fs/btrfs/relocation.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "volumes.h"
  29 #include "locking.h"
  30 #include "btrfs_inode.h"
  31 #include "async-thread.h"
  32 #include "free-space-cache.h"
  33 #include "inode-map.h"
  34
  35 /*
  36  * backref_node, mapping_node and tree_block start with this
  37  */
  38 struct tree_entry {
  39         struct rb_node rb_node;
  40         u64 bytenr;
  41 };
  42
  43 /*
  44  * present a tree block in the backref cache
  45  */
  46 struct backref_node {
  47         struct rb_node rb_node;
  48         u64 bytenr;
  49
  50         u64 new_bytenr;
  51         /* objectid of tree block owner, can be not uptodate */
  52         u64 owner;
  53         /* link to pending, changed or detached list */
  54         struct list_head list;
  55         /* list of upper level blocks reference this block */
  56         struct list_head upper;
  57         /* list of child blocks in the cache */
  58         struct list_head lower;
  59         /* NULL if this node is not tree root */
  60         struct btrfs_root *root;
  61         /* extent buffer got by COW the block */
  62         struct extent_buffer *eb;
  63         /* level of tree block */
  64         unsigned int level:8;
  65         /* is the block in non-reference counted tree */
  66         unsigned int cowonly:1;
  67         /* 1 if no child node in the cache */
  68         unsigned int lowest:1;
  69         /* is the extent buffer locked */
  70         unsigned int locked:1;
  71         /* has the block been processed */
  72         unsigned int processed:1;
  73         /* have backrefs of this block been checked */
  74         unsigned int checked:1;
  75         /*
  76          * 1 if corresponding block has been cowed but some upper
  77          * level block pointers may not point to the new location
  78          */
  79         unsigned int pending:1;
  80         /*
  81          * 1 if the backref node isn't connected to any other
  82          * backref node.
  83          */
  84         unsigned int detached:1;
  85 };
  86
  87 /*
  88  * present a block pointer in the backref cache
  89  */
  90 struct backref_edge {
  91         struct list_head list[2];
  92         struct backref_node *node[2];
  93 };
  94
  95 #define LOWER   0
  96 #define UPPER   1
  97
  98 struct backref_cache {
  99         /* red black tree of all backref nodes in the cache */
 100         struct rb_root rb_root;
 101         /* for passing backref nodes to btrfs_reloc_cow_block */
 102         struct backref_node *path[BTRFS_MAX_LEVEL];
 103         /*
 104          * list of blocks that have been cowed but some block
 105          * pointers in upper level blocks may not reflect the
 106          * new location
 107          */
 108         struct list_head pending[BTRFS_MAX_LEVEL];
 109         /* list of backref nodes with no child node */
 110         struct list_head leaves;
 111         /* list of blocks that have been cowed in current transaction */
 112         struct list_head changed;
 113         /* list of detached backref node. */
 114         struct list_head detached;
 115
 116         u64 last_trans;
 117
 118         int nr_nodes;
 119         int nr_edges;
 120 };
 121
 122 /*
 123  * map address of tree root to tree
 124  */
 125 struct mapping_node {
 126         struct rb_node rb_node;
 127         u64 bytenr;
 128         void *data;
 129 };
 130
 131 struct mapping_tree {
 132         struct rb_root rb_root;
 133         spinlock_t lock;
 134 };
 135
 136 /*
 137  * present a tree block to process
 138  */
 139 struct tree_block {
 140         struct rb_node rb_node;
 141         u64 bytenr;
 142         struct btrfs_key key;
 143         unsigned int level:8;
 144         unsigned int key_ready:1;
 145 };
 146
 147 #define MAX_EXTENTS 128
 148
 149 struct file_extent_cluster {
 150         u64 start;
 151         u64 end;
 152         u64 boundary[MAX_EXTENTS];
 153         unsigned int nr;
 154 };
 155
 156 struct reloc_control {
 157         /* block group to relocate */
 158         struct btrfs_block_group_cache *block_group;
 159         /* extent tree */
 160         struct btrfs_root *extent_root;
 161         /* inode for moving data */
 162         struct inode *data_inode;
 163
 164         struct btrfs_block_rsv *block_rsv;
 165
 166         struct backref_cache backref_cache;
 167
 168         struct file_extent_cluster cluster;
 169         /* tree blocks have been processed */
 170         struct extent_io_tree processed_blocks;
 171         /* map start of tree root to corresponding reloc tree */
 172         struct mapping_tree reloc_root_tree;
 173         /* list of reloc trees */
 174         struct list_head reloc_roots;
 175         /* size of metadata reservation for merging reloc trees */
 176         u64 merging_rsv_size;
 177         /* size of relocated tree nodes */
 178         u64 nodes_relocated;
 179
 180         u64 search_start;
 181         u64 extents_found;
 182
 183         unsigned int stage:8;
 184         unsigned int create_reloc_tree:1;
 185         unsigned int merge_reloc_tree:1;
 186         unsigned int found_file_extent:1;
 187         unsigned int commit_transaction:1;
 188 };
 189
 190 /* stages of data relocation */
 191 #define MOVE_DATA_EXTENTS       0
 192 #define UPDATE_DATA_PTRS        1
 193
 194 static void remove_backref_node(struct backref_cache *cache,
 195                                 struct backref_node *node);
 196 static void __mark_block_processed(struct reloc_control *rc,
 197                                    struct backref_node *node);
 198
 199 static void mapping_tree_init(struct mapping_tree *tree)
 200 {
 201         tree->rb_root = RB_ROOT;
 202         spin_lock_init(&tree->lock);
 203 }
 204
 205 static void backref_cache_init(struct backref_cache *cache)
 206 {
 207         int i;
 208         cache->rb_root = RB_ROOT;
 209         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 210                 INIT_LIST_HEAD(&cache->pending[i]);
 211         INIT_LIST_HEAD(&cache->changed);
 212         INIT_LIST_HEAD(&cache->detached);
 213         INIT_LIST_HEAD(&cache->leaves);
 214 }
 215
 216 static void backref_cache_cleanup(struct backref_cache *cache)
 217 {
 218         struct backref_node *node;
 219         int i;
 220
 221         while (!list_empty(&cache->detached)) {
 222                 node = list_entry(cache->detached.next,
 223                                   struct backref_node, list);
 224                 remove_backref_node(cache, node);
 225         }
 226
 227         while (!list_empty(&cache->leaves)) {
 228                 node = list_entry(cache->leaves.next,
 229                                   struct backref_node, lower);
 230                 remove_backref_node(cache, node);
 231         }
 232
 233         cache->last_trans = 0;
 234
 235         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
 236                 BUG_ON(!list_empty(&cache->pending[i]));
 237         BUG_ON(!list_empty(&cache->changed));
 238         BUG_ON(!list_empty(&cache->detached));
 239         BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
 240         BUG_ON(cache->nr_nodes);
 241         BUG_ON(cache->nr_edges);
 242 }
 243
 244 static struct backref_node *alloc_backref_node(struct backref_cache *cache)
 245 {
 246         struct backref_node *node;
 247
 248         node = kzalloc(sizeof(*node), GFP_NOFS);
 249         if (node) {
 250                 INIT_LIST_HEAD(&node->list);
 251                 INIT_LIST_HEAD(&node->upper);
 252                 INIT_LIST_HEAD(&node->lower);
 253                 RB_CLEAR_NODE(&node->rb_node);
 254                 cache->nr_nodes++;
 255         }
 256         return node;
 257 }
 258
 259 static void free_backref_node(struct backref_cache *cache,
 260                               struct backref_node *node)
 261 {
 262         if (node) {
 263                 cache->nr_nodes--;
 264                 kfree(node);
 265         }
 266 }
 267
 268 static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
 269 {
 270         struct backref_edge *edge;
 271
 272         edge = kzalloc(sizeof(*edge), GFP_NOFS);
 273         if (edge)
 274                 cache->nr_edges++;
 275         return edge;
 276 }
 277
 278 static void free_backref_edge(struct backref_cache *cache,
 279                               struct backref_edge *edge)
 280 {
 281         if (edge) {
 282                 cache->nr_edges--;
 283                 kfree(edge);
 284         }
 285 }
 286
 287 static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
 288                                    struct rb_node *node)
 289 {
 290         struct rb_node **p = &root->rb_node;
 291         struct rb_node *parent = NULL;
 292         struct tree_entry *entry;
 293
 294         while (*p) {
 295                 parent = *p;
 296                 entry = rb_entry(parent, struct tree_entry, rb_node);
 297
 298                 if (bytenr < entry->bytenr)
 299                         p = &(*p)->rb_left;
 300                 else if (bytenr > entry->bytenr)
 301                         p = &(*p)->rb_right;
 302                 else
 303                         return parent;
 304         }
 305
 306         rb_link_node(node, parent, p);
 307         rb_insert_color(node, root);
 308         return NULL;
 309 }
 310
 311 static struct rb_node *tree_search(struct rb_root *root, u64 bytenr)
 312 {
 313         struct rb_node *n = root->rb_node;
 314         struct tree_entry *entry;
 315
 316         while (n) {
 317                 entry = rb_entry(n, struct tree_entry, rb_node);
 318
 319                 if (bytenr < entry->bytenr)
 320                         n = n->rb_left;
 321                 else if (bytenr > entry->bytenr)
 322                         n = n->rb_right;
 323                 else
 324                         return n;
 325         }
 326         return NULL;
 327 }
 328
 329 /*
 330  * walk up backref nodes until reach node presents tree root
 331  */
 332 static struct backref_node *walk_up_backref(struct backref_node *node,
 333                                             struct backref_edge *edges[],
 334                                             int *index)
 335 {
 336         struct backref_edge *edge;
 337         int idx = *index;
 338
 339         while (!list_empty(&node->upper)) {
 340                 edge = list_entry(node->upper.next,
 341                                   struct backref_edge, list[LOWER]);
 342                 edges[idx++] = edge;
 343                 node = edge->node[UPPER];
 344         }
 345         BUG_ON(node->detached);
 346         *index = idx;
 347         return node;
 348 }
 349
 350 /*
 351  * walk down backref nodes to find start of next reference path
 352  */
 353 static struct backref_node *walk_down_backref(struct backref_edge *edges[],
 354                                               int *index)
 355 {
 356         struct backref_edge *edge;
 357         struct backref_node *lower;
 358         int idx = *index;
 359
 360         while (idx > 0) {
 361                 edge = edges[idx - 1];
 362                 lower = edge->node[LOWER];
 363                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
 364                         idx--;
 365                         continue;
 366                 }
 367                 edge = list_entry(edge->list[LOWER].next,
 368                                   struct backref_edge, list[LOWER]);
 369                 edges[idx - 1] = edge;
 370                 *index = idx;
 371                 return edge->node[UPPER];
 372         }
 373         *index = 0;
 374         return NULL;
 375 }
 376
 377 static void unlock_node_buffer(struct backref_node *node)
 378 {
 379         if (node->locked) {
 380                 btrfs_tree_unlock(node->eb);
 381                 node->locked = 0;
 382         }
 383 }
 384
 385 static void drop_node_buffer(struct backref_node *node)
 386 {
 387         if (node->eb) {
 388                 unlock_node_buffer(node);
 389                 free_extent_buffer(node->eb);
 390                 node->eb = NULL;
 391         }
 392 }
 393
 394 static void drop_backref_node(struct backref_cache *tree,
 395                               struct backref_node *node)
 396 {
 397         BUG_ON(!list_empty(&node->upper));
 398
 399         drop_node_buffer(node);
 400         list_del(&node->list);
 401         list_del(&node->lower);
 402         if (!RB_EMPTY_NODE(&node->rb_node))
 403                 rb_erase(&node->rb_node, &tree->rb_root);
 404         free_backref_node(tree, node);
 405 }
 406
 407 /*
 408  * remove a backref node from the backref cache
 409  */
 410 static void remove_backref_node(struct backref_cache *cache,
 411                                 struct backref_node *node)
 412 {
 413         struct backref_node *upper;
 414         struct backref_edge *edge;
 415
 416         if (!node)
 417                 return;
 418
 419         BUG_ON(!node->lowest && !node->detached);
 420         while (!list_empty(&node->upper)) {
 421                 edge = list_entry(node->upper.next, struct backref_edge,
 422                                   list[LOWER]);
 423                 upper = edge->node[UPPER];
 424                 list_del(&edge->list[LOWER]);
 425                 list_del(&edge->list[UPPER]);
 426                 free_backref_edge(cache, edge);
 427
 428                 if (RB_EMPTY_NODE(&upper->rb_node)) {
 429                         BUG_ON(!list_empty(&node->upper));
 430                         drop_backref_node(cache, node);
 431                         node = upper;
 432                         node->lowest = 1;
 433                         continue;
 434                 }
 435                 /*
 436                  * add the node to leaf node list if no other
 437                  * child block cached.
 438                  */
 439                 if (list_empty(&upper->lower)) {
 440                         list_add_tail(&upper->lower, &cache->leaves);
 441                         upper->lowest = 1;
 442                 }
 443         }
 444
 445         drop_backref_node(cache, node);
 446 }
 447
 448 static void update_backref_node(struct backref_cache *cache,
 449                                 struct backref_node *node, u64 bytenr)
 450 {
 451         struct rb_node *rb_node;
 452         rb_erase(&node->rb_node, &cache->rb_root);
 453         node->bytenr = bytenr;
 454         rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
 455         BUG_ON(rb_node);
 456 }
 457
 458 /*
 459  * update backref cache after a transaction commit
 460  */
 461 static int update_backref_cache(struct btrfs_trans_handle *trans,
 462                                 struct backref_cache *cache)
 463 {
 464         struct backref_node *node;
 465         int level = 0;
 466
 467         if (cache->last_trans == 0) {
 468                 cache->last_trans = trans->transid;
 469                 return 0;
 470         }
 471
 472         if (cache->last_trans == trans->transid)
 473                 return 0;
 474
 475         /*
 476          * detached nodes are used to avoid unnecessary backref
 477          * lookup. transaction commit changes the extent tree.
 478          * so the detached nodes are no longer useful.
 479          */
 480         while (!list_empty(&cache->detached)) {
 481                 node = list_entry(cache->detached.next,
 482                                   struct backref_node, list);
 483                 remove_backref_node(cache, node);
 484         }
 485
 486         while (!list_empty(&cache->changed)) {
 487                 node = list_entry(cache->changed.next,
 488                                   struct backref_node, list);
 489                 list_del_init(&node->list);
 490                 BUG_ON(node->pending);
 491                 update_backref_node(cache, node, node->new_bytenr);
 492         }
 493
 494         /*
 495          * some nodes can be left in the pending list if there were
 496          * errors during processing the pending nodes.
 497          */
 498         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
 499                 list_for_each_entry(node, &cache->pending[level], list) {
 500                         BUG_ON(!node->pending);
 501                         if (node->bytenr == node->new_bytenr)
 502                                 continue;
 503                         update_backref_node(cache, node, node->new_bytenr);
 504                 }
 505         }
 506
 507         cache->last_trans = 0;
 508         return 1;
 509 }
 510
 511 static int should_ignore_root(struct btrfs_root *root)
 512 {
 513         struct btrfs_root *reloc_root;
 514
 515         if (!root->ref_cows)
 516                 return 0;
 517
 518         reloc_root = root->reloc_root;
 519         if (!reloc_root)
 520                 return 0;
 521
 522         if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
 523             root->fs_info->running_transaction->transid - 1)
 524                 return 0;
 525         /*
 526          * if there is reloc tree and it was created in previous
 527          * transaction backref lookup can find the reloc tree,
 528          * so backref node for the fs tree root is useless for
 529          * relocation.
