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
 465                         ref2 = list_entry(pos2, struct __prelim_ref, list);
 466
 467                         if (mode == 1) {
 468                                 if (!ref_for_same_block(ref1, ref2))
 469                                         continue;
 470                                 if (!ref1->parent && ref2->parent) {
 471                                         xchg = ref1;
 472                                         ref1 = ref2;
 473                                         ref2 = xchg;
 474                                 }
 475                                 ref1->count += ref2->count;
 476                         } else {
 477                                 if (ref1->parent != ref2->parent)
 478                                         continue;
 479                                 ref1->count += ref2->count;
 480                         }
 481                         list_del(&ref2->list);
 482                         kfree(ref2);
 483                 }
 484
 485         }
 486         return 0;
 487 }
 488
 489 /*
 490  * add all currently queued delayed refs from this head whose seq nr is
 491  * smaller or equal that seq to the list
 492  */
 493 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 494                               struct list_head *prefs)
 495 {
 496         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 497         struct rb_node *n = &head->node.rb_node;
 498         struct btrfs_key key;
 499         struct btrfs_key op_key = {0};
 500         int sgn;
 501         int ret = 0;
 502
 503         if (extent_op && extent_op->update_key)
 504                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 505
 506         while ((n = rb_prev(n))) {
 507                 struct btrfs_delayed_ref_node *node;
 508                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 509                                 rb_node);
 510                 if (node->bytenr != head->node.bytenr)
 511                         break;
 512                 WARN_ON(node->is_head);
 513
 514                 if (node->seq > seq)
 515                         continue;
 516
 517                 switch (node->action) {
 518                 case BTRFS_ADD_DELAYED_EXTENT:
 519                 case BTRFS_UPDATE_DELAYED_HEAD:
 520                         WARN_ON(1);
 521                         continue;
 522                 case BTRFS_ADD_DELAYED_REF:
 523                         sgn = 1;
 524                         break;
 525                 case BTRFS_DROP_DELAYED_REF:
 526                         sgn = -1;
 527                         break;
 528                 default:
 529                         BUG_ON(1);
 530                 }
 531                 switch (node->type) {
 532                 case BTRFS_TREE_BLOCK_REF_KEY: {
 533                         struct btrfs_delayed_tree_ref *ref;
 534
 535                         ref = btrfs_delayed_node_to_tree_ref(node);
 536                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
 537                                                ref->level + 1, 0, node->bytenr,
 538                                                node->ref_mod * sgn);
 539                         break;
 540                 }
 541                 case BTRFS_SHARED_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, NULL,
 546                                                ref->level + 1, ref->parent,
 547                                                node->bytenr,
 548                                                node->ref_mod * sgn);
 549                         break;
 550                 }
 551                 case BTRFS_EXTENT_DATA_REF_KEY: {
 552                         struct btrfs_delayed_data_ref *ref;
 553                         ref = btrfs_delayed_node_to_data_ref(node);
 554
 555                         key.objectid = ref->objectid;
 556                         key.type = BTRFS_EXTENT_DATA_KEY;
 557                         key.offset = ref->offset;
 558                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
 559                                                node->bytenr,
 560                                                node->ref_mod * sgn);
 561                         break;
 562                 }
 563                 case BTRFS_SHARED_DATA_REF_KEY: {
 564                         struct btrfs_delayed_data_ref *ref;
 565
 566                         ref = btrfs_delayed_node_to_data_ref(node);
 567
 568                         key.objectid = ref->objectid;
 569                         key.type = BTRFS_EXTENT_DATA_KEY;
 570                         key.offset = ref->offset;
 571                         ret = __add_prelim_ref(prefs, ref->root, &key, 0,
 572                                                ref->parent, node->bytenr,
 573                                                node->ref_mod * sgn);
 574                         break;
 575                 }
 576                 default:
 577                         WARN_ON(1);
 578                 }
 579                 BUG_ON(ret);
 580         }
 581
 582         return 0;
 583 }
 584
 585 /*
 586  * add all inline backrefs for bytenr to the list
 587  */
 588 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
 589                              struct btrfs_path *path, u64 bytenr,
 590                              int *info_level, struct list_head *prefs)
 591 {
 592         int ret = 0;
 593         int slot;
 594         struct extent_buffer *leaf;
 595         struct btrfs_key key;
 596         unsigned long ptr;
 597         unsigned long end;
 598         struct btrfs_extent_item *ei;
 599         u64 flags;
 600         u64 item_size;
 601
 602         /*
 603          * enumerate all inline refs
 604          */
 605         leaf = path->nodes[0];
 606         slot = path->slots[0];
 607
 608         item_size = btrfs_item_size_nr(leaf, slot);
 609         BUG_ON(item_size < sizeof(*ei));
 610
 611         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 612         flags = btrfs_extent_flags(leaf, ei);
 613
 614         ptr = (unsigned long)(ei + 1);
 615         end = (unsigned long)ei + item_size;
 616
 617         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 618                 struct btrfs_tree_block_info *info;
 619
 620                 info = (struct btrfs_tree_block_info *)ptr;
 621                 *info_level = btrfs_tree_block_level(leaf, info);
 622                 ptr += sizeof(struct btrfs_tree_block_info);
 623                 BUG_ON(ptr > end);
 624         } else {
 625                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 626         }
 627
 628         while (ptr < end) {
 629                 struct btrfs_extent_inline_ref *iref;
 630                 u64 offset;
 631                 int type;
 632
 633                 iref = (struct btrfs_extent_inline_ref *)ptr;
 634                 type = btrfs_extent_inline_ref_type(leaf, iref);
 635                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 636
 637                 switch (type) {
 638                 case BTRFS_SHARED_BLOCK_REF_KEY:
 639                         ret = __add_prelim_ref(prefs, 0, NULL,
 640                                                 *info_level + 1, offset,
