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                 } else if (!extent_buffer_uptodate(eb)) {
 793                         free_pref(ref);
 794                         free_extent_buffer(eb);
 795                         return -EIO;
 796                 }
 797                 if (lock)
 798                         btrfs_tree_read_lock(eb);
 799                 if (btrfs_header_level(eb) == 0)
 800                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 801                 else
 802                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 803                 if (lock)
 804                         btrfs_tree_read_unlock(eb);
 805                 free_extent_buffer(eb);
 806                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 807                 cond_resched();
 808         }
 809         return 0;
 810 }
 811
 812 /*
 813  * add all currently queued delayed refs from this head whose seq nr is
 814  * smaller or equal that seq to the list
 815  */
 816 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 817                             struct btrfs_delayed_ref_head *head, u64 seq,
 818                             struct preftrees *preftrees, struct share_check *sc)
 819 {
 820         struct btrfs_delayed_ref_node *node;
 821         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 822         struct btrfs_key key;
 823         struct btrfs_key tmp_op_key;
 824         struct rb_node *n;
 825         int count;
 826         int ret = 0;
 827
 828         if (extent_op && extent_op->update_key)
 829                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 830
 831         spin_lock(&head->lock);
 832         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 833                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 834                                 ref_node);
 835                 if (node->seq > seq)
 836                         continue;
 837
 838                 switch (node->action) {
 839                 case BTRFS_ADD_DELAYED_EXTENT:
 840                 case BTRFS_UPDATE_DELAYED_HEAD:
 841                         WARN_ON(1);
 842                         continue;
 843                 case BTRFS_ADD_DELAYED_REF:
 844                         count = node->ref_mod;
 845                         break;
 846                 case BTRFS_DROP_DELAYED_REF:
 847                         count = node->ref_mod * -1;
 848                         break;
 849                 default:
 850                         BUG();
 851                 }
 852                 switch (node->type) {
 853                 case BTRFS_TREE_BLOCK_REF_KEY: {
 854                         /* NORMAL INDIRECT METADATA backref */
 855                         struct btrfs_delayed_tree_ref *ref;
 856
 857                         ref = btrfs_delayed_node_to_tree_ref(node);
 858                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 859                                                &tmp_op_key, ref->level + 1,
 860                                                node->bytenr, count, sc,
 861                                                GFP_ATOMIC);
 862                         break;
 863                 }
 864                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 865                         /* SHARED DIRECT METADATA backref */
 866                         struct btrfs_delayed_tree_ref *ref;
 867
 868                         ref = btrfs_delayed_node_to_tree_ref(node);
 869
 870                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 871                                              ref->parent, node->bytenr, count,
 872                                              sc, GFP_ATOMIC);
 873                         break;
 874                 }
 875                 case BTRFS_EXTENT_DATA_REF_KEY: {
 876                         /* NORMAL INDIRECT DATA backref */
 877                         struct btrfs_delayed_data_ref *ref;
 878                         ref = btrfs_delayed_node_to_data_ref(node);
 879
 880                         key.objectid = ref->objectid;
 881                         key.type = BTRFS_EXTENT_DATA_KEY;
 882                         key.offset = ref->offset;
 883
 884                         /*
 885                          * Found a inum that doesn't match our known inum, we
 886                          * know it's shared.
 887                          */
 888                         if (sc && sc->inum && ref->objectid != sc->inum) {
 889                                 ret = BACKREF_FOUND_SHARED;
 890                                 goto out;
 891                         }
 892
 893                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 894                                                &key, 0, node->bytenr, count, sc,
 895                                                GFP_ATOMIC);
 896                         break;
 897                 }
 898                 case BTRFS_SHARED_DATA_REF_KEY: {
 899                         /* SHARED DIRECT FULL backref */
 900                         struct btrfs_delayed_data_ref *ref;
 901
 902                         ref = btrfs_delayed_node_to_data_ref(node);
 903
 904                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 905                                              node->bytenr, count, sc,
 906                                              GFP_ATOMIC);
 907                         break;
 908                 }
 909                 default:
 910                         WARN_ON(1);
 911                 }
 912                 /*
 913                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 914                  * refs have been checked.
 915                  */
 916                 if (ret && (ret != BACKREF_FOUND_SHARED))
 917                         break;
 918         }
 919         if (!ret)
 920                 ret = extent_is_shared(sc);
 921 out:
 922         spin_unlock(&head->lock);
 923         return ret;
 924 }
 925
 926 /*
 927  * add all inline backrefs for bytenr to the list
 928  *
 929  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 930  */
 931 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 932                            struct btrfs_path *path, u64 bytenr,
 933                            int *info_level, struct preftrees *preftrees,
 934                            struct share_check *sc)
 935 {
 936         int ret = 0;
 937         int slot;
 938         struct extent_buffer *leaf;
 939         struct btrfs_key key;
 940         struct btrfs_key found_key;
 941         unsigned long ptr;
 942         unsigned long end;
 943         struct btrfs_extent_item *ei;
 944         u64 flags;
 945         u64 item_size;
 946
 947         /*
 948          * enumerate all inline refs
 949          */
 950         leaf = path->nodes[0];
 951         slot = path->slots[0];
 952
 953         item_size = btrfs_item_size(leaf, slot);
 954         BUG_ON(item_size < sizeof(*ei));
 955
 956         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 957         flags = btrfs_extent_flags(leaf, ei);
 958         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 959
 960         ptr = (unsigned long)(ei + 1);
 961         end = (unsigned long)ei + item_size;
 962
 963         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 964             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 965                 struct btrfs_tree_block_info *info;
 966
 967                 info = (struct btrfs_tree_block_info *)ptr;
 968                 *info_level = btrfs_tree_block_level(leaf, info);
 969                 ptr += sizeof(struct btrfs_tree_block_info);
 970                 BUG_ON(ptr > end);
 971         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 972                 *info_level = found_key.offset;
 973         } else {
 974                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 975         }
 976
 977         while (ptr < end) {
 978                 struct btrfs_extent_inline_ref *iref;
 979                 u64 offset;
 980                 int type;
 981
 982                 iref = (struct btrfs_extent_inline_ref *)ptr;
 983                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 984                                                         BTRFS_REF_TYPE_ANY);
 985                 if (type == BTRFS_REF_TYPE_INVALID)
 986                         return -EUCLEAN;
 987
 988                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 989
 990                 switch (type) {
 991                 case BTRFS_SHARED_BLOCK_REF_KEY:
 992                         ret = add_direct_ref(fs_info, preftrees,
 993                                              *info_level + 1, offset,
 994                                              bytenr, 1, NULL, GFP_NOFS);
 995                         break;
 996                 case BTRFS_SHARED_DATA_REF_KEY: {
 997                         struct btrfs_shared_data_ref *sdref;
 998                         int count;
 999
1000                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1001                         count = btrfs_shared_data_ref_count(leaf, sdref);
1002
1003                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
1004                                              bytenr, count, sc, GFP_NOFS);
1005                         break;
1006                 }
1007                 case BTRFS_TREE_BLOCK_REF_KEY:
1008                         ret = add_indirect_ref(fs_info, preftrees, offset,
1009                                                NULL, *info_level + 1,
1010                                                bytenr, 1, NULL, GFP_NOFS);
1011                         break;
1012                 case BTRFS_EXTENT_DATA_REF_KEY: {
1013                         struct btrfs_extent_data_ref *dref;
1014                         int count;
1015                         u64 root;
1016
1017                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1018                         count = btrfs_extent_data_ref_count(leaf, dref);
1019                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1020                                                                       dref);
1021                         key.type = BTRFS_EXTENT_DATA_KEY;
1022                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1023
1024                         if (sc && sc->inum && key.objectid != sc->inum) {
1025                                 ret = BACKREF_FOUND_SHARED;
1026                                 break;
1027                         }
1028
1029                         root = btrfs_extent_data_ref_root(leaf, dref);
1030
1031                         ret = add_indirect_ref(fs_info, preftrees, root,
1032                                                &key, 0, bytenr, count,
1033                                                sc, GFP_NOFS);
1034                         break;
1035                 }
1036                 default:
1037                         WARN_ON(1);
1038                 }
1039                 if (ret)
1040                         return ret;
1041                 ptr += btrfs_extent_inline_ref_size(type);
1042         }
1043
1044         return 0;
1045 }
1046
1047 /*
1048  * add all non-inline backrefs for bytenr to the list
1049  *
1050  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1051  */
1052 static int add_keyed_refs(struct btrfs_root *extent_root,
1053                           struct btrfs_path *path, u64 bytenr,
1054                           int info_level, struct preftrees *preftrees,
1055                           struct share_check *sc)
1056 {
1057         struct btrfs_fs_info *fs_info = extent_root->fs_info;
1058         int ret;
1059         int slot;
1060         struct extent_buffer *leaf;
1061         struct btrfs_key key;
1062
1063         while (1) {
1064                 ret = btrfs_next_item(extent_root, path);
1065                 if (ret < 0)
1066                         break;
1067                 if (ret) {
1068                         ret = 0;
1069                         break;
1070                 }
1071
1072                 slot = path->slots[0];
1073                 leaf = path->nodes[0];
1074                 btrfs_item_key_to_cpu(leaf, &key, slot);
1075
1076                 if (key.objectid != bytenr)
1077                         break;
1078                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1079                         continue;
1080                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1081                         break;
1082
1083                 switch (key.type) {
1084                 case BTRFS_SHARED_BLOCK_REF_KEY:
1085                         /* SHARED DIRECT METADATA backref */
1086                         ret = add_direct_ref(fs_info, preftrees,
1087                                              info_level + 1, key.offset,
1088                                              bytenr, 1, NULL, GFP_NOFS);
1089                         break;
1090                 case BTRFS_SHARED_DATA_REF_KEY: {
1091                         /* SHARED DIRECT FULL backref */
1092                         struct btrfs_shared_data_ref *sdref;
1093                         int count;
1094
1095                         sdref = btrfs_item_ptr(leaf, slot,
1096                                               struct btrfs_shared_data_ref);
1097                         count = btrfs_shared_data_ref_count(leaf, sdref);
1098                         ret = add_direct_ref(fs_info, preftrees, 0,
1099                                              key.offset, bytenr, count,
1100                                              sc, GFP_NOFS);
1101                         break;
1102                 }
1103                 case BTRFS_TREE_BLOCK_REF_KEY:
1104                         /* NORMAL INDIRECT METADATA backref */
1105                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1106                                                NULL, info_level + 1, bytenr,
1107                                                1, NULL, GFP_NOFS);
1108                         break;
1109                 case BTRFS_EXTENT_DATA_REF_KEY: {
1110                         /* NORMAL INDIRECT DATA backref */
1111                         struct btrfs_extent_data_ref *dref;
1112                         int count;
1113                         u64 root;
1114
1115                         dref = btrfs_item_ptr(leaf, slot,
1116                                               struct btrfs_extent_data_ref);
1117                         count = btrfs_extent_data_ref_count(leaf, dref);
1118                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1119                                                                       dref);
1120                         key.type = BTRFS_EXTENT_DATA_KEY;
1121                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1122
1123                         if (sc && sc->inum && key.objectid != sc->inum) {
1124                                 ret = BACKREF_FOUND_SHARED;
1125                                 break;
1126                         }
1127
1128                         root = btrfs_extent_data_ref_root(leaf, dref);
1129                         ret = add_indirect_ref(fs_info, preftrees, root,
1130                                                &key, 0, bytenr, count,
1131                                                sc, GFP_NOFS);
1132                         break;
1133                 }
1134                 default:
1135                         WARN_ON(1);
1136                 }
1137                 if (ret)
1138                         return ret;
1139
1140         }
1141
1142         return ret;
1143 }
1144
1145 /*
1146  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1147  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1148  * indirect refs to their parent bytenr.
