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 offset = 0;
  40         struct extent_inode_elem *e;
  41
  42         if (!btrfs_file_extent_compression(eb, fi) &&
  43             !btrfs_file_extent_encryption(eb, fi) &&
  44             !btrfs_file_extent_other_encoding(eb, fi)) {
  45                 u64 data_offset;
  46                 u64 data_len;
  47
  48                 data_offset = btrfs_file_extent_offset(eb, fi);
  49                 data_len = btrfs_file_extent_num_bytes(eb, fi);
  50
  51                 if (extent_item_pos < data_offset ||
  52                     extent_item_pos >= data_offset + data_len)
  53                         return 1;
  54                 offset = extent_item_pos - data_offset;
  55         }
  56
  57         e = kmalloc(sizeof(*e), GFP_NOFS);
  58         if (!e)
  59                 return -ENOMEM;
  60
  61         e->next = *eie;
  62         e->inum = key->objectid;
  63         e->offset = key->offset + offset;
  64         *eie = e;
  65
  66         return 0;
  67 }
  68
  69 static void free_inode_elem_list(struct extent_inode_elem *eie)
  70 {
  71         struct extent_inode_elem *eie_next;
  72
  73         for (; eie; eie = eie_next) {
  74                 eie_next = eie->next;
  75                 kfree(eie);
  76         }
  77 }
  78
  79 static int find_extent_in_eb(struct extent_buffer *eb, u64 wanted_disk_byte,
  80                                 u64 extent_item_pos,
  81                                 struct extent_inode_elem **eie)
  82 {
  83         u64 disk_byte;
  84         struct btrfs_key key;
  85         struct btrfs_file_extent_item *fi;
  86         int slot;
  87         int nritems;
  88         int extent_type;
  89         int ret;
  90
  91         /*
  92          * from the shared data ref, we only have the leaf but we need
  93          * the key. thus, we must look into all items and see that we
  94          * find one (some) with a reference to our extent item.
  95          */
  96         nritems = btrfs_header_nritems(eb);
  97         for (slot = 0; slot < nritems; ++slot) {
  98                 btrfs_item_key_to_cpu(eb, &key, slot);
  99                 if (key.type != BTRFS_EXTENT_DATA_KEY)
 100                         continue;
 101                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 102                 extent_type = btrfs_file_extent_type(eb, fi);
 103                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 104                         continue;
 105                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 106                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 107                 if (disk_byte != wanted_disk_byte)
 108                         continue;
 109
 110                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
 111                 if (ret < 0)
 112                         return ret;
 113         }
 114
 115         return 0;
 116 }
 117
 118 /*
 119  * this structure records all encountered refs on the way up to the root
 120  */
 121 struct __prelim_ref {
 122         struct list_head list;
 123         u64 root_id;
 124         struct btrfs_key key_for_search;
 125         int level;
 126         int count;
 127         struct extent_inode_elem *inode_list;
 128         u64 parent;
 129         u64 wanted_disk_byte;
 130 };
 131
 132 static struct kmem_cache *btrfs_prelim_ref_cache;
 133
 134 int __init btrfs_prelim_ref_init(void)
 135 {
 136         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 137                                         sizeof(struct __prelim_ref),
 138                                         0,
 139                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
 140                                         NULL);
 141         if (!btrfs_prelim_ref_cache)
 142                 return -ENOMEM;
 143         return 0;
 144 }
 145
 146 void btrfs_prelim_ref_exit(void)
 147 {
 148         if (btrfs_prelim_ref_cache)
 149                 kmem_cache_destroy(btrfs_prelim_ref_cache);
 150 }
 151
 152 /*
 153  * the rules for all callers of this function are:
 154  * - obtaining the parent is the goal
 155  * - if you add a key, you must know that it is a correct key
 156  * - if you cannot add the parent or a correct key, then we will look into the
 157  *   block later to set a correct key
 158  *
 159  * delayed refs
 160  * ============
 161  *        backref type | shared | indirect | shared | indirect
 162  * information         |   tree |     tree |   data |     data
 163  * --------------------+--------+----------+--------+----------
 164  *      parent logical |    y   |     -    |    -   |     -
 165  *      key to resolve |    -   |     y    |    y   |     y
 166  *  tree block logical |    -   |     -    |    -   |     -
 167  *  root for resolving |    y   |     y    |    y   |     y
 168  *
 169  * - column 1:       we've the parent -> done
 170  * - column 2, 3, 4: we use the key to find the parent
 171  *
 172  * on disk refs (inline or keyed)
 173  * ==============================
 174  *        backref type | shared | indirect | shared | indirect
 175  * information         |   tree |     tree |   data |     data
 176  * --------------------+--------+----------+--------+----------
 177  *      parent logical |    y   |     -    |    y   |     -
 178  *      key to resolve |    -   |     -    |    -   |     y
 179  *  tree block logical |    y   |     y    |    y   |     y
 180  *  root for resolving |    -   |     y    |    y   |     y
 181  *
 182  * - column 1, 3: we've the parent -> done
 183  * - column 2:    we take the first key from the block to find the parent
 184  *                (see __add_missing_keys)
 185  * - column 4:    we use the key to find the parent
 186  *
 187  * additional information that's available but not required to find the parent
 188  * block might help in merging entries to gain some speed.
 189  */
 190
 191 static int __add_prelim_ref(struct list_head *head, u64 root_id,
 192                             struct btrfs_key *key, int level,
 193                             u64 parent, u64 wanted_disk_byte, int count,
 194                             gfp_t gfp_mask)
 195 {
 196         struct __prelim_ref *ref;
 197
 198         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 199                 return 0;
 200
 201         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 202         if (!ref)
 203                 return -ENOMEM;
 204
 205         ref->root_id = root_id;
 206         if (key)
 207                 ref->key_for_search = *key;
 208         else
 209                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 210
 211         ref->inode_list = NULL;
 212         ref->level = level;
 213         ref->count = count;
 214         ref->parent = parent;
 215         ref->wanted_disk_byte = wanted_disk_byte;
 216         list_add_tail(&ref->list, head);
 217
 218         return 0;
 219 }
 220
 221 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 222                            struct ulist *parents, struct __prelim_ref *ref,
 223                            int level, u64 time_seq, const u64 *extent_item_pos)
 224 {
 225         int ret = 0;
 226         int slot;
 227         struct extent_buffer *eb;
 228         struct btrfs_key key;
 229         struct btrfs_key *key_for_search = &ref->key_for_search;
 230         struct btrfs_file_extent_item *fi;
 231         struct extent_inode_elem *eie = NULL, *old = NULL;
 232         u64 disk_byte;
 233         u64 wanted_disk_byte = ref->wanted_disk_byte;
 234         u64 count = 0;
 235
 236         if (level != 0) {
 237                 eb = path->nodes[level];
 238                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 239                 if (ret < 0)
 240                         return ret;
 241                 return 0;
 242         }
 243
 244         /*
 245          * We normally enter this function with the path already pointing to
 246          * the first item to check. But sometimes, we may enter it with
 247          * slot==nritems. In that case, go to the next leaf before we continue.
