fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  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/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 /*
 330  * walk up backref nodes until reach node presents tree root
 331  */
 332 static struct backref_node *walk_up_backref(struct backref_node *node,
 333                                             struct backref_edge *edges[],
 334                                             int *index)
 335 {
 336         struct backref_edge *edge;
 337         int idx = *index;
 338
 339         while (!list_empty(&node->upper)) {
 340                 edge = list_entry(node->upper.next,
 341                                   struct backref_edge, list[LOWER]);
 342                 edges[idx++] = edge;
 343                 node = edge->node[UPPER];
 344         }
 345         BUG_ON(node->detached);
 346         *index = idx;
 347         return node;
 348 }
 349
 350 /*
 351  * walk down backref nodes to find start of next reference path
 352  */
 353 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 354                                               int *index)
 355 {
 356         struct backref_edge *edge;
 357         struct backref_node *lower;
 358         int idx = *index;
 359
 360         while (idx > 0) {
 361                 edge = edges[idx - 1];
 362                 lower = edge->node[LOWER];
 363                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 364                         idx--;
 365                         continue;
 366                 }
 367                 edge = list_entry(edge->list[LOWER].next,
 368                                   struct backref_edge, list[LOWER]);
 369                 edges[idx - 1] = edge;
 370                 *index = idx;
 371                 return edge->node[UPPER];
 372         }
 373         *index = 0;
 374         return NULL;
 375 }
 376
 377 static void unlock_node_buffer(struct backref_node *node)
 378 {
 379         if (node->locked) {
 380                 btrfs_tree_unlock(node->eb);
 381                 node->locked = 0;
 382         }
 383 }
 384
 385 static void drop_node_buffer(struct backref_node *node)
 386 {
 387         if (node->eb) {
 388                 unlock_node_buffer(node);
 389                 free_extent_buffer(node->eb);
 390                 node->eb = NULL;
 391         }
 392 }
 393
 394 static void drop_backref_node(struct backref_cache *tree,
 395                               struct backref_node *node)
 396 {
 397         BUG_ON(!list_empty(&node->upper));
 398
 399         drop_node_buffer(node);
 400         list_del(&node->list);
 401         list_del(&node->lower);
 402         if (!RB_EMPTY_NODE(&node->rb_node))
 403                 rb_erase(&node->rb_node, &tree->rb_root);
 404         free_backref_node(tree, node);
 405 }
 406
 407 /*
 408  * remove a backref node from the backref cache
 409  */
 410 static void remove_backref_node(struct backref_cache *cache,
 411                                 struct backref_node *node)
 412 {
 413         struct backref_node *upper;
 414         struct backref_edge *edge;
 415
 416         if (!node)
 417                 return;
 418
 419         BUG_ON(!node->lowest && !node->detached);
 420         while (!list_empty(&node->upper)) {
 421                 edge = list_entry(node->upper.next, struct backref_edge,
 422                                   list[LOWER]);
 423                 upper = edge->node[UPPER];
 424                 list_del(&edge->list[LOWER]);
 425                 list_del(&edge->list[UPPER]);
 426                 free_backref_edge(cache, edge);
 427
 428                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 429                         BUG_ON(!list_empty(&node->upper));
 430                         drop_backref_node(cache, node);
 431                         node = upper;
 432                         node->lowest = 1;
 433                         continue;
 434                 }
 435                 /*
 436                  * add the node to leaf node list if no other
 437                  * child block cached.
 438                  */
 439                 if (list_empty(&upper->lower)) {
 440                         list_add_tail(&upper->lower, &cache->leaves);
 441                         upper->lowest = 1;
 442                 }
 443         }
 444
 445         drop_backref_node(cache, node);
 446 }
 447
 448 static void update_backref_node(struct backref_cache *cache,
 449                                 struct backref_node *node, u64 bytenr)
 450 {
 451         struct rb_node *rb_node;
 452         rb_erase(&node->rb_node, &cache->rb_root);
 453         node->bytenr = bytenr;
 454         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 455         BUG_ON(rb_node);
 456 }
 457
 458 /*
 459  * update backref cache after a transaction commit
 460  */
 461 static int update_backref_cache(struct btrfs_trans_handle *trans,
 462                                 struct backref_cache *cache)
 463 {
 464         struct backref_node *node;
 465         int level = 0;
 466
 467         if (cache->last_trans == 0) {
 468                 cache->last_trans = trans->transid;
 469                 return 0;
 470         }
 471
 472         if (cache->last_trans == trans->transid)
 473                 return 0;
 474
 475         /*
 476          * detached nodes are used to avoid unnecessary backref
 477          * lookup. transaction commit changes the extent tree.
 478          * so the detached nodes are no longer useful.
 479          */
 480         while (!list_empty(&cache->detached)) {
 481                 node = list_entry(cache->detached.next,
 482                                   struct backref_node, list);
 483                 remove_backref_node(cache, node);
 484         }
 485
 486         while (!list_empty(&cache->changed)) {
 487                 node = list_entry(cache->changed.next,
 488                                   struct backref_node, list);
 489                 list_del_init(&node->list);
 490                 BUG_ON(node->pending);
 491                 update_backref_node(cache, node, node->new_bytenr);
 492         }
 493
 494         /*
 495          * some nodes can be left in the pending list if there were
 496          * errors during processing the pending nodes.
 497          */
 498         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 499                 list_for_each_entry(node, &cache->pending[level], list) {
 500                         BUG_ON(!node->pending);
 501                         if (node->bytenr == node->new_bytenr)
 502                                 continue;
 503                         update_backref_node(cache, node, node->new_bytenr);
 504                 }
 505         }
 506
 507         cache->last_trans = 0;
 508         return 1;
 509 }
 510
 511
 512 static int should_ignore_root(struct btrfs_root *root)
 513 {
 514         struct btrfs_root *reloc_root;
 515
 516         if (!root->ref_cows)
 517                 return 0;
 518
 519         reloc_root = root->reloc_root;
 520         if (!reloc_root)
 521                 return 0;
 522
 523         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 524             root->fs_info->running_transaction->transid - 1)
 525                 return 0;
 526         /*
 527          * if there is reloc tree and it was created in previous
 528          * transaction backref lookup can find the reloc tree,
 529          * so backref node for the fs tree root is useless for
 530          * relocation.
 531          */
 532         return 1;
 533 }
 534 /*
 535  * find reloc tree by address of tree root
 536  */
 537 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 538                                           u64 bytenr)
 539 {
 540         struct rb_node *rb_node;
 541         struct mapping_node *node;
 542         struct btrfs_root *root = NULL;
 543
 544         spin_lock(&rc->reloc_root_tree.lock);
 545         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 546         if (rb_node) {
 547                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 548                 root = (struct btrfs_root *)node->data;
 549         }
 550         spin_unlock(&rc->reloc_root_tree.lock);
 551         return root;
 552 }
 553
 554 static int is_cowonly_root(u64 root_objectid)
 555 {
 556         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 557             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 558             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 559             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 560             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 561             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 562                 return 1;
 563         return 0;
 564 }
 565
 566 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 567                                         u64 root_objectid)
 568 {
 569         struct btrfs_key key;
 570
 571         key.objectid = root_objectid;
 572         key.type = BTRFS_ROOT_ITEM_KEY;
 573         if (is_cowonly_root(root_objectid))
 574                 key.offset = 0;
 575         else
 576                 key.offset = (u64)-1;
 577
 578         return btrfs_read_fs_root_no_name(fs_info, &key);
 579 }
 580
 581 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 582 static noinline_for_stack
 583 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 584                                   struct extent_buffer *leaf,
 585                                   struct btrfs_extent_ref_v0 *ref0)
 586 {
 587         struct btrfs_root *root;
 588         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 589         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 590
 591         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 592
 593         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 594         BUG_ON(IS_ERR(root));
 595
 596         if (root->ref_cows &&
 597             generation != btrfs_root_generation(&root->root_item))
 598                 return NULL;
 599
 600         return root;
 601 }
 602 #endif
 603
 604 static noinline_for_stack
 605 int find_inline_backref(struct extent_buffer *leaf, int slot,
 606                         unsigned long *ptr, unsigned long *end)
 607 {
 608         struct btrfs_extent_item *ei;
 609         struct btrfs_tree_block_info *bi;
 610         u32 item_size;
 611
 612         item_size = btrfs_item_size_nr(leaf, slot);
 613 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 614         if (item_size < sizeof(*ei)) {
 615                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 616                 return 1;
 617         }
 618 #endif
 619         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 620         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 621                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 622
 623         if (item_size <= sizeof(*ei) + sizeof(*bi)) {
 624                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 625                 return 1;
 626         }
 627
 628         bi = (struct btrfs_tree_block_info *)(ei + 1);
 629         *ptr = (unsigned long)(bi + 1);
 630         *end = (unsigned long)ei + item_size;
 631         return 0;
 632 }
 633
 634 /*
 635  * build backref tree for a given tree block. root of the backref tree
 636  * corresponds the tree block, leaves of the backref tree correspond
 637  * roots of b-trees that reference the tree block.
 638  *
 639  * the basic idea of this function is check backrefs of a given block
 640  * to find upper level blocks that refernece the block, and then check
 641  * bakcrefs of these upper level blocks recursively. the recursion stop
 642  * when tree root is reached or backrefs for the block is cached.
 643  *
 644  * NOTE: if we find backrefs for a block are cached, we know backrefs
 645  * for all upper level blocks that directly/indirectly reference the
 646  * block are also cached.
 647  */
 648 static noinline_for_stack
 649 struct backref_node *build_backref_tree(struct reloc_control *rc,
 650                                         struct btrfs_key *node_key,
 651                                         int level, u64 bytenr)
 652 {
 653         struct backref_cache *cache = &rc->backref_cache;
 654         struct btrfs_path *path1;
 655         struct btrfs_path *path2;
 656         struct extent_buffer *eb;
 657         struct btrfs_root *root;
 658         struct backref_node *cur;
 659         struct backref_node *upper;
 660         struct backref_node *lower;
 661         struct backref_node *node = NULL;
 662         struct backref_node *exist = NULL;
 663         struct backref_edge *edge;
 664         struct rb_node *rb_node;
 665         struct btrfs_key key;
 666         unsigned long end;
 667         unsigned long ptr;
 668         LIST_HEAD(list);
 669         LIST_HEAD(useless);
 670         int cowonly;
 671         int ret;
 672         int err = 0;
 673
 674         path1 = btrfs_alloc_path();
 675         path2 = btrfs_alloc_path();
 676         if (!path1 || !path2) {
 677                 err = -ENOMEM;
 678                 goto out;
 679         }
 680         path1->reada = 1;
 681         path2->reada = 2;
 682
 683         node = alloc_backref_node(cache);
 684         if (!node) {
 685                 err = -ENOMEM;
 686                 goto out;
 687         }
 688
 689         node->bytenr = bytenr;
 690         node->level = level;
 691         node->lowest = 1;
 692         cur = node;
 693 again:
 694         end = 0;
 695         ptr = 0;
 696         key.objectid = cur->bytenr;
 697         key.type = BTRFS_EXTENT_ITEM_KEY;
 698         key.offset = (u64)-1;
 699
 700         path1->search_commit_root = 1;
 701         path1->skip_locking = 1;
 702         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 703                                 0, 0);
 704         if (ret < 0) {
 705                 err = ret;
 706                 goto out;
 707         }
 708         BUG_ON(!ret || !path1->slots[0]);
 709
 710         path1->slots[0]--;
 711
 712         WARN_ON(cur->checked);
 713         if (!list_empty(&cur->upper)) {
 714                 /*
 715                  * the backref was added previously when processing
 716                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 717                  */
 718                 BUG_ON(!list_is_singular(&cur->upper));
 719                 edge = list_entry(cur->upper.next, struct backref_edge,
 720                                   list[LOWER]);
 721                 BUG_ON(!list_empty(&edge->list[UPPER]));
 722                 exist = edge->node[UPPER];
 723                 /*
 724                  * add the upper level block to pending list if we need
 725                  * check its backrefs
 726                  */
 727                 if (!exist->checked)
 728                         list_add_tail(&edge->list[UPPER], &list);
 729         } else {
 730                 exist = NULL;
 731         }
 732
 733         while (1) {
 734                 cond_resched();
 735                 eb = path1->nodes[0];
 736
 737                 if (ptr >= end) {
 738                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 739                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 740                                 if (ret < 0) {
 741                                         err = ret;
 742                                         goto out;
 743                                 }
 744                                 if (ret > 0)
 745                                         break;
 746                                 eb = path1->nodes[0];
 747                         }
 748
 749                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 750                         if (key.objectid != cur->bytenr) {
 751                                 WARN_ON(exist);
 752                                 break;
 753                         }
 754
 755                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 756                                 ret = find_inline_backref(eb, path1->slots[0],
 757                                                           &ptr, &end);
 758                                 if (ret)
 759                                         goto next;
 760                         }
 761                 }
 762
 763                 if (ptr < end) {
 764                         /* update key for inline back ref */
 765                         struct btrfs_extent_inline_ref *iref;
 766                         iref = (struct btrfs_extent_inline_ref *)ptr;
 767                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 768                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 769                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 770                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 771                 }
 772
 773                 if (exist &&
 774                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 775                       exist->owner == key.offset) ||
 776                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 777                       exist->bytenr == key.offset))) {
 778                         exist = NULL;
 779                         goto next;
 780                 }
 781
 782 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 783                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 784                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 785                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 786                                 struct btrfs_extent_ref_v0 *ref0;
 787                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 788                                                 struct btrfs_extent_ref_v0);
 789                                 if (key.objectid == key.offset) {
 790                                         root = find_tree_root(rc, eb, ref0);
 791                                         if (root && !should_ignore_root(root))
 792                                                 cur->root = root;
 793                                         else
 794                                                 list_add(&cur->list, &useless);
 795                                         break;
 796                                 }
 797                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 798                                                                       ref0)))
 799                                         cur->cowonly = 1;
 800                         }
 801 #else
 802                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 803                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 804 #endif
 805                         if (key.objectid == key.offset) {
 806                                 /*
 807                                  * only root blocks of reloc trees use
 808                                  * backref of this type.
 809                                  */
 810                                 root = find_reloc_root(rc, cur->bytenr);
 811                                 BUG_ON(!root);
 812                                 cur->root = root;
 813                                 break;
 814                         }
 815
 816                         edge = alloc_backref_edge(cache);
 817                         if (!edge) {
 818                                 err = -ENOMEM;
 819                                 goto out;
 820                         }
 821                         rb_node = tree_search(&cache->rb_root, key.offset);
 822                         if (!rb_node) {
 823                                 upper = alloc_backref_node(cache);
 824                                 if (!upper) {
 825                                         free_backref_edge(cache, edge);
 826                                         err = -ENOMEM;
 827                                         goto out;
 828                                 }
 829                                 upper->bytenr = key.offset;
 830                                 upper->level = cur->level + 1;
 831                                 /*
 832                                  *  backrefs for the upper level block isn't
 833                                  *  cached, add the block to pending list
 834                                  */
 835                                 list_add_tail(&edge->list[UPPER], &list);
 836                         } else {
 837                                 upper = rb_entry(rb_node, struct backref_node,
 838                                                  rb_node);
 839                                 BUG_ON(!upper->checked);
 840                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 841                         }
 842                         list_add_tail(&edge->list[LOWER], &cur->upper);
 843                         edge->node[LOWER] = cur;
 844                         edge->node[UPPER] = upper;
 845
 846                         goto next;
 847                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 848                         goto next;
 849                 }
 850
 851                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 852                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 853                 if (IS_ERR(root)) {
 854                         err = PTR_ERR(root);
 855                         goto out;
 856                 }
 857
 858                 if (!root->ref_cows)
 859                         cur->cowonly = 1;
 860
 861                 if (btrfs_root_level(&root->root_item) == cur->level) {
 862                         /* tree root */
 863                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 864                                cur->bytenr);
 865                         if (should_ignore_root(root))
 866                                 list_add(&cur->list, &useless);
 867                         else
 868                                 cur->root = root;
 869                         break;
 870                 }
 871
 872                 level = cur->level + 1;
 873
 874                 /*
 875                  * searching the tree to find upper level blocks
 876                  * reference the block.
