fs/btrfs/backref.c

   1 /*
   2  * Copyright (C) 2011 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/mm.h>
  20 #include <linux/rbtree.h>
  21 #include <trace/events/btrfs.h>
  22 #include "ctree.h"
  23 #include "disk-io.h"
  24 #include "backref.h"
  25 #include "ulist.h"
  26 #include "transaction.h"
  27 #include "delayed-ref.h"
  28 #include "locking.h"
  29
  30 /* Just an arbitrary number so we can be sure this happened */
  31 #define BACKREF_FOUND_SHARED 6
  32
  33 struct extent_inode_elem {
  34         u64 inum;
  35         u64 offset;
  36         struct extent_inode_elem *next;
  37 };
  38
  39 static int check_extent_in_eb(const struct btrfs_key *key,
  40                               const struct extent_buffer *eb,
  41                               const struct btrfs_file_extent_item *fi,
  42                               u64 extent_item_pos,
  43                               struct extent_inode_elem **eie)
  44 {
  45         u64 offset = 0;
  46         struct extent_inode_elem *e;
  47
  48         if (!btrfs_file_extent_compression(eb, fi) &&
  49             !btrfs_file_extent_encryption(eb, fi) &&
  50             !btrfs_file_extent_other_encoding(eb, fi)) {
  51                 u64 data_offset;
  52                 u64 data_len;
  53
  54                 data_offset = btrfs_file_extent_offset(eb, fi);
  55                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  56
  57                 if (extent_item_pos < data_offset ||
  58                     extent_item_pos >= data_offset + data_len)
  59                         return 1;
  60                 offset = extent_item_pos - data_offset;
  61         }
  62
  63         e = kmalloc(sizeof(*e), GFP_NOFS);
  64         if (!e)
  65                 return -ENOMEM;
  66
  67         e->next = *eie;
  68         e->inum = key->objectid;
  69         e->offset = key->offset + offset;
  70         *eie = e;
  71
  72         return 0;
  73 }
  74
  75 static void free_inode_elem_list(struct extent_inode_elem *eie)
  76 {
  77         struct extent_inode_elem *eie_next;
  78
  79         for (; eie; eie = eie_next) {
  80                 eie_next = eie->next;
  81                 kfree(eie);
  82         }
  83 }
  84
  85 static int find_extent_in_eb(const struct extent_buffer *eb,
  86                              u64 wanted_disk_byte, u64 extent_item_pos,
  87                              struct extent_inode_elem **eie)
  88 {
  89         u64 disk_byte;
  90         struct btrfs_key key;
  91         struct btrfs_file_extent_item *fi;
  92         int slot;
  93         int nritems;
  94         int extent_type;
  95         int ret;
  96
  97         /*
  98          * from the shared data ref, we only have the leaf but we need
  99          * the key. thus, we must look into all items and see that we
 100          * find one (some) with a reference to our extent item.
 101          */
 102         nritems = btrfs_header_nritems(eb);
 103         for (slot = 0; slot < nritems; ++slot) {
 104                 btrfs_item_key_to_cpu(eb, &key, slot);
 105                 if (key.type != BTRFS_EXTENT_DATA_KEY)
 106                         continue;
 107                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 108                 extent_type = btrfs_file_extent_type(eb, fi);
 109                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 110                         continue;
 111                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 112                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 113                 if (disk_byte != wanted_disk_byte)
 114                         continue;
 115
 116                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
 117                 if (ret < 0)
 118                         return ret;
 119         }
 120
 121         return 0;
 122 }
 123
 124 struct preftree {
 125         struct rb_root root;
 126         unsigned int count;
 127 };
 128
 129 #define PREFTREE_INIT   { .root = RB_ROOT, .count = 0 }
 130
 131 struct preftrees {
 132         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
 133         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
 134         struct preftree indirect_missing_keys;
 135 };
 136
 137 /*
 138  * Checks for a shared extent during backref search.
 139  *
 140  * The share_count tracks prelim_refs (direct and indirect) having a
 141  * ref->count >0:
 142  *  - incremented when a ref->count transitions to >0
 143  *  - decremented when a ref->count transitions to <1
 144  */
 145 struct share_check {
 146         u64 root_objectid;
 147         u64 inum;
 148         int share_count;
 149 };
 150
 151 static inline int extent_is_shared(struct share_check *sc)
 152 {
 153         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
 154 }
 155
 156 static struct kmem_cache *btrfs_prelim_ref_cache;
 157
 158 int __init btrfs_prelim_ref_init(void)
 159 {
 160         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 161                                         sizeof(struct prelim_ref),
 162                                         0,
 163                                         SLAB_MEM_SPREAD,
 164                                         NULL);
 165         if (!btrfs_prelim_ref_cache)
 166                 return -ENOMEM;
 167         return 0;
 168 }
 169
 170 void btrfs_prelim_ref_exit(void)
 171 {
 172         kmem_cache_destroy(btrfs_prelim_ref_cache);
 173 }
 174
 175 static void free_pref(struct prelim_ref *ref)
 176 {
 177         kmem_cache_free(btrfs_prelim_ref_cache, ref);
 178 }
 179
 180 /*
 181  * Return 0 when both refs are for the same block (and can be merged).
 182  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
 183  * indicates a 'higher' block.
 184  */
 185 static int prelim_ref_compare(struct prelim_ref *ref1,
 186                               struct prelim_ref *ref2)
 187 {
 188         if (ref1->level < ref2->level)
 189                 return -1;
 190         if (ref1->level > ref2->level)
 191                 return 1;
 192         if (ref1->root_id < ref2->root_id)
 193                 return -1;
 194         if (ref1->root_id > ref2->root_id)
 195                 return 1;
 196         if (ref1->key_for_search.type < ref2->key_for_search.type)
 197                 return -1;
 198         if (ref1->key_for_search.type > ref2->key_for_search.type)
 199                 return 1;
 200         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
 201                 return -1;
 202         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
 203                 return 1;
 204         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
 205                 return -1;
 206         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
 207                 return 1;
 208         if (ref1->parent < ref2->parent)
 209                 return -1;
 210         if (ref1->parent > ref2->parent)
 211                 return 1;
 212
 213         return 0;
 214 }
 215
 216 void update_share_count(struct share_check *sc, int oldcount, int newcount)
 217 {
 218         if ((!sc) || (oldcount == 0 && newcount < 1))
 219                 return;
 220
 221         if (oldcount > 0 && newcount < 1)
 222                 sc->share_count--;
 223         else if (oldcount < 1 && newcount > 0)
 224                 sc->share_count++;
 225 }
 226
 227 /*
 228  * Add @newref to the @root rbtree, merging identical refs.
 229  *
 230  * Callers should assume that newref has been freed after calling.
 231  */
 232 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
 233                               struct preftree *preftree,
 234                               struct prelim_ref *newref,
 235                               struct share_check *sc)
 236 {
 237         struct rb_root *root;
 238         struct rb_node **p;
 239         struct rb_node *parent = NULL;
 240         struct prelim_ref *ref;
 241         int result;
 242
 243         root = &preftree->root;
 244         p = &root->rb_node;
 245
 246         while (*p) {
 247                 parent = *p;
 248                 ref = rb_entry(parent, struct prelim_ref, rbnode);
 249                 result = prelim_ref_compare(ref, newref);
 250                 if (result < 0) {
 251                         p = &(*p)->rb_left;
 252                 } else if (result > 0) {
 253                         p = &(*p)->rb_right;
 254                 } else {
 255                         /* Identical refs, merge them and free @newref */
 256                         struct extent_inode_elem *eie = ref->inode_list;
 257
 258                         while (eie && eie->next)
 259                                 eie = eie->next;
 260
 261                         if (!eie)
 262                                 ref->inode_list = newref->inode_list;
 263                         else
 264                                 eie->next = newref->inode_list;
 265                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
 266                                                      preftree->count);
 267                         /*
 268                          * A delayed ref can have newref->count < 0.
 269                          * The ref->count is updated to follow any
 270                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
 271                          */
 272                         update_share_count(sc, ref->count,
 273                                            ref->count + newref->count);
 274                         ref->count += newref->count;
 275                         free_pref(newref);
 276                         return;
 277                 }
 278         }
 279
 280         update_share_count(sc, 0, newref->count);
 281         preftree->count++;
 282         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
 283         rb_link_node(&newref->rbnode, parent, p);
 284         rb_insert_color(&newref->rbnode, root);
 285 }
 286
 287 /*
 288  * Release the entire tree.  We don't care about internal consistency so
 289  * just free everything and then reset the tree root.
