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 <linux/rbtree.h>
  21 #include "ctree.h"
  22 #include "disk-io.h"
  23 #include "backref.h"
  24 #include "ulist.h"
  25 #include "transaction.h"
  26 #include "delayed-ref.h"
  27 #include "locking.h"
  28
  29 /* Just an arbitrary number so we can be sure this happened */
  30 #define BACKREF_FOUND_SHARED 6
  31
  32 struct extent_inode_elem {
  33         u64 inum;
  34         u64 offset;
  35         struct extent_inode_elem *next;
  36 };
  37
  38 /*
  39  * ref_root is used as the root of the ref tree that hold a collection
  40  * of unique references.
  41  */
  42 struct ref_root {
  43         struct rb_root rb_root;
  44
  45         /*
  46          * The unique_refs represents the number of ref_nodes with a positive
  47          * count stored in the tree. Even if a ref_node (the count is greater
  48          * than one) is added, the unique_refs will only increase by one.
  49          */
  50         unsigned int unique_refs;
  51 };
  52
  53 /* ref_node is used to store a unique reference to the ref tree. */
  54 struct ref_node {
  55         struct rb_node rb_node;
  56
  57         /* For NORMAL_REF, otherwise all these fields should be set to 0 */
  58         u64 root_id;
  59         u64 object_id;
  60         u64 offset;
  61
  62         /* For SHARED_REF, otherwise parent field should be set to 0 */
  63         u64 parent;
  64
  65         /* Ref to the ref_mod of btrfs_delayed_ref_node */
  66         int ref_mod;
  67 };
  68
  69 /* Dynamically allocate and initialize a ref_root */
  70 static struct ref_root *ref_root_alloc(void)
  71 {
  72         struct ref_root *ref_tree;
  73
  74         ref_tree = kmalloc(sizeof(*ref_tree), GFP_NOFS);
  75         if (!ref_tree)
  76                 return NULL;
  77
  78         ref_tree->rb_root = RB_ROOT;
  79         ref_tree->unique_refs = 0;
  80
  81         return ref_tree;
  82 }
  83
  84 /* Free all nodes in the ref tree, and reinit ref_root */
  85 static void ref_root_fini(struct ref_root *ref_tree)
  86 {
  87         struct ref_node *node;
  88         struct rb_node *next;
  89
  90         while ((next = rb_first(&ref_tree->rb_root)) != NULL) {
  91                 node = rb_entry(next, struct ref_node, rb_node);
  92                 rb_erase(next, &ref_tree->rb_root);
  93                 kfree(node);
  94         }
  95
  96         ref_tree->rb_root = RB_ROOT;
  97         ref_tree->unique_refs = 0;
  98 }
  99
 100 static void ref_root_free(struct ref_root *ref_tree)
 101 {
 102         if (!ref_tree)
 103                 return;
 104
 105         ref_root_fini(ref_tree);
 106         kfree(ref_tree);
 107 }
 108
 109 /*
 110  * Compare ref_node with (root_id, object_id, offset, parent)
 111  *
 112  * The function compares two ref_node a and b. It returns an integer less
 113  * than, equal to, or greater than zero , respectively, to be less than, to
 114  * equal, or be greater than b.
 115  */
 116 static int ref_node_cmp(struct ref_node *a, struct ref_node *b)
 117 {
 118         if (a->root_id < b->root_id)
 119                 return -1;
 120         else if (a->root_id > b->root_id)
 121                 return 1;
 122
 123         if (a->object_id < b->object_id)
 124                 return -1;
 125         else if (a->object_id > b->object_id)
 126                 return 1;
 127
 128         if (a->offset < b->offset)
 129                 return -1;
 130         else if (a->offset > b->offset)
 131                 return 1;
 132
 133         if (a->parent < b->parent)
 134                 return -1;
 135         else if (a->parent > b->parent)
 136                 return 1;
 137
 138         return 0;
 139 }
 140
 141 /*
 142  * Search ref_node with (root_id, object_id, offset, parent) in the tree
 143  *
 144  * if found, the pointer of the ref_node will be returned;
 145  * if not found, NULL will be returned and pos will point to the rb_node for
 146  * insert, pos_parent will point to pos'parent for insert;
 147 */
 148 static struct ref_node *__ref_tree_search(struct ref_root *ref_tree,
 149                                           struct rb_node ***pos,
 150                                           struct rb_node **pos_parent,
 151                                           u64 root_id, u64 object_id,
 152                                           u64 offset, u64 parent)
 153 {
 154         struct ref_node *cur = NULL;
 155         struct ref_node entry;
 156         int ret;
 157
 158         entry.root_id = root_id;
 159         entry.object_id = object_id;
 160         entry.offset = offset;
 161         entry.parent = parent;
 162
 163         *pos = &ref_tree->rb_root.rb_node;
 164
 165         while (**pos) {
 166                 *pos_parent = **pos;
 167                 cur = rb_entry(*pos_parent, struct ref_node, rb_node);
 168
 169                 ret = ref_node_cmp(cur, &entry);
 170                 if (ret > 0)
 171                         *pos = &(**pos)->rb_left;
 172                 else if (ret < 0)
 173                         *pos = &(**pos)->rb_right;
 174                 else
 175                         return cur;
 176         }
 177
 178         return NULL;
 179 }
 180
 181 /*
 182  * Insert a ref_node to the ref tree
 183  * @pos used for specifiy the position to insert
 184  * @pos_parent for specifiy pos's parent
 185  *
 186  * success, return 0;
 187  * ref_node already exists, return -EEXIST;
 188 */
 189 static int ref_tree_insert(struct ref_root *ref_tree, struct rb_node **pos,
 190                            struct rb_node *pos_parent, struct ref_node *ins)
 191 {
 192         struct rb_node **p = NULL;
 193         struct rb_node *parent = NULL;
 194         struct ref_node *cur = NULL;
 195
 196         if (!pos) {
 197                 cur = __ref_tree_search(ref_tree, &p, &parent, ins->root_id,
 198                                         ins->object_id, ins->offset,
 199                                         ins->parent);
 200                 if (cur)
 201                         return -EEXIST;
 202         } else {
 203                 p = pos;
 204                 parent = pos_parent;
 205         }
 206
 207         rb_link_node(&ins->rb_node, parent, p);
 208         rb_insert_color(&ins->rb_node, &ref_tree->rb_root);
 209
 210         return 0;
 211 }
 212
 213 /* Erase and free ref_node, caller should update ref_root->unique_refs */
 214 static void ref_tree_remove(struct ref_root *ref_tree, struct ref_node *node)
 215 {
 216         rb_erase(&node->rb_node, &ref_tree->rb_root);
 217         kfree(node);
 218 }
 219
 220 /*
 221  * Update ref_root->unique_refs
 222  *
 223  * Call __ref_tree_search
 224  *      1. if ref_node doesn't exist, ref_tree_insert this node, and update
 225  *      ref_root->unique_refs:
 226  *              if ref_node->ref_mod > 0, ref_root->unique_refs++;
 227  *              if ref_node->ref_mod < 0, do noting;
 228  *
 229  *      2. if ref_node is found, then get origin ref_node->ref_mod, and update
 230  *      ref_node->ref_mod.
 231  *              if ref_node->ref_mod is equal to 0,then call ref_tree_remove
 232  *
 233  *              according to origin_mod and new_mod, update ref_root->items
 234  *              +----------------+--------------+-------------+
 235  *              |                |new_count <= 0|new_count > 0|
 236  *              +----------------+--------------+-------------+
 237  *              |origin_count < 0|       0      |      1      |
 238  *              +----------------+--------------+-------------+
 239  *              |origin_count > 0|      -1      |      0      |
 240  *              +----------------+--------------+-------------+
 241  *
 242  * In case of allocation failure, -ENOMEM is returned and the ref_tree stays
 243  * unaltered.
