs. Tree blocks are
1297          * read in as needed. Unique entries are added to the ulist, and
1298          * the list of found roots is updated.
1299          *
1300          * We release the entire tree in one go before returning.
1301          */
1302         node = rb_first_cached(&preftrees.direct.root);
1303         while (node) {
1304                 ref = rb_entry(node, struct prelim_ref, rbnode);
1305                 node = rb_next(&ref->rbnode);
1306                 /*
1307                  * ref->count < 0 can happen here if there are delayed
1308                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1309                  * prelim_ref_insert() relies on this when merging
1310                  * identical refs to keep the overall count correct.
1311                  * prelim_ref_insert() will merge only those refs
1312                  * which compare identically.  Any refs having
1313                  * e.g. different offsets would not be merged,
1314                  * and would retain their original ref->count < 0.
1315                  */
1316                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1317                         if (sc && sc->root_objectid &&
1318                             ref->root_id != sc->root_objectid) {
1319                                 ret = BACKREF_FOUND_SHARED;
1320                                 goto out;
1321                         }
1322
1323                         /* no parent == root of tree */
1324                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1325                         if (ret < 0)
1326                                 goto out;
1327                 }
1328                 if (ref->count && ref->parent) {
1329                         if (extent_item_pos && !ref->inode_list &&
1330                             ref->level == 0) {
1331                                 struct extent_buffer *eb;
1332
1333                                 eb = read_tree_block(fs_info, ref->parent, 0,
1334                                                      ref->level, NULL);
1335                                 if (IS_ERR(eb)) {
1336                                         ret = PTR_ERR(eb);
1337                                         goto out;
1338                                 } else if (!extent_buffer_uptodate(eb)) {
1339                                         free_extent_buffer(eb);
1340                                         ret = -EIO;
1341                                         goto out;
1342                                 }
1343
1344                                 if (!path->skip_locking) {
1345                                         btrfs_tree_read_lock(eb);
1346                                         btrfs_set_lock_blocking_read(eb);
1347                                 }
1348                                 ret = find_extent_in_eb(eb, bytenr,
1349                                                         *extent_item_pos, &eie, ignore_offset);
1350                                 if (!path->skip_locking)
1351                                         btrfs_tree_read_unlock_blocking(eb);
1352                                 free_extent_buffer(eb);
1353                                 if (ret < 0)
1354                                         goto out;
1355                                 ref->inode_list = eie;
1356                         }
1357                         ret = ulist_add_merge_ptr(refs, ref->parent,
1358                                                   ref->inode_list,
1359                                                   (void **)&eie, GFP_NOFS);
1360                         if (ret < 0)
1361                                 goto out;
1362                         if (!ret && extent_item_pos) {
1363                                 /*
1364                                  * we've recorded that parent, so we must extend
1365                                  * its inode list here
1366                                  */
1367                                 BUG_ON(!eie);
1368                                 while (eie->next)
1369                                         eie = eie->next;
1370                                 eie->next = ref->inode_list;
1371                         }
1372                         eie = NULL;
1373                 }
1374                 cond_resched();
1375         }
1376
1377 out:
1378         btrfs_free_path(path);
1379
1380         prelim_release(&preftrees.direct);
1381         prelim_release(&preftrees.indirect);
1382         prelim_release(&preftrees.indirect_missing_keys);
1383
1384         if (ret < 0)
1385                 free_inode_elem_list(eie);
1386         return ret;
1387 }
1388
1389 static void free_leaf_list(struct ulist *blocks)
1390 {
1391         struct ulist_node *node = NULL;
1392         struct extent_inode_elem *eie;
1393         struct ulist_iterator uiter;
1394
1395         ULIST_ITER_INIT(&uiter);
1396         while ((node = ulist_next(blocks, &uiter))) {
1397                 if (!node->aux)
1398                         continue;
1399                 eie = unode_aux_to_inode_list(node);
1400                 free_inode_elem_list(eie);
1401                 node->aux = 0;
1402         }
1403
1404         ulist_free(blocks);
1405 }
1406
1407 /*
1408  * Finds all leafs with a reference to the specified combination of bytenr and
1409  * offset. key_list_head will point to a list of corresponding keys (caller must
1410  * free each list element). The leafs will be stored in the leafs ulist, which
1411  * must be freed with ulist_free.
1412  *
1413  * returns 0 on success, <0 on error
1414  */
1415 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1416                          struct btrfs_fs_info *fs_info, u64 bytenr,
1417                          u64 time_seq, struct ulist **leafs,
1418                          const u64 *extent_item_pos, bool ignore_offset)
1419 {
1420         int ret;
1421
1422         *leafs = ulist_alloc(GFP_NOFS);
1423         if (!*leafs)
1424                 return -ENOMEM;
1425
1426         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1427                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1428         if (ret < 0 && ret != -ENOENT) {
1429                 free_leaf_list(*leafs);
1430                 return ret;
1431         }
1432
1433         return 0;
1434 }
1435
1436 /*
1437  * walk all backrefs for a given extent to find all roots that reference this
1438  * extent. Walking a backref means finding all extents that reference this
1439  * extent and in turn walk the backrefs of those, too. Naturally this is a
1440  * recursive process, but here it is implemented in an iterative fashion: We
1441  * find all referencing extents for the extent in question and put them on a
1442  * list. In turn, we find all referencing extents for those, further appending
1443  * to the list. The way we iterate the list allows adding more elements after
1444  * the current while iterating. The process stops when we reach the end of the
1445  * list. Found roots are added to the roots list.
1446  *
1447  * returns 0 on success, < 0 on error.
1448  */
1449 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1450                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1451                                      u64 time_seq, struct ulist **roots,
1452                                      bool ignore_offset)
1453 {
1454         struct ulist *tmp;
1455         struct ulist_node *node = NULL;
1456         struct ulist_iterator uiter;
1457         int ret;
1458
1459         tmp = ulist_alloc(GFP_NOFS);
1460         if (!tmp)
1461                 return -ENOMEM;
1462         *roots = ulist_alloc(GFP_NOFS);
1463         if (!*roots) {
1464                 ulist_free(tmp);
1465                 return -ENOMEM;
1466         }
1467
1468         ULIST_ITER_INIT(&uiter);
1469         while (1) {
1470                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1471                                         tmp, *roots, NULL, NULL, ignore_offset);
1472                 if (ret < 0 && ret != -ENOENT) {
1473                         ulist_free(tmp);
1474                         ulist_free(*roots);
1475                         *roots = NULL;
1476                         return ret;
1477                 }
1478                 node = ulist_next(tmp, &uiter);
1479                 if (!node)
1480                         break;
1481                 bytenr = node->val;
1482                 cond_resched();
1483         }
1484
1485         ulist_free(tmp);
1486         return 0;
1487 }
1488
1489 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1490                          struct btrfs_fs_info *fs_info, u64 bytenr,
1491                          u64 time_seq, struct ulist **roots,
1492                          bool ignore_offset)
1493 {
1494         int ret;
1495
1496         if (!trans)
1497                 down_read(&fs_info->commit_root_sem);
1498         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1499                                         time_seq, roots, ignore_offset);
1500         if (!trans)
1501                 up_read(&fs_info->commit_root_sem);
1502         return ret;
1503 }
1504
1505 /**
1506  * btrfs_check_shared - tell us whether an extent is shared
1507  *
1508  * btrfs_check_shared uses the backref walking code but will short
1509  * circuit as soon as it finds a root or inode that doesn't match the
1510  * one passed in. This provides a significant performance benefit for
1511  * callers (such as fiemap) which want to know whether the extent is
1512  * shared but do not need a ref count.
