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