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
  17 /* Just an arbitrary number so we can be sure this happened */
  18 #define BACKREF_FOUND_SHARED 6
  19
  20 struct extent_inode_elem {
  21         u64 inum;
  22         u64 offset;
  23         struct extent_inode_elem *next;
  24 };
  25
  26 static int check_extent_in_eb(const struct btrfs_key *key,
  27                               const struct extent_buffer *eb,
  28                               const struct btrfs_file_extent_item *fi,
  29                               u64 extent_item_pos,
  30                               struct extent_inode_elem **eie,
  31                               bool ignore_offset)
  32 {
  33         u64 offset = 0;
  34         struct extent_inode_elem *e;
  35
  36         if (!ignore_offset &&
  37             !btrfs_file_extent_compression(eb, fi) &&
  38             !btrfs_file_extent_encryption(eb, fi) &&
  39             !btrfs_file_extent_other_encoding(eb, fi)) {
  40                 u64 data_offset;
  41                 u64 data_len;
  42
  43                 data_offset = btrfs_file_extent_offset(eb, fi);
  44                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  45
  46                 if (extent_item_pos < data_offset ||
  47                     extent_item_pos >= data_offset + data_len)
  48                         return 1;
  49                 offset = extent_item_pos - data_offset;
  50         }
  51
  52         e = kmalloc(sizeof(*e), GFP_NOFS);
  53         if (!e)
  54                 return -ENOMEM;
  55
  56         e->next = *eie;
  57         e->inum = key->objectid;
  58         e->offset = key->offset + offset;
  59         *eie = e;
  60
  61         return 0;
  62 }
  63
  64 static void free_inode_elem_list(struct extent_inode_elem *eie)
  65 {
  66         struct extent_inode_elem *eie_next;
  67
  68         for (; eie; eie = eie_next) {
  69                 eie_next = eie->next;
  70                 kfree(eie);
  71         }
  72 }
  73
  74 static int find_extent_in_eb(const struct extent_buffer *eb,
  75                              u64 wanted_disk_byte, u64 extent_item_pos,
  76                              struct extent_inode_elem **eie,
  77                              bool ignore_offset)
  78 {
  79         u64 disk_byte;
  80         struct btrfs_key key;
  81         struct btrfs_file_extent_item *fi;
  82         int slot;
  83         int nritems;
  84         int extent_type;
  85         int ret;
  86
  87         /*
  88          * from the shared data ref, we only have the leaf but we need
  89          * the key. thus, we must look into all items and see that we
  90          * find one (some) with a reference to our extent item.
  91          */
  92         nritems = btrfs_header_nritems(eb);
  93         for (slot = 0; slot < nritems; ++slot) {
  94                 btrfs_item_key_to_cpu(eb, &key, slot);
  95                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  96                         continue;
  97                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  98                 extent_type = btrfs_file_extent_type(eb, fi);
  99                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 100                         continue;
 101                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 102                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 103                 if (disk_byte != wanted_disk_byte)
 104                         continue;
 105
 106                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
 107                 if (ret < 0)
 108                         return ret;
 109         }
 110
 111         return 0;
 112 }
 113
 114 struct preftree {
 115         struct rb_root_cached root;
 116         unsigned int count;
 117 };
 118
 119 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
 120
 121 struct preftrees {
 122         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 123         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 124         struct preftree indirect_missing_keys;
 125 };
 126
 127 /*
 128  * Checks for a shared extent during backref search.
 129  *
 130  * The share_count tracks prelim_refs (direct and indirect) having a
 131  * ref->count >0:
 132  *  - incremented when a ref->count transitions to >0
 133  *  - decremented when a ref->count transitions to <1
 134  */
 135 struct share_check {
 136         u64 root_objectid;
 137         u64 inum;
 138         int share_count;
 139 };
 140
 141 static inline int extent_is_shared(struct share_check *sc)
 142 {
 143         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 144 }
 145
 146 static struct kmem_cache *btrfs_prelim_ref_cache;
 147
 148 int __init btrfs_prelim_ref_init(void)
 149 {
 150         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 151                                         sizeof(struct prelim_ref),
 152                                         0,
 153                                         SLAB_MEM_SPREAD,
 154                                         NULL);
 155         if (!btrfs_prelim_ref_cache)
 156                 return -ENOMEM;
 157         return 0;
 158 }
 159
 160 void __cold btrfs_prelim_ref_exit(void)
 161 {
 162         kmem_cache_destroy(btrfs_prelim_ref_cache);
 163 }
 164
 165 static void free_pref(struct prelim_ref *ref)
 166 {
 167         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 168 }
 169
 170 /*
 171  * Return 0 when both refs are for the same block (and can be merged).
 172  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 173  * indicates a 'higher' block.
 174  */
 175 static int prelim_ref_compare(struct prelim_ref *ref1,
 176                               struct prelim_ref *ref2)
 177 {
 178         if (ref1->level < ref2->level)
 179                 return -1;
 180         if (ref1->level > ref2->level)
 181                 return 1;
 182         if (ref1->root_id < ref2->root_id)
 183                 return -1;
 184         if (ref1->root_id > ref2->root_id)
 185                 return 1;
 186         if (ref1->key_for_search.type < ref2->key_for_search.type)
 187                 return -1;
 188         if (ref1->key_for_search.type > ref2->key_for_search.type)
 189                 return 1;
 190         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 191                 return -1;
 192         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 193                 return 1;
 194         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 195                 return -1;
 196         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 197                 return 1;
 198         if (ref1->parent < ref2->parent)
 199                 return -1;
 200         if (ref1->parent > ref2->parent)
 201                 return 1;
 202
 203         return 0;
 204 }
 205
 206 static void update_share_count(struct share_check *sc, int oldcount,
 207                                int newcount)
 208 {
 209         if ((!sc) || (oldcount == 0 && newcount < 1))
 210                 return;
 211
 212         if (oldcount > 0 && newcount < 1)
 213                 sc->share_count--;
 214         else if (oldcount < 1 && newcount > 0)
 215                 sc->share_count++;
 216 }
 217
 218 /*
 219  * Add @newref to the @root rbtree, merging identical refs.
 220  *
 221  * Callers should assume that newref has been freed after calling.
 222  */
 223 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 224                               struct preftree *preftree,
 225                               struct prelim_ref *newref,
 226                               struct share_check *sc)
 227 {
 228         struct rb_root_cached *root;
 229         struct rb_node **p;
 230         struct rb_node *parent = NULL;
 231         struct prelim_ref *ref;
 232         int result;
 233         bool leftmost = true;
 234
 235         root = &preftree->root;
 236         p = &root->rb_root.rb_node;
 237
 238         while (*p) {
 239                 parent = *p;
 240                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 241                 result = prelim_ref_compare(ref, newref);
 242                 if (result < 0) {
 243                         p = &(*p)->rb_left;
 244                 } else if (result > 0) {
 245                         p = &(*p)->rb_right;
 246                         leftmost = false;
 247                 } else {
 248                         /* Identical refs, merge them and free @newref */
 249                         struct extent_inode_elem *eie = ref->inode_list;
 250
 251                         while (eie && eie->next)
 252                                 eie = eie->next;
 253
 254                         if (!eie)
 255                                 ref->inode_list = newref->inode_list;
 256                         else
 257                                 eie->next = newref->inode_list;
 258                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 259                                                      preftree->count);
 260                         /*
 261                          * A delayed ref can have newref->count < 0.
 262                          * The ref->count is updated to follow any
 263                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 264                          */
 265                         update_share_count(sc, ref->count,
 266                                            ref->count + newref->count);
 267                         ref->count += newref->count;
 268                         free_pref(newref);
 269                         return;
 270                 }
 271         }
 272
 273         update_share_count(sc, 0, newref->count);
 274         preftree->count++;
 275         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 276         rb_link_node(&newref->rbnode, parent, p);
 277         rb_insert_color_cached(&newref->rbnode, root, leftmost);
 278 }
 279
 280 /*
 281  * Release the entire tree.  We don't care about internal consistency so
 282  * just free everything and then reset the tree root.
