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/vmalloc.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "backref.h"
  23 #include "ulist.h"
  24 #include "transaction.h"
  25 #include "delayed-ref.h"
  26 #include "locking.h"
  27
  28 struct extent_inode_elem {
  29         u64 inum;
  30         u64 offset;
  31         struct extent_inode_elem *next;
  32 };
  33
  34 static int check_extent_in_eb(struct btrfs_key *key, struct extent_buffer *eb,
  35                                 struct btrfs_file_extent_item *fi,
  36                                 u64 extent_item_pos,
  37                                 struct extent_inode_elem **eie)
  38 {
  39         u64 data_offset;
  40         u64 data_len;
  41         struct extent_inode_elem *e;
  42
  43         data_offset = btrfs_file_extent_offset(eb, fi);
  44         data_len = btrfs_file_extent_num_bytes(eb, fi);
  45
  46         if (extent_item_pos < data_offset ||
  47             extent_item_pos >= data_offset + data_len)
  48                 return 1;
  49
  50         e = kmalloc(sizeof(*e), GFP_NOFS);
  51         if (!e)
  52                 return -ENOMEM;
  53
  54         e->next = *eie;
  55         e->inum = key->objectid;
  56         e->offset = key->offset + (extent_item_pos - data_offset);
  57         *eie = e;
  58
  59         return 0;
  60 }
  61
  62 static int find_extent_in_eb(struct extent_buffer *eb, u64 wanted_disk_byte,
  63                                 u64 extent_item_pos,
  64                                 struct extent_inode_elem **eie)
  65 {
  66         u64 disk_byte;
  67         struct btrfs_key key;
  68         struct btrfs_file_extent_item *fi;
  69         int slot;
  70         int nritems;
  71         int extent_type;
  72         int ret;
  73
  74         /*
  75          * from the shared data ref, we only have the leaf but we need
  76          * the key. thus, we must look into all items and see that we
  77          * find one (some) with a reference to our extent item.
  78          */
  79         nritems = btrfs_header_nritems(eb);
  80         for (slot = 0; slot < nritems; ++slot) {
  81                 btrfs_item_key_to_cpu(eb, &key, slot);
  82                 if (key.type != BTRFS_EXTENT_DATA_KEY)
  83                         continue;
  84                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
  85                 extent_type = btrfs_file_extent_type(eb, fi);
  86                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
  87                         continue;
  88                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
  89                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
  90                 if (disk_byte != wanted_disk_byte)
  91                         continue;
  92
  93                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
  94                 if (ret < 0)
  95                         return ret;
  96         }
  97
  98         return 0;
  99 }
 100
 101 /*
 102  * this structure records all encountered refs on the way up to the root
 103  */
 104 struct __prelim_ref {
 105         struct list_head list;
 106         u64 root_id;
 107         struct btrfs_key key_for_search;
 108         int level;
 109         int count;
 110         struct extent_inode_elem *inode_list;
 111         u64 parent;
 112         u64 wanted_disk_byte;
 113 };
 114
 115 /*
 116  * the rules for all callers of this function are:
 117  * - obtaining the parent is the goal
 118  * - if you add a key, you must know that it is a correct key
 119  * - if you cannot add the parent or a correct key, then we will look into the
 120  *   block later to set a correct key
 121  *
 122  * delayed refs
 123  * ============
 124  *        backref type | shared | indirect | shared | indirect
 125  * information         |   tree |     tree |   data |     data
 126  * --------------------+--------+----------+--------+----------
 127  *      parent logical |    y   |     -    |    -   |     -
 128  *      key to resolve |    -   |     y    |    y   |     y
 129  *  tree block logical |    -   |     -    |    -   |     -
 130  *  root for resolving |    y   |     y    |    y   |     y
 131  *
 132  * - column 1:       we've the parent -> done
 133  * - column 2, 3, 4: we use the key to find the parent
 134  *
 135  * on disk refs (inline or keyed)
 136  * ==============================
 137  *        backref type | shared | indirect | shared | indirect
 138  * information         |   tree |     tree |   data |     data
 139  * --------------------+--------+----------+--------+----------
 140  *      parent logical |    y   |     -    |    y   |     -
 141  *      key to resolve |    -   |     -    |    -   |     y
 142  *  tree block logical |    y   |     y    |    y   |     y
 143  *  root for resolving |    -   |     y    |    y   |     y
 144  *
 145  * - column 1, 3: we've the parent -> done
 146  * - column 2:    we take the first key from the block to find the parent
 147  *                (see __add_missing_keys)
 148  * - column 4:    we use the key to find the parent
 149  *
 150  * additional information that's available but not required to find the parent
 151  * block might help in merging entries to gain some speed.
 152  */
 153
 154 static int __add_prelim_ref(struct list_head *head, u64 root_id,
 155                             struct btrfs_key *key, int level,
 156                             u64 parent, u64 wanted_disk_byte, int count)
 157 {
 158         struct __prelim_ref *ref;
 159
 160         /* in case we're adding delayed refs, we're holding the refs spinlock */
 161         ref = kmalloc(sizeof(*ref), GFP_ATOMIC);
 162         if (!ref)
 163                 return -ENOMEM;
 164
 165         ref->root_id = root_id;
 166         if (key)
 167                 ref->key_for_search = *key;
 168         else
 169                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 170
 171         ref->inode_list = NULL;
 172         ref->level = level;
 173         ref->count = count;
 174         ref->parent = parent;
 175         ref->wanted_disk_byte = wanted_disk_byte;
 176         list_add_tail(&ref->list, head);
 177
 178         return 0;
 179 }
 180
 181 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 182                                 struct ulist *parents, int level,
 183                                 struct btrfs_key *key_for_search, u64 time_seq,
 184                                 u64 wanted_disk_byte,
 185                                 const u64 *extent_item_pos)
 186 {
 187         int ret = 0;
 188         int slot;
 189         struct extent_buffer *eb;
 190         struct btrfs_key key;
 191         struct btrfs_file_extent_item *fi;
 192         struct extent_inode_elem *eie = NULL;
 193         u64 disk_byte;
 194
 195         if (level != 0) {
 196                 eb = path->nodes[level];
 197                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 198                 if (ret < 0)
 199                         return ret;
 200                 return 0;
 201         }
 202
 203         /*
 204          * We normally enter this function with the path already pointing to
 205          * the first item to check. But sometimes, we may enter it with
 206          * slot==nritems. In that case, go to the next leaf before we continue.