 244  * Success, return 0
 245  */
 246 static int ref_tree_add(struct ref_root *ref_tree, u64 root_id, u64 object_id,
 247                         u64 offset, u64 parent, int count)
 248 {
 249         struct ref_node *node = NULL;
 250         struct rb_node **pos = NULL;
 251         struct rb_node *pos_parent = NULL;
 252         int origin_count;
 253         int ret;
 254
 255         if (!count)
 256                 return 0;
 257
 258         node = __ref_tree_search(ref_tree, &pos, &pos_parent, root_id,
 259                                  object_id, offset, parent);
 260         if (node == NULL) {
 261                 node = kmalloc(sizeof(*node), GFP_NOFS);
 262                 if (!node)
 263                         return -ENOMEM;
 264
 265                 node->root_id = root_id;
 266                 node->object_id = object_id;
 267                 node->offset = offset;
 268                 node->parent = parent;
 269                 node->ref_mod = count;
 270
 271                 ret = ref_tree_insert(ref_tree, pos, pos_parent, node);
 272                 ASSERT(!ret);
 273                 if (ret) {
 274                         kfree(node);
 275                         return ret;
 276                 }
 277
 278                 ref_tree->unique_refs += node->ref_mod > 0 ? 1 : 0;
 279
 280                 return 0;
 281         }
 282
 283         origin_count = node->ref_mod;
 284         node->ref_mod += count;
 285
 286         if (node->ref_mod > 0)
 287                 ref_tree->unique_refs += origin_count > 0 ? 0 : 1;
 288         else if (node->ref_mod <= 0)
 289                 ref_tree->unique_refs += origin_count > 0 ? -1 : 0;
 290
 291         if (!node->ref_mod)
 292                 ref_tree_remove(ref_tree, node);
 293
 294         return 0;
 295 }
 296
 297 static int check_extent_in_eb(struct btrfs_key *key, struct extent_buffer *eb,
 298                                 struct btrfs_file_extent_item *fi,
 299                                 u64 extent_item_pos,
 300                                 struct extent_inode_elem **eie)
 301 {
 302         u64 offset = 0;
 303         struct extent_inode_elem *e;
 304
 305         if (!btrfs_file_extent_compression(eb, fi) &&
 306             !btrfs_file_extent_encryption(eb, fi) &&
 307             !btrfs_file_extent_other_encoding(eb, fi)) {
 308                 u64 data_offset;
 309                 u64 data_len;
 310
 311                 data_offset = btrfs_file_extent_offset(eb, fi);
 312                 data_len = btrfs_file_extent_num_bytes(eb, fi);
 313
 314                 if (extent_item_pos < data_offset ||
 315                     extent_item_pos >= data_offset + data_len)
 316                         return 1;
 317                 offset = extent_item_pos - data_offset;
 318         }
 319
 320         e = kmalloc(sizeof(*e), GFP_NOFS);
 321         if (!e)
 322                 return -ENOMEM;
 323
 324         e->next = *eie;
 325         e->inum = key->objectid;
 326         e->offset = key->offset + offset;
 327         *eie = e;
 328
 329         return 0;
 330 }
 331
 332 static void free_inode_elem_list(struct extent_inode_elem *eie)
 333 {
 334         struct extent_inode_elem *eie_next;
 335
 336         for (; eie; eie = eie_next) {
 337                 eie_next = eie->next;
 338                 kfree(eie);
 339         }
 340 }
 341
 342 static int find_extent_in_eb(struct extent_buffer *eb, u64 wanted_disk_byte,
 343                                 u64 extent_item_pos,
 344                                 struct extent_inode_elem **eie)
 345 {
 346         u64 disk_byte;
 347         struct btrfs_key key;
 348         struct btrfs_file_extent_item *fi;
 349         int slot;
 350         int nritems;
 351         int extent_type;
 352         int ret;
 353
 354         /*
 355          * from the shared data ref, we only have the leaf but we need
 356          * the key. thus, we must look into all items and see that we
 357          * find one (some) with a reference to our extent item.
 358          */
 359         nritems = btrfs_header_nritems(eb);
 360         for (slot = 0; slot < nritems; ++slot) {
 361                 btrfs_item_key_to_cpu(eb, &key, slot);
 362                 if (key.type != BTRFS_EXTENT_DATA_KEY)
 363                         continue;
 364                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 365                 extent_type = btrfs_file_extent_type(eb, fi);
 366                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
 367                         continue;
 368                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
 369                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 370                 if (disk_byte != wanted_disk_byte)
 371                         continue;
 372
 373                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
 374                 if (ret < 0)
 375                         return ret;
 376         }
 377
 378         return 0;
 379 }
 380
 381 /*
 382  * this structure records all encountered refs on the way up to the root
 383  */
 384 struct __prelim_ref {
 385         struct list_head list;
 386         u64 root_id;
 387         struct btrfs_key key_for_search;
 388         int level;
 389         int count;
 390         struct extent_inode_elem *inode_list;
 391         u64 parent;
 392         u64 wanted_disk_byte;
 393 };
 394
 395 static struct kmem_cache *btrfs_prelim_ref_cache;
 396
 397 int __init btrfs_prelim_ref_init(void)
 398 {
 399         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
 400                                         sizeof(struct __prelim_ref),
 401                                         0,
 402                                         SLAB_MEM_SPREAD,
 403                                         NULL);
 404         if (!btrfs_prelim_ref_cache)
 405                 return -ENOMEM;
 406         return 0;
 407 }
 408
 409 void btrfs_prelim_ref_exit(void)
 410 {
 411         kmem_cache_destroy(btrfs_prelim_ref_cache);
 412 }
 413
 414 /*
 415  * the rules for all callers of this function are:
 416  * - obtaining the parent is the goal
 417  * - if you add a key, you must know that it is a correct key
 418  * - if you cannot add the parent or a correct key, then we will look into the
 419  *   block later to set a correct key
 420  *
 421  * delayed refs
 422  * ============
 423  *        backref type | shared | indirect | shared | indirect
 424  * information         |   tree |     tree |   data |     data
 425  * --------------------+--------+----------+--------+----------
 426  *      parent logical |    y   |     -    |    -   |     -
 427  *      key to resolve |    -   |     y    |    y   |     y
 428  *  tree block logical |    -   |     -    |    -   |     -
 429  *  root for resolving |    y   |     y    |    y   |     y
 430  *
 431  * - column 1:       we've the parent -> done
 432  * - column 2, 3, 4: we use the key to find the parent
 433  *
 434  * on disk refs (inline or keyed)
 435  * ==============================
 436  *        backref type | shared | indirect | shared | indirect
 437  * information         |   tree |     tree |   data |     data
 438  * --------------------+--------+----------+--------+----------
 439  *      parent logical |    y   |     -    |    y   |     -
 440  *      key to resolve |    -   |     -    |    -   |     y
 441  *  tree block logical |    y   |     y    |    y   |     y
 442  *  root for resolving |    -   |     y    |    y   |     y
 443  *
 444  * - column 1, 3: we've the parent -> done
 445  * - column 2:    we take the first key from the block to find the parent
 446  *                (see __add_missing_keys)
 447  * - column 4:    we use the key to find the parent
 448  *
 449  * additional information that's available but not required to find the parent
 450  * block might help in merging entries to gain some speed.
 451  */
 452
 453 static int __add_prelim_ref(struct list_head *head, u64 root_id,
 454                             struct btrfs_key *key, int level,
 455                             u64 parent, u64 wanted_disk_byte, int count,
 456                             gfp_t gfp_mask)
 457 {
 458         struct __prelim_ref *ref;
 459
 460         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
 461                 return 0;
 462
 463         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
 464         if (!ref)
 465                 return -ENOMEM;
 466
 467         ref->root_id = root_id;
 468         if (key) {
 469                 ref->key_for_search = *key;
 470                 /*
 471                  * We can often find data backrefs with an offset that is too
 472                  * large (>= LLONG_MAX, maximum allowed file offset) due to
 473                  * underflows when subtracting a file's offset with the data
 474                  * offset of its corresponding extent data item. This can
 475                  * happen for example in the clone ioctl.
 476                  * So if we detect such case we set the search key's offset to
 477                  * zero to make sure we will find the matching file extent item
 478                  * at add_all_parents(), otherwise we will miss it because the
 479                  * offset taken form the backref is much larger then the offset
 480                  * of the file extent item. This can make us scan a very large
 481                  * number of file extent items, but at least it will not make
 482                  * us miss any.
 483                  * This is an ugly workaround for a behaviour that should have
 484                  * never existed, but it does and a fix for the clone ioctl
 485                  * would touch a lot of places, cause backwards incompatibility
 486                  * and would not fix the problem for extents cloned with older
 487                  * kernels.
 488                  */
 489                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
 490                     ref->key_for_search.offset >= LLONG_MAX)
 491                         ref->key_for_search.offset = 0;
 492         } else {
 493                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
 494         }
 495
 496         ref->inode_list = NULL;
 497         ref->level = level;
 498         ref->count = count;
 499         ref->parent = parent;
 500         ref->wanted_disk_byte = wanted_disk_byte;
 501         list_add_tail(&ref->list, head);
 502
 503         return 0;
 504 }
 505
 506 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
 507                            struct ulist *parents, struct __prelim_ref *ref,
 508                            int level, u64 time_seq, const u64 *extent_item_pos,
 509                            u64 total_refs)
 510 {
 511         int ret = 0;
 512         int slot;
 513         struct extent_buffer *eb;
 514         struct btrfs_key key;
 515         struct btrfs_key *key_for_search = &ref->key_for_search;
 516         struct btrfs_file_extent_item *fi;
 517         struct extent_inode_elem *eie = NULL, *old = NULL;
 518         u64 disk_byte;
 519         u64 wanted_disk_byte = ref->wanted_disk_byte;
 520         u64 count = 0;
 521
 522         if (level != 0) {
 523                 eb = path->nodes[level];
 524                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
 525                 if (ret < 0)
 526                         return ret;
 527                 return 0;
 528         }
 529
 530         /*
 531          * We normally enter this function with the path already pointing to
 532          * the first item to check. But sometimes, we may enter it with
 533          * slot==nritems. In that case, go to the next leaf before we continue.
