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, bool lock)
 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                 if (lock)
 746                         btrfs_tree_read_lock(eb);
 747                 if (btrfs_header_level(eb) == 0)
 748                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 749                 else
 750                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 751                 if (lock)
 752                         btrfs_tree_read_unlock(eb);
 753                 free_extent_buffer(eb);
 754                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
 755                 cond_resched();
 756         }
 757         return 0;
 758 }
 759
 760 /*
 761  * add all currently queued delayed refs from this head whose seq nr is
 762  * smaller or equal that seq to the list
 763  */
 764 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
 765                             struct btrfs_delayed_ref_head *head, u64 seq,
 766                             struct preftrees *preftrees, u64 *total_refs,
 767                             struct share_check *sc)
 768 {
 769         struct btrfs_delayed_ref_node *node;
 770         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 771         struct btrfs_key key;
 772         struct btrfs_key tmp_op_key;
 773         struct btrfs_key *op_key = NULL;
 774         int count;
 775         int ret = 0;
 776
 777         if (extent_op && extent_op->update_key) {
 778                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
 779                 op_key = &tmp_op_key;
 780         }
 781
 782         spin_lock(&head->lock);
 783         list_for_each_entry(node, &head->ref_list, list) {
 784                 if (node->seq > seq)
 785                         continue;
 786
 787                 switch (node->action) {
 788                 case BTRFS_ADD_DELAYED_EXTENT:
 789                 case BTRFS_UPDATE_DELAYED_HEAD:
 790                         WARN_ON(1);
 791                         continue;
 792                 case BTRFS_ADD_DELAYED_REF:
 793                         count = node->ref_mod;
 794                         break;
 795                 case BTRFS_DROP_DELAYED_REF:
 796                         count = node->ref_mod * -1;
 797                         break;
 798                 default:
 799                         BUG_ON(1);
 800                 }
 801                 *total_refs += count;
 802                 switch (node->type) {
 803                 case BTRFS_TREE_BLOCK_REF_KEY: {
 804                         /* NORMAL INDIRECT METADATA backref */
 805                         struct btrfs_delayed_tree_ref *ref;
 806
 807                         ref = btrfs_delayed_node_to_tree_ref(node);
 808                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 809                                                &tmp_op_key, ref->level + 1,
 810                                                node->bytenr, count, sc,
 811                                                GFP_ATOMIC);
 812                         break;
 813                 }
 814                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 815                         /* SHARED DIRECT METADATA backref */
 816                         struct btrfs_delayed_tree_ref *ref;
 817
 818                         ref = btrfs_delayed_node_to_tree_ref(node);
 819
 820                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
 821                                              ref->parent, node->bytenr, count,
 822                                              sc, GFP_ATOMIC);
 823                         break;
 824                 }
 825                 case BTRFS_EXTENT_DATA_REF_KEY: {
 826                         /* NORMAL INDIRECT DATA backref */
 827                         struct btrfs_delayed_data_ref *ref;
 828                         ref = btrfs_delayed_node_to_data_ref(node);
 829
 830                         key.objectid = ref->objectid;
 831                         key.type = BTRFS_EXTENT_DATA_KEY;
 832                         key.offset = ref->offset;
 833
 834                         /*
 835                          * Found a inum that doesn't match our known inum, we
 836                          * know it's shared.
 837                          */
 838                         if (sc && sc->inum && ref->objectid != sc->inum) {
 839                                 ret = BACKREF_FOUND_SHARED;
 840                                 goto out;
 841                         }
 842
 843                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
 844                                                &key, 0, node->bytenr, count, sc,
 845                                                GFP_ATOMIC);
 846                         break;
 847                 }
 848                 case BTRFS_SHARED_DATA_REF_KEY: {
 849                         /* SHARED DIRECT FULL backref */
 850                         struct btrfs_delayed_data_ref *ref;
 851
 852                         ref = btrfs_delayed_node_to_data_ref(node);
 853
 854                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
 855                                              node->bytenr, count, sc,
 856                                              GFP_ATOMIC);
 857                         break;
 858                 }
 859                 default:
 860                         WARN_ON(1);
 861                 }
 862                 /*
 863                  * We must ignore BACKREF_FOUND_SHARED until all delayed
 864                  * refs have been checked.
 865                  */
 866                 if (ret && (ret != BACKREF_FOUND_SHARED))
 867                         break;
 868         }
 869         if (!ret)
 870                 ret = extent_is_shared(sc);
 871 out:
 872         spin_unlock(&head->lock);
 873         return ret;
 874 }
 875
 876 /*
 877  * add all inline backrefs for bytenr to the list
 878  *
 879  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
 880  */
 881 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
 882                            struct btrfs_path *path, u64 bytenr,
 883                            int *info_level, struct preftrees *preftrees,
 884                            u64 *total_refs, struct share_check *sc)
 885 {
 886         int ret = 0;
 887         int slot;
 888         struct extent_buffer *leaf;
 889         struct btrfs_key key;
 890         struct btrfs_key found_key;
 891         unsigned long ptr;
 892         unsigned long end;
 893         struct btrfs_extent_item *ei;
 894         u64 flags;
 895         u64 item_size;
 896
 897         /*
 898          * enumerate all inline refs
 899          */
 900         leaf = path->nodes[0];
 901         slot = path->slots[0];
 902
 903         item_size = btrfs_item_size_nr(leaf, slot);
 904         BUG_ON(item_size < sizeof(*ei));
 905
 906         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 907         flags = btrfs_extent_flags(leaf, ei);
 908         *total_refs += btrfs_extent_refs(leaf, ei);
 909         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 910
 911         ptr = (unsigned long)(ei + 1);
 912         end = (unsigned long)ei + item_size;
 913
 914         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 915             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 916                 struct btrfs_tree_block_info *info;
 917
 918                 info = (struct btrfs_tree_block_info *)ptr;
 919                 *info_level = btrfs_tree_block_level(leaf, info);
 920                 ptr += sizeof(struct btrfs_tree_block_info);
 921                 BUG_ON(ptr > end);
 922         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 923                 *info_level = found_key.offset;
 924         } else {
 925                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 926         }
 927
 928         while (ptr < end) {
 929                 struct btrfs_extent_inline_ref *iref;
 930                 u64 offset;
 931                 int type;
