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->tree_root, ref->wanted_disk_byte,
 792                                      0);
 793                 if (IS_ERR(eb)) {
 794                         return PTR_ERR(eb);
 795                 } else if (!extent_buffer_uptodate(eb)) {
 796                         free_extent_buffer(eb);
 797                         return -EIO;
 798                 }
 799                 btrfs_tree_read_lock(eb);
 800                 if (btrfs_header_level(eb) == 0)
 801                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
 802                 else
 803                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
 804                 btrfs_tree_read_unlock(eb);
 805                 free_extent_buffer(eb);
 806         }
 807         return 0;
 808 }
 809
 810 /*
 811  * merge backrefs and adjust counts accordingly
 812  *
 813  * mode = 1: merge identical keys, if key is set
 814  *    FIXME: if we add more keys in __add_prelim_ref, we can merge more here.
 815  *           additionally, we could even add a key range for the blocks we
 816  *           looked into to merge even more (-> replace unresolved refs by those
 817  *           having a parent).
 818  * mode = 2: merge identical parents
 819  */
 820 static void __merge_refs(struct list_head *head, int mode)
 821 {
 822         struct __prelim_ref *pos1;
 823
 824         list_for_each_entry(pos1, head, list) {
 825                 struct __prelim_ref *pos2 = pos1, *tmp;
 826
 827                 list_for_each_entry_safe_continue(pos2, tmp, head, list) {
 828                         struct __prelim_ref *ref1 = pos1, *ref2 = pos2;
 829                         struct extent_inode_elem *eie;
 830
 831                         if (!ref_for_same_block(ref1, ref2))
 832                                 continue;
 833                         if (mode == 1) {
 834                                 if (!ref1->parent && ref2->parent)
 835                                         swap(ref1, ref2);
 836                         } else {
 837                                 if (ref1->parent != ref2->parent)
 838                                         continue;
 839                         }
 840
 841                         eie = ref1->inode_list;
 842                         while (eie && eie->next)
 843                                 eie = eie->next;
 844                         if (eie)
 845                                 eie->next = ref2->inode_list;
 846                         else
 847                                 ref1->inode_list = ref2->inode_list;
 848                         ref1->count += ref2->count;
 849
 850                         list_del(&ref2->list);
 851                         kmem_cache_free(btrfs_prelim_ref_cache, ref2);
 852                         cond_resched();
 853                 }
 854
 855         }
 856 }
 857
 858 /*
 859  * add all currently queued delayed refs from this head whose seq nr is
 860  * smaller or equal that seq to the list
 861  */
 862 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 863                               struct list_head *prefs, u64 *total_refs,
 864                               u64 inum)
 865 {
 866         struct btrfs_delayed_ref_node *node;
 867         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
 868         struct btrfs_key key;
 869         struct btrfs_key op_key = {0};
 870         int sgn;
 871         int ret = 0;
 872
 873         if (extent_op && extent_op->update_key)
 874                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 875
 876         spin_lock(&head->lock);
 877         list_for_each_entry(node, &head->ref_list, list) {
 878                 if (node->seq > seq)
 879                         continue;
 880
 881                 switch (node->action) {
 882                 case BTRFS_ADD_DELAYED_EXTENT:
 883                 case BTRFS_UPDATE_DELAYED_HEAD:
 884                         WARN_ON(1);
 885                         continue;
 886                 case BTRFS_ADD_DELAYED_REF:
 887                         sgn = 1;
 888                         break;
 889                 case BTRFS_DROP_DELAYED_REF:
 890                         sgn = -1;
 891                         break;
 892                 default:
 893                         BUG_ON(1);
 894                 }
 895                 *total_refs += (node->ref_mod * sgn);
 896                 switch (node->type) {
 897                 case BTRFS_TREE_BLOCK_REF_KEY: {
 898                         struct btrfs_delayed_tree_ref *ref;
 899
 900                         ref = btrfs_delayed_node_to_tree_ref(node);
 901                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
 902                                                ref->level + 1, 0, node->bytenr,
 903                                                node->ref_mod * sgn, GFP_ATOMIC);
 904                         break;
 905                 }
 906                 case BTRFS_SHARED_BLOCK_REF_KEY: {
 907                         struct btrfs_delayed_tree_ref *ref;
 908
 909                         ref = btrfs_delayed_node_to_tree_ref(node);
 910                         ret = __add_prelim_ref(prefs, 0, NULL,
 911                                                ref->level + 1, ref->parent,
 912                                                node->bytenr,
 913                                                node->ref_mod * sgn, GFP_ATOMIC);
 914                         break;
 915                 }
 916                 case BTRFS_EXTENT_DATA_REF_KEY: {
 917                         struct btrfs_delayed_data_ref *ref;
 918                         ref = btrfs_delayed_node_to_data_ref(node);
 919
 920                         key.objectid = ref->objectid;
 921                         key.type = BTRFS_EXTENT_DATA_KEY;
 922                         key.offset = ref->offset;
 923
 924                         /*
 925                          * Found a inum that doesn't match our known inum, we
 926                          * know it's shared.