 530          */
 531         return 1;
 532 }
 533
 534 /*
 535  * find reloc tree by address of tree root
 536  */
 537 static struct btrfs_root *find_reloc_root(struct reloc_control *rc,
 538                                           u64 bytenr)
 539 {
 540         struct rb_node *rb_node;
 541         struct mapping_node *node;
 542         struct btrfs_root *root = NULL;
 543
 544         spin_lock(&rc->reloc_root_tree.lock);
 545         rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr);
 546         if (rb_node) {
 547                 node = rb_entry(rb_node, struct mapping_node, rb_node);
 548                 root = (struct btrfs_root *)node->data;
 549         }
 550         spin_unlock(&rc->reloc_root_tree.lock);
 551         return root;
 552 }
 553
 554 static int is_cowonly_root(u64 root_objectid)
 555 {
 556         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
 557             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
 558             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
 559             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
 560             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
 561             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
 562                 return 1;
 563         return 0;
 564 }
 565
 566 static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info,
 567                                         u64 root_objectid)
 568 {
 569         struct btrfs_key key;
 570
 571         key.objectid = root_objectid;
 572         key.type = BTRFS_ROOT_ITEM_KEY;
 573         if (is_cowonly_root(root_objectid))
 574                 key.offset = 0;
 575         else
 576                 key.offset = (u64)-1;
 577
 578         return btrfs_read_fs_root_no_name(fs_info, &key);
 579 }
 580
 581 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 582 static noinline_for_stack
 583 struct btrfs_root *find_tree_root(struct reloc_control *rc,
 584                                   struct extent_buffer *leaf,
 585                                   struct btrfs_extent_ref_v0 *ref0)
 586 {
 587         struct btrfs_root *root;
 588         u64 root_objectid = btrfs_ref_root_v0(leaf, ref0);
 589         u64 generation = btrfs_ref_generation_v0(leaf, ref0);
 590
 591         BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID);
 592
 593         root = read_fs_root(rc->extent_root->fs_info, root_objectid);
 594         BUG_ON(IS_ERR(root));
 595
 596         if (root->ref_cows &&
 597             generation != btrfs_root_generation(&root->root_item))
 598                 return NULL;
 599
 600         return root;
 601 }
 602 #endif
 603
 604 static noinline_for_stack
 605 int find_inline_backref(struct extent_buffer *leaf, int slot,
 606                         unsigned long *ptr, unsigned long *end)
 607 {
 608         struct btrfs_extent_item *ei;
 609         struct btrfs_tree_block_info *bi;
 610         u32 item_size;
 611
 612         item_size = btrfs_item_size_nr(leaf, slot);
 613 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 614         if (item_size < sizeof(*ei)) {
 615                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
 616                 return 1;
 617         }
 618 #endif
 619         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 620         WARN_ON(!(btrfs_extent_flags(leaf, ei) &
 621                   BTRFS_EXTENT_FLAG_TREE_BLOCK));
 622
 623         if (item_size <= sizeof(*ei) + sizeof(*bi)) {
 624                 WARN_ON(item_size < sizeof(*ei) + sizeof(*bi));
 625                 return 1;
 626         }
 627
 628         bi = (struct btrfs_tree_block_info *)(ei + 1);
 629         *ptr = (unsigned long)(bi + 1);
 630         *end = (unsigned long)ei + item_size;
 631         return 0;
 632 }
 633
 634 /*
 635  * build backref tree for a given tree block. root of the backref tree
 636  * corresponds the tree block, leaves of the backref tree correspond
 637  * roots of b-trees that reference the tree block.
 638  *
 639  * the basic idea of this function is check backrefs of a given block
 640  * to find upper level blocks that refernece the block, and then check
 641  * bakcrefs of these upper level blocks recursively. the recursion stop
 642  * when tree root is reached or backrefs for the block is cached.
 643  *
 644  * NOTE: if we find backrefs for a block are cached, we know backrefs
 645  * for all upper level blocks that directly/indirectly reference the
 646  * block are also cached.
 647  */
 648 static noinline_for_stack
 649 struct backref_node *build_backref_tree(struct reloc_control *rc,
 650                                         struct btrfs_key *node_key,
 651                                         int level, u64 bytenr)
 652 {
 653         struct backref_cache *cache = &rc->backref_cache;
 654         struct btrfs_path *path1;
 655         struct btrfs_path *path2;
 656         struct extent_buffer *eb;
 657         struct btrfs_root *root;
 658         struct backref_node *cur;
 659         struct backref_node *upper;
 660         struct backref_node *lower;
 661         struct backref_node *node = NULL;
 662         struct backref_node *exist = NULL;
 663         struct backref_edge *edge;
 664         struct rb_node *rb_node;
 665         struct btrfs_key key;
 666         unsigned long end;
 667         unsigned long ptr;
 668         LIST_HEAD(list);
 669         LIST_HEAD(useless);
 670         int cowonly;
 671         int ret;
 672         int err = 0;
 673
 674         path1 = btrfs_alloc_path();
 675         path2 = btrfs_alloc_path();
 676         if (!path1 || !path2) {
 677                 err = -ENOMEM;
 678                 goto out;
 679         }
 680
 681         node = alloc_backref_node(cache);
 682         if (!node) {
 683                 err = -ENOMEM;
 684                 goto out;
 685         }
 686
 687         node->bytenr = bytenr;
 688         node->level = level;
 689         node->lowest = 1;
 690         cur = node;
 691 again:
 692         end = 0;
 693         ptr = 0;
 694         key.objectid = cur->bytenr;
 695         key.type = BTRFS_EXTENT_ITEM_KEY;
 696         key.offset = (u64)-1;
 697
 698         path1->search_commit_root = 1;
 699         path1->skip_locking = 1;
 700         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1,
 701                                 0, 0);
 702         if (ret < 0) {
 703                 err = ret;
 704                 goto out;
 705         }
 706         BUG_ON(!ret || !path1->slots[0]);
 707
 708         path1->slots[0]--;
 709
 710         WARN_ON(cur->checked);
 711         if (!list_empty(&cur->upper)) {
 712                 /*
 713                  * the backref was added previously when processsing
 714                  * backref of type BTRFS_TREE_BLOCK_REF_KEY
 715                  */
 716                 BUG_ON(!list_is_singular(&cur->upper));
 717                 edge = list_entry(cur->upper.next, struct backref_edge,
 718                                   list[LOWER]);
 719                 BUG_ON(!list_empty(&edge->list[UPPER]));
 720                 exist = edge->node[UPPER];
 721                 /*
 722                  * add the upper level block to pending list if we need
 723                  * check its backrefs
 724                  */
 725                 if (!exist->checked)
 726                         list_add_tail(&edge->list[UPPER], &list);
 727         } else {
 728                 exist = NULL;
 729         }
 730
 731         while (1) {
 732                 cond_resched();
 733                 eb = path1->nodes[0];
 734
 735                 if (ptr >= end) {
 736                         if (path1->slots[0] >= btrfs_header_nritems(eb)) {
 737                                 ret = btrfs_next_leaf(rc->extent_root, path1);
 738                                 if (ret < 0) {
 739                                         err = ret;
 740                                         goto out;
 741                                 }
 742                                 if (ret > 0)
 743                                         break;
 744                                 eb = path1->nodes[0];
 745                         }
 746
 747                         btrfs_item_key_to_cpu(eb, &key, path1->slots[0]);
 748                         if (key.objectid != cur->bytenr) {
 749                                 WARN_ON(exist);
 750                                 break;
 751                         }
 752
 753                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 754                                 ret = find_inline_backref(eb, path1->slots[0],
 755                                                           &ptr, &end);
 756                                 if (ret)
 757                                         goto next;
 758                         }
 759                 }
 760
 761                 if (ptr < end) {
 762                         /* update key for inline back ref */
 763                         struct btrfs_extent_inline_ref *iref;
 764                         iref = (struct btrfs_extent_inline_ref *)ptr;
 765                         key.type = btrfs_extent_inline_ref_type(eb, iref);
 766                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
 767                         WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY &&
 768                                 key.type != BTRFS_SHARED_BLOCK_REF_KEY);
 769                 }
 770
 771                 if (exist &&
 772                     ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
 773                       exist->owner == key.offset) ||
 774                      (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
 775                       exist->bytenr == key.offset))) {
 776                         exist = NULL;
 777                         goto next;
 778                 }
 779
 780 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 781                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
 782                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
 783                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
 784                                 struct btrfs_extent_ref_v0 *ref0;
 785                                 ref0 = btrfs_item_ptr(eb, path1->slots[0],
 786                                                 struct btrfs_extent_ref_v0);
 787                                 if (key.objectid == key.offset) {
 788                                         root = find_tree_root(rc, eb, ref0);
 789                                         if (root && !should_ignore_root(root))
 790                                                 cur->root = root;
 791                                         else
 792                                                 list_add(&cur->list, &useless);
 793                                         break;
 794                                 }
 795                                 if (is_cowonly_root(btrfs_ref_root_v0(eb,
 796                                                                       ref0)))
 797                                         cur->cowonly = 1;
 798                         }
 799 #else
 800                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
 801                 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
 802 #endif
 803                         if (key.objectid == key.offset) {
 804                                 /*
 805                                  * only root blocks of reloc trees use
 806                                  * backref of this type.
 807                                  */
 808                                 root = find_reloc_root(rc, cur->bytenr);
 809                                 BUG_ON(!root);
 810                                 cur->root = root;
 811                                 break;
 812                         }
 813
 814                         edge = alloc_backref_edge(cache);
 815                         if (!edge) {
 816                                 err = -ENOMEM;
 817                                 goto out;
 818                         }
 819                         rb_node = tree_search(&cache->rb_root, key.offset);
 820                         if (!rb_node) {
 821                                 upper = alloc_backref_node(cache);
 822                                 if (!upper) {
 823                                         free_backref_edge(cache, edge);
 824                                         err = -ENOMEM;
 825                                         goto out;
 826                                 }
 827                                 upper->bytenr = key.offset;
 828                                 upper->level = cur->level + 1;
 829                                 /*
 830                                  *  backrefs for the upper level block isn't
 831                                  *  cached, add the block to pending list
 832                                  */
 833                                 list_add_tail(&edge->list[UPPER], &list);
 834                         } else {
 835                                 upper = rb_entry(rb_node, struct backref_node,
 836                                                  rb_node);
 837                                 BUG_ON(!upper->checked);
 838                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 839                         }
 840                         list_add_tail(&edge->list[LOWER], &cur->upper);
 841                         edge->node[LOWER] = cur;
 842                         edge->node[UPPER] = upper;
 843
 844                         goto next;
 845                 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
 846                         goto next;
 847                 }
 848
 849                 /* key.type == BTRFS_TREE_BLOCK_REF_KEY */
 850                 root = read_fs_root(rc->extent_root->fs_info, key.offset);
 851                 if (IS_ERR(root)) {
 852                         err = PTR_ERR(root);
 853                         goto out;
 854                 }
 855
 856                 if (!root->ref_cows)
 857                         cur->cowonly = 1;
 858
 859                 if (btrfs_root_level(&root->root_item) == cur->level) {
 860                         /* tree root */
 861                         BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 862                                cur->bytenr);
 863                         if (should_ignore_root(root))
 864                                 list_add(&cur->list, &useless);
 865                         else
 866                                 cur->root = root;
 867                         break;
 868                 }
 869
 870                 level = cur->level + 1;
 871
 872                 /*
 873                  * searching the tree to find upper level blocks
 874                  * reference the block.
 875                  */
 876                 path2->search_commit_root = 1;
 877                 path2->skip_locking = 1;
 878                 path2->lowest_level = level;
 879                 ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0);
 880                 path2->lowest_level = 0;
 881                 if (ret < 0) {
 882                         err = ret;
 883                         goto out;
 884                 }
 885                 if (ret > 0 && path2->slots[level] > 0)
 886                         path2->slots[level]--;
 887
 888                 eb = path2->nodes[level];
 889                 WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) !=
 890                         cur->bytenr);
 891
 892                 lower = cur;
 893                 for (; level < BTRFS_MAX_LEVEL; level++) {
 894                         if (!path2->nodes[level]) {
 895                                 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
 896                                        lower->bytenr);
 897                                 if (should_ignore_root(root))
 898                                         list_add(&lower->list, &useless);
 899                                 else
 900                                         lower->root = root;
 901                                 break;
 902                         }
 903
 904                         edge = alloc_backref_edge(cache);
 905                         if (!edge) {
 906                                 err = -ENOMEM;
 907                                 goto out;
 908                         }
 909
 910                         eb = path2->nodes[level];
 911                         rb_node = tree_search(&cache->rb_root, eb->start);
 912                         if (!rb_node) {
 913                                 upper = alloc_backref_node(cache);
 914                                 if (!upper) {
 915                                         free_backref_edge(cache, edge);
 916                                         err = -ENOMEM;
 917                                         goto out;
 918                                 }
 919                                 upper->bytenr = eb->start;
 920                                 upper->owner = btrfs_header_owner(eb);
 921                                 upper->level = lower->level + 1;
 922                                 if (!root->ref_cows)
 923                                         upper->cowonly = 1;
 924
 925                                 /*
 926                                  * if we know the block isn't shared
 927                                  * we can void checking its backrefs.
 928                                  */
 929                                 if (btrfs_block_can_be_shared(root, eb))
 930                                         upper->checked = 0;
 931                                 else
 932                                         upper->checked = 1;
 933
 934                                 /*
 935                                  * add the block to pending list if we
 936                                  * need check its backrefs. only block
 937                                  * at 'cur->level + 1' is added to the
 938                                  * tail of pending list. this guarantees
 939                                  * we check backrefs from lower level
 940                                  * blocks to upper level blocks.
 941                                  */
 942                                 if (!upper->checked &&
 943                                     level == cur->level + 1) {
 944                                         list_add_tail(&edge->list[UPPER],
 945                                                       &list);
 946                                 } else
 947                                         INIT_LIST_HEAD(&edge->list[UPPER]);
 948                         } else {
 949                                 upper = rb_entry(rb_node, struct backref_node,
 950                                                  rb_node);
 951                                 BUG_ON(!upper->checked);
 952                                 INIT_LIST_HEAD(&edge->list[UPPER]);
 953                                 if (!upper->owner)
 954                                         upper->owner = btrfs_header_owner(eb);
 955                         }
 956                         list_add_tail(&edge->list[LOWER], &lower->upper);
 957                         edge->node[LOWER] = lower;
 958                         edge->node[UPPER] = upper;
 959
 960                         if (rb_node)
 961                                 break;
 962                         lower = upper;
 963                         upper = NULL;
 964                 }
 965                 btrfs_release_path(root, path2);
 966 next:
 967                 if (ptr < end) {
 968                         ptr += btrfs_extent_inline_ref_size(key.type);
 969                         if (ptr >= end) {
 970                                 WARN_ON(ptr > end);
 971                                 ptr = 0;
 972                                 end = 0;
 973                         }
 974                 }
 975                 if (ptr >= end)
 976                         path1->slots[0]++;
 977         }
 978         btrfs_release_path(rc->extent_root, path1);
 979
 980         cur->checked = 1;
 981         WARN_ON(exist);
 982
 983         /* the pending list isn't empty, take the first block to process */
 984         if (!list_empty(&list)) {
 985                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
 986                 list_del_init(&edge->list[UPPER]);
 987                 cur = edge->node[UPPER];
 988                 goto again;
 989         }
 990
 991         /*
 992          * everything goes well, connect backref nodes and insert backref nodes
 993          * into the cache.
 994          */
 995         BUG_ON(!node->checked);
 996         cowonly = node->cowonly;
 997         if (!cowonly) {
 998                 rb_node = tree_insert(&cache->rb_root, node->bytenr,
 999                                       &node->rb_node);
1000                 BUG_ON(rb_node);
1001                 list_add_tail(&node->lower, &cache->leaves);
1002         }
1003
1004         list_for_each_entry(edge, &node->upper, list[LOWER])
1005                 list_add_tail(&edge->list[UPPER], &list);
1006
1007         while (!list_empty(&list)) {
1008                 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
1009                 list_del_init(&edge->list[UPPER]);
1010                 upper = edge->node[UPPER];
1011                 if (upper->detached) {
1012                         list_del(&edge->list[LOWER]);
1013                         lower = edge->node[LOWER];
1014                         free_backref_edge(cache, edge);
1015                         if (list_empty(&lower->upper))
1016                                 list_add(&lower->list, &useless);
1017                         continue;
1018                 }
1019
1020                 if (!RB_EMPTY_NODE(&upper->rb_node)) {
1021                         if (upper->lowest) {
1022                                 list_del_init(&upper->lower);
1023                                 upper->lowest = 0;
1024                         }
1025
1026                         list_add_tail(&edge->list[UPPER], &upper->lower);
1027                         continue;
1028                 }
1029
1030                 BUG_ON(!upper->checked);
1031                 BUG_ON(cowonly != upper->cowonly);
1032                 if (!cowonly) {
1033                         rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1034                                               &upper->rb_node);
1035                         BUG_ON(rb_node);
1036                 }
1037
1038                 list_add_tail(&edge->list[UPPER], &upper->lower);
1039
1040                 list_for_each_entry(edge, &upper->upper, list[LOWER])
1041                         list_add_tail(&edge->list[UPPER], &list);
1042         }
1043         /*
1044          * process useless backref nodes. backref nodes for tree leaves
1045          * are deleted from the cache. backref nodes for upper level
1046          * tree blocks are left in the cache to avoid unnecessary backref
1047          * lookup.
1048          */
1049         while (!list_empty(&useless)) {
1050                 upper = list_entry(useless.next, struct backref_node, list);
1051                 list_del_init(&upper->list);
1052                 BUG_ON(!list_empty(&upper->upper));
1053                 if (upper == node)
1054                         node = NULL;
1055                 if (upper->lowest) {
1056                         list_del_init(&upper->lower);
1057                         upper->lowest = 0;
1058                 }
1059                 while (!list_empty(&upper->lower)) {
1060                         edge = list_entry(upper->lower.next,
1061                                           struct backref_edge, list[UPPER]);
1062                         list_del(&edge->list[UPPER]);
1063                         list_del(&edge->list[LOWER]);
1064                         lower = edge->node[LOWER];
1065                         free_backref_edge(cache, edge);
1066
1067                         if (list_empty(&lower->upper))
1068                                 list_add(&lower->list, &useless);
1069                 }
1070                 __mark_block_processed(rc, upper);
1071                 if (upper->level > 0) {
1072                         list_add(&upper->list, &cache->detached);
1073                         upper->detached = 1;
1074                 } else {
1075                         rb_erase(&upper->rb_node, &cache->rb_root);
1076                         free_backref_node(cache, upper);
1077                 }
1078         }
1079 out:
1080         btrfs_free_path(path1);
1081         btrfs_free_path(path2);
1082         if (err) {
1083                 while (!list_empty(&useless)) {
1084                         lower = list_entry(useless.next,
1085                                            struct backref_node, upper);
1086                         list_del_init(&lower->upper);
1087                 }
1088                 upper = node;
1089                 INIT_LIST_HEAD(&list);
1090                 while (upper) {
1091                         if (RB_EMPTY_NODE(&upper->rb_node)) {
1092                                 list_splice_tail(&upper->upper, &list);
1093                                 free_backref_node(cache, upper);
1094                         }
1095
1096                         if (list_empty(&list))
1097                                 break;
1098
1099                         edge = list_entry(list.next, struct backref_edge,
1100                                           list[LOWER]);
1101                         list_del(&edge->list[LOWER]);
1102                         upper = edge->node[UPPER];
1103                         free_backref_edge(cache, edge);
1104                 }
1105                 return ERR_PTR(err);
1106         }
1107         BUG_ON(node && node->detached);
1108         return node;
1109 }
1110
1111 /*
1112  * helper to add backref node for the newly created snapshot.