 932
 933                 iref = (struct btrfs_extent_inline_ref *)ptr;
 934                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
 935                                                         BTRFS_REF_TYPE_ANY);
 936                 if (type == BTRFS_REF_TYPE_INVALID)
 937                         return -EINVAL;
 938
 939                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 940
 941                 switch (type) {
 942                 case BTRFS_SHARED_BLOCK_REF_KEY:
 943                         ret = add_direct_ref(fs_info, preftrees,
 944                                              *info_level + 1, offset,
 945                                              bytenr, 1, NULL, GFP_NOFS);
 946                         break;
 947                 case BTRFS_SHARED_DATA_REF_KEY: {
 948                         struct btrfs_shared_data_ref *sdref;
 949                         int count;
 950
 951                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 952                         count = btrfs_shared_data_ref_count(leaf, sdref);
 953
 954                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
 955                                              bytenr, count, sc, GFP_NOFS);
 956                         break;
 957                 }
 958                 case BTRFS_TREE_BLOCK_REF_KEY:
 959                         ret = add_indirect_ref(fs_info, preftrees, offset,
 960                                                NULL, *info_level + 1,
 961                                                bytenr, 1, NULL, GFP_NOFS);
 962                         break;
 963                 case BTRFS_EXTENT_DATA_REF_KEY: {
 964                         struct btrfs_extent_data_ref *dref;
 965                         int count;
 966                         u64 root;
 967
 968                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 969                         count = btrfs_extent_data_ref_count(leaf, dref);
 970                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 971                                                                       dref);
 972                         key.type = BTRFS_EXTENT_DATA_KEY;
 973                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 974
 975                         if (sc && sc->inum && key.objectid != sc->inum) {
 976                                 ret = BACKREF_FOUND_SHARED;
 977                                 break;
 978                         }
 979
 980                         root = btrfs_extent_data_ref_root(leaf, dref);
 981
 982                         ret = add_indirect_ref(fs_info, preftrees, root,
 983                                                &key, 0, bytenr, count,
 984                                                sc, GFP_NOFS);
 985                         break;
 986                 }
 987                 default:
 988                         WARN_ON(1);
 989                 }
 990                 if (ret)
 991                         return ret;
 992                 ptr += btrfs_extent_inline_ref_size(type);
 993         }
 994
 995         return 0;
 996 }
 997
 998 /*
 999  * add all non-inline backrefs for bytenr to the list
1000  *
1001  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1002  */
1003 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
1004                           struct btrfs_path *path, u64 bytenr,
1005                           int info_level, struct preftrees *preftrees,
1006                           struct share_check *sc)
1007 {
1008         struct btrfs_root *extent_root = fs_info->extent_root;
1009         int ret;
1010         int slot;
1011         struct extent_buffer *leaf;
1012         struct btrfs_key key;
1013
1014         while (1) {
1015                 ret = btrfs_next_item(extent_root, path);
1016                 if (ret < 0)
1017                         break;
1018                 if (ret) {
1019                         ret = 0;
1020                         break;
1021                 }
1022
1023                 slot = path->slots[0];
1024                 leaf = path->nodes[0];
1025                 btrfs_item_key_to_cpu(leaf, &key, slot);
1026
1027                 if (key.objectid != bytenr)
1028                         break;
1029                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1030                         continue;
1031                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1032                         break;
1033
1034                 switch (key.type) {
1035                 case BTRFS_SHARED_BLOCK_REF_KEY:
1036                         /* SHARED DIRECT METADATA backref */
1037                         ret = add_direct_ref(fs_info, preftrees,
1038                                              info_level + 1, key.offset,
1039                                              bytenr, 1, NULL, GFP_NOFS);
1040                         break;
1041                 case BTRFS_SHARED_DATA_REF_KEY: {
1042                         /* SHARED DIRECT FULL backref */
1043                         struct btrfs_shared_data_ref *sdref;
1044                         int count;
1045
1046                         sdref = btrfs_item_ptr(leaf, slot,
1047                                               struct btrfs_shared_data_ref);
1048                         count = btrfs_shared_data_ref_count(leaf, sdref);
1049                         ret = add_direct_ref(fs_info, preftrees, 0,
1050                                              key.offset, bytenr, count,
1051                                              sc, GFP_NOFS);
1052                         break;
1053                 }
1054                 case BTRFS_TREE_BLOCK_REF_KEY:
1055                         /* NORMAL INDIRECT METADATA backref */
1056                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1057                                                NULL, info_level + 1, bytenr,
1058                                                1, NULL, GFP_NOFS);
1059                         break;
1060                 case BTRFS_EXTENT_DATA_REF_KEY: {
1061                         /* NORMAL INDIRECT DATA backref */
1062                         struct btrfs_extent_data_ref *dref;
1063                         int count;
1064                         u64 root;
1065
1066                         dref = btrfs_item_ptr(leaf, slot,
1067                                               struct btrfs_extent_data_ref);
1068                         count = btrfs_extent_data_ref_count(leaf, dref);
1069                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1070                                                                       dref);
1071                         key.type = BTRFS_EXTENT_DATA_KEY;
1072                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1073
1074                         if (sc && sc->inum && key.objectid != sc->inum) {
1075                                 ret = BACKREF_FOUND_SHARED;
1076                                 break;
1077                         }
1078
1079                         root = btrfs_extent_data_ref_root(leaf, dref);
1080                         ret = add_indirect_ref(fs_info, preftrees, root,
1081                                                &key, 0, bytenr, count,
1082                                                sc, GFP_NOFS);
1083                         break;
1084                 }
1085                 default:
1086                         WARN_ON(1);
1087                 }
1088                 if (ret)
1089                         return ret;
1090
1091         }
1092
1093         return ret;
1094 }
1095
1096 /*
1097  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1098  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1099  * indirect refs to their parent bytenr.
1100  * When roots are found, they're added to the roots list
1101  *
1102  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1103  * much like trans == NULL case, the difference only lies in it will not
1104  * commit root.
1105  * The special case is for qgroup to search roots in commit_transaction().