 534          */
 535         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 536                 if (time_seq == (u64)-1)
 537                         ret = btrfs_next_leaf(root, path);
 538                 else
 539                         ret = btrfs_next_old_leaf(root, path, time_seq);
 540         }
 541
 542         while (!ret && count < total_refs) {
 543                 eb = path->nodes[0];
 544                 slot = path->slots[0];
 545
 546                 btrfs_item_key_to_cpu(eb, &key, slot);
 547
 548                 if (key.objectid != key_for_search->objectid ||
 549                     key.type != BTRFS_EXTENT_DATA_KEY)
 550                         break;
 551
 552                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 553                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
 554
 555                 if (disk_byte == wanted_disk_byte) {
 556                         eie = NULL;
 557                         old = NULL;
 558                         count++;
 559                         if (extent_item_pos) {
 560                                 ret = check_extent_in_eb(&key, eb, fi,
 561                                                 *extent_item_pos,
 562                                                 &eie);
 563                                 if (ret < 0)
 564                                         break;
 565                         }
 566                         if (ret > 0)
 567                                 goto next;
 568                         ret = ulist_add_merge_ptr(parents, eb->start,
 569                                                   eie, (void **)&old, GFP_NOFS);
 570                         if (ret < 0)
 571                                 break;
 572                         if (!ret && extent_item_pos) {
 573                                 while (old->next)
 574                                         old = old->next;
 575                                 old->next = eie;
 576                         }
 577                         eie = NULL;
 578                 }
 579 next:
 580                 if (time_seq == (u64)-1)
 581                         ret = btrfs_next_item(root, path);
 582                 else
 583                         ret = btrfs_next_old_item(root, path, time_seq);
 584         }
 585
 586         if (ret > 0)
 587                 ret = 0;
 588         else if (ret < 0)
 589                 free_inode_elem_list(eie);
 590         return ret;
 591 }
 592
 593 /*
 594  * resolve an indirect backref in the form (root_id, key, level)
 595  * to a logical address
 596  */
 597 static int __resolve_indirect_ref(struct btrfs_fs_info *fs_info,
 598                                   struct btrfs_path *path, u64 time_seq,
 599                                   struct __prelim_ref *ref,
 600                                   struct ulist *parents,
 601                                   const u64 *extent_item_pos, u64 total_refs)
 602 {
 603         struct btrfs_root *root;
 604         struct btrfs_key root_key;
 605         struct extent_buffer *eb;
 606         int ret = 0;
 607         int root_level;
 608         int level = ref->level;
 609         int index;
 610
 611         root_key.objectid = ref->root_id;
 612         root_key.type = BTRFS_ROOT_ITEM_KEY;
 613         root_key.offset = (u64)-1;
 614
 615         index = srcu_read_lock(&fs_info->subvol_srcu);
 616
 617         root = btrfs_get_fs_root(fs_info, &root_key, false);
 618         if (IS_ERR(root)) {
 619                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 620                 ret = PTR_ERR(root);
 621                 goto out;
 622         }
 623
 624         if (btrfs_is_testing(fs_info)) {
 625                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 626                 ret = -ENOENT;
 627                 goto out;
 628         }
 629
 630         if (path->search_commit_root)
 631                 root_level = btrfs_header_level(root->commit_root);
 632         else if (time_seq == (u64)-1)
 633                 root_level = btrfs_header_level(root->node);
 634         else
 635                 root_level = btrfs_old_root_level(root, time_seq);
 636
 637         if (root_level + 1 == level) {
 638                 srcu_read_unlock(&fs_info->subvol_srcu, index);
 639                 goto out;
 640         }
 641
 642         path->lowest_level = level;
 643         if (time_seq == (u64)-1)
 644                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
 645                                         0, 0);
 646         else
 647                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
 648                                             time_seq);
 649
 650         /* root node has been locked, we can release @subvol_srcu safely here */
 651         srcu_read_unlock(&fs_info->subvol_srcu, index);
 652
 653         btrfs_debug(fs_info,
 654                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
 655                  ref->root_id, level, ref->count, ret,
 656                  ref->key_for_search.objectid, ref->key_for_search.type,
 657                  ref->key_for_search.offset);
 658         if (ret < 0)
 659                 goto out;
 660
 661         eb = path->nodes[level];
 662         while (!eb) {
 663                 if (WARN_ON(!level)) {
 664                         ret = 1;
 665                         goto out;
 666                 }
 667                 level--;
 668                 eb = path->nodes[level];
 669         }
 670
 671         ret = add_all_parents(root, path, parents, ref, level, time_seq,
 672                               extent_item_pos, total_refs);
 673 out:
 674         path->lowest_level = 0;
 675         btrfs_release_path(path);
 676         return ret;
 677 }
 678
 679 /*
 680  * resolve all indirect backrefs from the list
 681  */
 682 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
 683                                    struct btrfs_path *path, u64 time_seq,
 684                                    struct list_head *head,
 685                                    const u64 *extent_item_pos, u64 total_refs,
 686                                    u64 root_objectid)
 687 {
 688         int err;
 689         int ret = 0;
 690         struct __prelim_ref *ref;
 691         struct __prelim_ref *ref_safe;
 692         struct __prelim_ref *new_ref;
 693         struct ulist *parents;
 694         struct ulist_node *node;
 695         struct ulist_iterator uiter;
 696
 697         parents = ulist_alloc(GFP_NOFS);
 698         if (!parents)
 699                 return -ENOMEM;
 700
 701         /*
 702          * _safe allows us to insert directly after the current item without
 703          * iterating over the newly inserted items.
 704          * we're also allowed to re-assign ref during iteration.
 705          */
 706         list_for_each_entry_safe(ref, ref_safe, head, list) {
 707                 if (ref->parent)        /* already direct */
 708                         continue;
 709                 if (ref->count == 0)
 710                         continue;
 711                 if (root_objectid && ref->root_id != root_objectid) {
 712                         ret = BACKREF_FOUND_SHARED;
 713                         goto out;
 714                 }
 715                 err = __resolve_indirect_ref(fs_info, path, time_seq, ref,
 716                                              parents, extent_item_pos,
 717                                              total_refs);
 718                 /*
 719                  * we can only tolerate ENOENT,otherwise,we should catch error
 720                  * and return directly.
 721                  */
 722                 if (err == -ENOENT) {
 723                         continue;
 724                 } else if (err) {
 725                         ret = err;
 726                         goto out;
 727                 }
 728
 729                 /* we put the first parent into the ref at hand */
 730                 ULIST_ITER_INIT(&uiter);
 731                 node = ulist_next(parents, &uiter);
 732                 ref->parent = node ? node->val : 0;
 733                 ref->inode_list = node ?
 734                         (struct extent_inode_elem *)(uintptr_t)node->aux : NULL;
 735
 736                 /* additional parents require new refs being added here */
 737                 while ((node = ulist_next(parents, &uiter))) {
 738                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
 739                                                    GFP_NOFS);
 740                         if (!new_ref) {
 741                                 ret = -ENOMEM;
 742                                 goto out;
 743                         }
 744                         memcpy(new_ref, ref, sizeof(*ref));
 745                         new_ref->parent = node->val;
 746                         new_ref->inode_list = (struct extent_inode_elem *)
 747                                                         (uintptr_t)node->aux;
 748                         list_add(&new_ref->list, &ref->list);
 749                 }
 750                 ulist_reinit(parents);
 751         }
 752 out:
 753         ulist_free(parents);
 754         return ret;
 755 }
 756
 757 static inline int ref_for_same_block(struct __prelim_ref *ref1,
 758                                      struct __prelim_ref *ref2)
 759 {
 760         if (ref1->level != ref2->level)
 761                 return 0;
 762         if (ref1->root_id != ref2->root_id)
 763                 return 0;
 764         if (ref1->key_for_search.type != ref2->key_for_search.type)
 765                 return 0;
 766         if (ref1->key_for_search.objectid != ref2->key_for_search.objectid)
 767                 return 0;
 768         if (ref1->key_for_search.offset != ref2->key_for_search.offset)
 769                 return 0;
 770         if (ref1->parent != ref2->parent)
 771                 return 0;
 772
 773         return 1;
 774 }
 775
 776 /*
 777  * read tree blocks and add keys where required.
 778  */
 779 static int __add_missing_keys(struct btrfs_fs_info *fs_info,
 780                               struct list_head *head)
 781 {
 782         struct __prelim_ref *ref;
 783         struct extent_buffer *eb;
 784
 785         list_for_each_entry(ref, head, list) {
 786                 if (ref->parent)
 787                         continue;
 788                 if (ref->key_for_search.type)
 789                         continue;
 790                 BUG_ON(!ref->wanted_disk_byte);
 791                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0);
 792                 if (IS_ERR(eb)) {
 793                         return PTR_ERR(eb);
 794                 } else if (!extent_buffer_uptodate(eb)) {
 795                         free_extent_buffer(eb);
 796                         return -EIO;
 797                 }
 798                 btrfs_tree_read_lock(eb);
 799                 if (btrfs_header_level(eb) == 0)
 800                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 801                 else
 802                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 803                 btrfs_tree_read_unlock(eb);
 804                 free_extent_buffer(eb);
 805         }
 806         return 0;
 807 }
 808
 809 /*
 810  * merge backrefs and adjust counts accordingly
 811  *
 812  * mode = 1: merge identical keys, if key is set
 813  *    FIXME: if we add more keys in __add_prelim_ref, we can merge more here.
 814  *           additionally, we could even add a key range for the blocks we
 815  *           looked into to merge even more (-> replace unresolved refs by those
 816  *           having a parent).