1113  * the backref node is created by cloning backref node that
1114  * corresponds to root of source tree
1115  */
1116 static int clone_backref_node(struct btrfs_trans_handle *trans,
1117                               struct reloc_control *rc,
1118                               struct btrfs_root *src,
1119                               struct btrfs_root *dest)
1120 {
1121         struct btrfs_root *reloc_root = src->reloc_root;
1122         struct backref_cache *cache = &rc->backref_cache;
1123         struct backref_node *node = NULL;
1124         struct backref_node *new_node;
1125         struct backref_edge *edge;
1126         struct backref_edge *new_edge;
1127         struct rb_node *rb_node;
1128
1129         if (cache->last_trans > 0)
1130                 update_backref_cache(trans, cache);
1131
1132         rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1133         if (rb_node) {
1134                 node = rb_entry(rb_node, struct backref_node, rb_node);
1135                 if (node->detached)
1136                         node = NULL;
1137                 else
1138                         BUG_ON(node->new_bytenr != reloc_root->node->start);
1139         }
1140
1141         if (!node) {
1142                 rb_node = tree_search(&cache->rb_root,
1143                                       reloc_root->commit_root->start);
1144                 if (rb_node) {
1145                         node = rb_entry(rb_node, struct backref_node,
1146                                         rb_node);
1147                         BUG_ON(node->detached);
1148                 }
1149         }
1150
1151         if (!node)
1152                 return 0;
1153
1154         new_node = alloc_backref_node(cache);
1155         if (!new_node)
1156                 return -ENOMEM;
1157
1158         new_node->bytenr = dest->node->start;
1159         new_node->level = node->level;
1160         new_node->lowest = node->lowest;
1161         new_node->checked = 1;
1162         new_node->root = dest;
1163
1164         if (!node->lowest) {
1165                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1166                         new_edge = alloc_backref_edge(cache);
1167                         if (!new_edge)
1168                                 goto fail;
1169
1170                         new_edge->node[UPPER] = new_node;
1171                         new_edge->node[LOWER] = edge->node[LOWER];
1172                         list_add_tail(&new_edge->list[UPPER],
1173                                       &new_node->lower);
1174                 }
1175         }
1176
1177         rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1178                               &new_node->rb_node);
1179         BUG_ON(rb_node);
1180
1181         if (!new_node->lowest) {
1182                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1183                         list_add_tail(&new_edge->list[LOWER],
1184                                       &new_edge->node[LOWER]->upper);
1185                 }
1186         }
1187         return 0;
1188 fail:
1189         while (!list_empty(&new_node->lower)) {
1190                 new_edge = list_entry(new_node->lower.next,
1191                                       struct backref_edge, list[UPPER]);
1192                 list_del(&new_edge->list[UPPER]);
1193                 free_backref_edge(cache, new_edge);
1194         }
1195         free_backref_node(cache, new_node);
1196         return -ENOMEM;
1197 }
1198
1199 /*
1200  * helper to add 'address of tree root -> reloc tree' mapping
1201  */
1202 static int __add_reloc_root(struct btrfs_root *root)
1203 {
1204         struct rb_node *rb_node;
1205         struct mapping_node *node;
1206         struct reloc_control *rc = root->fs_info->reloc_ctl;
1207
1208         node = kmalloc(sizeof(*node), GFP_NOFS);
1209         BUG_ON(!node);
1210
1211         node->bytenr = root->node->start;
1212         node->data = root;
1213
1214         spin_lock(&rc->reloc_root_tree.lock);
1215         rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1216                               node->bytenr, &node->rb_node);
1217         spin_unlock(&rc->reloc_root_tree.lock);
1218         BUG_ON(rb_node);
1219
1220         list_add_tail(&root->root_list, &rc->reloc_roots);
1221         return 0;
1222 }
1223
1224 /*
1225  * helper to update/delete the 'address of tree root -> reloc tree'
1226  * mapping
1227  */
1228 static int __update_reloc_root(struct btrfs_root *root, int del)
1229 {
1230         struct rb_node *rb_node;
1231         struct mapping_node *node = NULL;
1232         struct reloc_control *rc = root->fs_info->reloc_ctl;
1233
1234         spin_lock(&rc->reloc_root_tree.lock);
1235         rb_node = tree_search(&rc->reloc_root_tree.rb_root,
1236                               root->commit_root->start);
1237         if (rb_node) {
1238                 node = rb_entry(rb_node, struct mapping_node, rb_node);
1239                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
1240         }
1241         spin_unlock(&rc->reloc_root_tree.lock);
1242
1243         BUG_ON((struct btrfs_root *)node->data != root);
1244
1245         if (!del) {
1246                 spin_lock(&rc->reloc_root_tree.lock);
1247                 node->bytenr = root->node->start;
1248                 rb_node = tree_insert(&rc->reloc_root_tree.rb_root,
1249                                       node->bytenr, &node->rb_node);
1250                 spin_unlock(&rc->reloc_root_tree.lock);
1251                 BUG_ON(rb_node);
1252         } else {
1253                 list_del_init(&root->root_list);
1254                 kfree(node);
1255         }
1256         return 0;
1257 }
1258
1259 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
1260                                         struct btrfs_root *root, u64 objectid)
1261 {
1262         struct btrfs_root *reloc_root;
1263         struct extent_buffer *eb;
1264         struct btrfs_root_item *root_item;
1265         struct btrfs_key root_key;
1266         int ret;
1267
1268         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
1269         BUG_ON(!root_item);
1270
1271         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
1272         root_key.type = BTRFS_ROOT_ITEM_KEY;
1273         root_key.offset = objectid;
1274
1275         if (root->root_key.objectid == objectid) {
1276                 /* called by btrfs_init_reloc_root */
1277                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1278                                       BTRFS_TREE_RELOC_OBJECTID);
1279                 BUG_ON(ret);
1280
1281                 btrfs_set_root_last_snapshot(&root->root_item,
1282                                              trans->transid - 1);
1283         } else {
1284                 /*
1285                  * called by btrfs_reloc_post_snapshot_hook.
1286                  * the source tree is a reloc tree, all tree blocks
1287                  * modified after it was created have RELOC flag
1288                  * set in their headers. so it's OK to not update
1289                  * the 'last_snapshot'.
1290                  */
1291                 ret = btrfs_copy_root(trans, root, root->node, &eb,
1292                                       BTRFS_TREE_RELOC_OBJECTID);
1293                 BUG_ON(ret);
1294         }
1295
1296         memcpy(root_item, &root->root_item, sizeof(*root_item));
1297         btrfs_set_root_bytenr(root_item, eb->start);
1298         btrfs_set_root_level(root_item, btrfs_header_level(eb));
1299         btrfs_set_root_generation(root_item, trans->transid);
1300
1301         if (root->root_key.objectid == objectid) {
1302                 btrfs_set_root_refs(root_item, 0);
1303                 memset(&root_item->drop_progress, 0,
1304                        sizeof(struct btrfs_disk_key));
1305                 root_item->drop_level = 0;
1306         }
1307
1308         btrfs_tree_unlock(eb);
1309         free_extent_buffer(eb);
1310
1311         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1312                                 &root_key, root_item);
1313         BUG_ON(ret);
1314         kfree(root_item);
1315
1316         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
1317                                                  &root_key);
1318         BUG_ON(IS_ERR(reloc_root));
1319         reloc_root->last_trans = trans->transid;
1320         return reloc_root;
1321 }
1322
1323 /*
1324  * create reloc tree for a given fs tree. reloc tree is just a
1325  * snapshot of the fs tree with special root objectid.
1326  */
1327 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1328                           struct btrfs_root *root)
1329 {
1330         struct btrfs_root *reloc_root;
1331         struct reloc_control *rc = root->fs_info->reloc_ctl;
1332         int clear_rsv = 0;
1333
1334         if (root->reloc_root) {
1335                 reloc_root = root->reloc_root;
1336                 reloc_root->last_trans = trans->transid;
1337                 return 0;
1338         }
1339
1340         if (!rc || !rc->create_reloc_tree ||
1341             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1342                 return 0;
1343
1344         if (!trans->block_rsv) {
1345                 trans->block_rsv = rc->block_rsv;
1346                 clear_rsv = 1;
1347         }
1348         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1349         if (clear_rsv)
1350                 trans->block_rsv = NULL;
1351
1352         __add_reloc_root(reloc_root);
1353         root->reloc_root = reloc_root;
1354         return 0;
1355 }
1356
1357 /*
1358  * update root item of reloc tree
1359  */
1360 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
1361                             struct btrfs_root *root)
1362 {
1363         struct btrfs_root *reloc_root;
1364         struct btrfs_root_item *root_item;
1365         int del = 0;
1366         int ret;
1367
1368         if (!root->reloc_root)
1369                 return 0;
1370
1371         reloc_root = root->reloc_root;
1372         root_item = &reloc_root->root_item;
1373
1374         if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1375             btrfs_root_refs(root_item) == 0) {
1376                 root->reloc_root = NULL;
1377                 del = 1;
1378         }
1379
1380         __update_reloc_root(reloc_root, del);
1381
1382         if (reloc_root->commit_root != reloc_root->node) {
1383                 btrfs_set_root_node(root_item, reloc_root->node);
1384                 free_extent_buffer(reloc_root->commit_root);
1385                 reloc_root->commit_root = btrfs_root_node(reloc_root);
1386         }
1387
1388         ret = btrfs_update_root(trans, root->fs_info->tree_root,
1389                                 &reloc_root->root_key, root_item);
1390         BUG_ON(ret);
1391         return 0;
1392 }
1393
1394 /*
1395  * helper to find first cached inode with inode number >= objectid
1396  * in a subvolume
1397  */
1398 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
1399 {
1400         struct rb_node *node;
1401         struct rb_node *prev;
1402         struct btrfs_inode *entry;
1403         struct inode *inode;
1404
1405         spin_lock(&root->inode_lock);
1406 again:
1407         node = root->inode_tree.rb_node;
1408         prev = NULL;
1409         while (node) {
1410                 prev = node;
1411                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1412
1413                 if (objectid < entry->vfs_inode.i_ino)
1414                         node = node->rb_left;
1415                 else if (objectid > entry->vfs_inode.i_ino)
1416                         node = node->rb_right;
1417                 else
1418                         break;
1419         }
1420         if (!node) {
1421                 while (prev) {
1422                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
1423                         if (objectid <= entry->vfs_inode.i_ino) {
1424                                 node = prev;
1425                                 break;
1426                         }
1427                         prev = rb_next(prev);
1428                 }
1429         }
1430         while (node) {
1431                 entry = rb_entry(node, struct btrfs_inode, rb_node);
1432                 inode = igrab(&entry->vfs_inode);
1433                 if (inode) {
1434                         spin_unlock(&root->inode_lock);
1435                         return inode;
1436                 }
1437
1438                 objectid = entry->vfs_inode.i_ino + 1;
1439                 if (cond_resched_lock(&root->inode_lock))
1440                         goto again;
1441
1442                 node = rb_next(node);
1443         }
1444         spin_unlock(&root->inode_lock);
1445         return NULL;
1446 }
1447
1448 static int in_block_group(u64 bytenr,
1449                           struct btrfs_block_group_cache *block_group)
1450 {
1451         if (bytenr >= block_group->key.objectid &&
1452             bytenr < block_group->key.objectid + block_group->key.offset)
1453                 return 1;
1454         return 0;
1455 }
1456
1457 /*
1458  * get new location of data
1459  */
1460 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1461                             u64 bytenr, u64 num_bytes)
1462 {
1463         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
1464         struct btrfs_path *path;
1465         struct btrfs_file_extent_item *fi;
1466         struct extent_buffer *leaf;
1467         int ret;
1468
1469         path = btrfs_alloc_path();
1470         if (!path)
1471                 return -ENOMEM;
1472
1473         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1474         ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino,
1475                                        bytenr, 0);
1476         if (ret < 0)
1477                 goto out;
1478         if (ret > 0) {
1479                 ret = -ENOENT;
1480                 goto out;
1481         }
1482
1483         leaf = path->nodes[0];
1484         fi = btrfs_item_ptr(leaf, path->slots[0],
1485                             struct btrfs_file_extent_item);
1486
1487         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1488                btrfs_file_extent_compression(leaf, fi) ||
1489                btrfs_file_extent_encryption(leaf, fi) ||
1490                btrfs_file_extent_other_encoding(leaf, fi));
1491
1492         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1493                 ret = 1;
1494                 goto out;
1495         }
1496
1497         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1498         ret = 0;
1499 out:
1500         btrfs_free_path(path);
1501         return ret;
1502 }
1503
1504 /*
1505  * update file extent items in the tree leaf to point to
1506  * the new locations.
1507  */
1508 static noinline_for_stack
1509 int replace_file_extents(struct btrfs_trans_handle *trans,
1510                          struct reloc_control *rc,
1511                          struct btrfs_root *root,
1512                          struct extent_buffer *leaf)
1513 {
1514         struct btrfs_key key;
1515         struct btrfs_file_extent_item *fi;
1516         struct inode *inode = NULL;
1517         u64 parent;
1518         u64 bytenr;
1519         u64 new_bytenr = 0;
1520         u64 num_bytes;
1521         u64 end;
1522         u32 nritems;
1523         u32 i;
1524         int ret;
1525         int first = 1;
1526         int dirty = 0;
1527
1528         if (rc->stage != UPDATE_DATA_PTRS)
1529                 return 0;
1530
1531         /* reloc trees always use full backref */
1532         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1533                 parent = leaf->start;
1534         else
1535                 parent = 0;
1536
1537         nritems = btrfs_header_nritems(leaf);
1538         for (i = 0; i < nritems; i++) {
1539                 cond_resched();
1540                 btrfs_item_key_to_cpu(leaf, &key, i);
1541                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1542                         continue;
1543                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1544                 if (btrfs_file_extent_type(leaf, fi) ==
1545                     BTRFS_FILE_EXTENT_INLINE)
1546                         continue;
1547                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1548                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1549                 if (bytenr == 0)
1550                         continue;
1551                 if (!in_block_group(bytenr, rc->block_group))
1552                         continue;
1553
1554                 /*
1555                  * if we are modifying block in fs tree, wait for readpage
1556                  * to complete and drop the extent cache
1557                  */
1558                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1559                         if (first) {
1560                                 inode = find_next_inode(root, key.objectid);
1561                                 first = 0;
1562                         } else if (inode && inode->i_ino < key.objectid) {
1563                                 btrfs_add_delayed_iput(inode);
1564                                 inode = find_next_inode(root, key.objectid);
1565                         }
1566                         if (inode && inode->i_ino == key.objectid) {
1567                                 end = key.offset +
1568                                       btrfs_file_extent_num_bytes(leaf, fi);
1569                                 WARN_ON(!IS_ALIGNED(key.offset,
1570                                                     root->sectorsize));
1571                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1572                                 end--;
1573                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1574                                                       key.offset, end,
1575                                                       GFP_NOFS);
1576                                 if (!ret)
1577                                         continue;
1578
1579                                 btrfs_drop_extent_cache(inode, key.offset, end,
1580                                                         1);
1581                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1582                                               key.offset, end, GFP_NOFS);
1583                         }
1584                 }
1585
1586                 ret = get_new_location(rc->data_inode, &new_bytenr,
1587                                        bytenr, num_bytes);
1588                 if (ret > 0) {
1589                         WARN_ON(1);
1590                         continue;
1591                 }
1592                 BUG_ON(ret < 0);
1593
1594                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1595                 dirty = 1;
1596
1597                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1598                 ret = btrfs_inc_extent_ref(trans, root, new_bytenr,
1599                                            num_bytes, parent,
1600                                            btrfs_header_owner(leaf),
1601                                            key.objectid, key.offset);
1602                 BUG_ON(ret);
1603
1604                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1605                                         parent, btrfs_header_owner(leaf),
1606                                         key.objectid, key.offset);
1607                 BUG_ON(ret);
1608         }
1609         if (dirty)
1610                 btrfs_mark_buffer_dirty(leaf);
1611         if (inode)
1612                 btrfs_add_delayed_iput(inode);
1613         return 0;
1614 }
1615
1616 static noinline_for_stack
1617 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1618                      struct btrfs_path *path, int level)
1619 {
1620         struct btrfs_disk_key key1;
1621         struct btrfs_disk_key key2;
1622         btrfs_node_key(eb, &key1, slot);
1623         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1624         return memcmp(&key1, &key2, sizeof(key1));
1625 }
1626
1627 /*
1628  * try to replace tree blocks in fs tree with the new blocks
1629  * in reloc tree. tree blocks haven't been modified since the
1630  * reloc tree was create can be replaced.
