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         }
1178
1179         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1180                               &new_node->rb_node);
1181         BUG_ON(rb_node);
1182
1183         if (!new_node->lowest) {
1184                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1185                         list_add_tail(&new_edge->list[LOWER],
1186                                       &new_edge->node[LOWER]->upper);
1187                 }
1188         }
1189         return 0;
1190 fail:
1191         while (!list_empty(&new_node->lower)) {
1192                 new_edge = list_entry(new_node->lower.next,
1193                                       struct backref_edge, list[UPPER]);
1194                 list_del(&new_edge->list[UPPER]);
1195                 free_backref_edge(cache, new_edge);
1196         }
1197         free_backref_node(cache, new_node);
1198         return -ENOMEM;
1199 }
1200
1201 /*
1202  * helper to add 'address of tree root -> reloc tree' mapping
1203  */
1204 static int __add_reloc_root(struct btrfs_root *root)
1205 {
1206         struct rb_node *rb_node;
1207         struct mapping_node *node;
1208         struct reloc_control *rc = root->fs_info->reloc_ctl;
1209
1210         node = kmalloc(sizeof(*node), GFP_NOFS);
1211         BUG_ON(!node);
1212
1213         node->bytenr = root->node->start;
1214         node->data = root;
1215
1216         spin_lock(&rc->reloc_root_tree.lock);
1217         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1218                               node->bytenr, &node->rb_node);
1219         spin_unlock(&rc->reloc_root_tree.lock);
1220         BUG_ON(rb_node);
1221
1222         list_add_tail(&root->root_list, &rc->reloc_roots);
1223         return 0;
1224 }
1225
1226 /*
1227  * helper to update/delete the 'address of tree root -> reloc tree'
1228  * mapping
1229  */
1230 static int __update_reloc_root(struct btrfs_root *root, int del)
1231 {
1232         struct rb_node *rb_node;
1233         struct mapping_node *node = NULL;
1234         struct reloc_control *rc = root->fs_info->reloc_ctl;
1235
1236         spin_lock(&rc->reloc_root_tree.lock);
1237         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1238                               root->commit_root->start);
1239         if (rb_node) {
1240                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1241                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1242         }
1243         spin_unlock(&rc->reloc_root_tree.lock);
1244
1245         BUG_ON((struct btrfs_root *)node->data != root);
1246
1247         if (!del) {
1248                 spin_lock(&rc->reloc_root_tree.lock);
1249                 node->bytenr = root->node->start;
1250                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1251                                       node->bytenr, &node->rb_node);
1252                 spin_unlock(&rc->reloc_root_tree.lock);
1253                 BUG_ON(rb_node);
1254         } else {
1255                 list_del_init(&root->root_list);
1256                 kfree(node);
1257         }
1258         return 0;
1259 }
1260
1261 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1262                                         struct btrfs_root *root, u64 objectid)
1263 {
1264         struct btrfs_root *reloc_root;
1265         struct extent_buffer *eb;
1266         struct btrfs_root_item *root_item;
1267         struct btrfs_key root_key;
1268         int ret;
1269
1270         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1271         BUG_ON(!root_item);
1272
1273         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1274         root_key.type = BTRFS_ROOT_ITEM_KEY;
1275         root_key.offset = objectid;
1276
1277         if (root->root_key.objectid == objectid) {
1278                 /* called by btrfs_init_reloc_root */
1279                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1280                                       BTRFS_TREE_RELOC_OBJECTID);
1281                 BUG_ON(ret);
1282
1283                 btrfs_set_root_last_snapshot(&root->root_item,
1284                                              trans->transid - 1);
1285         } else {
1286                 /*
1287                  * called by btrfs_reloc_post_snapshot_hook.
1288                  * the source tree is a reloc tree, all tree blocks
1289                  * modified after it was created have RELOC flag
1290                  * set in their headers. so it's OK to not update
1291                  * the 'last_snapshot'.
1292                  */
1293                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1294                                       BTRFS_TREE_RELOC_OBJECTID);
1295                 BUG_ON(ret);
1296         }
1297
1298         memcpy(root_item, &root->root_item, sizeof(*root_item));
1299         btrfs_set_root_bytenr(root_item, eb->start);
1300         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1301         btrfs_set_root_generation(root_item, trans->transid);
1302
1303         if (root->root_key.objectid == objectid) {
1304                 btrfs_set_root_refs(root_item, 0);
1305                 memset(&root_item->drop_progress, 0,
1306                        sizeof(struct btrfs_disk_key));
1307                 root_item->drop_level = 0;
1308         }
1309
1310         btrfs_tree_unlock(eb);
1311         free_extent_buffer(eb);
1312
1313         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1314                                 &root_key, root_item);
1315         BUG_ON(ret);
1316         kfree(root_item);
1317
1318         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1319                                                  &root_key);
1320         BUG_ON(IS_ERR(reloc_root));
1321         reloc_root->last_trans = trans->transid;
1322         return reloc_root;
1323 }
1324
1325 /*
1326  * create reloc tree for a given fs tree. reloc tree is just a
1327  * snapshot of the fs tree with special root objectid.
1328  */
1329 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1330                           struct btrfs_root *root)
1331 {
1332         struct btrfs_root *reloc_root;
1333         struct reloc_control *rc = root->fs_info->reloc_ctl;
1334         int clear_rsv = 0;
1335
1336         if (root->reloc_root) {
1337                 reloc_root = root->reloc_root;
1338                 reloc_root->last_trans = trans->transid;
1339                 return 0;
1340         }
1341
1342         if (!rc || !rc->create_reloc_tree ||
1343             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1344                 return 0;
1345
1346         if (!trans->block_rsv) {
1347                 trans->block_rsv = rc->block_rsv;
1348                 clear_rsv = 1;
1349         }
1350         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1351         if (clear_rsv)
1352                 trans->block_rsv = NULL;
1353
1354         __add_reloc_root(reloc_root);
1355         root->reloc_root = reloc_root;
1356         return 0;
1357 }
1358
1359 /*
1360  * update root item of reloc tree
1361  */
1362 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1363                             struct btrfs_root *root)
1364 {
1365         struct btrfs_root *reloc_root;
1366         struct btrfs_root_item *root_item;
1367         int del = 0;
1368         int ret;
1369
1370         if (!root->reloc_root)
1371                 return 0;
1372
1373         reloc_root = root->reloc_root;
1374         root_item = &reloc_root->root_item;
1375
1376         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1377             btrfs_root_refs(root_item) == 0) {
1378                 root->reloc_root = NULL;
1379                 del = 1;
1380         }
1381
1382         __update_reloc_root(reloc_root, del);
1383
1384         if (reloc_root->commit_root != reloc_root->node) {
1385                 btrfs_set_root_node(root_item, reloc_root->node);
1386                 free_extent_buffer(reloc_root->commit_root);
1387                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1388         }
1389
1390         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1391                                 &reloc_root->root_key, root_item);
1392         BUG_ON(ret);
1393         return 0;
1394 }
1395
1396 /*
1397  * helper to find first cached inode with inode number >= objectid
1398  * in a subvolume
1399  */
1400 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1401 {
1402         struct rb_node *node;
1403         struct rb_node *prev;
1404         struct btrfs_inode *entry;
1405         struct inode *inode;
1406
1407         spin_lock(&root->inode_lock);
1408 again:
1409         node = root->inode_tree.rb_node;
1410         prev = NULL;
1411         while (node) {
1412                 prev = node;
1413                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1414
1415                 if (objectid < btrfs_ino(&entry->vfs_inode))
1416                         node = node->rb_left;
1417                 else if (objectid > btrfs_ino(&entry->vfs_inode))
1418                         node = node->rb_right;
1419                 else
1420                         break;
1421         }
1422         if (!node) {
1423                 while (prev) {
1424                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1425                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
1426                                 node = prev;
1427                                 break;
1428                         }
1429                         prev = rb_next(prev);
1430                 }
1431         }
1432         while (node) {
1433                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1434                 inode = igrab(&entry->vfs_inode);
1435                 if (inode) {
1436                         spin_unlock(&root->inode_lock);
1437                         return inode;
1438                 }
1439
1440                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
1441                 if (cond_resched_lock(&root->inode_lock))
1442                         goto again;
1443
1444                 node = rb_next(node);
1445         }
1446         spin_unlock(&root->inode_lock);
1447         return NULL;
1448 }
1449
1450 static int in_block_group(u64 bytenr,
1451                           struct btrfs_block_group_cache *block_group)
1452 {
1453         if (bytenr >= block_group->key.objectid &&
1454             bytenr < block_group->key.objectid + block_group->key.offset)
1455                 return 1;
1456         return 0;
1457 }
1458
1459 /*
1460  * get new location of data
1461  */
1462 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1463                             u64 bytenr, u64 num_bytes)
1464 {
1465         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1466         struct btrfs_path *path;
1467         struct btrfs_file_extent_item *fi;
1468         struct extent_buffer *leaf;
1469         int ret;
1470
1471         path = btrfs_alloc_path();
1472         if (!path)
1473                 return -ENOMEM;
1474
1475         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1476         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(reloc_inode),
1477                                        bytenr, 0);
1478         if (ret < 0)
1479                 goto out;
1480         if (ret > 0) {
1481                 ret = -ENOENT;
1482                 goto out;
1483         }
1484
1485         leaf = path->nodes[0];
1486         fi = btrfs_item_ptr(leaf, path->slots[0],
1487                             struct btrfs_file_extent_item);
1488
1489         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1490                btrfs_file_extent_compression(leaf, fi) ||
1491                btrfs_file_extent_encryption(leaf, fi) ||
1492                btrfs_file_extent_other_encoding(leaf, fi));
1493
1494         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1495                 ret = 1;
1496                 goto out;
1497         }
1498
1499         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1500         ret = 0;
1501 out:
1502         btrfs_free_path(path);
1503         return ret;
1504 }
1505
1506 /*
1507  * update file extent items in the tree leaf to point to
1508  * the new locations.
1509  */
1510 static noinline_for_stack
1511 int replace_file_extents(struct btrfs_trans_handle *trans,
1512                          struct reloc_control *rc,
1513                          struct btrfs_root *root,
1514                          struct extent_buffer *leaf)
1515 {
1516         struct btrfs_key key;
1517         struct btrfs_file_extent_item *fi;
1518         struct inode *inode = NULL;
1519         u64 parent;
1520         u64 bytenr;
1521         u64 new_bytenr = 0;
1522         u64 num_bytes;
1523         u64 end;
1524         u32 nritems;
1525         u32 i;
1526         int ret;
1527         int first = 1;
1528         int dirty = 0;
1529
1530         if (rc->stage != UPDATE_DATA_PTRS)
1531                 return 0;
1532
1533         /* reloc trees always use full backref */
1534         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1535                 parent = leaf->start;
1536         else
1537                 parent = 0;
1538
1539         nritems = btrfs_header_nritems(leaf);
1540         for (i = 0; i < nritems; i++) {
1541                 cond_resched();
1542                 btrfs_item_key_to_cpu(leaf, &key, i);
1543                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1544                         continue;
1545                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1546                 if (btrfs_file_extent_type(leaf, fi) ==
1547                     BTRFS_FILE_EXTENT_INLINE)
1548                         continue;
1549                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1550                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1551                 if (bytenr == 0)
1552                         continue;
1553                 if (!in_block_group(bytenr, rc->block_group))
1554                         continue;
1555
1556                 /*
1557                  * if we are modifying block in fs tree, wait for readpage
1558                  * to complete and drop the extent cache
1559                  */
1560                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1561                         if (first) {
1562                                 inode = find_next_inode(root, key.objectid);
1563                                 first = 0;
1564                         } else if (inode && btrfs_ino(inode) < key.objectid) {
1565                                 btrfs_add_delayed_iput(inode);
1566                                 inode = find_next_inode(root, key.objectid);
1567                         }
1568                         if (inode && btrfs_ino(inode) == key.objectid) {
1569                                 end = key.offset +
1570                                       btrfs_file_extent_num_bytes(leaf, fi);
1571                                 WARN_ON(!IS_ALIGNED(key.offset,
1572                                                     root->sectorsize));
1573                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1574                                 end--;
1575                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1576                                                       key.offset, end,
1577                                                       GFP_NOFS);
1578                                 if (!ret)
1579                                         continue;
1580
1581                                 btrfs_drop_extent_cache(inode, key.offset, end,
1582                                                         1);
1583                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1584                                               key.offset, end, GFP_NOFS);
1585                         }
1586                 }
1587
1588                 ret = get_new_location(rc->data_inode, &new_bytenr,
1589                                        bytenr, num_bytes);
1590                 if (ret > 0) {
1591                         WARN_ON(1);
1592                         continue;
1593                 }
1594                 BUG_ON(ret < 0);
1595
1596                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1597                 dirty = 1;
1598
1599                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1600                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1601                                            num_bytes, parent,
1602                                            btrfs_header_owner(leaf),
1603                                            key.objectid, key.offset);
1604                 BUG_ON(ret);
1605
1606                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1607                                         parent, btrfs_header_owner(leaf),
1608                                         key.objectid, key.offset);
1609                 BUG_ON(ret);
1610         }
1611         if (dirty)
1612                 btrfs_mark_buffer_dirty(leaf);
1613         if (inode)
1614                 btrfs_add_delayed_iput(inode);
1615         return 0;
1616 }
1617
1618 static noinline_for_stack
1619 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1620                      struct btrfs_path *path, int level)
1621 {
1622         struct btrfs_disk_key key1;
1623         struct btrfs_disk_key key2;
1624         btrfs_node_key(eb, &key1, slot);
1625         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1626         return memcmp(&key1, &key2, sizeof(key1));
1627 }
1628
1629 /*
1630  * try to replace tree blocks in fs tree with the new blocks
1631  * in reloc tree. tree blocks haven't been modified since the
1632  * reloc tree was create can be replaced.