 877                  */
 878                 path2->search_commit_root = 1;
 879                 path2->skip_locking = 1;
 880                 path2->lowest_level = level;
 881                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 882                 path2->lowest_level = 0;
 883                 if (ret < 0) {
 884                         err = ret;
 885                         goto out;
 886                 }
 887                 if (ret > 0 && path2->slots[level] > 0)
 888                         path2->slots[level]--;
 889
 890                 eb = path2->nodes[level];
 891                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 892                         cur->bytenr);
 893
 894                 lower = cur;
 895                 for (; level < BTRFS_MAX_LEVEL; level++) {
 896                         if (!path2->nodes[level]) {
 897                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 898                                        lower->bytenr);
 899                                 if (should_ignore_root(root))
 900                                         list_add(&lower->list, &useless);
 901                                 else
 902                                         lower->root = root;
 903                                 break;
 904                         }
 905
 906                         edge = alloc_backref_edge(cache);
 907                         if (!edge) {
 908                                 err = -ENOMEM;
 909                                 goto out;
 910                         }
 911
 912                         eb = path2->nodes[level];
 913                         rb_node = tree_search(&cache->rb_root, eb->start);
 914                         if (!rb_node) {
 915                                 upper = alloc_backref_node(cache);
 916                                 if (!upper) {
 917                                         free_backref_edge(cache, edge);
 918                                         err = -ENOMEM;
 919                                         goto out;
 920                                 }
 921                                 upper->bytenr = eb->start;
 922                                 upper->owner = btrfs_header_owner(eb);
 923                                 upper->level = lower->level + 1;
 924                                 if (!root->ref_cows)
 925                                         upper->cowonly = 1;
 926
 927                                 /*
 928                                  * if we know the block isn't shared
 929                                  * we can void checking its backrefs.
 930                                  */
 931                                 if (btrfs_block_can_be_shared(root, eb))
 932                                         upper->checked = 0;
 933                                 else
 934                                         upper->checked = 1;
 935
 936                                 /*
 937                                  * add the block to pending list if we
 938                                  * need check its backrefs. only block
 939                                  * at 'cur->level + 1' is added to the
 940                                  * tail of pending list. this guarantees
 941                                  * we check backrefs from lower level
 942                                  * blocks to upper level blocks.
 943                                  */
 944                                 if (!upper->checked &&
 945                                     level == cur->level + 1) {
 946                                         list_add_tail(&edge->list[UPPER],
 947                                                       &list);
 948                                 } else
 949                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 950                         } else {
 951                                 upper = rb_entry(rb_node, struct backref_node,
 952                                                  rb_node);
 953                                 BUG_ON(!upper->checked);
 954                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 955                                 if (!upper->owner)
 956                                         upper->owner = btrfs_header_owner(eb);
 957                         }
 958                         list_add_tail(&edge->list[LOWER], &lower->upper);
 959                         edge->node[LOWER] = lower;
 960                         edge->node[UPPER] = upper;
 961
 962                         if (rb_node)
 963                                 break;
 964                         lower = upper;
 965                         upper = NULL;
 966                 }
 967                 btrfs_release_path(path2);
 968 next:
 969                 if (ptr < end) {
 970                         ptr += btrfs_extent_inline_ref_size(key.type);
 971                         if (ptr >= end) {
 972                                 WARN_ON(ptr > end);
 973                                 ptr = 0;
 974                                 end = 0;
 975                         }
 976                 }
 977                 if (ptr >= end)
 978                         path1->slots[0]++;
 979         }
 980         btrfs_release_path(path1);
 981
 982         cur->checked = 1;
 983         WARN_ON(exist);
 984
 985         /* the pending list isn't empty, take the first block to process */
 986         if (!list_empty(&list)) {
 987                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
 988                 list_del_init(&edge->list[UPPER]);
 989                 cur = edge->node[UPPER];
 990                 goto again;
 991         }
 992
 993         /*
 994          * everything goes well, connect backref nodes and insert backref nodes
 995          * into the cache.
 996          */
 997         BUG_ON(!node->checked);
 998         cowonly = node->cowonly;
 999         if (!cowonly) {
1000                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
1001                                       &node->rb_node);
1002                 BUG_ON(rb_node);
1003                 list_add_tail(&node->lower, &cache->leaves);
1004         }
1005
1006         list_for_each_entry(edge, &node->upper, list[LOWER])
1007                 list_add_tail(&edge->list[UPPER], &list);
1008
1009         while (!list_empty(&list)) {
1010                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1011                 list_del_init(&edge->list[UPPER]);
1012                 upper = edge->node[UPPER];
1013                 if (upper->detached) {
1014                         list_del(&edge->list[LOWER]);
1015                         lower = edge->node[LOWER];
1016                         free_backref_edge(cache, edge);
1017                         if (list_empty(&lower->upper))
1018                                 list_add(&lower->list, &useless);
1019                         continue;
1020                 }
1021
1022                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1023                         if (upper->lowest) {
1024                                 list_del_init(&upper->lower);
1025                                 upper->lowest = 0;
1026                         }
1027
1028                         list_add_tail(&edge->list[UPPER], &upper->lower);
1029                         continue;
1030                 }
1031
1032                 BUG_ON(!upper->checked);
1033                 BUG_ON(cowonly != upper->cowonly);
1034                 if (!cowonly) {
1035                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1036                                               &upper->rb_node);
1037                         BUG_ON(rb_node);
1038                 }
1039
1040                 list_add_tail(&edge->list[UPPER], &upper->lower);
1041
1042                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1043                         list_add_tail(&edge->list[UPPER], &list);
1044         }
1045         /*
1046          * process useless backref nodes. backref nodes for tree leaves
1047          * are deleted from the cache. backref nodes for upper level
1048          * tree blocks are left in the cache to avoid unnecessary backref
1049          * lookup.
1050          */
1051         while (!list_empty(&useless)) {
1052                 upper = list_entry(useless.next, struct backref_node, list);
1053                 list_del_init(&upper->list);
1054                 BUG_ON(!list_empty(&upper->upper));
1055                 if (upper == node)
1056                         node = NULL;
1057                 if (upper->lowest) {
1058                         list_del_init(&upper->lower);
1059                         upper->lowest = 0;
1060                 }
1061                 while (!list_empty(&upper->lower)) {
1062                         edge = list_entry(upper->lower.next,
1063                                           struct backref_edge, list[UPPER]);
1064                         list_del(&edge->list[UPPER]);
1065                         list_del(&edge->list[LOWER]);
1066                         lower = edge->node[LOWER];
1067                         free_backref_edge(cache, edge);
1068
1069                         if (list_empty(&lower->upper))
1070                                 list_add(&lower->list, &useless);
1071                 }
1072                 __mark_block_processed(rc, upper);
1073                 if (upper->level > 0) {
1074                         list_add(&upper->list, &cache->detached);
1075                         upper->detached = 1;
1076                 } else {
1077                         rb_erase(&upper->rb_node, &cache->rb_root);
1078                         free_backref_node(cache, upper);
1079                 }
1080         }
1081 out:
1082         btrfs_free_path(path1);
1083         btrfs_free_path(path2);
1084         if (err) {
1085                 while (!list_empty(&useless)) {
1086                         lower = list_entry(useless.next,
1087                                            struct backref_node, upper);
1088                         list_del_init(&lower->upper);
1089                 }
1090                 upper = node;
1091                 INIT_LIST_HEAD(&list);
1092                 while (upper) {
1093                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1094                                 list_splice_tail(&upper->upper, &list);
1095                                 free_backref_node(cache, upper);
1096                         }
1097
1098                         if (list_empty(&list))
1099                                 break;
1100
1101                         edge = list_entry(list.next, struct backref_edge,
1102                                           list[LOWER]);
1103                         list_del(&edge->list[LOWER]);
1104                         upper = edge->node[UPPER];
1105                         free_backref_edge(cache, edge);
1106                 }
1107                 return ERR_PTR(err);
1108         }
1109         BUG_ON(node && node->detached);
1110         return node;
1111 }
1112
1113 /*
1114  * helper to add backref node for the newly created snapshot.
1115  * the backref node is created by cloning backref node that
1116  * corresponds to root of source tree
1117  */
1118 static int clone_backref_node(struct btrfs_trans_handle *trans,
1119                               struct reloc_control *rc,
1120                               struct btrfs_root *src,
1121                               struct btrfs_root *dest)
1122 {
1123         struct btrfs_root *reloc_root = src->reloc_root;
1124         struct backref_cache *cache = &rc->backref_cache;
1125         struct backref_node *node = NULL;
1126         struct backref_node *new_node;
1127         struct backref_edge *edge;
1128         struct backref_edge *new_edge;
1129         struct rb_node *rb_node;
1130
1131         if (cache->last_trans > 0)
1132                 update_backref_cache(trans, cache);
1133
1134         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1135         if (rb_node) {
1136                 node = rb_entry(rb_node, struct backref_node, rb_node);
1137                 if (node->detached)
1138                         node = NULL;
1139                 else
1140                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1141         }
1142
1143         if (!node) {
1144                 rb_node = tree_search(&cache->rb_root,
1145                                       reloc_root->commit_root->start);
1146                 if (rb_node) {
1147                         node = rb_entry(rb_node, struct backref_node,
1148                                         rb_node);
1149                         BUG_ON(node->detached);
1150                 }
1151         }
1152
1153         if (!node)
1154                 return 0;
1155
1156         new_node = alloc_backref_node(cache);
1157         if (!new_node)
1158                 return -ENOMEM;
1159
1160         new_node->bytenr = dest->node->start;
1161         new_node->level = node->level;
1162         new_node->lowest = node->lowest;
1163         new_node->checked = 1;
1164         new_node->root = dest;
1165
1166         if (!node->lowest) {
1167                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1168                         new_edge = alloc_backref_edge(cache);
1169                         if (!new_edge)
1170                                 goto fail;
1171
1172                         new_edge->node[UPPER] = new_node;
1173                         new_edge->node[LOWER] = edge->node[LOWER];
1174                         list_add_tail(&new_edge->list[UPPER],
1175                                       &new_node->lower);
1176                 }
1177         } else {
1178                 list_add_tail(&new_node->lower, &cache->leaves);
1179         }
1180
1181         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1182                               &new_node->rb_node);
1183         BUG_ON(rb_node);
1184
1185         if (!new_node->lowest) {
1186                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1187                         list_add_tail(&new_edge->list[LOWER],
1188                                       &new_edge->node[LOWER]->upper);
1189                 }
1190         }
1191         return 0;
1192 fail:
1193         while (!list_empty(&new_node->lower)) {
1194                 new_edge = list_entry(new_node->lower.next,
1195                                       struct backref_edge, list[UPPER]);
1196                 list_del(&new_edge->list[UPPER]);
1197                 free_backref_edge(cache, new_edge);
1198         }
1199         free_backref_node(cache, new_node);
1200         return -ENOMEM;
1201 }
1202
1203 /*
1204  * helper to add 'address of tree root -> reloc tree' mapping
1205  */
1206 static int __add_reloc_root(struct btrfs_root *root)
1207 {
1208         struct rb_node *rb_node;
1209         struct mapping_node *node;
1210         struct reloc_control *rc = root->fs_info->reloc_ctl;
1211
1212         node = kmalloc(sizeof(*node), GFP_NOFS);
1213         BUG_ON(!node);
1214
1215         node->bytenr = root->node->start;
1216         node->data = root;
1217
1218         spin_lock(&rc->reloc_root_tree.lock);
1219         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1220                               node->bytenr, &node->rb_node);
1221         spin_unlock(&rc->reloc_root_tree.lock);
1222         BUG_ON(rb_node);
1223
1224         list_add_tail(&root->root_list, &rc->reloc_roots);
1225         return 0;
1226 }
1227
1228 /*
1229  * helper to update/delete the 'address of tree root -> reloc tree'
1230  * mapping
1231  */
1232 static int __update_reloc_root(struct btrfs_root *root, int del)
1233 {
1234         struct rb_node *rb_node;
1235         struct mapping_node *node = NULL;
1236         struct reloc_control *rc = root->fs_info->reloc_ctl;
1237
1238         spin_lock(&rc->reloc_root_tree.lock);
1239         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1240                               root->commit_root->start);
1241         if (rb_node) {
1242                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1243                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1244         }
1245         spin_unlock(&rc->reloc_root_tree.lock);
1246
1247         BUG_ON((struct btrfs_root *)node->data != root);
1248
1249         if (!del) {
1250                 spin_lock(&rc->reloc_root_tree.lock);
1251                 node->bytenr = root->node->start;
1252                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1253                                       node->bytenr, &node->rb_node);
1254                 spin_unlock(&rc->reloc_root_tree.lock);
1255                 BUG_ON(rb_node);
1256         } else {
1257                 list_del_init(&root->root_list);
1258                 kfree(node);
1259         }
1260         return 0;
1261 }
1262
1263 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1264                                         struct btrfs_root *root, u64 objectid)
1265 {
1266         struct btrfs_root *reloc_root;
1267         struct extent_buffer *eb;
1268         struct btrfs_root_item *root_item;
1269         struct btrfs_key root_key;
1270         int ret;
1271
1272         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1273         BUG_ON(!root_item);
1274
1275         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1276         root_key.type = BTRFS_ROOT_ITEM_KEY;
1277         root_key.offset = objectid;
1278
1279         if (root->root_key.objectid == objectid) {
1280                 /* called by btrfs_init_reloc_root */
1281                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1282                                       BTRFS_TREE_RELOC_OBJECTID);
1283                 BUG_ON(ret);
1284
1285                 btrfs_set_root_last_snapshot(&root->root_item,
1286                                              trans->transid - 1);
1287         } else {
1288                 /*
1289                  * called by btrfs_reloc_post_snapshot_hook.
1290                  * the source tree is a reloc tree, all tree blocks
1291                  * modified after it was created have RELOC flag
1292                  * set in their headers. so it's OK to not update
1293                  * the 'last_snapshot'.