1631  *
1632  * if a block was replaced, level of the block + 1 is returned.
1633  * if no block got replaced, 0 is returned. if there are other
1634  * errors, a negative error number is returned.
1635  */
1636 static noinline_for_stack
1637 int replace_path(struct btrfs_trans_handle *trans,
1638                  struct btrfs_root *dest, struct btrfs_root *src,
1639                  struct btrfs_path *path, struct btrfs_key *next_key,
1640                  int lowest_level, int max_level)
1641 {
1642         struct extent_buffer *eb;
1643         struct extent_buffer *parent;
1644         struct btrfs_key key;
1645         u64 old_bytenr;
1646         u64 new_bytenr;
1647         u64 old_ptr_gen;
1648         u64 new_ptr_gen;
1649         u64 last_snapshot;
1650         u32 blocksize;
1651         int cow = 0;
1652         int level;
1653         int ret;
1654         int slot;
1655
1656         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1657         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1658
1659         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1660 again:
1661         slot = path->slots[lowest_level];
1662         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1663
1664         eb = btrfs_lock_root_node(dest);
1665         btrfs_set_lock_blocking(eb);
1666         level = btrfs_header_level(eb);
1667
1668         if (level < lowest_level) {
1669                 btrfs_tree_unlock(eb);
1670                 free_extent_buffer(eb);
1671                 return 0;
1672         }
1673
1674         if (cow) {
1675                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1676                 BUG_ON(ret);
1677         }
1678         btrfs_set_lock_blocking(eb);
1679
1680         if (next_key) {
1681                 next_key->objectid = (u64)-1;
1682                 next_key->type = (u8)-1;
1683                 next_key->offset = (u64)-1;
1684         }
1685
1686         parent = eb;
1687         while (1) {
1688                 level = btrfs_header_level(parent);
1689                 BUG_ON(level < lowest_level);
1690
1691                 ret = btrfs_bin_search(parent, &key, level, &slot);
1692                 if (ret && slot > 0)
1693                         slot--;
1694
1695                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1696                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1697
1698                 old_bytenr = btrfs_node_blockptr(parent, slot);
1699                 blocksize = btrfs_level_size(dest, level - 1);
1700                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1701
1702                 if (level <= max_level) {
1703                         eb = path->nodes[level];
1704                         new_bytenr = btrfs_node_blockptr(eb,
1705                                                         path->slots[level]);
1706                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1707                                                         path->slots[level]);
1708                 } else {
1709                         new_bytenr = 0;
1710                         new_ptr_gen = 0;
1711                 }
1712
1713                 if (new_bytenr > 0 && new_bytenr == old_bytenr) {
1714                         WARN_ON(1);
1715                         ret = level;
1716                         break;
1717                 }
1718
1719                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1720                     memcmp_node_keys(parent, slot, path, level)) {
1721                         if (level <= lowest_level) {
1722                                 ret = 0;
1723                                 break;
1724                         }
1725
1726                         eb = read_tree_block(dest, old_bytenr, blocksize,
1727                                              old_ptr_gen);
1728                         BUG_ON(!eb);
1729                         btrfs_tree_lock(eb);
1730                         if (cow) {
1731                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1732                                                       slot, &eb);
1733                                 BUG_ON(ret);
1734                         }
1735                         btrfs_set_lock_blocking(eb);
1736
1737                         btrfs_tree_unlock(parent);
1738                         free_extent_buffer(parent);
1739
1740                         parent = eb;
1741                         continue;
1742                 }
1743
1744                 if (!cow) {
1745                         btrfs_tree_unlock(parent);
1746                         free_extent_buffer(parent);
1747                         cow = 1;
1748                         goto again;
1749                 }
1750
1751                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1752                                       path->slots[level]);
1753                 btrfs_release_path(src, path);
1754
1755                 path->lowest_level = level;
1756                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1757                 path->lowest_level = 0;
1758                 BUG_ON(ret);
1759
1760                 /*
1761                  * swap blocks in fs tree and reloc tree.
1762                  */
1763                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1764                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1765                 btrfs_mark_buffer_dirty(parent);
1766
1767                 btrfs_set_node_blockptr(path->nodes[level],
1768                                         path->slots[level], old_bytenr);
1769                 btrfs_set_node_ptr_generation(path->nodes[level],
1770                                               path->slots[level], old_ptr_gen);
1771                 btrfs_mark_buffer_dirty(path->nodes[level]);
1772
1773                 ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize,
1774                                         path->nodes[level]->start,
1775                                         src->root_key.objectid, level - 1, 0);
1776                 BUG_ON(ret);
1777                 ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize,
1778                                         0, dest->root_key.objectid, level - 1,
1779                                         0);
1780                 BUG_ON(ret);
1781
1782                 ret = btrfs_free_extent(trans, src, new_bytenr, blocksize,
1783                                         path->nodes[level]->start,
1784                                         src->root_key.objectid, level - 1, 0);
1785                 BUG_ON(ret);
1786
1787                 ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize,
1788                                         0, dest->root_key.objectid, level - 1,
1789                                         0);
1790                 BUG_ON(ret);
1791
1792                 btrfs_unlock_up_safe(path, 0);
1793
1794                 ret = level;
1795                 break;
1796         }
1797         btrfs_tree_unlock(parent);
1798         free_extent_buffer(parent);
1799         return ret;
1800 }
1801
1802 /*
1803  * helper to find next relocated block in reloc tree
1804  */
1805 static noinline_for_stack
1806 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1807                        int *level)
1808 {
1809         struct extent_buffer *eb;
1810         int i;
1811         u64 last_snapshot;
1812         u32 nritems;
1813
1814         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1815
1816         for (i = 0; i < *level; i++) {
1817                 free_extent_buffer(path->nodes[i]);
1818                 path->nodes[i] = NULL;
1819         }
1820
1821         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1822                 eb = path->nodes[i];
1823                 nritems = btrfs_header_nritems(eb);
1824                 while (path->slots[i] + 1 < nritems) {
1825                         path->slots[i]++;
1826                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1827                             last_snapshot)
1828                                 continue;
1829
1830                         *level = i;
1831                         return 0;
1832                 }
1833                 free_extent_buffer(path->nodes[i]);
1834                 path->nodes[i] = NULL;
1835         }
1836         return 1;
1837 }
1838
1839 /*
1840  * walk down reloc tree to find relocated block of lowest level
1841  */
1842 static noinline_for_stack
1843 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1844                          int *level)
1845 {
1846         struct extent_buffer *eb = NULL;
1847         int i;
1848         u64 bytenr;
1849         u64 ptr_gen = 0;
1850         u64 last_snapshot;
1851         u32 blocksize;
1852         u32 nritems;
1853
1854         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1855
1856         for (i = *level; i > 0; i--) {
1857                 eb = path->nodes[i];
1858                 nritems = btrfs_header_nritems(eb);
1859                 while (path->slots[i] < nritems) {
1860                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1861                         if (ptr_gen > last_snapshot)
1862                                 break;
1863                         path->slots[i]++;
1864                 }
1865                 if (path->slots[i] >= nritems) {
1866                         if (i == *level)
1867                                 break;
1868                         *level = i + 1;
1869                         return 0;
1870                 }
1871                 if (i == 1) {
1872                         *level = i;
1873                         return 0;
1874                 }
1875
1876                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1877                 blocksize = btrfs_level_size(root, i - 1);
1878                 eb = read_tree_block(root, bytenr, blocksize, ptr_gen);
1879                 BUG_ON(btrfs_header_level(eb) != i - 1);
1880                 path->nodes[i - 1] = eb;
1881                 path->slots[i - 1] = 0;
1882         }
1883         return 1;
1884 }
1885
1886 /*
1887  * invalidate extent cache for file extents whose key in range of
1888  * [min_key, max_key)
1889  */
1890 static int invalidate_extent_cache(struct btrfs_root *root,
1891                                    struct btrfs_key *min_key,
1892                                    struct btrfs_key *max_key)
1893 {
1894         struct inode *inode = NULL;
1895         u64 objectid;
1896         u64 start, end;
1897
1898         objectid = min_key->objectid;
1899         while (1) {
1900                 cond_resched();
1901                 iput(inode);
1902
1903                 if (objectid > max_key->objectid)
1904                         break;
1905
1906                 inode = find_next_inode(root, objectid);
1907                 if (!inode)
1908                         break;
1909
1910                 if (inode->i_ino > max_key->objectid) {
1911                         iput(inode);
1912                         break;
1913                 }
1914
1915                 objectid = inode->i_ino + 1;
1916                 if (!S_ISREG(inode->i_mode))
1917                         continue;
1918
1919                 if (unlikely(min_key->objectid == inode->i_ino)) {
1920                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1921                                 continue;
1922                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1923                                 start = 0;
1924                         else {
1925                                 start = min_key->offset;
1926                                 WARN_ON(!IS_ALIGNED(start, root->sectorsize));
1927                         }
1928                 } else {
1929                         start = 0;
1930                 }
1931
1932                 if (unlikely(max_key->objectid == inode->i_ino)) {
1933                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1934                                 continue;
1935                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1936                                 end = (u64)-1;
1937                         } else {
1938                                 if (max_key->offset == 0)
1939                                         continue;
1940                                 end = max_key->offset;
1941                                 WARN_ON(!IS_ALIGNED(end, root->sectorsize));
1942                                 end--;
1943                         }
1944                 } else {
1945                         end = (u64)-1;
1946                 }
1947
1948                 /* the lock_extent waits for readpage to complete */
1949                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1950                 btrfs_drop_extent_cache(inode, start, end, 1);
1951                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
1952         }
1953         return 0;
1954 }
1955
1956 static int find_next_key(struct btrfs_path *path, int level,
1957                          struct btrfs_key *key)
1958
1959 {
1960         while (level < BTRFS_MAX_LEVEL) {
1961                 if (!path->nodes[level])
1962                         break;
1963                 if (path->slots[level] + 1 <
1964                     btrfs_header_nritems(path->nodes[level])) {
1965                         btrfs_node_key_to_cpu(path->nodes[level], key,
1966                                               path->slots[level] + 1);
1967                         return 0;
1968                 }
1969                 level++;
1970         }
1971         return 1;
1972 }
1973
1974 /*
1975  * merge the relocated tree blocks in reloc tree with corresponding
1976  * fs tree.
1977  */
1978 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1979                                                struct btrfs_root *root)
1980 {
1981         LIST_HEAD(inode_list);
1982         struct btrfs_key key;
1983         struct btrfs_key next_key;
1984         struct btrfs_trans_handle *trans;
1985         struct btrfs_root *reloc_root;
1986         struct btrfs_root_item *root_item;
1987         struct btrfs_path *path;
1988         struct extent_buffer *leaf;
1989         unsigned long nr;
1990         int level;
1991         int max_level;
1992         int replaced = 0;
1993         int ret;
1994         int err = 0;
1995         u32 min_reserved;
1996
1997         path = btrfs_alloc_path();
1998         if (!path)
1999                 return -ENOMEM;
2000
2001         reloc_root = root->reloc_root;
2002         root_item = &reloc_root->root_item;
2003
2004         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
2005                 level = btrfs_root_level(root_item);
2006                 extent_buffer_get(reloc_root->node);
2007                 path->nodes[level] = reloc_root->node;
2008                 path->slots[level] = 0;
2009         } else {
2010                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
2011
2012                 level = root_item->drop_level;
2013                 BUG_ON(level == 0);
2014                 path->lowest_level = level;
2015                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
2016                 path->lowest_level = 0;
2017                 if (ret < 0) {
2018                         btrfs_free_path(path);
2019                         return ret;
2020                 }
2021
2022                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
2023                                       path->slots[level]);
2024                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
2025
2026                 btrfs_unlock_up_safe(path, 0);
2027         }
2028
2029         min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2030         memset(&next_key, 0, sizeof(next_key));
2031
2032         while (1) {
2033                 trans = btrfs_start_transaction(root, 0);
2034                 BUG_ON(IS_ERR(trans));
2035                 trans->block_rsv = rc->block_rsv;
2036
2037                 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
2038                                             min_reserved, 0);
2039                 if (ret) {
2040                         BUG_ON(ret != -EAGAIN);
2041                         ret = btrfs_commit_transaction(trans, root);
2042                         BUG_ON(ret);
2043                         continue;
2044                 }
2045
2046                 replaced = 0;
2047                 max_level = level;
2048
2049                 ret = walk_down_reloc_tree(reloc_root, path, &level);
2050                 if (ret < 0) {
2051                         err = ret;
2052                         goto out;
2053                 }
2054                 if (ret > 0)
2055                         break;
2056
2057                 if (!find_next_key(path, level, &key) &&
2058                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
2059                         ret = 0;
2060                 } else {
2061                         ret = replace_path(trans, root, reloc_root, path,
2062                                            &next_key, level, max_level);
2063                 }
2064                 if (ret < 0) {
2065                         err = ret;
2066                         goto out;
2067                 }
2068
2069                 if (ret > 0) {
2070                         level = ret;
2071                         btrfs_node_key_to_cpu(path->nodes[level], &key,
2072                                               path->slots[level]);
2073                         replaced = 1;
2074                 }
2075
2076                 ret = walk_up_reloc_tree(reloc_root, path, &level);
2077                 if (ret > 0)
2078                         break;
2079
2080                 BUG_ON(level == 0);
2081                 /*
2082                  * save the merging progress in the drop_progress.
2083                  * this is OK since root refs == 1 in this case.
2084                  */
2085                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
2086                                path->slots[level]);
2087                 root_item->drop_level = level;
2088
2089                 nr = trans->blocks_used;
2090                 btrfs_end_transaction_throttle(trans, root);
2091
2092                 btrfs_btree_balance_dirty(root, nr);
2093
2094                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
2095                         invalidate_extent_cache(root, &key, &next_key);
2096         }
2097
2098         /*
2099          * handle the case only one block in the fs tree need to be
2100          * relocated and the block is tree root.