 817  * mode = 2: merge identical parents
 818  */
 819 static void __merge_refs(struct list_head *head, int mode)
 820 {
 821         struct __prelim_ref *pos1;
 822
 823         list_for_each_entry(pos1, head, list) {
 824                 struct __prelim_ref *pos2 = pos1, *tmp;
 825
 826                 list_for_each_entry_safe_continue(pos2, tmp, head, list) {
 827                         struct __prelim_ref *ref1 = pos1, *ref2 = pos2;
 828                         struct extent_inode_elem *eie;
 829
 830                         if (!ref_for_same_block(ref1, ref2))
 831                                 continue;
 832                         if (mode == 1) {
 833                                 if (!ref1->parent && ref2->parent)
 834                                         swap(ref1, ref2);
 835                         } else {
 836                                 if (ref1->parent != ref2->parent)
 837                                         continue;
 838                         }
 839
 840                         eie = ref1->inode_list;
 841                         while (eie && eie->next)
 842                                 eie = eie->next;
 843                         if (eie)
 844                                 eie->next = ref2->inode_list;
 845                         else
 846                                 ref1->inode_list = ref2->inode_list;
 847                         ref1->count += ref2->count;
 848
 849                         list_del(&ref2->list);
 850                         kmem_cache_free(btrfs_prelim_ref_cache, ref2);
 851                         cond_resched();
 852                 }
 853
 854         }
 855 }
 856
 857 /*
 858  * add all currently queued delayed refs from this head whose seq nr is
 859  * smaller or equal that seq to the list
 860  */
 861 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 862                               struct list_head *prefs, u64 *total_refs,
 863                               u64 inum)
 864 {
 865         struct btrfs_delayed_ref_node *node;
 866         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 867         struct btrfs_key key;
 868         struct btrfs_key op_key = {0};
 869         int sgn;
 870         int ret = 0;
 871
 872         if (extent_op && extent_op->update_key)
 873                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 874
 875         spin_lock(&head->lock);
 876         list_for_each_entry(node, &head->ref_list, list) {
 877                 if (node->seq > seq)
 878                         continue;
 879
 880                 switch (node->action) {
 881                 case BTRFS_ADD_DELAYED_EXTENT:
 882                 case BTRFS_UPDATE_DELAYED_HEAD:
 883                         WARN_ON(1);
 884                         continue;
 885                 case BTRFS_ADD_DELAYED_REF:
 886                         sgn = 1;
 887                         break;
 888                 case BTRFS_DROP_DELAYED_REF:
 889                         sgn = -1;
 890                         break;
 891                 default:
 892                         BUG_ON(1);
 893                 }
 894                 *total_refs += (node->ref_mod * sgn);
 895                 switch (node->type) {
 896                 case BTRFS_TREE_BLOCK_REF_KEY: {
 897                         struct btrfs_delayed_tree_ref *ref;
 898
 899                         ref = btrfs_delayed_node_to_tree_ref(node);
 900                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
 901                                                ref->level + 1, 0, node->bytenr,
 902                                                node->ref_mod * sgn, GFP_ATOMIC);
 903                         break;
 904                 }
 905                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 906                         struct btrfs_delayed_tree_ref *ref;
 907
 908                         ref = btrfs_delayed_node_to_tree_ref(node);
 909                         ret = __add_prelim_ref(prefs, 0, NULL,
 910                                                ref->level + 1, ref->parent,
 911                                                node->bytenr,
 912                                                node->ref_mod * sgn, GFP_ATOMIC);
 913                         break;
 914                 }
 915                 case BTRFS_EXTENT_DATA_REF_KEY: {
 916                         struct btrfs_delayed_data_ref *ref;
 917                         ref = btrfs_delayed_node_to_data_ref(node);
 918
 919                         key.objectid = ref->objectid;
 920                         key.type = BTRFS_EXTENT_DATA_KEY;
 921                         key.offset = ref->offset;
 922
 923                         /*
 924                          * Found a inum that doesn't match our known inum, we
 925                          * know it's shared.
 926                          */
 927                         if (inum && ref->objectid != inum) {
 928                                 ret = BACKREF_FOUND_SHARED;
 929                                 break;
 930                         }
 931
 932                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
 933                                                node->bytenr,
 934                                                node->ref_mod * sgn, GFP_ATOMIC);
 935                         break;
 936                 }
 937                 case BTRFS_SHARED_DATA_REF_KEY: {
 938                         struct btrfs_delayed_data_ref *ref;
 939
 940                         ref = btrfs_delayed_node_to_data_ref(node);
 941                         ret = __add_prelim_ref(prefs, 0, NULL, 0,
 942                                                ref->parent, node->bytenr,
 943                                                node->ref_mod * sgn, GFP_ATOMIC);
 944                         break;
 945                 }
 946                 default:
 947                         WARN_ON(1);
 948                 }
 949                 if (ret)
 950                         break;
 951         }
 952         spin_unlock(&head->lock);
 953         return ret;
 954 }
 955
 956 /*
 957  * add all inline backrefs for bytenr to the list
 958  */
 959 static int __add_inline_refs(struct btrfs_path *path, u64 bytenr,
 960                              int *info_level, struct list_head *prefs,
 961                              struct ref_root *ref_tree,
 962                              u64 *total_refs, u64 inum)
 963 {
 964         int ret = 0;
 965         int slot;
 966         struct extent_buffer *leaf;
 967         struct btrfs_key key;
 968         struct btrfs_key found_key;
 969         unsigned long ptr;
 970         unsigned long end;
 971         struct btrfs_extent_item *ei;
 972         u64 flags;
 973         u64 item_size;
 974
 975         /*
 976          * enumerate all inline refs
 977          */
 978         leaf = path->nodes[0];
 979         slot = path->slots[0];
 980
 981         item_size = btrfs_item_size_nr(leaf, slot);
 982         BUG_ON(item_size < sizeof(*ei));
 983
 984         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 985         flags = btrfs_extent_flags(leaf, ei);
 986         *total_refs += btrfs_extent_refs(leaf, ei);
 987         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 988
 989         ptr = (unsigned long)(ei + 1);
 990         end = (unsigned long)ei + item_size;
 991
 992         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 993             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 994                 struct btrfs_tree_block_info *info;
 995
 996                 info = (struct btrfs_tree_block_info *)ptr;
 997                 *info_level = btrfs_tree_block_level(leaf, info);
 998                 ptr += sizeof(struct btrfs_tree_block_info);
 999                 BUG_ON(ptr > end);
1000         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1001                 *info_level = found_key.offset;
1002         } else {
1003                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1004         }
1005
1006         while (ptr < end) {
1007                 struct btrfs_extent_inline_ref *iref;
1008                 u64 offset;
1009                 int type;
1010
1011                 iref = (struct btrfs_extent_inline_ref *)ptr;
1012                 type = btrfs_extent_inline_ref_type(leaf, iref);
1013                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1014
1015                 switch (type) {
1016                 case BTRFS_SHARED_BLOCK_REF_KEY:
1017                         ret = __add_prelim_ref(prefs, 0, NULL,
1018                                                 *info_level + 1, offset,
1019                                                 bytenr, 1, GFP_NOFS);
1020                         break;
1021                 case BTRFS_SHARED_DATA_REF_KEY: {
1022                         struct btrfs_shared_data_ref *sdref;
1023                         int count;
1024
1025                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1026                         count = btrfs_shared_data_ref_count(leaf, sdref);
1027                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
1028                                                bytenr, count, GFP_NOFS);
1029                         if (ref_tree) {
1030                                 if (!ret)
1031                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1032                                                            bytenr, count);
1033                                 if (!ret && ref_tree->unique_refs > 1)
1034                                         ret = BACKREF_FOUND_SHARED;
1035                         }
1036                         break;
1037                 }
1038                 case BTRFS_TREE_BLOCK_REF_KEY:
1039                         ret = __add_prelim_ref(prefs, offset, NULL,
1040                                                *info_level + 1, 0,
1041                                                bytenr, 1, GFP_NOFS);
1042                         break;
1043                 case BTRFS_EXTENT_DATA_REF_KEY: {
1044                         struct btrfs_extent_data_ref *dref;
1045                         int count;
1046                         u64 root;
1047
1048                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1049                         count = btrfs_extent_data_ref_count(leaf, dref);
1050                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1051                                                                       dref);
1052                         key.type = BTRFS_EXTENT_DATA_KEY;
1053                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1054
1055                         if (inum && key.objectid != inum) {
1056                                 ret = BACKREF_FOUND_SHARED;
1057                                 break;
1058                         }
1059
1060                         root = btrfs_extent_data_ref_root(leaf, dref);
1061                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1062                                                bytenr, count, GFP_NOFS);
1063                         if (ref_tree) {
1064                                 if (!ret)
1065                                         ret = ref_tree_add(ref_tree, root,
1066                                                            key.objectid,
1067                                                            key.offset, 0,
1068                                                            count);
1069                                 if (!ret && ref_tree->unique_refs > 1)
1070                                         ret = BACKREF_FOUND_SHARED;
1071                         }
1072                         break;
1073                 }
1074                 default:
1075                         WARN_ON(1);
1076                 }
1077                 if (ret)
1078                         return ret;
1079                 ptr += btrfs_extent_inline_ref_size(type);
1080         }
1081
1082         return 0;
1083 }
1084
1085 /*
1086  * add all non-inline backrefs for bytenr to the list
1087  */
1088 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
1089                             struct btrfs_path *path, u64 bytenr,
1090                             int info_level, struct list_head *prefs,
1091                             struct ref_root *ref_tree, u64 inum)
1092 {
1093         struct btrfs_root *extent_root = fs_info->extent_root;
1094         int ret;
1095         int slot;
1096         struct extent_buffer *leaf;
1097         struct btrfs_key key;
1098
1099         while (1) {
1100                 ret = btrfs_next_item(extent_root, path);
1101                 if (ret < 0)
1102                         break;
1103                 if (ret) {
1104                         ret = 0;
1105                         break;
1106                 }
1107
1108                 slot = path->slots[0];
1109                 leaf = path->nodes[0];
1110                 btrfs_item_key_to_cpu(leaf, &key, slot);
1111
1112                 if (key.objectid != bytenr)
1113                         break;
1114                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1115                         continue;
1116                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1117                         break;
1118
1119                 switch (key.type) {
1120                 case BTRFS_SHARED_BLOCK_REF_KEY:
1121                         ret = __add_prelim_ref(prefs, 0, NULL,
1122                                                 info_level + 1, key.offset,
1123                                                 bytenr, 1, GFP_NOFS);
1124                         break;
1125                 case BTRFS_SHARED_DATA_REF_KEY: {
1126                         struct btrfs_shared_data_ref *sdref;
1127                         int count;
1128
1129                         sdref = btrfs_item_ptr(leaf, slot,
1130                                               struct btrfs_shared_data_ref);
1131                         count = btrfs_shared_data_ref_count(leaf, sdref);
1132                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
1133                                                 bytenr, count, GFP_NOFS);
1134                         if (ref_tree) {
1135                                 if (!ret)
1136                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1137                                                            bytenr, count);
1138                                 if (!ret && ref_tree->unique_refs > 1)
1139                                         ret = BACKREF_FOUND_SHARED;
1140                         }
1141                         break;
1142                 }
1143                 case BTRFS_TREE_BLOCK_REF_KEY:
1144                         ret = __add_prelim_ref(prefs, key.offset, NULL,
1145                                                info_level + 1, 0,
1146                                                bytenr, 1, GFP_NOFS);
1147                         break;
1148                 case BTRFS_EXTENT_DATA_REF_KEY: {
1149                         struct btrfs_extent_data_ref *dref;
1150                         int count;
1151                         u64 root;
1152
1153                         dref = btrfs_item_ptr(leaf, slot,
1154                                               struct btrfs_extent_data_ref);
1155                         count = btrfs_extent_data_ref_count(leaf, dref);
1156                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1157                                                                       dref);
1158                         key.type = BTRFS_EXTENT_DATA_KEY;
1159                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1160
1161                         if (inum && key.objectid != inum) {
1162                                 ret = BACKREF_FOUND_SHARED;
1163                                 break;
1164                         }
1165
1166                         root = btrfs_extent_data_ref_root(leaf, dref);
1167                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1168                                                bytenr, count, GFP_NOFS);
1169                         if (ref_tree) {
1170                                 if (!ret)
1171                                         ret = ref_tree_add(ref_tree, root,
1172                                                            key.objectid,
1173                                                            key.offset, 0,
1174                                                            count);
1175                                 if (!ret && ref_tree->unique_refs > 1)
1176                                         ret = BACKREF_FOUND_SHARED;
1177                         }
1178                         break;
1179                 }
1180                 default:
1181                         WARN_ON(1);
1182                 }
1183                 if (ret)
1184                         return ret;
1185
1186         }
1187
1188         return ret;
1189 }
1190
1191 /*
1192  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1193  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1194  * indirect refs to their parent bytenr.