1294                  */
1295                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1296                                       BTRFS_TREE_RELOC_OBJECTID);
1297                 BUG_ON(ret);
1298         }
1299
1300         memcpy(root_item, &root->root_item, sizeof(*root_item));
1301         btrfs_set_root_bytenr(root_item, eb->start);
1302         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1303         btrfs_set_root_generation(root_item, trans->transid);
1304
1305         if (root->root_key.objectid == objectid) {
1306                 btrfs_set_root_refs(root_item, 0);
1307                 memset(&root_item->drop_progress, 0,
1308                        sizeof(struct btrfs_disk_key));
1309                 root_item->drop_level = 0;
1310         }
1311
1312         btrfs_tree_unlock(eb);
1313         free_extent_buffer(eb);
1314
1315         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1316                                 &root_key, root_item);
1317         BUG_ON(ret);
1318         kfree(root_item);
1319
1320         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1321                                                  &root_key);
1322         BUG_ON(IS_ERR(reloc_root));
1323         reloc_root->last_trans = trans->transid;
1324         return reloc_root;
1325 }
1326
1327 /*
1328  * create reloc tree for a given fs tree. reloc tree is just a
1329  * snapshot of the fs tree with special root objectid.
1330  */
1331 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1332                           struct btrfs_root *root)
1333 {
1334         struct btrfs_root *reloc_root;
1335         struct reloc_control *rc = root->fs_info->reloc_ctl;
1336         int clear_rsv = 0;
1337
1338         if (root->reloc_root) {
1339                 reloc_root = root->reloc_root;
1340                 reloc_root->last_trans = trans->transid;
1341                 return 0;
1342         }
1343
1344         if (!rc || !rc->create_reloc_tree ||
1345             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1346                 return 0;
1347
1348         if (!trans->block_rsv) {
1349                 trans->block_rsv = rc->block_rsv;
1350                 clear_rsv = 1;
1351         }
1352         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1353         if (clear_rsv)
1354                 trans->block_rsv = NULL;
1355
1356         __add_reloc_root(reloc_root);
1357         root->reloc_root = reloc_root;
1358         return 0;
1359 }
1360
1361 /*
1362  * update root item of reloc tree
1363  */
1364 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1365                             struct btrfs_root *root)
1366 {
1367         struct btrfs_root *reloc_root;
1368         struct btrfs_root_item *root_item;
1369         int del = 0;
1370         int ret;
1371
1372         if (!root->reloc_root)
1373                 goto out;
1374
1375         reloc_root = root->reloc_root;
1376         root_item = &reloc_root->root_item;
1377
1378         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1379             btrfs_root_refs(root_item) == 0) {
1380                 root->reloc_root = NULL;
1381                 del = 1;
1382         }
1383
1384         __update_reloc_root(reloc_root, del);
1385
1386         if (reloc_root->commit_root != reloc_root->node) {
1387                 btrfs_set_root_node(root_item, reloc_root->node);
1388                 free_extent_buffer(reloc_root->commit_root);
1389                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1390         }
1391
1392         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1393                                 &reloc_root->root_key, root_item);
1394         BUG_ON(ret);
1395
1396 out:
1397         return 0;
1398 }
1399
1400 /*
1401  * helper to find first cached inode with inode number >= objectid
1402  * in a subvolume
1403  */
1404 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1405 {
1406         struct rb_node *node;
1407         struct rb_node *prev;
1408         struct btrfs_inode *entry;
1409         struct inode *inode;
1410
1411         spin_lock(&root->inode_lock);
1412 again:
1413         node = root->inode_tree.rb_node;
1414         prev = NULL;
1415         while (node) {
1416                 prev = node;
1417                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1418
1419                 if (objectid < btrfs_ino(&entry->vfs_inode))
1420                         node = node->rb_left;
1421                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1422                         node = node->rb_right;
1423                 else
1424                         break;
1425         }
1426         if (!node) {
1427                 while (prev) {
1428                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1429                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1430                                 node = prev;
1431                                 break;
1432                         }
1433                         prev = rb_next(prev);
1434                 }
1435         }
1436         while (node) {
1437                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1438                 inode = igrab(&entry->vfs_inode);
1439                 if (inode) {
1440                         spin_unlock(&root->inode_lock);
1441                         return inode;
1442                 }
1443
1444                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1445                 if (cond_resched_lock(&root->inode_lock))
1446                         goto again;
1447
1448                 node = rb_next(node);
1449         }
1450         spin_unlock(&root->inode_lock);
1451         return NULL;
1452 }
1453
1454 static int in_block_group(u64 bytenr,
1455                           struct btrfs_block_group_cache *block_group)
1456 {
1457         if (bytenr >= block_group->key.objectid &&
1458             bytenr < block_group->key.objectid + block_group->key.offset)
1459                 return 1;
1460         return 0;
1461 }
1462
1463 /*
1464  * get new location of data
1465  */
1466 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1467                             u64 bytenr, u64 num_bytes)
1468 {
1469         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1470         struct btrfs_path *path;
1471         struct btrfs_file_extent_item *fi;
1472         struct extent_buffer *leaf;
1473         int ret;
1474
1475         path = btrfs_alloc_path();
1476         if (!path)
1477                 return -ENOMEM;
1478
1479         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1480         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1481                                        bytenr, 0);
1482         if (ret < 0)
1483                 goto out;
1484         if (ret > 0) {
1485                 ret = -ENOENT;
1486                 goto out;
1487         }
1488
1489         leaf = path->nodes[0];
1490         fi = btrfs_item_ptr(leaf, path->slots[0],
1491                             struct btrfs_file_extent_item);
1492
1493         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1494                btrfs_file_extent_compression(leaf, fi) ||
1495                btrfs_file_extent_encryption(leaf, fi) ||
1496                btrfs_file_extent_other_encoding(leaf, fi));
1497
1498         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1499                 ret = 1;
1500                 goto out;
1501         }
1502
1503         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1504         ret = 0;
1505 out:
1506         btrfs_free_path(path);
1507         return ret;
1508 }
1509
1510 /*
1511  * update file extent items in the tree leaf to point to
1512  * the new locations.
1513  */
1514 static noinline_for_stack
1515 int replace_file_extents(struct btrfs_trans_handle *trans,
1516                          struct reloc_control *rc,
1517                          struct btrfs_root *root,
1518                          struct extent_buffer *leaf)
1519 {
1520         struct btrfs_key key;
1521         struct btrfs_file_extent_item *fi;
1522         struct inode *inode = NULL;
1523         u64 parent;
1524         u64 bytenr;
1525         u64 new_bytenr = 0;
1526         u64 num_bytes;
1527         u64 end;
1528         u32 nritems;
1529         u32 i;
1530         int ret;
1531         int first = 1;
1532         int dirty = 0;
1533
1534         if (rc->stage != UPDATE_DATA_PTRS)
1535                 return 0;
1536
1537         /* reloc trees always use full backref */
1538         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1539                 parent = leaf->start;
1540         else
1541                 parent = 0;
1542
1543         nritems = btrfs_header_nritems(leaf);
1544         for (i = 0; i < nritems; i++) {
1545                 cond_resched();
1546                 btrfs_item_key_to_cpu(leaf, &key, i);
1547                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1548                         continue;
1549                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1550                 if (btrfs_file_extent_type(leaf, fi) ==
1551                     BTRFS_FILE_EXTENT_INLINE)
1552                         continue;
1553                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1554                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1555                 if (bytenr == 0)
1556                         continue;
1557                 if (!in_block_group(bytenr, rc->block_group))
1558                         continue;
1559
1560                 /*
1561                  * if we are modifying block in fs tree, wait for readpage
1562                  * to complete and drop the extent cache
1563                  */
1564                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1565                         if (first) {
1566                                 inode = find_next_inode(root, key.objectid);
1567                                 first = 0;
1568                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1569                                 btrfs_add_delayed_iput(inode);
1570                                 inode = find_next_inode(root, key.objectid);
1571                         }
1572                         if (inode && btrfs_ino(inode) == key.objectid) {
1573                                 end = key.offset +
1574                                       btrfs_file_extent_num_bytes(leaf, fi);
1575                                 WARN_ON(!IS_ALIGNED(key.offset,
1576                                                     root->sectorsize));
1577                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1578                                 end--;
1579                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1580                                                       key.offset, end,
1581                                                       GFP_NOFS);
1582                                 if (!ret)
1583                                         continue;
1584
1585                                 btrfs_drop_extent_cache(inode, key.offset, end,
1586                                                         1);
1587                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1588                                               key.offset, end, GFP_NOFS);
1589                         }
1590                 }
1591
1592                 ret = get_new_location(rc->data_inode, &new_bytenr,
1593                                        bytenr, num_bytes);
1594                 if (ret > 0) {
1595                         WARN_ON(1);
1596                         continue;
1597                 }
1598                 BUG_ON(ret < 0);
1599
1600                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1601                 dirty = 1;
1602
1603                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1604                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1605                                            num_bytes, parent,
1606                                            btrfs_header_owner(leaf),
1607                                            key.objectid, key.offset);
1608                 BUG_ON(ret);
1609
1610                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1611                                         parent, btrfs_header_owner(leaf),
1612                                         key.objectid, key.offset);
1613                 BUG_ON(ret);
1614         }
1615         if (dirty)
1616                 btrfs_mark_buffer_dirty(leaf);
1617         if (inode)
1618                 btrfs_add_delayed_iput(inode);
1619         return 0;
1620 }
1621
1622 static noinline_for_stack
1623 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1624                      struct btrfs_path *path, int level)
1625 {
1626         struct btrfs_disk_key key1;
1627         struct btrfs_disk_key key2;
1628         btrfs_node_key(eb, &key1, slot);
1629         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1630         return memcmp(&key1, &key2, sizeof(key1));
1631 }
1632
1633 /*
1634  * try to replace tree blocks in fs tree with the new blocks
1635  * in reloc tree. tree blocks haven't been modified since the
1636  * reloc tree was create can be replaced.
1637  *
1638  * if a block was replaced, level of the block + 1 is returned.
1639  * if no block got replaced, 0 is returned. if there are other
1640  * errors, a negative error number is returned.
1641  */
1642 static noinline_for_stack
1643 int replace_path(struct btrfs_trans_handle *trans,
1644                  struct btrfs_root *dest, struct btrfs_root *src,
1645                  struct btrfs_path *path, struct btrfs_key *next_key,
1646                  int lowest_level, int max_level)
1647 {
1648         struct extent_buffer *eb;
1649         struct extent_buffer *parent;
1650         struct btrfs_key key;
1651         u64 old_bytenr;
1652         u64 new_bytenr;
1653         u64 old_ptr_gen;
1654         u64 new_ptr_gen;
1655         u64 last_snapshot;
1656         u32 blocksize;
1657         int cow = 0;
1658         int level;
1659         int ret;
1660         int slot;
1661
1662         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1663         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1664
1665         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1666 again:
1667         slot = path->slots[lowest_level];
1668         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1669
1670         eb = btrfs_lock_root_node(dest);
1671         btrfs_set_lock_blocking(eb);
1672         level = btrfs_header_level(eb);
1673
1674         if (level < lowest_level) {
1675                 btrfs_tree_unlock(eb);
1676                 free_extent_buffer(eb);
1677                 return 0;
1678         }
1679
1680         if (cow) {
1681                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1682                 BUG_ON(ret);
1683         }
1684         btrfs_set_lock_blocking(eb);
1685
1686         if (next_key) {
1687                 next_key->objectid = (u64)-1;
1688                 next_key->type = (u8)-1;
1689                 next_key->offset = (u64)-1;
1690         }
1691
1692         parent = eb;
1693         while (1) {
1694                 level = btrfs_header_level(parent);
1695                 BUG_ON(level < lowest_level);
1696
1697                 ret = btrfs_bin_search(parent, &key, level, &slot);
1698                 if (ret && slot > 0)
1699                         slot--;
1700
1701                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1702                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1703
1704                 old_bytenr = btrfs_node_blockptr(parent, slot);
1705                 blocksize = btrfs_level_size(dest, level - 1);
1706                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1707
1708                 if (level <= max_level) {
1709                         eb = path->nodes[level];
1710                         new_bytenr = btrfs_node_blockptr(eb,
1711                                                         path->slots[level]);
1712                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1713                                                         path->slots[level]);
1714                 } else {
1715                         new_bytenr = 0;
1716                         new_ptr_gen = 0;
1717                 }
1718
1719                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1720                         WARN_ON(1);
1721                         ret = level;
1722                         break;
1723                 }
1724
1725                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1726                     memcmp_node_keys(parent, slot, path, level)) {
1727                         if (level <= lowest_level) {
1728                                 ret = 0;
1729                                 break;
1730                         }
1731
1732                         eb = read_tree_block(dest, old_bytenr, blocksize,
1733                                              old_ptr_gen);
1734                         BUG_ON(!eb);
1735                         btrfs_tree_lock(eb);
1736                         if (cow) {
1737                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1738                                                       slot, &eb);
1739                                 BUG_ON(ret);
1740                         }
1741                         btrfs_set_lock_blocking(eb);
1742
1743                         btrfs_tree_unlock(parent);
1744                         free_extent_buffer(parent);
1745
1746                         parent = eb;
1747                         continue;
1748                 }
1749
1750                 if (!cow) {
1751                         btrfs_tree_unlock(parent);
1752                         free_extent_buffer(parent);
1753                         cow = 1;
1754                         goto again;
1755                 }
1756
1757                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1758                                       path->slots[level]);
1759                 btrfs_release_path(path);
1760
1761                 path->lowest_level = level;
1762                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1763                 path->lowest_level = 0;
1764                 BUG_ON(ret);
1765
1766                 /*
1767                  * swap blocks in fs tree and reloc tree.