1106  *
1107  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1108  * shared extent is detected.
1109  *
1110  * Otherwise this returns 0 for success and <0 for an error.
1111  *
1112  * FIXME some caching might speed things up
1113  */
1114 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1115                              struct btrfs_fs_info *fs_info, u64 bytenr,
1116                              u64 time_seq, struct ulist *refs,
1117                              struct ulist *roots, const u64 *extent_item_pos,
1118                              struct share_check *sc)
1119 {
1120         struct btrfs_key key;
1121         struct btrfs_path *path;
1122         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1123         struct btrfs_delayed_ref_head *head;
1124         int info_level = 0;
1125         int ret;
1126         struct prelim_ref *ref;
1127         struct rb_node *node;
1128         struct extent_inode_elem *eie = NULL;
1129         /* total of both direct AND indirect refs! */
1130         u64 total_refs = 0;
1131         struct preftrees preftrees = {
1132                 .direct = PREFTREE_INIT,
1133                 .indirect = PREFTREE_INIT,
1134                 .indirect_missing_keys = PREFTREE_INIT
1135         };
1136
1137         key.objectid = bytenr;
1138         key.offset = (u64)-1;
1139         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1140                 key.type = BTRFS_METADATA_ITEM_KEY;
1141         else
1142                 key.type = BTRFS_EXTENT_ITEM_KEY;
1143
1144         path = btrfs_alloc_path();
1145         if (!path)
1146                 return -ENOMEM;
1147         if (!trans) {
1148                 path->search_commit_root = 1;
1149                 path->skip_locking = 1;
1150         }
1151
1152         if (time_seq == SEQ_LAST)
1153                 path->skip_locking = 1;
1154
1155         /*
1156          * grab both a lock on the path and a lock on the delayed ref head.
1157          * We need both to get a consistent picture of how the refs look
1158          * at a specified point in time
1159          */
1160 again:
1161         head = NULL;
1162
1163         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1164         if (ret < 0)
1165                 goto out;
1166         if (ret == 0) {
1167                 /* This shouldn't happen, indicates a bug or fs corruption. */
1168                 ASSERT(ret != 0);
1169                 ret = -EUCLEAN;
1170                 goto out;
1171         }
1172
1173 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1174         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1175             time_seq != SEQ_LAST) {
1176 #else
1177         if (trans && time_seq != SEQ_LAST) {
1178 #endif
1179                 /*
1180                  * look if there are updates for this ref queued and lock the
1181                  * head
1182                  */
1183                 delayed_refs = &trans->transaction->delayed_refs;
1184                 spin_lock(&delayed_refs->lock);
1185                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1186                 if (head) {
1187                         if (!mutex_trylock(&head->mutex)) {
1188                                 refcount_inc(&head->node.refs);
1189                                 spin_unlock(&delayed_refs->lock);
1190
1191                                 btrfs_release_path(path);
1192
1193                                 /*
1194                                  * Mutex was contended, block until it's
1195                                  * released and try again
1196                                  */
1197                                 mutex_lock(&head->mutex);
1198                                 mutex_unlock(&head->mutex);
1199                                 btrfs_put_delayed_ref(&head->node);
1200                                 goto again;
1201                         }
1202                         spin_unlock(&delayed_refs->lock);
1203                         ret = add_delayed_refs(fs_info, head, time_seq,
1204                                                &preftrees, &total_refs, sc);
1205                         mutex_unlock(&head->mutex);
1206                         if (ret)
1207                                 goto out;
1208                 } else {
1209                         spin_unlock(&delayed_refs->lock);
1210                 }
1211         }
1212
1213         if (path->slots[0]) {
1214                 struct extent_buffer *leaf;
1215                 int slot;
1216
1217                 path->slots[0]--;
1218                 leaf = path->nodes[0];
1219                 slot = path->slots[0];
1220                 btrfs_item_key_to_cpu(leaf, &key, slot);
1221                 if (key.objectid == bytenr &&
1222                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1223                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1224                         ret = add_inline_refs(fs_info, path, bytenr,
1225                                               &info_level, &preftrees,
1226                                               &total_refs, sc);
1227                         if (ret)
1228                                 goto out;
1229                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1230                                              &preftrees, sc);
1231                         if (ret)
1232                                 goto out;
1233                 }
1234         }
1235
1236         btrfs_release_path(path);
1237
1238         ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1239         if (ret)
1240                 goto out;
1241
1242         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root));
1243
1244         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1245                                     extent_item_pos, total_refs, sc);
1246         if (ret)
1247                 goto out;
1248
1249         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root));
1250
1251         /*
1252          * This walks the tree of merged and resolved refs. Tree blocks are
1253          * read in as needed. Unique entries are added to the ulist, and
1254          * the list of found roots is updated.
1255          *
1256          * We release the entire tree in one go before returning.
1257          */
1258         node = rb_first(&preftrees.direct.root);
1259         while (node) {
1260                 ref = rb_entry(node, struct prelim_ref, rbnode);
1261                 node = rb_next(&ref->rbnode);
1262                 /*
1263                  * ref->count < 0 can happen here if there are delayed
1264                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1265                  * prelim_ref_insert() relies on this when merging
1266                  * identical refs to keep the overall count correct.
1267                  * prelim_ref_insert() will merge only those refs
1268                  * which compare identically.  Any refs having
1269                  * e.g. different offsets would not be merged,
1270                  * and would retain their original ref->count < 0.