 283  */
 284 static void prelim_release(struct preftree *preftree)
 285 {
 286         struct prelim_ref *ref, *next_ref;
 287
 288         rbtree_postorder_for_each_entry_safe(ref, next_ref,
 289                                              &preftree->root.rb_root, rbnode)
 290                 free_pref(ref);
 291
 292         preftree->root = RB_ROOT_CACHED;
 293         preftree->count = 0;
 294 }
 295
 296 /*
 297  * the rules for all callers of this function are:
 298  * - obtaining the parent is the goal
 299  * - if you add a key, you must know that it is a correct key
 300  * - if you cannot add the parent or a correct key, then we will look into the
 301  *   block later to set a correct key
 302  *
 303  * delayed refs
 304  * ============
 305  *        backref type | shared | indirect | shared | indirect
 306  * information         |   tree |     tree |   data |     data
 307  * --------------------+--------+----------+--------+----------
 308  *      parent logical |    y   |     -    |    -   |     -
 309  *      key to resolve |    -   |     y    |    y   |     y
 310  *  tree block logical |    -   |     -    |    -   |     -
 311  *  root for resolving |    y   |     y    |    y   |     y
 312  *
 313  * - column 1:       we've the parent -> done
 314  * - column 2, 3, 4: we use the key to find the parent
 315  *
 316  * on disk refs (inline or keyed)
 317  * ==============================
 318  *        backref type | shared | indirect | shared | indirect
 319  * information         |   tree |     tree |   data |     data
 320  * --------------------+--------+----------+--------+----------
 321  *      parent logical |    y   |     -    |    y   |     -
 322  *      key to resolve |    -   |     -    |    -   |     y
 323  *  tree block logical |    y   |     y    |    y   |     y
 324  *  root for resolving |    -   |     y    |    y   |     y
 325  *
 326  * - column 1, 3: we've the parent -> done
 327  * - column 2:    we take the first key from the block to find the parent
 328  *                (see add_missing_keys)
 329  * - column 4:    we use the key to find the parent
 330  *
 331  * additional information that's available but not required to find the parent
 332  * block might help in merging entries to gain some speed.
 333  */
 334 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 335                           struct preftree *preftree, u64 root_id,
 336                           const struct btrfs_key *key, int level, u64 parent,
 337                           u64 wanted_disk_byte, int count,
 338                           struct share_check *sc, gfp_t gfp_mask)
 339 {
 340         struct prelim_ref *ref;
 341
 342         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 343                 return 0;
 344
 345         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 346         if (!ref)
 347                 return -ENOMEM;
 348
 349         ref->root_id = root_id;
 350         if (key) {
 351                 ref->key_for_search = *key;
 352                 /*
 353                  * We can often find data backrefs with an offset that is too
 354                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 355                  * underflows when subtracting a file's offset with the data
 356                  * offset of its corresponding extent data item. This can
 357                  * happen for example in the clone ioctl.
 358                  * So if we detect such case we set the search key's offset to
 359                  * zero to make sure we will find the matching file extent item
 360                  * at add_all_parents(), otherwise we will miss it because the
 361                  * offset taken form the backref is much larger then the offset
 362                  * of the file extent item. This can make us scan a very large
 363                  * number of file extent items, but at least it will not make
 364                  * us miss any.
 365                  * This is an ugly workaround for a behaviour that should have
 366                  * never existed, but it does and a fix for the clone ioctl
 367                  * would touch a lot of places, cause backwards incompatibility
 368                  * and would not fix the problem for extents cloned with older
 369                  * kernels.
 370                  */
 371                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 372                     ref->key_for_search.offset >= LLONG_MAX)
 373                         ref->key_for_search.offset = 0;
 374         } else {
 375                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 376         }
 377
 378         ref->inode_list = NULL;
 379         ref->level = level;
 380         ref->count = count;
 381         ref->parent = parent;
 382         ref->wanted_disk_byte = wanted_disk_byte;
 383         prelim_ref_insert(fs_info, preftree, ref, sc);
 384         return extent_is_shared(sc);
 385 }
 386
 387 /* direct refs use root == 0, key == NULL */
 388 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 389                           struct preftrees *preftrees, int level, u64 parent,
 390                           u64 wanted_disk_byte, int count,
 391                           struct share_check *sc, gfp_t gfp_mask)
 392 {
 393         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 394                               parent, wanted_disk_byte, count, sc, gfp_mask);
 395 }
 396
 397 /* indirect refs use parent == 0 */
 398 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 399                             struct preftrees *preftrees, u64 root_id,
 400                             const struct btrfs_key *key, int level,
 401                             u64 wanted_disk_byte, int count,
 402                             struct share_check *sc, gfp_t gfp_mask)
 403 {
 404         struct preftree *tree = &preftrees->indirect;
 405
 406         if (!key)
 407                 tree = &preftrees->indirect_missing_keys;
 408         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 409                               wanted_disk_byte, count, sc, gfp_mask);
 410 }
 411
 412 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 413                            struct ulist *parents, struct prelim_ref *ref,
 414                            int level, u64 time_seq, const u64 *extent_item_pos,
 415                            u64 total_refs, bool ignore_offset)
 416 {
 417         int ret = 0;
 418         int slot;
 419         struct extent_buffer *eb;
 420         struct btrfs_key key;
 421         struct btrfs_key *key_for_search = &ref->key_for_search;
 422         struct btrfs_file_extent_item *fi;
 423         struct extent_inode_elem *eie = NULL, *old = NULL;
 424         u64 disk_byte;
 425         u64 wanted_disk_byte = ref->wanted_disk_byte;
 426         u64 count = 0;
 427
 428         if (level != 0) {
 429                 eb = path->nodes[level];
 430                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 431                 if (ret < 0)
 432                         return ret;
 433                 return 0;
 434         }
 435
 436         /*
 437          * We normally enter this function with the path already pointing to
 438          * the first item to check. But sometimes, we may enter it with
 439          * slot==nritems. In that case, go to the next leaf before we continue.
 440          */
 441         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 442                 if (time_seq == SEQ_LAST)
 443                         ret = btrfs_next_leaf(root, path);
 444                 else
 445                         ret = btrfs_next_old_leaf(root, path, time_seq);
 446         }
 447
 448         while (!ret && count < total_refs) {
 449                 eb = path->nodes[0];
 450                 slot = path->slots[0];
 451
 452                 btrfs_item_key_to_cpu(eb, &key, slot);
 453
 454                 if (key.objectid != key_for_search->objectid ||
 455                     key.type != BTRFS_EXTENT_DATA_KEY)
 456                         break;
 457
 458                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 459                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 460
 461                 if (disk_byte == wanted_disk_byte) {
 462                         eie = NULL;
 463                         old = NULL;
 464                         count++;
 465                         if (extent_item_pos) {
 466                                 ret = check_extent_in_eb(&key, eb, fi,
 467                                                 *extent_item_pos,
 468                                                 &eie, ignore_offset);
 469                                 if (ret < 0)
 470                                         break;
 471                         }
 472                         if (ret > 0)
 473                                 goto next;
 474                         ret = ulist_add_merge_ptr(parents, eb->start,
 475                                                   eie, (void **)&old, GFP_NOFS);
 476                         if (ret < 0)
 477                                 break;
 478                         if (!ret && extent_item_pos) {
 479                                 while (old->next)
 480                                         old = old->next;
 481                                 old->next = eie;
 482                         }
 483                         eie = NULL;
 484                 }
 485 next:
 486                 if (time_seq == SEQ_LAST)
 487                         ret = btrfs_next_item(root, path);
 488                 else
 489                         ret = btrfs_next_old_item(root, path, time_seq);
 490         }
 491
 492         if (ret > 0)
 493                 ret = 0;
 494         else if (ret < 0)
 495                 free_inode_elem_list(eie);
 496         return ret;
 497 }
 498
 499 /*
 500  * resolve an indirect backref in the form (root_id, key, level)
 501  * to a logical address
 502  */
 503 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 504                                 struct btrfs_path *path, u64 time_seq,
 505                                 struct prelim_ref *ref, struct ulist *parents,
 506                                 const u64 *extent_item_pos, u64 total_refs,
 507                                 bool ignore_offset)
 508 {
 509         struct btrfs_root *root;
 510         struct btrfs_key root_key;
 511         struct extent_buffer *eb;
 512         int ret = 0;
 513         int root_level;
 514         int level = ref->level;
 515         int index;
 516
 517         root_key.objectid = ref->root_id;
 518         root_key.type = BTRFS_ROOT_ITEM_KEY;
 519         root_key.offset = (u64)-1;
 520
 521         index = srcu_read_lock(&fs_info->subvol_srcu);
 522
 523         root = btrfs_get_fs_root(fs_info, &root_key, false);
 524         if (IS_ERR(root)) {
 525                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 526                 ret = PTR_ERR(root);
 527                 goto out;
 528         }
 529
 530         if (btrfs_is_testing(fs_info)) {
 531                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 532                 ret = -ENOENT;
 533                 goto out;
 534         }
 535
 536         if (path->search_commit_root)
 537                 root_level = btrfs_header_level(root->commit_root);
 538         else if (time_seq == SEQ_LAST)
 539                 root_level = btrfs_header_level(root->node);
 540         else
 541                 root_level = btrfs_old_root_level(root, time_seq);
 542
 543         if (root_level + 1 == level) {
 544                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 545                 goto out;
 546         }
 547
 548         path->lowest_level = level;
 549         if (time_seq == SEQ_LAST)
 550                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 551                                         0, 0);
 552         else
 553                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 554                                             time_seq);
 555
 556         /* root node has been locked, we can release @subvol_srcu safely here */
 557         srcu_read_unlock(&fs_info->subvol_srcu, index);
 558
 559         btrfs_debug(fs_info,
 560                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 561                  ref->root_id, level, ref->count, ret,
 562                  ref->key_for_search.objectid, ref->key_for_search.type,
 563                  ref->key_for_search.offset);
 564         if (ret < 0)
 565                 goto out;
 566
 567         eb = path->nodes[level];
 568         while (!eb) {
 569                 if (WARN_ON(!level)) {
 570                         ret = 1;
 571                         goto out;
 572                 }
 573                 level--;
 574                 eb = path->nodes[level];
 575         }
 576
 577         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 578                               extent_item_pos, total_refs, ignore_offset);
 579 out:
 580         path->lowest_level = 0;
 581         btrfs_release_path(path);
 582         return ret;
 583 }
 584
 585 static struct extent_inode_elem *
 586 unode_aux_to_inode_list(struct ulist_node *node)
 587 {
 588         if (!node)
 589                 return NULL;
 590         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 591 }
 592
 593 /*
 594  * We maintain three separate rbtrees: one for direct refs, one for
 595  * indirect refs which have a key, and one for indirect refs which do not
 596  * have a key. Each tree does merge on insertion.