1195  * When roots are found, they're added to the roots list
1196  *
1197  * NOTE: This can return values > 0
1198  *
1199  * If time_seq is set to (u64)-1, it will not search delayed_refs, and behave
1200  * much like trans == NULL case, the difference only lies in it will not
1201  * commit root.
1202  * The special case is for qgroup to search roots in commit_transaction().
1203  *
1204  * If check_shared is set to 1, any extent has more than one ref item, will
1205  * be returned BACKREF_FOUND_SHARED immediately.
1206  *
1207  * FIXME some caching might speed things up
1208  */
1209 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1210                              struct btrfs_fs_info *fs_info, u64 bytenr,
1211                              u64 time_seq, struct ulist *refs,
1212                              struct ulist *roots, const u64 *extent_item_pos,
1213                              u64 root_objectid, u64 inum, int check_shared)
1214 {
1215         struct btrfs_key key;
1216         struct btrfs_path *path;
1217         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1218         struct btrfs_delayed_ref_head *head;
1219         int info_level = 0;
1220         int ret;
1221         struct list_head prefs_delayed;
1222         struct list_head prefs;
1223         struct __prelim_ref *ref;
1224         struct extent_inode_elem *eie = NULL;
1225         struct ref_root *ref_tree = NULL;
1226         u64 total_refs = 0;
1227
1228         INIT_LIST_HEAD(&prefs);
1229         INIT_LIST_HEAD(&prefs_delayed);
1230
1231         key.objectid = bytenr;
1232         key.offset = (u64)-1;
1233         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1234                 key.type = BTRFS_METADATA_ITEM_KEY;
1235         else
1236                 key.type = BTRFS_EXTENT_ITEM_KEY;
1237
1238         path = btrfs_alloc_path();
1239         if (!path)
1240                 return -ENOMEM;
1241         if (!trans) {
1242                 path->search_commit_root = 1;
1243                 path->skip_locking = 1;
1244         }
1245
1246         if (time_seq == (u64)-1)
1247                 path->skip_locking = 1;
1248
1249         /*
1250          * grab both a lock on the path and a lock on the delayed ref head.
1251          * We need both to get a consistent picture of how the refs look
1252          * at a specified point in time
1253          */
1254 again:
1255         head = NULL;
1256
1257         if (check_shared) {
1258                 if (!ref_tree) {
1259                         ref_tree = ref_root_alloc();
1260                         if (!ref_tree) {
1261                                 ret = -ENOMEM;
1262                                 goto out;
1263                         }
1264                 } else {
1265                         ref_root_fini(ref_tree);
1266                 }
1267         }
1268
1269         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1270         if (ret < 0)
1271                 goto out;
1272         BUG_ON(ret == 0);
1273
1274 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1275         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1276             time_seq != (u64)-1) {
1277 #else
1278         if (trans && time_seq != (u64)-1) {
1279 #endif
1280                 /*
1281                  * look if there are updates for this ref queued and lock the
1282                  * head
1283                  */
1284                 delayed_refs = &trans->transaction->delayed_refs;
1285                 spin_lock(&delayed_refs->lock);
1286                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1287                 if (head) {
1288                         if (!mutex_trylock(&head->mutex)) {
1289                                 atomic_inc(&head->node.refs);
1290                                 spin_unlock(&delayed_refs->lock);
1291
1292                                 btrfs_release_path(path);
1293
1294                                 /*
1295                                  * Mutex was contended, block until it's
1296                                  * released and try again
1297                                  */
1298                                 mutex_lock(&head->mutex);
1299                                 mutex_unlock(&head->mutex);
1300                                 btrfs_put_delayed_ref(&head->node);
1301                                 goto again;
1302                         }
1303                         spin_unlock(&delayed_refs->lock);
1304                         ret = __add_delayed_refs(head, time_seq,
1305                                                  &prefs_delayed, &total_refs,
1306                                                  inum);
1307                         mutex_unlock(&head->mutex);
1308                         if (ret)
1309                                 goto out;
1310                 } else {
1311                         spin_unlock(&delayed_refs->lock);
1312                 }
1313
1314                 if (check_shared && !list_empty(&prefs_delayed)) {
1315                         /*
1316                          * Add all delay_ref to the ref_tree and check if there
1317                          * are multiple ref items added.
1318                          */
1319                         list_for_each_entry(ref, &prefs_delayed, list) {
1320                                 if (ref->key_for_search.type) {
1321                                         ret = ref_tree_add(ref_tree,
1322                                                 ref->root_id,
1323                                                 ref->key_for_search.objectid,
1324                                                 ref->key_for_search.offset,
1325                                                 0, ref->count);
1326                                         if (ret)
1327                                                 goto out;
1328                                 } else {
1329                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1330                                                      ref->parent, ref->count);
1331                                         if (ret)
1332                                                 goto out;
1333                                 }
1334
1335                         }
1336
1337                         if (ref_tree->unique_refs > 1) {
1338                                 ret = BACKREF_FOUND_SHARED;
1339                                 goto out;
1340                         }
1341
1342                 }
1343         }
1344
1345         if (path->slots[0]) {
1346                 struct extent_buffer *leaf;
1347                 int slot;
1348
1349                 path->slots[0]--;
1350                 leaf = path->nodes[0];
1351                 slot = path->slots[0];
1352                 btrfs_item_key_to_cpu(leaf, &key, slot);
1353                 if (key.objectid == bytenr &&
1354                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1355                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1356                         ret = __add_inline_refs(path, bytenr,
1357                                                 &info_level, &prefs,
1358                                                 ref_tree, &total_refs,
1359                                                 inum);
1360                         if (ret)
1361                                 goto out;
1362                         ret = __add_keyed_refs(fs_info, path, bytenr,
1363                                                info_level, &prefs,
1364                                                ref_tree, inum);
1365                         if (ret)
1366                                 goto out;
1367                 }
1368         }
1369         btrfs_release_path(path);
1370
1371         list_splice_init(&prefs_delayed, &prefs);
1372
1373         ret = __add_missing_keys(fs_info, &prefs);
1374         if (ret)
1375                 goto out;
1376
1377         __merge_refs(&prefs, 1);
1378
1379         ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
1380                                       extent_item_pos, total_refs,
1381                                       root_objectid);
1382         if (ret)
1383                 goto out;
1384
1385         __merge_refs(&prefs, 2);
1386
1387         while (!list_empty(&prefs)) {
1388                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1389                 WARN_ON(ref->count < 0);
1390                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1391                         if (root_objectid && ref->root_id != root_objectid) {
1392                                 ret = BACKREF_FOUND_SHARED;
1393                                 goto out;
1394                         }
1395
1396                         /* no parent == root of tree */
1397                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1398                         if (ret < 0)
1399                                 goto out;
1400                 }
1401                 if (ref->count && ref->parent) {
1402                         if (extent_item_pos && !ref->inode_list &&
1403                             ref->level == 0) {
1404                                 struct extent_buffer *eb;
1405
1406                                 eb = read_tree_block(fs_info, ref->parent, 0);
1407                                 if (IS_ERR(eb)) {
1408                                         ret = PTR_ERR(eb);
1409                                         goto out;
1410                                 } else if (!extent_buffer_uptodate(eb)) {
1411                                         free_extent_buffer(eb);
1412                                         ret = -EIO;
1413                                         goto out;
1414                                 }
1415                                 btrfs_tree_read_lock(eb);
1416                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1417                                 ret = find_extent_in_eb(eb, bytenr,
1418                                                         *extent_item_pos, &eie);
1419                                 btrfs_tree_read_unlock_blocking(eb);
1420                                 free_extent_buffer(eb);
1421                                 if (ret < 0)
1422                                         goto out;
1423                                 ref->inode_list = eie;
1424                         }
1425                         ret = ulist_add_merge_ptr(refs, ref->parent,
1426                                                   ref->inode_list,
1427                                                   (void **)&eie, GFP_NOFS);
1428                         if (ret < 0)
1429                                 goto out;
1430                         if (!ret && extent_item_pos) {
1431                                 /*
1432                                  * we've recorded that parent, so we must extend
1433                                  * its inode list here
1434                                  */
1435                                 BUG_ON(!eie);
1436                                 while (eie->next)
1437                                         eie = eie->next;
1438                                 eie->next = ref->inode_list;
1439                         }
1440                         eie = NULL;
1441                 }
1442                 list_del(&ref->list);
1443                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1444         }
1445
1446 out:
1447         btrfs_free_path(path);
1448         ref_root_free(ref_tree);
1449         while (!list_empty(&prefs)) {
1450                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1451                 list_del(&ref->list);
1452                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1453         }
1454         while (!list_empty(&prefs_delayed)) {
1455                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
1456                                        list);
1457                 list_del(&ref->list);
1458                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1459         }
1460         if (ret < 0)
1461                 free_inode_elem_list(eie);
1462         return ret;
1463 }
1464
1465 static void free_leaf_list(struct ulist *blocks)
1466 {
1467         struct ulist_node *node = NULL;
1468         struct extent_inode_elem *eie;
1469         struct ulist_iterator uiter;
1470
1471         ULIST_ITER_INIT(&uiter);
1472         while ((node = ulist_next(blocks, &uiter))) {
1473                 if (!node->aux)
1474                         continue;
1475                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
1476                 free_inode_elem_list(eie);
1477                 node->aux = 0;
1478         }
1479
1480         ulist_free(blocks);
1481 }
1482
1483 /*
1484  * Finds all leafs with a reference to the specified combination of bytenr and
1485  * offset. key_list_head will point to a list of corresponding keys (caller must
1486  * free each list element). The leafs will be stored in the leafs ulist, which
1487  * must be freed with ulist_free.