1633  *
1634  * if a block was replaced, level of the block + 1 is returned.
1635  * if no block got replaced, 0 is returned. if there are other
1636  * errors, a negative error number is returned.
1637  */
1638 static noinline_for_stack
1639 int replace_path(struct btrfs_trans_handle *trans,
1640                  struct btrfs_root *dest, struct btrfs_root *src,
1641                  struct btrfs_path *path, struct btrfs_key *next_key,
1642                  int lowest_level, int max_level)
1643 {
1644         struct extent_buffer *eb;
1645         struct extent_buffer *parent;
1646         struct btrfs_key key;
1647         u64 old_bytenr;
1648         u64 new_bytenr;
1649         u64 old_ptr_gen;
1650         u64 new_ptr_gen;
1651         u64 last_snapshot;
1652         u32 blocksize;
1653         int cow = 0;
1654         int level;
1655         int ret;
1656         int slot;
1657
1658         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1659         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1660
1661         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1662 again:
1663         slot = path->slots[lowest_level];
1664         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1665
1666         eb = btrfs_lock_root_node(dest);
1667         btrfs_set_lock_blocking(eb);
1668         level = btrfs_header_level(eb);
1669
1670         if (level < lowest_level) {
1671                 btrfs_tree_unlock(eb);
1672                 free_extent_buffer(eb);
1673                 return 0;
1674         }
1675
1676         if (cow) {
1677                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1678                 BUG_ON(ret);
1679         }
1680         btrfs_set_lock_blocking(eb);
1681
1682         if (next_key) {
1683                 next_key->objectid = (u64)-1;
1684                 next_key->type = (u8)-1;
1685                 next_key->offset = (u64)-1;
1686         }
1687
1688         parent = eb;
1689         while (1) {
1690                 level = btrfs_header_level(parent);
1691                 BUG_ON(level < lowest_level);
1692
1693                 ret = btrfs_bin_search(parent, &key, level, &slot);
1694                 if (ret && slot > 0)
1695                         slot--;
1696
1697                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1698                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1699
1700                 old_bytenr = btrfs_node_blockptr(parent, slot);
1701                 blocksize = btrfs_level_size(dest, level - 1);
1702                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1703
1704                 if (level <= max_level) {
1705                         eb = path->nodes[level];
1706                         new_bytenr = btrfs_node_blockptr(eb,
1707                                                         path->slots[level]);
1708                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1709                                                         path->slots[level]);
1710                 } else {
1711                         new_bytenr = 0;
1712                         new_ptr_gen = 0;
1713                 }
1714
1715                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1716                         WARN_ON(1);
1717                         ret = level;
1718                         break;
1719                 }
1720
1721                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1722                     memcmp_node_keys(parent, slot, path, level)) {
1723                         if (level <= lowest_level) {
1724                                 ret = 0;
1725                                 break;
1726                         }
1727
1728                         eb = read_tree_block(dest, old_bytenr, blocksize,
1729                                              old_ptr_gen);
1730                         BUG_ON(!eb);
1731                         btrfs_tree_lock(eb);
1732                         if (cow) {
1733                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1734                                                       slot, &eb);
1735                                 BUG_ON(ret);
1736                         }
1737                         btrfs_set_lock_blocking(eb);
1738
1739                         btrfs_tree_unlock(parent);
1740                         free_extent_buffer(parent);
1741
1742                         parent = eb;
1743                         continue;
1744                 }
1745
1746                 if (!cow) {
1747                         btrfs_tree_unlock(parent);
1748                         free_extent_buffer(parent);
1749                         cow = 1;
1750                         goto again;
1751                 }
1752
1753                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1754                                       path->slots[level]);
1755                 btrfs_release_path(path);
1756
1757                 path->lowest_level = level;
1758                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1759                 path->lowest_level = 0;
1760                 BUG_ON(ret);
1761
1762                 /*
1763                  * swap blocks in fs tree and reloc tree.
1764                  */
1765                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1766                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1767                 btrfs_mark_buffer_dirty(parent);
1768
1769                 btrfs_set_node_blockptr(path->nodes[level],
1770                                         path->slots[level], old_bytenr);
1771                 btrfs_set_node_ptr_generation(path->nodes[level],
1772                                               path->slots[level], old_ptr_gen);
1773                 btrfs_mark_buffer_dirty(path->nodes[level]);
1774
1775                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1776                                         path->nodes[level]->start,
1777                                         src->root_key.objectid, level - 1, 0);
1778                 BUG_ON(ret);
1779                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1780                                         0, dest->root_key.objectid, level - 1,
1781                                         0);
1782                 BUG_ON(ret);
1783
1784                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1785                                         path->nodes[level]->start,
1786                                         src->root_key.objectid, level - 1, 0);
1787                 BUG_ON(ret);
1788
1789                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1790                                         0, dest->root_key.objectid, level - 1,
1791                                         0);
1792                 BUG_ON(ret);
1793
1794                 btrfs_unlock_up_safe(path, 0);
1795
1796                 ret = level;
1797                 break;
1798         }
1799         btrfs_tree_unlock(parent);
1800         free_extent_buffer(parent);
1801         return ret;
1802 }
1803
1804 /*
1805  * helper to find next relocated block in reloc tree
1806  */
1807 static noinline_for_stack
1808 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1809                        int *level)
1810 {
1811         struct extent_buffer *eb;
1812         int i;
1813         u64 last_snapshot;
1814         u32 nritems;
1815
1816         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1817
1818         for (i = 0; i < *level; i++) {
1819                 free_extent_buffer(path->nodes[i]);
1820                 path->nodes[i] = NULL;
1821         }
1822
1823         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1824                 eb = path->nodes[i];
1825                 nritems = btrfs_header_nritems(eb);
1826                 while (path->slots[i] + 1 < nritems) {
1827                         path->slots[i]++;
1828                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1829                             last_snapshot)
1830                                 continue;
1831
1832                         *level = i;
1833                         return 0;
1834                 }
1835                 free_extent_buffer(path->nodes[i]);
1836                 path->nodes[i] = NULL;
1837         }
1838         return 1;
1839 }
1840
1841 /*
1842  * walk down reloc tree to find relocated block of lowest level
1843  */
1844 static noinline_for_stack
1845 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1846                          int *level)
1847 {
1848         struct extent_buffer *eb = NULL;
1849         int i;
1850         u64 bytenr;
1851         u64 ptr_gen = 0;
1852         u64 last_snapshot;
1853         u32 blocksize;
1854         u32 nritems;
1855
1856         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1857
1858         for (i = *level; i > 0; i--) {
1859                 eb = path->nodes[i];
1860                 nritems = btrfs_header_nritems(eb);
1861                 while (path->slots[i] < nritems) {
1862                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1863                         if (ptr_gen > last_snapshot)
1864                                 break;
1865                         path->slots[i]++;
1866                 }
1867                 if (path->slots[i] >= nritems) {
1868                         if (i == *level)
1869                                 break;
1870                         *level = i + 1;
1871                         return 0;
1872                 }
1873                 if (i == 1) {
1874                         *level = i;
1875                         return 0;
1876                 }
1877
1878                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1879                 blocksize = btrfs_level_size(root, i - 1);
1880                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1881                 BUG_ON(btrfs_header_level(eb) != i - 1);
1882                 path->nodes[i - 1] = eb;
1883                 path->slots[i - 1] = 0;
1884         }
1885         return 1;
1886 }
1887
1888 /*
1889  * invalidate extent cache for file extents whose key in range of
1890  * [min_key, max_key)
1891  */
1892 static int invalidate_extent_cache(struct btrfs_root *root,
1893                                    struct btrfs_key *min_key,
1894                                    struct btrfs_key *max_key)
1895 {
1896         struct inode *inode = NULL;
1897         u64 objectid;
1898         u64 start, end;
1899         u64 ino;
1900
1901         objectid = min_key->objectid;
1902         while (1) {
1903                 cond_resched();
1904                 iput(inode);
1905
1906                 if (objectid > max_key->objectid)
1907                         break;
1908
1909                 inode = find_next_inode(root, objectid);
1910                 if (!inode)
1911                         break;
1912                 ino = btrfs_ino(inode);
1913
1914                 if (ino > max_key->objectid) {
1915                         iput(inode);
1916                         break;
1917                 }
1918
1919                 objectid = ino + 1;
1920                 if (!S_ISREG(inode->i_mode))
1921                         continue;
1922
1923                 if (unlikely(min_key->objectid == ino)) {
1924                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1925                                 continue;
1926                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1927                                 start = 0;
1928                         else {
1929                                 start = min_key->offset;
1930                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1931                         }
1932                 } else {
1933                         start = 0;
1934                 }
1935
1936                 if (unlikely(max_key->objectid == ino)) {
1937                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1938                                 continue;
1939                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1940                                 end = (u64)-1;
1941                         } else {
1942                                 if (max_key->offset == 0)
1943                                         continue;
1944                                 end = max_key->offset;
1945                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1946                                 end--;
1947                         }
1948                 } else {
1949                         end = (u64)-1;
1950                 }
1951
1952                 /* the lock_extent waits for readpage to complete */
1953                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1954                 btrfs_drop_extent_cache(inode, start, end, 1);
1955                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1956         }
1957         return 0;
1958 }
1959
1960 static int find_next_key(struct btrfs_path *path, int level,
1961                          struct btrfs_key *key)
1962
1963 {
1964         while (level < BTRFS_MAX_LEVEL) {
1965                 if (!path->nodes[level])
1966                         break;
1967                 if (path->slots[level] + 1 <
1968                     btrfs_header_nritems(path->nodes[level])) {
1969                         btrfs_node_key_to_cpu(path->nodes[level], key,
1970                                               path->slots[level] + 1);
1971                         return 0;
1972                 }
1973                 level++;
1974         }
1975         return 1;
1976 }
1977
1978 /*
1979  * merge the relocated tree blocks in reloc tree with corresponding
1980  * fs tree.
