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 #include "tree-mod-log.h"
  18
  19 /* Just an arbitrary number so we can be sure this happened */
  20 #define BACKREF_FOUND_SHARED 6
  21
  22 struct extent_inode_elem {
  23         u64 inum;
  24         u64 offset;
  25         struct extent_inode_elem *next;
  26 };
  27
  28 static int check_extent_in_eb(const struct btrfs_key *key,
  29                               const struct extent_buffer *eb,
  30                               const struct btrfs_file_extent_item *fi,
  31                               u64 extent_item_pos,
  32                               struct extent_inode_elem **eie,
  33                               bool ignore_offset)
  34 {
  35         u64 offset = 0;
  36         struct extent_inode_elem *e;
  37
  38         if (!ignore_offset &&
  39             !btrfs_file_extent_compression(eb, fi) &&
  40             !btrfs_file_extent_encryption(eb, fi) &&
  41             !btrfs_file_extent_other_encoding(eb, fi)) {
  42                 u64 data_offset;
  43                 u64 data_len;
  44
  45                 data_offset = btrfs_file_extent_offset(eb, fi);
  46                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  47
  48                 if (extent_item_pos < data_offset ||
  49                     extent_item_pos >= data_offset + data_len)
  50                         return 1;
  51                 offset = extent_item_pos - data_offset;
  52         }
  53
  54         e = kmalloc(sizeof(*e), GFP_NOFS);
  55         if (!e)
  56                 return -ENOMEM;
  57
  58         e->next = *eie;
  59         e->inum = key->objectid;
  60         e->offset = key->offset + offset;
  61         *eie = e;
  62
  63         return 0;
  64 }
  65
  66 static void free_inode_elem_list(struct extent_inode_elem *eie)
  67 {
  68         struct extent_inode_elem *eie_next;
  69
  70         for (; eie; eie = eie_next) {
  71                 eie_next = eie->next;
  72                 kfree(eie);
  73         }
  74 }
  75
  76 static int find_extent_in_eb(const struct extent_buffer *eb,
  77                              u64 wanted_disk_byte, u64 extent_item_pos,
  78                              struct extent_inode_elem **eie,
  79                              bool ignore_offset)
  80 {
  81         u64 disk_byte;
  82         struct btrfs_key key;
  83         struct btrfs_file_extent_item *fi;
  84         int slot;
  85         int nritems;
  86         int extent_type;
  87         int ret;
  88
  89         /*
  90          * from the shared data ref, we only have the leaf but we need
  91          * the key. thus, we must look into all items and see that we
  92          * find one (some) with a reference to our extent item.
  93          */
  94         nritems = btrfs_header_nritems(eb);
  95         for (slot = 0; slot < nritems; ++slot) {
  96                 btrfs_item_key_to_cpu(eb, &key, slot);
  97                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  98                         continue;
  99                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 100                 extent_type = btrfs_file_extent_type(eb, fi);
 101                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 102                         continue;
 103                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 104                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 105                 if (disk_byte != wanted_disk_byte)
 106                         continue;
 107
 108                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 109                 if (ret < 0)
 110                         return ret;
 111         }
 112
 113         return 0;
 114 }
 115
 116 struct preftree {
 117         struct rb_root_cached root;
 118         unsigned int count;
 119 };
 120
 121 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 122
 123 struct preftrees {
 124         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 125         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 126         struct preftree indirect_missing_keys;
 127 };
 128
 129 /*
 130  * Checks for a shared extent during backref search.
 131  *
 132  * The share_count tracks prelim_refs (direct and indirect) having a
 133  * ref->count >0:
 134  *  - incremented when a ref->count transitions to >0
 135  *  - decremented when a ref->count transitions to <1
 136  */
 137 struct share_check {
 138         u64 root_objectid;
 139         u64 inum;
 140         int share_count;
 141 };
 142
 143 static inline int extent_is_shared(struct share_check *sc)
 144 {
 145         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 146 }
 147
 148 static struct kmem_cache *btrfs_prelim_ref_cache;
 149
 150 int __init btrfs_prelim_ref_init(void)
 151 {
 152         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 153                                         sizeof(struct prelim_ref),
 154                                         0,
 155                                         SLAB_MEM_SPREAD,
 156                                         NULL);
 157         if (!btrfs_prelim_ref_cache)
 158                 return -ENOMEM;
 159         return 0;
 160 }
 161
 162 void __cold btrfs_prelim_ref_exit(void)
 163 {
 164         kmem_cache_destroy(btrfs_prelim_ref_cache);
 165 }
 166
 167 static void free_pref(struct prelim_ref *ref)
 168 {
 169         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 170 }
 171
 172 /*
 173  * Return 0 when both refs are for the same block (and can be merged).
 174  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 175  * indicates a 'higher' block.
 176  */
 177 static int prelim_ref_compare(struct prelim_ref *ref1,
 178                               struct prelim_ref *ref2)
 179 {
 180         if (ref1->level < ref2->level)
 181                 return -1;
 182         if (ref1->level > ref2->level)
 183                 return 1;
 184         if (ref1->root_id < ref2->root_id)
 185                 return -1;
 186         if (ref1->root_id > ref2->root_id)
 187                 return 1;
 188         if (ref1->key_for_search.type < ref2->key_for_search.type)
 189                 return -1;
 190         if (ref1->key_for_search.type > ref2->key_for_search.type)
 191                 return 1;
 192         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 193                 return -1;
 194         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 195                 return 1;
 196         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 197                 return -1;
 198         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 199                 return 1;
 200         if (ref1->parent < ref2->parent)
 201                 return -1;
 202         if (ref1->parent > ref2->parent)
 203                 return 1;
 204
 205         return 0;
 206 }
 207
 208 static void update_share_count(struct share_check *sc, int oldcount,
 209                                int newcount)
 210 {
 211         if ((!sc) || (oldcount == 0 && newcount < 1))
 212                 return;
 213
 214         if (oldcount > 0 && newcount < 1)
 215                 sc->share_count--;
 216         else if (oldcount < 1 && newcount > 0)
 217                 sc->share_count++;
 218 }
 219
 220 /*
 221  * Add @newref to the @root rbtree, merging identical refs.
 222  *
 223  * Callers should assume that newref has been freed after calling.
 224  */
 225 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 226                               struct preftree *preftree,
 227                               struct prelim_ref *newref,
 228                               struct share_check *sc)
 229 {
 230         struct rb_root_cached *root;
 231         struct rb_node **p;
 232         struct rb_node *parent = NULL;
 233         struct prelim_ref *ref;
 234         int result;
 235         bool leftmost = true;
 236
 237         root = &preftree->root;
 238         p = &root->rb_root.rb_node;
 239
 240         while (*p) {
 241                 parent = *p;
 242                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 243                 result = prelim_ref_compare(ref, newref);
 244                 if (result < 0) {
 245                         p = &(*p)->rb_left;
 246                 } else if (result > 0) {
 247                         p = &(*p)->rb_right;
 248                         leftmost = false;
 249                 } else {
 250                         /* Identical refs, merge them and free @newref */
 251                         struct extent_inode_elem *eie = ref->inode_list;
 252
 253                         while (eie && eie->next)
 254                                 eie = eie->next;
 255
 256                         if (!eie)
 257                                 ref->inode_list = newref->inode_list;
 258                         else
 259                                 eie->next = newref->inode_list;
 260                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 261                                                      preftree->count);
 262                         /*
 263                          * A delayed ref can have newref->count < 0.
 264                          * The ref->count is updated to follow any
 265                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 266                          */
 267                         update_share_count(sc, ref->count,
 268                                            ref->count + newref->count);
 269                         ref->count += newref->count;
 270                         free_pref(newref);
 271                         return;
 272                 }
 273         }
 274
 275         update_share_count(sc, 0, newref->count);
 276         preftree->count++;
 277         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 278         rb_link_node(&newref->rbnode, parent, p);
 279         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 280 }
 281
 282 /*
 283  * Release the entire tree.  We don't care about internal consistency so
 284  * just free everything and then reset the tree root.
 285  */
 286 static void prelim_release(struct preftree *preftree)
 287 {
 288         struct prelim_ref *ref, *next_ref;
 289
 290         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 291                                              &preftree->root.rb_root, rbnode)
 292                 free_pref(ref);
 293
 294         preftree->root = RB_ROOT_CACHED;
 295         preftree->count = 0;
 296 }
 297
 298 /*
 299  * the rules for all callers of this function are:
 300  * - obtaining the parent is the goal
 301  * - if you add a key, you must know that it is a correct key
 302  * - if you cannot add the parent or a correct key, then we will look into the
 303  *   block later to set a correct key
 304  *
 305  * delayed refs
 306  * ============
 307  *        backref type | shared | indirect | shared | indirect
 308  * information         |   tree |     tree |   data |     data
 309  * --------------------+--------+----------+--------+----------
 310  *      parent logical |    y   |     -    |    -   |     -
 311  *      key to resolve |    -   |     y    |    y   |     y
 312  *  tree block logical |    -   |     -    |    -   |     -
 313  *  root for resolving |    y   |     y    |    y   |     y
 314  *
 315  * - column 1:       we've the parent -> done
 316  * - column 2, 3, 4: we use the key to find the parent
 317  *
 318  * on disk refs (inline or keyed)
 319  * ==============================
 320  *        backref type | shared | indirect | shared | indirect
 321  * information         |   tree |     tree |   data |     data
 322  * --------------------+--------+----------+--------+----------
 323  *      parent logical |    y   |     -    |    y   |     -
 324  *      key to resolve |    -   |     -    |    -   |     y
 325  *  tree block logical |    y   |     y    |    y   |     y
 326  *  root for resolving |    -   |     y    |    y   |     y
 327  *
 328  * - column 1, 3: we've the parent -> done
 329  * - column 2:    we take the first key from the block to find the parent
 330  *                (see add_missing_keys)
 331  * - column 4:    we use the key to find the parent
 332  *
 333  * additional information that's available but not required to find the parent
 334  * block might help in merging entries to gain some speed.
 335  */
 336 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 337                           struct preftree *preftree, u64 root_id,
 338                           const struct btrfs_key *key, int level, u64 parent,
 339                           u64 wanted_disk_byte, int count,
 340                           struct share_check *sc, gfp_t gfp_mask)
 341 {
 342         struct prelim_ref *ref;
 343
 344         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 345                 return 0;
 346
 347         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 348         if (!ref)
 349                 return -ENOMEM;
 350
 351         ref->root_id = root_id;
 352         if (key)
 353                 ref->key_for_search = *key;
 354         else
 355                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 356
 357         ref->inode_list = NULL;
 358         ref->level = level;
 359         ref->count = count;
 360         ref->parent = parent;
 361         ref->wanted_disk_byte = wanted_disk_byte;
 362         prelim_ref_insert(fs_info, preftree, ref, sc);
 363         return extent_is_shared(sc);
 364 }
 365
 366 /* direct refs use root == 0, key == NULL */
 367 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 368                           struct preftrees *preftrees, int level, u64 parent,
 369                           u64 wanted_disk_byte, int count,
 370                           struct share_check *sc, gfp_t gfp_mask)
 371 {
 372         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 373                               parent, wanted_disk_byte, count, sc, gfp_mask);
 374 }
 375
 376 /* indirect refs use parent == 0 */
 377 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 378                             struct preftrees *preftrees, u64 root_id,
 379                             const struct btrfs_key *key, int level,
 380                             u64 wanted_disk_byte, int count,
 381                             struct share_check *sc, gfp_t gfp_mask)
 382 {
 383         struct preftree *tree = &preftrees->indirect;
 384
 385         if (!key)
 386                 tree = &preftrees->indirect_missing_keys;
 387         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 388                               wanted_disk_byte, count, sc, gfp_mask);
 389 }
 390
 391 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
 392 {
 393         struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
 394         struct rb_node *parent = NULL;
 395         struct prelim_ref *ref = NULL;
 396         struct prelim_ref target = {};
 397         int result;
 398
 399         target.parent = bytenr;
 400
 401         while (*p) {
 402                 parent = *p;
 403                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 404                 result = prelim_ref_compare(ref, &target);
 405
 406                 if (result < 0)
 407                         p = &(*p)->rb_left;
 408                 else if (result > 0)
 409                         p = &(*p)->rb_right;
 410                 else
 411                         return 1;
 412         }
 413         return 0;
 414 }
 415
 416 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 417                            struct ulist *parents,
 418                            struct preftrees *preftrees, struct prelim_ref *ref,
 419                            int level, u64 time_seq, const u64 *extent_item_pos,
 420                            bool ignore_offset)
 421 {
 422         int ret = 0;
 423         int slot;
 424         struct extent_buffer *eb;
 425         struct btrfs_key key;
 426         struct btrfs_key *key_for_search = &ref->key_for_search;
 427         struct btrfs_file_extent_item *fi;
 428         struct extent_inode_elem *eie = NULL, *old = NULL;
 429         u64 disk_byte;
 430         u64 wanted_disk_byte = ref->wanted_disk_byte;
 431         u64 count = 0;
 432         u64 data_offset;
 433
 434         if (level != 0) {
 435                 eb = path->nodes[level];
 436                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 437                 if (ret < 0)
 438                         return ret;
 439                 return 0;
 440         }
 441
 442         /*
 443          * 1. We normally enter this function with the path already pointing to
 444          *    the first item to check. But sometimes, we may enter it with
 445          *    slot == nritems.