 597  *
 598  * Once all of the references are located, we iterate over the tree of
 599  * indirect refs with missing keys. An appropriate key is located and
 600  * the ref is moved onto the tree for indirect refs. After all missing
 601  * keys are thus located, we iterate over the indirect ref tree, resolve
 602  * each reference, and then insert the resolved reference onto the
 603  * direct tree (merging there too).
 604  *
 605  * New backrefs (i.e., for parent nodes) are added to the appropriate
 606  * rbtree as they are encountered. The new backrefs are subsequently
 607  * resolved as above.
 608  */
 609 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 610                                  struct btrfs_path *path, u64 time_seq,
 611                                  struct preftrees *preftrees,
 612                                  const u64 *extent_item_pos, u64 total_refs,
 613                                  struct share_check *sc, bool ignore_offset)
 614 {
 615         int err;
 616         int ret = 0;
 617         struct ulist *parents;
 618         struct ulist_node *node;
 619         struct ulist_iterator uiter;
 620         struct rb_node *rnode;
 621
 622         parents = ulist_alloc(GFP_NOFS);
 623         if (!parents)
 624                 return -ENOMEM;
 625
 626         /*
 627          * We could trade memory usage for performance here by iterating
 628          * the tree, allocating new refs for each insertion, and then
 629          * freeing the entire indirect tree when we're done.  In some test
 630          * cases, the tree can grow quite large (~200k objects).
 631          */
 632         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
 633                 struct prelim_ref *ref;
 634
 635                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 636                 if (WARN(ref->parent,
 637                          "BUG: direct ref found in indirect tree")) {
 638                         ret = -EINVAL;
 639                         goto out;
 640                 }
 641
 642                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
 643                 preftrees->indirect.count--;
 644
 645                 if (ref->count == 0) {
 646                         free_pref(ref);
 647                         continue;
 648                 }
 649
 650                 if (sc && sc->root_objectid &&
 651                     ref->root_id != sc->root_objectid) {
 652                         free_pref(ref);
 653                         ret = BACKREF_FOUND_SHARED;
 654                         goto out;
 655                 }
 656                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 657                                            parents, extent_item_pos,
 658                                            total_refs, ignore_offset);
 659                 /*
 660                  * we can only tolerate ENOENT,otherwise,we should catch error
 661                  * and return directly.
 662                  */
 663                 if (err == -ENOENT) {
 664                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 665                                           NULL);
 666                         continue;
 667                 } else if (err) {
 668                         free_pref(ref);
 669                         ret = err;
 670                         goto out;
 671                 }
 672
 673                 /* we put the first parent into the ref at hand */
 674                 ULIST_ITER_INIT(&uiter);
 675                 node = ulist_next(parents, &uiter);
 676                 ref->parent = node ? node->val : 0;
 677                 ref->inode_list = unode_aux_to_inode_list(node);
 678
 679                 /* Add a prelim_ref(s) for any other parent(s). */
 680                 while ((node = ulist_next(parents, &uiter))) {
 681                         struct prelim_ref *new_ref;
 682
 683                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 684                                                    GFP_NOFS);
 685                         if (!new_ref) {
 686                                 free_pref(ref);
 687                                 ret = -ENOMEM;
 688                                 goto out;
 689                         }
 690                         memcpy(new_ref, ref, sizeof(*ref));
 691                         new_ref->parent = node->val;
 692                         new_ref->inode_list = unode_aux_to_inode_list(node);
 693                         prelim_ref_insert(fs_info, &preftrees->direct,
 694                                           new_ref, NULL);
 695                 }
 696
 697                 /*
 698                  * Now it's a direct ref, put it in the direct tree. We must
 699                  * do this last because the ref could be merged/freed here.
 700                  */
 701                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 702
 703                 ulist_reinit(parents);
 704                 cond_resched();
 705         }
 706 out:
 707         ulist_free(parents);
 708         return ret;
 709 }
 710
 711 /*
 712  * read tree blocks and add keys where required.
 713  */
 714 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 715                             struct preftrees *preftrees, bool lock)
 716 {
 717         struct prelim_ref *ref;
 718         struct extent_buffer *eb;
 719         struct preftree *tree = &preftrees->indirect_missing_keys;
 720         struct rb_node *node;
 721
 722         while ((node = rb_first_cached(&tree->root))) {
 723                 ref = rb_entry(node, struct prelim_ref, rbnode);
 724                 rb_erase_cached(node, &tree->root);
 725
 726                 BUG_ON(ref->parent);    /* should not be a direct ref */
 727                 BUG_ON(ref->key_for_search.type);
 728                 BUG_ON(!ref->wanted_disk_byte);
 729
 730                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
 731                                      ref->level - 1, NULL);
 732                 if (IS_ERR(eb)) {
 733                         free_pref(ref);
 734                         return PTR_ERR(eb);
 735                 } else if (!extent_buffer_uptodate(eb)) {
 736                         free_pref(ref);
 737                         free_extent_buffer(eb);
 738                         return -EIO;
 739                 }
 740                 if (lock)
 741                         btrfs_tree_read_lock(eb);
 742                 if (btrfs_header_level(eb) == 0)
 743                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 744                 else
 745                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 746                 if (lock)
 747                         btrfs_tree_read_unlock(eb);
 748                 free_extent_buffer(eb);
 749                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 750                 cond_resched();
 751         }
 752         return 0;
 753 }
 754
 755 /*
 756  * add all currently queued delayed refs from this head whose seq nr is
 757  * smaller or equal that seq to the list
 758  */
 759 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 760                             struct btrfs_delayed_ref_head *head, u64 seq,
 761                             struct preftrees *preftrees, u64 *total_refs,
 762                             struct share_check *sc)
 763 {
 764         struct btrfs_delayed_ref_node *node;
 765         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 766         struct btrfs_key key;
 767         struct btrfs_key tmp_op_key;
 768         struct rb_node *n;
 769         int count;
 770         int ret = 0;
 771
 772         if (extent_op && extent_op->update_key)
 773                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 774
 775         spin_lock(&head->lock);
 776         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
 777                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 778                                 ref_node);
 779                 if (node->seq > seq)
 780                         continue;
 781
 782                 switch (node->action) {
 783                 case BTRFS_ADD_DELAYED_EXTENT:
 784                 case BTRFS_UPDATE_DELAYED_HEAD:
 785                         WARN_ON(1);
 786                         continue;
 787                 case BTRFS_ADD_DELAYED_REF:
 788                         count = node->ref_mod;
 789                         break;
 790                 case BTRFS_DROP_DELAYED_REF:
 791                         count = node->ref_mod * -1;
 792                         break;
 793                 default:
 794                         BUG_ON(1);
 795                 }
 796                 *total_refs += count;
 797                 switch (node->type) {
 798                 case BTRFS_TREE_BLOCK_REF_KEY: {
 799                         /* NORMAL INDIRECT METADATA backref */
 800                         struct btrfs_delayed_tree_ref *ref;
 801
 802                         ref = btrfs_delayed_node_to_tree_ref(node);
 803                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 804                                                &tmp_op_key, ref->level + 1,
 805                                                node->bytenr, count, sc,
 806                                                GFP_ATOMIC);
 807                         break;
 808                 }
 809                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 810                         /* SHARED DIRECT METADATA backref */
 811                         struct btrfs_delayed_tree_ref *ref;
 812
 813                         ref = btrfs_delayed_node_to_tree_ref(node);
 814
 815                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 816                                              ref->parent, node->bytenr, count,
 817                                              sc, GFP_ATOMIC);
 818                         break;
 819                 }
 820                 case BTRFS_EXTENT_DATA_REF_KEY: {
 821                         /* NORMAL INDIRECT DATA backref */
 822                         struct btrfs_delayed_data_ref *ref;
 823                         ref = btrfs_delayed_node_to_data_ref(node);
 824
 825                         key.objectid = ref->objectid;
 826                         key.type = BTRFS_EXTENT_DATA_KEY;
 827                         key.offset = ref->offset;
 828
 829                         /*
 830                          * Found a inum that doesn't match our known inum, we
 831                          * know it's shared.