1981  */
1982 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1983                                                struct btrfs_root *root)
1984 {
1985         LIST_HEAD(inode_list);
1986         struct btrfs_key key;
1987         struct btrfs_key next_key;
1988         struct btrfs_trans_handle *trans;
1989         struct btrfs_root *reloc_root;
1990         struct btrfs_root_item *root_item;
1991         struct btrfs_path *path;
1992         struct extent_buffer *leaf;
1993         unsigned long nr;
1994         int level;
1995         int max_level;
1996         int replaced = 0;
1997         int ret;
1998         int err = 0;
1999         u32 min_reserved;
2000
2001         path = btrfs_alloc_path();
2002         if (!path)
2003                 return -ENOMEM;
2004         path->reada = 1;
2005
2006         reloc_root = root->reloc_root;
2007         root_item = &reloc_root->root_item;
2008
2009         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2010                 level = btrfs_root_level(root_item);
2011                 extent_buffer_get(reloc_root->node);
2012                 path->nodes[level] = reloc_root->node;
2013                 path->slots[level] = 0;
2014         } else {
2015                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2016
2017                 level = root_item->drop_level;
2018                 BUG_ON(level == 0);
2019                 path->lowest_level = level;
2020                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2021                 path->lowest_level = 0;
2022                 if (ret < 0) {
2023                         btrfs_free_path(path);
2024                         return ret;
2025                 }
2026
2027                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2028                                       path->slots[level]);
2029                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2030
2031                 btrfs_unlock_up_safe(path, 0);
2032         }
2033
2034         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2035         memset(&next_key, 0, sizeof(next_key));
2036
2037         while (1) {
2038                 trans = btrfs_start_transaction(root, 0);
2039                 BUG_ON(IS_ERR(trans));
2040                 trans->block_rsv = rc->block_rsv;
2041
2042                 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
2043                                             min_reserved, 0);
2044                 if (ret) {
2045                         BUG_ON(ret != -EAGAIN);
2046                         ret = btrfs_commit_transaction(trans, root);
2047                         BUG_ON(ret);
2048                         continue;
2049                 }
2050
2051                 replaced = 0;
2052                 max_level = level;
2053
2054                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2055                 if (ret < 0) {
2056                         err = ret;
2057                         goto out;
2058                 }
2059                 if (ret > 0)
2060                         break;
2061
2062                 if (!find_next_key(path, level, &key) &&
2063                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2064                         ret = 0;
2065                 } else {
2066                         ret = replace_path(trans, root, reloc_root, path,
2067                                            &next_key, level, max_level);
2068                 }
2069                 if (ret < 0) {
2070                         err = ret;
2071                         goto out;
2072                 }
2073
2074                 if (ret > 0) {
2075                         level = ret;
2076                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2077                                               path->slots[level]);
2078                         replaced = 1;
2079                 }
2080
2081                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2082                 if (ret > 0)
2083                         break;
2084
2085                 BUG_ON(level == 0);
2086                 /*
2087                  * save the merging progress in the drop_progress.
2088                  * this is OK since root refs == 1 in this case.
2089                  */
2090                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2091                                path->slots[level]);
2092                 root_item->drop_level = level;
2093
2094                 nr = trans->blocks_used;
2095                 btrfs_end_transaction_throttle(trans, root);
2096
2097                 btrfs_btree_balance_dirty(root, nr);
2098
2099                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2100                         invalidate_extent_cache(root, &key, &next_key);
2101         }
2102
2103         /*
2104          * handle the case only one block in the fs tree need to be
2105          * relocated and the block is tree root.
2106          */
2107         leaf = btrfs_lock_root_node(root);
2108         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2109         btrfs_tree_unlock(leaf);
2110         free_extent_buffer(leaf);
2111         if (ret < 0)
2112                 err = ret;
2113 out:
2114         btrfs_free_path(path);
2115
2116         if (err == 0) {
2117                 memset(&root_item->drop_progress, 0,
2118                        sizeof(root_item->drop_progress));
2119                 root_item->drop_level = 0;
2120                 btrfs_set_root_refs(root_item, 0);
2121                 btrfs_update_reloc_root(trans, root);
2122         }
2123
2124         nr = trans->blocks_used;
2125         btrfs_end_transaction_throttle(trans, root);
2126
2127         btrfs_btree_balance_dirty(root, nr);
2128
2129         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2130                 invalidate_extent_cache(root, &key, &next_key);
2131
2132         return err;
2133 }
2134
2135 static noinline_for_stack
2136 int prepare_to_merge(struct reloc_control *rc, int err)
2137 {
2138         struct btrfs_root *root = rc->extent_root;
2139         struct btrfs_root *reloc_root;
2140         struct btrfs_trans_handle *trans;
2141         LIST_HEAD(reloc_roots);
2142         u64 num_bytes = 0;
2143         int ret;
2144
2145         spin_lock(&root->fs_info->trans_lock);
2146         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2147         rc->merging_rsv_size += rc->nodes_relocated * 2;
2148         spin_unlock(&root->fs_info->trans_lock);
2149 again:
2150         if (!err) {
2151                 num_bytes = rc->merging_rsv_size;
2152                 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2153                                           num_bytes);
2154                 if (ret)
2155                         err = ret;
2156         }
2157
2158         trans = btrfs_join_transaction(rc->extent_root);
2159         if (IS_ERR(trans)) {
2160                 if (!err)
2161                         btrfs_block_rsv_release(rc->extent_root,
2162                                                 rc->block_rsv, num_bytes);
2163                 return PTR_ERR(trans);
2164         }
2165
2166         if (!err) {
2167                 if (num_bytes != rc->merging_rsv_size) {
2168                         btrfs_end_transaction(trans, rc->extent_root);
2169                         btrfs_block_rsv_release(rc->extent_root,
2170                                                 rc->block_rsv, num_bytes);
2171                         goto again;
2172                 }
2173         }
2174
2175         rc->merge_reloc_tree = 1;
2176
2177         while (!list_empty(&rc->reloc_roots)) {
2178                 reloc_root = list_entry(rc->reloc_roots.next,
2179                                         struct btrfs_root, root_list);
2180                 list_del_init(&reloc_root->root_list);
2181
2182                 root = read_fs_root(reloc_root->fs_info,
2183                                     reloc_root->root_key.offset);
2184                 BUG_ON(IS_ERR(root));
2185                 BUG_ON(root->reloc_root != reloc_root);
2186
2187                 /*
2188                  * set reference count to 1, so btrfs_recover_relocation
2189                  * knows it should resumes merging
2190                  */
2191                 if (!err)
2192                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2193                 btrfs_update_reloc_root(trans, root);
2194
2195                 list_add(&reloc_root->root_list, &reloc_roots);
2196         }
2197
2198         list_splice(&reloc_roots, &rc->reloc_roots);
2199
2200         if (!err)
2201                 btrfs_commit_transaction(trans, rc->extent_root);
2202         else
2203                 btrfs_end_transaction(trans, rc->extent_root);
2204         return err;
2205 }
2206
2207 static noinline_for_stack
2208 int merge_reloc_roots(struct reloc_control *rc)
2209 {
2210         struct btrfs_root *root;
2211         struct btrfs_root *reloc_root;
2212         LIST_HEAD(reloc_roots);
2213         int found = 0;
2214         int ret;
2215 again:
2216         root = rc->extent_root;
2217         spin_lock(&root->fs_info->trans_lock);
2218         list_splice_init(&rc->reloc_roots, &reloc_roots);
2219         spin_unlock(&root->fs_info->trans_lock);
2220
2221         while (!list_empty(&reloc_roots)) {
2222                 found = 1;
2223                 reloc_root = list_entry(reloc_roots.next,
2224                                         struct btrfs_root, root_list);
2225
2226                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2227                         root = read_fs_root(reloc_root->fs_info,
2228                                             reloc_root->root_key.offset);
2229                         BUG_ON(IS_ERR(root));
2230                         BUG_ON(root->reloc_root != reloc_root);
2231
2232                         ret = merge_reloc_root(rc, root);
2233                         BUG_ON(ret);
2234                 } else {
2235                         list_del_init(&reloc_root->root_list);
2236                 }
2237                 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
2238         }
2239
2240         if (found) {
2241                 found = 0;
2242                 goto again;
2243         }
2244         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2245         return 0;
2246 }
2247
2248 static void free_block_list(struct rb_root *blocks)
2249 {
2250         struct tree_block *block;
2251         struct rb_node *rb_node;
2252         while ((rb_node = rb_first(blocks))) {
2253                 block = rb_entry(rb_node, struct tree_block, rb_node);
2254                 rb_erase(rb_node, blocks);
2255                 kfree(block);
2256         }
2257 }
2258
2259 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2260                                       struct btrfs_root *reloc_root)
2261 {
2262         struct btrfs_root *root;
2263
2264         if (reloc_root->last_trans == trans->transid)
2265                 return 0;
2266
2267         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2268         BUG_ON(IS_ERR(root));
2269         BUG_ON(root->reloc_root != reloc_root);
2270
2271         return btrfs_record_root_in_trans(trans, root);
2272 }
2273
2274 static noinline_for_stack
2275 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2276                                      struct reloc_control *rc,
2277                                      struct backref_node *node,
2278                                      struct backref_edge *edges[], int *nr)
2279 {
2280         struct backref_node *next;
2281         struct btrfs_root *root;
2282         int index = 0;
2283
2284         next = node;
2285         while (1) {
2286                 cond_resched();
2287                 next = walk_up_backref(next, edges, &index);
2288                 root = next->root;
2289                 BUG_ON(!root);
2290                 BUG_ON(!root->ref_cows);
2291
2292                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2293                         record_reloc_root_in_trans(trans, root);
2294                         break;
2295                 }
2296
2297                 btrfs_record_root_in_trans(trans, root);
2298                 root = root->reloc_root;
2299
2300                 if (next->new_bytenr != root->node->start) {
2301                         BUG_ON(next->new_bytenr);
2302                         BUG_ON(!list_empty(&next->list));
2303                         next->new_bytenr = root->node->start;
2304                         next->root = root;
2305                         list_add_tail(&next->list,
2306                                       &rc->backref_cache.changed);
2307                         __mark_block_processed(rc, next);
2308                         break;
2309                 }
2310
2311                 WARN_ON(1);
2312                 root = NULL;
2313                 next = walk_down_backref(edges, &index);
2314                 if (!next || next->level <= node->level)
2315                         break;
2316         }
2317         if (!root)
2318                 return NULL;
2319
2320         *nr = index;
2321         next = node;
2322         /* setup backref node path for btrfs_reloc_cow_block */
2323         while (1) {
2324                 rc->backref_cache.path[next->level] = next;
2325                 if (--index < 0)
2326                         break;
2327                 next = edges[index]->node[UPPER];
2328         }
2329         return root;
2330 }
2331
2332 /*
2333  * select a tree root for relocation. return NULL if the block
2334  * is reference counted. we should use do_relocation() in this
2335  * case. return a tree root pointer if the block isn't reference
2336  * counted. return -ENOENT if the block is root of reloc tree.