1768                  */
1769                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1770                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1771                 btrfs_mark_buffer_dirty(parent);
1772
1773                 btrfs_set_node_blockptr(path->nodes[level],
1774                                         path->slots[level], old_bytenr);
1775                 btrfs_set_node_ptr_generation(path->nodes[level],
1776                                               path->slots[level], old_ptr_gen);
1777                 btrfs_mark_buffer_dirty(path->nodes[level]);
1778
1779                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1780                                         path->nodes[level]->start,
1781                                         src->root_key.objectid, level - 1, 0);
1782                 BUG_ON(ret);
1783                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1784                                         0, dest->root_key.objectid, level - 1,
1785                                         0);
1786                 BUG_ON(ret);
1787
1788                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1789                                         path->nodes[level]->start,
1790                                         src->root_key.objectid, level - 1, 0);
1791                 BUG_ON(ret);
1792
1793                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1794                                         0, dest->root_key.objectid, level - 1,
1795                                         0);
1796                 BUG_ON(ret);
1797
1798                 btrfs_unlock_up_safe(path, 0);
1799
1800                 ret = level;
1801                 break;
1802         }
1803         btrfs_tree_unlock(parent);
1804         free_extent_buffer(parent);
1805         return ret;
1806 }
1807
1808 /*
1809  * helper to find next relocated block in reloc tree
1810  */
1811 static noinline_for_stack
1812 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1813                        int *level)
1814 {
1815         struct extent_buffer *eb;
1816         int i;
1817         u64 last_snapshot;
1818         u32 nritems;
1819
1820         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1821
1822         for (i = 0; i < *level; i++) {
1823                 free_extent_buffer(path->nodes[i]);
1824                 path->nodes[i] = NULL;
1825         }
1826
1827         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1828                 eb = path->nodes[i];
1829                 nritems = btrfs_header_nritems(eb);
1830                 while (path->slots[i] + 1 < nritems) {
1831                         path->slots[i]++;
1832                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1833                             last_snapshot)
1834                                 continue;
1835
1836                         *level = i;
1837                         return 0;
1838                 }
1839                 free_extent_buffer(path->nodes[i]);
1840                 path->nodes[i] = NULL;
1841         }
1842         return 1;
1843 }
1844
1845 /*
1846  * walk down reloc tree to find relocated block of lowest level
1847  */
1848 static noinline_for_stack
1849 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1850                          int *level)
1851 {
1852         struct extent_buffer *eb = NULL;
1853         int i;
1854         u64 bytenr;
1855         u64 ptr_gen = 0;
1856         u64 last_snapshot;
1857         u32 blocksize;
1858         u32 nritems;
1859
1860         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1861
1862         for (i = *level; i > 0; i--) {
1863                 eb = path->nodes[i];
1864                 nritems = btrfs_header_nritems(eb);
1865                 while (path->slots[i] < nritems) {
1866                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1867                         if (ptr_gen > last_snapshot)
1868                                 break;
1869                         path->slots[i]++;
1870                 }
1871                 if (path->slots[i] >= nritems) {
1872                         if (i == *level)
1873                                 break;
1874                         *level = i + 1;
1875                         return 0;
1876                 }
1877                 if (i == 1) {
1878                         *level = i;
1879                         return 0;
1880                 }
1881
1882                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1883                 blocksize = btrfs_level_size(root, i - 1);
1884                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1885                 BUG_ON(btrfs_header_level(eb) != i - 1);
1886                 path->nodes[i - 1] = eb;
1887                 path->slots[i - 1] = 0;
1888         }
1889         return 1;
1890 }
1891
1892 /*
1893  * invalidate extent cache for file extents whose key in range of
1894  * [min_key, max_key)
1895  */
1896 static int invalidate_extent_cache(struct btrfs_root *root,
1897                                    struct btrfs_key *min_key,
1898                                    struct btrfs_key *max_key)
1899 {
1900         struct inode *inode = NULL;
1901         u64 objectid;
1902         u64 start, end;
1903         u64 ino;
1904
1905         objectid = min_key->objectid;
1906         while (1) {
1907                 cond_resched();
1908                 iput(inode);
1909
1910                 if (objectid > max_key->objectid)
1911                         break;
1912
1913                 inode = find_next_inode(root, objectid);
1914                 if (!inode)
1915                         break;
1916                 ino = btrfs_ino(inode);
1917
1918                 if (ino > max_key->objectid) {
1919                         iput(inode);
1920                         break;
1921                 }
1922
1923                 objectid = ino + 1;
1924                 if (!S_ISREG(inode->i_mode))
1925                         continue;
1926
1927                 if (unlikely(min_key->objectid == ino)) {
1928                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1929                                 continue;
1930                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1931                                 start = 0;
1932                         else {
1933                                 start = min_key->offset;
1934                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1935                         }
1936                 } else {
1937                         start = 0;
1938                 }
1939
1940                 if (unlikely(max_key->objectid == ino)) {
1941                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1942                                 continue;
1943                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1944                                 end = (u64)-1;
1945                         } else {
1946                                 if (max_key->offset == 0)
1947                                         continue;
1948                                 end = max_key->offset;
1949                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1950                                 end--;
1951                         }
1952                 } else {
1953                         end = (u64)-1;
1954                 }
1955
1956                 /* the lock_extent waits for readpage to complete */
1957                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1958                 btrfs_drop_extent_cache(inode, start, end, 1);
1959                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1960         }
1961         return 0;
1962 }
1963
1964 static int find_next_key(struct btrfs_path *path, int level,
1965                          struct btrfs_key *key)
1966
1967 {
1968         while (level < BTRFS_MAX_LEVEL) {
1969                 if (!path->nodes[level])
1970                         break;
1971                 if (path->slots[level] + 1 <
1972                     btrfs_header_nritems(path->nodes[level])) {
1973                         btrfs_node_key_to_cpu(path->nodes[level], key,
1974                                               path->slots[level] + 1);
1975                         return 0;
1976                 }
1977                 level++;
1978         }
1979         return 1;
1980 }
1981
1982 /*
1983  * merge the relocated tree blocks in reloc tree with corresponding
1984  * fs tree.
1985  */
1986 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1987                                                struct btrfs_root *root)
1988 {
1989         LIST_HEAD(inode_list);
1990         struct btrfs_key key;
1991         struct btrfs_key next_key;
1992         struct btrfs_trans_handle *trans;
1993         struct btrfs_root *reloc_root;
1994         struct btrfs_root_item *root_item;
1995         struct btrfs_path *path;
1996         struct extent_buffer *leaf;
1997         unsigned long nr;
1998         int level;
1999         int max_level;
2000         int replaced = 0;
2001         int ret;
2002         int err = 0;
2003         u32 min_reserved;
2004
2005         path = btrfs_alloc_path();
2006         if (!path)
2007                 return -ENOMEM;
2008         path->reada = 1;
2009
2010         reloc_root = root->reloc_root;
2011         root_item = &reloc_root->root_item;
2012
2013         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2014                 level = btrfs_root_level(root_item);
2015                 extent_buffer_get(reloc_root->node);
2016                 path->nodes[level] = reloc_root->node;
2017                 path->slots[level] = 0;
2018         } else {
2019                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2020
2021                 level = root_item->drop_level;
2022                 BUG_ON(level == 0);
2023                 path->lowest_level = level;
2024                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2025                 path->lowest_level = 0;
2026                 if (ret < 0) {
2027                         btrfs_free_path(path);
2028                         return ret;
2029                 }
2030
2031                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2032                                       path->slots[level]);
2033                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2034
2035                 btrfs_unlock_up_safe(path, 0);
2036         }
2037
2038         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2039         memset(&next_key, 0, sizeof(next_key));
2040
2041         while (1) {
2042                 trans = btrfs_start_transaction(root, 0);
2043                 BUG_ON(IS_ERR(trans));
2044                 trans->block_rsv = rc->block_rsv;
2045
2046                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved);
2047                 if (ret) {
2048                         BUG_ON(ret != -EAGAIN);
2049                         ret = btrfs_commit_transaction(trans, root);
2050                         BUG_ON(ret);
2051                         continue;
2052                 }
2053
2054                 replaced = 0;
2055                 max_level = level;
2056
2057                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2058                 if (ret < 0) {
2059                         err = ret;
2060                         goto out;
2061                 }
2062                 if (ret > 0)
2063                         break;
2064
2065                 if (!find_next_key(path, level, &key) &&
2066                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2067                         ret = 0;
2068                 } else {
2069                         ret = replace_path(trans, root, reloc_root, path,
2070                                            &next_key, level, max_level);
2071                 }
2072                 if (ret < 0) {
2073                         err = ret;
2074                         goto out;
2075                 }
2076
2077                 if (ret > 0) {
2078                         level = ret;
2079                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2080                                               path->slots[level]);
2081                         replaced = 1;
2082                 }
2083
2084                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2085                 if (ret > 0)
2086                         break;
2087
2088                 BUG_ON(level == 0);
2089                 /*
2090                  * save the merging progress in the drop_progress.
2091                  * this is OK since root refs == 1 in this case.
2092                  */
2093                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2094                                path->slots[level]);
2095                 root_item->drop_level = level;
2096
2097                 nr = trans->blocks_used;
2098                 btrfs_end_transaction_throttle(trans, root);
2099
2100                 btrfs_btree_balance_dirty(root, nr);
2101
2102                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2103                         invalidate_extent_cache(root, &key, &next_key);
2104         }
2105
2106         /*
2107          * handle the case only one block in the fs tree need to be
2108          * relocated and the block is tree root.
2109          */
2110         leaf = btrfs_lock_root_node(root);
2111         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2112         btrfs_tree_unlock(leaf);
2113         free_extent_buffer(leaf);
2114         if (ret < 0)
2115                 err = ret;
2116 out:
2117         btrfs_free_path(path);
2118
2119         if (err == 0) {
2120                 memset(&root_item->drop_progress, 0,
2121                        sizeof(root_item->drop_progress));
2122                 root_item->drop_level = 0;
2123                 btrfs_set_root_refs(root_item, 0);
2124                 btrfs_update_reloc_root(trans, root);
2125         }
2126
2127         nr = trans->blocks_used;
2128         btrfs_end_transaction_throttle(trans, root);
2129
2130         btrfs_btree_balance_dirty(root, nr);
2131
2132         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2133                 invalidate_extent_cache(root, &key, &next_key);
2134
2135         return err;
2136 }
2137
2138 static noinline_for_stack
2139 int prepare_to_merge(struct reloc_control *rc, int err)
2140 {
2141         struct btrfs_root *root = rc->extent_root;
2142         struct btrfs_root *reloc_root;
2143         struct btrfs_trans_handle *trans;
2144         LIST_HEAD(reloc_roots);
2145         u64 num_bytes = 0;
2146         int ret;
2147
2148         mutex_lock(&root->fs_info->reloc_mutex);
2149         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2150         rc->merging_rsv_size += rc->nodes_relocated * 2;
2151         mutex_unlock(&root->fs_info->reloc_mutex);
2152
2153 again:
2154         if (!err) {
2155                 num_bytes = rc->merging_rsv_size;
2156                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes);
2157                 if (ret)
2158                         err = ret;
2159         }
2160
2161         trans = btrfs_join_transaction(rc->extent_root);
2162         if (IS_ERR(trans)) {
2163                 if (!err)
2164                         btrfs_block_rsv_release(rc->extent_root,
2165                                                 rc->block_rsv, num_bytes);
2166                 return PTR_ERR(trans);
2167         }
2168
2169         if (!err) {
2170                 if (num_bytes != rc->merging_rsv_size) {
2171                         btrfs_end_transaction(trans, rc->extent_root);
2172                         btrfs_block_rsv_release(rc->extent_root,
2173                                                 rc->block_rsv, num_bytes);
2174                         goto again;
2175                 }
2176         }
2177
2178         rc->merge_reloc_tree = 1;
2179
2180         while (!list_empty(&rc->reloc_roots)) {
2181                 reloc_root = list_entry(rc->reloc_roots.next,
2182                                         struct btrfs_root, root_list);
2183                 list_del_init(&reloc_root->root_list);
2184
2185                 root = read_fs_root(reloc_root->fs_info,
2186                                     reloc_root->root_key.offset);
2187                 BUG_ON(IS_ERR(root));
2188                 BUG_ON(root->reloc_root != reloc_root);
2189
2190                 /*
2191                  * set reference count to 1, so btrfs_recover_relocation
2192                  * knows it should resumes merging
2193                  */
2194                 if (!err)
2195                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2196                 btrfs_update_reloc_root(trans, root);
2197
2198                 list_add(&reloc_root->root_list, &reloc_roots);
2199         }
2200
2201         list_splice(&reloc_roots, &rc->reloc_roots);
2202
2203         if (!err)
2204                 btrfs_commit_transaction(trans, rc->extent_root);
2205         else
2206                 btrfs_end_transaction(trans, rc->extent_root);
2207         return err;
2208 }
2209
2210 static noinline_for_stack
2211 int merge_reloc_roots(struct reloc_control *rc)
2212 {
2213         struct btrfs_root *root;
2214         struct btrfs_root *reloc_root;
2215         LIST_HEAD(reloc_roots);
2216         int found = 0;
2217         int ret;
2218 again:
2219         root = rc->extent_root;
2220
2221         /*
2222          * this serializes us with btrfs_record_root_in_transaction,
2223          * we have to make sure nobody is in the middle of
2224          * adding their roots to the list while we are
2225          * doing this splice
2226          */
2227         mutex_lock(&root->fs_info->reloc_mutex);
2228         list_splice_init(&rc->reloc_roots, &reloc_roots);
2229         mutex_unlock(&root->fs_info->reloc_mutex);
2230
2231         while (!list_empty(&reloc_roots)) {
2232                 found = 1;
2233                 reloc_root = list_entry(reloc_roots.next,
2234                                         struct btrfs_root, root_list);
2235
2236                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2237                         root = read_fs_root(reloc_root->fs_info,
2238                                             reloc_root->root_key.offset);
2239                         BUG_ON(IS_ERR(root));
2240                         BUG_ON(root->reloc_root != reloc_root);
2241
2242                         ret = merge_reloc_root(rc, root);
2243                         BUG_ON(ret);
2244                 } else {
2245                         list_del_init(&reloc_root->root_list);
2246                 }
2247                 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
2248         }
2249
2250         if (found) {
2251                 found = 0;
2252                 goto again;
2253         }
2254         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2255         return 0;
2256 }
2257
2258 static void free_block_list(struct rb_root *blocks)
2259 {
2260         struct tree_block *block;
2261         struct rb_node *rb_node;
2262         while ((rb_node = rb_first(blocks))) {
2263                 block = rb_entry(rb_node, struct tree_block, rb_node);
2264                 rb_erase(rb_node, blocks);
2265                 kfree(block);
2266         }
2267 }
2268
2269 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2270                                       struct btrfs_root *reloc_root)
2271 {
2272         struct btrfs_root *root;
2273
2274         if (reloc_root->last_trans == trans->transid)
2275                 return 0;
2276
2277         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2278         BUG_ON(IS_ERR(root));
2279         BUG_ON(root->reloc_root != reloc_root);
2280
2281         return btrfs_record_root_in_trans(trans, root);
2282 }
2283
2284 static noinline_for_stack
2285 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2286                                      struct reloc_control *rc,
2287                                      struct backref_node *node,
2288                                      struct backref_edge *edges[], int *nr)
2289 {
2290         struct backref_node *next;
2291         struct btrfs_root *root;
2292         int index = 0;
2293
2294         next = node;
2295         while (1) {
2296                 cond_resched();
2297                 next = walk_up_backref(next, edges, &index);
2298                 root = next->root;
2299                 BUG_ON(!root);
2300                 BUG_ON(!root->ref_cows);
2301
2302                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2303                         record_reloc_root_in_trans(trans, root);
2304                         break;
2305                 }
2306
2307                 btrfs_record_root_in_trans(trans, root);
2308                 root = root->reloc_root;
2309
2310                 if (next->new_bytenr != root->node->start) {
2311                         BUG_ON(next->new_bytenr);
2312                         BUG_ON(!list_empty(&next->list));
2313                         next->new_bytenr = root->node->start;
2314                         next->root = root;
2315                         list_add_tail(&next->list,
2316                                       &rc->backref_cache.changed);
2317                         __mark_block_processed(rc, next);
2318                         break;
2319                 }
2320
2321                 WARN_ON(1);
2322                 root = NULL;
2323                 next = walk_down_backref(edges, &index);
2324                 if (!next || next->level <= node->level)
2325                         break;
2326         }
2327         if (!root)
2328                 return NULL;
2329
2330         *nr = index;
2331         next = node;
2332         /* setup backref node path for btrfs_reloc_cow_block */
2333         while (1) {
2334                 rc->backref_cache.path[next->level] = next;
2335                 if (--index < 0)
2336                         break;
2337                 next = edges[index]->node[UPPER];
2338         }
2339         return root;
2340 }
2341
2342 /*
2343  * select a tree root for relocation. return NULL if the block
2344  * is reference counted. we should use do_relocation() in this
2345  * case. return a tree root pointer if the block isn't reference
2346  * counted. return -ENOENT if the block is root of reloc tree.