 446          * 2. We are searching for normal backref but bytenr of this leaf
 447          *    matches shared data backref
 448          * 3. The leaf owner is not equal to the root we are searching
 449          *
 450          * For these cases, go to the next leaf before we continue.
 451          */
 452         eb = path->nodes[0];
 453         if (path->slots[0] >= btrfs_header_nritems(eb) ||
 454             is_shared_data_backref(preftrees, eb->start) ||
 455             ref->root_id != btrfs_header_owner(eb)) {
 456                 if (time_seq == BTRFS_SEQ_LAST)
 457                         ret = btrfs_next_leaf(root, path);
 458                 else
 459                         ret = btrfs_next_old_leaf(root, path, time_seq);
 460         }
 461
 462         while (!ret && count < ref->count) {
 463                 eb = path->nodes[0];
 464                 slot = path->slots[0];
 465
 466                 btrfs_item_key_to_cpu(eb, &key, slot);
 467
 468                 if (key.objectid != key_for_search->objectid ||
 469                     key.type != BTRFS_EXTENT_DATA_KEY)
 470                         break;
 471
 472                 /*
 473                  * We are searching for normal backref but bytenr of this leaf
 474                  * matches shared data backref, OR
 475                  * the leaf owner is not equal to the root we are searching for
 476                  */
 477                 if (slot == 0 &&
 478                     (is_shared_data_backref(preftrees, eb->start) ||
 479                      ref->root_id != btrfs_header_owner(eb))) {
 480                         if (time_seq == BTRFS_SEQ_LAST)
 481                                 ret = btrfs_next_leaf(root, path);
 482                         else
 483                                 ret = btrfs_next_old_leaf(root, path, time_seq);
 484                         continue;
 485                 }
 486                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 487                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 488                 data_offset = btrfs_file_extent_offset(eb, fi);
 489
 490                 if (disk_byte == wanted_disk_byte) {
 491                         eie = NULL;
 492                         old = NULL;
 493                         if (ref->key_for_search.offset == key.offset - data_offset)
 494                                 count++;
 495                         else
 496                                 goto next;
 497                         if (extent_item_pos) {
 498                                 ret = check_extent_in_eb(&key, eb, fi,
 499                                                 *extent_item_pos,
 500                                                 &eie, ignore_offset);
 501                                 if (ret < 0)
 502                                         break;
 503                         }
 504                         if (ret > 0)
 505                                 goto next;
 506                         ret = ulist_add_merge_ptr(parents, eb->start,
 507                                                   eie, (void **)&old, GFP_NOFS);
 508                         if (ret < 0)
 509                                 break;
 510                         if (!ret && extent_item_pos) {
 511                                 while (old->next)
 512                                         old = old->next;
 513                                 old->next = eie;
 514                         }
 515                         eie = NULL;
 516                 }
 517 next:
 518                 if (time_seq == BTRFS_SEQ_LAST)
 519                         ret = btrfs_next_item(root, path);
 520                 else
 521                         ret = btrfs_next_old_item(root, path, time_seq);
 522         }
 523
 524         if (ret > 0)
 525                 ret = 0;
 526         else if (ret < 0)
 527                 free_inode_elem_list(eie);
 528         return ret;
 529 }
 530
 531 /*
 532  * resolve an indirect backref in the form (root_id, key, level)
 533  * to a logical address
 534  */
 535 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 536                                 struct btrfs_path *path, u64 time_seq,
 537                                 struct preftrees *preftrees,
 538                                 struct prelim_ref *ref, struct ulist *parents,
 539                                 const u64 *extent_item_pos, bool ignore_offset)
 540 {
 541         struct btrfs_root *root;
 542         struct extent_buffer *eb;
 543         int ret = 0;
 544         int root_level;
 545         int level = ref->level;
 546         struct btrfs_key search_key = ref->key_for_search;
 547
 548         /*
 549          * If we're search_commit_root we could possibly be holding locks on
 550          * other tree nodes.  This happens when qgroups does backref walks when
 551          * adding new delayed refs.  To deal with this we need to look in cache
 552          * for the root, and if we don't find it then we need to search the
 553          * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
 554          * here.
 555          */
 556         if (path->search_commit_root)
 557                 root = btrfs_get_fs_root_commit_root(fs_info, path, ref->root_id);
 558         else
 559                 root = btrfs_get_fs_root(fs_info, ref->root_id, false);
 560         if (IS_ERR(root)) {
 561                 ret = PTR_ERR(root);
 562                 goto out_free;
 563         }
 564
 565         if (!path->search_commit_root &&
 566             test_bit(BTRFS_ROOT_DELETING, &root->state)) {
 567                 ret = -ENOENT;
 568                 goto out;
 569         }
 570
 571         if (btrfs_is_testing(fs_info)) {
 572                 ret = -ENOENT;
 573                 goto out;
 574         }
 575
 576         if (path->search_commit_root)
 577                 root_level = btrfs_header_level(root->commit_root);
 578         else if (time_seq == BTRFS_SEQ_LAST)
 579                 root_level = btrfs_header_level(root->node);
 580         else
 581                 root_level = btrfs_old_root_level(root, time_seq);
 582
 583         if (root_level + 1 == level)
 584                 goto out;
 585
 586         /*
 587          * We can often find data backrefs with an offset that is too large
 588          * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
 589          * subtracting a file's offset with the data offset of its
 590          * corresponding extent data item. This can happen for example in the
 591          * clone ioctl.
 592          *
 593          * So if we detect such case we set the search key's offset to zero to
 594          * make sure we will find the matching file extent item at
 595          * add_all_parents(), otherwise we will miss it because the offset
 596          * taken form the backref is much larger then the offset of the file
 597          * extent item. This can make us scan a very large number of file
 598          * extent items, but at least it will not make us miss any.
 599          *
 600          * This is an ugly workaround for a behaviour that should have never
 601          * existed, but it does and a fix for the clone ioctl would touch a lot
 602          * of places, cause backwards incompatibility and would not fix the
 603          * problem for extents cloned with older kernels.
 604          */
 605         if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
 606             search_key.offset >= LLONG_MAX)
 607                 search_key.offset = 0;
 608         path->lowest_level = level;
 609         if (time_seq == BTRFS_SEQ_LAST)
 610                 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 611         else
 612                 ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
 613
 614         btrfs_debug(fs_info,
 615                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 616                  ref->root_id, level, ref->count, ret,
 617                  ref->key_for_search.objectid, ref->key_for_search.type,
 618                  ref->key_for_search.offset);
 619         if (ret < 0)
 620                 goto out;
 621
 622         eb = path->nodes[level];
 623         while (!eb) {
 624                 if (WARN_ON(!level)) {
 625                         ret = 1;
 626                         goto out;
 627                 }
 628                 level--;
 629                 eb = path->nodes[level];
 630         }
 631
 632         ret = add_all_parents(root, path, parents, preftrees, ref, level,
 633                               time_seq, extent_item_pos, ignore_offset);
 634 out:
 635         btrfs_put_root(root);
 636 out_free:
 637         path->lowest_level = 0;
 638         btrfs_release_path(path);
 639         return ret;
 640 }
 641
 642 static struct extent_inode_elem *
 643 unode_aux_to_inode_list(struct ulist_node *node)
 644 {
 645         if (!node)
 646                 return NULL;
 647         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 648 }
 649
 650 /*
 651  * We maintain three separate rbtrees: one for direct refs, one for
 652  * indirect refs which have a key, and one for indirect refs which do not
 653  * have a key. Each tree does merge on insertion.
 654  *
 655  * Once all of the references are located, we iterate over the tree of
 656  * indirect refs with missing keys. An appropriate key is located and
 657  * the ref is moved onto the tree for indirect refs. After all missing
 658  * keys are thus located, we iterate over the indirect ref tree, resolve
 659  * each reference, and then insert the resolved reference onto the
 660  * direct tree (merging there too).
 661  *
 662  * New backrefs (i.e., for parent nodes) are added to the appropriate
 663  * rbtree as they are encountered. The new backrefs are subsequently
 664  * resolved as above.
 665  */
 666 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 667                                  struct btrfs_path *path, u64 time_seq,
 668                                  struct preftrees *preftrees,
 669                                  const u64 *extent_item_pos,
 670                                  struct share_check *sc, bool ignore_offset)
 671 {
 672         int err;
 673         int ret = 0;
 674         struct ulist *parents;
 675         struct ulist_node *node;
 676         struct ulist_iterator uiter;
 677         struct rb_node *rnode;
 678
 679         parents = ulist_alloc(GFP_NOFS);
 680         if (!parents)
 681                 return -ENOMEM;
 682
 683         /*
 684          * We could trade memory usage for performance here by iterating
 685          * the tree, allocating new refs for each insertion, and then
 686          * freeing the entire indirect tree when we're done.  In some test
 687          * cases, the tree can grow quite large (~200k objects).
 688          */
 689         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 690                 struct prelim_ref *ref;
 691
 692                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 693                 if (WARN(ref->parent,
 694                          "BUG: direct ref found in indirect tree")) {
 695                         ret = -EINVAL;
 696                         goto out;
 697                 }
 698
 699                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 700                 preftrees->indirect.count--;
 701
 702                 if (ref->count == 0) {
 703                         free_pref(ref);
 704                         continue;
 705                 }
 706
 707                 if (sc && sc->root_objectid &&
 708                     ref->root_id != sc->root_objectid) {
 709                         free_pref(ref);
 710                         ret = BACKREF_FOUND_SHARED;
 711                         goto out;
 712                 }
 713                 err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
 714                                            ref, parents, extent_item_pos,
 715                                            ignore_offset);
 716                 /*
 717                  * we can only tolerate ENOENT,otherwise,we should catch error
 718                  * and return directly.
 719                  */
 720                 if (err == -ENOENT) {
 721                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 722                                           NULL);
 723                         continue;
 724                 } else if (err) {
 725                         free_pref(ref);
 726                         ret = err;
 727                         goto out;
 728                 }
 729
 730                 /* we put the first parent into the ref at hand */
 731                 ULIST_ITER_INIT(&uiter);
 732                 node = ulist_next(parents, &uiter);
 733                 ref->parent = node ? node->val : 0;
 734                 ref->inode_list = unode_aux_to_inode_list(node);
 735
 736                 /* Add a prelim_ref(s) for any other parent(s). */
 737                 while ((node = ulist_next(parents, &uiter))) {
 738                         struct prelim_ref *new_ref;
 739
 740                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 741                                                    GFP_NOFS);
 742                         if (!new_ref) {
 743                                 free_pref(ref);
 744                                 ret = -ENOMEM;
 745                                 goto out;
 746                         }
 747                         memcpy(new_ref, ref, sizeof(*ref));
 748                         new_ref->parent = node->val;
 749                         new_ref->inode_list = unode_aux_to_inode_list(node);
 750                         prelim_ref_insert(fs_info, &preftrees->direct,
 751                                           new_ref, NULL);
 752                 }
 753
 754                 /*
 755                  * Now it's a direct ref, put it in the direct tree. We must
 756                  * do this last because the ref could be merged/freed here.
 757                  */
 758                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 759
 760                 ulist_reinit(parents);
 761                 cond_resched();
 762         }
 763 out:
 764         ulist_free(parents);
 765         return ret;
 766 }
 767
 768 /*
 769  * read tree blocks and add keys where required.