 927                          */
 928                         if (inum && ref->objectid != inum) {
 929                                 ret = BACKREF_FOUND_SHARED;
 930                                 break;
 931                         }
 932
 933                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
 934                                                node->bytenr,
 935                                                node->ref_mod * sgn, GFP_ATOMIC);
 936                         break;
 937                 }
 938                 case BTRFS_SHARED_DATA_REF_KEY: {
 939                         struct btrfs_delayed_data_ref *ref;
 940
 941                         ref = btrfs_delayed_node_to_data_ref(node);
 942                         ret = __add_prelim_ref(prefs, 0, NULL, 0,
 943                                                ref->parent, node->bytenr,
 944                                                node->ref_mod * sgn, GFP_ATOMIC);
 945                         break;
 946                 }
 947                 default:
 948                         WARN_ON(1);
 949                 }
 950                 if (ret)
 951                         break;
 952         }
 953         spin_unlock(&head->lock);
 954         return ret;
 955 }
 956
 957 /*
 958  * add all inline backrefs for bytenr to the list
 959  */
 960 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
 961                              struct btrfs_path *path, u64 bytenr,
 962                              int *info_level, struct list_head *prefs,
 963                              struct ref_root *ref_tree,
 964                              u64 *total_refs, u64 inum)
 965 {
 966         int ret = 0;
 967         int slot;
 968         struct extent_buffer *leaf;
 969         struct btrfs_key key;
 970         struct btrfs_key found_key;
 971         unsigned long ptr;
 972         unsigned long end;
 973         struct btrfs_extent_item *ei;
 974         u64 flags;
 975         u64 item_size;
 976
 977         /*
 978          * enumerate all inline refs
 979          */
 980         leaf = path->nodes[0];
 981         slot = path->slots[0];
 982
 983         item_size = btrfs_item_size_nr(leaf, slot);
 984         BUG_ON(item_size < sizeof(*ei));
 985
 986         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 987         flags = btrfs_extent_flags(leaf, ei);
 988         *total_refs += btrfs_extent_refs(leaf, ei);
 989         btrfs_item_key_to_cpu(leaf, &found_key, slot);
 990
 991         ptr = (unsigned long)(ei + 1);
 992         end = (unsigned long)ei + item_size;
 993
 994         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
 995             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 996                 struct btrfs_tree_block_info *info;
 997
 998                 info = (struct btrfs_tree_block_info *)ptr;
 999                 *info_level = btrfs_tree_block_level(leaf, info);
1000                 ptr += sizeof(struct btrfs_tree_block_info);
1001                 BUG_ON(ptr > end);
1002         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1003                 *info_level = found_key.offset;
1004         } else {
1005                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1006         }
1007
1008         while (ptr < end) {
1009                 struct btrfs_extent_inline_ref *iref;
1010                 u64 offset;
1011                 int type;
1012
1013                 iref = (struct btrfs_extent_inline_ref *)ptr;
1014                 type = btrfs_extent_inline_ref_type(leaf, iref);
1015                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1016
1017                 switch (type) {
1018                 case BTRFS_SHARED_BLOCK_REF_KEY:
1019                         ret = __add_prelim_ref(prefs, 0, NULL,
1020                                                 *info_level + 1, offset,
1021                                                 bytenr, 1, GFP_NOFS);
1022                         break;
1023                 case BTRFS_SHARED_DATA_REF_KEY: {
1024                         struct btrfs_shared_data_ref *sdref;
1025                         int count;
1026
1027                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1028                         count = btrfs_shared_data_ref_count(leaf, sdref);
1029                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
1030                                                bytenr, count, GFP_NOFS);
1031                         if (ref_tree) {
1032                                 if (!ret)
1033                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1034                                                            bytenr, count);
1035                                 if (!ret && ref_tree->unique_refs > 1)
1036                                         ret = BACKREF_FOUND_SHARED;
1037                         }
1038                         break;
1039                 }
1040                 case BTRFS_TREE_BLOCK_REF_KEY:
1041                         ret = __add_prelim_ref(prefs, offset, NULL,
1042                                                *info_level + 1, 0,
1043                                                bytenr, 1, GFP_NOFS);
1044                         break;
1045                 case BTRFS_EXTENT_DATA_REF_KEY: {
1046                         struct btrfs_extent_data_ref *dref;
1047                         int count;
1048                         u64 root;
1049
1050                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1051                         count = btrfs_extent_data_ref_count(leaf, dref);
1052                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1053                                                                       dref);
1054                         key.type = BTRFS_EXTENT_DATA_KEY;
1055                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1056
1057                         if (inum && key.objectid != inum) {
1058                                 ret = BACKREF_FOUND_SHARED;
1059                                 break;
1060                         }
1061
1062                         root = btrfs_extent_data_ref_root(leaf, dref);
1063                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1064                                                bytenr, count, GFP_NOFS);
1065                         if (ref_tree) {
1066                                 if (!ret)
1067                                         ret = ref_tree_add(ref_tree, root,
1068                                                            key.objectid,
1069                                                            key.offset, 0,
1070                                                            count);
1071                                 if (!ret && ref_tree->unique_refs > 1)
1072                                         ret = BACKREF_FOUND_SHARED;
1073                         }
1074                         break;
1075                 }
1076                 default:
1077                         WARN_ON(1);
1078                 }
1079                 if (ret)
1080                         return ret;
1081                 ptr += btrfs_extent_inline_ref_size(type);
1082         }
1083
1084         return 0;
1085 }
1086
1087 /*
1088  * add all non-inline backrefs for bytenr to the list
1089  */
1090 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
1091                             struct btrfs_path *path, u64 bytenr,
1092                             int info_level, struct list_head *prefs,
1093                             struct ref_root *ref_tree, u64 inum)
1094 {
1095         struct btrfs_root *extent_root = fs_info->extent_root;
1096         int ret;
1097         int slot;
1098         struct extent_buffer *leaf;
1099         struct btrfs_key key;
1100
1101         while (1) {
1102                 ret = btrfs_next_item(extent_root, path);
1103                 if (ret < 0)
1104                         break;
1105                 if (ret) {
1106                         ret = 0;
1107                         break;
1108                 }
1109
1110                 slot = path->slots[0];
1111                 leaf = path->nodes[0];
1112                 btrfs_item_key_to_cpu(leaf, &key, slot);
1113
1114                 if (key.objectid != bytenr)
1115                         break;
1116                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1117                         continue;
1118                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1119                         break;
1120
1121                 switch (key.type) {
1122                 case BTRFS_SHARED_BLOCK_REF_KEY:
1123                         ret = __add_prelim_ref(prefs, 0, NULL,
1124                                                 info_level + 1, key.offset,
1125                                                 bytenr, 1, GFP_NOFS);
1126                         break;
1127                 case BTRFS_SHARED_DATA_REF_KEY: {
1128                         struct btrfs_shared_data_ref *sdref;
1129                         int count;
1130
1131                         sdref = btrfs_item_ptr(leaf, slot,
1132                                               struct btrfs_shared_data_ref);
1133                         count = btrfs_shared_data_ref_count(leaf, sdref);
1134                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
1135                                                 bytenr, count, GFP_NOFS);
1136                         if (ref_tree) {
1137                                 if (!ret)
1138                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1139                                                            bytenr, count);
1140                                 if (!ret && ref_tree->unique_refs > 1)
1141                                         ret = BACKREF_FOUND_SHARED;
1142                         }
1143                         break;
1144                 }
1145                 case BTRFS_TREE_BLOCK_REF_KEY:
1146                         ret = __add_prelim_ref(prefs, key.offset, NULL,
1147                                                info_level + 1, 0,
1148                                                bytenr, 1, GFP_NOFS);
1149                         break;
1150                 case BTRFS_EXTENT_DATA_REF_KEY: {
1151                         struct btrfs_extent_data_ref *dref;
1152                         int count;
1153                         u64 root;
1154
1155                         dref = btrfs_item_ptr(leaf, slot,
1156                                               struct btrfs_extent_data_ref);
1157                         count = btrfs_extent_data_ref_count(leaf, dref);
1158                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1159                                                                       dref);
1160                         key.type = BTRFS_EXTENT_DATA_KEY;
1161                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1162
1163                         if (inum && key.objectid != inum) {
1164                                 ret = BACKREF_FOUND_SHARED;
1165                                 break;
1166                         }
1167
1168                         root = btrfs_extent_data_ref_root(leaf, dref);
1169                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1170                                                bytenr, count, GFP_NOFS);
1171                         if (ref_tree) {
1172                                 if (!ret)
1173                                         ret = ref_tree_add(ref_tree, root,
1174                                                            key.objectid,
1175                                                            key.offset, 0,
1176                                                            count);
1177                                 if (!ret && ref_tree->unique_refs > 1)
1178                                         ret = BACKREF_FOUND_SHARED;
1179                         }
1180                         break;
1181                 }
1182                 default:
1183                         WARN_ON(1);
1184                 }
1185                 if (ret)
1186                         return ret;
1187
1188         }
1189
1190         return ret;
1191 }
1192
1193 /*
1194  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1195  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1196  * indirect refs to their parent bytenr.