 207          */
 208         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0]))
 209                 ret = btrfs_next_old_leaf(root, path, time_seq);
 210
 211         while (!ret) {
 212                 eb = path->nodes[0];
 213                 slot = path->slots[0];
 214
 215                 btrfs_item_key_to_cpu(eb, &key, slot);
 216
 217                 if (key.objectid != key_for_search->objectid ||
 218                     key.type != BTRFS_EXTENT_DATA_KEY)
 219                         break;
 220
 221                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 222                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 223
 224                 if (disk_byte == wanted_disk_byte) {
 225                         eie = NULL;
 226                         if (extent_item_pos) {
 227                                 ret = check_extent_in_eb(&key, eb, fi,
 228                                                 *extent_item_pos,
 229                                                 &eie);
 230                                 if (ret < 0)
 231                                         break;
 232                         }
 233                         if (!ret) {
 234                                 ret = ulist_add(parents, eb->start,
 235                                                 (uintptr_t)eie, GFP_NOFS);
 236                                 if (ret < 0)
 237                                         break;
 238                                 if (!extent_item_pos) {
 239                                         ret = btrfs_next_old_leaf(root, path,
 240                                                         time_seq);
 241                                         continue;
 242                                 }
 243                         }
 244                 }
 245                 ret = btrfs_next_old_item(root, path, time_seq);
 246         }
 247
 248         if (ret > 0)
 249                 ret = 0;
 250         return ret;
 251 }
 252
 253 /*
 254  * resolve an indirect backref in the form (root_id, key, level)
 255  * to a logical address
 256  */
 257 static int __resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 258                                         int search_commit_root,
 259                                         u64 time_seq,
 260                                         struct __prelim_ref *ref,
 261                                         struct ulist *parents,
 262                                         const u64 *extent_item_pos)
 263 {
 264         struct btrfs_path *path;
 265         struct btrfs_root *root;
 266         struct btrfs_key root_key;
 267         struct extent_buffer *eb;
 268         int ret = 0;
 269         int root_level;
 270         int level = ref->level;
 271
 272         path = btrfs_alloc_path();
 273         if (!path)
 274                 return -ENOMEM;
 275         path->search_commit_root = !!search_commit_root;
 276
 277         root_key.objectid = ref->root_id;
 278         root_key.type = BTRFS_ROOT_ITEM_KEY;
 279         root_key.offset = (u64)-1;
 280         root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 281         if (IS_ERR(root)) {
 282                 ret = PTR_ERR(root);
 283                 goto out;
 284         }
 285
 286         root_level = btrfs_old_root_level(root, time_seq);
 287
 288         if (root_level + 1 == level)
 289                 goto out;
 290
 291         path->lowest_level = level;
 292         ret = btrfs_search_old_slot(root, &ref->key_for_search, path, time_seq);
 293         pr_debug("search slot in root %llu (level %d, ref count %d) returned "
 294                  "%d for key (%llu %u %llu)\n",
 295                  (unsigned long long)ref->root_id, level, ref->count, ret,
 296                  (unsigned long long)ref->key_for_search.objectid,
 297                  ref->key_for_search.type,
 298                  (unsigned long long)ref->key_for_search.offset);
 299         if (ret < 0)
 300                 goto out;
 301
 302         eb = path->nodes[level];
 303         while (!eb) {
 304                 if (!level) {
 305                         WARN_ON(1);
 306                         ret = 1;
 307                         goto out;
 308                 }
 309                 level--;
 310                 eb = path->nodes[level];
 311         }
 312
 313         ret = add_all_parents(root, path, parents, level, &ref->key_for_search,
 314                                 time_seq, ref->wanted_disk_byte,
 315                                 extent_item_pos);
 316 out:
 317         btrfs_free_path(path);
 318         return ret;
 319 }
 320
 321 /*
 322  * resolve all indirect backrefs from the list
 323  */
 324 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 325                                    int search_commit_root, u64 time_seq,
 326                                    struct list_head *head,
 327                                    const u64 *extent_item_pos)
 328 {
 329         int err;
 330         int ret = 0;
 331         struct __prelim_ref *ref;
 332         struct __prelim_ref *ref_safe;
 333         struct __prelim_ref *new_ref;
 334         struct ulist *parents;
 335         struct ulist_node *node;
 336         struct ulist_iterator uiter;
 337
 338         parents = ulist_alloc(GFP_NOFS);
 339         if (!parents)
 340                 return -ENOMEM;
 341
 342         /*
 343          * _safe allows us to insert directly after the current item without
 344          * iterating over the newly inserted items.
 345          * we're also allowed to re-assign ref during iteration.
 346          */
 347         list_for_each_entry_safe(ref, ref_safe, head, list) {
 348                 if (ref->parent)        /* already direct */
 349                         continue;
 350                 if (ref->count == 0)
 351                         continue;
 352                 err = __resolve_indirect_ref(fs_info, search_commit_root,
 353                                              time_seq, ref, parents,
 354                                              extent_item_pos);
 355                 if (err) {
 356                         if (ret == 0)
 357                                 ret = err;
 358                         continue;
 359                 }
 360
 361                 /* we put the first parent into the ref at hand */
 362                 ULIST_ITER_INIT(&uiter);
 363                 node = ulist_next(parents, &uiter);
 364                 ref->parent = node ? node->val : 0;
 365                 ref->inode_list = node ?
 366                         (struct extent_inode_elem *)(uintptr_t)node->aux : 0;
 367
 368                 /* additional parents require new refs being added here */
 369                 while ((node = ulist_next(parents, &uiter))) {
 370                         new_ref = kmalloc(sizeof(*new_ref), GFP_NOFS);
 371                         if (!new_ref) {
 372                                 ret = -ENOMEM;
 373                                 break;
 374                         }
 375                         memcpy(new_ref, ref, sizeof(*ref));
 376                         new_ref->parent = node->val;
 377                         new_ref->inode_list = (struct extent_inode_elem *)
 378                                                         (uintptr_t)node->aux;
 379                         list_add(&new_ref->list, &ref->list);
 380                 }
 381                 ulist_reinit(parents);
 382         }
 383
 384         ulist_free(parents);
 385         return ret;
 386 }
 387
 388 static inline int ref_for_same_block(struct __prelim_ref *ref1,
 389                                      struct __prelim_ref *ref2)
 390 {
 391         if (ref1->level != ref2->level)
 392                 return 0;
 393         if (ref1->root_id != ref2->root_id)
 394                 return 0;
 395         if (ref1->key_for_search.type != ref2->key_for_search.type)
 396                 return 0;
 397         if (ref1->key_for_search.objectid != ref2->key_for_search.objectid)
 398                 return 0;
 399         if (ref1->key_for_search.offset != ref2->key_for_search.offset)
 400                 return 0;
 401         if (ref1->parent != ref2->parent)
 402                 return 0;
 403
 404         return 1;
 405 }
 406
 407 /*
 408  * read tree blocks and add keys where required.
 409  */
 410 static int __add_missing_keys(struct btrfs_fs_info *fs_info,
 411                               struct list_head *head)
 412 {
 413         struct list_head *pos;
 414         struct extent_buffer *eb;
 415
 416         list_for_each(pos, head) {
 417                 struct __prelim_ref *ref;
 418                 ref = list_entry(pos, struct __prelim_ref, list);
 419
 420                 if (ref->parent)
 421                         continue;
 422                 if (ref->key_for_search.type)
 423                         continue;
 424                 BUG_ON(!ref->wanted_disk_byte);
 425                 eb = read_tree_block(fs_info->tree_root, ref->wanted_disk_byte,
 426                                      fs_info->tree_root->leafsize, 0);
 427                 BUG_ON(!eb);
 428                 btrfs_tree_read_lock(eb);
 429                 if (btrfs_header_level(eb) == 0)
 430                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 431                 else
 432                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 433                 btrfs_tree_read_unlock(eb);
 434                 free_extent_buffer(eb);
 435         }
 436         return 0;
 437 }
 438
 439 /*
 440  * merge two lists of backrefs and adjust counts accordingly
 441  *
 442  * mode = 1: merge identical keys, if key is set
 443  *    FIXME: if we add more keys in __add_prelim_ref, we can merge more here.
 444  *           additionally, we could even add a key range for the blocks we
 445  *           looked into to merge even more (-> replace unresolved refs by those
 446  *           having a parent).