1149  * When roots are found, they're added to the roots list
1150  *
1151  * If time_seq is set to BTRFS_SEQ_LAST, it will not search delayed_refs, and
1152  * behave much like trans == NULL case, the difference only lies in it will not
1153  * commit root.
1154  * The special case is for qgroup to search roots in commit_transaction().
1155  *
1156  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1157  * shared extent is detected.
1158  *
1159  * Otherwise this returns 0 for success and <0 for an error.
1160  *
1161  * If ignore_offset is set to false, only extent refs whose offsets match
1162  * extent_item_pos are returned.  If true, every extent ref is returned
1163  * and extent_item_pos is ignored.
1164  *
1165  * FIXME some caching might speed things up
1166  */
1167 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1168                              struct btrfs_fs_info *fs_info, u64 bytenr,
1169                              u64 time_seq, struct ulist *refs,
1170                              struct ulist *roots, const u64 *extent_item_pos,
1171                              struct share_check *sc, bool ignore_offset)
1172 {
1173         struct btrfs_root *root = btrfs_extent_root(fs_info, bytenr);
1174         struct btrfs_key key;
1175         struct btrfs_path *path;
1176         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1177         struct btrfs_delayed_ref_head *head;
1178         int info_level = 0;
1179         int ret;
1180         struct prelim_ref *ref;
1181         struct rb_node *node;
1182         struct extent_inode_elem *eie = NULL;
1183         struct preftrees preftrees = {
1184                 .direct = PREFTREE_INIT,
1185                 .indirect = PREFTREE_INIT,
1186                 .indirect_missing_keys = PREFTREE_INIT
1187         };
1188
1189         key.objectid = bytenr;
1190         key.offset = (u64)-1;
1191         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1192                 key.type = BTRFS_METADATA_ITEM_KEY;
1193         else
1194                 key.type = BTRFS_EXTENT_ITEM_KEY;
1195
1196         path = btrfs_alloc_path();
1197         if (!path)
1198                 return -ENOMEM;
1199         if (!trans) {
1200                 path->search_commit_root = 1;
1201                 path->skip_locking = 1;
1202         }
1203
1204         if (time_seq == BTRFS_SEQ_LAST)
1205                 path->skip_locking = 1;
1206
1207 again:
1208         head = NULL;
1209
1210         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1211         if (ret < 0)
1212                 goto out;
1213         if (ret == 0) {
1214                 /* This shouldn't happen, indicates a bug or fs corruption. */
1215                 ASSERT(ret != 0);
1216                 ret = -EUCLEAN;
1217                 goto out;
1218         }
1219
1220         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1221             time_seq != BTRFS_SEQ_LAST) {
1222                 /*
1223                  * We have a specific time_seq we care about and trans which
1224                  * means we have the path lock, we need to grab the ref head and
1225                  * lock it so we have a consistent view of the refs at the given
1226                  * time.
1227                  */
1228                 delayed_refs = &trans->transaction->delayed_refs;
1229                 spin_lock(&delayed_refs->lock);
1230                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1231                 if (head) {
1232                         if (!mutex_trylock(&head->mutex)) {
1233                                 refcount_inc(&head->refs);
1234                                 spin_unlock(&delayed_refs->lock);
1235
1236                                 btrfs_release_path(path);
1237
1238                                 /*
1239                                  * Mutex was contended, block until it's
1240                                  * released and try again
1241                                  */
1242                                 mutex_lock(&head->mutex);
1243                                 mutex_unlock(&head->mutex);
1244                                 btrfs_put_delayed_ref_head(head);
1245                                 goto again;
1246                         }
1247                         spin_unlock(&delayed_refs->lock);
1248                         ret = add_delayed_refs(fs_info, head, time_seq,
1249                                                &preftrees, sc);
1250                         mutex_unlock(&head->mutex);
1251                         if (ret)
1252                                 goto out;
1253                 } else {
1254                         spin_unlock(&delayed_refs->lock);
1255                 }
1256         }
1257
1258         if (path->slots[0]) {
1259                 struct extent_buffer *leaf;
1260                 int slot;
1261
1262                 path->slots[0]--;
1263                 leaf = path->nodes[0];
1264                 slot = path->slots[0];
1265                 btrfs_item_key_to_cpu(leaf, &key, slot);
1266                 if (key.objectid == bytenr &&
1267                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1268                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1269                         ret = add_inline_refs(fs_info, path, bytenr,
1270                                               &info_level, &preftrees, sc);
1271                         if (ret)
1272                                 goto out;
1273                         ret = add_keyed_refs(root, path, bytenr, info_level,
1274                                              &preftrees, sc);
1275                         if (ret)
1276                                 goto out;
1277                 }
1278         }
1279
1280         btrfs_release_path(path);
1281
1282         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1283         if (ret)
1284                 goto out;
1285
1286         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1287
1288         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1289                                     extent_item_pos, sc, ignore_offset);
1290         if (ret)
1291                 goto out;
1292
1293         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1294
1295         /*
1296          * This walks the tree of merged and resolved refs. Tree blocks are
1297          * read in as needed. Unique entries are added to the ulist, and
1298          * the list of found roots is updated.
1299          *
1300          * We release the entire tree in one go before returning.
1301          */
1302         node = rb_first_cached(&preftrees.direct.root);
1303         while (node) {
1304                 ref = rb_entry(node, struct prelim_ref, rbnode);
1305                 node = rb_next(&ref->rbnode);
1306                 /*
1307                  * ref->count < 0 can happen here if there are delayed
1308                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1309                  * prelim_ref_insert() relies on this when merging
1310                  * identical refs to keep the overall count correct.
1311                  * prelim_ref_insert() will merge only those refs
1312                  * which compare identically.  Any refs having
1313                  * e.g. different offsets would not be merged,
1314                  * and would retain their original ref->count < 0.
1315                  */
1316                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1317                         if (sc && sc->root_objectid &&
1318                             ref->root_id != sc->root_objectid) {
1319                                 ret = BACKREF_FOUND_SHARED;
1320                                 goto out;
1321                         }
1322
1323                         /* no parent == root of tree */
1324                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1325                         if (ret < 0)
1326                                 goto out;
1327                 }
1328                 if (ref->count && ref->parent) {
1329                         if (extent_item_pos && !ref->inode_list &&
1330                             ref->level == 0) {
1331                                 struct extent_buffer *eb;
1332
1333                                 eb = read_tree_block(fs_info, ref->parent, 0,
1334                                                      0, ref->level, NULL);
1335                                 if (IS_ERR(eb)) {
1336                                         ret = PTR_ERR(eb);
1337                                         goto out;
1338                                 } else if (!extent_buffer_uptodate(eb)) {
1339                                         free_extent_buffer(eb);
1340                                         ret = -EIO;
1341                                         goto out;
1342                                 }
1343
1344                                 if (!path->skip_locking)
1345                                         btrfs_tree_read_lock(eb);
1346                                 ret = find_extent_in_eb(eb, bytenr,
1347                                                         *extent_item_pos, &eie, ignore_offset);
1348                                 if (!path->skip_locking)
1349                                         btrfs_tree_read_unlock(eb);
1350                                 free_extent_buffer(eb);
1351                                 if (ret < 0)
1352                                         goto out;
1353                                 ref->inode_list = eie;
1354                         }
1355                         ret = ulist_add_merge_ptr(refs, ref->parent,
1356                                                   ref->inode_list,
1357                                                   (void **)&eie, GFP_NOFS);
1358                         if (ret < 0)
1359                                 goto out;
1360                         if (!ret && extent_item_pos) {
1361                                 /*
1362                                  * We've recorded that parent, so we must extend
1363                                  * its inode list here.