1197  * When roots are found, they're added to the roots list
1198  *
1199  * NOTE: This can return values > 0
1200  *
1201  * If time_seq is set to (u64)-1, it will not search delayed_refs, and behave
1202  * much like trans == NULL case, the difference only lies in it will not
1203  * commit root.
1204  * The special case is for qgroup to search roots in commit_transaction().
1205  *
1206  * If check_shared is set to 1, any extent has more than one ref item, will
1207  * be returned BACKREF_FOUND_SHARED immediately.
1208  *
1209  * FIXME some caching might speed things up
1210  */
1211 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1212                              struct btrfs_fs_info *fs_info, u64 bytenr,
1213                              u64 time_seq, struct ulist *refs,
1214                              struct ulist *roots, const u64 *extent_item_pos,
1215                              u64 root_objectid, u64 inum, int check_shared)
1216 {
1217         struct btrfs_key key;
1218         struct btrfs_path *path;
1219         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1220         struct btrfs_delayed_ref_head *head;
1221         int info_level = 0;
1222         int ret;
1223         struct list_head prefs_delayed;
1224         struct list_head prefs;
1225         struct __prelim_ref *ref;
1226         struct extent_inode_elem *eie = NULL;
1227         struct ref_root *ref_tree = NULL;
1228         u64 total_refs = 0;
1229
1230         INIT_LIST_HEAD(&prefs);
1231         INIT_LIST_HEAD(&prefs_delayed);
1232
1233         key.objectid = bytenr;
1234         key.offset = (u64)-1;
1235         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1236                 key.type = BTRFS_METADATA_ITEM_KEY;
1237         else
1238                 key.type = BTRFS_EXTENT_ITEM_KEY;
1239
1240         path = btrfs_alloc_path();
1241         if (!path)
1242                 return -ENOMEM;
1243         if (!trans) {
1244                 path->search_commit_root = 1;
1245                 path->skip_locking = 1;
1246         }
1247
1248         if (time_seq == (u64)-1)
1249                 path->skip_locking = 1;
1250
1251         /*
1252          * grab both a lock on the path and a lock on the delayed ref head.
1253          * We need both to get a consistent picture of how the refs look
1254          * at a specified point in time
1255          */
1256 again:
1257         head = NULL;
1258
1259         if (check_shared) {
1260                 if (!ref_tree) {
1261                         ref_tree = ref_root_alloc();
1262                         if (!ref_tree) {
1263                                 ret = -ENOMEM;
1264                                 goto out;
1265                         }
1266                 } else {
1267                         ref_root_fini(ref_tree);
1268                 }
1269         }
1270
1271         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1272         if (ret < 0)
1273                 goto out;
1274         if (ret == 0) {
1275                 /* This shouldn't happen, indicates a bug or fs corruption. */
1276                 ASSERT(ret != 0);
1277                 ret = -EUCLEAN;
1278                 goto out;
1279         }
1280
1281 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1282         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1283             time_seq != (u64)-1) {
1284 #else
1285         if (trans && time_seq != (u64)-1) {
1286 #endif
1287                 /*
1288                  * look if there are updates for this ref queued and lock the
1289                  * head
1290                  */
1291                 delayed_refs = &trans->transaction->delayed_refs;
1292                 spin_lock(&delayed_refs->lock);
1293                 head = btrfs_find_delayed_ref_head(trans, bytenr);
1294                 if (head) {
1295                         if (!mutex_trylock(&head->mutex)) {
1296                                 atomic_inc(&head->node.refs);
1297                                 spin_unlock(&delayed_refs->lock);
1298
1299                                 btrfs_release_path(path);
1300
1301                                 /*
1302                                  * Mutex was contended, block until it's
1303                                  * released and try again
1304                                  */
1305                                 mutex_lock(&head->mutex);
1306                                 mutex_unlock(&head->mutex);
1307                                 btrfs_put_delayed_ref(&head->node);
1308                                 goto again;
1309                         }
1310                         spin_unlock(&delayed_refs->lock);
1311                         ret = __add_delayed_refs(head, time_seq,
1312                                                  &prefs_delayed, &total_refs,
1313                                                  inum);
1314                         mutex_unlock(&head->mutex);
1315                         if (ret)
1316                                 goto out;
1317                 } else {
1318                         spin_unlock(&delayed_refs->lock);
1319                 }
1320
1321                 if (check_shared && !list_empty(&prefs_delayed)) {
1322                         /*
1323                          * Add all delay_ref to the ref_tree and check if there
1324                          * are multiple ref items added.