 290  */
 291 static void prelim_release(struct preftree *preftree)
 292 {
 293         struct prelim_ref *ref, *next_ref;
 294
 295         rbtree_postorder_for_each_entry_safe(ref, next_ref, &preftree->root,
 296                                              rbnode)
 297                 free_pref(ref);
 298
 299         preftree->root = RB_ROOT;
 300         preftree->count = 0;
 301 }
 302
 303 /*
 304  * the rules for all callers of this function are:
 305  * - obtaining the parent is the goal
 306  * - if you add a key, you must know that it is a correct key
 307  * - if you cannot add the parent or a correct key, then we will look into the
 308  *   block later to set a correct key
 309  *
 310  * delayed refs
 311  * ============
 312  *        backref type | shared | indirect | shared | indirect
 313  * information         |   tree |     tree |   data |     data
 314  * --------------------+--------+----------+--------+----------
 315  *      parent logical |    y   |     -    |    -   |     -
 316  *      key to resolve |    -   |     y    |    y   |     y
 317  *  tree block logical |    -   |     -    |    -   |     -
 318  *  root for resolving |    y   |     y    |    y   |     y
 319  *
 320  * - column 1:       we've the parent -> done
 321  * - column 2, 3, 4: we use the key to find the parent
 322  *
 323  * on disk refs (inline or keyed)
 324  * ==============================
 325  *        backref type | shared | indirect | shared | indirect
 326  * information         |   tree |     tree |   data |     data
 327  * --------------------+--------+----------+--------+----------
 328  *      parent logical |    y   |     -    |    y   |     -
 329  *      key to resolve |    -   |     -    |    -   |     y
 330  *  tree block logical |    y   |     y    |    y   |     y
 331  *  root for resolving |    -   |     y    |    y   |     y
 332  *
 333  * - column 1, 3: we've the parent -> done
 334  * - column 2:    we take the first key from the block to find the parent
 335  *                (see add_missing_keys)
 336  * - column 4:    we use the key to find the parent
 337  *
 338  * additional information that's available but not required to find the parent
 339  * block might help in merging entries to gain some speed.
 340  */
 341 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
 342                           struct preftree *preftree, u64 root_id,
 343                           const struct btrfs_key *key, int level, u64 parent,
 344                           u64 wanted_disk_byte, int count,
 345                           struct share_check *sc, gfp_t gfp_mask)
 346 {
 347         struct prelim_ref *ref;
 348
 349         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 350                 return 0;
 351
 352         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 353         if (!ref)
 354                 return -ENOMEM;
 355
 356         ref->root_id = root_id;
 357         if (key) {
 358                 ref->key_for_search = *key;
 359                 /*
 360                  * We can often find data backrefs with an offset that is too
 361                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 362                  * underflows when subtracting a file's offset with the data
 363                  * offset of its corresponding extent data item. This can
 364                  * happen for example in the clone ioctl.
 365                  * So if we detect such case we set the search key's offset to
 366                  * zero to make sure we will find the matching file extent item
 367                  * at add_all_parents(), otherwise we will miss it because the
 368                  * offset taken form the backref is much larger then the offset
 369                  * of the file extent item. This can make us scan a very large
 370                  * number of file extent items, but at least it will not make
 371                  * us miss any.
 372                  * This is an ugly workaround for a behaviour that should have
 373                  * never existed, but it does and a fix for the clone ioctl
 374                  * would touch a lot of places, cause backwards incompatibility
 375                  * and would not fix the problem for extents cloned with older
 376                  * kernels.
 377                  */
 378                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 379                     ref->key_for_search.offset >= LLONG_MAX)
 380                         ref->key_for_search.offset = 0;
 381         } else {
 382                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 383         }
 384
 385         ref->inode_list = NULL;
 386         ref->level = level;
 387         ref->count = count;
 388         ref->parent = parent;
 389         ref->wanted_disk_byte = wanted_disk_byte;
 390         prelim_ref_insert(fs_info, preftree, ref, sc);
 391         return extent_is_shared(sc);
 392 }
 393
 394 /* direct refs use root == 0, key == NULL */
 395 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
 396                           struct preftrees *preftrees, int level, u64 parent,
 397                           u64 wanted_disk_byte, int count,
 398                           struct share_check *sc, gfp_t gfp_mask)
 399 {
 400         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
 401                               parent, wanted_disk_byte, count, sc, gfp_mask);
 402 }
 403
 404 /* indirect refs use parent == 0 */
 405 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
 406                             struct preftrees *preftrees, u64 root_id,
 407                             const struct btrfs_key *key, int level,
 408                             u64 wanted_disk_byte, int count,
 409                             struct share_check *sc, gfp_t gfp_mask)
 410 {
 411         struct preftree *tree = &preftrees->indirect;
 412
 413         if (!key)
 414                 tree = &preftrees->indirect_missing_keys;
 415         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
 416                               wanted_disk_byte, count, sc, gfp_mask);
 417 }
 418
 419 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 420                            struct ulist *parents, struct prelim_ref *ref,
 421                            int level, u64 time_seq, const u64 *extent_item_pos,
 422                            u64 total_refs)
 423 {
 424         int ret = 0;
 425         int slot;
 426         struct extent_buffer *eb;
 427         struct btrfs_key key;
 428         struct btrfs_key *key_for_search = &ref->key_for_search;
 429         struct btrfs_file_extent_item *fi;
 430         struct extent_inode_elem *eie = NULL, *old = NULL;
 431         u64 disk_byte;
 432         u64 wanted_disk_byte = ref->wanted_disk_byte;
 433         u64 count = 0;
 434
 435         if (level != 0) {
 436                 eb = path->nodes[level];
 437                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 438                 if (ret < 0)
 439                         return ret;
 440                 return 0;
 441         }
 442
 443         /*
 444          * We normally enter this function with the path already pointing to
 445          * the first item to check. But sometimes, we may enter it with
 446          * slot==nritems. In that case, go to the next leaf before we continue.
 447          */
 448         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 449                 if (time_seq == SEQ_LAST)
 450                         ret = btrfs_next_leaf(root, path);
 451                 else
 452                         ret = btrfs_next_old_leaf(root, path, time_seq);
 453         }
 454
 455         while (!ret && count < total_refs) {
 456                 eb = path->nodes[0];
 457                 slot = path->slots[0];
 458
 459                 btrfs_item_key_to_cpu(eb, &key, slot);
 460
 461                 if (key.objectid != key_for_search->objectid ||
 462                     key.type != BTRFS_EXTENT_DATA_KEY)
 463                         break;
 464
 465                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 466                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 467
 468                 if (disk_byte == wanted_disk_byte) {
 469                         eie = NULL;
 470                         old = NULL;
 471                         count++;
 472                         if (extent_item_pos) {
 473                                 ret = check_extent_in_eb(&key, eb, fi,
 474                                                 *extent_item_pos,
 475                                                 &eie);
 476                                 if (ret < 0)
 477                                         break;
 478                         }
 479                         if (ret > 0)
 480                                 goto next;
 481                         ret = ulist_add_merge_ptr(parents, eb->start,
 482                                                   eie, (void **)&old, GFP_NOFS);
 483                         if (ret < 0)
 484                                 break;
 485                         if (!ret && extent_item_pos) {
 486                                 while (old->next)
 487                                         old = old->next;
 488                                 old->next = eie;
 489                         }
 490                         eie = NULL;
 491                 }
 492 next:
 493                 if (time_seq == SEQ_LAST)
 494                         ret = btrfs_next_item(root, path);
 495                 else
 496                         ret = btrfs_next_old_item(root, path, time_seq);
 497         }
 498
 499         if (ret > 0)
 500                 ret = 0;
 501         else if (ret < 0)
 502                 free_inode_elem_list(eie);
 503         return ret;
 504 }
 505
 506 /*
 507  * resolve an indirect backref in the form (root_id, key, level)
 508  * to a logical address
 509  */
 510 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 511                                 struct btrfs_path *path, u64 time_seq,
 512                                 struct prelim_ref *ref, struct ulist *parents,
 513                                 const u64 *extent_item_pos, u64 total_refs)
 514 {
 515         struct btrfs_root *root;
 516         struct btrfs_key root_key;
 517         struct extent_buffer *eb;
 518         int ret = 0;
 519         int root_level;
 520         int level = ref->level;
 521         int index;
 522
 523         root_key.objectid = ref->root_id;
 524         root_key.type = BTRFS_ROOT_ITEM_KEY;
 525         root_key.offset = (u64)-1;
 526
 527         index = srcu_read_lock(&fs_info->subvol_srcu);
 528
 529         root = btrfs_get_fs_root(fs_info, &root_key, false);
 530         if (IS_ERR(root)) {
 531                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 532                 ret = PTR_ERR(root);
 533                 goto out;
 534         }
 535
 536         if (btrfs_is_testing(fs_info)) {
 537                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 538                 ret = -ENOENT;
 539                 goto out;
 540         }
 541
 542         if (path->search_commit_root)
 543                 root_level = btrfs_header_level(root->commit_root);
 544         else if (time_seq == SEQ_LAST)
 545                 root_level = btrfs_header_level(root->node);
 546         else
 547                 root_level = btrfs_old_root_level(root, time_seq);
 548
 549         if (root_level + 1 == level) {
 550                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 551                 goto out;
 552         }
 553
 554         path->lowest_level = level;
 555         if (time_seq == SEQ_LAST)
 556                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 557                                         0, 0);
 558         else
 559                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 560                                             time_seq);
 561
 562         /* root node has been locked, we can release @subvol_srcu safely here */
 563         srcu_read_unlock(&fs_info->subvol_srcu, index);
 564
 565         btrfs_debug(fs_info,
 566                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 567                  ref->root_id, level, ref->count, ret,
 568                  ref->key_for_search.objectid, ref->key_for_search.type,
 569                  ref->key_for_search.offset);
 570         if (ret < 0)
 571                 goto out;
 572
 573         eb = path->nodes[level];
 574         while (!eb) {
 575                 if (WARN_ON(!level)) {
 576                         ret = 1;
 577                         goto out;
 578                 }
 579                 level--;
 580                 eb = path->nodes[level];
 581         }
 582
 583         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 584                               extent_item_pos, total_refs);
 585 out:
 586         path->lowest_level = 0;
 587         btrfs_release_path(path);
 588         return ret;
 589 }
 590
 591 static struct extent_inode_elem *
 592 unode_aux_to_inode_list(struct ulist_node *node)
 593 {
 594         if (!node)
 595                 return NULL;
 596         return (struct extent_inode_elem *)(uintptr_t)node->aux;
 597 }
 598
 599 /*
 600  * We maintain three seperate rbtrees: one for direct refs, one for
 601  * indirect refs which have a key, and one for indirect refs which do not
 602  * have a key. Each tree does merge on insertion.