 447  * mode = 2: merge identical parents
 448  */
 449 static int __merge_refs(struct list_head *head, int mode)
 450 {
 451         struct list_head *pos1;
 452
 453         list_for_each(pos1, head) {
 454                 struct list_head *n2;
 455                 struct list_head *pos2;
 456                 struct __prelim_ref *ref1;
 457
 458                 ref1 = list_entry(pos1, struct __prelim_ref, list);
 459
 460                 for (pos2 = pos1->next, n2 = pos2->next; pos2 != head;
 461                      pos2 = n2, n2 = pos2->next) {
 462                         struct __prelim_ref *ref2;
 463                         struct __prelim_ref *xchg;
 464                         struct extent_inode_elem *eie;
 465
 466                         ref2 = list_entry(pos2, struct __prelim_ref, list);
 467
 468                         if (mode == 1) {
 469                                 if (!ref_for_same_block(ref1, ref2))
 470                                         continue;
 471                                 if (!ref1->parent && ref2->parent) {
 472                                         xchg = ref1;
 473                                         ref1 = ref2;
 474                                         ref2 = xchg;
 475                                 }
 476                         } else {
 477                                 if (ref1->parent != ref2->parent)
 478                                         continue;
 479                         }
 480
 481                         eie = ref1->inode_list;
 482                         while (eie && eie->next)
 483                                 eie = eie->next;
 484                         if (eie)
 485                                 eie->next = ref2->inode_list;
 486                         else
 487                                 ref1->inode_list = ref2->inode_list;
 488                         ref1->count += ref2->count;
 489
 490                         list_del(&ref2->list);
 491                         kfree(ref2);
 492                 }
 493
 494         }
 495         return 0;
 496 }
 497
 498 /*
 499  * add all currently queued delayed refs from this head whose seq nr is
 500  * smaller or equal that seq to the list
 501  */
 502 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 503                               struct list_head *prefs)
 504 {
 505         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 506         struct rb_node *n = &head->node.rb_node;
 507         struct btrfs_key key;
 508         struct btrfs_key op_key = {0};
 509         int sgn;
 510         int ret = 0;
 511
 512         if (extent_op && extent_op->update_key)
 513                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 514
 515         while ((n = rb_prev(n))) {
 516                 struct btrfs_delayed_ref_node *node;
 517                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 518                                 rb_node);
 519                 if (node->bytenr != head->node.bytenr)
 520                         break;
 521                 WARN_ON(node->is_head);
 522
 523                 if (node->seq > seq)
 524                         continue;
 525
 526                 switch (node->action) {
 527                 case BTRFS_ADD_DELAYED_EXTENT:
 528                 case BTRFS_UPDATE_DELAYED_HEAD:
 529                         WARN_ON(1);
 530                         continue;
 531                 case BTRFS_ADD_DELAYED_REF:
 532                         sgn = 1;
 533                         break;
 534                 case BTRFS_DROP_DELAYED_REF:
 535                         sgn = -1;
 536                         break;
 537                 default:
 538                         BUG_ON(1);
 539                 }
 540                 switch (node->type) {
 541                 case BTRFS_TREE_BLOCK_REF_KEY: {
 542                         struct btrfs_delayed_tree_ref *ref;
 543
 544                         ref = btrfs_delayed_node_to_tree_ref(node);
 545                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
 546                                                ref->level + 1, 0, node->bytenr,
 547                                                node->ref_mod * sgn);
 548                         break;
 549                 }
 550                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 551                         struct btrfs_delayed_tree_ref *ref;
 552
 553                         ref = btrfs_delayed_node_to_tree_ref(node);
 554                         ret = __add_prelim_ref(prefs, ref->root, NULL,
 555                                                ref->level + 1, ref->parent,
 556                                                node->bytenr,
 557                                                node->ref_mod * sgn);
 558                         break;
 559                 }
 560                 case BTRFS_EXTENT_DATA_REF_KEY: {
 561                         struct btrfs_delayed_data_ref *ref;
 562                         ref = btrfs_delayed_node_to_data_ref(node);
 563
 564                         key.objectid = ref->objectid;
 565                         key.type = BTRFS_EXTENT_DATA_KEY;
 566                         key.offset = ref->offset;
 567                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
 568                                                node->bytenr,
 569                                                node->ref_mod * sgn);
 570                         break;
 571                 }
 572                 case BTRFS_SHARED_DATA_REF_KEY: {
 573                         struct btrfs_delayed_data_ref *ref;
 574
 575                         ref = btrfs_delayed_node_to_data_ref(node);
 576
 577                         key.objectid = ref->objectid;
 578                         key.type = BTRFS_EXTENT_DATA_KEY;
 579                         key.offset = ref->offset;
 580                         ret = __add_prelim_ref(prefs, ref->root, &key, 0,
 581                                                ref->parent, node->bytenr,
 582                                                node->ref_mod * sgn);
 583                         break;
 584                 }
 585                 default:
 586                         WARN_ON(1);
 587                 }
 588                 BUG_ON(ret);
 589         }
 590
 591         return 0;
 592 }
 593
 594 /*
 595  * add all inline backrefs for bytenr to the list
 596  */
 597 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
 598                              struct btrfs_path *path, u64 bytenr,
 599                              int *info_level, struct list_head *prefs)
 600 {
 601         int ret = 0;
 602         int slot;
 603         struct extent_buffer *leaf;
 604         struct btrfs_key key;
 605         unsigned long ptr;
 606         unsigned long end;
 607         struct btrfs_extent_item *ei;
 608         u64 flags;
 609         u64 item_size;
 610
 611         /*
 612          * enumerate all inline refs
 613          */
 614         leaf = path->nodes[0];
 615         slot = path->slots[0];
 616
 617         item_size = btrfs_item_size_nr(leaf, slot);
 618         BUG_ON(item_size < sizeof(*ei));
 619
 620         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 621         flags = btrfs_extent_flags(leaf, ei);
 622
 623         ptr = (unsigned long)(ei + 1);
 624         end = (unsigned long)ei + item_size;
 625
 626         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 627                 struct btrfs_tree_block_info *info;
 628
 629                 info = (struct btrfs_tree_block_info *)ptr;
 630                 *info_level = btrfs_tree_block_level(leaf, info);
 631                 ptr += sizeof(struct btrfs_tree_block_info);
 632                 BUG_ON(ptr > end);
 633         } else {
 634                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 635         }
 636
 637         while (ptr < end) {
 638                 struct btrfs_extent_inline_ref *iref;
 639                 u64 offset;
 640                 int type;
 641
 642                 iref = (struct btrfs_extent_inline_ref *)ptr;
 643                 type = btrfs_extent_inline_ref_type(leaf, iref);
 644                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 645
 646                 switch (type) {
 647                 case BTRFS_SHARED_BLOCK_REF_KEY:
 648                         ret = __add_prelim_ref(prefs, 0, NULL,
 649                                                 *info_level + 1, offset,
 650                                                 bytenr, 1);
 651                         break;
 652                 case BTRFS_SHARED_DATA_REF_KEY: {
 653                         struct btrfs_shared_data_ref *sdref;
 654                         int count;
 655
 656                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 657                         count = btrfs_shared_data_ref_count(leaf, sdref);
 658                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
 659                                                bytenr, count);
 660                         break;
 661                 }
 662                 case BTRFS_TREE_BLOCK_REF_KEY:
 663                         ret = __add_prelim_ref(prefs, offset, NULL,
 664                                                *info_level + 1, 0,
 665                                                bytenr, 1);
 666                         break;
 667                 case BTRFS_EXTENT_DATA_REF_KEY: {
 668                         struct btrfs_extent_data_ref *dref;
 669                         int count;
 670                         u64 root;
 671
 672                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 673                         count = btrfs_extent_data_ref_count(leaf, dref);
 674                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 675                                                                       dref);