 641                                                 bytenr, 1);
 642                         break;
 643                 case BTRFS_SHARED_DATA_REF_KEY: {
 644                         struct btrfs_shared_data_ref *sdref;
 645                         int count;
 646
 647                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 648                         count = btrfs_shared_data_ref_count(leaf, sdref);
 649                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
 650                                                bytenr, count);
 651                         break;
 652                 }
 653                 case BTRFS_TREE_BLOCK_REF_KEY:
 654                         ret = __add_prelim_ref(prefs, offset, NULL,
 655                                                *info_level + 1, 0,
 656                                                bytenr, 1);
 657                         break;
 658                 case BTRFS_EXTENT_DATA_REF_KEY: {
 659                         struct btrfs_extent_data_ref *dref;
 660                         int count;
 661                         u64 root;
 662
 663                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 664                         count = btrfs_extent_data_ref_count(leaf, dref);
 665                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 666                                                                       dref);
 667                         key.type = BTRFS_EXTENT_DATA_KEY;
 668                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 669                         root = btrfs_extent_data_ref_root(leaf, dref);
 670                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 671                                                bytenr, count);
 672                         break;
 673                 }
 674                 default:
 675                         WARN_ON(1);
 676                 }
 677                 BUG_ON(ret);
 678                 ptr += btrfs_extent_inline_ref_size(type);
 679         }
 680
 681         return 0;
 682 }
 683
 684 /*
 685  * add all non-inline backrefs for bytenr to the list
 686  */
 687 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
 688                             struct btrfs_path *path, u64 bytenr,
 689                             int info_level, struct list_head *prefs)
 690 {
 691         struct btrfs_root *extent_root = fs_info->extent_root;
 692         int ret;
 693         int slot;
 694         struct extent_buffer *leaf;
 695         struct btrfs_key key;
 696
 697         while (1) {
 698                 ret = btrfs_next_item(extent_root, path);
 699                 if (ret < 0)
 700                         break;
 701                 if (ret) {
 702                         ret = 0;
 703                         break;
 704                 }
 705
 706                 slot = path->slots[0];
 707                 leaf = path->nodes[0];
 708                 btrfs_item_key_to_cpu(leaf, &key, slot);
 709
 710                 if (key.objectid != bytenr)
 711                         break;
 712                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
 713                         continue;
 714                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
 715                         break;
 716
 717                 switch (key.type) {
 718                 case BTRFS_SHARED_BLOCK_REF_KEY:
 719                         ret = __add_prelim_ref(prefs, 0, NULL,
 720                                                 info_level + 1, key.offset,
 721                                                 bytenr, 1);
 722                         break;
 723                 case BTRFS_SHARED_DATA_REF_KEY: {
 724                         struct btrfs_shared_data_ref *sdref;
 725                         int count;
 726
 727                         sdref = btrfs_item_ptr(leaf, slot,
 728                                               struct btrfs_shared_data_ref);
 729                         count = btrfs_shared_data_ref_count(leaf, sdref);
 730                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
 731                                                 bytenr, count);
 732                         break;
 733                 }
 734                 case BTRFS_TREE_BLOCK_REF_KEY:
 735                         ret = __add_prelim_ref(prefs, key.offset, NULL,
 736                                                info_level + 1, 0,
 737                                                bytenr, 1);
 738                         break;
 739                 case BTRFS_EXTENT_DATA_REF_KEY: {
 740                         struct btrfs_extent_data_ref *dref;
 741                         int count;
 742                         u64 root;
 743
 744                         dref = btrfs_item_ptr(leaf, slot,
 745                                               struct btrfs_extent_data_ref);
 746                         count = btrfs_extent_data_ref_count(leaf, dref);
 747                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 748                                                                       dref);
 749                         key.type = BTRFS_EXTENT_DATA_KEY;
 750                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 751                         root = btrfs_extent_data_ref_root(leaf, dref);
 752                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 753                                                bytenr, count);
 754                         break;
 755                 }
 756                 default:
 757                         WARN_ON(1);
 758                 }
 759                 BUG_ON(ret);
 760         }
 761
 762         return ret;
 763 }
 764
 765 /*
 766  * this adds all existing backrefs (inline backrefs, backrefs and delayed
 767  * refs) for the given bytenr to the refs list, merges duplicates and resolves
 768  * indirect refs to their parent bytenr.
 769  * When roots are found, they're added to the roots list
 770  *
 771  * FIXME some caching might speed things up
 772  */
 773 static int find_parent_nodes(struct btrfs_trans_handle *trans,
 774                              struct btrfs_fs_info *fs_info, u64 bytenr,
 775                              u64 time_seq, struct ulist *refs,
 776                              struct ulist *roots, const u64 *extent_item_pos)
 777 {
 778         struct btrfs_key key;
 779         struct btrfs_path *path;
 780         struct btrfs_delayed_ref_root *delayed_refs = NULL;
 781         struct btrfs_delayed_ref_head *head;
 782         int info_level = 0;
 783         int ret;
 784         int search_commit_root = (trans == BTRFS_BACKREF_SEARCH_COMMIT_ROOT);
 785         struct list_head prefs_delayed;
 786         struct list_head prefs;
 787         struct __prelim_ref *ref;
 788
 789         INIT_LIST_HEAD(&prefs);
 790         INIT_LIST_HEAD(&prefs_delayed);
 791
 792         key.objectid = bytenr;
 793         key.type = BTRFS_EXTENT_ITEM_KEY;
 794         key.offset = (u64)-1;
 795
 796         path = btrfs_alloc_path();
 797         if (!path)
 798                 return -ENOMEM;
 799         path->search_commit_root = !!search_commit_root;
 800
 801         /*
 802          * grab both a lock on the path and a lock on the delayed ref head.