1488  *
1489  * returns 0 on success, <0 on error
1490  */
1491 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1492                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1493                                 u64 time_seq, struct ulist **leafs,
1494                                 const u64 *extent_item_pos)
1495 {
1496         int ret;
1497
1498         *leafs = ulist_alloc(GFP_NOFS);
1499         if (!*leafs)
1500                 return -ENOMEM;
1501
1502         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1503                                 *leafs, NULL, extent_item_pos, 0, 0, 0);
1504         if (ret < 0 && ret != -ENOENT) {
1505                 free_leaf_list(*leafs);
1506                 return ret;
1507         }
1508
1509         return 0;
1510 }
1511
1512 /*
1513  * walk all backrefs for a given extent to find all roots that reference this
1514  * extent. Walking a backref means finding all extents that reference this
1515  * extent and in turn walk the backrefs of those, too. Naturally this is a
1516  * recursive process, but here it is implemented in an iterative fashion: We
1517  * find all referencing extents for the extent in question and put them on a
1518  * list. In turn, we find all referencing extents for those, further appending
1519  * to the list. The way we iterate the list allows adding more elements after
1520  * the current while iterating. The process stops when we reach the end of the
1521  * list. Found roots are added to the roots list.
1522  *
1523  * returns 0 on success, < 0 on error.
1524  */
1525 static int __btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1526                                   struct btrfs_fs_info *fs_info, u64 bytenr,
1527                                   u64 time_seq, struct ulist **roots)
1528 {
1529         struct ulist *tmp;
1530         struct ulist_node *node = NULL;
1531         struct ulist_iterator uiter;
1532         int ret;
1533
1534         tmp = ulist_alloc(GFP_NOFS);
1535         if (!tmp)
1536                 return -ENOMEM;
1537         *roots = ulist_alloc(GFP_NOFS);
1538         if (!*roots) {
1539                 ulist_free(tmp);
1540                 return -ENOMEM;
1541         }
1542
1543         ULIST_ITER_INIT(&uiter);
1544         while (1) {
1545                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1546                                         tmp, *roots, NULL, 0, 0, 0);
1547                 if (ret < 0 && ret != -ENOENT) {
1548                         ulist_free(tmp);
1549                         ulist_free(*roots);
1550                         return ret;
1551                 }
1552                 node = ulist_next(tmp, &uiter);
1553                 if (!node)
1554                         break;
1555                 bytenr = node->val;
1556                 cond_resched();
1557         }
1558
1559         ulist_free(tmp);
1560         return 0;
1561 }
1562
1563 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1564                          struct btrfs_fs_info *fs_info, u64 bytenr,
1565                          u64 time_seq, struct ulist **roots)
1566 {
1567         int ret;
1568
1569         if (!trans)
1570                 down_read(&fs_info->commit_root_sem);
1571         ret = __btrfs_find_all_roots(trans, fs_info, bytenr, time_seq, roots);
1572         if (!trans)
1573                 up_read(&fs_info->commit_root_sem);
1574         return ret;
1575 }
1576
1577 /**
1578  * btrfs_check_shared - tell us whether an extent is shared
1579  *
1580  * @trans: optional trans handle
1581  *
1582  * btrfs_check_shared uses the backref walking code but will short
1583  * circuit as soon as it finds a root or inode that doesn't match the
1584  * one passed in. This provides a significant performance benefit for
1585  * callers (such as fiemap) which want to know whether the extent is
1586  * shared but do not need a ref count.
1587  *
1588  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1589  */
1590 int btrfs_check_shared(struct btrfs_trans_handle *trans,
1591                        struct btrfs_fs_info *fs_info, u64 root_objectid,
1592                        u64 inum, u64 bytenr)
1593 {
1594         struct ulist *tmp = NULL;
1595         struct ulist *roots = NULL;
1596         struct ulist_iterator uiter;
1597         struct ulist_node *node;
1598         struct seq_list elem = SEQ_LIST_INIT(elem);
1599         int ret = 0;
1600
1601         tmp = ulist_alloc(GFP_NOFS);
1602         roots = ulist_alloc(GFP_NOFS);
1603         if (!tmp || !roots) {
1604                 ulist_free(tmp);
1605                 ulist_free(roots);
1606                 return -ENOMEM;
1607         }
1608
1609         if (trans)
1610                 btrfs_get_tree_mod_seq(fs_info, &elem);
1611         else
1612                 down_read(&fs_info->commit_root_sem);
1613         ULIST_ITER_INIT(&uiter);
1614         while (1) {
1615                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1616                                         roots, NULL, root_objectid, inum, 1);
1617                 if (ret == BACKREF_FOUND_SHARED) {
1618                         /* this is the only condition under which we return 1 */
1619                         ret = 1;
1620                         break;
1621                 }
1622                 if (ret < 0 && ret != -ENOENT)
1623                         break;
1624                 ret = 0;
1625                 node = ulist_next(tmp, &uiter);
1626                 if (!node)
1627                         break;
1628                 bytenr = node->val;
1629                 cond_resched();
1630         }
1631         if (trans)
1632                 btrfs_put_tree_mod_seq(fs_info, &elem);
1633         else
1634                 up_read(&fs_info->commit_root_sem);
1635         ulist_free(tmp);
1636         ulist_free(roots);
1637         return ret;
1638 }
1639
1640 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1641                           u64 start_off, struct btrfs_path *path,
1642                           struct btrfs_inode_extref **ret_extref,
1643                           u64 *found_off)
1644 {
1645         int ret, slot;
1646         struct btrfs_key key;
1647         struct btrfs_key found_key;
1648         struct btrfs_inode_extref *extref;
1649         struct extent_buffer *leaf;
1650         unsigned long ptr;
1651
1652         key.objectid = inode_objectid;
1653         key.type = BTRFS_INODE_EXTREF_KEY;
1654         key.offset = start_off;
1655
1656         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1657         if (ret < 0)
1658                 return ret;
1659
1660         while (1) {
1661                 leaf = path->nodes[0];
1662                 slot = path->slots[0];
1663                 if (slot >= btrfs_header_nritems(leaf)) {
1664                         /*
1665                          * If the item at offset is not found,
1666                          * btrfs_search_slot will point us to the slot
1667                          * where it should be inserted. In our case
1668                          * that will be the slot directly before the
1669                          * next INODE_REF_KEY_V2 item. In the case
1670                          * that we're pointing to the last slot in a
1671                          * leaf, we must move one leaf over.
1672                          */
1673                         ret = btrfs_next_leaf(root, path);
1674                         if (ret) {
1675                                 if (ret >= 1)
1676                                         ret = -ENOENT;
1677                                 break;
1678                         }
1679                         continue;
1680                 }
1681
1682                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1683
1684                 /*
1685                  * Check that we're still looking at an extended ref key for
1686                  * this particular objectid. If we have different
1687                  * objectid or type then there are no more to be found
1688                  * in the tree and we can exit.
1689                  */
1690                 ret = -ENOENT;
1691                 if (found_key.objectid != inode_objectid)
1692                         break;
1693                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1694                         break;
1695
1696                 ret = 0;
1697                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1698                 extref = (struct btrfs_inode_extref *)ptr;
1699                 *ret_extref = extref;
1700                 if (found_off)
1701                         *found_off = found_key.offset;
1702                 break;
1703         }
1704
1705         return ret;
1706 }
1707
1708 /*
1709  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1710  * Elements of the path are separated by '/' and the path is guaranteed to be
1711  * 0-terminated. the path is only given within the current file system.
1712  * Therefore, it never starts with a '/'. the caller is responsible to provide
1713  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1714  * the start point of the resulting string is returned. this pointer is within
1715  * dest, normally.