 248          */
 249         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0]))
 250                 ret = btrfs_next_old_leaf(root, path, time_seq);
 251
 252         while (!ret && count < ref->count) {
 253                 eb = path->nodes[0];
 254                 slot = path->slots[0];
 255
 256                 btrfs_item_key_to_cpu(eb, &key, slot);
 257
 258                 if (key.objectid != key_for_search->objectid ||
 259                     key.type != BTRFS_EXTENT_DATA_KEY)
 260                         break;
 261
 262                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 263                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 264
 265                 if (disk_byte == wanted_disk_byte) {
 266                         eie = NULL;
 267                         old = NULL;
 268                         count++;
 269                         if (extent_item_pos) {
 270                                 ret = check_extent_in_eb(&key, eb, fi,
 271                                                 *extent_item_pos,
 272                                                 &eie);
 273                                 if (ret < 0)
 274                                         break;
 275                         }
 276                         if (ret > 0)
 277                                 goto next;
 278                         ret = ulist_add_merge(parents, eb->start,
 279                                               (uintptr_t)eie,
 280                                               (u64 *)&old, GFP_NOFS);
 281                         if (ret < 0)
 282                                 break;
 283                         if (!ret && extent_item_pos) {
 284                                 while (old->next)
 285                                         old = old->next;
 286                                 old->next = eie;
 287                         }
 288                         eie = NULL;
 289                 }
 290 next:
 291                 ret = btrfs_next_old_item(root, path, time_seq);
 292         }
 293
 294         if (ret > 0)
 295                 ret = 0;
 296         else if (ret < 0)
 297                 free_inode_elem_list(eie);
 298         return ret;
 299 }
 300
 301 /*
 302  * resolve an indirect backref in the form (root_id, key, level)
 303  * to a logical address
 304  */
 305 static int __resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 306                                   struct btrfs_path *path, u64 time_seq,
 307                                   struct __prelim_ref *ref,
 308                                   struct ulist *parents,
 309                                   const u64 *extent_item_pos)
 310 {
 311         struct btrfs_root *root;
 312         struct btrfs_key root_key;
 313         struct extent_buffer *eb;
 314         int ret = 0;
 315         int root_level;
 316         int level = ref->level;
 317         int index;
 318
 319         root_key.objectid = ref->root_id;
 320         root_key.type = BTRFS_ROOT_ITEM_KEY;
 321         root_key.offset = (u64)-1;
 322
 323         index = srcu_read_lock(&fs_info->subvol_srcu);
 324
 325         root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 326         if (IS_ERR(root)) {
 327                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 328                 ret = PTR_ERR(root);
 329                 goto out;
 330         }
 331
 332         root_level = btrfs_old_root_level(root, time_seq);
 333
 334         if (root_level + 1 == level) {
 335                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 336                 goto out;
 337         }
 338
 339         path->lowest_level = level;
 340         ret = btrfs_search_old_slot(root, &ref->key_for_search, path, time_seq);
 341
 342         /* root node has been locked, we can release @subvol_srcu safely here */
 343         srcu_read_unlock(&fs_info->subvol_srcu, index);
 344
 345         pr_debug("search slot in root %llu (level %d, ref count %d) returned "
 346                  "%d for key (%llu %u %llu)\n",
 347                  ref->root_id, level, ref->count, ret,
 348                  ref->key_for_search.objectid, ref->key_for_search.type,
 349                  ref->key_for_search.offset);
 350         if (ret < 0)
 351                 goto out;
 352
 353         eb = path->nodes[level];
 354         while (!eb) {
 355                 if (WARN_ON(!level)) {
 356                         ret = 1;
 357                         goto out;
 358                 }
 359                 level--;
 360                 eb = path->nodes[level];
 361         }
 362
 363         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 364                               extent_item_pos);
 365 out:
 366         path->lowest_level = 0;
 367         btrfs_release_path(path);
 368         return ret;
 369 }
 370
 371 /*
 372  * resolve all indirect backrefs from the list
 373  */
 374 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 375                                    struct btrfs_path *path, u64 time_seq,
 376                                    struct list_head *head,
 377                                    const u64 *extent_item_pos)
 378 {
 379         int err;
 380         int ret = 0;
 381         struct __prelim_ref *ref;
 382         struct __prelim_ref *ref_safe;
 383         struct __prelim_ref *new_ref;
 384         struct ulist *parents;
 385         struct ulist_node *node;
 386         struct ulist_iterator uiter;
 387
 388         parents = ulist_alloc(GFP_NOFS);
 389         if (!parents)
 390                 return -ENOMEM;
 391
 392         /*
 393          * _safe allows us to insert directly after the current item without
 394          * iterating over the newly inserted items.
 395          * we're also allowed to re-assign ref during iteration.
 396          */
 397         list_for_each_entry_safe(ref, ref_safe, head, list) {
 398                 if (ref->parent)        /* already direct */
 399                         continue;
 400                 if (ref->count == 0)
 401                         continue;
 402                 err = __resolve_indirect_ref(fs_info, path, time_seq, ref,
 403                                              parents, extent_item_pos);
 404                 /*
 405                  * we can only tolerate ENOENT,otherwise,we should catch error
 406                  * and return directly.
 407                  */
 408                 if (err == -ENOENT) {
 409                         continue;
 410                 } else if (err) {
 411                         ret = err;
 412                         goto out;
 413                 }
 414
 415                 /* we put the first parent into the ref at hand */
 416                 ULIST_ITER_INIT(&uiter);
 417                 node = ulist_next(parents, &uiter);
 418                 ref->parent = node ? node->val : 0;
 419                 ref->inode_list = node ?
 420                         (struct extent_inode_elem *)(uintptr_t)node->aux : NULL;
 421
 422                 /* additional parents require new refs being added here */
 423                 while ((node = ulist_next(parents, &uiter))) {
 424                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 425                                                    GFP_NOFS);
 426                         if (!new_ref) {
 427                                 ret = -ENOMEM;
 428                                 goto out;
 429                         }
 430                         memcpy(new_ref, ref, sizeof(*ref));
 431                         new_ref->parent = node->val;
 432                         new_ref->inode_list = (struct extent_inode_elem *)
 433                                                         (uintptr_t)node->aux;
 434                         list_add(&new_ref->list, &ref->list);
 435                 }
 436                 ulist_reinit(parents);
 437         }
 438 out:
 439         ulist_free(parents);
 440         return ret;
 441 }
 442
 443 static inline int ref_for_same_block(struct __prelim_ref *ref1,
 444                                      struct __prelim_ref *ref2)
 445 {
 446         if (ref1->level != ref2->level)
 447                 return 0;
 448         if (ref1->root_id != ref2->root_id)
 449                 return 0;
 450         if (ref1->key_for_search.type != ref2->key_for_search.type)
 451                 return 0;
 452         if (ref1->key_for_search.objectid != ref2->key_for_search.objectid)
 453                 return 0;
 454         if (ref1->key_for_search.offset != ref2->key_for_search.offset)
 455                 return 0;
 456         if (ref1->parent != ref2->parent)
 457                 return 0;
 458
 459         return 1;
 460 }
 461
 462 /*
 463  * read tree blocks and add keys where required.
 464  */
 465 static int __add_missing_keys(struct btrfs_fs_info *fs_info,
 466                               struct list_head *head)
 467 {
 468         struct list_head *pos;
 469         struct extent_buffer *eb;
 470
 471         list_for_each(pos, head) {
 472                 struct __prelim_ref *ref;
 473                 ref = list_entry(pos, struct __prelim_ref, list);
 474
 475                 if (ref->parent)
 476                         continue;
 477                 if (ref->key_for_search.type)
 478                         continue;
 479                 BUG_ON(!ref->wanted_disk_byte);
 480                 eb = read_tree_block(fs_info->tree_root, ref->wanted_disk_byte,
 481                                      fs_info->tree_root->leafsize, 0);
 482                 if (!eb || !extent_buffer_uptodate(eb)) {
 483                         free_extent_buffer(eb);
 484                         return -EIO;
 485                 }
 486                 btrfs_tree_read_lock(eb);
 487                 if (btrfs_header_level(eb) == 0)
 488                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 489                 else
 490                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 491                 btrfs_tree_read_unlock(eb);
 492                 free_extent_buffer(eb);
 493         }
 494         return 0;
 495 }
 496
 497 /*
 498  * merge two lists of backrefs and adjust counts accordingly
 499  *
 500  * mode = 1: merge identical keys, if key is set
 501  *    FIXME: if we add more keys in __add_prelim_ref, we can merge more here.