 803          * We need both to get a consistent picture of how the refs look
 804          * at a specified point in time
 805          */
 806 again:
 807         head = NULL;
 808
 809         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 810         if (ret < 0)
 811                 goto out;
 812         BUG_ON(ret == 0);
 813
 814         if (trans != BTRFS_BACKREF_SEARCH_COMMIT_ROOT) {
 815                 /*
 816                  * look if there are updates for this ref queued and lock the
 817                  * head
 818                  */
 819                 delayed_refs = &trans->transaction->delayed_refs;
 820                 spin_lock(&delayed_refs->lock);
 821                 head = btrfs_find_delayed_ref_head(trans, bytenr);
 822                 if (head) {
 823                         if (!mutex_trylock(&head->mutex)) {
 824                                 atomic_inc(&head->node.refs);
 825                                 spin_unlock(&delayed_refs->lock);
 826
 827                                 btrfs_release_path(path);
 828
 829                                 /*
 830                                  * Mutex was contended, block until it's
 831                                  * released and try again
 832                                  */
 833                                 mutex_lock(&head->mutex);
 834                                 mutex_unlock(&head->mutex);
 835                                 btrfs_put_delayed_ref(&head->node);
 836                                 goto again;
 837                         }
 838                         ret = __add_delayed_refs(head, time_seq,
 839                                                  &prefs_delayed);
 840                         mutex_unlock(&head->mutex);
 841                         if (ret) {
 842                                 spin_unlock(&delayed_refs->lock);
 843                                 goto out;
 844                         }
 845                 }
 846                 spin_unlock(&delayed_refs->lock);
 847         }
 848
 849         if (path->slots[0]) {
 850                 struct extent_buffer *leaf;
 851                 int slot;
 852
 853                 path->slots[0]--;
 854                 leaf = path->nodes[0];
 855                 slot = path->slots[0];
 856                 btrfs_item_key_to_cpu(leaf, &key, slot);
 857                 if (key.objectid == bytenr &&
 858                     key.type == BTRFS_EXTENT_ITEM_KEY) {
 859                         ret = __add_inline_refs(fs_info, path, bytenr,
 860                                                 &info_level, &prefs);
 861                         if (ret)
 862                                 goto out;
 863                         ret = __add_keyed_refs(fs_info, path, bytenr,
 864                                                info_level, &prefs);
 865                         if (ret)
 866                                 goto out;
 867                 }
 868         }
 869         btrfs_release_path(path);
 870
 871         list_splice_init(&prefs_delayed, &prefs);
 872
 873         ret = __add_missing_keys(fs_info, &prefs);
 874         if (ret)
 875                 goto out;
 876
 877         ret = __merge_refs(&prefs, 1);
 878         if (ret)
 879                 goto out;
 880
 881         ret = __resolve_indirect_refs(fs_info, search_commit_root, time_seq,
 882                                       &prefs, extent_item_pos);
 883         if (ret)
 884                 goto out;
 885
 886         ret = __merge_refs(&prefs, 2);
 887         if (ret)
 888                 goto out;
 889
 890         while (!list_empty(&prefs)) {
 891                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
 892                 list_del(&ref->list);
 893                 WARN_ON(ref->count < 0);
 894                 if (ref->count && ref->root_id && ref->parent == 0) {
 895                         /* no parent == root of tree */
 896                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
 897                         BUG_ON(ret < 0);
 898                 }
 899                 if (ref->count && ref->parent) {
 900                         struct extent_inode_elem *eie = NULL;
 901                         if (extent_item_pos && !ref->inode_list) {
 902                                 u32 bsz;
 903                                 struct extent_buffer *eb;
 904                                 bsz = btrfs_level_size(fs_info->extent_root,
 905                                                         info_level);
 906                                 eb = read_tree_block(fs_info->extent_root,
 907                                                            ref->parent, bsz, 0);
 908                                 BUG_ON(!eb);
 909                                 ret = find_extent_in_eb(eb, bytenr,
 910                                                         *extent_item_pos, &eie);
 911                                 ref->inode_list = eie;
 912                                 free_extent_buffer(eb);
 913                         }
 914                         ret = ulist_add_merge(refs, ref->parent,
 915                                               (uintptr_t)ref->inode_list,
 916                                               (u64 *)&eie, GFP_NOFS);
 917                         if (!ret && extent_item_pos) {
 918                                 /*
 919                                  * we've recorded that parent, so we must extend
 920                                  * its inode list here
 921                                  */
 922                                 BUG_ON(!eie);
 923                                 while (eie->next)
 924                                         eie = eie->next;
 925                                 eie->next = ref->inode_list;
 926                         }
 927                         BUG_ON(ret < 0);
 928                 }
 929                 kfree(ref);
 930         }
 931
 932 out:
 933         btrfs_free_path(path);
 934         while (!list_empty(&prefs)) {
 935                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
 936                 list_del(&ref->list);
 937                 kfree(ref);
 938         }
 939         while (!list_empty(&prefs_delayed)) {
 940                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
 941                                        list);
 942                 list_del(&ref->list);
 943                 kfree(ref);
 944         }
 945
 946         return ret;
 947 }
 948
 949 static void free_leaf_list(struct ulist *blocks)
 950 {
 951         struct ulist_node *node = NULL;
 952         struct extent_inode_elem *eie;
 953         struct extent_inode_elem *eie_next;
 954         struct ulist_iterator uiter;
 955
 956         ULIST_ITER_INIT(&uiter);
 957         while ((node = ulist_next(blocks, &uiter))) {
 958                 if (!node->aux)
 959                         continue;
 960                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
 961                 for (; eie; eie = eie_next) {
 962                         eie_next = eie->next;
 963                         kfree(eie);
 964                 }
 965                 node->aux = 0;
 966         }
 967
 968         ulist_free(blocks);
 969 }
 970
 971 /*
 972  * Finds all leafs with a reference to the specified combination of bytenr and
 973  * offset. key_list_head will point to a list of corresponding keys (caller must
 974  * free each list element). The leafs will be stored in the leafs ulist, which
 975  * must be freed with ulist_free.