1271                  */
1272                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1273                         if (sc && sc->root_objectid &&
1274                             ref->root_id != sc->root_objectid) {
1275                                 ret = BACKREF_FOUND_SHARED;
1276                                 goto out;
1277                         }
1278
1279                         /* no parent == root of tree */
1280                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1281                         if (ret < 0)
1282                                 goto out;
1283                 }
1284                 if (ref->count && ref->parent) {
1285                         if (extent_item_pos && !ref->inode_list &&
1286                             ref->level == 0) {
1287                                 struct extent_buffer *eb;
1288
1289                                 eb = read_tree_block(fs_info, ref->parent, 0);
1290                                 if (IS_ERR(eb)) {
1291                                         ret = PTR_ERR(eb);
1292                                         goto out;
1293                                 } else if (!extent_buffer_uptodate(eb)) {
1294                                         free_extent_buffer(eb);
1295                                         ret = -EIO;
1296                                         goto out;
1297                                 }
1298                                 if (!path->skip_locking) {
1299                                         btrfs_tree_read_lock(eb);
1300                                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1301                                 }
1302                                 ret = find_extent_in_eb(eb, bytenr,
1303                                                         *extent_item_pos, &eie);
1304                                 if (!path->skip_locking)
1305                                         btrfs_tree_read_unlock_blocking(eb);
1306                                 free_extent_buffer(eb);
1307                                 if (ret < 0)
1308                                         goto out;
1309                                 ref->inode_list = eie;
1310                         }
1311                         ret = ulist_add_merge_ptr(refs, ref->parent,
1312                                                   ref->inode_list,
1313                                                   (void **)&eie, GFP_NOFS);
1314                         if (ret < 0)
1315                                 goto out;
1316                         if (!ret && extent_item_pos) {
1317                                 /*
1318                                  * We've recorded that parent, so we must extend
1319                                  * its inode list here.
1320                                  *
1321                                  * However if there was corruption we may not
1322                                  * have found an eie, return an error in this
1323                                  * case.
1324                                  */
1325                                 ASSERT(eie);
1326                                 if (!eie) {
1327                                         ret = -EUCLEAN;
1328                                         goto out;
1329                                 }
1330                                 while (eie->next)
1331                                         eie = eie->next;
1332                                 eie->next = ref->inode_list;
1333                         }
1334                         eie = NULL;
1335                 }
1336                 cond_resched();
1337         }
1338
1339 out:
1340         btrfs_free_path(path);
1341
1342         prelim_release(&preftrees.direct);
1343         prelim_release(&preftrees.indirect);
1344         prelim_release(&preftrees.indirect_missing_keys);
1345
1346         if (ret < 0)
1347                 free_inode_elem_list(eie);
1348         return ret;
1349 }
1350
1351 static void free_leaf_list(struct ulist *blocks)
1352 {
1353         struct ulist_node *node = NULL;
1354         struct extent_inode_elem *eie;
1355         struct ulist_iterator uiter;
1356
1357         ULIST_ITER_INIT(&uiter);
1358         while ((node = ulist_next(blocks, &uiter))) {
1359                 if (!node->aux)
1360                         continue;
1361                 eie = unode_aux_to_inode_list(node);
1362                 free_inode_elem_list(eie);
1363                 node->aux = 0;
1364         }
1365
1366         ulist_free(blocks);
1367 }
1368
1369 /*
1370  * Finds all leafs with a reference to the specified combination of bytenr and
1371  * offset. key_list_head will point to a list of corresponding keys (caller must
1372  * free each list element). The leafs will be stored in the leafs ulist, which
1373  * must be freed with ulist_free.
1374  *
1375  * returns 0 on success, <0 on error
1376  */
1377 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1378                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1379                                 u64 time_seq, struct ulist **leafs,
1380                                 const u64 *extent_item_pos)
1381 {
1382         int ret;
1383
1384         *leafs = ulist_alloc(GFP_NOFS);
1385         if (!*leafs)
1386                 return -ENOMEM;
1387
1388         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1389                                 *leafs, NULL, extent_item_pos, NULL);
1390         if (ret < 0 && ret != -ENOENT) {
1391                 free_leaf_list(*leafs);
1392                 return ret;
1393         }
1394
1395         return 0;
1396 }
1397
1398 /*
1399  * walk all backrefs for a given extent to find all roots that reference this
1400  * extent. Walking a backref means finding all extents that reference this
1401  * extent and in turn walk the backrefs of those, too. Naturally this is a
1402  * recursive process, but here it is implemented in an iterative fashion: We
1403  * find all referencing extents for the extent in question and put them on a
1404  * list. In turn, we find all referencing extents for those, further appending
1405  * to the list. The way we iterate the list allows adding more elements after
1406  * the current while iterating. The process stops when we reach the end of the
1407  * list. Found roots are added to the roots list.
1408  *
1409  * returns 0 on success, < 0 on error.
1410  */
1411 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1412                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1413                                      u64 time_seq, struct ulist **roots)
1414 {
1415         struct ulist *tmp;
1416         struct ulist_node *node = NULL;
1417         struct ulist_iterator uiter;
1418         int ret;
1419
1420         tmp = ulist_alloc(GFP_NOFS);
1421         if (!tmp)
1422                 return -ENOMEM;
1423         *roots = ulist_alloc(GFP_NOFS);
1424         if (!*roots) {
1425                 ulist_free(tmp);
1426                 return -ENOMEM;
1427         }
1428
1429         ULIST_ITER_INIT(&uiter);
1430         while (1) {
1431                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1432                                         tmp, *roots, NULL, NULL);
1433                 if (ret < 0 && ret != -ENOENT) {
1434                         ulist_free(tmp);
1435                         ulist_free(*roots);
1436                         *roots = NULL;
1437                         return ret;
1438                 }
1439                 node = ulist_next(tmp, &uiter);
1440                 if (!node)
1441                         break;
1442                 bytenr = node->val;
1443                 cond_resched();
1444         }
1445
1446         ulist_free(tmp);
1447         return 0;
1448 }
1449
1450 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1451                          struct btrfs_fs_info *fs_info, u64 bytenr,
1452                          u64 time_seq, struct ulist **roots)
1453 {
1454         int ret;
1455
1456         if (!trans)
1457                 down_read(&fs_info->commit_root_sem);
1458         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1459                                         time_seq, roots);
1460         if (!trans)
1461                 up_read(&fs_info->commit_root_sem);
1462         return ret;
1463 }
1464
1465 /**
1466  * btrfs_check_shared - tell us whether an extent is shared
1467  *
1468  * btrfs_check_shared uses the backref walking code but will short
1469  * circuit as soon as it finds a root or inode that doesn't match the
1470  * one passed in. This provides a significant performance benefit for
1471  * callers (such as fiemap) which want to know whether the extent is
1472  * shared but do not need a ref count.