 502  *           additionally, we could even add a key range for the blocks we
 503  *           looked into to merge even more (-> replace unresolved refs by those
 504  *           having a parent).
 505  * mode = 2: merge identical parents
 506  */
 507 static void __merge_refs(struct list_head *head, int mode)
 508 {
 509         struct list_head *pos1;
 510
 511         list_for_each(pos1, head) {
 512                 struct list_head *n2;
 513                 struct list_head *pos2;
 514                 struct __prelim_ref *ref1;
 515
 516                 ref1 = list_entry(pos1, struct __prelim_ref, list);
 517
 518                 for (pos2 = pos1->next, n2 = pos2->next; pos2 != head;
 519                      pos2 = n2, n2 = pos2->next) {
 520                         struct __prelim_ref *ref2;
 521                         struct __prelim_ref *xchg;
 522                         struct extent_inode_elem *eie;
 523
 524                         ref2 = list_entry(pos2, struct __prelim_ref, list);
 525
 526                         if (mode == 1) {
 527                                 if (!ref_for_same_block(ref1, ref2))
 528                                         continue;
 529                                 if (!ref1->parent && ref2->parent) {
 530                                         xchg = ref1;
 531                                         ref1 = ref2;
 532                                         ref2 = xchg;
 533                                 }
 534                         } else {
 535                                 if (ref1->parent != ref2->parent)
 536                                         continue;
 537                         }
 538
 539                         eie = ref1->inode_list;
 540                         while (eie && eie->next)
 541                                 eie = eie->next;
 542                         if (eie)
 543                                 eie->next = ref2->inode_list;
 544                         else
 545                                 ref1->inode_list = ref2->inode_list;
 546                         ref1->count += ref2->count;
 547
 548                         list_del(&ref2->list);
 549                         kmem_cache_free(btrfs_prelim_ref_cache, ref2);
 550                 }
 551
 552         }
 553 }
 554
 555 /*
 556  * add all currently queued delayed refs from this head whose seq nr is
 557  * smaller or equal that seq to the list
 558  */
 559 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 560                               struct list_head *prefs)
 561 {
 562         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 563         struct rb_node *n = &head->node.rb_node;
 564         struct btrfs_key key;
 565         struct btrfs_key op_key = {0};
 566         int sgn;
 567         int ret = 0;
 568
 569         if (extent_op && extent_op->update_key)
 570                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 571
 572         spin_lock(&head->lock);
 573         n = rb_first(&head->ref_root);
 574         while (n) {
 575                 struct btrfs_delayed_ref_node *node;
 576                 node = rb_entry(n, struct btrfs_delayed_ref_node,
 577                                 rb_node);
 578                 n = rb_next(n);
 579                 if (node->seq > seq)
 580                         continue;
 581
 582                 switch (node->action) {
 583                 case BTRFS_ADD_DELAYED_EXTENT:
 584                 case BTRFS_UPDATE_DELAYED_HEAD:
 585                         WARN_ON(1);
 586                         continue;
 587                 case BTRFS_ADD_DELAYED_REF:
 588                         sgn = 1;
 589                         break;
 590                 case BTRFS_DROP_DELAYED_REF:
 591                         sgn = -1;
 592                         break;
 593                 default:
 594                         BUG_ON(1);
 595                 }
 596                 switch (node->type) {
 597                 case BTRFS_TREE_BLOCK_REF_KEY: {
 598                         struct btrfs_delayed_tree_ref *ref;
 599
 600                         ref = btrfs_delayed_node_to_tree_ref(node);
 601                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
 602                                                ref->level + 1, 0, node->bytenr,
 603                                                node->ref_mod * sgn, GFP_ATOMIC);
 604                         break;
 605                 }
 606                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 607                         struct btrfs_delayed_tree_ref *ref;
 608
 609                         ref = btrfs_delayed_node_to_tree_ref(node);
 610                         ret = __add_prelim_ref(prefs, ref->root, NULL,
 611                                                ref->level + 1, ref->parent,
 612                                                node->bytenr,
 613                                                node->ref_mod * sgn, GFP_ATOMIC);
 614                         break;
 615                 }
 616                 case BTRFS_EXTENT_DATA_REF_KEY: {
 617                         struct btrfs_delayed_data_ref *ref;
 618                         ref = btrfs_delayed_node_to_data_ref(node);
 619
 620                         key.objectid = ref->objectid;
 621                         key.type = BTRFS_EXTENT_DATA_KEY;
 622                         key.offset = ref->offset;
 623                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
 624                                                node->bytenr,
 625                                                node->ref_mod * sgn, GFP_ATOMIC);
 626                         break;
 627                 }
 628                 case BTRFS_SHARED_DATA_REF_KEY: {
 629                         struct btrfs_delayed_data_ref *ref;
 630
 631                         ref = btrfs_delayed_node_to_data_ref(node);
 632
 633                         key.objectid = ref->objectid;
 634                         key.type = BTRFS_EXTENT_DATA_KEY;
 635                         key.offset = ref->offset;
 636                         ret = __add_prelim_ref(prefs, ref->root, &key, 0,
 637                                                ref->parent, node->bytenr,
 638                                                node->ref_mod * sgn, GFP_ATOMIC);
 639                         break;
 640                 }
 641                 default:
 642                         WARN_ON(1);
 643                 }
 644                 if (ret)
 645                         break;
 646         }
 647         spin_unlock(&head->lock);
 648         return ret;
 649 }
 650
 651 /*
 652  * add all inline backrefs for bytenr to the list
 653  */
 654 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
 655                              struct btrfs_path *path, u64 bytenr,
 656                              int *info_level, struct list_head *prefs)
 657 {
 658         int ret = 0;
 659         int slot;
 660         struct extent_buffer *leaf;
 661         struct btrfs_key key;
 662         struct btrfs_key found_key;
 663         unsigned long ptr;
 664         unsigned long end;
 665         struct btrfs_extent_item *ei;
 666         u64 flags;
 667         u64 item_size;
 668
 669         /*
 670          * enumerate all inline refs
 671          */
 672         leaf = path->nodes[0];
 673         slot = path->slots[0];
 674
 675         item_size = btrfs_item_size_nr(leaf, slot);
 676         BUG_ON(item_size < sizeof(*ei));
 677
 678         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 679         flags = btrfs_extent_flags(leaf, ei);
 680         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 681
 682         ptr = (unsigned long)(ei + 1);
 683         end = (unsigned long)ei + item_size;
 684
 685         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 686             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 687                 struct btrfs_tree_block_info *info;
 688
 689                 info = (struct btrfs_tree_block_info *)ptr;
 690                 *info_level = btrfs_tree_block_level(leaf, info);
 691                 ptr += sizeof(struct btrfs_tree_block_info);
 692                 BUG_ON(ptr > end);
 693         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
 694                 *info_level = found_key.offset;