2347  */
2348 static noinline_for_stack
2349 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2350                                    struct backref_node *node)
2351 {
2352         struct backref_node *next;
2353         struct btrfs_root *root;
2354         struct btrfs_root *fs_root = NULL;
2355         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2356         int index = 0;
2357
2358         next = node;
2359         while (1) {
2360                 cond_resched();
2361                 next = walk_up_backref(next, edges, &index);
2362                 root = next->root;
2363                 BUG_ON(!root);
2364
2365                 /* no other choice for non-references counted tree */
2366                 if (!root->ref_cows)
2367                         return root;
2368
2369                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2370                         fs_root = root;
2371
2372                 if (next != node)
2373                         return NULL;
2374
2375                 next = walk_down_backref(edges, &index);
2376                 if (!next || next->level <= node->level)
2377                         break;
2378         }
2379
2380         if (!fs_root)
2381                 return ERR_PTR(-ENOENT);
2382         return fs_root;
2383 }
2384
2385 static noinline_for_stack
2386 u64 calcu_metadata_size(struct reloc_control *rc,
2387                         struct backref_node *node, int reserve)
2388 {
2389         struct backref_node *next = node;
2390         struct backref_edge *edge;
2391         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2392         u64 num_bytes = 0;
2393         int index = 0;
2394
2395         BUG_ON(reserve && node->processed);
2396
2397         while (next) {
2398                 cond_resched();
2399                 while (1) {
2400                         if (next->processed && (reserve || next != node))
2401                                 break;
2402
2403                         num_bytes += btrfs_level_size(rc->extent_root,
2404                                                       next->level);
2405
2406                         if (list_empty(&next->upper))
2407                                 break;
2408
2409                         edge = list_entry(next->upper.next,
2410                                           struct backref_edge, list[LOWER]);
2411                         edges[index++] = edge;
2412                         next = edge->node[UPPER];
2413                 }
2414                 next = walk_down_backref(edges, &index);
2415         }
2416         return num_bytes;
2417 }
2418
2419 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2420                                   struct reloc_control *rc,
2421                                   struct backref_node *node)
2422 {
2423         struct btrfs_root *root = rc->extent_root;
2424         u64 num_bytes;
2425         int ret;
2426
2427         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2428
2429         trans->block_rsv = rc->block_rsv;
2430         ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes);
2431         if (ret) {
2432                 if (ret == -EAGAIN)
2433                         rc->commit_transaction = 1;
2434                 return ret;
2435         }
2436
2437         return 0;
2438 }
2439
2440 static void release_metadata_space(struct reloc_control *rc,
2441                                    struct backref_node *node)
2442 {
2443         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2444         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2445 }
2446
2447 /*
2448  * relocate a block tree, and then update pointers in upper level
2449  * blocks that reference the block to point to the new location.
2450  *
2451  * if called by link_to_upper, the block has already been relocated.
2452  * in that case this function just updates pointers.
2453  */
2454 static int do_relocation(struct btrfs_trans_handle *trans,
2455                          struct reloc_control *rc,
2456                          struct backref_node *node,
2457                          struct btrfs_key *key,
2458                          struct btrfs_path *path, int lowest)
2459 {
2460         struct backref_node *upper;
2461         struct backref_edge *edge;
2462         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2463         struct btrfs_root *root;
2464         struct extent_buffer *eb;
2465         u32 blocksize;
2466         u64 bytenr;
2467         u64 generation;
2468         int nr;
2469         int slot;
2470         int ret;
2471         int err = 0;
2472
2473         BUG_ON(lowest && node->eb);
2474
2475         path->lowest_level = node->level + 1;
2476         rc->backref_cache.path[node->level] = node;
2477         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2478                 cond_resched();
2479
2480                 upper = edge->node[UPPER];
2481                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2482                 BUG_ON(!root);
2483
2484                 if (upper->eb && !upper->locked) {
2485                         if (!lowest) {
2486                                 ret = btrfs_bin_search(upper->eb, key,
2487                                                        upper->level, &slot);
2488                                 BUG_ON(ret);
2489                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2490                                 if (node->eb->start == bytenr)
2491                                         goto next;
2492                         }
2493                         drop_node_buffer(upper);
2494                 }
2495
2496                 if (!upper->eb) {
2497                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2498                         if (ret < 0) {
2499                                 err = ret;
2500                                 break;
2501                         }
2502                         BUG_ON(ret > 0);
2503
2504                         if (!upper->eb) {
2505                                 upper->eb = path->nodes[upper->level];
2506                                 path->nodes[upper->level] = NULL;
2507                         } else {
2508                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2509                         }
2510
2511                         upper->locked = 1;
2512                         path->locks[upper->level] = 0;
2513
2514                         slot = path->slots[upper->level];
2515                         btrfs_release_path(path);
2516                 } else {
2517                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2518                                                &slot);
2519                         BUG_ON(ret);
2520                 }
2521
2522                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2523                 if (lowest) {
2524                         BUG_ON(bytenr != node->bytenr);
2525                 } else {
2526                         if (node->eb->start == bytenr)
2527                                 goto next;
2528                 }
2529
2530                 blocksize = btrfs_level_size(root, node->level);
2531                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2532                 eb = read_tree_block(root, bytenr, blocksize, generation);
2533                 if (!eb) {
2534                         err = -EIO;
2535                         goto next;
2536                 }
2537                 btrfs_tree_lock(eb);
2538                 btrfs_set_lock_blocking(eb);
2539
2540                 if (!node->eb) {
2541                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2542                                               slot, &eb);
2543                         btrfs_tree_unlock(eb);
2544                         free_extent_buffer(eb);
2545                         if (ret < 0) {
2546                                 err = ret;
2547                                 goto next;
2548                         }
2549                         BUG_ON(node->eb != eb);
2550                 } else {
2551                         btrfs_set_node_blockptr(upper->eb, slot,
2552                                                 node->eb->start);
2553                         btrfs_set_node_ptr_generation(upper->eb, slot,
2554                                                       trans->transid);
2555                         btrfs_mark_buffer_dirty(upper->eb);
2556
2557                         ret = btrfs_inc_extent_ref(trans, root,
2558                                                 node->eb->start, blocksize,
2559                                                 upper->eb->start,
2560                                                 btrfs_header_owner(upper->eb),
2561                                                 node->level, 0);
2562                         BUG_ON(ret);
2563
2564                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2565                         BUG_ON(ret);
2566                 }
2567 next:
2568                 if (!upper->pending)
2569                         drop_node_buffer(upper);
2570                 else
2571                         unlock_node_buffer(upper);
2572                 if (err)
2573                         break;
2574         }
2575
2576         if (!err && node->pending) {
2577                 drop_node_buffer(node);
2578                 list_move_tail(&node->list, &rc->backref_cache.changed);
2579                 node->pending = 0;
2580         }
2581
2582         path->lowest_level = 0;
2583         BUG_ON(err == -ENOSPC);
2584         return err;
2585 }
2586
2587 static int link_to_upper(struct btrfs_trans_handle *trans,
2588                          struct reloc_control *rc,
2589                          struct backref_node *node,
2590                          struct btrfs_path *path)
2591 {
2592         struct btrfs_key key;
2593
2594         btrfs_node_key_to_cpu(node->eb, &key, 0);
2595         return do_relocation(trans, rc, node, &key, path, 0);
2596 }
2597
2598 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2599                                 struct reloc_control *rc,
2600                                 struct btrfs_path *path, int err)
2601 {
2602         LIST_HEAD(list);
2603         struct backref_cache *cache = &rc->backref_cache;
2604         struct backref_node *node;
2605         int level;
2606         int ret;
2607
2608         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2609                 while (!list_empty(&cache->pending[level])) {
2610                         node = list_entry(cache->pending[level].next,
2611                                           struct backref_node, list);
2612                         list_move_tail(&node->list, &list);
2613                         BUG_ON(!node->pending);
2614
2615                         if (!err) {
2616                                 ret = link_to_upper(trans, rc, node, path);
2617                                 if (ret < 0)
2618                                         err = ret;
2619                         }
2620                 }
2621                 list_splice_init(&list, &cache->pending[level]);
2622         }
2623         return err;
2624 }
2625
2626 static void mark_block_processed(struct reloc_control *rc,
2627                                  u64 bytenr, u32 blocksize)
2628 {
2629         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2630                         EXTENT_DIRTY, GFP_NOFS);
2631 }
2632
2633 static void __mark_block_processed(struct reloc_control *rc,
2634                                    struct backref_node *node)
2635 {
2636         u32 blocksize;
2637         if (node->level == 0 ||
2638             in_block_group(node->bytenr, rc->block_group)) {
2639                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2640                 mark_block_processed(rc, node->bytenr, blocksize);
2641         }
2642         node->processed = 1;
2643 }
2644
2645 /*
2646  * mark a block and all blocks directly/indirectly reference the block
2647  * as processed.
2648  */
2649 static void update_processed_blocks(struct reloc_control *rc,
2650                                     struct backref_node *node)
2651 {
2652         struct backref_node *next = node;
2653         struct backref_edge *edge;
2654         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2655         int index = 0;
2656
2657         while (next) {
2658                 cond_resched();
2659                 while (1) {
2660                         if (next->processed)
2661                                 break;
2662
2663                         __mark_block_processed(rc, next);
2664
2665                         if (list_empty(&next->upper))
2666                                 break;
2667
2668                         edge = list_entry(next->upper.next,
2669                                           struct backref_edge, list[LOWER]);
2670                         edges[index++] = edge;
2671                         next = edge->node[UPPER];
2672                 }
2673                 next = walk_down_backref(edges, &index);
2674         }
2675 }
2676
2677 static int tree_block_processed(u64 bytenr, u32 blocksize,
2678                                 struct reloc_control *rc)
2679 {
2680         if (test_range_bit(&rc->processed_blocks, bytenr,
2681                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2682                 return 1;
2683         return 0;
2684 }
2685
2686 static int get_tree_block_key(struct reloc_control *rc,
2687                               struct tree_block *block)
2688 {
2689         struct extent_buffer *eb;
2690
2691         BUG_ON(block->key_ready);
2692         eb = read_tree_block(rc->extent_root, block->bytenr,
2693                              block->key.objectid, block->key.offset);
2694         BUG_ON(!eb);
2695         WARN_ON(btrfs_header_level(eb) != block->level);
2696         if (block->level == 0)
2697                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2698         else
2699                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2700         free_extent_buffer(eb);
2701         block->key_ready = 1;
2702         return 0;
2703 }
2704
2705 static int reada_tree_block(struct reloc_control *rc,
2706                             struct tree_block *block)
2707 {
2708         BUG_ON(block->key_ready);
2709         readahead_tree_block(rc->extent_root, block->bytenr,
2710                              block->key.objectid, block->key.offset);
2711         return 0;
2712 }
2713
2714 /*
2715  * helper function to relocate a tree block
2716  */
2717 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2718                                 struct reloc_control *rc,
2719                                 struct backref_node *node,
2720                                 struct btrfs_key *key,
2721                                 struct btrfs_path *path)
2722 {
2723         struct btrfs_root *root;
2724         int release = 0;
2725         int ret = 0;
2726
2727         if (!node)
2728                 return 0;
2729
2730         BUG_ON(node->processed);
2731         root = select_one_root(trans, node);
2732         if (root == ERR_PTR(-ENOENT)) {
2733                 update_processed_blocks(rc, node);
2734                 goto out;
2735         }
2736
2737         if (!root || root->ref_cows) {
2738                 ret = reserve_metadata_space(trans, rc, node);
2739                 if (ret)
2740                         goto out;
2741                 release = 1;
2742         }
2743
2744         if (root) {
2745                 if (root->ref_cows) {
2746                         BUG_ON(node->new_bytenr);
2747                         BUG_ON(!list_empty(&node->list));
2748                         btrfs_record_root_in_trans(trans, root);
2749                         root = root->reloc_root;
2750                         node->new_bytenr = root->node->start;
2751                         node->root = root;
2752                         list_add_tail(&node->list, &rc->backref_cache.changed);
2753                 } else {
2754                         path->lowest_level = node->level;
2755                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2756                         btrfs_release_path(path);
2757                         if (ret > 0)
2758                                 ret = 0;
2759                 }
2760                 if (!ret)
2761                         update_processed_blocks(rc, node);
2762         } else {
2763                 ret = do_relocation(trans, rc, node, key, path, 1);
2764         }
2765 out:
2766         if (ret || node->level == 0 || node->cowonly) {
2767                 if (release)
2768                         release_metadata_space(rc, node);
2769                 remove_backref_node(&rc->backref_cache, node);
2770         }
2771         return ret;
2772 }
2773
2774 /*
2775  * relocate a list of blocks
2776  */
2777 static noinline_for_stack
2778 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2779                          struct reloc_control *rc, struct rb_root *blocks)
2780 {
2781         struct backref_node *node;
2782         struct btrfs_path *path;
2783         struct tree_block *block;
2784         struct rb_node *rb_node;
2785         int ret;
2786         int err = 0;
2787
2788         path = btrfs_alloc_path();
2789         if (!path)
2790                 return -ENOMEM;
2791
2792         rb_node = rb_first(blocks);
2793         while (rb_node) {
2794                 block = rb_entry(rb_node, struct tree_block, rb_node);
2795                 if (!block->key_ready)
2796                         reada_tree_block(rc, block);
2797                 rb_node = rb_next(rb_node);
2798         }
2799
2800         rb_node = rb_first(blocks);
2801         while (rb_node) {
2802                 block = rb_entry(rb_node, struct tree_block, rb_node);
2803                 if (!block->key_ready)
2804                         get_tree_block_key(rc, block);
2805                 rb_node = rb_next(rb_node);
2806         }
2807
2808         rb_node = rb_first(blocks);
2809         while (rb_node) {
2810                 block = rb_entry(rb_node, struct tree_block, rb_node);
2811
2812                 node = build_backref_tree(rc, &block->key,
2813                                           block->level, block->bytenr);
2814                 if (IS_ERR(node)) {
2815                         err = PTR_ERR(node);
2816                         goto out;
2817                 }