 770  */
 771 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 772                             struct preftrees *preftrees, bool lock)
 773 {
 774         struct prelim_ref *ref;
 775         struct extent_buffer *eb;
 776         struct preftree *tree = &preftrees->indirect_missing_keys;
 777         struct rb_node *node;
 778
 779         while ((node = rb_first_cached(&tree->root))) {
 780                 ref = rb_entry(node, struct prelim_ref, rbnode);
 781                 rb_erase_cached(node, &tree->root);
 782
 783                 BUG_ON(ref->parent);    /* should not be a direct ref */
 784                 BUG_ON(ref->key_for_search.type);
 785                 BUG_ON(!ref->wanted_disk_byte);
 786
 787                 eb = read_tree_block(fs_info, ref->wanted_disk_byte,
 788                                      ref->root_id, 0, ref->level - 1, NULL);
 789                 if (IS_ERR(eb)) {
 790                         free_pref(ref);
 791                         return PTR_ERR(eb);
 792                 }
 793                 if (!extent_buffer_uptodate(eb)) {
 794                         free_pref(ref);
 795                         free_extent_buffer(eb);
 796                         return -EIO;
 797                 }
 798
 799                 if (lock)
 800                         btrfs_tree_read_lock(eb);
 801                 if (btrfs_header_level(eb) == 0)
 802                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 803                 else
 804                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 805                 if (lock)
 806                         btrfs_tree_read_unlock(eb);
 807                 free_extent_buffer(eb);
 808                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 809                 cond_resched();
 810         }
 811         return 0;
 812 }
 813
 814 /*
 815  * add all currently queued delayed refs from this head whose seq nr is
 816  * smaller or equal that seq to the list
 817  */
 818 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 819                             struct btrfs_delayed_ref_head *head, u64 seq,
 820                             struct preftrees *preftrees, struct share_check *sc)
 821 {
 822         struct btrfs_delayed_ref_node *node;
 823         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 824         struct btrfs_key key;
 825         struct btrfs_key tmp_op_key;
 826         struct rb_node *n;
 827         int count;
 828         int ret = 0;
 829
 830         if (extent_op && extent_op->update_key)
 831                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 832
 833         spin_lock(&head->lock);
 834         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 835                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 836                                 ref_node);
 837                 if (node->seq > seq)
 838                         continue;
 839
 840                 switch (node->action) {
 841                 case BTRFS_ADD_DELAYED_EXTENT:
 842                 case BTRFS_UPDATE_DELAYED_HEAD:
 843                         WARN_ON(1);
 844                         continue;
 845                 case BTRFS_ADD_DELAYED_REF:
 846                         count = node->ref_mod;
 847                         break;
 848                 case BTRFS_DROP_DELAYED_REF:
 849                         count = node->ref_mod * -1;
 850                         break;
 851                 default:
 852                         BUG();
 853                 }
 854                 switch (node->type) {
 855                 case BTRFS_TREE_BLOCK_REF_KEY: {
 856                         /* NORMAL INDIRECT METADATA backref */
 857                         struct btrfs_delayed_tree_ref *ref;
 858
 859                         ref = btrfs_delayed_node_to_tree_ref(node);
 860                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 861                                                &tmp_op_key, ref->level + 1,
 862                                                node->bytenr, count, sc,
 863                                                GFP_ATOMIC);
 864                         break;
 865                 }
 866                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 867                         /* SHARED DIRECT METADATA backref */
 868                         struct btrfs_delayed_tree_ref *ref;
 869
 870                         ref = btrfs_delayed_node_to_tree_ref(node);
 871
 872                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 873                                              ref->parent, node->bytenr, count,
 874                                              sc, GFP_ATOMIC);
 875                         break;
 876                 }
 877                 case BTRFS_EXTENT_DATA_REF_KEY: {
 878                         /* NORMAL INDIRECT DATA backref */
 879                         struct btrfs_delayed_data_ref *ref;
 880                         ref = btrfs_delayed_node_to_data_ref(node);
 881
 882                         key.objectid = ref->objectid;
 883                         key.type = BTRFS_EXTENT_DATA_KEY;
 884                         key.offset = ref->offset;
 885
 886                         /*
 887                          * Found a inum that doesn't match our known inum, we
 888                          * know it's shared.
 889                          */
 890                         if (sc && sc->inum && ref->objectid != sc->inum) {
 891                                 ret = BACKREF_FOUND_SHARED;
 892                                 goto out;
 893                         }
 894
 895                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 896                                                &key, 0, node->bytenr, count, sc,
 897                                                GFP_ATOMIC);
 898                         break;
 899                 }
 900                 case BTRFS_SHARED_DATA_REF_KEY: {
 901                         /* SHARED DIRECT FULL backref */
 902                         struct btrfs_delayed_data_ref *ref;
 903
 904                         ref = btrfs_delayed_node_to_data_ref(node);
 905
 906                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 907                                              node->bytenr, count, sc,
 908                                              GFP_ATOMIC);
 909                         break;
 910                 }
 911                 default:
 912                         WARN_ON(1);
 913                 }
 914                 /*
 915                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 916                  * refs have been checked.
 917                  */
 918                 if (ret && (ret != BACKREF_FOUND_SHARED))
 919                         break;
 920         }
 921         if (!ret)
 922                 ret = extent_is_shared(sc);
 923 out:
 924         spin_unlock(&head->lock);
 925         return ret;
 926 }
 927
 928 /*
 929  * add all inline backrefs for bytenr to the list
 930  *
 931  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 932  */
 933 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 934                            struct btrfs_path *path, u64 bytenr,
 935                            int *info_level, struct preftrees *preftrees,
 936                            struct share_check *sc)
 937 {
 938         int ret = 0;
 939         int slot;
 940         struct extent_buffer *leaf;
 941         struct btrfs_key key;
 942         struct btrfs_key found_key;
 943         unsigned long ptr;
 944         unsigned long end;
 945         struct btrfs_extent_item *ei;
 946         u64 flags;
 947         u64 item_size;
 948
 949         /*
 950          * enumerate all inline refs
 951          */
 952         leaf = path->nodes[0];
 953         slot = path->slots[0];
 954
 955         item_size = btrfs_item_size(leaf, slot);
 956         BUG_ON(item_size < sizeof(*ei));
 957
 958         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 959         flags = btrfs_extent_flags(leaf, ei);
 960         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 961
 962         ptr = (unsigned long)(ei + 1);
 963         end = (unsigned long)ei + item_size;
 964
 965         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 966             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 967                 struct btrfs_tree_block_info *info;
 968
 969                 info = (struct btrfs_tree_block_info *)ptr;
 970                 *info_level = btrfs_tree_block_level(leaf, info);
 971                 ptr += sizeof(struct btrfs_tree_block_info);
 972                 BUG_ON(ptr > end);
 973         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 974                 *info_level = found_key.offset;
 975         } else {
 976                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 977         }
 978
 979         while (ptr < end) {
 980                 struct btrfs_extent_inline_ref *iref;
 981                 u64 offset;
 982                 int type;
 983
 984                 iref = (struct btrfs_extent_inline_ref *)ptr;
 985                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 986                                                         BTRFS_REF_TYPE_ANY);
 987                 if (type == BTRFS_REF_TYPE_INVALID)
 988                         return -EUCLEAN;
 989
 990                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 991
 992                 switch (type) {
 993                 case BTRFS_SHARED_BLOCK_REF_KEY:
 994                         ret = add_direct_ref(fs_info, preftrees,
 995                                              *info_level + 1, offset,
 996                                              bytenr, 1, NULL, GFP_NOFS);
 997                         break;
 998                 case BTRFS_SHARED_DATA_REF_KEY: {
 999                         struct btrfs_shared_data_ref *sdref;
1000                         int count;
1001
1002                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1003                         count = btrfs_shared_data_ref_count(leaf, sdref);
1004
1005                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1006                                              bytenr, count, sc, GFP_NOFS);
1007                         break;
1008                 }
1009                 case BTRFS_TREE_BLOCK_REF_KEY:
1010                         ret = add_indirect_ref(fs_info, preftrees, offset,
1011                                                NULL, *info_level + 1,
1012                                                bytenr, 1, NULL, GFP_NOFS);
1013                         break;
1014                 case BTRFS_EXTENT_DATA_REF_KEY: {
1015                         struct btrfs_extent_data_ref *dref;
1016                         int count;
1017                         u64 root;
1018
1019                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1020                         count = btrfs_extent_data_ref_count(leaf, dref);
1021                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1022                                                                       dref);
1023                         key.type = BTRFS_EXTENT_DATA_KEY;
1024                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1025
1026                         if (sc && sc->inum && key.objectid != sc->inum) {
1027                                 ret = BACKREF_FOUND_SHARED;
1028                                 break;
1029                         }
1030
1031                         root = btrfs_extent_data_ref_root(leaf, dref);
1032
1033                         ret = add_indirect_ref(fs_info, preftrees, root,
1034                                                &key, 0, bytenr, count,
1035                                                sc, GFP_NOFS);
1036                         break;
1037                 }
1038                 default:
1039                         WARN_ON(1);
1040                 }
1041                 if (ret)
1042                         return ret;
1043                 ptr += btrfs_extent_inline_ref_size(type);
1044         }
1045
1046         return 0;
1047 }
1048
1049 /*
1050  * add all non-inline backrefs for bytenr to the list
1051  *
1052  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1053  */
1054 static int add_keyed_refs(struct btrfs_root *extent_root,
1055                           struct btrfs_path *path, u64 bytenr,
1056                           int info_level, struct preftrees *preftrees,
1057                           struct share_check *sc)
1058 {
1059         struct btrfs_fs_info *fs_info = extent_root->fs_info;
1060         int ret;
1061         int slot;
1062         struct extent_buffer *leaf;
1063         struct btrfs_key key;
1064
1065         while (1) {
1066                 ret = btrfs_next_item(extent_root, path);
1067                 if (ret < 0)
1068                         break;
1069                 if (ret) {
1070                         ret = 0;
1071                         break;
1072                 }
1073
1074                 slot = path->slots[0];
1075                 leaf = path->nodes[0];
1076                 btrfs_item_key_to_cpu(leaf, &key, slot);
1077
1078                 if (key.objectid != bytenr)
1079                         break;
1080                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1081                         continue;
1082                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1083                         break;
1084
1085                 switch (key.type) {
1086                 case BTRFS_SHARED_BLOCK_REF_KEY:
1087                         /* SHARED DIRECT METADATA backref */
1088                         ret = add_direct_ref(fs_info, preftrees,
1089                                              info_level + 1, key.offset,
1090                                              bytenr, 1, NULL, GFP_NOFS);
1091                         break;
1092                 case BTRFS_SHARED_DATA_REF_KEY: {
1093                         /* SHARED DIRECT FULL backref */
1094                         struct btrfs_shared_data_ref *sdref;
1095                         int count;
1096
1097                         sdref = btrfs_item_ptr(leaf, slot,
1098                                               struct btrfs_shared_data_ref);
1099                         count = btrfs_shared_data_ref_count(leaf, sdref);
1100                         ret = add_direct_ref(fs_info, preftrees, 0,
1101                                              key.offset, bytenr, count,
1102                                              sc, GFP_NOFS);
1103                         break;
1104                 }
1105                 case BTRFS_TREE_BLOCK_REF_KEY:
1106                         /* NORMAL INDIRECT METADATA backref */
1107                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1108                                                NULL, info_level + 1, bytenr,
1109                                                1, NULL, GFP_NOFS);
1110                         break;
1111                 case BTRFS_EXTENT_DATA_REF_KEY: {
1112                         /* NORMAL INDIRECT DATA backref */
1113                         struct btrfs_extent_data_ref *dref;
1114                         int count;
1115                         u64 root;
1116
1117                         dref = btrfs_item_ptr(leaf, slot,
1118                                               struct btrfs_extent_data_ref);
1119                         count = btrfs_extent_data_ref_count(leaf, dref);
1120                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1121                                                                       dref);
1122                         key.type = BTRFS_EXTENT_DATA_KEY;
1123                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1124
1125                         if (sc && sc->inum && key.objectid != sc->inum) {
1126                                 ret = BACKREF_FOUND_SHARED;
1127                                 break;
1128                         }
1129
1130                         root = btrfs_extent_data_ref_root(leaf, dref);
1131                         ret = add_indirect_ref(fs_info, preftrees, root,
1132                                                &key, 0, bytenr, count,
1133                                                sc, GFP_NOFS);
1134                         break;
1135                 }
1136                 default:
1137                         WARN_ON(1);
1138                 }
1139                 if (ret)
1140                         return ret;
1141
1142         }
1143
1144         return ret;
1145 }
1146
1147 /*
1148  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1149  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1150  * indirect refs to their parent bytenr.
1151  * When roots are found, they're added to the roots list
1152  *
1153  * If time_seq is set to BTRFS_SEQ_LAST, it will not search delayed_refs, and
1154  * behave much like trans == NULL case, the difference only lies in it will not
1155  * commit root.