1325                          */
1326                         list_for_each_entry(ref, &prefs_delayed, list) {
1327                                 if (ref->key_for_search.type) {
1328                                         ret = ref_tree_add(ref_tree,
1329                                                 ref->root_id,
1330                                                 ref->key_for_search.objectid,
1331                                                 ref->key_for_search.offset,
1332                                                 0, ref->count);
1333                                         if (ret)
1334                                                 goto out;
1335                                 } else {
1336                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1337                                                      ref->parent, ref->count);
1338                                         if (ret)
1339                                                 goto out;
1340                                 }
1341
1342                         }
1343
1344                         if (ref_tree->unique_refs > 1) {
1345                                 ret = BACKREF_FOUND_SHARED;
1346                                 goto out;
1347                         }
1348
1349                 }
1350         }
1351
1352         if (path->slots[0]) {
1353                 struct extent_buffer *leaf;
1354                 int slot;
1355
1356                 path->slots[0]--;
1357                 leaf = path->nodes[0];
1358                 slot = path->slots[0];
1359                 btrfs_item_key_to_cpu(leaf, &key, slot);
1360                 if (key.objectid == bytenr &&
1361                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1362                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1363                         ret = __add_inline_refs(fs_info, path, bytenr,
1364                                                 &info_level, &prefs,
1365                                                 ref_tree, &total_refs,
1366                                                 inum);
1367                         if (ret)
1368                                 goto out;
1369                         ret = __add_keyed_refs(fs_info, path, bytenr,
1370                                                info_level, &prefs,
1371                                                ref_tree, inum);
1372                         if (ret)
1373                                 goto out;
1374                 }
1375         }
1376         btrfs_release_path(path);
1377
1378         list_splice_init(&prefs_delayed, &prefs);
1379
1380         ret = __add_missing_keys(fs_info, &prefs);
1381         if (ret)
1382                 goto out;
1383
1384         __merge_refs(&prefs, 1);
1385
1386         ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
1387                                       extent_item_pos, total_refs,
1388                                       root_objectid);
1389         if (ret)
1390                 goto out;
1391
1392         __merge_refs(&prefs, 2);
1393
1394         while (!list_empty(&prefs)) {
1395                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1396                 WARN_ON(ref->count < 0);
1397                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1398                         if (root_objectid && ref->root_id != root_objectid) {
1399                                 ret = BACKREF_FOUND_SHARED;
1400                                 goto out;
1401                         }
1402
1403                         /* no parent == root of tree */
1404                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1405                         if (ret < 0)
1406                                 goto out;
1407                 }
1408                 if (ref->count && ref->parent) {
1409                         if (extent_item_pos && !ref->inode_list &&
1410                             ref->level == 0) {
1411                                 struct extent_buffer *eb;
1412
1413                                 eb = read_tree_block(fs_info->extent_root,
1414                                                            ref->parent, 0);
1415                                 if (IS_ERR(eb)) {
1416                                         ret = PTR_ERR(eb);
1417                                         goto out;
1418                                 } else if (!extent_buffer_uptodate(eb)) {
1419                                         free_extent_buffer(eb);
1420                                         ret = -EIO;
1421                                         goto out;
1422                                 }
1423                                 btrfs_tree_read_lock(eb);
1424                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1425                                 ret = find_extent_in_eb(eb, bytenr,
1426                                                         *extent_item_pos, &eie);
1427                                 btrfs_tree_read_unlock_blocking(eb);
1428                                 free_extent_buffer(eb);
1429                                 if (ret < 0)
1430                                         goto out;
1431                                 ref->inode_list = eie;
1432                         }
1433                         ret = ulist_add_merge_ptr(refs, ref->parent,
1434                                                   ref->inode_list,
1435                                                   (void **)&eie, GFP_NOFS);
1436                         if (ret < 0)
1437                                 goto out;
1438                         if (!ret && extent_item_pos) {
1439                                 /*
1440                                  * We've recorded that parent, so we must extend
1441                                  * its inode list here.
1442                                  *
1443                                  * However if there was corruption we may not
1444                                  * have found an eie, return an error in this
1445                                  * case.
1446                                  */
1447                                 ASSERT(eie);
1448                                 if (!eie) {
1449                                         ret = -EUCLEAN;
1450                                         goto out;
1451                                 }
1452                                 while (eie->next)
1453                                         eie = eie->next;
1454                                 eie->next = ref->inode_list;
1455                         }
1456                         eie = NULL;
1457                 }
1458                 list_del(&ref->list);
1459                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1460         }
1461
1462 out:
1463         btrfs_free_path(path);
1464         ref_root_free(ref_tree);
1465         while (!list_empty(&prefs)) {
1466                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1467                 list_del(&ref->list);
1468                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1469         }
1470         while (!list_empty(&prefs_delayed)) {
1471                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
1472                                        list);
1473                 list_del(&ref->list);
1474                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1475         }
1476         if (ret < 0)
1477                 free_inode_elem_list(eie);
1478         return ret;
1479 }
1480
1481 static void free_leaf_list(struct ulist *blocks)
1482 {
1483         struct ulist_node *node = NULL;
1484         struct extent_inode_elem *eie;
1485         struct ulist_iterator uiter;
1486
1487         ULIST_ITER_INIT(&uiter);
1488         while ((node = ulist_next(blocks, &uiter))) {
1489                 if (!node->aux)
1490                         continue;
1491                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
1492                 free_inode_elem_list(eie);
1493                 node->aux = 0;
1494         }
1495
1496         ulist_free(blocks);
1497 }
1498
1499 /*
1500  * Finds all leafs with a reference to the specified combination of bytenr and
1501  * offset. key_list_head will point to a list of corresponding keys (caller must
1502  * free each list element). The leafs will be stored in the leafs ulist, which
1503  * must be freed with ulist_free.
1504  *
1505  * returns 0 on success, <0 on error
1506  */
1507 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1508                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1509                                 u64 time_seq, struct ulist **leafs,
1510                                 const u64 *extent_item_pos)
1511 {
1512         int ret;
1513
1514         *leafs = ulist_alloc(GFP_NOFS);
1515         if (!*leafs)
1516                 return -ENOMEM;
1517
1518         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1519                                 *leafs, NULL, extent_item_pos, 0, 0, 0);
1520         if (ret < 0 && ret != -ENOENT) {
1521                 free_leaf_list(*leafs);
1522                 return ret;
1523         }
1524
1525         return 0;
1526 }
1527
1528 /*
1529  * walk all backrefs for a given extent to find all roots that reference this
1530  * extent. Walking a backref means finding all extents that reference this
1531  * extent and in turn walk the backrefs of those, too. Naturally this is a
1532  * recursive process, but here it is implemented in an iterative fashion: We
1533  * find all referencing extents for the extent in question and put them on a
1534  * list. In turn, we find all referencing extents for those, further appending
1535  * to the list. The way we iterate the list allows adding more elements after
1536  * the current while iterating. The process stops when we reach the end of the
1537  * list. Found roots are added to the roots list.
1538  *
1539  * returns 0 on success, < 0 on error.