 832                          */
 833                         if (sc && sc->inum && ref->objectid != sc->inum) {
 834                                 ret = BACKREF_FOUND_SHARED;
 835                                 goto out;
 836                         }
 837
 838                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 839                                                &key, 0, node->bytenr, count, sc,
 840                                                GFP_ATOMIC);
 841                         break;
 842                 }
 843                 case BTRFS_SHARED_DATA_REF_KEY: {
 844                         /* SHARED DIRECT FULL backref */
 845                         struct btrfs_delayed_data_ref *ref;
 846
 847                         ref = btrfs_delayed_node_to_data_ref(node);
 848
 849                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 850                                              node->bytenr, count, sc,
 851                                              GFP_ATOMIC);
 852                         break;
 853                 }
 854                 default:
 855                         WARN_ON(1);
 856                 }
 857                 /*
 858                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 859                  * refs have been checked.
 860                  */
 861                 if (ret && (ret != BACKREF_FOUND_SHARED))
 862                         break;
 863         }
 864         if (!ret)
 865                 ret = extent_is_shared(sc);
 866 out:
 867         spin_unlock(&head->lock);
 868         return ret;
 869 }
 870
 871 /*
 872  * add all inline backrefs for bytenr to the list
 873  *
 874  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 875  */
 876 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 877                            struct btrfs_path *path, u64 bytenr,
 878                            int *info_level, struct preftrees *preftrees,
 879                            u64 *total_refs, struct share_check *sc)
 880 {
 881         int ret = 0;
 882         int slot;
 883         struct extent_buffer *leaf;
 884         struct btrfs_key key;
 885         struct btrfs_key found_key;
 886         unsigned long ptr;
 887         unsigned long end;
 888         struct btrfs_extent_item *ei;
 889         u64 flags;
 890         u64 item_size;
 891
 892         /*
 893          * enumerate all inline refs
 894          */
 895         leaf = path->nodes[0];
 896         slot = path->slots[0];
 897
 898         item_size = btrfs_item_size_nr(leaf, slot);
 899         BUG_ON(item_size < sizeof(*ei));
 900
 901         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 902         flags = btrfs_extent_flags(leaf, ei);
 903         *total_refs += btrfs_extent_refs(leaf, ei);
 904         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 905
 906         ptr = (unsigned long)(ei + 1);
 907         end = (unsigned long)ei + item_size;
 908
 909         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 910             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 911                 struct btrfs_tree_block_info *info;
 912
 913                 info = (struct btrfs_tree_block_info *)ptr;
 914                 *info_level = btrfs_tree_block_level(leaf, info);
 915                 ptr += sizeof(struct btrfs_tree_block_info);
 916                 BUG_ON(ptr > end);
 917         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 918                 *info_level = found_key.offset;
 919         } else {
 920                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 921         }
 922
 923         while (ptr < end) {
 924                 struct btrfs_extent_inline_ref *iref;
 925                 u64 offset;
 926                 int type;
 927
 928                 iref = (struct btrfs_extent_inline_ref *)ptr;
 929                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 930                                                         BTRFS_REF_TYPE_ANY);
 931                 if (type == BTRFS_REF_TYPE_INVALID)
 932                         return -EUCLEAN;
 933
 934                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 935
 936                 switch (type) {
 937                 case BTRFS_SHARED_BLOCK_REF_KEY:
 938                         ret = add_direct_ref(fs_info, preftrees,
 939                                              *info_level + 1, offset,
 940                                              bytenr, 1, NULL, GFP_NOFS);
 941                         break;
 942                 case BTRFS_SHARED_DATA_REF_KEY: {
 943                         struct btrfs_shared_data_ref *sdref;
 944                         int count;
 945
 946                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 947                         count = btrfs_shared_data_ref_count(leaf, sdref);
 948
 949                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 950                                              bytenr, count, sc, GFP_NOFS);
 951                         break;
 952                 }
 953                 case BTRFS_TREE_BLOCK_REF_KEY:
 954                         ret = add_indirect_ref(fs_info, preftrees, offset,
 955                                                NULL, *info_level + 1,
 956                                                bytenr, 1, NULL, GFP_NOFS);
 957                         break;
 958                 case BTRFS_EXTENT_DATA_REF_KEY: {
 959                         struct btrfs_extent_data_ref *dref;
 960                         int count;
 961                         u64 root;
 962
 963                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 964                         count = btrfs_extent_data_ref_count(leaf, dref);
 965                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 966                                                                       dref);
 967                         key.type = BTRFS_EXTENT_DATA_KEY;
 968                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 969
 970                         if (sc && sc->inum && key.objectid != sc->inum) {
 971                                 ret = BACKREF_FOUND_SHARED;
 972                                 break;
 973                         }
 974
 975                         root = btrfs_extent_data_ref_root(leaf, dref);
 976
 977                         ret = add_indirect_ref(fs_info, preftrees, root,
 978                                                &key, 0, bytenr, count,
 979                                                sc, GFP_NOFS);
 980                         break;
 981                 }
 982                 default:
 983                         WARN_ON(1);
 984                 }
 985                 if (ret)
 986                         return ret;
 987                 ptr += btrfs_extent_inline_ref_size(type);
 988         }
 989
 990         return 0;
 991 }
 992
 993 /*
 994  * add all non-inline backrefs for bytenr to the list
 995  *
 996  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 997  */
 998 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
 999                           struct btrfs_path *path, u64 bytenr,
1000                           int info_level, struct preftrees *preftrees,
1001                           struct share_check *sc)
1002 {
1003         struct btrfs_root *extent_root = fs_info->extent_root;
1004         int ret;
1005         int slot;
1006         struct extent_buffer *leaf;
1007         struct btrfs_key key;
1008
1009         while (1) {
1010                 ret = btrfs_next_item(extent_root, path);
1011                 if (ret < 0)
1012                         break;
1013                 if (ret) {
1014                         ret = 0;
1015                         break;
1016                 }
1017
1018                 slot = path->slots[0];
1019                 leaf = path->nodes[0];
1020                 btrfs_item_key_to_cpu(leaf, &key, slot);
1021
1022                 if (key.objectid != bytenr)
1023                         break;
1024                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1025                         continue;
1026                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1027                         break;
1028
1029                 switch (key.type) {
1030                 case BTRFS_SHARED_BLOCK_REF_KEY:
1031                         /* SHARED DIRECT METADATA backref */
1032                         ret = add_direct_ref(fs_info, preftrees,
1033                                              info_level + 1, key.offset,
1034                                              bytenr, 1, NULL, GFP_NOFS);
1035                         break;
1036                 case BTRFS_SHARED_DATA_REF_KEY: {
1037                         /* SHARED DIRECT FULL backref */
1038                         struct btrfs_shared_data_ref *sdref;
1039                         int count;
1040
1041                         sdref = btrfs_item_ptr(leaf, slot,
1042                                               struct btrfs_shared_data_ref);
1043                         count = btrfs_shared_data_ref_count(leaf, sdref);
1044                         ret = add_direct_ref(fs_info, preftrees, 0,
1045                                              key.offset, bytenr, count,
1046                                              sc, GFP_NOFS);
1047                         break;
1048                 }
1049                 case BTRFS_TREE_BLOCK_REF_KEY:
1050                         /* NORMAL INDIRECT METADATA backref */
1051                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1052                                                NULL, info_level + 1, bytenr,
1053                                                1, NULL, GFP_NOFS);
1054                         break;
1055                 case BTRFS_EXTENT_DATA_REF_KEY: {
1056                         /* NORMAL INDIRECT DATA backref */
1057                         struct btrfs_extent_data_ref *dref;
1058                         int count;
1059                         u64 root;
1060
1061                         dref = btrfs_item_ptr(leaf, slot,
1062                                               struct btrfs_extent_data_ref);
1063                         count = btrfs_extent_data_ref_count(leaf, dref);
1064                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1065                                                                       dref);
1066                         key.type = BTRFS_EXTENT_DATA_KEY;
1067                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1068
1069                         if (sc && sc->inum && key.objectid != sc->inum) {
1070                                 ret = BACKREF_FOUND_SHARED;
1071                                 break;
1072                         }
1073
1074                         root = btrfs_extent_data_ref_root(leaf, dref);
1075                         ret = add_indirect_ref(fs_info, preftrees, root,
1076                                                &key, 0, bytenr, count,
1077                                                sc, GFP_NOFS);
1078                         break;
1079                 }
1080                 default:
1081                         WARN_ON(1);
1082                 }
1083                 if (ret)
1084                         return ret;
1085
1086         }
1087
1088         return ret;
1089 }
1090
1091 /*
1092  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1093  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1094  * indirect refs to their parent bytenr.
1095  * When roots are found, they're added to the roots list
1096  *
1097  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1098  * much like trans == NULL case, the difference only lies in it will not
1099  * commit root.
1100  * The special case is for qgroup to search roots in commit_transaction().
1101  *
1102  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1103  * shared extent is detected.
1104  *
1105  * Otherwise this returns 0 for success and <0 for an error.