1156  * The special case is for qgroup to search roots in commit_transaction().
1157  *
1158  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1159  * shared extent is detected.
1160  *
1161  * Otherwise this returns 0 for success and <0 for an error.
1162  *
1163  * If ignore_offset is set to false, only extent refs whose offsets match
1164  * extent_item_pos are returned.  If true, every extent ref is returned
1165  * and extent_item_pos is ignored.
1166  *
1167  * FIXME some caching might speed things up
1168  */
1169 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1170                              struct btrfs_fs_info *fs_info, u64 bytenr,
1171                              u64 time_seq, struct ulist *refs,
1172                              struct ulist *roots, const u64 *extent_item_pos,
1173                              struct share_check *sc, bool ignore_offset)
1174 {
1175         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
1176         struct btrfs_key key;
1177         struct btrfs_path *path;
1178         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1179         struct btrfs_delayed_ref_head *head;
1180         int info_level = 0;
1181         int ret;
1182         struct prelim_ref *ref;
1183         struct rb_node *node;
1184         struct extent_inode_elem *eie = NULL;
1185         struct preftrees preftrees = {
1186                 .direct = PREFTREE_INIT,
1187                 .indirect = PREFTREE_INIT,
1188                 .indirect_missing_keys = PREFTREE_INIT
1189         };
1190
1191         key.objectid = bytenr;
1192         key.offset = (u64)-1;
1193         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1194                 key.type = BTRFS_METADATA_ITEM_KEY;
1195         else
1196                 key.type = BTRFS_EXTENT_ITEM_KEY;
1197
1198         path = btrfs_alloc_path();
1199         if (!path)
1200                 return -ENOMEM;
1201         if (!trans) {
1202                 path->search_commit_root = 1;
1203                 path->skip_locking = 1;
1204         }
1205
1206         if (time_seq == BTRFS_SEQ_LAST)
1207                 path->skip_locking = 1;
1208
1209 again:
1210         head = NULL;
1211
1212         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1213         if (ret < 0)
1214                 goto out;
1215         if (ret == 0) {
1216                 /* This shouldn't happen, indicates a bug or fs corruption. */
1217                 ASSERT(ret != 0);
1218                 ret = -EUCLEAN;
1219                 goto out;
1220         }
1221
1222         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1223             time_seq != BTRFS_SEQ_LAST) {
1224                 /*
1225                  * We have a specific time_seq we care about and trans which
1226                  * means we have the path lock, we need to grab the ref head and
1227                  * lock it so we have a consistent view of the refs at the given
1228                  * time.
1229                  */
1230                 delayed_refs = &trans->transaction->delayed_refs;
1231                 spin_lock(&delayed_refs->lock);
1232                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1233                 if (head) {
1234                         if (!mutex_trylock(&head->mutex)) {
1235                                 refcount_inc(&head->refs);
1236                                 spin_unlock(&delayed_refs->lock);
1237
1238                                 btrfs_release_path(path);
1239
1240                                 /*
1241                                  * Mutex was contended, block until it's
1242                                  * released and try again
1243                                  */
1244                                 mutex_lock(&head->mutex);
1245                                 mutex_unlock(&head->mutex);
1246                                 btrfs_put_delayed_ref_head(head);
1247                                 goto again;
1248                         }
1249                         spin_unlock(&delayed_refs->lock);
1250                         ret = add_delayed_refs(fs_info, head, time_seq,
1251                                                &preftrees, sc);
1252                         mutex_unlock(&head->mutex);
1253                         if (ret)
1254                                 goto out;
1255                 } else {
1256                         spin_unlock(&delayed_refs->lock);
1257                 }
1258         }
1259
1260         if (path->slots[0]) {
1261                 struct extent_buffer *leaf;
1262                 int slot;
1263
1264                 path->slots[0]--;
1265                 leaf = path->nodes[0];
1266                 slot = path->slots[0];
1267                 btrfs_item_key_to_cpu(leaf, &key, slot);
1268                 if (key.objectid == bytenr &&
1269                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1270                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1271                         ret = add_inline_refs(fs_info, path, bytenr,
1272                                               &info_level, &preftrees, sc);
1273                         if (ret)
1274                                 goto out;
1275                         ret = add_keyed_refs(root, path, bytenr, info_level,
1276                                              &preftrees, sc);
1277                         if (ret)
1278                                 goto out;
1279                 }
1280         }
1281
1282         btrfs_release_path(path);
1283
1284         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1285         if (ret)
1286                 goto out;
1287
1288         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1289
1290         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1291                                     extent_item_pos, sc, ignore_offset);
1292         if (ret)
1293                 goto out;
1294
1295         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1296
1297         /*
1298          * This walks the tree of merged and resolved refs. Tree blocks are
1299          * read in as needed. Unique entries are added to the ulist, and
1300          * the list of found roots is updated.
1301          *
1302          * We release the entire tree in one go before returning.
1303          */
1304         node = rb_first_cached(&preftrees.direct.root);
1305         while (node) {
1306                 ref = rb_entry(node, struct prelim_ref, rbnode);
1307                 node = rb_next(&ref->rbnode);
1308                 /*
1309                  * ref->count < 0 can happen here if there are delayed
1310                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1311                  * prelim_ref_insert() relies on this when merging
1312                  * identical refs to keep the overall count correct.
1313                  * prelim_ref_insert() will merge only those refs
1314                  * which compare identically.  Any refs having
1315                  * e.g. different offsets would not be merged,
1316                  * and would retain their original ref->count < 0.
1317                  */
1318                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1319                         if (sc && sc->root_objectid &&
1320                             ref->root_id != sc->root_objectid) {
1321                                 ret = BACKREF_FOUND_SHARED;
1322                                 goto out;
1323                         }
1324
1325                         /* no parent == root of tree */
1326                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1327                         if (ret < 0)
1328                                 goto out;
1329                 }
1330                 if (ref->count && ref->parent) {
1331                         if (extent_item_pos && !ref->inode_list &&
1332                             ref->level == 0) {
1333                                 struct extent_buffer *eb;
1334
1335                                 eb = read_tree_block(fs_info, ref->parent, 0,
1336                                                      0, ref->level, NULL);
1337                                 if (IS_ERR(eb)) {
1338                                         ret = PTR_ERR(eb);
1339                                         goto out;
1340                                 }
1341                                 if (!extent_buffer_uptodate(eb)) {
1342                                         free_extent_buffer(eb);
1343                                         ret = -EIO;
1344                                         goto out;
1345                                 }
1346
1347                                 if (!path->skip_locking)
1348                                         btrfs_tree_read_lock(eb);
1349                                 ret = find_extent_in_eb(eb, bytenr,
1350                                                         *extent_item_pos, &eie, ignore_offset);
1351                                 if (!path->skip_locking)
1352                                         btrfs_tree_read_unlock(eb);
1353                                 free_extent_buffer(eb);
1354                                 if (ret < 0)
1355                                         goto out;
1356                                 ref->inode_list = eie;
1357                         }
1358                         ret = ulist_add_merge_ptr(refs, ref->parent,
1359                                                   ref->inode_list,
1360                                                   (void **)&eie, GFP_NOFS);
1361                         if (ret < 0)
1362                                 goto out;
1363                         if (!ret && extent_item_pos) {
1364                                 /*
1365                                  * We've recorded that parent, so we must extend
1366                                  * its inode list here.
1367                                  *
1368                                  * However if there was corruption we may not
1369                                  * have found an eie, return an error in this
1370                                  * case.
1371                                  */
1372                                 ASSERT(eie);
1373                                 if (!eie) {
1374                                         ret = -EUCLEAN;
1375                                         goto out;
1376                                 }
1377                                 while (eie->next)
1378                                         eie = eie->next;
1379                                 eie->next = ref->inode_list;
1380                         }
1381                         eie = NULL;
1382                 }
1383                 cond_resched();
1384         }
1385
1386 out:
1387         btrfs_free_path(path);
1388
1389         prelim_release(&preftrees.direct);
1390         prelim_release(&preftrees.indirect);
1391         prelim_release(&preftrees.indirect_missing_keys);
1392
1393         if (ret < 0)
1394                 free_inode_elem_list(eie);
1395         return ret;
1396 }
1397
1398 static void free_leaf_list(struct ulist *blocks)
1399 {
1400         struct ulist_node *node = NULL;
1401         struct extent_inode_elem *eie;
1402         struct ulist_iterator uiter;
1403
1404         ULIST_ITER_INIT(&uiter);
1405         while ((node = ulist_next(blocks, &uiter))) {
1406                 if (!node->aux)
1407                         continue;
1408                 eie = unode_aux_to_inode_list(node);
1409                 free_inode_elem_list(eie);
1410                 node->aux = 0;
1411         }
1412
1413         ulist_free(blocks);
1414 }
1415
1416 /*
1417  * Finds all leafs with a reference to the specified combination of bytenr and
1418  * offset. key_list_head will point to a list of corresponding keys (caller must
1419  * free each list element). The leafs will be stored in the leafs ulist, which
1420  * must be freed with ulist_free.
1421  *
1422  * returns 0 on success, <0 on error
1423  */
1424 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1425                          struct btrfs_fs_info *fs_info, u64 bytenr,
1426                          u64 time_seq, struct ulist **leafs,
1427                          const u64 *extent_item_pos, bool ignore_offset)
1428 {
1429         int ret;
1430
1431         *leafs = ulist_alloc(GFP_NOFS);
1432         if (!*leafs)
1433                 return -ENOMEM;
1434
1435         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1436                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1437         if (ret < 0 && ret != -ENOENT) {
1438                 free_leaf_list(*leafs);
1439                 return ret;
1440         }
1441
1442         return 0;
1443 }
1444
1445 /*
1446  * walk all backrefs for a given extent to find all roots that reference this
1447  * extent. Walking a backref means finding all extents that reference this
1448  * extent and in turn walk the backrefs of those, too. Naturally this is a
1449  * recursive process, but here it is implemented in an iterative fashion: We
1450  * find all referencing extents for the extent in question and put them on a
1451  * list. In turn, we find all referencing extents for those, further appending
1452  * to the list. The way we iterate the list allows adding more elements after
1453  * the current while iterating. The process stops when we reach the end of the
1454  * list. Found roots are added to the roots list.
1455  *
1456  * returns 0 on success, < 0 on error.
1457  */
1458 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1459                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1460                                      u64 time_seq, struct ulist **roots,
1461                                      bool ignore_offset)
1462 {
1463         struct ulist *tmp;
1464         struct ulist_node *node = NULL;
1465         struct ulist_iterator uiter;
1466         int ret;
1467
1468         tmp = ulist_alloc(GFP_NOFS);
1469         if (!tmp)
1470                 return -ENOMEM;
1471         *roots = ulist_alloc(GFP_NOFS);
1472         if (!*roots) {
1473                 ulist_free(tmp);
1474                 return -ENOMEM;
1475         }
1476
1477         ULIST_ITER_INIT(&uiter);
1478         while (1) {
1479                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1480                                         tmp, *roots, NULL, NULL, ignore_offset);
1481                 if (ret < 0 && ret != -ENOENT) {
1482                         ulist_free(tmp);
1483                         ulist_free(*roots);
1484                         *roots = NULL;
1485                         return ret;
1486                 }
1487                 node = ulist_next(tmp, &uiter);
1488                 if (!node)
1489                         break;
1490                 bytenr = node->val;
1491                 cond_resched();
1492         }
1493
1494         ulist_free(tmp);
1495         return 0;
1496 }
1497
1498 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1499                          struct btrfs_fs_info *fs_info, u64 bytenr,
1500                          u64 time_seq, struct ulist **roots,
1501                          bool skip_commit_root_sem)
1502 {
1503         int ret;
1504
1505         if (!trans && !skip_commit_root_sem)
1506                 down_read(&fs_info->commit_root_sem);
1507         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1508                                         time_seq, roots, false);
1509         if (!trans && !skip_commit_root_sem)
1510                 up_read(&fs_info->commit_root_sem);
1511         return ret;
1512 }
1513
1514 /**
1515  * Check if an extent is shared or not
1516  *
1517  * @root:   root inode belongs to
1518  * @inum:   inode number of the inode whose extent we are checking
1519  * @bytenr: logical bytenr of the extent we are checking
1520  * @roots:  list of roots this extent is shared among
1521  * @tmp:    temporary list used for iteration
1522  *
1523  * btrfs_check_shared uses the backref walking code but will short
1524  * circuit as soon as it finds a root or inode that doesn't match the
1525  * one passed in. This provides a significant performance benefit for
1526  * callers (such as fiemap) which want to know whether the extent is
1527  * shared but do not need a ref count.