1540  */
1541 static int __btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1542                                   struct btrfs_fs_info *fs_info, u64 bytenr,
1543                                   u64 time_seq, struct ulist **roots)
1544 {
1545         struct ulist *tmp;
1546         struct ulist_node *node = NULL;
1547         struct ulist_iterator uiter;
1548         int ret;
1549
1550         tmp = ulist_alloc(GFP_NOFS);
1551         if (!tmp)
1552                 return -ENOMEM;
1553         *roots = ulist_alloc(GFP_NOFS);
1554         if (!*roots) {
1555                 ulist_free(tmp);
1556                 return -ENOMEM;
1557         }
1558
1559         ULIST_ITER_INIT(&uiter);
1560         while (1) {
1561                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1562                                         tmp, *roots, NULL, 0, 0, 0);
1563                 if (ret < 0 && ret != -ENOENT) {
1564                         ulist_free(tmp);
1565                         ulist_free(*roots);
1566                         *roots = NULL;
1567                         return ret;
1568                 }
1569                 node = ulist_next(tmp, &uiter);
1570                 if (!node)
1571                         break;
1572                 bytenr = node->val;
1573                 cond_resched();
1574         }
1575
1576         ulist_free(tmp);
1577         return 0;
1578 }
1579
1580 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1581                          struct btrfs_fs_info *fs_info, u64 bytenr,
1582                          u64 time_seq, struct ulist **roots)
1583 {
1584         int ret;
1585
1586         if (!trans)
1587                 down_read(&fs_info->commit_root_sem);
1588         ret = __btrfs_find_all_roots(trans, fs_info, bytenr, time_seq, roots);
1589         if (!trans)
1590                 up_read(&fs_info->commit_root_sem);
1591         return ret;
1592 }
1593
1594 /**
1595  * btrfs_check_shared - tell us whether an extent is shared
1596  *
1597  * @trans: optional trans handle
1598  *
1599  * btrfs_check_shared uses the backref walking code but will short
1600  * circuit as soon as it finds a root or inode that doesn't match the
1601  * one passed in. This provides a significant performance benefit for
1602  * callers (such as fiemap) which want to know whether the extent is
1603  * shared but do not need a ref count.
1604  *
1605  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1606  */
1607 int btrfs_check_shared(struct btrfs_trans_handle *trans,
1608                        struct btrfs_fs_info *fs_info, u64 root_objectid,
1609                        u64 inum, u64 bytenr)
1610 {
1611         struct ulist *tmp = NULL;
1612         struct ulist *roots = NULL;
1613         struct ulist_iterator uiter;
1614         struct ulist_node *node;
1615         struct seq_list elem = SEQ_LIST_INIT(elem);
1616         int ret = 0;
1617
1618         tmp = ulist_alloc(GFP_NOFS);
1619         roots = ulist_alloc(GFP_NOFS);
1620         if (!tmp || !roots) {
1621                 ulist_free(tmp);
1622                 ulist_free(roots);
1623                 return -ENOMEM;
1624         }
1625
1626         if (trans)
1627                 btrfs_get_tree_mod_seq(fs_info, &elem);
1628         else
1629                 down_read(&fs_info->commit_root_sem);
1630         ULIST_ITER_INIT(&uiter);
1631         while (1) {
1632                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1633                                         roots, NULL, root_objectid, inum, 1);
1634                 if (ret == BACKREF_FOUND_SHARED) {
1635                         /* this is the only condition under which we return 1 */
1636                         ret = 1;
1637                         break;
1638                 }
1639                 if (ret < 0 && ret != -ENOENT)
1640                         break;
1641                 ret = 0;
1642                 node = ulist_next(tmp, &uiter);
1643                 if (!node)
1644                         break;
1645                 bytenr = node->val;
1646                 cond_resched();
1647         }
1648         if (trans)
1649                 btrfs_put_tree_mod_seq(fs_info, &elem);
1650         else
1651                 up_read(&fs_info->commit_root_sem);
1652         ulist_free(tmp);
1653         ulist_free(roots);
1654         return ret;
1655 }
1656
1657 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1658                           u64 start_off, struct btrfs_path *path,
1659                           struct btrfs_inode_extref **ret_extref,
1660                           u64 *found_off)
1661 {
1662         int ret, slot;
1663         struct btrfs_key key;
1664         struct btrfs_key found_key;
1665         struct btrfs_inode_extref *extref;
1666         struct extent_buffer *leaf;
1667         unsigned long ptr;
1668
1669         key.objectid = inode_objectid;
1670         key.type = BTRFS_INODE_EXTREF_KEY;
1671         key.offset = start_off;
1672
1673         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1674         if (ret < 0)
1675                 return ret;
1676
1677         while (1) {
1678                 leaf = path->nodes[0];
1679                 slot = path->slots[0];
1680                 if (slot >= btrfs_header_nritems(leaf)) {
1681                         /*
1682                          * If the item at offset is not found,
1683                          * btrfs_search_slot will point us to the slot
1684                          * where it should be inserted. In our case
1685                          * that will be the slot directly before the
1686                          * next INODE_REF_KEY_V2 item. In the case
1687                          * that we're pointing to the last slot in a
1688                          * leaf, we must move one leaf over.
1689                          */
1690                         ret = btrfs_next_leaf(root, path);
1691                         if (ret) {
1692                                 if (ret >= 1)
1693                                         ret = -ENOENT;
1694                                 break;
1695                         }
1696                         continue;
1697                 }
1698
1699                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1700
1701                 /*
1702                  * Check that we're still looking at an extended ref key for
1703                  * this particular objectid. If we have different
1704                  * objectid or type then there are no more to be found
1705                  * in the tree and we can exit.
1706                  */
1707                 ret = -ENOENT;
1708                 if (found_key.objectid != inode_objectid)
1709                         break;
1710                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1711                         break;
1712
1713                 ret = 0;
1714                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1715                 extref = (struct btrfs_inode_extref *)ptr;
1716                 *ret_extref = extref;
1717                 if (found_off)
1718                         *found_off = found_key.offset;
1719                 break;
1720         }
1721
1722         return ret;
1723 }
1724
1725 /*
1726  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1727  * Elements of the path are separated by '/' and the path is guaranteed to be
1728  * 0-terminated. the path is only given within the current file system.
1729  * Therefore, it never starts with a '/'. the caller is responsible to provide
1730  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1731  * the start point of the resulting string is returned. this pointer is within
1732  * dest, normally.
1733  * in case the path buffer would overflow, the pointer is decremented further
1734  * as if output was written to the buffer, though no more output is actually
1735  * generated. that way, the caller can determine how much space would be
1736  * required for the path to fit into the buffer. in that case, the returned
1737  * value will be smaller than dest. callers must check this!