 676                         key.type = BTRFS_EXTENT_DATA_KEY;
 677                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 678                         root = btrfs_extent_data_ref_root(leaf, dref);
 679                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 680                                                bytenr, count);
 681                         break;
 682                 }
 683                 default:
 684                         WARN_ON(1);
 685                 }
 686                 BUG_ON(ret);
 687                 ptr += btrfs_extent_inline_ref_size(type);
 688         }
 689
 690         return 0;
 691 }
 692
 693 /*
 694  * add all non-inline backrefs for bytenr to the list
 695  */
 696 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
 697                             struct btrfs_path *path, u64 bytenr,
 698                             int info_level, struct list_head *prefs)
 699 {
 700         struct btrfs_root *extent_root = fs_info->extent_root;
 701         int ret;
 702         int slot;
 703         struct extent_buffer *leaf;
 704         struct btrfs_key key;
 705
 706         while (1) {
 707                 ret = btrfs_next_item(extent_root, path);
 708                 if (ret < 0)
 709                         break;
 710                 if (ret) {
 711                         ret = 0;
 712                         break;
 713                 }
 714
 715                 slot = path->slots[0];
 716                 leaf = path->nodes[0];
 717                 btrfs_item_key_to_cpu(leaf, &key, slot);
 718
 719                 if (key.objectid != bytenr)
 720                         break;
 721                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
 722                         continue;
 723                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
 724                         break;
 725
 726                 switch (key.type) {
 727                 case BTRFS_SHARED_BLOCK_REF_KEY:
 728                         ret = __add_prelim_ref(prefs, 0, NULL,
 729                                                 info_level + 1, key.offset,
 730                                                 bytenr, 1);
 731                         break;
 732                 case BTRFS_SHARED_DATA_REF_KEY: {
 733                         struct btrfs_shared_data_ref *sdref;
 734                         int count;
 735
 736                         sdref = btrfs_item_ptr(leaf, slot,
 737                                               struct btrfs_shared_data_ref);
 738                         count = btrfs_shared_data_ref_count(leaf, sdref);
 739                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
 740                                                 bytenr, count);
 741                         break;
 742                 }
 743                 case BTRFS_TREE_BLOCK_REF_KEY:
 744                         ret = __add_prelim_ref(prefs, key.offset, NULL,
 745                                                info_level + 1, 0,
 746                                                bytenr, 1);
 747                         break;
 748                 case BTRFS_EXTENT_DATA_REF_KEY: {
 749                         struct btrfs_extent_data_ref *dref;
 750                         int count;
 751                         u64 root;
 752
 753                         dref = btrfs_item_ptr(leaf, slot,
 754                                               struct btrfs_extent_data_ref);
 755                         count = btrfs_extent_data_ref_count(leaf, dref);
 756                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 757                                                                       dref);
 758                         key.type = BTRFS_EXTENT_DATA_KEY;
 759                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 760                         root = btrfs_extent_data_ref_root(leaf, dref);
 761                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 762                                                bytenr, count);
 763                         break;
 764                 }
 765                 default:
 766                         WARN_ON(1);
 767                 }
 768                 BUG_ON(ret);
 769         }
 770
 771         return ret;
 772 }
 773
 774 /*
 775  * this adds all existing backrefs (inline backrefs, backrefs and delayed
 776  * refs) for the given bytenr to the refs list, merges duplicates and resolves
 777  * indirect refs to their parent bytenr.
 778  * When roots are found, they're added to the roots list
 779  *
 780  * FIXME some caching might speed things up
 781  */
 782 static int find_parent_nodes(struct btrfs_trans_handle *trans,
 783                              struct btrfs_fs_info *fs_info, u64 bytenr,
 784                              u64 time_seq, struct ulist *refs,
 785                              struct ulist *roots, const u64 *extent_item_pos)
 786 {
 787         struct btrfs_key key;
 788         struct btrfs_path *path;
 789         struct btrfs_delayed_ref_root *delayed_refs = NULL;
 790         struct btrfs_delayed_ref_head *head;
 791         int info_level = 0;
 792         int ret;
 793         int search_commit_root = (trans == BTRFS_BACKREF_SEARCH_COMMIT_ROOT);
 794         struct list_head prefs_delayed;
 795         struct list_head prefs;
 796         struct __prelim_ref *ref;
 797
 798         INIT_LIST_HEAD(&prefs);
 799         INIT_LIST_HEAD(&prefs_delayed);
 800
 801         key.objectid = bytenr;
 802         key.type = BTRFS_EXTENT_ITEM_KEY;
 803         key.offset = (u64)-1;
 804
 805         path = btrfs_alloc_path();
 806         if (!path)
 807                 return -ENOMEM;
 808         path->search_commit_root = !!search_commit_root;
 809
 810         /*
 811          * grab both a lock on the path and a lock on the delayed ref head.
 812          * We need both to get a consistent picture of how the refs look
 813          * at a specified point in time
 814          */
 815 again:
 816         head = NULL;
 817
 818         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 819         if (ret < 0)
 820                 goto out;
 821         BUG_ON(ret == 0);
 822
 823         if (trans != BTRFS_BACKREF_SEARCH_COMMIT_ROOT) {
 824                 /*
 825                  * look if there are updates for this ref queued and lock the
 826                  * head
 827                  */
 828                 delayed_refs = &trans->transaction->delayed_refs;
 829                 spin_lock(&delayed_refs->lock);
 830                 head = btrfs_find_delayed_ref_head(trans, bytenr);
 831                 if (head) {
 832                         if (!mutex_trylock(&head->mutex)) {
 833                                 atomic_inc(&head->node.refs);
 834                                 spin_unlock(&delayed_refs->lock);
 835
 836                                 btrfs_release_path(path);
 837
 838                                 /*
 839                                  * Mutex was contended, block until it's
 840                                  * released and try again
 841                                  */
 842                                 mutex_lock(&head->mutex);
 843                                 mutex_unlock(&head->mutex);
 844                                 btrfs_put_delayed_ref(&head->node);
 845                                 goto again;
 846                         }
 847                         ret = __add_delayed_refs(head, time_seq,
 848                                                  &prefs_delayed);
 849                         mutex_unlock(&head->mutex);
 850                         if (ret) {
 851                                 spin_unlock(&delayed_refs->lock);
 852                                 goto out;
 853                         }
 854                 }
 855                 spin_unlock(&delayed_refs->lock);
 856         }
 857
 858         if (path->slots[0]) {
 859                 struct extent_buffer *leaf;
 860                 int slot;
 861
 862                 path->slots[0]--;
 863                 leaf = path->nodes[0];
 864                 slot = path->slots[0];
 865                 btrfs_item_key_to_cpu(leaf, &key, slot);
 866                 if (key.objectid == bytenr &&
 867                     key.type == BTRFS_EXTENT_ITEM_KEY) {
 868                         ret = __add_inline_refs(fs_info, path, bytenr,
 869                                                 &info_level, &prefs);
 870                         if (ret)
 871                                 goto out;
 872                         ret = __add_keyed_refs(fs_info, path, bytenr,
 873                                                info_level, &prefs);
 874                         if (ret)
 875                                 goto out;
 876                 }
 877         }
 878         btrfs_release_path(path);
 879
 880         list_splice_init(&prefs_delayed, &prefs);
 881
 882         ret = __add_missing_keys(fs_info, &prefs);
 883         if (ret)
 884                 goto out;
 885
 886         ret = __merge_refs(&prefs, 1);
 887         if (ret)
 888                 goto out;
 889
 890         ret = __resolve_indirect_refs(fs_info, search_commit_root, time_seq,
 891                                       &prefs, extent_item_pos);
 892         if (ret)
 893                 goto out;
 894
 895         ret = __merge_refs(&prefs, 2);
 896         if (ret)
 897                 goto out;
 898
 899         while (!list_empty(&prefs)) {
 900                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
 901                 list_del(&ref->list);
 902                 WARN_ON(ref->count < 0);
 903                 if (ref->count && ref->root_id && ref->parent == 0) {
 904                         /* no parent == root of tree */
 905                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