1528  *
1529  * This attempts to attach to the running transaction in order to account for
1530  * delayed refs, but continues on even when no running transaction exists.
1531  *
1532  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1533  */
1534 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1535                 struct ulist *roots, struct ulist *tmp)
1536 {
1537         struct btrfs_fs_info *fs_info = root->fs_info;
1538         struct btrfs_trans_handle *trans;
1539         struct ulist_iterator uiter;
1540         struct ulist_node *node;
1541         struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1542         int ret = 0;
1543         struct share_check shared = {
1544                 .root_objectid = root->root_key.objectid,
1545                 .inum = inum,
1546                 .share_count = 0,
1547         };
1548
1549         ulist_init(roots);
1550         ulist_init(tmp);
1551
1552         trans = btrfs_join_transaction_nostart(root);
1553         if (IS_ERR(trans)) {
1554                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1555                         ret = PTR_ERR(trans);
1556                         goto out;
1557                 }
1558                 trans = NULL;
1559                 down_read(&fs_info->commit_root_sem);
1560         } else {
1561                 btrfs_get_tree_mod_seq(fs_info, &elem);
1562         }
1563
1564         ULIST_ITER_INIT(&uiter);
1565         while (1) {
1566                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1567                                         roots, NULL, &shared, false);
1568                 if (ret == BACKREF_FOUND_SHARED) {
1569                         /* this is the only condition under which we return 1 */
1570                         ret = 1;
1571                         break;
1572                 }
1573                 if (ret < 0 && ret != -ENOENT)
1574                         break;
1575                 ret = 0;
1576                 node = ulist_next(tmp, &uiter);
1577                 if (!node)
1578                         break;
1579                 bytenr = node->val;
1580                 shared.share_count = 0;
1581                 cond_resched();
1582         }
1583
1584         if (trans) {
1585                 btrfs_put_tree_mod_seq(fs_info, &elem);
1586                 btrfs_end_transaction(trans);
1587         } else {
1588                 up_read(&fs_info->commit_root_sem);
1589         }
1590 out:
1591         ulist_release(roots);
1592         ulist_release(tmp);
1593         return ret;
1594 }
1595
1596 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1597                           u64 start_off, struct btrfs_path *path,
1598                           struct btrfs_inode_extref **ret_extref,
1599                           u64 *found_off)
1600 {
1601         int ret, slot;
1602         struct btrfs_key key;
1603         struct btrfs_key found_key;
1604         struct btrfs_inode_extref *extref;
1605         const struct extent_buffer *leaf;
1606         unsigned long ptr;
1607
1608         key.objectid = inode_objectid;
1609         key.type = BTRFS_INODE_EXTREF_KEY;
1610         key.offset = start_off;
1611
1612         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1613         if (ret < 0)
1614                 return ret;
1615
1616         while (1) {
1617                 leaf = path->nodes[0];
1618                 slot = path->slots[0];
1619                 if (slot >= btrfs_header_nritems(leaf)) {
1620                         /*
1621                          * If the item at offset is not found,
1622                          * btrfs_search_slot will point us to the slot
1623                          * where it should be inserted. In our case
1624                          * that will be the slot directly before the
1625                          * next INODE_REF_KEY_V2 item. In the case
1626                          * that we're pointing to the last slot in a
1627                          * leaf, we must move one leaf over.
1628                          */
1629                         ret = btrfs_next_leaf(root, path);
1630                         if (ret) {
1631                                 if (ret >= 1)
1632                                         ret = -ENOENT;
1633                                 break;
1634                         }
1635                         continue;
1636                 }
1637
1638                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1639
1640                 /*
1641                  * Check that we're still looking at an extended ref key for
1642                  * this particular objectid. If we have different
1643                  * objectid or type then there are no more to be found
1644                  * in the tree and we can exit.
1645                  */
1646                 ret = -ENOENT;
1647                 if (found_key.objectid != inode_objectid)
1648                         break;
1649                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1650                         break;
1651
1652                 ret = 0;
1653                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1654                 extref = (struct btrfs_inode_extref *)ptr;
1655                 *ret_extref = extref;
1656                 if (found_off)
1657                         *found_off = found_key.offset;
1658                 break;
1659         }
1660
1661         return ret;
1662 }
1663
1664 /*
1665  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1666  * Elements of the path are separated by '/' and the path is guaranteed to be
1667  * 0-terminated. the path is only given within the current file system.
1668  * Therefore, it never starts with a '/'. the caller is responsible to provide
1669  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1670  * the start point of the resulting string is returned. this pointer is within
1671  * dest, normally.
1672  * in case the path buffer would overflow, the pointer is decremented further
1673  * as if output was written to the buffer, though no more output is actually
1674  * generated. that way, the caller can determine how much space would be
1675  * required for the path to fit into the buffer. in that case, the returned
1676  * value will be smaller than dest. callers must check this!
1677  */
1678 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1679                         u32 name_len, unsigned long name_off,
1680                         struct extent_buffer *eb_in, u64 parent,
1681                         char *dest, u32 size)
1682 {
1683         int slot;
1684         u64 next_inum;
1685         int ret;
1686         s64 bytes_left = ((s64)size) - 1;
1687         struct extent_buffer *eb = eb_in;
1688         struct btrfs_key found_key;
1689         struct btrfs_inode_ref *iref;
1690
1691         if (bytes_left >= 0)
1692                 dest[bytes_left] = '\0';
1693
1694         while (1) {
1695                 bytes_left -= name_len;
1696                 if (bytes_left >= 0)
1697                         read_extent_buffer(eb, dest + bytes_left,
1698                                            name_off, name_len);
1699                 if (eb != eb_in) {
1700                         if (!path->skip_locking)
1701                                 btrfs_tree_read_unlock(eb);
1702                         free_extent_buffer(eb);
1703                 }
1704                 ret = btrfs_find_item(fs_root, path, parent, 0,
1705                                 BTRFS_INODE_REF_KEY, &found_key);
1706                 if (ret > 0)
1707                         ret = -ENOENT;
1708                 if (ret)
1709                         break;
1710
1711                 next_inum = found_key.offset;
1712
1713                 /* regular exit ahead */
1714                 if (parent == next_inum)
1715                         break;
1716
1717                 slot = path->slots[0];
1718                 eb = path->nodes[0];
1719                 /* make sure we can use eb after releasing the path */
1720                 if (eb != eb_in) {
1721                         path->nodes[0] = NULL;
1722                         path->locks[0] = 0;
1723                 }
1724                 btrfs_release_path(path);
1725                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1726
1727                 name_len = btrfs_inode_ref_name_len(eb, iref);
1728                 name_off = (unsigned long)(iref + 1);
1729
1730                 parent = next_inum;
1731                 --bytes_left;
1732                 if (bytes_left >= 0)
1733                         dest[bytes_left] = '/';
1734         }
1735
1736         btrfs_release_path(path);
1737
1738         if (ret)
1739                 return ERR_PTR(ret);
1740
1741         return dest + bytes_left;
1742 }
1743
1744 /*
1745  * this makes the path point to (logical EXTENT_ITEM *)
1746  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1747  * tree blocks and <0 on error.
1748  */
1749 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1750                         struct btrfs_path *path, struct btrfs_key *found_key,
1751                         u64 *flags_ret)
1752 {
1753         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
1754         int ret;
1755         u64 flags;
1756         u64 size = 0;
1757         u32 item_size;
1758         const struct extent_buffer *eb;
1759         struct btrfs_extent_item *ei;
1760         struct btrfs_key key;
1761
1762         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1763                 key.type = BTRFS_METADATA_ITEM_KEY;
1764         else
1765                 key.type = BTRFS_EXTENT_ITEM_KEY;
1766         key.objectid = logical;
1767         key.offset = (u64)-1;
1768
1769         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1770         if (ret < 0)
1771                 return ret;
1772
1773         ret = btrfs_previous_extent_item(extent_root, path, 0);
1774         if (ret) {
1775                 if (ret > 0)
1776                         ret = -ENOENT;
1777                 return ret;
1778         }
1779         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1780         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1781                 size = fs_info->nodesize;
1782         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1783                 size = found_key->offset;
1784
1785         if (found_key->objectid > logical ||
1786             found_key->objectid + size <= logical) {
1787                 btrfs_debug(fs_info,
1788                         "logical %llu is not within any extent", logical);
1789                 return -ENOENT;
1790         }
1791
1792         eb = path->nodes[0];
1793         item_size = btrfs_item_size(eb, path->slots[0]);
1794         BUG_ON(item_size < sizeof(*ei));
1795
1796         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1797         flags = btrfs_extent_flags(eb, ei);
1798
1799         btrfs_debug(fs_info,
1800                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1801                  logical, logical - found_key->objectid, found_key->objectid,
1802                  found_key->offset, flags, item_size);
1803
1804         WARN_ON(!flags_ret);
1805         if (flags_ret) {
1806                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1807                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1808                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1809                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1810                 else
1811                         BUG();
1812                 return 0;
1813         }
1814
1815         return -EIO;
1816 }
1817
1818 /*
1819  * helper function to iterate extent inline refs. ptr must point to a 0 value
1820  * for the first call and may be modified. it is used to track state.
1821  * if more refs exist, 0 is returned and the next call to
1822  * get_extent_inline_ref must pass the modified ptr parameter to get the
1823  * next ref. after the last ref was processed, 1 is returned.
1824  * returns <0 on error
1825  */
1826 static int get_extent_inline_ref(unsigned long *ptr,
1827                                  const struct extent_buffer *eb,
1828                                  const struct btrfs_key *key,
1829                                  const struct btrfs_extent_item *ei,
1830                                  u32 item_size,
1831                                  struct btrfs_extent_inline_ref **out_eiref,
1832                                  int *out_type)
1833 {
1834         unsigned long end;
1835         u64 flags;
1836         struct btrfs_tree_block_info *info;
1837
1838         if (!*ptr) {
1839                 /* first call */
1840                 flags = btrfs_extent_flags(eb, ei);
1841                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1842                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1843                                 /* a skinny metadata extent */
1844                                 *out_eiref =
1845                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1846                         } else {
1847                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1848                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1849                                 *out_eiref =
1850                                    (struct btrfs_extent_inline_ref *)(info + 1);
1851                         }
1852                 } else {
1853                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1854                 }
1855                 *ptr = (unsigned long)*out_eiref;
1856                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1857                         return -ENOENT;
1858         }
1859
1860         end = (unsigned long)ei + item_size;
1861         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1862         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1863                                                      BTRFS_REF_TYPE_ANY);
1864         if (*out_type == BTRFS_REF_TYPE_INVALID)
1865                 return -EUCLEAN;
1866
1867         *ptr += btrfs_extent_inline_ref_size(*out_type);
1868         WARN_ON(*ptr > end);
1869         if (*ptr == end)
1870                 return 1; /* last */
1871
1872         return 0;
1873 }
1874
1875 /*
1876  * reads the tree block backref for an extent. tree level and root are returned
1877  * through out_level and out_root. ptr must point to a 0 value for the first
1878  * call and may be modified (see get_extent_inline_ref comment).
1879  * returns 0 if data was provided, 1 if there was no more data to provide or
1880  * <0 on error.
1881  */
1882 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1883                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1884                             u32 item_size, u64 *out_root, u8 *out_level)
1885 {
1886         int ret;
1887         int type;
1888         struct btrfs_extent_inline_ref *eiref;
1889
1890         if (*ptr == (unsigned long)-1)
1891                 return 1;
1892
1893         while (1) {
1894                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1895                                               &eiref, &type);
1896                 if (ret < 0)
1897                         return ret;
1898
1899                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1900                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1901                         break;
1902
1903                 if (ret == 1)
1904                         return 1;
1905         }
1906
1907         /* we can treat both ref types equally here */
1908         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1909
1910         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1911                 struct btrfs_tree_block_info *info;
1912
1913                 info = (struct btrfs_tree_block_info *)(ei + 1);
1914                 *out_level = btrfs_tree_block_level(eb, info);
1915         } else {
1916                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1917                 *out_level = (u8)key->offset;
1918         }
1919
1920         if (ret == 1)
1921                 *ptr = (unsigned long)-1;
1922
1923         return 0;
1924 }
1925
1926 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1927                              struct extent_inode_elem *inode_list,
1928                              u64 root, u64 extent_item_objectid,
1929                              iterate_extent_inodes_t *iterate, void *ctx)
1930 {
1931         struct extent_inode_elem *eie;
1932         int ret = 0;
1933
1934         for (eie = inode_list; eie; eie = eie->next) {
1935                 btrfs_debug(fs_info,
1936                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1937                             extent_item_objectid, eie->inum,
1938                             eie->offset, root);
1939                 ret = iterate(eie->inum, eie->offset, root, ctx);
1940                 if (ret) {
1941                         btrfs_debug(fs_info,
1942                                     "stopping iteration for %llu due to ret=%d",
1943                                     extent_item_objectid, ret);
1944                         break;
1945                 }
1946         }
1947
1948         return ret;
1949 }
1950
1951 /*
1952  * calls iterate() for every inode that references the extent identified by
1953  * the given parameters.