1513  *
1514  * This attempts to attach to the running transaction in order to account for
1515  * delayed refs, but continues on even when no running transaction exists.
1516  *
1517  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1518  */
1519 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1520                 struct ulist *roots, struct ulist *tmp)
1521 {
1522         struct btrfs_fs_info *fs_info = root->fs_info;
1523         struct btrfs_trans_handle *trans;
1524         struct ulist_iterator uiter;
1525         struct ulist_node *node;
1526         struct seq_list elem = SEQ_LIST_INIT(elem);
1527         int ret = 0;
1528         struct share_check shared = {
1529                 .root_objectid = root->root_key.objectid,
1530                 .inum = inum,
1531                 .share_count = 0,
1532         };
1533
1534         ulist_init(roots);
1535         ulist_init(tmp);
1536
1537         trans = btrfs_join_transaction_nostart(root);
1538         if (IS_ERR(trans)) {
1539                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1540                         ret = PTR_ERR(trans);
1541                         goto out;
1542                 }
1543                 trans = NULL;
1544                 down_read(&fs_info->commit_root_sem);
1545         } else {
1546                 btrfs_get_tree_mod_seq(fs_info, &elem);
1547         }
1548
1549         ULIST_ITER_INIT(&uiter);
1550         while (1) {
1551                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1552                                         roots, NULL, &shared, false);
1553                 if (ret == BACKREF_FOUND_SHARED) {
1554                         /* this is the only condition under which we return 1 */
1555                         ret = 1;
1556                         break;
1557                 }
1558                 if (ret < 0 && ret != -ENOENT)
1559                         break;
1560                 ret = 0;
1561                 node = ulist_next(tmp, &uiter);
1562                 if (!node)
1563                         break;
1564                 bytenr = node->val;
1565                 shared.share_count = 0;
1566                 cond_resched();
1567         }
1568
1569         if (trans) {
1570                 btrfs_put_tree_mod_seq(fs_info, &elem);
1571                 btrfs_end_transaction(trans);
1572         } else {
1573                 up_read(&fs_info->commit_root_sem);
1574         }
1575 out:
1576         ulist_release(roots);
1577         ulist_release(tmp);
1578         return ret;
1579 }
1580
1581 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1582                           u64 start_off, struct btrfs_path *path,
1583                           struct btrfs_inode_extref **ret_extref,
1584                           u64 *found_off)
1585 {
1586         int ret, slot;
1587         struct btrfs_key key;
1588         struct btrfs_key found_key;
1589         struct btrfs_inode_extref *extref;
1590         const struct extent_buffer *leaf;
1591         unsigned long ptr;
1592
1593         key.objectid = inode_objectid;
1594         key.type = BTRFS_INODE_EXTREF_KEY;
1595         key.offset = start_off;
1596
1597         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1598         if (ret < 0)
1599                 return ret;
1600
1601         while (1) {
1602                 leaf = path->nodes[0];
1603                 slot = path->slots[0];
1604                 if (slot >= btrfs_header_nritems(leaf)) {
1605                         /*
1606                          * If the item at offset is not found,
1607                          * btrfs_search_slot will point us to the slot
1608                          * where it should be inserted. In our case
1609                          * that will be the slot directly before the
1610                          * next INODE_REF_KEY_V2 item. In the case
1611                          * that we're pointing to the last slot in a
1612                          * leaf, we must move one leaf over.
1613                          */
1614                         ret = btrfs_next_leaf(root, path);
1615                         if (ret) {
1616                                 if (ret >= 1)
1617                                         ret = -ENOENT;
1618                                 break;
1619                         }
1620                         continue;
1621                 }
1622
1623                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1624
1625                 /*
1626                  * Check that we're still looking at an extended ref key for
1627                  * this particular objectid. If we have different
1628                  * objectid or type then there are no more to be found
1629                  * in the tree and we can exit.
1630                  */
1631                 ret = -ENOENT;
1632                 if (found_key.objectid != inode_objectid)
1633                         break;
1634                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1635                         break;
1636
1637                 ret = 0;
1638                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1639                 extref = (struct btrfs_inode_extref *)ptr;
1640                 *ret_extref = extref;
1641                 if (found_off)
1642                         *found_off = found_key.offset;
1643                 break;
1644         }
1645
1646         return ret;
1647 }
1648
1649 /*
1650  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1651  * Elements of the path are separated by '/' and the path is guaranteed to be
1652  * 0-terminated. the path is only given within the current file system.
1653  * Therefore, it never starts with a '/'. the caller is responsible to provide
1654  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1655  * the start point of the resulting string is returned. this pointer is within
1656  * dest, normally.
1657  * in case the path buffer would overflow, the pointer is decremented further
1658  * as if output was written to the buffer, though no more output is actually
1659  * generated. that way, the caller can determine how much space would be
1660  * required for the path to fit into the buffer. in that case, the returned
1661  * value will be smaller than dest. callers must check this!
1662  */
1663 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1664                         u32 name_len, unsigned long name_off,
1665                         struct extent_buffer *eb_in, u64 parent,
1666                         char *dest, u32 size)
1667 {
1668         int slot;
1669         u64 next_inum;
1670         int ret;
1671         s64 bytes_left = ((s64)size) - 1;
1672         struct extent_buffer *eb = eb_in;
1673         struct btrfs_key found_key;
1674         int leave_spinning = path->leave_spinning;
1675         struct btrfs_inode_ref *iref;
1676
1677         if (bytes_left >= 0)
1678                 dest[bytes_left] = '\0';
1679
1680         path->leave_spinning = 1;
1681         while (1) {
1682                 bytes_left -= name_len;
1683                 if (bytes_left >= 0)
1684                         read_extent_buffer(eb, dest + bytes_left,
1685                                            name_off, name_len);
1686                 if (eb != eb_in) {
1687                         if (!path->skip_locking)
1688                                 btrfs_tree_read_unlock_blocking(eb);
1689                         free_extent_buffer(eb);
1690                 }
1691                 ret = btrfs_find_item(fs_root, path, parent, 0,
1692                                 BTRFS_INODE_REF_KEY, &found_key);
1693                 if (ret > 0)
1694                         ret = -ENOENT;
1695                 if (ret)
1696                         break;
1697
1698                 next_inum = found_key.offset;
1699
1700                 /* regular exit ahead */
1701                 if (parent == next_inum)
1702                         break;
1703
1704                 slot = path->slots[0];
1705                 eb = path->nodes[0];
1706                 /* make sure we can use eb after releasing the path */
1707                 if (eb != eb_in) {
1708                         if (!path->skip_locking)
1709                                 btrfs_set_lock_blocking_read(eb);
1710                         path->nodes[0] = NULL;
1711                         path->locks[0] = 0;
1712                 }
1713                 btrfs_release_path(path);
1714                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1715
1716                 name_len = btrfs_inode_ref_name_len(eb, iref);
1717                 name_off = (unsigned long)(iref + 1);
1718
1719                 parent = next_inum;
1720                 --bytes_left;
1721                 if (bytes_left >= 0)
1722                         dest[bytes_left] = '/';
1723         }
1724
1725         btrfs_release_path(path);
1726         path->leave_spinning = leave_spinning;
1727
1728         if (ret)
1729                 return ERR_PTR(ret);
1730
1731         return dest + bytes_left;
1732 }
1733
1734 /*
1735  * this makes the path point to (logical EXTENT_ITEM *)
1736  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1737  * tree blocks and <0 on error.