1473  *
1474  * This attempts to attach to the running transaction in order to account for
1475  * delayed refs, but continues on even when no running transaction exists.
1476  *
1477  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1478  */
1479 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1480 {
1481         struct btrfs_fs_info *fs_info = root->fs_info;
1482         struct btrfs_trans_handle *trans;
1483         struct ulist *tmp = NULL;
1484         struct ulist *roots = NULL;
1485         struct ulist_iterator uiter;
1486         struct ulist_node *node;
1487         struct seq_list elem = SEQ_LIST_INIT(elem);
1488         int ret = 0;
1489         struct share_check shared = {
1490                 .root_objectid = root->objectid,
1491                 .inum = inum,
1492                 .share_count = 0,
1493         };
1494
1495         tmp = ulist_alloc(GFP_NOFS);
1496         roots = ulist_alloc(GFP_NOFS);
1497         if (!tmp || !roots) {
1498                 ret = -ENOMEM;
1499                 goto out;
1500         }
1501
1502         trans = btrfs_attach_transaction(root);
1503         if (IS_ERR(trans)) {
1504                 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1505                         ret = PTR_ERR(trans);
1506                         goto out;
1507                 }
1508                 trans = NULL;
1509                 down_read(&fs_info->commit_root_sem);
1510         } else {
1511                 btrfs_get_tree_mod_seq(fs_info, &elem);
1512         }
1513
1514         ULIST_ITER_INIT(&uiter);
1515         while (1) {
1516                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1517                                         roots, NULL, &shared);
1518                 if (ret == BACKREF_FOUND_SHARED) {
1519                         /* this is the only condition under which we return 1 */
1520                         ret = 1;
1521                         break;
1522                 }
1523                 if (ret < 0 && ret != -ENOENT)
1524                         break;
1525                 ret = 0;
1526                 node = ulist_next(tmp, &uiter);
1527                 if (!node)
1528                         break;
1529                 bytenr = node->val;
1530                 shared.share_count = 0;
1531                 cond_resched();
1532         }
1533
1534         if (trans) {
1535                 btrfs_put_tree_mod_seq(fs_info, &elem);
1536                 btrfs_end_transaction(trans);
1537         } else {
1538                 up_read(&fs_info->commit_root_sem);
1539         }
1540 out:
1541         ulist_free(tmp);
1542         ulist_free(roots);
1543         return ret;
1544 }
1545
1546 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1547                           u64 start_off, struct btrfs_path *path,
1548                           struct btrfs_inode_extref **ret_extref,
1549                           u64 *found_off)
1550 {
1551         int ret, slot;
1552         struct btrfs_key key;
1553         struct btrfs_key found_key;
1554         struct btrfs_inode_extref *extref;
1555         const struct extent_buffer *leaf;
1556         unsigned long ptr;
1557
1558         key.objectid = inode_objectid;
1559         key.type = BTRFS_INODE_EXTREF_KEY;
1560         key.offset = start_off;
1561
1562         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1563         if (ret < 0)
1564                 return ret;
1565
1566         while (1) {
1567                 leaf = path->nodes[0];
1568                 slot = path->slots[0];
1569                 if (slot >= btrfs_header_nritems(leaf)) {
1570                         /*
1571                          * If the item at offset is not found,
1572                          * btrfs_search_slot will point us to the slot
1573                          * where it should be inserted. In our case
1574                          * that will be the slot directly before the
1575                          * next INODE_REF_KEY_V2 item. In the case
1576                          * that we're pointing to the last slot in a
1577                          * leaf, we must move one leaf over.
1578                          */
1579                         ret = btrfs_next_leaf(root, path);
1580                         if (ret) {
1581                                 if (ret >= 1)
1582                                         ret = -ENOENT;
1583                                 break;
1584                         }
1585                         continue;
1586                 }
1587
1588                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1589
1590                 /*
1591                  * Check that we're still looking at an extended ref key for
1592                  * this particular objectid. If we have different
1593                  * objectid or type then there are no more to be found
1594                  * in the tree and we can exit.
1595                  */
1596                 ret = -ENOENT;
1597                 if (found_key.objectid != inode_objectid)
1598                         break;
1599                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1600                         break;
1601
1602                 ret = 0;
1603                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1604                 extref = (struct btrfs_inode_extref *)ptr;
1605                 *ret_extref = extref;
1606                 if (found_off)
1607                         *found_off = found_key.offset;
1608                 break;
1609         }
1610
1611         return ret;
1612 }
1613
1614 /*
1615  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1616  * Elements of the path are separated by '/' and the path is guaranteed to be
1617  * 0-terminated. the path is only given within the current file system.
1618  * Therefore, it never starts with a '/'. the caller is responsible to provide
1619  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1620  * the start point of the resulting string is returned. this pointer is within
1621  * dest, normally.
1622  * in case the path buffer would overflow, the pointer is decremented further
1623  * as if output was written to the buffer, though no more output is actually
1624  * generated. that way, the caller can determine how much space would be
1625  * required for the path to fit into the buffer. in that case, the returned
1626  * value will be smaller than dest. callers must check this!