 976  *
 977  * returns 0 on success, <0 on error
 978  */
 979 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
 980                                 struct btrfs_fs_info *fs_info, u64 bytenr,
 981                                 u64 time_seq, struct ulist **leafs,
 982                                 const u64 *extent_item_pos)
 983 {
 984         struct ulist *tmp;
 985         int ret;
 986
 987         tmp = ulist_alloc(GFP_NOFS);
 988         if (!tmp)
 989                 return -ENOMEM;
 990         *leafs = ulist_alloc(GFP_NOFS);
 991         if (!*leafs) {
 992                 ulist_free(tmp);
 993                 return -ENOMEM;
 994         }
 995
 996         ret = find_parent_nodes(trans, fs_info, bytenr,
 997                                 time_seq, *leafs, tmp, extent_item_pos);
 998         ulist_free(tmp);
 999
1000         if (ret < 0 && ret != -ENOENT) {
1001                 free_leaf_list(*leafs);
1002                 return ret;
1003         }
1004
1005         return 0;
1006 }
1007
1008 /*
1009  * walk all backrefs for a given extent to find all roots that reference this
1010  * extent. Walking a backref means finding all extents that reference this
1011  * extent and in turn walk the backrefs of those, too. Naturally this is a
1012  * recursive process, but here it is implemented in an iterative fashion: We
1013  * find all referencing extents for the extent in question and put them on a
1014  * list. In turn, we find all referencing extents for those, further appending
1015  * to the list. The way we iterate the list allows adding more elements after
1016  * the current while iterating. The process stops when we reach the end of the
1017  * list. Found roots are added to the roots list.
1018  *
1019  * returns 0 on success, < 0 on error.
1020  */
1021 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1022                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1023                                 u64 time_seq, struct ulist **roots)
1024 {
1025         struct ulist *tmp;
1026         struct ulist_node *node = NULL;
1027         struct ulist_iterator uiter;
1028         int ret;
1029
1030         tmp = ulist_alloc(GFP_NOFS);
1031         if (!tmp)
1032                 return -ENOMEM;
1033         *roots = ulist_alloc(GFP_NOFS);
1034         if (!*roots) {
1035                 ulist_free(tmp);
1036                 return -ENOMEM;
1037         }
1038
1039         ULIST_ITER_INIT(&uiter);
1040         while (1) {
1041                 ret = find_parent_nodes(trans, fs_info, bytenr,
1042                                         time_seq, tmp, *roots, NULL);
1043                 if (ret < 0 && ret != -ENOENT) {
1044                         ulist_free(tmp);
1045                         ulist_free(*roots);
1046                         return ret;
1047                 }
1048                 node = ulist_next(tmp, &uiter);
1049                 if (!node)
1050                         break;
1051                 bytenr = node->val;
1052         }
1053
1054         ulist_free(tmp);
1055         return 0;
1056 }
1057
1058
1059 static int __inode_info(u64 inum, u64 ioff, u8 key_type,
1060                         struct btrfs_root *fs_root, struct btrfs_path *path,
1061                         struct btrfs_key *found_key)
1062 {
1063         int ret;
1064         struct btrfs_key key;
1065         struct extent_buffer *eb;
1066
1067         key.type = key_type;
1068         key.objectid = inum;
1069         key.offset = ioff;
1070
1071         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
1072         if (ret < 0)
1073                 return ret;
1074
1075         eb = path->nodes[0];
1076         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
1077                 ret = btrfs_next_leaf(fs_root, path);
1078                 if (ret)
1079                         return ret;
1080                 eb = path->nodes[0];
1081         }
1082
1083         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
1084         if (found_key->type != key.type || found_key->objectid != key.objectid)
1085                 return 1;
1086
1087         return 0;
1088 }
1089
1090 /*
1091  * this makes the path point to (inum INODE_ITEM ioff)
1092  */
1093 int inode_item_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1094                         struct btrfs_path *path)
1095 {
1096         struct btrfs_key key;
1097         return __inode_info(inum, ioff, BTRFS_INODE_ITEM_KEY, fs_root, path,
1098                                 &key);
1099 }
1100
1101 static int inode_ref_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1102                                 struct btrfs_path *path,
1103                                 struct btrfs_key *found_key)
1104 {
1105         return __inode_info(inum, ioff, BTRFS_INODE_REF_KEY, fs_root, path,
1106                                 found_key);
1107 }
1108
1109 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1110                           u64 start_off, struct btrfs_path *path,
1111                           struct btrfs_inode_extref **ret_extref,
1112                           u64 *found_off)
1113 {
1114         int ret, slot;
1115         struct btrfs_key key;
1116         struct btrfs_key found_key;
1117         struct btrfs_inode_extref *extref;
1118         struct extent_buffer *leaf;
1119         unsigned long ptr;
1120
1121         key.objectid = inode_objectid;
1122         btrfs_set_key_type(&key, BTRFS_INODE_EXTREF_KEY);
1123         key.offset = start_off;
1124
1125         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1126         if (ret < 0)
1127                 return ret;
1128
1129         while (1) {
1130                 leaf = path->nodes[0];
1131                 slot = path->slots[0];
1132                 if (slot >= btrfs_header_nritems(leaf)) {
1133                         /*
1134                          * If the item at offset is not found,
1135                          * btrfs_search_slot will point us to the slot
1136                          * where it should be inserted. In our case
1137                          * that will be the slot directly before the
1138                          * next INODE_REF_KEY_V2 item. In the case
1139                          * that we're pointing to the last slot in a
1140                          * leaf, we must move one leaf over.