 695         } else {
 696                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
 697         }
 698
 699         while (ptr < end) {
 700                 struct btrfs_extent_inline_ref *iref;
 701                 u64 offset;
 702                 int type;
 703
 704                 iref = (struct btrfs_extent_inline_ref *)ptr;
 705                 type = btrfs_extent_inline_ref_type(leaf, iref);
 706                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 707
 708                 switch (type) {
 709                 case BTRFS_SHARED_BLOCK_REF_KEY:
 710                         ret = __add_prelim_ref(prefs, 0, NULL,
 711                                                 *info_level + 1, offset,
 712                                                 bytenr, 1, GFP_NOFS);
 713                         break;
 714                 case BTRFS_SHARED_DATA_REF_KEY: {
 715                         struct btrfs_shared_data_ref *sdref;
 716                         int count;
 717
 718                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
 719                         count = btrfs_shared_data_ref_count(leaf, sdref);
 720                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
 721                                                bytenr, count, GFP_NOFS);
 722                         break;
 723                 }
 724                 case BTRFS_TREE_BLOCK_REF_KEY:
 725                         ret = __add_prelim_ref(prefs, offset, NULL,
 726                                                *info_level + 1, 0,
 727                                                bytenr, 1, GFP_NOFS);
 728                         break;
 729                 case BTRFS_EXTENT_DATA_REF_KEY: {
 730                         struct btrfs_extent_data_ref *dref;
 731                         int count;
 732                         u64 root;
 733
 734                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 735                         count = btrfs_extent_data_ref_count(leaf, dref);
 736                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 737                                                                       dref);
 738                         key.type = BTRFS_EXTENT_DATA_KEY;
 739                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 740                         root = btrfs_extent_data_ref_root(leaf, dref);
 741                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 742                                                bytenr, count, GFP_NOFS);
 743                         break;
 744                 }
 745                 default:
 746                         WARN_ON(1);
 747                 }
 748                 if (ret)
 749                         return ret;
 750                 ptr += btrfs_extent_inline_ref_size(type);
 751         }
 752
 753         return 0;
 754 }
 755
 756 /*
 757  * add all non-inline backrefs for bytenr to the list
 758  */
 759 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
 760                             struct btrfs_path *path, u64 bytenr,
 761                             int info_level, struct list_head *prefs)
 762 {
 763         struct btrfs_root *extent_root = fs_info->extent_root;
 764         int ret;
 765         int slot;
 766         struct extent_buffer *leaf;
 767         struct btrfs_key key;
 768
 769         while (1) {
 770                 ret = btrfs_next_item(extent_root, path);
 771                 if (ret < 0)
 772                         break;
 773                 if (ret) {
 774                         ret = 0;
 775                         break;
 776                 }
 777
 778                 slot = path->slots[0];
 779                 leaf = path->nodes[0];
 780                 btrfs_item_key_to_cpu(leaf, &key, slot);
 781
 782                 if (key.objectid != bytenr)
 783                         break;
 784                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
 785                         continue;
 786                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
 787                         break;
 788
 789                 switch (key.type) {
 790                 case BTRFS_SHARED_BLOCK_REF_KEY:
 791                         ret = __add_prelim_ref(prefs, 0, NULL,
 792                                                 info_level + 1, key.offset,
 793                                                 bytenr, 1, GFP_NOFS);
 794                         break;
 795                 case BTRFS_SHARED_DATA_REF_KEY: {
 796                         struct btrfs_shared_data_ref *sdref;
 797                         int count;
 798
 799                         sdref = btrfs_item_ptr(leaf, slot,
 800                                               struct btrfs_shared_data_ref);
 801                         count = btrfs_shared_data_ref_count(leaf, sdref);
 802                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
 803                                                 bytenr, count, GFP_NOFS);
 804                         break;
 805                 }
 806                 case BTRFS_TREE_BLOCK_REF_KEY:
 807                         ret = __add_prelim_ref(prefs, key.offset, NULL,
 808                                                info_level + 1, 0,
 809                                                bytenr, 1, GFP_NOFS);
 810                         break;
 811                 case BTRFS_EXTENT_DATA_REF_KEY: {
 812                         struct btrfs_extent_data_ref *dref;
 813                         int count;
 814                         u64 root;
 815
 816                         dref = btrfs_item_ptr(leaf, slot,
 817                                               struct btrfs_extent_data_ref);
 818                         count = btrfs_extent_data_ref_count(leaf, dref);
 819                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
 820                                                                       dref);
 821                         key.type = BTRFS_EXTENT_DATA_KEY;
 822                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
 823                         root = btrfs_extent_data_ref_root(leaf, dref);
 824                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
 825                                                bytenr, count, GFP_NOFS);
 826                         break;
 827                 }
 828                 default:
 829                         WARN_ON(1);
 830                 }
 831                 if (ret)
 832                         return ret;
 833
 834         }
 835
 836         return ret;
 837 }
 838
 839 /*
 840  * this adds all existing backrefs (inline backrefs, backrefs and delayed
 841  * refs) for the given bytenr to the refs list, merges duplicates and resolves
 842  * indirect refs to their parent bytenr.
 843  * When roots are found, they're added to the roots list
 844  *
 845  * FIXME some caching might speed things up
 846  */
 847 static int find_parent_nodes(struct btrfs_trans_handle *trans,
 848                              struct btrfs_fs_info *fs_info, u64 bytenr,
 849                              u64 time_seq, struct ulist *refs,
 850                              struct ulist *roots, const u64 *extent_item_pos)
 851 {
 852         struct btrfs_key key;
 853         struct btrfs_path *path;
 854         struct btrfs_delayed_ref_root *delayed_refs = NULL;
 855         struct btrfs_delayed_ref_head *head;
 856         int info_level = 0;
 857         int ret;
 858         struct list_head prefs_delayed;
 859         struct list_head prefs;
 860         struct __prelim_ref *ref;
 861         struct extent_inode_elem *eie = NULL;
 862
 863         INIT_LIST_HEAD(&prefs);
 864         INIT_LIST_HEAD(&prefs_delayed);
 865
 866         key.objectid = bytenr;
 867         key.offset = (u64)-1;
 868         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
 869                 key.type = BTRFS_METADATA_ITEM_KEY;
 870         else
 871                 key.type = BTRFS_EXTENT_ITEM_KEY;
 872
 873         path = btrfs_alloc_path();
 874         if (!path)
 875                 return -ENOMEM;
 876         if (!trans)
 877                 path->search_commit_root = 1;
 878
 879         /*
 880          * grab both a lock on the path and a lock on the delayed ref head.