1106  *
1107  * If ignore_offset is set to false, only extent refs whose offsets match
1108  * extent_item_pos are returned.  If true, every extent ref is returned
1109  * and extent_item_pos is ignored.
1110  *
1111  * FIXME some caching might speed things up
1112  */
1113 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1114                              struct btrfs_fs_info *fs_info, u64 bytenr,
1115                              u64 time_seq, struct ulist *refs,
1116                              struct ulist *roots, const u64 *extent_item_pos,
1117                              struct share_check *sc, bool ignore_offset)
1118 {
1119         struct btrfs_key key;
1120         struct btrfs_path *path;
1121         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1122         struct btrfs_delayed_ref_head *head;
1123         int info_level = 0;
1124         int ret;
1125         struct prelim_ref *ref;
1126         struct rb_node *node;
1127         struct extent_inode_elem *eie = NULL;
1128         /* total of both direct AND indirect refs! */
1129         u64 total_refs = 0;
1130         struct preftrees preftrees = {
1131                 .direct = PREFTREE_INIT,
1132                 .indirect = PREFTREE_INIT,
1133                 .indirect_missing_keys = PREFTREE_INIT
1134         };
1135
1136         key.objectid = bytenr;
1137         key.offset = (u64)-1;
1138         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1139                 key.type = BTRFS_METADATA_ITEM_KEY;
1140         else
1141                 key.type = BTRFS_EXTENT_ITEM_KEY;
1142
1143         path = btrfs_alloc_path();
1144         if (!path)
1145                 return -ENOMEM;
1146         if (!trans) {
1147                 path->search_commit_root = 1;
1148                 path->skip_locking = 1;
1149         }
1150
1151         if (time_seq == SEQ_LAST)
1152                 path->skip_locking = 1;
1153
1154         /*
1155          * grab both a lock on the path and a lock on the delayed ref head.
1156          * We need both to get a consistent picture of how the refs look
1157          * at a specified point in time
1158          */
1159 again:
1160         head = NULL;
1161
1162         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1163         if (ret < 0)
1164                 goto out;
1165         BUG_ON(ret == 0);
1166
1167 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1168         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1169             time_seq != SEQ_LAST) {
1170 #else
1171         if (trans && time_seq != SEQ_LAST) {
1172 #endif
1173                 /*
1174                  * look if there are updates for this ref queued and lock the
1175                  * head
1176                  */
1177                 delayed_refs = &trans->transaction->delayed_refs;
1178                 spin_lock(&delayed_refs->lock);
1179                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1180                 if (head) {
1181                         if (!mutex_trylock(&head->mutex)) {
1182                                 refcount_inc(&head->refs);
1183                                 spin_unlock(&delayed_refs->lock);
1184
1185                                 btrfs_release_path(path);
1186
1187                                 /*
1188                                  * Mutex was contended, block until it's
1189                                  * released and try again
1190                                  */
1191                                 mutex_lock(&head->mutex);
1192                                 mutex_unlock(&head->mutex);
1193                                 btrfs_put_delayed_ref_head(head);
1194                                 goto again;
1195                         }
1196                         spin_unlock(&delayed_refs->lock);
1197                         ret = add_delayed_refs(fs_info, head, time_seq,
1198                                                &preftrees, &total_refs, sc);
1199                         mutex_unlock(&head->mutex);
1200                         if (ret)
1201                                 goto out;
1202                 } else {
1203                         spin_unlock(&delayed_refs->lock);
1204                 }
1205         }
1206
1207         if (path->slots[0]) {
1208                 struct extent_buffer *leaf;
1209                 int slot;
1210
1211                 path->slots[0]--;
1212                 leaf = path->nodes[0];
1213                 slot = path->slots[0];
1214                 btrfs_item_key_to_cpu(leaf, &key, slot);
1215                 if (key.objectid == bytenr &&
1216                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1217                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1218                         ret = add_inline_refs(fs_info, path, bytenr,
1219                                               &info_level, &preftrees,
1220                                               &total_refs, sc);
1221                         if (ret)
1222                                 goto out;
1223                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1224                                              &preftrees, sc);
1225                         if (ret)
1226                                 goto out;
1227                 }
1228         }
1229
1230         btrfs_release_path(path);
1231
1232         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1233         if (ret)
1234                 goto out;
1235
1236         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1237
1238         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1239                                     extent_item_pos, total_refs, sc, ignore_offset);
1240         if (ret)
1241                 goto out;
1242
1243         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1244
1245         /*
1246          * This walks the tree of merged and resolved refs. Tree blocks are
1247          * read in as needed. Unique entries are added to the ulist, and
1248          * the list of found roots is updated.
1249          *
1250          * We release the entire tree in one go before returning.
1251          */
1252         node = rb_first_cached(&preftrees.direct.root);
1253         while (node) {
1254                 ref = rb_entry(node, struct prelim_ref, rbnode);
1255                 node = rb_next(&ref->rbnode);
1256                 /*
1257                  * ref->count < 0 can happen here if there are delayed
1258                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1259                  * prelim_ref_insert() relies on this when merging
1260                  * identical refs to keep the overall count correct.
1261                  * prelim_ref_insert() will merge only those refs
1262                  * which compare identically.  Any refs having
1263                  * e.g. different offsets would not be merged,
1264                  * and would retain their original ref->count < 0.
1265                  */
1266                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1267                         if (sc && sc->root_objectid &&
1268                             ref->root_id != sc->root_objectid) {
1269                                 ret = BACKREF_FOUND_SHARED;
1270                                 goto out;
1271                         }
1272
1273                         /* no parent == root of tree */
1274                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1275                         if (ret < 0)
1276                                 goto out;
1277                 }
1278                 if (ref->count && ref->parent) {
1279                         if (extent_item_pos && !ref->inode_list &&
1280                             ref->level == 0) {
1281                                 struct extent_buffer *eb;
1282
1283                                 eb = read_tree_block(fs_info, ref->parent, 0,
1284                                                      ref->level, NULL);
1285                                 if (IS_ERR(eb)) {
1286                                         ret = PTR_ERR(eb);
1287                                         goto out;
1288                                 } else if (!extent_buffer_uptodate(eb)) {
1289                                         free_extent_buffer(eb);
1290                                         ret = -EIO;
1291                                         goto out;
1292                                 }
1293
1294                                 if (!path->skip_locking) {
1295                                         btrfs_tree_read_lock(eb);
1296                                         btrfs_set_lock_blocking_read(eb);
1297                                 }
1298                                 ret = find_extent_in_eb(eb, bytenr,
1299                                                         *extent_item_pos, &eie, ignore_offset);
1300                                 if (!path->skip_locking)
1301                                         btrfs_tree_read_unlock_blocking(eb);
1302                                 free_extent_buffer(eb);
1303                                 if (ret < 0)
1304                                         goto out;
1305                                 ref->inode_list = eie;
1306                         }
1307                         ret = ulist_add_merge_ptr(refs, ref->parent,
1308                                                   ref->inode_list,
1309                                                   (void **)&eie, GFP_NOFS);
1310                         if (ret < 0)
1311                                 goto out;
1312                         if (!ret && extent_item_pos) {
1313                                 /*
1314                                  * we've recorded that parent, so we must extend
1315                                  * its inode list here
1316                                  */
1317                                 BUG_ON(!eie);
1318                                 while (eie->next)
1319                                         eie = eie->next;
1320                                 eie->next = ref->inode_list;
1321                         }
1322                         eie = NULL;
1323                 }
1324                 cond_resched();
1325         }
1326
1327 out:
1328         btrfs_free_path(path);
1329
1330         prelim_release(&preftrees.direct);
1331         prelim_release(&preftrees.indirect);
1332         prelim_release(&preftrees.indirect_missing_keys);
1333
1334         if (ret < 0)
1335                 free_inode_elem_list(eie);
1336         return ret;
1337 }
1338
1339 static void free_leaf_list(struct ulist *blocks)
1340 {
1341         struct ulist_node *node = NULL;
1342         struct extent_inode_elem *eie;
1343         struct ulist_iterator uiter;
1344
1345         ULIST_ITER_INIT(&uiter);
1346         while ((node = ulist_next(blocks, &uiter))) {
1347                 if (!node->aux)
1348                         continue;
1349                 eie = unode_aux_to_inode_list(node);
1350                 free_inode_elem_list(eie);
1351                 node->aux = 0;
1352         }
1353
1354         ulist_free(blocks);
1355 }
1356
1357 /*
1358  * Finds all leafs with a reference to the specified combination of bytenr and
1359  * offset. key_list_head will point to a list of corresponding keys (caller must
1360  * free each list element). The leafs will be stored in the leafs ulist, which
1361  * must be freed with ulist_free.