1738  */
1739 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1740                         struct btrfs_path *path, struct btrfs_key *found_key,
1741                         u64 *flags_ret)
1742 {
1743         int ret;
1744         u64 flags;
1745         u64 size = 0;
1746         u32 item_size;
1747         const struct extent_buffer *eb;
1748         struct btrfs_extent_item *ei;
1749         struct btrfs_key key;
1750
1751         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1752                 key.type = BTRFS_METADATA_ITEM_KEY;
1753         else
1754                 key.type = BTRFS_EXTENT_ITEM_KEY;
1755         key.objectid = logical;
1756         key.offset = (u64)-1;
1757
1758         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1759         if (ret < 0)
1760                 return ret;
1761
1762         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1763         if (ret) {
1764                 if (ret > 0)
1765                         ret = -ENOENT;
1766                 return ret;
1767         }
1768         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1769         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1770                 size = fs_info->nodesize;
1771         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1772                 size = found_key->offset;
1773
1774         if (found_key->objectid > logical ||
1775             found_key->objectid + size <= logical) {
1776                 btrfs_debug(fs_info,
1777                         "logical %llu is not within any extent", logical);
1778                 return -ENOENT;
1779         }
1780
1781         eb = path->nodes[0];
1782         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1783         BUG_ON(item_size < sizeof(*ei));
1784
1785         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1786         flags = btrfs_extent_flags(eb, ei);
1787
1788         btrfs_debug(fs_info,
1789                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1790                  logical, logical - found_key->objectid, found_key->objectid,
1791                  found_key->offset, flags, item_size);
1792
1793         WARN_ON(!flags_ret);
1794         if (flags_ret) {
1795                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1796                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1797                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1798                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1799                 else
1800                         BUG();
1801                 return 0;
1802         }
1803
1804         return -EIO;
1805 }
1806
1807 /*
1808  * helper function to iterate extent inline refs. ptr must point to a 0 value
1809  * for the first call and may be modified. it is used to track state.
1810  * if more refs exist, 0 is returned and the next call to
1811  * get_extent_inline_ref must pass the modified ptr parameter to get the
1812  * next ref. after the last ref was processed, 1 is returned.
1813  * returns <0 on error
1814  */
1815 static int get_extent_inline_ref(unsigned long *ptr,
1816                                  const struct extent_buffer *eb,
1817                                  const struct btrfs_key *key,
1818                                  const struct btrfs_extent_item *ei,
1819                                  u32 item_size,
1820                                  struct btrfs_extent_inline_ref **out_eiref,
1821                                  int *out_type)
1822 {
1823         unsigned long end;
1824         u64 flags;
1825         struct btrfs_tree_block_info *info;
1826
1827         if (!*ptr) {
1828                 /* first call */
1829                 flags = btrfs_extent_flags(eb, ei);
1830                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1831                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1832                                 /* a skinny metadata extent */
1833                                 *out_eiref =
1834                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1835                         } else {
1836                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1837                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1838                                 *out_eiref =
1839                                    (struct btrfs_extent_inline_ref *)(info + 1);
1840                         }
1841                 } else {
1842                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1843                 }
1844                 *ptr = (unsigned long)*out_eiref;
1845                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1846                         return -ENOENT;
1847         }
1848
1849         end = (unsigned long)ei + item_size;
1850         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1851         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1852                                                      BTRFS_REF_TYPE_ANY);
1853         if (*out_type == BTRFS_REF_TYPE_INVALID)
1854                 return -EUCLEAN;
1855
1856         *ptr += btrfs_extent_inline_ref_size(*out_type);
1857         WARN_ON(*ptr > end);
1858         if (*ptr == end)
1859                 return 1; /* last */
1860
1861         return 0;
1862 }
1863
1864 /*
1865  * reads the tree block backref for an extent. tree level and root are returned
1866  * through out_level and out_root. ptr must point to a 0 value for the first
1867  * call and may be modified (see get_extent_inline_ref comment).
1868  * returns 0 if data was provided, 1 if there was no more data to provide or
1869  * <0 on error.
1870  */
1871 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1872                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1873                             u32 item_size, u64 *out_root, u8 *out_level)
1874 {
1875         int ret;
1876         int type;
1877         struct btrfs_extent_inline_ref *eiref;
1878
1879         if (*ptr == (unsigned long)-1)
1880                 return 1;
1881
1882         while (1) {
1883                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1884                                               &eiref, &type);
1885                 if (ret < 0)
1886                         return ret;
1887
1888                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1889                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1890                         break;
1891
1892                 if (ret == 1)
1893                         return 1;
1894         }
1895
1896         /* we can treat both ref types equally here */
1897         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1898
1899         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1900                 struct btrfs_tree_block_info *info;
1901
1902                 info = (struct btrfs_tree_block_info *)(ei + 1);
1903                 *out_level = btrfs_tree_block_level(eb, info);
1904         } else {
1905                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1906                 *out_level = (u8)key->offset;
1907         }
1908
1909         if (ret == 1)
1910                 *ptr = (unsigned long)-1;
1911
1912         return 0;
1913 }
1914
1915 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1916                              struct extent_inode_elem *inode_list,
1917                              u64 root, u64 extent_item_objectid,
1918                              iterate_extent_inodes_t *iterate, void *ctx)
1919 {
1920         struct extent_inode_elem *eie;
1921         int ret = 0;
1922
1923         for (eie = inode_list; eie; eie = eie->next) {
1924                 btrfs_debug(fs_info,
1925                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1926                             extent_item_objectid, eie->inum,
1927                             eie->offset, root);
1928                 ret = iterate(eie->inum, eie->offset, root, ctx);
1929                 if (ret) {
1930                         btrfs_debug(fs_info,
1931                                     "stopping iteration for %llu due to ret=%d",
1932                                     extent_item_objectid, ret);
1933                         break;
1934                 }
1935         }
1936
1937         return ret;
1938 }
1939
1940 /*
1941  * calls iterate() for every inode that references the extent identified by
1942  * the given parameters.
1943  * when the iterator function returns a non-zero value, iteration stops.