1364                                  *
1365                                  * However if there was corruption we may not
1366                                  * have found an eie, return an error in this
1367                                  * case.
1368                                  */
1369                                 ASSERT(eie);
1370                                 if (!eie) {
1371                                         ret = -EUCLEAN;
1372                                         goto out;
1373                                 }
1374                                 while (eie->next)
1375                                         eie = eie->next;
1376                                 eie->next = ref->inode_list;
1377                         }
1378                         eie = NULL;
1379                 }
1380                 cond_resched();
1381         }
1382
1383 out:
1384         btrfs_free_path(path);
1385
1386         prelim_release(&preftrees.direct);
1387         prelim_release(&preftrees.indirect);
1388         prelim_release(&preftrees.indirect_missing_keys);
1389
1390         if (ret < 0)
1391                 free_inode_elem_list(eie);
1392         return ret;
1393 }
1394
1395 static void free_leaf_list(struct ulist *blocks)
1396 {
1397         struct ulist_node *node = NULL;
1398         struct extent_inode_elem *eie;
1399         struct ulist_iterator uiter;
1400
1401         ULIST_ITER_INIT(&uiter);
1402         while ((node = ulist_next(blocks, &uiter))) {
1403                 if (!node->aux)
1404                         continue;
1405                 eie = unode_aux_to_inode_list(node);
1406                 free_inode_elem_list(eie);
1407                 node->aux = 0;
1408         }
1409
1410         ulist_free(blocks);
1411 }
1412
1413 /*
1414  * Finds all leafs with a reference to the specified combination of bytenr and
1415  * offset. key_list_head will point to a list of corresponding keys (caller must
1416  * free each list element). The leafs will be stored in the leafs ulist, which
1417  * must be freed with ulist_free.
1418  *
1419  * returns 0 on success, <0 on error
1420  */
1421 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1422                          struct btrfs_fs_info *fs_info, u64 bytenr,
1423                          u64 time_seq, struct ulist **leafs,
1424                          const u64 *extent_item_pos, bool ignore_offset)
1425 {
1426         int ret;
1427
1428         *leafs = ulist_alloc(GFP_NOFS);
1429         if (!*leafs)
1430                 return -ENOMEM;
1431
1432         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1433                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1434         if (ret < 0 && ret != -ENOENT) {
1435                 free_leaf_list(*leafs);
1436                 return ret;
1437         }
1438
1439         return 0;
1440 }
1441
1442 /*
1443  * walk all backrefs for a given extent to find all roots that reference this
1444  * extent. Walking a backref means finding all extents that reference this
1445  * extent and in turn walk the backrefs of those, too. Naturally this is a
1446  * recursive process, but here it is implemented in an iterative fashion: We
1447  * find all referencing extents for the extent in question and put them on a
1448  * list. In turn, we find all referencing extents for those, further appending
1449  * to the list. The way we iterate the list allows adding more elements after
1450  * the current while iterating. The process stops when we reach the end of the
1451  * list. Found roots are added to the roots list.
1452  *
1453  * returns 0 on success, < 0 on error.
1454  */
1455 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1456                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1457                                      u64 time_seq, struct ulist **roots,
1458                                      bool ignore_offset)
1459 {
1460         struct ulist *tmp;
1461         struct ulist_node *node = NULL;
1462         struct ulist_iterator uiter;
1463         int ret;
1464
1465         tmp = ulist_alloc(GFP_NOFS);
1466         if (!tmp)
1467                 return -ENOMEM;
1468         *roots = ulist_alloc(GFP_NOFS);
1469         if (!*roots) {
1470                 ulist_free(tmp);
1471                 return -ENOMEM;
1472         }
1473
1474         ULIST_ITER_INIT(&uiter);
1475         while (1) {
1476                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1477                                         tmp, *roots, NULL, NULL, ignore_offset);
1478                 if (ret < 0 && ret != -ENOENT) {
1479                         ulist_free(tmp);
1480                         ulist_free(*roots);
1481                         *roots = NULL;
1482                         return ret;
1483                 }
1484                 node = ulist_next(tmp, &uiter);
1485                 if (!node)
1486                         break;
1487                 bytenr = node->val;
1488                 cond_resched();
1489         }
1490
1491         ulist_free(tmp);
1492         return 0;
1493 }
1494
1495 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1496                          struct btrfs_fs_info *fs_info, u64 bytenr,
1497                          u64 time_seq, struct ulist **roots,
1498                          bool skip_commit_root_sem)
1499 {
1500         int ret;
1501
1502         if (!trans && !skip_commit_root_sem)
1503                 down_read(&fs_info->commit_root_sem);
1504         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1505                                         time_seq, roots, false);
1506         if (!trans && !skip_commit_root_sem)
1507                 up_read(&fs_info->commit_root_sem);
1508         return ret;
1509 }
1510
1511 /**
1512  * Check if an extent is shared or not
1513  *
1514  * @root:   root inode belongs to
1515  * @inum:   inode number of the inode whose extent we are checking
1516  * @bytenr: logical bytenr of the extent we are checking
1517  * @roots:  list of roots this extent is shared among
1518  * @tmp:    temporary list used for iteration
1519  *
1520  * btrfs_check_shared uses the backref walking code but will short
1521  * circuit as soon as it finds a root or inode that doesn't match the
1522  * one passed in. This provides a significant performance benefit for
1523  * callers (such as fiemap) which want to know whether the extent is
1524  * shared but do not need a ref count.
1525  *
1526  * This attempts to attach to the running transaction in order to account for
1527  * delayed refs, but continues on even when no running transaction exists.
1528  *
1529  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1530  */
1531 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1532                 struct ulist *roots, struct ulist *tmp)
1533 {
1534         struct btrfs_fs_info *fs_info = root->fs_info;
1535         struct btrfs_trans_handle *trans;
1536         struct ulist_iterator uiter;
1537         struct ulist_node *node;
1538         struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1539         int ret = 0;
1540         struct share_check shared = {
1541                 .root_objectid = root->root_key.objectid,
1542                 .inum = inum,
1543                 .share_count = 0,
1544         };
1545
1546         ulist_init(roots);
1547         ulist_init(tmp);
1548
1549         trans = btrfs_join_transaction_nostart(root);
1550         if (IS_ERR(trans)) {
1551                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1552                         ret = PTR_ERR(trans);
1553                         goto out;
1554                 }
1555                 trans = NULL;
1556                 down_read(&fs_info->commit_root_sem);
1557         } else {
1558                 btrfs_get_tree_mod_seq(fs_info, &elem);
1559         }
1560
1561         ULIST_ITER_INIT(&uiter);
1562         while (1) {
1563                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1564                                         roots, NULL, &shared, false);
1565                 if (ret == BACKREF_FOUND_SHARED) {
1566                         /* this is the only condition under which we return 1 */
1567                         ret = 1;
1568                         break;
1569                 }
1570                 if (ret < 0 && ret != -ENOENT)
1571                         break;
1572                 ret = 0;
1573                 node = ulist_next(tmp, &uiter);
1574                 if (!node)
1575                         break;
1576                 bytenr = node->val;
1577                 shared.share_count = 0;
1578                 cond_resched();
1579         }
1580
1581         if (trans) {
1582                 btrfs_put_tree_mod_seq(fs_info, &elem);
1583                 btrfs_end_transaction(trans);
1584         } else {
1585                 up_read(&fs_info->commit_root_sem);
1586         }
1587 out:
1588         ulist_release(roots);
1589         ulist_release(tmp);
1590         return ret;
1591 }
1592
1593 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1594                           u64 start_off, struct btrfs_path *path,
1595                           struct btrfs_inode_extref **ret_extref,
1596                           u64 *found_off)
1597 {
1598         int ret, slot;
1599         struct btrfs_key key;
1600         struct btrfs_key found_key;
1601         struct btrfs_inode_extref *extref;
1602         const struct extent_buffer *leaf;
1603         unsigned long ptr;
1604
1605         key.objectid = inode_objectid;
1606         key.type = BTRFS_INODE_EXTREF_KEY;
1607         key.offset = start_off;
1608
1609         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1610         if (ret < 0)
1611                 return ret;
1612
1613         while (1) {
1614                 leaf = path->nodes[0];
1615                 slot = path->slots[0];
1616                 if (slot >= btrfs_header_nritems(leaf)) {
1617                         /*
1618                          * If the item at offset is not found,
1619                          * btrfs_search_slot will point us to the slot
1620                          * where it should be inserted. In our case
1621                          * that will be the slot directly before the
1622                          * next INODE_REF_KEY_V2 item. In the case
1623                          * that we're pointing to the last slot in a
1624                          * leaf, we must move one leaf over.
1625                          */
1626                         ret = btrfs_next_leaf(root, path);
1627                         if (ret) {
1628                                 if (ret >= 1)
1629                                         ret = -ENOENT;
1630                                 break;
1631                         }
1632                         continue;
1633                 }
1634
1635                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1636
1637                 /*
1638                  * Check that we're still looking at an extended ref key for
1639                  * this particular objectid. If we have different
1640                  * objectid or type then there are no more to be found
1641                  * in the tree and we can exit.