2101          */
2102         leaf = btrfs_lock_root_node(root);
2103         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf);
2104         btrfs_tree_unlock(leaf);
2105         free_extent_buffer(leaf);
2106         if (ret < 0)
2107                 err = ret;
2108 out:
2109         btrfs_free_path(path);
2110
2111         if (err == 0) {
2112                 memset(&root_item->drop_progress, 0,
2113                        sizeof(root_item->drop_progress));
2114                 root_item->drop_level = 0;
2115                 btrfs_set_root_refs(root_item, 0);
2116                 btrfs_update_reloc_root(trans, root);
2117         }
2118
2119         nr = trans->blocks_used;
2120         btrfs_end_transaction_throttle(trans, root);
2121
2122         btrfs_btree_balance_dirty(root, nr);
2123
2124         if (replaced && rc->stage == UPDATE_DATA_PTRS)
2125                 invalidate_extent_cache(root, &key, &next_key);
2126
2127         return err;
2128 }
2129
2130 static noinline_for_stack
2131 int prepare_to_merge(struct reloc_control *rc, int err)
2132 {
2133         struct btrfs_root *root = rc->extent_root;
2134         struct btrfs_root *reloc_root;
2135         struct btrfs_trans_handle *trans;
2136         LIST_HEAD(reloc_roots);
2137         u64 num_bytes = 0;
2138         int ret;
2139
2140         mutex_lock(&root->fs_info->trans_mutex);
2141         rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2142         rc->merging_rsv_size += rc->nodes_relocated * 2;
2143         mutex_unlock(&root->fs_info->trans_mutex);
2144 again:
2145         if (!err) {
2146                 num_bytes = rc->merging_rsv_size;
2147                 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2148                                           num_bytes);
2149                 if (ret)
2150                         err = ret;
2151         }
2152
2153         trans = btrfs_join_transaction(rc->extent_root, 1);
2154         if (IS_ERR(trans)) {
2155                 if (!err)
2156                         btrfs_block_rsv_release(rc->extent_root,
2157                                                 rc->block_rsv, num_bytes);
2158                 return PTR_ERR(trans);
2159         }
2160
2161         if (!err) {
2162                 if (num_bytes != rc->merging_rsv_size) {
2163                         btrfs_end_transaction(trans, rc->extent_root);
2164                         btrfs_block_rsv_release(rc->extent_root,
2165                                                 rc->block_rsv, num_bytes);
2166                         goto again;
2167                 }
2168         }
2169
2170         rc->merge_reloc_tree = 1;
2171
2172         while (!list_empty(&rc->reloc_roots)) {
2173                 reloc_root = list_entry(rc->reloc_roots.next,
2174                                         struct btrfs_root, root_list);
2175                 list_del_init(&reloc_root->root_list);
2176
2177                 root = read_fs_root(reloc_root->fs_info,
2178                                     reloc_root->root_key.offset);
2179                 BUG_ON(IS_ERR(root));
2180                 BUG_ON(root->reloc_root != reloc_root);
2181
2182                 /*
2183                  * set reference count to 1, so btrfs_recover_relocation
2184                  * knows it should resumes merging
2185                  */
2186                 if (!err)
2187                         btrfs_set_root_refs(&reloc_root->root_item, 1);
2188                 btrfs_update_reloc_root(trans, root);
2189
2190                 list_add(&reloc_root->root_list, &reloc_roots);
2191         }
2192
2193         list_splice(&reloc_roots, &rc->reloc_roots);
2194
2195         if (!err)
2196                 btrfs_commit_transaction(trans, rc->extent_root);
2197         else
2198                 btrfs_end_transaction(trans, rc->extent_root);
2199         return err;
2200 }
2201
2202 static noinline_for_stack
2203 int merge_reloc_roots(struct reloc_control *rc)
2204 {
2205         struct btrfs_root *root;
2206         struct btrfs_root *reloc_root;
2207         LIST_HEAD(reloc_roots);
2208         int found = 0;
2209         int ret;
2210 again:
2211         root = rc->extent_root;
2212         mutex_lock(&root->fs_info->trans_mutex);
2213         list_splice_init(&rc->reloc_roots, &reloc_roots);
2214         mutex_unlock(&root->fs_info->trans_mutex);
2215
2216         while (!list_empty(&reloc_roots)) {
2217                 found = 1;
2218                 reloc_root = list_entry(reloc_roots.next,
2219                                         struct btrfs_root, root_list);
2220
2221                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
2222                         root = read_fs_root(reloc_root->fs_info,
2223                                             reloc_root->root_key.offset);
2224                         BUG_ON(IS_ERR(root));
2225                         BUG_ON(root->reloc_root != reloc_root);
2226
2227                         ret = merge_reloc_root(rc, root);
2228                         BUG_ON(ret);
2229                 } else {
2230                         list_del_init(&reloc_root->root_list);
2231                 }
2232                 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
2233         }
2234
2235         if (found) {
2236                 found = 0;
2237                 goto again;
2238         }
2239         BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
2240         return 0;
2241 }
2242
2243 static void free_block_list(struct rb_root *blocks)
2244 {
2245         struct tree_block *block;
2246         struct rb_node *rb_node;
2247         while ((rb_node = rb_first(blocks))) {
2248                 block = rb_entry(rb_node, struct tree_block, rb_node);
2249                 rb_erase(rb_node, blocks);
2250                 kfree(block);
2251         }
2252 }
2253
2254 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2255                                       struct btrfs_root *reloc_root)
2256 {
2257         struct btrfs_root *root;
2258
2259         if (reloc_root->last_trans == trans->transid)
2260                 return 0;
2261
2262         root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset);
2263         BUG_ON(IS_ERR(root));
2264         BUG_ON(root->reloc_root != reloc_root);
2265
2266         return btrfs_record_root_in_trans(trans, root);
2267 }
2268
2269 static noinline_for_stack
2270 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2271                                      struct reloc_control *rc,
2272                                      struct backref_node *node,
2273                                      struct backref_edge *edges[], int *nr)
2274 {
2275         struct backref_node *next;
2276         struct btrfs_root *root;
2277         int index = 0;
2278
2279         next = node;
2280         while (1) {
2281                 cond_resched();
2282                 next = walk_up_backref(next, edges, &index);
2283                 root = next->root;
2284                 BUG_ON(!root);
2285                 BUG_ON(!root->ref_cows);
2286
2287                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2288                         record_reloc_root_in_trans(trans, root);
2289                         break;
2290                 }
2291
2292                 btrfs_record_root_in_trans(trans, root);
2293                 root = root->reloc_root;
2294
2295                 if (next->new_bytenr != root->node->start) {
2296                         BUG_ON(next->new_bytenr);
2297                         BUG_ON(!list_empty(&next->list));
2298                         next->new_bytenr = root->node->start;
2299                         next->root = root;
2300                         list_add_tail(&next->list,
2301                                       &rc->backref_cache.changed);
2302                         __mark_block_processed(rc, next);
2303                         break;
2304                 }
2305
2306                 WARN_ON(1);
2307                 root = NULL;
2308                 next = walk_down_backref(edges, &index);
2309                 if (!next || next->level <= node->level)
2310                         break;
2311         }
2312         if (!root)
2313                 return NULL;
2314
2315         *nr = index;
2316         next = node;
2317         /* setup backref node path for btrfs_reloc_cow_block */
2318         while (1) {
2319                 rc->backref_cache.path[next->level] = next;
2320                 if (--index < 0)
2321                         break;
2322                 next = edges[index]->node[UPPER];
2323         }
2324         return root;
2325 }
2326
2327 /*
2328  * select a tree root for relocation. return NULL if the block
2329  * is reference counted. we should use do_relocation() in this
2330  * case. return a tree root pointer if the block isn't reference
2331  * counted. return -ENOENT if the block is root of reloc tree.
2332  */
2333 static noinline_for_stack
2334 struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
2335                                    struct backref_node *node)
2336 {
2337         struct backref_node *next;
2338         struct btrfs_root *root;
2339         struct btrfs_root *fs_root = NULL;
2340         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2341         int index = 0;
2342
2343         next = node;
2344         while (1) {
2345                 cond_resched();
2346                 next = walk_up_backref(next, edges, &index);
2347                 root = next->root;
2348                 BUG_ON(!root);
2349
2350                 /* no other choice for non-refernce counted tree */
2351                 if (!root->ref_cows)
2352                         return root;
2353
2354                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2355                         fs_root = root;
2356
2357                 if (next != node)
2358                         return NULL;
2359
2360                 next = walk_down_backref(edges, &index);
2361                 if (!next || next->level <= node->level)
2362                         break;
2363         }
2364
2365         if (!fs_root)
2366                 return ERR_PTR(-ENOENT);
2367         return fs_root;
2368 }
2369
2370 static noinline_for_stack
2371 u64 calcu_metadata_size(struct reloc_control *rc,
2372                         struct backref_node *node, int reserve)
2373 {
2374         struct backref_node *next = node;
2375         struct backref_edge *edge;
2376         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2377         u64 num_bytes = 0;
2378         int index = 0;
2379
2380         BUG_ON(reserve && node->processed);
2381
2382         while (next) {
2383                 cond_resched();
2384                 while (1) {
2385                         if (next->processed && (reserve || next != node))
2386                                 break;
2387
2388                         num_bytes += btrfs_level_size(rc->extent_root,
2389                                                       next->level);
2390
2391                         if (list_empty(&next->upper))
2392                                 break;
2393
2394                         edge = list_entry(next->upper.next,
2395                                           struct backref_edge, list[LOWER]);
2396                         edges[index++] = edge;
2397                         next = edge->node[UPPER];
2398                 }
2399                 next = walk_down_backref(edges, &index);
2400         }
2401         return num_bytes;
2402 }
2403
2404 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2405                                   struct reloc_control *rc,
2406                                   struct backref_node *node)
2407 {
2408         struct btrfs_root *root = rc->extent_root;
2409         u64 num_bytes;
2410         int ret;
2411
2412         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2413
2414         trans->block_rsv = rc->block_rsv;
2415         ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes);
2416         if (ret) {
2417                 if (ret == -EAGAIN)
2418                         rc->commit_transaction = 1;
2419                 return ret;
2420         }
2421
2422         return 0;
2423 }
2424
2425 static void release_metadata_space(struct reloc_control *rc,
2426                                    struct backref_node *node)
2427 {
2428         u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2429         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
2430 }
2431
2432 /*
2433  * relocate a block tree, and then update pointers in upper level
2434  * blocks that reference the block to point to the new location.
2435  *
2436  * if called by link_to_upper, the block has already been relocated.
2437  * in that case this function just updates pointers.
2438  */
2439 static int do_relocation(struct btrfs_trans_handle *trans,
2440                          struct reloc_control *rc,
2441                          struct backref_node *node,
2442                          struct btrfs_key *key,
2443                          struct btrfs_path *path, int lowest)
2444 {
2445         struct backref_node *upper;
2446         struct backref_edge *edge;
2447         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2448         struct btrfs_root *root;
2449         struct extent_buffer *eb;
2450         u32 blocksize;
2451         u64 bytenr;
2452         u64 generation;
2453         int nr;
2454         int slot;
2455         int ret;
2456         int err = 0;
2457
2458         BUG_ON(lowest && node->eb);
2459
2460         path->lowest_level = node->level + 1;
2461         rc->backref_cache.path[node->level] = node;
2462         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2463                 cond_resched();
2464
2465                 upper = edge->node[UPPER];
2466                 root = select_reloc_root(trans, rc, upper, edges, &nr);
2467                 BUG_ON(!root);
2468
2469                 if (upper->eb && !upper->locked) {
2470                         if (!lowest) {
2471                                 ret = btrfs_bin_search(upper->eb, key,
2472                                                        upper->level, &slot);
2473                                 BUG_ON(ret);
2474                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2475                                 if (node->eb->start == bytenr)
2476                                         goto next;
2477                         }
2478                         drop_node_buffer(upper);
2479                 }
2480
2481                 if (!upper->eb) {
2482                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2483                         if (ret < 0) {
2484                                 err = ret;
2485                                 break;
2486                         }
2487                         BUG_ON(ret > 0);
2488
2489                         if (!upper->eb) {
2490                                 upper->eb = path->nodes[upper->level];
2491                                 path->nodes[upper->level] = NULL;
2492                         } else {
2493                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2494                         }
2495
2496                         upper->locked = 1;
2497                         path->locks[upper->level] = 0;
2498
2499                         slot = path->slots[upper->level];
2500                         btrfs_release_path(NULL, path);
2501                 } else {
2502                         ret = btrfs_bin_search(upper->eb, key, upper->level,
2503                                                &slot);
2504                         BUG_ON(ret);
2505                 }
2506
2507                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2508                 if (lowest) {
2509                         BUG_ON(bytenr != node->bytenr);
2510                 } else {
2511                         if (node->eb->start == bytenr)
2512                                 goto next;
2513                 }
2514
2515                 blocksize = btrfs_level_size(root, node->level);
2516                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2517                 eb = read_tree_block(root, bytenr, blocksize, generation);
2518                 if (!eb) {
2519                         err = -EIO;
2520                         goto next;
2521                 }
2522                 btrfs_tree_lock(eb);
2523                 btrfs_set_lock_blocking(eb);
2524
2525                 if (!node->eb) {
2526                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2527                                               slot, &eb);
2528                         btrfs_tree_unlock(eb);
2529                         free_extent_buffer(eb);
2530                         if (ret < 0) {
2531                                 err = ret;
2532                                 goto next;
2533                         }
2534                         BUG_ON(node->eb != eb);
2535                 } else {
2536                         btrfs_set_node_blockptr(upper->eb, slot,
2537                                                 node->eb->start);
2538                         btrfs_set_node_ptr_generation(upper->eb, slot,
2539                                                       trans->transid);
2540                         btrfs_mark_buffer_dirty(upper->eb);
2541
2542                         ret = btrfs_inc_extent_ref(trans, root,
2543                                                 node->eb->start, blocksize,
2544                                                 upper->eb->start,
2545                                                 btrfs_header_owner(upper->eb),
2546                                                 node->level, 0);
2547                         BUG_ON(ret);
2548
2549                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2550                         BUG_ON(ret);
2551                 }
2552 next:
2553                 if (!upper->pending)
2554                         drop_node_buffer(upper);
2555                 else
2556                         unlock_node_buffer(upper);
2557                 if (err)
2558                         break;
2559         }
2560
2561         if (!err && node->pending) {
2562                 drop_node_buffer(node);
2563                 list_move_tail(&node->list, &rc->backref_cache.changed);
2564                 node->pending = 0;
2565         }
2566
2567         path->lowest_level = 0;
2568         BUG_ON(err == -ENOSPC);
2569         return err;
2570 }
2571
2572 static int link_to_upper(struct btrfs_trans_handle *trans,
2573                          struct reloc_control *rc,
2574                          struct backref_node *node,
2575                          struct btrfs_path *path)
2576 {
2577         struct btrfs_key key;
2578
2579         btrfs_node_key_to_cpu(node->eb, &key, 0);
2580         return do_relocation(trans, rc, node, &key, path, 0);
2581 }
2582
2583 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2584                                 struct reloc_control *rc,
2585                                 struct btrfs_path *path, int err)
2586 {
2587         LIST_HEAD(list);
2588         struct backref_cache *cache = &rc->backref_cache;
2589         struct backref_node *node;
2590         int level;
2591         int ret;
2592
2593         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2594                 while (!list_empty(&cache->pending[level])) {
2595                         node = list_entry(cache->pending[level].next,
2596                                           struct backref_node, list);
2597                         list_move_tail(&node->list, &list);
2598                         BUG_ON(!node->pending);
2599
2600                         if (!err) {
2601                                 ret = link_to_upper(trans, rc, node, path);
2602                                 if (ret < 0)
2603                                         err = ret;
2604                         }
2605                 }
2606                 list_splice_init(&list, &cache->pending[level]);
2607         }
2608         return err;
2609 }
2610
2611 static void mark_block_processed(struct reloc_control *rc,
2612                                  u64 bytenr, u32 blocksize)
2613 {
2614         set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2615                         EXTENT_DIRTY, GFP_NOFS);
2616 }
2617
2618 static void __mark_block_processed(struct reloc_control *rc,
2619                                    struct backref_node *node)
2620 {
2621         u32 blocksize;
2622         if (node->level == 0 ||
2623             in_block_group(node->bytenr, rc->block_group)) {
2624                 blocksize = btrfs_level_size(rc->extent_root, node->level);
2625                 mark_block_processed(rc, node->bytenr, blocksize);
2626         }
2627         node->processed = 1;
2628 }
2629
2630 /*
2631  * mark a block and all blocks directly/indirectly reference the block
2632  * as processed.