1944  */
1945 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1946                                 u64 extent_item_objectid, u64 extent_item_pos,
1947                                 int search_commit_root,
1948                                 iterate_extent_inodes_t *iterate, void *ctx,
1949                                 bool ignore_offset)
1950 {
1951         int ret;
1952         struct btrfs_trans_handle *trans = NULL;
1953         struct ulist *refs = NULL;
1954         struct ulist *roots = NULL;
1955         struct ulist_node *ref_node = NULL;
1956         struct ulist_node *root_node = NULL;
1957         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1958         struct ulist_iterator ref_uiter;
1959         struct ulist_iterator root_uiter;
1960
1961         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1962                         extent_item_objectid);
1963
1964         if (!search_commit_root) {
1965                 trans = btrfs_attach_transaction(fs_info->extent_root);
1966                 if (IS_ERR(trans)) {
1967                         if (PTR_ERR(trans) != -ENOENT &&
1968                             PTR_ERR(trans) != -EROFS)
1969                                 return PTR_ERR(trans);
1970                         trans = NULL;
1971                 }
1972         }
1973
1974         if (trans)
1975                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1976         else
1977                 down_read(&fs_info->commit_root_sem);
1978
1979         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1980                                    tree_mod_seq_elem.seq, &refs,
1981                                    &extent_item_pos, ignore_offset);
1982         if (ret)
1983                 goto out;
1984
1985         ULIST_ITER_INIT(&ref_uiter);
1986         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1987                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1988                                                 tree_mod_seq_elem.seq, &roots,
1989                                                 ignore_offset);
1990                 if (ret)
1991                         break;
1992                 ULIST_ITER_INIT(&root_uiter);
1993                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1994                         btrfs_debug(fs_info,
1995                                     "root %llu references leaf %llu, data list %#llx",
1996                                     root_node->val, ref_node->val,
1997                                     ref_node->aux);
1998                         ret = iterate_leaf_refs(fs_info,
1999                                                 (struct extent_inode_elem *)
2000                                                 (uintptr_t)ref_node->aux,
2001                                                 root_node->val,
2002                                                 extent_item_objectid,
2003                                                 iterate, ctx);
2004                 }
2005                 ulist_free(roots);
2006         }
2007
2008         free_leaf_list(refs);
2009 out:
2010         if (trans) {
2011                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2012                 btrfs_end_transaction(trans);
2013         } else {
2014                 up_read(&fs_info->commit_root_sem);
2015         }
2016
2017         return ret;
2018 }
2019
2020 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2021                                 struct btrfs_path *path,
2022                                 iterate_extent_inodes_t *iterate, void *ctx,
2023                                 bool ignore_offset)
2024 {
2025         int ret;
2026         u64 extent_item_pos;
2027         u64 flags = 0;
2028         struct btrfs_key found_key;
2029         int search_commit_root = path->search_commit_root;
2030
2031         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2032         btrfs_release_path(path);
2033         if (ret < 0)
2034                 return ret;
2035         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2036                 return -EINVAL;
2037
2038         extent_item_pos = logical - found_key.objectid;
2039         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2040                                         extent_item_pos, search_commit_root,
2041                                         iterate, ctx, ignore_offset);
2042
2043         return ret;
2044 }
2045
2046 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2047                               struct extent_buffer *eb, void *ctx);
2048
2049 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2050                               struct btrfs_path *path,
2051                               iterate_irefs_t *iterate, void *ctx)
2052 {
2053         int ret = 0;
2054         int slot;
2055         u32 cur;
2056         u32 len;
2057         u32 name_len;
2058         u64 parent = 0;
2059         int found = 0;
2060         struct extent_buffer *eb;
2061         struct btrfs_item *item;
2062         struct btrfs_inode_ref *iref;
2063         struct btrfs_key found_key;
2064
2065         while (!ret) {
2066                 ret = btrfs_find_item(fs_root, path, inum,
2067                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2068                                 &found_key);
2069
2070                 if (ret < 0)
2071                         break;
2072                 if (ret) {
2073                         ret = found ? 0 : -ENOENT;
2074                         break;
2075                 }
2076                 ++found;
2077
2078                 parent = found_key.offset;
2079                 slot = path->slots[0];
2080                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2081                 if (!eb) {
2082                         ret = -ENOMEM;
2083                         break;
2084                 }
2085                 btrfs_release_path(path);
2086
2087                 item = btrfs_item_nr(slot);
2088                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2089
2090                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2091                         name_len = btrfs_inode_ref_name_len(eb, iref);
2092                         /* path must be released before calling iterate()! */
2093                         btrfs_debug(fs_root->fs_info,
2094                                 "following ref at offset %u for inode %llu in tree %llu",
2095                                 cur, found_key.objectid,
2096                                 fs_root->root_key.objectid);
2097                         ret = iterate(parent, name_len,
2098                                       (unsigned long)(iref + 1), eb, ctx);
2099                         if (ret)
2100                                 break;
2101                         len = sizeof(*iref) + name_len;
2102                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2103                 }
2104                 free_extent_buffer(eb);
2105         }
2106
2107         btrfs_release_path(path);
2108
2109         return ret;
2110 }
2111
2112 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2113                                  struct btrfs_path *path,
2114                                  iterate_irefs_t *iterate, void *ctx)
2115 {
2116         int ret;
2117         int slot;
2118         u64 offset = 0;
2119         u64 parent;
2120         int found = 0;
2121         struct extent_buffer *eb;
2122         struct btrfs_inode_extref *extref;
2123         u32 item_size;
2124         u32 cur_offset;
2125         unsigned long ptr;
2126
2127         while (1) {
2128                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2129                                             &offset);
2130                 if (ret < 0)
2131                         break;
2132                 if (ret) {
2133                         ret = found ? 0 : -ENOENT;
2134                         break;
2135                 }
2136                 ++found;
2137
2138                 slot = path->slots[0];
2139                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2140                 if (!eb) {
2141                         ret = -ENOMEM;
2142                         break;
2143                 }
2144                 btrfs_release_path(path);
2145
2146                 item_size = btrfs_item_size_nr(eb, slot);
2147                 ptr = btrfs_item_ptr_offset(eb, slot);
2148                 cur_offset = 0;
2149
2150                 while (cur_offset < item_size) {
2151                         u32 name_len;
2152
2153                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2154                         parent = btrfs_inode_extref_parent(eb, extref);
2155                         name_len = btrfs_inode_extref_name_len(eb, extref);
2156                         ret = iterate(parent, name_len,
2157                                       (unsigned long)&extref->name, eb, ctx);
2158                         if (ret)
2159                                 break;
2160
2161                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2162                         cur_offset += sizeof(*extref);
2163                 }
2164                 free_extent_buffer(eb);
2165
2166                 offset++;
2167         }
2168
2169         btrfs_release_path(path);
2170
2171         return ret;
2172 }
2173
2174 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2175                          struct btrfs_path *path, iterate_irefs_t *iterate,
2176                          void *ctx)
2177 {
2178         int ret;
2179         int found_refs = 0;
2180
2181         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2182         if (!ret)
2183                 ++found_refs;
2184         else if (ret != -ENOENT)
2185                 return ret;
2186
2187         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2188         if (ret == -ENOENT && found_refs)
2189                 return 0;
2190
2191         return ret;
2192 }
2193
2194 /*
2195  * returns 0 if the path could be dumped (probably truncated)
2196  * returns <0 in case of an error
2197  */
2198 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2199                          struct extent_buffer *eb, void *ctx)
2200 {
2201         struct inode_fs_paths *ipath = ctx;
2202         char *fspath;
2203         char *fspath_min;
2204         int i = ipath->fspath->elem_cnt;
2205         const int s_ptr = sizeof(char *);
2206         u32 bytes_left;
2207
2208         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2209                                         ipath->fspath->bytes_left - s_ptr : 0;
2210
2211         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2212         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2213                                    name_off, eb, inum, fspath_min, bytes_left);
2214         if (IS_ERR(fspath))
2215                 return PTR_ERR(fspath);
2216
2217         if (fspath > fspath_min) {
2218                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2219                 ++ipath->fspath->elem_cnt;
2220                 ipath->fspath->bytes_left = fspath - fspath_min;
2221         } else {
2222                 ++ipath->fspath->elem_missed;
2223                 ipath->fspath->bytes_missing += fspath_min - fspath;
2224                 ipath->fspath->bytes_left = 0;
2225         }
2226
2227         return 0;
2228 }
2229
2230 /*
2231  * this dumps all file system paths to the inode into the ipath struct, provided
2232  * is has been created large enough. each path is zero-terminated and accessed
2233  * from ipath->fspath->val[i].