 906                         BUG_ON(ret < 0);
 907                 }
 908                 if (ref->count && ref->parent) {
 909                         struct extent_inode_elem *eie = NULL;
 910                         if (extent_item_pos && !ref->inode_list) {
 911                                 u32 bsz;
 912                                 struct extent_buffer *eb;
 913                                 bsz = btrfs_level_size(fs_info->extent_root,
 914                                                         info_level);
 915                                 eb = read_tree_block(fs_info->extent_root,
 916                                                            ref->parent, bsz, 0);
 917                                 BUG_ON(!eb);
 918                                 ret = find_extent_in_eb(eb, bytenr,
 919                                                         *extent_item_pos, &eie);
 920                                 ref->inode_list = eie;
 921                                 free_extent_buffer(eb);
 922                         }
 923                         ret = ulist_add_merge(refs, ref->parent,
 924                                               (uintptr_t)ref->inode_list,
 925                                               (u64 *)&eie, GFP_NOFS);
 926                         if (!ret && extent_item_pos) {
 927                                 /*
 928                                  * we've recorded that parent, so we must extend
 929                                  * its inode list here
 930                                  */
 931                                 BUG_ON(!eie);
 932                                 while (eie->next)
 933                                         eie = eie->next;
 934                                 eie->next = ref->inode_list;
 935                         }
 936                         BUG_ON(ret < 0);
 937                 }
 938                 kfree(ref);
 939         }
 940
 941 out:
 942         btrfs_free_path(path);
 943         while (!list_empty(&prefs)) {
 944                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
 945                 list_del(&ref->list);
 946                 kfree(ref);
 947         }
 948         while (!list_empty(&prefs_delayed)) {
 949                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
 950                                        list);
 951                 list_del(&ref->list);
 952                 kfree(ref);
 953         }
 954
 955         return ret;
 956 }
 957
 958 static void free_leaf_list(struct ulist *blocks)
 959 {
 960         struct ulist_node *node = NULL;
 961         struct extent_inode_elem *eie;
 962         struct extent_inode_elem *eie_next;
 963         struct ulist_iterator uiter;
 964
 965         ULIST_ITER_INIT(&uiter);
 966         while ((node = ulist_next(blocks, &uiter))) {
 967                 if (!node->aux)
 968                         continue;
 969                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
 970                 for (; eie; eie = eie_next) {
 971                         eie_next = eie->next;
 972                         kfree(eie);
 973                 }
 974                 node->aux = 0;
 975         }
 976
 977         ulist_free(blocks);
 978 }
 979
 980 /*
 981  * Finds all leafs with a reference to the specified combination of bytenr and
 982  * offset. key_list_head will point to a list of corresponding keys (caller must
 983  * free each list element). The leafs will be stored in the leafs ulist, which
 984  * must be freed with ulist_free.
 985  *
 986  * returns 0 on success, <0 on error
 987  */
 988 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
 989                                 struct btrfs_fs_info *fs_info, u64 bytenr,
 990                                 u64 time_seq, struct ulist **leafs,
 991                                 const u64 *extent_item_pos)
 992 {
 993         struct ulist *tmp;
 994         int ret;
 995
 996         tmp = ulist_alloc(GFP_NOFS);
 997         if (!tmp)
 998                 return -ENOMEM;
 999         *leafs = ulist_alloc(GFP_NOFS);
1000         if (!*leafs) {
1001                 ulist_free(tmp);
1002                 return -ENOMEM;
1003         }
1004
1005         ret = find_parent_nodes(trans, fs_info, bytenr,
1006                                 time_seq, *leafs, tmp, extent_item_pos);
1007         ulist_free(tmp);
1008
1009         if (ret < 0 && ret != -ENOENT) {
1010                 free_leaf_list(*leafs);
1011                 return ret;
1012         }
1013
1014         return 0;
1015 }
1016
1017 /*
1018  * walk all backrefs for a given extent to find all roots that reference this
1019  * extent. Walking a backref means finding all extents that reference this
1020  * extent and in turn walk the backrefs of those, too. Naturally this is a
1021  * recursive process, but here it is implemented in an iterative fashion: We
1022  * find all referencing extents for the extent in question and put them on a
1023  * list. In turn, we find all referencing extents for those, further appending
1024  * to the list. The way we iterate the list allows adding more elements after
1025  * the current while iterating. The process stops when we reach the end of the
1026  * list. Found roots are added to the roots list.
1027  *
1028  * returns 0 on success, < 0 on error.
1029  */
1030 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1031                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1032                                 u64 time_seq, struct ulist **roots)
1033 {
1034         struct ulist *tmp;
1035         struct ulist_node *node = NULL;
1036         struct ulist_iterator uiter;
1037         int ret;
1038
1039         tmp = ulist_alloc(GFP_NOFS);
1040         if (!tmp)
1041                 return -ENOMEM;
1042         *roots = ulist_alloc(GFP_NOFS);
1043         if (!*roots) {
1044                 ulist_free(tmp);
1045                 return -ENOMEM;
1046         }
1047
1048         ULIST_ITER_INIT(&uiter);
1049         while (1) {
1050                 ret = find_parent_nodes(trans, fs_info, bytenr,
1051                                         time_seq, tmp, *roots, NULL);
1052                 if (ret < 0 && ret != -ENOENT) {
1053                         ulist_free(tmp);
1054                         ulist_free(*roots);
1055                         return ret;
1056                 }
1057                 node = ulist_next(tmp, &uiter);
1058                 if (!node)
1059                         break;
1060                 bytenr = node->val;
1061         }
1062
1063         ulist_free(tmp);
1064         return 0;
1065 }
1066
1067
1068 static int __inode_info(u64 inum, u64 ioff, u8 key_type,
1069                         struct btrfs_root *fs_root, struct btrfs_path *path,
1070                         struct btrfs_key *found_key)
1071 {
1072         int ret;
1073         struct btrfs_key key;
1074         struct extent_buffer *eb;
1075
1076         key.type = key_type;
1077         key.objectid = inum;
1078         key.offset = ioff;
1079
1080         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1081         if (ret < 0)
1082                 return ret;
1083
1084         eb = path->nodes[0];
1085         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1086                 ret = btrfs_next_leaf(fs_root, path);
1087                 if (ret)
1088                         return ret;
1089                 eb = path->nodes[0];
1090         }
1091
1092         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1093         if (found_key->type != key.type || found_key->objectid != key.objectid)
1094                 return 1;
1095
1096         return 0;
1097 }
1098
1099 /*
1100  * this makes the path point to (inum INODE_ITEM ioff)
1101  */
1102 int inode_item_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1103                         struct btrfs_path *path)
1104 {
1105         struct btrfs_key key;
1106         return __inode_info(inum, ioff, BTRFS_INODE_ITEM_KEY, fs_root, path,
1107                                 &key);
1108 }
1109
1110 static int inode_ref_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1111                                 struct btrfs_path *path,
1112                                 struct btrfs_key *found_key)
1113 {
1114         return __inode_info(inum, ioff, BTRFS_INODE_REF_KEY, fs_root, path,
1115                                 found_key);
1116 }
1117
1118 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1119                           u64 start_off, struct btrfs_path *path,
1120                           struct btrfs_inode_extref **ret_extref,
1121                           u64 *found_off)
1122 {
1123         int ret, slot;
1124         struct btrfs_key key;
1125         struct btrfs_key found_key;
1126         struct btrfs_inode_extref *extref;
1127         struct extent_buffer *leaf;
1128         unsigned long ptr;
1129
1130         key.objectid = inode_objectid;
1131         btrfs_set_key_type(&key, BTRFS_INODE_EXTREF_KEY);
1132         key.offset = start_off;
1133
1134         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1135         if (ret < 0)
1136                 return ret;
1137
1138         while (1) {
1139                 leaf = path->nodes[0];
1140                 slot = path->slots[0];
1141                 if (slot >= btrfs_header_nritems(leaf)) {
1142                         /*
1143                          * If the item at offset is not found,
1144                          * btrfs_search_slot will point us to the slot
1145                          * where it should be inserted. In our case
1146                          * that will be the slot directly before the
1147                          * next INODE_REF_KEY_V2 item. In the case
1148                          * that we're pointing to the last slot in a
1149                          * leaf, we must move one leaf over.