2633  */
2634 static void update_processed_blocks(struct reloc_control *rc,
2635                                     struct backref_node *node)
2636 {
2637         struct backref_node *next = node;
2638         struct backref_edge *edge;
2639         struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2640         int index = 0;
2641
2642         while (next) {
2643                 cond_resched();
2644                 while (1) {
2645                         if (next->processed)
2646                                 break;
2647
2648                         __mark_block_processed(rc, next);
2649
2650                         if (list_empty(&next->upper))
2651                                 break;
2652
2653                         edge = list_entry(next->upper.next,
2654                                           struct backref_edge, list[LOWER]);
2655                         edges[index++] = edge;
2656                         next = edge->node[UPPER];
2657                 }
2658                 next = walk_down_backref(edges, &index);
2659         }
2660 }
2661
2662 static int tree_block_processed(u64 bytenr, u32 blocksize,
2663                                 struct reloc_control *rc)
2664 {
2665         if (test_range_bit(&rc->processed_blocks, bytenr,
2666                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2667                 return 1;
2668         return 0;
2669 }
2670
2671 static int get_tree_block_key(struct reloc_control *rc,
2672                               struct tree_block *block)
2673 {
2674         struct extent_buffer *eb;
2675
2676         BUG_ON(block->key_ready);
2677         eb = read_tree_block(rc->extent_root, block->bytenr,
2678                              block->key.objectid, block->key.offset);
2679         BUG_ON(!eb);
2680         WARN_ON(btrfs_header_level(eb) != block->level);
2681         if (block->level == 0)
2682                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2683         else
2684                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2685         free_extent_buffer(eb);
2686         block->key_ready = 1;
2687         return 0;
2688 }
2689
2690 static int reada_tree_block(struct reloc_control *rc,
2691                             struct tree_block *block)
2692 {
2693         BUG_ON(block->key_ready);
2694         readahead_tree_block(rc->extent_root, block->bytenr,
2695                              block->key.objectid, block->key.offset);
2696         return 0;
2697 }
2698
2699 /*
2700  * helper function to relocate a tree block
2701  */
2702 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2703                                 struct reloc_control *rc,
2704                                 struct backref_node *node,
2705                                 struct btrfs_key *key,
2706                                 struct btrfs_path *path)
2707 {
2708         struct btrfs_root *root;
2709         int release = 0;
2710         int ret = 0;
2711
2712         if (!node)
2713                 return 0;
2714
2715         BUG_ON(node->processed);
2716         root = select_one_root(trans, node);
2717         if (root == ERR_PTR(-ENOENT)) {
2718                 update_processed_blocks(rc, node);
2719                 goto out;
2720         }
2721
2722         if (!root || root->ref_cows) {
2723                 ret = reserve_metadata_space(trans, rc, node);
2724                 if (ret)
2725                         goto out;
2726                 release = 1;
2727         }
2728
2729         if (root) {
2730                 if (root->ref_cows) {
2731                         BUG_ON(node->new_bytenr);
2732                         BUG_ON(!list_empty(&node->list));
2733                         btrfs_record_root_in_trans(trans, root);
2734                         root = root->reloc_root;
2735                         node->new_bytenr = root->node->start;
2736                         node->root = root;
2737                         list_add_tail(&node->list, &rc->backref_cache.changed);
2738                 } else {
2739                         path->lowest_level = node->level;
2740                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2741                         btrfs_release_path(root, path);
2742                         if (ret > 0)
2743                                 ret = 0;
2744                 }
2745                 if (!ret)
2746                         update_processed_blocks(rc, node);
2747         } else {
2748                 ret = do_relocation(trans, rc, node, key, path, 1);
2749         }
2750 out:
2751         if (ret || node->level == 0 || node->cowonly) {
2752                 if (release)
2753                         release_metadata_space(rc, node);
2754                 remove_backref_node(&rc->backref_cache, node);
2755         }
2756         return ret;
2757 }
2758
2759 /*
2760  * relocate a list of blocks
2761  */
2762 static noinline_for_stack
2763 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2764                          struct reloc_control *rc, struct rb_root *blocks)
2765 {
2766         struct backref_node *node;
2767         struct btrfs_path *path;
2768         struct tree_block *block;
2769         struct rb_node *rb_node;
2770         int ret;
2771         int err = 0;
2772
2773         path = btrfs_alloc_path();
2774         if (!path)
2775                 return -ENOMEM;
2776
2777         rb_node = rb_first(blocks);
2778         while (rb_node) {
2779                 block = rb_entry(rb_node, struct tree_block, rb_node);
2780                 if (!block->key_ready)
2781                         reada_tree_block(rc, block);
2782                 rb_node = rb_next(rb_node);
2783         }
2784
2785         rb_node = rb_first(blocks);
2786         while (rb_node) {
2787                 block = rb_entry(rb_node, struct tree_block, rb_node);
2788                 if (!block->key_ready)
2789                         get_tree_block_key(rc, block);
2790                 rb_node = rb_next(rb_node);
2791         }
2792
2793         rb_node = rb_first(blocks);
2794         while (rb_node) {
2795                 block = rb_entry(rb_node, struct tree_block, rb_node);
2796
2797                 node = build_backref_tree(rc, &block->key,
2798                                           block->level, block->bytenr);
2799                 if (IS_ERR(node)) {
2800                         err = PTR_ERR(node);
2801                         goto out;
2802                 }
2803
2804                 ret = relocate_tree_block(trans, rc, node, &block->key,
2805                                           path);
2806                 if (ret < 0) {
2807                         if (ret != -EAGAIN || rb_node == rb_first(blocks))
2808                                 err = ret;
2809                         goto out;
2810                 }
2811                 rb_node = rb_next(rb_node);
2812         }
2813 out:
2814         free_block_list(blocks);
2815         err = finish_pending_nodes(trans, rc, path, err);
2816
2817         btrfs_free_path(path);
2818         return err;
2819 }
2820
2821 static noinline_for_stack
2822 int prealloc_file_extent_cluster(struct inode *inode,
2823                                  struct file_extent_cluster *cluster)
2824 {
2825         u64 alloc_hint = 0;
2826         u64 start;
2827         u64 end;
2828         u64 offset = BTRFS_I(inode)->index_cnt;
2829         u64 num_bytes;
2830         int nr = 0;
2831         int ret = 0;
2832
2833         BUG_ON(cluster->start != cluster->boundary[0]);
2834         mutex_lock(&inode->i_mutex);
2835
2836         ret = btrfs_check_data_free_space(inode, cluster->end +
2837                                           1 - cluster->start);
2838         if (ret)
2839                 goto out;
2840
2841         while (nr < cluster->nr) {
2842                 start = cluster->boundary[nr] - offset;
2843                 if (nr + 1 < cluster->nr)
2844                         end = cluster->boundary[nr + 1] - 1 - offset;
2845                 else
2846                         end = cluster->end - offset;
2847
2848                 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2849                 num_bytes = end + 1 - start;
2850                 ret = btrfs_prealloc_file_range(inode, 0, start,
2851                                                 num_bytes, num_bytes,
2852                                                 end + 1, &alloc_hint);
2853                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2854                 if (ret)
2855                         break;
2856                 nr++;
2857         }
2858         btrfs_free_reserved_data_space(inode, cluster->end +
2859                                        1 - cluster->start);
2860 out:
2861         mutex_unlock(&inode->i_mutex);
2862         return ret;
2863 }
2864
2865 static noinline_for_stack
2866 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2867                          u64 block_start)
2868 {
2869         struct btrfs_root *root = BTRFS_I(inode)->root;
2870         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2871         struct extent_map *em;
2872         int ret = 0;
2873
2874         em = alloc_extent_map(GFP_NOFS);
2875         if (!em)
2876                 return -ENOMEM;
2877
2878         em->start = start;
2879         em->len = end + 1 - start;
2880         em->block_len = em->len;
2881         em->block_start = block_start;
2882         em->bdev = root->fs_info->fs_devices->latest_bdev;
2883         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2884
2885         lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2886         while (1) {
2887                 write_lock(&em_tree->lock);
2888                 ret = add_extent_mapping(em_tree, em);
2889                 write_unlock(&em_tree->lock);
2890                 if (ret != -EEXIST) {
2891                         free_extent_map(em);
2892                         break;
2893                 }
2894                 btrfs_drop_extent_cache(inode, start, end, 0);
2895         }
2896         unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2897         return ret;
2898 }
2899
2900 static int relocate_file_extent_cluster(struct inode *inode,
2901                                         struct file_extent_cluster *cluster)
2902 {
2903         u64 page_start;
2904         u64 page_end;
2905         u64 offset = BTRFS_I(inode)->index_cnt;
2906         unsigned long index;
2907         unsigned long last_index;
2908         struct page *page;
2909         struct file_ra_state *ra;
2910         int nr = 0;
2911         int ret = 0;
2912
2913         if (!cluster->nr)
2914                 return 0;
2915
2916         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2917         if (!ra)
2918                 return -ENOMEM;
2919
2920         ret = prealloc_file_extent_cluster(inode, cluster);
2921         if (ret)
2922                 goto out;
2923
2924         file_ra_state_init(ra, inode->i_mapping);
2925
2926         ret = setup_extent_mapping(inode, cluster->start - offset,
2927                                    cluster->end - offset, cluster->start);
2928         if (ret)
2929                 goto out;
2930
2931         index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2932         last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2933         while (index <= last_index) {
2934                 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2935                 if (ret)
2936                         goto out;
2937
2938                 page = find_lock_page(inode->i_mapping, index);
2939                 if (!page) {
2940                         page_cache_sync_readahead(inode->i_mapping,
2941                                                   ra, NULL, index,
2942                                                   last_index + 1 - index);
2943                         page = grab_cache_page(inode->i_mapping, index);
2944                         if (!page) {
2945                                 btrfs_delalloc_release_metadata(inode,
2946                                                         PAGE_CACHE_SIZE);
2947                                 ret = -ENOMEM;
2948                                 goto out;
2949                         }
2950                 }
2951
2952                 if (PageReadahead(page)) {
2953                         page_cache_async_readahead(inode->i_mapping,
2954                                                    ra, NULL, page, index,
2955                                                    last_index + 1 - index);
2956                 }
2957
2958                 if (!PageUptodate(page)) {
2959                         btrfs_readpage(NULL, page);
2960                         lock_page(page);
2961                         if (!PageUptodate(page)) {
2962                                 unlock_page(page);
2963                                 page_cache_release(page);
2964                                 btrfs_delalloc_release_metadata(inode,
2965                                                         PAGE_CACHE_SIZE);
2966                                 ret = -EIO;
2967                                 goto out;
2968                         }
2969                 }
2970
2971                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2972                 page_end = page_start + PAGE_CACHE_SIZE - 1;
2973
2974                 lock_extent(&BTRFS_I(inode)->io_tree,
2975                             page_start, page_end, GFP_NOFS);
2976
2977                 set_page_extent_mapped(page);
2978
2979                 if (nr < cluster->nr &&
2980                     page_start + offset == cluster->boundary[nr]) {
2981                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2982                                         page_start, page_end,
2983                                         EXTENT_BOUNDARY, GFP_NOFS);
2984                         nr++;
2985                 }
2986
2987                 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2988                 set_page_dirty(page);
2989
2990                 unlock_extent(&BTRFS_I(inode)->io_tree,
2991                               page_start, page_end, GFP_NOFS);
2992                 unlock_page(page);
2993                 page_cache_release(page);
2994
2995                 index++;
2996                 balance_dirty_pages_ratelimited(inode->i_mapping);
2997                 btrfs_throttle(BTRFS_I(inode)->root);
2998         }
2999         WARN_ON(nr != cluster->nr);
3000 out:
3001         kfree(ra);
3002         return ret;
3003 }
3004
3005 static noinline_for_stack
3006 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
3007                          struct file_extent_cluster *cluster)
3008 {
3009         int ret;
3010
3011         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
3012                 ret = relocate_file_extent_cluster(inode, cluster);
3013                 if (ret)
3014                         return ret;
3015                 cluster->nr = 0;
3016         }
3017
3018         if (!cluster->nr)
3019                 cluster->start = extent_key->objectid;
3020         else
3021                 BUG_ON(cluster->nr >= MAX_EXTENTS);
3022         cluster->end = extent_key->objectid + extent_key->offset - 1;
3023         cluster->boundary[cluster->nr] = extent_key->objectid;
3024         cluster->nr++;
3025
3026         if (cluster->nr >= MAX_EXTENTS) {
3027                 ret = relocate_file_extent_cluster(inode, cluster);
3028                 if (ret)
3029                         return ret;
3030                 cluster->nr = 0;
3031         }
3032         return 0;
3033 }
3034
3035 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3036 static int get_ref_objectid_v0(struct reloc_control *rc,
3037                                struct btrfs_path *path,
3038                                struct btrfs_key *extent_key,
3039                                u64 *ref_objectid, int *path_change)
3040 {
3041         struct btrfs_key key;
3042         struct extent_buffer *leaf;
3043         struct btrfs_extent_ref_v0 *ref0;
3044         int ret;
3045         int slot;
3046
3047         leaf = path->nodes[0];
3048         slot = path->slots[0];
3049         while (1) {
3050                 if (slot >= btrfs_header_nritems(leaf)) {
3051                         ret = btrfs_next_leaf(rc->extent_root, path);
3052                         if (ret < 0)
3053                                 return ret;
3054                         BUG_ON(ret > 0);
3055                         leaf = path->nodes[0];
3056                         slot = path->slots[0];
3057                         if (path_change)
3058                                 *path_change = 1;
3059                 }
3060                 btrfs_item_key_to_cpu(leaf, &key, slot);
3061                 if (key.objectid != extent_key->objectid)
3062                         return -ENOENT;
3063
3064                 if (key.type != BTRFS_EXTENT_REF_V0_KEY) {
3065                         slot++;
3066                         continue;
3067                 }
3068                 ref0 = btrfs_item_ptr(leaf, slot,
3069                                 struct btrfs_extent_ref_v0);
3070                 *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0);
3071                 break;
3072         }
3073         return 0;
3074 }
3075 #endif
3076
3077 /*
3078  * helper to add a tree block to the list.
3079  * the major work is getting the generation and level of the block
3080  */
3081 static int add_tree_block(struct reloc_control *rc,
3082                           struct btrfs_key *extent_key,
3083                           struct btrfs_path *path,
3084                           struct rb_root *blocks)
3085 {
3086         struct extent_buffer *eb;
3087         struct btrfs_extent_item *ei;
3088         struct btrfs_tree_block_info *bi;
3089         struct tree_block *block;
3090         struct rb_node *rb_node;
3091         u32 item_size;
3092         int level = -1;
3093         int generation;
3094
3095         eb =  path->nodes[0];
3096         item_size = btrfs_item_size_nr(eb, path->slots[0]);
3097
3098         if (item_size >= sizeof(*ei) + sizeof(*bi)) {
3099                 ei = btrfs_item_ptr(eb, path->slots[0],
3100                                 struct btrfs_extent_item);
3101                 bi = (struct btrfs_tree_block_info *)(ei + 1);
3102                 generation = btrfs_extent_generation(eb, ei);
3103                 level = btrfs_tree_block_level(eb, bi);
3104         } else {
3105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3106                 u64 ref_owner;
3107                 int ret;
3108
3109                 BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3110                 ret = get_ref_objectid_v0(rc, path, extent_key,
3111                                           &ref_owner, NULL);
3112                 if (ret < 0)
3113                         return ret;
3114                 BUG_ON(ref_owner >= BTRFS_MAX_LEVEL);
3115                 level = (int)ref_owner;
3116                 /* FIXME: get real generation */
3117                 generation = 0;
3118 #else
3119                 BUG();
3120 #endif
3121         }
3122
3123         btrfs_release_path(rc->extent_root, path);
3124
3125         BUG_ON(level == -1);
3126
3127         block = kmalloc(sizeof(*block), GFP_NOFS);
3128         if (!block)
3129                 return -ENOMEM;
3130
3131         block->bytenr = extent_key->objectid;
3132         block->key.objectid = extent_key->offset;
3133         block->key.offset = generation;
3134         block->level = level;
3135         block->key_ready = 0;
3136
3137         rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
3138         BUG_ON(rb_node);
3139
3140         return 0;
3141 }
3142
3143 /*
3144  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
3145  */
3146 static int __add_tree_block(struct reloc_control *rc,
3147                             u64 bytenr, u32 blocksize,
3148                             struct rb_root *blocks)
3149 {
3150         struct btrfs_path *path;
3151         struct btrfs_key key;
3152         int ret;
3153
3154         if (tree_block_processed(bytenr, blocksize, rc))
3155                 return 0;
3156
3157         if (tree_search(blocks, bytenr))
3158                 return 0;
3159
3160         path = btrfs_alloc_path();
3161         if (!path)
3162                 return -ENOMEM;
3163
3164         key.objectid = bytenr;
3165         key.type = BTRFS_EXTENT_ITEM_KEY;
3166         key.offset = blocksize;
3167
3168         path->search_commit_root = 1;
3169         path->skip_locking = 1;
3170         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
3171         if (ret < 0)
3172                 goto out;
3173         BUG_ON(ret);
3174
3175         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3176         ret = add_tree_block(rc, &key, path, blocks);
3177 out:
3178         btrfs_free_path(path);
3179         return ret;
3180 }
3181
3182 /*
3183  * helper to check if the block use full backrefs for pointers in it
3184  */
3185 static int block_use_full_backref(struct reloc_control *rc,
3186                                   struct extent_buffer *eb)
3187 {
3188         u64 flags;
3189         int ret;
3190
3191         if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) ||
3192             btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
3193                 return 1;
3194
3195         ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
3196                                        eb->start, eb->len, NULL, &flags);
3197         BUG_ON(ret);
3198
3199         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
3200                 ret = 1;
3201         else
3202                 ret = 0;
3203         return ret;
3204 }
3205
3206 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
3207                                     struct inode *inode, u64 ino)
3208 {
3209         struct btrfs_key key;
3210         struct btrfs_path *path;
3211         struct btrfs_root *root = fs_info->tree_root;
3212         struct btrfs_trans_handle *trans;
3213         unsigned long nr;
3214         int ret = 0;
3215
3216         if (inode)
3217                 goto truncate;
3218
3219         key.objectid = ino;
3220         key.type = BTRFS_INODE_ITEM_KEY;
3221         key.offset = 0;
3222
3223         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
3224         if (!inode || IS_ERR(inode) || is_bad_inode(inode)) {
3225                 if (inode && !IS_ERR(inode))
3226                         iput(inode);
3227                 return -ENOENT;
3228         }
3229
3230 truncate:
3231         path = btrfs_alloc_path();
3232         if (!path) {
3233                 ret = -ENOMEM;
3234                 goto out;
3235         }
3236
3237         trans = btrfs_join_transaction(root, 0);
3238         if (IS_ERR(trans)) {
3239                 btrfs_free_path(path);
3240                 ret = PTR_ERR(trans);
3241                 goto out;
3242         }
3243
3244         ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
3245
3246         btrfs_free_path(path);
3247         nr = trans->blocks_used;
3248         btrfs_end_transaction(trans, root);
3249         btrfs_btree_balance_dirty(root, nr);
3250 out:
3251         iput(inode);
3252         return ret;
3253 }
3254
3255 /*
3256  * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY
3257  * this function scans fs tree to find blocks reference the data extent
3258  */
3259 static int find_data_references(struct reloc_control *rc,
3260                                 struct btrfs_key *extent_key,
3261                                 struct extent_buffer *leaf,
3262                                 struct btrfs_extent_data_ref *ref,
3263                                 struct rb_root *blocks)
3264 {
3265         struct btrfs_path *path;
3266         struct tree_block *block;
3267         struct btrfs_root *root;
3268         struct btrfs_file_extent_item *fi;
3269         struct rb_node *rb_node;
3270         struct btrfs_key key;
3271         u64 ref_root;
3272         u64 ref_objectid;
3273         u64 ref_offset;
3274         u32 ref_count;
3275         u32 nritems;
3276         int err = 0;
3277         int added = 0;
3278         int counted;
3279         int ret;
3280
3281         ref_root = btrfs_extent_data_ref_root(leaf, ref);
3282         ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref);
3283         ref_offset = btrfs_extent_data_ref_offset(leaf, ref);
3284         ref_count = btrfs_extent_data_ref_count(leaf, ref);
3285
3286         /*
3287          * This is an extent belonging to the free space cache, lets just delete
3288          * it and redo the search.