2234  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2235  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2236  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2237  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2238  * have been needed to return all paths.
2239  */
2240 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2241 {
2242         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2243                              inode_to_path, ipath);
2244 }
2245
2246 struct btrfs_data_container *init_data_container(u32 total_bytes)
2247 {
2248         struct btrfs_data_container *data;
2249         size_t alloc_bytes;
2250
2251         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2252         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2253         if (!data)
2254                 return ERR_PTR(-ENOMEM);
2255
2256         if (total_bytes >= sizeof(*data)) {
2257                 data->bytes_left = total_bytes - sizeof(*data);
2258                 data->bytes_missing = 0;
2259         } else {
2260                 data->bytes_missing = sizeof(*data) - total_bytes;
2261                 data->bytes_left = 0;
2262         }
2263
2264         data->elem_cnt = 0;
2265         data->elem_missed = 0;
2266
2267         return data;
2268 }
2269
2270 /*
2271  * allocates space to return multiple file system paths for an inode.
2272  * total_bytes to allocate are passed, note that space usable for actual path
2273  * information will be total_bytes - sizeof(struct inode_fs_paths).
2274  * the returned pointer must be freed with free_ipath() in the end.
2275  */
2276 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2277                                         struct btrfs_path *path)
2278 {
2279         struct inode_fs_paths *ifp;
2280         struct btrfs_data_container *fspath;
2281
2282         fspath = init_data_container(total_bytes);
2283         if (IS_ERR(fspath))
2284                 return ERR_CAST(fspath);
2285
2286         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2287         if (!ifp) {
2288                 kvfree(fspath);
2289                 return ERR_PTR(-ENOMEM);
2290         }
2291
2292         ifp->btrfs_path = path;
2293         ifp->fspath = fspath;
2294         ifp->fs_root = fs_root;
2295
2296         return ifp;
2297 }
2298
2299 void free_ipath(struct inode_fs_paths *ipath)
2300 {
2301         if (!ipath)
2302                 return;
2303         kvfree(ipath->fspath);
2304         kfree(ipath);
2305 }
2306
2307 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2308                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2309 {
2310         struct btrfs_backref_iter *ret;
2311
2312         ret = kzalloc(sizeof(*ret), gfp_flag);
2313         if (!ret)
2314                 return NULL;
2315
2316         ret->path = btrfs_alloc_path();
2317         if (!ret->path) {
2318                 kfree(ret);
2319                 return NULL;
2320         }
2321
2322         /* Current backref iterator only supports iteration in commit root */
2323         ret->path->search_commit_root = 1;
2324         ret->path->skip_locking = 1;
2325         ret->fs_info = fs_info;
2326
2327         return ret;
2328 }
2329
2330 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2331 {
2332         struct btrfs_fs_info *fs_info = iter->fs_info;
2333         struct btrfs_path *path = iter->path;
2334         struct btrfs_extent_item *ei;
2335         struct btrfs_key key;
2336         int ret;
2337
2338         key.objectid = bytenr;
2339         key.type = BTRFS_METADATA_ITEM_KEY;
2340         key.offset = (u64)-1;
2341         iter->bytenr = bytenr;
2342
2343         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
2344         if (ret < 0)
2345                 return ret;
2346         if (ret == 0) {
2347                 ret = -EUCLEAN;
2348                 goto release;
2349         }
2350         if (path->slots[0] == 0) {
2351                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2352                 ret = -EUCLEAN;
2353                 goto release;
2354         }
2355         path->slots[0]--;
2356
2357         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2358         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2359              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2360                 ret = -ENOENT;
2361                 goto release;
2362         }
2363         memcpy(&iter->cur_key, &key, sizeof(key));
2364         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2365                                                     path->slots[0]);
2366         iter->end_ptr = (u32)(iter->item_ptr +
2367                         btrfs_item_size_nr(path->nodes[0], path->slots[0]));
2368         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2369                             struct btrfs_extent_item);
2370
2371         /*
2372          * Only support iteration on tree backref yet.
2373          *
2374          * This is an extra precaution for non skinny-metadata, where
2375          * EXTENT_ITEM is also used for tree blocks, that we can only use
2376          * extent flags to determine if it's a tree block.
2377          */
2378         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2379                 ret = -ENOTSUPP;
2380                 goto release;
2381         }
2382         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2383
2384         /* If there is no inline backref, go search for keyed backref */
2385         if (iter->cur_ptr >= iter->end_ptr) {
2386                 ret = btrfs_next_item(fs_info->extent_root, path);
2387
2388                 /* No inline nor keyed ref */
2389                 if (ret > 0) {
2390                         ret = -ENOENT;
2391                         goto release;
2392                 }
2393                 if (ret < 0)
2394                         goto release;
2395
2396                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2397                                 path->slots[0]);
2398                 if (iter->cur_key.objectid != bytenr ||
2399                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2400                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2401                         ret = -ENOENT;
2402                         goto release;
2403                 }
2404                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2405                                                            path->slots[0]);
2406                 iter->item_ptr = iter->cur_ptr;
2407                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size_nr(
2408                                       path->nodes[0], path->slots[0]));
2409         }
2410
2411         return 0;
2412 release:
2413         btrfs_backref_iter_release(iter);
2414         return ret;
2415 }
2416
2417 /*
2418  * Go to the next backref item of current bytenr, can be either inlined or
2419  * keyed.
2420  *
2421  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2422  *
2423  * Return 0 if we get next backref without problem.
2424  * Return >0 if there is no extra backref for this bytenr.
2425  * Return <0 if there is something wrong happened.