1150                          */
1151                         ret = btrfs_next_leaf(root, path);
1152                         if (ret) {
1153                                 if (ret >= 1)
1154                                         ret = -ENOENT;
1155                                 break;
1156                         }
1157                         continue;
1158                 }
1159
1160                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1161
1162                 /*
1163                  * Check that we're still looking at an extended ref key for
1164                  * this particular objectid. If we have different
1165                  * objectid or type then there are no more to be found
1166                  * in the tree and we can exit.
1167                  */
1168                 ret = -ENOENT;
1169                 if (found_key.objectid != inode_objectid)
1170                         break;
1171                 if (btrfs_key_type(&found_key) != BTRFS_INODE_EXTREF_KEY)
1172                         break;
1173
1174                 ret = 0;
1175                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1176                 extref = (struct btrfs_inode_extref *)ptr;
1177                 *ret_extref = extref;
1178                 if (found_off)
1179                         *found_off = found_key.offset;
1180                 break;
1181         }
1182
1183         return ret;
1184 }
1185
1186 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1187                         u32 name_len, unsigned long name_off,
1188                         struct extent_buffer *eb_in, u64 parent,
1189                         char *dest, u32 size)
1190 {
1191         int slot;
1192         u64 next_inum;
1193         int ret;
1194         s64 bytes_left = ((s64)size) - 1;
1195         struct extent_buffer *eb = eb_in;
1196         struct btrfs_key found_key;
1197         int leave_spinning = path->leave_spinning;
1198         struct btrfs_inode_ref *iref;
1199
1200         if (bytes_left >= 0)
1201                 dest[bytes_left] = '\0';
1202
1203         path->leave_spinning = 1;
1204         while (1) {
1205                 bytes_left -= name_len;
1206                 if (bytes_left >= 0)
1207                         read_extent_buffer(eb, dest + bytes_left,
1208                                            name_off, name_len);
1209                 if (eb != eb_in) {
1210                         btrfs_tree_read_unlock_blocking(eb);
1211                         free_extent_buffer(eb);
1212                 }
1213                 ret = inode_ref_info(parent, 0, fs_root, path, &found_key);
1214                 if (ret > 0)
1215                         ret = -ENOENT;
1216                 if (ret)
1217                         break;
1218
1219                 next_inum = found_key.offset;
1220
1221                 /* regular exit ahead */
1222                 if (parent == next_inum)
1223                         break;
1224
1225                 slot = path->slots[0];
1226                 eb = path->nodes[0];
1227                 /* make sure we can use eb after releasing the path */
1228                 if (eb != eb_in) {
1229                         atomic_inc(&eb->refs);
1230                         btrfs_tree_read_lock(eb);
1231                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1232                 }
1233                 btrfs_release_path(path);
1234                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1235
1236                 name_len = btrfs_inode_ref_name_len(eb, iref);
1237                 name_off = (unsigned long)(iref + 1);
1238
1239                 parent = next_inum;
1240                 --bytes_left;
1241                 if (bytes_left >= 0)
1242                         dest[bytes_left] = '/';
1243         }
1244
1245         btrfs_release_path(path);
1246         path->leave_spinning = leave_spinning;
1247
1248         if (ret)
1249                 return ERR_PTR(ret);
1250
1251         return dest + bytes_left;
1252 }
1253
1254 /*
1255  * this iterates to turn a btrfs_inode_ref into a full filesystem path. elements
1256  * of the path are separated by '/' and the path is guaranteed to be
1257  * 0-terminated. the path is only given within the current file system.
1258  * Therefore, it never starts with a '/'. the caller is responsible to provide
1259  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1260  * the start point of the resulting string is returned. this pointer is within
1261  * dest, normally.
1262  * in case the path buffer would overflow, the pointer is decremented further
1263  * as if output was written to the buffer, though no more output is actually
1264  * generated. that way, the caller can determine how much space would be
1265  * required for the path to fit into the buffer. in that case, the returned
1266  * value will be smaller than dest. callers must check this!
1267  */
1268 char *btrfs_iref_to_path(struct btrfs_root *fs_root,
1269                          struct btrfs_path *path,
1270                          struct btrfs_inode_ref *iref,
1271                          struct extent_buffer *eb_in, u64 parent,
1272                          char *dest, u32 size)
1273 {
1274         return btrfs_ref_to_path(fs_root, path,
1275                                  btrfs_inode_ref_name_len(eb_in, iref),
1276                                  (unsigned long)(iref + 1),
1277                                  eb_in, parent, dest, size);
1278 }
1279
1280 /*
1281  * this makes the path point to (logical EXTENT_ITEM *)
1282  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1283  * tree blocks and <0 on error.
1284  */
1285 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1286                         struct btrfs_path *path, struct btrfs_key *found_key,
1287                         u64 *flags_ret)
1288 {
1289         int ret;
1290         u64 flags;
1291         u32 item_size;
1292         struct extent_buffer *eb;
1293         struct btrfs_extent_item *ei;
1294         struct btrfs_key key;
1295
1296         key.type = BTRFS_EXTENT_ITEM_KEY;
1297         key.objectid = logical;
1298         key.offset = (u64)-1;
1299
1300         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1301         if (ret < 0)
1302                 return ret;
1303         ret = btrfs_previous_item(fs_info->extent_root, path,
1304                                         0, BTRFS_EXTENT_ITEM_KEY);
1305         if (ret < 0)
1306                 return ret;
1307
1308         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1309         if (found_key->type != BTRFS_EXTENT_ITEM_KEY ||
1310             found_key->objectid > logical ||
1311             found_key->objectid + found_key->offset <= logical) {
1312                 pr_debug("logical %llu is not within any extent\n",
1313                          (unsigned long long)logical);
1314                 return -ENOENT;
1315         }
1316
1317         eb = path->nodes[0];
1318         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1319         BUG_ON(item_size < sizeof(*ei));
1320
1321         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1322         flags = btrfs_extent_flags(eb, ei);
1323
1324         pr_debug("logical %llu is at position %llu within the extent (%llu "
1325                  "EXTENT_ITEM %llu) flags %#llx size %u\n",
1326                  (unsigned long long)logical,
1327                  (unsigned long long)(logical - found_key->objectid),
1328                  (unsigned long long)found_key->objectid,
1329                  (unsigned long long)found_key->offset,
1330                  (unsigned long long)flags, item_size);
1331
1332         WARN_ON(!flags_ret);
1333         if (flags_ret) {
1334                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1335                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1336                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1337                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1338                 else
1339                         BUG_ON(1);
1340                 return 0;
1341         }
1342
1343         return -EIO;
1344 }
1345
1346 /*
1347  * helper function to iterate extent inline refs. ptr must point to a 0 value
1348  * for the first call and may be modified. it is used to track state.