 881          * We need both to get a consistent picture of how the refs look
 882          * at a specified point in time
 883          */
 884 again:
 885         head = NULL;
 886
 887         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
 888         if (ret < 0)
 889                 goto out;
 890         BUG_ON(ret == 0);
 891
 892         if (trans) {
 893                 /*
 894                  * look if there are updates for this ref queued and lock the
 895                  * head
 896                  */
 897                 delayed_refs = &trans->transaction->delayed_refs;
 898                 spin_lock(&delayed_refs->lock);
 899                 head = btrfs_find_delayed_ref_head(trans, bytenr);
 900                 if (head) {
 901                         if (!mutex_trylock(&head->mutex)) {
 902                                 atomic_inc(&head->node.refs);
 903                                 spin_unlock(&delayed_refs->lock);
 904
 905                                 btrfs_release_path(path);
 906
 907                                 /*
 908                                  * Mutex was contended, block until it's
 909                                  * released and try again
 910                                  */
 911                                 mutex_lock(&head->mutex);
 912                                 mutex_unlock(&head->mutex);
 913                                 btrfs_put_delayed_ref(&head->node);
 914                                 goto again;
 915                         }
 916                         spin_unlock(&delayed_refs->lock);
 917                         ret = __add_delayed_refs(head, time_seq,
 918                                                  &prefs_delayed);
 919                         mutex_unlock(&head->mutex);
 920                         if (ret)
 921                                 goto out;
 922                 } else {
 923                         spin_unlock(&delayed_refs->lock);
 924                 }
 925         }
 926
 927         if (path->slots[0]) {
 928                 struct extent_buffer *leaf;
 929                 int slot;
 930
 931                 path->slots[0]--;
 932                 leaf = path->nodes[0];
 933                 slot = path->slots[0];
 934                 btrfs_item_key_to_cpu(leaf, &key, slot);
 935                 if (key.objectid == bytenr &&
 936                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
 937                      key.type == BTRFS_METADATA_ITEM_KEY)) {
 938                         ret = __add_inline_refs(fs_info, path, bytenr,
 939                                                 &info_level, &prefs);
 940                         if (ret)
 941                                 goto out;
 942                         ret = __add_keyed_refs(fs_info, path, bytenr,
 943                                                info_level, &prefs);
 944                         if (ret)
 945                                 goto out;
 946                 }
 947         }
 948         btrfs_release_path(path);
 949
 950         list_splice_init(&prefs_delayed, &prefs);
 951
 952         ret = __add_missing_keys(fs_info, &prefs);
 953         if (ret)
 954                 goto out;
 955
 956         __merge_refs(&prefs, 1);
 957
 958         ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
 959                                       extent_item_pos);
 960         if (ret)
 961                 goto out;
 962
 963         __merge_refs(&prefs, 2);
 964
 965         while (!list_empty(&prefs)) {
 966                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
 967                 WARN_ON(ref->count < 0);
 968                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
 969                         /* no parent == root of tree */
 970                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
 971                         if (ret < 0)
 972                                 goto out;
 973                 }
 974                 if (ref->count && ref->parent) {
 975                         if (extent_item_pos && !ref->inode_list) {
 976                                 u32 bsz;
 977                                 struct extent_buffer *eb;
 978                                 bsz = btrfs_level_size(fs_info->extent_root,
 979                                                         info_level);
 980                                 eb = read_tree_block(fs_info->extent_root,
 981                                                            ref->parent, bsz, 0);
 982                                 if (!eb || !extent_buffer_uptodate(eb)) {
 983                                         free_extent_buffer(eb);
 984                                         ret = -EIO;
 985                                         goto out;
 986                                 }
 987                                 ret = find_extent_in_eb(eb, bytenr,
 988                                                         *extent_item_pos, &eie);
 989                                 free_extent_buffer(eb);
 990                                 if (ret < 0)
 991                                         goto out;
 992                                 ref->inode_list = eie;
 993                         }
 994                         ret = ulist_add_merge(refs, ref->parent,
 995                                               (uintptr_t)ref->inode_list,
 996                                               (u64 *)&eie, GFP_NOFS);
 997                         if (ret < 0)
 998                                 goto out;
 999                         if (!ret && extent_item_pos) {
1000                                 /*
1001                                  * we've recorded that parent, so we must extend
1002                                  * its inode list here
1003                                  */
1004                                 BUG_ON(!eie);
1005                                 while (eie->next)
1006                                         eie = eie->next;
1007                                 eie->next = ref->inode_list;
1008                         }
1009                         eie = NULL;
1010                 }
1011                 list_del(&ref->list);
1012                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1013         }
1014
1015 out:
1016         btrfs_free_path(path);
1017         while (!list_empty(&prefs)) {
1018                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1019                 list_del(&ref->list);
1020                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1021         }
1022         while (!list_empty(&prefs_delayed)) {
1023                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
1024                                        list);
1025                 list_del(&ref->list);
1026                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1027         }
1028         if (ret < 0)
1029                 free_inode_elem_list(eie);
1030         return ret;
1031 }
1032
1033 static void free_leaf_list(struct ulist *blocks)
1034 {
1035         struct ulist_node *node = NULL;
1036         struct extent_inode_elem *eie;
1037         struct ulist_iterator uiter;
1038
1039         ULIST_ITER_INIT(&uiter);
1040         while ((node = ulist_next(blocks, &uiter))) {
1041                 if (!node->aux)
1042                         continue;
1043                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
1044                 free_inode_elem_list(eie);
1045                 node->aux = 0;
1046         }
1047
1048         ulist_free(blocks);
1049 }
1050
1051 /*
1052  * Finds all leafs with a reference to the specified combination of bytenr and
1053  * offset. key_list_head will point to a list of corresponding keys (caller must
1054  * free each list element). The leafs will be stored in the leafs ulist, which
1055  * must be freed with ulist_free.
1056  *
1057  * returns 0 on success, <0 on error
1058  */
1059 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1060                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1061                                 u64 time_seq, struct ulist **leafs,
1062                                 const u64 *extent_item_pos)
1063 {
1064         int ret;
1065
1066         *leafs = ulist_alloc(GFP_NOFS);
1067         if (!*leafs)
1068                 return -ENOMEM;
1069
1070         ret = find_parent_nodes(trans, fs_info, bytenr,
1071                                 time_seq, *leafs, NULL, extent_item_pos);
1072         if (ret < 0 && ret != -ENOENT) {
1073                 free_leaf_list(*leafs);
1074                 return ret;
1075         }
1076
1077         return 0;
1078 }
1079
1080 /*
1081  * walk all backrefs for a given extent to find all roots that reference this
1082  * extent. Walking a backref means finding all extents that reference this
1083  * extent and in turn walk the backrefs of those, too. Naturally this is a
1084  * recursive process, but here it is implemented in an iterative fashion: We
1085  * find all referencing extents for the extent in question and put them on a
1086  * list. In turn, we find all referencing extents for those, further appending
1087  * to the list. The way we iterate the list allows adding more elements after
1088  * the current while iterating. The process stops when we reach the end of the
1089  * list. Found roots are added to the roots list.
1090  *
1091  * returns 0 on success, < 0 on error.