1642                  */
1643                 ret = -ENOENT;
1644                 if (found_key.objectid != inode_objectid)
1645                         break;
1646                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1647                         break;
1648
1649                 ret = 0;
1650                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1651                 extref = (struct btrfs_inode_extref *)ptr;
1652                 *ret_extref = extref;
1653                 if (found_off)
1654                         *found_off = found_key.offset;
1655                 break;
1656         }
1657
1658         return ret;
1659 }
1660
1661 /*
1662  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1663  * Elements of the path are separated by '/' and the path is guaranteed to be
1664  * 0-terminated. the path is only given within the current file system.
1665  * Therefore, it never starts with a '/'. the caller is responsible to provide
1666  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1667  * the start point of the resulting string is returned. this pointer is within
1668  * dest, normally.
1669  * in case the path buffer would overflow, the pointer is decremented further
1670  * as if output was written to the buffer, though no more output is actually
1671  * generated. that way, the caller can determine how much space would be
1672  * required for the path to fit into the buffer. in that case, the returned
1673  * value will be smaller than dest. callers must check this!
1674  */
1675 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1676                         u32 name_len, unsigned long name_off,
1677                         struct extent_buffer *eb_in, u64 parent,
1678                         char *dest, u32 size)
1679 {
1680         int slot;
1681         u64 next_inum;
1682         int ret;
1683         s64 bytes_left = ((s64)size) - 1;
1684         struct extent_buffer *eb = eb_in;
1685         struct btrfs_key found_key;
1686         struct btrfs_inode_ref *iref;
1687
1688         if (bytes_left >= 0)
1689                 dest[bytes_left] = '\0';
1690
1691         while (1) {
1692                 bytes_left -= name_len;
1693                 if (bytes_left >= 0)
1694                         read_extent_buffer(eb, dest + bytes_left,
1695                                            name_off, name_len);
1696                 if (eb != eb_in) {
1697                         if (!path->skip_locking)
1698                                 btrfs_tree_read_unlock(eb);
1699                         free_extent_buffer(eb);
1700                 }
1701                 ret = btrfs_find_item(fs_root, path, parent, 0,
1702                                 BTRFS_INODE_REF_KEY, &found_key);
1703                 if (ret > 0)
1704                         ret = -ENOENT;
1705                 if (ret)
1706                         break;
1707
1708                 next_inum = found_key.offset;
1709
1710                 /* regular exit ahead */
1711                 if (parent == next_inum)
1712                         break;
1713
1714                 slot = path->slots[0];
1715                 eb = path->nodes[0];
1716                 /* make sure we can use eb after releasing the path */
1717                 if (eb != eb_in) {
1718                         path->nodes[0] = NULL;
1719                         path->locks[0] = 0;
1720                 }
1721                 btrfs_release_path(path);
1722                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1723
1724                 name_len = btrfs_inode_ref_name_len(eb, iref);
1725                 name_off = (unsigned long)(iref + 1);
1726
1727                 parent = next_inum;
1728                 --bytes_left;
1729                 if (bytes_left >= 0)
1730                         dest[bytes_left] = '/';
1731         }
1732
1733         btrfs_release_path(path);
1734
1735         if (ret)
1736                 return ERR_PTR(ret);
1737
1738         return dest + bytes_left;
1739 }
1740
1741 /*
1742  * this makes the path point to (logical EXTENT_ITEM *)
1743  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1744  * tree blocks and <0 on error.
1745  */
1746 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1747                         struct btrfs_path *path, struct btrfs_key *found_key,
1748                         u64 *flags_ret)
1749 {
1750         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
1751         int ret;
1752         u64 flags;
1753         u64 size = 0;
1754         u32 item_size;
1755         const struct extent_buffer *eb;
1756         struct btrfs_extent_item *ei;
1757         struct btrfs_key key;
1758
1759         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1760                 key.type = BTRFS_METADATA_ITEM_KEY;
1761         else
1762                 key.type = BTRFS_EXTENT_ITEM_KEY;
1763         key.objectid = logical;
1764         key.offset = (u64)-1;
1765
1766         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1767         if (ret < 0)
1768                 return ret;
1769
1770         ret = btrfs_previous_extent_item(extent_root, path, 0);
1771         if (ret) {
1772                 if (ret > 0)
1773                         ret = -ENOENT;
1774                 return ret;
1775         }
1776         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1777         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1778                 size = fs_info->nodesize;
1779         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1780                 size = found_key->offset;
1781
1782         if (found_key->objectid > logical ||
1783             found_key->objectid + size <= logical) {
1784                 btrfs_debug(fs_info,
1785                         "logical %llu is not within any extent", logical);
1786                 return -ENOENT;
1787         }
1788
1789         eb = path->nodes[0];
1790         item_size = btrfs_item_size(eb, path->slots[0]);
1791         BUG_ON(item_size < sizeof(*ei));
1792
1793         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1794         flags = btrfs_extent_flags(eb, ei);
1795
1796         btrfs_debug(fs_info,
1797                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1798                  logical, logical - found_key->objectid, found_key->objectid,
1799                  found_key->offset, flags, item_size);
1800
1801         WARN_ON(!flags_ret);
1802         if (flags_ret) {
1803                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1804                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1805                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1806                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1807                 else
1808                         BUG();
1809                 return 0;
1810         }
1811
1812         return -EIO;
1813 }
1814
1815 /*
1816  * helper function to iterate extent inline refs. ptr must point to a 0 value
1817  * for the first call and may be modified. it is used to track state.
1818  * if more refs exist, 0 is returned and the next call to
1819  * get_extent_inline_ref must pass the modified ptr parameter to get the
1820  * next ref. after the last ref was processed, 1 is returned.
1821  * returns <0 on error
1822  */
1823 static int get_extent_inline_ref(unsigned long *ptr,
1824                                  const struct extent_buffer *eb,
1825                                  const struct btrfs_key *key,
1826                                  const struct btrfs_extent_item *ei,
1827                                  u32 item_size,
1828                                  struct btrfs_extent_inline_ref **out_eiref,
1829                                  int *out_type)
1830 {
1831         unsigned long end;
1832         u64 flags;
1833         struct btrfs_tree_block_info *info;
1834
1835         if (!*ptr) {
1836                 /* first call */
1837                 flags = btrfs_extent_flags(eb, ei);
1838                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1839                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1840                                 /* a skinny metadata extent */
1841                                 *out_eiref =
1842                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1843                         } else {
1844                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1845                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1846                                 *out_eiref =
1847                                    (struct btrfs_extent_inline_ref *)(info + 1);
1848                         }
1849                 } else {
1850                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1851                 }
1852                 *ptr = (unsigned long)*out_eiref;
1853                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1854                         return -ENOENT;
1855         }
1856
1857         end = (unsigned long)ei + item_size;
1858         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1859         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1860                                                      BTRFS_REF_TYPE_ANY);
1861         if (*out_type == BTRFS_REF_TYPE_INVALID)
1862                 return -EUCLEAN;
1863
1864         *ptr += btrfs_extent_inline_ref_size(*out_type);
1865         WARN_ON(*ptr > end);
1866         if (*ptr == end)
1867                 return 1; /* last */
1868
1869         return 0;
1870 }
1871
1872 /*
1873  * reads the tree block backref for an extent. tree level and root are returned
1874  * through out_level and out_root. ptr must point to a 0 value for the first
1875  * call and may be modified (see get_extent_inline_ref comment).
1876  * returns 0 if data was provided, 1 if there was no more data to provide or
1877  * <0 on error.
1878  */
1879 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1880                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1881                             u32 item_size, u64 *out_root, u8 *out_level)
1882 {
1883         int ret;
1884         int type;
1885         struct btrfs_extent_inline_ref *eiref;
1886
1887         if (*ptr == (unsigned long)-1)
1888                 return 1;
1889
1890         while (1) {
1891                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1892                                               &eiref, &type);
1893                 if (ret < 0)
1894                         return ret;
1895
1896                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1897                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1898                         break;
1899
1900                 if (ret == 1)
1901                         return 1;
1902         }
1903
1904         /* we can treat both ref types equally here */
1905         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1906
1907         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1908                 struct btrfs_tree_block_info *info;
1909
1910                 info = (struct btrfs_tree_block_info *)(ei + 1);
1911                 *out_level = btrfs_tree_block_level(eb, info);
1912         } else {
1913                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1914                 *out_level = (u8)key->offset;
1915         }
1916
1917         if (ret == 1)
1918                 *ptr = (unsigned long)-1;
1919
1920         return 0;
1921 }
1922
1923 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1924                              struct extent_inode_elem *inode_list,
1925                              u64 root, u64 extent_item_objectid,
1926                              iterate_extent_inodes_t *iterate, void *ctx)
1927 {
1928         struct extent_inode_elem *eie;
1929         int ret = 0;
1930
1931         for (eie = inode_list; eie; eie = eie->next) {
1932                 btrfs_debug(fs_info,
1933                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1934                             extent_item_objectid, eie->inum,
1935                             eie->offset, root);
1936                 ret = iterate(eie->inum, eie->offset, root, ctx);
1937                 if (ret) {
1938                         btrfs_debug(fs_info,
1939                                     "stopping iteration for %llu due to ret=%d",
1940                                     extent_item_objectid, ret);
1941                         break;
1942                 }
1943         }
1944
1945         return ret;
1946 }
1947
1948 /*
1949  * calls iterate() for every inode that references the extent identified by
1950  * the given parameters.
1951  * when the iterator function returns a non-zero value, iteration stops.