1349  * if more refs exist, 0 is returned and the next call to
1350  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1351  * next ref. after the last ref was processed, 1 is returned.
1352  * returns <0 on error
1353  */
1354 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1355                                 struct btrfs_extent_item *ei, u32 item_size,
1356                                 struct btrfs_extent_inline_ref **out_eiref,
1357                                 int *out_type)
1358 {
1359         unsigned long end;
1360         u64 flags;
1361         struct btrfs_tree_block_info *info;
1362
1363         if (!*ptr) {
1364                 /* first call */
1365                 flags = btrfs_extent_flags(eb, ei);
1366                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1367                         info = (struct btrfs_tree_block_info *)(ei + 1);
1368                         *out_eiref =
1369                                 (struct btrfs_extent_inline_ref *)(info + 1);
1370                 } else {
1371                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1372                 }
1373                 *ptr = (unsigned long)*out_eiref;
1374                 if ((void *)*ptr >= (void *)ei + item_size)
1375                         return -ENOENT;
1376         }
1377
1378         end = (unsigned long)ei + item_size;
1379         *out_eiref = (struct btrfs_extent_inline_ref *)*ptr;
1380         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1381
1382         *ptr += btrfs_extent_inline_ref_size(*out_type);
1383         WARN_ON(*ptr > end);
1384         if (*ptr == end)
1385                 return 1; /* last */
1386
1387         return 0;
1388 }
1389
1390 /*
1391  * reads the tree block backref for an extent. tree level and root are returned
1392  * through out_level and out_root. ptr must point to a 0 value for the first
1393  * call and may be modified (see __get_extent_inline_ref comment).
1394  * returns 0 if data was provided, 1 if there was no more data to provide or
1395  * <0 on error.
1396  */
1397 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1398                                 struct btrfs_extent_item *ei, u32 item_size,
1399                                 u64 *out_root, u8 *out_level)
1400 {
1401         int ret;
1402         int type;
1403         struct btrfs_tree_block_info *info;
1404         struct btrfs_extent_inline_ref *eiref;
1405
1406         if (*ptr == (unsigned long)-1)
1407                 return 1;
1408
1409         while (1) {
1410                 ret = __get_extent_inline_ref(ptr, eb, ei, item_size,
1411                                                 &eiref, &type);
1412                 if (ret < 0)
1413                         return ret;
1414
1415                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1416                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1417                         break;
1418
1419                 if (ret == 1)
1420                         return 1;
1421         }
1422
1423         /* we can treat both ref types equally here */
1424         info = (struct btrfs_tree_block_info *)(ei + 1);
1425         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1426         *out_level = btrfs_tree_block_level(eb, info);
1427
1428         if (ret == 1)
1429                 *ptr = (unsigned long)-1;
1430
1431         return 0;
1432 }
1433
1434 static int iterate_leaf_refs(struct extent_inode_elem *inode_list,
1435                                 u64 root, u64 extent_item_objectid,
1436                                 iterate_extent_inodes_t *iterate, void *ctx)
1437 {
1438         struct extent_inode_elem *eie;
1439         int ret = 0;
1440
1441         for (eie = inode_list; eie; eie = eie->next) {
1442                 pr_debug("ref for %llu resolved, key (%llu EXTEND_DATA %llu), "
1443                          "root %llu\n", extent_item_objectid,
1444                          eie->inum, eie->offset, root);
1445                 ret = iterate(eie->inum, eie->offset, root, ctx);
1446                 if (ret) {
1447                         pr_debug("stopping iteration for %llu due to ret=%d\n",
1448                                  extent_item_objectid, ret);
1449                         break;
1450                 }
1451         }
1452
1453         return ret;
1454 }
1455
1456 /*
1457  * calls iterate() for every inode that references the extent identified by
1458  * the given parameters.
1459  * when the iterator function returns a non-zero value, iteration stops.
1460  */
1461 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1462                                 u64 extent_item_objectid, u64 extent_item_pos,
1463                                 int search_commit_root,
1464                                 iterate_extent_inodes_t *iterate, void *ctx)
1465 {
1466         int ret;
1467         struct list_head data_refs = LIST_HEAD_INIT(data_refs);
1468         struct list_head shared_refs = LIST_HEAD_INIT(shared_refs);
1469         struct btrfs_trans_handle *trans;
1470         struct ulist *refs = NULL;
1471         struct ulist *roots = NULL;
1472         struct ulist_node *ref_node = NULL;
1473         struct ulist_node *root_node = NULL;
1474         struct seq_list tree_mod_seq_elem = {};
1475         struct ulist_iterator ref_uiter;
1476         struct ulist_iterator root_uiter;
1477
1478         pr_debug("resolving all inodes for extent %llu\n",
1479                         extent_item_objectid);
1480
1481         if (search_commit_root) {
1482                 trans = BTRFS_BACKREF_SEARCH_COMMIT_ROOT;
1483         } else {
1484                 trans = btrfs_join_transaction(fs_info->extent_root);
1485                 if (IS_ERR(trans))
1486                         return PTR_ERR(trans);
1487                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1488         }
1489
1490         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1491                                    tree_mod_seq_elem.seq, &refs,
1492                                    &extent_item_pos);
1493         if (ret)
1494                 goto out;
1495
1496         ULIST_ITER_INIT(&ref_uiter);
1497         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1498                 ret = btrfs_find_all_roots(trans, fs_info, ref_node->val,
1499                                            tree_mod_seq_elem.seq, &roots);
1500                 if (ret)
1501                         break;
1502                 ULIST_ITER_INIT(&root_uiter);
1503                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1504                         pr_debug("root %llu references leaf %llu, data list "
1505                                  "%#llx\n", root_node->val, ref_node->val,
1506                                  (long long)ref_node->aux);
1507                         ret = iterate_leaf_refs((struct extent_inode_elem *)
1508                                                 (uintptr_t)ref_node->aux,
1509                                                 root_node->val,
1510                                                 extent_item_objectid,
1511                                                 iterate, ctx);
1512                 }
1513                 ulist_free(roots);
1514                 roots = NULL;
1515         }
1516
1517         free_leaf_list(refs);
1518         ulist_free(roots);
1519 out:
1520         if (!search_commit_root) {
1521                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1522                 btrfs_end_transaction(trans, fs_info->extent_root);
1523         }
1524
1525         return ret;
1526 }
1527
1528 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1529                                 struct btrfs_path *path,
1530                                 iterate_extent_inodes_t *iterate, void *ctx)
1531 {
1532         int ret;
1533         u64 extent_item_pos;
1534         u64 flags = 0;
1535         struct btrfs_key found_key;
1536         int search_commit_root = path->search_commit_root;
1537
1538         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1539         btrfs_release_path(path);
1540         if (ret < 0)
1541                 return ret;
1542         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1543                 return -EINVAL;
1544
1545         extent_item_pos = logical - found_key.objectid;
1546         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1547                                         extent_item_pos, search_commit_root,
1548                                         iterate, ctx);
1549
1550         return ret;
1551 }
1552
1553 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1554                               struct extent_buffer *eb, void *ctx);
1555
1556 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1557                               struct btrfs_path *path,
1558                               iterate_irefs_t *iterate, void *ctx)
1559 {
1560         int ret = 0;
1561         int slot;
1562         u32 cur;
1563         u32 len;
1564         u32 name_len;
1565         u64 parent = 0;
1566         int found = 0;
1567         struct extent_buffer *eb;
1568         struct btrfs_item *item;
1569         struct btrfs_inode_ref *iref;
1570         struct btrfs_key found_key;
1571
1572         while (!ret) {
1573                 path->leave_spinning = 1;