1738  */
1739 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1740                         u32 name_len, unsigned long name_off,
1741                         struct extent_buffer *eb_in, u64 parent,
1742                         char *dest, u32 size)
1743 {
1744         int slot;
1745         u64 next_inum;
1746         int ret;
1747         s64 bytes_left = ((s64)size) - 1;
1748         struct extent_buffer *eb = eb_in;
1749         struct btrfs_key found_key;
1750         int leave_spinning = path->leave_spinning;
1751         struct btrfs_inode_ref *iref;
1752
1753         if (bytes_left >= 0)
1754                 dest[bytes_left] = '\0';
1755
1756         path->leave_spinning = 1;
1757         while (1) {
1758                 bytes_left -= name_len;
1759                 if (bytes_left >= 0)
1760                         read_extent_buffer(eb, dest + bytes_left,
1761                                            name_off, name_len);
1762                 if (eb != eb_in) {
1763                         if (!path->skip_locking)
1764                                 btrfs_tree_read_unlock_blocking(eb);
1765                         free_extent_buffer(eb);
1766                 }
1767                 ret = btrfs_find_item(fs_root, path, parent, 0,
1768                                 BTRFS_INODE_REF_KEY, &found_key);
1769                 if (ret > 0)
1770                         ret = -ENOENT;
1771                 if (ret)
1772                         break;
1773
1774                 next_inum = found_key.offset;
1775
1776                 /* regular exit ahead */
1777                 if (parent == next_inum)
1778                         break;
1779
1780                 slot = path->slots[0];
1781                 eb = path->nodes[0];
1782                 /* make sure we can use eb after releasing the path */
1783                 if (eb != eb_in) {
1784                         if (!path->skip_locking)
1785                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1786                         path->nodes[0] = NULL;
1787                         path->locks[0] = 0;
1788                 }
1789                 btrfs_release_path(path);
1790                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1791
1792                 name_len = btrfs_inode_ref_name_len(eb, iref);
1793                 name_off = (unsigned long)(iref + 1);
1794
1795                 parent = next_inum;
1796                 --bytes_left;
1797                 if (bytes_left >= 0)
1798                         dest[bytes_left] = '/';
1799         }
1800
1801         btrfs_release_path(path);
1802         path->leave_spinning = leave_spinning;
1803
1804         if (ret)
1805                 return ERR_PTR(ret);
1806
1807         return dest + bytes_left;
1808 }
1809
1810 /*
1811  * this makes the path point to (logical EXTENT_ITEM *)
1812  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1813  * tree blocks and <0 on error.
1814  */
1815 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1816                         struct btrfs_path *path, struct btrfs_key *found_key,
1817                         u64 *flags_ret)
1818 {
1819         int ret;
1820         u64 flags;
1821         u64 size = 0;
1822         u32 item_size;
1823         struct extent_buffer *eb;
1824         struct btrfs_extent_item *ei;
1825         struct btrfs_key key;
1826
1827         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1828                 key.type = BTRFS_METADATA_ITEM_KEY;
1829         else
1830                 key.type = BTRFS_EXTENT_ITEM_KEY;
1831         key.objectid = logical;
1832         key.offset = (u64)-1;
1833
1834         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1835         if (ret < 0)
1836                 return ret;
1837
1838         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1839         if (ret) {
1840                 if (ret > 0)
1841                         ret = -ENOENT;
1842                 return ret;
1843         }
1844         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1845         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1846                 size = fs_info->extent_root->nodesize;
1847         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1848                 size = found_key->offset;
1849
1850         if (found_key->objectid > logical ||
1851             found_key->objectid + size <= logical) {
1852                 btrfs_debug(fs_info,
1853                         "logical %llu is not within any extent", logical);
1854                 return -ENOENT;
1855         }
1856
1857         eb = path->nodes[0];
1858         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1859         BUG_ON(item_size < sizeof(*ei));
1860
1861         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1862         flags = btrfs_extent_flags(eb, ei);
1863
1864         btrfs_debug(fs_info,
1865                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1866                  logical, logical - found_key->objectid, found_key->objectid,
1867                  found_key->offset, flags, item_size);
1868
1869         WARN_ON(!flags_ret);
1870         if (flags_ret) {
1871                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1872                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1873                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1874                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1875                 else
1876                         BUG_ON(1);
1877                 return 0;
1878         }
1879
1880         return -EIO;
1881 }
1882
1883 /*
1884  * helper function to iterate extent inline refs. ptr must point to a 0 value
1885  * for the first call and may be modified. it is used to track state.
1886  * if more refs exist, 0 is returned and the next call to
1887  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1888  * next ref. after the last ref was processed, 1 is returned.
1889  * returns <0 on error
1890  */
1891 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1892                                    struct btrfs_key *key,
1893                                    struct btrfs_extent_item *ei, u32 item_size,
1894                                    struct btrfs_extent_inline_ref **out_eiref,
1895                                    int *out_type)
1896 {
1897         unsigned long end;
1898         u64 flags;
1899         struct btrfs_tree_block_info *info;
1900
1901         if (!*ptr) {
1902                 /* first call */
1903                 flags = btrfs_extent_flags(eb, ei);
1904                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1905                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1906                                 /* a skinny metadata extent */
1907                                 *out_eiref =
1908                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1909                         } else {
1910                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1911                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1912                                 *out_eiref =
1913                                    (struct btrfs_extent_inline_ref *)(info + 1);
1914                         }
1915                 } else {
1916                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1917                 }
1918                 *ptr = (unsigned long)*out_eiref;
1919                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1920                         return -ENOENT;
1921         }
1922
1923         end = (unsigned long)ei + item_size;
1924         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1925         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1926
1927         *ptr += btrfs_extent_inline_ref_size(*out_type);
1928         WARN_ON(*ptr > end);
1929         if (*ptr == end)
1930                 return 1; /* last */
1931
1932         return 0;
1933 }
1934
1935 /*
1936  * reads the tree block backref for an extent. tree level and root are returned
1937  * through out_level and out_root. ptr must point to a 0 value for the first
1938  * call and may be modified (see __get_extent_inline_ref comment).
1939  * returns 0 if data was provided, 1 if there was no more data to provide or
1940  * <0 on error.