1716  * in case the path buffer would overflow, the pointer is decremented further
1717  * as if output was written to the buffer, though no more output is actually
1718  * generated. that way, the caller can determine how much space would be
1719  * required for the path to fit into the buffer. in that case, the returned
1720  * value will be smaller than dest. callers must check this!
1721  */
1722 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1723                         u32 name_len, unsigned long name_off,
1724                         struct extent_buffer *eb_in, u64 parent,
1725                         char *dest, u32 size)
1726 {
1727         int slot;
1728         u64 next_inum;
1729         int ret;
1730         s64 bytes_left = ((s64)size) - 1;
1731         struct extent_buffer *eb = eb_in;
1732         struct btrfs_key found_key;
1733         int leave_spinning = path->leave_spinning;
1734         struct btrfs_inode_ref *iref;
1735
1736         if (bytes_left >= 0)
1737                 dest[bytes_left] = '\0';
1738
1739         path->leave_spinning = 1;
1740         while (1) {
1741                 bytes_left -= name_len;
1742                 if (bytes_left >= 0)
1743                         read_extent_buffer(eb, dest + bytes_left,
1744                                            name_off, name_len);
1745                 if (eb != eb_in) {
1746                         if (!path->skip_locking)
1747                                 btrfs_tree_read_unlock_blocking(eb);
1748                         free_extent_buffer(eb);
1749                 }
1750                 ret = btrfs_find_item(fs_root, path, parent, 0,
1751                                 BTRFS_INODE_REF_KEY, &found_key);
1752                 if (ret > 0)
1753                         ret = -ENOENT;
1754                 if (ret)
1755                         break;
1756
1757                 next_inum = found_key.offset;
1758
1759                 /* regular exit ahead */
1760                 if (parent == next_inum)
1761                         break;
1762
1763                 slot = path->slots[0];
1764                 eb = path->nodes[0];
1765                 /* make sure we can use eb after releasing the path */
1766                 if (eb != eb_in) {
1767                         if (!path->skip_locking)
1768                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1769                         path->nodes[0] = NULL;
1770                         path->locks[0] = 0;
1771                 }
1772                 btrfs_release_path(path);
1773                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1774
1775                 name_len = btrfs_inode_ref_name_len(eb, iref);
1776                 name_off = (unsigned long)(iref + 1);
1777
1778                 parent = next_inum;
1779                 --bytes_left;
1780                 if (bytes_left >= 0)
1781                         dest[bytes_left] = '/';
1782         }
1783
1784         btrfs_release_path(path);
1785         path->leave_spinning = leave_spinning;
1786
1787         if (ret)
1788                 return ERR_PTR(ret);
1789
1790         return dest + bytes_left;
1791 }
1792
1793 /*
1794  * this makes the path point to (logical EXTENT_ITEM *)
1795  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1796  * tree blocks and <0 on error.
1797  */
1798 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1799                         struct btrfs_path *path, struct btrfs_key *found_key,
1800                         u64 *flags_ret)
1801 {
1802         int ret;
1803         u64 flags;
1804         u64 size = 0;
1805         u32 item_size;
1806         struct extent_buffer *eb;
1807         struct btrfs_extent_item *ei;
1808         struct btrfs_key key;
1809
1810         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1811                 key.type = BTRFS_METADATA_ITEM_KEY;
1812         else
1813                 key.type = BTRFS_EXTENT_ITEM_KEY;
1814         key.objectid = logical;
1815         key.offset = (u64)-1;
1816
1817         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1818         if (ret < 0)
1819                 return ret;
1820
1821         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1822         if (ret) {
1823                 if (ret > 0)
1824                         ret = -ENOENT;
1825                 return ret;
1826         }
1827         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1828         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1829                 size = fs_info->nodesize;
1830         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1831                 size = found_key->offset;
1832
1833         if (found_key->objectid > logical ||
1834             found_key->objectid + size <= logical) {
1835                 btrfs_debug(fs_info,
1836                         "logical %llu is not within any extent", logical);
1837                 return -ENOENT;
1838         }
1839
1840         eb = path->nodes[0];
1841         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1842         BUG_ON(item_size < sizeof(*ei));
1843
1844         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1845         flags = btrfs_extent_flags(eb, ei);
1846
1847         btrfs_debug(fs_info,
1848                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1849                  logical, logical - found_key->objectid, found_key->objectid,
1850                  found_key->offset, flags, item_size);
1851
1852         WARN_ON(!flags_ret);
1853         if (flags_ret) {
1854                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1855                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1856                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1857                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1858                 else
1859                         BUG_ON(1);
1860                 return 0;
1861         }
1862
1863         return -EIO;
1864 }
1865
1866 /*
1867  * helper function to iterate extent inline refs. ptr must point to a 0 value
1868  * for the first call and may be modified. it is used to track state.
1869  * if more refs exist, 0 is returned and the next call to
1870  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1871  * next ref. after the last ref was processed, 1 is returned.
1872  * returns <0 on error
1873  */
1874 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1875                                    struct btrfs_key *key,
1876                                    struct btrfs_extent_item *ei, u32 item_size,
1877                                    struct btrfs_extent_inline_ref **out_eiref,
1878                                    int *out_type)
1879 {
1880         unsigned long end;
1881         u64 flags;
1882         struct btrfs_tree_block_info *info;
1883
1884         if (!*ptr) {
1885                 /* first call */
1886                 flags = btrfs_extent_flags(eb, ei);
1887                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1888                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1889                                 /* a skinny metadata extent */
1890                                 *out_eiref =
1891                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1892                         } else {
1893                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1894                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1895                                 *out_eiref =
1896                                    (struct btrfs_extent_inline_ref *)(info + 1);
1897                         }
1898                 } else {
1899                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1900                 }
1901                 *ptr = (unsigned long)*out_eiref;
1902                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1903                         return -ENOENT;
1904         }
1905
1906         end = (unsigned long)ei + item_size;
1907         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1908         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1909
1910         *ptr += btrfs_extent_inline_ref_size(*out_type);
1911         WARN_ON(*ptr > end);
1912         if (*ptr == end)
1913                 return 1; /* last */
1914
1915         return 0;
1916 }
1917
1918 /*
1919  * reads the tree block backref for an extent. tree level and root are returned
1920  * through out_level and out_root. ptr must point to a 0 value for the first
1921  * call and may be modified (see __get_extent_inline_ref comment).
1922  * returns 0 if data was provided, 1 if there was no more data to provide or
1923  * <0 on error.
1924  */
1925 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1926                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1927                             u32 item_size, u64 *out_root, u8 *out_level)
1928 {
1929         int ret;
1930         int type;
1931         struct btrfs_extent_inline_ref *eiref;
1932
1933         if (*ptr == (unsigned long)-1)
1934                 return 1;
1935
1936         while (1) {
1937                 ret = __get_extent_inline_ref(ptr, eb, key, ei, item_size,
1938                                               &eiref, &type);
1939                 if (ret < 0)
1940                         return ret;
1941
1942                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1943                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1944                         break;
1945
1946                 if (ret == 1)
1947                         return 1;
1948         }
1949
1950         /* we can treat both ref types equally here */
1951         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1952
1953         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1954                 struct btrfs_tree_block_info *info;
1955
1956                 info = (struct btrfs_tree_block_info *)(ei + 1);
1957                 *out_level = btrfs_tree_block_level(eb, info);
1958         } else {
1959                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1960                 *out_level = (u8)key->offset;
1961         }
1962
1963         if (ret == 1)
1964                 *ptr = (unsigned long)-1;
1965
1966         return 0;
1967 }
1968
1969 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1970                              struct extent_inode_elem *inode_list,
1971                              u64 root, u64 extent_item_objectid,
1972                              iterate_extent_inodes_t *iterate, void *ctx)
1973 {
1974         struct extent_inode_elem *eie;
1975         int ret = 0;
1976
1977         for (eie = inode_list; eie; eie = eie->next) {
1978                 btrfs_debug(fs_info,
1979                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1980                             extent_item_objectid, eie->inum,
1981                             eie->offset, root);
1982                 ret = iterate(eie->inum, eie->offset, root, ctx);
1983                 if (ret) {
1984                         btrfs_debug(fs_info,
1985                                     "stopping iteration for %llu due to ret=%d",
1986                                     extent_item_objectid, ret);
1987                         break;
1988                 }
1989         }
1990
1991         return ret;
1992 }
1993
1994 /*
1995  * calls iterate() for every inode that references the extent identified by
1996  * the given parameters.
1997  * when the iterator function returns a non-zero value, iteration stops.