1952  */
1953 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1954                                 u64 extent_item_objectid, u64 extent_item_pos,
1955                                 int search_commit_root,
1956                                 iterate_extent_inodes_t *iterate, void *ctx,
1957                                 bool ignore_offset)
1958 {
1959         int ret;
1960         struct btrfs_trans_handle *trans = NULL;
1961         struct ulist *refs = NULL;
1962         struct ulist *roots = NULL;
1963         struct ulist_node *ref_node = NULL;
1964         struct ulist_node *root_node = NULL;
1965         struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
1966         struct ulist_iterator ref_uiter;
1967         struct ulist_iterator root_uiter;
1968
1969         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1970                         extent_item_objectid);
1971
1972         if (!search_commit_root) {
1973                 trans = btrfs_attach_transaction(fs_info->tree_root);
1974                 if (IS_ERR(trans)) {
1975                         if (PTR_ERR(trans) != -ENOENT &&
1976                             PTR_ERR(trans) != -EROFS)
1977                                 return PTR_ERR(trans);
1978                         trans = NULL;
1979                 }
1980         }
1981
1982         if (trans)
1983                 btrfs_get_tree_mod_seq(fs_info, &seq_elem);
1984         else
1985                 down_read(&fs_info->commit_root_sem);
1986
1987         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1988                                    seq_elem.seq, &refs,
1989                                    &extent_item_pos, ignore_offset);
1990         if (ret)
1991                 goto out;
1992
1993         ULIST_ITER_INIT(&ref_uiter);
1994         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1995                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1996                                                 seq_elem.seq, &roots,
1997                                                 ignore_offset);
1998                 if (ret)
1999                         break;
2000                 ULIST_ITER_INIT(&root_uiter);
2001                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2002                         btrfs_debug(fs_info,
2003                                     "root %llu references leaf %llu, data list %#llx",
2004                                     root_node->val, ref_node->val,
2005                                     ref_node->aux);
2006                         ret = iterate_leaf_refs(fs_info,
2007                                                 (struct extent_inode_elem *)
2008                                                 (uintptr_t)ref_node->aux,
2009                                                 root_node->val,
2010                                                 extent_item_objectid,
2011                                                 iterate, ctx);
2012                 }
2013                 ulist_free(roots);
2014         }
2015
2016         free_leaf_list(refs);
2017 out:
2018         if (trans) {
2019                 btrfs_put_tree_mod_seq(fs_info, &seq_elem);
2020                 btrfs_end_transaction(trans);
2021         } else {
2022                 up_read(&fs_info->commit_root_sem);
2023         }
2024
2025         return ret;
2026 }
2027
2028 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2029                                 struct btrfs_path *path,
2030                                 iterate_extent_inodes_t *iterate, void *ctx,
2031                                 bool ignore_offset)
2032 {
2033         int ret;
2034         u64 extent_item_pos;
2035         u64 flags = 0;
2036         struct btrfs_key found_key;
2037         int search_commit_root = path->search_commit_root;
2038
2039         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2040         btrfs_release_path(path);
2041         if (ret < 0)
2042                 return ret;
2043         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2044                 return -EINVAL;
2045
2046         extent_item_pos = logical - found_key.objectid;
2047         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2048                                         extent_item_pos, search_commit_root,
2049                                         iterate, ctx, ignore_offset);
2050
2051         return ret;
2052 }
2053
2054 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2055                               struct extent_buffer *eb, void *ctx);
2056
2057 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2058                               struct btrfs_path *path,
2059                               iterate_irefs_t *iterate, void *ctx)
2060 {
2061         int ret = 0;
2062         int slot;
2063         u32 cur;
2064         u32 len;
2065         u32 name_len;
2066         u64 parent = 0;
2067         int found = 0;
2068         struct extent_buffer *eb;
2069         struct btrfs_inode_ref *iref;
2070         struct btrfs_key found_key;
2071
2072         while (!ret) {
2073                 ret = btrfs_find_item(fs_root, path, inum,
2074                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2075                                 &found_key);
2076
2077                 if (ret < 0)
2078                         break;
2079                 if (ret) {
2080                         ret = found ? 0 : -ENOENT;
2081                         break;
2082                 }
2083                 ++found;
2084
2085                 parent = found_key.offset;
2086                 slot = path->slots[0];
2087                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2088                 if (!eb) {
2089                         ret = -ENOMEM;
2090                         break;
2091                 }
2092                 btrfs_release_path(path);
2093
2094                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2095
2096                 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2097                         name_len = btrfs_inode_ref_name_len(eb, iref);
2098                         /* path must be released before calling iterate()! */
2099                         btrfs_debug(fs_root->fs_info,
2100                                 "following ref at offset %u for inode %llu in tree %llu",
2101                                 cur, found_key.objectid,
2102                                 fs_root->root_key.objectid);
2103                         ret = iterate(parent, name_len,
2104                                       (unsigned long)(iref + 1), eb, ctx);
2105                         if (ret)
2106                                 break;
2107                         len = sizeof(*iref) + name_len;
2108                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2109                 }
2110                 free_extent_buffer(eb);
2111         }
2112
2113         btrfs_release_path(path);
2114
2115         return ret;
2116 }
2117
2118 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2119                                  struct btrfs_path *path,
2120                                  iterate_irefs_t *iterate, void *ctx)
2121 {
2122         int ret;
2123         int slot;
2124         u64 offset = 0;
2125         u64 parent;
2126         int found = 0;
2127         struct extent_buffer *eb;
2128         struct btrfs_inode_extref *extref;
2129         u32 item_size;
2130         u32 cur_offset;
2131         unsigned long ptr;
2132
2133         while (1) {
2134                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2135                                             &offset);
2136                 if (ret < 0)
2137                         break;
2138                 if (ret) {
2139                         ret = found ? 0 : -ENOENT;
2140                         break;
2141                 }
2142                 ++found;
2143
2144                 slot = path->slots[0];
2145                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2146                 if (!eb) {
2147                         ret = -ENOMEM;
2148                         break;
2149                 }
2150                 btrfs_release_path(path);
2151
2152                 item_size = btrfs_item_size(eb, slot);
2153                 ptr = btrfs_item_ptr_offset(eb, slot);
2154                 cur_offset = 0;
2155
2156                 while (cur_offset < item_size) {
2157                         u32 name_len;
2158
2159                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2160                         parent = btrfs_inode_extref_parent(eb, extref);
2161                         name_len = btrfs_inode_extref_name_len(eb, extref);
2162                         ret = iterate(parent, name_len,
2163                                       (unsigned long)&extref->name, eb, ctx);
2164                         if (ret)
2165                                 break;
2166
2167                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2168                         cur_offset += sizeof(*extref);
2169                 }
2170                 free_extent_buffer(eb);
2171
2172                 offset++;
2173         }
2174
2175         btrfs_release_path(path);
2176
2177         return ret;
2178 }
2179
2180 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2181                          struct btrfs_path *path, iterate_irefs_t *iterate,
2182                          void *ctx)
2183 {
2184         int ret;
2185         int found_refs = 0;
2186
2187         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2188         if (!ret)
2189                 ++found_refs;
2190         else if (ret != -ENOENT)
2191                 return ret;
2192
2193         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2194         if (ret == -ENOENT && found_refs)
2195                 return 0;
2196
2197         return ret;
2198 }
2199
2200 /*
2201  * returns 0 if the path could be dumped (probably truncated)
2202  * returns <0 in case of an error
2203  */
2204 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2205                          struct extent_buffer *eb, void *ctx)
2206 {
2207         struct inode_fs_paths *ipath = ctx;
2208         char *fspath;
2209         char *fspath_min;
2210         int i = ipath->fspath->elem_cnt;
2211         const int s_ptr = sizeof(char *);
2212         u32 bytes_left;
2213
2214         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2215                                         ipath->fspath->bytes_left - s_ptr : 0;
2216
2217         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2218         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2219                                    name_off, eb, inum, fspath_min, bytes_left);
2220         if (IS_ERR(fspath))
2221                 return PTR_ERR(fspath);
2222
2223         if (fspath > fspath_min) {
2224                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2225                 ++ipath->fspath->elem_cnt;
2226                 ipath->fspath->bytes_left = fspath - fspath_min;
2227         } else {
2228                 ++ipath->fspath->elem_missed;
2229                 ipath->fspath->bytes_missing += fspath_min - fspath;
2230                 ipath->fspath->bytes_left = 0;
2231         }
2232
2233         return 0;
2234 }
2235
2236 /*
2237  * this dumps all file system paths to the inode into the ipath struct, provided
2238  * is has been created large enough. each path is zero-terminated and accessed
2239  * from ipath->fspath->val[i].
2240  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2241  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2242  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2243  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2244  * have been needed to return all paths.
2245  */
2246 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2247 {
2248         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2249                              inode_to_path, ipath);
2250 }
2251
2252 struct btrfs_data_container *init_data_container(u32 total_bytes)
2253 {
2254         struct btrfs_data_container *data;
2255         size_t alloc_bytes;
2256
2257         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2258         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2259         if (!data)
2260                 return ERR_PTR(-ENOMEM);
2261
2262         if (total_bytes >= sizeof(*data)) {
2263                 data->bytes_left = total_bytes - sizeof(*data);
2264                 data->bytes_missing = 0;
2265         } else {
2266                 data->bytes_missing = sizeof(*data) - total_bytes;
2267                 data->bytes_left = 0;
2268         }
2269
2270         data->elem_cnt = 0;
2271         data->elem_missed = 0;
2272
2273         return data;
2274 }
2275
2276 /*
2277  * allocates space to return multiple file system paths for an inode.