 603  *
 604  * Once all of the references are located, we iterate over the tree of
 605  * indirect refs with missing keys. An appropriate key is located and
 606  * the ref is moved onto the tree for indirect refs. After all missing
 607  * keys are thus located, we iterate over the indirect ref tree, resolve
 608  * each reference, and then insert the resolved reference onto the
 609  * direct tree (merging there too).
 610  *
 611  * New backrefs (i.e., for parent nodes) are added to the appropriate
 612  * rbtree as they are encountered. The new backrefs are subsequently
 613  * resolved as above.
 614  */
 615 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 616                                  struct btrfs_path *path, u64 time_seq,
 617                                  struct preftrees *preftrees,
 618                                  const u64 *extent_item_pos, u64 total_refs,
 619                                  struct share_check *sc)
 620 {
 621         int err;
 622         int ret = 0;
 623         struct ulist *parents;
 624         struct ulist_node *node;
 625         struct ulist_iterator uiter;
 626         struct rb_node *rnode;
 627
 628         parents = ulist_alloc(GFP_NOFS);
 629         if (!parents)
 630                 return -ENOMEM;
 631
 632         /*
 633          * We could trade memory usage for performance here by iterating
 634          * the tree, allocating new refs for each insertion, and then
 635          * freeing the entire indirect tree when we're done.  In some test
 636          * cases, the tree can grow quite large (~200k objects).
 637          */
 638         while ((rnode = rb_first(&preftrees->indirect.root))) {
 639                 struct prelim_ref *ref;
 640
 641                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
 642                 if (WARN(ref->parent,
 643                          "BUG: direct ref found in indirect tree")) {
 644                         ret = -EINVAL;
 645                         goto out;
 646                 }
 647
 648                 rb_erase(&ref->rbnode, &preftrees->indirect.root);
 649                 preftrees->indirect.count--;
 650
 651                 if (ref->count == 0) {
 652                         free_pref(ref);
 653                         continue;
 654                 }
 655
 656                 if (sc && sc->root_objectid &&
 657                     ref->root_id != sc->root_objectid) {
 658                         free_pref(ref);
 659                         ret = BACKREF_FOUND_SHARED;
 660                         goto out;
 661                 }
 662                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
 663                                            parents, extent_item_pos,
 664                                            total_refs);
 665                 /*
 666                  * we can only tolerate ENOENT,otherwise,we should catch error
 667                  * and return directly.
 668                  */
 669                 if (err == -ENOENT) {
 670                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
 671                                           NULL);
 672                         continue;
 673                 } else if (err) {
 674                         free_pref(ref);
 675                         ret = err;
 676                         goto out;
 677                 }
 678
 679                 /* we put the first parent into the ref at hand */
 680                 ULIST_ITER_INIT(&uiter);
 681                 node = ulist_next(parents, &uiter);
 682                 ref->parent = node ? node->val : 0;
 683                 ref->inode_list = unode_aux_to_inode_list(node);
 684
 685                 /* Add a prelim_ref(s) for any other parent(s). */
 686                 while ((node = ulist_next(parents, &uiter))) {
 687                         struct prelim_ref *new_ref;
 688
 689                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 690                                                    GFP_NOFS);
 691                         if (!new_ref) {
 692                                 free_pref(ref);
 693                                 ret = -ENOMEM;
 694                                 goto out;
 695                         }
 696                         memcpy(new_ref, ref, sizeof(*ref));
 697                         new_ref->parent = node->val;
 698                         new_ref->inode_list = unode_aux_to_inode_list(node);
 699                         prelim_ref_insert(fs_info, &preftrees->direct,
 700                                           new_ref, NULL);
 701                 }
 702
 703                 /*
 704                  * Now it's a direct ref, put it in the the direct tree. We must
 705                  * do this last because the ref could be merged/freed here.
 706                  */
 707                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
 708
 709                 ulist_reinit(parents);
 710                 cond_resched();
 711         }
 712 out:
 713         ulist_free(parents);
 714         return ret;
 715 }
 716
 717 /*
 718  * read tree blocks and add keys where required.
 719  */
 720 static int add_missing_keys(struct btrfs_fs_info *fs_info,
 721                             struct preftrees *preftrees)
 722 {
 723         struct prelim_ref *ref;
 724         struct extent_buffer *eb;
 725         struct preftree *tree = &preftrees->indirect_missing_keys;
 726         struct rb_node *node;
 727
 728         while ((node = rb_first(&tree->root))) {
 729                 ref = rb_entry(node, struct prelim_ref, rbnode);
 730                 rb_erase(node, &tree->root);
 731
 732                 BUG_ON(ref->parent);    /* should not be a direct ref */
 733                 BUG_ON(ref->key_for_search.type);
 734                 BUG_ON(!ref->wanted_disk_byte);
 735
 736                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0);
 737                 if (IS_ERR(eb)) {
 738                         free_pref(ref);
 739                         return PTR_ERR(eb);
 740                 } else if (!extent_buffer_uptodate(eb)) {
 741                         free_pref(ref);
 742                         free_extent_buffer(eb);
 743                         return -EIO;
 744                 }
 745                 btrfs_tree_read_lock(eb);
 746                 if (btrfs_header_level(eb) == 0)
 747                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 748                 else
 749                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 750                 btrfs_tree_read_unlock(eb);
 751                 free_extent_buffer(eb);
 752                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 753                 cond_resched();
 754         }
 755         return 0;
 756 }
 757
 758 /*
 759  * add all currently queued delayed refs from this head whose seq nr is
 760  * smaller or equal that seq to the list
 761  */
 762 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 763                             struct btrfs_delayed_ref_head *head, u64 seq,
 764                             struct preftrees *preftrees, u64 *total_refs,
 765                             struct share_check *sc)
 766 {
 767         struct btrfs_delayed_ref_node *node;
 768         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 769         struct btrfs_key key;
 770         struct btrfs_key tmp_op_key;
 771         struct btrfs_key *op_key = NULL;
 772         int count;
 773         int ret = 0;
 774
 775         if (extent_op && extent_op->update_key) {
 776                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 777                 op_key = &tmp_op_key;
 778         }
 779
 780         spin_lock(&head->lock);
 781         list_for_each_entry(node, &head->ref_list, list) {
 782                 if (node->seq > seq)
 783                         continue;
 784
 785                 switch (node->action) {
 786                 case BTRFS_ADD_DELAYED_EXTENT:
 787                 case BTRFS_UPDATE_DELAYED_HEAD:
 788                         WARN_ON(1);
 789                         continue;
 790                 case BTRFS_ADD_DELAYED_REF:
 791                         count = node->ref_mod;
 792                         break;
 793                 case BTRFS_DROP_DELAYED_REF:
 794                         count = node->ref_mod * -1;
 795                         break;
 796                 default:
 797                         BUG_ON(1);
 798                 }
 799                 *total_refs += count;
 800                 switch (node->type) {
 801                 case BTRFS_TREE_BLOCK_REF_KEY: {
 802                         /* NORMAL INDIRECT METADATA backref */
 803                         struct btrfs_delayed_tree_ref *ref;
 804
 805                         ref = btrfs_delayed_node_to_tree_ref(node);
 806                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 807                                                &tmp_op_key, ref->level + 1,
 808                                                node->bytenr, count, sc,
 809                                                GFP_ATOMIC);
 810                         break;
 811                 }
 812                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 813                         /* SHARED DIRECT METADATA backref */
 814                         struct btrfs_delayed_tree_ref *ref;
 815
 816                         ref = btrfs_delayed_node_to_tree_ref(node);
 817
 818                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 819                                              ref->parent, node->bytenr, count,
 820                                              sc, GFP_ATOMIC);
 821                         break;
 822                 }
 823                 case BTRFS_EXTENT_DATA_REF_KEY: {
 824                         /* NORMAL INDIRECT DATA backref */
 825                         struct btrfs_delayed_data_ref *ref;
 826                         ref = btrfs_delayed_node_to_data_ref(node);
 827
 828                         key.objectid = ref->objectid;
 829                         key.type = BTRFS_EXTENT_DATA_KEY;
 830                         key.offset = ref->offset;
 831
 832                         /*
 833                          * Found a inum that doesn't match our known inum, we
 834                          * know it's shared.