1627  */
1628 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1629                         u32 name_len, unsigned long name_off,
1630                         struct extent_buffer *eb_in, u64 parent,
1631                         char *dest, u32 size)
1632 {
1633         int slot;
1634         u64 next_inum;
1635         int ret;
1636         s64 bytes_left = ((s64)size) - 1;
1637         struct extent_buffer *eb = eb_in;
1638         struct btrfs_key found_key;
1639         int leave_spinning = path->leave_spinning;
1640         struct btrfs_inode_ref *iref;
1641
1642         if (bytes_left >= 0)
1643                 dest[bytes_left] = '\0';
1644
1645         path->leave_spinning = 1;
1646         while (1) {
1647                 bytes_left -= name_len;
1648                 if (bytes_left >= 0)
1649                         read_extent_buffer(eb, dest + bytes_left,
1650                                            name_off, name_len);
1651                 if (eb != eb_in) {
1652                         if (!path->skip_locking)
1653                                 btrfs_tree_read_unlock_blocking(eb);
1654                         free_extent_buffer(eb);
1655                 }
1656                 ret = btrfs_find_item(fs_root, path, parent, 0,
1657                                 BTRFS_INODE_REF_KEY, &found_key);
1658                 if (ret > 0)
1659                         ret = -ENOENT;
1660                 if (ret)
1661                         break;
1662
1663                 next_inum = found_key.offset;
1664
1665                 /* regular exit ahead */
1666                 if (parent == next_inum)
1667                         break;
1668
1669                 slot = path->slots[0];
1670                 eb = path->nodes[0];
1671                 /* make sure we can use eb after releasing the path */
1672                 if (eb != eb_in) {
1673                         if (!path->skip_locking)
1674                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1675                         path->nodes[0] = NULL;
1676                         path->locks[0] = 0;
1677                 }
1678                 btrfs_release_path(path);
1679                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1680
1681                 name_len = btrfs_inode_ref_name_len(eb, iref);
1682                 name_off = (unsigned long)(iref + 1);
1683
1684                 parent = next_inum;
1685                 --bytes_left;
1686                 if (bytes_left >= 0)
1687                         dest[bytes_left] = '/';
1688         }
1689
1690         btrfs_release_path(path);
1691         path->leave_spinning = leave_spinning;
1692
1693         if (ret)
1694                 return ERR_PTR(ret);
1695
1696         return dest + bytes_left;
1697 }
1698
1699 /*
1700  * this makes the path point to (logical EXTENT_ITEM *)
1701  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1702  * tree blocks and <0 on error.
1703  */
1704 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1705                         struct btrfs_path *path, struct btrfs_key *found_key,
1706                         u64 *flags_ret)
1707 {
1708         int ret;
1709         u64 flags;
1710         u64 size = 0;
1711         u32 item_size;
1712         const struct extent_buffer *eb;
1713         struct btrfs_extent_item *ei;
1714         struct btrfs_key key;
1715
1716         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1717                 key.type = BTRFS_METADATA_ITEM_KEY;
1718         else
1719                 key.type = BTRFS_EXTENT_ITEM_KEY;
1720         key.objectid = logical;
1721         key.offset = (u64)-1;
1722
1723         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1724         if (ret < 0)
1725                 return ret;
1726
1727         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1728         if (ret) {
1729                 if (ret > 0)
1730                         ret = -ENOENT;
1731                 return ret;
1732         }
1733         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1734         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1735                 size = fs_info->nodesize;
1736         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1737                 size = found_key->offset;
1738
1739         if (found_key->objectid > logical ||
1740             found_key->objectid + size <= logical) {
1741                 btrfs_debug(fs_info,
1742                         "logical %llu is not within any extent", logical);
1743                 return -ENOENT;
1744         }
1745
1746         eb = path->nodes[0];
1747         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1748         BUG_ON(item_size < sizeof(*ei));
1749
1750         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1751         flags = btrfs_extent_flags(eb, ei);
1752
1753         btrfs_debug(fs_info,
1754                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1755                  logical, logical - found_key->objectid, found_key->objectid,
1756                  found_key->offset, flags, item_size);
1757
1758         WARN_ON(!flags_ret);
1759         if (flags_ret) {
1760                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1761                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1762                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1763                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1764                 else
1765                         BUG_ON(1);
1766                 return 0;
1767         }
1768
1769         return -EIO;
1770 }
1771
1772 /*
1773  * helper function to iterate extent inline refs. ptr must point to a 0 value
1774  * for the first call and may be modified. it is used to track state.
1775  * if more refs exist, 0 is returned and the next call to
1776  * get_extent_inline_ref must pass the modified ptr parameter to get the
1777  * next ref. after the last ref was processed, 1 is returned.
1778  * returns <0 on error
1779  */
1780 static int get_extent_inline_ref(unsigned long *ptr,
1781                                  const struct extent_buffer *eb,
1782                                  const struct btrfs_key *key,
1783                                  const struct btrfs_extent_item *ei,
1784                                  u32 item_size,
1785                                  struct btrfs_extent_inline_ref **out_eiref,
1786                                  int *out_type)
1787 {
1788         unsigned long end;
1789         u64 flags;
1790         struct btrfs_tree_block_info *info;
1791
1792         if (!*ptr) {
1793                 /* first call */
1794                 flags = btrfs_extent_flags(eb, ei);
1795                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1796                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1797                                 /* a skinny metadata extent */
1798                                 *out_eiref =
1799                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1800                         } else {
1801                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1802                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1803                                 *out_eiref =
1804                                    (struct btrfs_extent_inline_ref *)(info + 1);
1805                         }
1806                 } else {
1807                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1808                 }
1809                 *ptr = (unsigned long)*out_eiref;
1810                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1811                         return -ENOENT;
1812         }
1813
1814         end = (unsigned long)ei + item_size;
1815         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1816         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1817                                                      BTRFS_REF_TYPE_ANY);
1818         if (*out_type == BTRFS_REF_TYPE_INVALID)
1819                 return -EINVAL;
1820
1821         *ptr += btrfs_extent_inline_ref_size(*out_type);
1822         WARN_ON(*ptr > end);
1823         if (*ptr == end)
1824                 return 1; /* last */
1825
1826         return 0;
1827 }
1828
1829 /*
1830  * reads the tree block backref for an extent. tree level and root are returned
1831  * through out_level and out_root. ptr must point to a 0 value for the first
1832  * call and may be modified (see get_extent_inline_ref comment).
1833  * returns 0 if data was provided, 1 if there was no more data to provide or
1834  * <0 on error.