2278  * total_bytes to allocate are passed, note that space usable for actual path
2279  * information will be total_bytes - sizeof(struct inode_fs_paths).
2280  * the returned pointer must be freed with free_ipath() in the end.
2281  */
2282 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2283                                         struct btrfs_path *path)
2284 {
2285         struct inode_fs_paths *ifp;
2286         struct btrfs_data_container *fspath;
2287
2288         fspath = init_data_container(total_bytes);
2289         if (IS_ERR(fspath))
2290                 return ERR_CAST(fspath);
2291
2292         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2293         if (!ifp) {
2294                 kvfree(fspath);
2295                 return ERR_PTR(-ENOMEM);
2296         }
2297
2298         ifp->btrfs_path = path;
2299         ifp->fspath = fspath;
2300         ifp->fs_root = fs_root;
2301
2302         return ifp;
2303 }
2304
2305 void free_ipath(struct inode_fs_paths *ipath)
2306 {
2307         if (!ipath)
2308                 return;
2309         kvfree(ipath->fspath);
2310         kfree(ipath);
2311 }
2312
2313 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2314                 struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2315 {
2316         struct btrfs_backref_iter *ret;
2317
2318         ret = kzalloc(sizeof(*ret), gfp_flag);
2319         if (!ret)
2320                 return NULL;
2321
2322         ret->path = btrfs_alloc_path();
2323         if (!ret->path) {
2324                 kfree(ret);
2325                 return NULL;
2326         }
2327
2328         /* Current backref iterator only supports iteration in commit root */
2329         ret->path->search_commit_root = 1;
2330         ret->path->skip_locking = 1;
2331         ret->fs_info = fs_info;
2332
2333         return ret;
2334 }
2335
2336 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2337 {
2338         struct btrfs_fs_info *fs_info = iter->fs_info;
2339         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2340         struct btrfs_path *path = iter->path;
2341         struct btrfs_extent_item *ei;
2342         struct btrfs_key key;
2343         int ret;
2344
2345         key.objectid = bytenr;
2346         key.type = BTRFS_METADATA_ITEM_KEY;
2347         key.offset = (u64)-1;
2348         iter->bytenr = bytenr;
2349
2350         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2351         if (ret < 0)
2352                 return ret;
2353         if (ret == 0) {
2354                 ret = -EUCLEAN;
2355                 goto release;
2356         }
2357         if (path->slots[0] == 0) {
2358                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2359                 ret = -EUCLEAN;
2360                 goto release;
2361         }
2362         path->slots[0]--;
2363
2364         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2365         if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2366              key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2367                 ret = -ENOENT;
2368                 goto release;
2369         }
2370         memcpy(&iter->cur_key, &key, sizeof(key));
2371         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2372                                                     path->slots[0]);
2373         iter->end_ptr = (u32)(iter->item_ptr +
2374                         btrfs_item_size(path->nodes[0], path->slots[0]));
2375         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2376                             struct btrfs_extent_item);
2377
2378         /*
2379          * Only support iteration on tree backref yet.
2380          *
2381          * This is an extra precaution for non skinny-metadata, where
2382          * EXTENT_ITEM is also used for tree blocks, that we can only use
2383          * extent flags to determine if it's a tree block.
2384          */
2385         if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2386                 ret = -ENOTSUPP;
2387                 goto release;
2388         }
2389         iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2390
2391         /* If there is no inline backref, go search for keyed backref */
2392         if (iter->cur_ptr >= iter->end_ptr) {
2393                 ret = btrfs_next_item(extent_root, path);
2394
2395                 /* No inline nor keyed ref */
2396                 if (ret > 0) {
2397                         ret = -ENOENT;
2398                         goto release;
2399                 }
2400                 if (ret < 0)
2401                         goto release;
2402
2403                 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2404                                 path->slots[0]);
2405                 if (iter->cur_key.objectid != bytenr ||
2406                     (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2407                      iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2408                         ret = -ENOENT;
2409                         goto release;
2410                 }
2411                 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2412                                                            path->slots[0]);
2413                 iter->item_ptr = iter->cur_ptr;
2414                 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2415                                       path->nodes[0], path->slots[0]));
2416         }
2417
2418         return 0;
2419 release:
2420         btrfs_backref_iter_release(iter);
2421         return ret;
2422 }
2423
2424 /*
2425  * Go to the next backref item of current bytenr, can be either inlined or
2426  * keyed.
2427  *
2428  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2429  *
2430  * Return 0 if we get next backref without problem.
2431  * Return >0 if there is no extra backref for this bytenr.
2432  * Return <0 if there is something wrong happened.
2433  */
2434 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2435 {
2436         struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2437         struct btrfs_root *extent_root;
2438         struct btrfs_path *path = iter->path;
2439         struct btrfs_extent_inline_ref *iref;
2440         int ret;
2441         u32 size;
2442
2443         if (btrfs_backref_iter_is_inline_ref(iter)) {
2444                 /* We're still inside the inline refs */
2445                 ASSERT(iter->cur_ptr < iter->end_ptr);
2446
2447                 if (btrfs_backref_has_tree_block_info(iter)) {
2448                         /* First tree block info */
2449                         size = sizeof(struct btrfs_tree_block_info);
2450                 } else {
2451                         /* Use inline ref type to determine the size */
2452                         int type;
2453
2454                         iref = (struct btrfs_extent_inline_ref *)
2455                                 ((unsigned long)iter->cur_ptr);
2456                         type = btrfs_extent_inline_ref_type(eb, iref);
2457
2458                         size = btrfs_extent_inline_ref_size(type);
2459                 }
2460                 iter->cur_ptr += size;
2461                 if (iter->cur_ptr < iter->end_ptr)
2462                         return 0;
2463
2464                 /* All inline items iterated, fall through */
2465         }
2466
2467         /* We're at keyed items, there is no inline item, go to the next one */
2468         extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2469         ret = btrfs_next_item(extent_root, iter->path);
2470         if (ret)
2471                 return ret;
2472
2473         btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2474         if (iter->cur_key.objectid != iter->bytenr ||
2475             (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2476              iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2477                 return 1;
2478         iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2479                                         path->slots[0]);
2480         iter->cur_ptr = iter->item_ptr;
2481         iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
2482                                                 path->slots[0]);
2483         return 0;
2484 }
2485
2486 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2487                               struct btrfs_backref_cache *cache, int is_reloc)
2488 {
2489         int i;
2490
2491         cache->rb_root = RB_ROOT;
2492         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2493                 INIT_LIST_HEAD(&cache->pending[i]);
2494         INIT_LIST_HEAD(&cache->changed);
2495         INIT_LIST_HEAD(&cache->detached);
2496         INIT_LIST_HEAD(&cache->leaves);
2497         INIT_LIST_HEAD(&cache->pending_edge);
2498         INIT_LIST_HEAD(&cache->useless_node);
2499         cache->fs_info = fs_info;
2500         cache->is_reloc = is_reloc;
2501 }
2502
2503 struct btrfs_backref_node *btrfs_backref_alloc_node(
2504                 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2505 {
2506         struct btrfs_backref_node *node;
2507
2508         ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2509         node = kzalloc(sizeof(*node), GFP_NOFS);
2510         if (!node)
2511                 return node;
2512
2513         INIT_LIST_HEAD(&node->list);
2514         INIT_LIST_HEAD(&node->upper);
2515         INIT_LIST_HEAD(&node->lower);
2516         RB_CLEAR_NODE(&node->rb_node);
2517         cache->nr_nodes++;
2518         node->level = level;
2519         node->bytenr = bytenr;
2520
2521         return node;
2522 }
2523
2524 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2525                 struct btrfs_backref_cache *cache)
2526 {
2527         struct btrfs_backref_edge *edge;
2528
2529         edge = kzalloc(sizeof(*edge), GFP_NOFS);
2530         if (edge)
2531                 cache->nr_edges++;
2532         return edge;
2533 }
2534
2535 /*
2536  * Drop the backref node from cache, also cleaning up all its
2537  * upper edges and any uncached nodes in the path.
2538  *
2539  * This cleanup happens bottom up, thus the node should either
2540  * be the lowest node in the cache or a detached node.
2541  */
2542 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2543                                 struct btrfs_backref_node *node)
2544 {
2545         struct btrfs_backref_node *upper;
2546         struct btrfs_backref_edge *edge;
2547
2548         if (!node)
2549                 return;
2550
2551         BUG_ON(!node->lowest && !node->detached);
2552         while (!list_empty(&node->upper)) {
2553                 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2554                                   list[LOWER]);
2555                 upper = edge->node[UPPER];
2556                 list_del(&edge->list[LOWER]);
2557                 list_del(&edge->list[UPPER]);
2558                 btrfs_backref_free_edge(cache, edge);
2559
2560                 /*
2561                  * Add the node to leaf node list if no other child block
2562                  * cached.