 835                          */
 836                         if (sc && sc->inum && ref->objectid != sc->inum) {
 837                                 ret = BACKREF_FOUND_SHARED;
 838                                 goto out;
 839                         }
 840
 841                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 842                                                &key, 0, node->bytenr, count, sc,
 843                                                GFP_ATOMIC);
 844                         break;
 845                 }
 846                 case BTRFS_SHARED_DATA_REF_KEY: {
 847                         /* SHARED DIRECT FULL backref */
 848                         struct btrfs_delayed_data_ref *ref;
 849
 850                         ref = btrfs_delayed_node_to_data_ref(node);
 851
 852                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 853                                              node->bytenr, count, sc,
 854                                              GFP_ATOMIC);
 855                         break;
 856                 }
 857                 default:
 858                         WARN_ON(1);
 859                 }
 860                 /*
 861                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 862                  * refs have been checked.
 863                  */
 864                 if (ret && (ret != BACKREF_FOUND_SHARED))
 865                         break;
 866         }
 867         if (!ret)
 868                 ret = extent_is_shared(sc);
 869 out:
 870         spin_unlock(&head->lock);
 871         return ret;
 872 }
 873
 874 /*
 875  * add all inline backrefs for bytenr to the list
 876  *
 877  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 878  */
 879 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 880                            struct btrfs_path *path, u64 bytenr,
 881                            int *info_level, struct preftrees *preftrees,
 882                            u64 *total_refs, struct share_check *sc)
 883 {
 884         int ret = 0;
 885         int slot;
 886         struct extent_buffer *leaf;
 887         struct btrfs_key key;
 888         struct btrfs_key found_key;
 889         unsigned long ptr;
 890         unsigned long end;
 891         struct btrfs_extent_item *ei;
 892         u64 flags;
 893         u64 item_size;
 894
 895         /*
 896          * enumerate all inline refs
 897          */
 898         leaf = path->nodes[0];
 899         slot = path->slots[0];
 900
 901         item_size = btrfs_item_size_nr(leaf, slot);
 902         BUG_ON(item_size < sizeof(*ei));
 903
 904         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 905         flags = btrfs_extent_flags(leaf, ei);
 906         *total_refs += btrfs_extent_refs(leaf, ei);
 907         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 908
 909         ptr = (unsigned long)(ei + 1);
 910         end = (unsigned long)ei + item_size;
 911
 912         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 913             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 914                 struct btrfs_tree_block_info *info;
 915
 916                 info = (struct btrfs_tree_block_info *)ptr;
 917                 *info_level = btrfs_tree_block_level(leaf, info);
 918                 ptr += sizeof(struct btrfs_tree_block_info);
 919                 BUG_ON(ptr > end);
 920         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 921                 *info_level = found_key.offset;
 922         } else {
 923                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 924         }
 925
 926         while (ptr < end) {
 927                 struct btrfs_extent_inline_ref *iref;
 928                 u64 offset;
 929                 int type;
 930
 931                 iref = (struct btrfs_extent_inline_ref *)ptr;
 932                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 933                                                         BTRFS_REF_TYPE_ANY);
 934                 if (type == BTRFS_REF_TYPE_INVALID)
 935                         return -EINVAL;
 936
 937                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 938
 939                 switch (type) {
 940                 case BTRFS_SHARED_BLOCK_REF_KEY:
 941                         ret = add_direct_ref(fs_info, preftrees,
 942                                              *info_level + 1, offset,
 943                                              bytenr, 1, NULL, GFP_NOFS);
 944                         break;
 945                 case BTRFS_SHARED_DATA_REF_KEY: {
 946                         struct btrfs_shared_data_ref *sdref;
 947                         int count;
 948
 949                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 950                         count = btrfs_shared_data_ref_count(leaf, sdref);
 951
 952                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 953                                              bytenr, count, sc, GFP_NOFS);
 954                         break;
 955                 }
 956                 case BTRFS_TREE_BLOCK_REF_KEY:
 957                         ret = add_indirect_ref(fs_info, preftrees, offset,
 958                                                NULL, *info_level + 1,
 959                                                bytenr, 1, NULL, GFP_NOFS);
 960                         break;
 961                 case BTRFS_EXTENT_DATA_REF_KEY: {
 962                         struct btrfs_extent_data_ref *dref;
 963                         int count;
 964                         u64 root;
 965
 966                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 967                         count = btrfs_extent_data_ref_count(leaf, dref);
 968                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 969                                                                       dref);
 970                         key.type = BTRFS_EXTENT_DATA_KEY;
 971                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 972
 973                         if (sc && sc->inum && key.objectid != sc->inum) {
 974                                 ret = BACKREF_FOUND_SHARED;
 975                                 break;
 976                         }
 977
 978                         root = btrfs_extent_data_ref_root(leaf, dref);
 979
 980                         ret = add_indirect_ref(fs_info, preftrees, root,
 981                                                &key, 0, bytenr, count,
 982                                                sc, GFP_NOFS);
 983                         break;
 984                 }
 985                 default:
 986                         WARN_ON(1);
 987                 }
 988                 if (ret)
 989                         return ret;
 990                 ptr += btrfs_extent_inline_ref_size(type);
 991         }
 992
 993         return 0;
 994 }
 995
 996 /*
 997  * add all non-inline backrefs for bytenr to the list
 998  *
 999  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1000  */
1001 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1002                           struct btrfs_path *path, u64 bytenr,
1003                           int info_level, struct preftrees *preftrees,
1004                           struct share_check *sc)
1005 {
1006         struct btrfs_root *extent_root = fs_info->extent_root;
1007         int ret;
1008         int slot;
1009         struct extent_buffer *leaf;
1010         struct btrfs_key key;
1011
1012         while (1) {
1013                 ret = btrfs_next_item(extent_root, path);
1014                 if (ret < 0)
1015                         break;
1016                 if (ret) {
1017                         ret = 0;
1018                         break;
1019                 }
1020
1021                 slot = path->slots[0];
1022                 leaf = path->nodes[0];
1023                 btrfs_item_key_to_cpu(leaf, &key, slot);
1024
1025                 if (key.objectid != bytenr)
1026                         break;
1027                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1028                         continue;
1029                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1030                         break;
1031
1032                 switch (key.type) {
1033                 case BTRFS_SHARED_BLOCK_REF_KEY:
1034                         /* SHARED DIRECT METADATA backref */
1035                         ret = add_direct_ref(fs_info, preftrees,
1036                                              info_level + 1, key.offset,
1037                                              bytenr, 1, NULL, GFP_NOFS);
1038                         break;
1039                 case BTRFS_SHARED_DATA_REF_KEY: {
1040                         /* SHARED DIRECT FULL backref */
1041                         struct btrfs_shared_data_ref *sdref;
1042                         int count;
1043
1044                         sdref = btrfs_item_ptr(leaf, slot,
1045                                               struct btrfs_shared_data_ref);
1046                         count = btrfs_shared_data_ref_count(leaf, sdref);
1047                         ret = add_direct_ref(fs_info, preftrees, 0,
1048                                              key.offset, bytenr, count,
1049                                              sc, GFP_NOFS);
1050                         break;
1051                 }
1052                 case BTRFS_TREE_BLOCK_REF_KEY:
1053                         /* NORMAL INDIRECT METADATA backref */
1054                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1055                                                NULL, info_level + 1, bytenr,
1056                                                1, NULL, GFP_NOFS);
1057                         break;
1058                 case BTRFS_EXTENT_DATA_REF_KEY: {
1059                         /* NORMAL INDIRECT DATA backref */
1060                         struct btrfs_extent_data_ref *dref;
1061                         int count;
1062                         u64 root;
1063
1064                         dref = btrfs_item_ptr(leaf, slot,
1065                                               struct btrfs_extent_data_ref);
1066                         count = btrfs_extent_data_ref_count(leaf, dref);
1067                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1068                                                                       dref);
1069                         key.type = BTRFS_EXTENT_DATA_KEY;
1070                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1071
1072                         if (sc && sc->inum && key.objectid != sc->inum) {
1073                                 ret = BACKREF_FOUND_SHARED;
1074                                 break;
1075                         }
1076
1077                         root = btrfs_extent_data_ref_root(leaf, dref);
1078                         ret = add_indirect_ref(fs_info, preftrees, root,
1079                                                &key, 0, bytenr, count,
1080                                                sc, GFP_NOFS);
1081                         break;
1082                 }
1083                 default:
1084                         WARN_ON(1);
1085                 }
1086                 if (ret)
1087                         return ret;
1088
1089         }
1090
1091         return ret;
1092 }
1093
1094 /*
1095  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1096  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1097  * indirect refs to their parent bytenr.