1362  *
1363  * returns 0 on success, <0 on error
1364  */
1365 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1366                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1367                                 u64 time_seq, struct ulist **leafs,
1368                                 const u64 *extent_item_pos, bool ignore_offset)
1369 {
1370         int ret;
1371
1372         *leafs = ulist_alloc(GFP_NOFS);
1373         if (!*leafs)
1374                 return -ENOMEM;
1375
1376         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1377                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1378         if (ret < 0 && ret != -ENOENT) {
1379                 free_leaf_list(*leafs);
1380                 return ret;
1381         }
1382
1383         return 0;
1384 }
1385
1386 /*
1387  * walk all backrefs for a given extent to find all roots that reference this
1388  * extent. Walking a backref means finding all extents that reference this
1389  * extent and in turn walk the backrefs of those, too. Naturally this is a
1390  * recursive process, but here it is implemented in an iterative fashion: We
1391  * find all referencing extents for the extent in question and put them on a
1392  * list. In turn, we find all referencing extents for those, further appending
1393  * to the list. The way we iterate the list allows adding more elements after
1394  * the current while iterating. The process stops when we reach the end of the
1395  * list. Found roots are added to the roots list.
1396  *
1397  * returns 0 on success, < 0 on error.
1398  */
1399 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1400                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1401                                      u64 time_seq, struct ulist **roots,
1402                                      bool ignore_offset)
1403 {
1404         struct ulist *tmp;
1405         struct ulist_node *node = NULL;
1406         struct ulist_iterator uiter;
1407         int ret;
1408
1409         tmp = ulist_alloc(GFP_NOFS);
1410         if (!tmp)
1411                 return -ENOMEM;
1412         *roots = ulist_alloc(GFP_NOFS);
1413         if (!*roots) {
1414                 ulist_free(tmp);
1415                 return -ENOMEM;
1416         }
1417
1418         ULIST_ITER_INIT(&uiter);
1419         while (1) {
1420                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1421                                         tmp, *roots, NULL, NULL, ignore_offset);
1422                 if (ret < 0 && ret != -ENOENT) {
1423                         ulist_free(tmp);
1424                         ulist_free(*roots);
1425                         return ret;
1426                 }
1427                 node = ulist_next(tmp, &uiter);
1428                 if (!node)
1429                         break;
1430                 bytenr = node->val;
1431                 cond_resched();
1432         }
1433
1434         ulist_free(tmp);
1435         return 0;
1436 }
1437
1438 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1439                          struct btrfs_fs_info *fs_info, u64 bytenr,
1440                          u64 time_seq, struct ulist **roots,
1441                          bool ignore_offset)
1442 {
1443         int ret;
1444
1445         if (!trans)
1446                 down_read(&fs_info->commit_root_sem);
1447         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1448                                         time_seq, roots, ignore_offset);
1449         if (!trans)
1450                 up_read(&fs_info->commit_root_sem);
1451         return ret;
1452 }
1453
1454 /**
1455  * btrfs_check_shared - tell us whether an extent is shared
1456  *
1457  * btrfs_check_shared uses the backref walking code but will short
1458  * circuit as soon as it finds a root or inode that doesn't match the
1459  * one passed in. This provides a significant performance benefit for
1460  * callers (such as fiemap) which want to know whether the extent is
1461  * shared but do not need a ref count.
1462  *
1463  * This attempts to allocate a transaction in order to account for
1464  * delayed refs, but continues on even when the alloc fails.
1465  *
1466  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1467  */
1468 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1469 {
1470         struct btrfs_fs_info *fs_info = root->fs_info;
1471         struct btrfs_trans_handle *trans;
1472         struct ulist *tmp = NULL;
1473         struct ulist *roots = NULL;
1474         struct ulist_iterator uiter;
1475         struct ulist_node *node;
1476         struct seq_list elem = SEQ_LIST_INIT(elem);
1477         int ret = 0;
1478         struct share_check shared = {
1479                 .root_objectid = root->root_key.objectid,
1480                 .inum = inum,
1481                 .share_count = 0,
1482         };
1483
1484         tmp = ulist_alloc(GFP_NOFS);
1485         roots = ulist_alloc(GFP_NOFS);
1486         if (!tmp || !roots) {
1487                 ulist_free(tmp);
1488                 ulist_free(roots);
1489                 return -ENOMEM;
1490         }
1491
1492         trans = btrfs_join_transaction(root);
1493         if (IS_ERR(trans)) {
1494                 trans = NULL;
1495                 down_read(&fs_info->commit_root_sem);
1496         } else {
1497                 btrfs_get_tree_mod_seq(fs_info, &elem);
1498         }
1499
1500         ULIST_ITER_INIT(&uiter);
1501         while (1) {
1502                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1503                                         roots, NULL, &shared, false);
1504                 if (ret == BACKREF_FOUND_SHARED) {
1505                         /* this is the only condition under which we return 1 */
1506                         ret = 1;
1507                         break;
1508                 }
1509                 if (ret < 0 && ret != -ENOENT)
1510                         break;
1511                 ret = 0;
1512                 node = ulist_next(tmp, &uiter);
1513                 if (!node)
1514                         break;
1515                 bytenr = node->val;
1516                 shared.share_count = 0;
1517                 cond_resched();
1518         }
1519
1520         if (trans) {
1521                 btrfs_put_tree_mod_seq(fs_info, &elem);
1522                 btrfs_end_transaction(trans);
1523         } else {
1524                 up_read(&fs_info->commit_root_sem);
1525         }
1526         ulist_free(tmp);
1527         ulist_free(roots);
1528         return ret;
1529 }
1530
1531 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1532                           u64 start_off, struct btrfs_path *path,
1533                           struct btrfs_inode_extref **ret_extref,
1534                           u64 *found_off)
1535 {
1536         int ret, slot;
1537         struct btrfs_key key;
1538         struct btrfs_key found_key;
1539         struct btrfs_inode_extref *extref;
1540         const struct extent_buffer *leaf;
1541         unsigned long ptr;
1542
1543         key.objectid = inode_objectid;
1544         key.type = BTRFS_INODE_EXTREF_KEY;
1545         key.offset = start_off;
1546
1547         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1548         if (ret < 0)
1549                 return ret;
1550
1551         while (1) {
1552                 leaf = path->nodes[0];
1553                 slot = path->slots[0];
1554                 if (slot >= btrfs_header_nritems(leaf)) {
1555                         /*
1556                          * If the item at offset is not found,
1557                          * btrfs_search_slot will point us to the slot
1558                          * where it should be inserted. In our case
1559                          * that will be the slot directly before the
1560                          * next INODE_REF_KEY_V2 item. In the case
1561                          * that we're pointing to the last slot in a
1562                          * leaf, we must move one leaf over.
1563                          */
1564                         ret = btrfs_next_leaf(root, path);
1565                         if (ret) {
1566                                 if (ret >= 1)
1567                                         ret = -ENOENT;
1568                                 break;
1569                         }
1570                         continue;
1571                 }
1572
1573                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1574
1575                 /*
1576                  * Check that we're still looking at an extended ref key for
1577                  * this particular objectid. If we have different
1578                  * objectid or type then there are no more to be found
1579                  * in the tree and we can exit.
1580                  */
1581                 ret = -ENOENT;
1582                 if (found_key.objectid != inode_objectid)
1583                         break;
1584                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1585                         break;
1586
1587                 ret = 0;
1588                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1589                 extref = (struct btrfs_inode_extref *)ptr;
1590                 *ret_extref = extref;
1591                 if (found_off)
1592                         *found_off = found_key.offset;
1593                 break;
1594         }
1595
1596         return ret;
1597 }
1598
1599 /*
1600  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1601  * Elements of the path are separated by '/' and the path is guaranteed to be
1602  * 0-terminated. the path is only given within the current file system.
1603  * Therefore, it never starts with a '/'. the caller is responsible to provide
1604  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1605  * the start point of the resulting string is returned. this pointer is within
1606  * dest, normally.
1607  * in case the path buffer would overflow, the pointer is decremented further
1608  * as if output was written to the buffer, though no more output is actually
1609  * generated. that way, the caller can determine how much space would be
1610  * required for the path to fit into the buffer. in that case, the returned
1611  * value will be smaller than dest. callers must check this!
