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