2426  */
2427 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2428 {
2429         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2430         struct btrfs_path *path = iter->path;
2431         struct btrfs_extent_inline_ref *iref;
2432         int ret;
2433         u32 size;
2434
2435         if (btrfs_backref_iter_is_inline_ref(iter)) {
2436                 /* We're still inside the inline refs */
2437                 ASSERT(iter->cur_ptr < iter->end_ptr);
2438
2439                 if (btrfs_backref_has_tree_block_info(iter)) {
2440                         /* First tree block info */
2441                         size = sizeof(struct btrfs_tree_block_info);
2442                 } else {
2443                         /* Use inline ref type to determine the size */
2444                         int type;
2445
2446                         iref = (struct btrfs_extent_inline_ref *)
2447                                 ((unsigned long)iter->cur_ptr);
2448                         type = btrfs_extent_inline_ref_type(eb, iref);
2449
2450                         size = btrfs_extent_inline_ref_size(type);
2451                 }
2452                 iter->cur_ptr += size;
2453                 if (iter->cur_ptr < iter->end_ptr)
2454                         return 0;
2455
2456                 /* All inline items iterated, fall through */
2457         }
2458
2459         /* We're at keyed items, there is no inline item, go to the next one */
2460         ret = btrfs_next_item(iter->fs_info->extent_root, iter->path);
2461         if (ret)
2462                 return ret;
2463
2464         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2465         if (iter->cur_key.objectid != iter->bytenr ||
2466             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2467              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2468                 return 1;
2469         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2470                                         path->slots[0]);
2471         iter->cur_ptr = iter->item_ptr;
2472         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size_nr(path->nodes[0],
2473                                                 path->slots[0]);
2474         return 0;
2475 }
2476
2477 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2478                               struct btrfs_backref_cache *cache, int is_reloc)
2479 {
2480         int i;
2481
2482         cache->rb_root = RB_ROOT;
2483         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2484                 INIT_LIST_HEAD(&cache->pending[i]);
2485         INIT_LIST_HEAD(&cache->changed);
2486         INIT_LIST_HEAD(&cache->detached);
2487         INIT_LIST_HEAD(&cache->leaves);
2488         INIT_LIST_HEAD(&cache->pending_edge);
2489         INIT_LIST_HEAD(&cache->useless_node);
2490         cache->fs_info = fs_info;
2491         cache->is_reloc = is_reloc;
2492 }
2493
2494 struct btrfs_backref_node *btrfs_backref_alloc_node(
2495                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2496 {
2497         struct btrfs_backref_node *node;
2498
2499         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2500         node = kzalloc(sizeof(*node), GFP_NOFS);
2501         if (!node)
2502                 return node;
2503
2504         INIT_LIST_HEAD(&node->list);
2505         INIT_LIST_HEAD(&node->upper);
2506         INIT_LIST_HEAD(&node->lower);
2507         RB_CLEAR_NODE(&node->rb_node);
2508         cache->nr_nodes++;
2509         node->level = level;
2510         node->bytenr = bytenr;
2511
2512         return node;
2513 }
2514
2515 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2516                 struct btrfs_backref_cache *cache)
2517 {
2518         struct btrfs_backref_edge *edge;
2519
2520         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2521         if (edge)
2522                 cache->nr_edges++;
2523         return edge;
2524 }
2525
2526 /*
2527  * Drop the backref node from cache, also cleaning up all its
2528  * upper edges and any uncached nodes in the path.
2529  *
2530  * This cleanup happens bottom up, thus the node should either
2531  * be the lowest node in the cache or a detached node.
2532  */
2533 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2534                                 struct btrfs_backref_node *node)
2535 {
2536         struct btrfs_backref_node *upper;
2537         struct btrfs_backref_edge *edge;
2538
2539         if (!node)
2540                 return;
2541
2542         BUG_ON(!node->lowest && !node->detached);
2543         while (!list_empty(&node->upper)) {
2544                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2545                                   list[LOWER]);
2546                 upper = edge->node[UPPER];
2547                 list_del(&edge->list[LOWER]);
2548                 list_del(&edge->list[UPPER]);
2549                 btrfs_backref_free_edge(cache, edge);
2550
2551                 if (RB_EMPTY_NODE(&upper->rb_node)) {
2552                         BUG_ON(!list_empty(&node->upper));
2553                         btrfs_backref_drop_node(cache, node);
2554                         node = upper;
2555                         node->lowest = 1;
2556                         continue;
2557                 }
2558                 /*
2559                  * Add the node to leaf node list if no other child block
2560                  * cached.
2561                  */
2562                 if (list_empty(&upper->lower)) {
2563                         list_add_tail(&upper->lower, &cache->leaves);
2564                         upper->lowest = 1;
2565                 }
2566         }
2567
2568         btrfs_backref_drop_node(cache, node);
2569 }
2570
2571 /*
2572  * Release all nodes/edges from current cache
2573  */
2574 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2575 {
2576         struct btrfs_backref_node *node;
2577         int i;
2578
2579         while (!list_empty(&cache->detached)) {
2580                 node = list_entry(cache->detached.next,
2581                                   struct btrfs_backref_node, list);
2582                 btrfs_backref_cleanup_node(cache, node);
2583         }
2584
2585         while (!list_empty(&cache->leaves)) {
2586                 node = list_entry(cache->leaves.next,
2587                                   struct btrfs_backref_node, lower);
2588                 btrfs_backref_cleanup_node(cache, node);
2589         }
2590
2591         cache->last_trans = 0;
2592
2593         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2594                 ASSERT(list_empty(&cache->pending[i]));
2595         ASSERT(list_empty(&cache->pending_edge));
2596         ASSERT(list_empty(&cache->useless_node));
2597         ASSERT(list_empty(&cache->changed));
2598         ASSERT(list_empty(&cache->detached));
2599         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2600         ASSERT(!cache->nr_nodes);
2601         ASSERT(!cache->nr_edges);
2602 }
2603
2604 /*
2605  * Handle direct tree backref
2606  *
2607  * Direct tree backref means, the backref item shows its parent bytenr
2608  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2609  *
2610  * @ref_key:    The converted backref key.
2611  *              For keyed backref, it's the item key.
2612  *              For inlined backref, objectid is the bytenr,
2613  *              type is btrfs_inline_ref_type, offset is
2614  *              btrfs_inline_ref_offset.
2615  */
2616 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2617                                       struct btrfs_key *ref_key,
2618                                       struct btrfs_backref_node *cur)
2619 {
2620         struct btrfs_backref_edge *edge;
2621         struct btrfs_backref_node *upper;
2622         struct rb_node *rb_node;
2623
2624         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2625
2626         /* Only reloc root uses backref pointing to itself */
2627         if (ref_key->objectid == ref_key->offset) {
2628                 struct btrfs_root *root;
2629
2630                 cur->is_reloc_root = 1;
2631                 /* Only reloc backref cache cares about a specific root */
2632                 if (cache->is_reloc) {
2633                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2634                         if (WARN_ON(!root))
2635                                 return -ENOENT;
2636                         cur->root = root;
2637                 } else {
2638                         /*
2639                          * For generic purpose backref cache, reloc root node
2640                          * is useless.
2641                          */
2642                         list_add(&cur->list, &cache->useless_node);
2643                 }
2644                 return 0;
2645         }
2646
2647         edge = btrfs_backref_alloc_edge(cache);
2648         if (!edge)
2649                 return -ENOMEM;
2650
2651         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2652         if (!rb_node) {
2653                 /* Parent node not yet cached */
2654                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2655                                            cur->level + 1);
2656                 if (!upper) {
2657                         btrfs_backref_free_edge(cache, edge);
2658                         return -ENOMEM;
2659                 }
2660
2661                 /*
2662                  *  Backrefs for the upper level block isn't cached, add the
2663                  *  block to pending list
2664                  */
2665                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2666         } else {
2667                 /* Parent node already cached */
2668                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2669                 ASSERT(upper->checked);
2670                 INIT_LIST_HEAD(&edge->list[UPPER]);
2671         }
2672         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2673         return 0;
2674 }
2675
2676 /*
2677  * Handle indirect tree backref
2678  *
2679  * Indirect tree backref means, we only know which tree the node belongs to.
2680  * We still need to do a tree search to find out the parents. This is for
2681  * TREE_BLOCK_REF backref (keyed or inlined).
2682  *
2683  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2684  * @tree_key:   The first key of this tree block.
2685  * @path:       A clean (released) path, to avoid allocating path everytime
2686  *              the function get called.