1098  * When roots are found, they're added to the roots list
1099  *
1100  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1101  * much like trans == NULL case, the difference only lies in it will not
1102  * commit root.
1103  * The special case is for qgroup to search roots in commit_transaction().
1104  *
1105  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1106  * shared extent is detected.
1107  *
1108  * Otherwise this returns 0 for success and <0 for an error.
1109  *
1110  * FIXME some caching might speed things up
1111  */
1112 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1113                              struct btrfs_fs_info *fs_info, u64 bytenr,
1114                              u64 time_seq, struct ulist *refs,
1115                              struct ulist *roots, const u64 *extent_item_pos,
1116                              struct share_check *sc)
1117 {
1118         struct btrfs_key key;
1119         struct btrfs_path *path;
1120         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1121         struct btrfs_delayed_ref_head *head;
1122         int info_level = 0;
1123         int ret;
1124         struct prelim_ref *ref;
1125         struct rb_node *node;
1126         struct extent_inode_elem *eie = NULL;
1127         /* total of both direct AND indirect refs! */
1128         u64 total_refs = 0;
1129         struct preftrees preftrees = {
1130                 .direct = PREFTREE_INIT,
1131                 .indirect = PREFTREE_INIT,
1132                 .indirect_missing_keys = PREFTREE_INIT
1133         };
1134
1135         key.objectid = bytenr;
1136         key.offset = (u64)-1;
1137         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1138                 key.type = BTRFS_METADATA_ITEM_KEY;
1139         else
1140                 key.type = BTRFS_EXTENT_ITEM_KEY;
1141
1142         path = btrfs_alloc_path();
1143         if (!path)
1144                 return -ENOMEM;
1145         if (!trans) {
1146                 path->search_commit_root = 1;
1147                 path->skip_locking = 1;
1148         }
1149
1150         if (time_seq == SEQ_LAST)
1151                 path->skip_locking = 1;
1152
1153         /*
1154          * grab both a lock on the path and a lock on the delayed ref head.
1155          * We need both to get a consistent picture of how the refs look
1156          * at a specified point in time
1157          */
1158 again:
1159         head = NULL;
1160
1161         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1162         if (ret < 0)
1163                 goto out;
1164         BUG_ON(ret == 0);
1165
1166 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1167         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1168             time_seq != SEQ_LAST) {
1169 #else
1170         if (trans && time_seq != SEQ_LAST) {
1171 #endif
1172                 /*
1173                  * look if there are updates for this ref queued and lock the
1174                  * head
1175                  */
1176                 delayed_refs = &trans->transaction->delayed_refs;
1177                 spin_lock(&delayed_refs->lock);
1178                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1179                 if (head) {
1180                         if (!mutex_trylock(&head->mutex)) {
1181                                 refcount_inc(&head->node.refs);
1182                                 spin_unlock(&delayed_refs->lock);
1183
1184                                 btrfs_release_path(path);
1185
1186                                 /*
1187                                  * Mutex was contended, block until it's
1188                                  * released and try again
1189                                  */
1190                                 mutex_lock(&head->mutex);
1191                                 mutex_unlock(&head->mutex);
1192                                 btrfs_put_delayed_ref(&head->node);
1193                                 goto again;
1194                         }
1195                         spin_unlock(&delayed_refs->lock);
1196                         ret = add_delayed_refs(fs_info, head, time_seq,
1197                                                &preftrees, &total_refs, sc);
1198                         mutex_unlock(&head->mutex);
1199                         if (ret)
1200                                 goto out;
1201                 } else {
1202                         spin_unlock(&delayed_refs->lock);
1203                 }
1204         }
1205
1206         if (path->slots[0]) {
1207                 struct extent_buffer *leaf;
1208                 int slot;
1209
1210                 path->slots[0]--;
1211                 leaf = path->nodes[0];
1212                 slot = path->slots[0];
1213                 btrfs_item_key_to_cpu(leaf, &key, slot);
1214                 if (key.objectid == bytenr &&
1215                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1216                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1217                         ret = add_inline_refs(fs_info, path, bytenr,
1218                                               &info_level, &preftrees,
1219                                               &total_refs, sc);
1220                         if (ret)
1221                                 goto out;
1222                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1223                                              &preftrees, sc);
1224                         if (ret)
1225                                 goto out;
1226                 }
1227         }
1228
1229         btrfs_release_path(path);
1230
1231         ret = add_missing_keys(fs_info, &preftrees);
1232         if (ret)
1233                 goto out;
1234
1235         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1236
1237         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1238                                     extent_item_pos, total_refs, sc);
1239         if (ret)
1240                 goto out;
1241
1242         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1243
1244         /*
1245          * This walks the tree of merged and resolved refs. Tree blocks are
1246          * read in as needed. Unique entries are added to the ulist, and
1247          * the list of found roots is updated.
1248          *
1249          * We release the entire tree in one go before returning.
1250          */
1251         node = rb_first(&preftrees.direct.root);
1252         while (node) {
1253                 ref = rb_entry(node, struct prelim_ref, rbnode);
1254                 node = rb_next(&ref->rbnode);
1255                 /*
1256                  * ref->count < 0 can happen here if there are delayed
1257                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1258                  * prelim_ref_insert() relies on this when merging
1259                  * identical refs to keep the overall count correct.
1260                  * prelim_ref_insert() will merge only those refs
1261                  * which compare identically.  Any refs having
1262                  * e.g. different offsets would not be merged,
1263                  * and would retain their original ref->count < 0.
1264                  */
1265                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1266                         if (sc && sc->root_objectid &&
1267                             ref->root_id != sc->root_objectid) {
1268                                 ret = BACKREF_FOUND_SHARED;
1269                                 goto out;
1270                         }
1271
1272                         /* no parent == root of tree */
1273                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1274                         if (ret < 0)
1275                                 goto out;
1276                 }
1277                 if (ref->count && ref->parent) {
1278                         if (extent_item_pos && !ref->inode_list &&
1279                             ref->level == 0) {
1280                                 struct extent_buffer *eb;
1281
1282                                 eb = read_tree_block(fs_info, ref->parent, 0);
1283                                 if (IS_ERR(eb)) {
1284                                         ret = PTR_ERR(eb);
1285                                         goto out;
1286                                 } else if (!extent_buffer_uptodate(eb)) {
1287                                         free_extent_buffer(eb);
1288                                         ret = -EIO;
1289                                         goto out;
1290                                 }
1291                                 btrfs_tree_read_lock(eb);
1292                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1293                                 ret = find_extent_in_eb(eb, bytenr,
1294                                                         *extent_item_pos, &eie);
1295                                 btrfs_tree_read_unlock_blocking(eb);
1296                                 free_extent_buffer(eb);
1297                                 if (ret < 0)
1298                                         goto out;
1299                                 ref->inode_list = eie;
1300                         }
1301                         ret = ulist_add_merge_ptr(refs, ref->parent,
1302                                                   ref->inode_list,
1303                                                   (void **)&eie, GFP_NOFS);
1304                         if (ret < 0)
1305                                 goto out;
1306                         if (!ret && extent_item_pos) {
1307                                 /*
1308                                  * we've recorded that parent, so we must extend
1309                                  * its inode list here
1310                                  */
1311                                 BUG_ON(!eie);
1312                                 while (eie->next)
1313                                         eie = eie->next;
1314                                 eie->next = ref->inode_list;
1315                         }
1316                         eie = NULL;
1317                 }
1318                 cond_resched();
1319         }
1320
1321 out:
1322         btrfs_free_path(path);
1323
1324         prelim_release(&preftrees.direct);
1325         prelim_release(&preftrees.indirect);
1326         prelim_release(&preftrees.indirect_missing_keys);
1327
1328         if (ret < 0)
1329                 free_inode_elem_list(eie);
1330         return ret;
1331 }
1332
1333 static void free_leaf_list(struct ulist *blocks)
1334 {
1335         struct ulist_node *node = NULL;
1336         struct extent_inode_elem *eie;
1337         struct ulist_iterator uiter;
1338
1339         ULIST_ITER_INIT(&uiter);
1340         while ((node = ulist_next(blocks, &uiter))) {
1341                 if (!node->aux)
1342                         continue;
1343                 eie = unode_aux_to_inode_list(node);
1344                 free_inode_elem_list(eie);
1345                 node->aux = 0;
1346         }
1347
1348         ulist_free(blocks);
1349 }
1350
1351 /*
1352  * Finds all leafs with a reference to the specified combination of bytenr and
1353  * offset. key_list_head will point to a list of corresponding keys (caller must
1354  * free each list element). The leafs will be stored in the leafs ulist, which
1355  * must be freed with ulist_free.