1612  */
1613 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1614                         u32 name_len, unsigned long name_off,
1615                         struct extent_buffer *eb_in, u64 parent,
1616                         char *dest, u32 size)
1617 {
1618         int slot;
1619         u64 next_inum;
1620         int ret;
1621         s64 bytes_left = ((s64)size) - 1;
1622         struct extent_buffer *eb = eb_in;
1623         struct btrfs_key found_key;
1624         int leave_spinning = path->leave_spinning;
1625         struct btrfs_inode_ref *iref;
1626
1627         if (bytes_left >= 0)
1628                 dest[bytes_left] = '\0';
1629
1630         path->leave_spinning = 1;
1631         while (1) {
1632                 bytes_left -= name_len;
1633                 if (bytes_left >= 0)
1634                         read_extent_buffer(eb, dest + bytes_left,
1635                                            name_off, name_len);
1636                 if (eb != eb_in) {
1637                         if (!path->skip_locking)
1638                                 btrfs_tree_read_unlock_blocking(eb);
1639                         free_extent_buffer(eb);
1640                 }
1641                 ret = btrfs_find_item(fs_root, path, parent, 0,
1642                                 BTRFS_INODE_REF_KEY, &found_key);
1643                 if (ret > 0)
1644                         ret = -ENOENT;
1645                 if (ret)
1646                         break;
1647
1648                 next_inum = found_key.offset;
1649
1650                 /* regular exit ahead */
1651                 if (parent == next_inum)
1652                         break;
1653
1654                 slot = path->slots[0];
1655                 eb = path->nodes[0];
1656                 /* make sure we can use eb after releasing the path */
1657                 if (eb != eb_in) {
1658                         if (!path->skip_locking)
1659                                 btrfs_set_lock_blocking_read(eb);
1660                         path->nodes[0] = NULL;
1661                         path->locks[0] = 0;
1662                 }
1663                 btrfs_release_path(path);
1664                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1665
1666                 name_len = btrfs_inode_ref_name_len(eb, iref);
1667                 name_off = (unsigned long)(iref + 1);
1668
1669                 parent = next_inum;
1670                 --bytes_left;
1671                 if (bytes_left >= 0)
1672                         dest[bytes_left] = '/';
1673         }
1674
1675         btrfs_release_path(path);
1676         path->leave_spinning = leave_spinning;
1677
1678         if (ret)
1679                 return ERR_PTR(ret);
1680
1681         return dest + bytes_left;
1682 }
1683
1684 /*
1685  * this makes the path point to (logical EXTENT_ITEM *)
1686  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1687  * tree blocks and <0 on error.
1688  */
1689 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1690                         struct btrfs_path *path, struct btrfs_key *found_key,
1691                         u64 *flags_ret)
1692 {
1693         int ret;
1694         u64 flags;
1695         u64 size = 0;
1696         u32 item_size;
1697         const struct extent_buffer *eb;
1698         struct btrfs_extent_item *ei;
1699         struct btrfs_key key;
1700
1701         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1702                 key.type = BTRFS_METADATA_ITEM_KEY;
1703         else
1704                 key.type = BTRFS_EXTENT_ITEM_KEY;
1705         key.objectid = logical;
1706         key.offset = (u64)-1;
1707
1708         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1709         if (ret < 0)
1710                 return ret;
1711
1712         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1713         if (ret) {
1714                 if (ret > 0)
1715                         ret = -ENOENT;
1716                 return ret;
1717         }
1718         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1719         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1720                 size = fs_info->nodesize;
1721         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1722                 size = found_key->offset;
1723
1724         if (found_key->objectid > logical ||
1725             found_key->objectid + size <= logical) {
1726                 btrfs_debug(fs_info,
1727                         "logical %llu is not within any extent", logical);
1728                 return -ENOENT;
1729         }
1730
1731         eb = path->nodes[0];
1732         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1733         BUG_ON(item_size < sizeof(*ei));
1734
1735         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1736         flags = btrfs_extent_flags(eb, ei);
1737
1738         btrfs_debug(fs_info,
1739                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1740                  logical, logical - found_key->objectid, found_key->objectid,
1741                  found_key->offset, flags, item_size);
1742
1743         WARN_ON(!flags_ret);
1744         if (flags_ret) {
1745                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1746                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1747                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1748                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1749                 else
1750                         BUG_ON(1);
1751                 return 0;
1752         }
1753
1754         return -EIO;
1755 }
1756
1757 /*
1758  * helper function to iterate extent inline refs. ptr must point to a 0 value
1759  * for the first call and may be modified. it is used to track state.
1760  * if more refs exist, 0 is returned and the next call to
1761  * get_extent_inline_ref must pass the modified ptr parameter to get the
1762  * next ref. after the last ref was processed, 1 is returned.
1763  * returns <0 on error
1764  */
1765 static int get_extent_inline_ref(unsigned long *ptr,
1766                                  const struct extent_buffer *eb,
1767                                  const struct btrfs_key *key,
1768                                  const struct btrfs_extent_item *ei,
1769                                  u32 item_size,
1770                                  struct btrfs_extent_inline_ref **out_eiref,
1771                                  int *out_type)
1772 {
1773         unsigned long end;
1774         u64 flags;
1775         struct btrfs_tree_block_info *info;
1776
1777         if (!*ptr) {
1778                 /* first call */
1779                 flags = btrfs_extent_flags(eb, ei);
1780                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1781                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1782                                 /* a skinny metadata extent */
1783                                 *out_eiref =
1784                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1785                         } else {
1786                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1787                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1788                                 *out_eiref =
1789                                    (struct btrfs_extent_inline_ref *)(info + 1);
1790                         }
1791                 } else {
1792                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1793                 }
1794                 *ptr = (unsigned long)*out_eiref;
1795                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1796                         return -ENOENT;
1797         }
1798
1799         end = (unsigned long)ei + item_size;
1800         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1801         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1802                                                      BTRFS_REF_TYPE_ANY);
1803         if (*out_type == BTRFS_REF_TYPE_INVALID)
1804                 return -EUCLEAN;
1805
1806         *ptr += btrfs_extent_inline_ref_size(*out_type);
1807         WARN_ON(*ptr > end);
1808         if (*ptr == end)
1809                 return 1; /* last */
1810
1811         return 0;
1812 }
1813
1814 /*
1815  * reads the tree block backref for an extent. tree level and root are returned
1816  * through out_level and out_root. ptr must point to a 0 value for the first
1817  * call and may be modified (see get_extent_inline_ref comment).
1818  * returns 0 if data was provided, 1 if there was no more data to provide or
1819  * <0 on error.
1820  */
1821 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1822                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1823                             u32 item_size, u64 *out_root, u8 *out_level)
1824 {
1825         int ret;
1826         int type;
1827         struct btrfs_extent_inline_ref *eiref;
1828
1829         if (*ptr == (unsigned long)-1)
1830                 return 1;
1831
1832         while (1) {
1833                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1834                                               &eiref, &type);
1835                 if (ret < 0)
1836                         return ret;
1837
1838                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1839                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1840                         break;
1841
1842                 if (ret == 1)
1843                         return 1;
1844         }
1845
1846         /* we can treat both ref types equally here */
1847         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1848
1849         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1850                 struct btrfs_tree_block_info *info;
1851
1852                 info = (struct btrfs_tree_block_info *)(ei + 1);
1853                 *out_level = btrfs_tree_block_level(eb, info);
1854         } else {
1855                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1856                 *out_level = (u8)key->offset;
1857         }
1858
1859         if (ret == 1)
1860                 *ptr = (unsigned long)-1;
1861
1862         return 0;
1863 }
1864
1865 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1866                              struct extent_inode_elem *inode_list,
1867                              u64 root, u64 extent_item_objectid,
1868                              iterate_extent_inodes_t *iterate, void *ctx)
1869 {
1870         struct extent_inode_elem *eie;
1871         int ret = 0;
1872
1873         for (eie = inode_list; eie; eie = eie->next) {
1874                 btrfs_debug(fs_info,
1875                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1876                             extent_item_objectid, eie->inum,
1877                             eie->offset, root);
1878                 ret = iterate(eie->inum, eie->offset, root, ctx);
1879                 if (ret) {
1880                         btrfs_debug(fs_info,
1881                                     "stopping iteration for %llu due to ret=%d",
1882                                     extent_item_objectid, ret);
1883                         break;
1884                 }
1885         }
1886
1887         return ret;
1888 }
1889
1890 /*
1891  * calls iterate() for every inode that references the extent identified by
1892  * the given parameters.
1893  * when the iterator function returns a non-zero value, iteration stops.