1941  */
1942 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1943                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1944                             u32 item_size, u64 *out_root, u8 *out_level)
1945 {
1946         int ret;
1947         int type;
1948         struct btrfs_extent_inline_ref *eiref;
1949
1950         if (*ptr == (unsigned long)-1)
1951                 return 1;
1952
1953         while (1) {
1954                 ret = __get_extent_inline_ref(ptr, eb, key, ei, item_size,
1955                                               &eiref, &type);
1956                 if (ret < 0)
1957                         return ret;
1958
1959                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1960                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1961                         break;
1962
1963                 if (ret == 1)
1964                         return 1;
1965         }
1966
1967         /* we can treat both ref types equally here */
1968         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1969
1970         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1971                 struct btrfs_tree_block_info *info;
1972
1973                 info = (struct btrfs_tree_block_info *)(ei + 1);
1974                 *out_level = btrfs_tree_block_level(eb, info);
1975         } else {
1976                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1977                 *out_level = (u8)key->offset;
1978         }
1979
1980         if (ret == 1)
1981                 *ptr = (unsigned long)-1;
1982
1983         return 0;
1984 }
1985
1986 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1987                              struct extent_inode_elem *inode_list,
1988                              u64 root, u64 extent_item_objectid,
1989                              iterate_extent_inodes_t *iterate, void *ctx)
1990 {
1991         struct extent_inode_elem *eie;
1992         int ret = 0;
1993
1994         for (eie = inode_list; eie; eie = eie->next) {
1995                 btrfs_debug(fs_info,
1996                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1997                             extent_item_objectid, eie->inum,
1998                             eie->offset, root);
1999                 ret = iterate(eie->inum, eie->offset, root, ctx);
2000                 if (ret) {
2001                         btrfs_debug(fs_info,
2002                                     "stopping iteration for %llu due to ret=%d",
2003                                     extent_item_objectid, ret);
2004                         break;
2005                 }
2006         }
2007
2008         return ret;
2009 }
2010
2011 /*
2012  * calls iterate() for every inode that references the extent identified by
2013  * the given parameters.
2014  * when the iterator function returns a non-zero value, iteration stops.
2015  */
2016 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
2017                                 u64 extent_item_objectid, u64 extent_item_pos,
2018                                 int search_commit_root,
2019                                 iterate_extent_inodes_t *iterate, void *ctx)
2020 {
2021         int ret;
2022         struct btrfs_trans_handle *trans = NULL;
2023         struct ulist *refs = NULL;
2024         struct ulist *roots = NULL;
2025         struct ulist_node *ref_node = NULL;
2026         struct ulist_node *root_node = NULL;
2027         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
2028         struct ulist_iterator ref_uiter;
2029         struct ulist_iterator root_uiter;
2030
2031         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
2032                         extent_item_objectid);
2033
2034         if (!search_commit_root) {
2035                 trans = btrfs_attach_transaction(fs_info->extent_root);
2036                 if (IS_ERR(trans)) {
2037                         if (PTR_ERR(trans) != -ENOENT &&
2038                             PTR_ERR(trans) != -EROFS)
2039                                 return PTR_ERR(trans);
2040                         trans = NULL;
2041                 }
2042         }
2043
2044         if (trans)
2045                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2046         else
2047                 down_read(&fs_info->commit_root_sem);
2048
2049         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2050                                    tree_mod_seq_elem.seq, &refs,
2051                                    &extent_item_pos);
2052         if (ret)
2053                 goto out;
2054
2055         ULIST_ITER_INIT(&ref_uiter);
2056         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2057                 ret = __btrfs_find_all_roots(trans, fs_info, ref_node->val,
2058                                              tree_mod_seq_elem.seq, &roots);
2059                 if (ret)
2060                         break;
2061                 ULIST_ITER_INIT(&root_uiter);
2062                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2063                         btrfs_debug(fs_info,
2064                                     "root %llu references leaf %llu, data list %#llx",
2065                                     root_node->val, ref_node->val,
2066                                     ref_node->aux);
2067                         ret = iterate_leaf_refs(fs_info,
2068                                                 (struct extent_inode_elem *)
2069                                                 (uintptr_t)ref_node->aux,
2070                                                 root_node->val,
2071                                                 extent_item_objectid,
2072                                                 iterate, ctx);
2073                 }
2074                 ulist_free(roots);
2075         }
2076
2077         free_leaf_list(refs);
2078 out:
2079         if (trans) {
2080                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2081                 btrfs_end_transaction(trans, fs_info->extent_root);
2082         } else {
2083                 up_read(&fs_info->commit_root_sem);
2084         }
2085
2086         return ret;
2087 }
2088
2089 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2090                                 struct btrfs_path *path,
2091                                 iterate_extent_inodes_t *iterate, void *ctx)
2092 {
2093         int ret;
2094         u64 extent_item_pos;
2095         u64 flags = 0;
2096         struct btrfs_key found_key;
2097         int search_commit_root = path->search_commit_root;
2098
2099         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2100         btrfs_release_path(path);
2101         if (ret < 0)
2102                 return ret;
2103         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2104                 return -EINVAL;
2105
2106         extent_item_pos = logical - found_key.objectid;
2107         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2108                                         extent_item_pos, search_commit_root,
2109                                         iterate, ctx);
2110
2111         return ret;
2112 }
2113
2114 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2115                               struct extent_buffer *eb, void *ctx);
2116
2117 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2118                               struct btrfs_path *path,
2119                               iterate_irefs_t *iterate, void *ctx)
2120 {
2121         int ret = 0;
2122         int slot;
2123         u32 cur;
2124         u32 len;
2125         u32 name_len;
2126         u64 parent = 0;
2127         int found = 0;
2128         struct extent_buffer *eb;
2129         struct btrfs_item *item;
2130         struct btrfs_inode_ref *iref;
2131         struct btrfs_key found_key;
2132
2133         while (!ret) {
2134                 ret = btrfs_find_item(fs_root, path, inum,
2135                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2136                                 &found_key);
2137
2138                 if (ret < 0)
2139                         break;
2140                 if (ret) {
2141                         ret = found ? 0 : -ENOENT;
2142                         break;
2143                 }
2144                 ++found;
2145
2146                 parent = found_key.offset;
2147                 slot = path->slots[0];
2148                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2149                 if (!eb) {
2150                         ret = -ENOMEM;
2151                         break;
2152                 }