2337  */
2338 static noinline_for_stack
2339 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2340                                    struct backref_node *node)
2341 {
2342         struct backref_node *next;
2343         struct btrfs_root *root;
2344         struct btrfs_root *fs_root = NULL;
2345         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2346         int index = 0;
2347
2348         next = node;
2349         while (1) {
2350                 cond_resched();
2351                 next = walk_up_backref(next, edges, &index);
2352                 root = next->root;
2353                 BUG_ON(!root);
2354
2355                 /* no other choice for non-references counted tree */
2356                 if (!root->ref_cows)
2357                         return root;
2358
2359                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2360                         fs_root = root;
2361
2362                 if (next != node)
2363                         return NULL;
2364
2365                 next = walk_down_backref(edges, &index);
2366                 if (!next || next->level <= node->level)
2367                         break;
2368         }
2369
2370         if (!fs_root)
2371                 return ERR_PTR(-ENOENT);
2372         return fs_root;
2373 }
2374
2375 static noinline_for_stack
2376 u64 calcu_metadata_size(struct reloc_control *rc,
2377                         struct backref_node *node, int reserve)
2378 {
2379         struct backref_node *next = node;
2380         struct backref_edge *edge;
2381         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2382         u64 num_bytes = 0;
2383         int index = 0;
2384
2385         BUG_ON(reserve && node->processed);
2386
2387         while (next) {
2388                 cond_resched();
2389                 while (1) {
2390                         if (next->processed && (reserve || next != node))
2391                                 break;
2392
2393                         num_bytes += btrfs_level_size(rc->extent_root,
2394                                                       next->level);
2395
2396                         if (list_empty(&next->upper))
2397                                 break;
2398
2399                         edge = list_entry(next->upper.next,
2400                                           struct backref_edge, list[LOWER]);
2401                         edges[index++] = edge;
2402                         next = edge->node[UPPER];
2403                 }
2404                 next = walk_down_backref(edges, &index);
2405         }
2406         return num_bytes;
2407 }
2408
2409 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2410                                   struct reloc_control *rc,
2411                                   struct backref_node *node)
2412 {
2413         struct btrfs_root *root = rc->extent_root;
2414         u64 num_bytes;
2415         int ret;
2416
2417         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2418
2419         trans->block_rsv = rc->block_rsv;
2420         ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes);
2421         if (ret) {
2422                 if (ret == -EAGAIN)
2423                         rc->commit_transaction = 1;
2424                 return ret;
2425         }
2426
2427         return 0;
2428 }
2429
2430 static void release_metadata_space(struct reloc_control *rc,
2431                                    struct backref_node *node)
2432 {
2433         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2434         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2435 }
2436
2437 /*
2438  * relocate a block tree, and then update pointers in upper level
2439  * blocks that reference the block to point to the new location.
2440  *
2441  * if called by link_to_upper, the block has already been relocated.
2442  * in that case this function just updates pointers.
2443  */
2444 static int do_relocation(struct btrfs_trans_handle *trans,
2445                          struct reloc_control *rc,
2446                          struct backref_node *node,
2447                          struct btrfs_key *key,
2448                          struct btrfs_path *path, int lowest)
2449 {
2450         struct backref_node *upper;
2451         struct backref_edge *edge;
2452         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2453         struct btrfs_root *root;
2454         struct extent_buffer *eb;
2455         u32 blocksize;
2456         u64 bytenr;
2457         u64 generation;
2458         int nr;
2459         int slot;
2460         int ret;
2461         int err = 0;
2462
2463         BUG_ON(lowest && node->eb);
2464
2465         path->lowest_level = node->level + 1;
2466         rc->backref_cache.path[node->level] = node;
2467         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2468                 cond_resched();
2469
2470                 upper = edge->node[UPPER];
2471                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2472                 BUG_ON(!root);
2473
2474                 if (upper->eb && !upper->locked) {
2475                         if (!lowest) {
2476                                 ret = btrfs_bin_search(upper->eb, key,
2477                                                        upper->level, &slot);
2478                                 BUG_ON(ret);
2479                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2480                                 if (node->eb->start == bytenr)
2481                                         goto next;
2482                         }
2483                         drop_node_buffer(upper);
2484                 }
2485
2486                 if (!upper->eb) {
2487                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2488                         if (ret < 0) {
2489                                 err = ret;
2490                                 break;
2491                         }
2492                         BUG_ON(ret > 0);
2493
2494                         if (!upper->eb) {
2495                                 upper->eb = path->nodes[upper->level];
2496                                 path->nodes[upper->level] = NULL;
2497                         } else {
2498                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2499                         }
2500
2501                         upper->locked = 1;
2502                         path->locks[upper->level] = 0;
2503
2504                         slot = path->slots[upper->level];
2505                         btrfs_release_path(path);
2506                 } else {
2507                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2508                                                &slot);
2509                         BUG_ON(ret);
2510                 }
2511
2512                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2513                 if (lowest) {
2514                         BUG_ON(bytenr != node->bytenr);
2515                 } else {
2516                         if (node->eb->start == bytenr)
2517                                 goto next;
2518                 }
2519
2520                 blocksize = btrfs_level_size(root, node->level);
2521                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2522                 eb = read_tree_block(root, bytenr, blocksize, generation);
2523                 if (!eb) {
2524                         err = -EIO;
2525                         goto next;
2526                 }
2527                 btrfs_tree_lock(eb);
2528                 btrfs_set_lock_blocking(eb);
2529
2530                 if (!node->eb) {
2531                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2532                                               slot, &eb);
2533                         btrfs_tree_unlock(eb);
2534                         free_extent_buffer(eb);
2535                         if (ret < 0) {
2536                                 err = ret;
2537                                 goto next;
2538                         }
2539                         BUG_ON(node->eb != eb);
2540                 } else {
2541                         btrfs_set_node_blockptr(upper->eb, slot,
2542                                                 node->eb->start);
2543                         btrfs_set_node_ptr_generation(upper->eb, slot,
2544                                                       trans->transid);
2545                         btrfs_mark_buffer_dirty(upper->eb);
2546
2547                         ret = btrfs_inc_extent_ref(trans, root,
2548                                                 node->eb->start, blocksize,
2549                                                 upper->eb->start,
2550                                                 btrfs_header_owner(upper->eb),
2551                                                 node->level, 0);
2552                         BUG_ON(ret);
2553
2554                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2555                         BUG_ON(ret);
2556                 }
2557 next:
2558                 if (!upper->pending)
2559                         drop_node_buffer(upper);
2560                 else
2561                         unlock_node_buffer(upper);
2562                 if (err)
2563                         break;
2564         }
2565
2566         if (!err && node->pending) {
2567                 drop_node_buffer(node);
2568                 list_move_tail(&node->list, &rc->backref_cache.changed);
2569                 node->pending = 0;
2570         }
2571
2572         path->lowest_level = 0;
2573         BUG_ON(err == -ENOSPC);
2574         return err;
2575 }
2576
2577 static int link_to_upper(struct btrfs_trans_handle *trans,
2578                          struct reloc_control *rc,
2579                          struct backref_node *node,
2580                          struct btrfs_path *path)
2581 {
2582         struct btrfs_key key;
2583
2584         btrfs_node_key_to_cpu(node->eb, &key, 0);
2585         return do_relocation(trans, rc, node, &key, path, 0);
2586 }
2587
2588 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2589                                 struct reloc_control *rc,
2590                                 struct btrfs_path *path, int err)
2591 {
2592         LIST_HEAD(list);
2593         struct backref_cache *cache = &rc->backref_cache;
2594         struct backref_node *node;
2595         int level;
2596         int ret;
2597
2598         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2599                 while (!list_empty(&cache->pending[level])) {
2600                         node = list_entry(cache->pending[level].next,
2601                                           struct backref_node, list);
2602                         list_move_tail(&node->list, &list);
2603                         BUG_ON(!node->pending);
2604
2605                         if (!err) {
2606                                 ret = link_to_upper(trans, rc, node, path);
2607                                 if (ret < 0)
2608                                         err = ret;
2609                         }
2610                 }
2611                 list_splice_init(&list, &cache->pending[level]);
2612         }
2613         return err;
2614 }
2615
2616 static void mark_block_processed(struct reloc_control *rc,
2617                                  u64 bytenr, u32 blocksize)
2618 {
2619         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2620                         EXTENT_DIRTY, GFP_NOFS);
2621 }
2622
2623 static void __mark_block_processed(struct reloc_control *rc,
2624                                    struct backref_node *node)
2625 {
2626         u32 blocksize;
2627         if (node->level == 0 ||
2628             in_block_group(node->bytenr, rc->block_group)) {
2629                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2630                 mark_block_processed(rc, node->bytenr, blocksize);
2631         }
2632         node->processed = 1;
2633 }
2634
2635 /*
2636  * mark a block and all blocks directly/indirectly reference the block
2637  * as processed.
2638  */
2639 static void update_processed_blocks(struct reloc_control *rc,
2640                                     struct backref_node *node)
2641 {
2642         struct backref_node *next = node;
2643         struct backref_edge *edge;
2644         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2645         int index = 0;
2646
2647         while (next) {
2648                 cond_resched();
2649                 while (1) {
2650                         if (next->processed)
2651                                 break;
2652
2653                         __mark_block_processed(rc, next);
2654
2655                         if (list_empty(&next->upper))
2656                                 break;
2657
2658                         edge = list_entry(next->upper.next,
2659                                           struct backref_edge, list[LOWER]);
2660                         edges[index++] = edge;
2661                         next = edge->node[UPPER];
2662                 }
2663                 next = walk_down_backref(edges, &index);
2664         }
2665 }
2666
2667 static int tree_block_processed(u64 bytenr, u32 blocksize,
2668                                 struct reloc_control *rc)
2669 {
2670         if (test_range_bit(&rc->processed_blocks, bytenr,
2671                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2672                 return 1;
2673         return 0;
2674 }
2675
2676 static int get_tree_block_key(struct reloc_control *rc,
2677                               struct tree_block *block)
2678 {
2679         struct extent_buffer *eb;
2680
2681         BUG_ON(block->key_ready);
2682         eb = read_tree_block(rc->extent_root, block->bytenr,
2683                              block->key.objectid, block->key.offset);
2684         BUG_ON(!eb);
2685         WARN_ON(btrfs_header_level(eb) != block->level);
2686         if (block->level == 0)
2687                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2688         else
2689                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2690         free_extent_buffer(eb);
2691         block->key_ready = 1;
2692         return 0;
2693 }
2694
2695 static int reada_tree_block(struct reloc_control *rc,
2696                             struct tree_block *block)
2697 {
2698         BUG_ON(block->key_ready);
2699         readahead_tree_block(rc->extent_root, block->bytenr,
2700                              block->key.objectid, block->key.offset);
2701         return 0;
2702 }
2703
2704 /*
2705  * helper function to relocate a tree block
2706  */
2707 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2708                                 struct reloc_control *rc,