1092  */
1093 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1094                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1095                                 u64 time_seq, struct ulist **roots)
1096 {
1097         struct ulist *tmp;
1098         struct ulist_node *node = NULL;
1099         struct ulist_iterator uiter;
1100         int ret;
1101
1102         tmp = ulist_alloc(GFP_NOFS);
1103         if (!tmp)
1104                 return -ENOMEM;
1105         *roots = ulist_alloc(GFP_NOFS);
1106         if (!*roots) {
1107                 ulist_free(tmp);
1108                 return -ENOMEM;
1109         }
1110
1111         ULIST_ITER_INIT(&uiter);
1112         while (1) {
1113                 ret = find_parent_nodes(trans, fs_info, bytenr,
1114                                         time_seq, tmp, *roots, NULL);
1115                 if (ret < 0 && ret != -ENOENT) {
1116                         ulist_free(tmp);
1117                         ulist_free(*roots);
1118                         return ret;
1119                 }
1120                 node = ulist_next(tmp, &uiter);
1121                 if (!node)
1122                         break;
1123                 bytenr = node->val;
1124                 cond_resched();
1125         }
1126
1127         ulist_free(tmp);
1128         return 0;
1129 }
1130
1131 /*
1132  * this makes the path point to (inum INODE_ITEM ioff)
1133  */
1134 int inode_item_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1135                         struct btrfs_path *path)
1136 {
1137         struct btrfs_key key;
1138         return btrfs_find_item(fs_root, path, inum, ioff,
1139                         BTRFS_INODE_ITEM_KEY, &key);
1140 }
1141
1142 static int inode_ref_info(u64 inum, u64 ioff, struct btrfs_root *fs_root,
1143                                 struct btrfs_path *path,
1144                                 struct btrfs_key *found_key)
1145 {
1146         return btrfs_find_item(fs_root, path, inum, ioff,
1147                         BTRFS_INODE_REF_KEY, found_key);
1148 }
1149
1150 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1151                           u64 start_off, struct btrfs_path *path,
1152                           struct btrfs_inode_extref **ret_extref,
1153                           u64 *found_off)
1154 {
1155         int ret, slot;
1156         struct btrfs_key key;
1157         struct btrfs_key found_key;
1158         struct btrfs_inode_extref *extref;
1159         struct extent_buffer *leaf;
1160         unsigned long ptr;
1161
1162         key.objectid = inode_objectid;
1163         btrfs_set_key_type(&key, BTRFS_INODE_EXTREF_KEY);
1164         key.offset = start_off;
1165
1166         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1167         if (ret < 0)
1168                 return ret;
1169
1170         while (1) {
1171                 leaf = path->nodes[0];
1172                 slot = path->slots[0];
1173                 if (slot >= btrfs_header_nritems(leaf)) {
1174                         /*
1175                          * If the item at offset is not found,
1176                          * btrfs_search_slot will point us to the slot
1177                          * where it should be inserted. In our case
1178                          * that will be the slot directly before the
1179                          * next INODE_REF_KEY_V2 item. In the case
1180                          * that we're pointing to the last slot in a
1181                          * leaf, we must move one leaf over.
1182                          */
1183                         ret = btrfs_next_leaf(root, path);
1184                         if (ret) {
1185                                 if (ret >= 1)
1186                                         ret = -ENOENT;
1187                                 break;
1188                         }
1189                         continue;
1190                 }
1191
1192                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1193
1194                 /*
1195                  * Check that we're still looking at an extended ref key for
1196                  * this particular objectid. If we have different
1197                  * objectid or type then there are no more to be found
1198                  * in the tree and we can exit.
1199                  */
1200                 ret = -ENOENT;
1201                 if (found_key.objectid != inode_objectid)
1202                         break;
1203                 if (btrfs_key_type(&found_key) != BTRFS_INODE_EXTREF_KEY)
1204                         break;
1205
1206                 ret = 0;
1207                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1208                 extref = (struct btrfs_inode_extref *)ptr;
1209                 *ret_extref = extref;
1210                 if (found_off)
1211                         *found_off = found_key.offset;
1212                 break;
1213         }
1214
1215         return ret;
1216 }
1217
1218 /*
1219  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1220  * Elements of the path are separated by '/' and the path is guaranteed to be
1221  * 0-terminated. the path is only given within the current file system.
1222  * Therefore, it never starts with a '/'. the caller is responsible to provide
1223  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1224  * the start point of the resulting string is returned. this pointer is within
1225  * dest, normally.
1226  * in case the path buffer would overflow, the pointer is decremented further
1227  * as if output was written to the buffer, though no more output is actually
1228  * generated. that way, the caller can determine how much space would be
1229  * required for the path to fit into the buffer. in that case, the returned
1230  * value will be smaller than dest. callers must check this!
1231  */
1232 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1233                         u32 name_len, unsigned long name_off,
1234                         struct extent_buffer *eb_in, u64 parent,
1235                         char *dest, u32 size)
1236 {
1237         int slot;
1238         u64 next_inum;
1239         int ret;
1240         s64 bytes_left = ((s64)size) - 1;
1241         struct extent_buffer *eb = eb_in;
1242         struct btrfs_key found_key;
1243         int leave_spinning = path->leave_spinning;
1244         struct btrfs_inode_ref *iref;
1245
1246         if (bytes_left >= 0)
1247                 dest[bytes_left] = '\0';
1248
1249         path->leave_spinning = 1;
1250         while (1) {
1251                 bytes_left -= name_len;
1252                 if (bytes_left >= 0)
1253                         read_extent_buffer(eb, dest + bytes_left,
1254                                            name_off, name_len);
1255                 if (eb != eb_in) {
1256                         btrfs_tree_read_unlock_blocking(eb);
1257                         free_extent_buffer(eb);
1258                 }
1259                 ret = inode_ref_info(parent, 0, fs_root, path, &found_key);
1260                 if (ret > 0)
1261                         ret = -ENOENT;
1262                 if (ret)
1263                         break;
1264
1265                 next_inum = found_key.offset;
1266
1267                 /* regular exit ahead */
1268                 if (parent == next_inum)
1269                         break;
1270
1271                 slot = path->slots[0];
1272                 eb = path->nodes[0];
1273                 /* make sure we can use eb after releasing the path */
1274                 if (eb != eb_in) {
1275                         atomic_inc(&eb->refs);
1276                         btrfs_tree_read_lock(eb);
1277                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1278                 }
1279                 btrfs_release_path(path);
1280                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1281
1282                 name_len = btrfs_inode_ref_name_len(eb, iref);
1283                 name_off = (unsigned long)(iref + 1);
1284
1285                 parent = next_inum;
1286                 --bytes_left;
1287                 if (bytes_left >= 0)
1288                         dest[bytes_left] = '/';
1289         }
1290
1291         btrfs_release_path(path);
1292         path->leave_spinning = leave_spinning;
1293
1294         if (ret)
1295                 return ERR_PTR(ret);
1296
1297         return dest + bytes_left;
1298 }
1299
1300 /*
1301  * this makes the path point to (logical EXTENT_ITEM *)
1302  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1303  * tree blocks and <0 on error.
1304  */
1305 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1306                         struct btrfs_path *path, struct btrfs_key *found_key,
1307                         u64 *flags_ret)
1308 {
1309         int ret;
1310         u64 flags;
1311         u64 size = 0;
1312         u32 item_size;
1313         struct extent_buffer *eb;
1314         struct btrfs_extent_item *ei;
1315         struct btrfs_key key;
1316
1317         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1318                 key.type = BTRFS_METADATA_ITEM_KEY;
1319         else
1320                 key.type = BTRFS_EXTENT_ITEM_KEY;
1321         key.objectid = logical;
1322         key.offset = (u64)-1;
1323
1324         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1325         if (ret < 0)
1326                 return ret;
1327
1328         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1329         if (ret) {
1330                 if (ret > 0)
1331                         ret = -ENOENT;
1332                 return ret;
1333         }
1334         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1335         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1336                 size = fs_info->extent_root->leafsize;
1337         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1338                 size = found_key->offset;
1339
1340         if (found_key->objectid > logical ||
1341             found_key->objectid + size <= logical) {
1342                 pr_debug("logical %llu is not within any extent\n", logical);
1343                 return -ENOENT;
1344         }
1345
1346         eb = path->nodes[0];
1347         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1348         BUG_ON(item_size < sizeof(*ei));
1349
1350         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1351         flags = btrfs_extent_flags(eb, ei);
1352
1353         pr_debug("logical %llu is at position %llu within the extent (%llu "
1354                  "EXTENT_ITEM %llu) flags %#llx size %u\n",
1355                  logical, logical - found_key->objectid, found_key->objectid,
1356                  found_key->offset, flags, item_size);
1357
1358         WARN_ON(!flags_ret);
1359         if (flags_ret) {
1360                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1361                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1362                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1363                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1364                 else
1365                         BUG_ON(1);
1366                 return 0;
1367         }
1368
1369         return -EIO;
1370 }
1371
1372 /*
1373  * helper function to iterate extent inline refs. ptr must point to a 0 value
1374  * for the first call and may be modified. it is used to track state.