1141                          */
1142                         ret = btrfs_next_leaf(root, path);
1143                         if (ret) {
1144                                 if (ret >= 1)
1145                                         ret = -ENOENT;
1146                                 break;
1147                         }
1148                         continue;
1149                 }
1150
1151                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1152
1153                 /*
1154                  * Check that we're still looking at an extended ref key for
1155                  * this particular objectid. If we have different
1156                  * objectid or type then there are no more to be found
1157                  * in the tree and we can exit.
1158                  */
1159                 ret = -ENOENT;
1160                 if (found_key.objectid != inode_objectid)
1161                         break;
1162                 if (btrfs_key_type(&found_key) != BTRFS_INODE_EXTREF_KEY)
1163                         break;
1164
1165                 ret = 0;
1166                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1167                 extref = (struct btrfs_inode_extref *)ptr;
1168                 *ret_extref = extref;
1169                 if (found_off)
1170                         *found_off = found_key.offset;
1171                 break;
1172         }
1173
1174         return ret;
1175 }
1176
1177 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1178                         u32 name_len, unsigned long name_off,
1179                         struct extent_buffer *eb_in, u64 parent,
1180                         char *dest, u32 size)
1181 {
1182         int slot;
1183         u64 next_inum;
1184         int ret;
1185         s64 bytes_left = ((s64)size) - 1;
1186         struct extent_buffer *eb = eb_in;
1187         struct btrfs_key found_key;
1188         int leave_spinning = path->leave_spinning;
1189         struct btrfs_inode_ref *iref;
1190
1191         if (bytes_left >= 0)
1192                 dest[bytes_left] = '\0';
1193
1194         path->leave_spinning = 1;
1195         while (1) {
1196                 bytes_left -= name_len;
1197                 if (bytes_left >= 0)
1198                         read_extent_buffer(eb, dest + bytes_left,
1199                                            name_off, name_len);
1200                 if (eb != eb_in) {
1201                         btrfs_tree_read_unlock_blocking(eb);
1202                         free_extent_buffer(eb);
1203                 }
1204                 ret = inode_ref_info(parent, 0, fs_root, path, &found_key);
1205                 if (ret > 0)
1206                         ret = -ENOENT;
1207                 if (ret)
1208                         break;
1209
1210                 next_inum = found_key.offset;
1211
1212                 /* regular exit ahead */
1213                 if (parent == next_inum)
1214                         break;
1215
1216                 slot = path->slots[0];
1217                 eb = path->nodes[0];
1218                 /* make sure we can use eb after releasing the path */
1219                 if (eb != eb_in) {
1220                         atomic_inc(&eb->refs);
1221                         btrfs_tree_read_lock(eb);
1222                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1223                 }
1224                 btrfs_release_path(path);
1225                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1226
1227                 name_len = btrfs_inode_ref_name_len(eb, iref);
1228                 name_off = (unsigned long)(iref + 1);
1229
1230                 parent = next_inum;
1231                 --bytes_left;
1232                 if (bytes_left >= 0)
1233                         dest[bytes_left] = '/';
1234         }
1235
1236         btrfs_release_path(path);
1237         path->leave_spinning = leave_spinning;
1238
1239         if (ret)
1240                 return ERR_PTR(ret);
1241
1242         return dest + bytes_left;
1243 }
1244
1245 /*
1246  * this iterates to turn a btrfs_inode_ref into a full filesystem path. elements
1247  * of the path are separated by '/' and the path is guaranteed to be
1248  * 0-terminated. the path is only given within the current file system.
1249  * Therefore, it never starts with a '/'. the caller is responsible to provide
1250  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1251  * the start point of the resulting string is returned. this pointer is within
1252  * dest, normally.
1253  * in case the path buffer would overflow, the pointer is decremented further
1254  * as if output was written to the buffer, though no more output is actually
1255  * generated. that way, the caller can determine how much space would be
1256  * required for the path to fit into the buffer. in that case, the returned
1257  * value will be smaller than dest. callers must check this!
1258  */
1259 char *btrfs_iref_to_path(struct btrfs_root *fs_root,
1260                          struct btrfs_path *path,
1261                          struct btrfs_inode_ref *iref,
1262                          struct extent_buffer *eb_in, u64 parent,
1263                          char *dest, u32 size)
1264 {
1265         return btrfs_ref_to_path(fs_root, path,
1266                                  btrfs_inode_ref_name_len(eb_in, iref),
1267                                  (unsigned long)(iref + 1),
1268                                  eb_in, parent, dest, size);
1269 }
1270
1271 /*
1272  * this makes the path point to (logical EXTENT_ITEM *)
1273  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1274  * tree blocks and <0 on error.