2709                                 struct backref_node *node,
2710                                 struct btrfs_key *key,
2711                                 struct btrfs_path *path)
2712 {
2713         struct btrfs_root *root;
2714         int release = 0;
2715         int ret = 0;
2716
2717         if (!node)
2718                 return 0;
2719
2720         BUG_ON(node->processed);
2721         root = select_one_root(trans, node);
2722         if (root == ERR_PTR(-ENOENT)) {
2723                 update_processed_blocks(rc, node);
2724                 goto out;
2725         }
2726
2727         if (!root || root->ref_cows) {
2728                 ret = reserve_metadata_space(trans, rc, node);
2729                 if (ret)
2730                         goto out;
2731                 release = 1;
2732         }
2733
2734         if (root) {
2735                 if (root->ref_cows) {
2736                         BUG_ON(node->new_bytenr);
2737                         BUG_ON(!list_empty(&node->list));
2738                         btrfs_record_root_in_trans(trans, root);
2739                         root = root->reloc_root;
2740                         node->new_bytenr = root->node->start;
2741                         node->root = root;
2742                         list_add_tail(&node->list, &rc->backref_cache.changed);
2743                 } else {
2744                         path->lowest_level = node->level;
2745                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2746                         btrfs_release_path(path);
2747                         if (ret > 0)
2748                                 ret = 0;
2749                 }
2750                 if (!ret)
2751                         update_processed_blocks(rc, node);
2752         } else {
2753                 ret = do_relocation(trans, rc, node, key, path, 1);
2754         }
2755 out:
2756         if (ret || node->level == 0 || node->cowonly) {
2757                 if (release)
2758                         release_metadata_space(rc, node);
2759                 remove_backref_node(&rc->backref_cache, node);
2760         }
2761         return ret;
2762 }
2763
2764 /*
2765  * relocate a list of blocks
2766  */
2767 static noinline_for_stack
2768 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2769                          struct reloc_control *rc, struct rb_root *blocks)
2770 {
2771         struct backref_node *node;
2772         struct btrfs_path *path;
2773         struct tree_block *block;
2774         struct rb_node *rb_node;
2775         int ret;
2776         int err = 0;
2777
2778         path = btrfs_alloc_path();
2779         if (!path)
2780                 return -ENOMEM;
2781
2782         rb_node = rb_first(blocks);
2783         while (rb_node) {
2784                 block = rb_entry(rb_node, struct tree_block, rb_node);
2785                 if (!block->key_ready)
2786                         reada_tree_block(rc, block);
2787                 rb_node = rb_next(rb_node);
2788         }
2789
2790         rb_node = rb_first(blocks);
2791         while (rb_node) {
2792                 block = rb_entry(rb_node, struct tree_block, rb_node);
2793                 if (!block->key_ready)
2794                         get_tree_block_key(rc, block);
2795                 rb_node = rb_next(rb_node);
2796         }
2797
2798         rb_node = rb_first(blocks);
2799         while (rb_node) {
2800                 block = rb_entry(rb_node, struct tree_block, rb_node);
2801
2802                 node = build_backref_tree(rc, &block->key,
2803                                           block->level, block->bytenr);
2804                 if (IS_ERR(node)) {
2805                         err = PTR_ERR(node);
2806                         goto out;
2807                 }
2808
2809                 ret = relocate_tree_block(trans, rc, node, &block->key,
2810                                           path);
2811                 if (ret < 0) {
2812                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2813                                 err = ret;
2814                         goto out;
2815                 }
2816                 rb_node = rb_next(rb_node);
2817         }
2818 out:
2819         free_block_list(blocks);
2820         err = finish_pending_nodes(trans, rc, path, err);
2821
2822         btrfs_free_path(path);
2823         return err;
2824 }
2825
2826 static noinline_for_stack
2827 int prealloc_file_extent_cluster(struct inode *inode,
2828                                  struct file_extent_cluster *cluster)
2829 {
2830         u64 alloc_hint = 0;
2831         u64 start;
2832         u64 end;
2833         u64 offset = BTRFS_I(inode)->index_cnt;
2834         u64 num_bytes;
2835         int nr = 0;
2836         int ret = 0;
2837
2838         BUG_ON(cluster->start != cluster->boundary[0]);
2839         mutex_lock(&inode->i_mutex);
2840
2841         ret = btrfs_check_data_free_space(inode, cluster->end +
2842                                           1 - cluster->start);
2843         if (ret)
2844                 goto out;
2845
2846         while (nr < cluster->nr) {
2847                 start = cluster->boundary[nr] - offset;
2848                 if (nr + 1 < cluster->nr)
2849                         end = cluster->boundary[nr + 1] - 1 - offset;
2850                 else
2851                         end = cluster->end - offset;
2852
2853                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2854                 num_bytes = end + 1 - start;
2855                 ret = btrfs_prealloc_file_range(inode, 0, start,
2856                                                 num_bytes, num_bytes,
2857                                                 end + 1, &alloc_hint);
2858                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2859                 if (ret)
2860                         break;
2861                 nr++;
2862         }
2863         btrfs_free_reserved_data_space(inode, cluster->end +
2864                                        1 - cluster->start);
2865 out:
2866         mutex_unlock(&inode->i_mutex);
2867         return ret;
2868 }
2869
2870 static noinline_for_stack
2871 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2872                          u64 block_start)
2873 {
2874         struct btrfs_root *root = BTRFS_I(inode)->root;
2875         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2876         struct extent_map *em;
2877         int ret = 0;
2878
2879         em = alloc_extent_map();
2880         if (!em)
2881                 return -ENOMEM;
2882
2883         em->start = start;
2884         em->len = end + 1 - start;
2885         em->block_len = em->len;
2886         em->block_start = block_start;
2887         em->bdev = root->fs_info->fs_devices->latest_bdev;
2888         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2889
2890         lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2891         while (1) {
2892                 write_lock(&em_tree->lock);
2893                 ret = add_extent_mapping(em_tree, em);
2894                 write_unlock(&em_tree->lock);
2895                 if (ret != -EEXIST) {
2896                         free_extent_map(em);
2897                         break;
2898                 }
2899                 btrfs_drop_extent_cache(inode, start, end, 0);
2900         }
2901         unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2902         return ret;
2903 }
2904
2905 static int relocate_file_extent_cluster(struct inode *inode,
2906                                         struct file_extent_cluster *cluster)
2907 {
2908         u64 page_start;
2909         u64 page_end;
2910         u64 offset = BTRFS_I(inode)->index_cnt;
2911         unsigned long index;
2912         unsigned long last_index;
2913         struct page *page;
2914         struct file_ra_state *ra;
2915         int nr = 0;
2916         int ret = 0;
2917
2918         if (!cluster->nr)
2919                 return 0;
2920
2921         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2922         if (!ra)
2923                 return -ENOMEM;
2924
2925         ret = prealloc_file_extent_cluster(inode, cluster);
2926         if (ret)
2927                 goto out;
2928
2929         file_ra_state_init(ra, inode->i_mapping);
2930
2931         ret = setup_extent_mapping(inode, cluster->start - offset,
2932                                    cluster->end - offset, cluster->start);
2933         if (ret)
2934                 goto out;
2935
2936         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2937         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2938         while (index <= last_index) {
2939                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2940                 if (ret)
2941                         goto out;
2942
2943                 page = find_lock_page(inode->i_mapping, index);
2944                 if (!page) {
2945                         page_cache_sync_readahead(inode->i_mapping,
2946                                                   ra, NULL, index,
2947                                                   last_index + 1 - index);
2948                         page = grab_cache_page(inode->i_mapping, index);
2949                         if (!page) {
2950                                 btrfs_delalloc_release_metadata(inode,
2951                                                         PAGE_CACHE_SIZE);
2952                                 ret = -ENOMEM;
2953                                 goto out;
2954                         }
2955                 }
2956
2957                 if (PageReadahead(page)) {
2958                         page_cache_async_readahead(inode->i_mapping,
2959                                                    ra, NULL, page, index,
2960                                                    last_index + 1 - index);
2961                 }
2962
2963                 if (!PageUptodate(page)) {
2964                         btrfs_readpage(NULL, page);
2965                         lock_page(page);
2966                         if (!PageUptodate(page)) {
2967                                 unlock_page(page);
2968                                 page_cache_release(page);
2969                                 btrfs_delalloc_release_metadata(inode,
2970                                                         PAGE_CACHE_SIZE);
2971                                 ret = -EIO;
2972                                 goto out;
2973                         }
2974                 }
2975
2976                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2977                 page_end = page_start + PAGE_CACHE_SIZE - 1;
2978
2979                 lock_extent(&BTRFS_I(inode)->io_tree,
2980                             page_start, page_end, GFP_NOFS);
2981
2982                 set_page_extent_mapped(page);
2983
2984                 if (nr < cluster->nr &&
2985                     page_start + offset == cluster->boundary[nr]) {
2986                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2987                                         page_start, page_end,
2988                                         EXTENT_BOUNDARY, GFP_NOFS);
2989                         nr++;
2990                 }
2991
2992                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2993                 set_page_dirty(page);
2994
2995                 unlock_extent(&BTRFS_I(inode)->io_tree,
2996                               page_start, page_end, GFP_NOFS);
2997                 unlock_page(page);
2998                 page_cache_release(page);
2999
3000                 index++;
3001                 balance_dirty_pages_ratelimited(inode->i_mapping);
3002                 btrfs_throttle(BTRFS_I(inode)->root);
3003         }
3004         WARN_ON(nr != cluster->nr);
3005 out:
3006         kfree(ra);
3007         return ret;
3008 }
3009
3010 static noinline_for_stack
3011 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3012                          struct file_extent_cluster *cluster)
3013 {
3014         int ret;
3015
3016         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3017                 ret = relocate_file_extent_cluster(inode, cluster);
3018                 if (ret)
3019                         return ret;
3020                 cluster->nr = 0;
3021         }
3022
3023         if (!cluster->nr)
3024                 cluster->start = extent_key->objectid;
3025         else
3026                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3027         cluster->end = extent_key->objectid + extent_key->offset - 1;
3028         cluster->boundary[cluster->nr] = extent_key->objectid;
3029         cluster->nr++;
3030
3031         if (cluster->nr >= MAX_EXTENTS) {
3032                 ret = relocate_file_extent_cluster(inode, cluster);
3033                 if (ret)
3034                         return ret;
3035                 cluster->nr = 0;
3036         }
3037         return 0;
3038 }
3039
3040 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3041 static int get_ref_objectid_v0(struct reloc_control *rc,
3042                                struct btrfs_path *path,
3043                                struct btrfs_key *extent_key,
3044                                u64 *ref_objectid, int *path_change)
3045 {
3046         struct btrfs_key key;
3047         struct extent_buffer *leaf;
3048         struct btrfs_extent_ref_v0 *ref0;
3049         int ret;
3050         int slot;
3051
3052         leaf = path->nodes[0];
3053         slot = path->slots[0];
3054         while (1) {
3055                 if (slot >= btrfs_header_nritems(leaf)) {
3056                         ret = btrfs_next_leaf(rc->extent_root, path);
3057                         if (ret < 0)
3058                                 return ret;
3059                         BUG_ON(ret > 0);
3060                         leaf = path->nodes[0];
3061                         slot = path->slots[0];
3062                         if (path_change)
3063                                 *path_change = 1;
3064                 }
3065                 btrfs_item_key_to_cpu(leaf, &key, slot);
3066                 if (key.objectid != extent_key->objectid)
3067                         return -ENOENT;
3068
3069                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3070                         slot++;
3071                         continue;
3072                 }
3073                 ref0 = btrfs_item_ptr(leaf, slot,
3074                                 struct btrfs_extent_ref_v0);
3075                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3076                 break;
3077         }
3078         return 0;
3079 }
3080 #endif
3081
3082 /*
3083  * helper to add a tree block to the list.