1375  * if more refs exist, 0 is returned and the next call to
1376  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1377  * next ref. after the last ref was processed, 1 is returned.
1378  * returns <0 on error
1379  */
1380 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1381                                 struct btrfs_extent_item *ei, u32 item_size,
1382                                 struct btrfs_extent_inline_ref **out_eiref,
1383                                 int *out_type)
1384 {
1385         unsigned long end;
1386         u64 flags;
1387         struct btrfs_tree_block_info *info;
1388
1389         if (!*ptr) {
1390                 /* first call */
1391                 flags = btrfs_extent_flags(eb, ei);
1392                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1393                         info = (struct btrfs_tree_block_info *)(ei + 1);
1394                         *out_eiref =
1395                                 (struct btrfs_extent_inline_ref *)(info + 1);
1396                 } else {
1397                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1398                 }
1399                 *ptr = (unsigned long)*out_eiref;
1400                 if ((void *)*ptr >= (void *)ei + item_size)
1401                         return -ENOENT;
1402         }
1403
1404         end = (unsigned long)ei + item_size;
1405         *out_eiref = (struct btrfs_extent_inline_ref *)*ptr;
1406         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1407
1408         *ptr += btrfs_extent_inline_ref_size(*out_type);
1409         WARN_ON(*ptr > end);
1410         if (*ptr == end)
1411                 return 1; /* last */
1412
1413         return 0;
1414 }
1415
1416 /*
1417  * reads the tree block backref for an extent. tree level and root are returned
1418  * through out_level and out_root. ptr must point to a 0 value for the first
1419  * call and may be modified (see __get_extent_inline_ref comment).
1420  * returns 0 if data was provided, 1 if there was no more data to provide or
1421  * <0 on error.
1422  */
1423 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1424                                 struct btrfs_extent_item *ei, u32 item_size,
1425                                 u64 *out_root, u8 *out_level)
1426 {
1427         int ret;
1428         int type;
1429         struct btrfs_tree_block_info *info;
1430         struct btrfs_extent_inline_ref *eiref;
1431
1432         if (*ptr == (unsigned long)-1)
1433                 return 1;
1434
1435         while (1) {
1436                 ret = __get_extent_inline_ref(ptr, eb, ei, item_size,
1437                                                 &eiref, &type);
1438                 if (ret < 0)
1439                         return ret;
1440
1441                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1442                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1443                         break;
1444
1445                 if (ret == 1)
1446                         return 1;
1447         }
1448
1449         /* we can treat both ref types equally here */
1450         info = (struct btrfs_tree_block_info *)(ei + 1);
1451         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1452         *out_level = btrfs_tree_block_level(eb, info);
1453
1454         if (ret == 1)
1455                 *ptr = (unsigned long)-1;
1456
1457         return 0;
1458 }
1459
1460 static int iterate_leaf_refs(struct extent_inode_elem *inode_list,
1461                                 u64 root, u64 extent_item_objectid,
1462                                 iterate_extent_inodes_t *iterate, void *ctx)
1463 {
1464         struct extent_inode_elem *eie;
1465         int ret = 0;
1466
1467         for (eie = inode_list; eie; eie = eie->next) {
1468                 pr_debug("ref for %llu resolved, key (%llu EXTEND_DATA %llu), "
1469                          "root %llu\n", extent_item_objectid,
1470                          eie->inum, eie->offset, root);
1471                 ret = iterate(eie->inum, eie->offset, root, ctx);
1472                 if (ret) {
1473                         pr_debug("stopping iteration for %llu due to ret=%d\n",
1474                                  extent_item_objectid, ret);
1475                         break;
1476                 }
1477         }
1478
1479         return ret;
1480 }
1481
1482 /*
1483  * calls iterate() for every inode that references the extent identified by
1484  * the given parameters.
1485  * when the iterator function returns a non-zero value, iteration stops.
1486  */
1487 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1488                                 u64 extent_item_objectid, u64 extent_item_pos,
1489                                 int search_commit_root,
1490                                 iterate_extent_inodes_t *iterate, void *ctx)
1491 {
1492         int ret;
1493         struct btrfs_trans_handle *trans = NULL;
1494         struct ulist *refs = NULL;
1495         struct ulist *roots = NULL;
1496         struct ulist_node *ref_node = NULL;
1497         struct ulist_node *root_node = NULL;
1498         struct seq_list tree_mod_seq_elem = {};
1499         struct ulist_iterator ref_uiter;
1500         struct ulist_iterator root_uiter;
1501
1502         pr_debug("resolving all inodes for extent %llu\n",
1503                         extent_item_objectid);
1504
1505         if (!search_commit_root) {
1506                 trans = btrfs_join_transaction(fs_info->extent_root);
1507                 if (IS_ERR(trans))
1508                         return PTR_ERR(trans);
1509                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1510         }
1511
1512         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1513                                    tree_mod_seq_elem.seq, &refs,
1514                                    &extent_item_pos);
1515         if (ret)
1516                 goto out;
1517
1518         ULIST_ITER_INIT(&ref_uiter);
1519         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1520                 ret = btrfs_find_all_roots(trans, fs_info, ref_node->val,
1521                                            tree_mod_seq_elem.seq, &roots);
1522                 if (ret)
1523                         break;
1524                 ULIST_ITER_INIT(&root_uiter);
1525                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1526                         pr_debug("root %llu references leaf %llu, data list "
1527                                  "%#llx\n", root_node->val, ref_node->val,
1528                                  ref_node->aux);
1529                         ret = iterate_leaf_refs((struct extent_inode_elem *)
1530                                                 (uintptr_t)ref_node->aux,
1531                                                 root_node->val,
1532                                                 extent_item_objectid,
1533                                                 iterate, ctx);
1534                 }
1535                 ulist_free(roots);
1536         }
1537
1538         free_leaf_list(refs);
1539 out:
1540         if (!search_commit_root) {
1541                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1542                 btrfs_end_transaction(trans, fs_info->extent_root);
1543         }
1544
1545         return ret;
1546 }
1547
1548 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1549                                 struct btrfs_path *path,
1550                                 iterate_extent_inodes_t *iterate, void *ctx)
1551 {
1552         int ret;
1553         u64 extent_item_pos;
1554         u64 flags = 0;
1555         struct btrfs_key found_key;
1556         int search_commit_root = path->search_commit_root;
1557
1558         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1559         btrfs_release_path(path);
1560         if (ret < 0)
1561                 return ret;
1562         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1563                 return -EINVAL;
1564
1565         extent_item_pos = logical - found_key.objectid;
1566         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1567                                         extent_item_pos, search_commit_root,
1568                                         iterate, ctx);
1569
1570         return ret;
1571 }
1572
1573 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1574                               struct extent_buffer *eb, void *ctx);
1575
1576 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1577                               struct btrfs_path *path,
1578                               iterate_irefs_t *iterate, void *ctx)
1579 {
1580         int ret = 0;
1581         int slot;
1582         u32 cur;
1583         u32 len;
1584         u32 name_len;
1585         u64 parent = 0;
1586         int found = 0;
1587         struct extent_buffer *eb;
1588         struct btrfs_item *item;
1589         struct btrfs_inode_ref *iref;
1590         struct btrfs_key found_key;
1591
1592         while (!ret) {
1593                 ret = inode_ref_info(inum, parent ? parent+1 : 0, fs_root, path,
1594                                      &found_key);
1595                 if (ret < 0)
1596                         break;
1597                 if (ret) {