2818
2819                 ret = relocate_tree_block(trans, rc, node, &block->key,
2820                                           path);
2821                 if (ret < 0) {
2822                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2823                                 err = ret;
2824                         goto out;
2825                 }
2826                 rb_node = rb_next(rb_node);
2827         }
2828 out:
2829         free_block_list(blocks);
2830         err = finish_pending_nodes(trans, rc, path, err);
2831
2832         btrfs_free_path(path);
2833         return err;
2834 }
2835
2836 static noinline_for_stack
2837 int prealloc_file_extent_cluster(struct inode *inode,
2838                                  struct file_extent_cluster *cluster)
2839 {
2840         u64 alloc_hint = 0;
2841         u64 start;
2842         u64 end;
2843         u64 offset = BTRFS_I(inode)->index_cnt;
2844         u64 num_bytes;
2845         int nr = 0;
2846         int ret = 0;
2847
2848         BUG_ON(cluster->start != cluster->boundary[0]);
2849         mutex_lock(&inode->i_mutex);
2850
2851         ret = btrfs_check_data_free_space(inode, cluster->end +
2852                                           1 - cluster->start);
2853         if (ret)
2854                 goto out;
2855
2856         while (nr < cluster->nr) {
2857                 start = cluster->boundary[nr] - offset;
2858                 if (nr + 1 < cluster->nr)
2859                         end = cluster->boundary[nr + 1] - 1 - offset;
2860                 else
2861                         end = cluster->end - offset;
2862
2863                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2864                 num_bytes = end + 1 - start;
2865                 ret = btrfs_prealloc_file_range(inode, 0, start,
2866                                                 num_bytes, num_bytes,
2867                                                 end + 1, &alloc_hint);
2868                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2869                 if (ret)
2870                         break;
2871                 nr++;
2872         }
2873         btrfs_free_reserved_data_space(inode, cluster->end +
2874                                        1 - cluster->start);
2875 out:
2876         mutex_unlock(&inode->i_mutex);
2877         return ret;
2878 }
2879
2880 static noinline_for_stack
2881 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2882                          u64 block_start)
2883 {
2884         struct btrfs_root *root = BTRFS_I(inode)->root;
2885         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2886         struct extent_map *em;
2887         int ret = 0;
2888
2889         em = alloc_extent_map();
2890         if (!em)
2891                 return -ENOMEM;
2892
2893         em->start = start;
2894         em->len = end + 1 - start;
2895         em->block_len = em->len;
2896         em->block_start = block_start;
2897         em->bdev = root->fs_info->fs_devices->latest_bdev;
2898         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2899
2900         lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2901         while (1) {
2902                 write_lock(&em_tree->lock);
2903                 ret = add_extent_mapping(em_tree, em);
2904                 write_unlock(&em_tree->lock);
2905                 if (ret != -EEXIST) {
2906                         free_extent_map(em);
2907                         break;
2908                 }
2909                 btrfs_drop_extent_cache(inode, start, end, 0);
2910         }
2911         unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2912         return ret;
2913 }
2914
2915 static int relocate_file_extent_cluster(struct inode *inode,
2916                                         struct file_extent_cluster *cluster)
2917 {
2918         u64 page_start;
2919         u64 page_end;
2920         u64 offset = BTRFS_I(inode)->index_cnt;
2921         unsigned long index;
2922         unsigned long last_index;
2923         struct page *page;
2924         struct file_ra_state *ra;
2925         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2926         int nr = 0;
2927         int ret = 0;
2928
2929         if (!cluster->nr)
2930                 return 0;
2931
2932         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2933         if (!ra)
2934                 return -ENOMEM;
2935
2936         ret = prealloc_file_extent_cluster(inode, cluster);
2937         if (ret)
2938                 goto out;
2939
2940         file_ra_state_init(ra, inode->i_mapping);
2941
2942         ret = setup_extent_mapping(inode, cluster->start - offset,
2943                                    cluster->end - offset, cluster->start);
2944         if (ret)
2945                 goto out;
2946
2947         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2948         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2949         while (index <= last_index) {
2950                 mutex_lock(&inode->i_mutex);
2951                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2952                 mutex_unlock(&inode->i_mutex);
2953                 if (ret)
2954                         goto out;
2955
2956                 page = find_lock_page(inode->i_mapping, index);
2957                 if (!page) {
2958                         page_cache_sync_readahead(inode->i_mapping,
2959                                                   ra, NULL, index,
2960                                                   last_index + 1 - index);
2961                         page = find_or_create_page(inode->i_mapping, index,
2962                                                    mask);
2963                         if (!page) {
2964                                 btrfs_delalloc_release_metadata(inode,
2965                                                         PAGE_CACHE_SIZE);
2966                                 ret = -ENOMEM;
2967                                 goto out;
2968                         }
2969                 }
2970
2971                 if (PageReadahead(page)) {
2972                         page_cache_async_readahead(inode->i_mapping,
2973                                                    ra, NULL, page, index,
2974                                                    last_index + 1 - index);
2975                 }
2976
2977                 if (!PageUptodate(page)) {
2978                         btrfs_readpage(NULL, page);
2979                         lock_page(page);
2980                         if (!PageUptodate(page)) {
2981                                 unlock_page(page);
2982                                 page_cache_release(page);
2983                                 btrfs_delalloc_release_metadata(inode,
2984                                                         PAGE_CACHE_SIZE);
2985                                 ret = -EIO;
2986                                 goto out;
2987                         }
2988                 }
2989
2990                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2991                 page_end = page_start + PAGE_CACHE_SIZE - 1;
2992
2993                 lock_extent(&BTRFS_I(inode)->io_tree,
2994                             page_start, page_end, GFP_NOFS);
2995
2996                 set_page_extent_mapped(page);
2997
2998                 if (nr < cluster->nr &&
2999                     page_start + offset == cluster->boundary[nr]) {
3000                         set_extent_bits(&BTRFS_I(inode)->io_tree,
3001                                         page_start, page_end,
3002                                         EXTENT_BOUNDARY, GFP_NOFS);
3003                         nr++;
3004                 }
3005
3006                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
3007                 set_page_dirty(page);
3008
3009                 unlock_extent(&BTRFS_I(inode)->io_tree,
3010                               page_start, page_end, GFP_NOFS);
3011                 unlock_page(page);
3012                 page_cache_release(page);
3013
3014                 index++;
3015                 balance_dirty_pages_ratelimited(inode->i_mapping);
3016                 btrfs_throttle(BTRFS_I(inode)->root);
3017         }
3018         WARN_ON(nr != cluster->nr);
3019 out:
3020         kfree(ra);
3021         return ret;
3022 }
3023
3024 static noinline_for_stack
3025 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3026                          struct file_extent_cluster *cluster)
3027 {
3028         int ret;
3029
3030         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3031                 ret = relocate_file_extent_cluster(inode, cluster);
3032                 if (ret)
3033                         return ret;
3034                 cluster->nr = 0;
3035         }
3036
3037         if (!cluster->nr)
3038                 cluster->start = extent_key->objectid;
3039         else
3040                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3041         cluster->end = extent_key->objectid + extent_key->offset - 1;
3042         cluster->boundary[cluster->nr] = extent_key->objectid;
3043         cluster->nr++;
3044
3045         if (cluster->nr >= MAX_EXTENTS) {
3046                 ret = relocate_file_extent_cluster(inode, cluster);
3047                 if (ret)
3048                         return ret;
3049                 cluster->nr = 0;
3050         }
3051         return 0;
3052 }
3053
3054 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3055 static int get_ref_objectid_v0(struct reloc_control *rc,
3056                                struct btrfs_path *path,
3057                                struct btrfs_key *extent_key,
3058                                u64 *ref_objectid, int *path_change)
3059 {
3060         struct btrfs_key key;
3061         struct extent_buffer *leaf;
3062         struct btrfs_extent_ref_v0 *ref0;
3063         int ret;
3064         int slot;
3065
3066         leaf = path->nodes[0];
3067         slot = path->slots[0];
3068         while (1) {
3069                 if (slot >= btrfs_header_nritems(leaf)) {
3070                         ret = btrfs_next_leaf(rc->extent_root, path);
3071                         if (ret < 0)
3072                                 return ret;
3073                         BUG_ON(ret > 0);
3074                         leaf = path->nodes[0];
3075                         slot = path->slots[0];
3076                         if (path_change)
3077                                 *path_change = 1;
3078                 }
3079                 btrfs_item_key_to_cpu(leaf, &key, slot);
3080                 if (key.objectid != extent_key->objectid)
3081                         return -ENOENT;
3082
3083                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3084                         slot++;
3085                         continue;
3086                 }
3087                 ref0 = btrfs_item_ptr(leaf, slot,
3088                                 struct btrfs_extent_ref_v0);
3089                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3090                 break;
3091         }
3092         return 0;
3093 }
3094 #endif
3095
3096 /*
3097  * helper to add a tree block to the list.
3098  * the major work is getting the generation and level of the block
3099  */
3100 static int add_tree_block(struct reloc_control *rc,
3101                           struct btrfs_key *extent_key,
3102                           struct btrfs_path *path,
3103                           struct rb_root *blocks)
3104 {
3105         struct extent_buffer *eb;
3106         struct btrfs_extent_item *ei;
3107         struct btrfs_tree_block_info *bi;
3108         struct tree_block *block;
3109         struct rb_node *rb_node;
3110         u32 item_size;
3111         int level = -1;
3112         int generation;
3113
3114         eb =  path->nodes[0];
3115         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3116
3117         if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3118                 ei = btrfs_item_ptr(eb, path->slots[0],
3119                                 struct btrfs_extent_item);
3120                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3121                 generation = btrfs_extent_generation(eb, ei);
3122                 level = btrfs_tree_block_level(eb, bi);
3123         } else {
3124 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3125                 u64 ref_owner;
3126                 int ret;
3127
3128                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3129                 ret = get_ref_objectid_v0(rc, path, extent_key,
3130                                           &ref_owner, NULL);
3131                 if (ret < 0)
3132                         return ret;
3133                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3134                 level = (int)ref_owner;
3135                 /* FIXME: get real generation */
3136                 generation = 0;
3137 #else
3138                 BUG();
3139 #endif
3140         }
3141
3142         btrfs_release_path(path);
3143
3144         BUG_ON(level == -1);
3145
3146         block = kmalloc(sizeof(*block), GFP_NOFS);
3147         if (!block)
3148                 return -ENOMEM;
3149
3150         block->bytenr = extent_key->objectid;
3151         block->key.objectid = extent_key->offset;
3152         block->key.offset = generation;
3153         block->level = level;
3154         block->key_ready = 0;
3155
3156         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3157         BUG_ON(rb_node);
3158
3159         return 0;
3160 }
3161
3162 /*
3163  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3164  */
3165 static int __add_tree_block(struct reloc_control *rc,
3166                             u64 bytenr, u32 blocksize,
3167                             struct rb_root *blocks)
3168 {
3169         struct btrfs_path *path;
3170         struct btrfs_key key;
3171         int ret;
3172
3173         if (tree_block_processed(bytenr, blocksize, rc))
3174                 return 0;
3175
3176         if (tree_search(blocks, bytenr))
3177                 return 0;
3178
3179         path = btrfs_alloc_path();
3180         if (!path)
3181                 return -ENOMEM;
3182
3183         key.objectid = bytenr;
3184         key.type = BTRFS_EXTENT_ITEM_KEY;
3185         key.offset = blocksize;
3186
3187         path->search_commit_root = 1;
3188         path->skip_locking = 1;
3189         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3190         if (ret < 0)
3191                 goto out;
3192         BUG_ON(ret);
3193
3194         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3195         ret = add_tree_block(rc, &key, path, blocks);
3196 out:
3197         btrfs_free_path(path);
3198         return ret;
3199 }
3200
3201 /*
3202  * helper to check if the block use full backrefs for pointers in it
3203  */
3204 static int block_use_full_backref(struct reloc_control *rc,
3205                                   struct extent_buffer *eb)
3206 {
3207         u64 flags;
3208         int ret;
3209
3210         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3211             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3212                 return 1;
3213
3214         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3215                                        eb->start, eb->len, NULL, &flags);
3216         BUG_ON(ret);
3217
3218         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3219                 ret = 1;
3220         else
3221                 ret = 0;
3222         return ret;
3223 }
3224
3225 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3226                                     struct inode *inode, u64 ino)
3227 {
3228         struct btrfs_key key;
3229         struct btrfs_path *path;
3230         struct btrfs_root *root = fs_info->tree_root;
3231         struct btrfs_trans_handle *trans;
3232         unsigned long nr;
3233         int ret = 0;
3234
3235         if (inode)
3236                 goto truncate;
3237
3238         key.objectid = ino;
3239         key.type = BTRFS_INODE_ITEM_KEY;
3240         key.offset = 0;
3241
3242         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3243         if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) {
3244                 if (inode && !IS_ERR(inode))
3245                         iput(inode);
3246                 return -ENOENT;
3247         }
3248
3249 truncate:
3250         path = btrfs_alloc_path();
3251         if (!path) {
3252                 ret = -ENOMEM;
3253                 goto out;
3254         }
3255
3256         trans = btrfs_join_transaction(root);
3257         if (IS_ERR(trans)) {
3258                 btrfs_free_path(path);
3259                 ret = PTR_ERR(trans);
3260                 goto out;
3261         }
3262
3263         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3264
3265         btrfs_free_path(path);
3266         nr = trans->blocks_used;
3267         btrfs_end_transaction(trans, root);
3268         btrfs_btree_balance_dirty(root, nr);
3269 out:
3270         iput(inode);
3271         return ret;
3272 }
3273
3274 /*
3275  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3276  * this function scans fs tree to find blocks reference the data extent
3277  */
3278 static int find_data_references(struct reloc_control *rc,
3279                                 struct btrfs_key *extent_key,
3280                                 struct extent_buffer *leaf,
3281                                 struct btrfs_extent_data_ref *ref,
3282                                 struct rb_root *blocks)
3283 {
3284         struct btrfs_path *path;
3285         struct tree_block *block;
3286         struct btrfs_root *root;
3287         struct btrfs_file_extent_item *fi;
3288         struct rb_node *rb_node;
3289         struct btrfs_key key;
3290         u64 ref_root;
3291         u64 ref_objectid;
3292         u64 ref_offset;
3293         u32 ref_count;
3294         u32 nritems;
3295         int err = 0;
3296         int added = 0;
3297         int counted;
3298         int ret;
3299
3300         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3301         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3302         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3303         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3304
3305         /*
3306          * This is an extent belonging to the free space cache, lets just delete
3307          * it and redo the search.
