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