3084  * the major work is getting the generation and level of the block
3085  */
3086 static int add_tree_block(struct reloc_control *rc,
3087                           struct btrfs_key *extent_key,
3088                           struct btrfs_path *path,
3089                           struct rb_root *blocks)
3090 {
3091         struct extent_buffer *eb;
3092         struct btrfs_extent_item *ei;
3093         struct btrfs_tree_block_info *bi;
3094         struct tree_block *block;
3095         struct rb_node *rb_node;
3096         u32 item_size;
3097         int level = -1;
3098         int generation;
3099
3100         eb =  path->nodes[0];
3101         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3102
3103         if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3104                 ei = btrfs_item_ptr(eb, path->slots[0],
3105                                 struct btrfs_extent_item);
3106                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3107                 generation = btrfs_extent_generation(eb, ei);
3108                 level = btrfs_tree_block_level(eb, bi);
3109         } else {
3110 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3111                 u64 ref_owner;
3112                 int ret;
3113
3114                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3115                 ret = get_ref_objectid_v0(rc, path, extent_key,
3116                                           &ref_owner, NULL);
3117                 if (ret < 0)
3118                         return ret;
3119                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3120                 level = (int)ref_owner;
3121                 /* FIXME: get real generation */
3122                 generation = 0;
3123 #else
3124                 BUG();
3125 #endif
3126         }
3127
3128         btrfs_release_path(path);
3129
3130         BUG_ON(level == -1);
3131
3132         block = kmalloc(sizeof(*block), GFP_NOFS);
3133         if (!block)
3134                 return -ENOMEM;
3135
3136         block->bytenr = extent_key->objectid;
3137         block->key.objectid = extent_key->offset;
3138         block->key.offset = generation;
3139         block->level = level;
3140         block->key_ready = 0;
3141
3142         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3143         BUG_ON(rb_node);
3144
3145         return 0;
3146 }
3147
3148 /*
3149  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3150  */
3151 static int __add_tree_block(struct reloc_control *rc,
3152                             u64 bytenr, u32 blocksize,
3153                             struct rb_root *blocks)
3154 {
3155         struct btrfs_path *path;
3156         struct btrfs_key key;
3157         int ret;
3158
3159         if (tree_block_processed(bytenr, blocksize, rc))
3160                 return 0;
3161
3162         if (tree_search(blocks, bytenr))
3163                 return 0;
3164
3165         path = btrfs_alloc_path();
3166         if (!path)
3167                 return -ENOMEM;
3168
3169         key.objectid = bytenr;
3170         key.type = BTRFS_EXTENT_ITEM_KEY;
3171         key.offset = blocksize;
3172
3173         path->search_commit_root = 1;
3174         path->skip_locking = 1;
3175         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3176         if (ret < 0)
3177                 goto out;
3178         BUG_ON(ret);
3179
3180         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3181         ret = add_tree_block(rc, &key, path, blocks);
3182 out:
3183         btrfs_free_path(path);
3184         return ret;
3185 }
3186
3187 /*
3188  * helper to check if the block use full backrefs for pointers in it
3189  */
3190 static int block_use_full_backref(struct reloc_control *rc,
3191                                   struct extent_buffer *eb)
3192 {
3193         u64 flags;
3194         int ret;
3195
3196         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3197             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3198                 return 1;
3199
3200         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3201                                        eb->start, eb->len, NULL, &flags);
3202         BUG_ON(ret);
3203
3204         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3205                 ret = 1;
3206         else
3207                 ret = 0;
3208         return ret;
3209 }
3210
3211 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3212                                     struct inode *inode, u64 ino)
3213 {
3214         struct btrfs_key key;
3215         struct btrfs_path *path;
3216         struct btrfs_root *root = fs_info->tree_root;
3217         struct btrfs_trans_handle *trans;
3218         unsigned long nr;
3219         int ret = 0;
3220
3221         if (inode)
3222                 goto truncate;
3223
3224         key.objectid = ino;
3225         key.type = BTRFS_INODE_ITEM_KEY;
3226         key.offset = 0;
3227
3228         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3229         if (IS_ERR_OR_NULL(inode) || is_bad_inode(inode)) {
3230                 if (inode && !IS_ERR(inode))
3231                         iput(inode);
3232                 return -ENOENT;
3233         }
3234
3235 truncate:
3236         path = btrfs_alloc_path();
3237         if (!path) {
3238                 ret = -ENOMEM;
3239                 goto out;
3240         }
3241
3242         trans = btrfs_join_transaction(root);
3243         if (IS_ERR(trans)) {
3244                 btrfs_free_path(path);
3245                 ret = PTR_ERR(trans);
3246                 goto out;
3247         }
3248
3249         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3250
3251         btrfs_free_path(path);
3252         nr = trans->blocks_used;
3253         btrfs_end_transaction(trans, root);
3254         btrfs_btree_balance_dirty(root, nr);
3255 out:
3256         iput(inode);
3257         return ret;
3258 }
3259
3260 /*
3261  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3262  * this function scans fs tree to find blocks reference the data extent
3263  */
3264 static int find_data_references(struct reloc_control *rc,
3265                                 struct btrfs_key *extent_key,
3266                                 struct extent_buffer *leaf,
3267                                 struct btrfs_extent_data_ref *ref,
3268                                 struct rb_root *blocks)
3269 {
3270         struct btrfs_path *path;
3271         struct tree_block *block;
3272         struct btrfs_root *root;
3273         struct btrfs_file_extent_item *fi;
3274         struct rb_node *rb_node;
3275         struct btrfs_key key;
3276         u64 ref_root;
3277         u64 ref_objectid;
3278         u64 ref_offset;
3279         u32 ref_count;
3280         u32 nritems;
3281         int err = 0;
3282         int added = 0;
3283         int counted;
3284         int ret;
3285
3286         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3287         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3288         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3289         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3290
3291         /*
3292          * This is an extent belonging to the free space cache, lets just delete
3293          * it and redo the search.
3294          */
3295         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3296                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3297                                                NULL, ref_objectid);
3298                 if (ret != -ENOENT)
3299                         return ret;
3300                 ret = 0;
3301         }
3302
3303         path = btrfs_alloc_path();
3304         if (!path)
3305                 return -ENOMEM;
3306         path->reada = 1;
3307
3308         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3309         if (IS_ERR(root)) {
3310                 err = PTR_ERR(root);
3311                 goto out;
3312         }
3313
3314         key.objectid = ref_objectid;
3315         key.offset = ref_offset;
3316         key.type = BTRFS_EXTENT_DATA_KEY;
3317
3318         path->search_commit_root = 1;
3319         path->skip_locking = 1;
3320         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3321         if (ret < 0) {
3322                 err = ret;
3323                 goto out;
3324         }
3325
3326         leaf = path->nodes[0];
3327         nritems = btrfs_header_nritems(leaf);
3328         /*
3329          * the references in tree blocks that use full backrefs
3330          * are not counted in
3331          */
3332         if (block_use_full_backref(rc, leaf))
3333                 counted = 0;
3334         else
3335                 counted = 1;
3336         rb_node = tree_search(blocks, leaf->start);
3337         if (rb_node) {
3338                 if (counted)
3339                         added = 1;
3340                 else
3341                         path->slots[0] = nritems;
3342         }
3343
3344         while (ref_count > 0) {
3345                 while (path->slots[0] >= nritems) {
3346                         ret = btrfs_next_leaf(root, path);
3347                         if (ret < 0) {
3348                                 err = ret;
3349                                 goto out;
3350                         }
3351                         if (ret > 0) {
3352                                 WARN_ON(1);
3353                                 goto out;
3354                         }
3355
3356                         leaf = path->nodes[0];
3357                         nritems = btrfs_header_nritems(leaf);
3358                         added = 0;
3359
3360                         if (block_use_full_backref(rc, leaf))
3361                                 counted = 0;
3362                         else
3363                                 counted = 1;
3364                         rb_node = tree_search(blocks, leaf->start);
3365                         if (rb_node) {
3366                                 if (counted)
3367                                         added = 1;
3368                                 else
3369                                         path->slots[0] = nritems;
3370                         }
3371                 }
3372
3373                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3374                 if (key.objectid != ref_objectid ||
3375                     key.type != BTRFS_EXTENT_DATA_KEY) {
3376                         WARN_ON(1);
3377                         break;
3378                 }
3379
3380                 fi = btrfs_item_ptr(leaf, path->slots[0],
3381                                     struct btrfs_file_extent_item);
3382
3383                 if (btrfs_file_extent_type(leaf, fi) ==
3384                     BTRFS_FILE_EXTENT_INLINE)
3385                         goto next;
3386
3387                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3388                     extent_key->objectid)
3389                         goto next;
3390
3391                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3392                 if (key.offset != ref_offset)
3393                         goto next;
3394
3395                 if (counted)
3396                         ref_count--;
3397                 if (added)
3398                         goto next;
3399
3400                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3401                         block = kmalloc(sizeof(*block), GFP_NOFS);
3402                         if (!block) {
3403                                 err = -ENOMEM;
3404                                 break;
3405                         }
3406                         block->bytenr = leaf->start;
3407                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3408                         block->level = 0;
3409                         block->key_ready = 1;
3410                         rb_node = tree_insert(blocks, block->bytenr,
3411                                               &block->rb_node);
3412                         BUG_ON(rb_node);
3413                 }
3414                 if (counted)
3415                         added = 1;
3416                 else
3417                         path->slots[0] = nritems;
3418 next:
3419                 path->slots[0]++;
3420
3421         }
3422 out:
3423         btrfs_free_path(path);
3424         return err;
3425 }
3426
3427 /*
3428  * hepler to find all tree blocks that reference a given data extent
3429  */
3430 static noinline_for_stack
3431 int add_data_references(struct reloc_control *rc,
3432                         struct btrfs_key *extent_key,
3433                         struct btrfs_path *path,
3434                         struct rb_root *blocks)
3435 {
3436         struct btrfs_key key;
3437         struct extent_buffer *eb;
3438         struct btrfs_extent_data_ref *dref;
3439         struct btrfs_extent_inline_ref *iref;
3440         unsigned long ptr;
3441         unsigned long end;
3442         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3443         int ret;
3444         int err = 0;
3445
3446         eb = path->nodes[0];
3447         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3448         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3449 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3450         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3451                 ptr = end;
3452         else
3453 #endif
3454                 ptr += sizeof(struct btrfs_extent_item);
3455
3456         while (ptr < end) {
3457                 iref = (struct btrfs_extent_inline_ref *)ptr;
3458                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3459                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3460                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3461                         ret = __add_tree_block(rc, key.offset, blocksize,
3462                                                blocks);
3463                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3464                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3465                         ret = find_data_references(rc, extent_key,
3466                                                    eb, dref, blocks);
3467                 } else {
3468                         BUG();
3469                 }
3470                 ptr += btrfs_extent_inline_ref_size(key.type);
3471         }
3472         WARN_ON(ptr > end);
3473
3474         while (1) {
3475                 cond_resched();
3476                 eb = path->nodes[0];
3477                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3478                         ret = btrfs_next_leaf(rc->extent_root, path);
3479                         if (ret < 0) {
3480                                 err = ret;
3481                                 break;
3482                         }
3483                         if (ret > 0)
3484                                 break;
3485                         eb = path->nodes[0];
3486                 }
3487
3488                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3489                 if (key.objectid != extent_key->objectid)
3490                         break;
3491
3492 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3493                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3494                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3495 #else
3496                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3497                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3498 #endif
3499                         ret = __add_tree_block(rc, key.offset, blocksize,
3500                                                blocks);
3501                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3502                         dref = btrfs_item_ptr(eb, path->slots[0],
3503                                               struct btrfs_extent_data_ref);
3504                         ret = find_data_references(rc, extent_key,
3505                                                    eb, dref, blocks);
3506                 } else {
3507                         ret = 0;
3508                 }
3509                 if (ret) {
3510                         err = ret;
3511                         break;
3512                 }
3513                 path->slots[0]++;
3514         }
3515         btrfs_release_path(path);
3516         if (err)
3517                 free_block_list(blocks);
3518         return err;
3519 }
3520
3521 /*
3522  * hepler to find next unprocessed extent
3523  */
3524 static noinline_for_stack
3525 int find_next_extent(struct btrfs_trans_handle *trans,
3526                      struct reloc_control *rc, struct btrfs_path *path,
3527                      struct btrfs_key *extent_key)
3528 {
3529         struct btrfs_key key;
3530         struct extent_buffer *leaf;
3531         u64 start, end, last;
3532         int ret;
3533
3534         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3535         while (1) {
3536                 cond_resched();
3537                 if (rc->search_start >= last) {
3538                         ret = 1;
3539                         break;
3540                 }
3541
3542                 key.objectid = rc->search_start;
3543                 key.type = BTRFS_EXTENT_ITEM_KEY;
3544                 key.offset = 0;
3545
3546                 path->search_commit_root = 1;
3547                 path->skip_locking = 1;
3548                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3549                                         0, 0);
3550                 if (ret < 0)
3551                         break;
3552 next:
3553                 leaf = path->nodes[0];
3554                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3555                         ret = btrfs_next_leaf(rc->extent_root, path);
3556                         if (ret != 0)
3557                                 break;
3558                         leaf = path->nodes[0];
3559                 }
3560
3561                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3562                 if (key.objectid >= last) {
3563                         ret = 1;
3564                         break;
3565                 }
3566
3567                 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3568                     key.objectid + key.offset <= rc->search_start) {
3569                         path->slots[0]++;
3570                         goto next;
3571                 }
3572
3573                 ret = find_first_extent_bit(&rc->processed_blocks,
3574                                             key.objectid, &start, &end,
3575                                             EXTENT_DIRTY);
3576
3577                 if (ret == 0 && start <= key.objectid) {
3578                         btrfs_release_path(path);
3579                         rc->search_start = end + 1;
3580                 } else {
3581                         rc->search_start = key.objectid + key.offset;
3582                         memcpy(extent_key, &key, sizeof(key));
3583                         return 0;
3584                 }
3585         }
3586         btrfs_release_path(path);
3587         return ret;
3588 }
3589
3590 static void set_reloc_control(struct reloc_control *rc)
3591 {
3592         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3593         spin_lock(&fs_info->trans_lock);
3594         fs_info->reloc_ctl = rc;
3595         spin_unlock(&fs_info->trans_lock);
3596 }
3597
3598 static void unset_reloc_control(struct reloc_control *rc)
3599 {
3600         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3601         spin_lock(&fs_info->trans_lock);
3602         fs_info->reloc_ctl = NULL;
3603         spin_unlock(&fs_info->trans_lock);
3604 }
3605
3606 static int check_extent_flags(u64 flags)
3607 {
3608         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3609             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3610                 return 1;
3611         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3612             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3613                 return 1;
3614         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3615             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3616                 return 1;
3617         return 0;
3618 }
3619
3620 static noinline_for_stack
3621 int prepare_to_relocate(struct reloc_control *rc)
3622 {
3623         struct btrfs_trans_handle *trans;
3624         int ret;
3625
3626         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3627         if (!rc->block_rsv)
3628                 return -ENOMEM;
3629
3630         /*
3631          * reserve some space for creating reloc trees.