2153                 extent_buffer_get(eb);
2154                 btrfs_tree_read_lock(eb);
2155                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2156                 btrfs_release_path(path);
2157
2158                 item = btrfs_item_nr(slot);
2159                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2160
2161                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2162                         name_len = btrfs_inode_ref_name_len(eb, iref);
2163                         /* path must be released before calling iterate()! */
2164                         btrfs_debug(fs_root->fs_info,
2165                                 "following ref at offset %u for inode %llu in tree %llu",
2166                                 cur, found_key.objectid, fs_root->objectid);
2167                         ret = iterate(parent, name_len,
2168                                       (unsigned long)(iref + 1), eb, ctx);
2169                         if (ret)
2170                                 break;
2171                         len = sizeof(*iref) + name_len;
2172                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2173                 }
2174                 btrfs_tree_read_unlock_blocking(eb);
2175                 free_extent_buffer(eb);
2176         }
2177
2178         btrfs_release_path(path);
2179
2180         return ret;
2181 }
2182
2183 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2184                                  struct btrfs_path *path,
2185                                  iterate_irefs_t *iterate, void *ctx)
2186 {
2187         int ret;
2188         int slot;
2189         u64 offset = 0;
2190         u64 parent;
2191         int found = 0;
2192         struct extent_buffer *eb;
2193         struct btrfs_inode_extref *extref;
2194         u32 item_size;
2195         u32 cur_offset;
2196         unsigned long ptr;
2197
2198         while (1) {
2199                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2200                                             &offset);
2201                 if (ret < 0)
2202                         break;
2203                 if (ret) {
2204                         ret = found ? 0 : -ENOENT;
2205                         break;
2206                 }
2207                 ++found;
2208
2209                 slot = path->slots[0];
2210                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2211                 if (!eb) {
2212                         ret = -ENOMEM;
2213                         break;
2214                 }
2215                 extent_buffer_get(eb);
2216
2217                 btrfs_tree_read_lock(eb);
2218                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2219                 btrfs_release_path(path);
2220
2221                 item_size = btrfs_item_size_nr(eb, slot);
2222                 ptr = btrfs_item_ptr_offset(eb, slot);
2223                 cur_offset = 0;
2224
2225                 while (cur_offset < item_size) {
2226                         u32 name_len;
2227
2228                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2229                         parent = btrfs_inode_extref_parent(eb, extref);
2230                         name_len = btrfs_inode_extref_name_len(eb, extref);
2231                         ret = iterate(parent, name_len,
2232                                       (unsigned long)&extref->name, eb, ctx);
2233                         if (ret)
2234                                 break;
2235
2236                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2237                         cur_offset += sizeof(*extref);
2238                 }
2239                 btrfs_tree_read_unlock_blocking(eb);
2240                 free_extent_buffer(eb);
2241
2242                 offset++;
2243         }
2244
2245         btrfs_release_path(path);
2246
2247         return ret;
2248 }
2249
2250 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2251                          struct btrfs_path *path, iterate_irefs_t *iterate,
2252                          void *ctx)
2253 {
2254         int ret;
2255         int found_refs = 0;
2256
2257         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2258         if (!ret)
2259                 ++found_refs;
2260         else if (ret != -ENOENT)
2261                 return ret;
2262
2263         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2264         if (ret == -ENOENT && found_refs)
2265                 return 0;
2266
2267         return ret;
2268 }
2269
2270 /*
2271  * returns 0 if the path could be dumped (probably truncated)
2272  * returns <0 in case of an error
2273  */
2274 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2275                          struct extent_buffer *eb, void *ctx)
2276 {
2277         struct inode_fs_paths *ipath = ctx;
2278         char *fspath;
2279         char *fspath_min;
2280         int i = ipath->fspath->elem_cnt;
2281         const int s_ptr = sizeof(char *);
2282         u32 bytes_left;
2283
2284         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2285                                         ipath->fspath->bytes_left - s_ptr : 0;
2286
2287         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2288         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2289                                    name_off, eb, inum, fspath_min, bytes_left);
2290         if (IS_ERR(fspath))
2291                 return PTR_ERR(fspath);
2292
2293         if (fspath > fspath_min) {
2294                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2295                 ++ipath->fspath->elem_cnt;
2296                 ipath->fspath->bytes_left = fspath - fspath_min;
2297         } else {
2298                 ++ipath->fspath->elem_missed;
2299                 ipath->fspath->bytes_missing += fspath_min - fspath;
2300                 ipath->fspath->bytes_left = 0;
2301         }
2302
2303         return 0;
2304 }
2305
2306 /*
2307  * this dumps all file system paths to the inode into the ipath struct, provided
2308  * is has been created large enough. each path is zero-terminated and accessed
2309  * from ipath->fspath->val[i].
2310  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2311  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2312  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2313  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2314  * have been needed to return all paths.
2315  */
2316 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2317 {
2318         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2319                              inode_to_path, ipath);
2320 }
2321
2322 struct btrfs_data_container *init_data_container(u32 total_bytes)
2323 {
2324         struct btrfs_data_container *data;
2325         size_t alloc_bytes;
2326
2327         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2328         data = vmalloc(alloc_bytes);
2329         if (!data)
2330                 return ERR_PTR(-ENOMEM);
2331
2332         if (total_bytes >= sizeof(*data)) {
2333                 data->bytes_left = total_bytes - sizeof(*data);
2334                 data->bytes_missing = 0;
2335         } else {
2336                 data->bytes_missing = sizeof(*data) - total_bytes;
2337                 data->bytes_left = 0;
2338         }
2339
2340         data->elem_cnt = 0;
2341         data->elem_missed = 0;
2342
2343         return data;
2344 }
2345
2346 /*
2347  * allocates space to return multiple file system paths for an inode.
2348  * total_bytes to allocate are passed, note that space usable for actual path
2349  * information will be total_bytes - sizeof(struct inode_fs_paths).
2350  * the returned pointer must be freed with free_ipath() in the end.
2351  */
2352 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2353                                         struct btrfs_path *path)
2354 {
2355         struct inode_fs_paths *ifp;
2356         struct btrfs_data_container *fspath;
2357
2358         fspath = init_data_container(total_bytes);
2359         if (IS_ERR(fspath))
2360                 return (void *)fspath;
2361
2362         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
2363         if (!ifp) {
2364                 vfree(fspath);
2365                 return ERR_PTR(-ENOMEM);
2366         }
2367
2368         ifp->btrfs_path = path;
2369         ifp->fspath = fspath;
2370         ifp->fs_root = fs_root;
2371
2372         return ifp;
2373 }
2374
2375 void free_ipath(struct inode_fs_paths *ipath)
2376 {
2377         if (!ipath)
2378                 return;
2379         vfree(ipath->fspath);
2380         kfree(ipath);
2381 }