3289          */
3290         if (ref_root == BTRFS_ROOT_TREE_OBJECTID) {
3291                 ret = delete_block_group_cache(rc->extent_root->fs_info,
3292                                                NULL, ref_objectid);
3293                 if (ret != -ENOENT)
3294                         return ret;
3295                 ret = 0;
3296         }
3297
3298         path = btrfs_alloc_path();
3299         if (!path)
3300                 return -ENOMEM;
3301
3302         root = read_fs_root(rc->extent_root->fs_info, ref_root);
3303         if (IS_ERR(root)) {
3304                 err = PTR_ERR(root);
3305                 goto out;
3306         }
3307
3308         key.objectid = ref_objectid;
3309         key.offset = ref_offset;
3310         key.type = BTRFS_EXTENT_DATA_KEY;
3311
3312         path->search_commit_root = 1;
3313         path->skip_locking = 1;
3314         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3315         if (ret < 0) {
3316                 err = ret;
3317                 goto out;
3318         }
3319
3320         leaf = path->nodes[0];
3321         nritems = btrfs_header_nritems(leaf);
3322         /*
3323          * the references in tree blocks that use full backrefs
3324          * are not counted in
3325          */
3326         if (block_use_full_backref(rc, leaf))
3327                 counted = 0;
3328         else
3329                 counted = 1;
3330         rb_node = tree_search(blocks, leaf->start);
3331         if (rb_node) {
3332                 if (counted)
3333                         added = 1;
3334                 else
3335                         path->slots[0] = nritems;
3336         }
3337
3338         while (ref_count > 0) {
3339                 while (path->slots[0] >= nritems) {
3340                         ret = btrfs_next_leaf(root, path);
3341                         if (ret < 0) {
3342                                 err = ret;
3343                                 goto out;
3344                         }
3345                         if (ret > 0) {
3346                                 WARN_ON(1);
3347                                 goto out;
3348                         }
3349
3350                         leaf = path->nodes[0];
3351                         nritems = btrfs_header_nritems(leaf);
3352                         added = 0;
3353
3354                         if (block_use_full_backref(rc, leaf))
3355                                 counted = 0;
3356                         else
3357                                 counted = 1;
3358                         rb_node = tree_search(blocks, leaf->start);
3359                         if (rb_node) {
3360                                 if (counted)
3361                                         added = 1;
3362                                 else
3363                                         path->slots[0] = nritems;
3364                         }
3365                 }
3366
3367                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3368                 if (key.objectid != ref_objectid ||
3369                     key.type != BTRFS_EXTENT_DATA_KEY) {
3370                         WARN_ON(1);
3371                         break;
3372                 }
3373
3374                 fi = btrfs_item_ptr(leaf, path->slots[0],
3375                                     struct btrfs_file_extent_item);
3376
3377                 if (btrfs_file_extent_type(leaf, fi) ==
3378                     BTRFS_FILE_EXTENT_INLINE)
3379                         goto next;
3380
3381                 if (btrfs_file_extent_disk_bytenr(leaf, fi) !=
3382                     extent_key->objectid)
3383                         goto next;
3384
3385                 key.offset -= btrfs_file_extent_offset(leaf, fi);
3386                 if (key.offset != ref_offset)
3387                         goto next;
3388
3389                 if (counted)
3390                         ref_count--;
3391                 if (added)
3392                         goto next;
3393
3394                 if (!tree_block_processed(leaf->start, leaf->len, rc)) {
3395                         block = kmalloc(sizeof(*block), GFP_NOFS);
3396                         if (!block) {
3397                                 err = -ENOMEM;
3398                                 break;
3399                         }
3400                         block->bytenr = leaf->start;
3401                         btrfs_item_key_to_cpu(leaf, &block->key, 0);
3402                         block->level = 0;
3403                         block->key_ready = 1;
3404                         rb_node = tree_insert(blocks, block->bytenr,
3405                                               &block->rb_node);
3406                         BUG_ON(rb_node);
3407                 }
3408                 if (counted)
3409                         added = 1;
3410                 else
3411                         path->slots[0] = nritems;
3412 next:
3413                 path->slots[0]++;
3414
3415         }
3416 out:
3417         btrfs_free_path(path);
3418         return err;
3419 }
3420
3421 /*
3422  * hepler to find all tree blocks that reference a given data extent
3423  */
3424 static noinline_for_stack
3425 int add_data_references(struct reloc_control *rc,
3426                         struct btrfs_key *extent_key,
3427                         struct btrfs_path *path,
3428                         struct rb_root *blocks)
3429 {
3430         struct btrfs_key key;
3431         struct extent_buffer *eb;
3432         struct btrfs_extent_data_ref *dref;
3433         struct btrfs_extent_inline_ref *iref;
3434         unsigned long ptr;
3435         unsigned long end;
3436         u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3437         int ret;
3438         int err = 0;
3439
3440         eb = path->nodes[0];
3441         ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3442         end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
3443 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3444         if (ptr + sizeof(struct btrfs_extent_item_v0) == end)
3445                 ptr = end;
3446         else
3447 #endif
3448                 ptr += sizeof(struct btrfs_extent_item);
3449
3450         while (ptr < end) {
3451                 iref = (struct btrfs_extent_inline_ref *)ptr;
3452                 key.type = btrfs_extent_inline_ref_type(eb, iref);
3453                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3454                         key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3455                         ret = __add_tree_block(rc, key.offset, blocksize,
3456                                                blocks);
3457                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3458                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
3459                         ret = find_data_references(rc, extent_key,
3460                                                    eb, dref, blocks);
3461                 } else {
3462                         BUG();
3463                 }
3464                 ptr += btrfs_extent_inline_ref_size(key.type);
3465         }
3466         WARN_ON(ptr > end);
3467
3468         while (1) {
3469                 cond_resched();
3470                 eb = path->nodes[0];
3471                 if (path->slots[0] >= btrfs_header_nritems(eb)) {
3472                         ret = btrfs_next_leaf(rc->extent_root, path);
3473                         if (ret < 0) {
3474                                 err = ret;
3475                                 break;
3476                         }
3477                         if (ret > 0)
3478                                 break;
3479                         eb = path->nodes[0];
3480                 }
3481
3482                 btrfs_item_key_to_cpu(eb, &key, path->slots[0]);
3483                 if (key.objectid != extent_key->objectid)
3484                         break;
3485
3486 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3487                 if (key.type == BTRFS_SHARED_DATA_REF_KEY ||
3488                     key.type == BTRFS_EXTENT_REF_V0_KEY) {
3489 #else
3490                 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
3491                 if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3492 #endif
3493                         ret = __add_tree_block(rc, key.offset, blocksize,
3494                                                blocks);
3495                 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3496                         dref = btrfs_item_ptr(eb, path->slots[0],
3497                                               struct btrfs_extent_data_ref);
3498                         ret = find_data_references(rc, extent_key,
3499                                                    eb, dref, blocks);
3500                 } else {
3501                         ret = 0;
3502                 }
3503                 if (ret) {
3504                         err = ret;
3505                         break;
3506                 }
3507                 path->slots[0]++;
3508         }
3509         btrfs_release_path(rc->extent_root, path);
3510         if (err)
3511                 free_block_list(blocks);
3512         return err;
3513 }
3514
3515 /*
3516  * hepler to find next unprocessed extent
3517  */
3518 static noinline_for_stack
3519 int find_next_extent(struct btrfs_trans_handle *trans,
3520                      struct reloc_control *rc, struct btrfs_path *path,
3521                      struct btrfs_key *extent_key)
3522 {
3523         struct btrfs_key key;
3524         struct extent_buffer *leaf;
3525         u64 start, end, last;
3526         int ret;
3527
3528         last = rc->block_group->key.objectid + rc->block_group->key.offset;
3529         while (1) {
3530                 cond_resched();
3531                 if (rc->search_start >= last) {
3532                         ret = 1;
3533                         break;
3534                 }
3535
3536                 key.objectid = rc->search_start;
3537                 key.type = BTRFS_EXTENT_ITEM_KEY;
3538                 key.offset = 0;
3539
3540                 path->search_commit_root = 1;
3541                 path->skip_locking = 1;
3542                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3543                                         0, 0);
3544                 if (ret < 0)
3545                         break;
3546 next:
3547                 leaf = path->nodes[0];
3548                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3549                         ret = btrfs_next_leaf(rc->extent_root, path);
3550                         if (ret != 0)
3551                                 break;
3552                         leaf = path->nodes[0];
3553                 }
3554
3555                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3556                 if (key.objectid >= last) {
3557                         ret = 1;
3558                         break;
3559                 }
3560
3561                 if (key.type != BTRFS_EXTENT_ITEM_KEY ||
3562                     key.objectid + key.offset <= rc->search_start) {
3563                         path->slots[0]++;
3564                         goto next;
3565                 }
3566
3567                 ret = find_first_extent_bit(&rc->processed_blocks,
3568                                             key.objectid, &start, &end,
3569                                             EXTENT_DIRTY);
3570
3571                 if (ret == 0 && start <= key.objectid) {
3572                         btrfs_release_path(rc->extent_root, path);
3573                         rc->search_start = end + 1;
3574                 } else {
3575                         rc->search_start = key.objectid + key.offset;
3576                         memcpy(extent_key, &key, sizeof(key));
3577                         return 0;
3578                 }
3579         }
3580         btrfs_release_path(rc->extent_root, path);
3581         return ret;
3582 }
3583
3584 static void set_reloc_control(struct reloc_control *rc)
3585 {
3586         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3587         mutex_lock(&fs_info->trans_mutex);
3588         fs_info->reloc_ctl = rc;
3589         mutex_unlock(&fs_info->trans_mutex);
3590 }
3591
3592 static void unset_reloc_control(struct reloc_control *rc)
3593 {
3594         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3595         mutex_lock(&fs_info->trans_mutex);
3596         fs_info->reloc_ctl = NULL;
3597         mutex_unlock(&fs_info->trans_mutex);
3598 }
3599
3600 static int check_extent_flags(u64 flags)
3601 {
3602         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3603             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3604                 return 1;
3605         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3606             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3607                 return 1;
3608         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3609             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3610                 return 1;
3611         return 0;
3612 }
3613
3614 static noinline_for_stack
3615 int prepare_to_relocate(struct reloc_control *rc)
3616 {
3617         struct btrfs_trans_handle *trans;
3618         int ret;
3619
3620         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3621         if (!rc->block_rsv)
3622                 return -ENOMEM;
3623
3624         /*
3625          * reserve some space for creating reloc trees.
3626          * btrfs_init_reloc_root will use them when there
3627          * is no reservation in transaction handle.
3628          */
3629         ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3630                                   rc->extent_root->nodesize * 256);
3631         if (ret)
3632                 return ret;
3633
3634         rc->block_rsv->refill_used = 1;
3635         btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3636
3637         memset(&rc->cluster, 0, sizeof(rc->cluster));
3638         rc->search_start = rc->block_group->key.objectid;
3639         rc->extents_found = 0;
3640         rc->nodes_relocated = 0;
3641         rc->merging_rsv_size = 0;
3642
3643         rc->create_reloc_tree = 1;
3644         set_reloc_control(rc);
3645
3646         trans = btrfs_join_transaction(rc->extent_root, 1);
3647         BUG_ON(IS_ERR(trans));
3648         btrfs_commit_transaction(trans, rc->extent_root);
3649         return 0;
3650 }
3651
3652 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3653 {
3654         struct rb_root blocks = RB_ROOT;
3655         struct btrfs_key key;
3656         struct btrfs_trans_handle *trans = NULL;
3657         struct btrfs_path *path;
3658         struct btrfs_extent_item *ei;
3659         unsigned long nr;
3660         u64 flags;
3661         u32 item_size;
3662         int ret;
3663         int err = 0;
3664         int progress = 0;
3665
3666         path = btrfs_alloc_path();
3667         if (!path)
3668                 return -ENOMEM;
3669
3670         ret = prepare_to_relocate(rc);
3671         if (ret) {
3672                 err = ret;
3673                 goto out_free;
3674         }
3675
3676         while (1) {
3677                 progress++;
3678                 trans = btrfs_start_transaction(rc->extent_root, 0);
3679                 BUG_ON(IS_ERR(trans));
3680 restart:
3681                 if (update_backref_cache(trans, &rc->backref_cache)) {
3682                         btrfs_end_transaction(trans, rc->extent_root);
3683                         continue;
3684                 }
3685
3686                 ret = find_next_extent(trans, rc, path, &key);
3687                 if (ret < 0)
3688                         err = ret;
3689                 if (ret != 0)
3690                         break;
3691
3692                 rc->extents_found++;
3693
3694                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3695                                     struct btrfs_extent_item);
3696                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3697                 if (item_size >= sizeof(*ei)) {
3698                         flags = btrfs_extent_flags(path->nodes[0], ei);
3699                         ret = check_extent_flags(flags);
3700                         BUG_ON(ret);
3701
3702                 } else {
3703 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3704                         u64 ref_owner;
3705                         int path_change = 0;
3706
3707                         BUG_ON(item_size !=
3708                                sizeof(struct btrfs_extent_item_v0));
3709                         ret = get_ref_objectid_v0(rc, path, &key, &ref_owner,
3710                                                   &path_change);
3711                         if (ref_owner < BTRFS_FIRST_FREE_OBJECTID)
3712                                 flags = BTRFS_EXTENT_FLAG_TREE_BLOCK;
3713                         else
3714                                 flags = BTRFS_EXTENT_FLAG_DATA;
3715
3716                         if (path_change) {
3717                                 btrfs_release_path(rc->extent_root, path);
3718
3719                                 path->search_commit_root = 1;
3720                                 path->skip_locking = 1;
3721                                 ret = btrfs_search_slot(NULL, rc->extent_root,
3722                                                         &key, path, 0, 0);
3723                                 if (ret < 0) {
3724                                         err = ret;
3725                                         break;
3726                                 }
3727                                 BUG_ON(ret > 0);
3728                         }
3729 #else
3730                         BUG();
3731 #endif
3732                 }
3733
3734                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3735                         ret = add_tree_block(rc, &key, path, &blocks);
3736                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3737                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3738                         ret = add_data_references(rc, &key, path, &blocks);
3739                 } else {
3740                         btrfs_release_path(rc->extent_root, path);
3741                         ret = 0;
3742                 }
3743                 if (ret < 0) {
3744                         err = ret;
3745                         break;
3746                 }
3747
3748                 if (!RB_EMPTY_ROOT(&blocks)) {
3749                         ret = relocate_tree_blocks(trans, rc, &blocks);
3750                         if (ret < 0) {
3751                                 if (ret != -EAGAIN) {
3752                                         err = ret;
3753                                         break;
3754                                 }
3755                                 rc->extents_found--;
3756                                 rc->search_start = key.objectid;
3757                         }
3758                 }
3759
3760                 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3761                                             rc->block_rsv, 0, 5);
3762                 if (ret < 0) {
3763                         if (ret != -EAGAIN) {
3764                                 err = ret;
3765                                 WARN_ON(1);
3766                                 break;
3767                         }
3768                         rc->commit_transaction = 1;
3769                 }
3770
3771                 if (rc->commit_transaction) {
3772                         rc->commit_transaction = 0;
3773                         ret = btrfs_commit_transaction(trans, rc->extent_root);
3774                         BUG_ON(ret);
3775                 } else {
3776                         nr = trans->blocks_used;
3777                         btrfs_end_transaction_throttle(trans, rc->extent_root);
3778                         btrfs_btree_balance_dirty(rc->extent_root, nr);
3779                 }
3780                 trans = NULL;
3781
3782                 if (rc->stage == MOVE_DATA_EXTENTS &&
3783                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3784                         rc->found_file_extent = 1;
3785                         ret = relocate_data_extent(rc->data_inode,
3786                                                    &key, &rc->cluster);
3787                         if (ret < 0) {
3788                                 err = ret;
3789                                 break;
3790                         }
3791                 }
3792         }
3793         if (trans && progress && err == -ENOSPC) {
3794                 ret = btrfs_force_chunk_alloc(trans, rc->extent_root,
3795                                               rc->block_group->flags);
3796                 if (ret == 0) {
3797                         err = 0;
3798                         progress = 0;
3799                         goto restart;
3800                 }
3801         }
3802
3803         btrfs_release_path(rc->extent_root, path);
3804         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3805                           GFP_NOFS);
3806
3807         if (trans) {
3808                 nr = trans->blocks_used;
3809                 btrfs_end_transaction_throttle(trans, rc->extent_root);
3810                 btrfs_btree_balance_dirty(rc->extent_root, nr);
3811         }
3812
3813         if (!err) {
3814                 ret = relocate_file_extent_cluster(rc->data_inode,
3815                                                    &rc->cluster);
3816                 if (ret < 0)
3817                         err = ret;
3818         }
3819
3820         rc->create_reloc_tree = 0;
3821         set_reloc_control(rc);
3822
3823         backref_cache_cleanup(&rc->backref_cache);
3824         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3825
3826         err = prepare_to_merge(rc, err);
3827
3828         merge_reloc_roots(rc);
3829
3830         rc->merge_reloc_tree = 0;
3831         unset_reloc_control(rc);
3832         btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3833
3834         /* get rid of pinned extents */
3835         trans = btrfs_join_transaction(rc->extent_root, 1);
3836         if (IS_ERR(trans))
3837                 err = PTR_ERR(trans);
3838         else
3839                 btrfs_commit_transaction(trans, rc->extent_root);
3840 out_free:
3841         btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3842         btrfs_free_path(path);
3843         return err;
3844 }
3845
3846 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3847                                  struct btrfs_root *root, u64 objectid)
3848 {
3849         struct btrfs_path *path;
3850         struct btrfs_inode_item *item;
3851         struct extent_buffer *leaf;
3852         int ret;
3853
3854         path = btrfs_alloc_path();
3855         if (!path)
3856                 return -ENOMEM;
3857
3858         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3859         if (ret)
3860                 goto out;
3861
3862         leaf = path->nodes[0];
3863         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3864         memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item));
3865         btrfs_set_inode_generation(leaf, item, 1);
3866         btrfs_set_inode_size(leaf, item, 0);
3867         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3868         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3869                                           BTRFS_INODE_PREALLOC);
3870         btrfs_mark_buffer_dirty(leaf);
3871         btrfs_release_path(root, path);
3872 out:
3873         btrfs_free_path(path);
3874         return ret;
3875 }
3876
3877 /*
3878  * helper to create inode for data relocation.