3308          */
3309         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3310                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3311                                                NULL, ref_objectid);
3312                 if (ret != -ENOENT)
3313                         return ret;
3314                 ret = 0;
3315         }
3316
3317         path = btrfs_alloc_path();
3318         if (!path)
3319                 return -ENOMEM;
3320         path->reada = 1;
3321
3322         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3323         if (IS_ERR(root)) {
3324                 err = PTR_ERR(root);
3325                 goto out;
3326         }
3327
3328         key.objectid = ref_objectid;
3329         key.type = BTRFS_EXTENT_DATA_KEY;
3330         if (ref_offset > ((u64)-1 << 32))
3331                 key.offset = 0;
3332         else
3333                 key.offset = ref_offset;
3334
3335         path->search_commit_root = 1;
3336         path->skip_locking = 1;
3337         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3338         if (ret < 0) {
3339                 err = ret;
3340                 goto out;
3341         }
3342
3343         leaf = path->nodes[0];
3344         nritems = btrfs_header_nritems(leaf);
3345         /*
3346          * the references in tree blocks that use full backrefs
3347          * are not counted in
3348          */
3349         if (block_use_full_backref(rc, leaf))
3350                 counted = 0;
3351         else
3352                 counted = 1;
3353         rb_node = tree_search(blocks, leaf->start);
3354         if (rb_node) {
3355                 if (counted)
3356                         added = 1;
3357                 else
3358                         path->slots[0] = nritems;
3359         }
3360
3361         while (ref_count > 0) {
3362                 while (path->slots[0] >= nritems) {
3363                         ret = btrfs_next_leaf(root, path);
3364                         if (ret < 0) {
3365                                 err = ret;
3366                                 goto out;
3367                         }
3368                         if (ret > 0) {
3369                                 WARN_ON(1);
3370                                 goto out;
3371                         }
3372
3373                         leaf = path->nodes[0];
3374                         nritems = btrfs_header_nritems(leaf);
3375                         added = 0;
3376
3377                         if (block_use_full_backref(rc, leaf))
3378                                 counted = 0;
3379                         else
3380                                 counted = 1;
3381                         rb_node = tree_search(blocks, leaf->start);
3382                         if (rb_node) {
3383                                 if (counted)
3384                                         added = 1;
3385                                 else
3386                                         path->slots[0] = nritems;
3387                         }
3388                 }
3389
3390                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3391                 if (key.objectid != ref_objectid ||
3392                     key.type != BTRFS_EXTENT_DATA_KEY) {
3393                         WARN_ON(1);
3394                         break;
3395                 }
3396
3397                 fi = btrfs_item_ptr(leaf, path->slots[0],
3398                                     struct btrfs_file_extent_item);
3399
3400                 if (btrfs_file_extent_type(leaf, fi) ==
3401                     BTRFS_FILE_EXTENT_INLINE)
3402                         goto next;
3403
3404                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3405                     extent_key->objectid)
3406                         goto next;
3407
3408                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3409                 if (key.offset != ref_offset)
3410                         goto next;
3411
3412                 if (counted)
3413                         ref_count--;
3414                 if (added)
3415                         goto next;
3416
3417                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3418                         block = kmalloc(sizeof(*block), GFP_NOFS);
3419                         if (!block) {
3420                                 err = -ENOMEM;
3421                                 break;
3422                         }
3423                         block->bytenr = leaf->start;
3424                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3425                         block->level = 0;
3426                         block->key_ready = 1;
3427                         rb_node = tree_insert(blocks, block->bytenr,
3428                                               &block->rb_node);
3429                         BUG_ON(rb_node);
3430                 }
3431                 if (counted)
3432                         added = 1;
3433                 else
3434                         path->slots[0] = nritems;
3435 next:
3436                 path->slots[0]++;
3437
3438         }
3439 out:
3440         btrfs_free_path(path);
3441         return err;
3442 }
3443
3444 /*
3445  * hepler to find all tree blocks that reference a given data extent
3446  */
3447 static noinline_for_stack
3448 int add_data_references(struct reloc_control *rc,
3449                         struct btrfs_key *extent_key,
3450                         struct btrfs_path *path,
3451                         struct rb_root *blocks)
3452 {
3453         struct btrfs_key key;
3454         struct extent_buffer *eb;
3455         struct btrfs_extent_data_ref *dref;
3456         struct btrfs_extent_inline_ref *iref;
3457         unsigned long ptr;
3458         unsigned long end;
3459         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3460         int ret;
3461         int err = 0;
3462
3463         eb = path->nodes[0];
3464         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3465         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3466 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3467         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3468                 ptr = end;
3469         else
3470 #endif
3471                 ptr += sizeof(struct btrfs_extent_item);
3472
3473         while (ptr < end) {
3474                 iref = (struct btrfs_extent_inline_ref *)ptr;
3475                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3476                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3477                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3478                         ret = __add_tree_block(rc, key.offset, blocksize,
3479                                                blocks);
3480                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3481                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3482                         ret = find_data_references(rc, extent_key,
3483                                                    eb, dref, blocks);
3484                 } else {
3485                         BUG();
3486                 }
3487                 ptr += btrfs_extent_inline_ref_size(key.type);
3488         }
3489         WARN_ON(ptr > end);
3490
3491         while (1) {
3492                 cond_resched();
3493                 eb = path->nodes[0];
3494                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3495                         ret = btrfs_next_leaf(rc->extent_root, path);
3496                         if (ret < 0) {
3497                                 err = ret;
3498                                 break;
3499                         }
3500                         if (ret > 0)
3501                                 break;
3502                         eb = path->nodes[0];
3503                 }
3504
3505                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3506                 if (key.objectid != extent_key->objectid)
3507                         break;
3508
3509 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3510                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3511                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3512 #else
3513                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3514                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3515 #endif
3516                         ret = __add_tree_block(rc, key.offset, blocksize,
3517                                                blocks);
3518                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3519                         dref = btrfs_item_ptr(eb, path->slots[0],
3520                                               struct btrfs_extent_data_ref);
3521                         ret = find_data_references(rc, extent_key,
3522                                                    eb, dref, blocks);
3523                 } else {
3524                         ret = 0;
3525                 }
3526                 if (ret) {
3527                         err = ret;
3528                         break;
3529                 }
3530                 path->slots[0]++;
3531         }
3532         btrfs_release_path(path);
3533         if (err)
3534                 free_block_list(blocks);
3535         return err;
3536 }
3537
3538 /*
3539  * hepler to find next unprocessed extent
3540  */
3541 static noinline_for_stack
3542 int find_next_extent(struct btrfs_trans_handle *trans,
3543                      struct reloc_control *rc, struct btrfs_path *path,
3544                      struct btrfs_key *extent_key)
3545 {
3546         struct btrfs_key key;
3547         struct extent_buffer *leaf;
3548         u64 start, end, last;
3549         int ret;
3550
3551         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3552         while (1) {
3553                 cond_resched();
3554                 if (rc->search_start >= last) {
3555                         ret = 1;
3556                         break;
3557                 }
3558
3559                 key.objectid = rc->search_start;
3560                 key.type = BTRFS_EXTENT_ITEM_KEY;
3561                 key.offset = 0;
3562
3563                 path->search_commit_root = 1;
3564                 path->skip_locking = 1;
3565                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3566                                         0, 0);
3567                 if (ret < 0)
3568                         break;
3569 next:
3570                 leaf = path->nodes[0];
3571                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3572                         ret = btrfs_next_leaf(rc->extent_root, path);
3573                         if (ret != 0)
3574                                 break;
3575                         leaf = path->nodes[0];
3576                 }
3577
3578                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3579                 if (key.objectid >= last) {
3580                         ret = 1;
3581                         break;
3582                 }
3583
3584                 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3585                     key.objectid + key.offset <= rc->search_start) {
3586                         path->slots[0]++;
3587                         goto next;
3588                 }
3589
3590                 ret = find_first_extent_bit(&rc->processed_blocks,
3591                                             key.objectid, &start, &end,
3592                                             EXTENT_DIRTY);
3593
3594                 if (ret == 0 && start <= key.objectid) {
3595                         btrfs_release_path(path);
3596                         rc->search_start = end + 1;
3597                 } else {
3598                         rc->search_start = key.objectid + key.offset;
3599                         memcpy(extent_key, &key, sizeof(key));
3600                         return 0;
3601                 }
3602         }
3603         btrfs_release_path(path);
3604         return ret;
3605 }
3606
3607 static void set_reloc_control(struct reloc_control *rc)
3608 {
3609         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3610
3611         mutex_lock(&fs_info->reloc_mutex);
3612         fs_info->reloc_ctl = rc;
3613         mutex_unlock(&fs_info->reloc_mutex);
3614 }
3615
3616 static void unset_reloc_control(struct reloc_control *rc)
3617 {
3618         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3619
3620         mutex_lock(&fs_info->reloc_mutex);
3621         fs_info->reloc_ctl = NULL;
3622         mutex_unlock(&fs_info->reloc_mutex);
3623 }
3624
3625 static int check_extent_flags(u64 flags)
3626 {
3627         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3628             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3629                 return 1;
3630         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3631             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3632                 return 1;
3633         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3634             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3635                 return 1;
3636         return 0;
3637 }
3638
3639 static noinline_for_stack
3640 int prepare_to_relocate(struct reloc_control *rc)
3641 {
3642         struct btrfs_trans_handle *trans;
3643         int ret;
3644
3645         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3646         if (!rc->block_rsv)
3647                 return -ENOMEM;
3648
3649         /*
3650          * reserve some space for creating reloc trees.
3651          * btrfs_init_reloc_root will use them when there
3652          * is no reservation in transaction handle.
3653          */
3654         ret = btrfs_block_rsv_add(rc->extent_root, rc->block_rsv,
3655                                   rc->extent_root->nodesize * 256);
3656         if (ret)
3657                 return ret;
3658
3659         memset(&rc->cluster, 0, sizeof(rc->cluster));
3660         rc->search_start = rc->block_group->key.objectid;
3661         rc->extents_found = 0;
3662         rc->nodes_relocated = 0;
3663         rc->merging_rsv_size = 0;
3664
3665         rc->create_reloc_tree = 1;
3666         set_reloc_control(rc);
3667
3668         trans = btrfs_join_transaction(rc->extent_root);
3669         BUG_ON(IS_ERR(trans));
3670         btrfs_commit_transaction(trans, rc->extent_root);
3671         return 0;
3672 }
3673
3674 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3675 {
3676         struct rb_root blocks = RB_ROOT;
3677         struct btrfs_key key;
3678         struct btrfs_trans_handle *trans = NULL;
3679         struct btrfs_path *path;
3680         struct btrfs_extent_item *ei;
3681         unsigned long nr;
3682         u64 flags;
3683         u32 item_size;
3684         int ret;
3685         int err = 0;
3686         int progress = 0;
3687
3688         path = btrfs_alloc_path();
3689         if (!path)
3690                 return -ENOMEM;
3691         path->reada = 1;
3692
3693         ret = prepare_to_relocate(rc);
3694         if (ret) {
3695                 err = ret;
3696                 goto out_free;
3697         }
3698
3699         while (1) {
3700                 progress++;
3701                 trans = btrfs_start_transaction(rc->extent_root, 0);
3702                 BUG_ON(IS_ERR(trans));
3703 restart:
3704                 if (update_backref_cache(trans, &rc->backref_cache)) {
3705                         btrfs_end_transaction(trans, rc->extent_root);
3706                         continue;
3707                 }
3708
3709                 ret = find_next_extent(trans, rc, path, &key);
3710                 if (ret < 0)
3711                         err = ret;
3712                 if (ret != 0)
3713                         break;
3714
3715                 rc->extents_found++;
3716
3717                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3718                                     struct btrfs_extent_item);
3719                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3720                 if (item_size >= sizeof(*ei)) {
3721                         flags = btrfs_extent_flags(path->nodes[0], ei);
3722                         ret = check_extent_flags(flags);
3723                         BUG_ON(ret);
3724
3725                 } else {
3726 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3727                         u64 ref_owner;
3728                         int path_change = 0;
3729
3730                         BUG_ON(item_size !=
3731                                sizeof(struct btrfs_extent_item_v0));
3732                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3733                                                   &path_change);
3734                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3735                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3736                         else
3737                                 flags = BTRFS_EXTENT_FLAG_DATA;
3738
3739                         if (path_change) {
3740                                 btrfs_release_path(path);
3741
3742                                 path->search_commit_root = 1;
3743                                 path->skip_locking = 1;
3744                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3745                                                         &key, path, 0, 0);
3746                                 if (ret < 0) {
3747                                         err = ret;
3748                                         break;
3749                                 }
3750                                 BUG_ON(ret > 0);
3751                         }
3752 #else
3753                         BUG();
3754 #endif
3755                 }
3756
3757                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3758                         ret = add_tree_block(rc, &key, path, &blocks);
3759                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3760                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3761                         ret = add_data_references(rc, &key, path, &blocks);
3762                 } else {
3763                         btrfs_release_path(path);
3764                         ret = 0;
3765                 }
3766                 if (ret < 0) {
3767                         err = ret;
3768                         break;
3769                 }
3770
3771                 if (!RB_EMPTY_ROOT(&blocks)) {
3772                         ret = relocate_tree_blocks(trans, rc, &blocks);
3773                         if (ret < 0) {
3774                                 if (ret != -EAGAIN) {
3775                                         err = ret;
3776                                         break;
3777                                 }
3778                                 rc->extents_found--;
3779                                 rc->search_start = key.objectid;
3780                         }
3781                 }
3782
3783                 ret = btrfs_block_rsv_check(rc->extent_root, rc->block_rsv, 5);
3784                 if (ret < 0) {
3785                         if (ret != -EAGAIN) {
3786                                 err = ret;
3787                                 WARN_ON(1);
3788                                 break;
3789                         }
3790                         rc->commit_transaction = 1;
3791                 }
3792
3793                 if (rc->commit_transaction) {
3794                         rc->commit_transaction = 0;
3795                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3796                         BUG_ON(ret);
3797                 } else {
3798                         nr = trans->blocks_used;
3799                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3800                         btrfs_btree_balance_dirty(rc->extent_root, nr);
3801                 }
3802                 trans = NULL;
3803
3804                 if (rc->stage == MOVE_DATA_EXTENTS &&
3805                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3806                         rc->found_file_extent = 1;
3807                         ret = relocate_data_extent(rc->data_inode,
3808                                                    &key, &rc->cluster);
3809                         if (ret < 0) {
3810                                 err = ret;
3811                                 break;
3812                         }
3813                 }
3814         }
3815         if (trans && progress && err == -ENOSPC) {
3816                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3817                                               rc->block_group->flags);
3818                 if (ret == 0) {
3819                         err = 0;
3820                         progress = 0;
3821                         goto restart;
3822                 }
3823         }
3824
3825         btrfs_release_path(path);
3826         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3827                           GFP_NOFS);
3828
3829         if (trans) {
3830                 nr = trans->blocks_used;
3831                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3832                 btrfs_btree_balance_dirty(rc->extent_root, nr);
3833         }
3834
3835         if (!err) {
3836                 ret = relocate_file_extent_cluster(rc->data_inode,
3837                                                    &rc->cluster);
3838                 if (ret < 0)
3839                         err = ret;
3840         }
3841
3842         rc->create_reloc_tree = 0;
3843         set_reloc_control(rc);
3844
3845         backref_cache_cleanup(&rc->backref_cache);
3846         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3847
3848         err = prepare_to_merge(rc, err);
3849
3850         merge_reloc_roots(rc);
3851
3852         rc->merge_reloc_tree = 0;
3853         unset_reloc_control(rc);
3854         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3855
3856         /* get rid of pinned extents */
3857         trans = btrfs_join_transaction(rc->extent_root);
3858         if (IS_ERR(trans))
3859                 err = PTR_ERR(trans);
3860         else
3861                 btrfs_commit_transaction(trans, rc->extent_root);
3862 out_free:
3863         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3864         btrfs_free_path(path);
3865         return err;
3866 }
3867
3868 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3869                                  struct btrfs_root *root, u64 objectid)
3870 {
3871         struct btrfs_path *path;
3872         struct btrfs_inode_item *item;
3873         struct extent_buffer *leaf;
3874         int ret;
3875
3876         path = btrfs_alloc_path();
3877         if (!path)
3878                 return -ENOMEM;
3879
3880         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3881         if (ret)
3882                 goto out;
3883
3884         leaf = path->nodes[0];
3885         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3886         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3887         btrfs_set_inode_generation(leaf, item, 1);
3888         btrfs_set_inode_size(leaf, item, 0);
3889         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3890         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3891                                           BTRFS_INODE_PREALLOC);
3892         btrfs_mark_buffer_dirty(leaf);
3893         btrfs_release_path(path);
3894 out:
3895         btrfs_free_path(path);
3896         return ret;
3897 }
3898
3899 /*
3900  * helper to create inode for data relocation.