1275  */
1276 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1277                         struct btrfs_path *path, struct btrfs_key *found_key,
1278                         u64 *flags_ret)
1279 {
1280         int ret;
1281         u64 flags;
1282         u32 item_size;
1283         struct extent_buffer *eb;
1284         struct btrfs_extent_item *ei;
1285         struct btrfs_key key;
1286
1287         key.type = BTRFS_EXTENT_ITEM_KEY;
1288         key.objectid = logical;
1289         key.offset = (u64)-1;
1290
1291         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1292         if (ret < 0)
1293                 return ret;
1294         ret = btrfs_previous_item(fs_info->extent_root, path,
1295                                         0, BTRFS_EXTENT_ITEM_KEY);
1296         if (ret < 0)
1297                 return ret;
1298
1299         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1300         if (found_key->type != BTRFS_EXTENT_ITEM_KEY ||
1301             found_key->objectid > logical ||
1302             found_key->objectid + found_key->offset <= logical) {
1303                 pr_debug("logical %llu is not within any extent\n",
1304                          (unsigned long long)logical);
1305                 return -ENOENT;
1306         }
1307
1308         eb = path->nodes[0];
1309         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1310         BUG_ON(item_size < sizeof(*ei));
1311
1312         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1313         flags = btrfs_extent_flags(eb, ei);
1314
1315         pr_debug("logical %llu is at position %llu within the extent (%llu "
1316                  "EXTENT_ITEM %llu) flags %#llx size %u\n",
1317                  (unsigned long long)logical,
1318                  (unsigned long long)(logical - found_key->objectid),
1319                  (unsigned long long)found_key->objectid,
1320                  (unsigned long long)found_key->offset,
1321                  (unsigned long long)flags, item_size);
1322
1323         WARN_ON(!flags_ret);
1324         if (flags_ret) {
1325                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1326                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1327                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1328                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1329                 else
1330                         BUG_ON(1);
1331                 return 0;
1332         }
1333
1334         return -EIO;
1335 }
1336
1337 /*
1338  * helper function to iterate extent inline refs. ptr must point to a 0 value
1339  * for the first call and may be modified. it is used to track state.
1340  * if more refs exist, 0 is returned and the next call to
1341  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1342  * next ref. after the last ref was processed, 1 is returned.
1343  * returns <0 on error
1344  */
1345 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1346                                 struct btrfs_extent_item *ei, u32 item_size,
1347                                 struct btrfs_extent_inline_ref **out_eiref,
1348                                 int *out_type)
1349 {
1350         unsigned long end;
1351         u64 flags;
1352         struct btrfs_tree_block_info *info;
1353
1354         if (!*ptr) {
1355                 /* first call */
1356                 flags = btrfs_extent_flags(eb, ei);
1357                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1358                         info = (struct btrfs_tree_block_info *)(ei + 1);
1359                         *out_eiref =
1360                                 (struct btrfs_extent_inline_ref *)(info + 1);
1361                 } else {
1362                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1363                 }
1364                 *ptr = (unsigned long)*out_eiref;
1365                 if ((void *)*ptr >= (void *)ei + item_size)
1366                         return -ENOENT;
1367         }
1368
1369         end = (unsigned long)ei + item_size;
1370         *out_eiref = (struct btrfs_extent_inline_ref *)*ptr;
1371         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1372
1373         *ptr += btrfs_extent_inline_ref_size(*out_type);
1374         WARN_ON(*ptr > end);
1375         if (*ptr == end)
1376                 return 1; /* last */
1377
1378         return 0;
1379 }
1380
1381 /*
1382  * reads the tree block backref for an extent. tree level and root are returned
1383  * through out_level and out_root. ptr must point to a 0 value for the first
1384  * call and may be modified (see __get_extent_inline_ref comment).
1385  * returns 0 if data was provided, 1 if there was no more data to provide or
1386  * <0 on error.
1387  */
1388 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1389                                 struct btrfs_extent_item *ei, u32 item_size,
1390                                 u64 *out_root, u8 *out_level)
1391 {
1392         int ret;
1393         int type;
1394         struct btrfs_tree_block_info *info;
1395         struct btrfs_extent_inline_ref *eiref;
1396
1397         if (*ptr == (unsigned long)-1)
1398                 return 1;
1399
1400         while (1) {
1401                 ret = __get_extent_inline_ref(ptr, eb, ei, item_size,
1402                                                 &eiref, &type);
1403                 if (ret < 0)
1404                         return ret;
1405
1406                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1407                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1408                         break;
1409
1410                 if (ret == 1)
1411                         return 1;
1412         }
1413
1414         /* we can treat both ref types equally here */
1415         info = (struct btrfs_tree_block_info *)(ei + 1);
1416         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1417         *out_level = btrfs_tree_block_level(eb, info);
1418
1419         if (ret == 1)
1420                 *ptr = (unsigned long)-1;
1421
1422         return 0;
1423 }
1424
1425 static int iterate_leaf_refs(struct extent_inode_elem *inode_list,
1426                                 u64 root, u64 extent_item_objectid,
1427                                 iterate_extent_inodes_t *iterate, void *ctx)
1428 {
1429         struct extent_inode_elem *eie;
1430         int ret = 0;
1431
1432         for (eie = inode_list; eie; eie = eie->next) {
1433                 pr_debug("ref for %llu resolved, key (%llu EXTEND_DATA %llu), "
1434                          "root %llu\n", extent_item_objectid,
1435                          eie->inum, eie->offset, root);
1436                 ret = iterate(eie->inum, eie->offset, root, ctx);
1437                 if (ret) {
1438                         pr_debug("stopping iteration for %llu due to ret=%d\n",
1439                                  extent_item_objectid, ret);
1440                         break;
1441                 }
1442         }
1443
1444         return ret;
1445 }
1446
1447 /*
1448  * calls iterate() for every inode that references the extent identified by
1449  * the given parameters.
1450  * when the iterator function returns a non-zero value, iteration stops.