1835  */
1836 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1837                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1838                             u32 item_size, u64 *out_root, u8 *out_level)
1839 {
1840         int ret;
1841         int type;
1842         struct btrfs_extent_inline_ref *eiref;
1843
1844         if (*ptr == (unsigned long)-1)
1845                 return 1;
1846
1847         while (1) {
1848                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1849                                               &eiref, &type);
1850                 if (ret < 0)
1851                         return ret;
1852
1853                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1854                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1855                         break;
1856
1857                 if (ret == 1)
1858                         return 1;
1859         }
1860
1861         /* we can treat both ref types equally here */
1862         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1863
1864         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1865                 struct btrfs_tree_block_info *info;
1866
1867                 info = (struct btrfs_tree_block_info *)(ei + 1);
1868                 *out_level = btrfs_tree_block_level(eb, info);
1869         } else {
1870                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1871                 *out_level = (u8)key->offset;
1872         }
1873
1874         if (ret == 1)
1875                 *ptr = (unsigned long)-1;
1876
1877         return 0;
1878 }
1879
1880 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1881                              struct extent_inode_elem *inode_list,
1882                              u64 root, u64 extent_item_objectid,
1883                              iterate_extent_inodes_t *iterate, void *ctx)
1884 {
1885         struct extent_inode_elem *eie;
1886         int ret = 0;
1887
1888         for (eie = inode_list; eie; eie = eie->next) {
1889                 btrfs_debug(fs_info,
1890                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1891                             extent_item_objectid, eie->inum,
1892                             eie->offset, root);
1893                 ret = iterate(eie->inum, eie->offset, root, ctx);
1894                 if (ret) {
1895                         btrfs_debug(fs_info,
1896                                     "stopping iteration for %llu due to ret=%d",
1897                                     extent_item_objectid, ret);
1898                         break;
1899                 }
1900         }
1901
1902         return ret;
1903 }
1904
1905 /*
1906  * calls iterate() for every inode that references the extent identified by
1907  * the given parameters.
1908  * when the iterator function returns a non-zero value, iteration stops.
1909  */
1910 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1911                                 u64 extent_item_objectid, u64 extent_item_pos,
1912                                 int search_commit_root,
1913                                 iterate_extent_inodes_t *iterate, void *ctx)
1914 {
1915         int ret;
1916         struct btrfs_trans_handle *trans = NULL;
1917         struct ulist *refs = NULL;
1918         struct ulist *roots = NULL;
1919         struct ulist_node *ref_node = NULL;
1920         struct ulist_node *root_node = NULL;
1921         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1922         struct ulist_iterator ref_uiter;
1923         struct ulist_iterator root_uiter;
1924
1925         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1926                         extent_item_objectid);
1927
1928         if (!search_commit_root) {
1929                 trans = btrfs_attach_transaction(fs_info->extent_root);
1930                 if (IS_ERR(trans)) {
1931                         if (PTR_ERR(trans) != -ENOENT &&
1932                             PTR_ERR(trans) != -EROFS)
1933                                 return PTR_ERR(trans);
1934                         trans = NULL;
1935                 }
1936         }
1937
1938         if (trans)
1939                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1940         else
1941                 down_read(&fs_info->commit_root_sem);
1942
1943         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1944                                    tree_mod_seq_elem.seq, &refs,
1945                                    &extent_item_pos);
1946         if (ret)
1947                 goto out;
1948
1949         ULIST_ITER_INIT(&ref_uiter);
1950         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1951                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1952                                                 tree_mod_seq_elem.seq, &roots);
1953                 if (ret)
1954                         break;
1955                 ULIST_ITER_INIT(&root_uiter);
1956                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1957                         btrfs_debug(fs_info,
1958                                     "root %llu references leaf %llu, data list %#llx",
1959                                     root_node->val, ref_node->val,
1960                                     ref_node->aux);
1961                         ret = iterate_leaf_refs(fs_info,
1962                                                 (struct extent_inode_elem *)
1963                                                 (uintptr_t)ref_node->aux,
1964                                                 root_node->val,
1965                                                 extent_item_objectid,
1966                                                 iterate, ctx);
1967                 }
1968                 ulist_free(roots);
1969         }
1970
1971         free_leaf_list(refs);
1972 out:
1973         if (trans) {
1974                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1975                 btrfs_end_transaction(trans);
1976         } else {
1977                 up_read(&fs_info->commit_root_sem);
1978         }
1979
1980         return ret;
1981 }
1982
1983 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1984                                 struct btrfs_path *path,
1985                                 iterate_extent_inodes_t *iterate, void *ctx)
1986 {
1987         int ret;
1988         u64 extent_item_pos;
1989         u64 flags = 0;
1990         struct btrfs_key found_key;
1991         int search_commit_root = path->search_commit_root;
1992
1993         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1994         btrfs_release_path(path);
1995         if (ret < 0)
1996                 return ret;
1997         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1998                 return -EINVAL;
1999
2000         extent_item_pos = logical - found_key.objectid;
2001         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2002                                         extent_item_pos, search_commit_root,
2003                                         iterate, ctx);
2004
2005         return ret;
2006 }
2007
2008 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2009                               struct extent_buffer *eb, void *ctx);
2010
2011 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2012                               struct btrfs_path *path,
2013                               iterate_irefs_t *iterate, void *ctx)
2014 {
2015         int ret = 0;
2016         int slot;
2017         u32 cur;
2018         u32 len;
2019         u32 name_len;
2020         u64 parent = 0;
2021         int found = 0;
2022         struct extent_buffer *eb;
2023         struct btrfs_item *item;
2024         struct btrfs_inode_ref *iref;
2025         struct btrfs_key found_key;
2026
2027         while (!ret) {
2028                 ret = btrfs_find_item(fs_root, path, inum,
2029                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2030                                 &found_key);
2031
2032                 if (ret < 0)
2033                         break;
2034                 if (ret) {
2035                         ret = found ? 0 : -ENOENT;
2036                         break;
2037                 }
2038                 ++found;
2039
2040                 parent = found_key.offset;
2041                 slot = path->slots[0];
2042                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2043                 if (!eb) {
2044                         ret = -ENOMEM;
2045                         break;
2046                 }
2047                 extent_buffer_get(eb);