3879  * the inode is in data relocation tree and its link count is 0
3880  */
3881 static noinline_for_stack
3882 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3883                                  struct btrfs_block_group_cache *group)
3884 {
3885         struct inode *inode = NULL;
3886         struct btrfs_trans_handle *trans;
3887         struct btrfs_root *root;
3888         struct btrfs_key key;
3889         unsigned long nr;
3890         u64 objectid = BTRFS_FIRST_FREE_OBJECTID;
3891         int err = 0;
3892
3893         root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID);
3894         if (IS_ERR(root))
3895                 return ERR_CAST(root);
3896
3897         trans = btrfs_start_transaction(root, 6);
3898         if (IS_ERR(trans))
3899                 return ERR_CAST(trans);
3900
3901         err = btrfs_find_free_objectid(root, &objectid);
3902         if (err)
3903                 goto out;
3904
3905         err = __insert_orphan_inode(trans, root, objectid);
3906         BUG_ON(err);
3907
3908         key.objectid = objectid;
3909         key.type = BTRFS_INODE_ITEM_KEY;
3910         key.offset = 0;
3911         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
3912         BUG_ON(IS_ERR(inode) || is_bad_inode(inode));
3913         BTRFS_I(inode)->index_cnt = group->key.objectid;
3914
3915         err = btrfs_orphan_add(trans, inode);
3916 out:
3917         nr = trans->blocks_used;
3918         btrfs_end_transaction(trans, root);
3919         btrfs_btree_balance_dirty(root, nr);
3920         if (err) {
3921                 if (inode)
3922                         iput(inode);
3923                 inode = ERR_PTR(err);
3924         }
3925         return inode;
3926 }
3927
3928 static struct reloc_control *alloc_reloc_control(void)
3929 {
3930         struct reloc_control *rc;
3931
3932         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3933         if (!rc)
3934                 return NULL;
3935
3936         INIT_LIST_HEAD(&rc->reloc_roots);
3937         backref_cache_init(&rc->backref_cache);
3938         mapping_tree_init(&rc->reloc_root_tree);
3939         extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3940         return rc;
3941 }
3942
3943 /*
3944  * function to relocate all extents in a block group.
3945  */
3946 int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3947 {
3948         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3949         struct reloc_control *rc;
3950         struct inode *inode;
3951         struct btrfs_path *path;
3952         int ret;
3953         int rw = 0;
3954         int err = 0;
3955
3956         rc = alloc_reloc_control();
3957         if (!rc)
3958                 return -ENOMEM;
3959
3960         rc->extent_root = extent_root;
3961
3962         rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3963         BUG_ON(!rc->block_group);
3964
3965         if (!rc->block_group->ro) {
3966                 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3967                 if (ret) {
3968                         err = ret;
3969                         goto out;
3970                 }
3971                 rw = 1;
3972         }
3973
3974         path = btrfs_alloc_path();
3975         if (!path) {
3976                 err = -ENOMEM;
3977                 goto out;
3978         }
3979
3980         inode = lookup_free_space_inode(fs_info->tree_root, rc->block_group,
3981                                         path);
3982         btrfs_free_path(path);
3983
3984         if (!IS_ERR(inode))
3985                 ret = delete_block_group_cache(fs_info, inode, 0);
3986         else
3987                 ret = PTR_ERR(inode);
3988
3989         if (ret && ret != -ENOENT) {
3990                 err = ret;
3991                 goto out;
3992         }
3993
3994         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3995         if (IS_ERR(rc->data_inode)) {
3996                 err = PTR_ERR(rc->data_inode);
3997                 rc->data_inode = NULL;
3998                 goto out;
3999         }
4000
4001         printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n",
4002                (unsigned long long)rc->block_group->key.objectid,
4003                (unsigned long long)rc->block_group->flags);
4004
4005         btrfs_start_delalloc_inodes(fs_info->tree_root, 0);
4006         btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
4007
4008         while (1) {
4009                 mutex_lock(&fs_info->cleaner_mutex);
4010
4011                 btrfs_clean_old_snapshots(fs_info->tree_root);
4012                 ret = relocate_block_group(rc);
4013
4014                 mutex_unlock(&fs_info->cleaner_mutex);
4015                 if (ret < 0) {
4016                         err = ret;
4017                         goto out;
4018                 }
4019
4020                 if (rc->extents_found == 0)
4021                         break;
4022
4023                 printk(KERN_INFO "btrfs: found %llu extents\n",
4024                         (unsigned long long)rc->extents_found);
4025
4026                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
4027                         btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1);
4028                         invalidate_mapping_pages(rc->data_inode->i_mapping,
4029                                                  0, -1);
4030                         rc->stage = UPDATE_DATA_PTRS;
4031                 }
4032         }
4033
4034         filemap_write_and_wait_range(fs_info->btree_inode->i_mapping,
4035                                      rc->block_group->key.objectid,
4036                                      rc->block_group->key.objectid +
4037                                      rc->block_group->key.offset - 1);
4038
4039         WARN_ON(rc->block_group->pinned > 0);
4040         WARN_ON(rc->block_group->reserved > 0);
4041         WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
4042 out:
4043         if (err && rw)
4044                 btrfs_set_block_group_rw(extent_root, rc->block_group);
4045         iput(rc->data_inode);
4046         btrfs_put_block_group(rc->block_group);
4047         kfree(rc);
4048         return err;
4049 }
4050
4051 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
4052 {
4053         struct btrfs_trans_handle *trans;
4054         int ret;
4055
4056         trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
4057         BUG_ON(IS_ERR(trans));
4058
4059         memset(&root->root_item.drop_progress, 0,
4060                 sizeof(root->root_item.drop_progress));
4061         root->root_item.drop_level = 0;
4062         btrfs_set_root_refs(&root->root_item, 0);
4063         ret = btrfs_update_root(trans, root->fs_info->tree_root,
4064                                 &root->root_key, &root->root_item);
4065         BUG_ON(ret);
4066
4067         ret = btrfs_end_transaction(trans, root->fs_info->tree_root);
4068         BUG_ON(ret);
4069         return 0;
4070 }
4071
4072 /*
4073  * recover relocation interrupted by system crash.
4074  *
4075  * this function resumes merging reloc trees with corresponding fs trees.
4076  * this is important for keeping the sharing of tree blocks
4077  */
4078 int btrfs_recover_relocation(struct btrfs_root *root)
4079 {
4080         LIST_HEAD(reloc_roots);
4081         struct btrfs_key key;
4082         struct btrfs_root *fs_root;
4083         struct btrfs_root *reloc_root;
4084         struct btrfs_path *path;
4085         struct extent_buffer *leaf;
4086         struct reloc_control *rc = NULL;
4087         struct btrfs_trans_handle *trans;
4088         int ret;
4089         int err = 0;
4090
4091         path = btrfs_alloc_path();
4092         if (!path)
4093                 return -ENOMEM;
4094
4095         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
4096         key.type = BTRFS_ROOT_ITEM_KEY;
4097         key.offset = (u64)-1;
4098
4099         while (1) {
4100                 ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key,
4101                                         path, 0, 0);
4102                 if (ret < 0) {
4103                         err = ret;
4104                         goto out;
4105                 }
4106                 if (ret > 0) {
4107                         if (path->slots[0] == 0)
4108                                 break;
4109                         path->slots[0]--;
4110                 }
4111                 leaf = path->nodes[0];
4112                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4113                 btrfs_release_path(root->fs_info->tree_root, path);
4114
4115                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
4116                     key.type != BTRFS_ROOT_ITEM_KEY)
4117                         break;
4118
4119                 reloc_root = btrfs_read_fs_root_no_radix(root, &key);
4120                 if (IS_ERR(reloc_root)) {
4121                         err = PTR_ERR(reloc_root);
4122                         goto out;
4123                 }
4124
4125                 list_add(&reloc_root->root_list, &reloc_roots);
4126
4127                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
4128                         fs_root = read_fs_root(root->fs_info,
4129                                                reloc_root->root_key.offset);
4130                         if (IS_ERR(fs_root)) {
4131                                 ret = PTR_ERR(fs_root);
4132                                 if (ret != -ENOENT) {
4133                                         err = ret;
4134                                         goto out;
4135                                 }
4136                                 mark_garbage_root(reloc_root);
4137                         }
4138                 }
4139
4140                 if (key.offset == 0)
4141                         break;
4142
4143                 key.offset--;
4144         }
4145         btrfs_release_path(root->fs_info->tree_root, path);
4146
4147         if (list_empty(&reloc_roots))
4148                 goto out;
4149
4150         rc = alloc_reloc_control();
4151         if (!rc) {
4152                 err = -ENOMEM;
4153                 goto out;
4154         }
4155
4156         rc->extent_root = root->fs_info->extent_root;
4157
4158         set_reloc_control(rc);
4159
4160         trans = btrfs_join_transaction(rc->extent_root, 1);
4161         if (IS_ERR(trans)) {
4162                 unset_reloc_control(rc);
4163                 err = PTR_ERR(trans);
4164                 goto out_free;
4165         }
4166
4167         rc->merge_reloc_tree = 1;
4168
4169         while (!list_empty(&reloc_roots)) {
4170                 reloc_root = list_entry(reloc_roots.next,
4171                                         struct btrfs_root, root_list);
4172                 list_del(&reloc_root->root_list);
4173
4174                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
4175                         list_add_tail(&reloc_root->root_list,
4176                                       &rc->reloc_roots);
4177                         continue;
4178                 }
4179
4180                 fs_root = read_fs_root(root->fs_info,
4181                                        reloc_root->root_key.offset);
4182                 BUG_ON(IS_ERR(fs_root));
4183
4184                 __add_reloc_root(reloc_root);
4185                 fs_root->reloc_root = reloc_root;
4186         }
4187
4188         btrfs_commit_transaction(trans, rc->extent_root);
4189
4190         merge_reloc_roots(rc);
4191
4192         unset_reloc_control(rc);
4193
4194         trans = btrfs_join_transaction(rc->extent_root, 1);
4195         if (IS_ERR(trans))
4196                 err = PTR_ERR(trans);
4197         else
4198                 btrfs_commit_transaction(trans, rc->extent_root);
4199 out_free:
4200         kfree(rc);
4201 out:
4202         while (!list_empty(&reloc_roots)) {
4203                 reloc_root = list_entry(reloc_roots.next,
4204                                         struct btrfs_root, root_list);
4205                 list_del(&reloc_root->root_list);
4206                 free_extent_buffer(reloc_root->node);
4207                 free_extent_buffer(reloc_root->commit_root);
4208                 kfree(reloc_root);
4209         }
4210         btrfs_free_path(path);
4211
4212         if (err == 0) {
4213                 /* cleanup orphan inode in data relocation tree */
4214                 fs_root = read_fs_root(root->fs_info,
4215                                        BTRFS_DATA_RELOC_TREE_OBJECTID);
4216                 if (IS_ERR(fs_root))
4217                         err = PTR_ERR(fs_root);
4218                 else
4219                         err = btrfs_orphan_cleanup(fs_root);
4220         }
4221         return err;
4222 }
4223
4224 /*
4225  * helper to add ordered checksum for data relocation.
4226  *
4227  * cloning checksum properly handles the nodatasum extents.
4228  * it also saves CPU time to re-calculate the checksum.
4229  */
4230 int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
4231 {
4232         struct btrfs_ordered_sum *sums;
4233         struct btrfs_sector_sum *sector_sum;
4234         struct btrfs_ordered_extent *ordered;
4235         struct btrfs_root *root = BTRFS_I(inode)->root;
4236         size_t offset;
4237         int ret;
4238         u64 disk_bytenr;
4239         LIST_HEAD(list);
4240
4241         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
4242         BUG_ON(ordered->file_offset != file_pos || ordered->len != len);
4243
4244         disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt;
4245         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr,
4246                                        disk_bytenr + len - 1, &list);
4247
4248         while (!list_empty(&list)) {
4249                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
4250                 list_del_init(&sums->list);
4251
4252                 sector_sum = sums->sums;
4253                 sums->bytenr = ordered->start;
4254
4255                 offset = 0;
4256                 while (offset < sums->len) {
4257                         sector_sum->bytenr += ordered->start - disk_bytenr;
4258                         sector_sum++;
4259                         offset += root->sectorsize;
4260                 }
4261
4262                 btrfs_add_ordered_sum(inode, ordered, sums);
4263         }
4264         btrfs_put_ordered_extent(ordered);
4265         return ret;
4266 }
4267
4268 void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4269                            struct btrfs_root *root, struct extent_buffer *buf,
4270                            struct extent_buffer *cow)
4271 {
4272         struct reloc_control *rc;
4273         struct backref_node *node;
4274         int first_cow = 0;
4275         int level;
4276         int ret;
4277
4278         rc = root->fs_info->reloc_ctl;
4279         if (!rc)
4280                 return;
4281
4282         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4283                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4284
4285         level = btrfs_header_level(buf);
4286         if (btrfs_header_generation(buf) <=
4287             btrfs_root_last_snapshot(&root->root_item))
4288                 first_cow = 1;
4289
4290         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4291             rc->create_reloc_tree) {
4292                 WARN_ON(!first_cow && level == 0);
4293
4294                 node = rc->backref_cache.path[level];
4295                 BUG_ON(node->bytenr != buf->start &&
4296                        node->new_bytenr != buf->start);
4297
4298                 drop_node_buffer(node);
4299                 extent_buffer_get(cow);
4300                 node->eb = cow;
4301                 node->new_bytenr = cow->start;
4302
4303                 if (!node->pending) {
4304                         list_move_tail(&node->list,
4305                                        &rc->backref_cache.pending[level]);
4306                         node->pending = 1;
4307                 }
4308
4309                 if (first_cow)
4310                         __mark_block_processed(rc, node);
4311
4312                 if (first_cow && level > 0)
4313                         rc->nodes_relocated += buf->len;
4314         }
4315
4316         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4317                 ret = replace_file_extents(trans, rc, root, cow);
4318                 BUG_ON(ret);
4319         }
4320 }
4321
4322 /*
4323  * called before creating snapshot. it calculates metadata reservation
4324  * requried for relocating tree blocks in the snapshot
4325  */
4326 void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4327                               struct btrfs_pending_snapshot *pending,
4328                               u64 *bytes_to_reserve)
4329 {
4330         struct btrfs_root *root;
4331         struct reloc_control *rc;
4332
4333         root = pending->root;
4334         if (!root->reloc_root)
4335                 return;
4336
4337         rc = root->fs_info->reloc_ctl;
4338         if (!rc->merge_reloc_tree)
4339                 return;
4340
4341         root = root->reloc_root;
4342         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4343         /*
4344          * relocation is in the stage of merging trees. the space
4345          * used by merging a reloc tree is twice the size of
4346          * relocated tree nodes in the worst case. half for cowing
4347          * the reloc tree, half for cowing the fs tree. the space
4348          * used by cowing the reloc tree will be freed after the
4349          * tree is dropped. if we create snapshot, cowing the fs
4350          * tree may use more space than it frees. so we need
4351          * reserve extra space.
4352          */
4353         *bytes_to_reserve += rc->nodes_relocated;
4354 }
4355
4356 /*
4357  * called after snapshot is created. migrate block reservation
4358  * and create reloc root for the newly created snapshot
4359  */
4360 void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4361                                struct btrfs_pending_snapshot *pending)
4362 {
4363         struct btrfs_root *root = pending->root;
4364         struct btrfs_root *reloc_root;
4365         struct btrfs_root *new_root;
4366         struct reloc_control *rc;
4367         int ret;
4368
4369         if (!root->reloc_root)
4370                 return;
4371
4372         rc = root->fs_info->reloc_ctl;
4373         rc->merging_rsv_size += rc->nodes_relocated;
4374
4375         if (rc->merge_reloc_tree) {
4376                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4377                                               rc->block_rsv,
4378                                               rc->nodes_relocated);
4379                 BUG_ON(ret);
4380         }
4381
4382         new_root = pending->snap;
4383         reloc_root = create_reloc_root(trans, root->reloc_root,
4384                                        new_root->root_key.objectid);
4385
4386         __add_reloc_root(reloc_root);
4387         new_root->reloc_root = reloc_root;
4388
4389         if (rc->create_reloc_tree) {
4390                 ret = clone_backref_node(trans, rc, root, reloc_root);
4391                 BUG_ON(ret);
4392         }
4393 }