2687  */
2688 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2689                                         struct btrfs_path *path,
2690                                         struct btrfs_key *ref_key,
2691                                         struct btrfs_key *tree_key,
2692                                         struct btrfs_backref_node *cur)
2693 {
2694         struct btrfs_fs_info *fs_info = cache->fs_info;
2695         struct btrfs_backref_node *upper;
2696         struct btrfs_backref_node *lower;
2697         struct btrfs_backref_edge *edge;
2698         struct extent_buffer *eb;
2699         struct btrfs_root *root;
2700         struct rb_node *rb_node;
2701         int level;
2702         bool need_check = true;
2703         int ret;
2704
2705         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2706         if (IS_ERR(root))
2707                 return PTR_ERR(root);
2708         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2709                 cur->cowonly = 1;
2710
2711         if (btrfs_root_level(&root->root_item) == cur->level) {
2712                 /* Tree root */
2713                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2714                 /*
2715                  * For reloc backref cache, we may ignore reloc root.  But for
2716                  * general purpose backref cache, we can't rely on
2717                  * btrfs_should_ignore_reloc_root() as it may conflict with
2718                  * current running relocation and lead to missing root.
2719                  *
2720                  * For general purpose backref cache, reloc root detection is
2721                  * completely relying on direct backref (key->offset is parent
2722                  * bytenr), thus only do such check for reloc cache.
2723                  */
2724                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2725                         btrfs_put_root(root);
2726                         list_add(&cur->list, &cache->useless_node);
2727                 } else {
2728                         cur->root = root;
2729                 }
2730                 return 0;
2731         }
2732
2733         level = cur->level + 1;
2734
2735         /* Search the tree to find parent blocks referring to the block */
2736         path->search_commit_root = 1;
2737         path->skip_locking = 1;
2738         path->lowest_level = level;
2739         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2740         path->lowest_level = 0;
2741         if (ret < 0) {
2742                 btrfs_put_root(root);
2743                 return ret;
2744         }
2745         if (ret > 0 && path->slots[level] > 0)
2746                 path->slots[level]--;
2747
2748         eb = path->nodes[level];
2749         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2750                 btrfs_err(fs_info,
2751 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2752                           cur->bytenr, level - 1, root->root_key.objectid,
2753                           tree_key->objectid, tree_key->type, tree_key->offset);
2754                 btrfs_put_root(root);
2755                 ret = -ENOENT;
2756                 goto out;
2757         }
2758         lower = cur;
2759
2760         /* Add all nodes and edges in the path */
2761         for (; level < BTRFS_MAX_LEVEL; level++) {
2762                 if (!path->nodes[level]) {
2763                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2764                                lower->bytenr);
2765                         /* Same as previous should_ignore_reloc_root() call */
2766                         if (btrfs_should_ignore_reloc_root(root) &&
2767                             cache->is_reloc) {
2768                                 btrfs_put_root(root);
2769                                 list_add(&lower->list, &cache->useless_node);
2770                         } else {
2771                                 lower->root = root;
2772                         }
2773                         break;
2774                 }
2775
2776                 edge = btrfs_backref_alloc_edge(cache);
2777                 if (!edge) {
2778                         btrfs_put_root(root);
2779                         ret = -ENOMEM;
2780                         goto out;
2781                 }
2782
2783                 eb = path->nodes[level];
2784                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2785                 if (!rb_node) {
2786                         upper = btrfs_backref_alloc_node(cache, eb->start,
2787                                                          lower->level + 1);
2788                         if (!upper) {
2789                                 btrfs_put_root(root);
2790                                 btrfs_backref_free_edge(cache, edge);
2791                                 ret = -ENOMEM;
2792                                 goto out;
2793                         }
2794                         upper->owner = btrfs_header_owner(eb);
2795                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2796                                 upper->cowonly = 1;
2797
2798                         /*
2799                          * If we know the block isn't shared we can avoid
2800                          * checking its backrefs.
2801                          */
2802                         if (btrfs_block_can_be_shared(root, eb))
2803                                 upper->checked = 0;
2804                         else
2805                                 upper->checked = 1;
2806
2807                         /*
2808                          * Add the block to pending list if we need to check its
2809                          * backrefs, we only do this once while walking up a
2810                          * tree as we will catch anything else later on.
2811                          */
2812                         if (!upper->checked && need_check) {
2813                                 need_check = false;
2814                                 list_add_tail(&edge->list[UPPER],
2815                                               &cache->pending_edge);
2816                         } else {
2817                                 if (upper->checked)
2818                                         need_check = true;
2819                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2820                         }
2821                 } else {
2822                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2823                                          rb_node);
2824                         ASSERT(upper->checked);
2825                         INIT_LIST_HEAD(&edge->list[UPPER]);
2826                         if (!upper->owner)
2827                                 upper->owner = btrfs_header_owner(eb);
2828                 }
2829                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2830
2831                 if (rb_node) {
2832                         btrfs_put_root(root);
2833                         break;
2834                 }
2835                 lower = upper;
2836                 upper = NULL;
2837         }
2838 out:
2839         btrfs_release_path(path);
2840         return ret;
2841 }
2842
2843 /*
2844  * Add backref node @cur into @cache.
2845  *
2846  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2847  *       links aren't yet bi-directional. Needs to finish such links.
2848  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2849  *
2850  * @path:       Released path for indirect tree backref lookup
2851  * @iter:       Released backref iter for extent tree search
2852  * @node_key:   The first key of the tree block
2853  */
2854 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2855                                 struct btrfs_path *path,
2856                                 struct btrfs_backref_iter *iter,
2857                                 struct btrfs_key *node_key,
2858                                 struct btrfs_backref_node *cur)
2859 {
2860         struct btrfs_fs_info *fs_info = cache->fs_info;
2861         struct btrfs_backref_edge *edge;
2862         struct btrfs_backref_node *exist;
2863         int ret;
2864
2865         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2866         if (ret < 0)
2867                 return ret;
2868         /*
2869          * We skip the first btrfs_tree_block_info, as we don't use the key
2870          * stored in it, but fetch it from the tree block
2871          */
2872         if (btrfs_backref_has_tree_block_info(iter)) {
2873                 ret = btrfs_backref_iter_next(iter);
2874                 if (ret < 0)
2875                         goto out;
2876                 /* No extra backref? This means the tree block is corrupted */
2877                 if (ret > 0) {
2878                         ret = -EUCLEAN;
2879                         goto out;
2880                 }
2881         }
2882         WARN_ON(cur->checked);
2883         if (!list_empty(&cur->upper)) {
2884                 /*
2885                  * The backref was added previously when processing backref of
2886                  * type BTRFS_TREE_BLOCK_REF_KEY
2887                  */
2888                 ASSERT(list_is_singular(&cur->upper));
2889                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2890                                   list[LOWER]);
2891                 ASSERT(list_empty(&edge->list[UPPER]));
2892                 exist = edge->node[UPPER];
2893                 /*
2894                  * Add the upper level block to pending list if we need check
2895                  * its backrefs
2896                  */
2897                 if (!exist->checked)
2898                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2899         } else {
2900                 exist = NULL;
2901         }
2902
2903         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2904                 struct extent_buffer *eb;
2905                 struct btrfs_key key;
2906                 int type;
2907
2908                 cond_resched();
2909                 eb = btrfs_backref_get_eb(iter);
2910
2911                 key.objectid = iter->bytenr;
2912                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2913                         struct btrfs_extent_inline_ref *iref;
2914
2915                         /* Update key for inline backref */
2916                         iref = (struct btrfs_extent_inline_ref *)
2917                                 ((unsigned long)iter->cur_ptr);
2918                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2919                                                         BTRFS_REF_TYPE_BLOCK);
2920                         if (type == BTRFS_REF_TYPE_INVALID) {
2921                                 ret = -EUCLEAN;
2922                                 goto out;
2923                         }
2924                         key.type = type;
2925                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2926                 } else {
2927                         key.type = iter->cur_key.type;
2928                         key.offset = iter->cur_key.offset;
2929                 }
2930
2931                 /*
2932                  * Parent node found and matches current inline ref, no need to
2933                  * rebuild this node for this inline ref
2934                  */
2935                 if (exist &&
2936                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2937                       exist->owner == key.offset) ||
2938                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2939                       exist->bytenr == key.offset))) {
2940                         exist = NULL;
2941                         continue;
2942                 }
2943
2944                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2945                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2946                         ret = handle_direct_tree_backref(cache, &key, cur);
2947                         if (ret < 0)
2948                                 goto out;
2949                         continue;
2950                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2951                         ret = -EINVAL;
2952                         btrfs_print_v0_err(fs_info);
2953                         btrfs_handle_fs_error(fs_info, ret, NULL);
2954                         goto out;
2955                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2956                         continue;
2957                 }
2958
2959                 /*
2960                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2961                  * means the root objectid. We need to search the tree to get
2962                  * its parent bytenr.