3632          * btrfs_init_reloc_root will use them when there
3633          * is no reservation in transaction handle.
3634          */
3635         ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3636                                   rc->extent_root->nodesize * 256);
3637         if (ret)
3638                 return ret;
3639
3640         rc->block_rsv->refill_used = 1;
3641         btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3642
3643         memset(&rc->cluster, 0, sizeof(rc->cluster));
3644         rc->search_start = rc->block_group->key.objectid;
3645         rc->extents_found = 0;
3646         rc->nodes_relocated = 0;
3647         rc->merging_rsv_size = 0;
3648
3649         rc->create_reloc_tree = 1;
3650         set_reloc_control(rc);
3651
3652         trans = btrfs_join_transaction(rc->extent_root);
3653         BUG_ON(IS_ERR(trans));
3654         btrfs_commit_transaction(trans, rc->extent_root);
3655         return 0;
3656 }
3657
3658 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3659 {
3660         struct rb_root blocks = RB_ROOT;
3661         struct btrfs_key key;
3662         struct btrfs_trans_handle *trans = NULL;
3663         struct btrfs_path *path;
3664         struct btrfs_extent_item *ei;
3665         unsigned long nr;
3666         u64 flags;
3667         u32 item_size;
3668         int ret;
3669         int err = 0;
3670         int progress = 0;
3671
3672         path = btrfs_alloc_path();
3673         if (!path)
3674                 return -ENOMEM;
3675         path->reada = 1;
3676
3677         ret = prepare_to_relocate(rc);
3678         if (ret) {
3679                 err = ret;
3680                 goto out_free;
3681         }
3682
3683         while (1) {
3684                 progress++;
3685                 trans = btrfs_start_transaction(rc->extent_root, 0);
3686                 BUG_ON(IS_ERR(trans));
3687 restart:
3688                 if (update_backref_cache(trans, &rc->backref_cache)) {
3689                         btrfs_end_transaction(trans, rc->extent_root);
3690                         continue;
3691                 }
3692
3693                 ret = find_next_extent(trans, rc, path, &key);
3694                 if (ret < 0)
3695                         err = ret;
3696                 if (ret != 0)
3697                         break;
3698
3699                 rc->extents_found++;
3700
3701                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3702                                     struct btrfs_extent_item);
3703                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3704                 if (item_size >= sizeof(*ei)) {
3705                         flags = btrfs_extent_flags(path->nodes[0], ei);
3706                         ret = check_extent_flags(flags);
3707                         BUG_ON(ret);
3708
3709                 } else {
3710 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3711                         u64 ref_owner;
3712                         int path_change = 0;
3713
3714                         BUG_ON(item_size !=
3715                                sizeof(struct btrfs_extent_item_v0));
3716                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3717                                                   &path_change);
3718                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3719                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3720                         else
3721                                 flags = BTRFS_EXTENT_FLAG_DATA;
3722
3723                         if (path_change) {
3724                                 btrfs_release_path(path);
3725
3726                                 path->search_commit_root = 1;
3727                                 path->skip_locking = 1;
3728                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3729                                                         &key, path, 0, 0);
3730                                 if (ret < 0) {
3731                                         err = ret;
3732                                         break;
3733                                 }
3734                                 BUG_ON(ret > 0);
3735                         }
3736 #else
3737                         BUG();
3738 #endif
3739                 }
3740
3741                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3742                         ret = add_tree_block(rc, &key, path, &blocks);
3743                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3744                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3745                         ret = add_data_references(rc, &key, path, &blocks);
3746                 } else {
3747                         btrfs_release_path(path);
3748                         ret = 0;
3749                 }
3750                 if (ret < 0) {
3751                         err = ret;
3752                         break;
3753                 }
3754
3755                 if (!RB_EMPTY_ROOT(&blocks)) {
3756                         ret = relocate_tree_blocks(trans, rc, &blocks);
3757                         if (ret < 0) {
3758                                 if (ret != -EAGAIN) {
3759                                         err = ret;
3760                                         break;
3761                                 }
3762                                 rc->extents_found--;
3763                                 rc->search_start = key.objectid;
3764                         }
3765                 }
3766
3767                 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3768                                             rc->block_rsv, 0, 5);
3769                 if (ret < 0) {
3770                         if (ret != -EAGAIN) {
3771                                 err = ret;
3772                                 WARN_ON(1);
3773                                 break;
3774                         }
3775                         rc->commit_transaction = 1;
3776                 }
3777
3778                 if (rc->commit_transaction) {
3779                         rc->commit_transaction = 0;
3780                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3781                         BUG_ON(ret);
3782                 } else {
3783                         nr = trans->blocks_used;
3784                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3785                         btrfs_btree_balance_dirty(rc->extent_root, nr);
3786                 }
3787                 trans = NULL;
3788
3789                 if (rc->stage == MOVE_DATA_EXTENTS &&
3790                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3791                         rc->found_file_extent = 1;
3792                         ret = relocate_data_extent(rc->data_inode,
3793                                                    &key, &rc->cluster);
3794                         if (ret < 0) {
3795                                 err = ret;
3796                                 break;
3797                         }
3798                 }
3799         }
3800         if (trans && progress && err == -ENOSPC) {
3801                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3802                                               rc->block_group->flags);
3803                 if (ret == 0) {
3804                         err = 0;
3805                         progress = 0;
3806                         goto restart;
3807                 }
3808         }
3809
3810         btrfs_release_path(path);
3811         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3812                           GFP_NOFS);
3813
3814         if (trans) {
3815                 nr = trans->blocks_used;
3816                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3817                 btrfs_btree_balance_dirty(rc->extent_root, nr);
3818         }
3819
3820         if (!err) {
3821                 ret = relocate_file_extent_cluster(rc->data_inode,
3822                                                    &rc->cluster);
3823                 if (ret < 0)
3824                         err = ret;
3825         }
3826
3827         rc->create_reloc_tree = 0;
3828         set_reloc_control(rc);
3829
3830         backref_cache_cleanup(&rc->backref_cache);
3831         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3832
3833         err = prepare_to_merge(rc, err);
3834
3835         merge_reloc_roots(rc);
3836
3837         rc->merge_reloc_tree = 0;
3838         unset_reloc_control(rc);
3839         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3840
3841         /* get rid of pinned extents */
3842         trans = btrfs_join_transaction(rc->extent_root);
3843         if (IS_ERR(trans))
3844                 err = PTR_ERR(trans);
3845         else
3846                 btrfs_commit_transaction(trans, rc->extent_root);
3847 out_free:
3848         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3849         btrfs_free_path(path);
3850         return err;
3851 }
3852
3853 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3854                                  struct btrfs_root *root, u64 objectid)
3855 {
3856         struct btrfs_path *path;
3857         struct btrfs_inode_item *item;
3858         struct extent_buffer *leaf;
3859         int ret;
3860
3861         path = btrfs_alloc_path();
3862         if (!path)
3863                 return -ENOMEM;
3864
3865         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3866         if (ret)
3867                 goto out;
3868
3869         leaf = path->nodes[0];
3870         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3871         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3872         btrfs_set_inode_generation(leaf, item, 1);
3873         btrfs_set_inode_size(leaf, item, 0);
3874         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3875         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3876                                           BTRFS_INODE_PREALLOC);
3877         btrfs_mark_buffer_dirty(leaf);
3878         btrfs_release_path(path);
3879 out:
3880         btrfs_free_path(path);
3881         return ret;
3882 }
3883
3884 /*
3885  * helper to create inode for data relocation.
3886  * the inode is in data relocation tree and its link count is 0
3887  */
3888 static noinline_for_stack
3889 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3890                                  struct btrfs_block_group_cache *group)
3891 {
3892         struct inode *inode = NULL;
3893         struct btrfs_trans_handle *trans;
3894         struct btrfs_root *root;
3895         struct btrfs_key key;
3896         unsigned long nr;
3897         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3898         int err = 0;
3899
3900         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3901         if (IS_ERR(root))
3902                 return ERR_CAST(root);
3903
3904         trans = btrfs_start_transaction(root, 6);
3905         if (IS_ERR(trans))
3906                 return ERR_CAST(trans);
3907
3908         err = btrfs_find_free_objectid(root, &objectid);
3909         if (err)
3910                 goto out;
3911
3912         err = __insert_orphan_inode(trans, root, objectid);
3913         BUG_ON(err);
3914
3915         key.objectid = objectid;
3916         key.type = BTRFS_INODE_ITEM_KEY;
3917         key.offset = 0;
3918         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3919         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3920         BTRFS_I(inode)->index_cnt = group->key.objectid;
3921
3922         err = btrfs_orphan_add(trans, inode);
3923 out:
3924         nr = trans->blocks_used;
3925         btrfs_end_transaction(trans, root);
3926         btrfs_btree_balance_dirty(root, nr);
3927         if (err) {
3928                 if (inode)
3929                         iput(inode);
3930                 inode = ERR_PTR(err);
3931         }
3932         return inode;
3933 }
3934
3935 static struct reloc_control *alloc_reloc_control(void)
3936 {
3937         struct reloc_control *rc;
3938
3939         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3940         if (!rc)
3941                 return NULL;
3942
3943         INIT_LIST_HEAD(&rc->reloc_roots);
3944         backref_cache_init(&rc->backref_cache);
3945         mapping_tree_init(&rc->reloc_root_tree);
3946         extent_io_tree_init(&rc->processed_blocks, NULL);
3947         return rc;
3948 }
3949
3950 /*
3951  * function to relocate all extents in a block group.