1598                         ret = found ? 0 : -ENOENT;
1599                         break;
1600                 }
1601                 ++found;
1602
1603                 parent = found_key.offset;
1604                 slot = path->slots[0];
1605                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
1606                 if (!eb) {
1607                         ret = -ENOMEM;
1608                         break;
1609                 }
1610                 extent_buffer_get(eb);
1611                 btrfs_tree_read_lock(eb);
1612                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1613                 btrfs_release_path(path);
1614
1615                 item = btrfs_item_nr(slot);
1616                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1617
1618                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
1619                         name_len = btrfs_inode_ref_name_len(eb, iref);
1620                         /* path must be released before calling iterate()! */
1621                         pr_debug("following ref at offset %u for inode %llu in "
1622                                  "tree %llu\n", cur, found_key.objectid,
1623                                  fs_root->objectid);
1624                         ret = iterate(parent, name_len,
1625                                       (unsigned long)(iref + 1), eb, ctx);
1626                         if (ret)
1627                                 break;
1628                         len = sizeof(*iref) + name_len;
1629                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
1630                 }
1631                 btrfs_tree_read_unlock_blocking(eb);
1632                 free_extent_buffer(eb);
1633         }
1634
1635         btrfs_release_path(path);
1636
1637         return ret;
1638 }
1639
1640 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
1641                                  struct btrfs_path *path,
1642                                  iterate_irefs_t *iterate, void *ctx)
1643 {
1644         int ret;
1645         int slot;
1646         u64 offset = 0;
1647         u64 parent;
1648         int found = 0;
1649         struct extent_buffer *eb;
1650         struct btrfs_inode_extref *extref;
1651         struct extent_buffer *leaf;
1652         u32 item_size;
1653         u32 cur_offset;
1654         unsigned long ptr;
1655
1656         while (1) {
1657                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
1658                                             &offset);
1659                 if (ret < 0)
1660                         break;
1661                 if (ret) {
1662                         ret = found ? 0 : -ENOENT;
1663                         break;
1664                 }
1665                 ++found;
1666
1667                 slot = path->slots[0];
1668                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
1669                 if (!eb) {
1670                         ret = -ENOMEM;
1671                         break;
1672                 }
1673                 extent_buffer_get(eb);
1674
1675                 btrfs_tree_read_lock(eb);
1676                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1677                 btrfs_release_path(path);
1678
1679                 leaf = path->nodes[0];
1680                 item_size = btrfs_item_size_nr(leaf, slot);
1681                 ptr = btrfs_item_ptr_offset(leaf, slot);
1682                 cur_offset = 0;
1683
1684                 while (cur_offset < item_size) {
1685                         u32 name_len;
1686
1687                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
1688                         parent = btrfs_inode_extref_parent(eb, extref);
1689                         name_len = btrfs_inode_extref_name_len(eb, extref);
1690                         ret = iterate(parent, name_len,
1691                                       (unsigned long)&extref->name, eb, ctx);
1692                         if (ret)
1693                                 break;
1694
1695                         cur_offset += btrfs_inode_extref_name_len(leaf, extref);
1696                         cur_offset += sizeof(*extref);
1697                 }
1698                 btrfs_tree_read_unlock_blocking(eb);
1699                 free_extent_buffer(eb);
1700
1701                 offset++;
1702         }
1703
1704         btrfs_release_path(path);
1705
1706         return ret;
1707 }
1708
1709 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
1710                          struct btrfs_path *path, iterate_irefs_t *iterate,
1711                          void *ctx)
1712 {
1713         int ret;
1714         int found_refs = 0;
1715
1716         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
1717         if (!ret)
1718                 ++found_refs;
1719         else if (ret != -ENOENT)
1720                 return ret;
1721
1722         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
1723         if (ret == -ENOENT && found_refs)
1724                 return 0;
1725
1726         return ret;
1727 }
1728
1729 /*
1730  * returns 0 if the path could be dumped (probably truncated)
1731  * returns <0 in case of an error
1732  */
1733 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
1734                          struct extent_buffer *eb, void *ctx)
1735 {
1736         struct inode_fs_paths *ipath = ctx;
1737         char *fspath;
1738         char *fspath_min;
1739         int i = ipath->fspath->elem_cnt;
1740         const int s_ptr = sizeof(char *);
1741         u32 bytes_left;
1742
1743         bytes_left = ipath->fspath->bytes_left > s_ptr ?
1744                                         ipath->fspath->bytes_left - s_ptr : 0;
1745
1746         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
1747         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
1748                                    name_off, eb, inum, fspath_min, bytes_left);
1749         if (IS_ERR(fspath))
1750                 return PTR_ERR(fspath);
1751
1752         if (fspath > fspath_min) {
1753                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
1754                 ++ipath->fspath->elem_cnt;
1755                 ipath->fspath->bytes_left = fspath - fspath_min;
1756         } else {
1757                 ++ipath->fspath->elem_missed;
1758                 ipath->fspath->bytes_missing += fspath_min - fspath;
1759                 ipath->fspath->bytes_left = 0;
1760         }
1761
1762         return 0;
1763 }
1764
1765 /*
1766  * this dumps all file system paths to the inode into the ipath struct, provided
1767  * is has been created large enough. each path is zero-terminated and accessed
1768  * from ipath->fspath->val[i].
1769  * when it returns, there are ipath->fspath->elem_cnt number of paths available
1770  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
1771  * number of missed paths in recored in ipath->fspath->elem_missed, otherwise,
1772  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
1773  * have been needed to return all paths.
1774  */
1775 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
1776 {
1777         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
1778                              inode_to_path, ipath);
1779 }
1780
1781 struct btrfs_data_container *init_data_container(u32 total_bytes)
1782 {
1783         struct btrfs_data_container *data;
1784         size_t alloc_bytes;
1785
1786         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
1787         data = vmalloc(alloc_bytes);
1788         if (!data)
1789                 return ERR_PTR(-ENOMEM);
1790
1791         if (total_bytes >= sizeof(*data)) {
1792                 data->bytes_left = total_bytes - sizeof(*data);
1793                 data->bytes_missing = 0;
1794         } else {
1795                 data->bytes_missing = sizeof(*data) - total_bytes;
1796                 data->bytes_left = 0;
1797         }
1798
1799         data->elem_cnt = 0;
1800         data->elem_missed = 0;
1801
1802         return data;
1803 }
1804
1805 /*
1806  * allocates space to return multiple file system paths for an inode.
1807  * total_bytes to allocate are passed, note that space usable for actual path
1808  * information will be total_bytes - sizeof(struct inode_fs_paths).
1809  * the returned pointer must be freed with free_ipath() in the end.
1810  */
1811 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
1812                                         struct btrfs_path *path)
1813 {
1814         struct inode_fs_paths *ifp;
1815         struct btrfs_data_container *fspath;
1816
1817         fspath = init_data_container(total_bytes);
1818         if (IS_ERR(fspath))
1819                 return (void *)fspath;
1820
1821         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
1822         if (!ifp) {
1823                 kfree(fspath);
1824                 return ERR_PTR(-ENOMEM);
1825         }
1826
1827         ifp->btrfs_path = path;
1828         ifp->fspath = fspath;
1829         ifp->fs_root = fs_root;
1830
1831         return ifp;
1832 }
1833
1834 void free_ipath(struct inode_fs_paths *ipath)
1835 {
1836         if (!ipath)
1837                 return;
1838         vfree(ipath->fspath);
1839         kfree(ipath);
1840 }