1451  */
1452 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1453                                 u64 extent_item_objectid, u64 extent_item_pos,
1454                                 int search_commit_root,
1455                                 iterate_extent_inodes_t *iterate, void *ctx)
1456 {
1457         int ret;
1458         struct list_head data_refs = LIST_HEAD_INIT(data_refs);
1459         struct list_head shared_refs = LIST_HEAD_INIT(shared_refs);
1460         struct btrfs_trans_handle *trans;
1461         struct ulist *refs = NULL;
1462         struct ulist *roots = NULL;
1463         struct ulist_node *ref_node = NULL;
1464         struct ulist_node *root_node = NULL;
1465         struct seq_list tree_mod_seq_elem = {};
1466         struct ulist_iterator ref_uiter;
1467         struct ulist_iterator root_uiter;
1468
1469         pr_debug("resolving all inodes for extent %llu\n",
1470                         extent_item_objectid);
1471
1472         if (search_commit_root) {
1473                 trans = BTRFS_BACKREF_SEARCH_COMMIT_ROOT;
1474         } else {
1475                 trans = btrfs_join_transaction(fs_info->extent_root);
1476                 if (IS_ERR(trans))
1477                         return PTR_ERR(trans);
1478                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1479         }
1480
1481         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1482                                    tree_mod_seq_elem.seq, &refs,
1483                                    &extent_item_pos);
1484         if (ret)
1485                 goto out;
1486
1487         ULIST_ITER_INIT(&ref_uiter);
1488         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1489                 ret = btrfs_find_all_roots(trans, fs_info, ref_node->val,
1490                                            tree_mod_seq_elem.seq, &roots);
1491                 if (ret)
1492                         break;
1493                 ULIST_ITER_INIT(&root_uiter);
1494                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1495                         pr_debug("root %llu references leaf %llu, data list "
1496                                  "%#llx\n", root_node->val, ref_node->val,
1497                                  (long long)ref_node->aux);
1498                         ret = iterate_leaf_refs((struct extent_inode_elem *)
1499                                                 (uintptr_t)ref_node->aux,
1500                                                 root_node->val,
1501                                                 extent_item_objectid,
1502                                                 iterate, ctx);
1503                 }
1504                 ulist_free(roots);
1505                 roots = NULL;
1506         }
1507
1508         free_leaf_list(refs);
1509         ulist_free(roots);
1510 out:
1511         if (!search_commit_root) {
1512                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1513                 btrfs_end_transaction(trans, fs_info->extent_root);
1514         }
1515
1516         return ret;
1517 }
1518
1519 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1520                                 struct btrfs_path *path,
1521                                 iterate_extent_inodes_t *iterate, void *ctx)
1522 {
1523         int ret;
1524         u64 extent_item_pos;
1525         u64 flags = 0;
1526         struct btrfs_key found_key;
1527         int search_commit_root = path->search_commit_root;
1528
1529         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1530         btrfs_release_path(path);
1531         if (ret < 0)
1532                 return ret;
1533         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1534                 return -EINVAL;
1535
1536         extent_item_pos = logical - found_key.objectid;
1537         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1538                                         extent_item_pos, search_commit_root,
1539                                         iterate, ctx);
1540
1541         return ret;
1542 }
1543
1544 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1545                               struct extent_buffer *eb, void *ctx);
1546
1547 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1548                               struct btrfs_path *path,
1549                               iterate_irefs_t *iterate, void *ctx)
1550 {
1551         int ret = 0;
1552         int slot;
1553         u32 cur;
1554         u32 len;
1555         u32 name_len;
1556         u64 parent = 0;
1557         int found = 0;
1558         struct extent_buffer *eb;
1559         struct btrfs_item *item;
1560         struct btrfs_inode_ref *iref;
1561         struct btrfs_key found_key;
1562
1563         while (!ret) {
1564                 path->leave_spinning = 1;
1565                 ret = inode_ref_info(inum, parent ? parent+1 : 0, fs_root, path,
1566                                      &found_key);
1567                 if (ret < 0)
1568                         break;
1569                 if (ret) {
1570                         ret = found ? 0 : -ENOENT;
1571                         break;
1572                 }
1573                 ++found;
1574
1575                 parent = found_key.offset;
1576                 slot = path->slots[0];
1577                 eb = path->nodes[0];
1578                 /* make sure we can use eb after releasing the path */
1579                 atomic_inc(&eb->refs);
1580                 btrfs_tree_read_lock(eb);
1581                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1582                 btrfs_release_path(path);
1583
1584                 item = btrfs_item_nr(eb, slot);
1585                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1586
1587                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
1588                         name_len = btrfs_inode_ref_name_len(eb, iref);
1589                         /* path must be released before calling iterate()! */
1590                         pr_debug("following ref at offset %u for inode %llu in "
1591                                  "tree %llu\n", cur,
1592                                  (unsigned long long)found_key.objectid,
1593                                  (unsigned long long)fs_root->objectid);
1594                         ret = iterate(parent, name_len,
1595                                       (unsigned long)(iref + 1), eb, ctx);
1596                         if (ret)
1597                                 break;
1598                         len = sizeof(*iref) + name_len;
1599                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
1600                 }
1601                 btrfs_tree_read_unlock_blocking(eb);
1602                 free_extent_buffer(eb);
1603         }
1604
1605         btrfs_release_path(path);
1606
1607         return ret;
1608 }
1609
1610 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
1611                                  struct btrfs_path *path,
1612                                  iterate_irefs_t *iterate, void *ctx)
1613 {
1614         int ret;