1954  * when the iterator function returns a non-zero value, iteration stops.
1955  */
1956 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1957                                 u64 extent_item_objectid, u64 extent_item_pos,
1958                                 int search_commit_root,
1959                                 iterate_extent_inodes_t *iterate, void *ctx,
1960                                 bool ignore_offset)
1961 {
1962         int ret;
1963         struct btrfs_trans_handle *trans = NULL;
1964         struct ulist *refs = NULL;
1965         struct ulist *roots = NULL;
1966         struct ulist_node *ref_node = NULL;
1967         struct ulist_node *root_node = NULL;
1968         struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
1969         struct ulist_iterator ref_uiter;
1970         struct ulist_iterator root_uiter;
1971
1972         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1973                         extent_item_objectid);
1974
1975         if (!search_commit_root) {
1976                 trans = btrfs_attach_transaction(fs_info->tree_root);
1977                 if (IS_ERR(trans)) {
1978                         if (PTR_ERR(trans) != -ENOENT &&
1979                             PTR_ERR(trans) != -EROFS)
1980                                 return PTR_ERR(trans);
1981                         trans = NULL;
1982                 }
1983         }
1984
1985         if (trans)
1986                 btrfs_get_tree_mod_seq(fs_info, &seq_elem);
1987         else
1988                 down_read(&fs_info->commit_root_sem);
1989
1990         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1991                                    seq_elem.seq, &refs,
1992                                    &extent_item_pos, ignore_offset);
1993         if (ret)
1994                 goto out;
1995
1996         ULIST_ITER_INIT(&ref_uiter);
1997         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1998                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1999                                                 seq_elem.seq, &roots,
2000                                                 ignore_offset);
2001                 if (ret)
2002                         break;
2003                 ULIST_ITER_INIT(&root_uiter);
2004                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2005                         btrfs_debug(fs_info,
2006                                     "root %llu references leaf %llu, data list %#llx",
2007                                     root_node->val, ref_node->val,
2008                                     ref_node->aux);
2009                         ret = iterate_leaf_refs(fs_info,
2010                                                 (struct extent_inode_elem *)
2011                                                 (uintptr_t)ref_node->aux,
2012                                                 root_node->val,
2013                                                 extent_item_objectid,
2014                                                 iterate, ctx);
2015                 }
2016                 ulist_free(roots);
2017         }
2018
2019         free_leaf_list(refs);
2020 out:
2021         if (trans) {
2022                 btrfs_put_tree_mod_seq(fs_info, &seq_elem);
2023                 btrfs_end_transaction(trans);
2024         } else {
2025                 up_read(&fs_info->commit_root_sem);
2026         }
2027
2028         return ret;
2029 }
2030
2031 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2032                                 struct btrfs_path *path,
2033                                 iterate_extent_inodes_t *iterate, void *ctx,
2034                                 bool ignore_offset)
2035 {
2036         int ret;
2037         u64 extent_item_pos;
2038         u64 flags = 0;
2039         struct btrfs_key found_key;
2040         int search_commit_root = path->search_commit_root;
2041
2042         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2043         btrfs_release_path(path);
2044         if (ret < 0)
2045                 return ret;
2046         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2047                 return -EINVAL;
2048
2049         extent_item_pos = logical - found_key.objectid;
2050         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2051                                         extent_item_pos, search_commit_root,
2052                                         iterate, ctx, ignore_offset);
2053
2054         return ret;
2055 }
2056
2057 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2058                               struct extent_buffer *eb, void *ctx);
2059
2060 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2061                               struct btrfs_path *path,
2062                               iterate_irefs_t *iterate, void *ctx)
2063 {
2064         int ret = 0;
2065         int slot;
2066         u32 cur;
2067         u32 len;
2068         u32 name_len;
2069         u64 parent = 0;
2070         int found = 0;
2071         struct extent_buffer *eb;
2072         struct btrfs_inode_ref *iref;
2073         struct btrfs_key found_key;
2074
2075         while (!ret) {
2076                 ret = btrfs_find_item(fs_root, path, inum,
2077                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2078                                 &found_key);
2079
2080                 if (ret < 0)
2081                         break;
2082                 if (ret) {
2083                         ret = found ? 0 : -ENOENT;
2084                         break;
2085                 }
2086                 ++found;
2087
2088                 parent = found_key.offset;
2089                 slot = path->slots[0];
2090                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2091                 if (!eb) {
2092                         ret = -ENOMEM;
2093                         break;
2094                 }
2095                 btrfs_release_path(path);
2096
2097                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2098
2099                 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2100                         name_len = btrfs_inode_ref_name_len(eb, iref);
2101                         /* path must be released before calling iterate()! */
2102                         btrfs_debug(fs_root->fs_info,
2103                                 "following ref at offset %u for inode %llu in tree %llu",
2104                                 cur, found_key.objectid,
2105                                 fs_root->root_key.objectid);
2106                         ret = iterate(parent, name_len,
2107                                       (unsigned long)(iref + 1), eb, ctx);
2108                         if (ret)
2109                                 break;
2110                         len = sizeof(*iref) + name_len;
2111                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2112                 }
2113                 free_extent_buffer(eb);
2114         }
2115
2116         btrfs_release_path(path);
2117
2118         return ret;
2119 }
2120
2121 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2122                                  struct btrfs_path *path,
2123                                  iterate_irefs_t *iterate, void *ctx)
2124 {
2125         int ret;
2126         int slot;
2127         u64 offset = 0;
2128         u64 parent;
2129         int found = 0;
2130         struct extent_buffer *eb;
2131         struct btrfs_inode_extref *extref;
2132         u32 item_size;
2133         u32 cur_offset;
2134         unsigned long ptr;
2135
2136         while (1) {
2137                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2138                                             &offset);
2139                 if (ret < 0)
2140                         break;
2141                 if (ret) {
2142                         ret = found ? 0 : -ENOENT;
2143                         break;
2144                 }
2145                 ++found;
2146
2147                 slot = path->slots[0];
2148                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2149                 if (!eb) {
2150                         ret = -ENOMEM;
2151                         break;
2152                 }
2153                 btrfs_release_path(path);
2154
2155                 item_size = btrfs_item_size(eb, slot);
2156                 ptr = btrfs_item_ptr_offset(eb, slot);
2157                 cur_offset = 0;
2158
2159                 while (cur_offset < item_size) {
2160                         u32 name_len;
2161
2162                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2163                         parent = btrfs_inode_extref_parent(eb, extref);
2164                         name_len = btrfs_inode_extref_name_len(eb, extref);
2165                         ret = iterate(parent, name_len,
2166                                       (unsigned long)&extref->name, eb, ctx);
2167                         if (ret)
2168                                 break;
2169
2170                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2171                         cur_offset += sizeof(*extref);
2172                 }
2173                 free_extent_buffer(eb);
2174
2175                 offset++;
2176         }
2177
2178         btrfs_release_path(path);
2179
2180         return ret;
2181 }
2182
2183 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2184                          struct btrfs_path *path, iterate_irefs_t *iterate,
2185                          void *ctx)
2186 {
2187         int ret;
2188         int found_refs = 0;
2189
2190         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2191         if (!ret)
2192                 ++found_refs;
2193         else if (ret != -ENOENT)
2194                 return ret;
2195
2196         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2197         if (ret == -ENOENT && found_refs)
2198                 return 0;
2199
2200         return ret;
2201 }
2202
2203 /*
2204  * returns 0 if the path could be dumped (probably truncated)
2205  * returns <0 in case of an error
2206  */
2207 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2208                          struct extent_buffer *eb, void *ctx)
2209 {
2210         struct inode_fs_paths *ipath = ctx;
2211         char *fspath;
2212         char *fspath_min;
2213         int i = ipath->fspath->elem_cnt;
2214         const int s_ptr = sizeof(char *);
2215         u32 bytes_left;
2216
2217         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2218                                         ipath->fspath->bytes_left - s_ptr : 0;
2219
2220         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2221         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2222                                    name_off, eb, inum, fspath_min, bytes_left);
2223         if (IS_ERR(fspath))
2224                 return PTR_ERR(fspath);
2225
2226         if (fspath > fspath_min) {
2227                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2228                 ++ipath->fspath->elem_cnt;
2229                 ipath->fspath->bytes_left = fspath - fspath_min;
2230         } else {
2231                 ++ipath->fspath->elem_missed;
2232                 ipath->fspath->bytes_missing += fspath_min - fspath;
2233                 ipath->fspath->bytes_left = 0;
2234         }
2235
2236         return 0;
2237 }
2238
2239 /*
2240  * this dumps all file system paths to the inode into the ipath struct, provided
2241  * is has been created large enough. each path is zero-terminated and accessed
2242  * from ipath->fspath->val[i].
2243  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2244  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2245  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2246  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2247  * have been needed to return all paths.
2248  */
2249 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2250 {
2251         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2252                              inode_to_path, ipath);
2253 }
2254
2255 struct btrfs_data_container *init_data_container(u32 total_bytes)
2256 {
2257         struct btrfs_data_container *data;
2258         size_t alloc_bytes;
2259
2260         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2261         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2262         if (!data)
2263                 return ERR_PTR(-ENOMEM);
2264
2265         if (total_bytes >= sizeof(*data)) {
2266                 data->bytes_left = total_bytes - sizeof(*data);
2267                 data->bytes_missing = 0;
2268         } else {
2269                 data->bytes_missing = sizeof(*data) - total_bytes;
2270                 data->bytes_left = 0;
2271         }
2272
2273         data->elem_cnt = 0;
2274         data->elem_missed = 0;
2275
2276         return data;
2277 }
2278
2279 /*
2280  * allocates space to return multiple file system paths for an inode.
2281  * total_bytes to allocate are passed, note that space usable for actual path
2282  * information will be total_bytes - sizeof(struct inode_fs_paths).
2283  * the returned pointer must be freed with free_ipath() in the end.
2284  */
2285 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2286                                         struct btrfs_path *path)
2287 {
2288         struct inode_fs_paths *ifp;
2289         struct btrfs_data_container *fspath;
2290
2291         fspath = init_data_container(total_bytes);
2292         if (IS_ERR(fspath))
2293                 return ERR_CAST(fspath);
2294
2295         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2296         if (!ifp) {
2297                 kvfree(fspath);
2298                 return ERR_PTR(-ENOMEM);
2299         }
2300
2301         ifp->btrfs_path = path;
2302         ifp->fspath = fspath;
2303         ifp->fs_root = fs_root;
2304
2305         return ifp;
2306 }
2307
2308 void free_ipath(struct inode_fs_paths *ipath)
2309 {
2310         if (!ipath)
2311                 return;
2312         kvfree(ipath->fspath);
2313         kfree(ipath);
2314 }
2315
2316 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2317                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2318 {
2319         struct btrfs_backref_iter *ret;
2320
2321         ret = kzalloc(sizeof(*ret), gfp_flag);
2322         if (!ret)
2323                 return NULL;
2324
2325         ret->path = btrfs_alloc_path();
2326         if (!ret->path) {
2327                 kfree(ret);
2328                 return NULL;
2329         }
2330
2331         /* Current backref iterator only supports iteration in commit root */
2332         ret->path->search_commit_root = 1;
2333         ret->path->skip_locking = 1;
2334         ret->fs_info = fs_info;
2335
2336         return ret;
2337 }
2338
2339 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2340 {
2341         struct btrfs_fs_info *fs_info = iter->fs_info;
2342         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2343         struct btrfs_path *path = iter->path;
2344         struct btrfs_extent_item *ei;
2345         struct btrfs_key key;
2346         int ret;
2347
2348         key.objectid = bytenr;
2349         key.type = BTRFS_METADATA_ITEM_KEY;
2350         key.offset = (u64)-1;
2351         iter->bytenr = bytenr;
2352
2353         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2354         if (ret < 0)
2355                 return ret;
2356         if (ret == 0) {
2357                 ret = -EUCLEAN;
2358                 goto release;
2359         }
2360         if (path->slots[0] == 0) {
2361                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2362                 ret = -EUCLEAN;
2363                 goto release;
2364         }
2365         path->slots[0]--;
2366
2367         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2368         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2369              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2370                 ret = -ENOENT;
2371                 goto release;
2372         }
2373         memcpy(&iter->cur_key, &key, sizeof(key));
2374         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2375                                                     path->slots[0]);
2376         iter->end_ptr = (u32)(iter->item_ptr +
2377                         btrfs_item_size(path->nodes[0], path->slots[0]));
2378         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2379                             struct btrfs_extent_item);
2380
2381         /*
2382          * Only support iteration on tree backref yet.