1356  *
1357  * returns 0 on success, <0 on error
1358  */
1359 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1360                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1361                                 u64 time_seq, struct ulist **leafs,
1362                                 const u64 *extent_item_pos)
1363 {
1364         int ret;
1365
1366         *leafs = ulist_alloc(GFP_NOFS);
1367         if (!*leafs)
1368                 return -ENOMEM;
1369
1370         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1371                                 *leafs, NULL, extent_item_pos, NULL);
1372         if (ret < 0 && ret != -ENOENT) {
1373                 free_leaf_list(*leafs);
1374                 return ret;
1375         }
1376
1377         return 0;
1378 }
1379
1380 /*
1381  * walk all backrefs for a given extent to find all roots that reference this
1382  * extent. Walking a backref means finding all extents that reference this
1383  * extent and in turn walk the backrefs of those, too. Naturally this is a
1384  * recursive process, but here it is implemented in an iterative fashion: We
1385  * find all referencing extents for the extent in question and put them on a
1386  * list. In turn, we find all referencing extents for those, further appending
1387  * to the list. The way we iterate the list allows adding more elements after
1388  * the current while iterating. The process stops when we reach the end of the
1389  * list. Found roots are added to the roots list.
1390  *
1391  * returns 0 on success, < 0 on error.
1392  */
1393 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1394                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1395                                      u64 time_seq, struct ulist **roots)
1396 {
1397         struct ulist *tmp;
1398         struct ulist_node *node = NULL;
1399         struct ulist_iterator uiter;
1400         int ret;
1401
1402         tmp = ulist_alloc(GFP_NOFS);
1403         if (!tmp)
1404                 return -ENOMEM;
1405         *roots = ulist_alloc(GFP_NOFS);
1406         if (!*roots) {
1407                 ulist_free(tmp);
1408                 return -ENOMEM;
1409         }
1410
1411         ULIST_ITER_INIT(&uiter);
1412         while (1) {
1413                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1414                                         tmp, *roots, NULL, NULL);
1415                 if (ret < 0 && ret != -ENOENT) {
1416                         ulist_free(tmp);
1417                         ulist_free(*roots);
1418                         return ret;
1419                 }
1420                 node = ulist_next(tmp, &uiter);
1421                 if (!node)
1422                         break;
1423                 bytenr = node->val;
1424                 cond_resched();
1425         }
1426
1427         ulist_free(tmp);
1428         return 0;
1429 }
1430
1431 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1432                          struct btrfs_fs_info *fs_info, u64 bytenr,
1433                          u64 time_seq, struct ulist **roots)
1434 {
1435         int ret;
1436
1437         if (!trans)
1438                 down_read(&fs_info->commit_root_sem);
1439         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1440                                         time_seq, roots);
1441         if (!trans)
1442                 up_read(&fs_info->commit_root_sem);
1443         return ret;
1444 }
1445
1446 /**
1447  * btrfs_check_shared - tell us whether an extent is shared
1448  *
1449  * btrfs_check_shared uses the backref walking code but will short
1450  * circuit as soon as it finds a root or inode that doesn't match the
1451  * one passed in. This provides a significant performance benefit for
1452  * callers (such as fiemap) which want to know whether the extent is
1453  * shared but do not need a ref count.
1454  *
1455  * This attempts to allocate a transaction in order to account for
1456  * delayed refs, but continues on even when the alloc fails.
1457  *
1458  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1459  */
1460 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1461 {
1462         struct btrfs_fs_info *fs_info = root->fs_info;
1463         struct btrfs_trans_handle *trans;
1464         struct ulist *tmp = NULL;
1465         struct ulist *roots = NULL;
1466         struct ulist_iterator uiter;
1467         struct ulist_node *node;
1468         struct seq_list elem = SEQ_LIST_INIT(elem);
1469         int ret = 0;
1470         struct share_check shared = {
1471                 .root_objectid = root->objectid,
1472                 .inum = inum,
1473                 .share_count = 0,
1474         };
1475
1476         tmp = ulist_alloc(GFP_NOFS);
1477         roots = ulist_alloc(GFP_NOFS);
1478         if (!tmp || !roots) {
1479                 ulist_free(tmp);
1480                 ulist_free(roots);
1481                 return -ENOMEM;
1482         }
1483
1484         trans = btrfs_join_transaction(root);
1485         if (IS_ERR(trans)) {
1486                 trans = NULL;
1487                 down_read(&fs_info->commit_root_sem);
1488         } else {
1489                 btrfs_get_tree_mod_seq(fs_info, &elem);
1490         }
1491
1492         ULIST_ITER_INIT(&uiter);
1493         while (1) {
1494                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1495                                         roots, NULL, &shared);
1496                 if (ret == BACKREF_FOUND_SHARED) {
1497                         /* this is the only condition under which we return 1 */
1498                         ret = 1;
1499                         break;
1500                 }
1501                 if (ret < 0 && ret != -ENOENT)
1502                         break;
1503                 ret = 0;
1504                 node = ulist_next(tmp, &uiter);
1505                 if (!node)
1506                         break;
1507                 bytenr = node->val;
1508                 shared.share_count = 0;
1509                 cond_resched();
1510         }
1511
1512         if (trans) {
1513                 btrfs_put_tree_mod_seq(fs_info, &elem);
1514                 btrfs_end_transaction(trans);
1515         } else {
1516                 up_read(&fs_info->commit_root_sem);
1517         }
1518         ulist_free(tmp);
1519         ulist_free(roots);
1520         return ret;
1521 }
1522
1523 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1524                           u64 start_off, struct btrfs_path *path,
1525                           struct btrfs_inode_extref **ret_extref,
1526                           u64 *found_off)
1527 {
1528         int ret, slot;
1529         struct btrfs_key key;
1530         struct btrfs_key found_key;
1531         struct btrfs_inode_extref *extref;
1532         const struct extent_buffer *leaf;
1533         unsigned long ptr;
1534
1535         key.objectid = inode_objectid;
1536         key.type = BTRFS_INODE_EXTREF_KEY;
1537         key.offset = start_off;
1538
1539         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1540         if (ret < 0)
1541                 return ret;
1542
1543         while (1) {
1544                 leaf = path->nodes[0];
1545                 slot = path->slots[0];
1546                 if (slot >= btrfs_header_nritems(leaf)) {
1547                         /*
1548                          * If the item at offset is not found,
1549                          * btrfs_search_slot will point us to the slot
1550                          * where it should be inserted. In our case
1551                          * that will be the slot directly before the
1552                          * next INODE_REF_KEY_V2 item. In the case
1553                          * that we're pointing to the last slot in a
1554                          * leaf, we must move one leaf over.
1555                          */
1556                         ret = btrfs_next_leaf(root, path);
1557                         if (ret) {
1558                                 if (ret >= 1)
1559                                         ret = -ENOENT;
1560                                 break;
1561                         }
1562                         continue;
1563                 }
1564
1565                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1566
1567                 /*
1568                  * Check that we're still looking at an extended ref key for
1569                  * this particular objectid. If we have different
1570                  * objectid or type then there are no more to be found
1571                  * in the tree and we can exit.
1572                  */
1573                 ret = -ENOENT;
1574                 if (found_key.objectid != inode_objectid)
1575                         break;
1576                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1577                         break;
1578
1579                 ret = 0;
1580                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1581                 extref = (struct btrfs_inode_extref *)ptr;
1582                 *ret_extref = extref;
1583                 if (found_off)
1584                         *found_off = found_key.offset;
1585                 break;
1586         }
1587
1588         return ret;
1589 }
1590
1591 /*
1592  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1593  * Elements of the path are separated by '/' and the path is guaranteed to be
1594  * 0-terminated. the path is only given within the current file system.
1595  * Therefore, it never starts with a '/'. the caller is responsible to provide
1596  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1597  * the start point of the resulting string is returned. this pointer is within
1598  * dest, normally.
1599  * in case the path buffer would overflow, the pointer is decremented further
1600  * as if output was written to the buffer, though no more output is actually
1601  * generated. that way, the caller can determine how much space would be
1602  * required for the path to fit into the buffer. in that case, the returned
1603  * value will be smaller than dest. callers must check this!