3901  * the inode is in data relocation tree and its link count is 0
3902  */
3903 static noinline_for_stack
3904 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3905                                  struct btrfs_block_group_cache *group)
3906 {
3907         struct inode *inode = NULL;
3908         struct btrfs_trans_handle *trans;
3909         struct btrfs_root *root;
3910         struct btrfs_key key;
3911         unsigned long nr;
3912         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3913         int err = 0;
3914
3915         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3916         if (IS_ERR(root))
3917                 return ERR_CAST(root);
3918
3919         trans = btrfs_start_transaction(root, 6);
3920         if (IS_ERR(trans))
3921                 return ERR_CAST(trans);
3922
3923         err = btrfs_find_free_objectid(root, &objectid);
3924         if (err)
3925                 goto out;
3926
3927         err = __insert_orphan_inode(trans, root, objectid);
3928         BUG_ON(err);
3929
3930         key.objectid = objectid;
3931         key.type = BTRFS_INODE_ITEM_KEY;
3932         key.offset = 0;
3933         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3934         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3935         BTRFS_I(inode)->index_cnt = group->key.objectid;
3936
3937         err = btrfs_orphan_add(trans, inode);
3938 out:
3939         nr = trans->blocks_used;
3940         btrfs_end_transaction(trans, root);
3941         btrfs_btree_balance_dirty(root, nr);
3942         if (err) {
3943                 if (inode)
3944                         iput(inode);
3945                 inode = ERR_PTR(err);
3946         }
3947         return inode;
3948 }
3949
3950 static struct reloc_control *alloc_reloc_control(void)
3951 {
3952         struct reloc_control *rc;
3953
3954         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3955         if (!rc)
3956                 return NULL;
3957
3958         INIT_LIST_HEAD(&rc->reloc_roots);
3959         backref_cache_init(&rc->backref_cache);
3960         mapping_tree_init(&rc->reloc_root_tree);
3961         extent_io_tree_init(&rc->processed_blocks, NULL);
3962         return rc;
3963 }
3964
3965 /*
3966  * function to relocate all extents in a block group.
3967  */
3968 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3969 {
3970         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3971         struct reloc_control *rc;
3972         struct inode *inode;
3973         struct btrfs_path *path;
3974         int ret;
3975         int rw = 0;
3976         int err = 0;
3977
3978         rc = alloc_reloc_control();
3979         if (!rc)
3980                 return -ENOMEM;
3981
3982         rc->extent_root = extent_root;
3983
3984         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3985         BUG_ON(!rc->block_group);
3986
3987         if (!rc->block_group->ro) {
3988                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3989                 if (ret) {
3990                         err = ret;
3991                         goto out;
3992                 }
3993                 rw = 1;
3994         }
3995
3996         path = btrfs_alloc_path();
3997         if (!path) {
3998                 err = -ENOMEM;
3999                 goto out;
4000         }
4001
4002         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
4003                                         path);
4004         btrfs_free_path(path);
4005
4006         if (!IS_ERR(inode))
4007                 ret = delete_block_group_cache(fs_info, inode, 0);
4008         else
4009                 ret = PTR_ERR(inode);
4010
4011         if (ret && ret != -ENOENT) {
4012                 err = ret;
4013                 goto out;
4014         }
4015
4016         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4017         if (IS_ERR(rc->data_inode)) {
4018                 err = PTR_ERR(rc->data_inode);
4019                 rc->data_inode = NULL;
4020                 goto out;
4021         }
4022
4023         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4024                (unsigned long long)rc->block_group->key.objectid,
4025                (unsigned long long)rc->block_group->flags);
4026
4027         btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4028         btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
4029
4030         while (1) {
4031                 mutex_lock(&fs_info->cleaner_mutex);
4032
4033                 btrfs_clean_old_snapshots(fs_info->tree_root);
4034                 ret = relocate_block_group(rc);
4035
4036                 mutex_unlock(&fs_info->cleaner_mutex);
4037                 if (ret < 0) {
4038                         err = ret;
4039                         goto out;
4040                 }
4041
4042                 if (rc->extents_found == 0)
4043                         break;
4044
4045                 printk(KERN_INFO "btrfs: found %llu extents\n",
4046                         (unsigned long long)rc->extents_found);
4047
4048                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4049                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4050                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4051                                                  0, -1);
4052                         rc->stage = UPDATE_DATA_PTRS;
4053                 }
4054         }
4055
4056         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4057                                      rc->block_group->key.objectid,
4058                                      rc->block_group->key.objectid +
4059                                      rc->block_group->key.offset - 1);
4060
4061         WARN_ON(rc->block_group->pinned > 0);
4062         WARN_ON(rc->block_group->reserved > 0);
4063         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4064 out:
4065         if (err && rw)
4066                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4067         iput(rc->data_inode);
4068         btrfs_put_block_group(rc->block_group);
4069         kfree(rc);
4070         return err;
4071 }
4072
4073 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4074 {
4075         struct btrfs_trans_handle *trans;
4076         int ret;
4077
4078         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4079         BUG_ON(IS_ERR(trans));
4080
4081         memset(&root->root_item.drop_progress, 0,
4082                 sizeof(root->root_item.drop_progress));
4083         root->root_item.drop_level = 0;
4084         btrfs_set_root_refs(&root->root_item, 0);
4085         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4086                                 &root->root_key, &root->root_item);
4087         BUG_ON(ret);
4088
4089         ret = btrfs_end_transaction(trans, root->fs_info->tree_root);
4090         BUG_ON(ret);
4091         return 0;
4092 }
4093
4094 /*
4095  * recover relocation interrupted by system crash.
4096  *
4097  * this function resumes merging reloc trees with corresponding fs trees.
4098  * this is important for keeping the sharing of tree blocks
4099  */
4100 int btrfs_recover_relocation(struct btrfs_root *root)
4101 {
4102         LIST_HEAD(reloc_roots);
4103         struct btrfs_key key;
4104         struct btrfs_root *fs_root;
4105         struct btrfs_root *reloc_root;
4106         struct btrfs_path *path;
4107         struct extent_buffer *leaf;
4108         struct reloc_control *rc = NULL;
4109         struct btrfs_trans_handle *trans;
4110         int ret;
4111         int err = 0;
4112
4113         path = btrfs_alloc_path();
4114         if (!path)
4115                 return -ENOMEM;
4116         path->reada = -1;
4117
4118         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4119         key.type = BTRFS_ROOT_ITEM_KEY;
4120         key.offset = (u64)-1;
4121
4122         while (1) {
4123                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4124                                         path, 0, 0);
4125                 if (ret < 0) {
4126                         err = ret;
4127                         goto out;
4128                 }
4129                 if (ret > 0) {
4130                         if (path->slots[0] == 0)
4131                                 break;
4132                         path->slots[0]--;
4133                 }
4134                 leaf = path->nodes[0];
4135                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4136                 btrfs_release_path(path);
4137
4138                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4139                     key.type != BTRFS_ROOT_ITEM_KEY)
4140                         break;
4141
4142                 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4143                 if (IS_ERR(reloc_root)) {
4144                         err = PTR_ERR(reloc_root);
4145                         goto out;
4146                 }
4147
4148                 list_add(&reloc_root->root_list, &reloc_roots);
4149
4150                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4151                         fs_root = read_fs_root(root->fs_info,
4152                                                reloc_root->root_key.offset);
4153                         if (IS_ERR(fs_root)) {
4154                                 ret = PTR_ERR(fs_root);
4155                                 if (ret != -ENOENT) {
4156                                         err = ret;
4157                                         goto out;
4158                                 }
4159                                 mark_garbage_root(reloc_root);
4160                         }
4161                 }
4162
4163                 if (key.offset == 0)
4164                         break;
4165
4166                 key.offset--;
4167         }
4168         btrfs_release_path(path);
4169
4170         if (list_empty(&reloc_roots))
4171                 goto out;
4172
4173         rc = alloc_reloc_control();
4174         if (!rc) {
4175                 err = -ENOMEM;
4176                 goto out;
4177         }
4178
4179         rc->extent_root = root->fs_info->extent_root;
4180
4181         set_reloc_control(rc);
4182
4183         trans = btrfs_join_transaction(rc->extent_root);
4184         if (IS_ERR(trans)) {
4185                 unset_reloc_control(rc);
4186                 err = PTR_ERR(trans);
4187                 goto out_free;
4188         }
4189
4190         rc->merge_reloc_tree = 1;
4191
4192         while (!list_empty(&reloc_roots)) {
4193                 reloc_root = list_entry(reloc_roots.next,
4194                                         struct btrfs_root, root_list);
4195                 list_del(&reloc_root->root_list);
4196
4197                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4198                         list_add_tail(&reloc_root->root_list,
4199                                       &rc->reloc_roots);
4200                         continue;
4201                 }
4202
4203                 fs_root = read_fs_root(root->fs_info,
4204                                        reloc_root->root_key.offset);
4205                 BUG_ON(IS_ERR(fs_root));
4206
4207                 __add_reloc_root(reloc_root);
4208                 fs_root->reloc_root = reloc_root;
4209         }
4210
4211         btrfs_commit_transaction(trans, rc->extent_root);
4212
4213         merge_reloc_roots(rc);
4214
4215         unset_reloc_control(rc);
4216
4217         trans = btrfs_join_transaction(rc->extent_root);
4218         if (IS_ERR(trans))
4219                 err = PTR_ERR(trans);
4220         else
4221                 btrfs_commit_transaction(trans, rc->extent_root);
4222 out_free:
4223         kfree(rc);
4224 out:
4225         while (!list_empty(&reloc_roots)) {
4226                 reloc_root = list_entry(reloc_roots.next,
4227                                         struct btrfs_root, root_list);
4228                 list_del(&reloc_root->root_list);
4229                 free_extent_buffer(reloc_root->node);
4230                 free_extent_buffer(reloc_root->commit_root);
4231                 kfree(reloc_root);
4232         }
4233         btrfs_free_path(path);
4234
4235         if (err == 0) {
4236                 /* cleanup orphan inode in data relocation tree */
4237                 fs_root = read_fs_root(root->fs_info,
4238                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4239                 if (IS_ERR(fs_root))
4240                         err = PTR_ERR(fs_root);
4241                 else
4242                         err = btrfs_orphan_cleanup(fs_root);
4243         }
4244         return err;
4245 }
4246
4247 /*
4248  * helper to add ordered checksum for data relocation.
4249  *
4250  * cloning checksum properly handles the nodatasum extents.
4251  * it also saves CPU time to re-calculate the checksum.
4252  */
4253 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4254 {
4255         struct btrfs_ordered_sum *sums;
4256         struct btrfs_sector_sum *sector_sum;
4257         struct btrfs_ordered_extent *ordered;
4258         struct btrfs_root *root = BTRFS_I(inode)->root;
4259         size_t offset;
4260         int ret;
4261         u64 disk_bytenr;
4262         LIST_HEAD(list);
4263
4264         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4265         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4266
4267         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4268         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4269                                        disk_bytenr + len - 1, &list, 0);
4270
4271         while (!list_empty(&list)) {
4272                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4273                 list_del_init(&sums->list);
4274
4275                 sector_sum = sums->sums;
4276                 sums->bytenr = ordered->start;
4277
4278                 offset = 0;
4279                 while (offset < sums->len) {
4280                         sector_sum->bytenr += ordered->start - disk_bytenr;
4281                         sector_sum++;
4282                         offset += root->sectorsize;
4283                 }
4284
4285                 btrfs_add_ordered_sum(inode, ordered, sums);
4286         }
4287         btrfs_put_ordered_extent(ordered);
4288         return ret;
4289 }
4290
4291 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4292                            struct btrfs_root *root, struct extent_buffer *buf,
4293                            struct extent_buffer *cow)
4294 {
4295         struct reloc_control *rc;
4296         struct backref_node *node;
4297         int first_cow = 0;
4298         int level;
4299         int ret;
4300
4301         rc = root->fs_info->reloc_ctl;
4302         if (!rc)
4303                 return;
4304
4305         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4306                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4307
4308         level = btrfs_header_level(buf);
4309         if (btrfs_header_generation(buf) <=
4310             btrfs_root_last_snapshot(&root->root_item))
4311                 first_cow = 1;
4312
4313         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4314             rc->create_reloc_tree) {
4315                 WARN_ON(!first_cow && level == 0);
4316
4317                 node = rc->backref_cache.path[level];
4318                 BUG_ON(node->bytenr != buf->start &&
4319                        node->new_bytenr != buf->start);
4320
4321                 drop_node_buffer(node);
4322                 extent_buffer_get(cow);
4323                 node->eb = cow;
4324                 node->new_bytenr = cow->start;
4325
4326                 if (!node->pending) {
4327                         list_move_tail(&node->list,
4328                                        &rc->backref_cache.pending[level]);
4329                         node->pending = 1;
4330                 }
4331
4332                 if (first_cow)
4333                         __mark_block_processed(rc, node);
4334
4335                 if (first_cow && level > 0)
4336                         rc->nodes_relocated += buf->len;
4337         }
4338
4339         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4340                 ret = replace_file_extents(trans, rc, root, cow);
4341                 BUG_ON(ret);
4342         }
4343 }
4344
4345 /*
4346  * called before creating snapshot. it calculates metadata reservation
4347  * requried for relocating tree blocks in the snapshot
4348  */
4349 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4350                               struct btrfs_pending_snapshot *pending,
4351                               u64 *bytes_to_reserve)
4352 {
4353         struct btrfs_root *root;
4354         struct reloc_control *rc;
4355
4356         root = pending->root;
4357         if (!root->reloc_root)
4358                 return;
4359
4360         rc = root->fs_info->reloc_ctl;
4361         if (!rc->merge_reloc_tree)
4362                 return;
4363
4364         root = root->reloc_root;
4365         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4366         /*
4367          * relocation is in the stage of merging trees. the space
4368          * used by merging a reloc tree is twice the size of
4369          * relocated tree nodes in the worst case. half for cowing
4370          * the reloc tree, half for cowing the fs tree. the space
4371          * used by cowing the reloc tree will be freed after the
4372          * tree is dropped. if we create snapshot, cowing the fs
4373          * tree may use more space than it frees. so we need
4374          * reserve extra space.
4375          */
4376         *bytes_to_reserve += rc->nodes_relocated;
4377 }
4378
4379 /*
4380  * called after snapshot is created. migrate block reservation
4381  * and create reloc root for the newly created snapshot
4382  */
4383 void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4384                                struct btrfs_pending_snapshot *pending)
4385 {
4386         struct btrfs_root *root = pending->root;
4387         struct btrfs_root *reloc_root;
4388         struct btrfs_root *new_root;
4389         struct reloc_control *rc;
4390         int ret;
4391
4392         if (!root->reloc_root)
4393                 return;
4394
4395         rc = root->fs_info->reloc_ctl;
4396         rc->merging_rsv_size += rc->nodes_relocated;
4397
4398         if (rc->merge_reloc_tree) {
4399                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4400                                               rc->block_rsv,
4401                                               rc->nodes_relocated);
4402                 BUG_ON(ret);
4403         }
4404
4405         new_root = pending->snap;
4406         reloc_root = create_reloc_root(trans, root->reloc_root,
4407                                        new_root->root_key.objectid);
4408
4409         __add_reloc_root(reloc_root);
4410         new_root->reloc_root = reloc_root;
4411
4412         if (rc->create_reloc_tree) {
4413                 ret = clone_backref_node(trans, rc, root, reloc_root);
4414                 BUG_ON(ret);
4415         }
4416 }