2048                 btrfs_tree_read_lock(eb);
2049                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2050                 btrfs_release_path(path);
2051
2052                 item = btrfs_item_nr(slot);
2053                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2054
2055                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2056                         name_len = btrfs_inode_ref_name_len(eb, iref);
2057                         /* path must be released before calling iterate()! */
2058                         btrfs_debug(fs_root->fs_info,
2059                                 "following ref at offset %u for inode %llu in tree %llu",
2060                                 cur, found_key.objectid, fs_root->objectid);
2061                         ret = iterate(parent, name_len,
2062                                       (unsigned long)(iref + 1), eb, ctx);
2063                         if (ret)
2064                                 break;
2065                         len = sizeof(*iref) + name_len;
2066                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2067                 }
2068                 btrfs_tree_read_unlock_blocking(eb);
2069                 free_extent_buffer(eb);
2070         }
2071
2072         btrfs_release_path(path);
2073
2074         return ret;
2075 }
2076
2077 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2078                                  struct btrfs_path *path,
2079                                  iterate_irefs_t *iterate, void *ctx)
2080 {
2081         int ret;
2082         int slot;
2083         u64 offset = 0;
2084         u64 parent;
2085         int found = 0;
2086         struct extent_buffer *eb;
2087         struct btrfs_inode_extref *extref;
2088         u32 item_size;
2089         u32 cur_offset;
2090         unsigned long ptr;
2091
2092         while (1) {
2093                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2094                                             &offset);
2095                 if (ret < 0)
2096                         break;
2097                 if (ret) {
2098                         ret = found ? 0 : -ENOENT;
2099                         break;
2100                 }
2101                 ++found;
2102
2103                 slot = path->slots[0];
2104                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2105                 if (!eb) {
2106                         ret = -ENOMEM;
2107                         break;
2108                 }
2109                 extent_buffer_get(eb);
2110
2111                 btrfs_tree_read_lock(eb);
2112                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2113                 btrfs_release_path(path);
2114
2115                 item_size = btrfs_item_size_nr(eb, slot);
2116                 ptr = btrfs_item_ptr_offset(eb, slot);
2117                 cur_offset = 0;
2118
2119                 while (cur_offset < item_size) {
2120                         u32 name_len;
2121
2122                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2123                         parent = btrfs_inode_extref_parent(eb, extref);
2124                         name_len = btrfs_inode_extref_name_len(eb, extref);
2125                         ret = iterate(parent, name_len,
2126                                       (unsigned long)&extref->name, eb, ctx);
2127                         if (ret)
2128                                 break;
2129
2130                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2131                         cur_offset += sizeof(*extref);
2132                 }
2133                 btrfs_tree_read_unlock_blocking(eb);
2134                 free_extent_buffer(eb);
2135
2136                 offset++;
2137         }
2138
2139         btrfs_release_path(path);
2140
2141         return ret;
2142 }
2143
2144 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2145                          struct btrfs_path *path, iterate_irefs_t *iterate,
2146                          void *ctx)
2147 {
2148         int ret;
2149         int found_refs = 0;
2150
2151         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2152         if (!ret)
2153                 ++found_refs;
2154         else if (ret != -ENOENT)
2155                 return ret;
2156
2157         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2158         if (ret == -ENOENT && found_refs)
2159                 return 0;
2160
2161         return ret;
2162 }
2163
2164 /*
2165  * returns 0 if the path could be dumped (probably truncated)
2166  * returns <0 in case of an error
2167  */
2168 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2169                          struct extent_buffer *eb, void *ctx)
2170 {
2171         struct inode_fs_paths *ipath = ctx;
2172         char *fspath;
2173         char *fspath_min;
2174         int i = ipath->fspath->elem_cnt;
2175         const int s_ptr = sizeof(char *);
2176         u32 bytes_left;
2177
2178         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2179                                         ipath->fspath->bytes_left - s_ptr : 0;
2180
2181         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2182         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2183                                    name_off, eb, inum, fspath_min, bytes_left);
2184         if (IS_ERR(fspath))
2185                 return PTR_ERR(fspath);
2186
2187         if (fspath > fspath_min) {
2188                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2189                 ++ipath->fspath->elem_cnt;
2190                 ipath->fspath->bytes_left = fspath - fspath_min;
2191         } else {
2192                 ++ipath->fspath->elem_missed;
2193                 ipath->fspath->bytes_missing += fspath_min - fspath;
2194                 ipath->fspath->bytes_left = 0;
2195         }
2196
2197         return 0;
2198 }
2199
2200 /*
2201  * this dumps all file system paths to the inode into the ipath struct, provided
2202  * is has been created large enough. each path is zero-terminated and accessed
2203  * from ipath->fspath->val[i].
2204  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2205  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2206  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2207  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2208  * have been needed to return all paths.
2209  */
2210 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2211 {
2212         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2213                              inode_to_path, ipath);
2214 }
2215
2216 struct btrfs_data_container *init_data_container(u32 total_bytes)
2217 {
2218         struct btrfs_data_container *data;
2219         size_t alloc_bytes;
2220
2221         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2222         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2223         if (!data)
2224                 return ERR_PTR(-ENOMEM);
2225
2226         if (total_bytes >= sizeof(*data)) {
2227                 data->bytes_left = total_bytes - sizeof(*data);
2228                 data->bytes_missing = 0;
2229         } else {
2230                 data->bytes_missing = sizeof(*data) - total_bytes;
2231                 data->bytes_left = 0;
2232         }
2233
2234         data->elem_cnt = 0;
2235         data->elem_missed = 0;
2236
2237         return data;
2238 }
2239
2240 /*
2241  * allocates space to return multiple file system paths for an inode.
2242  * total_bytes to allocate are passed, note that space usable for actual path
2243  * information will be total_bytes - sizeof(struct inode_fs_paths).
2244  * the returned pointer must be freed with free_ipath() in the end.
2245  */
2246 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2247                                         struct btrfs_path *path)
2248 {
2249         struct inode_fs_paths *ifp;
2250         struct btrfs_data_container *fspath;
2251
2252         fspath = init_data_container(total_bytes);
2253         if (IS_ERR(fspath))
2254                 return (void *)fspath;
2255
2256         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2257         if (!ifp) {
2258                 kvfree(fspath);
2259                 return ERR_PTR(-ENOMEM);
2260         }
2261
2262         ifp->btrfs_path = path;
2263         ifp->fspath = fspath;
2264         ifp->fs_root = fs_root;
2265
2266         return ifp;
2267 }
2268
2269 void free_ipath(struct inode_fs_paths *ipath)
2270 {
2271         if (!ipath)
2272                 return;
2273         kvfree(ipath->fspath);
2274         kfree(ipath);
2275 }