1604  */
1605 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1606                         u32 name_len, unsigned long name_off,
1607                         struct extent_buffer *eb_in, u64 parent,
1608                         char *dest, u32 size)
1609 {
1610         int slot;
1611         u64 next_inum;
1612         int ret;
1613         s64 bytes_left = ((s64)size) - 1;
1614         struct extent_buffer *eb = eb_in;
1615         struct btrfs_key found_key;
1616         int leave_spinning = path->leave_spinning;
1617         struct btrfs_inode_ref *iref;
1618
1619         if (bytes_left >= 0)
1620                 dest[bytes_left] = '\0';
1621
1622         path->leave_spinning = 1;
1623         while (1) {
1624                 bytes_left -= name_len;
1625                 if (bytes_left >= 0)
1626                         read_extent_buffer(eb, dest + bytes_left,
1627                                            name_off, name_len);
1628                 if (eb != eb_in) {
1629                         if (!path->skip_locking)
1630                                 btrfs_tree_read_unlock_blocking(eb);
1631                         free_extent_buffer(eb);
1632                 }
1633                 ret = btrfs_find_item(fs_root, path, parent, 0,
1634                                 BTRFS_INODE_REF_KEY, &found_key);
1635                 if (ret > 0)
1636                         ret = -ENOENT;
1637                 if (ret)
1638                         break;
1639
1640                 next_inum = found_key.offset;
1641
1642                 /* regular exit ahead */
1643                 if (parent == next_inum)
1644                         break;
1645
1646                 slot = path->slots[0];
1647                 eb = path->nodes[0];
1648                 /* make sure we can use eb after releasing the path */
1649                 if (eb != eb_in) {
1650                         if (!path->skip_locking)
1651                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1652                         path->nodes[0] = NULL;
1653                         path->locks[0] = 0;
1654                 }
1655                 btrfs_release_path(path);
1656                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1657
1658                 name_len = btrfs_inode_ref_name_len(eb, iref);
1659                 name_off = (unsigned long)(iref + 1);
1660
1661                 parent = next_inum;
1662                 --bytes_left;
1663                 if (bytes_left >= 0)
1664                         dest[bytes_left] = '/';
1665         }
1666
1667         btrfs_release_path(path);
1668         path->leave_spinning = leave_spinning;
1669
1670         if (ret)
1671                 return ERR_PTR(ret);
1672
1673         return dest + bytes_left;
1674 }
1675
1676 /*
1677  * this makes the path point to (logical EXTENT_ITEM *)
1678  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1679  * tree blocks and <0 on error.
1680  */
1681 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1682                         struct btrfs_path *path, struct btrfs_key *found_key,
1683                         u64 *flags_ret)
1684 {
1685         int ret;
1686         u64 flags;
1687         u64 size = 0;
1688         u32 item_size;
1689         const struct extent_buffer *eb;
1690         struct btrfs_extent_item *ei;
1691         struct btrfs_key key;
1692
1693         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1694                 key.type = BTRFS_METADATA_ITEM_KEY;
1695         else
1696                 key.type = BTRFS_EXTENT_ITEM_KEY;
1697         key.objectid = logical;
1698         key.offset = (u64)-1;
1699
1700         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1701         if (ret < 0)
1702                 return ret;
1703
1704         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1705         if (ret) {
1706                 if (ret > 0)
1707                         ret = -ENOENT;
1708                 return ret;
1709         }
1710         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1711         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1712                 size = fs_info->nodesize;
1713         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1714                 size = found_key->offset;
1715
1716         if (found_key->objectid > logical ||
1717             found_key->objectid + size <= logical) {
1718                 btrfs_debug(fs_info,
1719                         "logical %llu is not within any extent", logical);
1720                 return -ENOENT;
1721         }
1722
1723         eb = path->nodes[0];
1724         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1725         BUG_ON(item_size < sizeof(*ei));
1726
1727         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1728         flags = btrfs_extent_flags(eb, ei);
1729
1730         btrfs_debug(fs_info,
1731                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1732                  logical, logical - found_key->objectid, found_key->objectid,
1733                  found_key->offset, flags, item_size);
1734
1735         WARN_ON(!flags_ret);
1736         if (flags_ret) {
1737                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1738                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1739                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1740                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1741                 else
1742                         BUG_ON(1);
1743                 return 0;
1744         }
1745
1746         return -EIO;
1747 }
1748
1749 /*
1750  * helper function to iterate extent inline refs. ptr must point to a 0 value
1751  * for the first call and may be modified. it is used to track state.
1752  * if more refs exist, 0 is returned and the next call to
1753  * get_extent_inline_ref must pass the modified ptr parameter to get the
1754  * next ref. after the last ref was processed, 1 is returned.
1755  * returns <0 on error
1756  */
1757 static int get_extent_inline_ref(unsigned long *ptr,
1758                                  const struct extent_buffer *eb,
1759                                  const struct btrfs_key *key,
1760                                  const struct btrfs_extent_item *ei,
1761                                  u32 item_size,
1762                                  struct btrfs_extent_inline_ref **out_eiref,
1763                                  int *out_type)
1764 {
1765         unsigned long end;
1766         u64 flags;
1767         struct btrfs_tree_block_info *info;
1768
1769         if (!*ptr) {
1770                 /* first call */
1771                 flags = btrfs_extent_flags(eb, ei);
1772                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1773                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1774                                 /* a skinny metadata extent */
1775                                 *out_eiref =
1776                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1777                         } else {
1778                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1779                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1780                                 *out_eiref =
1781                                    (struct btrfs_extent_inline_ref *)(info + 1);
1782                         }
1783                 } else {
1784                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1785                 }
1786                 *ptr = (unsigned long)*out_eiref;
1787                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1788                         return -ENOENT;
1789         }
1790
1791         end = (unsigned long)ei + item_size;
1792         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1793         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1794                                                      BTRFS_REF_TYPE_ANY);
1795         if (*out_type == BTRFS_REF_TYPE_INVALID)
1796                 return -EINVAL;
1797
1798         *ptr += btrfs_extent_inline_ref_size(*out_type);
1799         WARN_ON(*ptr > end);
1800         if (*ptr == end)
1801                 return 1; /* last */
1802
1803         return 0;
1804 }
1805
1806 /*
1807  * reads the tree block backref for an extent. tree level and root are returned
1808  * through out_level and out_root. ptr must point to a 0 value for the first
1809  * call and may be modified (see get_extent_inline_ref comment).
1810  * returns 0 if data was provided, 1 if there was no more data to provide or
1811  * <0 on error.
1812  */
1813 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1814                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1815                             u32 item_size, u64 *out_root, u8 *out_level)
1816 {
1817         int ret;
1818         int type;
1819         struct btrfs_extent_inline_ref *eiref;
1820
1821         if (*ptr == (unsigned long)-1)
1822                 return 1;
1823
1824         while (1) {
1825                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1826                                               &eiref, &type);
1827                 if (ret < 0)
1828                         return ret;
1829
1830                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1831                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1832                         break;
1833
1834                 if (ret == 1)
1835                         return 1;
1836         }
1837
1838         /* we can treat both ref types equally here */
1839         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1840
1841         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1842                 struct btrfs_tree_block_info *info;
1843
1844                 info = (struct btrfs_tree_block_info *)(ei + 1);
1845                 *out_level = btrfs_tree_block_level(eb, info);
1846         } else {
1847                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1848                 *out_level = (u8)key->offset;
1849         }
1850
1851         if (ret == 1)
1852                 *ptr = (unsigned long)-1;
1853
1854         return 0;
1855 }
1856
1857 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1858                              struct extent_inode_elem *inode_list,
1859                              u64 root, u64 extent_item_objectid,
1860                              iterate_extent_inodes_t *iterate, void *ctx)
1861 {
1862         struct extent_inode_elem *eie;
1863         int ret = 0;
1864
1865         for (eie = inode_list; eie; eie = eie->next) {
1866                 btrfs_debug(fs_info,
1867                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1868                             extent_item_objectid, eie->inum,
1869                             eie->offset, root);
1870                 ret = iterate(eie->inum, eie->offset, root, ctx);
1871                 if (ret) {
1872                         btrfs_debug(fs_info,
1873                                     "stopping iteration for %llu due to ret=%d",
1874                                     extent_item_objectid, ret);
1875                         break;
1876                 }
1877         }
1878
1879         return ret;
1880 }
1881
1882 /*
1883  * calls iterate() for every inode that references the extent identified by
1884  * the given parameters.
1885  * when the iterator function returns a non-zero value, iteration stops.