2563                  */
2564                 if (list_empty(&upper->lower)) {
2565                         list_add_tail(&upper->lower, &cache->leaves);
2566                         upper->lowest = 1;
2567                 }
2568         }
2569
2570         btrfs_backref_drop_node(cache, node);
2571 }
2572
2573 /*
2574  * Release all nodes/edges from current cache
2575  */
2576 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2577 {
2578         struct btrfs_backref_node *node;
2579         int i;
2580
2581         while (!list_empty(&cache->detached)) {
2582                 node = list_entry(cache->detached.next,
2583                                   struct btrfs_backref_node, list);
2584                 btrfs_backref_cleanup_node(cache, node);
2585         }
2586
2587         while (!list_empty(&cache->leaves)) {
2588                 node = list_entry(cache->leaves.next,
2589                                   struct btrfs_backref_node, lower);
2590                 btrfs_backref_cleanup_node(cache, node);
2591         }
2592
2593         cache->last_trans = 0;
2594
2595         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2596                 ASSERT(list_empty(&cache->pending[i]));
2597         ASSERT(list_empty(&cache->pending_edge));
2598         ASSERT(list_empty(&cache->useless_node));
2599         ASSERT(list_empty(&cache->changed));
2600         ASSERT(list_empty(&cache->detached));
2601         ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2602         ASSERT(!cache->nr_nodes);
2603         ASSERT(!cache->nr_edges);
2604 }
2605
2606 /*
2607  * Handle direct tree backref
2608  *
2609  * Direct tree backref means, the backref item shows its parent bytenr
2610  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2611  *
2612  * @ref_key:    The converted backref key.
2613  *              For keyed backref, it's the item key.
2614  *              For inlined backref, objectid is the bytenr,
2615  *              type is btrfs_inline_ref_type, offset is
2616  *              btrfs_inline_ref_offset.
2617  */
2618 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2619                                       struct btrfs_key *ref_key,
2620                                       struct btrfs_backref_node *cur)
2621 {
2622         struct btrfs_backref_edge *edge;
2623         struct btrfs_backref_node *upper;
2624         struct rb_node *rb_node;
2625
2626         ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2627
2628         /* Only reloc root uses backref pointing to itself */
2629         if (ref_key->objectid == ref_key->offset) {
2630                 struct btrfs_root *root;
2631
2632                 cur->is_reloc_root = 1;
2633                 /* Only reloc backref cache cares about a specific root */
2634                 if (cache->is_reloc) {
2635                         root = find_reloc_root(cache->fs_info, cur->bytenr);
2636                         if (!root)
2637                                 return -ENOENT;
2638                         cur->root = root;
2639                 } else {
2640                         /*
2641                          * For generic purpose backref cache, reloc root node
2642                          * is useless.
2643                          */
2644                         list_add(&cur->list, &cache->useless_node);
2645                 }
2646                 return 0;
2647         }
2648
2649         edge = btrfs_backref_alloc_edge(cache);
2650         if (!edge)
2651                 return -ENOMEM;
2652
2653         rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2654         if (!rb_node) {
2655                 /* Parent node not yet cached */
2656                 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2657                                            cur->level + 1);
2658                 if (!upper) {
2659                         btrfs_backref_free_edge(cache, edge);
2660                         return -ENOMEM;
2661                 }
2662
2663                 /*
2664                  *  Backrefs for the upper level block isn't cached, add the
2665                  *  block to pending list
2666                  */
2667                 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2668         } else {
2669                 /* Parent node already cached */
2670                 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2671                 ASSERT(upper->checked);
2672                 INIT_LIST_HEAD(&edge->list[UPPER]);
2673         }
2674         btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2675         return 0;
2676 }
2677
2678 /*
2679  * Handle indirect tree backref
2680  *
2681  * Indirect tree backref means, we only know which tree the node belongs to.
2682  * We still need to do a tree search to find out the parents. This is for
2683  * TREE_BLOCK_REF backref (keyed or inlined).
2684  *
2685  * @ref_key:    The same as @ref_key in  handle_direct_tree_backref()
2686  * @tree_key:   The first key of this tree block.
2687  * @path:       A clean (released) path, to avoid allocating path every time
2688  *              the function get called.
2689  */
2690 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2691                                         struct btrfs_path *path,
2692                                         struct btrfs_key *ref_key,
2693                                         struct btrfs_key *tree_key,
2694                                         struct btrfs_backref_node *cur)
2695 {
2696         struct btrfs_fs_info *fs_info = cache->fs_info;
2697         struct btrfs_backref_node *upper;
2698         struct btrfs_backref_node *lower;
2699         struct btrfs_backref_edge *edge;
2700         struct extent_buffer *eb;
2701         struct btrfs_root *root;
2702         struct rb_node *rb_node;
2703         int level;
2704         bool need_check = true;
2705         int ret;
2706
2707         root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2708         if (IS_ERR(root))
2709                 return PTR_ERR(root);
2710         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2711                 cur->cowonly = 1;
2712
2713         if (btrfs_root_level(&root->root_item) == cur->level) {
2714                 /* Tree root */
2715                 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2716                 /*
2717                  * For reloc backref cache, we may ignore reloc root.  But for
2718                  * general purpose backref cache, we can't rely on
2719                  * btrfs_should_ignore_reloc_root() as it may conflict with
2720                  * current running relocation and lead to missing root.
2721                  *
2722                  * For general purpose backref cache, reloc root detection is
2723                  * completely relying on direct backref (key->offset is parent
2724                  * bytenr), thus only do such check for reloc cache.
2725                  */
2726                 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2727                         btrfs_put_root(root);
2728                         list_add(&cur->list, &cache->useless_node);
2729                 } else {
2730                         cur->root = root;
2731                 }
2732                 return 0;
2733         }
2734
2735         level = cur->level + 1;
2736
2737         /* Search the tree to find parent blocks referring to the block */
2738         path->search_commit_root = 1;
2739         path->skip_locking = 1;
2740         path->lowest_level = level;
2741         ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2742         path->lowest_level = 0;
2743         if (ret < 0) {
2744                 btrfs_put_root(root);
2745                 return ret;
2746         }
2747         if (ret > 0 && path->slots[level] > 0)
2748                 path->slots[level]--;
2749
2750         eb = path->nodes[level];
2751         if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2752                 btrfs_err(fs_info,
2753 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2754                           cur->bytenr, level - 1, root->root_key.objectid,
2755                           tree_key->objectid, tree_key->type, tree_key->offset);
2756                 btrfs_put_root(root);
2757                 ret = -ENOENT;
2758                 goto out;
2759         }
2760         lower = cur;
2761
2762         /* Add all nodes and edges in the path */
2763         for (; level < BTRFS_MAX_LEVEL; level++) {
2764                 if (!path->nodes[level]) {
2765                         ASSERT(btrfs_root_bytenr(&root->root_item) ==
2766                                lower->bytenr);
2767                         /* Same as previous should_ignore_reloc_root() call */
2768                         if (btrfs_should_ignore_reloc_root(root) &&
2769                             cache->is_reloc) {
2770                                 btrfs_put_root(root);
2771                                 list_add(&lower->list, &cache->useless_node);
2772                         } else {
2773                                 lower->root = root;
2774                         }
2775                         break;
2776                 }
2777
2778                 edge = btrfs_backref_alloc_edge(cache);
2779                 if (!edge) {
2780                         btrfs_put_root(root);
2781                         ret = -ENOMEM;
2782                         goto out;
2783                 }
2784
2785                 eb = path->nodes[level];
2786                 rb_node = rb_simple_search(&cache->rb_root, eb->start);
2787                 if (!rb_node) {
2788                         upper = btrfs_backref_alloc_node(cache, eb->start,
2789                                                          lower->level + 1);
2790                         if (!upper) {
2791                                 btrfs_put_root(root);
2792                                 btrfs_backref_free_edge(cache, edge);
2793                                 ret = -ENOMEM;
2794                                 goto out;
2795                         }
2796                         upper->owner = btrfs_header_owner(eb);
2797                         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2798                                 upper->cowonly = 1;
2799
2800                         /*
2801                          * If we know the block isn't shared we can avoid
2802                          * checking its backrefs.
2803                          */
2804                         if (btrfs_block_can_be_shared(root, eb))
2805                                 upper->checked = 0;
2806                         else
2807                                 upper->checked = 1;
2808
2809                         /*
2810                          * Add the block to pending list if we need to check its
2811                          * backrefs, we only do this once while walking up a
2812                          * tree as we will catch anything else later on.
2813                          */
2814                         if (!upper->checked && need_check) {
2815                                 need_check = false;
2816                                 list_add_tail(&edge->list[UPPER],
2817                                               &cache->pending_edge);
2818                         } else {
2819                                 if (upper->checked)
2820                                         need_check = true;
2821                                 INIT_LIST_HEAD(&edge->list[UPPER]);
2822                         }
2823                 } else {
2824                         upper = rb_entry(rb_node, struct btrfs_backref_node,
2825                                          rb_node);
2826                         ASSERT(upper->checked);
2827                         INIT_LIST_HEAD(&edge->list[UPPER]);
2828                         if (!upper->owner)
2829                                 upper->owner = btrfs_header_owner(eb);
2830                 }
2831                 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2832
2833                 if (rb_node) {
2834                         btrfs_put_root(root);
2835                         break;
2836                 }
2837                 lower = upper;
2838                 upper = NULL;
2839         }
2840 out:
2841         btrfs_release_path(path);
2842         return ret;
2843 }
2844
2845 /*
2846  * Add backref node @cur into @cache.
2847  *
2848  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2849  *       links aren't yet bi-directional. Needs to finish such links.
2850  *       Use btrfs_backref_finish_upper_links() to finish such linkage.