2963                  */
2964                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2965                                                    cur);
2966                 if (ret < 0)
2967                         goto out;
2968         }
2969         ret = 0;
2970         cur->checked = 1;
2971         WARN_ON(exist);
2972 out:
2973         btrfs_backref_iter_release(iter);
2974         return ret;
2975 }
2976
2977 /*
2978  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
2979  */
2980 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
2981                                      struct btrfs_backref_node *start)
2982 {
2983         struct list_head *useless_node = &cache->useless_node;
2984         struct btrfs_backref_edge *edge;
2985         struct rb_node *rb_node;
2986         LIST_HEAD(pending_edge);
2987
2988         ASSERT(start->checked);
2989
2990         /* Insert this node to cache if it's not COW-only */
2991         if (!start->cowonly) {
2992                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
2993                                            &start->rb_node);
2994                 if (rb_node)
2995                         btrfs_backref_panic(cache->fs_info, start->bytenr,
2996                                             -EEXIST);
2997                 list_add_tail(&start->lower, &cache->leaves);
2998         }
2999
3000         /*
3001          * Use breadth first search to iterate all related edges.
3002          *
3003          * The starting points are all the edges of this node
3004          */
3005         list_for_each_entry(edge, &start->upper, list[LOWER])
3006                 list_add_tail(&edge->list[UPPER], &pending_edge);
3007
3008         while (!list_empty(&pending_edge)) {
3009                 struct btrfs_backref_node *upper;
3010                 struct btrfs_backref_node *lower;
3011
3012                 edge = list_first_entry(&pending_edge,
3013                                 struct btrfs_backref_edge, list[UPPER]);
3014                 list_del_init(&edge->list[UPPER]);
3015                 upper = edge->node[UPPER];
3016                 lower = edge->node[LOWER];
3017
3018                 /* Parent is detached, no need to keep any edges */
3019                 if (upper->detached) {
3020                         list_del(&edge->list[LOWER]);
3021                         btrfs_backref_free_edge(cache, edge);
3022
3023                         /* Lower node is orphan, queue for cleanup */
3024                         if (list_empty(&lower->upper))
3025                                 list_add(&lower->list, useless_node);
3026                         continue;
3027                 }
3028
3029                 /*
3030                  * All new nodes added in current build_backref_tree() haven't
3031                  * been linked to the cache rb tree.
3032                  * So if we have upper->rb_node populated, this means a cache
3033                  * hit. We only need to link the edge, as @upper and all its
3034                  * parents have already been linked.
3035                  */
3036                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3037                         if (upper->lowest) {
3038                                 list_del_init(&upper->lower);
3039                                 upper->lowest = 0;
3040                         }
3041
3042                         list_add_tail(&edge->list[UPPER], &upper->lower);
3043                         continue;
3044                 }
3045
3046                 /* Sanity check, we shouldn't have any unchecked nodes */
3047                 if (!upper->checked) {
3048                         ASSERT(0);
3049                         return -EUCLEAN;
3050                 }
3051
3052                 /* Sanity check, COW-only node has non-COW-only parent */
3053                 if (start->cowonly != upper->cowonly) {
3054                         ASSERT(0);
3055                         return -EUCLEAN;
3056                 }
3057
3058                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3059                 if (!upper->cowonly) {
3060                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3061                                                    &upper->rb_node);
3062                         if (rb_node) {
3063                                 btrfs_backref_panic(cache->fs_info,
3064                                                 upper->bytenr, -EEXIST);
3065                                 return -EUCLEAN;
3066                         }
3067                 }
3068
3069                 list_add_tail(&edge->list[UPPER], &upper->lower);
3070
3071                 /*
3072                  * Also queue all the parent edges of this uncached node
3073                  * to finish the upper linkage
3074                  */
3075                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3076                         list_add_tail(&edge->list[UPPER], &pending_edge);
3077         }
3078         return 0;
3079 }
3080
3081 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3082                                  struct btrfs_backref_node *node)
3083 {
3084         struct btrfs_backref_node *lower;
3085         struct btrfs_backref_node *upper;
3086         struct btrfs_backref_edge *edge;
3087
3088         while (!list_empty(&cache->useless_node)) {
3089                 lower = list_first_entry(&cache->useless_node,
3090                                    struct btrfs_backref_node, list);
3091                 list_del_init(&lower->list);
3092         }
3093         while (!list_empty(&cache->pending_edge)) {
3094                 edge = list_first_entry(&cache->pending_edge,
3095                                 struct btrfs_backref_edge, list[UPPER]);
3096                 list_del(&edge->list[UPPER]);
3097                 list_del(&edge->list[LOWER]);
3098                 lower = edge->node[LOWER];
3099                 upper = edge->node[UPPER];
3100                 btrfs_backref_free_edge(cache, edge);
3101
3102                 /*
3103                  * Lower is no longer linked to any upper backref nodes and
3104                  * isn't in the cache, we can free it ourselves.
3105                  */
3106                 if (list_empty(&lower->upper) &&
3107                     RB_EMPTY_NODE(&lower->rb_node))
3108                         list_add(&lower->list, &cache->useless_node);
3109
3110                 if (!RB_EMPTY_NODE(&upper->rb_node))
3111                         continue;
3112
3113                 /* Add this guy's upper edges to the list to process */
3114                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3115                         list_add_tail(&edge->list[UPPER],
3116                                       &cache->pending_edge);
3117                 if (list_empty(&upper->upper))
3118                         list_add(&upper->list, &cache->useless_node);
3119         }
3120
3121         while (!list_empty(&cache->useless_node)) {
3122                 lower = list_first_entry(&cache->useless_node,
3123                                    struct btrfs_backref_node, list);
3124                 list_del_init(&lower->list);
3125                 if (lower == node)
3126                         node = NULL;
3127                 btrfs_backref_free_node(cache, lower);
3128         }
3129
3130         btrfs_backref_cleanup_node(cache, node);
3131         ASSERT(list_empty(&cache->useless_node) &&
3132                list_empty(&cache->pending_edge));
3133 }