1886  */
1887 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1888                                 u64 extent_item_objectid, u64 extent_item_pos,
1889                                 int search_commit_root,
1890                                 iterate_extent_inodes_t *iterate, void *ctx)
1891 {
1892         int ret;
1893         struct btrfs_trans_handle *trans = NULL;
1894         struct ulist *refs = NULL;
1895         struct ulist *roots = NULL;
1896         struct ulist_node *ref_node = NULL;
1897         struct ulist_node *root_node = NULL;
1898         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1899         struct ulist_iterator ref_uiter;
1900         struct ulist_iterator root_uiter;
1901
1902         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1903                         extent_item_objectid);
1904
1905         if (!search_commit_root) {
1906                 trans = btrfs_join_transaction(fs_info->extent_root);
1907                 if (IS_ERR(trans))
1908                         return PTR_ERR(trans);
1909                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1910         } else {
1911                 down_read(&fs_info->commit_root_sem);
1912         }
1913
1914         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1915                                    tree_mod_seq_elem.seq, &refs,
1916                                    &extent_item_pos);
1917         if (ret)
1918                 goto out;
1919
1920         ULIST_ITER_INIT(&ref_uiter);
1921         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1922                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1923                                                 tree_mod_seq_elem.seq, &roots);
1924                 if (ret)
1925                         break;
1926                 ULIST_ITER_INIT(&root_uiter);
1927                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1928                         btrfs_debug(fs_info,
1929                                     "root %llu references leaf %llu, data list %#llx",
1930                                     root_node->val, ref_node->val,
1931                                     ref_node->aux);
1932                         ret = iterate_leaf_refs(fs_info,
1933                                                 (struct extent_inode_elem *)
1934                                                 (uintptr_t)ref_node->aux,
1935                                                 root_node->val,
1936                                                 extent_item_objectid,
1937                                                 iterate, ctx);
1938                 }
1939                 ulist_free(roots);
1940         }
1941
1942         free_leaf_list(refs);
1943 out:
1944         if (!search_commit_root) {
1945                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1946                 btrfs_end_transaction(trans);
1947         } else {
1948                 up_read(&fs_info->commit_root_sem);
1949         }
1950
1951         return ret;
1952 }
1953
1954 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1955                                 struct btrfs_path *path,
1956                                 iterate_extent_inodes_t *iterate, void *ctx)
1957 {
1958         int ret;
1959         u64 extent_item_pos;
1960         u64 flags = 0;
1961         struct btrfs_key found_key;
1962         int search_commit_root = path->search_commit_root;
1963
1964         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1965         btrfs_release_path(path);
1966         if (ret < 0)
1967                 return ret;
1968         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1969                 return -EINVAL;
1970
1971         extent_item_pos = logical - found_key.objectid;
1972         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1973                                         extent_item_pos, search_commit_root,
1974                                         iterate, ctx);
1975
1976         return ret;
1977 }
1978
1979 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1980                               struct extent_buffer *eb, void *ctx);
1981
1982 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1983                               struct btrfs_path *path,
1984                               iterate_irefs_t *iterate, void *ctx)
1985 {
1986         int ret = 0;
1987         int slot;
1988         u32 cur;
1989         u32 len;
1990         u32 name_len;
1991         u64 parent = 0;
1992         int found = 0;
1993         struct extent_buffer *eb;
1994         struct btrfs_item *item;
1995         struct btrfs_inode_ref *iref;
1996         struct btrfs_key found_key;
1997
1998         while (!ret) {
1999                 ret = btrfs_find_item(fs_root, path, inum,
2000                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2001                                 &found_key);
2002
2003                 if (ret < 0)
2004                         break;
2005                 if (ret) {
2006                         ret = found ? 0 : -ENOENT;
2007                         break;
2008                 }
2009                 ++found;
2010
2011                 parent = found_key.offset;
2012                 slot = path->slots[0];
2013                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2014                 if (!eb) {
2015                         ret = -ENOMEM;
2016                         break;
2017                 }
2018                 extent_buffer_get(eb);
2019                 btrfs_tree_read_lock(eb);
2020                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2021                 btrfs_release_path(path);
2022
2023                 item = btrfs_item_nr(slot);
2024                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2025
2026                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2027                         name_len = btrfs_inode_ref_name_len(eb, iref);
2028                         /* path must be released before calling iterate()! */
2029                         btrfs_debug(fs_root->fs_info,
2030                                 "following ref at offset %u for inode %llu in tree %llu",
2031                                 cur, found_key.objectid, fs_root->objectid);
2032                         ret = iterate(parent, name_len,
2033                                       (unsigned long)(iref + 1), eb, ctx);
2034                         if (ret)
2035                                 break;
2036                         len = sizeof(*iref) + name_len;
2037                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2038                 }
2039                 btrfs_tree_read_unlock_blocking(eb);
2040                 free_extent_buffer(eb);
2041         }
2042
2043         btrfs_release_path(path);
2044
2045         return ret;
2046 }
2047
2048 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2049                                  struct btrfs_path *path,
2050                                  iterate_irefs_t *iterate, void *ctx)
2051 {
2052         int ret;
2053         int slot;
2054         u64 offset = 0;
2055         u64 parent;
2056         int found = 0;
2057         struct extent_buffer *eb;
2058         struct btrfs_inode_extref *extref;
2059         u32 item_size;
2060         u32 cur_offset;
2061         unsigned long ptr;
2062
2063         while (1) {
2064                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2065                                             &offset);
2066                 if (ret < 0)
2067                         break;
2068                 if (ret) {
2069                         ret = found ? 0 : -ENOENT;
2070                         break;
2071                 }
2072                 ++found;
2073
2074                 slot = path->slots[0];
2075                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2076                 if (!eb) {
2077                         ret = -ENOMEM;
2078                         break;
2079                 }
2080                 extent_buffer_get(eb);
2081
2082                 btrfs_tree_read_lock(eb);
2083                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2084                 btrfs_release_path(path);
2085
2086                 item_size = btrfs_item_size_nr(eb, slot);
2087                 ptr = btrfs_item_ptr_offset(eb, slot);
2088                 cur_offset = 0;
2089
2090                 while (cur_offset < item_size) {
2091                         u32 name_len;
2092
2093                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2094                         parent = btrfs_inode_extref_parent(eb, extref);
2095                         name_len = btrfs_inode_extref_name_len(eb, extref);
2096                         ret = iterate(parent, name_len,
2097                                       (unsigned long)&extref->name, eb, ctx);
2098                         if (ret)
2099                                 break;
2100
2101                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2102                         cur_offset += sizeof(*extref);
2103                 }
2104                 btrfs_tree_read_unlock_blocking(eb);
2105                 free_extent_buffer(eb);
2106
2107                 offset++;
2108         }
2109
2110         btrfs_release_path(path);
2111
2112         return ret;
2113 }
2114
2115 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2116                          struct btrfs_path *path, iterate_irefs_t *iterate,
2117                          void *ctx)
2118 {
2119         int ret;
2120         int found_refs = 0;
2121
2122         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2123         if (!ret)
2124                 ++found_refs;
2125         else if (ret != -ENOENT)
2126                 return ret;
2127
2128         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2129         if (ret == -ENOENT && found_refs)
2130                 return 0;
2131
2132         return ret;
2133 }
2134
2135 /*
2136  * returns 0 if the path could be dumped (probably truncated)
2137  * returns <0 in case of an error
2138  */
2139 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2140                          struct extent_buffer *eb, void *ctx)
2141 {
2142         struct inode_fs_paths *ipath = ctx;
2143         char *fspath;
2144         char *fspath_min;
2145         int i = ipath->fspath->elem_cnt;
2146         const int s_ptr = sizeof(char *);
2147         u32 bytes_left;
2148
2149         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2150                                         ipath->fspath->bytes_left - s_ptr : 0;
2151
2152         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2153         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2154                                    name_off, eb, inum, fspath_min, bytes_left);
2155         if (IS_ERR(fspath))
2156                 return PTR_ERR(fspath);
2157
2158         if (fspath > fspath_min) {
2159                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2160                 ++ipath->fspath->elem_cnt;
2161                 ipath->fspath->bytes_left = fspath - fspath_min;
2162         } else {
2163                 ++ipath->fspath->elem_missed;
2164                 ipath->fspath->bytes_missing += fspath_min - fspath;
2165                 ipath->fspath->bytes_left = 0;
2166         }
2167
2168         return 0;
2169 }
2170
2171 /*
2172  * this dumps all file system paths to the inode into the ipath struct, provided
2173  * is has been created large enough. each path is zero-terminated and accessed
2174  * from ipath->fspath->val[i].
2175  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2176  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2177  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2178  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2179  * have been needed to return all paths.
2180  */
2181 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2182 {
2183         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2184                              inode_to_path, ipath);
2185 }
2186
2187 struct btrfs_data_container *init_data_container(u32 total_bytes)
2188 {
2189         struct btrfs_data_container *data;
2190         size_t alloc_bytes;
2191
2192         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2193         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2194         if (!data)
2195                 return ERR_PTR(-ENOMEM);
2196
2197         if (total_bytes >= sizeof(*data)) {
2198                 data->bytes_left = total_bytes - sizeof(*data);
2199                 data->bytes_missing = 0;
2200         } else {
2201                 data->bytes_missing = sizeof(*data) - total_bytes;
2202                 data->bytes_left = 0;
2203         }
2204
2205         data->elem_cnt = 0;
2206         data->elem_missed = 0;
2207
2208         return data;
2209 }
2210
2211 /*
2212  * allocates space to return multiple file system paths for an inode.
2213  * total_bytes to allocate are passed, note that space usable for actual path
2214  * information will be total_bytes - sizeof(struct inode_fs_paths).
2215  * the returned pointer must be freed with free_ipath() in the end.
2216  */
2217 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2218                                         struct btrfs_path *path)
2219 {
2220         struct inode_fs_paths *ifp;
2221         struct btrfs_data_container *fspath;
2222
2223         fspath = init_data_container(total_bytes);
2224         if (IS_ERR(fspath))
2225                 return (void *)fspath;
2226
2227         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2228         if (!ifp) {
2229                 kvfree(fspath);
2230                 return ERR_PTR(-ENOMEM);
2231         }
2232
2233         ifp->btrfs_path = path;
2234         ifp->fspath = fspath;
2235         ifp->fs_root = fs_root;
2236
2237         return ifp;
2238 }
2239
2240 void free_ipath(struct inode_fs_paths *ipath)
2241 {
2242         if (!ipath)
2243                 return;
2244         kvfree(ipath->fspath);
2245         kfree(ipath);
2246 }