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