3952  */
3953 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3954 {
3955         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3956         struct reloc_control *rc;
3957         struct inode *inode;
3958         struct btrfs_path *path;
3959         int ret;
3960         int rw = 0;
3961         int err = 0;
3962
3963         rc = alloc_reloc_control();
3964         if (!rc)
3965                 return -ENOMEM;
3966
3967         rc->extent_root = extent_root;
3968
3969         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3970         BUG_ON(!rc->block_group);
3971
3972         if (!rc->block_group->ro) {
3973                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3974                 if (ret) {
3975                         err = ret;
3976                         goto out;
3977                 }
3978                 rw = 1;
3979         }
3980
3981         path = btrfs_alloc_path();
3982         if (!path) {
3983                 err = -ENOMEM;
3984                 goto out;
3985         }
3986
3987         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
3988                                         path);
3989         btrfs_free_path(path);
3990
3991         if (!IS_ERR(inode))
3992                 ret = delete_block_group_cache(fs_info, inode, 0);
3993         else
3994                 ret = PTR_ERR(inode);
3995
3996         if (ret && ret != -ENOENT) {
3997                 err = ret;
3998                 goto out;
3999         }
4000
4001         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
4002         if (IS_ERR(rc->data_inode)) {
4003                 err = PTR_ERR(rc->data_inode);
4004                 rc->data_inode = NULL;
4005                 goto out;
4006         }
4007
4008         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4009                (unsigned long long)rc->block_group->key.objectid,
4010                (unsigned long long)rc->block_group->flags);
4011
4012         btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4013         btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
4014
4015         while (1) {
4016                 mutex_lock(&fs_info->cleaner_mutex);
4017
4018                 btrfs_clean_old_snapshots(fs_info->tree_root);
4019                 ret = relocate_block_group(rc);
4020
4021                 mutex_unlock(&fs_info->cleaner_mutex);
4022                 if (ret < 0) {
4023                         err = ret;
4024                         goto out;
4025                 }
4026
4027                 if (rc->extents_found == 0)
4028                         break;
4029
4030                 printk(KERN_INFO "btrfs: found %llu extents\n",
4031                         (unsigned long long)rc->extents_found);
4032
4033                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4034                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4035                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4036                                                  0, -1);
4037                         rc->stage = UPDATE_DATA_PTRS;
4038                 }
4039         }
4040
4041         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4042                                      rc->block_group->key.objectid,
4043                                      rc->block_group->key.objectid +
4044                                      rc->block_group->key.offset - 1);
4045
4046         WARN_ON(rc->block_group->pinned > 0);
4047         WARN_ON(rc->block_group->reserved > 0);
4048         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4049 out:
4050         if (err && rw)
4051                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4052         iput(rc->data_inode);
4053         btrfs_put_block_group(rc->block_group);
4054         kfree(rc);
4055         return err;
4056 }
4057
4058 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4059 {
4060         struct btrfs_trans_handle *trans;
4061         int ret;
4062
4063         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4064         BUG_ON(IS_ERR(trans));
4065
4066         memset(&root->root_item.drop_progress, 0,
4067                 sizeof(root->root_item.drop_progress));
4068         root->root_item.drop_level = 0;
4069         btrfs_set_root_refs(&root->root_item, 0);
4070         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4071                                 &root->root_key, &root->root_item);
4072         BUG_ON(ret);
4073
4074         ret = btrfs_end_transaction(trans, root->fs_info->tree_root);
4075         BUG_ON(ret);
4076         return 0;
4077 }
4078
4079 /*
4080  * recover relocation interrupted by system crash.
4081  *
4082  * this function resumes merging reloc trees with corresponding fs trees.
4083  * this is important for keeping the sharing of tree blocks
4084  */
4085 int btrfs_recover_relocation(struct btrfs_root *root)
4086 {
4087         LIST_HEAD(reloc_roots);
4088         struct btrfs_key key;
4089         struct btrfs_root *fs_root;
4090         struct btrfs_root *reloc_root;
4091         struct btrfs_path *path;
4092         struct extent_buffer *leaf;
4093         struct reloc_control *rc = NULL;
4094         struct btrfs_trans_handle *trans;
4095         int ret;
4096         int err = 0;
4097
4098         path = btrfs_alloc_path();
4099         if (!path)
4100                 return -ENOMEM;
4101         path->reada = -1;
4102
4103         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4104         key.type = BTRFS_ROOT_ITEM_KEY;
4105         key.offset = (u64)-1;
4106
4107         while (1) {
4108                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4109                                         path, 0, 0);
4110                 if (ret < 0) {
4111                         err = ret;
4112                         goto out;
4113                 }
4114                 if (ret > 0) {
4115                         if (path->slots[0] == 0)
4116                                 break;
4117                         path->slots[0]--;
4118                 }
4119                 leaf = path->nodes[0];
4120                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4121                 btrfs_release_path(path);
4122
4123                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4124                     key.type != BTRFS_ROOT_ITEM_KEY)
4125                         break;
4126
4127                 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4128                 if (IS_ERR(reloc_root)) {
4129                         err = PTR_ERR(reloc_root);
4130                         goto out;
4131                 }
4132
4133                 list_add(&reloc_root->root_list, &reloc_roots);
4134
4135                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4136                         fs_root = read_fs_root(root->fs_info,
4137                                                reloc_root->root_key.offset);
4138                         if (IS_ERR(fs_root)) {
4139                                 ret = PTR_ERR(fs_root);
4140                                 if (ret != -ENOENT) {
4141                                         err = ret;
4142                                         goto out;
4143                                 }
4144                                 mark_garbage_root(reloc_root);
4145                         }
4146                 }
4147
4148                 if (key.offset == 0)
4149                         break;
4150
4151                 key.offset--;
4152         }
4153         btrfs_release_path(path);
4154
4155         if (list_empty(&reloc_roots))
4156                 goto out;
4157
4158         rc = alloc_reloc_control();
4159         if (!rc) {
4160                 err = -ENOMEM;
4161                 goto out;
4162         }
4163
4164         rc->extent_root = root->fs_info->extent_root;
4165
4166         set_reloc_control(rc);
4167
4168         trans = btrfs_join_transaction(rc->extent_root);
4169         if (IS_ERR(trans)) {
4170                 unset_reloc_control(rc);
4171                 err = PTR_ERR(trans);
4172                 goto out_free;
4173         }
4174
4175         rc->merge_reloc_tree = 1;
4176
4177         while (!list_empty(&reloc_roots)) {
4178                 reloc_root = list_entry(reloc_roots.next,
4179                                         struct btrfs_root, root_list);
4180                 list_del(&reloc_root->root_list);
4181
4182                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4183                         list_add_tail(&reloc_root->root_list,
4184                                       &rc->reloc_roots);
4185                         continue;
4186                 }
4187
4188                 fs_root = read_fs_root(root->fs_info,
4189                                        reloc_root->root_key.offset);
4190                 BUG_ON(IS_ERR(fs_root));
4191
4192                 __add_reloc_root(reloc_root);
4193                 fs_root->reloc_root = reloc_root;
4194         }
4195
4196         btrfs_commit_transaction(trans, rc->extent_root);
4197
4198         merge_reloc_roots(rc);
4199
4200         unset_reloc_control(rc);
4201
4202         trans = btrfs_join_transaction(rc->extent_root);
4203         if (IS_ERR(trans))
4204                 err = PTR_ERR(trans);
4205         else
4206                 btrfs_commit_transaction(trans, rc->extent_root);
4207 out_free:
4208         kfree(rc);
4209 out:
4210         while (!list_empty(&reloc_roots)) {
4211                 reloc_root = list_entry(reloc_roots.next,
4212                                         struct btrfs_root, root_list);
4213                 list_del(&reloc_root->root_list);
4214                 free_extent_buffer(reloc_root->node);
4215                 free_extent_buffer(reloc_root->commit_root);
4216                 kfree(reloc_root);
4217         }
4218         btrfs_free_path(path);
4219
4220         if (err == 0) {
4221                 /* cleanup orphan inode in data relocation tree */
4222                 fs_root = read_fs_root(root->fs_info,
4223                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4224                 if (IS_ERR(fs_root))
4225                         err = PTR_ERR(fs_root);
4226                 else
4227                         err = btrfs_orphan_cleanup(fs_root);
4228         }
4229         return err;
4230 }
4231
4232 /*
4233  * helper to add ordered checksum for data relocation.
4234  *
4235  * cloning checksum properly handles the nodatasum extents.
4236  * it also saves CPU time to re-calculate the checksum.
4237  */
4238 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4239 {
4240         struct btrfs_ordered_sum *sums;
4241         struct btrfs_sector_sum *sector_sum;
4242         struct btrfs_ordered_extent *ordered;
4243         struct btrfs_root *root = BTRFS_I(inode)->root;
4244         size_t offset;
4245         int ret;
4246         u64 disk_bytenr;
4247         LIST_HEAD(list);
4248
4249         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4250         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4251
4252         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4253         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4254                                        disk_bytenr + len - 1, &list, 0);
4255
4256         while (!list_empty(&list)) {
4257                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4258                 list_del_init(&sums->list);
4259
4260                 sector_sum = sums->sums;
4261                 sums->bytenr = ordered->start;
4262
4263                 offset = 0;
4264                 while (offset < sums->len) {
4265                         sector_sum->bytenr += ordered->start - disk_bytenr;
4266                         sector_sum++;
4267                         offset += root->sectorsize;
4268                 }
4269
4270                 btrfs_add_ordered_sum(inode, ordered, sums);
4271         }
4272         btrfs_put_ordered_extent(ordered);
4273         return ret;
4274 }
4275
4276 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4277                            struct btrfs_root *root, struct extent_buffer *buf,
4278                            struct extent_buffer *cow)
4279 {
4280         struct reloc_control *rc;
4281         struct backref_node *node;
4282         int first_cow = 0;
4283         int level;
4284         int ret;
4285
4286         rc = root->fs_info->reloc_ctl;
4287         if (!rc)
4288                 return;
4289
4290         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4291                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4292
4293         level = btrfs_header_level(buf);
4294         if (btrfs_header_generation(buf) <=
4295             btrfs_root_last_snapshot(&root->root_item))
4296                 first_cow = 1;
4297
4298         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4299             rc->create_reloc_tree) {
4300                 WARN_ON(!first_cow && level == 0);
4301
4302                 node = rc->backref_cache.path[level];
4303                 BUG_ON(node->bytenr != buf->start &&
4304                        node->new_bytenr != buf->start);
4305
4306                 drop_node_buffer(node);
4307                 extent_buffer_get(cow);
4308                 node->eb = cow;
4309                 node->new_bytenr = cow->start;
4310
4311                 if (!node->pending) {
4312                         list_move_tail(&node->list,
4313                                        &rc->backref_cache.pending[level]);
4314                         node->pending = 1;
4315                 }
4316
4317                 if (first_cow)
4318                         __mark_block_processed(rc, node);
4319
4320                 if (first_cow && level > 0)
4321                         rc->nodes_relocated += buf->len;
4322         }
4323
4324         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4325                 ret = replace_file_extents(trans, rc, root, cow);
4326                 BUG_ON(ret);
4327         }
4328 }
4329
4330 /*
4331  * called before creating snapshot. it calculates metadata reservation
4332  * requried for relocating tree blocks in the snapshot
4333  */
4334 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4335                               struct btrfs_pending_snapshot *pending,
4336                               u64 *bytes_to_reserve)
4337 {
4338         struct btrfs_root *root;
4339         struct reloc_control *rc;
4340
4341         root = pending->root;
4342         if (!root->reloc_root)
4343                 return;
4344
4345         rc = root->fs_info->reloc_ctl;
4346         if (!rc->merge_reloc_tree)
4347                 return;
4348
4349         root = root->reloc_root;
4350         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4351         /*
4352          * relocation is in the stage of merging trees. the space
4353          * used by merging a reloc tree is twice the size of
4354          * relocated tree nodes in the worst case. half for cowing
4355          * the reloc tree, half for cowing the fs tree. the space
4356          * used by cowing the reloc tree will be freed after the
4357          * tree is dropped. if we create snapshot, cowing the fs
4358          * tree may use more space than it frees. so we need
4359          * reserve extra space.
4360          */
4361         *bytes_to_reserve += rc->nodes_relocated;
4362 }
4363
4364 /*
4365  * called after snapshot is created. migrate block reservation
4366  * and create reloc root for the newly created snapshot
4367  */
4368 void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4369                                struct btrfs_pending_snapshot *pending)
4370 {
4371         struct btrfs_root *root = pending->root;
4372         struct btrfs_root *reloc_root;
4373         struct btrfs_root *new_root;
4374         struct reloc_control *rc;
4375         int ret;
4376
4377         if (!root->reloc_root)
4378                 return;
4379
4380         rc = root->fs_info->reloc_ctl;
4381         rc->merging_rsv_size += rc->nodes_relocated;
4382
4383         if (rc->merge_reloc_tree) {
4384                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4385                                               rc->block_rsv,
4386                                               rc->nodes_relocated);
4387                 BUG_ON(ret);
4388         }
4389
4390         new_root = pending->snap;
4391         reloc_root = create_reloc_root(trans, root->reloc_root,
4392                                        new_root->root_key.objectid);
4393
4394         __add_reloc_root(reloc_root);
4395         new_root->reloc_root = reloc_root;
4396
4397         if (rc->create_reloc_tree) {
4398                 ret = clone_backref_node(trans, rc, root, reloc_root);
4399                 BUG_ON(ret);
4400         }
4401 }