1894  */
1895 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1896                                 u64 extent_item_objectid, u64 extent_item_pos,
1897                                 int search_commit_root,
1898                                 iterate_extent_inodes_t *iterate, void *ctx,
1899                                 bool ignore_offset)
1900 {
1901         int ret;
1902         struct btrfs_trans_handle *trans = NULL;
1903         struct ulist *refs = NULL;
1904         struct ulist *roots = NULL;
1905         struct ulist_node *ref_node = NULL;
1906         struct ulist_node *root_node = NULL;
1907         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1908         struct ulist_iterator ref_uiter;
1909         struct ulist_iterator root_uiter;
1910
1911         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1912                         extent_item_objectid);
1913
1914         if (!search_commit_root) {
1915                 trans = btrfs_join_transaction(fs_info->extent_root);
1916                 if (IS_ERR(trans))
1917                         return PTR_ERR(trans);
1918                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1919         } else {
1920                 down_read(&fs_info->commit_root_sem);
1921         }
1922
1923         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1924                                    tree_mod_seq_elem.seq, &refs,
1925                                    &extent_item_pos, ignore_offset);
1926         if (ret)
1927                 goto out;
1928
1929         ULIST_ITER_INIT(&ref_uiter);
1930         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1931                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1932                                                 tree_mod_seq_elem.seq, &roots,
1933                                                 ignore_offset);
1934                 if (ret)
1935                         break;
1936                 ULIST_ITER_INIT(&root_uiter);
1937                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1938                         btrfs_debug(fs_info,
1939                                     "root %llu references leaf %llu, data list %#llx",
1940                                     root_node->val, ref_node->val,
1941                                     ref_node->aux);
1942                         ret = iterate_leaf_refs(fs_info,
1943                                                 (struct extent_inode_elem *)
1944                                                 (uintptr_t)ref_node->aux,
1945                                                 root_node->val,
1946                                                 extent_item_objectid,
1947                                                 iterate, ctx);
1948                 }
1949                 ulist_free(roots);
1950         }
1951
1952         free_leaf_list(refs);
1953 out:
1954         if (!search_commit_root) {
1955                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1956                 btrfs_end_transaction(trans);
1957         } else {
1958                 up_read(&fs_info->commit_root_sem);
1959         }
1960
1961         return ret;
1962 }
1963
1964 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1965                                 struct btrfs_path *path,
1966                                 iterate_extent_inodes_t *iterate, void *ctx,
1967                                 bool ignore_offset)
1968 {
1969         int ret;
1970         u64 extent_item_pos;
1971         u64 flags = 0;
1972         struct btrfs_key found_key;
1973         int search_commit_root = path->search_commit_root;
1974
1975         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1976         btrfs_release_path(path);
1977         if (ret < 0)
1978                 return ret;
1979         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1980                 return -EINVAL;
1981
1982         extent_item_pos = logical - found_key.objectid;
1983         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1984                                         extent_item_pos, search_commit_root,
1985                                         iterate, ctx, ignore_offset);
1986
1987         return ret;
1988 }
1989
1990 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1991                               struct extent_buffer *eb, void *ctx);
1992
1993 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1994                               struct btrfs_path *path,
1995                               iterate_irefs_t *iterate, void *ctx)
1996 {
1997         int ret = 0;
1998         int slot;
1999         u32 cur;
2000         u32 len;
2001         u32 name_len;
2002         u64 parent = 0;
2003         int found = 0;
2004         struct extent_buffer *eb;
2005         struct btrfs_item *item;
2006         struct btrfs_inode_ref *iref;
2007         struct btrfs_key found_key;
2008
2009         while (!ret) {
2010                 ret = btrfs_find_item(fs_root, path, inum,
2011                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2012                                 &found_key);
2013
2014                 if (ret < 0)
2015                         break;
2016                 if (ret) {
2017                         ret = found ? 0 : -ENOENT;
2018                         break;
2019                 }
2020                 ++found;
2021
2022                 parent = found_key.offset;
2023                 slot = path->slots[0];
2024                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2025                 if (!eb) {
2026                         ret = -ENOMEM;
2027                         break;
2028                 }
2029                 btrfs_release_path(path);
2030
2031                 item = btrfs_item_nr(slot);
2032                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2033
2034                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2035                         name_len = btrfs_inode_ref_name_len(eb, iref);
2036                         /* path must be released before calling iterate()! */
2037                         btrfs_debug(fs_root->fs_info,
2038                                 "following ref at offset %u for inode %llu in tree %llu",
2039                                 cur, found_key.objectid,
2040                                 fs_root->root_key.objectid);
2041                         ret = iterate(parent, name_len,
2042                                       (unsigned long)(iref + 1), eb, ctx);
2043                         if (ret)
2044                                 break;
2045                         len = sizeof(*iref) + name_len;
2046                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2047                 }
2048                 free_extent_buffer(eb);
2049         }
2050
2051         btrfs_release_path(path);
2052
2053         return ret;
2054 }
2055
2056 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2057                                  struct btrfs_path *path,
2058                                  iterate_irefs_t *iterate, void *ctx)
2059 {
2060         int ret;
2061         int slot;
2062         u64 offset = 0;
2063         u64 parent;
2064         int found = 0;
2065         struct extent_buffer *eb;
2066         struct btrfs_inode_extref *extref;
2067         u32 item_size;
2068         u32 cur_offset;
2069         unsigned long ptr;
2070
2071         while (1) {
2072                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2073                                             &offset);
2074                 if (ret < 0)
2075                         break;
2076                 if (ret) {
2077                         ret = found ? 0 : -ENOENT;
2078                         break;
2079                 }
2080                 ++found;
2081
2082                 slot = path->slots[0];
2083                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2084                 if (!eb) {
2085                         ret = -ENOMEM;
2086                         break;
2087                 }
2088                 btrfs_release_path(path);
2089
2090                 item_size = btrfs_item_size_nr(eb, slot);
2091                 ptr = btrfs_item_ptr_offset(eb, slot);
2092                 cur_offset = 0;
2093
2094                 while (cur_offset < item_size) {
2095                         u32 name_len;
2096
2097                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2098                         parent = btrfs_inode_extref_parent(eb, extref);
2099                         name_len = btrfs_inode_extref_name_len(eb, extref);
2100                         ret = iterate(parent, name_len,
2101                                       (unsigned long)&extref->name, eb, ctx);
2102                         if (ret)
2103                                 break;
2104
2105                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2106                         cur_offset += sizeof(*extref);
2107                 }
2108                 free_extent_buffer(eb);
2109
2110                 offset++;
2111         }
2112
2113         btrfs_release_path(path);
2114
2115         return ret;
2116 }
2117
2118 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2119                          struct btrfs_path *path, iterate_irefs_t *iterate,
2120                          void *ctx)
2121 {
2122         int ret;
2123         int found_refs = 0;
2124
2125         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2126         if (!ret)
2127                 ++found_refs;
2128         else if (ret != -ENOENT)
2129                 return ret;
2130
2131         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2132         if (ret == -ENOENT && found_refs)
2133                 return 0;
2134
2135         return ret;
2136 }
2137
2138 /*
2139  * returns 0 if the path could be dumped (probably truncated)
2140  * returns <0 in case of an error
2141  */
2142 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2143                          struct extent_buffer *eb, void *ctx)
2144 {
2145         struct inode_fs_paths *ipath = ctx;
2146         char *fspath;
2147         char *fspath_min;
2148         int i = ipath->fspath->elem_cnt;
2149         const int s_ptr = sizeof(char *);
2150         u32 bytes_left;
2151
2152         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2153                                         ipath->fspath->bytes_left - s_ptr : 0;
2154
2155         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2156         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2157                                    name_off, eb, inum, fspath_min, bytes_left);
2158         if (IS_ERR(fspath))
2159                 return PTR_ERR(fspath);
2160
2161         if (fspath > fspath_min) {
2162                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2163                 ++ipath->fspath->elem_cnt;
2164                 ipath->fspath->bytes_left = fspath - fspath_min;
2165         } else {
2166                 ++ipath->fspath->elem_missed;
2167                 ipath->fspath->bytes_missing += fspath_min - fspath;
2168                 ipath->fspath->bytes_left = 0;
2169         }
2170
2171         return 0;
2172 }
2173
2174 /*
2175  * this dumps all file system paths to the inode into the ipath struct, provided
2176  * is has been created large enough. each path is zero-terminated and accessed
2177  * from ipath->fspath->val[i].
2178  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2179  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2180  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2181  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2182  * have been needed to return all paths.
2183  */
2184 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2185 {
2186         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2187                              inode_to_path, ipath);
2188 }
2189
2190 struct btrfs_data_container *init_data_container(u32 total_bytes)
2191 {
2192         struct btrfs_data_container *data;
2193         size_t alloc_bytes;
2194
2195         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2196         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2197         if (!data)
2198                 return ERR_PTR(-ENOMEM);
2199
2200         if (total_bytes >= sizeof(*data)) {
2201                 data->bytes_left = total_bytes - sizeof(*data);
2202                 data->bytes_missing = 0;
2203         } else {
2204                 data->bytes_missing = sizeof(*data) - total_bytes;
2205                 data->bytes_left = 0;
2206         }
2207
2208         data->elem_cnt = 0;
2209         data->elem_missed = 0;
2210
2211         return data;
2212 }
2213
2214 /*
2215  * allocates space to return multiple file system paths for an inode.
2216  * total_bytes to allocate are passed, note that space usable for actual path
2217  * information will be total_bytes - sizeof(struct inode_fs_paths).
2218  * the returned pointer must be freed with free_ipath() in the end.
2219  */
2220 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2221                                         struct btrfs_path *path)
2222 {
2223         struct inode_fs_paths *ifp;
2224         struct btrfs_data_container *fspath;
2225
2226         fspath = init_data_container(total_bytes);
2227         if (IS_ERR(fspath))
2228                 return ERR_CAST(fspath);
2229
2230         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2231         if (!ifp) {
2232                 kvfree(fspath);
2233                 return ERR_PTR(-ENOMEM);
2234         }
2235
2236         ifp->btrfs_path = path;
2237         ifp->fspath = fspath;
2238         ifp->fs_root = fs_root;
2239
2240         return ifp;
2241 }
2242
2243 void free_ipath(struct inode_fs_paths *ipath)
2244 {
2245         if (!ipath)
2246                 return;
2247         kvfree(ipath->fspath);
2248         kfree(ipath);
2249 }