1615         int slot;
1616         u64 offset = 0;
1617         u64 parent;
1618         int found = 0;
1619         struct extent_buffer *eb;
1620         struct btrfs_inode_extref *extref;
1621         struct extent_buffer *leaf;
1622         u32 item_size;
1623         u32 cur_offset;
1624         unsigned long ptr;
1625
1626         while (1) {
1627                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
1628                                             &offset);
1629                 if (ret < 0)
1630                         break;
1631                 if (ret) {
1632                         ret = found ? 0 : -ENOENT;
1633                         break;
1634                 }
1635                 ++found;
1636
1637                 slot = path->slots[0];
1638                 eb = path->nodes[0];
1639                 /* make sure we can use eb after releasing the path */
1640                 atomic_inc(&eb->refs);
1641
1642                 btrfs_tree_read_lock(eb);
1643                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1644                 btrfs_release_path(path);
1645
1646                 leaf = path->nodes[0];
1647                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1648                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1649                 cur_offset = 0;
1650
1651                 while (cur_offset < item_size) {
1652                         u32 name_len;
1653
1654                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
1655                         parent = btrfs_inode_extref_parent(eb, extref);
1656                         name_len = btrfs_inode_extref_name_len(eb, extref);
1657                         ret = iterate(parent, name_len,
1658                                       (unsigned long)&extref->name, eb, ctx);
1659                         if (ret)
1660                                 break;
1661
1662                         cur_offset += btrfs_inode_extref_name_len(leaf, extref);
1663                         cur_offset += sizeof(*extref);
1664                 }
1665                 btrfs_tree_read_unlock_blocking(eb);
1666                 free_extent_buffer(eb);
1667
1668                 offset++;
1669         }
1670
1671         btrfs_release_path(path);
1672
1673         return ret;
1674 }
1675
1676 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
1677                          struct btrfs_path *path, iterate_irefs_t *iterate,
1678                          void *ctx)
1679 {
1680         int ret;
1681         int found_refs = 0;
1682
1683         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
1684         if (!ret)
1685                 ++found_refs;
1686         else if (ret != -ENOENT)
1687                 return ret;
1688
1689         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
1690         if (ret == -ENOENT && found_refs)
1691                 return 0;
1692
1693         return ret;
1694 }
1695
1696 /*
1697  * returns 0 if the path could be dumped (probably truncated)
1698  * returns <0 in case of an error
1699  */
1700 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
1701                          struct extent_buffer *eb, void *ctx)
1702 {
1703         struct inode_fs_paths *ipath = ctx;
1704         char *fspath;
1705         char *fspath_min;
1706         int i = ipath->fspath->elem_cnt;
1707         const int s_ptr = sizeof(char *);
1708         u32 bytes_left;
1709
1710         bytes_left = ipath->fspath->bytes_left > s_ptr ?
1711                                         ipath->fspath->bytes_left - s_ptr : 0;
1712
1713         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
1714         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
1715                                    name_off, eb, inum, fspath_min, bytes_left);
1716         if (IS_ERR(fspath))
1717                 return PTR_ERR(fspath);
1718
1719         if (fspath > fspath_min) {
1720                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
1721                 ++ipath->fspath->elem_cnt;
1722                 ipath->fspath->bytes_left = fspath - fspath_min;
1723         } else {
1724                 ++ipath->fspath->elem_missed;
1725                 ipath->fspath->bytes_missing += fspath_min - fspath;
1726                 ipath->fspath->bytes_left = 0;
1727         }
1728
1729         return 0;
1730 }
1731
1732 /*
1733  * this dumps all file system paths to the inode into the ipath struct, provided
1734  * is has been created large enough. each path is zero-terminated and accessed
1735  * from ipath->fspath->val[i].
1736  * when it returns, there are ipath->fspath->elem_cnt number of paths available
1737  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
1738  * number of missed paths in recored in ipath->fspath->elem_missed, otherwise,
1739  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
1740  * have been needed to return all paths.
1741  */
1742 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
1743 {
1744         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
1745                              inode_to_path, ipath);
1746 }
1747
1748 struct btrfs_data_container *init_data_container(u32 total_bytes)
1749 {
1750         struct btrfs_data_container *data;
1751         size_t alloc_bytes;
1752
1753         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
1754         data = vmalloc(alloc_bytes);
1755         if (!data)
1756                 return ERR_PTR(-ENOMEM);
1757
1758         if (total_bytes >= sizeof(*data)) {
1759                 data->bytes_left = total_bytes - sizeof(*data);
1760                 data->bytes_missing = 0;
1761         } else {
1762                 data->bytes_missing = sizeof(*data) - total_bytes;
1763                 data->bytes_left = 0;
1764         }
1765
1766         data->elem_cnt = 0;
1767         data->elem_missed = 0;
1768
1769         return data;
1770 }
1771
1772 /*
1773  * allocates space to return multiple file system paths for an inode.
1774  * total_bytes to allocate are passed, note that space usable for actual path
1775  * information will be total_bytes - sizeof(struct inode_fs_paths).
1776  * the returned pointer must be freed with free_ipath() in the end.
1777  */
1778 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
1779                                         struct btrfs_path *path)
1780 {
1781         struct inode_fs_paths *ifp;
1782         struct btrfs_data_container *fspath;
1783
1784         fspath = init_data_container(total_bytes);
1785         if (IS_ERR(fspath))
1786                 return (void *)fspath;
1787
1788         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
1789         if (!ifp) {
1790                 kfree(fspath);
1791                 return ERR_PTR(-ENOMEM);
1792         }
1793
1794         ifp->btrfs_path = path;
1795         ifp->fspath = fspath;
1796         ifp->fs_root = fs_root;
1797
1798         return ifp;
1799 }
1800
1801 void free_ipath(struct inode_fs_paths *ipath)
1802 {
1803         if (!ipath)
1804                 return;
1805         vfree(ipath->fspath);
1806         kfree(ipath);
1807 }