2383          *
2384          * This is an extra precaution for non skinny-metadata, where
2385          * EXTENT_ITEM is also used for tree blocks, that we can only use
2386          * extent flags to determine if it's a tree block.
2387          */
2388         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2389                 ret = -ENOTSUPP;
2390                 goto release;
2391         }
2392         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2393
2394         /* If there is no inline backref, go search for keyed backref */
2395         if (iter->cur_ptr >= iter->end_ptr) {
2396                 ret = btrfs_next_item(extent_root, path);
2397
2398                 /* No inline nor keyed ref */
2399                 if (ret > 0) {
2400                         ret = -ENOENT;
2401                         goto release;
2402                 }
2403                 if (ret < 0)
2404                         goto release;
2405
2406                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2407                                 path->slots[0]);
2408                 if (iter->cur_key.objectid != bytenr ||
2409                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2410                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2411                         ret = -ENOENT;
2412                         goto release;
2413                 }
2414                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2415                                                            path->slots[0]);
2416                 iter->item_ptr = iter->cur_ptr;
2417                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2418                                       path->nodes[0], path->slots[0]));
2419         }
2420
2421         return 0;
2422 release:
2423         btrfs_backref_iter_release(iter);
2424         return ret;
2425 }
2426
2427 /*
2428  * Go to the next backref item of current bytenr, can be either inlined or
2429  * keyed.
2430  *
2431  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2432  *
2433  * Return 0 if we get next backref without problem.
2434  * Return >0 if there is no extra backref for this bytenr.
2435  * Return <0 if there is something wrong happened.
2436  */
2437 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2438 {
2439         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2440         struct btrfs_root *extent_root;
2441         struct btrfs_path *path = iter->path;
2442         struct btrfs_extent_inline_ref *iref;
2443         int ret;
2444         u32 size;
2445
2446         if (btrfs_backref_iter_is_inline_ref(iter)) {
2447                 /* We're still inside the inline refs */
2448                 ASSERT(iter->cur_ptr < iter->end_ptr);
2449
2450                 if (btrfs_backref_has_tree_block_info(iter)) {
2451                         /* First tree block info */
2452                         size = sizeof(struct btrfs_tree_block_info);
2453                 } else {
2454                         /* Use inline ref type to determine the size */
2455                         int type;
2456
2457                         iref = (struct btrfs_extent_inline_ref *)
2458                                 ((unsigned long)iter->cur_ptr);
2459                         type = btrfs_extent_inline_ref_type(eb, iref);
2460
2461                         size = btrfs_extent_inline_ref_size(type);
2462                 }
2463                 iter->cur_ptr += size;
2464                 if (iter->cur_ptr < iter->end_ptr)
2465                         return 0;
2466
2467                 /* All inline items iterated, fall through */
2468         }
2469
2470         /* We're at keyed items, there is no inline item, go to the next one */
2471         extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2472         ret = btrfs_next_item(extent_root, iter->path);
2473         if (ret)
2474                 return ret;
2475
2476         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2477         if (iter->cur_key.objectid != iter->bytenr ||
2478             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2479              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2480                 return 1;
2481         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2482                                         path->slots[0]);
2483         iter->cur_ptr = iter->item_ptr;
2484         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
2485                                                 path->slots[0]);
2486         return 0;
2487 }
2488
2489 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2490                               struct btrfs_backref_cache *cache, int is_reloc)
2491 {
2492         int i;
2493
2494         cache->rb_root = RB_ROOT;
2495         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2496                 INIT_LIST_HEAD(&cache->pending[i]);
2497         INIT_LIST_HEAD(&cache->changed);
2498         INIT_LIST_HEAD(&cache->detached);
2499         INIT_LIST_HEAD(&cache->leaves);
2500         INIT_LIST_HEAD(&cache->pending_edge);
2501         INIT_LIST_HEAD(&cache->useless_node);
2502         cache->fs_info = fs_info;
2503         cache->is_reloc = is_reloc;
2504 }
2505
2506 struct btrfs_backref_node *btrfs_backref_alloc_node(
2507                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2508 {
2509         struct btrfs_backref_node *node;
2510
2511         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2512         node = kzalloc(sizeof(*node), GFP_NOFS);
2513         if (!node)
2514                 return node;
2515
2516         INIT_LIST_HEAD(&node->list);
2517         INIT_LIST_HEAD(&node->upper);
2518         INIT_LIST_HEAD(&node->lower);
2519         RB_CLEAR_NODE(&node->rb_node);
2520         cache->nr_nodes++;
2521         node->level = level;
2522         node->bytenr = bytenr;
2523
2524         return node;
2525 }
2526
2527 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2528                 struct btrfs_backref_cache *cache)
2529 {
2530         struct btrfs_backref_edge *edge;
2531
2532         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2533         if (edge)
2534                 cache->nr_edges++;
2535         return edge;
2536 }
2537
2538 /*
2539  * Drop the backref node from cache, also cleaning up all its
2540  * upper edges and any uncached nodes in the path.
2541  *
2542  * This cleanup happens bottom up, thus the node should either
2543  * be the lowest node in the cache or a detached node.
2544  */
2545 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2546                                 struct btrfs_backref_node *node)
2547 {
2548         struct btrfs_backref_node *upper;
2549         struct btrfs_backref_edge *edge;
2550
2551         if (!node)
2552                 return;
2553
2554         BUG_ON(!node->lowest && !node->detached);
2555         while (!list_empty(&node->upper)) {
2556                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2557                                   list[LOWER]);
2558                 upper = edge->node[UPPER];
2559                 list_del(&edge->list[LOWER]);
2560                 list_del(&edge->list[UPPER]);
2561                 btrfs_backref_free_edge(cache, edge);
2562
2563                 /*
2564                  * Add the node to leaf node list if no other child block
2565                  * cached.
2566                  */
2567                 if (list_empty(&upper->lower)) {
2568                         list_add_tail(&upper->lower, &cache->leaves);
2569                         upper->lowest = 1;
2570                 }
2571         }
2572
2573         btrfs_backref_drop_node(cache, node);
2574 }
2575
2576 /*
2577  * Release all nodes/edges from current cache
2578  */
2579 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2580 {
2581         struct btrfs_backref_node *node;
2582         int i;
2583
2584         while (!list_empty(&cache->detached)) {
2585                 node = list_entry(cache->detached.next,
2586                                   struct btrfs_backref_node, list);
2587                 btrfs_backref_cleanup_node(cache, node);
2588         }
2589
2590         while (!list_empty(&cache->leaves)) {
2591                 node = list_entry(cache->leaves.next,
2592                                   struct btrfs_backref_node, lower);
2593                 btrfs_backref_cleanup_node(cache, node);
2594         }
2595
2596         cache->last_trans = 0;
2597
2598         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2599                 ASSERT(list_empty(&cache->pending[i]));
2600         ASSERT(list_empty(&cache->pending_edge));
2601         ASSERT(list_empty(&cache->useless_node));
2602         ASSERT(list_empty(&cache->changed));
2603         ASSERT(list_empty(&cache->detached));
2604         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2605         ASSERT(!cache->nr_nodes);
2606         ASSERT(!cache->nr_edges);
2607 }
2608
2609 /*
2610  * Handle direct tree backref
2611  *
2612  * Direct tree backref means, the backref item shows its parent bytenr
2613  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2614  *
2615  * @ref_key:    The converted backref key.
2616  *              For keyed backref, it's the item key.
2617  *              For inlined backref, objectid is the bytenr,
2618  *              type is btrfs_inline_ref_type, offset is
2619  *              btrfs_inline_ref_offset.
2620  */
2621 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2622                                       struct btrfs_key *ref_key,
2623                                       struct btrfs_backref_node *cur)
2624 {
2625         struct btrfs_backref_edge *edge;
2626         struct btrfs_backref_node *upper;
2627         struct rb_node *rb_node;
2628
2629         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2630
2631         /* Only reloc root uses backref pointing to itself */
2632         if (ref_key->objectid == ref_key->offset) {
2633                 struct btrfs_root *root;
2634
2635                 cur->is_reloc_root = 1;
2636                 /* Only reloc backref cache cares about a specific root */
2637                 if (cache->is_reloc) {
2638                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2639                         if (!root)
2640                                 return -ENOENT;
2641                         cur->root = root;
2642                 } else {
2643                         /*
2644                          * For generic purpose backref cache, reloc root node
2645                          * is useless.
2646                          */
2647                         list_add(&cur->list, &cache->useless_node);
2648                 }
2649                 return 0;
2650         }
2651
2652         edge = btrfs_backref_alloc_edge(cache);
2653         if (!edge)
2654                 return -ENOMEM;
2655
2656         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2657         if (!rb_node) {
2658                 /* Parent node not yet cached */
2659                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2660                                            cur->level + 1);
2661                 if (!upper) {
2662                         btrfs_backref_free_edge(cache, edge);
2663                         return -ENOMEM;
2664                 }
2665
2666                 /*
2667                  *  Backrefs for the upper level block isn't cached, add the
2668                  *  block to pending list
2669                  */
2670                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2671         } else {
2672                 /* Parent node already cached */
2673                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2674                 ASSERT(upper->checked);
2675                 INIT_LIST_HEAD(&edge->list[UPPER]);
2676         }
2677         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2678         return 0;
2679 }
2680
2681 /*
2682  * Handle indirect tree backref
2683  *
2684  * Indirect tree backref means, we only know which tree the node belongs to.
2685  * We still need to do a tree search to find out the parents. This is for
2686  * TREE_BLOCK_REF backref (keyed or inlined).
2687  *
2688  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2689  * @tree_key:   The first key of this tree block.
2690  * @path:       A clean (released) path, to avoid allocating path every time
2691  *              the function get called.
2692  */
2693 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2694                                         struct btrfs_path *path,
2695                                         struct btrfs_key *ref_key,
2696                                         struct btrfs_key *tree_key,
2697                                         struct btrfs_backref_node *cur)
2698 {
2699         struct btrfs_fs_info *fs_info = cache->fs_info;
2700         struct btrfs_backref_node *upper;
2701         struct btrfs_backref_node *lower;
2702         struct btrfs_backref_edge *edge;
2703         struct extent_buffer *eb;
2704         struct btrfs_root *root;
2705         struct rb_node *rb_node;
2706         int level;
2707         bool need_check = true;
2708         int ret;
2709
2710         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2711         if (IS_ERR(root))
2712                 return PTR_ERR(root);
2713         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2714                 cur->cowonly = 1;
2715
2716         if (btrfs_root_level(&root->root_item) == cur->level) {
2717                 /* Tree root */
2718                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2719                 /*
2720                  * For reloc backref cache, we may ignore reloc root.  But for
2721                  * general purpose backref cache, we can't rely on
2722                  * btrfs_should_ignore_reloc_root() as it may conflict with
2723                  * current running relocation and lead to missing root.
2724                  *
2725                  * For general purpose backref cache, reloc root detection is
2726                  * completely relying on direct backref (key->offset is parent
2727                  * bytenr), thus only do such check for reloc cache.