1998  */
1999 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
2000                                 u64 extent_item_objectid, u64 extent_item_pos,
2001                                 int search_commit_root,
2002                                 iterate_extent_inodes_t *iterate, void *ctx)
2003 {
2004         int ret;
2005         struct btrfs_trans_handle *trans = NULL;
2006         struct ulist *refs = NULL;
2007         struct ulist *roots = NULL;
2008         struct ulist_node *ref_node = NULL;
2009         struct ulist_node *root_node = NULL;
2010         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
2011         struct ulist_iterator ref_uiter;
2012         struct ulist_iterator root_uiter;
2013
2014         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
2015                         extent_item_objectid);
2016
2017         if (!search_commit_root) {
2018                 trans = btrfs_join_transaction(fs_info->extent_root);
2019                 if (IS_ERR(trans))
2020                         return PTR_ERR(trans);
2021                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2022         } else {
2023                 down_read(&fs_info->commit_root_sem);
2024         }
2025
2026         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2027                                    tree_mod_seq_elem.seq, &refs,
2028                                    &extent_item_pos);
2029         if (ret)
2030                 goto out;
2031
2032         ULIST_ITER_INIT(&ref_uiter);
2033         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2034                 ret = __btrfs_find_all_roots(trans, fs_info, ref_node->val,
2035                                              tree_mod_seq_elem.seq, &roots);
2036                 if (ret)
2037                         break;
2038                 ULIST_ITER_INIT(&root_uiter);
2039                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2040                         btrfs_debug(fs_info,
2041                                     "root %llu references leaf %llu, data list %#llx",
2042                                     root_node->val, ref_node->val,
2043                                     ref_node->aux);
2044                         ret = iterate_leaf_refs(fs_info,
2045                                                 (struct extent_inode_elem *)
2046                                                 (uintptr_t)ref_node->aux,
2047                                                 root_node->val,
2048                                                 extent_item_objectid,
2049                                                 iterate, ctx);
2050                 }
2051                 ulist_free(roots);
2052         }
2053
2054         free_leaf_list(refs);
2055 out:
2056         if (!search_commit_root) {
2057                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2058                 btrfs_end_transaction(trans);
2059         } else {
2060                 up_read(&fs_info->commit_root_sem);
2061         }
2062
2063         return ret;
2064 }
2065
2066 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2067                                 struct btrfs_path *path,
2068                                 iterate_extent_inodes_t *iterate, void *ctx)
2069 {
2070         int ret;
2071         u64 extent_item_pos;
2072         u64 flags = 0;
2073         struct btrfs_key found_key;
2074         int search_commit_root = path->search_commit_root;
2075
2076         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2077         btrfs_release_path(path);
2078         if (ret < 0)
2079                 return ret;
2080         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2081                 return -EINVAL;
2082
2083         extent_item_pos = logical - found_key.objectid;
2084         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2085                                         extent_item_pos, search_commit_root,
2086                                         iterate, ctx);
2087
2088         return ret;
2089 }
2090
2091 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2092                               struct extent_buffer *eb, void *ctx);
2093
2094 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2095                               struct btrfs_path *path,
2096                               iterate_irefs_t *iterate, void *ctx)
2097 {
2098         int ret = 0;
2099         int slot;
2100         u32 cur;
2101         u32 len;
2102         u32 name_len;
2103         u64 parent = 0;
2104         int found = 0;
2105         struct extent_buffer *eb;
2106         struct btrfs_item *item;
2107         struct btrfs_inode_ref *iref;
2108         struct btrfs_key found_key;
2109
2110         while (!ret) {
2111                 ret = btrfs_find_item(fs_root, path, inum,
2112                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2113                                 &found_key);
2114
2115                 if (ret < 0)
2116                         break;
2117                 if (ret) {
2118                         ret = found ? 0 : -ENOENT;
2119                         break;
2120                 }
2121                 ++found;
2122
2123                 parent = found_key.offset;
2124                 slot = path->slots[0];
2125                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2126                 if (!eb) {
2127                         ret = -ENOMEM;
2128                         break;
2129                 }
2130                 extent_buffer_get(eb);
2131                 btrfs_tree_read_lock(eb);
2132                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2133                 btrfs_release_path(path);
2134
2135                 item = btrfs_item_nr(slot);
2136                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2137
2138                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2139                         name_len = btrfs_inode_ref_name_len(eb, iref);
2140                         /* path must be released before calling iterate()! */
2141                         btrfs_debug(fs_root->fs_info,
2142                                 "following ref at offset %u for inode %llu in tree %llu",
2143                                 cur, found_key.objectid, fs_root->objectid);
2144                         ret = iterate(parent, name_len,
2145                                       (unsigned long)(iref + 1), eb, ctx);
2146                         if (ret)
2147                                 break;
2148                         len = sizeof(*iref) + name_len;
2149                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2150                 }
2151                 btrfs_tree_read_unlock_blocking(eb);
2152                 free_extent_buffer(eb);
2153         }
2154
2155         btrfs_release_path(path);
2156
2157         return ret;
2158 }
2159
2160 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2161                                  struct btrfs_path *path,
2162                                  iterate_irefs_t *iterate, void *ctx)
2163 {
2164         int ret;
2165         int slot;
2166         u64 offset = 0;
2167         u64 parent;
2168         int found = 0;
2169         struct extent_buffer *eb;
2170         struct btrfs_inode_extref *extref;
2171         u32 item_size;
2172         u32 cur_offset;
2173         unsigned long ptr;
2174
2175         while (1) {
2176                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2177                                             &offset);
2178                 if (ret < 0)
2179                         break;
2180                 if (ret) {
2181                         ret = found ? 0 : -ENOENT;
2182                         break;
2183                 }
2184                 ++found;
2185
2186                 slot = path->slots[0];
2187                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2188                 if (!eb) {
2189                         ret = -ENOMEM;
2190                         break;
2191                 }
2192                 extent_buffer_get(eb);
2193
2194                 btrfs_tree_read_lock(eb);
2195                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2196                 btrfs_release_path(path);
2197
2198                 item_size = btrfs_item_size_nr(eb, slot);
2199                 ptr = btrfs_item_ptr_offset(eb, slot);
2200                 cur_offset = 0;
2201
2202                 while (cur_offset < item_size) {
2203                         u32 name_len;
2204
2205                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2206                         parent = btrfs_inode_extref_parent(eb, extref);
2207                         name_len = btrfs_inode_extref_name_len(eb, extref);
2208                         ret = iterate(parent, name_len,
2209                                       (unsigned long)&extref->name, eb, ctx);
2210                         if (ret)
2211                                 break;
2212
2213                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2214                         cur_offset += sizeof(*extref);
2215                 }
2216                 btrfs_tree_read_unlock_blocking(eb);
2217                 free_extent_buffer(eb);
2218
2219                 offset++;
2220         }
2221
2222         btrfs_release_path(path);
2223
2224         return ret;
2225 }
2226
2227 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2228                          struct btrfs_path *path, iterate_irefs_t *iterate,
2229                          void *ctx)
2230 {
2231         int ret;
2232         int found_refs = 0;
2233
2234         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2235         if (!ret)
2236                 ++found_refs;
2237         else if (ret != -ENOENT)
2238                 return ret;
2239
2240         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2241         if (ret == -ENOENT && found_refs)
2242                 return 0;
2243
2244         return ret;
2245 }
2246
2247 /*
2248  * returns 0 if the path could be dumped (probably truncated)
2249  * returns <0 in case of an error
2250  */
2251 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2252                          struct extent_buffer *eb, void *ctx)
2253 {
2254         struct inode_fs_paths *ipath = ctx;
2255         char *fspath;
2256         char *fspath_min;
2257         int i = ipath->fspath->elem_cnt;
2258         const int s_ptr = sizeof(char *);
2259         u32 bytes_left;
2260
2261         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2262                                         ipath->fspath->bytes_left - s_ptr : 0;
2263
2264         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2265         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2266                                    name_off, eb, inum, fspath_min, bytes_left);
2267         if (IS_ERR(fspath))
2268                 return PTR_ERR(fspath);
2269
2270         if (fspath > fspath_min) {
2271                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2272                 ++ipath->fspath->elem_cnt;
2273                 ipath->fspath->bytes_left = fspath - fspath_min;
2274         } else {
2275                 ++ipath->fspath->elem_missed;
2276                 ipath->fspath->bytes_missing += fspath_min - fspath;
2277                 ipath->fspath->bytes_left = 0;
2278         }
2279
2280         return 0;
2281 }
2282
2283 /*
2284  * this dumps all file system paths to the inode into the ipath struct, provided
2285  * is has been created large enough. each path is zero-terminated and accessed
2286  * from ipath->fspath->val[i].
2287  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2288  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2289  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2290  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2291  * have been needed to return all paths.
2292  */
2293 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2294 {
2295         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2296                              inode_to_path, ipath);
2297 }
2298
2299 struct btrfs_data_container *init_data_container(u32 total_bytes)
2300 {
2301         struct btrfs_data_container *data;
2302         size_t alloc_bytes;
2303
2304         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2305         data = vmalloc(alloc_bytes);
2306         if (!data)
2307                 return ERR_PTR(-ENOMEM);
2308
2309         if (total_bytes >= sizeof(*data)) {
2310                 data->bytes_left = total_bytes - sizeof(*data);
2311                 data->bytes_missing = 0;
2312         } else {
2313                 data->bytes_missing = sizeof(*data) - total_bytes;
2314                 data->bytes_left = 0;
2315         }
2316
2317         data->elem_cnt = 0;
2318         data->elem_missed = 0;
2319
2320         return data;
2321 }
2322
2323 /*
2324  * allocates space to return multiple file system paths for an inode.
2325  * total_bytes to allocate are passed, note that space usable for actual path
2326  * information will be total_bytes - sizeof(struct inode_fs_paths).
2327  * the returned pointer must be freed with free_ipath() in the end.
2328  */
2329 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2330                                         struct btrfs_path *path)
2331 {
2332         struct inode_fs_paths *ifp;
2333         struct btrfs_data_container *fspath;
2334
2335         fspath = init_data_container(total_bytes);
2336         if (IS_ERR(fspath))
2337                 return (void *)fspath;
2338
2339         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
2340         if (!ifp) {
2341                 vfree(fspath);
2342                 return ERR_PTR(-ENOMEM);
2343         }
2344
2345         ifp->btrfs_path = path;
2346         ifp->fspath = fspath;
2347         ifp->fs_root = fs_root;
2348
2349         return ifp;
2350 }
2351
2352 void free_ipath(struct inode_fs_paths *ipath)
2353 {
2354         if (!ipath)
2355                 return;
2356         vfree(ipath->fspath);
2357         kfree(ipath);
2358 }