1574                 ret = inode_ref_info(inum, parent ? parent+1 : 0, fs_root, path,
1575                                      &found_key);
1576                 if (ret < 0)
1577                         break;
1578                 if (ret) {
1579                         ret = found ? 0 : -ENOENT;
1580                         break;
1581                 }
1582                 ++found;
1583
1584                 parent = found_key.offset;
1585                 slot = path->slots[0];
1586                 eb = path->nodes[0];
1587                 /* make sure we can use eb after releasing the path */
1588                 atomic_inc(&eb->refs);
1589                 btrfs_tree_read_lock(eb);
1590                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1591                 btrfs_release_path(path);
1592
1593                 item = btrfs_item_nr(eb, slot);
1594                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1595
1596                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
1597                         name_len = btrfs_inode_ref_name_len(eb, iref);
1598                         /* path must be released before calling iterate()! */
1599                         pr_debug("following ref at offset %u for inode %llu in "
1600                                  "tree %llu\n", cur,
1601                                  (unsigned long long)found_key.objectid,
1602                                  (unsigned long long)fs_root->objectid);
1603                         ret = iterate(parent, name_len,
1604                                       (unsigned long)(iref + 1), eb, ctx);
1605                         if (ret)
1606                                 break;
1607                         len = sizeof(*iref) + name_len;
1608                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
1609                 }
1610                 btrfs_tree_read_unlock_blocking(eb);
1611                 free_extent_buffer(eb);
1612         }
1613
1614         btrfs_release_path(path);
1615
1616         return ret;
1617 }
1618
1619 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
1620                                  struct btrfs_path *path,
1621                                  iterate_irefs_t *iterate, void *ctx)
1622 {
1623         int ret;
1624         int slot;
1625         u64 offset = 0;
1626         u64 parent;
1627         int found = 0;
1628         struct extent_buffer *eb;
1629         struct btrfs_inode_extref *extref;
1630         struct extent_buffer *leaf;
1631         u32 item_size;
1632         u32 cur_offset;
1633         unsigned long ptr;
1634
1635         while (1) {
1636                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
1637                                             &offset);
1638                 if (ret < 0)
1639                         break;
1640                 if (ret) {
1641                         ret = found ? 0 : -ENOENT;
1642                         break;
1643                 }
1644                 ++found;
1645
1646                 slot = path->slots[0];
1647                 eb = path->nodes[0];
1648                 /* make sure we can use eb after releasing the path */
1649                 atomic_inc(&eb->refs);
1650
1651                 btrfs_tree_read_lock(eb);
1652                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1653                 btrfs_release_path(path);
1654
1655                 leaf = path->nodes[0];
1656                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1657                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1658                 cur_offset = 0;
1659
1660                 while (cur_offset < item_size) {
1661                         u32 name_len;
1662
1663                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
1664                         parent = btrfs_inode_extref_parent(eb, extref);
1665                         name_len = btrfs_inode_extref_name_len(eb, extref);
1666                         ret = iterate(parent, name_len,
1667                                       (unsigned long)&extref->name, eb, ctx);
1668                         if (ret)
1669                                 break;
1670
1671                         cur_offset += btrfs_inode_extref_name_len(leaf, extref);
1672                         cur_offset += sizeof(*extref);
1673                 }
1674                 btrfs_tree_read_unlock_blocking(eb);
1675                 free_extent_buffer(eb);
1676
1677                 offset++;
1678         }
1679
1680         btrfs_release_path(path);
1681
1682         return ret;
1683 }
1684
1685 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
1686                          struct btrfs_path *path, iterate_irefs_t *iterate,
1687                          void *ctx)
1688 {
1689         int ret;
1690         int found_refs = 0;
1691
1692         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
1693         if (!ret)
1694                 ++found_refs;
1695         else if (ret != -ENOENT)
1696                 return ret;
1697
1698         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
1699         if (ret == -ENOENT && found_refs)
1700                 return 0;
1701
1702         return ret;
1703 }
1704
1705 /*
1706  * returns 0 if the path could be dumped (probably truncated)
1707  * returns <0 in case of an error
1708  */
1709 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
1710                          struct extent_buffer *eb, void *ctx)
1711 {
1712         struct inode_fs_paths *ipath = ctx;
1713         char *fspath;
1714         char *fspath_min;
1715         int i = ipath->fspath->elem_cnt;
1716         const int s_ptr = sizeof(char *);
1717         u32 bytes_left;
1718
1719         bytes_left = ipath->fspath->bytes_left > s_ptr ?
1720                                         ipath->fspath->bytes_left - s_ptr : 0;
1721
1722         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
1723         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
1724                                    name_off, eb, inum, fspath_min, bytes_left);
1725         if (IS_ERR(fspath))
1726                 return PTR_ERR(fspath);
1727
1728         if (fspath > fspath_min) {
1729                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
1730                 ++ipath->fspath->elem_cnt;
1731                 ipath->fspath->bytes_left = fspath - fspath_min;
1732         } else {
1733                 ++ipath->fspath->elem_missed;
1734                 ipath->fspath->bytes_missing += fspath_min - fspath;
1735                 ipath->fspath->bytes_left = 0;
1736         }
1737
1738         return 0;
1739 }
1740
1741 /*
1742  * this dumps all file system paths to the inode into the ipath struct, provided
1743  * is has been created large enough. each path is zero-terminated and accessed
1744  * from ipath->fspath->val[i].
1745  * when it returns, there are ipath->fspath->elem_cnt number of paths available
1746  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
1747  * number of missed paths in recored in ipath->fspath->elem_missed, otherwise,
1748  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
1749  * have been needed to return all paths.
1750  */
1751 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
1752 {
1753         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
1754                              inode_to_path, ipath);
1755 }
1756
1757 struct btrfs_data_container *init_data_container(u32 total_bytes)
1758 {
1759         struct btrfs_data_container *data;
1760         size_t alloc_bytes;
1761
1762         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
1763         data = vmalloc(alloc_bytes);
1764         if (!data)
1765                 return ERR_PTR(-ENOMEM);
1766
1767         if (total_bytes >= sizeof(*data)) {
1768                 data->bytes_left = total_bytes - sizeof(*data);
1769                 data->bytes_missing = 0;
1770         } else {
1771                 data->bytes_missing = sizeof(*data) - total_bytes;
1772                 data->bytes_left = 0;
1773         }
1774
1775         data->elem_cnt = 0;
1776         data->elem_missed = 0;
1777
1778         return data;
1779 }
1780
1781 /*
1782  * allocates space to return multiple file system paths for an inode.
1783  * total_bytes to allocate are passed, note that space usable for actual path
1784  * information will be total_bytes - sizeof(struct inode_fs_paths).
1785  * the returned pointer must be freed with free_ipath() in the end.
1786  */
1787 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
1788                                         struct btrfs_path *path)
1789 {
1790         struct inode_fs_paths *ifp;
1791         struct btrfs_data_container *fspath;
1792
1793         fspath = init_data_container(total_bytes);
1794         if (IS_ERR(fspath))
1795                 return (void *)fspath;
1796
1797         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
1798         if (!ifp) {
1799                 kfree(fspath);
1800                 return ERR_PTR(-ENOMEM);
1801         }
1802
1803         ifp->btrfs_path = path;
1804         ifp->fspath = fspath;
1805         ifp->fs_root = fs_root;
1806
1807         return ifp;
1808 }
1809
1810 void free_ipath(struct inode_fs_paths *ipath)
1811 {
1812         if (!ipath)
1813                 return;
1814         vfree(ipath->fspath);
1815         kfree(ipath);
1816 }