2728                  */
2729                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2730                         btrfs_put_root(root);
2731                         list_add(&cur->list, &cache->useless_node);
2732                 } else {
2733                         cur->root = root;
2734                 }
2735                 return 0;
2736         }
2737
2738         level = cur->level + 1;
2739
2740         /* Search the tree to find parent blocks referring to the block */
2741         path->search_commit_root = 1;
2742         path->skip_locking = 1;
2743         path->lowest_level = level;
2744         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2745         path->lowest_level = 0;
2746         if (ret < 0) {
2747                 btrfs_put_root(root);
2748                 return ret;
2749         }
2750         if (ret > 0 && path->slots[level] > 0)
2751                 path->slots[level]--;
2752
2753         eb = path->nodes[level];
2754         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2755                 btrfs_err(fs_info,
2756 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2757                           cur->bytenr, level - 1, root->root_key.objectid,
2758                           tree_key->objectid, tree_key->type, tree_key->offset);
2759                 btrfs_put_root(root);
2760                 ret = -ENOENT;
2761                 goto out;
2762         }
2763         lower = cur;
2764
2765         /* Add all nodes and edges in the path */
2766         for (; level < BTRFS_MAX_LEVEL; level++) {
2767                 if (!path->nodes[level]) {
2768                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2769                                lower->bytenr);
2770                         /* Same as previous should_ignore_reloc_root() call */
2771                         if (btrfs_should_ignore_reloc_root(root) &&
2772                             cache->is_reloc) {
2773                                 btrfs_put_root(root);
2774                                 list_add(&lower->list, &cache->useless_node);
2775                         } else {
2776                                 lower->root = root;
2777                         }
2778                         break;
2779                 }
2780
2781                 edge = btrfs_backref_alloc_edge(cache);
2782                 if (!edge) {
2783                         btrfs_put_root(root);
2784                         ret = -ENOMEM;
2785                         goto out;
2786                 }
2787
2788                 eb = path->nodes[level];
2789                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2790                 if (!rb_node) {
2791                         upper = btrfs_backref_alloc_node(cache, eb->start,
2792                                                          lower->level + 1);
2793                         if (!upper) {
2794                                 btrfs_put_root(root);
2795                                 btrfs_backref_free_edge(cache, edge);
2796                                 ret = -ENOMEM;
2797                                 goto out;
2798                         }
2799                         upper->owner = btrfs_header_owner(eb);
2800                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2801                                 upper->cowonly = 1;
2802
2803                         /*
2804                          * If we know the block isn't shared we can avoid
2805                          * checking its backrefs.
2806                          */
2807                         if (btrfs_block_can_be_shared(root, eb))
2808                                 upper->checked = 0;
2809                         else
2810                                 upper->checked = 1;
2811
2812                         /*
2813                          * Add the block to pending list if we need to check its
2814                          * backrefs, we only do this once while walking up a
2815                          * tree as we will catch anything else later on.
2816                          */
2817                         if (!upper->checked && need_check) {
2818                                 need_check = false;
2819                                 list_add_tail(&edge->list[UPPER],
2820                                               &cache->pending_edge);
2821                         } else {
2822                                 if (upper->checked)
2823                                         need_check = true;
2824                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2825                         }
2826                 } else {
2827                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2828                                          rb_node);
2829                         ASSERT(upper->checked);
2830                         INIT_LIST_HEAD(&edge->list[UPPER]);
2831                         if (!upper->owner)
2832                                 upper->owner = btrfs_header_owner(eb);
2833                 }
2834                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2835
2836                 if (rb_node) {
2837                         btrfs_put_root(root);
2838                         break;
2839                 }
2840                 lower = upper;
2841                 upper = NULL;
2842         }
2843 out:
2844         btrfs_release_path(path);
2845         return ret;
2846 }
2847
2848 /*
2849  * Add backref node @cur into @cache.
2850  *
2851  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2852  *       links aren't yet bi-directional. Needs to finish such links.
2853  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2854  *
2855  * @path:       Released path for indirect tree backref lookup
2856  * @iter:       Released backref iter for extent tree search
2857  * @node_key:   The first key of the tree block
2858  */
2859 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2860                                 struct btrfs_path *path,
2861                                 struct btrfs_backref_iter *iter,
2862                                 struct btrfs_key *node_key,
2863                                 struct btrfs_backref_node *cur)
2864 {
2865         struct btrfs_fs_info *fs_info = cache->fs_info;
2866         struct btrfs_backref_edge *edge;
2867         struct btrfs_backref_node *exist;
2868         int ret;
2869
2870         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2871         if (ret < 0)
2872                 return ret;
2873         /*
2874          * We skip the first btrfs_tree_block_info, as we don't use the key
2875          * stored in it, but fetch it from the tree block
2876          */
2877         if (btrfs_backref_has_tree_block_info(iter)) {
2878                 ret = btrfs_backref_iter_next(iter);
2879                 if (ret < 0)
2880                         goto out;
2881                 /* No extra backref? This means the tree block is corrupted */
2882                 if (ret > 0) {
2883                         ret = -EUCLEAN;
2884                         goto out;
2885                 }
2886         }
2887         WARN_ON(cur->checked);
2888         if (!list_empty(&cur->upper)) {
2889                 /*
2890                  * The backref was added previously when processing backref of
2891                  * type BTRFS_TREE_BLOCK_REF_KEY
2892                  */
2893                 ASSERT(list_is_singular(&cur->upper));
2894                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2895                                   list[LOWER]);
2896                 ASSERT(list_empty(&edge->list[UPPER]));
2897                 exist = edge->node[UPPER];
2898                 /*
2899                  * Add the upper level block to pending list if we need check
2900                  * its backrefs
2901                  */
2902                 if (!exist->checked)
2903                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2904         } else {
2905                 exist = NULL;
2906         }
2907
2908         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2909                 struct extent_buffer *eb;
2910                 struct btrfs_key key;
2911                 int type;
2912
2913                 cond_resched();
2914                 eb = btrfs_backref_get_eb(iter);
2915
2916                 key.objectid = iter->bytenr;
2917                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2918                         struct btrfs_extent_inline_ref *iref;
2919
2920                         /* Update key for inline backref */
2921                         iref = (struct btrfs_extent_inline_ref *)
2922                                 ((unsigned long)iter->cur_ptr);
2923                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2924                                                         BTRFS_REF_TYPE_BLOCK);
2925                         if (type == BTRFS_REF_TYPE_INVALID) {
2926                                 ret = -EUCLEAN;
2927                                 goto out;
2928                         }
2929                         key.type = type;
2930                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2931                 } else {
2932                         key.type = iter->cur_key.type;
2933                         key.offset = iter->cur_key.offset;
2934                 }
2935
2936                 /*
2937                  * Parent node found and matches current inline ref, no need to
2938                  * rebuild this node for this inline ref
2939                  */
2940                 if (exist &&
2941                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2942                       exist->owner == key.offset) ||
2943                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2944                       exist->bytenr == key.offset))) {
2945                         exist = NULL;
2946                         continue;
2947                 }
2948
2949                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2950                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2951                         ret = handle_direct_tree_backref(cache, &key, cur);
2952                         if (ret < 0)
2953                                 goto out;
2954                         continue;
2955                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2956                         ret = -EINVAL;
2957                         btrfs_print_v0_err(fs_info);
2958                         btrfs_handle_fs_error(fs_info, ret, NULL);
2959                         goto out;
2960                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2961                         continue;
2962                 }
2963
2964                 /*
2965                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2966                  * means the root objectid. We need to search the tree to get
2967                  * its parent bytenr.
2968                  */
2969                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2970                                                    cur);
2971                 if (ret < 0)
2972                         goto out;
2973         }
2974         ret = 0;
2975         cur->checked = 1;
2976         WARN_ON(exist);
2977 out:
2978         btrfs_backref_iter_release(iter);
2979         return ret;
2980 }
2981
2982 /*
2983  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
2984  */
2985 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
2986                                      struct btrfs_backref_node *start)
2987 {
2988         struct list_head *useless_node = &cache->useless_node;
2989         struct btrfs_backref_edge *edge;
2990         struct rb_node *rb_node;
2991         LIST_HEAD(pending_edge);
2992
2993         ASSERT(start->checked);
2994
2995         /* Insert this node to cache if it's not COW-only */
2996         if (!start->cowonly) {
2997                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
2998                                            &start->rb_node);
2999                 if (rb_node)
3000                         btrfs_backref_panic(cache->fs_info, start->bytenr,
3001                                             -EEXIST);
3002                 list_add_tail(&start->lower, &cache->leaves);
3003         }
3004
3005         /*
3006          * Use breadth first search to iterate all related edges.
3007          *
3008          * The starting points are all the edges of this node
3009          */
3010         list_for_each_entry(edge, &start->upper, list[LOWER])
3011                 list_add_tail(&edge->list[UPPER], &pending_edge);
3012
3013         while (!list_empty(&pending_edge)) {
3014                 struct btrfs_backref_node *upper;
3015                 struct btrfs_backref_node *lower;
3016
3017                 edge = list_first_entry(&pending_edge,
3018                                 struct btrfs_backref_edge, list[UPPER]);
3019                 list_del_init(&edge->list[UPPER]);
3020                 upper = edge->node[UPPER];
3021                 lower = edge->node[LOWER];
3022
3023                 /* Parent is detached, no need to keep any edges */
3024                 if (upper->detached) {
3025                         list_del(&edge->list[LOWER]);
3026                         btrfs_backref_free_edge(cache, edge);
3027
3028                         /* Lower node is orphan, queue for cleanup */
3029                         if (list_empty(&lower->upper))
3030                                 list_add(&lower->list, useless_node);
3031                         continue;
3032                 }
3033
3034                 /*
3035                  * All new nodes added in current build_backref_tree() haven't
3036                  * been linked to the cache rb tree.
3037                  * So if we have upper->rb_node populated, this means a cache
3038                  * hit. We only need to link the edge, as @upper and all its
3039                  * parents have already been linked.
3040                  */
3041                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3042                         if (upper->lowest) {
3043                                 list_del_init(&upper->lower);
3044                                 upper->lowest = 0;
3045                         }
3046
3047                         list_add_tail(&edge->list[UPPER], &upper->lower);
3048                         continue;
3049                 }
3050
3051                 /* Sanity check, we shouldn't have any unchecked nodes */
3052                 if (!upper->checked) {
3053                         ASSERT(0);
3054                         return -EUCLEAN;
3055                 }
3056
3057                 /* Sanity check, COW-only node has non-COW-only parent */
3058                 if (start->cowonly != upper->cowonly) {
3059                         ASSERT(0);
3060                         return -EUCLEAN;
3061                 }
3062
3063                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3064                 if (!upper->cowonly) {
3065                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3066                                                    &upper->rb_node);
3067                         if (rb_node) {
3068                                 btrfs_backref_panic(cache->fs_info,
3069                                                 upper->bytenr, -EEXIST);
3070                                 return -EUCLEAN;
3071                         }
3072                 }
3073
3074                 list_add_tail(&edge->list[UPPER], &upper->lower);
3075
3076                 /*
3077                  * Also queue all the parent edges of this uncached node
3078                  * to finish the upper linkage
3079                  */
3080                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3081                         list_add_tail(&edge->list[UPPER], &pending_edge);
3082         }
3083         return 0;
3084 }
3085
3086 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3087                                  struct btrfs_backref_node *node)
3088 {
3089         struct btrfs_backref_node *lower;
3090         struct btrfs_backref_node *upper;
3091         struct btrfs_backref_edge *edge;
3092
3093         while (!list_empty(&cache->useless_node)) {
3094                 lower = list_first_entry(&cache->useless_node,
3095                                    struct btrfs_backref_node, list);
3096                 list_del_init(&lower->list);
3097         }
3098         while (!list_empty(&cache->pending_edge)) {
3099                 edge = list_first_entry(&cache->pending_edge,
3100                                 struct btrfs_backref_edge, list[UPPER]);
3101                 list_del(&edge->list[UPPER]);
3102                 list_del(&edge->list[LOWER]);
3103                 lower = edge->node[LOWER];
3104                 upper = edge->node[UPPER];
3105                 btrfs_backref_free_edge(cache, edge);
3106
3107                 /*
3108                  * Lower is no longer linked to any upper backref nodes and
3109                  * isn't in the cache, we can free it ourselves.
3110                  */
3111                 if (list_empty(&lower->upper) &&
3112                     RB_EMPTY_NODE(&lower->rb_node))
3113                         list_add(&lower->list, &cache->useless_node);
3114
3115                 if (!RB_EMPTY_NODE(&upper->rb_node))
3116                         continue;
3117
3118                 /* Add this guy's upper edges to the list to process */
3119                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3120                         list_add_tail(&edge->list[UPPER],
3121                                       &cache->pending_edge);
3122                 if (list_empty(&upper->upper))
3123                         list_add(&upper->list, &cache->useless_node);
3124         }
3125
3126         while (!list_empty(&cache->useless_node)) {
3127                 lower = list_first_entry(&cache->useless_node,
3128                                    struct btrfs_backref_node, list);
3129                 list_del_init(&lower->list);
3130                 if (lower == node)
3131                         node = NULL;
3132                 btrfs_backref_drop_node(cache, lower);
3133         }
3134
3135         btrfs_backref_cleanup_node(cache, node);
3136         ASSERT(list_empty(&cache->useless_node) &&
3137                list_empty(&cache->pending_edge));
3138 }