2851  *
2852  * @path:       Released path for indirect tree backref lookup
2853  * @iter:       Released backref iter for extent tree search
2854  * @node_key:   The first key of the tree block
2855  */
2856 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2857                                 struct btrfs_path *path,
2858                                 struct btrfs_backref_iter *iter,
2859                                 struct btrfs_key *node_key,
2860                                 struct btrfs_backref_node *cur)
2861 {
2862         struct btrfs_fs_info *fs_info = cache->fs_info;
2863         struct btrfs_backref_edge *edge;
2864         struct btrfs_backref_node *exist;
2865         int ret;
2866
2867         ret = btrfs_backref_iter_start(iter, cur->bytenr);
2868         if (ret < 0)
2869                 return ret;
2870         /*
2871          * We skip the first btrfs_tree_block_info, as we don't use the key
2872          * stored in it, but fetch it from the tree block
2873          */
2874         if (btrfs_backref_has_tree_block_info(iter)) {
2875                 ret = btrfs_backref_iter_next(iter);
2876                 if (ret < 0)
2877                         goto out;
2878                 /* No extra backref? This means the tree block is corrupted */
2879                 if (ret > 0) {
2880                         ret = -EUCLEAN;
2881                         goto out;
2882                 }
2883         }
2884         WARN_ON(cur->checked);
2885         if (!list_empty(&cur->upper)) {
2886                 /*
2887                  * The backref was added previously when processing backref of
2888                  * type BTRFS_TREE_BLOCK_REF_KEY
2889                  */
2890                 ASSERT(list_is_singular(&cur->upper));
2891                 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2892                                   list[LOWER]);
2893                 ASSERT(list_empty(&edge->list[UPPER]));
2894                 exist = edge->node[UPPER];
2895                 /*
2896                  * Add the upper level block to pending list if we need check
2897                  * its backrefs
2898                  */
2899                 if (!exist->checked)
2900                         list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2901         } else {
2902                 exist = NULL;
2903         }
2904
2905         for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2906                 struct extent_buffer *eb;
2907                 struct btrfs_key key;
2908                 int type;
2909
2910                 cond_resched();
2911                 eb = btrfs_backref_get_eb(iter);
2912
2913                 key.objectid = iter->bytenr;
2914                 if (btrfs_backref_iter_is_inline_ref(iter)) {
2915                         struct btrfs_extent_inline_ref *iref;
2916
2917                         /* Update key for inline backref */
2918                         iref = (struct btrfs_extent_inline_ref *)
2919                                 ((unsigned long)iter->cur_ptr);
2920                         type = btrfs_get_extent_inline_ref_type(eb, iref,
2921                                                         BTRFS_REF_TYPE_BLOCK);
2922                         if (type == BTRFS_REF_TYPE_INVALID) {
2923                                 ret = -EUCLEAN;
2924                                 goto out;
2925                         }
2926                         key.type = type;
2927                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2928                 } else {
2929                         key.type = iter->cur_key.type;
2930                         key.offset = iter->cur_key.offset;
2931                 }
2932
2933                 /*
2934                  * Parent node found and matches current inline ref, no need to
2935                  * rebuild this node for this inline ref
2936                  */
2937                 if (exist &&
2938                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2939                       exist->owner == key.offset) ||
2940                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2941                       exist->bytenr == key.offset))) {
2942                         exist = NULL;
2943                         continue;
2944                 }
2945
2946                 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2947                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2948                         ret = handle_direct_tree_backref(cache, &key, cur);
2949                         if (ret < 0)
2950                                 goto out;
2951                         continue;
2952                 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2953                         ret = -EINVAL;
2954                         btrfs_print_v0_err(fs_info);
2955                         btrfs_handle_fs_error(fs_info, ret, NULL);
2956                         goto out;
2957                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2958                         continue;
2959                 }
2960
2961                 /*
2962                  * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2963                  * means the root objectid. We need to search the tree to get
2964                  * its parent bytenr.
2965                  */
2966                 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2967                                                    cur);
2968                 if (ret < 0)
2969                         goto out;
2970         }
2971         ret = 0;
2972         cur->checked = 1;
2973         WARN_ON(exist);
2974 out:
2975         btrfs_backref_iter_release(iter);
2976         return ret;
2977 }
2978
2979 /*
2980  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
2981  */
2982 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
2983                                      struct btrfs_backref_node *start)
2984 {
2985         struct list_head *useless_node = &cache->useless_node;
2986         struct btrfs_backref_edge *edge;
2987         struct rb_node *rb_node;
2988         LIST_HEAD(pending_edge);
2989
2990         ASSERT(start->checked);
2991
2992         /* Insert this node to cache if it's not COW-only */
2993         if (!start->cowonly) {
2994                 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
2995                                            &start->rb_node);
2996                 if (rb_node)
2997                         btrfs_backref_panic(cache->fs_info, start->bytenr,
2998                                             -EEXIST);
2999                 list_add_tail(&start->lower, &cache->leaves);
3000         }
3001
3002         /*
3003          * Use breadth first search to iterate all related edges.
3004          *
3005          * The starting points are all the edges of this node
3006          */
3007         list_for_each_entry(edge, &start->upper, list[LOWER])
3008                 list_add_tail(&edge->list[UPPER], &pending_edge);
3009
3010         while (!list_empty(&pending_edge)) {
3011                 struct btrfs_backref_node *upper;
3012                 struct btrfs_backref_node *lower;
3013
3014                 edge = list_first_entry(&pending_edge,
3015                                 struct btrfs_backref_edge, list[UPPER]);
3016                 list_del_init(&edge->list[UPPER]);
3017                 upper = edge->node[UPPER];
3018                 lower = edge->node[LOWER];
3019
3020                 /* Parent is detached, no need to keep any edges */
3021                 if (upper->detached) {
3022                         list_del(&edge->list[LOWER]);
3023                         btrfs_backref_free_edge(cache, edge);
3024
3025                         /* Lower node is orphan, queue for cleanup */
3026                         if (list_empty(&lower->upper))
3027                                 list_add(&lower->list, useless_node);
3028                         continue;
3029                 }
3030
3031                 /*
3032                  * All new nodes added in current build_backref_tree() haven't
3033                  * been linked to the cache rb tree.
3034                  * So if we have upper->rb_node populated, this means a cache
3035                  * hit. We only need to link the edge, as @upper and all its
3036                  * parents have already been linked.
3037                  */
3038                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3039                         if (upper->lowest) {
3040                                 list_del_init(&upper->lower);
3041                                 upper->lowest = 0;
3042                         }
3043
3044                         list_add_tail(&edge->list[UPPER], &upper->lower);
3045                         continue;
3046                 }
3047
3048                 /* Sanity check, we shouldn't have any unchecked nodes */
3049                 if (!upper->checked) {
3050                         ASSERT(0);
3051                         return -EUCLEAN;
3052                 }
3053
3054                 /* Sanity check, COW-only node has non-COW-only parent */
3055                 if (start->cowonly != upper->cowonly) {
3056                         ASSERT(0);
3057                         return -EUCLEAN;
3058                 }
3059
3060                 /* Only cache non-COW-only (subvolume trees) tree blocks */
3061                 if (!upper->cowonly) {
3062                         rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3063                                                    &upper->rb_node);
3064                         if (rb_node) {
3065                                 btrfs_backref_panic(cache->fs_info,
3066                                                 upper->bytenr, -EEXIST);
3067                                 return -EUCLEAN;
3068                         }
3069                 }
3070
3071                 list_add_tail(&edge->list[UPPER], &upper->lower);
3072
3073                 /*
3074                  * Also queue all the parent edges of this uncached node
3075                  * to finish the upper linkage
3076                  */
3077                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3078                         list_add_tail(&edge->list[UPPER], &pending_edge);
3079         }
3080         return 0;
3081 }
3082
3083 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3084                                  struct btrfs_backref_node *node)
3085 {
3086         struct btrfs_backref_node *lower;
3087         struct btrfs_backref_node *upper;
3088         struct btrfs_backref_edge *edge;
3089
3090         while (!list_empty(&cache->useless_node)) {
3091                 lower = list_first_entry(&cache->useless_node,
3092                                    struct btrfs_backref_node, list);
3093                 list_del_init(&lower->list);
3094         }
3095         while (!list_empty(&cache->pending_edge)) {
3096                 edge = list_first_entry(&cache->pending_edge,
3097                                 struct btrfs_backref_edge, list[UPPER]);
3098                 list_del(&edge->list[UPPER]);
3099                 list_del(&edge->list[LOWER]);
3100                 lower = edge->node[LOWER];
3101                 upper = edge->node[UPPER];
3102                 btrfs_backref_free_edge(cache, edge);
3103
3104                 /*
3105                  * Lower is no longer linked to any upper backref nodes and
3106                  * isn't in the cache, we can free it ourselves.
3107                  */
3108                 if (list_empty(&lower->upper) &&
3109                     RB_EMPTY_NODE(&lower->rb_node))
3110                         list_add(&lower->list, &cache->useless_node);
3111
3112                 if (!RB_EMPTY_NODE(&upper->rb_node))
3113                         continue;
3114
3115                 /* Add this guy's upper edges to the list to process */
3116                 list_for_each_entry(edge, &upper->upper, list[LOWER])
3117                         list_add_tail(&edge->list[UPPER],
3118                                       &cache->pending_edge);
3119                 if (list_empty(&upper->upper))
3120                         list_add(&upper->list, &cache->useless_node);
3121         }
3122
3123         while (!list_empty(&cache->useless_node)) {
3124                 lower = list_first_entry(&cache->useless_node,
3125                                    struct btrfs_backref_node, list);
3126                 list_del_init(&lower->list);
3127                 if (lower == node)
3128                         node = NULL;
3129                 btrfs_backref_drop_node(cache, lower);
3130         }
3131
3132         btrfs_backref_cleanup_node(cache, node);
3133         ASSERT(list_empty(&cache->useless_node) &&
3134                list_empty(&cache->pending_edge));
3135 }